The Endoscopic Classication of Representations: Orthogonal and Symplectic Groups James Arthur

Department of Mathematics, University of Toronto, 40 St. George Street, Toronto, Ontario M5S 2E4 E-mail address: arthur@math.toronto.edu

2000 Mathematics Subject Classication. Primary

Supported in part by NSERC Discovery Grant A3483.

Abstract.

Contents
Preface Chapter 1. Parameters 1.1. The automorphic Langlands group 1.2. Self-dual, nite dimensional representations 1.3. Representations of GLpN q 1.4. A substitute for global parameters 1.5. Statement of three main theorems Chapter 2. Local Transfer 2.1. Langlands-Shelstad-Kottwitz transfer 2.2. Characterization of the local classication 2.3. Normalized intertwining operators 2.4. Statement of the local intertwining relation 2.5. Relations with Whittaker models Chapter 3. Global Stabilization 3.1. The discrete part of the trace formula 3.2. Stabilization 3.3. Contribution of a parameter 3.4. A preliminary comparison 3.5. On the vanishing of coecients Chapter 4. The Standard Model 4.1. Statement of the stable multiplicity formula 4.2. On the global intertwining relation 4.3. Spectral terms 4.4. Endoscopic terms 4.5. The comparison 4.6. The two sign lemmas 4.7. On the global theorems 4.8. Remarks on general groups Chapter 5. A Study of Critical Cases 5.1. On the case of square integrable 5.2. On the case of elliptic 5.3. A supplementary parameter ` 5.4. Generic parameters with local constraints
v

vii 1 1 6 15 26 40 51 51 62 79 96 108 121 121 130 137 145 152 165 165 171 184 191 199 212 226 237 251 251 262 270 285

vi

CONTENTS

Chapter 6. The Local Classication 6.1. Local parameters 6.2. Construction of global representations 9 9 6.3. Construction of global parameters 6.4. The local intertwining relation for 6.5. Orthogonality relations for 6.6. Local packets for composite 6.7. Local packets for simple 6.8. Resolution Chapter 7. Local Nontempered Representations 7.1. Local parameters and duality 9 7.2. Construction of global parameters 7.3. The local intertwining relation for 7.4. Local packets for composite and simple 7.5. Some remarks on characters Chapter 8. The Global Classication 8.1. On the global theorems 8.2. Proof by contradiction 8.3. Reections on the results 8.4. Renements for even orthogonal groups 8.5. An approximation of the Langlands group Chapter 9. Inner Forms 9.1. Inner twists and centralizers 9.2. Statement of theorems for inner forms 9.3. Local correspondence for inner forms 9.4. Nontempered packets for inner forms 9.5. Completion of the proof Bibliography

299 299 307 319 330 343 353 365 374 385 385 394 401 413 427 441 441 451 467 482 500 511 511 511 511 511 511 513

Preface
This preface contains a brief overview of the volume. We start with a rough summary of the main theorems. We then give short descriptions of the contents of the various chapters. At the end of the preface, we will add a couple of remarks on the overall structure of the proof, notably our use of induction. The preface can serve as an introduction. The beginning of the actual text, in the form of the rst two or three sections of Chapter 1, represents a dierent sort of introduction. It will be our attempt to motivate what follows from a few basic principles. Automorphic representations for GLpN q have been important objects of study for many years. We recall that GLpN q, the general linear group of invertible pN N q-matrices, assigns a group GLpN, Rq to any commutative ring R with identity. For example, R could be a number eld F , or the ring A AF of ad`les over F . Automorphic representations of GLpN q are e the irreducible representations of GLpN, Aq that occur in the decomposition ` of its regular representation on L2 GLpN, F qzGLpN, Aq . This informal denition is made precise in [L6], and carries over to any connected reductive group G over F . The primary aim of the volume is to classify the automorphic representations of special orthogonal and symplectic groups G in terms of those of GLpN q. Our main tool will be the stable trace formula for G, which until recently was conditional on the fundamental lemma. The fundamental lemma has now been established in complete generality, and in all of its various forms. In particular, the stabilization of the trace formula is now known for any connected group. However, we will also require the stabilization of twisted trace formulas for GLpN q and SOp2nq. Since these have yet to be established, our results will still be conditional. A secondary purpose will be to lay foundations for the endoscopic study of more general groups G. It is reasonable to believe that the methods we introduce here extend to groups that Ramakrishnan has called quasiclassical. These would comprise the largest class of groups whose representations could ultimately be tied to those of general linear groups. Our third goal is expository. In adopting a style that is sometimes more discursive than strictly necessary, we have tried to place at least some of the techniques into perspective. We hope that there will be parts of the volume that are accessible to readers who are not experts in the subject.

vii

viii

PREFACE

Automorphic representations are interesting for many reasons, but among the most fundamental is the arithmetic data they carry. Recall that GLpN, Aq
v

GLpN, Fv q

is a restricted direct product, taken over (equivalence classes of) valuations v of F . An automorphic representation of GLpN q is a restricted direct product v ,
v

where v is an irreducible representation of GLpN, Fv q that is unramied for almost all v. We recall that v is unramied if v is nonarchimedean, and v contains the trivial representation of the hyperspecial maximal compact subgroup GLpN, ov q of integral points in GLpN, Fv q. The representation is then parametrized by a semisimple conjugacy class cv pq cpv q in the complex dual group GLpN q^ GLpN, Cq of GLpN q. (See [Bo, (6.4), (6.5)] for the precise assertion, as it applies to a general connected reductive group G.) It is the relations among the semisimple conjugacy classes cv pq that will contain the fundamental arithmetic information. There are three basic theorems for the group GLpN q that together give us a pretty clear understanding of its representations. The rst is local, while others, which actually predate the rst, are global. The rst theorem is the local Langlands correspondence for GLpN q. It was established for archimedean elds by Langlands, and more recently for padic (which is to say nonarchimedean) elds by Harris, Taylor and Henniart. It classies the irreducible representations of GLpN, Fv q at all places v by (equivalence classes of) semisimple, N -dimensional representations of the local Langlands group # WF v , v archimedean, LFv WFv SU p2q, otherwise. In particular, an unramied representation of GLpN, Fv q corresponds to an N -dimensional representation of LFv that is trivial on the product of SU p2q with the inertia subgroup IFv of the local Weil group WFv . It therefore corresponds to a semisimple representation of the cyclic quotient LFv {IFv SU p2q WFv {IFv Z, and hence a semisimple conjugacy class in GLpN, Cq, as above.

PREFACE

ix

The rst of the global theorems is due to Jacquet and Shalika. If is any smooth representation of GLpN, Aq, one can form the family of semisimple conjugacy classes ( cpq lim cv pq cpv q : v R S
S

in GLpN, Cq, dened up to a nite set of valuations S. In other words, cpq is an equivalence class of families, two such families being equivalent if they are equal for almost all v. The theorem of Jacquet and Shalika asserts that if an automorphic representation of GLpN q is restricted slightly to be isobaric [L7, 2], it is uniquely determined by cpq. This theorem can be regarded as a generalization of the theorem of strong multiplicity one for cuspidal automorphic representations of GLpN q. The other global theorem for GLpN q is due to Moeglin and Waldspurger. It characterizes the automorphic (relatively) discrete spectrum of GLpN q in terms of the set of cuspidal automorphic representations. Since Langlands general theory of Eisenstein series characterizes the full automorphic spectrum of any group G in terms of discrete spectra, this theorem characterizes the automorphic spectrum for GLpN q in terms of cuspidal automorphic representations. Combined with the rst global theorem, it classies the full automorphic spectrum of GLpN q explicitly in terms of families cpq, constructed from cuspidal automorphic representations of general linear groups. Our goal is to generalize these three theorems. As we shall see, however, there is very little that comes easily. It has been known for many years that the representations of groups other than GLpN q have more structure. In particular, they should separate naturally into L-packets, composed of representations with the same L-functions and -factors. This was demonstrated for the group G SLp2q Spp2q by Labesse and Langlands, in a paper [LL] that became a model for Langlandss conjectural theory of endoscopy [L8], [L10]. The simplest and most elegant way to formulate the theory of endoscopy is in terms of the global Langlands group LF . This is a hypothetical global analogue of the explicit local Langlands groups LFv dened above. It is thought to be a locally compact extension 1 KF LF WF 1 of WF by a compact connected group KF . (See [L7, 2], [K1, 9].) However, its existence is very deep, and could well turn out to be the nal theorem in the subject to be proved! One of our rst tasks, which we address in 1.4, will be to introduce makeshift objects to be used in place of LF . For simplicity, however, let us describe our results here in terms of LF . Our main results apply to the case that G is a quasisplit special orthogonal or symplectic group. They are stated as three theorems in 1.5. The proof of these theorems will then take up much of the rest of the volume.

PREFACE

Theorem 1.5.1 is the main local result. It contains a local Langlands parametrization of the irreducible representations of GpFv q, for any p-adic valuation v of F , as a disjoint union of nite L-packets v . These are indexed by local Langlands parameters, namely L-homomorphisms v : LFv
L

Gv

p from LFv to the local L-group L Gv G GalpF v {Fv q of G. The theorem includes a way to index the representations in an L-packet v by linear characters on a nite abelian group Sv attached to v . Since similar results for archimedean valuations v are already known from the work of Shelstad, we obtain a classication of the representations of each local group GpFv q. Theorem 1.5.1 also contains a somewhat less precise description of the representations of GpFv q that are local components of automorphic representations. These fall naturally into rather dierent packets v , indexed according to the conjectures of [A8] by L-homomorphisms (1) v : LFv SU p2q
L

Gv ,

with bounded image. The theorem includes the assertion, also conjectured in [A8], that the representations in these packets are all unitary. Theorem 1.5.2 is the main global result. As a rst approximation, it gives a rough decomposition (2) Rdisc Rdisc,

of the representation Rdisc of GpAq on the automorphic discrete spectrum ` L2 GpF qzGpAq . The indices can be thought of as L-homomorphisms disc (3) : LF SU p2q
L

of bounded image that do not factor through any proper parabolic subgroup p of the global L-group L G G GalpF {F q. They have localizations (1) dened by conjugacy classes of embeddings LFv LF , or rather the makeshift analogues of such embeddings that we formulate in 1.4. The localizations v of are unramied at almost v, and consequently lead to a family ( cpq lim cv pq cpv q : v R S
S

of equivalence classes of semisimple elements in L G. The rough decomposition (2) is of interest as it stands. It implies that G has no embedded eigenvalues, in the sense of unramied Hecke operators. In other words, the family cpq attached to any global parameter in the decomposition of the automorphic discrete spectrum is distinct from any family obtained from the continuous spectrum. This follows from the nature of the parameters in (3), and the application of the theorem of Jacquet-Shalika to the natural image of cpq in the appropriate complex general linear group.

PREFACE

xi

Theorem 1.5.2 also contains a ner decomposition (4) Rdisc, m pq ,

for any global parameter . The indices range over representations in the global packet of , dened as a restricted direct product of local packets provided by Theorem 1.5.1. The multiplicities m pq are given by an explicit reciprocity formula in terms of the nite abelian groups Sv , and their global analogue S . We thus obtain a decomposition of the automorphic discrete spectrum of G into irreducible representations of GpAq. We shall say that a parameter is generic if it is trivial on the factor SU p2q. The representations P , with generic and m pq 0, are the constituents of the automorphic discrete spectrum that are expected to satisfy the analogue for G of Ramanujans conjecture. It is not generic, the formula for m pq has an extra ingredient. It is a sign character on S , dened (1.5.6) in terms of symplectic root numbers. That the discrete spectrum should be governed by objects of such immediate arithmetic signicance seems quite striking. Theorem 1.5.2 has application to the question of multiplicity one. Suppose that is an irreducible constituent of the automorphic discrete spectrum of G that also lies in some generic global packet . We shall then show that the multiplicity of in the discrete spectrum equals 1 unless p G SOp2n, Cq, in which case the multiplicity is either 1 or 2, according to an explicit condition we shall give. In particular, if G equals either SOp2n ` 1q or Spp2nq, the automorphic representations in the discrete spectrum that are expected to satisfy Ramanujans conjecture all have multiplicity 1. Local results of Moeglin [M4] on nontempered p-adic packets suggest that similar results hold for all automorphic representations in the discrete spectrum. Theorems 1.5.1 and 1.5.2 are founded on the proof of several cases of Langlands principle of functoriality. In fact, our basic denitions will be derived from the functorial correspondence from G to GLpN q, relative to the standard representation of L G into GLpN, Cq. Otherwise said, our construction of representations of G will be formulated in terms of representations of GLpN q. The integer N of course equals 2n, 2n ` 1 and 2n as G ranges over the groups SOp2n ` 1q, Spp2nq and SOp2nq in the three innite families Bn , p Cn and Dn , with dual groups G being equal to Spp2n, Cq, SOp2n`1, Cq and SOp2n, Cq, respectively. The third case SOp2nq, which includes quasisplit outer twists, is complicated by the fact that it is really the nonconnected group Op2nq that is directly tied to GLpN q. This is what is responsible for the failure of multiplicity one described above. It is also the reason we have not yet specied the equivalence relation for the local and global parameters p (1) and (3). Let us now agree that they are to be taken up to G-conjugacy if G equals SOp2n ` 1q or Spp2nq, and up to conjugacy by Op2n, Cq, a group whose quotient r OutN pGq Op2n, Cq{SOp2n, Cq

xii

PREFACE

p acts by outer automorphism on G SOp2n, Cq, if G equals SOp2nq. This is the understanding on which the decomposition (2) holds. r In the text, we shall write pGv q for the set of equivalence classes of r r local parameters (1). The packet v attached to any v P pGv q will then r be composed of OutN pGv q-orbits of equivalence classes of irreducible repr resentations (with OutN pGq being trivial in case G equals SOp2n ` 1q or r Spp2nq). We will write pGq for the set of equivalence classes of general r global parameters , and 2 pGq for the subset of classes with the suppler mentary condition of (3). The packet of any P pGq will then be restricted direct product r r
v

of local packets. These are the objects that correspond to (isobaric) aur tomorphic representations of GLpN q. In particular, the global packets , r , are the objects rather than the individual (orbits of) representations in that retain the property of strong multiplicity one from GLpN q. Similarly, r r the global packets attached to parameters P 2 pGq retain the qualitative properties of automorphic discrete spectrum of GLpN q. They come with a sort of Jordan decomposition, in which the semisimple packets correspond to the generic global parameters , and contain the automorphic representations that are expected to satisfy the G-analogue of Ramanujans conjecture. In view of these comments, we see that Theorem 1.5.2 can be regarded as a simultaneous analogue for G of both of the global theorems for GLpN q. Theorem 1.5.3 is a global supplement to Theorem 1.5.2. Its rst assertion r applies to global parameters P pGq that are both generic and simple, in the sense that they correspond to cuspidal automorphic representations of p GLpN q. Theorem 1.5.3(a) asserts that the dual group G is orthogonal (resp. symplectic) if and only if the symmetric square L-function Lps, , S 2 q (resp. the skew-symmetric square L-function Lps, , 2 q) has a pole at s 1. `1 Theorem 1.5.3(b) asserts that the Rankin-Selberg -factor 2 , 1 2 r p equals 1 for any pair of generic simple parameters i P pGi q such that G1 p and G2 are either both orthogonal or both symplectic. These two assertions are automorphic analogues of well known properties of Artin L-functions and -factors. They are interesting in their own right. But they are also an essential part of our induction argument. We will need them in Chapter 4 to interpret the terms in the trace formula attached to compound parameters r P pGq. This completes our summary of the main theorems. The rst two sections of Chapter 1 contain further motivation, for the global Langlands group LF in 1.1, and the relations between representations of G and GLpN q in 1.2. In 1.3, we will recall the three basic theorems for GLpN q. Section 1.4

PREFACE

xiii

is given over to our makeshift substitutes for global Langlands parameters, as we have said, while 1.5 contains the formal statements of the theorems. As might be expected, the three theorems will have to be established together. The unied proof will take us down a long road, which starts in Chapter 2, and crosses many dierent landscapes before coming to an end nally in 8.2. The argument is ultimately founded on harmonic analysis, which is represented locally by orbital integrals and characters, and globally by the trace formula. This of course is at the heart of the theory of endoscopy. We refer the reader to the introductory remarks of individual sections, where we have tried to oer guidance and motivation. We shall be content here with a minimal outline of the main stages. Chapter 2 is devoted to local endoscopy. It contains a more precise formulation (Theorem 2.2.1) of the local Theorem 1.5.1. This provides for a r canonical construction of the local packets v in terms of twisted characters on GLpN q. Chapter 2 also includes the statement of Theorem 2.4.1, which we will call the local intertwining relation. This is closely related to Theorem 1.5.1 and its renement Theorem 2.2.1, and from a technical standpoint, can be regarded as our primary local result. It includes a delicate construction of signs, which will be critical for the interpretation of terms in the trace formula. Chapter 3 is devoted to global endoscopy. We will recall the discrete part of the trace formula in 3.1, and its stabilization in 3.2. We are speaking here of those spectral terms that are linear combinations of automorphic characters, and to which all of the other terms are ultimately dedicated. They are the only terms in the trace formula that will appear explicitly in this volume. In 3.5, we shall establish criteria for the vanishing of coecients in certain identities (Proposition 3.5.1, Corollary 3.5.3). We will use these criteria many times throughout the volume in drawing conclusions from the comparison of discrete spectral terms. In general, we will have to treat three separate cases of endoscopy. They are represented respectively by pairs pG, G1 q, where G is one of the groups to which Theorems 1.5.1,`1.5.2 and 1.5.3 apply and G1 is a corresponding r r endoscopic datum, pairs GpN q, G in which GpN q is the twisted general r linear group G0 pN q GLpN q and G is a corresponding twisted endoscopic r r r datum, and pairs pG, G1 q in which G is a twisted even orthogonal group 0 SOp2nq and G1 is again a corresponding twisted endoscopic datum. r r G r r The rst two cases will be our main concern. However, the third case pG, G1 q is also a necessary part of the story. Among other things, it is forced on us by the need to specify the signs in the local intertwining relation. For the most part, we will not try to treat the three cases uniformly as cases of the general theory of endoscopy. This might have been dicult, given that we have to deduce many local and global results along the way. At any rate, the separate treatment of the three cases gives our exposition a more concrete avour, if at the expense of some possible sacrice of eciency.

xiv

PREFACE

In Chapter 4, we shall study the comparison of trace formulas. Specifically, we will compare the contribution (4.1.1) of a parameter to the discrete part of the trace formula with the contribution (4.1.2) of to the corresponding endoscopic decomposition. We begin with the statement of Theorem 4.1.2, which we will call the stable multiplicity formula. This is closely related to Theorem 1.5.2, and from a technical standpoint again, is our primary global result. Together with the global intertwining relation (Corollary 4.2.1), which we state as a global corollary of Theorem 2.4.1, it governs how individual terms in trace formulas are related. Chapter 4 represents a standard model, in the sense that if we grant the analogues of the two primary theorems for general groups, it explains how the terms on the right hand sides of (4.1.1) and (4.1.2) match. This is discussed heuristically in Sections 4.7 and 4.8. However, the purpose of this volume is to derive Theorems 2.4.1 and 4.1.2 for our groups G from the standard model, and whatever else we can bring to bear on the problem. This is the perspective of Sections 4.3 and 4.4. In 4.5, we combine the analysis of these sections with a general induction hypothesis to deduce the stable multiplicity formula and the global intertwining relation for many . Section 4.6 contains the proof of two critical sign lemmas that are essential ingredients of the parallel Sections 4.3 and 4.4. Chapter 5 is the center of the volume. It is a bridge between the global discussion of Chapters 3 and 4 and the local discussion of Chapters 6 and 7. It also represents a transition from the general comparisons of Chapter 4 to the study of the remaining parameters needed to complete the induction hypotheses. These exceptional cases are the crux of the matter. In 5.2 and 5.3, we shall extract several identities from the standard model, in which we display the possible failure of Theorems 2.4.1 and 4.1.2 as correction r terms. Section 5.3 applies to the critical case of a parameter P 2 pGq, and calls for the introduction of a supplementary parameter ` . In 5.4, we shall resolve the global problems for families of parameters that are assumed to have certain rather technical local properties. Chapter 6 applies to generic local parameters. It contains a proof of the local Langlands classication for our groups G (modied by the outer automorphism in the case G SOp2nq). We will rst have to embed a given local parameter into a family of global parameters with the local constraints of 5.4. This will be the object of Sections 6.2 and 6.3, which rest ultimately on the simple form of the invariant trace formula for G. We will then have to extract the required local information from the global results obtained in 5.4. In 6.4, we will deduce the generic local intertwining relation from its global counterpart in 5.4. Then in 6.5, we will stabilize the orthogonality relations that are known to hold for elliptic tempered characters. This will allow us to quantify the contributions from the remaining elliptic tempered characters, the ones attached to square integrable representations. We will use the information so obtained in 6.6 and 6.7. In these sections, we shall establish Theorems 2.2.1 and 1.5.1 for the remaining square integrable

PREFACE

xv

r Langlands parameters P 2 pGq. Finally, in 6.8, we shall resolve the various hypotheses taken on at the beginning of Chapter 6. Chapter 7 applies to nongeneric local parameters. It contains the proof of the local theorems in general. In 7.2, we shall use the construction of 6.2 to embed a given nongeneric local parameter into a family of global parameters, but with local constraints that dier slightly from those of 5.4. We will then deduce special cases of the local theorems that apply to the places v with local constraints. These will follow from the local theorems for generic parameters, established in Chapter 6, and the duality operator of Aubert and Schneider-Stuhler, which we review in 7.1. We will then exploit our control over the places v to derive the local theorems at the localization 9 u that represents the original given parameter. We will nish the proof of the global theorems in the rst two sections of Chapter 8. Armed with the local theorems, and the resulting renements of the lemmas from Chapter 5, we will be able to establish almost all of the global results in 8.1. However, there will still to be one nal obstacle. It is r the case of a simple parameter P sim pGq, which among other things, will be essential for a resolution of our induction hypotheses. An examination of this case leads us to the initial impression that it will be resistant to all of our earlier techniques. However, we will then see that there is a way to treat it. We will introduce a second supplementary parameter, which appears ungainly at rst, but which, with the support of two rather intricate lemmas, takes us to a successful conclusion. Section 8.2 is the climax of our long running induction argument, as well as its most dicult point of application. Its nal resolution is what brings us to the end of the proof. We will then be free in 8.3 for some general reections that will give us some perspective on what has been established. In 8.4, we will sharpen our results for the groups SOp2nq, in which the outer automorphism creates some ambiguity. We will use the stabilized trace formula to construct the local and global L-packets predicted for these groups by the conjectural theory of endoscopy. In 8.5, we will describe an approximation L of the global Langlands group LF that is tailored to the F classical groups of this volume. It could potentially be used in place of the ad hoc global parameters of 1.4 to streamline the statements of the global theorems. We shall treat inner forms of orthogonal and symplectic groups in Chapter 9. The automorphic representation theory for inner twists is considerably easier for knowing what happens in the case of quasisplit groups. In particular, the stable multiplicity formula is already in place, since it applies only to the quasisplit case. What work remains is primarily local. We shall describe how to modify our earlier arguments in order to classify local and global representations of arbitrary orthogonal and symplectic groups. Having briey summarized the various chapters, we had best add some comment on our use of induction. As we have noted, induction is a central

xvi

PREFACE

part of the unied argument that will carry us from Chapter 2 to Section 8.2. We will have two kinds of hypotheses, both based on the positive integer N that indexes the underlying general linear group GLpN q. The rst kind includes various ad hoc assumptions, such as those implicit in some of our r denitions. For example, the global parameter sets pGq dened in 1.4 are based on the inductive application of two seed Theorems 1.4.1 and 1.4.2. The second will be the formal induction hypotheses introduced explicitly at the beginning of 4.3, and in more rened form at the beginning of 5.1. They assert essentially that the stated theorems are all valid for parameters of rank less than N . In particular, they include the informal hypotheses implicit in the denitions. We do not actually have to regard the earlier, informal assumptions as inductive. They really represent implicit appeals to stated theorems, in support of proofs of what amount to corollaries. In fact, from a logical standpoint, it is simpler to treat them as inductive assumptions only after we introduce the formal induction hypotheses in 4.3 and 5.1. For a little more discussion of this point, the reader can consult the two parallel Remarks following Corollaries 4.1.3 and 4.2.4. The induction hypotheses of 5.1 are formulated for an abstract family r of global parameters. They pertain to the parameters P F of rank r F less than N , and are supplemented also by a hypothesis (Assumption 5.1.1) r for certain parameters in F of rank equal to N . The results of Chapter r 5 will be applied three times, to three separate families F. These are the family of generic parameters with local constraints used to establish the local classication of Chapter 6, the family of nongeneric parameters with local constraints used to deduce the local results for non-tempered representations in Chapter 7, and the family of all global parameters used to establish the global theorems in the rst two sections of Chapter 8. In each of these cases, r the assumptions have to be resolved for the given family F. In the case of Chapter 6, the induction hypotheses are actually imposed in two stages. The local hypothesis at the beginning of 6.3 is needed to construct the family r F, on which we then impose the global part of the general hypothesis of Chapter 5 at the beginning of 6.4. The earlier induction hypotheses of 4.3 apply to general global parameters of rank less than N . They are used in 4.34.6 to deduce the global theorems for parameters that are highly reducible. Their general resolution comes only after the proof of the global theorems in 8.2. Our induction assumptions have of course to be distinguished from the general condition (Hypothesis 3.2.1) on which our results rely. This is the stabilization of the twisted trace formula for the two groups GLpN q and SOp2nq. As a part of the condition, we implicitly include twisted analogues of the two local results that have a role in the stabilization of the ordinary trace formula. These are the orthogonality relations for elliptic

PREFACE

xvii

tempered characters of [A10, Theorem 6.1], and the weak spectral transfer of tempered p-adic characters given by [A11, Theorems 6.1 and 6.2]. The stabilization of orthogonality relations in 6.5, which requires twisted orthogonality relations for GLpN q, will be an essential part of the local classication in Chapter 6. The two theorems from [A11] can be regarded as a partial generalization of the fundamental lemma for the full spherical Hecke algebra [Hal]. (Their global proof of course depends on the basic fundamental lemma for the unit, established by Ngo.) We will use them in combination with their twisted analogues in the proofs of Proposition 2.1.1 and Corollary 6.7.4. The rst of these gives the image of the twisted transfer of functions from GLpN q, which is needed in the proof of Proposition 3.5.1. The second gives a relation among tempered characters, which completes the local classication. There is one other local theorem whose twisted analogues for GLpN q and SOp2nq we shall also have to take for granted. It is Shelstads strong spectral transfer of tempered archimedean characters, which is to say, her endoscopic classication of representations of real groups. This of course is major result. Together with its two twisted analogues, it gives the archimedean cases of the local classication in Theorems 2.2.1 and 2.2.4. We shall combine it with a global argument in Chapter 6 to establish the p-adic form of these theorems. The general twisted form of Shelstads endoscopic classication appears to be within reach. It is likely to be established soon by some extension of recent work by Mezo [Me] and Shelstad [S7]. Finally, let me include a word on the notation. Because our main theorems require interlocking proofs, which consume a good part of the volume, there is always the risk of losing ones way. Until the end of 8.2, assertions as Theorems are generally stated with the understanding that their proofs will usually be taken up much later (unless of course they are simply quoted from some other source). On the other hand, assertions denoted Propositions, Lemmas or Corollaries represent results along the route, for which the reader can expect a timely proof. The actual mathematical notation might appear unconventional at times. I have tried to structure it so as to reect implicit symmetries in the various objects it represents. With luck, it might help a reader navigate the arguments without necessarily being aware of such symmetries. The three main theorems of the volume were described in [A18, 30]. I gave lecture courses on them in 19941995 at the Institute for Advanced Study and the University of Paris VII, and later in 2000, again at the Institute for Advanced Study. Parts of Chapter 4 were also treated heuristically in the earlier article [A9]. In writing this volume, I have added some topics to my original notes. These include the local Langlands classication for GLpN q, the treatment of inner twists in Chapter 9 and the remarks on Whittaker models in 8.3. I have also had to ll unforeseen gaps in the notes. For example, I did not realize that twisted endoscopy for SOp2nq was needed to formulate the local intertwining relation. In retrospect, it is

xviii

PREFACE

probably for the best that this second case of twisted endoscopy does have a role here, since it forces us to confront a general phenomenon in a concrete situation. I have tried to make this point explicit in 2.4 with the discussion surrounding the short exact sequence (2.4.10). In any case, I hope that I have accounted for most of the recent work on the subject, in the references and the text. There will no doubt be omissions. I most regret not being able to describe the results of Moeglin on the structure of p-adic packets r v ([M1][M4]). It is clearly an important problem to establish analogues of her results for archimedean packets. Acknowledgment: I am indebted to the referee, for a close and careful reading of an earlier draft of the manuscript (which went to the end of 8.2), and for many very helpful comments.

CHAPTER 1

Parameters
1.1. The automorphic Langlands group We begin with some general motivation. We shall review some of the fundamental ideas that underlie the theoretical foundations laid by Langlands. This will help us put our theorems into perspective. It will also lead naturally to a formulation of some of the essential objects with which we need to work. We take F to be a local or global eld of characteristic 0. In other words, F is a nite extension of either the real eld R or a p-adic eld Qp , or it is a nite extension of Q itself. Suppose that G is a connected reductive algebraic group over F , which to be concrete we take to be a classical matrix group. For example, we could let G be the general linear group GpN q GLpN q of invertible matrices of rank N over F . In his original paper [L2], Langlands introduced what later became known as the L-group of G. This object is a semidirect product
L

p G G F

p of a complex dual group G of G with the Galois group F GalpF {F q p of an algebraic closure F of F . The action of F on G (called an L-action) is determined by its action on a based root datum for G and a corresponding p splitting for G, according to the general theory of algebraic groups. (See [K1, 1.11.3].) It factors through the quotient E{F of F attached to any nite Galois extension E F over which G splits. We sometimes formulate the L-group by the simpler prescription
L

p G L GE{F G E{F ,

since this suces for many purposes. If G GpN q, for example, the action p p of F on G is trivial. Since G is just the complex general linear group GLpN, Cq in this case, one can often take
L

p G G GLpN, Cq.
1

1. PARAMETERS

Langlands conjectures [L2] predicate a fundamental role for the L-group in the representation theory of G. Among other things, Langlands conjectured the existence of a natural correspondence between two quite dierent kinds of objects. The domain consists of (continuous) L-homomorphisms : F
L

G,

p taken up to conjugation by G. (An L-homomorphism between two groups that bre over F is a homomorphism that commutes with the two projections onto F .) The codomain consists of irreducible representations of GpF q if F is local, and automorphic representations of GpAq if F is global, taken in each case up to the usual relation of equivalence of irreducible representations. Recall that if F is global, the adele ring is dened as a restricted tensor product A AF Fv
v

of completions Fv of F . In this case, the Langlands correspondence should satisfy the natural local-global compatibility condition. Namely, if v denotes the restriction of to the subgroup Fv of F (which is dened up to conjugacy), and is a restricted tensor product
v

v ,

v v ,

of representations that correspond to these localizations, then should correspond to . We refer the reader to the respective articles [F] and [L6] for a discussion of restricted direct products and automorphic representations. The correspondence , which remains conjectural, is to be understood in the literal sense of the word. For general G, it will not be a mapping. However, in the case G GLpN q, the correspondence should in fact reduce to a well dened, injective mapping. For local F , this is part of what has now been established. For global F , the injectivity would be a consequence of the required local-global compatibility condition and the theorem of strong multiplicity one, or rather its generalization in [JS] that we will recall in 1.3. However, the correspondence will very denitely not be surjective. In our initial attempts at motivation, we should not lose sight of the fact that the conjectural Langlands correspondence is very deep. For example, even though the mapping is known to exist for G GLp1q, it takes the form of the fundamental reciprocity laws of local and global class eld theory. Its generalization to GLpN q would amount to a formulation of nonabelian class eld theory.

1.1. THE AUTOMORPHIC LANGLANDS GROUP

Langlands actually proposed the correspondence with the Weil group WF in place of the Galois group F . We recall that WF is a locally compact group, which was dened separately for local and global F by Weil. It is equipped with a continuous homomorphism WF F , with dense image. If F is global, there is a commutative diagram WF v WF Fv F

for any completion Fv of F , with vertical embeddings dened up to conjugation. (See [T2].) In the Weil form of the Langlands correspondence, represents an L-homomorphism from WF to L G. The restriction mapping of continuous functions on F to continuous functions on WF is injective. For this reason, the conjectural Langlands correspondence for Weil groups is a generalization of its version for Galois groups. For G GLpN q, the Weil form of the conjectural correspondence again reduces to an injective mapping. (In the global case, one has to take to be an isobaric automorphic representation, a natural restriction introduced in [L7] that includes all the representations in the automorphic spectral decomposition of GLpN q.) If G GLp1q, it also becomes surjective. However, for nonabelian groups G, and in particular for GLpN q, the correspondence will again not be surjective. One of the purposes of Langlands article [L7] was to suggest the possibility of a larger group, which when used in place of the Weil group, would give rise to a bijective mapping for GLpN q. Langlands formulated the group as a complex, reductive, proalgebraic group, in the spirit of the complex form of Grothendiecks motivic Galois group. Kottwitz later pointed out that Langlands group ought to have an equivalent but simpler formulation as a locally compact group LF [K1]. It would come with a surjective mapping LF WF onto the Weil group, whose kernel KF should be compact and connected, and (in the optimistic view of some [A17]) even simply connected. If F is local, LF would take the simple form # WF , if F is archimedean, (1.1.1) LF WF SU p2q, if F is nonarchimedean. In this case, LF is actually a split extension of WF by a compact, simply connected group (namely, the trivial group t1u if F is archimedean and the three dimensional compact Lie group SU p2q SU p2, Rq if F is p-adic.) If F is global, LF remains hypothetical. Its existence is in fact one of the deepest

1. PARAMETERS

problems in the subject. Whatever form it does ultimately take, it ought to t into a larger commutative diagram LFv LF WFv WF Fv F

for any completion Fv , the vertical embedding on the left again being dened up to conjugation. The hypothetical formal structure of LF is thus compatible with an extension of the Langlands correspondence from WF to LF . This is what Langlands proposed in [L7] (for the proalgebraic form of LF ). The extension amounts to a hypothetical correspondence , in which now represents an L-homomorphism : LF L G, p taken again up to G-conjugacy. Here it is convenient to use the Weil form of the L-group L p G G WF , p for the action of WF on G inherited from F . An L-homomorphism between two groups over WF is again one that commutes with the two projections. There are some minor conditions on that are implicit here. For example, since WF and LF are no longer compact, one has to require that for any p P LF , the image of pq in G be semisimple. If G is not quasisplit, one generally also requires that be relevant to G, in the sense that if its image lies in a parabolic subgroup L P of L G, there is a corresponding parabolic subgroup P of G that is dened over F . Suppose again that G GLpN q. Then the hypothetical extended correspondence again reduces to an injective mapping. However, this time it should also be surjective (provided that for global F , we take the image to be the set of isobaric automorphic representations). If F is local archimedean, so that LF WF , the correspondence was established (for any G in fact) by Langlands [L11]. If F is a local p-adic eld, so that LF WF SU p2q, the correspondence was established for GLpN q by Harris and Taylor [HT] and Henniart [He1]. For global F , the correspondence for GLpN q is much deeper, and remains highly conjectural. We have introduced it here as a model to motivate the form of the theorems we seek for classical groups. If G is more general, the extended correspondence will not reduce to a mapping. It was to account for this circumstance that Langlands introduced what are now called L-packets. We recall that L-packets are supposed to be the equivalence classes for a natural relation that is weaker than the usual notion of equivalence of irreducible representations. (The supplementary relation is called L-equivalence, since it is arithmetic in nature, and is designed to preserve the L-functions and -factors of representations.) The extended correspondence is supposed to project to a well dened

1.1. THE AUTOMORPHIC LANGLANDS GROUP

mapping from the set of parameters to the set of L-packets. For G GLpN q, L-equivalence reduces to ordinary equivalence. The L-packets then contain one element each, which is the reason that the correspondence reduces to a mapping in this case. The general construction of L-packets is part of Langlands conjectural theory of endoscopy. It will be a central topic of investigation for this volume. We recall at this stage simply that the L-packet attached to a given will be intimately related to the centralizer ` p (1.1.2) S Cent Impq, G p in G of the image pLF q of , generally through its nite quotient (1.1.3)
0 p S S {S ZpGq .

0 Following standard notation, we have written S for the connected compop p nent of 1 in the complex reductive group S , ZpGq for the center of G, and p p ZpGq for the subgroup of invariants in ZpGq under the natural action of the Galois group F . For G GLpN q, the groups S have all to be connected. Each quotient S is therefore trivial. The implication for other groups G is that we will have to nd a way to introduce the centralizers S , even though we have no hope of constructing the automorphic Langlands group LF and the general parameters . There is a further matter that must also be taken into consideration. Suppose for example that F is global and that G GLpN q. The problem in this case is that the conjectural parametrization of automorphic representations by N -dimensional representations

p : LF G GLpN, Cq ` is not compatible with the spectral decomposition of L2 GpF qzGpAq . If is irreducible, is supposed to be a cuspidal automorphic representation. Any such representation is part of the discrete spectrum (taken modulo the center). However, there are also noncuspidal automorphic representations in the discrete spectrum. These come from residues of Eisenstein series, and include for example the trivial one-dimensional representation of GpAq. Such automorphic representations will correspond to certain reducible N dimensional representations of LF . How is one to account for them? The answer, it turns out, lies in the product of LF with the supplementary group SU p2q SU p2, Rq. The representations in the discrete automorphic spectrum of GLpN q should be attached to irreducible unitary N dimensional representations of this product. The local constituents of these automorphic representations should again be determined by the restriction of parameters, this time from the product LF SU p2q to its subgroups LFv SU p2q. Notice that if v is a p-adic valuation, the localization LFv SU p2q WFv SU p2q SU p2q

1. PARAMETERS

contains two SU p2q-factors. Each will have its own distinct role. In 1.3, we shall recall the general construction, and why it is the product LF SU p2q that governs the automorphic spectrum of GLpN q. Similar considerations should apply to a more general connected group G over any F . One would consider L-homomorphisms : LF SU p2q
L

G,

p with relatively compact image in G (the analogue of the unitary condition GLpN q). If F is global, the parameters should govern the automorphic spectrum of G. If F is local, they ought to determine corresponding local constituents. In either case, the relevant representations should occur in packets , which are larger and more complicated than L-packets, but which are better adapted to the spectral properties of automorphic representations. These packets should in turn be related to the centralizers ` p S Cent Impq, G and their quotients
0 p S S {S ZpGq .

For G GLpN q, the groups S remain connected. However, for other classical groups G we might wish to study, we must again be prepared to introduce parameters and centralizers S without reference to the global Langlands group LF . The objects of study in this volume will be orthogonal and symplectic groups G. Our general goal will be to classify the representations of such groups in terms of those of GLpN q. In the hypothetical setting of the discussion above, the problem includes being able to relate the parameters for G with those for GLpN q. As further motivation for what is to come, we shall consider this question in the next section. We shall analyze the self-dual, nite dimensional representations of a general group F . Among other things, this exercise will allow us to introduce endoscopic data, the internal objects for G that drive the classication, in concrete terms. 1.2. Self-dual, nite dimensional representations We continue to take F to be any local or global eld of characteristic 0. For this section, we let F denote a general, unspecied topological group. The reader can take F to be one of the groups F , WF or LF discussed in 1.1, or perhaps the product of one of these groups with SU p2q. We assume only that F is equipped with a continuous mapping F F , with connected kernel and dense image. We shall be looking at continuous, N -dimensional representations r : F GLpN, Cq. Any such r factors through the preimage of a nite quotient of F . We can therefore replace F by its preimage. In fact, one could simply take a large

1.2. SELF-DUAL, FINITE DIMENSIONAL REPRESENTATIONS

nite quotient of F in place of F , which for the purposes of the present exercise we could treat as an abstract nite group. We say that r is self-dual if it is equivalent to its contragredient representation r_ pq t rpq1 , P F , t x is the usual transpose mapping. In other words, the equivawhere x lence class of r is invariant under the standard outer automorphism pxq x_ t x1 , x P GLpN q, of GLpN q. This condition depends only on the inner class of . It remains the same if is replaced by any conjugate g pxq g 1 pxqg, g P GLpN q. We shall analyze the self-dual representations r in terms of orthogonal and symplectic subgroups of GLpN, Cq. We decompose a given representation r into a direct sum r
1 r1

r rr ,

for inequivalent irreducible representations rk : F GLpNk , Cq, and multiplicities

k

1 k r,
r Nr .

with N
1 N1

` `

The representation is self-dual if and only if there is an involution k k _ _ on the indices such that for any k, rk is equivalent to rk_ and k k_ . We shall say that r is elliptic if it satises the further constraint that for each k, k _ k and k 1. We shall concentrate on this case. Assume that r is elliptic. Then r r1 rr , for distinct irreducible, self-dual representations ri of F of degree Ni . If i is any index, we can write
_ ri pq Ai ri pqA1 , i

P F ,

for a xed element Ai P GLpNi , Cq. Applying the automorphism to each side of this equation, we then see that
_ ri pq A_ ri pA_ q1 pA_ Ai qri pqpA_ Ai q1 . i i i i

Since ri is irreducible, the product A_ Ai is a scalar matrix. We can therefore i write t Ai ci Ai , ci P C . If we take the transpose of each side of this equation, we see further that c2 1. Thus, ci equals `1 or 1, and the nonsingular matrix Ai is either i symmetric or skew-symmetric. The mapping x pA1 q t x Ai , i x P GLpN q,

1. PARAMETERS

of course represents the adjoint relative to the bilinear form dened by Ai . Therefore ri pq belongs to the corresponding orthogonal group OpAi , Cq or symplectic group SppAi , Cq, according to whether ci equals `1 or 1. Let us write IO and IS for the set of indices i such that ci equals `1 and 1 respectively. We then write ri pq, P F , r pq
iPI

A and N

iPI

Ai ,

iPI

Ni ,

for equal to O or S. Thus AO is a symmetric matrix in GLpNO , Cq, AS is a skew-symmetric matrix in GLpNS , Cq, and rO and rS are representations of F that take values in the respective groups OpAO , Cq and SppAS , Cq. We have established a canonical decomposition r rO rS of the self-dual representation r into orthogonal and symplectic components. It will be only the equivalence class of r that is relevant, so we are free to replace rpq by its conjugate B 1 rpqB by a matrix B P GLpN, Cq. This has the eect of replacing the matrix A AO AS by t BAB. In particular, we could take AO to be any symmetric matrix in GLpNO , Cq, and AS to be any skew-symmetric matrix in GLpNS , Cq. We may therefore put the orthogonal and symplectic groups that contain the images of rO and rS into standard form. It will be convenient to adopt a slightly dierent convention for these groups. As our standard orthogonal group in GLpN q, we take OpN q OpN, Jq, where 0 J JpN q 1 1 . .. 0

is the second diagonal in GLpN q. This is a group of two connected components, whose identity component is the special orthogonal group ( SOpN q x P OpN q : detpxq 1 .

1.2. SELF-DUAL, FINITE DIMENSIONAL REPRESENTATIONS

As the standard symplectic group in GLpN q, dened for N 2N 1 even, we take the connected group SppN q SppN, J 1 q for the skew-symmetric matrix J J pN q
1 1

0 JpN 1 q . JpN 1 q 0

The advantage of this formalism is that the set of diagonal matrices in either SOpN q or SppN q forms a maximal torus. Similarly, the set of upper triangular matrices in either group forms a Borel subgroup. The point is that if t t 1 , x P GLpN q, tx J x J J x J denotes the transpose of x about the second diagonal, the automorphism IntpJq : x JpxqJ 1 t x1 of GLpN q stabilizes standard Borel subgroup of upper triangular matrices. Notice that there is a related automorphism r r r r pN q IntpJq : x JpxqJ 1 dened by the matrix (1.2.1) r r J JpN q p1qN `1 0 1 1 , .. . 0

which stabilizes the standard splitting in GLpN q as well. Both of these automorphisms lie in the inner class of , and either one could have been used originally in place of . Returning to our discussion, we can arrange that A equals JpNO q J 1 pNS q. It is best to work with the matrix 1 0 JpNS q 1 , JpNO q JO,S JpNO , NS q NS 2NS , 1q JpNS 0 obtained from the obvious embedding of JpNO q J 1 pNS q into GLpN, Cq. The associated elliptic representation r from the given equivalence class then maps F to the corresponding subgroup of GLpN, Cq, namely the subgroup OpNO , Cq SppNS , Cq dened by the embedding y11 0 y12 px, yq 0 x 0 , y21 0 y22
1 1 where yij are the four pNS NS q-block components of the matrix y P SppNS , Cq.

10

1. PARAMETERS

The symplectic part rS of r is the simpler of the two. Its image is contained in the connected complex group p GS SppNS , Cq. This in turn is the dual group of the split classical group GS SOpNS ` 1q. The orthogonal part rO of r is complicated by fact that its image is contained only in the disconnected group OpNO , Cq. Its composition with the projection of OpNO , Cq onto the group OpNO , Cq{SOpNO , Cq Z{2Z of components yields a character on F of order 1 or 2. Since we are assuming that the kernel of the mapping of F to F is connected, can be identied with a character on the Galois group F of order 1 or 2. This in turn determines an extension E of F of degree 1 or 2. Suppose rst that NO is odd. In this case, the matrix pIq in OpNO q represents the nonidentity component, and the orthogonal group is a direct product OpNO , Cq SOpNO , Cq Z{2Z. We write p SOpNO , Cq GO , where GO is the split group SppNO 1q over F . We then use to identify the direct product L p pGqE{F GO E{F with a subgroup of OpNO , Cq, namely SOpNO , Cq or OpNO , Cq, according to whether has order 1 or 2. We thus obtain an embedding of the (restricted) L-group of GO into GLpNO , Cq. Assume next that N is even. In this case, the nonidentity component in OpN q acts by outer automorphism on SOpNO q. We write p SOpNO , Cq GO , where GO is now the corresponding quasisplit orthogonal group SOpNO , q over F dened by . In other words, GO is the split group SOpNO q if is trivial, and the non-split group obtained by twisting SOpNO q over E by the given outer automorphism if is nontrivial. Let wpNO q be the permutation r matrix in GLpNO q that interchanges the middle two coordinates, and leaves the other coordinates invariant. We take this element as a representative of the nonidentity component of OpNO , Cq. We then use to identify the semidirect product L p pGO qE{F GO E{F with a subgroup of OpNO , Cq, namely SOpNO , Cq or OpNO , Cq as before. We again obtain an embedding of the (restricted) L-group of GO into GLpNO , Cq.

1.2. SELF-DUAL, FINITE DIMENSIONAL REPRESENTATIONS

11

We have shown that the elliptic self-dual representation r factors through the embedded subgroup
L

GE{F

pGO qE{F L pGS qE{F

of GLpN, Cq attached to a quasisplit group G GO GS over F . The group G is called a -twisted endoscopic group for GLpN q. It is determined by r, and in fact by the decomposition N NO ` NS and the character G (of order 1 or 2) attached to r. The same is true of the L-embedding p O,S, : L G G F L GpN q GLpN, Cq F , obtained by inating the embedding above to the full L-groups. It is convenient to form the semidirect product r r G` pN q GLpN q xy GpN q xpN qy, r where xy and xpN qy are the groups of order 2 generated by the automorr q. We write phisms and pN r G0 pN q GLpN q 1 GpN q 1 for the identity component, which we can of course identify with the general linear group GLpN q, and r r (1.2.2) GpN q GLpN q GpN q pN q for the other connected component. Given r, and hence also the decomposition N NO ` NS , we can form the semisimple element
1 s sO,S JO,S ,

p r in the dual set GpN q of complex points GLpN, Cq . The complex group p GO GS , attached to r as above, is then the connected centralizer of p p G s in the group p p r GpN q G0 pN q GLpN, Cq. r The triplet pG, s, q is called an endoscopic datum for GpN q, since it becomes r a special case of the terminology of [KS, p. 16] if we replace GpN q with the ` 0 r r pN q, pN q . pair G The endoscopic datum pG, s, q we have introduced has the property of being elliptic. This is a consequence of our condition that the original selfdual representation r is elliptic. A general (nonelliptic) endoscopic datum r for GpN q is again a triplet pG, s, q, where G is a quasisplit group over F , s p r p is a semisimple element in GpN q of which G is the connected centralizer in p r r G0 pN q, and is an L-embedding of L G into the L-groups L G0 pN q L GpN q of GLpN q. (In the present setting, we are free to take either the Galois or p Weil form of the L-groups.) We require that equal the identity on G, and

12

1. PARAMETERS

p r that the projection onto G0 pN q of the image of lie in the full centralizer r of s. The endoscopic group G (or datum pG, s, q) for GpN q then said to be p , the subgroup of elements in the center ZpGq of G invariant p p elliptic if ZpGq under the action of the Galois group F , is nite. The notion of isomorphism between two general endoscopic data is dened in [KS, p. 18]. In the case at hand, it is given by an element g in the p dual group GpN q GLpN, Cq whose action by conjugation is compatible in a natural sense with the two endoscopic data. We write r AutN pGq AutGpN q pGq r for the group of isomorphisms of the endoscopic datum G to itself. The main role for this group is in its image r r OutN pGq AutN pGq{r N pGq Int in the group of outer automorphisms of the group G over F . (Following standard practice, we often let the endoscopic group G represent a full endoscopic datum pG, s, q, or even an isomorphism class of such data.) If G r represents one of the elliptic endoscopic data constructed above, OutN pGq is trivial if the integer NO is odd or zero. In the remaining case that NO is even r and positive, OutN pGq is a group of order 2, the nontrivial element being the outer automorphism induced by the nontrivial connected component of OpNO , Cq. We write ` r r EpN q E GpN q r for the set of isomorphism classes of endoscopic data for GpN q, and ` r r Eell pN q Eell GpN q r for the subset of classes in EpN q that are elliptic. The data pG, s, q, attached to equivalence classes of elliptic, self-dual representations r as above, r r form a complete set of representatives of Eell pN q. The set Eell pN q is thus parametrized by triplets pNO , NS , q, where NO ` NS N is a decomposition of N into nonnegative integers with NS even, and G is a character of F of order 1 or 2 with the property that 1 if NO 0, and 1 if NO 2. (The last constraint is required in order that the datum be elliptic.) The goal of this volume is to describe the representations of the classical groups G in terms of those of GLpN q. The general arguments will p be inductive. For this reason, the case in which G is either purely orthogonal or purely symplectic will have a special role. Accordingly, we write ` r r Esim pN q Esim GpN q r for the set of elements in Eell pN q that are simple, in the sense that one of the integers NO or NS vanishes. We then have a chain of sets (1.2.3) r r r Esim pN q Eell pN q EpN q,

1.2. SELF-DUAL, FINITE DIMENSIONAL REPRESENTATIONS

13

which are all nite if F is local, and all innite if F is global. r The elements G P EpN q are usually called twisted endoscopic data, since they are attached to the automorphism . We shall have to work also with ordinary (untwisted) endoscopic data, at least for the quasisplit orthogonal r and symplectic groups G that represent elements in Esim pN q. An endoscopic 1 for G is similar to what we have described above for GpN q. It r datum G 1 , s1 , 1 q, where G1 is a (connected) quasisplit group amounts to a triplet pG p p over F , s1 is a semisimple element in G of which G1 is the connected central1 is an L-embedding of L G1 into L G. We again require that p izer in G, and 1 equal the identity on G1 , and that its image lie in the centralizer of s1 in p L G. (See [LS1, (1.2)], a special case of the general denition in [KS], which we have specialized further to the case at hand.) There is again the notion of isomorphism of endoscopic data, which allows us to form the associated nite group OutG pG1 q AutG pG1 q{IntG pG1 q of outer automorphisms of any given G1 . We write EpGq for the set of isomorphism classes of endoscopic data G1 for G, and Eell pGq for the subset p of data that are elliptic, in the sense that ZpG1 q is nite. We then have a second chain of sets (1.2.4) Esim pGq Eell pGq EpGq, where Esim pGq tGu is the subset consisting of G alone. Similar denitions r apply to groups G that represent more general data in EpN q. rsim pN q. It suces to It is easy to describe the set Eell pGq, for any G P E p consider diagonal matrices s1 P G with eigenvalues 1. For example, in the p rst case that G SOpN ` 1q and G SppN, Cq (with N NS even), it is enough to take diagonal matrices of the form 2 I1 0 1 , I2 s1 2 0 I1
2 2 1 1 where I1 is the identity matrix of rank N1 N1 {2, and I2 is the identity 1 1 1 matrix of rank N2 . The set Eell pGq is parametrized by pairs pN1 , N2 q of 1 N 1 and N N 1 ` N 1 . The corresponding even integers with 0 N1 2 1 2 endoscopic groups are the split groups 1 1 G1 SOpN1 ` 1q SOpN2 ` 1q,

with dual groups 1 1 p p G1 SppN1 , Cq SppN2 , Cq SppN, Cq G.

1 1 The group OutG pG1 q is trivial in this case unless N1 N2 , in which case it has order 2. The other cases are similar. In the second case that G SppN 1q p and G SOpN, Cq (with N NO odd), Eell pGq is parametrized by pairs 1 , N 1 q of nonnegative even integers with N N 1 `pN 1 `1q, and characters pN1 2 1 2

14

1. PARAMETERS

1 on F with p 1 q2 1. The corresponding endoscopic groups are the quasisplit groups 1 1 G1 SOpN1 , 1 q SppN2 q, with dual groups
1 1 p p G1 SOpN1 , Cq SOpN2 ` 1, Cq SOpN, Cq G.

p In the third case that G SOpN, q and G SOpN, Cq (with N NO 1 1 even), Eell pGq is parametrized by pairs of even integers pN1 , N2 q with 1 N 1 and N N 1 ` N 1 , and pairs p 1 , 1 q of characters on 0 N1 F 2 1 2 1 2 1 1 1 1 with p1 q2 p2 q2 1 and 1 2 . The corresponding endoscopic groups are the quasisplit groups
1 1 1 1 G1 SOpN1 , 1 q SOpN2 , 2 q,

with dual groups

1 1 p p G1 SOpN1 , Cq SOpN2 , Cq SOpN, Cq G.

In the second and third cases, each character 1 has to be nontrivial if the corresponding integer N 1 equals 2, while if N 1 0, 1 must of course be trivial. In these cases, the group OutG pG1 q has order 2 unless N is even and 1 1 N1 N2 1, in which case it is a product of two groups of order 2, or 1 0, in which case the group is trivial. N1 p Observe that in the even orthogonal case, where G SOpN, Cq with N 1 P E pGq will not be able to isolate constituents even, the endoscopic data G ell ri of r of odd dimension. The discrepancy is made up by a third kind of endoscopic datum. These are the twisted endoscopic data for the even r orthogonal groups G SOpN, q in Esim pN q. For any such G, let ` r r r r (1.2.5) G G , Int wpN q , be the nonidentity component in the semi-direct product of G with the r group of order two generated by the outer automorphism of SOpN q. In r r r r this setting a (twisted) endoscopic datum is a triplet pG1 , s1 , 1 q, where G1 1 is a semisimple element in the dual set is a quasisplit group over F , s r p p r of which G1 is the connected centralizer in G, and 1 is an r r p r p G G r L-embedding of L G1 into L G, all being subject also to further conditions r r and denitions as above. The subset Eell pGq EpGq of isomorphism classes r of elliptic endoscopic data for G is parametrized by pairs of odd integers r 1 , N 1 q, with 0 N 1 N 1 and N N 1 ` N 1 , and pairs of characters r r r r r pN1 2 1 2 1 2 1 , 1 q on , with pr1 q2 pr1 q2 1 and 1 1 . The corresponding r1 r2 pr1 r2 1 2 F endoscopic groups are the quasisplit groups
1 r r1 G1 SppN1 q SppN2 q,

with dual groups p r r1 r1 p G1 SOpN1 , Cq SOpN2 , Cq SOpN, Cq G.

1.3. REPRESENTATION OF GLpN q

15

r The family Eell pGq will have to be part of our analysis. However, its role r will be subsidiary to that of the two primary families Eell pN q and Eell pGq. We have completed our brief study of elliptic self-dual representations r. Remember that we are regarding these objects as parameters, in the spirit of 1.1. We have seen that a parameter for GLpN q factors into a product of two parameters for quasisplit classical groups. The products are governed by r twisted endoscopic data G P Eell pN q. They can be rened further according to ordinary endoscopic data G1 P Eell pGq. Thus, while the parameters are not available (for lack of a global Langlands group LF ), the endoscopic data that control many of their properties are. Before we can study the ramications of this, we must rst formulate a makeshift substitute for the parameters attached to our classical groups. We shall do so in 1.4, after a review in 1.3 of the representations of GLpN q that will serve as parameters for this group. We have considered only the self dual representation r that are elliptic, since it is these objects that pertain directly to our theorems. Before going on, we might ask what happens if r is not elliptic. A moments reection reveals that any such r factors through subgroups of GLpN, Cq attached to r several data G P Eell pN q, in contrast to what we have seen in the elliptic case. This is because r also factors through a subgroup attached to a datum r M in the complement of Eell pN q in EpN q, and because any such M can be identied with a proper Levi subgroup of several G. The analysis of general self-dual representations r is therefore more complicated, though still not very dicult. It is best formulated in terms of the centralizers ` ` p r r r Sr pN q Sr GpN q Cent Imprq, GpN q and ` p Sr Sr pGq Cent Imprq, Gq of the images of r. We shall return to this matter briey in 1.4, and then more systematically as part of the general theory of Chapter 4. 1.3. Representations of GLpN q A general goal, for the present volume and beyond, is to classify representations of a broad class of groups in terms of those of general linear groups. What makes this useful is the fact that much of the representation theory of GLpN q is both well understood and relatively simple. We shall review what we need of the theory. Suppose rst that F is local. In this case, we can replace the abstract group F of the last section by the local Langlands group LF dened by (1.1.1). The local Langlands classication parametrizes irreducible representations of GLpN, F q in terms of N -dimensional representations : LF GLpN, Cq. Before we state it formally, we should recall a few basic notions.

16

1. PARAMETERS

Given a nite dimensional (semisimple, continuous) representation of LF , we can form the local L-function Lps, q, a meromorphic function of s P C. We can also form the local -factor ps, , F q, a monomial of the form abs which also depends on a nontrivial additive character F of F . If F is archimedean, we refer the reader to the denition in [T2, 3]. If F is p-adic, we extend analytically to a representation of the product of WF with the complexication SLp2, Cq of the subgroup SU p2q of LF . We can then form the representation 1 |w| 2 0 pwq w, , w P WF , 1 0 |w| 2 of WF , where |w| is the absolute value on WF , and the nilpotent matrix 1 1 N log 1, . 0 1 The pair V p , N q gives a representation of the Weil-Deligne group [T2, (4.1.3)], for which we dene an L-function
s Lps, q ZpV , qF q

and -factor
s ps, , F q pV , qF q, following notation in [T2, 4]. (We have written qF here for the order of the residue eld of F .) Of particular interest are the tensor product L-function

Lps, 1 2 q Lps, 1 b 2 q and -factor ps, 1 2 , F q ps, 1 b 2 , F q, attached to any pair of representations 1 and 2 of LF . We expect also to be able to attach local L-functions Lps, , rq and factors ps, , r, F q to any connected reductive group G over F , where ranges over irreducible representations of GpF q, and r is a nite dimensional representation of L G. For general G, this has been done in only the simplest of cases. However, if G is a product GpN1 q GpN2 q of general linear groups, there is a broader theory [JPS]. (See also [MW2, Appendice].) It applies to any representation 1 2 , in the case that r is the standard representation (1.3.1) of p G GLpN1 , Cq GLpN2 , Cq on the space of complex pN1 N2 q-matrices X. The theory yields functions Lps, 1 2 q Lps, , rq and ps, 1 2 , F q ps, , r, F q, rpg1 , g2 q : X g1 X t g2 , gi P GpNi q,

1.3. REPRESENTATION OF GLpN q

17

known as local Rankin-Selberg convolutions. The local classication for GLpN q is essentially characterized by being compatible with local Rankin-Selberg convolutions. It has other important properties as well. Some of these relate to supplementary conditions we can impose on the parameters as follows. Suppose for a moment that G is any connected group over F . We p write pGq for the set of G-orbits of (semisimple, continuous, G-relevant) L-homomorphisms : LF
L

G,

and pGq for the set of equivalence classes of irreducible (admissible) representations of GpF q. (See [Bo].) These sets come with parallel chains of subsets 2,bdd pGq bdd pGq pGq and 2,temp pGq temp pGq pGq. In the second chain, temp pGq denotes the set of tempered representations in pGq, and 2,temp pGq 2 pGq X temp pGq, where 2 pGq is the set of representations in pGq that are essentially square integrable, in the sense that after tensoring with the appropriate positive character on GpF q, they are square integrable modulo the centre of GpF q. Recall that temp pGq can be described informally as the set of represen` tations P pGq that occur in the spectral decomposition of L2 GpF q . Similarly, 2,temp pGq is the set of that occur in the discrete spectrum (taken modulo the center). In the rst chain, bdd pGq denotes the set of P pGq whose image in L G projects onto a relatively compact subset of p G, and 2,bdd pGq 2 pGq X bdd pGq, where 2 pGq is the set of parameters in pGq whose image does not lie in any proper parabolic subgroup L P of L G. In the` case G GpN q GLpN q of present concern, we write ` pN q GLpN q and pN q GLpN q , and follow similar notation for the corresponding subsets above. Then pN q can be identied with the set of equivalence classes of (semisimple, continuous) N -dimensional representations of LF . The subset sim pN q 2 pN q consists of those representations that are irreducible, while bdd pN q corresponds to representations that are unitary. On the other hand, the set unit pN q of unitary representations in pN q properly contains temp pN q, if N 2. It has an elegant classication [V2], [Tad1], but our point here is

18

1. PARAMETERS

that it is not parallel to the set of N -dimensional representations that are unitary. We do observe that 2,temp pN q 2 pN q X unit pN q 2,unit pN q, so the notions of tempered and unitary are the same for square integrable representations. If F is p-adic, we can write scusp,temp pN q (resp. scusp pN q) for the set of supercuspidal representations in 2,temp pN q (resp. 2 pN q). We can also write scusp,bdd pN q (resp. scusp pN q) for the set of in sim,bdd pN q (resp. sim pN q) that are trivial on the second factor SU p2q of LF . If F is archimedean, it is natural to take scusp,temp pN q and scusp pN q to be empty unless GLpN, F q is compact modulo the center (which is a silly way of saying that N 1), in which case we can take them to the correspondbe ` ` ` ` ing sets 2,temp GLp1q temp GLp1q and 2 GLp1q GLp1q . We thus have two parallel chains (1.3.2) and (1.3.3) scusp,temp pN q 2,temp pN q temp pN q pN q, scusp,bdd pN q sim,bdd pN q bdd pN q pN q,

for our given local eld F . The local classication for G GLpN q can now be formulated as follows. Theorem 1.3.1 (Langlands [L11], Harris-Taylor [HT], Henniart [He1]). There is a unique bijective correspondence from pN q onto pN q such that piq b b p detq,

for any character in the group p1q p1q, piiq det , _ _ ,

for the central character of , and piiiq

for the duality involutions _ on pN q and pN q, and such that if i i , then pivq and pvq ps, 1 2 , F q ps, 1 2 , F q. Lps, 1 2 q Lps, 1 2 q i P pNi q, i 1, 2,

Furthermore, the bijection is compatible with the two chains (1.3.2) and (1.3.3), in the sense that it maps each subset in (1.3.2) onto its counterpart in (1.3.3).

1.3. REPRESENTATION OF GLpN q

19

We now take F to be global. This brings us to the representations that will be the foundation of all that follows. They are the objects in the set ` Acusp pN q Acusp GLpN q of (equivalence classes of) unitary, cuspidal automorphic representations of GLpN q. Assume rst that G is an arbitrary connected group over the global eld F . To suppress the noncompact part of the center, one often works with the closed subgroup ( GpAq1 x P GpAq : |pxq| 1, P XpGqF of GpAq, where XpGqF is the additive group of characters of G dened over F . We recall that GpF qzGpAq1 has nite volume, and that there is a sequence ` ` ` 1 L2 L2 GpF qzGpAq1 L2 GpF qzGpAq1 , cusp GpF qzGpAq disc of embedded, rightGpAq1 -invariant Hilbert spaces. In particular, the space ` ` 1 L2 functions in L2 GpF qzGpAq1 is contained in cusp GpF qzGpAq` of cuspidal the subspace L2 GpF qzGpAq1 that decomposes under the action of GpAq1 disc into a direct sum of irreducible representations. We can then introduce a corresponding chain of subsets of irreducible automorphic representations Acusp pGq A2 pGq ApGq. By denition, Acusp pGq, A2 pGq and ApGq denote the subsets of irreducible unitary representations of GpAq whose restrictions to GpAq1 are irreducible constituents of the respective spaces L2 , L2 and L2 . We shall also cusp disc write A` pGq and A` pGq for the analogues of Acusp pGq and A2 pGq dened cusp 2 without the condition that be unitary. (The denition of ApGq here is somewhat informal, and will be used only for guidance. It can in fact be made precise at the singular points in the continuous spectrum where there might be some ambiguity.) We specialize again to the ` case G GLpN q, taken now over the global eld F . We write ApN q A GLpN q , with similar notation for the corresponding subsets above. Observe that GLpN, Aq1 is the group of adelic matrices x P GLpN, Aq whose determinant has absolute value 1. In this case, the set ApN q is easy to characterize. It is composed of the induced representations IP p1 b b r q, i P A2 pNi q, where P is the standard parabolic subgroup of block upper triangular matrices in G GLpN q corresponding to a partition pN1 , . . . , Nr q of N , with the standard Levi subgroup of block diagonal matrices MP GLpN1 q GLpNr q. This follows from the theory of Eisenstein series [L1], [L5], [A1] (valid for any G), and the fact [Be] (special to G GLpN q) that an induced

20

1. PARAMETERS

G representation IP pq is irreducible for any representation of MP pAq that is irreducible and unitary. Let us also write A` pN q for the set of induced representations as above, but with the components i now taken from the larger sets A` pNi q. These representations can be reducible at certain 2 points, although they typically remain irreducible. We thus have a chain of sets

(1.3.4)

Acusp pN q A2 pN q ApN q A` pN q

of representations, for our given global eld F . This is a rough global analogue of the local sequence (1.3.3). We have dropped the subscript temp in the global notation, since the local components of a constituent of ` L2 GLpN, F qzGLpN, Aq1 need not be tempered, and added the superscript ` to remind ourselves that A` pN q contains full induced representations, rather than irreducible quotients. There are two fundamental theorems on the automorphic representations of GLpN q that will be essential to us. The rst is the classication of automorphic representations by Jacquet and Shalika, while the second is the characterization by Moeglin and Waldspurger of the discrete spectrum in terms of cuspidal spectra. We shall review each in turn. Suppose again that G is an arbitrary connected group over the global eld F . An automorphic representation of G is among other things, a weakly continuous, irreducible representation of GpAq. As such, it can be written as a restricted tensor product v
v

of irreducible representations of the local groups GpFv q, almost all which are unramied [F]. Recall that if v is unramied, the group Gv G{Fv is unramied. This means that Fv is a p-adic eld, that Gv is quasisplit, and p that the action of WFv on G factors through the innite cyclic quotient WFv {IFv xFrobv y of WFv by the inertia subgroup IFv with canonical generator Frobv . Recall also that the local Langlands correspondence has long existed in this very p particular context. The G-orbit of homomorphisms v : LFv
L

p Gv G xFrobv y

in pGv q to which v corresponds factors through the quotient ` LFv { IFv SU p2q WFv {IFv of LFv . The resulting mapping v cpv q v pFrobv q is a bijection from the set of unramied representations of GpFv q (relative to any given hyperspecial maximal compact subgroup Kv GpFv q) and the p set of semisimple G-orbits in L Gv that project to the Frobenius generator

1.3. REPRESENTATION OF GLpN q

21

of WFv {IFv . Let cv pq be the image of cpv q in L G under the embedding of L Gv into L G that is dened canonically up to conjugation. In this way, the automorphic representation of G gives rise to a family of semisimple conjugacy classes ( cS pq cv pq : v R S in L G, where S is some nite set of valuations of F outside of which G is unramied. S Let Caut pGq be the set of families cS tcv : v R Su of semisimple conjugacy classes in L G obtained in this way. That is, cS cS pq, for some automorphic representation of G. We dene Caut pGq 1 to be the set of equivalence classes of such families, cS and pc1 qS being equivalent if cv equals c1 for almost all v. We then have a mapping v cpq from the set of automorphic representations of G onto Caut pGq. The families cS pq arise most often in the guise of (partial) global Lfunctions. Suppose that r is a nite dimensional representation of L G that is unramied outside of S. The corresponding incomplete L-function is given by an innite product LS ps, , rq Lps, v , rv q
vRS

of unramied local L-functions ` ` s 1 Lps, v , rv q det 1 r cpv q qv ,

v R S,

which converges for the real part of s large. It is at the ramied places v P S that the construction of local L-functions and -factors poses a challenge. As the reader is no doubt well aware, the completed L-function Lps, , rq Lps, v , rv q
v

is expected to have analytic continuation as a meromorphic function of s P C that satises the functional equation (1.3.5) Lps, , rq ps, , rqLp1 s, , r_ q, where r_ is the contragredient of r, and ps, , rq ps, v , rv , Fv q.
vRS

Here, Fv stands for the localization of a nontrivial additive character F on A{F that is unramied outside of S. We return again to the case G GLpN q. The elements in the set ` Caut pN q Caut GLpN q

22

1. PARAMETERS

can be identied with families of semisimple conjugacy classes in GLpN, Cq, dened up to the equivalence relation above. It is convenient to restrict the domain of the corresponding mapping cpq. Recall that general automorphic representations for GLpN q can be characterized as the irreducible constituents of induced representations IP p1 b b r q, i A` pNi q. cusp (See Proposition 1.2 of [L7], which applies to any G.) With this condition, the nonunitary induced representation can have many irreducible constituents. However, it does have a canonical constituent. This is the irreducible representation v , v ` where v P GLpN qv is obtained from the local Langlands parameter v 1,v r,v , i,v i,v , ` in GLpN qv by Theorem 1.3.1 (applied to both GLpN qv and its subgroups GLpNi qv ). Isobaric representations are the automorphic representations of GLpN q obtained in this way. The equivalence class of does not change if we reorder the cuspidal representations ti u. We formalize this property by writing (1.3.6) 1 r , i P A` pNi q, cusp in the notation [L7, 2] of Langlands, where the right hand side is regarded as a formal, unordered direct sum. Theorem 1.3.2 (Jacquet-Shalika [JS]). The mapping 1 r cpq, from the set of equivalence classes of isobaric automorphic representations of GLpN q to the set of elements c cpq in Caut pN q, is a bijection. Global Rankin-Selberg L-functions Lps, 1 2 q and -factors ps, 1 2 q are dened as in the local case. They correspond to automorphic representations 1 2 of a group G GLpN1 q GLpN2 q, p and the standard representation (1.3.1) of G. The analytic behaviour of these functions is now quite well understood [JPS], [MW2, Appendice]. In particular, Lps, 1 2 q has analytic continuation with functional equation (1.3.5). Moreover, if 1 and 2 are cuspidal, Lps, 1 2 q is an entire _ function of s unless N1 equals N2 , and 2 is equivalent to the representation 1 pg1 q | det g1 |s1 1 , g1 P GLpN1 , Aq, for some s1 P C, in which case Lps, 1 2 q has only a simple pole at s s1 . It is this property that is used to prove Theorem 1.3.2. Theorem 1.3.2 can be regarded as a characterization of (isobaric) automorphic representations of GLpN q in terms of simpler objects, families of semisimple conjugacy classes. However, it does not in itself characterize

1.3. REPRESENTATION OF GLpN q

23

the spectral properties of these representations. For example, it is not a priori clear that the representations in ApN q, or even its subset A2 pN q, are isobaric. The following theorem provides the necessary corroboration. Theorem 1.3.3 (Moeglin-Waldspurger [MW2]). The representations P A2 pN q are parametrized by the set of pairs pm, q, where N mn is divisible by m, and is a representation in Acusp pmq. If P is the standard parabolic subgroup corresponding to the partition pm, . . . , mq of N , and is the representation x px1 q| det x1 | of the Levi subgroup ( MP pAq x px1 , . . . , xn q : xi P GLpm, Aq , then is the unique irreducible quotient of the induced representation IP p q. Moreover, the restriction of to GLpN, Aq1 occurs in the discrete spectrum with multiplicity one. Consider the representation P A2 pN q described in the theorem. For any valuation v, its local component v is the Langlands quotient of the local component IP p,v q of IP p q. It then follows from the denitions that v is the irreducible representation of GLpN, Fv q that corresponds to the Langlands parameter of the induced representation IP p,v q. In other words, the automorphic representation is isobaric. We can therefore write (1.3.7) n1 n3 pn1q , 2 2 2 in the notation (1.3.6), for cuspidal automorphic representations piq : xi pxq|x|i , x P GLpm, Aq. Theorem 1.3.3 also provides a description of the automorphic representations in the larger set ApN q. For as we noted above, ApN q consists of the set of irreducible induced representations IP p1 b b r q, i A2 pNi q, where N N1 ` ` Nr is again a partition of N , and P is the corresponding standard parabolic subgroup of GLpN q. It is easy to see from its irreducibility that is also isobaric. We can therefore write (1.3.8) 1 r , i A2 pNi q. Observe that despite the notation, (1.3.8) diers slightly from the general isobaric representation (1.3.6). Its constituents i are more complex, since they are not cuspidal, but the associated induced representation is simpler, since it is irreducible. There is another way to view Theorem 1.3.3. Since F is global, the Langlands group LF is not available to us. If it were, we would expect its set pN q of (equivalence classes of) N -dimensional representations : LF GLpN, Cq
n1 2

b px2 q| det x2 |

n3 2

b b pxn q| det xn |

pn1q 2

24

1. PARAMETERS

to parametrize the isobaric automorphic representations of GLpN q. A very simple case, for example, is the pullback || to LF of the absolute value on the quotient ab WF GLp1, F qzGLp1, Aq of LF . This of course corresponds to the automorphic character on GLp1q given by the original absolute value. We would expect the subset sim,bdd pN q 2,bdd pN q of irreducible unitary representations in pN q to parametrize the unitary cuspidal automorphic representations of GLpN q. How then are we to account for the full discrete spectrum A2 pN q? A convenient way to do so is to take the product of LF with the group SU p2q SU p2, Rq. Let us write pN q temporarily for the set of (equivalence classes of) unitary N -dimension representations : LF SU p2q GLpN, Cq. According to Theorem 1.3.3, it would then be the subset sim pN q 2 pN q of irreducible representations in pN q that parametrizes A2 pN q. Indeed, any P sim pN q decomposes uniquely as a tensor product b of irreducible representations of LF and SU p2q. It therefore gives rise to a pair pm, q, in which P Acusp pmq represents a unitary cuspidal automorphic representation of GLpN q as well as the corresponding irreducible, unitary, m-dimensional representation of LF . Conversely, given any such pair, we again identify P Acusp pmq with the corresponding representation of LF , and we take to be the unique irreducible representation of SU p2q of degree n N m1 . We are identifying any nite dimensional representation of SU p2q with its analytic extension to SLp2, Cq. With this convention, we attach an N dimensional representation 1 0 |u| 2 : u u, , u P LF , 1 0 |u| 2 of LF to any P pN q. If b belongs to the subset sim pN q, decomposes as a direct sum u puq|u|
n1 2

puq|u|p

n1 q 2

to which we associate the induced representation IP p q of Theoreom 1.3.3. According to the rules of the hypothetical global correspondence from pN q to isobaric automorphic representations of GLpN q, it is the unique irreducible quotient of IP p q that is supposed to correspond to the parameter . This representation is unitary, as of course is implicit in Theorem 1.3.3. The mapping is thus an explicit realization of the

1.3. REPRESENTATION OF GLpN q

25

bijective correspondence from sim pN q to A2 pN q implied by the theorem of Moeglin and Waldspurger (and the existence of the group LF ). More generally, suppose that belongs to the larger set pN q. Then the isobaric representation attached to the parameter P pN q belongs to ApN q. It can be described in the familiar way as the irreducible unitary representation induced from a unitary representation of a Levi subgroup. The mapping becomes a bijection from pN q to ApN q. In general, the restriction of parameters is an injection from pN q into pN q. The role of the set pN q we have just dened in terms of the supplementary group SU p2q is thus to single out the subset of pN q that corresponds to the subset ApN q of globally tempered automorphic representations. We are also free to form larger sets ` pN q pN q and ` pN q sim pN q sim of representations of LF SU p2q by removing the condition that they be unitary. Any element P ` pN q is then a direct sum of irreducible representations i P ` pNi q. The components i should correspond to ausim tomorphic representations i i in A` pNi q. The corresponding induced 2 representation IP p1 b b r q then belongs to the set of A` pN q of (possibly reducible) representations of GLpN, Aq introduced above. Notice that the extended mapping from ` pN q to pN q is no longer injective. In particular, will not in general be equal to the automorphic representation corresponding to . However, the mapping will be a bijection from ` pN q to A` pN q. In the interests of symmetry, we can also write cusp pN q sim,bdd pN q for the set of representations that are trivial on the factor SU p2q. This gives us a chain of sets (1.3.9) cusp pN q sim pN q pN q ` pN q that is parallel to (1.3.4). We will then have a bijective correspondence that takes each set in (1.3.9) to its counterpart in (1.3.4). The global parameter sets in the chain (1.3.9) are hypothetical, depending as they do on the global Langlands group LF . However, their local analogues are not. They can be dened for the general linear group G0 pN q GLpN qv over any completion Fv of F . Replacing LF by LFv in v the denitions above, we obtain local parameter sets (1.3.10) cusp,v pN q sim,v pN q v pN q ` pN q. v We also obtain a restriction mapping v from the hypothetical global set ` pN q to the local set ` pN q. The generalized Ramanujan conjecture v

26

1. PARAMETERS

for GLpN q implies that this mapping takes pN q to the subset v pN q of ` pN q. However, the conjecture is not known. For this reason, we will v be forced to work with the larger local sets ` pN q in our study of global v spectra. Our purpose in introducing the hypothetical families of parameters (1.3.9) has been to persuade ourselves that they correspond to well dened families of automorphic representations (1.3.4). This will inform the discussion of the next section. There we shall revisit the constructions of the last section for orthogonal and symplectic groups, but with the objects (1.3.4) in place of the parameter sets (1.3.9). Notice that the left hand three sets in (1.3.4) contain only isobaric automorphic representations. They therefore correspond bijectively with three subsets (1.3.11) Csim pN q C2 pN q CpN q Caut pN q Csim pN q Ccusp pN q for the smallest of these sets, in part because it would be bijective with the hypothetical family sim,bdd pN q cusp pN q of irreducible unitary, N -dimensional representations of the group LF . The largest set Caut pN q would of course be bijective with the family pN q of all N -dimensional representations of LF . The sets (1.3.11) can all be expressed in terms of the smallest set Csim pN q (or rather its analogue Csim pmq for m N ). This is a consequence of Theorem 1.3.3, or if one prefers, its embodiment in the left hand three parameter in (1.3.9). The sets Csim pN q thus contain all the global information for general linear groups. The notation we have chosen is meant to reect the role of their elements as the simple building blocks. Our ultimate goal is to show that this fundamental role extends to orthogonal and symplectic groups. 1.4. A substitute for global parameters We can now resume the discussion from 1.2. We shall use the cuspidal automorphic representations of GLpN q as a substitute for the irreducible N dimensional representations of the hypothetical global Langlands group LF . This will allow us to construct objects that ultimately parametrize families of automorphic representations of classical groups. Assume that F is global. In the last section, we wrote 2 pN q temporarily for the set of equivalence classes of irreducible unitary representations of the group LF SU p2q. From now on, we let ` ` sim pN q sim GpN q sim GLpN q stand for the set of formal tensor products b , P Acusp pmq,

of Caut pN q under the mapping of Theorem 1.3.2. We have written

1.4. A SUBSTITUTE FOR GLOBAL PARAMETERS

27

where N mn and is the unique irreducible representation of SU p2q of degree n. Elements in this set are in bijective correspondence with the set of pairs pm, q of Theorem 1.3.3. We have therefore a canonical bijection from sim pN q onto the relative discrete, automorphic spectrum A2 pN q of GLpN q. For any such , we set cpq equal to cp q, an equivalence class of families in C2 pN q. Then ( cpq cv pq : v R S , for some nite set S S8 of valuations of F outside of which is unramied, and for semisimple conjugacy classes
n1

cv pq cv pq b cv pq cv pqqv 2 cv pqqv

n1 2

in GLpN, Cq. Here S8 denotes the set of archimedean valuations of F as usual, while qv is again the order of the residue eld of Fv . In 1.3, we also wrote pN q temporarily for the set of all unitary representations of LF SU p2q. Following the notation (1.3.8), we now let ` ` pN q GpN q GLpN q denote the set of formal, unordered direct sums (1.4.1)
1 1

r r ,

for positive integers k and distinct elements k k b k in sim pNk q. The ranks here are positive integers Nk mk nk such that N
1 N1

` `

r Nr

1 m 1 n1

` `

r mr nr .

For any , we take to be the irreducible induced representation ` IP 1 b b 1 b b r b b r loooooooomoooooooon loooooooomoooooooon

1 r

of GLpN, Aq, where P is the parabolic subgroup corresponding to the partition ` N1 , . . . , N1 , . . . , Nr , . . . , Nr looooomooooon looooomooooon
1 r

of N . We then have a bijection from pN q onto the full automorphic spectrum ApN q of GLpN q. Again we set cpq equal to the associated class cp q in CpN q. Then ( cpq cv pq : v R S , for a nite set S S8 outside of which each i is unramied, and semisimple conjugacy classes ` ` cv pq cv p1 q cv p1 q cv pr q cv pr q loooooooooooomoooooooooooon loooooooooooomoooooooooooon
1 r

in GLpN, Cq. Suppose that v is any valuation in F . If b belongs to sim pN q, the local component v of is an irreducible representation of GLpm, Fv q.

28

1. PARAMETERS

It corresponds to an m-dimensional representation of the local Langlands group LFv , whose tensor product v with belongs to the set ` pN q of v N -dimensional representations of LFv b SU p2q. More generally, if is any element (1.4.1) in the larger set pN q, the direct sum ` ` v 1,v 1,v r,v r,v looooooooomooooooooon looooooooomooooooooon
1 r

also belongs to the set ` pN q. There is thus a mapping v from v pN q to ` pN q. We emphasize again that we cannot say that v lies in v the smaller local set v pN q, since the generalized Ramanujan conjecture for GLpN q has not been established. We now have a set pN q of formal parameters to represent the automorphic spectrum ApN q of GLpN q. By Theorem 1.3.2, the elements in ApN q are faithfully represented by the (equivalence classes of) families cpq in CpN q. As we noted at the end of 1.3, and have seen explicitly here, these families are easily constructed in terms of cuspidal families c. Reiterating the conclusion from 1.3, the sets Csim pN q Ccusp pN q of cuspidal families are to be regarded as our fundamental data, in terms of which everything we hope to establish for classical groups can be formulated. We are now ready to formalize the constructions of 1.2. The outer automorphism : x x_ of GLpN q acts on the set pN q. It transforms a general parameter (1.4.1) in this set to its contragredient _
_ _ 1 1 r r _ _ _ _ 1 p1 b 1 q r pr b r q _ _ 1 p1 b 1 q r pr b r q,

where _ is the contragredient of the cuspidal automorphic representation k _ k of GLpmk q. We have written k k , since any irreducible repre_ sentation of SU p2q is self-dual. The contragredient of the associated automorphic representation P ApN q is then equal to _ . We introduce the subset ` ( r r (1.4.2) pN q GpN q P pN q : _ of self-dual parameters in pN q, a family that is naturally associated to the connected component r r GpN q GpN q pN q r in GpN q` GLpN q xy. Elements in this subset correspond to the repr resentations P ApN q that extend to the group GpN, Aq` generated by r GpN, Aq. There is also an action c tcv u c_ tc_ u v r of on the set of automorphic families CpN q. The subset pN q of pN q r q of self-dual families in CpN q. corresponds to the subset CpN

1.4. A SUBSTITUTE FOR GLOBAL PARAMETERS

29

It is sometimes convenient to write K for the set that parametrizes the simple constituents k in (1.4.1) of a general element P pN q. We shall usually think of K as an abstract indexing set, of which (1.4.1) represents r a (noncanonical) enumeration. If belongs to the subset pN q of pN q, _ there is an involution k k _ on K such that k_ k and k_ k . We can then write the indexing set as a disjoint union _ _ K I J J , J tj _ : j P J u, where I is the set of xed points of the involution, and J is some complementary subset that represents the orbits of order 2. The union of I and J of course represents the set tK u of all orbits of the involution on K . r Having dened the set pN q, we can form the subset r r sim pN q pN q X sim pN q, r of parameters (1.4.1) in pN q that are simple, in the sense that r 1 and r 1 1. An element in sim pN q is thus a self-dual parameter b in the set sim pN q. We shall also write r r r sim pN q cusp pN q pN q X cusp pN q r for the subset of parameters in sim pN q that are generic, in the sense that r is trivial. A simple generic parameter in pN q is thus a self-dual element in the basic set Acusp pN q of unitary cuspidal automorphic representations of GLpN q. These will be our fundamental building blocks for orthogonal and symplectic groups. Recall that the term simple was applied in 1.2 to endoscopic data. r It was used to designate the subset Esim pN q of elliptic endoscopic data r r G P Eell pN q that are not composite. We note here that any G in Esim pN q, or r indeed in the larger set Eell pN q, has the property that the split component of its center is trivial. This means that the group GpAq1 equals the full adelic group GpAq. r Theorem 1.4.1. Suppose that P sim pN q is a simple generic (forr mal) global parameter. Then there is a unique G P Eell pN q (taken as an isomorphism class of endoscopic data pG , s , q) such that ` cpq cpq , for some representation in A2 pGq. Moreover, G is simple. The assertion here is quite transparent. From among all the twisted, r r elliptic endoscopic data G P Eell pN q for GpN q, there is exactly one that is the source of the conjugacy class data of . This is of course what we would expect from the discussion of 1.2. Theorem 1.4.1 is the rst of a collection of theorems for orthogonal and symplectic groups, the main body of which will be formulated in the next section. We have stated it here as a seed theorem, which will be needed

30

1. PARAMETERS

to dene the objects on which the others depend. All of the theorems will be established together. The proof will be based on a long induction argument, which is to be carried throughout the course of the volume. The implication for Theorem 1.4.1 is that the denitions it yields will ultimately have to be inductive. We remind the reader that our results will also be conditional on the stabilization of the discrete part of the twisted trace formula, a hypothesis we will state formally in 3.2. Suppose that N is xed, and that Theorem 1.4.1 holds with N replaced by any integer m N . What can we say about an arbitrary parameter ` r r in the set pN q GpN q ? We can certainly write (1.4.3) i i j pj j _ q .
iPI jPJ

If i belongs to I , we apply Theorem 1.4.1 to the simple generic factor r i P sim pmi q of i i b i . This gives a canonical datum pGi , si , i q in r Epmi q, which it will be convenient to denote by Hi . If j belongs to J , we simply set Gj GLpmj q. We thus obtain a connected reductive group Hk over F for any index k in either I or J , or equivalently, in the set tK u of orbits of the involution on K . Let L Hk be the Galois form of its L-group. We can then form the bre product (1.4.4) L p L Hk F q
kPtK u

of these groups over F . This is to be our substitute for the global Langlands group in our study of automorphic representations attached to . To make matters slightly more transparent, we have formulated it in algebraic form, as an extension of the pronite group F by a complex reductive group rather than an extension of WF by a compact topological group. For this reason, we shall work with the Galois forms of L-groups throughout much of the volume, rather than their Weil form. This leads to no diculty in the framework of orthogonal and symplectic groups we will study. If an index k i in (1.4.1) belongs to I , we have the standard embedding ` i i : L Hi L GLpmi q GLpmi , Cq F r that is part of the endoscopic datum Gi . If k j belongs to J , we dene a standard embedding ` r j : L Hj L GLp2mj q GLp2mj , Cq F by setting r j phj q phj t h1 q , j We then dene an L-embedding r : L SLp2, Cq
L

p hj P Hj GLpmj , Cq, P F .

GLpN q GLpN, Cq F

1.4. A SUBSTITUTE FOR GLOBAL PARAMETERS

31

as the direct sum r (1.4.5)

iPI

i pri

b i q
jPJ

j prj

b j q .

Our use of SLp2, Cq here in place of SU p2q is purely notational, and is in keeping with our construction of L as a complex proalgebraic group. We r are of course free to interpret the embedding also as an N -dimensional representation of L SLp2, Cq. With either interpretation, we shall be prir marily interested in the equivalence class of , as a GLpN, Cq conjugacy class ` of homomorphisms from L SLp2, Cq to either GLpN, Cq or L GLpN q . r r Suppose that G belongs to Esim pN q. We write pGq for the set of eler r ments P pN q such that factors through L G. By this, we mean that there exists an L-homomorphism r G : L SLp2, Cq L G such that r r G , ` where is the embedding of L G into L GpN q that is part of the twisted r endoscopic datum represented by G. Since and are to be regarded as r GLpN, Cq-conjugacy classes of homomorphisms, G is determined up to the p stabilizer in GLpN, Cq of its image, a group that contains G. The quotient of p r this group by G equals OutN pGq, the group of outer automorphisms of the endoscopic datum G, described in 1.2. It is a nite group, which is actually trivial unless G is an even orthogonal group, in which case it equals Z{2Z. p In the latter case though, there can be two G-orbits of homomorphisms G rG . It is for this reason that we write pGq in place of the r in the class of more familiar symbol pGq. r More generally, suppose that G belongs to the larger set Eell pN q, or even r r the full set EpN q of endoscopic data for GpN q. As a group over F , G equals a product G G

(1.4.6)

of groups G that can range over (quasisplit) connected orthogonal and symplectic groups and (split) general linear groups. We dene the set of parameters for G as the product r r pGq pG q,

r where pG q equals pG q if G is orthogonal or symplectic, and equals pG q if G is a general linear group. After a little reection, we see that r r an element in pGq can be identied with a pair p, G q, for a parameter r q and an L-embedding P pN r G : L SLp2, Cq L G

32

1. PARAMETERS

p that satises (1.4.6), and is dened as a G-orbit only up to the action of r OutN pGq. The projection r p, G q is not generally injective. This is in contrast to the injective embedding of r r pGq into pN q for simple G, which is an implicit aspect of our original denition. (It is also in contrast to the special case of elliptic parameters r r P ell pN q and elliptic endoscopic data G P Eell pN q, which we will describe in a moment.) However, we shall still sometimes allow ourselves to denote r elements in the more general sets pGq by when there is no danger of confusion. r Suppose that is any parameter in pN q. Let ` ` r p r r r S pN q S GpN q Cent Impq, GpN q and ` ` r p r r r r0 S pN q S pG0 pN q Cent Impq, G0 pN q Cent Impq, GLpN, Cq be the centralizers of the image ` r r Impq L SLp2, Cq p p p r r r r of in the respective components GpN q G0 pN q and G0 pN q. Then p r r0 S pN q is a reductive subgroup of G0 pN q, which acts simply transitively by r r r left or right translation on S pN q. Both S pN q and S 0 pN q are connected.

r r On the other hand, if G is a datum in EpN q and belongs to pGq, the centralizer ` r p S S pGq Cent ImpG q, G r of the image of G need not be connected. Its quotient
0 p S S pGq S {S ZpGqF

is a nite abelian group, which plays an essential role in the theory. Notice that there is a canonical element rG 1, 1 0 s 0 1 in S . Its image in S (which we denote also by s ) will be a part of the description of nontempered automorphic representations of G. The centralr r izers S pN q and S are in obvious analogy with the centralizers Sr pN q and Sr described at the end of 1.2. Their denition settles the question posed in 1.1 of how to introduce centralizers of parameters without recourse to the global Langlands group LF . We write ` r r ell pN q ell GpN q

1.4. A SUBSTITUTE FOR GLOBAL PARAMETERS

33

r for the subset of parameters P pN q such that the indexing set J is empty, and such that i 1 for each i P I . These objects are analogous to the self-dual representations r we called elliptic in 1.2. The discussion r of 1.2 applies here without change. It tells us that any P ell pN q has r r a unique source in one of the sets pGq. More precisely, let 2 pGq be the r r r r preimage of ell pN q in pGq, for any G P Eell pN q. The mapping from pGq r q then takes 2 pGq injectively onto a subset of ell pN q, which we r r to pN r r identify with 2 pGq, and ell pN q is then the disjoint union r r ell pN q 2 pGq
r GPEell pN q

of these subsets. We thus have parallel chains of parameter sets r r r sim pN q ell pN q pN q and r G P Eell pN q, r r r where sim pGq denotes the intersection of sim pN q with 2 pGq. Observe r 2 pGq is the subset of parameters P pGq such that the centralizer r that r S is nite, while sim pGq consists of those such that S equals the p minimal group ZpGqF . Suppose that 1 r r ell pN q. How do we determine the group G P Eell pN q such that r belongs to r lies in 2 pGq? To answer the question, we have to be able to write O S , where O i (1.4.7) r r r sim pGq 2 pGq pGqq,
iPIO

is the sum of those components of orthogonal type, S i

iPIS

is the sum of components of symplectic type, and I I,O I,S is the corresponding partition of the set I K of indices. Consider a general component i i b i r of the given . The irreducible cuspidal element i P sim pmi q has a central character i i of order 1 or 2, which we identify with a character on F . r It also gives rise to a datum Hi P Esim pmi q, according to Theorem 1.4.1. This p provides a complex, connected classical dual group Hi GLpmi , Cq. The ni dimensional representation i of SLp2, Cq determines the complex connected p classical group Ki GLpni , Cq that contains its image. By considering

34

1. PARAMETERS

p principal unipotent elements, for example, the reader can see that Ki is symplectic when ni is even, and orthogonal when ni is odd. The tensor product embedding p p Hi Ki GLpNi , Cq, Ni mi ni , p is then contained in a canonical classical group Gi GLpNi , Cq. In concrete p p p terms, Gi is symplectic if one of Hi or Ki is symplectic and the other is p i and Ki are both of the same type. This p orthogonal, and is orthogonal if H allows us to designate i as either orthogonal or symplectic, and thus gives rise to the required decomposition O S . As in 1.2, we set Ni , O, S, N
iPI,

r r r Then S lies in the subset 2 pGS q of ell pNS q, for the datum GS P Esim pNS q p S SppNS , Cq, while O lies in the subset 2 pGO q of r with dual group G r r p ell pNO q, for the datum GO P Esim pNO q with dual group GO SOpNO , Cq and character O pi qni i .
iPI,O iPI,O

r The original element therefore lies in 2 pGq, where G is the product r datum GO GS in Eell pN q. More generally, suppose that is a parameter (1.4.3) in the larger set r r pN q. Then can have several preimages in the various families pGq, rell pN q. How does one determine them explicitly? For a given datum GPE G GO GS , there are several ways to divide the components k among the two factors GO and GS . To describe the possibilities, it suces to determine r r whether a given belongs to the subset pGq of pN q attached to a xed r simple datum G P Esim pN q. p r The dual group G of G P Esim pN q is purely orthogonal or symplectic. The parameter comes with a partition of its index set K into subsets I , _ r r J and J . If j belongs to J , the corresponding component j pj j _ q r of the embedding takes values in a factor ` L Mj L GLpNj q GLpNj q looooooooooooooomooooooooooooooon
j

of a Levi subgroup M of As we have just seen (at least in the special r 2 pGq), the complementary set I has its own partition into case that P subsets I,O and I,S of orthogonal and symplectic type. It also indexes characters i of order 1 or 2. If an index i P I has even multiplicity 1 r r i 2 i , the component i i of takes values in another factor ` L Mi L loooooooooooooomoooooooooooooon GLpNi q GLpNi q
1 i

L G.

1.4. A SUBSTITUTE FOR GLOBAL PARAMETERS

35

of L M , whether i belongs to I,O or I,S . The indices j P J and i P I with i even therefore impose no constraints. However, if i P I with i odd, there is a supplementary copy of i to content with. In this case i p and G must be of the same type, either both symplectic or both orthogonal. p Moreover, if G is orthogonal, these factors also impose a constraint on the character G that goes with G. It must satisfy the identity pi q i . pi q i ni
iPI,O iPI,O

The last two conditions are both necessary and sucient for to belong to r the subset pGq. It is also easy to describe the centralizers S S pGq. Suppose that r G P Esim pN q, and that is a parameter (1.4.3) that lies in the subset r r pGq of . Every index j P J contributes a factor GLp j , Cq to S . The complementary indexing set has a separate partition ` I I pGq I pGq,
` p where I pGq (resp. I pGq) is the set of indices i P I such that i and G ` are of the same (resp. dierent) type. In other words, I pGq I,O and p I pGq I,S if G is orthogonal, while I ` pGq I,S and I pGq I,O

p if G is symplectic. It follows from what we have just observed that an index i P I pGq has even multiplicity i . It contributes a factor Spp i , Cq ` to S . Each index i P I pGq contributes a factor Op i , Cq to S , with the stipulation that a product (with multiplicities) of the determinants of these factors be 1. Let ` Op i , Cq
` iPI pGq

be the kernel of the character ` ` pGq : gi pdet gi qNi ,

i i

` gi P Op i , Cq, i P I pGq.

The centralizer S is then given by ` (1.4.8) S Op i , Cq

` iPI pGq

Spp i , Cq GLp j , Cq .
jPJ

iPI pGq

r Suppose in particular that belongs to 2 pGq. Then

` I pGq I t1, . . . , ru,

and each (1.4.9)

equals 1. In this case, the centralizer specializes to # r ` pZ{2Zqr , if each Ni is even, S Op1q pZ{2Zqr1 , otherwise. i1

36

1. PARAMETERS

Similar constructions apply to (ordinary) endoscopic data G1 P EpGq, or for that matter, iterated data G2 P EpG1 q, G3 P EpG2 q, and so on. As a group over F , G1 is again a product of elementary groups G1 for which the r sets pG1 q have been dened. We dene pG1 q to be the corresponding r product of these sets. A parameter 1 G1 in pG1 q can be identied with 1 q, for a parameter in pGq, and an L-embedding r r a pair p, r r 1 G1 : L SLp2, Cq that satises (1.4.10) r r 1 1 G , p and is dened as a G1 -orbit only up to the action of the nite group r r OutN pG1 q OutG pG1 q OutN pGq. Here, 1 : L G1 L G is the L-embedding attached to G1 . Notions introduced r earlier for have obvious meaning for parameters 1 P pG1 q here. For example, we have the centralizer ` p S1 S1 pG1 q Cent Imp 1 q, G and its abelian quotient
0 p S1 S1 pG1 q S1 {S1 ZpG1 qF , L

G1

for any such 1 . Any pair pG1 , 1 q, gives rise to a second pair p, sq, r G1 P EpGq, 1 P pG1 q, r P pGq, s P S ,

r where G is the parameter in pGq attached to 1 , and s is the image p of the semisimple element s1 P G1 that is part of the endoscopic datum p in G 1 . Conversely, suppose that is any parameter in pGq, and that s is a r G p 1 be the connected centralizer of s in G, p semisimple element in S . Let G p and let s1 be the preimage of s in G1 . The product ` p r G 1 G1 G L SLp2, Cq r p of G1 with the image of G is an L-subgroup of L G, for which the identity 1 is an L-homomorphism. We take G1 to be a quasisplit group embedding p for which G1 , with the L-action induced by G 1 , is a dual group. In the present situation, there is a natural way to identify G 1 with L G1 . The triplet pG1 , s1 , 1 q represented by G1 is then an endoscopic datum for G1 , in the r restricted sense of 1.2. Since s lies in the centralizer of the image of G , rG factors through L G1 . We obtain an L-embedding 1 of L SLp2, Cq r r into L G1 that satises (1.4.10), and hence an element 1 P pG1 q. The pair 1 , 1 q of the original sort. p, sq thus gives rise to a pair pG

1.4. A SUBSTITUTE FOR GLOBAL PARAMETERS

37

The correspondence (1.4.11) pG1 , 1 q p, sq is a general phenomenon. It applies to arbitrary endoscopic data, at least in the context of Langlands parameters. In particular, it has an obvious variant ` r for our other primary case in which pG, G1 q is replaced by GpN q, G . This leads to a more systematic way to view the discussion of 1.2. In general, the correspondence reduces many questions on the transfer of characters to a study of groups S . It will be part of the foundations we are calling the standard model in Chapter 4. The discussion of this section applies also to our supplementary case, r r that of a bitorsor G (1.2.4) for an even orthogonal group G P Esim pN q. We r r r r r write pGq pGq for the subset of -xed elements in pGq. These are r r the parameters in pGq such that the AutN pGq-orbit of L-embeddings r G : L SLp2, Cq L G r p contains only one G-orbit, or equivalently, such that G factors through the L-embedding r r 1 : L G1 L G r r of some twisted endoscopic datum G1 P Eell pGq. They comprise the minimal set in a chain r r r pGq pGq pN q pN q of subsets of our original family of parameters pN q for GLpN q. Observe that the subset r r r 2 pGq 2 pGq X pGq r r of pGq consists of those parameters (1.4.1) that lie in 2 pGq, and for which one of the irreducible ranks Ni is odd. In particular, the subset r r r r r sim pGq sim pGq X pGq sim pGq X 2 pGq r r of simple elements in pGq is empty, since N is even. For any P pGq, the centralizer ` p r r r r S S pGq Cent ImpG q, G is a bitorsor under S , whose quotient r r r r0 p S S pGq S {S ZpGqF is a bitorsor under the quotient S of S . It gives rise to an analogue of the r r correspondence (1.4.11), with pG, G1 q in place of pG, G1 q. We have so far devoted this section to global parameters. We have shown that their study is parallel to that of the nite dimensional representations in 1.2. We have also tried to demonstrate that objects attached to any can be computed explicitly, even if the details may be somewhat complicated. The details themselves need not be taken too seriously. When we resume this discussion in Chapter 4, we shall treat the relations among our global parameters in a more systematic fashion.

38

1. PARAMETERS

We shall now discuss the local parameters attached to a completion Fv of F . These are simpler, since they really do represent homomorphisms dened on the group LFv SU p2q. In particular, the remarks in 1.3 pertaining to the chain (1.3.10) apply here. Observe that the local analogue of the assertion of Theorem 1.4.1 is r elementary. For we have seen in 1.2 that any parameter v P sim,v pN q, r sim,v pN q of generic parameters and in particular, any v v in the subset r in sim,v pN q, factors through a unique local endoscopic datum Gv in the set ` r r Eell,v pN q Eell Gv pN q . Moreover, this datum is simple. In these more concrete terms, all of the global discussion above carries over to the completion Fv of F . In fact, it extends without change to the larger parameter set ( r ` pN q v P ` pN q : _ v
v v v

of self-dual, N -dimensional representations of LFv SU p2q that are not necessarily unitary. We thus attach chains of parameter sets r r r r (1.4.12) sim pGv q 2 pGv q pGv q ` pGv q r to local endoscopic data Gv P Ev pN q, following the local analogues of the global notation (1.4.7). They consist of L-homomorphisms v : LFv SU p2q
L

Gv ,

which can be composed with the endoscopic embedding ` v Gv : L Gv L GpN q . r r The resulting mapping from ` pGv q to ` pN q is injective if Gv is simple, v r ` pGv q of ` pGv q. This allows us in each r or if v belongs to the subset 2 of the two cases to identify the domain of the mapping with a subset of r ` pN q. For example, if Gv is simple, any parameter set in (1.4.12) equals v r r the intersection of ` pGv q with the relevant set of parameters for Gv pN q in the chain r r r r sim,v pN q ell,v pN q v pN q ` pN q. v r ` pGv q, the local centralizer For any parameter v in ` p Sv Sv pGv q Cent Impv q, Gv , and its abelian quotient
0 p Sv Sv pGv q Sv {Sv ZpGv qv ,

r are dened in the usual way. We also have the set ` pG1 q attached to any v 1 P EpG q, as well as the local analogue of the correspondence (1.4.11). We Gv v shall apply all of these local constructions at will, sometimes with the local eld being F itself (and the notation modied accordingly) rather than the completion Fv here of the global eld F .

1.4. A SUBSTITUTE FOR GLOBAL PARAMETERS

39

Finally, we will need to dene a localization mapping v from the r r global set pGq attached to any G P Esim pN q to the associated local set r ` pGv q. It would not be hard to formulate the localization as a mapping r r from pN q to ` pN q, or if we wanted to follow the general remarks for the v hypothetical group LF at the end of the last section, as a mapping between the larger sets pN q and ` pN q. To complete the denition, however, we v r r would need to know that it takes the subset pGq of pN q to the subset ` pG q of pN q. This property is not elementary. It is a consequence of r rv v a second seed theorem, which we state here as a complement to Theorem 1.4.1, but which like Theorem 1.4.1, will have to be proved inductively at the same time as our other theorems. r Theorem 1.4.2. Suppose that P sim pN q is simple generic, as in Theorem 1.4.1. Then the localization v of at any v, a priori an element in r r r the subset v pN q of generic parameters in ` pN q, lies in the subset pG,v q v r r of v pN q attached to the localization G,v of the global datum G P Esim pN q of Theorem 1.4.1. The theorem asserts that the N -dimensional representation v of LFv , which is attached by the local Langlands correspondence to the v-component of the cuspidal automorphic representation of GLpN q given by , factors through the local endoscopic embedding ,v :
L

G,v

GpN qv .

It allows us to identify v with an L-homomorphism v : LFv

L

G,v .

Here G is the endoscopic datum attached to by Theorem 1.4.1, and v is r determined as a mapping into L G,v up to the action of the group AutN pG q p . Like Theorem 1.4.1, this theorem will be proved by a long induction on G argument that includes the proof of our other results. Suppose that Theorems 1.4.1 and 1.4.2 hold with N replaced by any integer m N . Let us consider the group L attached to a general paramr eter (1.4.1) in pN q, and its relation to the local Langlands group LFv at L H is one of the factors (1.4.4) of L , H represents either a simple v. If k k rsim pmk q, or a general linear group GLpmk q. In the endoscopic datum in E rst case, Theorem 1.4.2 gives a conjugacy class of L-homomorphisms LFv
LH k

(1.4.13)

v F , F

r p which is determined up to the action of Autmk pHk q on Hk . In the second case, we obtain a similar assertion from the local Langlands correspondence

40

1. PARAMETERS

` r for GLpmk q, with Autmk pHk q equal to Int GLpmk , Cq . The bre product (1.4.4) then yields a conjugacy class of L-homomorphisms LFv L v F F

(1.4.14)

which is determined up to the L-action of the group r r (1.4.15) AutpL q Autmk pHk q
k

on L . r r Suppose that belongs to pGq, for some G P Esim pN q, or indeed, for r q. It then follows from this discussion that we any G in the general set EpN can identify the localization of with an L-homomorphism v : LFv SU p2q
L

Gv ,

which ts into a larger commutative diagram of L-homomorphisms

v

(1.4.16)

LFv p2q SU

LG

F v v

G L SLp2, Cq

LG

F .

The left hand vertical arrow is given by the mapping of LFv into L in (1.4.14) and the usual embedding of SU p2q into SLp2, Cq. Since v is essenr tially the restriction of the global embedding G to the image of LFv SU p2q, the global centralizer S is contained in Sv . It follows that there is a canonical mapping x xv , x P S , of S into Sv . 1.5. Statement of three main theorems We shall now state our main theorems. They apply to the quasisplit orthogonal and symplectic groups we have introduced, and provide what amounts to a classication of the representations of these groups in terms of representations of general linear groups. They will be conditional on the stabilization of the twisted trace formula, an assumption we shall state formally when we come to discuss the relevant part of the trace formula in Chapter 3. Recall that the groups under consideration represent simple endoscopic r r data G P Esim pN q, for the torsor GpN q attached to a general linear group

1.5. STATEMENT OF THREE MAIN THEOREMS

41

GpN q GLpN q. By presenting them in standard form in 1.2, we have imr plicitly attached extra structure to these groups. Each G P Esim pN q comes with a scheme structure over the ring of integers oF . It also comes with a standard Borel subgroup B, with maximal torus T , and a standard splitr ting for pT, Bq. We have also identied the group OutN pGq of outer automorphisms of G over F with the subgroup of F -automorphisms of G that preserve the given splitting. This group seems innocuous enough, being of order 1 or 2, but it does play an essential role. The rst theorem concerns the case that F is local. It classies representations of GpF q in terms of a construction dened by the endoscopic transfer of characters. We shall postpone the formal description of this construction until 2.1, as part of our general discussion of local endoscopic transfer. In the meantime, we shall be content to give a somewhat less precise statement of the theorem. We are assuming for the moment that F is local. We x a maximal compact subgroup KF of the group GF , which we take to be special if F is p-adic, and hyperspecial if G is unramied over F . The oF -scheme structure on G actually leads to a natural choice of KF . For example, if F is p-adic and G is split, we can take KF GpoF q. On the other hand, if F is padic, but G SOpN, G q is quasisplit but not split (so that N 2n is even), KF is a little more complicated. (See the diagrams in [Ti, 1.16].) In any case, given KF , we write HpGq for the corresponding Hecke algebra of smooth, left and right KF -nite functions of compact support on GpF q. r r We also write HpGq for the subalgebra of OutN pGq-invariant functions in HpGq. Because our results are ultimately tied to GLpN q, they will apply r to the representation theory of HpGq, which is slightly weaker than that of HpGq. r r We write pGq for the set of orbits of OutN pGq in pGq, which we r recall is the set of irreducible representations of GpF q. Then pGq equals r pGq unless G is an even orthogonal group SOp2n, G q, in which case pGq r contains orbits of order 2 as well as of order 1. Any element P pGq gives rise to a well dened character ` r fG pq tr pf q , f P HpGq, r on HpGq. Similar notation applies to any subset of pGq, such as for example the set unit pGq of unitary representations in pGq. We shall state the local r theorem in terms of packets that are sets over unit pGq. By a set over S, or an S-set, or even an S-packet, we mean simply a set S1 with a bration S1 S over S. Equivalently, S1 is a set that can be represented as a disjoint union r of subsets of S. Any function on S, such as the character fG pq on HpGq in r case S equals pGq, will be identied with its pullback to a function on S1 .

42

1. PARAMETERS

The order m1 : S N Y t0, 8u of the bres in S1 represents a multiplicity function, which makes S1 into a multiset on S in the sense of [Z, p. 169]. If S1 is multiplicity free, in that every element in S has multiplicity at most 1, it is just a subset of S. r Theorem 1.5.1. Assume that F is local and that G P Esim pN q. r r (a) For any local parameter P pGq, there is a nite set over r unit pGq, constructed from by endoscopic transfer, and equipped with a canonical mapping r x, y, P , r p from into the group S of characters on S such that x, y 1 if G and are unramied (relative to KF ). r r (b) If belongs to the subset bdd pGq of parameters in pGq that r are trivial on the factor SU p2q, the elements in are tempered and mulr p tiplicity free, and the corresponding mapping from to S is injective. r r Moreover, every element in temp pGq belongs to exactly one packet . Fir to S is bijective. p nally, if F is nonarchimedean, the mapping from r r The nite subsets of temp pGq in (b) represent the tempered Lr packets for GpF q. They are composed of (OutN pGq-orbits of) tempered p representations, which are parametrized by characters in S . Since the r theorem asserts that temp pGq is a disjoint union of these packets, it can be regarded as a classication of the irreducible, tempered representations of GpF q. r The more general sets of (a) represent local factors of automorphic r representations. They can contain (OutN pGq-orbits of) nontempered repp resentations, and typically only bre over S . It seems likely that these r packets are also multiplicity free (that is, subsets of unit pGq rather than multi-subsets). For nonarchimedean F , Moeglin has recently established r this fact [M4], using some of the properties of the tempered L-packets we will establish in the course of proving the theorem. However, the genr p r eral mapping from to S is still not injective, so the elements in are not parametrized by characters on S . One could rearrange the denition r simply by combining all the elements in that map to a given character. r With this equivalent formulation, the elements in would no longer be p irreducible, but they would map injectively to S . This was the point of view in the announcement of [A18, 30]. The objects of the theorem will be dened in a canonical way. We shall state the precise form of the construction in the next chapter. For archimedean F , the resulting strong form of the theorem includes special cases of results of Shelstad [S2] and [S3][S6] for tempered representations,

1.5. STATEMENT OF THREE MAIN THEOREMS

43

and of Adams, Barbasch and Vogan [ABV] for nontempered representations. In case F is nonarchimedean, part (b) of the theorem gives the local r Langlands correspondence for GpF q in case OutN pGq 1, and a slightly r weakened form of the correspondence if OutN pGq 1. In all cases, the theorem yields local Rankin-Selberg L-functions and -factors for representations of general quasisplit orthogonal and symplectic groups. For any pair of representations r i P temp pGi q, r Gi P Esim pNi q, i 1, 2,

we have only to take the corresponding pair of parameters ( r r i P bdd pGi q : i P i , i 1, 2 for the associated local packets. These objects can of course also be regarded as local parameters for general linear groups. They allow us to dene the local L-functions Lps, 1 2 q Lps, 1 2 q and -factors ps, 1 2 , F q ps, 1 2 , F q in terms of what is already known for general linear groups. We note that similar denitions apply if one of the groups Gi is replaced by a general linear group. In this case, they have been studied in terms of the LanglandsShahidi method. The remaining two theorems are global. However, their statement requires an extension of the local construction. We could easily formulate the r objects of Theorem 1.5.1 for parameters in the larger set ` pGq. However, it is instructive to introduce an intermediate set r r r pGq ` pGq ` pGq, unit which serves the purpose and is perhaps more natural. We rst introduce an analogous intermediate set pN q ` pN q ` pN q unit for GLpN q. Recall that the local set ` pN q consists of N -dimensional representations
1 1

r r ,

i P ` pNi q sim

of the locally compact group LF SU p2q. Any such is an irreducible representation of the Levi subgroup GLpN1 , F q 1 GLpNr , F q
r

of GLpN, F q, which is unitary modulo the centre. We write ` pN q for unit the set of P ` pN q such that the associated induced representation of GLpN, F q is irreducible and unitary. This set can be described explicitly in terms of the parametrization of unitary representations in [V2] and [Tad1].

44

1. PARAMETERS

It is the natural domain for the set of local factors of automorphic representations in the discrete spectrum of GLpN q. The associated set for G is then dened as the intersection r r ` pGq ` pGq X ` pN q. unit unit r Suppose that P ` pGq. Any such parameter gives rise to an unit r OutN pGq-orbit of standard parabolic subgroups P M N of G, a parameter r M P pM q and a point in the open chamber of P in the real vector space a XpM q b R. M F These objects are characterized by the property that equals the comr position of the twist M, P ` pM q of M by with the L-embedding L M L G. Recall that there is a canonical homomorphism ` HM : M pF q aM Hom XpM q , R . F The twist M, is the product of M with the central Langlands parameter that is dual to the unramied quasicharacter : m epHM pmqq , m P M pF q.

The Levi subgroup M G attached to is a product of several genr eral linear groups with a group G P Esim pN q, for some N N . The r N pP q of P in OutN pGq is isomorphic to OutN pG q. The r r stabilizer Out r obvious variant of Theorem 1.5.1(a) attaches objects M and x, M y to r r the parameter M . An element M P M is then an OutN pP q-orbit of representations P unit pM q. Let M, be the corresponding orbit of repr resentations M b , and let IP pM, q be the associated OutN pGq-orbit of induced representations. With this understanding, we dene the packet of as the family ( r r (1.5.1) IP pM, q : M P M , r a nite set that is bijective with M . We are not assuming at this point that the induced representations in question are irreducible or unitary. In r r other words, an element P is an OutN pGq-orbit of (possibly reducible, possibly nonunitary) representations of GpF q. However, the numbers tr pf q r and the operators pf q are nonetheless well dened for functions f P HpGq. x The G-centralizer S S pGq of the given is contained in M . It follows that S SM , from which we deduce that S SM . We can therefore dene the character (1.5.2) s xs, y xsM , M y, r P ,

for any element s sM in the group S SM . We thus obtain the local r objects of Theorem 1.5.1(a) for any parameter P ` pGq. The assertion unit (a) for these objects, which is to say the canonical construction by endoscopic

1.5. STATEMENT OF THREE MAIN THEOREMS

45

transfer that we will make precise in 2.2, will follow immediately from the case of M by analytic continuation. r Suppose now that the eld F is global. Our group G P Esim pN q then represents a global endoscopic datum. The rst global theorem is the central result. It gives a decomposition of the automorphic discrete spectrum of G r in terms of global parameters P 2 pGq and the local objects of Theorem 1.5.1(a) (or rather their extensions (1.5.1) and (1.5.2)). It is best stated in terms of the global Hecke algebra HpGq on GpAq with respect to the maximal compact subgroup K Kv .
v

Recall that HpGq is the space of linear combinations of products fv , fv P HpGv q,

v

such that fv is the characteristic function of Kv for almost all v. In other words, HpGq is the restricted tensor product of the local Hecke algebras r HpGv q. Let us write HpGq for the restricted tensor product of the local r symmetric Hecke algebras HpGv q. This is the subspace of functions in HpGq r that are invariant under each of the groups OutN pGv q. For any function f r in HpGq, and any admissible representation of GpAq of the form v ,
v

r the operator pf q depends only on the OutN pGv q-orbit of any component r v . The point is to describe the discrete spectrum as an HpGq-module. We are assuming the seed Theorems 1.4.1 and 1.4.2. The second of these r implies that the localization mapping v takes the global set pGq to r r ` pGv q of pN q. The local theorem we have just stated then the subset unit r attaches a local packet v to and v. We can thus attach a global packet ! ) r r (1.5.3) v : v P v , x, v y 1 for almost all v
v

r of (orbits of) representations of GpAq to any P pGq. Any representation v in the global packet determines a character (1.5.4) xx, y xxv , v y, x P S ,
v

on the global quotient S . r Theorem 1.5.2. Assume that F is global and that G P Esim pN q. Then r there is an HpGq-module isomorphism ` (1.5.5) L2 GpF qzGpAq m , disc
P2 pGq P p q

46

1. PARAMETERS

where m equals 1 or 2, while : S t1u is a linear character dened explicitly in terms of symplectic -factors, and r r p q is the subset of representations in the global packet such that the character x, y on S equals . We need to supplement the statement of Theorem 1.5.2 with a description of the integer m and the character . The rst of these is easy enough. The parameter comes with the L-embedding r G : L SLp2, Cq L G, r p determined up to the action of the group AutN pGq by conjugation on G. We r p r dene m for any P pGq to be the number of G-orbits in the AutN pGqrG . The theorem applies to a parameter orbit of 1 r r 2 pGq. In this case, it is easy to check that m equals 1, unless N is in p even, G equals SOpN, Cq, and the rank Ni of each of the components i of is also even, in which case m equals 2. The sign character is more interesting. But rst we make an observar tion on general L-functions. Suppose that P pN q is an arbitrary global parameter, and that r is an arbitrary nite dimensional representation of L , subject only to the condition that its equivalence class is stable under r the group AutpL q. Then r pulls back to a well dened representation rv of LFv , for any v. We can therefore dene the global L-function Lps, rq Lps, rv q
v

by an Euler product that converges for the real part of s large. Of course we do not know in general that it has analytic continuation and functional equation. We can still dene the global -factor as a nite product ps, r, F q ps, rv , Fv q,
v

where F is a nontrivial additive character on A{F . We cannot say that this function is independent of F in general. But if r is symplectic, by which we mean that it takes values in the symplectic subgroup of the underlying general linear group, the value of the local factor ps, rv , Fv q at s 1 is 2 known to equal `1 or 1, and to be independent of Fv . We therefore have a global sign ` ` 1 , r 1 , r, F 1 2 2 in this case, which is independent of F . r We shall dene if is a general parameter in pGq. We rst dene a representation : S L SLp2, Cq GLpq g

1.5. STATEMENT OF THREE MAIN THEOREMS

47

p on the Lie algebra g of G by setting ` r ps, g, hq Ad s G pg hq ,

s P S , pg, hq P L SLp2, Cq,

where Ad AdG is the adjoint representation of L G. Notice that since it is invariant under the Killing form on g, this representation is orthogonal, and hence self-dual. Let p b b q

be its decomposition into irreducible representations , and of the respective groups S , L and SLp2, Cq. We then dene 1 ` det psq , s P S , (1.5.6) pxq

where x is the image of s in S , and 1 denotes the product over those indices with symplectic and `1 (1.5.7) 2 , 1. (The denition (1.5.6) is equivalent to [A18, (30.16)], and is conjecturally equivalent to the slightly dierent formulations [A8, (8.4)] and [A9, (4.5)] in the earlier papers in which the general multiplicity formula (1.5.5) was postulated.) Theorem 1.5.2 thus asserts that there is an intimate relationship between the automorphic discrete spectrum of G and symplectic root numbers. Before we state the second global theorem, we recall a property of certain Rankin-Selberg L-functions. Suppose that P Acusp pN q is a cuspidal automorphic representation of GLpN q. Then there are two formal products (1.5.8) and (1.5.9) ps, , F q ps, , S 2 , F q ps, , 2 , F q, where S 2 (resp. 2 ) is the representation of GLpN, Cq on the space of symmetric (resp. skew-symmetric) pN N q-complex matrices. Each of S 2 and p p 2 is associated with a Siegel maximal parabolic subgroup P` G` , for a r split, simple endoscopic group G` P Esim pN` q with N` 2N . The dual p ` equals SppN` , Cq or SOpN` , Cq, according to which of the two group G representations S 2 are 2 we are considering. In each case, the represenx tation is given by the adjoint action of the Levi subgroup M` GLpN, Cq on the Lie algebra of the unipotent radical. Viewed in this way, one sees that the L-functions on the right hand side of (1.5.8) are among the cases of the Langlands-Shahidi method treated in [Sha4]. In both cases, the local L-functions and -factors have been constructed so that the formal products (1.5.8) and (1.5.9) become actual products, and so that the global L-functions have analytic continuation with functional equation (1.3.5). Lps, q Lps, , S 2 q Lps, , 2 q

48

1. PARAMETERS

r Suppose now that P sim pN q is cuspidal generic. In other words, is given by a representation P Acusp pN q that is self dual. Theorem 1.4.1 asserts that belongs to the subset r r r sim pGq sim pN q X pGq, r for a unique G P Esim pN q. We will need to understand how G is related to L-functions and -factors. The Rankin-Selberg L-function Lps, q Lps, q Lps, _ q has a pole of order 1 at s 1. It is known that neither of the corresponding factors Lps, , S 2 q and Lps, , 2 q on the right hand side of (1.5.8) has a zero at s 1. It follows that exactly one of them has a pole at s 1 (which will be of order 1). This motivates the rst assertion (a) of the second global theorem, which we can now state. Theorem 1.5.3. Assume that F is global. r r p (a) Suppose that G P Esim pN q and that belongs to sim pGq. Then G 2 q has is orthogonal if and only if the symmetric square L-function Lps, , S p a pole at s 1, while G is symplectic if and only if the skew-symmetric 2 q has a pole at s 1. L-function Lps, , r (b) Suppose that for i 1, 2, i belongs to sim pGi q, for simple endor scopic data Gi Esim pNi q. Then the corresponding Rankin-Selberg -factor satises ` 1 , 1 2 1, 2 p p if G1 and G2 are either both orthogonal or both symplectic. The assertion (a) of this theorem was suggested by the corresponding property for L-functions of representations of Galois groups and Weil groups. It seems to have been rst conjectured for automorphic representations by Jacquet and Shalika, at the early stages of their work on Rankin-Selberg L-functions. The assertion (b) is suggested by the corresponding property for -factors of orthogonal representations of Galois groups [FQ] and Weil groups [D1]. It was conjectured for automorphic representations in [A8, 8] and [A9, 4]. Observe that (a) gives an independent characterization of the group G of Theorem 1.4.1. It has been established for representations with Whittaker models by Cogdell, Kim, Piatetskii-Shapiro and Shahidi [CKPS1], [CKPS2]. In fact, apart from the uniqueness assertion of the rst seed Theorem 1.4.1, the statements of Theorem 1.4.1 and 1.4.2, together with that of (a) above, follow from the combined results of [CKPS2] and [GRS]. However, we will not be able to use these results. From our standpoint, the essence of the two seed theorems will be in two further statements that we have temporarily suppressed in the interests of simplicity. These give further characterizations of G and its completions Gv in terms of harmonic

1.5. STATEMENT OF THREE MAIN THEOREMS

49

analysis, and will be what really drives the general argument. They will be included in the assertions of Theorems 2.2.1 and 4.1.2, which we will come to in due course. The broader signicance of (b) will become clear later. This assertion has been proved in special cases by Lapid [Lap]. Its role for us will be rather similar to that of (a). Both assertions will be established in the course of proving the other theorems. In fact, they are both an inextricable part of the general induction argument by which all of the theorems will eventually be proved.

CHAPTER 2

Local Transfer
2.1. Langlands-Shelstad-Kottwitz transfer We now take the eld F to be local, a condition that will remain in place throughout Chapter 2. In the rst section, we shall review the endoscopic transfer of functions. This is based on the transfer factors of Langlands and Shelstad, and the twisted transfer factors of Kottwitz and Shelstad. It will serve as the foundation for the local classication we stated in 1.5. r Consider our GLpN q-coset GpN q. The twisted transfer factors of Kotr twitz and Shelstad for GpN q depend on a choice of F -rational automorphism of GLpN q within the given inner class. For this object, we choose the aur tomorphism pN q introduced in 1.2 that xes the standard splitting of ` r r r GLpN q. We then allow GpN q to stand for the pair G0 pN q, pN q , as well as the connected component r r G0 pN q pN q GLpN q pN q, r obtained from the pair. We recall that pN q lies in the inner class of the original automorphism pN q, and therefore gives the same coset in the r semidirect product G` pN q. The datum ` ` r r r r r GpN q G0 pN q, pN q G0 pN q, pN q, 1 is a special case of a general triplet pG0 , , q, where G0 is a connected reductive group over F , is an automorphism of G over F , and is a character on G0 pF q. The group generated by need not be nite. Its semidirect product with G0 therefore need not be an algebraic group. However, we can still form the connected variety G G0 over F , equipped with the two-sided G0 -action ` x1 py qx2 x1 ypx2 q , x P G0 ,

of G0 on G. The restriction of this action to the diagonal image of G0 can evidently be identied with the familiar action of G0 on itself by r conjugation. As in the case of GpN q, we shall also write G pG0 , , q.
51

52

2. LOCAL TRANSFER

In other words, we allow G to represent the underlying triplet as well as the associate G0 -bitorsor over F . For the moment, let us take G to be a general triplet over F . We can then form the general local Hecke module HpGq of G. It consists of the smooth, compactly supported functions on GpF q that are left and right Knite, relative to the two-sided action on GpF q of a suitable (xed) maximal compact subgroup K of G0 pF q. A semisimple element P G will be called strongly regular if its G0 -centralizer ( G _ pG0 q x P G0 : x1 x is an abelian group, whose group G pF q of rational points lies in the kernel of the character on G0 pF q. For any such , and any function f P HpGq, we form the invariant orbital integral 1 fG pq |Dpq| 2 f px1 xqpxqdx.
G pF qzG0 pF q

It is normalized here by the usual Weyl discriminant ` Dpq det p1 Adpqg0 {g , where g0 and g0 denote the Lie algebras of G0 and G0 respectively. We regard fG as a function on the set of strongly regular points, and write ( IpGq fG : f P HpGq for the G0 -invariant Hecke space of such functions. The functions in IpGq also have a spectral interpretation. Suppose that is a unitary extension to GpF q of an irreducible unitary representation p 0 , V q of G0 pF q. In other words, is a function from GpF q to the space of unitary operators on V such that (2.1.1) px1 xx2 q 0 px1 qpxq 0 px2 qpx2 q, x1 , x2 P G0 pF q. The set of such extensions of 0 is a U p1q-torsor, on which the character ` tr pf q tr f pxqpxqdx , f P HpGq,
GpF q

is equivariant. We set ` fG pq tr pf q . It can then be shown that either of the two functions tfG pqu or tfG pqu attached to f determines the other. We can therefore regard fG as a function of either or . The spectral interpretation is in some ways preferable. This is because the trace Paley-Wiener theorem ([BDK], [CD], [Ro1], [DM]) leads to a simple description of IpGq as a space of functions of . In general, an endoscopic datum G1 for G represents a 4-tuple 1 , G 1 , s1 , 1 q. According to Langlands, Shelstad and Kottwitz [LS1], [KS], pG one can attach a transfer factor to G1 . Recall that any comes with r some auxiliary data, namely a suitable central extension G1 G1 over F r1 : G 1 L G1 , and is then determined up to and an admissible L-embedding

2.1. LANGLANDS-SHELSTAD-KOTTWITZ TRANSFER

53

a complex multiplicative constant of absolute value 1. (See [KS, 2.2]. We shall review these notions later in the global context of 3.2.) The transfer factor is a function p 1 , q, where 1 is a strongly G-regular stable conjur gacy class in G1 pF q, and is a strongly regular orbit in GpF q under the 0 pF q by conjugation. It serves as the kernel function for the action of G transfer mapping, which sends functions f P HpGq to functions r G1 f 1 p 1 q f p 1 q p 1 , qfG pq

of

1.

We recall that

has an equivariance property r r f 1 pz 1 1 q 1 pz 1 q1 f 1 p 1 q, z 1 P C 1 pF q,

f1

r r where C 1 is the kernel of the projection G1 G1 , and 1 is a character on r r r 1 pF q that depends on the choice of 1 . C Suppose for example that G G0 is just a quasisplit group (with trivial r r character ), and that G1 G. If we set G1 equal to G and 1 equal to r the identity embedding of G 1 L G into L G1 L G, becomes a constant, which we can take to be 1. In this case, the transfer f G pq fG pq, f P HpGq,

of f is the stable orbital integral over a strongly regular, stable conjugacy class . It is a sum of invariant orbital integrals over the nite set of GF conjugacy classes in . We write SpGq for the space of functions of obtained in this way. More generally, if XG is a closed subgroup of GpF q, and is a character on XG , we write SpG, q for the associated space of 1 -equivariant functions G f pq f G pzqpzqdz, f P HpGq,
XG

of . A -equivariant linear form S on HpGq is called stable if its value at G f depends only on f . If this is so, we can identify S with the linear form (2.1.2) p G Spf q Spf q, f P HpGq,

` r r r on SpG, q. These conventions can then be applied with G1 , C 1 pF q, 1 in place of pG, XG , q. The LSK (Langlands-Shelstad-Kottwitz) conjecture asserts that for any G and G1 , any associated transfer factor , and any f P HpGq, the function r r f 1 belongs to SpG1 , 1 q. The fundamental lemma is a variant of this, which applies to the case that G is unramied. (Among other things [W6, 4.4], unramied means that F is p-adic, and that G0 is quasisplit and split over an unramied extension of F .) The fundamental lemma asserts that if f is the characteristic function of a product K, where K is a -stable, hyperspecial maximal compact subgroup of G0 pF q, then f 1 can be taken to be the image r r in SpG1 , 1 q of the characteristic function of a hyperspecial maximal compact

54

2. LOCAL TRANSFER

r subgroup of G1 pF q. Both of these longstanding conjectures have now been resolved. We include a couple of brief remarks about the proofs, some of which are quite recent. If F is archimedean, the results are due to Shelstad. In the case G G0 , she established the transfer theorem in terms of the somewhat ad hoc transfer factors she introduced for the purpose [S2]. These earlier transfer factors for real groups were the precursors of the systematic transfer factors for both real and p-adic groups in [LS1]. In a second paper, Langlands and Shelstad showed that the two transfer factors were in fact the same for real groups, by an indirect argument that used Shelstads original proof of the transfer theorem [LS2, Theorem 2.6.A]. Shelstad has recently expanded her earlier results [S3][S5], establishing the transfer conjecture (again for G G0 ) directly in terms of the transfer factors of [LS1]. She has recently completed a proof of the general archimedean transfer [S6], using the explicit specialization to real groups of the twisted transfer factors of [KS]. (See also [Re].) For nonarchimedean F , the fundamental lemma has long been a serious obstacle. Its recent proof by Ngo [N] was a breakthrough, which has now opened the way for progress on several fronts. The general proof follows the special case of G U pnq treated earlier by Laumon and Ngo [LN], and is based on the new geometric ideas that were introduced by Goresky, Kottwitz and MacPherson [GKM1], [GKM2]. The methods in all of these papers exploit the algebraic geometry over elds of positive characteristic. However, by the results of Waldspurger [W3] on independence of characteristic (which have also been established by the methods of motivic integration [CHL], following [CH]), they apply also to our eld F of characteristic 0. The paper of Ngo treats both the fundamental lemma for G G0 and the variant to which Waldspurger had reduced the general case [W6]. It therefore resolves the fundamental lemma for any G. As for the LSK conjecture, Waldspurger established some time ago that the case G G0 would follow from the fundamental lemma [W1]. His recent papers [W6], [W7] extend this implication to the general case. The general results of Waldspurger therefore yield the p-adic LKS conjecture in all cases. Our main focus in this volume will be on the three cases introduced in 1.2. In the rst, we take G pG0 , , q to be the pair ` r r r GpN q G0 pN q, pN q (with trivial) we were just discussing. In this case, ` r r G1 represents a twisted endoscopic datum G in the set EpN q E GpN q . We have already xed a homomorphism of L G into the dual group p G0 pN q GLpN, Cq, which of course determines an L-embedding of L G into L G0 pN q. This amounts to a canonical choice of auxiliary data. In the second r case, G and G1 are simply two objects G P Esim pN q and G1 P EpGq. Here we L G1 into L G, so there is again no need have also xed an L-embedding of of further auxiliary data. In the third case, we take G pG0 , , q to be the r r r pair G pG0 , q (with trivial) attached to an even orthogonal group G in

2.1. LANGLANDS-SHELSTAD-KOTTWITZ TRANSFER

55

r r r Esim pN q, and G1 to be a twisted endoscopic datum G1 in EpGq. Once again, r r we have an L-embedding of L G1 into L G0 , and no need for further auxiliary data. This case will generally be easier to manage once we have taken care of the others, and will often be treated as an addendum. The rst two cases on the other hand will have a central role in almost all of our arguments. In fact, we shall sometimes nd it convenient to treat the corresponding pairs ` r GpN q, G and pG, G1 q in tandem. Since the groups we are considering here are quasisplit, we can x canonical transfer factors. We shall use the Whittaker normalization of [KS, 5.3]. In general, a Whittaker datum consists of a rational Borel subgroup and a non-degenerate character on its group of rational points. In present circumstances, these objects can be chosen canonically, up to the choice of a xed nontrivial additive character F of F . r r Consider the GLpN q-bitorsor GpN q. We x a pN q-stable Whittaker ` r datum BpN q, pN q for G0 pN q by taking BpN q to be the standard Borel r subgroup of G0 pN q GLpN q, and pN q to be the non-degenerate character on the unipotent radical NBpN q pF q of BpN, F q attached to F . That is, pN, xq F px1,2 ` ` xn1,n q, x pxi,j q P NBpN q pF q.

r Given a twisted endoscopic datum G P EpN q, we take 1 `1 r r r N pN q SpN q 2 , F , F ` r to be the transfer factor assigned to BpN q, pN q in [KS]. Here SpN q is the transfer factor attached in [KS, p. 63] to the standard splitting SpN q r of G0 pN q GLpN q, F is the additive character used to dene pN q, and F is the representation of the Galois group F on the space X pT q b Q, where T is standard (diagonal) maximal torus in G. The local -factor the `1 2 , G , G is trivial unless G has aquasisplit group SOp2n, G q as a factor, `1 in which case it equals 2 , , G . (The general denition in [KS] also G `1 r r contains a factor 2 , N , F , where N is the representation of F on ` r T pN q pN q b Q attached to GpN q, but since the associated r the space X maximal torus T pN q BpN q is split, this factor equals 1). r A general group G P EpN q is a product of (at most two) quasisplit orthogonal and symplectic groups, together with a number of general linear groups. The discussion of 1.2 leads to a standard splitting for each factor, and hence a splitting S for the product G. We use this splitting to form a Whittaker datum pB, q for G. We then have a corresponding transfer factor ` 1 ` 1 1 2 , G , F S 1 , G , F 2 for any G1 P EpGq. Similar considerations give rise to a canonical transfer factor ` 1 ` 1 1 r r r r 2 , G , F S 1 , G , F 2 r

56

2. LOCAL TRANSFER

r r for the twisted component G of an even orthogonal group G P Esim pN q, and r r a twisted endoscopic datum G1 P EpGq. r Given the normalized transfer factors N , and with the knowledge that the KLS-conjecture holds in general, we have a canonical transfer mapping r r r r r f f G from HpN q to SpGq, for any G P EpN q. The function f G actur ally lies in the subspace of OutN pGq-invariant functions in SpGq, since the r function N is invariant under the action of this group on the rst variable. We also have a canonical transfer mapping f f 1 from HpGq to SpG1 q, for any G1 P EpGq, whose image has a similar symmetry property. Finally, r in the supplementary case of a twisted (even) orthogonal group G, we have r f 1 from HpGq to SpG1 q attached to any r r r a canonical transfer mapping f r r G1 P EpGq. r r r r If G P Esim pN q is simple, we follow 1.5 by dening HpGq, IpGq and SpGq r as the subspaces of OutN pGq-invariant functions in HpGq, IpGq and SpGq. r We extend this denition in the natural way to any G P EpN q, or indeed to any connected, quasisplit group that is a product of orthogonal, symplectic and general linear factors, by imposing the appropriate symmetry condition at the orthogonal factors of G. For example, if G1 G1 G 1 , 1 2 then r r r SpG1 q SpG1 q b SpG1 q. 1 2 r f G and f f 1 then take HpN q into SpGq and HpGq r r r r The mappings f r r r into SpG1 q. Observe, however, that if G P EpN q is not elliptic, OutN pGq is r larger than the symmetry group of the subspace SpGq. In particular, the r f G from HpN q to SpGq is not surjective in this case. r r r transfer mapping f Twisted transfer for GLpN q has a special role to play. It is what leads us to the essential objects for orthogonal and symplectic groups, in their r guise as simple endoscopic data G P Esim pN q, on which we will be able to base the local classication. To this end, we will need to know that the r r r r mapping f f G does take HpN q onto SpGq. The property is part of a broader characterization of the image of the collective transfer mapping G r r r r r f f , f P HpN q, G P EpN q,
G

r G1 P Esim pNi q, i 1, 2, i

which we will later also have to know. r We sometimes denote general elements in EpN q by M , since they can r be regarded as Levi subgroups of elliptic endoscopic groups G P Eell pN q. Recall that any such M comes with an L-embedding M of M L M into L GLpN q, which we can treat as an equivalence class of N -dimensional representations of L M . We may as well take G also to be a general element r in EpN q. Suppose that its N -dimensional representation G can be chosen so that it contains the image of (a representative) of M . In other

2.1. LANGLANDS-SHELSTAD-KOTTWITZ TRANSFER

57

words, M is the composition of an L-embedding L M L G with . This x p L-embedding identies M with a Levi subgroup of G, which is dual to an F -rational embedding M : M G as a Levi subgroup of G. We shall call M a Levi embedding of M into G. It is uniquely determined up to composition M M r r by elements P AutN pGq and M P AutN pM q. We note that for any M , the stable orbital integrals of a function f P HpGq over G-regular classes in M pF q give a function f M pfM qM in SpM q. Consider a family of functions ( r r (2.1.3) F f P HpGq : G P EpN q , r parametrized by the endoscopic data for GpN q. We shall say that F is a r compatible family if for any data G and M in EpN q, and any Levi embedding r r M of M into G, the associated functions f P HpGq and h P HpM q in F satisfy f M hM . r In other words, the function in SpM q attached to M equals the restriction of the corresponding function for G. In particular, the family of stable funcr tions tf G u attached to F is determined by the subset of functions f P HpGq r parametrized by the elliptic elements G P Eell pN q, since any M has a Levi embedding into some elliptic G. We could therefore have formulated F as r a family of functions parametrized by Eell pN q. Proposition 2.1.1. Suppose that F is any family of functions (2.1.3). r r Then F is a compatible family if and only if there is a function f P HpN q such that r r f G f G, G P EpN q. Proof. This is the twisted analogue (for GLpN q) of a property that plays a signicant role in the stabilization of the trace formula. For the general untwisted case, the proof is implicit in [S2] and [A11], and will be r given explicitly in [A24]. (See [A11, p. 329].) For the twisted group GpN q, the proof follows similar lines, based on work in progress by Shelstad and Mezo on twisted characters for real groups and Waldspurgers p-adic results [W4] on twisted endoscopy, which include the twisted analogue of a theorem from [W1]. However, since the lemma represents part of the stabilization of the twisted trace formula, which we will be adopting as a general hypothesis in the next chapter, we will not give a formal proof. We will be content simply to outline the main arguments. Suppose rst that F is p-adic. This is the more dicult case. Its proof has three steps, which we discuss briey in turn. It will be easier for us to

58

2. LOCAL TRANSFER

r refer to the untwisted case in [A11], even though the proof for GpN q is at least partly implicit in [W4, VVI]. r The rst step is to establish twisted analogues for GpN q of the adjoint relations [A11, (2.4), (2.5)] for geometric transfer factors. Let ` r r reg,ell pN q reg,ell GpN q be the set of G0 pN, F q-conjugacy orbits in GpN, F q that are strongly regular and elliptic. (We are using the natural twisted forms of these notions, for r which the reader can consult [KS, p. 28, 74].) If G belongs to Eell pN q, let r N -reg,ell pGq GpN q-reg,ell pGq r r be the set of stable conjugacy classes in GpF q that are GpN q-strongly regular and elliptic. We then set ` ` r r r r E pN q E GpN q N -reg,ell pGq{OutN pGq , reg,ell reg,ell
r GPEell pN q

where the right hand quotient denotes the set of orbits under the nite group r r OutN pGq. The twisted transfer factor for GpN q is a function p 1 , q p, q r r of P reg,ell pN q and 1 P N -reg,ell pN q, which depends only on the image 1 in E r of reg,ell pN q. The adjoint relations take the form p, qp, 1 q p, 1 q
r PE reg,ell pN q

and
r Preg,ell pN q

p, qp, 1 q p, 1 q,

r r for elements , 1 P reg,ell pN q and , 1 P E reg,ell pN q, where p, q p, q is the adjoint transfer factor, and p, q is the Kronecker delta. Their proof is similar to that of [A11, Lemma 2.2]. In the twisted case here, we use the equivariance relation @ D p, q invp, 1 q, p, 1 q of [KS, Theorem 5.1.D], and the local form of the bijection established in [KS, Lemma 7.2.A]. The second step is to establish that the transfer mapping G G r TE T : a aE a , a P Icusp pN q,
G G

2.1. LANGLANDS-SHELSTAD-KOTTWITZ TRANSFER

59

r from the invariant space Icusp pN q of cuspidal functions to its endoscopic counterpart ` OutN pGq r rE (2.1.4) Icusp pN q Scusp pGq ,
r GPEell pN q

` r r is an isomorphism. (Recall that a function a in the space IpN q I GpN q r is cuspidal if apq vanishes for every in the complement of reg,ell pN q in r the set reg pN q of all regular classes.) The proof of this fact is similar to that of its untwisted analogue [A11, Proposition 3.5]. One uses the adjoint transfer factor p, q to dene a candidate TE : aE a for the inverse mapping. The adjoint relations then tell us that the mapping rE TE is indeed the inverse of T E , once we know that it takes Icusp pN q to r Icusp pN q. This last fact relies on deeper results of Waldspurger. One has to use his general methods of twisted descent [W6], for twisted transfer factors and the corresponding transfer mappings, in place of those of Langlands and Shelstad [LS2]. This reduces the problem to a local assertion for (untwisted) p-adic Lie algebras, which one veries by inverting Waldspurgers kernel formula [W1, (1.2)], as in the proof of [A11, Proposition 3.5]. (See [W4, V.1].) The third step is to extend the isomorphism T E to the full space ` r r IpN q I GpN q . This is where we are taking the most for granted, namely r the extension to the twisted group GpN q of Theorem 6.2 of [A11]. We shall be brief. Following [A11, 1], we form the graded vector space ` W pM q , r Icusp pM q (2.1.5) Igr pN q
tM u

r where tM u ranges over orbits of (semistandard) Levi subsets of GpN q under the Weyl group G0 pN q N W0 W0 SN , and ` r W pM q W N pM q Norm A , G0 pN q {M 0
M

r is the relative Weyl for M . (Recall that M GpN q is dened [A4, 1] as a r Levi component of a parabolic subset of GpN q, with split component AM .) r q (which we will introduce formally The twisted tempered characters for GpN at the beginning of the next section) then provide a canonical isomorphism r r from IpN q onto Igr pN q, as in the untwisted case [A11, 4]. If G belongs to rell pN q, we can also form the graded vector space E ` W pM q (2.1.6) Sgr pGq Scusp pM q ,
tM u

60

2. LOCAL TRANSFER

G where tM u ranges over W0 -orbits of Levi subgroups of G, and W pM q is the relative Weyl group for M . The spectral construction from [A11, 5] then provides a canonical isomorphism from SpGq onto Sgr pGq. This assertion includes a nontrivial property of stability, and requires proof, unlike the r statement for GpN q above. The necessary justication is furnished by [A11, Theorem 6.1], the rst main theorem of [A11]. It is the second main theorem r [A11, Theorem 6.2] whose extension to GpN q we are taking for granted. The extended assertion is that the two isomorphisms are compatible with transfer. In other words, the diagram r IpN q Igr pN q r T G

(2.1.7)

SpGq Sgr pGq , in which the left hand vertical arrow is the (twisted) transfer mapping, and the right hand vertical arrow is built out of restrictions of transfer mappings to cuspidal functions, is commutative. The assertion of the proposition for our p-adic eld F is a consequence of these constructions. We introduce an invariant endoscopic space $ , & . r (2.1.8) I E pN q fG : f P F % r GPEell pN q

as the subspace of stable functions in `

r GPEell pN q

SpGq

dened by the stable images of compatible families F. We also form the graded endoscopic space ` E rE W pM q . (2.1.9) Igr pN q Icusp pM q
tM u

It is then not hard to construct a commutative diagram of isomorphisms

r IpN q Igr pN q r E . T r I E pN q Igr pN q rE

(2.1.10)

The upper horizontal isomorphism is as above. The right hand vertical isomorphism is a consequence of the isomorphisms for spaces of cuspidal functions we have described. The lower horizontal isomorphism follows from the denition of a compatible family, and the existence of the corresponding isomorphism in (2.1.7). The three isomorphisms together then tell us

2.1. LANGLANDS-SHELSTAD-KOTTWITZ TRANSFER

61

that the left hand vertical arrow is also an isomorphism, which from the corresponding arrow in (2.1.7) we see is the endoscopic transfer mapping TE T G.
r GPEell pN q

r The assertion of the proposition follows, for elements G P Eell pN q and hence r for any G P EpN q. Assume now that F is archimedean. In this case, we take for granted the extension of Shelstads theory of endoscopy, both geometric and spectral, to r the twisted group GpN q. (See [S6], [Me].) Then the archimedean forms of the spaces dened for p-adic F above all have spectral interpretations. They can each be regarded as a Paley-Wiener space of functions in an appropriate space of bounded, self-dual Langlands parameters : LF
L

r G0 pN q GLpN, Cq F .

Such parameters can of course be identied with self-dual, unitary, N dimensional representations of WF . They can be classied by the analysis of 1.2, or rather the extension of this analysis to parameters that are not elliptic. It is then not hard to deduce the archimedean form of the last commutative diagram of isomorphisms, and hence the assertion of the proposition. r Corollary 2.1.2. Suppose that G P Esim pN q is simple. (twisted) transfer mapping r r f f G , r takes HpN q onto the subspace ` OutN pGq r r SpGq SpGq of SpGq. Proof. We need to check that the transfer mapping T G in (2.1.7) takes r r r IpN q onto SpGq. We can either show directly that any function in SpGq is the image of a compatible family, or we can examine the lower horizontal mappings in (2.1.7) and (2.1.10). Taking the latter (less direct) course, we r compose the inverse of the mapping in (2.1.10) with the projection of I E pN q onto SpGq and the mapping in (2.1.7). This gives us a mapping (2.1.11) rE Igr pN q Sgr pGq. r r f P HpN q, Then the

We must check that its image is the space of invariants in Sgr pGq under the r natural action of OutN pGq. rE If we restrict the mapping (2.1.11) to a summand of Igr pN q from the decompositions (2.1.9) and (2.1.4), which is then mapped to the corresponding summand of Sgr pGq in (2.1.6), we obtain an injection from a subspace

62

2. LOCAL TRANSFER

of Scusp pM q into a larger subspace. The problem is simply to characterize its image. It amounts to establishing a isomorphism of groups r OutM pM qW pM , M q W pM qOutN pG, M q, M P Eell pM q,

r where W pM , M q is the stabilizer of M in W pM q, and OutM pG, M q is the r stabilizer of M in OutM pGq. This is a straightforward exercise, which we leave to the reader. For example, in the special case that M is the subgroup of diagonal matrices T tt : t_ tu in G, OutM pM q is trivial, and the product of groups on the right is indeed equal to the stabilizer of M T in W pM q SN . We note here that this property fails if G is composite, and therefore that the condition that G be simple is in fact necessary. Remarks. 1. One direction in the statement of the proposition is more r r r or less obvious. The fact that for any f P HpN q, the set tf G u comes from a compatible family F follows directly from the basic properties of transfer r factors. It is the converse, the existence of some f for any F, that is the main point. The argument we have given establishes both parts together. 2. The discussion from the proof of the corollary applies also to paramer r ters in the set pGq. It gives us another way to see that for any G P Esim pN q, r r the mapping from pGq to pN q is injective. We have of course always been r r treating pGq as a subset of pN q, a property that was already implicit in the discussion of 1.2. 2.2. Characterization of the local classication The rst of the three theorems stated in 1.5 applies to the local eld F . It includes a local classication of representations of a quasisplit orthogonal or symplectic group G. However, the description is purely qualitative. In this section, we shall make the assertion precise. We shall state a theorem that characterizes the representations in terms of the LSK-transfer of functions. The classication is based on the transfer of self-dual representations of r GpN q GLpN q, or rather the extensions of such representations to GpN q. Having earlier normalized the LSK-transfer factors, we shall now normalize the extensions of these representations. We need to show that any irreducible r self-dual representation of G0 pN, F q GLpN, F q has a canonical extension r to the group G` pN, F q. As was implicit for the transfer factors, we will use the theory of Whittaker models. ` Suppose rst that is tempered. Then has a BpN q, pN q -Whittaker ` functional , where BpN q, pN q is the standard Whittaker datum xed in the last section. By denition, is a nonzero linear form on the underlying

2.2. CHARACTERIZATION OF THE LOCAL CLASSIFICATION

63

space V8 of smooth vectors for such that ` pnqv pN, nqpvq, n P NBpN q pF q, v P V8 . It is unique up to a scalar multiple. We are assuming that is self-dual, r which is to say that the representation pN q is equivalent to . We can r r therefore choose a nontrivial intertwining operator I from to pN q, which is unique up to a nonzero scalar multiple. Since the Whittaker da` r r tum BpN q, pN q is pN q-stable, the linear form I on V is also a ` BpN q, pN q -Whittaker model for . It therefore equals c, for some ` r r c P C . We set pN q pN q equal to the operator c1 I. Then pN q is r r r q such that the unique intertwining operator from to pN r pN q. r It provides a unitary extension of to the group G` pN, F q generated by r r r GpN, F q. In particular, it gives an extension of to the bitorsor GpN, F q that satises (2.1.1). Suppose next that is replaced by a self-dual standard representation . Recall that is a (possibly reducible) representation, induced from the twist M, of an irreducible tempered representation of a Levi subgroup M pF q by a regular, positive real valued character. We can identify M with a block diagonal subgroup GLpm1 q GLpmk q of GLpN q, and with a representation 1 px1 q| det x1 |1 b b k pxk q| det xk |k , x P M pF q,

for irreducible tempered representations 1 , . . . , k and real numbers 1 k . It follows easily from the self-duality of , and the form we have chosen for , that
_ i k`1i ,

1 i k.

r r In particular, M is pN q-stable, and is isomorphic to pN q. By a minor variation of our discussion of the tempered representation above, we see that has a canonical extension to the group r M ` pF q M pF q xpN qy. Notice that the standard block upper triangular subgroup P in GLpN q of r r type pm1 , . . . , mk q is also pN q-stable. Its normalizer P ` in G` pN q can be ` pN q. The pullback from M ` pF q to regarded as a parabolic subgroup of G r P ` pF q of the extended representation can then be induced to G` pN, F q. r ` pN, F q. We thus obtain an extension of to G r We note that in the case of GLpN q here, there is a bijection from r the set pN q of self-dual Langlands parameters for GLpN q and the set of

64

2. LOCAL TRANSFER

r self-dual standard representations of GLpN, F q. Given P pN q, we shall write ` r r r r r fN pq tr pf q , f P HpN q, r r for the extension of we have just described. Observe that for of the form above, we can replace the vector p1 , . . . , k q in Rk by a complex multiple z, z P C. This gives us another self-dual standard representation z , with correspondr ing Langlands parameter z . The function fN pz q, dened for z near 1 in terms of the Whittaker model of M , extends to an entire function. Now if z r is purely imaginary, z is tempered, and fN pz q is also dened in terms of the Whittaker model of GLpN q. However, it follows easily from the general structure of induced Whittaker models [Shal], [CS] that the two functions r of z P iR are equal. In other words, fN pq can be obtained for general by analytic continuation from the case of tempered . Suppose nally that is a general irreducible, self-dual representation of GLpN, F q. Then is the Langlands quotient of a uniquely determined standard representation . The bijective correspondence between irreducible and standard representations implies that is also self-dual. The canonical r r r extension of to G` pN, F q then provides a canonical extension of r ` pN, F q. This is what we needed to construct. In the case that is to G r unitary, the restriction of this extension to the bitorsor GpN, F q represents a canonical extension from the general class (2.1.1). r Suppose now that is a parameter in pN q. As in the global remarks following the statement of Theorem 1.3.3, we dene the local Langlands r parameter P pN q by setting 1 |u| 2 0 puq u, , u P LF . 1 0 |u| 2 We then have the standard representation of GLpN, F q attached to , and its Langlands quotient . The irreducible representation is unitary and self-dual. It therefore has a canonical extension to the r r bitorsor GpN, F q. We write ` r r r r r (2.2.1) fN pq tr pf q , f P HpN q. Our interest will be in the transfer of this linear form to a twisted endoscopic group. r r Suppose that belongs to the subset pGq of pN q attached to a simple rsim pN q. Then can be identied with an L-homomorphism datum G P E r G : LF SU p2q L G. On the other hand, we have just seen that determines a linear form (2.2.1) ` r r on the space HpN q H GpN q . The next theorem uses this linear form to

2.2. CHARACTERIZATION OF THE LOCAL CLASSIFICATION

65

make the local correspondence of Theorem 1.5.1(a) precise. Specically, it r gives an explicit construction of the packet and the pairing x x , y of Theorem 1.5.1(a) in terms of (2.2.1), and the endoscopic transfer of functions. r Theorem 2.2.1. (a) Suppose that G P Eell pN q, and that belongs to r pGq. Then there is a unique stable linear form (2.2.2) f f G pq, r r f G pq fN pq, f G pq f S pS qf O pO q, G G S GO , S O , and f G f S f O, r f P SpG q, O, S, r f P HpGq,

r on HpGq with the general property (2.2.3) (2.2.4) in case r G P Esim pN q, r P pG q, r r f P HpN q, r f P HpGq,

together with a secondary property

are composite. r r (b) Suppose that G P Esim pN q. Then for every P pGq, there is a nite r over unit pGq, together with a mapping r set (2.2.5) x , y, r P ,

r p from to the group S of irreducible characters on S , with the following property. If s is a semisimple element in the centralizer S S pGq and pG1 , 1 q is the preimage of p, sq under the local version of the correspondence (1.4.11) in 1.4, then r (2.2.6) f 1 p 1 q xs x, y fG pq, f P HpGq,
r P

where x is the image of s in S . r Remarks. 1. If G P Esim pN q is simple, the uniqueness of the linear r r r form in (a) follows from (2.2.3), since the mapping f f G takes HpN q r rsim pN q is composite, the uniqueness follows from the onto SpGq. If G R E product formula (2.2.4). r 2. It is clear that (2.2.6) characterizes the packet and the pairing xx, y in (b). It therefore does provide a canonical formulation of the local correspondence of Theorem 1.5.1. 3. If G is composite, the symbol on the right hand of (2.2.3) is understood to be the image of the given parameter under (the not necessarily

66

2. LOCAL TRANSFER

r r injective) mapping of pGq to pN q. In this case, the identity (2.2.3) is not needed to characterize either the linear form in (a) or the local correspondence in (b). However, it will still have an important role in our proofs. 4. Theorem 2.2.1 will be proved at the same time as the theorems stated in Chapter 1, together also with further theorems we will state in due course. The argument, which will take up much of the rest of the volume, will rely on long term induction hypothesis based on the integer N . r 5. If F is archimedean and lies in the subset bdd pN q of generic r parameters in pGq, the assertions of the theorem are included in the general results of Shelstad [S2], [S3][S6], and their twisted analogues for GLpN q [Me], [S7] in preparation. Theorem 2.2.1 of course includes parameters that are not generic, in the sense that they are nontrivial on the second factor of LF SU p2q. One ultimately wants to reduce their study to the case that is generic. With this in mind, we shall describe how the twisted character on the right hand side of (2.2.3) can be expanded in terms of standard twisted characters. r r Let us write P pN q for the set of standard representations of GpN, F q. Elements in this set are induced objects r IP prM, q, r P pa q` , r P

r r r r where P M N is a standard parabolic subset [A4, 1] of GpN q, and M pF q of an irreducible tempered representation is an extension (2.1.1) to M r r r0 M of M 0 pF q. (The sets P and P pN q are of course dierent. The rst symbol P is an upper case p, while the second is supposed to be an upper r r r case .) By denition, P is the normalizer in GpN q of a -stable, standard r0 M 0 N 0 of G0 pN q GLpN q. Its Levi subset M r r parabolic subgroup P has a split component AM AM 0 , and an associated real vector space aM aM 0 with an open chamber a` r` and corresponding dual chamber aP r P q` pa q` . As in (2.1.1), there is a simply transitive action pa r P
P

r r u : M M,u ,

u P U p1q,

of the group U p1q on the set of extensions M of M , and therefore a correr r0 r r r r r0 sponding action u on the set of P P pN q attached to M and . We r note that any U p1q-orbit in P pN q contains a simply transitive pZ{2Zq-orbit. r r It consists of those such that M generates an irreducible representation of the group M ` pF q generated by M pF q. Within this pZ{2Zq-orbit, we have r the element dened by the Whittaker normalization above. r r r Any P P pN q has a canonical Langlands quotient . It is an exr ` 0 P G0 pN q r r r tension (2.1.1) to GpN, F q of an irreducible representation (except for the fact that 0 does not satisfy the unitary condition of (2.1.1)). r r r As in (2.1.1), the set of extensions attached to 0 in this way is a transir r tive U p1q-orbit. The mapping is then a U p1q-bijection from P pN q r

2.2. CHARACTERIZATION OF THE LOCAL CLASSIFICATION

67

r r onto the corresponding set pN q of extensions to GpN, F q of irreducible r r r representations of G0 pN, F q. We write tP pN qu and tpN qu for the set of r r U p1q-orbits in P pN q and pN q respectively. r A standard parabolic subset P comes with a canonical embedding of q` into the closure pa q` of the dual chamber of the its dual chamber pa r B P standard Borel subgroup B. For any standard representation P P pN q with r r r r Langlands quotient in pN q as above, we shall write r , 1 , 0 , 0 , r r 1 , P pa q` , B for the corresponding image of the point . In general, one writes if 1 is a nonnegative integral combination of simple roots of pB, AB q. r r This determines a partial order on each of the sets tP pN qu and tpN qu. We thus have two families ` r r r r r r r fN prq tr pf q , f P HpN q, P P pN q, and r r r r f P HpN q, P pN q, r r of G0 pN, F q-invariant linear forms on HpN q. Taken up to the action of U p1q, they represent two bases of the same vector space. More precisely, there are uniquely determined complex numbers ( r r r r r r mpr, q, npr, q : P P pN q, P pN q , with mpru , v q u mpr, qv 1 , r r and r r nprv , u q v npr, qu1 , such that (2.2.7) and (2.2.8) r fN prq
PtP pN qu r r

` r r r fN prq tr pf q ,

u, v P U p1q, u, v P U p1q, P P pN q, r r

r fN prq

PtpN qu r r

r r mpr, q fN prq,

r npr, q fN prq,

P pN q. r r

This is a special case of a general result, which can be formulated in the same manner for any triplet G pG0 , , q. (See [A7, Lemma 5.1], for example.) The rst expansion (2.2.7) follows from the decomposition of the standard r r representation 0 of G0 pN, F q attached to into irreducible representations. The second expansion (2.2.8) follows by an inversion of the rst, in which the coecient function mpr, q is treated as a unipotent matrix. r r Let us be more specic about the matrix tmpr, qu. We recall that mpr, q 1 if . In general, if mpr, q 0 for a given , the linear r r r r r

68

2. LOCAL TRANSFER

form satises , with equality holding only when . This is r r the property that is responsible for the matrix being unipotent. In addition, mpr, q vanishes unless and have the same central character r r r : F C , and more signicantly, the same innitesimal character : ZpN q C. Here, ZpN q denotes the center of the universal enveloping algebra of the r complexied Lie algebra of G0 pN, F q if F is archimedean, and the Bernstein 0 pN, F q if F is p-adic. It is an abelian C-algebra, with a two-sided r center of G r action on HpN q that is compatible with the associated two-sided action of r the Hecke algebra HpN q H0 pN q. We can therefore regard tmpr, qu as r a block diagonal matrix, whose blocks are parametrized by -stable pairs p, q. Since there are only nitely many irreducible or standard representar tions of G0 pN, F q with a given innitesimal character , the blocks represent unipotent matrices of nite rank. We obtain equivalent expansions if we x representatives in the two sets of U p1q-orbits. Let us do so by choosing the Whittaker extensions introduced r r above. The sets tP pN qu and tpN qu are then each bijective with the family r r r of Langlands parameters pN q. For any P pN q, we are thus taking to be the associated standard representation , and to be the associated r r r irreducible Langlands quotient . For any P pN q, we take to be r r r the associated representation . This gives a set of representatives for the r r subset of tpN qu indexed by pN q. The expansion (2.2.8) for elements in this subset can then be written r r r r (2.2.9) fN pq np, q fN pq, P pN q,
r PpN q

where r np, q npr , q. r Before we proceed to the initial study of Theorem 2.2.1, we shall describe a second well known feature of general Langlands parameters P pN q for GLpN q, their innitesimal characters. These are by denition the innitesimal characters of the standard representations of GLpN, F q (or the corresponding Langlands quotients , or indeed, any of the constituents of ). Our immediate motivation is the fact that the terms on the right hand side of (2.2.9) all have the same innitesimal character. Our remarks will r lead to information about the set pGq, which we will need for the analysis of (2.2.9). They will also be useful later on. Suppose rst that F is archimedean, and that : WF GLpN, Cq

2.2. CHARACTERIZATION OF THE LOCAL CLASSIFICATION

69

is a general Langlands parameter for GLpN q over F . The innitesimal character of is given by a semisimple adjoint orbit in the Lie algebra of GLpN, Cq. To describe it, we rst choose a representative of whose restriction to the subgroup C of WF takes values in the group of diagonal matrices. Following Langlands general notation [L11, p. 128], we write ` _ _ (2.2.10) _ pzq z x , y z x , y , z P C , P ZN , for complex diagonal matrices and with integral. The innitesimal character of is then equal to (the orbit of) , which we can regard as a multiset of N complex numbers. r Suppose for example that belongs to the subset ell pN q of pN q. The condition that be self-dual is equivalent to the relation t . The ellipticity condition implies further that is a direct sum of distinct, self-dual two dimensional representations of WF (together with a self-dual one-dimensional character of F in case N is odd), from which one sees that . It follows that ` _ z P C , _ P ZN , _ pzq pzz 1 qx , y ,
1 where P 2 ZN is now a diagonal half integral matrix with t . We r r know that belongs to 2 pGq, for a unique G P Eell pN q. The dual group of G is given by

p p p G GS GO SppNS , Cq SOpNO , Cq, where NO and NS are dened as the number of components of , a priori half integers, which are respectively integral and nonintegral. This is an immediate consequence of the explicit description of Langlands parameters for orthogonal and symplectic groups, which we will review in 6.1. The p group G is then determined by G and the fact that its quadratic character G equals the determinant of . Conversely, suppose that P pN q satises (2.2.10), for a diagonal matrix P 1 ZN with t . This does not imply that lies in 2 r ell pN q, even if the components of are all distinct. However, we can r still attach a canonical element G P Eell pN q to and , according to the prescription above. We can also choose a general endoscopic datum r M P EpN q such that is the image of a local Langlands parameter M in r 2 pM q. It follows from the various denitions that M can be identied with a Levi subgroup of G, and hence that is the image in pN q of a canonical r r element in pGq. If G P Esim pN q is simple, for example, pGq embeds into r pN q, so that itself represents a canonical element in the subset pGq of pN q. Suppose next that F is p-adic. The innitesimal character of a Langlands parameter : LF WF SU p2q GLpN, Cq

70

2. LOCAL TRANSFER

for GLpN q over F is often called the cuspidal support of or of . It can be regarded as a multiset r k ,
k1

built from segments k with

r k1

|k | N,

in the teminology of [Z]. Let us be more precise. We rst recall [Z, 3] that a segment of norm || m ms mu is a set ri , j s ti , i1 , . . . , j u, ` where P scusp GLpms q is a (not necessarily unitary) supercuspidal representation of GLpms , F q, while mu 1 i j is an integer that is a ` dierence of two given half integers i j, and P scusp GLpms q is dened by p qpxq pxq | det x| , x P GLpms , F q. (By a half integer, we again mean a number whose product with 2 is an integer!) The associated induced representation of GLpm, F q then has a unique subrepresentation . (See [Z, 3]. By implicitly ordering by decreasing half integers, as above, we have adopted the convention of Langlands rather than that of Zelevinsky. In other words, the special, essentially square integrable representation is a subrepresentation of the induced representation rather than a quotient.) Suppose that a is a multiset consisting of segments tk u of total norm N . We can then form the associated induced representation a IP p1 b b r q of GLpN, F q, and its corresponding Langlands quotient a . The segments are bijective with simple Langlands parameters P sim pmq. The associated mapping is the Langlands correspondence for the relative discrete series of GLpmq. The multisets a are bijective with general Langlands parameters a P pN q. The associated two mappings a a and a a in this case represent the general Langlands correspondence for irreducible and standard representations of GLpN q. Our description of the innitesimal character of a p-adic Langlands parameter P pN q will now be clear. We write a , for a multiset a whose elements are segments tk u. Then is the multiset obtained by taking the disjoint union over k of the sets k . For another description,

2.2. CHARACTERIZATION OF THE LOCAL CLASSIFICATION

71

which is a little closer to that of the archimedean case, we write as at the beginning of 1.3 for the restriction of to the subgroup , $ 1 . & 0 w, |w| 2 : w P WF 1 % 0 |w| 2 of WF SLp2, Cq. As an N -dimensional representation of WF , P pN q can then be regarded as a Langlands parameter that is trivial on the second factor of LF . The innitesimal character is then equal to the multiset a that corresponds to . As we observe from either description, is a multiset of supercuspidal representations , while a is a coarser multiset of segments k . r Suppose for example that belongs to the subset ell pN q of pN q. Then is a direct sum of distinct, irreducible, self-dual representations of LF , which in turn correspond to distinct, self-dual segments. Since the dual of a general segment ri , j s equals _ r _ j , _ i s, a self-dual segment may be written in the form (2.2.11) ri , i s ti , i1 , . . . , i u, for a self-dual (unitary) supercuspidal representation P scusp pmq, and a nonnegative half integer i. This segment is symplectic if is symplectic and i is an integer, or if is orthogonal and i is not an integer. It is orthogonal otherwise, that is, if is orthogonal and i is an integer, or if is symplectic and i is not an integer. (Recall that the self-dual representation was designated symplectic or orthogonal in 1.4 according to whether its corresponding Langlands parameter factors through the associated complex subgroup of GLpm, Cq.) The dual group of G is again a product p p p G GS GO SppNS , Cq SOpNO , Cq, given in this case by the partition of the segments of into those of orthogonal and symplectic type. The group G is then determined by the dual group p G and the character G . Conversely, suppose that P pN q is a parameter such that is a disjoint union of self-dual sets (2.2.11). This does not imply that lies in r ell pN q. However, as in the archimedean case, we can still attach a canonical r element G P Eell pN q to and , with the property that is the image r in pN q of a canonical parameter in pGq. In the p-adic case, is just the determinant of , and is therefore superuous here. However, we will continue to carry it in order to accommodate the Archimedean case. With these notions in mind, we consider again the expansion (2.2.9). r Each parameter P pN q on the right has an innitesimal character and r determinant . The parameter P pN q on the left is likewise equipped with data and . These parameters also come with the linear functionals

72

2. LOCAL TRANSFER

and . Given , we dene ( r r pN, q P pN q : p , q p , q, . The expansion (2.2.9) then remains valid if the sum is restricted to the r r subset pN, q of pN q. Lemma 2.2.2. Suppose that the assertion (a) of Theorem 2.2.1 holds in the special case that is generic. It is then valid for any . Proof. The assertion is assumed to hold if lies in the subset r r bdd pN q of generic parameters in pN q. Suppose that belongs to the r larger set of generic parameters in the general family ` pN q. In other r words, P pN q is a general self-dual Langlands parameter. We have observed that the associated (twisted) standard character r r fN pq fN pr q, r r f P HpN q,

can be obtained by analytic continuation from the tempered case. Since the left hand side of (2.2.3) is also an analytic function in the relevant complex variable, our assumption implies that the obvious variant of assertion (a) of the theorem holds if is replaced by any P pN q. Suppose rst that G is simple. In this case, we have only the denition r (2.2.3) to contend with. We must show that for the given element P pGq, r pq in (2.2.3) depends only on the image of the mapping the linear form fN r r r f f G . We shall use the expansion (2.2.9) for as an element in pN q. Given the supposition of the lemma, and its extension above to general r r parameters P pGq, we need only show that if the coecient np, q in r q, then lies in the subset pGq of r (2.2.9) is nonzero for some P pN r r r pN q. We rst note that belongs to the complement of ell pN q in pN q rM P pM q attached to a proper if and only if it is the image of a parameter r standard Levi subset M of GpN q. In this case, the terms in (2.2.9) are each induced from analogous terms for M . The required assertion then follows inductively from its analogue for the proper Levi subgroup M of G attached r to M . We may therefore assume that lies in the intersection of pGq with r r r ell pN q, which is just the subset 2 pGq of pGq. Another induction argu ment, combined with the knowledge we now have of the elliptic endoscopic data for G, reduces the problem further to the case that is simple. r r We are thus taking both G P Esim pN q and P sim pGq to be simple. Then b , N m n, for irreducible unitary representations and of LF and SU p2q of respective dimensions m and n. The innitesimal character of is by denition the innitesimal character of the irreducible representation

2.2. CHARACTERIZATION OF THE LOCAL CLASSIFICATION

73

in (2.2.1). It equals the tensor product of the innitesimal characters of and , dened in the obvious way, and is naturally compatible with the tensor product of the representations themselves. (The overlapping notation in and is unfortunate, but should not cause confusion.) If we compare the remarks for real and p-adic innitesimal characters above with the local r version of the discussion following (1.4.7), we see that the datum in Eell pN q attached to and equals G. But we have agreed that the r r sum in (2.2.9) can be taken over in the subset pN, q of pN q. For any such , the datum attached to and is therefore also equal r r to G. In particular, belongs to the subset pGq of pN q. This is what we had to show. r r We have established that for simple G P Esim pN q, and any P pGq, the set r r pG, q pN, q r r is contained in the subset pGq of pN q. The denition (2.2.3) is thus valid for G. It takes the form r r (2.2.12) f G pq np, q f G pq, f P HpGq,
r PpG,q

which we obtain from (2.2.9). r r Suppose now that G P Eell pN q and P pGq are composite, as in (2.2.4). In this case, the linear form (2.2.2) is dened by the product (2.2.4) and what we have established for the simple data GS and GO . The problem in this case is to show that (2.2.3) is valid. r It is again enough to assume that belongs to the subset ell pN q of r q. In other words, lies in the subset 2 pGq of pGq. We are now r r pN dealing with decompositions G GS GO , S O and N NS ` NO that are nontrivial. However, we shall treat them in the same way, by applyr r ing the expansion (2.2.9) with summed over the subset pN, q of pN q. r Any P pN, q has the same innitesimal character and determinant as r the original parameter P ell pN q. It follows from the earlier remarks for real and p-adic innitesimal characters that is the image of a canonical r parameter in pGq. In particular, we can identify with an element in r r r pGq, even though the mapping from pGq to pN q is not injective. With this interpretation, we obtain a canonical decomposition (2.2.13) S O , r r S P pNS , S q, O P pNO , O q,

r for any P pN, q. r It is convenient to write IpN, q for the nite dimensional space of comr r plex valued functions fN pq on the nite set pN, q. This space represents r q of GpN q, since the twisted r a quotient of the invariant Hecke space IpN trace Paley-Wiener theorem for GLpN q implies that the restriction mapr r ping from IpN q to IpN, q is surjective. We can identify it with the space

74

2. LOCAL TRANSFER

of functions on the set r r P pN, q tr : P pN, qu, r r r since fN pr q fN pq for any P pN, q. It is also canonically isomorphic to the space of functions on the set r r pN, q tr : P pN, qu, where we recall that like , represents the canonical Whittaker extension r r r to GpN, F q of the underlying representation of GLpN, F q. Indeed, the genr eral expansions (2.2.7) and (2.2.8) remain valid with the sets P pN, q and r r r r r pN, q in place of tP pM qu and tpN qu. For any function fN P IpN, q, r the corresponding function on pN, q is dened by this modication r r fN pr1 q npr1 , q fN pr q, r r 1 P pN, q,
r PpN,q

of (2.2.8). We dene a linear transformation (2.2.14) r r r IpN, q IpNS , S q b IpNO , O q

r r by mapping fN P IpN, q to the function r r fNS ,NO pS O q fN pq, S O ,

in the tensor product. The mapping is actually an isomorphism, by virtue r of the canonical decomposition (2.2.13) of any parameter P pN, q. It is r r closely related to the twisted transfer mapping f f G . More precisely, it is part of a commutative diagram of isomorphisms r IpN, q
>

r r IpNS , S q b IpNO , O q

r IpG, q

r r > IpGS , S q b IpNO , O q,

r r in which IpG, q is the quotient of IpGq dened by restriction of functions r r r to the subset pG, q pN, q of pGq. The left hand vertical arrow is the r r corresponding lift of the transfer mapping fN f G , while the right hand vertical arrow is its analogue for S O . This diagram is a consequence of the supposition of the lemma, or rather its extension to non-tempered generic parameters described at the beginning of the proof. We shall show that the isomorphism dual to (2.2.14) maps the product r S O , regarded as a linear form on the right hand tensor product, to r r the linear form on IpN, q. At the same time, we shall show that the r coecients in (2.2.9) satisfy the natural formula (2.2.15) np, q npS , S q npO , O q, r r r r P pN, q.

2.2. CHARACTERIZATION OF THE LOCAL CLASSIFICATION

75

It follows from the denition of (2.2.14) that the dual isomorphism takes r r r r r the linear form S O S O attached to any P pN, q to the r r linear form . Viewed directly in terms of standard representations, this correspondence is a composition (2.2.16) r S O S O IP pS O q , r r

where P is the standard parabolic subgroup of GLpN q of type pNS , NO q. r r The expansion (2.2.7) for is obtained from the decomposition of (as a standard representation of G` pN, F q) into irreducible constituents. Its r r analogues for S and O are of course obtained in the same way. The repr r r resentation is the Langlands quotient of , or more precisely, the r ` pN, F q of the Langlands quotient of that is compatible extension to G r r r with the Whittaker extension of , while S O is the Langlands r r quotient of S O . Using a local form of the global notation (1.3.6), we could write S O for the image of S O under the composition r r r r (2.2.16). But since the representations S and O are unitary, the induced representation IP pS O q is irreducible [Be]. It follows that S O , r r a priori a quotient of , is actually the Langlands quotient . This is what r r we wanted to show. Moreover, if we apply the dual of (2.2.14) to the terms in the product of the expansions (2.2.9) for S and O , we see that r r r npS , S q npO , O q fN pq, r r fN prS O q
r PpN,q

in the obvious notation. The expansion (2.2.9) itself can be written r r np, q fN pq. r fN pr q
r PpN,q

Since the left hand sides are equal, the required decomposition (2.2.15) of the coecients follows as well. We can now complete our proof of the second half of Lemma 2.2.2. r r Suppose that f is any function in HpN q. According to the denition (2.2.12) rG pq, we can write of f r r f G pq f G pS O q r r r npS , S q npO , O q f G pS O q,
S ,O

r r for a double sum over S P pNS , S q and O P pNO , O q. This in turn equals r npS , S q npO , O q fNS ,NO pS O q r r
S ,O

r PpN,q

r r np, q fN pq,

76

2. LOCAL TRANSFER

r by the denition of fNS ,NO , the commutative diagram above, and the cor ecient formula (2.2.15). The last sum is then equal fN pq, according to r r (2.2.9). We have established that f G pq equals fN pq, which is the required identity (2.2.3). Remarks. 1. My original proof was incorrectly based on the analogue of (2.2.9) for the connected group GLpN q. I thank the referee for pointing this out, and for suggestions that motivated the revised proof above. 2. The analogue of the decomposition (2.2.15) holds more generally for the coecients r npr, q, r P pN, q, P P pN, q, r r in (2.2.8). This is because each induced representation IP pS O q, r r r r S P pNS , S q, O P pNO , O q,

is irreducible, an observation I owe to the referee. The point is that there is no interaction between the innitesimal characters of S and O . To be precise, we can attach an ane Z-module to the innitesimal character of any P pN q by writing Zr s t : P , in the earlier notation for p-adic F , and Zr s ts ` : s P , P Zu if F is archimedean and is treated as a multiset of order N . The sets ZrS s and ZrO s are then disjoint, from which it follows by standard methods that the induced representation is irreducible. This presumably leads to a generalization of (2.2.3) that is tied to (2.2.8) (and a general representation ) rather than (2.2.9) (with the representation ). r r 3. The generic case of Theorem 2.2.1(a), which was our hypothesis for the lemma, will be established later by global means (Lemmas 5.4.2 and 6.6.3). It is possible that the general case of Theorem 2.2.1(a), the actual assertion of the lemma, could also be proved globally as a nongeneric supplement of Lemma 5.4.2. (See Lemmas 7.3.1 and 7.3.2, which represent nongeneric supplements of the other results in 5.4.) However, the local proof we have given here is perhaps more instructive. The following lemma gives another reduction of Theorem 2.2.1. This one is much easier, and has to some extent been implicit in our earlier discussion. Lemma 2.2.3. Suppose that Theorem 2.2.1 is valid if N is replaced by r any integer N N . Then it also holds for any parameter P pGq, for rsim pN q, such that the group S has innite center. GPE r Proof. Suppose that P pGq is as given. The centralizer in L G of any nontrivial central torus in S is then the L-group L M of a proper Levi subgroup M of G. We x M , and choose an L-homomorphism M from P Zu

2.2. CHARACTERIZATION OF THE LOCAL CLASSIFICATION

77

r LF SU p2q to L M whose image in pGq equals . The centralizer SM in x p M of the image of M is then equal to the original centralizer S in G of the image of . The Levi subgroup M is a product of general linear groups with a group r G P Esim pN q, for some N N . (See (2.3.4), for example.) It follows from the supposition of the lemma that the natural analogue of Theorem r 2.2.1 holds for M . This gives us a stable linear form hM pM q on SpM q, a r over unit pM q, and a mapping M x , M y from to r r packet M M p SM . We dene the corresponding objects for G by setting r f P HpGq, ( r r IP pM q : M P M , and x , y x , IP pM qy x , M y. The required character identity (2.2.6) then follows from the familiar descent formula ` r fG pq tr IP pM , f q fM pM q, f P HpGq. f G pq f M pM q,

P P PpM q,

Theorem 2.2.1 characterizes the local classication in terms of two kinds of local endoscopy, twisted endoscopy for GLpN q and ordinary endoscopy r for the group G P Esim pN q. Recall that we have also been thinking of a third case, that of twisted endoscopy for an even orthogonal group. This case is not needed for the statement of the classication. However, it has an indispensable role in the proof, and it also provides supplementary information about characters. We shall state what is needed as a theorem, to be established along with everything else. r Even orthogonal groups G P Esim pN q of course represent the case that r the group OutN pGq is nontrivial. They are responsible for our general nor r tation pGq and (rather than pGq and ), since it is in this case that r r an element in can be a nontrivial OutN pGq-orbit of irreducible representations (rather than a singleton). Twisted endoscopy is dened by the r nontrivial automorphism of order 2 in OutN pGq. It applies to the subset r of OutN pGq-xed parameters in pGq, or more accurately, those r r pGq r r P pGq that can be represented by an OutN pGq-xed L-homomorphism L G. For any such , we would expect the elements from LF SU p2q to r in to be singletons. In other words, they should consist of irreducible r representations of GpF q that extend to the group G` pF q generated by r q. For it is clear that the corresponding linear characters (2.2.5) on S GpF r` r extend to characters on the group S generated by S . However, we shall establish this only in case is generic.

78

2. LOCAL TRANSFER

r Theorem 2.2.4. Suppose that N is even, that G P Esim pN q is orthogonal r p r in the sense that G SOpN, Cq, and that lies in the subset pGq of r stable elements in pGq. r r (a) Suppose that s is a semisimple element in the G-twisted centralizer r S , and that r r r r r r pG1 , 1 q, G1 P EpGq, 1 P pG1 q, r is the preimage of p, sq under the analogue of the correspondence (1.4.11). Then we have an identity r r r r r r rr (2.2.17) f 1 p 1 q xs x, y fG prq, f P HpGq,
r P

r r r where x is the image of s in S , is any extension of to G` pF q, and r r r` r x, y is a corresponding extension of the linear character (2.2.5) to S such that the product r r rr xs x, y fG prq, x P S , r r in (2.2.17) depends only on (as an element in unit pGq). r r (b) If is generic, each P is an OutN pGq-stable representar r tion of GpF q, which consequently does have an extension to G` pF q. (In r r general, if P does not have an extension, we agree simply to set equal to 0). r Remarks. 1. The theorem relates twisted characters on GpF q with r r stable characters on the twisted endoscopic group G1 pF q. Recall that G1 is 1 , N 1 q whose sum equals N , r determined by a pair of odd positive integers pN1 r2 1 , 1 q on whose product equals the and a pair of quadratic characters pr1 r2 F quadratic character that denes G as a quasisplit group. The left hand r r r side of (2.2.17) is the value of f 1 , the transfer of f to SpG1 q, at a product of stable linear forms r r r r r 1 1 1 , 1 P pN 1 q,
1 2 i i

each dened by Theorem 2.2.1(a). On the right hand side, the sum is taken r over those P that as irreducible representations of GpF q have extenr sions to G` pF q. If does have an extension, there is in general no canonical choice (in contrast to the case of twisted GLpN q). However, the theorem asserts that the summand in (2.2.17) is independent of the choice, which is all we need. r 2. The extension x , y of x , y is uniquely determined by the extension r of . Given the condition that x , y depends only on as a representation r r of GpF q (which allows for the possibility that P occurs several times with the same linear character x, y), this easily established from (2.2.17). Indeed, if (2.2.17) is valid for two dierent extensions x , 1 y and x , 1 y1 r r r` , for a given 1 P with extension 1 to G` pF q, we can r r of (2.2.5) to S identify the corresponding summands on the right hand side. This implies

2.3. NORMALIZED INTERTWINING OPERATORS

79

rr r r that fG pr1 q vanishes for any f P HpGq, which is contrary to the basic properties of irreducible characters. r 3. We could have denoted the packet of by rather , since is r OutN pGq-stable. We shall often do so in the future. 2.3. Normalized intertwining operators Intertwining operators play an essential role in the proof of our theorems. Global intertwining operators are the main terms in the discrete part of the trace formula. It is important to understand their local factors in order to interpret the stabilization of this global object. Local intertwining operators also play a role in the local classication. They lead to a partial construction r of the packets and pairings x, y of Theorem 2.2.1, which reduces their r study to the case that belongs to 2 pGq. We shall devote the remaining three sections of the chapter to this topic. One of the main problems is to normalize local intertwining operators. This has been done quite generally by Shahidi [Sha4], for inducing representations with Whittaker models. However, our inducing representations will frequently not have Whittaker models, even when they are tempered. We will apply global methods to their study, both in the initial stage here and in arguments from later chapters. The problem is not simply to normalize intertwining operators, but to do so in a way that is compatible with endoscopic transfer. To focus on the dierent questions that arise, it will be best to separate the problem into three distinct steps. The rst is to normalize intertwining operators between representations induced from dierent parabolic subgroups. We begin by introducing a slightly nonstandard family of normalizing factors. Recall that if is an N -dimensional representation of the local Langlands group, we can form the local L-function Lps, q and -factor ps, , F q. The -factor follows the notation [T2, (3.6.4)] of Langlands, and depends on a choice of nontrivial additive character F of F . For p-adic F , it has the general form ps, , F q p, F q qF for a nonzero complex number ` p, F q 1 , , F , 2
nps 1 q 2

and an integer n np, F q. The integer n is the simpler of the two constants, since it can be expressed explicitly in terms of the Artin conductor of and the conductor of F [T2, (3.6.4), (3.6.5)]. The quotient ` 1 n (2.3.1) p, F q p0, , F q 1 , , F pqF q 2 2 is therefore a more elementary object than the -factor. r Suppose now that G represents an element in Esim pN q, and that M is a xed Levi subgroup of G. Intertwining operators will be attached to parameters and representations for M . For the time being, therefore, we

80

2. LOCAL TRANSFER

shall denote such objects by symbols , , , etc., that have hitherto been reserved for G. r In particular, suppose that P pM q is a Langlands parameter for M . The normalizing factors of include special values of local L-functions, which are only dened for parameters in general position. To take care of this, we let be a point in general position in the vector space a X pM qF b C. M,C The associated twist pwq pwq|w| , w P LF , of is then in general position. Suppose that P and P 1 belong to PpM q, the set of parabolic subgroups of G over F with Levi component M . We write P 1 |P for the adjoint representation of L M on the quotient pP 1 {pP 1 X pP , n n n p where pP 1 denotes the Lie algebra of the unipotent radical of P 1 . The comn position (2.3.2) _1 |P P

of with the contragredient of P 1 |P is of course a nite dimensional representation of LF . We dene a corresponding local normalizing factor rP 1 |P p q rP 1 |P p , F q as the quotient (2.3.3) Lp0, _1 |P q p_1 |P , F q1 Lp1, _1 |P q1 . P P P
1 1 M GLpN1 q GLpNr1 q G ,

Both M and have parallel decompositions (2.3.4) and r 1 1 1 , P pG q, 1 r 1 , . . . , N 1 and N such that for nonnegative integers N1 r1
1 1 2N1 ` ` 2Nr1 ` N N.

r G P Esim pN q,

(Keep in mind that each general linear factor of M has a diagonal embedding g g g _ into G.) We shall ultimately argue by induction on N . This means that if M is proper in G, we will be able to assume that the local Langlands correspondence of Theorem 1.5.1 holds for the factor G of M . r In particular, we will be able to assume that any P pM q belongs to a r unique L-packet , and thereby write (2.3.5) rP 1 |P p q rP 1 |P p q. In case F is archimedean, the notation here matches that of [A7], provided that F is taken to be the standard additive character [T2, (3.2.4), (3.2.5)]. For in this case, the general -factors are independent of s. The general

2.3. NORMALIZED INTERTWINING OPERATORS

81

quotient (2.3.1) is then equal to 1, and can be removed from the denition (2.3.3). The formula (2.3.5) requires a word of explanation. It is not a denition, since the right hand side is dened (2.3.3) in terms of local Artin L-functions and -factors, while the left hand side is the corresponding quotient Lp0, , _1 |P q p , _1 |P q1 Lp1, , _1 |P q1 P P P of representation theoretic objects. This in turn is dened as the quotient Lp0, , , _1 |P q p, , _1 |P q1 Lp1, , , _1 |P q1 P P P attached to the product
1 1 M 0 pN q GLpN1 q GLpNq q GLpN q

of general linear groups, where is the representation of M 0 pN, F q given by , and where the nite dimensional representation P 1 |P is identied with its transfer from L M to the dual group of M 0 pN q. The Rankin-Selberg constituents of each side of (2.3.5) are equal, thanks to the local classication for GLpN q. The symmetric and exterior square constituents are more subtle. However, it is known that their L-functions and -factors are equal [He2], even though at the time of writing, this has not been established for their -factors. The identity (2.3.5) is therefore valid. We need a more general formulation, which applies to nongeneric parameters . Suppose rst is an n-dimensional representation of the product of LF with SU p2q. Then extends to a representation of the product of LF with the complex group SLp2, Cq. Its pullback 1 |u| 2 0 puq u, , u P LF , 1 0 |u| 2 becomes an N -dimensional representation of LF . We dene Lps, q Lps, q, ps, , F q ps, , F q, and p, F q p , F q. In following standard notation here, we have had to use the symbol in two dierent roles. To minimize confusion, we shall always include the subscript F whenever denotes an additive character. r Suppose now that P pM q is a general parameter for the Levi subgroup 1 P PpM q as above, the meromorphic function M . For P, P (2.3.6) rP 1 |P p q rP 1 |P p q rP 1 |P p, q of P a depends implicitly on the additive character F . It will serve as M,C our general normalizing factor.

82

2. LOCAL TRANSFER

The basic unnormalized intertwining operators are dened as for example in [A7, (1.1)]. We take to be an irreducible unitary representation of M pF q, with twist pmq pmqepHM pmqq , m P M pF q,

by P a . We then have the familiar operator M,C JP 1 |P p q : HP pq HP 1 pq that intertwines the induced representations IP p q and IP 1 p q. It is dened as a meromorphic function of by analytic continuation of an integral, taken over the space (2.3.7) NP 1 pF q X NP pF qzNP 1 pF q,

which converges absolutely for Repq in an ane chamber in a . (See [A7, M 1].) We do have to specify the underlying invariant measure on (2.3.7). The standard splitting on the quasisplit group G determines a Chevalley basis, and hence an invariant F -valued dierential form of highest degree on the quotient (2.3.7). Its absolute value, together with the Haar measure on F that is self-dual with respect to F , then determines an invariant measure on (2.4.7). This is the measure that we take in the denition of JP 1 |P p q. It follows easily from [T2, (3.6.5)] that the dependence of JP 1 |P p q on F is parallel to that of the normalizing factor rP 1 |P p q. We leave the reader to check that the measure is compatible with the convention in [A7] for real groups. In other words, if F is archimedean and F is the standard additive character, the canonical measure dened on p. 31 of [A7] in terms of a certain bilinear form coincides with the measure on (2.3.7) chosen here. We will therefore be able to apply the results of [A7] without modication. Our version of the normalized intertwining operators will require an assumption of the kind mentioned above. We assume that M is proper in G, and that the local Theorem 1.5.1 holds if G is replaced by M . There is of course nothing new to prove for the general linear factors of M . The assumption refers specically to the factor G of M . It will later be treated as an induction hypothesis, applied to the integer N N . We can then r assume the existence of the packet of (OutN pG q-orbits) of representar tions of M pF q for every P pM q. We dene the normalized intertwining r r operator attached to any P pM q and P by (2.3.8) RP 1 |P p , q rP 1 |P p q1 JP 1 |P p q.

It is independent of the additive character F . We note that (2.3.8) diers in some respects from the standard normalized intertwining operator, conjectured in general in [L5, p. 281282], and established for archimedean F in [A7]. For example, the -factor in (2.3.3) diers from the -factor of [L5]. This is essentially because the remarks of

2.3. NORMALIZED INTERTWINING OPERATORS

83

[L5] apply to elements w in the relative Weyl group W pM q W G pM q NormpAM , Gq{M, or more precisely, representatives of such elements in GpF q, rather than to pairs pP, P 1 q of parabolic subgroups. (We have already noted that the factor reduces to 1 in the archimedean case of [A7].) Another point is that here represents an orbit of irreducible representations of M pF q under the group r r r OutN pM q OutP pM q OutN pG q. However, any use we make of (2.3.8) will be a context that is independent of the representative of the orbit. Finally, the normalizing factor (2.3.6) is dened in terms of the L-function and -factor of , rather than (which is to say, the Langlands parameter such that lies in ). If is not generic, the two possible normalizing factors can dier. The nonstandard form (2.3.6) turns out to be appropriate for the comparison of trace formulas. Proposition 2.3.1. Assume that the local and global theorems are valid if N is replaced by any integer N N . Then the operators (2.3.8) satisfy the multiplicative property (2.3.9) (2.3.10) RP 2 |P p , q RP 2 |P 1 p , qRP 1 |P p , q, RP 1 |P p , q RP |P 1 p , q. for any P , P 1 and P 2 in PpM q, as well as the adjoint condition In particular, RP 1 |P p , q is unitary and hence analytic if is purely imaginary, and the operator RP 1 |P p, q RP 1 |P p0 , 0 q is therefore dened. Proof. Suppose rst that is generic, which we recall means that r r it is trivial on the extra factor SU p2q. In this case, is an L-packet, and (2.3.8) is the normalized intertwining operator r RP 1 |P p q rP 1 |P p q1 JP 1 |P p q, P , studied for real groups in [A7]. Since the required assertions were established for real groups in [A7], we can assume that F is p-adic here. Moreover, by standard reductions [A7, p. 29], it suces to consider the case that represents an element in the set 2 pM q of representations of M pF q that are square integrable modulo the center. In fact, we shall see that it suces to consider the case that the general linear components of are supercuspidal. Imposing these conditions on , we write ` r (2.3.11) 1 1 1 , i scusp GLpN 1 q , P 2 pG q,
1 r i

for its decomposition relative to (2.3.4). To handle this basic case, we shall use global means. Having reserved 9 the symbol F in this chapter for our local eld, we write F for a global eld.

84

2. LOCAL TRANSFER

u 9 If u is any valuation of F , we write S8 puq S8 Y tuu and S8 S8 tuu, 9 where S8 is the set of archimedean valuations of F . We shall apply the 1 of . following two lemmas to the components and i

r Lemma 2.3.2. Given the local objects F, G P Esim pN q and 9 , G P Esim pN q 9 r P 2 pM q, we can nd corresponding global objects F 9 9 and P 2 pG q, together with a valuation u of F , with the following prop9 erties. 9 9 (i) pF, G , q pFu , G,u , ,u q 9 9 (ii) For any v R S8 puq, ,v has a vector xed by a special maximal 9 9 compact subgroup of G pFv q. We shall establish a stronger version of Lemma 2.3.2 in 6.2, as an application of the simple invariant trace formula. In the meantime, we shall take the lemma for granted. Lemma 2.3.3 (Henniart, Shahidi, Vignras). We can choose the objects e ` 9 F and u of the last lemma so that the other local factors i P scusp GLpNi1 q of also have global analogues. More precisely, for each i we can nd a 9 cuspidal automorphic representation i of GLpNi1 q over F such that i 9 i,u , and such that for any v R S8 puq, i,v has a vector xed by a maximal 9 9 1 , F q. 9v compact subgroup of GLpNi This lemma is a consequence of the fact that i has a Whittaker model. In the form stated here, it is Proposition 5.1 of [Sha4], which applies to a generic representation of any quasisplit group over F . We note, however, that need not be generic, which is why we are using Lemma 2.3.2 in place of the analogue of Lemma 2.3.3 for G . Applying the lemmas to the components of the representation (2.3.11), we obtain an automorphic representation 1 r 1 9 91 91 9 of the group 9 91 91 9 M M1 Mr 1 G , 9 9 Mi1 GLpNi1 {F q,

9 9 such that u . We can identify M with a Levi subgroup of a group G over 9 9 r 9 F , which represents a simple endoscopic datum in Esim pN q over F , and has 9 u G. We then have the global induced representation the property that G 9 IP pq, 9 P P PpM q,

and the global intertwining operator, dened by analytic continuation in of the product 1 9 9 MP 1 |P p q JPv |Pv pv, q,
v

9 9 from IP p q to IP 1 p q.

2.3. NORMALIZED INTERTWINING OPERATORS

85

The operator MP 1 |P p q here is a minor modication of the global op9 erator that is at the heart of Langlands theory of Eisenstein series. In this setting Langlands functional equation takes the form (2.3.12) MP 2 |P p q MP 2 |P 1 p qMP 1 |P p q. 9 9 9 (See for example [A3, 1].) We shall combine it with what can be obtained from the intertwining operators at places v u. We can write (2.3.13) 9 9 9 MP 1 |P p q rP 1 |P p q RP 1 |P p q,

for the normalized global intertwining operator 9 9 (2.3.14) RP 1 |P p q RP 1 |P pv, q,

v

and the global normalizing factor dened by analytic continuation in of a product rP 1 |P p q 9 rP 1 |P pv, q. 9
v

The factor of v in this last product is our canonical normalizing factor for 9 9 the generic parameter v . It is dened by the analogue for v, of (2.3.3) and (2.3.5). The existence of the corresponding L-function and -factor is guaranteed by the implicit assumption that Theorem 1.5.1 holds for the 9 9 localization G,v of G . We note that the measures and additive characters used to dene the local factors in (2.3.14) can be chosen to be compatible with corresponding global objects. This follows from the well known fact [T1, (3.3)] that if each Fv is obtained from a xed, nontrivial additive 9 9 9 9 character F on A{F , the associated product measure on A is self-dual 9 9 F. 9 (relative to F ) and assigns volume 1 to the quotient A{ 9 We claim that (2.3.15) RP 2 |P pv, q RP 2 |P 1 pv, qRP 1 |P pv, q, 9 9 9

9 for any valuation v u of F . For archimedean v, this is the concrete form of [A7, Theorem 2.1(R2 )], established as Proposition 3.1 in the Appendix 9 of [A7]. For p-adic v, we use the property (ii) of ,v in Lemma 2.3.2. Combined with the theory of p-adic spherical functions [Ma] and Whittaker models [C], it implies that ,v has a Whittaker model. Since for any i, 9 9 the generic representation i,v of GLpNi , Fv q also has a Whittaker model, 9 9 9 the same is true of the representation v of M pFv q. We can therefore apply 9 9 Shahidis results [Sha4] to the induced representation IP pv, q. They imply that Langlands conjectural normalizing factors have the desired properties. In particular, the formula (2.3.15) does hold for v, as claimed. (We will discuss Shahidis results further in 2.5.) The global normalizing factors have an expression 9 9 9 9 rP 1 |P p q Lp0, , _1 |P q p , _1 |P q1 Lp1, , _1 |P q1 P P P

86

2. LOCAL TRANSFER

in terms of global L-functions and -factors 9 9 p , _1 |P q p0, , _1 |P q P P We claim that they satisfy the identity (2.3.16) 9 9 9 rP 2 |P p q rP 2 |P 1 p qrP 1 |P p q. By standard reductions, it suces to consider (2.3.16) when the Levi 9 9 subgroup M of G is maximal. (See [A7, 2] for example.) In this case there 9 are only two groups P and P in PpM q, and it suces to take P 1 P and 2 P , since r 9 P P |P p q 1. But P |P is just the direct sum of the standard 1 9 (tensor product) representation of GLpN1 , Cq L G with either the sym1 metric square or skew symmetric square representation of GLpN1 , Cq. We 9 are assuming that the global theorems hold for G . This allows us to express _ q as the product of a Rankin-Selberg L-function with either a 9 Lps, , P |P symmetric square or skew symmetric square L-function. In particular, this L-function has analytic continuation and functional equation 9 9 9 Lps, , _|P q ps, , _|P qLp1 s, , P |P q. P P It follows that 9 9 9 rP |P p q Lp1, , P |P qLp1, , _|P q1 P Similarly, we have 9 9 rP |P p q Lp1, , P |P qLp1, , _|P q1 9 P `1
_ 9 2 , , P |P

`1

1 _ . 2 , , P 1 |P

`1

9 , , _|P . 2 P .

Since _|P P |P , the product of the four L-functions equals 1. Another P application of the functional equation tells us that the product of the two -factors also equals 1. The product of the two normalizing factors thus equals 1. In other words, the identity (2.3.16) holds in the special case at hand, and hence for any M , P 1 and P , as claimed. 9 Consider the product expression for the operator MP 1 |P p q given by (2.3.13) and (2.3.14). We need only combine the multiplicative property (2.3.12) of this operator with its analogues (2.3.16) and (2.3.15) for the scalar factor in (2.3.13) and the factors with v u in (2.3.14). It follows that the same property holds for the remaining factor, that corresponding to v u in (2.3.13). This is the required specialization RP 2 |P p q RP 2 |P 1 p qRP 1 |P p q of (2.3.9) to the case that is generic, and is of the form (2.3.11). It is easy to relax the conditions on . From the remarks leading to [A7, (2.2)], we see that the last product formula remains valid for representations (2.3.11) whose general linear ` components i are induced super cuspidal. Since any representation in 2 GLpNi q is a subrepresentation of such a i , and the corresponding normalizing factors can be related by [A7, r Proposition 6.2], the product formula holds for any P 2 pM q. Similar rer marks then allow us to extend the formula to any representation P pM q.

2.3. NORMALIZED INTERTWINING OPERATORS

87

(See [A7, 2], for example.) The required formula (2.3.9) is thus valid if r is any parameter in pM q. Suppose now that the local parameter for M is arbitrary. The normalizing factor in (2.3.8) is attached to rather than . This is the distinction noted prior to the statement of the proposition. It is pertinent because the r set to which belongs will not in general be an L-packet. In other words, the local Langlands parameter of may be dierent from the local Langlands parameter of . By our assumption on G , we are free to regard both and as local Langlands parameters for a product of general linear groups. The normalizing factor (2.3.6) to which the proposition applies is taken relative to . The extension of the required identity (2.3.9) from tempered representations to arbitrary Langlands quotients, which was established for any F on p. 30 of [A7], is implicitly based on the parameter . We must show that if (2.3.9) holds for the one set of normalizing factors, it also holds for the other. As Langlands parameters attached to a product of general linear groups, and correspond to irreducible representations and of the corresponding products of groups over F . The two representations are related by the algorithm of Zelevinsky, used in the proof of Lemma 2.2.2, (and its analogue [AH] for F R.) In other words, and are block equivalent, in the sense of D. Vogan. (See [A7, p. 42].) It then follows from [A7, Proposition 5.2] that the quotient rP 1 |P p , q rP 1 |P p q1 rP 1 |P p q of the two normalizing factors satises the analogue rP 2 |P p , q rP 2 |P 1 p , qrP 1 |P p , q of (2.3.9). This implies that the two formulations of the identity (2.3.9), the one here and the original version in [A7], are equivalent. The identity therefore holds as stated. We still have the adjoint condition (2.3.10) to verify. It is a straightforward matter to establish a general identity
JP 1 |P p q JP |P 1 p q,

for any irreducible representation of M pF q with adjoint . One rst takes to lie in the domain of absolute convergence of the integral that denes JP 1 |P p q. One obtains the formula in this case by a suitable change of variables in the quasi-invariant measure on P 1 pF qzGpF q determined by the inner product on HP pq. The formula for arbitrary then follows by r analytic continuation. In (2.3.10), lies in the packet , and is hence unitary by our assumption that Theorem 1.5.1 holds for G . (We of course know that the unitary condition holds also for the remaining, general linear factors of M .) In other words, . It is also not dicult to see that rP 1 |P p q rP |P 1 p q,

88

2. LOCAL TRANSFER

r for P pM q as in (2.3.10). In the special case that is generic, this follows from the explicit formulas in [T2] for the terms in (2.3.3). For r r a general parameter P pM q, we write the parameter P pM q as a r twist , where P temp pM q is the restriction of to the subgroup LF of LF SU p2q, and is a point in a . Since the Langlands parameter is M x M -conjugate to , we obtain rP 1 |P p q rP 1 |P p` q rP |P 1 p q rP |P 1 p q rP |P 1 p q, as claimed. The adjoint condition (2.3.10) then follows from the denition (2.3.8), since is assumed to be unitary. The identity (2.3.9) reduces to RP |P 1 p q RP 1 |P p q1 . when P 2 P . If we take to be purely imaginary in (2.3.10), we then see that RP 1 |P p , q RP |P 1 p , q RP |P 1 p , q RP 1 |P p , q1 . In other words RP 1 |P p q is unitary, as noted in the last assertion of the proposition. The proof of Proposition 2.3.1 is complete. We have constructed normalized intertwining operators RP 1 |P p, q between induced representations IP pq and IP 1 pq. This is just the rst step. It remains to convert these objects to self-intertwining operators of IP pq, which will be attached to elements w P W pM q that stabilize . Suppose that P P PpM q, P unit pM q and w P W pM q are xed. We can certainly choose a representative w of w in the normalizer N pM q of M r in GpF q. This gives us another representation r r pwqpmq pw1 mwq, m P M pF q,

of M pF q on the underlying space V of , which depends only on the image r of w in the quotient of N pM q by the center of M pF q. It also gives us an intertwining isomorphism r pw, q : HP 1 pq HP pwq from IP 1 pq to IP pwq by left translation ` r r (2.3.17) pw, q1 pxq 1 pw1 xq, where P 1 w1 P w1 P w1 . r r If w is equivalent to , we can in addition choose an intertwining isomorr phism pwq of V from w to . Its pointwise action on HP pwq intertwines 1 P HP 1 pq, x P GpF q,

2.3. NORMALIZED INTERTWINING OPERATORS

89

IP pwq with IP pq. The composition r pwq pw, q RP 1 |P p, q, r of intertwining maps HP pq HP 1 pq HP pwq HP pq will then be a self intertwining operator of IP pq. It is in the choices of w r r and pwq that we will need to be aware of the implications for endocopic transfer. We will deal with the rst at the end of this section, and the second in the beginning of the next. In each of these remaining two steps, we will also have to adjust the relevant intertwining map by its own scalar normalizing factor. r There is a natural way to choose the representative w of w. The group G comes with a splitting ` S T, B, tx u . We assume that M and P are standard, in the sense that they contain T and B respectively. Let wT be the representative of w in the Weyl group WF pG, T q that stabilizes the simple roots of pB X M, T q. We then take r r w wS to be the representative of wT in GpF q attached to the splitting S, as on p. 228 of [LS1]. Thus, w w1 wr , r r r where wT w1 wr is a reduced decomposition of wT relative to the simple roots of pG, T q, and r w exppX q exppX q exppX q, in the notation of [LS1]. r However, our representative w introduces another diculty. For it is well known that the mapping w w from W pM q to N pM q is not multiplicative r in w. The obstruction is the co-cycle of [LS1, Lemma 2.1.A], which plays a critical role in the construction transfer factors. To compensate for it, we ` 1 of must introduce the factor 2 , that was taken out of the original denition (2.3.1). For global reasons, we will use the representation theoretic -factor, given by the representation of the product M 0 pN, F q of general linear r groups. We are assuming now that belongs to the packet attached to r a given P pM q. Since M is proper, we can assume as before that the packet exists, and that is unitary. The choice of P allows us to identify x x x W pM q with the Weyl group W pM q of M . (We are regarding M here as a x complex group with an L-action of the Galois group F , so that W pM q x x is the group of -invariant elements in the quotient NormG pM q{M .q We p dene the -factor `1 w P W pM q. (2.3.18) P pw, q 2 , , _1 P |P , F , w r P ,

90

2. LOCAL TRANSFER

We know from the general results of Shahidi that we must also include a supplementary term. It is the -factor (2.3.19) pwq pw, F q a pF q, w P W pM q,
a

of [KeS, (4.1)] and [Sha4, (3.1)], in which a ranges over the reduced roots of pB, AB q such that wT a 0, and AB is the split component of B (or T ) over F . This term is built out of the complex numbers pE{F, F q attached to nite extensions E{F by Langlands [L3] to account for the behaviour of -factors under induction. In the case at hand, a pF q pFa {F, F q, where Fa is either a quadratic extension of F or F itself. This eld is given by ` Ga,sc ResFa {F SLp2q , where Ga,sc is the simply connected cover of the derived group of Ga,sc , the Levi subgroup of G of semisimple rank one attached to a. It is known that neither P pw, q or pwq are multiplicative in w. However, the next lemma tells us that their combined obstruction matches that r of P pw, q. Lemma 2.3.4. The product (2.3.20) pw, , q pwq1 P pw, q pw, q, r pw1 w, , q pw1 , w, wq pw, , q, w P W pM q, w1 , w P W pM q. satises the condition r r Proof. We consider general elements w1 and w in W pM q. Since w1 w 1w are two representatives in M pF q of w1 w, and since they preserve and w the same splitting of M , they satisfy r r w1 w w1 w zpw1 , wq, for a point zpw1 , wq in the center of M pF q. It follows that ` pw1 w, q zpw1 , wq pw1 , wq pw, q, r r where is the central character of . Suppose for a moment that M equals the minimal Levi subgroup T . The elements w1 and w then belong to the rational Weyl group W pT q WF pT q of T . We write B pw1 , wq for the set of roots P B of pB, T q such that w R B and w1 w P B . The corresponding sum pw1 , wq _ , P B pw1 , wq,

is a co-character in X pT q, which is xed by the Galois group F , since both w1 and w are dened over F . It then follows from [LS1, Lemma 2.1.A] that 1 zpw1 , wq p1qpw ,wq ,

2.3. NORMALIZED INTERTWINING OPERATORS

91

where the right hand side is understood as a point in ` T pF q X pT q b F . Suppose now that M is arbitrary, subject only to our requirement that P P PpM q contains B. For any w P W pM q, we have constructed the rational Weyl element wT P W pT q. The denitions tell us that zpw1 , wq p1qpw ,wq , where pw1 , wq zpw1 , wq
1 1 In other words, zpwT , wT q and pw1 , wq pwT , wT q. We claim 1 , wq belongs to the subgroup X pA q of co-characters of the split that pw M component AM of the center of M . To see this, consider an element u in M 1 the rational Weyl group W0 of pM, T q. If is any root in B pwT , wT q, u `obviously belongs to B . Moreover, we see that wT puq R B and 1 1 wT wT puq P B , since wT and wT wT each conjugate u to another element M . In other words, u also lies in pw 1 , w q. The Weyl group W M in W0 B T 0 T 1 thus acts by permutation on B pwT , wT q, and consequently xes the sum pw1 , wq. The claim then follows from the fact that pw1 , wq is also xed by . In particular, zpw1 , wq lies in the group AM pF q X pAM q b F . The split torus AM is the product of the centers of the general linear factors of M in (2.3.4). We can therefore write ` ` zpw1 , wq zpw1 , wq ,
1

w1 , w P W pM q,
1 P B pwT , wT q.

_ ,

where is the central character of the representation , since and are equal on these general linear factors. We have established an obstruction ` _ 1 pw1 w, q pw, q1 pw1 , wq1 p1q pw ,wq r r r for the operator pw, q to be multiplicative in w. Before we compare it with the corresponding obstructions for the other two factors in (2.3.20), let us rst take note of its L-group interpretation. p The complex dual torus AM of AM can be identied with the split x x part of the cocenter of M , namely the largest connected quotient of M on L M by conjugation is trivial. This provides the second which the action of of the two canonical isomorphisms p x X pAM q X pAM q X pM qF . x x Moreover, we can extend any character on M in the set X pM qF to a L M that is trivial on the semidirect factor W L M . If character on F of r P pM q is a Langlands parameter with central character on AM pF q, it is then easy to see that ` _ (2.3.21) px q p_ qpxq _ puq , x P F , _ P X pAM q,

92

2. LOCAL TRANSFER

where _ is identied with a character on L M , and u is any element in WF ab whose image in the abelian quotient F WF equals x. Let _ pw1 , wq _

be the natural decomposition of the character _ _ pw1 , wq into a sum over roots in the subset ( p p r pw1 , wq P r : w 0, w1 w 0 P P p p x of the reduced roots r of pP , AM q. The summand of thus equals P _ _ , P ,

where is the subset of roots of pG, T q whose coroot _ restricts to a positive multiple of on AM . The components _ of _ pw1 , wq each belong x to the group AM , by virtue of the argument we applied above to _ pw1 , wq. x r It follows that the obstruction for pw, q can be written as a product ` _ _ 1 (2.3.22) p1q pw ,wq p qp1q

p over P r pw1 , wq. P p For any P r , let _ be the adjoint representation of P the subspace p p_ , n n P ,

LM

on

of pP attached to . Th -term in (2.3.20) then has a decomposition n p P pw, q P _ , w 0. 1 , , , F , P 2

This is similar to the decomposition we have come to expect of normalizing factors. For example, it tells us that the obstruction for P pw, q to be multiplicative in w, which in general we write in the inverse form epw1 , wq P pw1 , wq P pw, q P pw1 w, q1 , is trivial if the length of w1 w equals the sum of the lengths of w1 and w. For arbitrary elements w1 and w in W pM q, we observe from this decomposition that 1 1 p P r pw1 , wq. epw1 , wq 2 , , , P 2 , , , F , P

For each , the product of the two representation theoretic -factors here equals the corresponding product of Artin -factors. It then follows from the general property [T2, (3.6.8)] of Artin -factors that 1 1 , , , F 2 , , , F pdet qp1q, 2

2.3. NORMALIZED INTERTWINING OPERATORS

93

where the right hand side is understood as the value of the determinant of ab at any u P WF whose image in F WF equals p1q. We can therefore write the obstruction for P pw, q as a product (2.3.23) epw1 , wq pdet qp1q

over P Suppose for example that G is split. Then the supplementary term p pwq in (2.3.20) is trivial. Moreover, for any P r , we can regard as a P x representation of the complex connected group M L M on p . Its weights n p are the characters _ parametrized by P , each of which occurs on T p with multiplicity 1. It follows that the restriction of to T has determinant x equal to the character _ . Since every element in M is conjugate to an x p element in T , the characters _ and det on M are equal. It follows that p p_ qp1q pdet qp1q, P r pw1 , wq, P

r pw1 , wq. P

so that the obstructions for the two primary terms in (2.3.20) match. The lemma is therefore valid in case G is split. r r Assume now that G P Esim pN q is a general element in Esim pN q. Our remaining concern is the nonsplit group G SOp2n, E{F q, where G E{F is the quadratic character on F attached to a xed quadratic extension E{F . This is the case in which the supplementary term pwq in (2.3.20) is nontrivial. We recall that pwq is given by a product (2.3.19) of -factors pFa {F, F q, taken over the reduced roots a of pB, AB q with wT a 0. Each a determines an orbit tu of roots of pG, T q under the Galois group F . The eld Fa is the extension of F that corresponds to the stabilizer of a in . It equals either F , in which case the associated factor pF {F, F q is trivial, or the quadratic extension E, in which case the factor satises pE{F, F q2 E{F p1q. (See [KeS, (2.7)].) It follows that the obstruction for pwq to be a character in w equals the product (2.3.24) pw1 wq pwq1 pw1 q1 E{F p1q

over P where is the number of roots a with Fa E such that the associated orbit tu is contained in . We need to combine this with the other two obstructions (2.3.23) and (2.3.22). r x For the general group G P Esim pN q, the characters _ and det on M p attached to any P P remain equal. However, we must also consider their x extensions to the nonconnected group L M M E{F . As an element _ in X pA q, following the general identity (2.3.21), _ has a canonical M extension that is trivial on the factor E{F of L M . The character det is

r pw1 , wq, P

94

2. LOCAL TRANSFER

the determinant of a representation that is dened a priori on L M . It is generally not trivial on E{F . p Fix the root P r . The quotient E{F of WF acts by permutation on P p_ in pP , in a way that is compatible with the natural the coroot spaces n n action of E{F on the roots P . To describe the extension of det to L M , it suces to consider the case that P is a maximal parabolic subgroup of the group G, which we are implicitly assuming equals SOp2n, E{F q. The Levi component M of P then takes the form M GLpmq SOp2n 2m, E{F q, The set of roots can be written as ( ij ptq ti ptj q1 : 1 i m 1, m j n , where t represents a diagonal matrix pti q in the standard maximal torus T of G. The group E{F acts simply transitively on any pair
` pi pin , in q,

m n.

i m 1,

and acts trivially on each of the remaining roots of . Consequently, the determinant of the restriction of to the factor E{F of L M equals a product of copies of the nontrivial character of E{F , taken over the set of pairs pi . It follows that det p1 q p1qm1 , for the nontrivial element P E{F . This will allow us to compare the characters _ and det on the factor E{F of L M . Moreover, it is clear from the denitions that m1 equals the integer assigned to in (2.3.24). In comparing the two primary obstructions (2.3.22) and (2.3.23), we need ab to choose an element u in the Weil group WF whose image in F WF equals p1q. We write puq m puq puq x where m puq belongs to M , and puq lies in the quotient E{F of WF . We are assuming here that G is not split, so that E is dened as a quadratic extension of F . We then observe that the element puq P t1u equals E{F p1q, a property that follows from the fact that E{F p1q equals 1 if and only if the element p1q P F lies in the image of the norm mapping p from E to F . For any P r pw1 , wq, we then have P ` pdet qp1q det puq ` ` det m puq det 1 puq ` _ m puq E{F p1q ` _ puq E{F p1q p_ qp1q E{F p1q .

2.3. NORMALIZED INTERTWINING OPERATORS

95

It remains only to take the product over of the numbers in each side of the identity we have obtained. We conclude that the obstruction (2.3.23) equals the product of the original obstruction (2.3.22) with the supplementary obstruction (2.3.24). It then follows from the denitions that the threefold product (2.3.20) is indeed multiplicative in w, as required. Remarks. 1. The operator pw, , q makes sense for an arbitrary pair pG, M q, under the appropriate hypothesis on the local Langlands correspondence for M . The proof of the lemma should carry over to the general case, although I have not checked the details. I thank the referee for pointing out that pwq is generally not a character in w. 2. It is interesting to see the local root numbers (2.3.18) appearing as explicit factors of the operators (2.3.20). They remind us of the global root numbers (1.5.7) used to construct the sign character (1.5.6). As orthogonal root numbers, however, they are less deep, and better understood [D1] than the symplectic root numbers (1.5.7) and their local analogues. We can think of Lemma 2.3.4 as a normalization of the miniature r intertwining operator pw, q. It is reminiscent of the main property (2.3.9) of Proposition 2.3.1, if somewhat simpler. We combine the two normalized operators by writing (2.3.25) RP pw, , q pw, , qRw1 P |P p, q, r RP pw, , q rP pw, q1 JP pw, q, JP pw, q pw, qJw1 P |P p q r r from HP p q to HP pw q, and a corresponding normalizing factor rP pw, q pwqP pw, q1 rw1 P |P p q that reduces to the product of pwq with the familiar quotient (2.3.27) Lp0, , , _1 P |P qp0, , , _1 P |P , F q1 Lp1, , , _1 P |P q1 . w w w We are using the -variant of (2.3.5) here, while continuing to suppress the implicit dependence on the additive character F from the notation. It follows from Proposition 2.3.1 and Lemma 2.3.4 that the normalized intertwining operator RP pw, , q : HP pq HP pwq satises the familiar co-cycle relation (2.3.28) RP pw1 w, , q RP pw1 , w, wqRP pw, , q, w1 , w P W pM q. w P W pM q.

This operator then has a decomposition (2.3.26)

for the unnormalized intertwining operator

This is essentially what has been conjectured in general by Langlands [L5, Appendix II].

96

2. LOCAL TRANSFER

2.4. Statement of the local intertwining relation The purpose of this section is to state a theorem that relates local intertwining operators with endoscopic transfer. The relation will have global implications for our later analysis of trace formulas. It will also be closely tied to Theorem 2.2.1, both as a supplement and as a part of the proof. r We are assuming that G belongs to Esim pN q, and that M is a proper Levi subgroup of G. There is still some unnished business from the last section. It is to convert the intertwining operator RP pw, , q from IP pq to IP pwq to a self-intertwining operator for IP pq. This is the last step in the normalization process, and the one that is most closely tied to endoscopy. r r Recall that lies in pM q, and that belongs to the packet . We will usually take w to be in the subgroup (2.4.1) W pq W pq X W pM q tw P W pM q : w , w u

x of elements w in W pM q W pM q that stabilize the equivalence classes of both and . We will then be able to introduce an intertwining operator r pwq from w to . r We are assuming that lies in the packet . Suppose that w lies in W pq, and that pwq is an intertwining operator from w to . We can r r think of as an extension of to the M pF q-bitorsor Mw pF q M pF q w, r ` in the sense of (2.1.1), or even an extension of to the group Mw pF q r generated by Mw pF q, if pwq is chosen so that its order equals that of w. The product (2.4.2) r r RP pw, , q pwq RP pw, , q is then a self-intertwining operator of IP pq, which of course depends on the r r choice of pwq. Suppose that the operators pwq on V could be chosen to be multiplicative in w. It would then follow from (2.3.28) and the denition of RP pw, , q that the mapping w RP pw, , q, r w P W pq, is a homomorphism from W pq to the space of intertwining operators of IP pq. However, there is no canonical choice for pwq in general. While r this does not necessarily preclude an ad hoc construction, it represents a problem that left unattended would soon lead to trouble. The solution is provided by an application of Theorem 2.2.4 to M . We rst recall the various nite groups that can be attached to the r parameter . We are regarding as an element in pM q. However, its L M into L G is an element in composition with the standard embedding of r pGq, which we shall also denote by . From these two objects, we obtain the two complex reductive groups S pM q S pGq S ,

2.4. STATEMENT OF THE LOCAL INTERTWINING RELATION

97

and their two nite quotients S pM q S pGq S . Recall that S is the group of connected components in the quotient p S S {ZpGq . It is clear that S pM q is the subgroup of elements in S that leave the complex torus ` 0 x AM ZpM qF x pointwise xed, and it is easily seen that S pM q is the image of S pM q in S . There are other interesting nite groups as well. We write N pG, M q for the normalizer of the complex torus AM in S and N pG, M q for the group x of components in the quotient p N pG, M q N pG, M q{ZpGq . Then S pM q is a normal subgroup of N pG, M q. Its quotient is isomorphic to the Weyl group W pG, M q W pS , AM q, x which is to say, the group of automorphisms of AM induced from S . We x 0 write W pG, M q to be the normal subgroup of automorphisms in W pG, M q that are induced from the connected component S 0 , and
0 R pG, M q W pG, M q{W pG, M q

for its quotient. On the other hand, there is a canonical injection from 0 W pG, M q into N pG, M q. This map is well dened, since the centralizer of 0 0 AM in S is connected, and is therefore contained in S pM q. In other words, x 0 pG, M q can also be regarded as a normal subgroup of N pG, M q. The W 0 quotient S pG, M q of N pG, M q by W pG, M q is a subgroup of S . We can summarize the relations among these nite groups in a commutative diagram 1 1

0 0 W pG, M q W pG, M q

>

(2.4.3)

>

S pM q }

>

N pG, M q

W pG, M q

>

>

S pM q

>

S pG, M q

>

R pG, M q

>

98

2. LOCAL TRANSFER

of short exact sequences. The dotted arrows stand for splittings determined by the parabolic subgroup
0 p p P pP X S q{ZpGqF

of S , or rather the chamber of P in the Lie algebra of AM . x Suppose that u belongs to N pG, M q. We write wu for the image of u in W pG, M q given by the horizontal short exact sequence at the center x of (2.4.3). Since wu stabilizes AM , it normalizes M , and can therefore be x x treated as an element in W pM q. The choice of P then allows us to identify wu with an element in W pM q. Recall that M is a product of groups (2.3.4), all of which all have standard splittings. We set r Mu Mwu pM, wu q, where we recall that wu is the automorphism in the inner class of w that r preserves the corresponding splitting of M . Then Mu is a twisted group, r and belongs to the corresponding subset pMu q of pMu q. As earlier, the centralizer quotient r S,u S pMu q is an S pM q-bitorsor. We shall write u for the element in this set dened r by u. r Assume that represents a wu -stable element in the packet r pM q. We can then choose an intertwining operator pwu q from r r wu to , from which we are led to the intertwining operator RP pwu , , q r of IP pq by (2.4.2). As we have said, there is generally no canonical choice for pwu q. However, M is a product (2.3.4) of groups to which either Ther orem 2.2.1 or Theorem 2.2.4 can be applied. In particular, the products xr, y pwu q u r r and (2.4.4) u r r xr, y RP pwu , , q

r can be dened, and are independent of the choice of pwu q. (The automorphism wu of M can also include a permutation of factors in (2.3.4), but the r extension of Theorem 2.2.1 to this slightly more general situation is clear. We have taken the liberty of expressing the assertion (2.2.6) of this theorem in the same form as its counterpart (2.2.17) in Theorem 2.2.4, with the extension x , y attached to the Whittaker extension of the general linear components of being trivial.) We therefore obtain a canonical linear form ` r (2.4.5) fG p, uq xr, y tr RP pwu , , qIP p, f q u r
r P

in f P HpGq. The convention here is similar to that of Theorem 2.2.4, in r that the summand on the right is zero if happens not to be wu -stable.

2.4. STATEMENT OF THE LOCAL INTERTWINING RELATION

99

We will later have to consider the role of u in the vertical split exact 0 sequence at the center of (2.4.3). We will then write xu and wu for the 0 pG, M q. images of u in the respective groups S pG, M q and W Suppose now that s is any semisimple element in the complex reductive group S . We write (2.4.6)
1 fG p, sq f 1 p 1 q

r for the linear form in f P HpGq attached to the preimage pG1 , 1 q of p, sq under the local form of the correspondence (1.4.11). Any such s projects to an element in the quotient S S pGq, which may or may not lie in the subgroup SpG, M q. Theorem 2.4.1 (Local intertwining relation for G). For any u in the local normalizer N pG, M q, the identity (2.4.7)
1 fG p, s sq fG p, uq,

r f P HpGq,

holds for any semisimple element s P S that projects onto the image xu of u in S pG, M q. r r The most important case is when lies in the subset 2 pM q of pM q. Then the group T AM x is a maximal torus in S , and S pG, M q equals the full group S . In this case, we write N N pG, M q, W W pG, M q,
0 0 W W pG, M q,

and R R pG, M q. The last group R can be regarded as the R-group of the parameter , since it is closely related to the reducibility of the induced representations IP pq, r P P PpM q, P pM q.

r We shall often treat strictly as an element in pGq in this case, and we shall write 1 S S pM q, r since M is uniquely determined up to the action of the group AutN pGq. Lemma 2.4.2. Suppose that for any M , Theorem 2.4.1 is valid whenever r r belongs to 2 pM q. Then it holds for any in the general set pM q.

100

2. LOCAL TRANSFER

The proof of the lemma will be straightforward, but it does depend on some further observations. We shall postpone it until the end of this section. In the meantime, we shall discuss an important implication of the local intertwining relation postulated by the theorem. We rst observe that the groups in (2.4.3) can be described in concrete r terms. For any P pGq, we write
1 1

r r ,

following the local form of the notation (1.4.1). As in the global case of 1.4, ` the subscripts k P K range over disjoint indexing sets I I pGq I pGq, r J and J _ . We recall that if i P I , i P sim pGi q is a simple parameter

r for an endoscopic datum Gi P Esim pNi q, while if j P J , j belongs to ` r the complement of sim pNi q in sim pNi q. The subset I pGq consists of p p those i such that G and Gi are of the same type (both orthogonal or both p p symplectic), while I pGq consists of those i such that G and Gi are of

opposite type. We then have ` (2.4.8) S Op i , Cq

` iPI pGq iPI pGq

GLp j , Cq Spp i , Cq
jPJ

as in (1.4.8), where pq` pq` denotes the kernel of the character ` ` gi P Op i , Cq, i P I pGq. : gi pdet gi qNi ,
i i

The identity component of this group obviously equals 0 S SOp i , Cq GLp j , Cq . Spp i , Cq
` iPI pGq iPI pGq

jPJ

Finally S is the quotient of S by a central subgroup, of order 2 if N is 0 even and order 1 if N is odd, which may or may not lie in S . Consider for simplicity the basic case that is the image of a parameter r 2 pM q. We can then identify the complex torus T A x with the in M 0 product of standard maximal tori in the simple factors of S . It is clear that 0 pS q is the group of connected components in the left hand factor p q` of S , and that its image in 0 pS q is the group S . The subgroup 1 S is given by the contribution of the subfactors Op i , Cq with i odd, while the quotient R is given by the subfactors Op i , Cq with i even. The group 0 N is an extension of S by the product W of standard Weyl groups of the 0. simple factors of S It is worth looking at one example in greater detail. Consider the lower horizontal short exact sequence (2.4.9)
1 1 S S R 1

2.4. STATEMENT OF THE LOCAL INTERTWINING RELATION

101

p from (2.4.3), in case N is even and G is orthogonal. We write

` I ` I pGq

for simplicity, and identify the group of connected components in the product Op i , Cq
iPI `

from (2.4.8) with the group of functions : I ` Z{2Z. The point is to account for the indices i P I ` in terms of the parity of both associated integers i and Ni . For example, the group 0 pS q corresponds to the subgroup ) ! 1 piq 1 , 1 P :
i

where the product is over the set of indices i with Ni odd. The group S 0 pS q corresponds to the quotient 1 { xy, where is the element in 1 (of order 1 or 2) such that piq 1 whenever ` ` i belongs to the set Io of indices with i odd. (The set Io X I `,o of i with 1 .) The group S 1 corresponds both i and Ni odd is even, so does lie in to the subgroup p q1 of P such that piq 1 whenever i belongs to ` the subset Ie of indices with i even. The R-group R then corresponds to 1 the associated quotient of , which is canonically isomorphic to the group R1 of functions ` : Ie Z{2Z, with the supplementary requirement that 1 piq 1,
i 1 1 1

I `,o

in the exceptional case that X I `,o is empty. The short exact sequence (2.4.9) can thus be identied with the more concrete exact sequence (2.4.10) 1 p q1 R1 1.
1 1

` Io

p The other two cases (with N odd or G symplectic) are similar, but simpler. In all three cases, the short exact sequence (2.4.9) (and its analogue p (2.4.10)) splits. However, if N is even and G is orthogonal there is generally no canonical splitting. The sign character (2.4.11) pq 2 piq, P R1 ,
i ` where the product is over the intersection Ie X I `,o , is a distinguishing feature of this case. It represents the obstruction to a canonical splitting of (2.4.10). Observe also that as an element in R , can act by outer r automorphism on the even orthogonal factor G P Eell pN q, as well as the

102

2. LOCAL TRANSFER

general linear factors. It is not hard to infer from our discussion that the action on G is nontrivial if and only if pq 1. The character (2.4.11) thus governs the ambiguity of both factors in the product (2.4.4). In fact, it is fair to say that this character is what ultimately forces us to deal with twisted endoscopy for even orthogonal groups, as stated in Theorem 2.2.4, if only to resolve the ambiguity in the denition (2.4.2). We now continue with our discussion of the local intertwining relation. The last lemma tells us that its general proof can be reduced to the case that r P 2 pM q. In dealing with this basic case, we shall make a slight change of notation. Following an earlier suggestion, we will generally take to be r simply a parameter in the set pGq. We can then choose a pair pM, M q, r r where M lies in the set 2 pM, q of parameters in 2 pM q whose image equals . As we have noted, the Levi subgroup M is uniquely determined r as an OutN pGq-orbit of G-conjugacy classes by . We are assuming that M is proper in G, which is to say that does not r belong to 2 pGq. We can therefore assume as usual that we have constructed r r the packet M of (OutN pM q-orbits of) representations M of M pF q. We shall combine this with the local intertwining relation to construct the packet r r of pOutN pGq-orbits) of representations of GpF q. The construction provides an important reduction of the local classication. It also gives us r an independent way to view the packet , in the more concrete terms of induced representations and intertwining operators. Its proof amounts to a reordering of some of the discussion of this section. We shall try to present it so as to be suggestive of more general situations. Proposition 2.4.3. Assume that for any proper Levi subgroup M of G, r Theorem 2.4.1 holds for parameters in 2 pM q. Then if is any parameter r 2 pGq in pGq, the packet and pairing xx, y of r r in the complement of Theorem 2.2.1 exist, and satisfy (2.2.6). r Proof. The packet M provides a representation (2.4.12) M pM b M q
pM ,M q 1 of the group S M pF q. We let M range here over a set of representatives r r of the OutN pM q-orbits that comprise the packet M and M range over the associated (1-dimensional) representations

M pxM q xxM , M y
1 of S . We are of course free to identify M with a representation of 1 r the group algebra CpS q b HpM q. Its restriction M to the subalgebra 1 r CpS q b HpM q is then independent of the choice of representatives. The overlapping notation is deliberate.

2.4. STATEMENT OF THE LOCAL INTERTWINING RELATION

103

The group N acts on M pF q by r r1 u : m wu mwu , m P M pF q, u P N ,

where the image wu of u in W is regarded as an element in W pM q. The associated semidirect product M pF q N

1 is then an extension of the group W by S M pF q. We claim that there is a canonical extension of the representation of M to M pF q N . To see this, we have rst to expand slightly on the discussion of the short exact sequence (2.4.10) above. The action of N on M pF q obviously stabilizes the orthogonal or symplectic factor G of M . Let N, be the p subgroup of elements that act on G by inner automorphism. If G is symp plectic or N is odd, for example, N, equals N . In the other case that G is orthogonal and N is even, N, is the kernel of the pullback to N of the sign character (2.4.11) on R . In all cases, there is a canonical splitting 1 N, S W, ,

where W, is the image of N, in W . It follows from the denitions that any constituent M of (2.4.12) has a canonical extension to M pF q W, . The representation M therefore has a canonical extension to the subgroup M pF q N, (of index 1 or 2) of M pF q N . The claim is that M can be extended canonically to the larger group M pF q N . Suppose that M is W -stable. Applying Theorems 2.2.4 and 2.2.1 separately to the components of M relative to the decomposition (2.3.4), as in the denition (2.4.4), we see that M b M has a canonical extension to M pF q N . In general, N (and W ) act by permutation on the set of M attached to a given M . If o is a nontrivial orbit (necessarily of order 2), we observe that as a representation of M pF q N, , the sum pM b M q
M Po

is the restriction of the representation of M pF q N induced from any of the summands. We conclude that the full sum (2.4.12) does have a canonical r extension to M pF q N , as claimed. We shall denote it by M , despite the further ambiguity in the notation. Let ` (2.4.13) IP pM q M b IP pM q
pM ,M q 1 be the representation of S GpF q induced parabolically from the represen1 tation M of S M pF q. Our canonical extension of M to M pF q N can then be combined with the intertwining operators RP pw, M , M q from the last section. A modest generalization of (2.4.2), in which the irreducible representation of M pF q is replaced by the (reducible) restriction of M

104

2. LOCAL TRANSFER

to M pF q, and the extension of to Mwu pF q is replaced by our canonical r r of to M pF q N , allows us to attach an intertwining extension M M operator (2.4.14) r r RP pu, M , M q M puq RP pwu , M , M q

to any u P N . This operator then commutes with the restriction of to GpF q. The product (2.4.15) r pu, gq RP pu, M , M q pgq, u P N , g P GpF q,

therefore gives a canonical extension of the representation to the product N GpF q. We consider the character ` r r tr pu, f q , u P N , f P HpGq, r of the restriction of to the product of N with the symmetric Hecke r algebra HpGq. Suppose that o is an N -orbit of representations M . If r the orbit contains one element, the restriction of the operator M puq in (2.4.14) to the subspace of pM b M q in (2.4.12) equals r r M puq b M pwu q, r where M puq xr, M y. It follows from (2.4.2) that the restriction of the inu r tertwining operator (2.4.14) to the corresponding subspace in (2.4.13) equals r M puqRP pwu , M , M q. r The contribution of this subspace to the character is therefore ` xr, M y tr RP pwu , M , M qIP pM , f q . u r r On the other hand, if o contains two elements, the operator M puq interchanges the subspaces of the two representations pM b M q. The restriction of (2.4.14) therefore interchanges the corresponding two subspaces of (2.4.13), so the contribution of these subspaces to the character vanishes. It follows from the denition (2.4.5) (with our notation now dictating a sum r over M P M in place of the sum over P in (2.4.5)) that ` r r (2.4.16) tr pu, f q fG p, uq, u P N , f P HpGq. We will be concerned with the restriction of to the subgroup S GpF q of N GpF q, relative to the splitting of the left hand vertical exact sequence in (2.4.3). The local intertwining relation suggests that 0 the representation is trivial on the complementary subgroup W . We will indeed establish this fact by global means later, so there will be no loss of information in the restriction. As a representation of S GpF q, has a decomposition p b q,
p,q

2.4. STATEMENT OF THE LOCAL INTERTWINING RELATION

105

for irreducible representations of S and of GpF q. The sum is of course r nite, but the constituents can have multiplicities. We dene the packet to be the disjoint union of the set of , each identied with it associated r r OutN pGq-orbit, that occur. Any P then comes with a corresponding character pxq xx, y, x P S , of the abelian group S . It follows that ` r r xx, yfG pq, x P S , f P HpGq. (2.4.17) tr px, f q
r P

r We have constructed the packet and pairing xx, y. To show that they satisfy (2.2.6), we apply our assumption that Theorem 2.4.1 is valid. Combined with (2.4.16) and (2.4.17), it tells us that
1 fG p, s sq fG p, uq ` ` r r tr pu, f q tr px, f q xx, y fG pq, r P

for elements s and u that project to the same point x in S . A similar iden1 tity holds for fG p, sq, provided that we replace x by the point s1 x s x. Since 1 fG p, sq f 1 p 1 q, we obtain r f 1 p 1 q xs x, y fG pq, f P HpGq.
r P

This is the required identity (2.2.6). Remarks. 1. It is only at the end of the proof that we appeal to our hypothesis that Theorem 2.4.1 holds for G. In particular, the actual construction (2.4.17) of the packet and pairing of Theorem 2.2.1 depends only on the application of Theorems 2.2.1 and 2.2.4 to the smaller group G . 2. The structure of the proof appears to be quite general. We shall add several comments on what might be expected if G is a general connected group. (i) There are certainly examples of quasisplit groups G for which the 1 nite groups S , S and R are nonabelian. This will make the associated representations M and more interesting. (In general, would continue to denote both a packet and a representation.) It seems likely that there will again be a canonical extension of the representation M of 1 S M pF q to the semidirect product M pF q N . However, there will also be new phenomena.

106

2. LOCAL TRANSFER

(ii) The stabilizer W pM q of a representation M P M could be proper in W . However, we would expect, at least if is generic (and in contrast to the hypothetical possibility treated in the proof of the proposition) that W pM q is also equal to the stabilizer of the corresponding irre1 ducible representation M of S attached to M . It is conceivable that the representation M will not have an extension to M pF q W pM q. The obstruction would presumably reduce to a complex valued 2-cocycle on the image R pM q of W pM q in R , of the kind considered in [A10, 3]. However, we could expect the same obstruction (or rather its inverse) to account for the failure of M to have an extension to the preimage N pM q of W pM q in N . In fact, there ought to be a canonical extension of the representation 1 M b M of S M pF q to the semidirect product M pF q N pM q. One could then induce this extended representation to the full group M pF qN . The direct sum of the representations so obtained, taken over the W -orbits tM u in the packet M , would be the canonical extension of the represen1 tation M from S M pF q to M pF q N . With this object in hand, one would then be able to construct the representation of S GpF q from r induced representations IP pM q and intertwining operators RP pw, M , q, as in the proof of the proposition. (iii) If G is not quasisplit, the situation is a little more complicated. In 1 this case, one would need to replace the groups S and S by extensions 1 S,sc and S,sc , dened as in [A19, 3]. With this proviso, the arguments for quasisplit groups would presumably carry over. We shall discuss inner twists of orthogonal and symplectic groups in Chapter 9.

r We return to our quasisplit classical group G P Esim pN q. We have still to establish the reduction asserted in Lemma 2.4.2. However, we shall rst state a variant of the local intertwining relation, which applies to a twisted r orthogonal group G. r r r r Recall that G pG0 , q, where G G0 is an even orthogonal group rsim pN q. The statement will be almost the same as that of Theorem in E 2.4.1. In particular, M is a proper Levi subgroup of G, and P P PpM q is a parabolic subgroup of G. (We do not assume that P is normalized by any r r element in G.) If lies in the packet , IP pq denotes the corresponding r r representation induced from P pF q to the group G` pF q generated by GpF q. The normalized intertwining operator (2.4.2) is then the earlier operator attached to G and M , but with one dierence. The element w must be r r taken here from the rational Weyl set W pG, M q induced from G. This implies that the four groups in the lower right hand block in (2.4.3) have r also to be formulated as sets, with G in place of G. For any u in the set r N pG, M q, we construct the objects wu , wu , Mu and xu as above, and we r r rr p, uq on HpGq as in (2.4.5). We also dene the r dene the linear form fG

2.4. STATEMENT OF THE LOCAL INTERTWINING RELATION

107

r r r r linear form f 1r p, sq on HpGq as in (2.4.6), for any semisimple element s in G r r the centralizer S S pGq. r r Theorem 2.4.4 (Local intertwining relation for G). For any u P N pG, M q, the identity r r r r r f P HpGq, (2.4.18) f 1 p, s sq f r p, uq,
r G G

r r r holds for any semisimple element s P S that projects to the image xu of u r M q. in S pG, This theorem joins our growing body of unproved assertions that will be established later by interlocking induction arguments. In the meantime, we assume as necessary that it holds if G is replaced by an even orthogonal r group G P Esim pN q, for any N N . r Proof of Lemma 2.4.2. We are given a local parameter P pM q, r where M is a Levi subgroup of the simple group G P Esim pN q. We are also given a point u P N pG, M q, and a point s P S whose image in the quotient S equals the image xu of u in the subgroup S pG, M q of S . We have to show that the linear form ` u r r fG p, uq xr, y tr RP pwu , , qIP p, f q
r P 1 equals its endoscopic counterpart fG p, s sq. The parameter is the image of a square integrable parameter r 1 P 2 pM1 q, for a Levi subgroup M1 of M . The point u is a coset in 0 the complex group S that normalizes the torus AM in S . It has a reprex 0 sentative that also normalizes the maximal torus T in S . In other words, we can choose a representative u1 of u in the group N1 , which one sees is 0 determined up to translation by the subgroup W1 pM, M1 q of N1 in the chain 0 0 W1 pM, M1 q W1 pG, M1 q N1 pG, M1 q N1 . The point s can of course be identied with an element s1 P S1 pGq, since the groups S1 pGq and S pGq are equal. Moreover, s1 projects to the image xu1 of u1 in the group S1 pGq S1 pG, M1 q, since xu1 equals xu . The information we are given from the lemma therefore tells us that the linear form ` u r r (2.4.19) fG p1 , u1 q xr1 , 1 y tr RP1 pwu1 , 1 , 1 qIP1 p1 , f q r 1 P1 1 equals its endoscopic counterpart fG p1 , s1 s1 q. Since s1 s1 equals s s, the 1 p , s s q and f 1 p, s sq are equal. We have therefore to linear forms fG 1 1 1 G verify that fG p1 , u1 q equals fG p, uq. The problem is simply to interpret the terms in the expression (2.4.19) for fG p1 , u1 q. Assuming an appropriate choice of M1 , we can arrange that

108

2. LOCAL TRANSFER

the parabolic subgroup P1 P PpM1 q is contained in P . The Weyl element in (2.4.19) then satises M wu1 wu1 wu ,
M where wu1 P W M pM1 q is a Weyl element for M , and the point wu P W pM q is identied with the Weyl element in W pM1 q that stabilizes the parabolic subgroup R1 P1 XM of M . The induced representation in (2.4.19) satises ` M IP1 p1 q IP IR1 p1 q ,

by induction in stages, while the intertwining operator has a corresponding decomposition

M M r r r RP1 pwu1 , 1 , 1 q RR1 pwu1 , 1 , 1 qRP pwu , , q.

We assume implicitly that the analogues of Theorems 2.2.1 and 2.2.4, and of Theorems 2.4.1 and 2.4.4, hold for the proper Levi subgroup M in place of G. They tell us that for any function h P HpMu q, the sum ` M M M r xr1 , 1 y tr RR1 pwu1 , 1 , 1 qIR1 p1 , hq u r
r 1 P1

equals the linear form h1 p1 , s1 s1 q M

r P

u r xr, y hM prq.

Finally, from general principles, we know that the trace in (2.4.19) bres as a product over the Hilbert space ` M HP1 p1 q HP HR1 p1 q . It follows that the expression for fG p1 , u1 q reduces to the earlier expression for fG p, uq, as required. We have now established preliminary reductions of our local theorems. Lemma 2.4.2 reduces the local intertwining relation of Theorem 2.4.1 to r the case of parameters in the set 2 pM q. Proposition 2.4.3 interprets the local, intertwining relation itself as an explicit construction of some of the r packets of Theorem 2.2.1, and hence reduces this theorem to parameters r in the set 2 pGq. Similar reductions, which we shall leave to the reader, apply to the assertions of Theorems 2.2.4 and 2.4.4. For the basic cases that remain, we will need global methods. We shall establish them, along with the supplementary local assertions of Theorem 1.5.1(b), in Chapters 6 and 7. 2.5. Relations with Whittaker models The theory of Whittaker models has been an important part of representation theory for many years. For example, it is at the heart of the theorem of multiplicity 1 for GLpN q [Shal], and its generalization Theorem 1.3.2 by Jacquet and Shalika. In the hands of Shahidi, Whittaker models have

2.5. RELATIONS WITH WHITTAKER MODELS

109

yielded a broader understanding of intertwining operators. We shall review a couple of his main results, for comparison with our earlier discussion. Suppose for the moment that G is an arbitrary quasisplit, connected reductive group over the local eld F . Suppose also that pB, T, tXk uq is a F -stable splitting of G over F , where Xk Xk , 1 k n, are xed root vectors for the simple roots of pB, T q. If F is a xed nontrivial additive character on F , the function puq F pu1 ` ` un q, u P NB pF q, is a nondegenerate character on the unipotent radical NB pF q of BpF q. As usual, tuk u are the coordinates of logpuq relative to the simple root vectors tXk u, or more correctly, any basis of root vectors of nB pF q that includes tXk u. Then represents a Whittaker datum pB, q for GpF q. We are interested in irreducible tempered representations P temp pGq of GpF q that have a pB, q-Whittaker model. As in the special case of G GLpN q from 2.2, a pB, q-Whittaker functional for is a nonzero linear form on the underlying space of smooth vector V8 for such that ` puqv puq pvq, u P NB pF q, v P V8 . We recall that the vector space spanned by the pB, q-Whittaker functionals for has dimension at most 1 [Shal], and that is said to be generic if the space is actually nonzero. For any such , the function ` W px, vq pxqv , x P GpF q, v P V, satises ` W x, pyqv pxy, vq, y P GpF q, and therefore represents an intertwining operator from to the representation INB pq of GpF q induced from the character of NB pF q. The Whittaker model of is the corresponding space of functions ( W p, q W pxq W px, vq : v P V8 of x, equipped with the representation of GpF q by right translation. Conversely, suppose that W p, q is a subrepresentation of INB pq that is equivalent to . If v W px, vq, v P V, x P GpF q, is a nontrivial intertwining operator from to W p, q, the linear form pvq W p1, vq, v P V8 , is a pB, q-Whittaker functional for . Whittaker models and Whittaker functionals are thus essentially the same. Our focus will be on the functionals. Suppose that M is standard Levi subgroup of G over F . Then M is the Levi component of a parabolic subgroup P M NP of G that contains B. The Whittaker datum pB, q for G restricts to a Whittaker datum pBM , M q

110

2. LOCAL TRANSFER

for M pF q, which can be expressed as above in terms of the splitting of M attached to that of G. Assume now that and are attached to M instead of G. That is, is an irreducible representation of M pF q, with a pBM , M qWhittaker functional . There is then a canonical Whittaker functional , pq for the induced representation IP pq. It is dened in terms of the Whittaker integral ` 1 h, pw n xq pn q1 dn , (2.5.1) W, px, h, q
N pF q

whose ingredients we recall. The vector h lies in the space HP,8 pq of smooth functions in HP pq. We are following the convention that HP pq is a Hilbert space of functions from K to V , which does not change if is replaced with its twist by a point P a . We have therefore to write M,C ` ` h, pxq MP pxq h KP pxq ep`P qpHP pxqq , x P GpF q, in the notation of [A7, p. 26], to obtain a vector in the usual space on which IP p q acts. The group N NP is the unipotent radical of the standard parabolic subgroup P M N that is adjoint to P , in the sense that
1 M w M w ,

w w wM ,

where w and wM are the longest elements in the restricted Weyl groups of G and M respectively. The Whittaker integral (2.5.1) converges absolutely for Repq in a certain chamber, and has analytic continuation as an entire function of . (See [CS, Proposition 2.1] and [Sha1, Proposition 3.1].) Its value W, px, , q is therefore a well dened intertwining operator from IP pq to INB pq. We are interested in the corresponding pB, q-Whittaker functional , pq : h , ph, q W, p1, h, q, h P HP,8 pq, for IP pq. The constructions from the earlier sections of this chapter pertain to the r special case of G P Esim pN q. To compare them with the results of Shahidi, we take P temp pM q to be tempered. We assume that Theorem 1.5.1 holds for the Levi subgroup M , which we take to be proper. Then belongs to r r the packet of a parameter P bdd pM q. We can denote the normalizing factor from the end of 2.3 by rP pw, q rP pw, q, w P W pM q, since is uniquely determined by . For the same reason, we write RP pw, q RP pw, , q for the normalized intertwining operator in (2.3.26). Suppose in addition that is generic, and is equipped with a pBM , M q-Whittaker functional . There is then a canonical choice for the intertwining operator pwq from w r r to . To see this, recall that the representative w of w in GpF q is dened by

2.5. RELATIONS WITH WHITTAKER MODELS

111

the splitting of G. It follows from [Sp, Proposition 11.2.11] that w preserves r the splitting of M , and therefore stabilizes the Whittaker functional . We can therefore choose pwq uniquely so that r (2.5.2) r pwq.

The denition (2.4.2) thus gives us a canonical self-intertwining operator r r RP pw, q RP pw, , q of IP pq, in the case P temp pM q is both generic and equivalent to w. In this section, we have taken G to be an arbitrary quasisplit group. With the assumption that is generic, Shahidi was nonetheless able to construct the local L and -functions that appear in the normalizing factors rP pw, q. This was a local ingredient of the paper [Sha4] that became known as the Langlands-Shahidi method, in which Langlandss original results [L4] were extended, and the analytic continuation and functional equation for a broad class of automorphic L-functions were established. In particular, the local denitions of 2.3 and 2.4 all carry over to the general group G, if P temp pM q is a representation that is generic. We therefore have the normalizing factor rP pw, q, and the normalized intertwining operator r RP pw, q rP pw, q1 JP pw, q, w P W pM q,

r in (2.3.26), to go with the basic unnormalized intertwining operator JP pw, q. By the general results of Harish-Chandra [Ha4] and the properties of rP pw, q established by Shahidi, these functions all have analytic continuation as meromorphic functions of P a . Moreover, if w belongs to the stabilizer M,C W pq of in W pM q, we obtain the canonical self-intertwining operator r r RP pw, q pwq RP pw, q of IP pq from (2.5.2). Theorem 2.5.1 (Shahidi). Suppose that G is a quasisplit group over F with Levi subgroup M , and that P temp pM q is generic. (a) The normalized intertwining operators Rpw, q RP pw, q, are unitary, and satisfy the relation. (2.5.3) Rpw1 w, q Rpw1 , wq Rpw, q, w1 , w P W pM q. w P W pM q,

(b) Suppose that w belongs to the subgroup W pq of W pM q. Then the canonical self intertwining operator RP pw, q satises the relation r (2.5.4) , pq , pq RP pw, q, r

if is a pBM , M q-Whittaker functional for .

112

2. LOCAL TRANSFER

The property (a) is [Sha4, Theorem 7.9]. Its proof is contained in several papers, which include [Sha1] and [Sha2] as well as [Sha4]. Since the argument is by induction on the length of w, the assertion has to be proved in slightly greater generality. Namely, w must be taken from the more general set W pM, M 1 q of Weyl elements that conjugate M to a second standard Levi subgroup M 1 , while w1 is taken from a second set W pM 1 , M 2 q. The operators in (2.5.3) are otherwise dened exactly as above, and the required identity takes the same form. Shahidis starting point is the meromorphic scalar valued function C, pw, q, dened by the relation ` r , p q C, pw, q , pw q Jpw, q given by the unnormalized intertwining operator r r Jpw, q JP pw, q. The existence of this function follows from the multiplicity 1 of Whittaker r functionals, and the fact that Jpw, q is meromorphic in . Since r Jpw1 w, q rpw1 , w, q Jpw1 , w q Jpw, q, r for a meromorphic scalar valued function rpw1 , w, q, one sees that rP pw1 , w, q C, pw1 w, q1 C, pw1 , w q C, pw, q. To split the 2-cycle rpw1 , w, q in terms of suitable normalizing factors, it then suces to compute the local coecients C, pw, q. Therein lies the problem. Shahidis construction of these objects eventually leads to the construction of L and -functions, which he uses to express C, pw, q and to dene normalizing factors rP pw, q such that the corresponding normalized operators satisfy (2.5.3). The general version of (2.5.4) must be written dierently. It takes the form (2.5.5) , pq , pwq Rpw, q, for a general element w P W pM, M 1 q. In the special case that M 1 M and w belongs to the subgroup W pq of W pM q W pM, M q, we have the intertwining operator pwq from w to that stabilizes . It follows from r the denitions that the right hand side of (2.5.5) equals r r r , pwq pwq1 RP pw, q , pq RP pw, q. Therefore (2.5.5) does reduce to (2.5.4) in this case. The proof of (2.5.5) is only implicit in [Sha4]. The right hand side of (2.5.5) equals the value at 0 of the operator ` r , pw q rP pw, q1 Jpw, q rP pw, q1 C, pw, q1 , p q. The problem is to show that the product rP pw, q C, pw, q

2.5. RELATIONS WITH WHITTAKER MODELS

113

is analytic at 0, with value at 0 equal to 1. We shall have to leave the reader to extract this fact from the formulas of [Sha4] and [Sha2], following the calculation from the special case in [KeS, 4]. We shall review the calculation in the paper [A27] in preparation. We will have a special interest in the following ve cases: (i) F arbitrary, G GLpN q, M arbitrary; r (ii) F C, G P Esim pN q, M arbitrary; r (iii) F arbitrary, G P Esim pN q, M GLprN {2sq; r (iv) F arbitrary, G P Esim pN q, M T minimal; r (v) F arbitrary, G P Esim pN q, Mder Spp2q. In cases (i)(iv), it is known that any P temp pM q is generic. In the last case (v), G is of the form Spp2nq, while M is an product of several copies of GLp1q with the group SLp2q Spp2q. From the general relation between representations of SLp2q and GLp2q described in the introduction of [LL], ` we see that any P temp SLp2q is generic, for some choice of Whittaker datum. The same is therefore true for any P temp pM q. This case will not be needed in any of our future arguments, unlike the other four. We have included it here for general perspective, and because it will arise in a natural setting later in Lemma 6.4.1. It follows that in all ve cases, the r canonical self intertwining operator RP pw, q is dened for any w P W pq, and satises (2.5.4). Corollary 2.5.2. Suppose that pF, G, M q is as in one of the cases (i) (v), that P temp pM q, and that w P W pq. Then if p, Vq is the unique r irreducible pB, q-generic subrepresentation of IP pq, the operator RP pw, q satises r RP pw, q , P V8 . In particular, if IP pq is irreducible, we have r RP pw, q 1. Proof. The restriction of RP pw, q to the irreducible subspace V of r r HP pq is a nonzero scalar. Since is generic, RP pw, q satises (2.5.4). Since the Whittaker functional , pq on HP,8 pq is supported on the subspace V, the scalar in question equals 1. We note that in the cases (i) and (ii), the induced representation IP pq is always irreducible. In (iii), suppose that corresponds to the Langlands parameter P 2 pM q, and that w P W pq is nontrivial. The centralizer S is then isomorphic to either Spp2, Cq or Op2, Cq. We will apply the corollary later to the case that the centralizer is Spp2, Cq, in the proof of Lemma r 5.4.6, after rst showing that the operator RP pw, q is a scalar. In (iv), our interest will be in the case that is trivial on the maximal compact subgroup of T pF q. The representation IP pq need not be irreducible in this case. However, the generic subrepresentation is easily identied. For it

114

2. LOCAL TRANSFER

follows from (2.5.1) (together with an approximation argument in case F is archimedean) that is the irreducible constituent of IP pq that contains the trivial representation of K. It is only in the last case (v) that the L-packets for M can be nontrivial. In the same paper, Shahidi conjectured that any tempered L-packet for the quasisplit group G has a pB, q-generic constituent [Sha4, Conjecture 9.4], motivated by the proof of the property for archimedean F that followed from the results of [Kos] and [V1]. Shelstad [S5] has recently established a strong form of this conjecture for real groups. Namely, if the transfer factors are normalized as in [KS, (5.3)], the pB, q-generic representation in a tempered packet is the unique representation for which the linear character r (2.2.5) is trivial. For nonarchimedean F and G P Esim pN q, Konno [Kon] showed that the conjecture would follow from local twisted endoscopy for G, which is to say, the assertion of Theorem 2.2.1. In 8.3, we will show that the strong form of the conjecture (apart from the uniqueness condition) is also valid in this case, after proving Theorem 2.2.1 in Chapter 6. In retrospect, we could have assumed Shahidis conjecture inductively for M , and then appealed to his conditional proof of Proposition 2.3.1 in [Sha4, 9]. I have retained the proof of the proposition in 2.3 for its dierent perspective, and its role in the transition to the self intertwining operators (2.4.4). We need to expand on the case (i) of GLpN q. It will be important to r establish an analogue for the component GpN q of the identity of Corollary 2.5.2. Of equal importance, but harder, is the task of proving an identity r for general parameters P pN q. Suppose that M is a standard Levi subgroup of GLpN q, and that : LF SU p2q
L

is a general parameter in pM q. Then corresponds to an irreducible unitary representation of M pF q, which is the Langlands quotient of a standard representation . We write W pM q for the Weyl set of outer r automorphisms of M induced from the component GpN q, and W pM q for pM q. Any element in W pM q then stabilizes and the stabilizer of in W as well as . We assume that the set W pM q is nonempty. This implies that the image of in pN q, which we again denote by , lies in the subset r pN q. Suppose that w belongs to W pM q. Then r w pN q w0 , r where pN q is the standard outer automorphism of GLpN q dened in 1.3, r and w0 belongs to a set W pM, M 1 q. Bearing in mind that pN q acts as an involution on the set of standard parabolic subgroups of GLpN q, we note that M 1 is the standard Levi subgroup that is paired with M . The representative r r r w pN q w0

2.5. RELATIONS WITH WHITTAKER MODELS

115

r of w in GpN, F q preserves the standard Whittaker datum pBM , M q for M . We would like to dene twisted intertwining operators by some variant of (2.4.2) and (2.3.25). The operator RP pw, q pw, q Rw1 P |P p q in (2.3.25) has to be slightly modied. For we cannot dene the operator pw, q Hw1 P p q HP pw q r in the product r pw, q P pw, q pw, q 1 , since the operation would not take values by a simple left translation by w r in HP pw q. We instead write ` ` rpw, q pxq w1 x pN q , r r r P Hw1 P p q, (2.5.6) in order that the right hand side make sense, as well as rpw, q P pw, q rpw, q. r The product r RP pw, q rpw, q Rw1 P |P p q is then an operator r RP pw, q : HP p q HP pw q. If is a pBM , M q-Whittaker functional for , we dene the intertwining operator pwq from w to uniquely by analytic continuation, and r the property r pwq . r This lifts to an intertwining operator pwq from w to . The product r r RP pw, q pwq RP pw, q r r is then a canonical operator on HP p q. We note that its construction is a variant of both the denition at the beginning of 2.2, and the denition of the operator in Theorem 2.5.1(b). r r The twisted operator RP pw, q does not intertwine IP p q with itself. Indeed, it follows from (2.5.6) and the other denitions that r r r RP pw, q : IP p q IP p q pN q. r r In other words, RP p, q intertwines IP p q with the representation r r pN q1 IP p q IP p q pN q. Recall that we have already introduced an operator with the same intertwining property. In 2.2, we dened an operator r r IP p , N q pN q, IP p q, with r r IP p , N q : IP p q IP p q pN q,

116

2. LOCAL TRANSFER

in terms of a Whittaker functional for IP p q. How are the two objects related? Theorem 2.5.3. The intertwining operator attached to P pM q and w P W pM q satises (2.5.7) r r RP pw, q IP p , N q. r

In case is generic, one can apply the formula (2.5.5) to the element w0 P W pM, M 1 q. The required identity (2.5.7) then follows without much diculty from the various denitions. The general case is more dicult, and as far as I know, has not been investigated. It requires further techniques, based on some version of minimal K-types. Rather than introduce them here, we shall leave the general proof of Theorem 2.5.3 for a separate paper [A26]. There is another way to interpret the identity (2.5.7). We begin by p r r writing S pN q as in 1.4 for the centralizer in GpN q of the image of . r r0 Then S pN q is a bitorsor under the centralizer S pN q of the image of p r r in G0 pN q. The analogue of the diagram (2.4.3) for the G0 G0 pN q torsor r G GpN q certainly makes sense. If we dene its objects in the natural way in terms of the bitorsor r S S pN q, we see that N pG, M q W pG, M q W pM q, and that this last set is a bitorsor under the group
0 W pG, M q W pM q.

All the other sets in the diagram are trivial. We shall use the diagram r to formulate the analogue for GpN q of the local intertwining relation of Theorem 2.4.1. The formalities of the process will be helpful in the next section in understanding the spectral terms of the twisted trace formula r for GLpN q. The actual intertwining relation for GpN q will be an essential ingredient of the global comparison in Chapter 4. The twisted intertwining relation can be most easily stated if we inate r IP p q to an induced representation of G` pN, F q. With this interpretation, IP p q acts on the larger Hilbert space r` r HP p q HP p q HP p q, r r where HP p q is the space of functions supported on the component GpN, F q. For any w P W pM q, we then obtain a linear transformation r r RP pw, q : HP p q HP p q by setting ` ` ` r r r r RP pw, q pxq RP pw, q xpN q1 ,

2.5. RELATIONS WITH WHITTAKER MODELS

117

r for P HP p q and x P GpN, F q. In other words, RP pw, q is dened r exactly as in (2.4.2) and (2.3.25), namely, with the translation operation r r pw, q dened by excluding the right translation by pN q from (2.5.6). If r r f belongs to the Hecke space HpN q, its integration against IP p q gives a linear transformation r r IP p , f q : HP p q HP p q r that we can compose with RP pw, q. Following (2.4.5), we set ` r r r r r (2.5.8) fN p, uq tr RP pw, q IP p , f q , f P HpN q, for any point u w in the set N pG, M q W pM q. r 0 pN, F q on HP p q also has an extension The representation IP p q of G r to the bitorsor GpN, F q. This is provided by the operator ` r r r IP p , N q IP , pN q , which then yields the linear form ` r r r fN pq tr IP p , f q , r r f P HpN q,

of (2.2.1). Observe that if ` ` ` r r IP p , N q pxq IP p , N q xpN q1 , r for P HP p q and x P GpN, F q, then ` ` r r r fN pq tr IP p , f q IP p , N q tr IP p , N q IP p , f q . It follows from Theorem 2.5.3 that r r (2.5.9) fN pq fN p, uq, u P W pM q.

r Suppose that s is a semisimple element in the torsor S pN q. The pair p, sq then has a preimage under the general correspondence (1.4.11), which r r consists of a datum G P EpN q, and a parameter in pGq that we continue to denote by . If the assertion (a) of Theorem 2.2.1 is valid, we have the linear form r r r r f G pq fN pq, f P HpN q, r on SpGq. Following (2.4.6), we then set (2.5.10) rG r fN p, sq f G pq, r r f P HpN q.

r Corollary 2.5.4 (Local intertwining relation for GpN q). Assume that the assertion (a) of Theorem 2.2.1 is valid for any pair pG, q, (2.5.11) r r G P EpN q, P pGq. r r f P HpN q,

Then if P pM q is as in Theorem 2.5.3, the identity rG r fN p, s sq fN p, uq,

r holds for any u P W pM q and any semisimple element s P S pGq.

118

2. LOCAL TRANSFER

rG r Proof. It follows from (2.5.10) and (2.5.9) that fN p, sq equals fN p, uq, for any s and u. The two linear forms are therefore independent of the points s and u. This is something we could have expected from general principles, r given that the set S pGq is connected. Its proof, together with that of the equality of the two linear forms, is thus an immediate consequence of Theorem 2.5.3. The required identity (2.5.11) obviously follows if we replace s by s s.

r We return briey to the case that G P Esim pN q is a quasisplit orthogonal or symplectic group. A central theme of Chapter 2 has been that the general self-intertwining operator RP pw, , q in (2.4.2) has no canonical denition. r x r Only when the operator is balanced with the pairing xru , y in (2.4.4) do we obtain a canonical object in general. We have seen in this section that the denition (2.4.2) can be made canonical in the special case that the parameter and the representation are generic. Before we move on, we remind ourselves of another case that leads to a canonical denition. To r r describe it, we take to be a general parameter in 2 pM q, with P pM q r being a general element in the corresponding packet . We are assuming that w lies in the stabilizer W pq of and . As an element in W pM q, w stabilizes the orthogonal or symplectic part G of M . Suppose that it acts on G by inner automorphism. This is always so p unless N is even and G is orthogonal, in which case it is equivalent to the condition that the value corresponding to w of the explicit sign character r (2.4.11) be 1. The representative w of w then commutes with G . We r can therefore dene the intertwining operator pwq from w to by asking that it stabilize the relevant Whittaker functional on each of the general linear factors of M . The product (2.4.2) is then a canonical self intertwining operator RP pw, , q of IP pq. It owes its existence to the fact that the r r extension x , y of the character x , y from S pM q to S pM q in (2.4.5) is canonical. 0 Suppose for example that w lies in the subgroup W of elements in 0 W induced by points in the connected group S pGq. The corresponding value of (2.4.11) is then 1, for trivial reasons, and we obtain a canonical r self intertwining operator RP pw, , q. The local intertwining relation for G suggests that RP pw, , q equals 1. We shall establish this fact later, in the r course of proving the intertwining relation. There is one other application of the theory of Whittaker models that we will need. It is a kind of converse of the theorem of Konno, for nonarchimedean F , which we will need at one point in the proof of the local intertwining relation. Since the results in [Kon] rely on the exponential map, we do have to assume that the residual characteristic charpF q of F is not equal to 2.

2.5. RELATIONS WITH WHITTAKER MODELS

119

Lemma 2.5.5. Suppose that F is nonarchimedean, and that the objects r r r G P Esim pN q, P pGq, M P 2 pM, q, M G and pBM , M q over F have the following three properties. (i) p charpF q 2. r (ii) The element M P M with x , M y 1 is the unique pBM , M qr generic representation in the packet M . (iii) For any pair ps, uq as in the statement of Theorem 2.4.1, there is a constant eps, uq such that r fG p, sq eps, uq f 1 p, uq, f P HpGq.
G

Then eps, uq 1 for every ps, uq. In other words, the local intertwining relation is valid for and M . Results of this nature are probably known. The proof in any case is not dicult, given Theorem 2.5.1(b) and the papers [Kon], [MW1] and [Rod2]. Since we need to move on, we shall leave the details for [A27]. We shall use Lemma 2.5.5 only in very special cases, which will come up in the proofs of Lemma 6.4.1 and Lemma 6.6.2.

CHAPTER 3

Global Stabilization
3.1. The discrete part of the trace formula We can now begin to set in place the means for proving the three main theorems. The methods are based on the trace formula. In recent years, it has been possible to compare many terms in the trace formula directly with their stable analogues for endoscopic groups. The resulting cancellation has led to a corresponding identity for the remaining terms. These are the essential spectral terms, the ones which carry the desired information about automorphic representations. They comprise what is called the discrete part of the trace formula. We shall review the discrete part of the trace formula in this section. The theorems we are trying to prove concern quasisplit orthogonal and symplectic groups. In the long term, we are of course interested in more general groups so it would be sensible to think as broadly as possible in discussing the general theory. We shall do so whenever practical, with the understanding that the results will ultimately be specialized to orthogonal and symplectic groups, and the corresponding global parameters dened in 1.4. We assume until further notice that the eld F is global. Suppose for the moment that G is a connected, reductive algebraic group over F . The discrete part of the trace formula for G is a linear form on the Hecke algebra HpGq. We recall that HpGq is the space of smooth, compactly supported, complex-valued functions on GpAq that are K-nite relative to the left and right actions of a suitable maximal compact subgroup K GpAq. It will be convenient to have this linear form depend on two other quantities, which we will use to account for two minor technical complications. The rst point concerns the nominal failure for G to have any discrete spectrum. This is governed by the split component AG of the center of G, or equivalently, the quotient of GpAq by GpAq1 . The subgroup GpAq1 of GpAq equals the kernel of the homomorphism HG : GpAq aG from GpAq onto the real vector space ` aG HomZ XpGqF , R , which is dened by the familiar prescription exHG pxq,y |pxq|, x P GpAq, P XpGqF .
121

122

3. GLOBAL STABILIZATION

It has a group theoretic complement A` AGQ pRq0 , G,8 GQ RF {Q pGq, in GpAq, where RF {Q pq denotes the restriction of scalars, and p q0 stands for the connected component of 1. In particular, the mapping HG restricts to a group theoretic isomorphism from A` onto aG . It gives rise to an G,8 isometric isomorphism ` ` L2 GpF qA` zGpAq L2 GpF qzGpAq1 . G,8 We are therefore free ` work withthe discrete spectrum of the left hand to space in place of L2 GpF qzGpAq1 . disc To allow exibility for induction arguments, it is useful to treat a slightly more ` general situation. We choose a closed subgroup XG of the full cen ` ter Z GpAq ` GpAq such that the product XG Z GpF q is closed and coof compact in Z GpAq . (In general, we write ZpHq for the center of any group H.) The quotient GpF qXG zGpAq GpF qzGpAq{XG then has nite invariant volume, as in the special case that XG A` . We G,8 also x a character on the quotient ` ` ` Z GpF q X XG zXG Z GpF q zZ GpF q XG , and write ` ` L2 GpF qzGpAq, L2 GpF qzGpAq, disc for the space of -equivariant functions on GpF qzGpAq that are squareintegrable modulo XG , and decompose discretely under the action of GpAq. The central character datum pXG , q for G gives rise to corresponding data for Levi subgroups M of G. Given M , we write A`,G for the kernel in M,8 A` of the composition M,8
M A` M,8 aM aG .

The product XM pA`,G qXG M,8 is then an extension of XG , to which we can pull back the character . We obtain a triplet pM, XM , q that satises the same conditions as pG, XG , q. ` In particular, we obtain a representation of M pAq on L2 M pF qzM pAq, disc that decomposes discretely. If P NP M belongs to the set PpM q of parabolic subgroups of G with Levi component M , we write
G IP pq IP pq

for the corresponding parabolically induced representation. It acts on the Hilbert space HP pq of left NP pAq-invariant functions on GpAq such that the function pmkq on M pAq K belongs to the space ` L2 M pF qzM pAq, b L2 pKq. disc

3.1. THE DISCRETE PART OF THE TRACE FORMULA

123

This discussion can of course be applied to the special case above, namely where XG A` , 1, and XM A` . G,8 M,8 The second point has to do with analytic estimates. W. Mller has u established [Mu] that the restriction of any operator IP p, f q f pxqIP p, xqdx, f P HpGq,
GpAq

to the discrete spectrum is of trace class. Rather than work with the estimates of Mller, however, we shall rely on the consequences in [A5] of what u was later called the multiplier convergence estimate [A14, 3]. These allow us to work with small subrepresentations of IP pxq, dened by restricted Archimedean innitesimal characters. We x a minimal Levi subgroup M0 of G, which we can assume is in good position relative to K. We can then construct the real vector space h ihK h0 in terms of the group M0,Q RF {Q pM0 q, as in [A2, 3]. Thus, h0 is the Lie algebra of a maximal real split torus in the real group M0,Q pRq M0 pF8 q, and hK is a Cartan subalgebra of the Lie algebra of the compact real group K X M0,Q pRq K X M0 pF8 q. The complexication hC of h is a Cartan subalgebra of the Lie algebra of the complex group GQ pCq, whose real form h is invariant under the complex Weyl group W . The role of the real space h hG is simply to control innitesimal characters. We represent the Archimedean innitesimal character of an irreducible representation of GpAq by a complex valued linear form ,R ` i,I on h. The imaginary part ,I , regarded as a W -orbit in h , is the object of interest. It has a well dened norm },I }, relative to a xed, W -invariant Hermitian metric on h . There is then a decomposition C IP pq IP,t pq
t0

of IP pq, where IP,t pq is the subrepresentation of IP pq composed of those irreducible constituents with },I } t. The endoscopic comparison of trace formulas reduces ultimately to a reciprocity law among the representations IP,t pq attached to dierent groups. The discrete part of the trace formula will be a linear form that depends on and t. We may as well build the dependence on into the test function f . Let HpG, q be the equivariant Hecke algebra of functions f on GpAq that satisfy f pxzq f pxqpzq1 , z P XG , and are compactly supported modulo XG . The operator IP,t p, f q f pxqIP,t p, xqdx
GpAq{XG

124

3. GLOBAL STABILIZATION

on HP, is then dened for f in HpG, q. In this setting, the discrete part of the trace formula is the linear form
G Idisc,t pf q Idisc,t pf q,

t 0, f P HpG, q,

dened by (3.1.1) M G |W0 ||W0 |1

M PL

wPW pM qreg

` | detpw 1qaG |1 tr MP,t pw, qIP,t p, f q .

M

The outer sum in (3.1.1) is over the nite set L LpM0 q of Levi subgroups of G that contain M0 . The inner sum is over the set of regular elements ( W pM qreg w P W pM q : detpw 1qaG 0
M

in the relative Weyl group W pM q W G pM q NormpAM , Gq{M for G and M . Here aG is the canonical complement of aG in aM , so M W pM qreg is the set of w whose xed subspace in aM equals the minimal M space aG . Following standard practice, we have written W0 W M pM0 q G W G pM q for the Weyl groups with respect to M . and W0 0 0 The remaining ingredient of (3.1.1) comes from the standard global intertwining operator MP pw, q that is at the heart of Langlands theory of Eisenstein series. We can dene MP pw, q as the value at 0 of a meromorphic composition (3.1.2) MP pw, q pwq MP 1 |P p q, P 1 w1 P, P paG q , M C of operator valued functions on HP pq. The factor pwq stands for the mapping from HP 1 pq to HP pq dened by left translation by any representative r w1 of w1 in GpQq, while P 1 w1 P is the parabolic subgroup w1 P w in r r PpM q. The other factor is the operator MP 1 |P p q : HP pq HP 1 pq, whose value ` MP 1 |P p q pxq, P HP pq, x P GpAq, is dened for the real part of in a certain cone in paG q by the familiar M intertwining integral pnxqep`P qpHP pnxqq dn ep`P 1 qpHP 1 pxqq .
NP 1 pAqXNP pAqzNP 1 pAq

P, P 1 P PpM q,

We recall that HP is the mapping from GpAq to aM dened by HP pnmkq HM pmq, n P NP pAq, m P M pAq, k P K, that P is the usual linear form on aM dened by half the sum of the roots (with multiplicity) of pP, AM q, and that puq puq epHM puqq , u P XM ,

3.1. THE DISCRETE PART OF THE TRACE FORMULA

125

is the twist of by . The operator MP pw, q leaves the subspace HP,t pq of HP pq invariant, and intertwines the two induced representations IP,t p q and IP,t pw q on this subspace. It has analytic continuation as a meromorphic function of P paG q , whose values at ipaG q are analytic and unitary. Thus, M C M MP pw, q restricts to a unitary operator MP,t pw, q on the Hilbert space HP,t pq, which intertwines the representation IP,t pq. This operator is the last of the terms in (3.1.1). It is also the most interesting. Its analysis includes the local intertwining relation stated in the last chapter, and will be an important aspect of future chapters. More generally, suppose that (3.1.3) G pG0 , , q

is an arbitrary triplet over F . Then G0 is a connected reductive group over F , is a semisimple automorphism of G over F , and in the global context here, is a character on G0 pAq that is trivial on the subgroup G0 pF q. As in the local case of 2.1, we also write G G0 more narrowly for the associated G0 -bitorsor over F , with distinguished point 1 . The Hecke module HpG, q is then a space of functions on GpAq dened exactly as in the case G G0 above. We note that there are two morphisms from G G to G0 . They are dened in the obvious way as formal algebraic operations on the bitorsor by # 1 y1 y2 1 px1 x2 q, 1 (3.1.4) 1 y1 y2 x1 x1 , 2 for two points yi xi in G. Our general convention of letting G stand for both a G0 -torsor and the underlying triplet should not lead to confusion. It is similar to our convention for endoscopy, in which a symbol G1 represents both an endoscopic group and the underlying endoscopic datum. The discrete part of the trace formula for G again takes the form (3.1.1), provided that the terms are properly interpreted. We recall how these terms can be understood in the general twisted case. For a start, the central character datum pXG , q has to be adapted to G (rather than G0 q. We write ( aG a 0 H P aG0 : AdpqH H aG0 G for the subspace of -xed vectors in the real vector space aG0 , and AG pA 0 q0 AG0 G

126

3. GLOBAL STABILIZATION

for the maximal F -split torus in the centralizer of G in G0 . We assume implicitly that is trivial on the subgroup A` AGQ pRq0 , G,8 GQ RF {Q pGq,

of G0 pAq, since the twisted trace formula for G otherwise becomes trivial. If A is any ring that contains F , we write ` ` ` Z GpAq Z G0 pAq Z G0 pAq for the centralizer of GpAq in G0 pAq. We then take ` XG Z GpAq to be any closed subgroup satisfying the conditions above, with the further requirement that it lie in the kernel of , and to be any character on ` the quotient of XG by its intersection with Z GpF q X XG . The standard example is of course the pair pXG , q pA` , 1q. G,8 For any such choice, we can form the equivariant Hecke module HpG, q of functions in GpAq, relative to pXG , q and a suitably xed maximal compact subgroup K G0 pAq. The general version of (3.1.1) will be a linear form in a function f in HpG, q. The general version of (3.1.1) is again based on a xed minimal Levi subgroup M0 G0 . Suppose that M belongs to the set L LpM0 q of Levi subgroups of G0 that contain M0 . The composition aM aG0 aG0 , , together with the natural isomorphism between aG and the space of covariants ( aG0 , aG0 { X AdpqX : X P aG0 in aG0 , gives a canonical linear projection from aM onto aG . We write aG M for its null space in aM . We then have the regular set ( G Wreg pM q Wreg pM q w P W pM q : detpw 1qaG 0 ,
M

in the Weyl set W pM q W G pM q NormpAM , Gq{M of outer automorphisms of M induced by the conjugation action of G on G0 . The set Wreg pM q is generally quite distinct from its untwisted analogue 0 G0 Wreg pM q Wreg pM q, dened for G0 above. These conventions account for the general indices of summation M and w, and the corresponding coecients | detpw 1qaG |1 ,
M

in the twisted interpretation of (3.1.1).

3.1. THE DISCRETE PART OF THE TRACE FORMULA

127

We observe that pXG , q again provides a central character datum pXM , q for any M P L. The rst component is the extension XM pA`,G qXG M,8 of XG , and the second component is the character on XM obtained by pulling back . The only dierence from the case G G0 above is that A`,G is M,8 now the kernel of the projection of A` onto the subspace aG attached to M,8 G (rather than G0 ). The remaining term is the trace in (3.1.1). Suppose that P belongs to the set PpM q of parabolic subgroups of G0 with Levi component M . We 0 write HP,t pq for the Hilbert space of the induced representation dened for G0 above. We reserve the symbol HP,t pq for its analogue for G, namely the space of complex-valued functions on GpAq such that for any y P GpAq, the function pxyq, x P G0 pAq, 0 belongs to HP,t pq. The operators in (3.1.1) attached to f P HpG, q and w P W pM q make sense in this more general context. However, they have now to be interpreted as linear transformations (3.1.5) and (3.1.6)
0 MP,t pw, q : HP,t pq HP,t pq. 0 IP,t p, f q : HP,t pq HP,t pq

The rst transformation equals

G IP,t p, f q IP,t p, f q GpAq{XG G f pyqIP,t p, yqdy,

for the induced mapping ` G IP p, yq pxq pxyqp1 xyq,

P HP,t pq, x P G0 pAq,

0 from HP,t pq to HP,t pq. (There is no need to include in the notation, as it is implicit in superscript G.) The second transformation is given by (3.1.2), 0 with the translation pwq dened now as a map from HP 1 pq to HP pq by (3.1.4). The product MP,t pw, qIP,t p, f q is then an operator on HP,t pq, for which the trace in (3.1.1) is dened. With this notation, all of the terms in the expression (3.1.1) makes sense. The discrete part of the trace formula for G pG0 , , q is the linear form G Idisc,t pf q Idisc,t pf q,

t 0, f P HpG, q,

dened by the general form of this expression. It is the main spectral component of a general formula usually known as the twisted trace formula. The twisted trace formula was established in [CLL] and [A5]. Our notation here diers from that of [A5] in three minor respects, which might be worth pointing out explicitly.

128

3. GLOBAL STABILIZATION

In [A5], G represents a component of a (nonconnected) reductive algebraic group, and is trivial. This is slightly less general that the triplet pG0 , , q here. However, the earlier arguments all extend without diculty. The given datum of [A5] is also less precise, in that it species the underlying automorphism of G0 only up to inner automorphism. The point of view here is more suitable to stablization, which we will discuss in the next section. 0 Secondly, the induced Hilbert spaces HP,t and HP,t do not take this form in [A5]. They reect the fact here that the G0 -torsor G need not be attached to a component of a reductive algebraic group. They also streamline the notation in all cases. Finally, we reiterate that the elements M P L here are Levi subgroups of G0 . The complementary geometric terms in the twisted trace formula require another notion, that of a Levi subset M of G, which we can recall from the local context of 2.2. To keep the two kinds of objects straight, we 0 wrote L0 LG pM0 q in [A5] for the set of Levi subgroups of G0 containing M0 and L LG pM0 q for the collection of Levi subsets of G, noting at the same time that there was an injective mapping M M 0 from L to L0 . In the present volume, the discrete part Idisc,t is the only term we will consider. Since it is based only on the Levi subgroups of G0 , we are free here to denote these objects by M , and to write L for the set of M that contain M0 . It is with this notation that the expression from Idisc,t pf q takes exactly the same form (3.1.1) in general as it does in the original case of G G0 . The starting point for the trace formula is the kernel of an integral operator. In the general case at hand, the operator is a composition IG p, f q MG p, q, f P HpG, q,

of what amount to specializations to M G0 of the transformations (3.1.5) ` and (3.1.6), but with the -equivariant Hilbert space L2 G0 pF qzG0 pAq, in place of ` 0 0 0 HG0 ,t pq L2 disc,t G pF qzG pAq, . The reader can check that the kernel of this operator equals Kpx, yq f px1 yqpyq, x, y P G0 pF qzG0 pAq.
PGpF q

As a general rule, we have not tried to be explicit about measures. In the global setting here, we note simply that the operator IP,t p, f q in the formula (3.1.1) for Idisc,t pf q depends on a choice of G0 pAq-invariant measure on GpAq{XG . This amounts to a choice of Haar measure on G0 pAq{XG . Observe that a Haar measure on the group XG determines a Haar measure on A` such that G,8 ` ` ` ` ` ` vol XG Z GpF q zZ GpAq vol A` Z GpF q zZ GpAq , G,8

3.1. THE DISCRETE PART OF THE TRACE FORMULA

129

` for any Haar measure on Z GpAq . The linear form Idisc,t pf q thus depends on a choice of G0 pAq-invariant measure on GpAq{A` , or equivalently, a G,8 Haar measure on G0 pAq{A` . G,8 Finally, we observe that the linear form Idisc,t pf q is admissible in f . In other words, it can be written as a linear combination of twisted characters ` tr pf q , 0 P 0 , where 0 is a nite set of irreducible unitary representations of G0 pAq, and is an extension of 0 to GpAq, in the sense of the global analogue of (2.1.1). The set 0 0 pf q depends on f , but only through its K-type. t To be more precise, let us agree that a Hecke type for G means a pair p8 , 8 q, where 8 is an open compact subgroup of G0 pA8 q and 8 is a nite set of irreducible representations of a maximal compact subgroup K8 of GpA8 q. We have written ( A8 a P A : av 0, if v R S8 Fv F8
vPS8

and ( A8 a P A : av 0, if v P S8 here for the respective subrings of archimedean and nite adeles in A. We also write ( AS a P A : av 0, if v P S and ( A8 a P A8 : av 0, if v R S , S if S is any set of valuations of F that contains the set S8 of archimedean valuations. We can then decompose 8 as a product 8 K S , where S S8 is S nite, 8 is an open compact subgroup of G0 pA8 q, and K S is a product over S S v P S of hyperspecial maximal compact subgroups Kv G0 pFv q. Assume that the product K8 8 is contained in the maximal compact subgroup K (and is in particular a subgroup of nite index in K). We shall say that p8 , 8 q is a Hecke type for f if f is bi-invariant under translation by 8 , and transforms under left and right translation by K8 according to representations in the set 8 . Any f P HpG, q of course has a Hecke type. The assertion above is that the nite set 0 depends only on a choice of Hecke type for f . This follows from the denition (3.1.1) of Idisc,t pf q and the theory of Eisenstein series, specically Langlands decomposition ` of L2 G0 pF qzG0 pAq in terms of residues of cuspidal Eisenstein series [L5, Chapter 7]. (See [A5, Lemma 4.1]. Once 8 has been xed, the archimedean innitesimal character of is controlled by the K8 -type 8 and the norm t },I } of its imaginary part.) Notice that for the extension of 0 , ,I 0 ,I is a linear form on the quotient h h0 h0 {tH P h0 : H AdpqHu of the real vector space h0 hG0 attached to G0 .

130

3. GLOBAL STABILIZATION

3.2. Stabilization In this section we shall formally state the condition on which our theoG rems depend. We have just described a linear form Idisc,t pf q on HpG, q. The condition is that it can be stabilized. There has been considerable progress on this problem in recent years, and a general solution in case G G0 is now in place. One can hope that its extension to twisted groups might soon be within reach. For our part here, we require a solution only for the two r r twisted groups G GpN q and G G, initially introduced in 1.2. We continue to suppose for the time being that G represents a general triplet pG0 , , q over the global eld F . We need to discuss what it means to G stabilize the linear form Idisc,t pf q. Let us rst review a few of the background notions from the beginning of [KS], which underlie both the local discussion of 2.1 and the global discussion we are about to undertake here. General endoscopic transfer has to be formulated in terms of several supplementary data. These include a connected quasisplit group G over F , equipped with an inner class of inner twistings : G0 G of the connected group G0 . Given G , one xes an F -automorphism of G that preserves some F -splitting, and is of the form Intpg q 1 , for some element g P G . (As usual, G denotes the simply connected sc sc cover of derived group of G .) The automorphism of G0 also induces p p an automorphism of the dual group G0 of G0 that preserves some F splitting, and is determined up to conjugation. One xes a 1-cocycle a from p p the Weil group WF of F to the center ZpG0 q of G0 that is the Langlands dual of the automorphic character on G0 pAq. This then gives rise to an L-automorphism
L

p : g w pgqa pwq1 w,
L G0

gw P

G0 ,

of the L-group

p of G0 . We use it to dene the dual G0 -bitorsor p p p G G G 0 L ,

on which the L-group L G0 acts by conjugation. These matters are discussed (with slightly dierent notation) in the early stages (2.1) of [KS]. So is the general notion of endoscopic datum G1 for G. As in the local case mentioned briey in 2.1, a general endoscopic datum G1 represents a 4-tuple pG1 , G 1 , s1 , 1 q. We recall that the rst component, denoted also by G1 , is a connected quasisplit group over F . The other p components are a split extension G 1 of WF by G1 , a semisimple element 1 in G, and an L-embedding 1 of G 1 into L G0 . These four components p s are required to satisfy the conditions (2.1.1)(2.1.4) of [KS], the most basic p being the assertion (2.1.4b) that 1 pG1 q equals the connected centralizer of p s1 in G0 . Similar denitions apply to the local completions Fv of F . In particular, an endoscopic datum G1 for G over F localizes to an endoscopic

3.2. STABILIZATION

131

datum G1 for Gv over Fv . The general notion of isomorphism between v endoscopic data is dened on p. 18 of [KS]. As in our earlier special cases, we write OutG pG1 q AutG pG1 q{IntG pG1 q for the group of outer automorphisms of G over F obtained from automorphisms of G as an endoscopic datum. We also write EpGq for the set of isomorphism classes of endoscopic data G1 that are locally relevant to G, by which we mean that for every v, G1 pFv q G1 pFv q contains elements that v are norms from GpFv q ([KS, p. 29]). Endoscopic data that are not locally relevant play no role in transfer, and can be ignored. For a given endoscopic datum G1 , the group G 1 need not be L-isomorphic to L G1 . Even if it is, there often is not a canonical isomorphism. To remedy r r this defect, one attaches an auxiliary datum pG1 , 1 q to G1 , as in the local r discussion of 2.1. We recall that G1 represents a central extension of G1 r 1 over F that is induced, in the sense that it is a product of by a torus C r r tori of the form RE{F pGm q, and 1 is an L-embedding of G 1 into L G1 . For r example, one can always take G1 to be a z-extension of G1 (See [KS, (2.2)].) x1 r The L-embedding 1 gives rise to a cohomology class in H 1 pF, r which Cq, 1 on C 1 pAq{C 1 pF q, by the global r r in turn yields an automorphic character r r Langlands correspondence for tori. (With a suitable choice of 1 , we can 1 is unitary.) assume that r r r The global transfer factor for G and G1 (and the auxiliary datum pG1 , 1 q) 1 , q of two ad`lic variables 1 and . These lie is a canonical function p e in the ad`lized varieties of strongly G-regular, stable conjugacy classes in e r G1 and G respectively. It again serves as the kernel of a transfer mapping, which takes functions f P HpGq to functions f 1 p 1 q p 1 , qfG pq

The global transfer factor can be written as a (noncanonical) product 1 (3.2.1) p 1 , q v pv , v q

v

of

1.

of local transfer factors. The transfer mapping therefore takes a decomposable function f fv
v

in HpGq to the decomposable function 1 1 f 1 p 1 q fv pv q

v

of The LSK transfer conjecture, now a theorem, tells us that for any 1 r r1 v, fv belongs to the space SpG1 , v q. The fundamental lemma tells us that v 1 r r1 for almost all v, fv is the image in SpG1 , v q of the characteristic function v

1.

132

3. GLOBAL STABILIZATION

r of a hyperspecial maximal compact subgroup of G1 pFv q. It follows that f 1 belongs to the restricted tensor product r r r r SpG1 , 1 q SpG1 , 1 q
v v v

of local stable Hecke algebras. The general situation is actually slightly more complicated than we have ` indicated. It comes with a cohomology class z 1 P H 1 F, ZpG q with sc values in the group of -covariants ( ZpG q ZpG q{ z pzq1 : z P ZpG q sc sc sc in the center of G [KS, Lemma 3.1.A]. For much of [KS], the authors sc assume that this class (or rather its local analogue) is trivial. They explain how to take care of the more general situation in (5.4), at the expense of r r replacing SpG1 , 1 q by a twisted stable Hecke algebra relative to an inner r r automorphism 1 of G1 over F . However, it seems likely that by choosing r r r pG1 , 1 q suitably, for example so that G1 is a z-extension of G1 with connected 1 be trivial. It ought to be easy to check this r center, one could arrange that point. I have not done so, since the most general case will not be our main r r focus here. Instead, let us simply assume as a condition on G that pG1 , 1 q r can be chosen for each G1 so that 1 is trivial, and therefore that any transfer r r mapping does indeed take values in the untwisted Hecke algebra SpG1 , 1 q. r r There is also another point. The auxiliary datum pG1 , 1 q gives rise to a surjective, ane linear mapping r1 h h0 h

r between the real vector spaces attached to G1 and G. This is dual to a mapping ,R ` i,I 1 ,R ` i1 ,I of archimedean innitesimal characters of corresponding irreducible tem1 1 r pered representations 8 and 8 . (The representation 8 of G1 pF8 q is 1 on C 1 pF q, while r 8 required to have central character 8 r 8 is understood to be an extension of a representation to GpF8 q of the form (2.1.1). They can be any corresponding representations in Shelstads endoscopic transfer of archimedean L-packets.) We can always modify the L-homomorphism r r 1 by tensoring it with an automorphic character on G1 pAq. For simplicity, we assume that this has been done so that the ane mapping from r1 to h h descends to a linear isomorphism from r1 to h. We can then assume that h },I } }1 ,I }1 , for a suitably Hermitian norm } }1 on h . (The alternative would be to C take } } to be an ane norm, in the sense that it is a translate in h of a C Hermitian norm, and } }1 to be an ane norm on r1 that is related to } } hC r r in a way that depends on pG1 , 1 q.)

3.2. STABILIZATION

133

In the last section, we worked with a central character datum pXG , q for G. To relate it to global transfer, we recall that for any G1 , there is a canonical injection of the group of covariants ( ZpG0 q ZpG0 q{ zpzq1 : z P ZpG0 q of ZpG0 q into ZpG1 q. (See [KS, p. 53].) We assume here that is trivial on the kernel of the restriction of this mapping to XG . Then pulls back to ` a character on the image of XG in Z G1 pAq , and hence also to a character (which we denote again by ) on its preimage r X1 XG 1 ` r r r in Z G1 pAq . Recall also that the automorphic character 1 on C 1 pAq at` ` r r r 1 , 1 q is actually dened on the preimage of Z G0 pAq in Z G1 pAq tached to pG r [KS, p. 53 and p. 112]. It therefore restricts to a character on X1 . The product r 1 r1 r 1 . The pair pX1 , 1 q will serve as our r r then represents a third character on X r central character datum for G1 . It follows from the discussion on p. 112 of [KS] that the global transfer of functions f f 1 can be dened if f belongs r r to HpG, q, and descends to a mapping from HpG, q to SpG1 , 1 q. By r 1 , 1 q is the space of strongly regular, stable orbital integrals of denition, SpG r r r functions in the equivariant global Hecke algebra HpG1 , 1 q. We can therefore 1 on HpG1 , 1 q with a linear form S 1 on r r p identify any stable linear form S r 1 , 1 q by the global analogue of the local prescription (2.1.2). SpG r For any G1 , we write ` p p p p p p p (3.2.2) ZpG1 q ZpG1 q { ZpG1 q X ZpGq ZpG1 q ZpGq{ZpGq, where p p ZpGq ZpG0 q
p

p p is the subgroup of -xed points in the center of G0 . The general endoscopic 1 is called elliptic if ZpG1 q is nite. We write E pGq as usual for the p datum G ell subset of elliptic isomorphism classes in EpGq. The elliptic classes represent the global endoscopic data that are used to stabilize the trace formula. In their stabilization of the regular elliptic terms in the twisted trace formula, Kottwitz and Shelstad introduced global coecients pG, G1 q, G1 P Eell pGq.

These objects are dened by an explicit formula [KS, p. 115], which generalizes the formula of Kottwitz [K1, Theorem 8.3.1] for the original coecients introduced by Langlands [L10] in case G G0 . They are also essentially the same as the provisional coecients dened in [A9, (3.3)]. These were introduced to describe a conjectural stabilization of the discrete part of the trace formula [A9, Hypothesis 3.1].

134

3. GLOBAL STABILIZATION

As an aside, we note that the discrete part of the trace formula is the spectral analogue of what was called the elliptic part of the trace formula, in the case G G0 studied in [A15]. The elliptic part of the trace formula is a sum of invariant orbital integrals, geometric terms that include semisimple orbital integrals. The semisimple elliptic part was stabilized for G G0 in [K2]. It is the regular (semisimple) elliptic part that was stabilized for any G in [KS], generalizing the rst results of Langlands [L10] for G G0 . We note that all of these partial stabilizations were originally dependent on the LSK-transfer conjecture and the fundamental lemma. They are now unconditional. What is meant by a stabilization of some part of the trace formula? For the discrete part (as in general), it is a decomposition p1 f P HpG, q, pG, G1 qSdisc,t pf 1 q, (3.2.3) Idisc,t pf q
G1 PEell pGq

of the given linear form in terms of stable linear forms

G 1 r r Sdisc,t Sdisc,t : HpG1 , 1 q C
1

attached to elliptic endoscopic data. The correspondence f f 1 f G

1

r r is the global transfer mapping from HpG, q to SpG1 , 1 q we have just de1 scribed. The stable linear form Sdisc,t is universal, in that it depends only r on G1 (and not G). The coecient pG, G1 q depends on G as well as G1 , but not on which part of the trace formula is being stabilized. The set of summation Eell pGq is generally innite in the global case here. However, the sum can be taken over a nite subset of Eell pGq, which depends only on a choice of nite set S S8 of valuations outside of which f is unramied. We note that if G G0 is quasisplit, the stable linear form G Sdisc,t Sdisc,t on GpAq is dened inductively by (3.2.3). The assertion in this case is simply that the dierence p1 Idisc,t pf q pG, G1 qSdisc,t pf 1 q, f P HpG, q,
G1 G

is stable. For any other G, there are no further denitions that can be made. The assertion then becomes an identity, which ties the original linear form Idisc,t pf q to the seemingly unrelated forms p1 f Sdisc,t pf 1 q. In general, the summand of G1 in (3.2.3) is supposed to be independent of the 1 r r r choice of pG1 , 1 q, while the linear form Sdisc,t on G1 pAq is to be independent r r of 1 . But the transfer mapping f f 1 does depend on 1 . This apparent contradiction can be resolved directly by induction, and an appeal to [A5, Lemma 4.3].

3.2. STABILIZATION

135

If G G0 is a connected group, the decomposition (3.2.3) was established in [A16], subject to a general condition on the fundamental lemma. As we have noted, the standard fundamental lemma has now been established. The condition in [A16] was actually a generalization of the fundamental lemma, which was conjectured [A12, Conjecture 5.1] for weighted orbital integrals. It was used in the stabilization of the general geometric terms, its role being roughly similar to that of the standard fundamental lemma in the analysis of the elliptic geometric terms. The weighted fundamental lemma has now been established by Chaudouard and Laumon [CL1], [CL2], using the global methods introduced by Ngo [N], and a reduction by Waldspurger [W7] to elds of positive characteristic. This removes the condition that had qualied the results for G G0 in [A16]. In particular, the decomposition (3.2.3) is now valid if G is any connected group. If G is a general triplet (3.1.3), the decomposition (3.2.3) remains conjectural. The problem is to extend the results of the papers [A14][A16] to the twisted case, or in other words, to stabilize the twisted trace formula. The linear form Idisc,t pf q whose stabilization (3.2.3) we seek is given by the explicit formula (3.1.1). However, there is no direct way to stabilize the terms in this formula without rst assuming the theorems we are trying to prove in this volume, (or rather their general analogues for G). One has instead to stabilize all the complementary terms in the twisted trace formula. Once this is done, the decomposition will follow from the twisted r trace formula itself, and the stable trace formula for each G1 (which was established for connected quasisplit groups in [A16]). Some of the techniques will certainly carry over to the twisted case without much change. Others will call for renement, and no doubt new ideas. Still, there is reason to be hopeful that the general stabilization can be established in the not too far distant future. In any case, it is not our intention in this volume to discuss the complementary terms in the (twisted) trace formula. We shall instead simply take the decomposition (3.2.3) as a hypothesis in the cases under consideration. We note that the setting for the papers [A14][A16] was actually slightly dierent from that of a connected reductive group G G0 . It pertains to what we called a K-group, a somewhat articial object consisting of a nite disjoint union of connected groups over F with extra structure. The point is that one can arrange for G to be one of these components. The identity (3.2.3) for G G0 follows from the results of [A16] (with minor adjustments for the slightly dierent kind of central character datum pXG , q) by extending a given function f on GpAq to be zero on the complementary adelic components of the K-group. Before we leave the general case, we recall the general formula

(3.2.4)

p pG, G1 q |0 pG q|1 kpG, G1 q|ZpG1 q |1 |OutG pG1 q|1

136

3. GLOBAL STABILIZATION

for the global coecients. Here, 0 pG q is the group of connected components in the complex group ` 0 p p G ZpGq X ZpG0 q , and ` 1 ` p p kpG, G1 q ker1 F, ZpG0 q ker1 F, ZpG1 q , where ker1 pF, q denotes the subgroup of locally trivial classes in the global p cohomology group H 1 pF, q. The group ZpG1 q given by (3.2.2) is nite, 1 is assumed to be elliptic. Notice that p q equals the group of since G 0 G p acting on the complex torus xed points of the automorphism , ` 0 ` p p 0 ZpG0 q { ZpGq . On the other hand, the dual automorphism acts as a linear isomorphism Adpq on the real vector space aG0 , for which aG is the subspace of pointwise p xed vectors. The order of the group of xed points of equals the absolute ` value of the determinant of Adpq 1 on the subspace aG0 of aG0 . (See G [SpS, II.17].) It follows that 0 pG q1 detp 1q G 1 . (3.2.5) a
G0

This factor in (3.2.4) does not occur in the corresponding formula on p. 115 of [KS]. (It was taken from the formula [A9, (3.5)], which was used in a provisional stabilization of general spectral terms.) The discrepancy can be traced to the choice of Haar measure on GpAq{GpF qA` on p. 74 of [KS]. G,8 With the conventions of [KS], the determinant of pw 1q in (3.1.1) would 0 have to be taken on the space aG rather than aG . M M We return to our special cases ` study. For the two principal cases, of r r r either G equals the triplet GpN q GpN q0 , pN q, 1 , or G G0 represents r an element in Esim pN q. It is pretty clear how to specialize the general notions above. In the rst case, we take the inner twist to be the identity r r r r morphism from GpN q0 onto GpN q GLpN q, and we set pN q pN q. p r r r We also set pN q L pN q pN q, regarded as an automorphism of eip r r ther GpN q0 GLpN, Cq or L GpN q0 GLpN, Cq F . In both cases, an 1 P E pGq (or rather its isomorphism class) takes the endoscopic datum G ell simplied form from Chapter 1. That is, G 1 is just the L-group L G1 and 1 is the canonical L-embedding of L G1 into L G0 that accompanies G1 as an endoscopic datum. Since there is no need for the supplementary datum r r r r pG1 , 1 q, we simply set G1 G1 and 1 1. There is also no need of central r character data. We can therefore take XG X1 1 in all cases. We also have the supplementary case, in which G is the triplet r r r r r G pG0 , , 1q attached to an even orthogonal group G0 P Esim pN q. Simr . There r ilar remarks apply here. In particular, we take 1 and

3.3. CONTRIBUTION OF A PARAMETER

137

r r is again no call for auxiliary datum, which leaves us free to have G1 and 1 r r r r r stand for components in an endoscopic datum pG1 , G 1 , s1 , 1 q for G. 1 q contains a quotient kpG, G1 q of The general formula (3.2.4) for pG, G orders of groups of locally trivial cohomology classes. In the three cases of p p interest, the action of F on either ZpG0 q or ZpG1 q factors through an abelian quotient of . It follows that these cohomology groups are trivial and that kpG, G1 q 1. The general formula therefore reduces to 1 1 p 1 (3.2.6) pG, G1 q 0 pG q ZpG1 q OutG pG1 q . r For example, in the twisted general linear group GpN q, we can write 1 1 ` 1 r p r rpN, Gq GpN q, G ZpGq OutN pGq ,
2

for any G P Eell pN q, since G Z{2Z in this case. The global transfer factor p 1 , q is canonical, but the local factors in (3.2.1) are generally not. However, in the quasisplit case where we are now working, the local transfer factors can be specied uniquely. In fact, as we saw in the explicit discussion of our three special cases in 2.1, there are two ways to normalize the local transfer factors. One depends on a choice of splitting over Fv , while the other depends on a choice of Whittaker datum over Fv . They dier by the -factor ` 1 , rv , Fv 2 of a virtual representation rv of v [KS, 5.3]. It is a consequence of the global denitions ([LS1, 6.26.3], [KS, 7.3]) that the product formula (3.2.1) is valid for quasisplit G if the local transfer factors are all normalized in terms of a common splitting over F . If on the other hand, they are all normalized in terms of a common global Whittaker datum, as has been the understanding for our three special cases here, the product changes by ` the global -factor 1 , r of a virtual representation r of . But it is a 2 consequence of the general construction in 5.3 of [KS] that the virtual representation r is orthogonal. It then follows from [FQ] that the global ` -factor 1 , r equals 1. The product formula (3.2.1) is therefore valid for 2 the Whittaker normalizations of the local transfer factors we have adopted. We shall state formally the hypothesis on which the results of this volume rest. Keep in mind that (3.2.3) has now been established in case G G0 , r and in particular, if G is any element in Esim pN q. The question therefore concerns our other two basic cases. r r Hypothesis 3.2.1. Suppose that G equals either GpN q or G, in the noG tation introduced initially in 1.2. Then the stabilization (3.2.3) of Idisc,t pf q holds for any t 0 and f P HpGq. 3.3. Contribution of a parameter The decomposition (3.2.3) is what will drive the proofs of our theorems. In the end, our results will come from the interplay of formulas obtained

138

3. GLOBAL STABILIZATION

r r by specializing (3.2.3) to the cases G GpN q and G P Esim pN q. We have r rst to bring the global parameters P pN q back into the discussion. In this section, we shall describe the contribution of to the terms in the specializations of (3.2.3). We may as well continue to work more broadly while this is still feasible. We do not have global parameters for general G, but we are free to replace them by Hecke eigenfamilies c for G. Some of this discussion will be rather formal. The reader might prefer to pass directly to Corollary 3.3.2, which contains its application to the parameters . For simplicity, suppose rst that G is a connected reductive group over the global eld F . In 1.3, we introduced the set Caut pGq of (equivalence classes of) families c of semisimple classes in L G. We may as well work with the subset Caut pG, q of classes in Caut pGq that are compatible with the central character datum pXG , q. Any class c P Caut pGq determines an ` unramied character v on Z GpFv q for almost all v, namely the central character of the unramied representation of GpFv q attached cv . Recall to ` that is a character on the closed subgroup XG of Z GpAq . We dene Caut pG, q somewhat ` articially`as theset of c in Caut pGq such` that extends to a character on Z GpF q zZ GpAq whose restriction to Z GpFv q equals v for almost any v. We need to see how the set Caut pG, q is related to the terms in the explicit expression (3.1.1) for the discrete part Idisc,t pf q, f P HpG, q,

of the trace formula. Consider the operator IP,t p, f q in this expression. It is isomorphic to a direct sum of induced representations of the form IP pM q, },I } t,

in which M is taken from the set of irreducible subrepresentations of ` L2 M pF qzM pAq, . For any such , the class cpq belongs to Caut pG, q. disc If c is an arbitrary class in Caut pG, q, we write IP,t,c p, f q IP, p, f q,
t: cpqcu

where IP pM q as above, and IP, pq is the subrepresentation of IP,t pq corresponding to . We also write MP,t,c pw, q for the restriction of the operator MP,t pw, q in (3.1.1) to the invariant subspace HP,t,c pq on which IP,t,c pq acts. Then ` ` tr MP,t pw, qIP,t p, f q tr MP,t,c pw, qIP,t,c p, f q .
c P Caut pG,q

It follows that (3.3.1) Idisc,t pf q

c P Caut pG,q

Idisc,t,c pf q,

3.3. CONTRIBUTION OF A PARAMETER

139

where Idisc,t,c pf q is the c-variant of Idisc,t pf q, obtained by replacing the trace in (3.1.1) by the summand of c in its decomposition above. Suppose now that G represents an arbitrary triplet pG0 , , q. To extend the decomposition to this case, we need only agree on the meaning of the associated set Caut pG, q. We dene it to be the subset of classes c P CpG0 , q that are compatible with and , in the sense that for almost all valuations v, the associated conjugacy classes cv satisfy p v pcv q cv cpv q. With this interpretation of Caut pG, q, the decomposition (3.3.1) remains valid as stated. We note that in general, the sum in (3.3.1) can be taken over a nite set that depends on f only through a choice of Hecke type. Recall that HpG, q is the 1 -equivariant Hecke module of GpAq, relative to a suitably chosen maximal compact subgroup K Kv
v

of

G0 pAq.

It is a direct limit HpG, q lim HpG, K S , q,

S

where is the space of functions in HpG, q that are biinvariant under the product KS Kv , Kv G0 pFv q,
vRS

HpG, K S , q

of hyperspecial maximal compact subgroups. The set Caut pG, q is an inverse limit S Caut pG, q lim Caut pG, q,
S

where is the relevant variant of the set dened in 1.3, composed of families of semisimple conjugacy classes
S Caut pG, q

cS tcv : v R Su. In both limits, S S8 represents a large nite set of valuations outside of which G is unramied. The two are related through the unramied Hecke algebra ` S S 8 Hun Hun pG0 q Cc K S zG0 pAS q{K S
S on G0 pAS q. On the one hand, an element cS P Caut pG, q determines a complex-valued character (3.3.2) h p S q, hpc h P HS , un

on

S Hun .

On the other, there is an action f fh ,

S f P HpG, K S , q, h P Hun ,

S of Hun on HpG, K S , q that is characterized by multipliers. Namely, if is an extension to GpAq of an irreducible unitary representation 0 of G0 pAq,

140

3. GLOBAL STABILIZATION

which is unramied outside of S and satises the global analogue of (2.1.1), then ` ` ` tr pfh q p cS pq tr pf q . h It is easy to describe the decomposition (3.3.1) in terms of eigenvalues S of Hun . Any function f P HpG, q belongs to HpG, K S , q, for some S. We x S and f , and consider the linear form Idisc,t pfh q,
S h P Hun , S on Hun . It follows from the expression (3.1.1) for Idisc,t pf q that this linear form is a nite sum of eigenforms. More precisely, we can write Idisc,t pf q Idisc,t,cS pf q, cS

where Idisc,t,cS is a linear form on HpG, K S , q such that (3.3.3) Idisc,t,cS pfh q p S qIdisc,t,cS pf q. hpc
S The sum is over a nite subset of Caut pG, q, which again depends on f only through a choice of Hecke type (necessarily of the form p8 , 8 K S q in this S case). If c belongs to Caut pG, q, we can then write (3.3.4) Idisc,t,c pf q Idisc,t,cS pf q, cS c S the sum being over the preimage of c in Caut pG, q. This is the summand of c on the right hand side of (3.3.1). Suppose that G1 P Eell pGq is an elliptic endoscopic datum, with G again r r being a general triplet (3.1.3). If pG1 , 1 q is an auxiliary datum for G1 , we r 1 q of Caut pG1 q. The properties of pG1 , 1 q are such that r r r have the subset Caut pG r r an element c1 P CpG1 , 1 q transfers to a family c of conjugacy classes for G. However, because we do not know the principle of functoriality for G and G1 , we cannot say that c lies in Caut pG, q. To sidestep this formal diculty, let us simply write S CA pG, q lim CA pG, q S

for the set of all equivalence classes of families cS of semisimple conjugacy classes in L G that are compatible with , and such that for each v, cv projects a Frobenius class in WFv at v. It follows from the various construcr r tions that there is a canonical mapping c1 c from CA pG1 , 1 q to CA pG, q. We can of course take CA pG, q to be the domain of summation in (3.1.1) if we set Idisc,t,c pf q 0 for c in the complement of Caut pG, q in CA pG, q. We would like to convert the decomposition (3.2.3) for Idisc,t pf q into a parallel decomposition for Idisc,t,c pf q. We need to be clear about our assumptions, since (3.2.3) remains conjectural in the twisted case. However, the decomposition is valid in case G G0 , and in particular if G G is quasisplit. This means that the stable linear forms on the right hand side of (3.2.3) are dened unconditionally for any G. As an assumption on

3.3. CONTRIBUTION OF A PARAMETER

141

G, (3.2.3) can therefore be regarded as a conjectural identity, both sides of which are well dened. Lemma 3.3.1. Assume that G satises (3.2.3). (a) Suppose that G equals G0 and is quasisplit. Then there is a decomposition (3.3.5) Sdisc,t pf q
c P CA pG,q

Sdisc,t,c pf q,

f P HpG, q,

G for stable linear forms Sdisc,t,c Sdisc,t,c that satisfy the analogues of (3.3.3) and (3.3.4), and vanish for all c outside a nite subset of CA pG, q that depends on f only through a choice of Hecke type. (b) Suppose that G is arbitrary. Then for any c P CA pG, q, there is a decomposition

(3.3.6)

Idisc,t,c pf q

G1 PE
ell pGq

pG, G1 q

c1 c

p1 Sdisc,t,c1 pf 1 q,

f P HpG, q,

r r where c1 is summed over the set of classes in CA pG1 , 1 q that map to c.

1 Proof. We prove that the stable linear form Sdisc,t exists for any G1 P Eell pGq. Assume inductively that if G1 G, this form satises the analogue for G1 of the assertion (a). Suppose rst that G G0 is quasisplit, as in (a). If f P HpG, K S , q S and cS P CA pG, q, we set

Sdisc,t,cS pf q Idisc,t,cS pf q

G1 G

pG, G1 q

c1,S cS

p1 Sdisc,t,c1,S pf 1 q,

S r r the inner sum being over elements c1,S in CA pG1 , 1 q that map to cS . If h S for G, we have belongs to the unramied Hecke algebra Hun

Sdisc,t,cS pfh q Idisc,t,cS pfh q

G1 G S

pG, G1 q
S

c1,S cS 1

p1 Sdisc,t,c1,S pfh q1
c1,S

p qIdisc,t,cS pf q p q hpc hpc p S qSdisc,t,cS pf q, hpc

G1

pG, G q

p1 Sdisc,t,c1,S pf 1 q

142

3. GLOBAL STABILIZATION

1 by the stable analogue for G1 of (3.3.3), and the identities pfh q1 fh1 and p 1 pc1,S q p S q. Letting cS vary over C S pG, q, we write h hpc A Sdisc,t,cS pf q cS

cS

Idisc,t,cS pf q

pG, G1 q

p1 Sdisc,t,c1,S pf 1 q

Idisc,t pf q
G1 G

pG, G q

cS G1 G 1 c1

c1,S cS p1 Sdisc,t,c1 pf 1 q

Idisc,t pf q
G1 G

p1 pG, G1 qSdisc,t pf 1 q

Sdisc,t pf q, by the analogues for G1 of (3.3.4) and (3.3.5) (which we assume inductively), and the fact that the double sum over c1,S and cS reduces to a simple sum r r over c1 P CA pG1 , 1 q. The triple sum over cS , G1 and pc1 qS above can be taken over a nite set that depends on f only through a choice of Hecke type. This follows from the corresponding property for Idisc,t pf q, the niteness of the sum over G1 , the uniformity properties of the transfer mapping f f 1 that follow from the two theorems in [A11, 6], and our induction hypothesis above for G1 . We set Sdisc,t,c pf q Sdisc,t,cS pf q.
cS c

The assertion (a) for G follows. This in turn completes the induction argument. Suppose that G is arbitrary, as in (b), and that f again belongs to HpG, K S , q. To establish (3.3.6), we consider the expression p1 (3.3.7) pG, G1 q Sdisc,t,c1,S pf 1 q
G1 PEell pGq c1,S cS S attached to any cS P CA pG, q. If f is replaced by its transform fh under S a general element h P Hun , the expression is multiplied by the factor p S q. hpc Recall that the linear form Idisc,t,cS pf q transforms the same way under the S S action of Hun , and that its sum over cS in CA pG, q equals Idisc,t pf q. But S of (3.3.7) equals the sum over c p1 pG, G1 qSdisc,t pf 1 q Idisc,t pf q G1

as well, by (3.2.3) and the analogues of (3.3.4) and (3.3.5) for G1 . Thus, both Idisc,t,cS pf q and (3.3.7) represent the cS -component of the linear form Idisc,t pf q, relative to its decomposition into eigenfunctions under the action S of Hun . Therefore (3.3.7) equals Idisc,t,cS pf q. The required decomposition (3.3.6) is then given by the sum of each side of this identity over those cS that map to the given class c P CA pG, q.

3.3. CONTRIBUTION OF A PARAMETER

143

r More generally, suppose that CA pG, q is the family of equivalence classes relative to some given equivalence relation on CA pG, q. If c now belongs to r CA pG, q, we write Idisc,t,c pf q for the sum of the corresponding linear forms attached to elements of CA pG, q in this equivalence class. We obtain the obvious variant Idisc,t pf q Idisc,t,c pf q
r c P CA pG,q

of the decomposition (3.3.1). Similarly, we have the stable variant Sdisc,t,c pf q Sdisc,t pf q
r c P CA pG,q

of (3.3.5), in case G G0 is quasisplit. If G1 belongs to Eell pGq, we can pull r r the equivalence relation back to CA pG1 , 1 q. We write 1 1 (3.3.8) Idisc,t,c pf 1 q Idisc,t,c1 pf 1 q
c1 c

and (3.3.9)
1 Sdisc,t,c pf 1 q

c1 c

1 Sdisc,t,c1 pf 1 q

r r r for any c P CA pG, q and f 1 P HpG1 , 1 q, the sum being over the elements r r in CA pG1 , 1 q that map to c. This is compatible with the convention above for G, in case c lies in the image of the map and is consequently identied r r r with a class in CA pG1 , 1 q. It is also compatible with the summation over c1 in the decomposition (3.3.6). We can therefore rewrite (3.3.6) slightly more simply as p1 (3.3.10) Idisc,t,c pf q pG, G1 qSdisc,t,c pf 1 q,
G1 PEell pGq

r for any c in CA pG, q. This is the form of the general decomposition we will specialize. Our discussion so far has been quite formal and considerably more r general than we need. It simplies in the two cases of G GpN q and rsim pN q that are our primary concern. We recall that in both of these GPE r r cases, we take 1, G1 G1 and 1 1. r Suppose rst that G equals GpN q. It follows from Theorems 1.3.2 and ` r r 1.3.3 that the mapping cpq from the set pN q GpN q to the ` ` r r r subset CpN q C GpN q of CA GpN q is a bijection. These theorems tell us r in addition that the mapping cS c is injective on the preimage of CpN q ` rA pN q CA GpN q , and hence that the sums over cS above are r in the set C r all superuous in this case. Any element P pN q also gives rise to an

144

3. GLOBAL STABILIZATION

archimedean innitesimal character, the norm of whose imaginary part we denote by tpq. With this notation, we write r r r r (3.3.11) Idisc, pf q Idisc,tpq,cpq pf q, f P HpN q. r Suppose next that G represents an element in Esim pN q. Following the general notation (3.3.8) and (3.3.9), we set (3.3.12) and (3.3.13) Sdisc, pf q Sdisc,tpq,cpq pf q, Idisc, pf q Idisc,tpq,cpq pf q

r for any P pN q and f P HpGq. We will obviously have a special interest r r in the case that lies in the subset pGq of parameters in pN q attached to G. We note, however, that the existence of this subset will require a running induction hypothesis based on the assertion of Theorem 1.4.1. The notation (3.3.12) and (3.3.13) holds more generally if G belongs to the larger r set Eell pN q. We shall state the specialization of (3.3.10) as a corollary of Lemma 3.3.1. It is of course Hypothesis 3.2.1 that tells us that the condition of the r lemma holds if G GpN q. r Corollary 3.3.2. Suppose that belongs to pN q. r (a) If G GpN q, we have r r r r pG rpN, GqSdisc, pf G q, f P HpN q. (3.3.14) Idisc, pf q
r GPEell pN q

r (b) If G P Esim pN q, we have (3.3.15) Idisc, pf q

G1 PE
ell pGq

p1 pG, G1 qSdisc, pf 1 q,

f P HpGq.

Suppose that G and are as in (b). The linear form Idisc, pf q on the left hand side of (3.3.15) then has a spectral expansion ` M G detpw 1q G 1 tr MP, pwqIP, pf q , |W0 ||W0 |1 a
M

M PL

wPW pM qreg

where MP, pwqIP, pf q MP,tpq,cpq pw, 1M qIP,tpq,cpq p1M , f q. It is understood here that the last expression vanishes if cpq lies in the r the complement of Caut pGq, and that 1M stands for the trivial character on ` AM,8 . The same formula obviously applies to the left hand side of (3.3.14) r r if we set G GpN q and f f . We have come at length to the decompositions (3.3.14) and (3.3.15) r that for any P pN q will serve as the foundation for our proofs. They

3.4. A PRELIMINARY COMPARISON

145

are the culmination of the general discussion of the last three sections. The discussion has been broader than necessary, in the hope that it might oer some general perspective, and in addition, to serve as a foundation for future investigations. We turn now to the business at hand. We shall begin a study of the implications of these decompositions for the proofs of our theorems. 3.4. A preliminary comparison Our interest will be focused on the decompositions (3.3.14) and (3.3.15) r attached to a parameter P pN q. One of our goals will be to establish transparent expansions of the two sides of each identity that we can compare. This will be the topic of Chapter 4. In the remaining two sections of this chapter, we shall establish a complementary result that is more modest. r Suppose rst that G equals the twisted general linear group GpN q. The classication of Theorems 1.3.2 and 1.3.3 then characterizes the automorphic spectrum in terms of self-dual families c. For we have observed that the r mapping ` cpq in this case is a bijection from pN q onto the subset r r r CpN q of CA GpN q . We recall that CpN q represents the set of self-dual representations in the automorphic spectrum of GLpN q. In particular, it r represents all the representations that occur in the formula (3.1.1) for GpN q. The summand of c in (3.3.1) therefore vanishes in this case unless c cpq r for some P pN q. The decomposition (3.3.1) thus reduces to r r r r (3.4.1) Idisc,t pf q Idisc, pf q, f P HpN q.
r tPpN q:tpqtu

We would like to establish a similar formula if G is any element in r r Esim pN q. Following the general convention in 3.3, we write CA pGq for the set of bres of the mapping ` r r CA pGq CA pN q CA GpN q . In this case, we will generally be working with the symmetric subalger r bra HpGq of HpGq. Recall from 1.5 that HpGq equals HpGq unless p G SOp2m, Cq, in which case it consists of functions that at each v are symmetric under the automorphism of G we have xed. The bres in CA pGq r of points in CA pGq have a similar description. However, so that there be no misunderstanding, we note that even if f belongs to the symmetric Hecke r algebra HpGq, the function Idisc,t,c pf q of c P CA pGq is not constant on the bres. Indeed, a bre can be innite, while this function has nite support. r We recall that if c is taken to be a class in CA pGq instead of CA pGq, Idisc,t,c pf q is the sum of the function over the bre of c in CA pGq. The following proposition will reduce the study of automorphic spectra r r of our groups G P Esim pN q to the subsets attached to parameters P pN q. Its proof over the next two sections will be a model for the more elaborate global comparisons that will occupy us later.

146

3. GLOBAL STABILIZATION

r r Proposition 3.4.1. Suppose that G P Esim pN q, f P HpGq, t 0 and r c P CA pGq. Then Idisc,t,c pf q 0 Sdisc,t,c pf q, unless ` pt, cq tpq, cpq , r for some P pN q. We shall begin the proof later in this section, and complete it in the next. Before doing so, we shall establish two important corollaries of the proposition, and comment on some related matters. The rst corollary follows immediately from the decompositions (3.3.1) r and (3.3.5) (stated in terms of CA pGq rather than CA pGq), and the denitions (3.3.12) and (3.3.13). r Corollary 3.4.2. For any t 0 and f P HpGq, we have G (3.4.2) Idisc,t pf q Idisc, pf q
r tPpN q: tpqtu

and (3.4.3)
G Sdisc,t pf q

r tPpN q: tpqtu

G Sdisc, pf q.

The other corollary bears directly on our main global theorem. For a r group G P Esim pN q, we dene an invariant subspace ` r t 0, c P CA pGq, L2 disc,t,c GpF qzGpAq , ` of L2 GpF qzGpAq in the obvious way. It is the direct sum disc mpq,
t: },I }t, cpqcu

of the irreducible representations of GpAq attached to t and c that occur ` in L2 GpF qzGpAq with positive multiplicity mpq. Then disc ` 2 ` L2 GpF qzGpAq Ldisc,t,c GpF qzGpAq . disc
t,c

r If belongs to pN q, we set ` ` 2 L2 disc, GpF qzGpAq Ldisc,tpq,cpq GpF qzGpAq . Corollary 3.4.3. If t 0 and c P CA pGq, then ` L2 disc,t,c GpF qzGpAq 0, unless ` pt, cq tpq, cpq ,

3.4. A PRELIMINARY COMPARISON

147

r for some P pN q. In particular, we have a decomposition ` ` (3.4.4) L2 GpF qzGpAq L2 disc disc, GpF qzGpAq .
r PpN q

Proof of Corollary 3.4.3. Assume that pt, cq is not of the form tpq, cpq . The proposition then tells us Idisc,t,c pf q vanishes. The formula for Idisc,t,c pf q given by the c-analogue of (3.1.1) is a sum over Levi subgroups M of G. If M is proper in G, it is a product of general linear r groups with a group G P Esim pN q, for some integer N N . We can assume inductively that the corollary holds if N is replaced by N . Since the same property holds for the general linear factors of M , the analogue of the corollary holds for M itself. The operator IP,t,c p1M , f q in the summand of M is induced from the pt, cq-component of the automorphic discrete spectrum of M , or more correctly, the component of the discrete spectrum given by some partition of pt, cq among the factors of M . This operator then vanishes, by our assumption on pt, q, and so therefore does the summand of M . The remaining term in the formula for Idisc,t,c pf q is the summand ` tr IG,t,c p1G , f q ` corresponding to M G. We have established that it vanishes. But IG,t,c pf q IG,t,c p1G , f q ` is the operator on the space L2 disc,t,c GpF qzGpAq by right convolution of f . Its trace equals the sum ` mpqtr pf q , },I } t, cpq c,

of the irreducible subrepresentations of this space. The function f belongs r only to the (locally) symmetric Hecke algebra HpGq. However, since the multiplicities mpq are all positive, it is clear that the sum cannot vanish r for all f P HpGq unless it also vanishes for all f P HpGq. It follows that mpq 0 for each . In other words, the sum is over an empty set, and the assertions of the corollary hold. It will be convenient to write (3.4.5)
G Rdisc, pf q IG,tpq,cpq pf q,

f P HpGq,

G r r for any G P Esim pN q and P pN q. Then Rdisc, stands for the regular ` representation of GpAq on L2 disc, GpF qzGpAq . The second corollary of (the yet unproven) Proposition 3.4.1 reduces the study of the automorphic G discrete spectrum of G to that of the subrepresentations Rdisc, . Given the statement of Theorem 1.5.2, we will obviously need to reduce the problem r r further to parameters that lie in the subset pGq of pN q. This question turns out to be surprisingly dicult. We will be able to give a partial answer below, after some elementary remarks on central characters. However, its

148

3. GLOBAL STABILIZATION

full resolution will not come before the end of our general induction argument in Chapter 8. r Recall that any G P Esim pN q comes with a character G of F of order 1 or 2. If this character is nontrivial, it determines G uniquely. The same is true if G 1 and N is odd. However, if G 1 and N is even, there r p is a second element G_ P Esim pN q with G_ 1. Its dual group G_ equals p SppN, Cq and SppN, Cq if G SOpN, Cq. In other words, p SOpN, Cq if G G_ is the split group SOpN q if G equals SOpN ` 1q, while G_ equals SOpN ` 1q if G is the split group SOpN q. r Any parameter P pN q also comes with a character on F of order 1 or 2. It is the id`le class character detp q, where is the self e dual automorphic representation of GLpN q attached to . If G , the Tchebotarev density theorem tells us that we can nd a valuation v, at which , G, and G are unramied, such that the local characters ,v v and G,v Gv are distinct. It follows from the denitions that the conjugacy class cpv q in GLpN, Cq does not meet the image of L Gv . This implies that (3.4.6)
G Rdisc, 0,

r G , G P Esim pN q.

r On the other hand, it follows from the construction of the subset pGq of r pN q in 1.4, which we will presently formalize with a running induction r hypothesis, that G if belongs to pGq. This gives a partial charr r acterization of the complement of pGq in pN q, for our simple endoscopic r datum G P Esim pN q. Our partial answer to the question posed above is that G r r vanishes for any in the complement of pGq in pN q, except posR
disc,

sibly in the case that 1, N is even and G is split (which is to say that G 1). It is this last case that contains major unresolved diculties. r r Suppose again that G , for some G P Esim pN q and P pN q. If v is a valuation such that ,v G,v as above, the conjugacy class cpv q in GLpN, Cq does not meet the image of L Mv in GLpN, Cq, for any Levi subgroup M of G. This implies that IP,tpq,cpq pf q 0, r P P PpM q, f P HpGq.

It follows from the analogue of (3.1.1) for that Idisc, pf q 0. Applying this assertion to the left hand side of (3.3.15), we see inductively that a G similar assertion holds for Sdisc, pf q. In other words,
G G (3.4.7) Idisc, pf q 0 Sdisc, pf q,

r r f P HpGq, G , G P Esim pN q.

There is one other remark to be made before we start the proof of the proposition. It concerns a minor point of possible confusion. Our convention of letting the same symbol denote both a group and the endoscopic datum it represents has usually been harmless. For we agreed in the general case 1 (3.2.3) that the linear form Sdisc,t depends on G1 only as a group (or more

3.4. A PRELIMINARY COMPARISON

149

r r correctly, on the pair pG1 , 1 q), and not on the other components G 1 , s1 and 1 of the endoscopic datum. However, the c-component (3.3.9) of S 1 disc,t (as 1 well as its analogue (3.3.8) for Idisc,t ) depends by denition on G1 as an r endoscopic datum. In the case G P Esim pN q we are considering here, this G means that the -component (3.3.13) of the stable linear form Sdisc,t Sdisc,t G (as well as its analogue (3.3.12) for Idisc,t Idisc,t ) depends nominally on G as an endoscopic datum. The question arises when N is odd. In this p case, G SppN 1q is the unique group with G SOpN, Cq, while the endoscopic datum is further parametrized by the quadratic character G . However, this distinction is still innocuous. For the -component (3.3.13) vanishes if G , and in the case that G , it remains unchanged if is replaced with a translate b by a quadratic character , and G is adjusted accordingly. Proof of Proposition 3.4.1. (First step): We assume inductively that the proposition holds if N is replaced by any integer N N . We r will let G P Esim pN q vary, so we need to x a general pair pt, cq, r t 0, c P CA pN q, ` that is independent of G. We assume that pt, cq is not of the form tpq, cpq r for any P pN q. We must establish that the linear forms Idisc,t,c pf q and r r Sdisc,t,c pf q both vanish for any G P Esim pN q and f P HpGq. We x G for the moment, and begin as in the proof of Corollary 3.4.3. Our induction hypothesis implies that the corollary also holds if N is replaced by N N . It follows that the pt, cq-component of the discrete spectrum of any proper Levi subgroup M of G vanishes. The operator IP,t,c pf q in the summand of M in the c-analogue of (3.1.1) therefore vanishes, and so then does the summand itself. Since the only remaining summand corresponds to M G, we see that ` (3.4.8) Idisc,t,c pf q tr IG,t,c pf q , for any f P HpGq. We can apply similar arguments to the decomposition (3.3.10) given by the c-analogue of (3.2.3). Consider an index of summation G1 P Eell pGq in this decomposition with G1 G. Then G1 is a proper product G1 G1 G1 , 1 2 G1 P Esim pNi1 q, Ni1 N. i

p1 The corresponding linear form Sdisc,t,c pf 1 q in (3.3.10) is dened in turn by a sum (3.3.9). It follows from our induction hypothesis and our condition on pt, cq that each of the summands in (3.3.9) vanish. The linear form in (3.3.10) thus vanishes, and so therefore does the summand of G1 . The remaining term in the decomposition (3.3.10) of Idisc,t,c pf q is the stable linear form Sdisc,t,c pf q. We have shown that (3.4.9) Idisc,t,c pf q Sdisc,t,c pf q.

150

3. GLOBAL STABILIZATION

r Next, we apply (3.3.10) again, but this time with GpN q in place of G. r r We have then to replace f by a function f in HpN q. The datum G plays 1 in (3.3.10), and is summed over the general indexing set the role of G ` r r r r Eell pN q Eell GpN q . If G lies in the complement of Esim pN q in Eell pN q, it equals a proper product r G G S GO , G P Esim pN q, N N. Arguing as above, we see from our induction hypothesis, the condition on pt, cq, and the denition (3.3.9) that r pG Sdisc,t,c pf G q 0 in this case. The sum on the right hand side of (3.3.10) can therefore be r r taken over G in the subset Esim pN q of Eell pN q. The left hand side of the rq. It vanishes by the remarks preceding identity is the linear form Idisc,t,c pf (3.4.1) and our condition on pt, cq. The identity becomes r pG rpN, GqSdisc,t,c pf G q 0. (3.4.10)
r GPEsim pN q

This completes the rst step. To state what we have established so far, we shall introduce a global form of the local object (2.1.3) from 2.1. We shall formulate it as a family of functions ( r r (3.4.11) F f P HpGq : G P Eell pN q parametrized by global endoscopic data that are elliptic, with the condition that f 0 for all but nitely many G. Let us say that F is a decomposable compatible family if for each v, there is a local compatible family of functions ( r r Fv fv P HpGv q : Gv P Ev pN q r on the local endoscopic groups Gv pFv q such that for any G P Eell pN q with rv pN q, the corresponding function satises completion Gv P E f fv .
v

We shall then say simply that F is a compatible family if there is a nite set of decomposable compatible families ( r r Fi fi P HpGq : G P Eell pN q , such that the functions attached to any G satisfy f fi . This is a natural analogue of the local denition. We observe from Proposir tion 2.1.1 that F is a compatible family if and only if there is a function f r in HpN q such that r r f G f G, G P Eell pN q.

3.4. A PRELIMINARY COMPARISON

151

r In other words, the function f P F attached to any G P Eell pN q has the same r r image in SpGq as f . The notion of Hecke type can obviously be formulated here. By a Hecke r type for Eell pN q, we shall mean a pair ` ( (3.4.12) S, p8 , 8 q , ( where S S8 is a nite set of valuations, and p8 , 8 q is a nite set of r Hecke types, one for each G P Eell pN q that is unramied outside of S, such that 8 8 K S , 8 GpA8 q, S S S for a hyperspecial maximal compact subgroup K S of GpAS q. Then (3.4.12) is a Hecke type for F if the function f attached to G vanishes whenever G ramies outside of S, and has Hecke type p8 , 8 q if G is unramied outside of S. Any compatible family F obviously has a Hecke type. r r Suppose that F is any compatible family (3.4.11), and that f P HpN q r is chosen so that f G f G for any G. We can then write ` rpN, Gqtr IG,t,c pf q rpN, GqSdisc,t,c pf q
r GPEsim pN q G

p rpN, GqSdisc,t,c pf G q r p rpN, GqSdisc,t,c pf G q,

by (3.4.8) and (3.4.9). According to (3.4.10), this last sum vanishes. We conclude that ` rpN, Gqtr IG,t,c pf q 0, f P F, (3.4.13)
r GPEsim pN q

if F is any compatible family (3.4.11). Since rpN, Gq 0, we can write the summand of G as a linear combination ` cG pqtr pf q , P unit pGq,

of irreducible unitary characters on GpAq with nonnegative coecients cG pq. We can then write (3.4.13) in the form ` (3.4.14) cG pqtr pf q 0, f P F,
r GPEell pN q Punit pGq

for nonnegative coecients cG pq, which actually vanish if G lies in the r complement of Esim pN q. The identity (3.4.14) is valid for any compatible family F, and the double sum can be taken over a nite set that depends on F only through the choice of a Hecke type.

152

3. GLOBAL STABILIZATION

In the next section we use the identity (3.4.14) to show that all of the coecients cG pq vanish. In fact, we will establish general vanishing properties, which will be a foundation for more subtle comparisons later on. The coecients are dened by the decomposition of the pt, cq-component IG,t,c of the representation of GpAq on the discrete spectrum. If they vanish, we see that ` ` tr IG,t,c pf q rpN, Gq1 cG pqtr pf q 0,

r for any G P Esim pN q and any f P HpGq. It then follows from (3.4.8) and (3.4.9) that Idisc,t,c pf q Sdisc,t,c pf q 0, as required. In other words, we will have a proof of Proposition 3.4.1 once we have shown that the coecients cG pq vanish. 3.5. On the vanishing of coecients We have to complete the proof of Proposition 3.4.1. In the last section, we reduced the problem to proving that the coecients cG pq in the identity (3.4.14) all vanish. We shall now deduce that in any such identity, which we recall is associated to the global eld F and the positive integer N , the coecients automatically vanish. Proposition 3.5.1. Suppose that there are nonnegative coecients cG pq, r G P Eell pN q, P pGq, f P F, such that for every global compatible family F (3.4.11), the function pG, q cG pqfG pq,

is supported on a nite set that depends only on a choice of Hecke type for F, and such that the double sum (3.5.1) cG pqfG pq, f P F,
G

vanishes. Then cG pq 0, for every G and . The proof of Proposition 3.5.1 is primarily local. In order to focus on the essential ideas, we shall revert briey to the local notation of Chapter 2. Thus, in contrast to the convention that has prevailed to this point in Chapter 3, and until further notice later in the section, we take F to be a r r r local eld. The sets EpN q, pGq, HpGq, and so on, are then to be understood as local objects over F . We shall take advantage of this interlude to recall some notions from local harmonic analysis, both for the proof and for later use. We shall review the theory of the representation theoretic R-group [A10, 13], which is founded on work of Harish-Chandra, Knapp, Stein and Zuckerman on intertwining operators (see [Ha4], [Kn], [KnS], [KnZ1] and [Si2]). For these remarks,

3.5. ON THE VANISHING OF COEFFICIENTS

153

we may as well allow G to be an arbitrary connected reductive group over F. As in the global notation of 3.1, we let L LG LG pM0 q, be the nite set of Levi subgroups of G that contain a xed minimal Levi G subgroup M0 , and we write W0 for the Weyl group W pM0 q of M0 . The set of GpF q-conjugacy classes of Levi subgroups of G is bijective with the set of G G W0 -orbits in L. We then write T pGq for the set of W0 -orbits of triplets (3.5.2) r pM, , rq, M P L, P 2 pM q, r P Rpq, where Rpq is the R-group of in G. We recall that Rpq is given by a short exact sequence 1 W 0 pq W pq Rpq 1, where W pq is the stabilizer of in the Weyl group W pM q of M , and W 0 pq is the Weyl group of a root system dened by the vanishing of Plancherel densities. Any choice of positive chamber for the root system determines a splitting of the sequence, and thereby allows us to identity Rpq with a subgroup of W pq. It is known how to construct general normalized intertwining operators RP pw, q : IP pq IP pq, w P W pq, from the basic intertwining integrals, which satisfy analogues of (2.3.9) and (2.3.10). (The focus of the discussion of 2.2.3 and 2.3.4 was more specic, namely to establish particular normalizations that would be compatible with endoscopy.) The group W 0 pq becomes the subgroup of elements w P W pq such that RP pw, q is a scalar. What is not known in general is whether these normalizations can be chosen to be multiplicative in w. The question really pertains to the R-group Rpq [A10, p. 91, Remark 1], and is equivalent to whether the irreducible representation has an extension to a semidirect product M pF q Rpq. One sidesteps the problem by introducing a nite central extension r 1 Z Rpq Rpq 1 of Rpq. It then becomes possible to attach normalized intertwining operar tors RP pr, q to elements r in Rpq such that r RP pzr, q pzq1 RP pr, q, z P Z , r P Rpq, for a xed character of Z , and so that the mapping r r RP pr, q, r P Rpq, r is a homomorphism from Rpq to the space of intertwining operators of rpGq for the set of W G -orbits of triplets of the form IP pq. We write T 0 r (3.5.2), but with r in the extension Rpq of Rpq. We then set ` (3.5.3) fG p q fG pr q tr RP pr, qIP p, f q , f P HpGq,

154

3. GLOBAL STABILIZATION

r for any element r in T pGq. r The groups Rpq and Rpq do not have to be abelian. However, we can ` r r still work with the set Rpq, of irreducible representations of Rpq whose Z -central character equals . For any such , the linear form ` (3.5.4) fG p q |Rpq|1 tr prq fG pr q, f P HpGq,
rPRpq

is the character of a subrepresentation of IP pq, which turns out to be irreducible if is tempered. r We write Ttemp pGq and Ttemp pGq for the subsets of triplets in the r respective sets T pGq and T pGq for which the second component of (3.5.2) belongs to 2,temp pM q, the subset of tempered representations in 2 pM q. For any such , we write pGq for the set of irreducible constituents of the induced tempered representations IP pq of GpF q. It is well known that the set temp pGq of all irreducible tempered representations of GpF q is a G disjoint union, over W0 -orbits of pairs pM, q, of the sets pGq. (See for example [A10, Proposition 1.1].) Moreover, the basis theorem of HarishChandra [Ha4] tells us that if belongs to 2,temp pM q, and tru is a set of r representatives in Rpq of the conjugacy classes in Rpq, the corresponding set of linear forms r fG pr q, f P HpGq, is a basis of the space spanned by the characters of representations in pGq. It follows that the mapping ` r , P Rpq, , ` r is a bijection from Rpq, to pGq. The full set temp pGq can thereG fore be identied with the family of W0 -orbits of triplets ` r pM, , q, M P L, P 2,temp pM q, P Rpq, . (See [A10, 3], where the set Ttemp pGq was denoted by T pGq. We need to reserve the symbol T pGq here for nontempered triplets.) One thus obtains a rough classication of temp pGq from local harmonic analysis. It is independent of the ner endoscopic classication, which for r simple endoscopic data G P Esim pN q is contained in the assertions of the local theorems. We will eventually relate the two classications in Chapter 6, as part of the proof of the local theorems. This will obviously entail letting r r represent an element M P M , with M P 2 pM, q being a preimage r of a generic parameter P pGq. The resulting relations will then conrm 0 pq W 0 , Rpq R , and therefore that the two that W pq W , W kinds of R-groups are the same. In the general nontempered case, we have to distinguish between irreducible representations P pGq and standard representations P P pGq. The general relations between the two kinds of representations are parallel

3.5. ON THE VANISHING OF COEFFICIENTS

155

to those between extensions P pN q and P P pN q discussed in 2.2. Let r r r r us review them. There is a bijection , P pGq,

between pGq and P pGq such that is the Langlands quotient of . The character of any has a decomposition fG pq np, q fG pq, f P HpGq,
PP pGq

into standard characters. The coecients np, q are uniquely determined integers, with np, q 1, which have nite support in for any . Both and determine real linear forms and . These objects lie in the dual closed chamber pa q` in 0 a0 aM0 attached to a preassigned minimal parabolic subgroup P0 P PpM0 q of G, and are a measure of the failure of the representations to be tempered. In particular, is the point in the relatively open cone pa q` pa q` , P 0 P P0

that represents the nonunitary part of , as a representation induced from P pF q. If np, q 0, then , in the usual sense that is a nonnegative integral combination of simple roots of pP0 , A0 q, with equality holding if and only if equals . (See [A7, Proposition 5.1], which is based on results in [BoW] and underlying ideas of Vogan.) Standard characters can in turn be decomposed into the virtual characr ters parametrized by T pGq. For it follows from the denition of a standard representation, and the general classication of temp pGq described above, that the set P pGq of standard representations of GpF q can be identied with G the family of W0 -orbits of triplets ` r (3.5.5) pM, , q, M P L, P 2 pM q, P Rpq, . We dene ` x, y |Rpq|1 tr prq , r for as in (3.5.5) and r as in the analogue for T pGq of (3.5.2). rpGq that are comWe also set x, y 0 for elements P P pGq and P T plementary, in the sense that their corresponding triplets do not share the G same W0 -orbit of pairs pM, q. It then follows that (3.5.6) fG pq np, qfG p q, P pGq, f P HpGq,
PT pGq

where np, q
PP pGq

np, qx, y,

r P T pGq.

156

3. GLOBAL STABILIZATION

r If belongs to T pGq, we set , for any element P P pGq with x, y 0. Then is a linear form in pa q` , with the property that 0 if np, q 0. Elliptic elements in T pGq have a special role. We recall that Tell pGq is G the subset of (W0 -orbits of) triplets r in (3.5.2) such that r is regular, in the sense that aG tH P aM : rH Hu. In general, we will write Greg Gstr-reg for the open connected subset of strongly regular points in G, and Greg,ell pF q Gstr-reg,ell pF q for the subset of elements in Greg pF q that are elliptic over F . Thus, Greg,ell pF q is the subset of elements in GpF q whose centralizer is a torus T such that T pF q{AG pF q is compact. (As in Chapter 2, we will generally use the simpler notation on the left, since we will not need to consider regular points that are not strongly regular.) Elements in Tell pGq have the important property that they are uniquely determined by the restriction of their virtual characters to Greg,ell pF q. To be more precise, let t u be a set of representatives of the set Tell pGq r r in its preimage Tell pGq in T pGq. Then the corresponding set of distributions ` 8 fG p q, f P Cc Greg,ell pF q , on Greg,ell pF q is a basis for the space spanned by the restriction to Greg,ell pF q of characters of representations in either of the sets pGq or P pGq. (See [Ka] and [A11]. This is also a consequence of the orthogonality relations in [A11, G 6].) It follows from the denitions that T pGq is a disjoint union, over W0 orbits of Levi subgroups M , of images in T pGq of the corresponding elliptic sets Tell pM q. For any P T pGq, we write M for a Levi subgroup such that lies in the image of Tell pM q. Suppose that P pGq. If the associated standard representation P P pGq corresponds to the triplet pM , , q under (3.5.5), we dene r an element in either T pGq or T pGq by the triplet pM , , 1q. The two elements and are such that both np, q and x , y are nonzero, and they satisfy . Conversely, suppose that P P pGq and P T pGq are elements such that both np, q and x, y are nonzero ( being identied here with a representative r in T pGq). Then . If these linear forms are all equal, M contains M (with M and M taken G to be suitable representatives of the associated W0 -orbits in L). Moreover,

3.5. ON THE VANISHING OF COEFFICIENTS

157

with the assumption that , the group M equals M if and only if and . We observe that in this case, we have np, q np, qx , y 1 |Rp q|1 dimp q 0. We now return to the case that G represents a twisted endoscopic datum for GLpN q. The statement of Proposition 3.5.1 makes sense over the local eld F . The set F has of course to be understood as a local comr patible family (2.1.3), with the implication that the global set Eell pN q is r replaced by the general local set EpN q. A Hecke type for F will be family r of objects, parametrized by the nite set of G P EpN q, consisting of open compact subgroups K0 GpF q if F is p-adic, and nite sets R of irreducible representations of maximal compact subgroups KR GpRq if F R. r Lemma 3.5.2. Proposition 3.5.1 is valid, stated with EpN q in place of rell pN q, over the local eld F . E Proof. We are given that the sum (3.5.7) cG pqfG pq,
G

r over G P EpN q and P pGq, vanishes for any compatible family F over the local eld F . The function f of course stands for the element of F indexed by G. We shall let F vary over the compatible families attached to a xed local Hecke type. By the general discussion above, we can write the sum (3.5.7) as cG pqnp, qfG p q
r GPEpN q PpGq PT pGq

G G |W0 p q||W0 |1 cG pqnp, qfM p q,

M PLG PTell pM q

G G where W0 p q is the stabilizer of pM, q in W0 . The coecient np, q in G -invariant function of P T pM q, rell the last expression is understood as a W0 which is to say, a function whose product with fM p q depends only on the image of in T pGq. The multiple sum can be taken over a nite set of indices pG, , M, q, which depends only on the given local Hecke type. A Levi subgroup M of G can be treated in its own right as an element r q. Conversely, for any M P EpN q, there will generally be several r in EpN r G P EpN q in which M embeds as a Levi subgroup. The best way to ` account rM pN q EM GpN q r for these groups is as elements in the twisted analogue E of the set EM 1 pGq dened (for any connected reductive group G and any endoscopic datum M 1 for G) on p. 227 of [A12]. We leave the reader to forr mulate the denition of EM pN q as a set of twisted endoscopic data pG, s, q

158

3. GLOBAL STABILIZATION

r for GpN q, which contain M as a Levi subgroup, and are taken up to translation of s by elements in the product ` ` p 0 p r ZpGq Z GpN q . ` p 0 (The inclusion of the connected component ZpGq is a departure from r the convention of [A12], and is designed to make the set EM pN q nite.) r We can then replace the last double sum over G P EpN q and M P LG by a r q and G P EM pN q, provided that we adjust the r double sum over M P EpN r coecients to compensate for the fact that several elements in EM pN q could r represent the same isomorphism class in EpN q. It follows that (3.5.7) equals cG pqnp, qG p qfM p q,
r r M PEpN q PTell pM q GPEM pN q PpGq

for positive constants G p q. r For any M P EpN q, the function fM p q in the last sum depends on G, since it comes from the function f P F attached to G. The family fM p q, r r r M P EpN q, G P EM pN q, P Tell pM q,

of such functions is indexed by F. As F varies, the corresponding set of families is a natural vector space. It is not hard to see from the denition of a compatible family (2.1.3), together with the trace Paley-Wiener theorem r [CD] [DM], that this vector space is a direct sum over M P EpN q of the rell pM q. It follows subspaces of (families of) functions that are supported on T that the sum cG pq np, q G p q fM p q
r PTell pM q GPEM pN q PpGq

vanishes for any M , and any compatible family F. r We x M P EpN q, and consider the function fM p q attached to G M . r By the denition (2.1.3), fM lies in the subspace of OutN pM q-invariant functions in IpM q. We will assume that fM also lies in the subspace Icusp pM q of cuspidal functions in IpM q, but this will be the only other constraint. It then follows from the trace Paley-Wiener theorem for M , and the linear independence of the set of distributions on Mstr-reg,ell pF q attached to Tell pM q, r that we can take fM p q to be any OutN pM q-invariant, 1 -equivariant func r tion in the natural Paley-Wiener space on Tell pM q. Since we are working with a preassigned Hecke type, we are actually appealing here to an implicit consequence of the trace Paley-Weiner theorem. Namely, if is a nite set of irreducible representations of the maximal compact subgroup KM of M pF q, and IpM q and HpM q are the corresponding subspaces of IpM q and HpM q, there is a suitable section IpM q HpM q .

3.5. ON THE VANISHING OF COEFFICIENTS

159

(See [A5, Lemma A.1], for example.) The only other constraint imposed r by the denition of F is that as G varies, the images f M in SpM q of the r functions f P HpGq are all equal. It follows from this that we can take the functions fM p q parametrized by G to be equal. In particular, we can r choose the common function fM p q so that it isolates any OutN pM q-orbit in Tell pM q. It follows that the sum (3.5.8) cM prq cG pq np, q G p q
r GPEM pN q PpGq Pr

r r r r vanishes for any M P EpN q and any OutN pM q-orbit in Tell pM q. It is instructive to introduce an equivalence relation on the set of pairs pG, q, For any pG, q, and any pair pM, q, r P Tell pM q, r G P EpN q, P pGq.

we set np, q equal to the coecient we have dened if G represents a r datum in Eell pN q, and 0 otherwise. With this understanding, we write pG, q pG1 , 1 q if there is a pair pM, q such that both np, q and np1 , q are nonzero. We can then dene -equivalence to be the equivalence relation generated by this basic relation. If P is any -equivalence class, we write P for the adjoint equivalence class in the set of pairs pM, q. It consists of those pM, q such that np, q is nonzero for some pG, q in P . One sees without diculty that the classes P and P are both nite. These two equivalence relations are natural variants of the familiar relations of block equivalence, which are dened on the two sets pGq and P pGq attached to any connected reductive group G over F . We shall use them in the analysis of (3.5.8). Assume that the assertion of the lemma is false. We can then x a -equivalence class P such that the subset ( P 1 pG, q P P : cG pq 0 is not empty. We dene 1 max } } : pG, q P P 1 and ( 1 min dimpM q : pG, q P P 1 , } } 1 . (As in the general setting of 3.2, the Hermitian norm } } on the space a 0 attached to G is dened in terms of a suitable, preassigned norm on the r space r attached to GpN q.) Let pG1 , 1 q be a xed pair in P 1 such that a0 ` }1 }, dimpM1 q p1 , 1 q. The pair pM 1 , 1 q pM1 , 1 q (

160

3. GLOBAL STABILIZATION

belongs to P , since we can treat 1 pM 1 , 1 , 1q r as an element in Tell pM 1 q. It also has the property that ` } 1 }, dimpM 1 q p1 , 1 q. r Moreover, G1 is represented by a datum in EM 1 pN q. We shall use the fact 1 q 0 to derive a contradiction. that cM 1 pr Suppose that pG, q is any pair in P 1 such that G is represented by a r r r datum in EM 1 pN q, and that is any element in the OutN pM 1 q-orbit 1 such that p, q 0. The latter condition implies that , that M M in case , and that in case both and M M hold. But it is clear that ` ` ` } }, dimpM q } 1 }, dimpM 1 q }1 }, dimpM1 q , all three pairs being equal to the extremal pair p1 , 1 q. It then follows from the denitions that the two relations and M M do indeed hold, so that . The coecient p, q is therefore strictly positive. The summand of G, and in the expression (3.5.8) for cM 1 pr1 q is consequently positive. The rst two sums in (3.5.8) reduce to a sum over pairs pG, q in P , according to the denition of -equivalence. We have just seen that a summand of pG, q and is either positive or zero. Since the summand of pG1 , 1 q and 1 is positive by assumption, the coecient cM 1 pr1 q itself is positive. This gives the desired contradiction, and completes the proof of Lemma 3.5.2. We now go back to the global setting. For the rest of the chapter (and indeed until the end of Chapter 5) we assume that the eld F is global. We need to establish Proposition 3.5.1. We shall do so by applying the local proof above to the completions Fv of F . Proof of Proposition 3.5.1. Let S S8 be a nite set of valuations r of the global eld F . We shall write ES pN q for the nite set of products r GS Gv , Gv P Ev pN q,
vPS

` r r where Ev pN q denotes the set E Gv pN q of twisted endoscopic data for GLpN q over Fv , as earlier. We can then form the set ! ) pGS q S v : v P pGv q
vPS

r of irreducible representations of GS . Let us also write Eell pN, Sq for the rell pN q that are unramied outside nite set of global endoscopic data G P E of S, and pG, Sq for the set of irreducible representations of GpAq that are unramied outside of S. For a given pair r pGS , S q, GS P ES pN q, S P pGS q,

3.5. ON THE VANISHING OF COEFFICIENTS

161

we then dene cGS pS q

pG,q

cG pq,

G P Eell pN, Sq, P pG, Sq.

The sum here is taken over the preimage of pGS , S q under the localization mapping pG, q pGS , S q, and can be restricted to a nite set, in view of the given condition on the Hecke type. We are also given that each coecient cG pq is nonnegative, so the same is true of cGS pS q. It therefore suces to show that cGS pS q vanishes for any S and any pGS , S q. Suppose that ! ) r r FS Fv fS fv : fv P HpGv q, Gv P Ev pN q
vPS v

is a product of local compatible families Fv . Then FS determines a decomposable global compatible family ( r F f P HpGq : G P Eell pN q r in the natural way. That is, f vanishes unless G lies in Eell pN, Sq, in which r case f is the product of the function fS P HpGS q with the characteristic function of a hyperspecial maximal compact subgroup K S of GpAS q. The expression (3.5.1) becomes a double sum fS P FS , (3.5.9) cGS pS qfS,GS pS q,
GS S

which can be taken over a nite set that depends only on a choice of Hecke type for FS . We are told that (3.5.9) vanishes. We must use this to deduce that each cGS pS q vanishes. We have simply to duplicate the proof of Lemma 3.5.2, amplifying the notation in the appropriate way (generally without comment). We nd that (3.5.9) equals (3.5.10) cGS pS qnpS , S qGS pS qfS,MS pS q,
M S S G S S

r r for sums over MS P ES pN q, S P Tell pMS q, GS P EMS pN q and S P pGS q, r and for positive constants GS pS q. It is understood here that EMS pN q rMv pN q (rather than the global stands for the product of the local sets E object dened on [A12, p. 241]). We can then take each fS,MS pS q to be r any general, cuspidal, OutN pMS q-invariant Paley-Wiener function that is r independent of GS . Varying fS , we conclude that for any MS P ES pN q, and r N pMS q, the sum r any orbit S of Out (3.5.11) cMS prS q cGS pS qnpS , S qGS pS q
r GS PEMS pN q S PpGS q S PrS

162

3. GLOBAL STABILIZATION

vanishes. Following the proof of the lemma further, we x a S -equivalence class ( r PS pGS , S q : GS P ES pN q, S P pGS q . To apply this part of the local argument, we enumerate the elements v in S as v1 , . . . , vn , and introduce a lexicographic order on the associated local objects }v } and dimpMv q. We take the alphabet A to be the set of pairs x p, q, for a nonnegative real number and a nonnegative integer. It has a linear order dened by x p, q x1 p1 , 1 q, if 1 , or if 1 and 1 . We take the dictionary DS to be the set AS An , with the associated lexicographic order. There is then a mapping ` pGS , S q wpGS , S q }v }, dimpMv q
vPS

from PS to DS , as well as a mapping pMS , S q wpMS , S q from the adjoint class to DS . We need to show that the set
1 PS pGS , S q P PS : cGS pS q 0 PS

`
vPS

}v }, dimpMv q

is empty. Suppose that it is not. We can then take the largest element
1 w1 wpG1 , S q, S 1 1 pG1 , S q P PS , S 1 1 in the image of PS in DS . The choice of pG1 , S q in the preimage also gives S us an associated pair 1 1 1 1 pMS , S q pMS , S q 1 in PS . Suppose that pGS , S q is any pair in PS , and that S is any element in 1 the orbit S of S such that npS , S q 0. It follows from the lexicographic r1 order that wpMS , S q wpGS , S q,

with equality if and only if S S . By denition, we have

1 1 1 wpMS , S q wpMS , S q wpG1 , S q w1 . S

It then follows from the denition of w1 that S S , and hence that pS , S q is strictly positive. The summand of GS , S and S in (3.5.11) is therefore positive. The rst two sums in (3.5.11) reduce to a double sum over pairs pGS , S q in PS . Any summand in (3.5.11) is consequently either positive or zero. 1 1 Since the summand of pG1 , S q and S is positive, the sum itself is positive, S 1 1 contradicting the fact that cMS prS q vanishes. This completes the proof of Proposition 3.5.1.

3.5. ON THE VANISHING OF COEFFICIENTS

163

The proof we have just completed is largely formal, though this may not be apparent in the general framework we have had to adopt. It is essentially a consequence of the decomposition of irreducible representations into standard representations. We will also need a more technical generalization of the proposition. We shall state it as a corollary, since its proof is the same. Corollary 3.5.3. Suppose that we have coecients cG pq that are as in the proposition, except that for a xed valuation v and a simple datum r G1 P Esim pN q, the given sum (3.5.1) equals a separate expression v (3.5.12) d1 pv , f1 qfv,G1 pv q,
v PT pG1,v q

in which the function f1 corresponding to G1 in the compatible family F is taken to be a product ` v v r r f1 fv f1 , fv P HpG1,v q, f1 P H GpAv q , and where the coecient
v d1 pv , f1 q,

v P T pG1,v q,

r is an OutN pG1,v q-invariant function of v , which is supported on a nite set that depends only on a choice of Hecke type for fv , and which equals 0 for v any v of the form pMv , v , 1q. Then the coecients cG pq and d1 pv , f1 q all vanish. Proof. Suppose that the coecients cG pq all vanish. Then the sum (3.5.1) vanishes for any F, and so therefore does the expression (3.5.12). G r r Since G1 is simple, the mapping fv fv 1 from HpG1,v q to SpG1,v q is surjective by Corollary 2.1.2. We can therefore choose the compatible family F so that the function fv,G1 pv q takes preassigned values on any nite set of v r OutN pG1,v q-orbits v . It follows that the coecients d1 pv , f1 q also vanish. We have thus only to show that each cG pq vanishes. As in the proof of the proposition, we x a nite set of valuations S S8 , which we can assume contains v. We then see that it is enough to show that the coecients cGS pS q all vanish. Following the next step in the proof of the proposition, we see that the original sum (3.5.1) takes the form (3.5.10), an expression that is therefore equal to (3.5.12). We can again choose the function fS,MS pS q in (3.5.10) as we wish, and in particular, so that it isolates r r a given OutN pMS q-orbit S . If the v component of any representative of this orbit is of the form pMv , v , 1q, the expression (3.5.12) vanishes, by the given v condition on the coecient d1 pv , f1 q. It follows that the coecient cMS prS q dened by (3.5.11) vanishes in this case. This is all we need to complete the argument in the proof of the propo1 1 sition. For it was the vanishing of cMS prS q, in the case of the chosen pair
1 1 1 1 pMS , S q pMS , S q,

164

3. GLOBAL STABILIZATION

1 from which we deduced that the coecients cMS pS q all vanish. But S is of the form 1 1 1 1 S S pMS , S , 1q, by our denition of maximality. The assertions of the corollary follow.

With the proof of Proposition 3.5.1, we have established that the coecients cG pq in the earlier identity (3.4.14) vanish. This completes the proof of Proposition 3.4.1, the problem we started with. In particular, the expansions of Corollaries 3.4.2 and 3.4.3 are valid. They tell us that for r any G P Esim pN q, the discrete spectrum and the discrete part of the trace r formula are both delimited to the original global parameters P pN q of Chapter 1. We can therefore concentrate on the contributions of these parameters. This leads to a ner analysis, which will occupy the rest of the volume, and in which Proposition 3.5.1 and its corollary will have a critical role.

CHAPTER 4

The Standard Model

4.1. Statement of the stable multiplicity formula We assume that G is as in the theorems stated in Chapter 1. It is therefore a connected, quasisplit orthogonal or symplectic group, which is to say that it represents an element in the set

` r r Esim pN q Esim GpN q ,

r r N 1, GpN q GLpN q ,

of isomorphism classes of simple, twisted endoscopic data for GLpN q over F . We will later need to broaden the discussion somewhat, in order to include r at least the case G GpN q. It will also be useful to add more general comments from time to time for added perspective, and to lay foundations for future study. However, we shall always be explicit whenever G represents r something other than the basic case of an element in Esim pN q. We also assume that the underlying eld F is global. We have seen that the only families c P Caut pN q that contribute to the discrete part of the trace r formula for G come from parameters P pN q. We expect that must r r q. But as we have noted, this actually belong to the subset pGq of pN fact is deep, and will be established only after a sustained analysis of other r questions. We must therefore work with a general parameter P pN q for the time being. r Observe that in even mentioning the set pGq, we are making the implicit assumption that Theorem 1.4.1 is valid. This was the seed theorem, which when applied to the simple generic constituents i of led us in 1.4 r to a denition of pGq. In this sense, Chapter 4 resembles Chapter 2 in relying on some cases of our stated theorems for its denitions and arguments. We recall that Chapter 3 was more direct, since it was independent of any part of the local and global theorems. At the beginning of 4.3, we shall take on some formal induction hypotheses, which replace these informal implicit assumptions, and which will carry the argument into later chapters. In 3.3, we described the spectral expansion (4.1.1) ` M G Idisc, pf q |W0 ||W0 |1 | detpw 1qaG |1 tr MP, pwqIP, pf q
M w 165
M

166

4. THE STANDARD MODEL

r for the contribution of our parameter P pN q. We also established an endoscopic expansion p1 (4.1.2) Idisc, pf q pG, G1 q Sdisc, pf 1 q
G1 PEell pGq

for this contribution. Our long term goal is to extract as much information as possible from the identity of these two expansions. In this section, we p1 shall state a formula for the linear form Sdisc, pf 1 q on the right hand side of (4.1.2). It will suce to describe the case that G1 G. We have therefore to state a putative formula for the stable linear form
G Sdisc, pf q Sdisc, pf q,

r f P HpGq.

The formula depends on local information yet to be established. Specically, it is based on the stable linear form postulated in Theorem 2.2.1(a), or rather a global product G fv , fv pv q, f (4.1.3) f G pq
v v

of these objects. Theorem 1.4.2 asserts that this product makes sense if r r lies in the subset pGq of pN q. However, the formula also has a global component, which will add to the complexity of its eventual proof. The stable linear form Sdisc, pf q is uniquely determined by the expansion (4.1.2). The stable linear form f G pq is uniquely determined by the conditions of Theorem 2.2.1(a). Our formula asserts that the rst linear form equals a multiple of the second by an explicitly given constant, which r vanishes unless lies in pGq. It is this constant that contains the global information. We will call it the stable multiplicity of . The global constant has three factors that will be familiar from the denitions of 1.5. They are the integer m , the inverse of the order of the nite group S , and a special value (4.1.4) G pq ps q G ps q of the sign character G . The point s P S was dened in 1.4 as the image of the nontrivial central element of SLp2, Cq. Since m vanishes r unless belongs to pGq, in which case the objects S and G make sense, the product of the three factors is well dened and vanishes unless lies r in pGq. There will also be a fourth factor. This is the number pS 0 q constructed from in [A9]. We shall review its denition. The factor pS 0 q is part of a general construction that applies to any complex (not necessarily connected) reductive group S. To allow for induction, we take S more generally to be any union of connected components in some complex reductive group. We write S 0 for the connected component of 1, and ZpSq CentpS, S 0 q

4.1. STATEMENT OF THE STABLE MULTIPLICITY FORMULA

167

for the centralizer of S in S 0 . Let T be a xed maximal torus in S 0 . We can then form the Weyl set W pSq NormpT, Sq{T of automorphisms of T induced from S. The Weyl group W 0 W pS 0 q of S 0 is of course an obvious special case. We write Wreg pSq for the subset of elements w P W pSq that are regular, in the sense that the xed point set of w in T is nite. This property holds for w if and only if the determinant detpw 1q is nonzero, where pw 1q is regarded as a linear transformation on the real vector space ` aT Hom XpT q, R . Finally, we dene the sign s0 pwq sgn0 pwq 1 of an element w P W to be the parity of the number of positive roots of pS 0 , T q mapped by w to negative roots. Given these various objects, we attach a real number (4.1.5) ipSq |W pSq|1 s0 pwq| detpw 1q|1
wPWreg pSq

to S. The number ipSq bears a formal resemblance to the spectral expansion (4.1.1) of Idisc, pf q. The rest of the construction amounts to the denition of a number epSq that bears a similar resemblance to the endoscopic expansion (4.1.2). This second number is dened inductively in terms of factors pS1 q attached to connected complex reductive groups S1 . Let us write Sss for the set of semisimple elements in S. For any s P Sss , we set (4.1.6) Ss Centps, S 0 q, the centralizer of s in the connected group S 0 . Then Ss is also a complex reductive group, with identity component
0 Ss pSs q0 Centps, S 0 q0 .

(The notation here diers from the convention used in some places, in which the symbol Ss is reserved for the identity component of the centralizer. We must also take care not to confuse elements s P Sss with the sign character in (4.1.5), which we will always denote with the superscript 0.) If is any subset of S that is invariant under conjugation by S 0 , we shall write Epq for the set of equivalence classes in ss X Sss , with the equivalence relation dened by setting s1 s

168

4. THE STANDARD MODEL

if s1 zs0 sps0 q1 , The essential case is the subset (4.1.7)

0 s0 P S 0 , z P ZpSs q0 .

( 0 Sell s P Sss : |ZpSs q| 8

of S. For among other things, the equivalence relation in Sell that denes the quotient Eell pSq EpSell q 0 -conjugacy. The resemblance here with the (more complicated) is simply S denition of endoscopic data [KS] is not accidental. The following proposition is a restatement of Theorem 8.1 of [A9]. Proposition 4.1.1. There are unique constants pS1 q, dened whenever S1 is a connected complex reductive group, such that for any S, the number 0 (4.1.8) epSq |0 pSs q|1 pSs q
sPEell pSq

equals ipSq, and such that (4.1.9) pS1 q pS1 {Z1 q|Z1 |1 ,

for any central subgroup Z1 of S1 . To see the uniqueness of the constants pS1 q, one notes that (4.1.8) implies that pS1 q equals 0 if the center ZpS1 q is innite. If ZpS1 q is nite, we dene pS1 q inductively from (4.2.8) by setting 0 pS1 q|ZpS1 q| ipS1 q |0 pS1,s q|1 pS1,s q,
s

` where s is summed over the complement of Eell ZpS1 q in Eell pS1 q. We refer the reader to [A9, 8] for the proof of the general identities (4.1.8) and (4.1.9). Remark. If S is a single connected component, it is a bitorsor under the actions of S 0 by left and right translation. The constructions actually make sense if S is taken to be any bitorsor, relative to a complex connected reductive group S 0 . Proposition 4.1.1 remains valid in this setting, since the group multiplication on S plays no role in the proof. More generally one could take S to any nite union of S 0 -torsors. r We return to our parameter P pN q. Recall that S stands for the p of the image of (some representative of) . As in the local centralizer in G discussion from 2.4, we shall often work directly with the quotient (4.1.10) p S S {ZpGq , F .
0

It is of course the group S 0 pS q S s {S s

4.1. STATEMENT OF THE STABLE MULTIPLICITY FORMULA

169

of connected components of S that governs the main assertion of our basic global Theorem 1.5.2. The fourth global factor is the number pS 0 q, dened for the connected reductive group S1 S 0 by the proposition. We can now state the formula for Sdisc, pf q. r Theorem 4.1.2 (Stable multiplicity formula). Given P pN q and rsim pN q, we have GPE (4.1.11) Sdisc, pf q m |S |1 pS 0 q G pq f G pq,

r for any f P HpGq. In particular, Sdisc, pf q vanishes unless belongs to the r r subset pGq of pN q. This represents one of the main results of the volume. It will have to be proved at the same time as the theorems stated in Chapter 1, and their local renements from Chapter 2. However, we will be able to apply it inductively to endoscopic groups G1 P Eell pGq that are proper. We write the specialization of the formula to G1 in slightly dierent terms. r r Recall that elements P pGq can be identied with AutN pGq-orbits of L-homomorphisms r G : L SLp2, Cq
L

G.

Following a suggestion from 1.4, we write pGq for the corresponding set p r of orbits under the subgroup G of AutN pGq. There is then a surjective r mapping G from pGq to pGq, for which the order of the bre of r equals the integer m . If f lies in HpGq and G maps to , we write f G pG q m1 f G pq. p The corresponding set pG1 q for G1 consists of G1 -orbits of L-homomorphisms 1 : L SLp2, Cq
L

G1 .

The composition of any such 1 with the endoscopic L-embedding 1 of L G1 r into L G is an element G in pGq, which in turn maps into pN q. This 1 q to pN q, whose image is contained in the subset r gives a mapping from pG r r pGq of pN q. We write pG1 , q pG1 , G, q r for the bre in pG1 q of our given element P pN q, a set that is empty r unless belongs to pGq. We want to specialize the formula of Theorem 4.1.2 to the linear form p1 Sdisc, pf 1 q in (4.1.2). It is a consequence of the denitions (3.3.9) and (3.3.13) that p1 p1 Sdisc, pf 1 q Sdisc,1 pf 1 q,
1

170

4. THE STANDARD MODEL

r where 1 is summed over those elements in pG1 q that map to . Since r this sum is over the image of pG1 , q in pG1 q, the specialization of the 1 yields the following corollary. theorem to G r r Corollary 4.1.3. Given G P Esim pN q, G1 P Eell pGq, P pN q and r f P HpGq, we have an expansion (4.1.12) where 1 p 1 q G p 1 q G1 ps1 q. Remark. Suppose that the endoscopic group G1 is proper, in the sense that it is distinct from G. In other words, it is a proper product r G 1 G 1 G1 , G1 P Esim pN 1 q, N 1 ` N 1 N,
1 2 i i 1 2
1 1

p1 Sdisc, pf 1 q

1 PpG1 ,q

|S1 |1 pS 0 1 q 1 p 1 q f 1 p 1 q,

of simple endoscopic groups. We note that the informal assumptions under which the corollary makes sense depend only on the positive integer N . They could be replaced by a formal induction hypothesis that the theorems we have stated so far, including Theorem 4.1.2, are valid if N is replaced by a smaller integer N . Corollary 4.1.3 would then hold for G1 . The group S given by (4.1.10) is of course insensitive to whether we r treat as an element in pGq or pGq. With either interpretation of , the group will be at the heart of our analysis of the linear form Idisc, pf q. As varies, the properties of S characterize an important chain of subsets of pGq. We write (4.1.13) where sim pGq t P pGq : |S | 1u, 2 pGq t P pGq : S is niteu, ell pGq t P pGq : S ,s is nite for some s P S ,ss u, and disc pGq t P pGq : ZpS q is niteu. r These sets are obviously stable under the action of the group OutN pGq (of order 1 or 2). The associated sets of orbits give a corresponding chain of r subsets of pGq, which can be regarded as a renement of the earlier chain (1.4.7) from 1.4. r The subsets (4.1.13) of pGq (or rather their analogues for pGq) have a direct bearing on how we interpret the corresponding linear forms Idisc, pf q. r r For example, disc pGq should be the subset of parameters P pGq such sim pGq 2 pGq ell pGq disc pGq pGq,

4.2. ON THE GLOBAL INTERTWINING RELATION

171

r that Idisc, pf q is nonzero. Elements in the smaller set ell pGq are charac1 P E pGq, the subset terized by the property that for some G ell 2 pG1 , q 2 pG1 q X pG1 , q of pG1 q is nonempty. In other words, should contribute to the discrete r r spectrum of G1 . The subset 2 pGq of ell pGq consists of those for which 1 G. This is of course to say that should contribute to the we can take G discrete spectrum of G, or equivalently, that the term with M G in the r expansion (4.1.1) of Idisc, pf q is nonzero. The smallest set sim pGq consists r 2 pGq such that for any G1 P Eell pGq distinct simply of the parameters P from G, the set 2 pG1 , q (or for that matter the set pG1 , q) is empty. These last remarks are intended as motivation for what follows. They oer some guidance for our eorts to compare the two expansions (4.1.1) and (4.1.2). Notice that the properties that dene the subsets in (4.1.13) are given entirely in terms of the group S . They make sense if S is replaced by any complex reductive group, or even a disjoint union S of bitorsors under a connected group S 0 . For example, one could take S to be bitorsor under a complex reductive (but not necessarily connected) group S , a set that can obviously be regarded as a nite disjoint union of bitorsors under the connected group S 0 pS q0 . It is necessary to formulate the denitions in this setting if one wants to replace G by an arbitrary triplet (3.1.2). 4.2. On the global intertwining relation r We continue to work with a xed group G P Esim pN q and a parameter r q, both taken over the global eld F . Our long term aim will be P pN to compare the spectral expansion (4.1.1) of Idisc, pf q with its endoscopic expansion (4.1.2). In the last section, we stated a formula for the linear form p1 Sdisc, pf 1 q, r f P HpGq,

in the endoscopic expansion. In this section, we shall establish an expansion for the analogous term ` r (4.2.1) tr MP, pwq IP, pf q , f P HpGq, in the spectral expansion. We shall also reformulate the local intertwining relation stated in Chapter 2 as a global relation. This will ultimately serve as a link between the two formulas. The trace (4.2.1) depends on xed elements M P L, P P PpM q and w P W pM qreg . The Levi subgroup M is a product of several general linear groups with an orthogonal or symplectic group of the same form as G, as in the local case (2.3.4). We can therefore construct the sets r r 2 pM q pM q of parameters as products of sets of the sort we have already dened. We r r r write M pGq for the image of 2 pM q in pGq under the natural mapping.

172

4. THE STANDARD MODEL

We can also form the set 2 pM q. For later reference, we write 2 pM, q 2 pM, G, q

r for the bre of our given element P pN q in 2 pM q, a set that is empty r r unless belongs to M pGq. We may as well also write 2 pM, q for the r 2 pM q. bre of in r Once again, the set pGq (or for that matter, any of corresponding sets attached to M ) depends on the implicit assumption that Theorem 1.4.1 is valid. The rules here are the same as in the last section. Each object we introduce is predicated on the validity of any cases of the stated theorems required for its existence. As we have noted, we will replace this implicit assumption with formal induction hypotheses at the beginning of 4.3. Our informal assumptions will include Theorem 1.5.2, since among other things, we will need to introduce a global analogue of the expression (2.4.5). In particular, we can assume that the relative discrete spectrum of M der composes into subspaces indexed by parameters in 2 pM q. The trace (4.2.1) comes from the contribution of to a representation induced from the disr crete spectrum of M . It thus vanishes if does not lie in M pGq. We shall r M pGq, and in particular, that it lies in therefore assume that belongs to r r the subset pGq of pN q. Our central algebraic object will continue to be the complex reductive group S . We digress to consider some of its global implications, including our formulation of the global intertwining relation. The group S gives rise to several nite groups, which are global analogues of the nite groups introduced in 2.4. For example, the basic com1 ponent group S 0 pS q has a normal subgroup S . This is isomorphic r to the component group SM attached to any M in the subset 2 pM, q of r 2 pM q that map to . Our assumption that M lies in 2 pM q r elements in means that the quotient ` p x p T ZpM q {ZpGq AM { AM X ZpGq , x x

(4.2.2) where

` x 0 AM ZpM q , x
0 is a maximal torus in S . The associated nite groups W and W of automorphisms can then be identied with the full Weyl groups of S 0 and

4.2. ON THE GLOBAL INTERTWINING RELATION

173

S . They take their places in the global version 1 1

0 0 W W

>

(4.2.3)

>

1 S

>

>

} 1
> 1 S >

>

>

1 1 of the commutative diagram (2.4.3) (or rather the special case of (2.4.3) in r which is represented by an element in 2 pM q). It is interesting to compare the general diagram (4.2.3) with its specialization to parameters in the subsets (4.1.13). The set ell pGq in (4.1.13), for example, is closely related to the R-group R in (4.2.3). One sees easily r that belongs to ell pGq if and only if there is an element w P R whose xed point set in T (relative to the embedding of R into W dened by a choice of Borel subgroup B S ) is nite. This condition in turn implies that S 0 T . Such elements will be an important part of our future study. Consider a point u from the group N N pG, T q{T at the center of the diagram. Following the local notation from 2.4, we shall write wu and xu for the images of x in W and S respectively. We can identify the rst point wu with an element in the group W pM q. This gives a global twisted group r Mu Mwu pM, wu q. We have also been following the later convention from 2.4, in which M r denotes an element in 2 pM q that maps to . This parameter actually lies u q of 2 pM q. We again write u (somewhat superuously) r r in the subset 2 pM for the image of u in the associated SM -bitorsor r SM ,u SM pMu q. This notation will be useful to describe the normalization of the global intertwining operator in (4.2.1). We are assuming that the analogue of Theorem 1.4.2 is valid for M . This implies that the localization M,v of M at any v belongs to the correspondr r ing set ` pMv , v q. The localization v of therefore belongs to ` pGv q.

174

4. THE STANDARD MODEL

It follows easily that S embeds into S v , and that there is a morphism of the diagram (4.2.3) into its local counterpart (2.4.3). In particular, any element u in N has a local image uv in the group Nv pGv , Mv q, for any valuation v. We can therefore dene the local linear form r fv,G pv , uv q, fv P HpGv q, by (2.4.5) (with Mv , Mv and uv taking the place here of the local symbols , and u in (2.4.5)). We obtain a global linear form fG p, uq, by setting (4.2.4) fG p, uq
v

r f P HpGq, f
v

fv,G pv , uv q,

fv ,

for a product that by the discussion of 2.5 can be taken over a nite set. It is clear from 2.4 that the denition has an equivalent global formulation ` p4.2.4q1 fG p, uq xr, M y tr RP pwu , M , M qIP pM , f q , u r r
r M PM

r where RP pwu , M , M q is a normalized global intertwining operator. This operator is dened as a product over v of its local analogues, constructed in Chapter 2 by (2.3.26), (2.4.2) and (2.4.4). In particular, it depends implic r u r itly on an extension M of M to the torsor Mu pAq. The pairing xr, M y represents the corresponding extension of the character x , M y on the group 1 r SM S to the bitorsor SM ,u , which is again dened as a product of its local analogues. Consider also a semisimple element s in the original group S . It has a local image sv in the group S v , for any valuation v. We can therefore dene the local linear form r f 1 pv , sv q, fv P HpGv q,
v,G

by (2.4.6) (with v and sv in place of and s). We obtain a global linear form 1 r fG p, sq, f P HpGq, by setting 1 1 (4.2.5) fG p, sq fv,G pv , sv q, f fv ,
v v

for a product that can again be taken over a nite set. The global formulation of this denition is of course p4.2.5q1
1 fG p, sq f 1 p 1 q,

where pG1 , 1 q is the preimage of p, sq under the global correspondence (1.4.11). The local intertwining relation stated as Theorem 2.4.1 applies to the local factors in (4.2.4) and (4.2.5). We can state the corresponding global identity for the products as a corollary of this theorem.

4.2. ON THE GLOBAL INTERTWINING RELATION

175

Corollary 4.2.1 (Global intertwining relation for G). For any u in the global normalizer N , the identity (4.2.6)
1 fG p, s sq fG p, uq,

r f P HpGq,

holds for any semisimple element s P S that projects onto the image xu of u in S . We remind ourselves that Theorem 2.4.1 has yet to be established. Its proof in fact will be one of our major concerns. The same therefore goes for the corollary. We have formulated it here to give our discussion some sense of direction, and in particular, to see how we will eventually relate the summands of (4.1.1) and (4.1.2). We note that the diagram (4.2.3) has an obvious analogue if S is replaced by any complex reductive group S. More generally, it makes sense if we take S to be a bitorsor under a complex reductive group S . (We write S here because we need to reserve the symbol S 0 pS q0 for the connected component of 1 in S .) The corresponding objects N, W , S and R in the diagram are then bitorsors under their analogues N , W , S and R for the group S . On the other hand, the objects W 0 and S 1 are the equal to the associated groups for S . In particular, S 1 is the projection onto the quotient S 0 pS q of a subgroup S 1 of S with S0 S1 S. For example, we can take r r r S S S pGq{ZpGq , r where G is a twisted orthogonal group over F , and belongs to the subset r of pGq. It would be easy to formulate a global intertwining relation r pGq r for G, as a corollary of Theorem 2.4.4 that is parallel to Corollary 4.2.1. We will not do so, since we do not need it. However, we will formulate a global r supplement of Theorem 1.5.2 for G. We shall state it here for application to the xed group M above, even though its proof will only come later. r r r r We are supposing that G pG0 , q, where G G0 is an even orthogonal r r group in Esim pN q over the global eld F . Consider a parameter P 2 pGq. ` 2 r For any function P Ldisc, GpF qzGpAq , and any y P GpAq, the function (4.2.7) r G r Rdisc, pyq pxq p1 xyq, x P GpAq, ` also belongs to L2 disc, GpF qzGpAq . We thus obtain a canonical extension G r r of the representation Rdisc, of GpAq to the group GpAq` generated by GpAq. Theorem 1.5.2, together with Theorem 2.2.1, describes the character of the original representation in terms of the transfer of characters from GLpN q. The theorem we are about to state, in combination with Theorem 2.2.4, plays a similar role in it extension. `

176

4. THE STANDARD MODEL

Suppose that lies in the subset r r r 2 pGq pGq X 2 pGq r of 2 pGq. Then the set r r S 0 pS q is a (nonempty) S -torsor. Any localization v of lies in the subset r r r r r r ` pGv q of ` pGv q, and determines a mapping x xv from S to Sv . r Suppose that is a representation in the global packet , equipped with ` (or equivalently, an intertwining operator r an extension to the group GpAq r of of order 2). We can then choose extensions of the local components pq r v of to the groups GpFv q` whose tensor product is compatible with the extension of . Theorem 2.2.4 tells us that there is a corresponding product x r x r x P S , r xr, y xrv , v y,
v

of extensions of linear characters x , v y, which determines an extension of r` r the linear character x , y on S to the group S generated by S . Assume that x , y equals the sign character On the one hand, has a canonical r` extension to S . This follows from the denition (1.5.6) and the fact that p the adjoint action of the complex group G SOpN, Cq on its Lie algebra has a canonical extension to the group OpN, Cq. On the other, Theorem G 1.5.2 asserts that occurs in the decomposition of Rdisc, . Moreover, we G r have just seen that Rdisc, has a canonical extension to GpAq` . The theorem stated below asserts that the restriction of this extension to the subspace of coincides with the extension of under the local correspondence of extensions postulated by Theorem 2.2.4. r r Theorem 4.2.2. Suppose that G, with G P Esim pN q, is a twisted orthogr onal group over F , and that belongs to 2 pGq. r r (a) If lies in the complement of 2 pGq in 2 pGq, we have ` G r r r r tr Rdisc, pf q 0, f P HpGq. r (b) Assume that lies in 2 pGq, and that is a representation in r such that the linear character x , y on S is equal to the global packet G 1 . Then the restriction to of the canonical extension of Rdisc, to r r GpAq` corresponds to the canonical extension of to S ` under the product

of local correspondences of extensions given by Theorem 2.2.4. We will now take up the expansion of the global trace (4.2.1). As an aid to the reader, we shall try to formulate it so that it is roughly parallel to its endoscopic counterpart, Corollary 4.1.3. In particular, we shall state it as a corollary of Theorem 4.2.2 and Theorem 1.5.2, just as its endoscopic analogue was stated as a corollary of Theorem 4.1.2. We recall that Theorem 4.1.2 was the stable multiplicity formula, while Theorem 1.5.2 is the ordinary

4.2. ON THE GLOBAL INTERTWINING RELATION

177

multiplicity formula (with its twisted supplement Theorem 4.2.2). We will want to apply the two corollaries in the next two sections, long before any of these three theorems have been proved. We will then have to apply them under a common induction hypothesis. Our proof of the expansion of (4.2.1) in the remainder of this section will be more complicated than that of its endoscopic counterpart. In addition to the two theorems on which it depends as a corollary, the proof also requires r an analogue of Theorem 4.2.2 for the original GLpN q-torsor GpN q. However, this is elementary, unlike Theorem 4.2.2, in the sense that we will be able to establish it on the spot. ` r r Consider a parameter in the global set sim pN q sim GpN q . The analogue ` ` N r r y P GpN, Aq, x P GLpN, Aq, (4.2.8) Rdisc, pyq pxq pN q1 xy , ` of (4.2.7), dened for P L2 disc, GLpN, F qzGLpN, Aq , is a canonical exN r tension of the representation Rdisc, of GLpN, Aq to the group GpN, Aq` . It corresponds to a canonical extension of the automorphic representation of GLpN, Aq attached to . On the other hand, for any completion v of , we also have the extension of the representation v of GLpN, Fv q described in 2.1. r Lemma 4.2.3. Assume that P sim pN q. Then the extension of the automorphic representation of GLpN, Aq dened by (4.2.8) equals the r tensor product of local extensions v pN q of representations v dened in terms of Whittaker functionals in 2.1.
N Proof. The denition (4.2.8) makes sense if Rdisc, is replaced by the N of GpN, Aq on any invariant space of funcright regular representation R tions on the quotient of GLpN, Aq by GLpN, F q. In particular, the operator ` r r RpN q RN pN q

r intertwines RN with RN pN q. It suces to show that the intertwining N r operator Rdisc, pN q of Rdisc, corresponds to the intertwining operator r v pN q r pN q
v

of
v

v .

We write b and N mn, where is a unitary cuspidal automorphic representation of GLpmq, and n is the irreducible representation of SLp2, Cq of dimension n. Let IP p, q be the induced representation of GLpN, Aq dened as in the statement of Theorem 1.3.3, but with a general vector P Cn in place

178

4. THE STANDARD MODEL

of the xed vector P

n1 n2 2 , 2 ,

pn1q 2

of exponents. It acts on a Hilbert space HP p q that is independent of . Given that is self-dual, we see that there are isomorphisms IP p, q IP p_ ,_ q IP p,_ q_ , in which _ pn , . . . , 1 q. It follows that there is a canonical intertwining isomorphism r r IP p, , N q : IP p, q IP p,_ q_ IP p,_ q pN q that is compatible with the corresponding two standard global Whittaker functionals (namely, the global analogues of (2.5.1) for and _ ) on the space HP p q. Suppose now that is in general position. Then for any 0 in the subspace HP p q of -nite vectors in HP p q, the Eisenstein series Ep, q : x Epx, , q is a well dened automorphic form on GLpN, Aq. Let RN p, q be the rer striction of RN , regarded as a representation of the Hecke algebra HpN q, to the space 0 Ap, q tEp, q : P HP p qu The mapping Epq : Ep, q is then an intertwining isomorphism between the irreducible representations r IP p, q and RN p, q of HpN q. Our third ingredient is the intertwinr ing mapping RpN q above. It takes Ap, q onto the space Ap_ , _ q _ q, as one sees readily from an examination of the relevant constant Ap, terms, and therefore represents an intertwining isomorphism r r RpN q : RN p, q RN p, _ q_ RN p, _ q pN q. The main point is to show that ` r r (4.2.9) RpN qEp, q E IP p, , N q, _ . The putative identity (4.2.9) asserts that the diagram
IP p, ,N q r IP p, q IP p,_ q pN q Epq Ep_ q r RpN q r RN p, q RN p, _ q pN q , is commutative. Since the four arrows each represent intertwining isomorphisms between irreducible representations, we can write r r Ep_ qIP p, , N q c RpN qEpq, r

for a multiplicative complex number c P C . We must prove that c 1.

4.2. ON THE GLOBAL INTERTWINING RELATION

179

The group GLpN q has a standard global Whittaker datum pB, q. It is dened as in the local discussion of 2.5, except that is a left NB pF qinvariant, nondegenerate character on NB pAq. The Eisenstein series in (4.2.9) has a canonical automorphic -Whittaker functional pEqp, q Epn, , q pnq dn.
NB pF qzNB pAq

A change of variables in this integral, combined with the denitions of r and pN q in terms of the standard splitting of GLpN q, tells us that ` ` r RpN qEp, q Ep, q . Let us write p, q for the -Whittaker functional for IP p, q with respect r to which we have normalized the isomorphism IP p, , N q. It is induced from the automorphic MP -Whittaker functional on the space of cuspidal functions on MP pAq, and is dened by the global analogues of the integral (2.5.1). We then have ` 0 r IP p, , N q, _ p, q, P HP p q, by denition. Moreover, one can show from the basic denition of Ep, q as an absolutely convergent series on an open set of that ` 0 Epq p, q, P HP p q. (See [Sha1, p. 351].) Combining these identities, we see that ` ` r r Ep_ qIP p, , N q IP p, , N q, _ ` p, q Epq ` r RpN qEpq . It follows that c 1, as required, and therefore that the identity (4.2.9) is valid. The embedding of the global Langlands quotient into the automorphic discrete spectrum is provided by residues of Eisenstein series. More precisely, it is the sum of iterated residues of Epx, , q at the point P _ P given by the general scheme in [L5, Chapter 7], and its specialization to GLpN q in [MW2], that yields the intertwining isomorphism from to N Rdisc, . We recall that the residue scheme is noncanonical. In particular, it could be replaced by its adjoint, in which the iterated residues are taken r r in _ rather than . We are trying to relate the operators RpN q and pN q. But pN q is the Langlands quotient of the value at P of the operator r r IP p, , N q. The lemma then follows from an application of the residue operation to the left hand side of (4.2.9), and its adjoint to the right hand side. We now consider the global trace (4.2.1). We shall expand it into an expression that is roughly parallel to the formula (4.1.12) of Corollary 4.1.3, and which, as we noted above, is a corollary of Theorems 4.2.2 and 1.5.2.

180

4. THE STANDARD MODEL

r Any parameter M P 2 pM q determines a subspace ` ` 2 0 L2 disc,M M pF qzM pAq Ldisc AM pRq M pF qzM pAq of the discrete spectrum. This subspace in turn has an M pAq-equivariant decomposition ` L2 mM pM qM , disc,M M pF qzM pAq
M

for irreducible representations M of M pAq such that cpM q maps to cpM q, and nonnegative multiplicities mM pM q. The induced representation in (4.2.1) therefore has a decomposition mM pM q IP pM , f q, IP, pf q
M M

r r where the sum over M can be restricted to the subset 2 pM, q of 2 pM q. Let us write MP pw, M q for the transfer of the intertwining operator MP, pwq in (4.2.1) to the space on which IP pM , f q acts. We can then express the global trace (4.2.1) as a double sum ` mM pM q tr MP pw, M q IP pM , f q ,
M M

with the understanding that the trace vanishes if wM is not equivalent to M . The original intertwining operator MP, pwq in (4.2.1) acts on the Hilbert space HP, of the induced representation IP, . It is dened by analytic continuation of Langlands original integral over Nw1 P pAq X NP pAqzNw1 P pAq. Since the denition also includes a left translation by w1 , the transfer r MP pw, M q JP pw, M q of MP, pwq to the Hilbert space of IP pM , f q is the global analogue of the r unnormalized operator JP pw, q (with M in place of ) in (2.3.26). The global form of the entire expression (2.3.26) is the product RP pw, M , M q rP pw, M q1 MP pw, M q, where rP pw, M q is the product of the global -factor pwq with the global normalizing factor 1 (4.2.10) Lp0, M , _1 P |P q p0, M , _1 P |P q1 Lp1, M , _1 P |P w w w dened by analytic continuation of a product of local normalizing factors (2.3.27). Since the global -factor equals 1, the global normalizing factor is in fact just equal to (4.2.10). In the discussion of Chapter 2, the next step was to transform the local object (2.3.26), which occurs on the right hand side of (2.4.2), into the selfintertwining operator on the left hand side of (2.4.2). In the global setting

4.2. ON THE GLOBAL INTERTWINING RELATION

181

here, RP pw, M , M q is already a self-intertwining operator, under the condition w P W pM q of (2.4.2). In fact, a global version of the intertwining operator pwq of (2.4.2) is built into the denition of RP pw, M , M q. It r comes from operators (4.2.7) and (4.2.8), applied to the relevant factors of r M , and taken at values y w and x 1 at which they become trivial. For the general linear factors of M , we must use Lemma 4.2.3 to verify that the analogue of (4.2.8) equals the tensor product of the local Whittaker extensions that are implicit in the global self-intertwining operator in (4.2.4)1 . We can therefore write MP pw, M q rP pw, M q RP pw, w , M q, r w P W pM q, in the global form of the notation (2.4.2). It remains to apply Theorems 1.5.2 and 4.2.2 to the other factor G of M . We have expressed the global trace (4.2.1) as a double sum over M and M . We rst apply Theorem 4.2.2(a) to the G -component of the global multiplicity mM pq that occurs in any summand. We see that the summand vanishes unless M belongs to the subset r 2 pMw , q pMw q X 2 pM, q r of w-xed elements in 2 pM, q. We can also restrict the second sum to r elements M in the packet M . It follows that (4.2.1) equals ` r mM pM q rP pw, M q tr RP pw, M , M q IP pM , f q .
r M P2 pMw ,q M PM

Next, we apply Theorem 1.5.2 to G . This gives us the multiplicity formula mM pM q mM |SM |1 M pxM q xxM , M y,
xM PSM

by Fourier inversion on the abelian group SM . The last step will be to apply Theorem 4.2.2(b) to G . The condition that M lies in 2 pMw , q means that w lies in the subgroup W of W pM q. (Like the other groups in the diagram (4.2.3), W is in fact dened in terms of an implicit representative M of .) Let us write r N pwq SM ,u , wu w, for the bre of w in N , relative to the horizontal exact sequence at the center of (4.2.3). We can then write u u r M pxM q xxM , M y M prqxr, M y,
xM PSM uPN pwq

r` where the summand on the right is an extension to the group SM ,u of the character on SM dened by the summand on the left. The extension must be trivial if the character on the left is trivial, but can otherwise be arbitrary, since both sums then automatically vanish. We take M prq to u be the canonical extension of M pxM q. If the linear character xxM , M y

182

4. THE STANDARD MODEL

on SM equals 1 , the pairing xr, M y will then have to be the canonical u r M 1 extension of M . Applying Theorem 4.2.2(b) to M in this case, we conclude that RP pw, M , M q is dened in terms of an implicit extension of M to r Mw pAq` that is compatible with the product of local extensions associated by Theorem 2.2.4 to xr, M y. In case xxM , M y is distinct from 1 , which u r M is to say that M does not occur in the discrete spectrum, we have only to r dene RP pw, M , M q by any product of local intertwining operators (2.4.2) that is compatible with the extension of xr, M y we have chosen. u r Having agreed upon these conventions, we take the sum over u P N pwq r outside the sum over M P M . It then follows from the denition (4.2.4) that the resulting sum ` r xr, M y tr RP pw, M , M q IP pM , f q u r
r M PM

equals fG p, uq. Combining this with the other terms discussed above, we obtain the expansion of (4.2.1) that represents our common corollary of Theorems 4.2.2 and 1.5.2. r r Corollary 4.2.4. Given G P Esim pN q, M P L, P M pGq and f P r HpGq, we have an expansion (4.2.11) |SM |1 rP pw, M q M prq fG p, uq u
M P2 pMw ,q uPN pwq

for the global trace ` tr MP, pwq IP, pf q , w P W pM q.

Remark. Suppose that the Levi subgroup M is proper in G. In other words, it is a product of general linear groups with a simple endoscopic group r G P Esim pN q that is a proper subgroup of G. As in Corollary 4.1.3, the informal assumptions under which this corollary makes sense depend only on the positive integer N . They could be replaced by a formal induction hypothesis that the theorems stated so far, including Theorems 1.5.2 and 4.2.2, are valid if N is replaced by a smaller integer N . Corollary 4.2.4 would then hold for M . The global intertwining relation (4.2.6) and the formula (4.2.11) for the global trace (4.2.1) are obviously designed to be used together. They have important implications for the two expansions (4.1.1) and (4.1.2), which we shall analyse over the next few sections. However, we should rst describe twisted versions of the formulas for GLpN q, since they will also be needed. r Suppose then that G equals the component GpN q, rather than an eler ment in Esim pN q. Then G represents the triplet ` ` r r r GpN q0 , pN q, 1 GLpN q, pN q, 1 ,

4.2. ON THE GLOBAL INTERTWINING RELATION

183

as in the general discussion of Chapter 3. We take to be a parameter in the set ` r r r pGq pN q GpN q , and f to be a function in the space ` r r r HpGq HpN q H GpN q . We then have the spectral expansion (4.1.1) (which has been established in general), and the endoscopic expansion (4.1.2) (provided by Hypothesis 3.2.1). One of our aims has been to arrange matters so that the twisted case r G GpN q does not require separate notation. It is for this reason that we write ` p S S pGq Cent Impq, G and p p S S {ZpG0 q S {ZpG0 q, p p p p where G is the G0 -torsor G0 . These two sets are bitorsors under the corresponding complex reductive groups p S S pG0 q CentpImpq, G0

and
p p S S {ZpG0 q S {ZpG0 q. Similarly, the quotient S 0 pS q is a bitorsor under the nite group S 0 pS q. Observe that S is a product of general linear groups, and that p p ZpG0 q ZpG0 q equals the group C , embedded diagonally in S . In par and S are both connected. Therefore S equals S 0 , ticular, the groups S so the sets S and S are both trivial in this case. Despite the fact that S is trivial, the constructions of 4.1 are still interesting when applied to the bitorsor S S . Proposition 4.1.1 gives a noteworthy relationship among the associated numbers ipSq pS q and 0 pSs q pS 0 q. The properties formulated in terms of S at the end ,s of 4.1 still make sense, and provide a chain of subsets (4.1.13) of the set r pGq pGq. In this case, the subsets sim pGq and 2 pGq are equal. The formula of Corollary 4.1.3 also makes sense as stated, that is, with G1 r representing any datum in the set Eell pGq Eell pN q. It is an immediate consequence of Theorem 4.1.2 (with G being as originally understood, an r element in Esim pN q.) The earlier discussion of this section also extends to the twisted case. The diagram (4.2.3) becomes simpler, since the sets in the lower horizontal exact sequence are now all trivial. It follows that the sets N and W

184

4. THE STANDARD MODEL

0 are equal, and are both bi-torsors under the nite group W . The families 2 pM, q, pM, q, pG1 , q, etc., all have obvious meaning. It is understood here that M is as in the twisted form of (4.1.1), a Levi subgroup of G0 GLpN q, which we assume is proper. Furthermore, the notation (4.2.4) and (4.2.5) continues to make sense. In the case of (4.2.4), we note r the packet M M attached to any M P 2 pM, q contains one representation M , and that the corresponding global factor xr, M y is trivial. u r The other denition (4.2.5) is formulated in terms of the correspondence

pG1 , 1 q p, sq,

G1 P EpGq, s P S ,ss ,

which is given by the obvious construction. The statements of Corollaries 4.2.1 and 4.2.4 also extend without change in notation. In the formula (4.2.11) of Corollary 4.2.4, the factors |SM | and M pxM q are both equal to 1, while rP pw, M q represents a global normalizing factor for G0 GLpN q. The derivation of this formula is similar. (The formula cannot be described as a corollary in this case, since Theorem 4.2.2 no longer has a role.) The proof of the twisted global intertwining relation (4.2.6), or rather its reduction to the local relation Corollary 2.5.4, also remains the same. We recall that Corollary 2.5.4 depends on the denitions provided by Theorem 2.2.1 (which of course still has to be proved). The discussion of this section thus holds uniformly if G is either an r r element in Esim pN q or if G equals GpN q itself. We shall apply it in this generality over the next few sections. 4.3. Spectral terms We are now at the stage where we can begin the proof of our results in earnest. There are essentially nine theorems. They are the original three theorems of classication, Theorems 1.5.1, 1.5.2 and 1.5.3 (which for their statements implicitly include Theorems 1.4.1 and 1.4.2), their local renements Theorems 2.2.1, 2.2.4, 2.4.1 and 2.4.4 from Chapter 2, and the global renements Theorems 4.1.2 and 4.2.2. We shall prove all of the theorems together. The argument will be a multilayered induction, which will take up much of the rest of the volume. We need to be clear about our induction assumptions. For a start, they are to be distinguished from our structural condition (Hypothesis 3.2.1) on the stabilization of the twisted trace formula. Our primary induction hypothesis has already been applied informally (and often implicitly) a number of times. It concerns the rank N . Recall that the nine theorems all apply to a group G in the set ` r r Esim pN q Esim GpN q , where r r GpN q GLpN q pN q, for some positive integer N . We x N . We then assume inductively that r the theorems all hold if G is replaced by any group G P Esim pN q (and

4.3. SPECTRAL TERMS

185

r r hence also for any G in one of the larger sets Eell pN q or EpN q), for any positive integer N with N N . We will later have occasion to take on more delicate induction hypotheses. We assume for the rest of this chapter that F is global. For any G r r G P Esim pN q and P pN q, we have been writing Rdisc, for the reprer sentation of the symmetric global Hecke algebra HpGq on the -component ` 2 Ldisc, GpF qzGpAq of the discrete spectrum. It follows from Corollary 3.4.3 (one of the results we have established, rather than just stated) that the r representation of HpGq on the full discrete spectrum is a direct sum over G r P pN q of the subrepresentations Rdisc, . Theorem 1.5.2 (one of our G stated, but yet unproven theorems) describes the decomposition of Rdisc, into irreducible representations, at least up to an understanding of the local G r constituents of the packet . It asserts that Rdisc, is zero if does not r r r belong to the subset 2 pGq of pN q, and that if does lie in 2 pGq, then G m pq , Rdisc,
r P

where # m pq m , 0 if x , y 1 , otherwise.

This is of course one of the assertions we can apply inductively (having already done so implicitly in the last section). Its validity for G, however, will be among the most dicult things we have to prove. r r Suppose for example that belongs to the subset pGq of pN q, but r lies in the complement of 2 pGq. Then factors through a parameter M P 2 pM q, for a proper Levi subgroup M of G. By our induction assumption, applied to the factors of M as a product of groups, the image r of M in 2 pM, q contributes to the automorphic discrete spectrum of M (taken modulo the center). The theory of Eisenstein series then tells us that contributes to the continuous spectrum of G. Theorem 1.5.2 includes the assertion that does not contribute to the discrete spectrum of G. In other words, the automorphic spectrum of G has no embedded eigenvalues, in the sense of unramied Hecke eigenvalues in Caut pGq. This question is obviously an important part of the theorem. It will be a focal point of the long term study we are about to undertake. For the next few sections, we shall be working from the slightly broader perspective reached at the end of the last section. That is, we assume that G r r is either an element in Esim pN q or equal to GpN q itself. The basic expansions (4.1.1) and (4.1.2), which are our starting points, were originally formulated in this generality. The rest of the discussion of 4.1 and 4.2 (apart from Theorem 4.1.2, which is expressly limited to the case that G is a connected quasisplit group) also applies.

186

4. THE STANDARD MODEL

We note that the operator

G Rdisc, pf q,

r f P HpGq,

r makes sense if G GpN q. It is dened by the extension (4.2.8). Equivalently, it is simply the composition of the two factors in (4.2.1), under the condition that M P G0 . In particular, its trace gives the corresponding contribution to the term with M G0 in (4.1.1). The other contribution to this term comes from the sole element w in Wreg pM q. We observe that | detpw 1qaG |1 | detp 1qaG |1 |G |1 ,
M G0

by (3.2.5), recalling at the same time that |G |1 |0 pG q|1 1 , 2

G r in the case G GpN q at hand. The operator Rdisc, pf q is straightforward in this case, since we know from Theorem 1.3.2 that GLpN q has no embedded r eigenvalues. In particular, the operator equals 0 unless belongs to sim pGq. We include it in the discussion in order to have uniform statements of the results. r r We x P pN q and f P HpGq. We are looking at the expansion (4.1.1), r r with G now allowed to be either an element in Esim pN q or GpN q itself. The 0 equals the product term with M G ` G G rdisc, pf q |G |1 tr Rdisc, pf q .

It follows that the dierence (4.3.1)

G Idisc, pf q rdisc, pf q

equals the sum of those terms in (4.1.1) with M G. The summand of M G depends only on the W0 -orbit of M . We can therefore sum over the set G -orbits in L distinct from G, provided that we multiply by the tM u of W0 number G M |W0 ||W0 |1 |W pM q|1 of elements in a given orbit. The dierence (4.3.1) therefore equals ` |W pM q|1 | detpw 1qaG |1 tr MP, pwq IP, pf q .
M

tM utGu

wPWreg pM q

We can apply Corollary 4.2.4, which is founded on our underlying induction hypothesis, to the factor ` tr MP, pwq IP, pf q . The dierence (4.3.1) becomes a fourfold sum, over tM u tGu, w P Wreg pM q, M P 2 pMw , q and u P N pwq, of the product of |W pM q|1 | detpw 1qaG |1
M

with |SM |1 rP pw, M q M prq fG p, uq. u

4.3. SPECTRAL TERMS

187

Despite its notation, the second factor is independent of the representative M of . We will therefore be able to remove the sum over the set of M , so long as we multiply the summand by its order. The result will be clearer if we rst rearrange the double sum over w and M . We start by writing the sum over M P 2 pMw , q as a double sum w , G q, the subset of parameters in over G in pG, q and M in 2 pM p 2 pMw , q that map to the G-orbit G . We then interchange the sums over w and G . This allows us to combine the resulting sums over w and M into a double sum over the set ( V pM , wq P 2 pM, G q Wreg pM q : Intpwq M M . Now, the projection pM , wq w gives a canonical bration V tW,reg u of V over the set of orbits of W by conjugation on W,reg . We claim p that the group W pM q acts transitively on the bres. Since G is a G-orbit, 0 , for any and since AM is a maximal torus in SM M P 2 pM, q, any two x p parameters in 2 pM, G q are conjugate by an element in G that stabilizes AM . Therefore W pM q acts transitively on 2 pM, G q. The claim follows x from the fact that the stabilizer in W pM q of any parameter in 2 pM, G q is isomorphic to W . The summand is constant on the bres. We can therefore replace the double sum over V with a constant multiple of a simple sum over tW,reg u. The scaling constant is slightly simpler if we take the simple sum over the set W,reg , rather than its quotient tW,reg u. With this change, we have then only to multiply the summand by the integer |2 pM, G q| |W pM q||W |1 . For any G , there is at most one orbit tM u such that the set 2 pM, G q is not empty. Dierent elements G P pG, q can give dierent orbits tM u, but the corresponding summands remain equal. We can therefore remove the exterior sum over G , provided that we multiply the summand by the second integer |pG, q| m . We conclude that the dierence (4.3.1) equals (4.3.2) m |W |1 | detpw 1qaG |1 |SM |1 rP pw, M q M prq fG p, uq, u
w
M

where w and u are summed over W,reg and N pwq respectively, and pM, M q is any pair with M G0 and M P 2 pMw , q. Of course, if there is no such pM, M q, the expression is understood to be 0. This is the case if r r r lies in either 2 pGq or in the complement of the subset pGq of pN q. r The case that G GpN q, which is now part of our analysis, is generally the simpler of the two. However, it does have one minor complication. This is because aG , the p`1q-eigenspace of the operator on aG0 , is a proper

188

4. THE STANDARD MODEL

subspace of aG0 . (In fact, it is clear that dim aG0 1 and dim aG 0.) We obtain a decomposition | detpw 1qaG | | detpw 1qaG0 | | detpw 1qaG |,
M M G0

for any w P W . Since the restriction of w to aG0 equals , we use (3.2.5) to write (4.3.3) | detpw 1qaG | | detp 1qaG | |G |
G0 G0

for the second factor on the right. The rst factor may be written as | detpw 1qaG0 | | detpw 1q|,
M

where the operator pw1q on the right is to be regarded as an endomorphism 0 of the real vector space aT , as in 4.1. Indeed, aG is isomorphic to the dual M of aT , since as a maximal torus in the complex connected group S 0 S , p 0 q . We substitute the resulting the group T is isomorphic to AM {ZpG x product into our expression (4.3.2) for the dierence (4.3.1). We also add a couple of housekeeping changes to the notation in (4.3.2). 1 The group SM is isomorphic to the group S in the diagram (4.2.3). It then follows from the relevant two short exact sequences in the diagram that
0 1 |W |1 |SM |1 |W |1 |S |1 |N |1 |S |1 |W |1 .

If u is any element in the coset N pwq, we set u 1 puq M prq and

G r pwq rP pw, M q,

since these functions depend only on . With these changes, we have now left ourselves with no reference to M in our expression for (4.3.1). This is not a problem, since the orbit of M is still implicit in the torsor S . However, we will need a marker to rule out the case that M G0 . We take care of this by setting 1 W,reg tw P W,reg : w 1u, since it is clear that #
1 W,reg

H,

if S is nite,

W,reg , otherwise.

Collecting the various terms, we see that the dierence (4.3.1) equals 0 G (4.3.4) C |W |1 | detpw 1q|1 r pwq 1 puq fG p, uq,
1 wPW,reg uPN pwq

where (4.3.5) C m |G |1 |S |1 .

4.3. SPECTRAL TERMS

189

r r Keep in mind that if does not belong to the subset pGq of , then m and C both vanish, and so therefore does the expression. The expression consequently makes sense, even though S is not dened in this case. There is a critical lemma, which gives a dierent way to express the G product of the two signs r pwq and 1 puq. To state it, we need to recall the original sign function s0 pwq s0 pwq on W that came up in (4.1.5). If only for reasons of symmetry, we shall write s0 puq s0 pwu q, for its pullback to N . Then s0 puq equals p1q raised to a power given by the number of roots of pB , T q mapped by wu to negative roots. This object depends on only through the group W . It it not to be confused with the deeper sign character pxq G pxq on S , dened by (1.5.6) in terms of symplectic root numbers attached to , or its twisted analogue 1 puq for M . It is of course also unrelated to the xed point s in S . Lemma 4.3.1. The sign characters satisfy the relation
G r pwu q 1 puq G pxu q s0 puq,

u P N .

We shall give the proof of this lemma in 4.6. Assuming it for now, we make the appropriate substitution in the last expression (4.3.4) for the dierence (4.3.1). The next step is to rearrange the double sum in (4.3.4) by using the vertical exact sequence of N in the diagram (4.2.3) in place of the horizontal exact sequence. Accordingly, we write N pxq, x P S ,

for the bre of x under the mapping from N to S . We also write W pxq for the bijective image of N pxq in W . We then let N,reg pxq and N1 ,reg pxq denote the preimages in N pxq of the respective subsets W,reg pxq W pxq X W,reg and
1 1 W,reg pxq W pxq X W,reg

of W pxq. With this notation, we obtain the following result. Lemma 4.3.2. The dierence
G Idisc, pf q rdisc, pf q

equals (4.3.6) C
xPS 0 G pxq |W |1

uPN1 ,reg pxq

s0 puq | detpwu 1q|1 fG p, uq,

r r for any P pN q and f P HpGq.

190

4. THE STANDARD MODEL

The existence of the pairing xxM , M y for M G0 , which is implicit in (4.2.4), is part of our basic induction hypothesis. Suppose that we happen to know also that it is dened when M G0 . Since is assumed to belong r r to M pGq in (4.2.4), this is only relevant to the case that P 2 pGq. It is essentially a local hypothesis, namely that for any G P 2 pG, q, the localization G,v belongs to ` pGv , v q, and the local pairing xxG,v , G,v y is dened. In particular, it is weaker than asking that the global multiplicity formula of Theorem 1.5.2 hold. However, it does allow us to form the dierence 0 G G xx, y fG pq, G pxq rdisc, pf q rdisc, pf q |G |1 m |S |1
xPS r P G between the (normalized) trace rdisc, pf q and its expected value. It follows easily from the denitions (4.2.4) and (4.3.5) that G 0 G G pxq fG p, xq, (4.3.7) rdisc, pf q rdisc, pf q C xPS

r in case P 2 pGq. In particular, the existence of the pairing xxM , M y for r M G0 amounts to the existence of the function fG p, xq for P 2 pGq. In this case, the second term in the dierence (4.3.7) is equal to the expresssion given by Lemma 4.3.2, but with the index of summation N,reg pxq in place of N1 ,reg pxq. This is of course because N,reg pxq equals the point txu when r belongs to 2 pGq. (The original indexing set N1 pxq is empty in this case.)
,reg

r If does not lie in 2 pGq, on the other hand, N1 ,reg pxq equals N,reg pxq. In this case, we simply set (4.3.8)
0 G rdisc, pf q G rdisc, pf q,

G since the expected value of the trace rdisc, pf q is then equal to 0. It follows G that the formula of Lemma 4.3.2 remains valid with 0 r,disc pf q in place of G r,disc pf q, and N,reg pxq in place of N1 ,reg pxq. Suppose that in addition to being well dened, the function fG p, uq depends only on the image x of u in S . This is implied by the global intertwining identity, but is obviously weaker. The function can then be taken outside the sum over u in N,reg pxq. Lemma 4.3.2, or rather its minor extension above, can then be formulated as follows.

Corollary 4.3.3. Suppose that the function fG p, xq fG p, uq, x P S , u P N pxq,

is dened , and depends only on x. Then the dierence

G Idisc, pf q 0 rdisc, pf q

4.4. ENDOSCOPIC TERMS

191

equals (4.3.9) where

0 i pxq |W |1

xPS

i pxq G pxq fG p, xq,

wPW,reg pxq

s0 pwq| detpw 1q|1 .

There is perhaps more notation in this section than is strictly necessary. As in other sections, it has been designed to suggest how various results t together, and thus oer some guidance to the course of the overall argument. In the case here, it is meant to emphasize the intrinsic symmetry between formulas in this section and those of the next. 4.4. Endoscopic terms This section is parallel to the last one. It is aimed at the terms in the endoscopic expansion (4.1.2) of Idisc, pf q. We shall derive a ner expansion of (4.1.2) that can be compared with the one we have established for (4.1.1). r We are working in the setting adopted in the last section. Then GpN q r r is xed, and G is allowed to denote either an element in Esim pN q or GpN q r q is also xed, as is the function of f P HpGq. r itself. The parameter P pN The terms in (4.1.2) are parametrized by elliptic endoscopic data 1 P E pGq. The subset E G sim pGq of Eell pGq is of particular interest, because ell its terms are not amenable to induction. In recognition of this, we set p1 sG pf q pG, G1 q Sdisc, pf 1 q. disc,
G1 PEsim pGq

If G1 belongs to the complement of Esim pGq in Eell pGq, it is represented by a product of groups to which we can apply our induction hypothesis. In particular, we can apply the stable multiplicity identity of Corollary 4.1.3 1 to the corresponding term Sdisc, pf 1 q in (4.1.2). We can of course also apply the specialization at the end of 3.2 of the general formula (3.2.4) to the coecient pG, G1 q in (4.1.2). We see that the dierence (4.4.1) equals (4.4.2)
G1

Idisc, pf q sG pf q disc, pG, G1 q

1

p1 Sdisc,1 pf 1 q,

where G1 is summed over the complement of Esim pGq in Eell pGq and 1 is summed over pG1 , q, while (4.4.3) and (4.4.4) p pG, G1 q |G |1 |ZpG1 q |1 |OutG pG1 q|1 . p1 Sdisc,1 pf 1 q |S1 |1 pS 0 1 q 1 p 1 q f 1 p 1 q

192

4. THE STANDARD MODEL

p We recall that pG1 , q is the set of G1 -orbits of L-embeddings of L G0 that map to , and that L SLp2, Cq to 1 p 1 q G1 ps1 q. r r r If does not lie in pGq, a subset of pN q that is proper if G P Esim pN q r q if G GpN q, each set pG1 , q is empty. The dierence r and is equal to pN (4.4.1) thus vanishes in this case. We can therefore assume that does lie r in pGq. The computations of the present section are again founded on the centralizer p S S {ZpG0 q . p This object is well dened (as a G0 -orbit) under our assumption that r P pGq. It is represented by the original complex reductive group (4.1.10) r if G P Esim pN q, and a bitorsor under the complex connected reductive group p S S pG0 q{ZpG0 q , r in case G GpN q. Any semisimple element s P S ,ss in S gives rise to a pair pG1 , 1 q, G1 P Eell pGq, 1 P pG1 , q, which maps to p, sq under the correspondence (1.4.11) of 1.4. The subset ( S ,ell s P S ,ss : |ZpS 0 q| 8 ,s of S ,ss consists of those elements s such that the endoscopic datum G1 G1 s belongs to Eell pGq. Our concern will be the subset ( 1 S1 ,ell s P S ,ell : Gs R Esim pGq r of S ,ell . In the case that G P Esim pN q, for example, S 1 ,ell is just the subset of elements s P S ,ell with s 1. Following notation from 4.1, we write E,ell EpS ,ell q and
1 E,ell EpS 1 q ,ell 0 for the set of orbits in S ,ell and S 1 ,ell respectively under the action of S by conjugation. These are of course contained in the nite set
1

E EpS q EpS ,ss q dened in 4.1. Our aim is to replace the double sum over G1 and 1 in 1 our expression for (4.4.1) by a simple sum over E,ell . This leads to some rescaling constants, which we need to compute. Consider the set of pairs ( 1 Ydisc, pGq y pG , sG q ,

4.4. ENDOSCOPIC TERMS

193

where G is an actual L-homomorphism from L SLp2, Cq to L G0 that maps to , and sG belongs to S 1 G ,ell . The centralizer set S 1 G ,ell here is obviously the analogue of S 1 , which is to say that it is dened directly ,ell for the L-homomorphism G rather than some unspecied representative of r the equivalence class of L-homomorphisms G attached to in 1.4. There 1 p is a natural left action of the group G0 on Yell pG, q, dened by conjugation on the two components of a pair y. We would like to replace the double sum over G1 and 1 in (4.4.2) by a sum over the corresponding set
1 p G0 zzYdisc, pGq 1 p of G0 -orbits in Ydisc, pGq. (We will denote general orbit spaces by a double bar, so we can reserve a single bar for the special case of cosets.) 1 Any pair y pG , sG q in Ydisc, pGq is the image of a unique pair 1 , 1 q, under the correspondence (1.4.11). It is understood here that x pG G1 is simply an elliptic (nonsimple) endoscopic datum for G, taken up to p p translation of the associated semisimple element s1 P G by ZpG0 q , and 1 is an actual L-homomorphism from L SLp2, Cq to L G1 that maps that to . However, it is the isomorphism class of pG1 , 1 q, as an element in Eell pGq pG1 , q, that indexes the double sum (4.2.2). This class maps to p a subset of the G0 -orbit of pG , sG q. The summand of pG1 , 1 q actually dep pends only on the G0 -orbit, a property that follows easily from the fact that r f belongs to the subspace HpGq of HpGq. We have therefore to count the p number of isomorphism classes pG1 , 1 q that map to the G0 -orbit of pG , sG q. p To describe the G0 -orbit of y pG , sG q, we rst note that the stabilizer p 0 is the group of G in G ` p SG SG pG0 q Cent ImpG q, G0 .

p It then follows that the stabilizer of y in G0 is the stabilizer ( ` ` Sy SG ,sG g P SG : gsG g 1 sG

p p of the ZpG0 q -coset sG in SG . The orbit of y in G0 is therefore bijective ` p with the quotient coset space G0 {Sy . p To describe the isomorphism classes pG1 , 1 q in the G0 -orbit of y, we note 1 in E pGq can also be treated as a G0 -orbit. Its stabilizer p that the class of G ell p in G0 equals the subgroup AutG pG1 q. (These assertions follow directly from the denitions in [KS, p. 1819], and the fact that we are regarding G1 as a p ZpG0 q -orbit of endoscopic data.) The natural mapping from the orbit of y to the class of G1 can then be identied with the projection ` p p G0 {Sy AutG pG1 q{G0 . ` In particular, the stabilizer of 1 in AutG pG1 q can be identied with Sy . 1 in pG1 , q is not its orbit under Aut pG1 q, but However, the class of G

194

4. THE STANDARD MODEL

p rather its orbit (by conjugation) under the subgroup G1 of AutG pG1 q. This p is the reason that there are several isomorphism classes in the G0 -orbit of 1 -orbit of 1 is the same as its orbit under the product p y. Now the G p p IntG pG1 q G1 ZpG0 q , p since ZpG0 q commutes with the image of 1 . We recall that IntG pG1 q is a normal subgroup of AutG pG1 q, whose quotient is the nite group OutG pG1 q. p It follows from these remarks that the set of pairs pG1 , 1 q in the G0 -orbit of y is bijective with the set ` ` ` AutG pG1 q{Sy IntG pG1 q OutG pG1 q{ Sy IntG pG1 q{IntG pG1 q . Writing
` ` ` p p Sy IntG pG1 q{IntG pG1 q Sy {Sy X G1 ZpG0 q ,

we see that the number of such pairs is equal to the quotient (4.4.5)
` ` p p |OutG pG1 q| |Sy {Sy X G1 ZpG0 q |1 .

We have established that the double sum over G1 and 1 in (4.4.2) can p be replaced by a sum over G0 -orbits of pairs y pG , sG q, provided that the summand is multiplied by (4.4.5). The next step is to replace the new sum by an iterated sum over G and sG . The outer sum will be over the set p pG, q of G0 -orbits of L-homomorphisms G that map to the given element r r P pGq. Since the function lies in HpGq, the corresponding summand is independent of G , and the outer sum collapes. In other words, we can identify G with , provided that we multiply by the order (4.4.6) m |pG, q|.

p The stabilizer of G in G0 is by denition the group

SG pG0 q S pG0 q S .

The remaining inner sum can therefore be taken over the nite set
1 SG pG0 qzzS 1 G,ell S zzS,ell

S zzS 1 ,ell of orbits sG s in S 1 , under action by conjugation of either of the groups ,ell p S or S S {ZpG0 q .

We prefer to take orbits of the connected component S 0 pS q0 of S rather than S . The stabilizer in S of an element s P S 1 ,ell is the centralizer ` p S ` S,s {ZpG0 q Centps, S q ,s of s in S . The S -orbit of s is therefore bijective with S {S ` . The ,s S 0 -orbit of s is bijective with S 0 {S ,s , where we recall that S ,s is the

4.4. ENDOSCOPIC TERMS

195

centralizer of s in S 0 . We can therefore sum s over the set

1 1 S 0 zzS 1 ,ell EpS ,ell q E,ell

of S 0 -orbits in S 1 , rather than the set S z S 1 ,ell ,ell above, provided that we multiply the summand by the quotient (4.4.7) |S ` {S ,s | |S {S 0 |1 . ,s The last few observations have been directed at the sum over G1 and 1 in (4.4.2). We have now established that this double sum can be replaced 1 by a simple sum over the set E,ell , provided that the summand is multiplied by the product of the factors (4.4.5), (4.4.6), and (4.4.7). The summand itself becomes the product of these three factors with the right hand sides of (4.4.4) and (4.4.3). We thus obtain an expanded formula for the dierence (4.4.1) with which we started. Multiplying the various factors together, we 1 see that (4.4.1) can be written as the sum over s in E,ell of the product of two expressions p p p |S ` {S ` X G1 ZpG0 q |1 |S1 |1 |ZpG1 q|1 |S ` {S ,s |
,s ,s ,s

and |G |1 m |S {S 0 |1 pS 0 1 q 1 p 1 q f 1 p 1 q, in which pG1 , 1 q maps to the pair y p, sq pG , sG q. We consider each of the two expressions in turn. To simplify the rst expression, we note that S 0 pS ` q0 Centps, S 0 q0 , ,s ,s and hence that p p |S1 | |0 pS 1 q| S ` X G1 {pS ` q0 ZpG1 q, ,s ,s p where G1 denotes the quotient p p p p p G1 ZpG0 q{ZpG0 q G1 {ZpG1 q. Consequently,
` ` p p |S,s {S,s X G1 ZpG0 q|1 |S1 |1 p p p 1 |S ` { S ,s X G1 |1 S ` X G1 {pS ` q0 ZpG1 q ,s ,s ,s

p |S ` {S 0 ZpG1 q|1 . ,s ,s The rst expression therefore equals p p |S ,s {S 0 ZpG1 q|1 |ZpG1 q|1 , ,s a product that also equals p |0 pS ,s |1 |S 0 X ZpG1 q|1 . ,s

196

4. THE STANDARD MODEL

The product of the rst three factors in the second expression equals the constant C of (4.3.5), since |S {S 0 | |S {S 0 | |S |. The fourth factor pS 0 1 q in the expression has been dened (4.1.8) if S 0 1 is replaced by any complex, connected reductive group S1 . We recall the property (4.1.9) from 4.1, which asserts that pS1 q pS1 {Z1 q |Z1 |1 , for any central subgroup Z1 of S1 . We can therefore write ` p pS 0 1 q S 0 {S 0 X ZpG1 q ,s ,s p pS 0 q |S 0 X ZpG1 q|. ,s ,s Taking the product of the resulting two expressions, we conclude that the dierence (4.4.1) equals |0 pS ,s q|1 pS 0 q 1 p 1 q f 1 p 1 q, (4.4.8) C ,s
1 sPE,ell

where pG1 , 1 q maps to the pair p, sq. As at this point in the discussion of the last section, we will need a lemma on signs. Its purpose is to express the sign 1 p 1 q G1 ps1 q explicitly in terms of p, sq. To state it, we write xs for the image of s in S . Lemma 4.4.1. The sign characters for G and G1 satisfy the relation G1 ps1 q G ps xs q, for any elements G1 , 1 and s such that pG1 , 1 q maps to p, sq. We shall give the proof of this lemma in 4.6, along with that of its predecessor Lemma 4.3.1 from the last section. Assuming it for now, we make the appropriate substitution in the expression (4.4.8) we have obtained 1 for (4.4.1). We also write f 1 p 1 q as fG p, sq, according to the denition (4.2.5). We thus nd that (4.4.1) equals 1 C |0 pS ,s q|1 pS 0 qG ps xs q fG p, sq. ,s
1 sPE,ell 1 1

The last expression more closely resembles that of Lemma 4.3.2 if we bre the sum over points x P S . Accordingly, we write E pxq, x P S , for the bre of x under the canonical mapping from E to S . We then set E,ell pxq E pxq X E,ell

4.4. ENDOSCOPIC TERMS

197

and
1 1 E,ell pxq E pxq X E,ell .

With this notation, we obtain the following result . Lemma 4.4.2. The dierence Idisc, pf q sG pf q disc, equals C
1 |0 pS ,s |1 pS 0 q G ps xq fG p, sq, ,s
1 xPS sPE,ell pxq

r r for any P pN q and f P HpGq. The existence of the linear form f 1 p 1 q, for any G1 in the complement of Esim pGq, is part of our basic induction hypothesis. Suppose that we happen to know also that f 1 p 1 q is dened for any G1 P Esim pGq and 1 P pG1 , q. This is primarily a local assumption, namely that for any valuation v, 1 the localization v belongs to ` pG1 , v q, and Theorem 2.2.1(a) holds for v 1 , 1 q. It is weaker than asking that the stable multiplicity formula of pGv v Theorem 4.1.2 hold for G1 and 1 . However, it does allow us to dene the dierence 0 p1 p1 |S1 |1 pS 0 1 q1 p 1 q f 1 p 1 q Sdisc, pf 1 q Sdisc, pf 1 q (4.4.9)
1 PpG1 ,q

p1 between Sdisc, pf 1 q and its expected value, as well as the sum 0 G p1 pG, G1 q 0 Sdisc, pf 1 q, (4.4.10) sdisc, pf q
G1 PEsim pGq

which represents the dierence between sG pf q and its expected value. disc, The dierence sG pf q 0 sG pf q disc, disc, equals the sum over G1 P Esim pGq and 1 P pG1 , q of the product of the right hand sides of (4.4.3) and (4.4.4). Using (4.2.5), (4.3.5) and other denitions above, we can write it as a sum similar to that of Lemma 4.4.2. Indeed, we have only to apply the discussion above to summands indexed by pairs pG1 , 1 q with G1 P Esim pGq. These indices are attached to pairs p, sq in which s belongs to the complement of S 1 ,ell in S ,ell . In particular, 1 p 1 q for G1 P E the existence of the linear forms f sim pGq amounts to the 1 existence of fG p, sq, as a function dened for s in the entire domain S ,ell . The arguments above carry over to these summands without change. It follows that the formula of Lemma 4.4.2 remains valid with 0 sG pf q in ,disc 1 place of sG pf q, and E,ell pxq in place of E,ell pxq. ,disc It is convenient to change variables in the resulting sum over s. The set of S 0 -orbits in S pxqell is invariant under translation by the point s . Since

198

4. THE STANDARD MODEL

s lies in the center of S , the centralizer S ,s equals S ,s s . A change of variables from s to s s therefore yields the expression 1 C |0 pS ,s q|1 pS 0 q G pxq fG p, s sq. ,s
xPS sPE,ell pxq

This is equal to the dierence between Idisc, pf q and 0 sG pf q. ,disc 1 Suppose that in addition to being well dened, the function fG p, sq depends only on the image x of s in S . This is implied by the global intertwining identity, but is again weaker. In fact, it is easily derived from 1 basic denitions, as we shall see in the next section. The function fG p, s sq above can then be taken outside the last sum over s. Lemma 4.4.2, or rather its minor extension above, can then be formulated as follows. Corollary 4.4.3. Suppose that the function
1 1 fG p, xq fG p, sq,

x P S , s P E pxq,

is dened, and depends only on x. Then the dierence Idisc, pf q 0 sG pf q disc, equals (4.4.11) where e pxq
sPE,ell pxq

xPS

1 e pxq G pxq fG p, s xq,

|0 pS ,s q|1 pS 0 q. ,s

In Corollaries 4.3.3 and 4.4.3, we now have two parallel expansions for the discrete part of the trace formula, or rather its -component Idisc, pf q. The comparison of these expansions will be a central preoccupation from this point on. As an elementary example of what is to come, let us consider the simplest of cases. r r Suppose that P pN q and G P Esim pGq are such that does not belong r to pGq. Then ` G 0 G G rdisc, pf q rdisc, pf q tr Rdisc, pf q and
0 G sdisc, pf q G G sG pf q Sdisc, pf q 0 Sdisc, pf q, disc, G since the expected values of rdisc, pf q and sG pf q are by denition equal disc, to 0. Strictly speaking, the premises of the two corollaries are not valid, r since the relevant functions are not dened in the case R pGq. However, the expansions of Lemmas 4.3.2 and 4.4.2 hold. The coecient C in these expansions vanishes, since the factor m in (4.3.5) equals 0. It follows that (4.4.12) ` G G G r r tr Rdisc, pf q Sdisc, pf q 0 Sdisc, pf q, R pGq, f P HpGq.

4.5. THE COMPARISON

199

G r In particular, the trace of Rdisc, pf q is a stable linear form on HpGq in this case. We of course expect that each term in the identity vanishes, but we are not yet ready to prove this. Observe that we could also have derived (4.4.12) directly from our induction hypotheses and the original expansions (4.1.1) and (4.1.2). In particular, the identity does not depend on the sign formulas of Lemmas 4.3.1 and 4.4.1.

4.5. The comparison We shall now compare the two expansions we have established. In particular, we shall try to show that for many , the conditions of the parallel expansions of Corollaries 4.3.3 and 4.4.3 are valid. We recall that belongs ` r r to the set pN q GpN q of parameters for the twisted general linear r r group GpN q GLpN q pN q over the global eld F . The most dicult r r case is when lies in ell pN q, which is to say that belongs to 2 pGq, for r some datum G (a priori unique) in Eell pN q. Its analysis will be taken up in later chapters. In this section, we assume that lies in the complement of r ell pN q. The two corollaries were based on the further assumption that two functions are dened, and depend only on a point x P S . Suppose that r belongs to pGq, where as in 4.3 and 4.4, G is either a given datum in rsim pN q or equal to GpN q. Our condition that lies in the complement r E r of ell pN q implies that the functions are at least dened. For the function fG p, uq of Corollary 4.3.3, this follows from our induction hypothesis that the pairing xr, M y in the denition p4.2.4q1 exists for the Levi subgroup u r 1 M G0 attached to . The function fG p, sq of Corollary 4.4.3 is also amenable to induction. Its existence reduces to the analogue of Theorem 2.2.1(a) for the group M G0 , as the reader can readily verify. Alterna1 tively, the existence of fG p, sq can be regarded as part of an analysis of the 1 p 1 q in the denition p4.2.5q1 that we will undertake presently linear form f (following the statements of formulas (4.5.7) and (4.5.8) below). The main part of the assumption underlying the two corollaries is that each of the two functions depends only on the image x of u or s in S . To see how the comparison works, we temporarily take on a broader assumption, which includes the premise of Corollaries 4.3.3 and 4.4.3. We suppose for the moment that the global intertwining relation holds for any G in either of the corollaries. In other words, we suppose that (4.5.1)
1 fG p, s xq fG p, xq,

r f P HpGq, x P S ,

r r r where G is either an element in Esim pN q with P pGq, or is equal to GpN q itself. According to Corollary 4.4.3, the dierence (4.5.2) Idisc, pf q 0 sG pf q, disc, r f P HpGq,

200

4. THE STANDARD MODEL

equals C
xPS 1 e pxq G pxq fG p, s xq.

Corollary 4.3.3 tells us that the dierence (4.5.3) equals C

xPS G Idisc, pf q 0 rdisc, pf q,

r f P HpGq,

i pxq G pxq fG p, xq.

We have now reached the point where we appeal to Proposition 4.1.1. This result, which we recall is a kind of miniature replica of the basic identity between (4.1.2) and (4.1.1), tells us that the coecients e pxq and i pxq in the two expressions are equal. It follows from (4.5.1) that the expressions themselves are equal. The two dierences (4.5.2) and (4.5.3) are therefore equal. We conclude from the denitions (4.4.10) and (4.3.8) that ` G p1 (4.5.4) |G |1 tr Rdisc, pf q pG, G1 q 0 Sdisc, pf q,
G1 PEsim pGq

r for any G and any f P HpGq. r Suppose rst that G is one of the groups in Esim pN q. Then G1 G is the only element in the set Esim pGq. The equation (4.5.4) reduces to ` G G (4.5.5) tr Rdisc, pf q 0 Sdisc, pf q,
G since |G | 1 in this case. In particular, the trace of Rdisc, pf q is a stable r linear form on HpGq. Notice that (4.5.5) has the same form as the identity (4.4.12) established at the end of the last section, in case does not lie r in pGq. Consequently, (4.5.5) is valid if G is replaced by any datum in rsim pN q. E r Next, suppose that G equals GpN q. The character ` G f P HpGq, tr Rdisc, pf q ,

r then vanishes, since our condition that R ell pN q means that cannot r q0 GLpN q. The general idencontribute to the discrete spectrum of GpN tity (4.5.4) reduces to p1 (4.5.6) pG, G1 q 0 Sdisc,1 pf 1 q 0
G1 PEsim pGq

r in this case. The groups G1 here represent the data in Esim pN q we have just treated, and for which we have established (under the condition (4.5.1)) that ` G1 0 p1 Sdisc, pf 1 q tr Rdisc, ph1 q , r r for any function h1 P HpGq whose image in SpG1 q equals f 1 . Continuing to suppose that the general condition (4.5.1) is valid, we combine (4.5.5) and (4.5.6), and then let G range over all of the elements

4.5. THE COMPARISON

201

r in Esim pN q. Appealing to Proposition 2.1.1, or rather its global extension described in 3.4, we see that ` G rpN, Gq tr Rdisc, pf q 0,
r GPEsim pN q

for any compatible family of functions ( r r f P HpGq : G P Eell pN q . r (The groups G that represent elements in the complement of Esim pN q in r Eell pN q are part of the original denition (3.4.11), but they play no role here. For since they are composite, we can assume inductively that the G corresponding representations Rdisc, are zero.) The coecients ` r rpN, Gq GpN q, G
G are positive. The coecients of the decompositon of Rdisc, pf q into irreducible representations are of course also positive. We can therefore apply Proposition 3.5.1, the basic vanishing property we have established for compatible families. It tells us that

(4.5.7)

G Rdisc, pf q 0,

r f P HpGq,

r for any G P Esim pN q. In other words, does not contribute to the discrete spectrum of G. This in turn implies that (4.5.8)
0 G Sdisc, pf q

0,

r f P HpGq.

G In other words, the formula of Theorem 4.1.2 for Sdisc, pf q is valid for any r G P Esim pN q. We shall see that the comparison just completed has broad application. r In particular, we shall show that for many parameters P pN q, the condir tion (4.5.1) is indeed valid for every G with P pGq. This will allow us to r conclude that the global theorems hold for the given parameter R ell pN q. The putative identity (4.5.1) includes the condition that each side is a well dened function of x P S . In other words, each side depends only on the image in S of the datum in terms of which it is dened. This is of course the premise on which the two Corollaries 4.3.3 and 4.4.3 were based. It is not dicult to verify much of it directly. We shall do so, as a necessary prelude to being able to establish cases of (4.5.1) by induction. 1 The left hand side fG p, s xq of (4.5.1) can be treated uniformly for r r G P Esim pN q or G GpN q. It is dened p4.2.5q1 in terms of a semisimple element s P S whose projection onto S equals x. It equals 1

f p 1q f

G1

p 1 q,

r f P HpGq,

where p G1 , 1 q is the preimage of the pair p, sq, s s s,

202

4. THE STANDARD MODEL

under the general correspondence pG1 , 1 q p, sq. We shall establish for any G that f 1 p 1 q depends only on the image x of s. Note rst that if s is replaced by an S 0 -conjugate s1 , pG1 , 1 q is replaced 1 by its image pG1 , 1 q under an isomorphism of endoscopic data. It follows 1 from the denitions that
1 1 f1 p1 q f 1 p 1 q, 1 f1 f G1 .
1

Since any automorphism of the complex reductive group S 0 stablilizes a conjugate of pT , B q, we can choose s1 so that Intps1 q stablilizes pT , B q itself. With this property, the representative sx s1 of x in S is determined up to a T -translate. In particular, the complex torus T ,x Centpsx , T q0 in T is uniquely determined by x. Now any point in T sx can be written in the form tsx t1 tx tsx tx t1 , t P T , tx P T ,x . It will therefore be enough to show that the linear form
1 1 1 1 fx px q f1 p1 q

remains unchanged under translation of sx by any element tx P T ,x . x0 p p The centralizer Mx of T ,x in G0 is a F -stable Levi subgroup of G0 x that contains the dual M of the underlying Levi subgroup M . Since the x0 automorphism Intpsx q centralizes T ,x , it stabilizes not only Mx , but also p0 x0 x0 some parabolic subgroup Px P PpMx q. The Mx -torsor x x0 Mx Mx Intpsx q p can therefore be treated as a Levi subset of G, which is in turn dual to a r Levi subset Mx of G. (Remember that we are including the case G GpN q in our analysis. As in 2.2, we are using terminology here that originated in [A5, 1].) Now p, sq is the image of a pair pMx , sMx q, r Mx P pMx q, sMx P SMx ,
0 attached to Mx under the L-embeddings L Mx L G0 and L Mx L G. 1 1 This pair is in turn the image of an endoscopic pair pMx , Mx q for Mx . In 1 for M , which can be particular, we obtain an elliptic endoscopic datum Mx x identied with a Levi subgroup of G1 . It follows that 1 1 f 1 p 1 q fMx pMx q,

where
1 1 fMx f Mx pf 1 qMx
1

4.5. THE COMPARISON

203

1 1 1 is the transfer of f to Mx . Since fMx pMx q does not change under translation 1 1 of sx by T ,x , the same is true of fx px q. We conclude that the function 1 1 1 fG p, xq f 1 p 1 q fx px q 1 does indeed depend only on x. So therefore does the left hand side fG p, s xq of (4.5.1). The right hand side fG p, xq of (4.5.1) is less clear-cut. Recall that it is dened p4.2.4q1 in terms of a point u in N whose projection xu onto S equals x. The bre of x under this projection is a torsor under the action 0 of the subgroup W of N . (We have to include the subscript , since the r case G GpN q remains a part of the discussion.) The dependence of fG p, xq on u is twofold. It depends on its projection w wu onto W , through the normalized intertwining operator in the trace ` tr RP pw, M , M q IP pM , f q , r

r r r and on the value at u of x , M y, an extension of the character x , M y 1 to S 1 u that in turn determines the associated extension rM of from S 1 W 0 is a normal subgroup of N , we can the representation M . Since S 0 r assume that each is constant on the W -orbit of u. That is, we can arrange that r r r r ru pw0 uq pw0 uq prq, w0 P W 0 .
0 1 r In particular, in the case that u lies in S W , we take to be the 1 with the trivial character on W 0 . product of the character x , M y on S We then conclude, from the earlier stages of the proof of Proposition 2.4.3 for example, that

r M pw0 wq M pwq, r and hence that the corresponding intertwining operators for G satisfy r r r RP pw0 w, M , M q RP pw0 , M , M q RP pw, M , M q,
0 w 0 P W .

The condition that fG p, xq depend only on x can at this point be seen to be equivalent to the requirement that the operators RP pw0 , M , M q equal 1 for r 0 all w0 P W . If the operators are all trivial, the condition on fG p, xq follows immediately from p4.2.4q1 and the conventions above. The other direction r is perhaps less clear, since f ranges only over the subspace HpGq of HpGq, 0 , , q is unitary. Since but it can be deduced from the fact that RP pw rM M we will ultimately establish the condition on fG p, xq by showing directly r that the operators RP pw0 , M , M q are all trivial, there will be no harm in assuming henceforth that the two conditions are indeed equivalent. We will not be in a position, though, to establish either condition in general for some time.

204

4. THE STANDARD MODEL

r r Suppose however that G GpN q. The global diagram (4.2.3) for GpN q maps into the local diagram (2.4.3) attached to any completion v . In partic0 0 0 ular, we have an injection w0 wv from W to Wv , and a decomposition 0 r r RP pw0 , M , q RP pwv , M,v , v q.
v

The factors on the right are local normalized intertwining operators for G0 GLpN q. It follows from Theorem 2.5.3 that they are all equal to the r identity. The global operator RP pw0 , M , q is therefore also equal to 1. The right hand side fG p, xq of (4.5.1) therefore does depend only on x, in r the case G GpN q. r Consider the other case that G P Esim pN q. As the Weyl group of 0 , T q, W 0 is generated by simple reections tw 0 u. Suppose that pS dimpT q 2.
0 Then any w centralizes a torus of positive dimension in T . The centralizer p x p x of this torus in G is a proper Levi subgroup M of G that contains M , and is dual to a Levi subgroup M of G that contains M . If is the image in 0 r pM q of the chosen parameter M P 2 pM, q, w belongs to the Weyl group of pS 0 , T q. We can assume inductively that 0 0 r r RP pw , M , M q RP XM pw , M , q 1. 0 r It follows that RP pw0 , M , M q equals 1 for any w0 P W . Thus, the right r hand side fG p, xq of (4.5.1) again depends only on x, in case G P Esim pN q and dimpT q 2. It is the complementary case that dimpT q 1 that causes diculty. We shall treat some (though not all) of it by a more sophisticated comparison argument later in the section. How do the groups S S pGq vary, as G ranges over elements r r G P Esim pN q with P pGq? Suppose for a moment that G is xed. The centralizer S S pGq is then equal to a product ` Spp i , Cq (4.5.9) Op i , Cq GLp j , Cq ,
` iPI pGq

iPI pGq

jPJ

r in the notation of 1.4. As we observed in 3.4, the condition that P pGq implies that the character G equals . If is nontrivial or N is odd, G r is the only datum in Esim pN q with G . If is trivial and N is even, r however, there is a second datum G_ P Esim pN q with G_ . The pa` _ q if and only if r rameter then belongs to pG i is even for each i P I pGq. ` ` With this condition, we have I pG_ q I pGq and I pG_ q I pGq, and _ _ the centralizer S S pG_ q equals ` (4.5.10) Spp i , Cq Op i , Cq GLp j , Cq .
` iPI pGq iPI pGq

jPJ

4.5. THE COMPARISON

205

r Proposition 4.5.1. Suppose that lies in the complement of ell pN q r r in pN q, and in addition, that it lies in the complement of ell pGq for every r G P Esim pN q. Then ` G G r f P HpGq, tr Rdisc, pf q 0 0 Sdisc, pf q, r for any G P Esim pN q. In other words, the global Theorems 1.5.2 and 4.1.2 hold for G and . Remarks. 1. The rst condition on has been a de facto hypothesis for the last three sections. The second condition is new, and has been imposed in order to avoid subtleties that will have to be treated later. It r does not imply the rst, since it does not preclude parameters P 2 pGq, rsim pN q in Eell pN q. r for elements G in the complement of E 2. There are other global theorems besides the two mentioned in the statement of the proposition that must be established in general. However, Theorem 1.5.3 applies to parameters that are simple (and generic), r and therefore lie in ell pN q. It is irrelevant to the case at hand. The same is true of the global Theorem 4.2.2. The remaining global assertion is the r intertwining relation of Corollary 4.2.1, for both the case G P Esim pN q and r G GpN q. This is just the putative identity (4.5.1). If we are able to establish it directly, we will thus obtain not just the statement of the proposition, but in fact all of the global results that pertain to . Proof. We shall try to establish (4.5.1) directly by induction, and thereby obtain the required assertion from the comparison above. We shall nd that this is possible for many, but not all of the given parameters . There will be some cases that require a more delicate comparison. To see where the problems arise, let us begin with any parameter in the compler ment of ell pN q, and see how far we can go. r r We have to verify (4.5.1) in both the case G P Esim pN q and G GpN q. The identity makes sense only if the group S S pGq is dened, which is r r to say that belongs to pGq. This condition always holds if G GpN q. r r However, it is quite possible that not lie in pGq for any G P Esim pN q. If this is the case, (4.5.5) is still valid for each G, as we saw in (4.4.12), and the comparison that yields the required identities (4.5.7) and (4.5.8) can be r carried out. We shall therefore assume implicitly that P pGq, for any of the data G P Esim pN q we consider below. Assume for the present that piq dimpT q 2, r G P Esim pN q,

as in the remarks prior to the statement of the proposition. Then as we have seen in all cases, the two sides of (4.5.1) each depend only on x. Moreover, the function on the left hand side satises the relation
1 1 1 1 fG p, xq f 1 p 1 q fx px q fMx px , xq,

206

4. THE STANDARD MODEL

r for Mx G and x P pMx q as in the earlier remarks. It follows easily from the denitions that Ms x Mx and s x x . The left hand side itself therefore satises 1 1 fG p, s xq fMx px , s xq. On the other hand, the earlier representative sx of x in S lies in the subset N of S . We take its T -coset ux to represent x in N , and on the right hand side of (4.5.1). The coset ux centralizes the complex torus T ,x Centpsx , T q0 in T . Since T ,x can be identitied with the -split part of the centralizer x of Mx , the right hand side satises the relation fG p, xq fMx px , xq. Assume also that piiq dimpT ,x q 1, r G P Esim pN q, x P S .

r Since the center of any G P Esim pN q is nite, Mx is a proper Levi subgroup r of G in this case. Consider the other case that G GpN q. The given r ell pN q means that either the multiplicity k condition that not lie in of some component k of is larger than 1, or the indexing set J J p is not empty. In the rst instance, the subgroup S of G0 has an Intpsx q stable factor GLp k , Cq of positive semisimple rank. In the second, S has a subgroup isomorphic to C C , whose factors are interchanged by Intpsx q. r It follows easily that the condition (ii) holds also in case G GpN q, and that Mx is again proper in G. We can therefore argue by induction. More precisely, we can assume inductively that the analogue of (4.5.1) for Mx holds in all cases. It follows that
1 1 fG p, s xq fMx px , s xq fMx px , xq fG p, xq.

The identity (4.5.1) therefore holds for G. We note for future reference that r in case G GpN q, (4.5.1) in fact holds for any in the complement of r ell pN q. We have established the required identity (4.5.1) for , provided that r the conditions (i) and (ii) on hold for any G in Esim pN q and any x in S S pGq. For which do the two conditions not hold? Consider the product formula (4.5.9) for the group S attached to a r r given G P Esim pN q with P pGq. The second and third components of the product are connected complex groups. We can therefore assume that the representative sx of x described above lies in the rst component, the nonconnected group dened by the product over I` pGq. To check the conditions (i) and (ii), we need only describe the rank of the complex group (4.5.9), and the rank of the centralizer of sx in (4.5.9). The two conditions clearly hold unless the third component, the product over J , is either trivial or consists of just one factor GLp1, Cq C . In the

4.5. THE COMPARISON

207

second case we have an innite central subgroup of S , whose centralizer is p a proper Levi subgroup of G. The identity (4.5.1) then follows by induction. r (We note that lies in the complement of disc pGq in this case, and that the global theorems also follow directly from the expansions for (4.5.2) and (4.5.3).) We may therefore assume that the indexing set J for is empty. The two conditions remain valid unless the second component, the product over I pGq, is either trivial or consists of one factor Spp2, Cq. Putting aside the second possibility for the moment, we suppose that the set I pGq is also empty. The conditions (i) and (ii) are then still valid unless i 2 for every ` ` i P I pGq, or i 3 for one i P I pGq and i 1 for the remaining indices i. Putting aside the second possibility again, we suppose that S `
` iPI pGq

` Op i , Cq ,

P t1, 2u.

` ` Recall that p q` stands for the kernel of a character pGq of order 1 or 2. If there were an element s si , si P Op i , Cq, i

in the kernel, with detpsi q 1 whenever i 2, would belong to the r r subset ell pGq of pGq, since the centralizer of s in S would then be nite. This outcome is expressly ruled out by the hypotheses of the proposition. On the other hand, if there is no such s, S has a central torus C . In this case, (4.5.1) again follows inductively from its reduction to a proper Levi subgroup of G. We have thus seen that (4.5.1) holds for any given that does not fall into one of the two cases set aside above. In the rst of these cases, the index set J is empty, while the product over I pGq in (4.5.9) reduces simply to Spp2, Cq. Observe that the conditions (i) and (ii) remain valid in this case ` unless the factors in the product over I pGq are all equal to Op1, Cq Z{2Z. In the second case, the index sets J and I pGq are both empty, while the ` product over I pGq reduces to a product of Op3, Cq with several copies of the group Op1, Cq Z{2Z. The two outstanding cases thus take the form # (4.5.11) and # (4.5.12) 31 2 r S SOp3, Cq pZ{2Zqr ,
1

21 2 r S SLp2, Cq pZ{2Zqr 1 ,
1

r 1,

r 1,

208

4. THE STANDARD MODEL

r where S S pGq for some G P Esim pN q, and # ` r, if 1, 1 r r 1, otherwise. r In these remaining cases, it is understood that G P Esim pN q is some r datum with P pGq. According to the discussion leading up to (4.5.10), G is uniquely determined by this condition, except for the case of (4.5.11) with r 1. In this case, however, there is a second element r r G_ P Esim pN q with P pG_ q. It then follows from (4.5.11) that the group ` ` _ S S pG_ q equals Op2, Cq . There are two possibilities. If the charr acter `,_ ` pG_ q is 1, S _ equals Op2, Cq, and lies in ell pG_ q, a

condition on that is ruled out by the proposition. We can therefore as`,_ _ sume that 1 if r 1 in (4.5.11). In this case, S C , and the _ follows inductively from its reduction to a proper analogue of (4.5.1) for G Levi subgroup of G_ . Let us summarize our progress to this point. We have established (4.5.1) for any that satises the conditions of the proposition, and that does not belong to one of the two exceptional cases (4.5.11) and (4.5.12). The required global theorems for such then follow from the general comparison at the beginning of the section. It remains to treat the exceptional cases by a more subtle comparison. Suppose that and G are as in (4.5.11). We cannot use the earlier comparison, because we do not know that the global identity (4.5.1) holds. However, we can still apply Corollary 4.4.3 to the dierence (4.5.2), since the required condition was established in general prior to the statement of this proposition. We can also apply Corollary 4.3.3 to the dierence (4.5.3). Here, we do not know that the function fG p, xq in the resulting expression depends only on the image x of the element u P N pxq used to dene it. However, in the case at hand, there is only one element w wx in the subset W,reg pxq of W pxq that indexes the internal sum in the expression (4.3.9) of Corollary 4.3.3, and hence only one element in u ux in the corresponding subset N,reg pxq of N pxq. We take the liberty of setting fG p, xq fG px, ux q in order to use the notation of Corollary 4.3.3. Substituting the identity e pxq i pxq, we see that the dierence ` G 0 G G G (4.5.13) sdisc, pf q 0 rdisc, pf q 0 Sdisc, pf q tr Rdisc, pf q equals C
xPS

` 1 i pxq G pxq fG p, xq fG p, s xq .

The constant C is dened in (4.3.5) as a product C m |G |1 |S |1 m |S |1

4.5. THE COMPARISON

209

of positive factors. The coecient i pxq satises 0 i pxq |W |1 s0 pwq | detpw 1q|1
wPW,reg pxq 0 |W |1 s0 pwx q | detpwx 1q|1 1 , 4 0 since |W | 2, | detpwx 1q| 2 and s0 pwx q 1, in this case. It is the minus sign here that will be the critical factor. From the form (4.5.11) of S , we observe that the R-group R is trivial. It follows that there is an isomorphism x xM from S onto SM , where M represents a xed element in 2 pM, q. The coecient G pxq then satises

G pxq M pxM q M pxM q, M a reduction one can infer either directly from the denition (1.5.6) and the x fact that we can identify x with a point xM in M . We can therefore write C i pxq G pxq n mM |SM |1 M pxM q, for the positive constant n
1 fG p, s xq 1 4

m m1 . M

The remaining terms in our expression for (4.5.13) are the linear forms and fG p, xq. It follows from the denitions p4.2.5q1 and (2.2.6) that
1 1 fG p, xq f 1 p 1 q f 1 pM q xs xM , M y fM pM q, r M PM 1 where pM 1 , M q is the endoscopic pair attached to pM , xM q. A similar formula holds with s x in place of x and xM in place of s xM , since

s ps xqM s ps xM q ps q2 xM xM . We can therefore write 1 C i pxq G pxq fG p, s xq

xPS

n mM |SM |1 n
r M PM

M pxM q xxM , M y fM pM q

xM PSM M P r M

` mM pM q tr IP pM , f q ,

by applying the multiplicity formula mM pM q mM |SM |1

xM PSM

M pxM q xxM , M y

210

4. THE STANDARD MODEL

of Theorem 1.5.2 inductively to the proper Levi subgroup M of G, and substituting the usual adjoint relation ` fM pM q tr IP pM , f q of fM . The outer linear form fG p, xq fG p, ux q is given by p4.2.4q1 . It can be written as ` r xxM , M y tr RP pwx , M , M q IP pM , f q ,
r M PM

u r since wx wu , and xxM , M y xr, M y if u ux . We can therefore write C i pxq G pxq fG p, xq

xPS

r M PM

` r mM pM q tr RP pwx , M , M q IP pM , f q

by again applying the multiplicity formula. We have expanded (4.5.13) into a (signed) sum of two rather similar expressions. The dierence (4.5.13) thus equals `` r (4.5.14) n mM pM q tr 1 RP pwx , M , M q IP pM , f q .
r M PM G It follows that the stable linear form 0 Sdisc, pf q on the right hand side of ` G (4.5.13) equals the sum of tr Rdisc, pf q with (4.5.14). r We have taken G to be the xed group in Esim pN q that is implicit in for a general group in E rsim pN q. If does (4.5.11). We therefore write G r not lie in pG q, the identity (4.4.12) (with G in place of G) is valid. This condition applies to any G G, if r 2. If r 1, we have seen that there r is one group G G with P pG q, namely G G_ . But we have also `,_ r seen that 1 in this case (given our assumption that R ell pG_ q), and that the analogue of (4.5.1) for G follows by induction. It thus follows from (4.5.5) that ` G 0 G r Sdisc, pf q tr Rdisc, pf q , f P HpG q,

r for any G P Esim pN q distinct from G. r We saw earlier in the proof that the analogue of (4.5.1) for GpN q holds for r ell pN q, and in particular, for the case (4.5.11) we are considering. any R It then follows from (4.5.6) that r pG rpN, G q 0 Sdisc, pf G q 0,
r G PEsim pN q

4.5. THE COMPARISON

211

r r for any f P HpN q. Substituting for each of the stable linear forms on the left, we see that the sum of the expression ` G rpN, G q tr Rdisc, pf q (4.5.15)
r G PEsim pN q

with (4.5.14) vanishes, for any compatible family of functions ( r f P HpG q : G P Esim pN q . r The global intertwining operator RP pwx , M , M q in (4.5.14) has square equal to 1. Its eigenvalues are therefore equal to 1 or 1. The expression (4.5.14) is in consequence a nonnegative sum of irreducible characters. The same is true of the other expression (4.5.15). We can therefore apply Proposition 3.5.1 to the sum of the two expressions. We conclude that r RP pwx , M , M q 1, and that ` G G tr Rdisc, pf q 0 Sdisc, pf q 0, r f P HpG q, r M P M pM q,

r for every G P Esim pN q. This completes the proof of the proposition for the rst exceptional case (4.5.11). The argument for the remaining exceptional case (4.5.12) is similar. It is r actually slightly simpler, since in this case G is the only element in Esim pN q r with P pGq. In particular, there is no distinction to be made between r 1 and r 1. The comparison is otherwise identical, and the required global theorems in this last case follow as above. Proposition 4.5.1 thus holds in all cases. Corollary 4.5.2. Suppose that is as in the proposition. Then the global intertwining relation (4.5.1) of Corollary 4.2.1 is valid unless falls into one of the two exceptional classes (4.5.11) or (4.5.12), in which case we have only the weaker identity ` 1 r (4.5.16) G pxq fG p, s xq fG p, xq 0, f P HpGq,
xPS

r r r for any G P Esim pN q with P pGq. If G GpN q, the identity (4.5.1) holds r without restriction, and indeed, for any R ell pN q. Proof. The assertions of the corollary were established in the course of proving the proposition. For example, we veried (4.5.1) in the generality above in order to apply the earlier comparison. In the case (4.5.11), the identity (4.5.16) follows from the fact, established at the end of the proof, that the dierence (4.5.13) vanishes. The case of (4.5.12) is similar. The most interesting aspect of the last proof is the treatment of the exceptional cases (4.5.11) and (4.5.12). For example, if r 1 in (4.5.11), G is r r the only datum in Esim pN q with P pGq. The same goes for (4.5.12), with

212

4. THE STANDARD MODEL

r r 1. If G P Esim pN q is distinct from G, and satises the further condition G G that G , we know a priori that R 0 and 0 S 0. However, this does not account for the case that G G_ , in which G 1 and N is even. We really do need the argument above to deal with this case, as well as to compensate for not knowing (4.5.1) in case G G. Looking back at the argument, the reader can see that we have been lucky with the sign s0 pwx q, in which wx was the unique element in W,reg pxq. If s0 pwx q had somehow turned out to equal 1 instead of 1, the argument would have failed, perhaps dooming the entire enterprise. The role of the sign s0 pwx q has broader implications. It is representa tive of a whole collection of arguments, of which our treatment of (4.5.11) and (4.5.12) is among the simplest and most transparent. The more elaborate comparisons will be applied in later chapters to the cases excluded by Propositon 4.5.1. In general, s0 pwx q is an essential ingredient of the interlocking sign Lemmas 4.3.1 and 4.4.1, and hence also Proposition 4.5.1, but which were instrumental in the general reductions of 4.3 and 4.4, and which have so far been taken for granted. This underlines the signicance of the next section, in which we will establish the two lemmas in general.

4.6. The two sign lemmas Lemmas 4.3.1 and 4.4.1 are parallel. They played critical roles in the corresponding expansions of 4.3 and 4.4, which in turn led to the general results of 4.5. Both lemmas concern the basic sign character G attached to in 1.5. The rst can be regarded as a reduction of the character to its analogue for a Levi subgroup, while the second amounts to a reduction to endoscopic groups. We shall prove them in this section, thereby completing the proof of the general reduction obtained in the last section. We revert to the setting of 4.34.4. That is, we take G to be either r r a group in Esim pN q or equal to GpN q itself. In the second case, L G and S are only bitorsors, under the respective groups L G0 and S . Following our earlier convention, we shall understand a representation of L G or S to be a morphism that extends to a representation of the group L G` or S ` generated by L G or S . For example, the canonical extension of G to the r quotient S of S in the case G GpN q is a character in this sense. That
` is, it extends to a character (of order 2) on the quotient S of S ` . We will need a certain amount of preliminary discussion before we can prove the two sign lemmas. This will be useful in its own right. For we shall see that it gives some insight into the meaning of the sign characters . r Suppose that P pGq is xed. We assume that Theorems 1.4.1 and 1.4.2 hold for the self-dual generic constituents of . The group L is then dened, and has local embeddings (1.4.14). The sign character is dened

4.6. THE TWO SIGN LEMMAS

213

(1.5.6) in terms of the irreducible constituents b b of the representation ` r ps, g, hq AdG s G pg, hq of the product (4.6.1) S L SLp2, Cq, p acting on the Lie algebra g of G0 . What we need here is a decomposition of into (possibly reducible) representations whose L -constituents have good L-functions. Recall that L is a bre product over k P tK u of L-groups L Gk . The factor L Gk comes in turn with a complex homomorphism into GLpmk , Cq. For any k, we write gk for the image in GLpmk , Cq of a point g P L under the composition L L Gk GLpmk , Cq. Recall also that tK u is the set of orbits of the involution k k _ on the set _ K I J J . We may as well identify tK u with the union of I and J , even though choice of the subset J of K is not canonical. We will then write
1 _ gk_ gk t gk ,

g P L ,

for any index P The decomposition of the representation gives rise to two general families of representations of L . The rst consists of the representations Rkk1 pgqX gk X t gk1 ,
_ Rkk1 Rk_ pk1 q_

k_

_ J .

g P L , k, k 1 P K ,

on complex pmk mk1 q-matrix space. They satisfy the obvious relations and Rkk1 Rk1 k , g P L , k, k 1 P K . and they include also the representations
1 Rkpk1 q_ pgqX gk X gk1 ,

Their L-functions are the standard Rankin-Selberg L-functions Lps, Rkk1 q Lps, k k1 q, where k k1 is the cuspidal automorphic representation of GLpmk q GLpmk1 q attached to the pk, k 1 q-component of . The other family consists of the symmetric square representations Sk pgq S 2 pgk q, k pgq 2 pgk q, g P L , k P K , g P L , k P K . and the skew-symmetric square representations

214

4. THE STANDARD MODEL

They satisfy the relations

_ Sk Sk_

and

_ k_ , k

and also have familiar L-functions. We shall say that a nite dimensional representation of L is standard if it belongs to one of these two general families. The standard representations that are self-dual consist of the following three sets ( Rkk1 : k, k 1 P I , ( Rkk_ : k P K , and ( S k , k : k P I . p We recall here that I is the subset of indices k P K such that Gk equals a classical group SOpmk , Cq or Sppmk , Cq, rather than GLpmk , Cq. Among these self-dual representations, the ones in the second and third sets are all p p orthogonal. Those in the rst set are orthogonal if Gk and Gk1 are of the p p same type, and symplectic if Gk and Gk1 are of opposite type. Observe that a standard representation could be reducible. If it is symplectic, however, it must be irreducible. The global L-function Lps, q attached to a standard representation of L has analytic continuation, with functional equation Lps, q ps, q Lp1 s, _ q. If belongs to the third general family above, it is understood that ps, q ps, , q is the automorphic -factor for the cuspidal automorphic representation of GLpmk q attached to , since it is not known that the arithmetic -factors are the same as the automorphic -factors in this case. However, if is self-dual and symplectic, it is an irreducible Rankin-Selberg representation, for which the two kinds of -factors are known to be equal. In particular, the -factor ps, q is a nite product of local arithmetic -factors, as in the denition of G . Lemma 4.6.1. There is a decomposition (4.6.2) p b b q,

P K ,

of relative to the product (4.6.1), for standard representations , irreducible representations and , and an indexing set K with an involution _ such that
_ _ .

4.6. THE TWO SIGN LEMMAS

215

Proof. It is clear that we can introduce the notion of a standard representation for the product A L SLp2, Cq by copying the denition above for L . One can then establish a decomposition p b q, P I ,

that satises the required properties of (4.6.2), but with standard representations of A in place of the tensor products b . This is a straightforward consequence of the denition (1.4.1) of and the structure of the Lie algebra g, which we leave to the reader. The exercise becomes _ a little more transparent if we take gk to be the second transpose t g 1 , which we can do because the assertion depends only on the inner class of the automorphism. The second step is to observe that any standard representation of A has a decomposition p b q,
PK,

for standard representation of L and irreducible representations of SLp2, Cq. If is a Rankin-Selberg representation Rkk1 for A , is the same Rankin-Selberg representation for L , while ranges over the irreducible constituents of the tensor product (4.6.3) k b k1 representation of SLp2, Cq. If is a standard representation Sk or k for A , alternates between the corresponding two kinds of standard representation of L , and ranges over appropriate constituents of (4.6.3) (with k 1 k). The lemma follows. Corollary 4.6.2. If a constituent of (4.6.2) is diagonal, in the sense that it is of the form Rkk , Rkk_ , Sk or k , the corresponding representation of SLp2, Cq is odd dimensional. Proof. The assertion follows from the proof of the lemma, and the fact that the irreducible constituents of the tensor product (4.6.3) are odd dimensional if k 1 k. We can use the lemma to express the sign character in terms of standard representations. We claim that ` pxq G pxq det psq , s P S,

where x xs is the image of s in S , and the product is over the set of `1 indices P K such that is symplectic, and satises L 2 , 1. The standard representations parametrized by this set are irreducible, so the corresponding representations of S are among the factors in the product (1.5.6) that denes G . There could be other factors in

216

4. THE STANDARD MODEL

(1.5.6), since the diagonal standard representations are reducible, and might have symplectic constituents even though they are not symplectic themselves. However, these supplementary factors occur in pairs and _ , whose contributions to (1.5.6) cancel. Indeed, either the corresponding _ subrepresentation of is not self-dual, in which case also contributes to the product, or is self-dual but not symplectic, since it is reducible, in which case its symplectic factors occur in pairs. The claim follows. We can in fact pare the product further by excluding those factors with ps q 1. These are the factors for which the irreducible representation of SLp2, Cq is odd dimensional, and hence orthogonal. For if the corresponding subrepresentation of is not self-dual, or if it occurs in with even multiplicity, we again obtain pairs of factors that cancel. If is self-dual, and has odd multiplicity, it is orthogonal (since is orthogonal). This implies that the representation of S is symplectic (since is symplectic and k is orthogonal). It thus has determinant 1, and does not contribute to the product. We can therefore write ` (4.6.4) pxs q det psq , s P S,
PK

where K is the subset of indices P K such that (i) `is symplectic, (ii) L 1 , 1, 2 and (iii) ps q 1. We shall work for a time with the restriction r ,1 AdG G

of to the subgroup A L SLp2, Cq of the product (4.6.1). Let M P 2 pM, q be a xed embedding from L to L M attached to , for some xed Levi subgroup M G0 with dual Levi subgroup L M L G0 . Then r AdG G AdG,M M , where AdG,M is the restriction of AdG to L M . We x a decomposition AdG,M M ph b h q
hPH

of this representation, which we assume is compatible with both the decomposition (4.6.2) of , and the root space decomposition x M g g g
p PM

of g with respect to the -split component ` x 0 AM ZpM q x

4.6. THE TWO SIGN LEMMAS

217

x of the center of M . The rst condition tells us that we can write h b h b , h P H , for a surjective mapping h from the indexing set H onto K . In the p second condition, ranges over the roots of pG0 , AM q (including 0), while x p M denotes the usual set of nonzero roots. This condition means that H is a disjoint union of sets H, such that if is the restriction of AdG,M to the subspace g of g, then ph b h q. (4.6.5) M
hPH, Let H, be the preimage of K in H, , and set , g gh ,
hPH,

where gh is the subspace of g on which the irreducible representation h b h h b h , of A acts. The bitorsor S does not generally act on the space g (or for that matter on the larger root space g ). However, the product S A does act on the sum g g, gh ,

hPH

where
H K

H,

is the preimage of in H . There is consequently a S -equivariant mor phism from S into the semisimple algebra EndA p q. We can therefore g write ` (4.6.6) pxs q det s, EndA p q , g s P S. We shall be concerned also with the subspace p ,0 m g of g , and its quotient p g {m
p PM

, g .

The group A of course acts on these two spaces. So does the normalizer N in S of the maximal torus ` p x p T ZpM q {ZpG0 q AM { AM X ZpG0 q x x in S . We write ` p M puq det n, EndA pm q , n P N ,

218

4. THE STANDARD MODEL

` G{M g p puq det n, EndA p {m q, where u is the projection of n onto the quotient N N {T . Then

and

n P N ,

M puq puq pxu q, u P N , where we recall that xu is the image of u in S . Consider the adjoint action of the product N A on the block diagonal p subalgebra m of g . The formula for M puq above matches an earlier M prq of the character M to S r ,u , which u denition of the extension
M M
M

G{M

appears for example in (4.2.11). We can therefore write M puq M prq 1 puq, M u in the notation introduced prior to the statement of Lemma 4.3.1. On the 1 , other hand, the subgroup S of N leaves invariant each of the spaces g . 1 on g and g , , are dual, their contributions to Since the actions of S
G{M

puq cancel. It follows that the value

G{M

pwu q

G{M

puq

depends only on the image wu of u in W . We have shown that our sign character pxq G pxq satises a decomposition formula (4.6.7) 1 puq
G{M

pwu q pxu q,

u P N .

In preparation for the proof of Lemma 4.3.1, we shall also investigate the global normalizing factor
G r pwq r pwq rP pw, M q,

given by (4.2.10). It equals the value at 0 of the quotient (4.6.8) Lp0, P,w M, q p0, P,w M, q1 Lp1, P,w M, q1 , where P,w _1 P |P is contragredient of the adjoint representation of L M w on w1 pP w{w1 pP w X pP , n n n and M, represents the formal twist of the L-homomorphism M by a point in the complex vector space a XpM qF b C. M,C We are now using the Artin notation for L-functions introduced in 1.5, or rather its extension to nite-dimensional representations of the product A , since the automorphic representation M in (4.2.10) will not be an explicit part of the discussion here. It is of course implicit in the parameter M , and we must indeed remember to give the -factor in (4.6.8) its automorphic interpretation.

4.6. THE TWO SIGN LEMMAS

219

As a global normalizing factor, we know that (4.6.8) is a meromorphic function of , which is analytic at 0. This is a consequence of the general theory of Eisenstein series, and the normalization of local intertwining operators from Chapter 2. However, we will need to be more explicit. We write ( p p p P,w P P : w R P p p p x as usual, where P M denotes the set of roots in pP , AM q. Then P,w _ ,
p PP,w p PP,w

since the Killing form on g determines an isomorphism from _ to . From the decomposition (4.6.5) of M into standard representations, we see that the L-function Lps, P,w M, q in (4.6.5) has analytic continuation with functional equation. It follows that (4.6.8) equals Lp1, _ M, q Lp1, P,w M, q1 . P,w We can thus write r pwq lim Lp1, _ M, q Lp1, P,w M, q1 P,w 0 ` lim Lp1, _ M, q Lp1, , M, q1
0 p PP,w

lim lim

p PP,w

` Lp1 pq, _ M q Lp1 ` pq, M q1 ` ppqqa ppqqa ,

p PP,w

where a denotes the order of pole of Lps, M q at s 1. It follows from (4.6.5) that Lps, M q Lps, h b h q.
hPH,

In particular, a where ah ords1 Lps, h b h q. Moreover, since Lps, h b h q can be regarded as a standard automorphic L-function (of type Rkk1 , Sk or k , but attached to unitary, not necessarily cuspidal representations), we have the adjoint relation ` L s, ph b h q_ Lps, h b h q. It follows that ` ah ords1 L s, ph b h q_ .
hPH,

ah ,

220

4. THE STANDARD MODEL

Since _ , the orders a and a are then equal. We conclude that p1qah . (4.6.9) r pwq
p PP,w hPH,

Observe that from the last adjoint relation, the contributions to (4.6.9) of distinct representations ph b h q_ and h b h cancel. The interior product _ in (4.6.9) need therefore be taken over only those h with h h . The question then is to determine the parity of ah , for any h P H, _ with h h . The irreducible representation h maps the diagonal matrix 1 1 diagp|t| 2 , |t| 2 q in SLp2, Cq to the diagonal matrix ` 1 1 1 diag |t| 2 pnh 1q , |t| 2 pnh 3q , . . . , |t| 2 pnh 1q , in GLpnh , Cq, where nh dimph q. It follows from the denitions that (4.6.10) Lps, h b h q
nh i1

` L s ` 1 pnh 2i ` 1q, h . 2

We must therefore describe the order of Lps, h q at any real half integer. As a standard (unitary, cuspidal) automorphic L-function of type Rkk1 , Sk or k , Lps, h q is nonzero whenever Repsq 1 or Repsq 0. Moreover, its only possible poles in this region are at s 1 and s 0. The relevant half integers in (4.6.10) are thus conned to the points 0, 1 , and 1. The question 2 separates into two cases. If dimph q is even, we have only to consider the order of Lps, h q at s 1 . If dimph q is odd, we need to consider its order 2 at the points s 0 and s 1. In both cases, we shall assume that the assertions of Theorem 1.5.3 can be applied inductively to the simple generic constituents of . Suppose rst that dimph q is even. Then h cannot be diagonal, in the sense of Corollary 4.6.2. It must consequently be an irreducible, selfdual Rankin-Selberg representation of type Rkk1 . Its L-function satises the functional equation (4.6.11) Lps, h q ps, h q Lp1 s, h q.

1 It follows that Lps, k q will have even or odd order at s 2 according `1 to whether 2 , h equals `1 1. The contribution of h to (4.6.9) is or `1 therefore just the sign of 2 , h . Observe that if h is orthogonal, Theorem 1.5.3(b) asserts that the sign is `1, and can be ignored. In other words, we can assume that h is symplectic, and therefore that h lies in the subset H, of H, . We thus obtain the contribution to the formula (4.6.9) for r pwq of those h with dimph q even. It equals the product (4.6.12) r pwq p1q|H, | . p PP,w

Suppose next that dimph q is odd. Then h is a self dual representation of L , which is typically diagonal, in the sense of Corollary 4.6.2. Its

4.6. THE TWO SIGN LEMMAS

221

L-function still satises the functional equation (4.6.11), but with the understanding that the -factor ps, h q is given its automorphic interpretation. In particular, the orders of Lps, h q at s 0 and s 1 are equal. It follows from (4.6.10) that the contribution of h to (4.6.9) is just 1 if dimph q is greater than 1, and is equal to the order of Lps, h q at s 1 if h is the trivial one dimensional representation of SLp2, Cq. Theorem 1.5.3(a) tells us this order, in case h is of the form Sk or k . It amounts to the assertion that Lps, h q has a pole at s 1, necessarily of order 1, if and only if the representation h of L contains the trivial representation. The same assertion has long been known in the other case that h is of Rankin-Selberg type Rkk1 . This gives us the contribution of h to (4.6.9) in case dimph q is odd. It equals p1q if the representation h b h of A contains the trivial 1-dimensional representation, and is p`1q otherwise. It is easy to characterize the total contribution to (4.6.9) of those h with dimph q odd. The representation h b h of A acts on a subspace g,h of of g. It contains the trivial representation if and only if the root space g g,h has nontrivial intersection with the Lie algebra of S . This is to say that is a root pS , T q, and that g,h contains the corresponding root p space. As and h range over P,w and H, , we obtain a factor p1q in the product formula (4.6.9) for r pwq for every positive root of pS , T q that is mapped by w to a negative root. The product of these factors is equal to the ordinary sign character s0 pwq on W introduced in 4.1. The character s0 pwq thus gives the contribution from those h with dimph q odd. This completes our analysis of the product (4.6.9) in terms of its even and G odd parts. We have shown that the global normalizing factor r pwq r pwq has a decomposition (4.6.13)
r pwq r pwq s0 pwq,

w P W .

Proof of Lemma 4.3.1. We must establish an identity

G r pwu q 1 puq pxu q s0 puq,

u P N ,

among the four signs, where s0 puq s0 pwu q. The decompositions (4.6.7) and (4.6.13) reduce the problem to showing that the two characters r and on W are equal. We are assuming that and M are as in 4.3. In fact we assume that M is proper in G0 , and therefore subject to our induction hypotheses, since the required assertion is trivial if M G0 . We x w P W . Let T ` be the group generated in S ` by the subset ,w T w of S , and let be the representation of this group on the space
V EndA p {m q. g p G{M

We can then write

G{M

` pwq det pwq ,

since the determinant is independent of the preimage of w in T ` . ,w

222

4. THE STANDARD MODEL

The group T ` is an extension by T of the cyclic group xwy generated ,w by w. Its representation theory is well understood. For example, there is a surjective mapping o, from the set of irreducible representations of T ` onto the set of xwy-orbits o of characters t t on T , such that ,w ` tr ptq t , t P T .
Po

It is a consequence of our denitions that ` tr ptq |H, | t ,

p PM

t P T .

This leads to a decomposition h ,

o hPH ,

p where o ranges over the xwy-orbits of roots P M , and h is an irreducible ` representation of T ,w that maps to o . We can therefore write ` G{M (4.6.14) pwq det h pwq .
o hPH ,

An orbit o is said to be symmetric if it contains the root pq. If is not symmetric, o and o are two distinct orbits of equal order. They give rise to pairwise contragredient sets of representations h , whose contributions to the formula (4.6.14) cancel. We can therefore restrict the outer product in (4.6.14) to orbits o that are symmetric. Other factors in the product (4.6.14) can also pair o. Recall that any index h P H, gives rise to an irreducible representation h bh of A . The `1 self dual representations h and h are both symplectic, since 2 , h 1 and h ps q 1, so their tensor product is orthogonal. On the other hand, the representation ph b h b h q
h

of A on g is invariant under the Killing form, and hence orthogonal. This does not mean that the individual representations h need to be orthogonal. However, the representations h that are not orthogonal will occur in pairs, whose contributions to (4.6.14) cancel. We can therefore take the interior product in (4.6.14) over the subset H,,o of indices h P H, such that h is orthogonal. Consider an orthogonal representation h ,
h P H,,o .

T` ,w

Elements of the Weyl group of a complex orthogonal group can be represented by permutation matrices. One sees from this that the image of T ` ,w

4.6. THE TWO SIGN LEMMAS

223

under h is the semidirect product of the image of T with the (cyclic) quotient of xwy of order |o |. It then follows that the restriction of h to xwy is isomorphic to the permutation representation of xwy on o . In particular, h pwq is a well dened operator, whose eigenvalues are the set of roots of 1 of order |o |. We are assuming that the orbit o is symmetric. It therefore has even order, from which it follows that ` det h pwq 1. The orthogonal representations thus each contribute a sign to the product (4.6.14). The formula reduces to G{M (4.6.15) pwq p1q|H,,o | .
o symmetric The analysis of the product formula (4.6.12) for r pwq is slightly dierent. Since |H,w | |H, |,

the factor of in the product depends only on the orbit o . If o is symmetric, we see after a moments thought that the integer p o X P,w equals the number of sign changes in the sequence p, w, . . . , q p p (relative to the order on M dened P ). Since this number is odd, the contribtution to (4.6.12) of all the roots in a symmetric orbit is the same as the contribution of any positive root in the orbit. If o is not symmetric, o and o are disjoint, and the integer p p o X P,w ` o X P,w equals the number of sign changes in the sequence p, w, . . . , q. Since the number is even in this case, the nonsymmetric orbits contribute nothing to (4.6.12). The product (4.6.12) therefore reduces to p1q|H, | .
o symmetric As we have noted, indices h in the complement of H,,o in H, occur in pairs, and consequently contribute nothing to the product. The right hand side of (4.6.12) therefore reduces to the right hand side of (4.6.15). This establishes the required identity

G{M

pwq r pwq,

and completes the proof of Lemma 4.3.1.

224

4. THE STANDARD MODEL

We observe that both assertions of Theorem 1.5.3 were needed to prove Lemma 4.3.1. They were each invoked in the preamble that led to the decomposition (4.6.13) of r pwq. For example, we used the assertion (b) on orthogonal -factors to obtain an exponent in (4.6.12) that included only indices h with h symplectic. This reduction was essential. Had we been forced also to include orthogonal Rankin-Selberg representations h (with a corresponding enlargement of the exponent |H, | in (4.6.12)), our product (4.6.12) would then not have matched the formula for we obtained in the proof of the lemma. Our application of both assertions (a) and (b) is of course predicated on our induction hypothesis. In the case of (b), we can assume that the assertion holds for any pair of simple generic parameters (4.6.16) r i P sim pGi q, r Gi P Epmi q, i 1, 2,
G{M

with m1 ` m2 N . We can actually manage with a weaker assumption on Theorem 1.5.3(b). Our proof of Lemma 4.3.1 for requires only that Theorem 1.5.3(b) apply to any relevant pair i i b i , i 1, 2, of distinct simple constituents of . In other words, when Theorem 1.5.3(b) implies that the contribution to (4.6.9) of a pair is 1, we require that the contribution actually be equal to 1. The pair p1 , 2 q gives rise to a RankinSelberg representation 1 2 of A , and corresponds to a root in (4.6.9). It is relevant to Theorem 1.5.3(b) only if the Rankin-Selberg representation 1 2 of L is orthogonal, and the tensor product 1 b 2 decomposes into irreducible representations SLp2, Cq of even dimension. In fact, we see from (4.6.10) and the proof of Lemma 4.6.1 that 1 b 2 must decompose into an odd number of such representations for the application of Theorem 1.5.3(b) to (4.6.9) to be nontrivial. We shall use this weaker form of the induction hypothesis in the next chapter. Proof of Lemma 4.4.1. This lemma will be easier to prove than the last. We have to show that G1 ps1 q ps xs q,
1

s P S ,ss ,

where pG1 , 1 q is the preimage of p, sq under the correspondence (1.4.11). We shall follow the general argument set out in [A9, p. 5153]. We begin with the formula ` pxs q G pxs q det psq
PK

obtained (4.6.4) from any decomposition p b b q

PK

4.6. THE TWO SIGN LEMMAS

225

that satises the conditions of Lemma 4.6.1. We could in fact weaken these conditions. For example, a decomposition in which the representations of S are allowed to be reducible gives the same formula for G pxs q. Notice that the point s belongs both the group SLp2, Cq and the cen tralizer S . Therefore ps q ps q 1,
for any in the subset K of K . It follows that G ps q p1qdimp q .
PK

We will of course be free to apply this formula if G is replaced by the endoscopic group G1 . Suppose that s, G1 and 1 are as in the putative identity. The dual p p group G1 is then the connected centralizer in G0 of the point s1 s. Its 1 is just the kernel of Adpsq in g. The parameter 1 for G1 is Lie algebra g dened by the L-embedding r 1 : A
L

G1 ,

r which in turn is obtained from G by restriction of the codomain to the L G1 of L G0 . It follows that 1 , the restriction of 1 to A , L-subgroup ,1 is simply the mapping given by restricting the representation ,1 to the subspace g1 of g. For any P K , let s be the representation of S 1 on the p`1q-eigenspace of psq. If K1 denotes the set of P K with s 0, we can then write ps b b q. 1
PK1

This decomposition satises the conditions of Lemma 4.6.1 (weakened to the extent that s could be reducible). It follows that 1 s G1 ps1 q p1qdimp q ,
PK1

where K1 is the intersection of K1 with K . The original formula (4.6.4) tells us that G pxs q equals the product of all the eigenvalues, counting multiplicities, of all of the operators ( psq : P K .

But the contragredient _ can be treated as an involution on the set of representations with P K . It follows that if is an eigenvalue not equal to p`1q or p1q, its inverse 1 is another eigenvalue, but with the same total multiplicity. Therefore G pxs q p1qmp1q ,

226

4. THE STANDARD MODEL

where mp1q is the total multiplicity of the eigenvalue p1q. By the same token, the number ` dimp q dimps q mp1q,
PK

being the sum of multiplicities of eigenvalues distinct from p1q, is an even integer. It follows that G ps xs q G ps q G pxs q p1qdimp q p1qmp1q
PK

p1qmp1q G1 ps1 q,
1

PK

p1qpdimp qdimp qq G1 ps1 q

as required. This completes the proof of Lemma 4.4.1.

4.7. On the global theorems The remaining two sections of the chapter represent a digression. The purpose of this section is to relate the global theorems we have stated with the trace formula. The discussion amounts to a series of informal remarks, which will help us to understand the formal proofs later on. In the next section, we shall sketch how our earlier comparison of spectral and endoscopic terms extends to general groups. A reader wishing to go on with the main argument, which we left o at the end of 4.5, can skip these sections and proceed directly to Chapter 5. This section is intended as more technical motivation for the global theorems stated in Chapter 1. Among other things, we shall see how the theorems are consequences of the stable multiplicity formula of Theorem 4.1.2. This will be for guidance only, since in the end we will have to establish all of the theorems together. However, it might explain how the multiplicity formula of Theorem 1.5.2, and the properties of poles and signs in Theorem 1.5.3, are in some sense implicit in the trace formula. The remarks of this section can also be regarded as an illustration of the general comparison r in the simplest situation the case of a parameter P ell pN q that was excluded earlier, and which is ultimately the hardest to prove. For this exercise, we shall assume whatever we need of the local theorems we have stated. We shall also suppose that the global seed Theorem 1.4.1 and its complement Theorem 1.4.2 are both valid, and that the resulting global r parameters for groups G P Esim pN q have the rough spectral properties that motivated the initial remarks in 1.2. Specically, we assume that there is no contribution to the discrete spectrum of G from any parameter that

4.7. ON THE GLOBAL THEOREMS

227

r does not lie in 2 pGq. We shall treat the ne spectral assertions of Theorem 1.5.2 and parts (a) and (b) of Theorem 1.5.3 in reverse order. We begin with a general global parameter (4.7.1) 1 r , r i P sim pNi q,

r in ell pN q. This was implicitly excluded from earlier sections, specically in the comparisons of 4.3 and 4.4 that led to Proposition 4.5.1. However, with our heuristic assumptions above, the comparisons are valid for , and in fact are quite transparent. They might give us a better overall view if we r retrace some of the steps. We shall begin with G GpN q and f equal to a r r function f P HpN q. Suppose for a moment that r 1. Then 1 b
0 is simple, and has a contribution Rdisc, to the relative discrete spectrum of G0 GLpN q. We note that the sign character G is trivial in this case, since as in Corollary 4.6.2, the tensor product of the unipotent part with itself is a sum of irreducible odd dimensional representations. We G0 G also recall that Rdisc, , the canonical extension of Rdisc, , is a product of the corresponding local extensions, according to the theory of local and global Whittaker models for GLpN q. Therefore ` G 1 G G 1 Idisc, pf q rdisc, pf q 2 tr Rdisc, pf q 2 fG pq.

r This can be regarded as an elementary analogue for G GpN q of either Theorem 1.5.2 or Theorem 4.2.2. We have used it, at least implicitly, in our earlier discussion in 4.3 and 4.4. G If r 1, Idisc, pf q equals the dierence (4.3.1) with which we began the comparison in 4.3, since does not contribute to the discrete spectrum of G r G GpN q. Therefore Idisc, pf q equals the expression (4.3.2) we obtained for the dierence. To analyze it, we take M to be the standard Levi subgroup of G0 GLpN q corresponding to the partition pN1 , . . . , Nr q. Then is the image of a parameter M P 2 pM, q. There is one element w in the set W . This element stabilizes M , and induces the standard outer automorphism on each of the factors GLpNi q of M . It follows that ` detpw 1q G 1 1 r . a 2
M

The other coecients in (4.3.2) satisfy m |W | |SM | 1, according to their denitions, and can be ignored. The bre N pwq of w in N consists of one element u. We obtain ` r fG p, uq tr RP pw, M , M q IP pM , f q fG pq,

228

4. THE STANDARD MODEL

again by the theory of Whittaker models for GLpN q, applied this time to both M and G. The expression (4.3.2) therefore reduces to ` 1 r r r G GpN q, f P HpGq. 2 rP pw, M q fG pq, We have shown that
N r r Idisc, pf q Idisc, pf q

` 1 r
2

r rP pw, M q fN pq, r

r r f P HpN q,

r r where is as in (4.7.1), P is a parabolic subgroup of G0 pN q with Levi compo` r nent M M , w is the unique element in the Weyl set W N W GpN q, M

r and rP pw, M q is the global normalizing factor for GpN q. We also have the r stable decomposition N r r pG rpN, Gq Sdisc, pf G q. Idisc, pf q
r GPEell pN q G Our heuristic assumptions, together with (4.4.12), tell us that Sdisc, vanr r ishes if does not belong to the subset 2 pGq of ell pN q. We can therefore r r x G to be the unique element in Eell pN q such that G does belong to 2 pGq. r The local Theorem 2.2.1, which we are also assuming, tells us that fN pq r equals f G pq. It follows that ` r r r pG r (4.7.2) Sdisc, pf G q rpN, Gq1 1 rP pw, M q f G pq, f P HpN q. r 2

Consider now the case that r 2 in (4.7.1), and that 1 and 2 are of opposite type. In other words, one is orthogonal and the other is symplectic. r The datum G with P 2 pGq is then of composite form G G1 G 2 , r Gi P Esim pNi q.

The stable multiplicity formula stated in Theorem 4.1.2 is global, and is not included in our heuristic assumptions. However, for our simple parameters r i P sim pGi q, the formula follows inductively from our discussion of the case r 1 above, with another appeal to Corollary 4.6.2 to conrm that the signs Gi pi q equal 1. The stable multiplicity formula therefore holds for the simple parameter of the product G. It takes the reduced form
G Sdisc, pf q m f G pq,

r f P HpGq,

since the coecients |S |1 , pS 0 q and G pq are all trivial in this case. Combining this with (4.7.2), we obtain ` 2 r r r r m f G pq rpN, Gq1 1 rP pw, M q f G pq, f P HpN q. r 2 It is easy to see that rpN, Gq1 ` 1 2
2

`1

2 |ZpGq

p |1 |OutN pGq|1 r

1 ` 1 2
2

r |OutN pGq| m ,

4.7. ON THE GLOBAL THEOREMS

229

p given the formula for rpN, Gq at the end of 3.2, the fact that |ZpGq | 2 for the composite G, and the denition in 1.5 of m . The resulting cancellation then gives r r r r f P HpN q. f G pq rP pw, M q f G pq, r We thus obtain an identity (4.7.3) rP pw, M q 1, r

in our case of r 2, and 1 and 2 of opposite type. Actually, we should be a little careful in the last step of the justication of (4.7.3), since the mapping r r r r f f G f 1 f 2 , r r f i P HpGi q,

r r does not take HpN q onto HpGq. However, an inductive application of Theorem 1.5.2 to G1 and G2 tells us that the linear form f G pq, r f P HpGq,

is a sum of characters. The conclusion (4.7.3) then follows as a simple special case of Proposition 3.5.1. The identity (4.7.3) leads to a proof of Theorem 1.5.3(b). To see this, suppose that r i P sim pmi q, i 1, 2, are simple generic parameters of the same type, either both orthogonal or both symplectic. Set 1 1 and 2 2 b 2 , where 2 is the irreducible two-dimensional representation of SLp2, Cq. The sum of 1 and 2 then satises the conditions above, with N m1 ` 2m2 . We calculated the global normalizing factor rP pw, M q in 4.6. In the case r r here, P is the standard maximal parabolic subgroup of GLpN q of type pN1 , N2 q pm1 , 2m2 q. The representation P ,w M of A in (4.6.8) acts r on the Lie algebra nP , and is the exterior tensor product of the orthogonal r Rankin-Selberg representation of L attached to 1 2 with the representation 2 of SLp2, Cq. It follows from a special case of (4.6.9), (4.6.10), and (4.6.11), specically the description below (4.6.11) of the contribution of the representation h to (4.6.9), that ` ` rP pw, M q 1 , q 1 , 1 2 . r 2 2 It then follows from (4.7.3) that `1 (4.7.4) 2 , 1 2 q 1. This is the assertion (b) of Theorem 1.5.3, with 1 2 and pm1 , m2 q in place of 1 2 and pN1 , N2 q. The assertion (a) of Theorem 1.5.3 is harder, even with the heuristic assumptions we have allowed ourselves. Its formal proof will be one of our

230

4. THE STANDARD MODEL

long term goals. We shall make only a few general remarks here, just to give a sense of the direction of the later proof. Theorem 1.5.3(a) concerns the case r 1. Suppose that ber longs to the set sim pGq of simple generic parameters, for a simple datum r G P Esim pN q, as in the assertion. The statement was of course motivated by the expected properties of global L-functions. If is orthogonal, according to the criterion of Theorem 1.4.1, the group L L is an extension p of WF by the complex orthogonal group G . In this case, the representar tion S 2 G of L contains the trivial representation, and the L-function 2 q would be expected to have a pole at s 1. Similarly, if is Lps, , S symplectic, the L-function Lps, , 2 q would be expected to have a pole at s 1. We will eventually have to establish these criteria by harmonic analysis and the trace formula. The technique relies on the introduction of a supplementary parameter ` r in pN` q, with N` 2N . r The global parameter ` does not lie in the elliptic set ell pN` q, since r it has a repeated factor. Let G` P Esim pN` q be the twisted endoscopic r p p datum such that ` lies in pG` q, and such that G` and G are of the same type, either both orthogonal or both symplectic. The quadratic characters ` and G` attached to ` and G` are both equal to 1. We r therefore have the companion datum G_ in Esim pN` q, and ` is also con` _ q. The groups G and G_ share a maxr tained in its parameter set pG` ` ` imal Levi subgroup M` . It is isomorphic to GLpN q, and comes with a canonical element `,M` P pM` , ` q (which we may as well identify with ` ). Let ` pG` q and _ pG_ q be the adjoint representations of ` ` x M` GLpN, Cq on the Lie algebras n` and n_ of the unipotent radicals ` p p of corresponding maximal parabolic subgroups of G` and G_ . (We caution ` _ does not denote the contragredient of here.) Then _ ourselves that ` ` ` p p is of the same type as G, namely S 2 if G is orthogonal and 2 if G is symplectic, while ` is of type opposite to G. On the other hand, the centralizer groups satisfy _ S` S` pG_ q Spp2, Cq ` and # Op2, Cq, if N is even, S` S` pG` q SOp2, Cq, if N is odd. _ intersects n_ in a 1-dimensional space, It follows that the Lie algebra of S` ` while the Lie algebra of S` intersects n` only at t0u. These properties are direct consequences of the structure of the Lie alp p gebras of G` and G_ . They reduce the problem to the behaviour of global ` normalizing factors. To be more specic, we have rst to note that the heuristic assumptions of this section preclude a contribution from ` to the

4.7. ON THE GLOBAL THEOREMS

231

discrete spectrum of either G` or G_ . The analogue for either of these ` groups of the expression (4.3.2) is therefore simply the discrete part of the corresponding trace formula, represented by the rst term in the dierence (4.3.1). Consider then the expression for
` _ _ _ Idisc,` pf` q Idisc,` pf` q,

G_

_ r f` P HpG_ q, `

given by the analogue of (4.3.2). The double sum in (4.3.2) collapses to one simple summand in this case, of which the various terms are easy to describe explicitly. In particular, by (4.6.9) and (4.6.10), the analogue of the global normalizing factor rP pw, M q reduces to the sign p1qa` ,
_

a_ ords1 Lps, _ ` q. ` `

This sign is supposed to equal p1q, since it comes from the L-function that is expected to have a simple pole. One can also describe all of the terms in the expression for N r r r f` P HpN` q, I ` pf` q,
disc,`

given by its analogue of (4.3.2). In this case, the relevant global normalizing _ r factor does in fact equal p1q. We choose f` and f` so that they transfer r to the same function in SpG_ q. It then follows easily from the global inter` _ (which reduces to the local intertwining relation we twining relation for G` r are assuming) and its analogue for GpN` q (which we already know) that the _ r analogues for f` and f` of the linear form fG p, uq in (4.3.2) are equal. We then see immediately that the two expressions are equal up to a multiple that is positive if and only if Lps, _ ` q does have a pole (necessarily ` simple) at s 1. A little more thought reveals that the absolute value of this multiple is actually equal to the coecient rpN` , G_ q. The required ` assertion (a) is therefore equivalent to the undetermined sign in an identity
N` _ _ r Idisc,` pf` q rpN` , G_ q Idisc,` pf` q `

being positive. How might one use the existence of this identity to determine the sign? We can write _ _ _ _ Idisc,` pf` q Sdisc,` pf` q, since the only elliptic endoscopic datum for G_ through which ` factors ` in G_ itself. For similar reasons, we can write `
N` r rG rG p p_ (4.7.5) Idisc,` pf` q rpN` , G` q Sdisc,` pf` ` q ` rpN` , G_ q Sdisc,` pf` ` q. `
_

Now the factor pS 0 ` q in the putative stable multiplicity formula for Sdisc,` vanishes, since the group S 0 ` S ` pG` q0 GLp1, Cq

232

4. THE STANDARD MODEL

has innite center. We therefore expect that Sdisc,` vanishes (as a linear r form on HpG` q). If it does, we see that
_ _ r Idisc,` pf` q rpN` , G_ q Idisc,` pf` q, ` _ r r given that f` and f` transfer to the same function in SpG_ q. In other ` words, the identity above will indeed have the required positive sign. Our heuristic assumptions do not include the stable multiplicity formula of Theorem 4.1.2. However, it is possible to establish that Sdisc,` vanishes from the assumptions we do have, and the expression for
` Idisc,` pf` q Idisc,` pf` q,

r f` P HpG` q,

given by the analogue of (4.3.2). If N is odd, the argument is easy. In this case, the group S ` S 0 ` has r disc pG` q of pG` q r innite center, and ` does not belong to the subset dened in 4.1, from which we see that Idisc,` pf` q vanishes. Since G G is not an endoscopic group for G` in this case (it is actually a twisted endoscopic group), ` does not factor through any proper endoscopic datum for G` . It follows that Sdisc,` pf` q Idisc,` pf` q 0, as required. If N is even, the heuristic argument is more complicated. In this case, the expression for Idisc,` pf` q contains a sign p1qa` , a` ords1 Lps, ` ` q, which equals p1q if Theorem 1.5.3(a) if false. One combines this informa_ _ _ tion, and the fact that the corresponding sign p1qa` for Idisc,` pf` q then equals `1, with the endoscopic decomposition p pf 1 q Idisc, pf` q Sdisc, pf` q ` pG` , G1 q S 1
` `

disc,`

r for G` and G1 G G, and its analogue (4.7.5) for GpN` q, G` and G_ . ` ` Since 1 G G Sdisc,` Sdisc, b Sdisc, , r an expression we can evaluate easily given that P sim pGq is simple, we nd that the rst decomposition gives a formula for Sdisc,` pf` q. An analysis of the three terms in the second decomposition then leads to a contradiction, namely that (4.7.5) cannot hold if the two signs are as above. In other words, Theorem 1.5.3(a) must be valid. These last remarks have been quite sketchy. They are intended to serve as a heuristic argument in support of assertion (a) of Theorem 1.5.3. There is no point in being more precise now, since we will later have to revisit the argument in more detail and greater generality. As we will see, the formal proof of the assertion will be much more dicult without our heuristic assumption that ` contributes nothing to the discrete spectra of G` and G_ . `

4.7. ON THE GLOBAL THEOREMS

233

Having discussed the two assertions of Theorem 1.5.3, we take as in (4.7.1), and assume that both assertions are valid (where applicable) for the simple constituents i of . This allows us to apply the two sign lemmas r proved in 4.6 to . We shall use the lemmas, both for G GpN q and r G P Esim pN q, to derive the stable multiplicity formula of Theorem 4.1.2. r With xed, we take G to be the unique datum in Eell pN q such that r r lies in the subset 2 pGq of ell pN q. We of course want to assume that G is simple, the case we are studying. The rst step is to apply Lemma 4.3.1, r with GpN q in place of G, to the global normalizing factor in (4.7.2). We obtain rN rN rP pw, M q r pwq r pwq 1 puq r r N pxu q s0 puq N pxu q, r where u is the unique element in the bre N pwq. We are using the fact that 1 puq M puq 1, following an earlier remark in the special case r 1, and that s0 puq s0 pwu q 1, r r r since the group S pN q0 is abelian. The point x xu is the unique element r in the set S pN q. It represents a point s P S pN q such that the pair pG, q is the preimage of p, sq under the correspondence (1.4.11). The next step ` r is to apply Lemma 4.4.1, with GpN q, G in place of pG, G1 q, to the sign N pxq
r GpN q

pxq. We obtain N pxu q N pxq N ps xq G ps q G pq,

r0 r since the point s lies in the connected group S pN q S pN q. The identity (4.7.2) then takes the form ` 1 r r r r r pG Sdisc, pf G q rpN, Gq1 2 G pq f G pq, f P HpN q. The remaining terms in the identity are easily treated. Given the formula p for rpN, Gq at the end of 3.2, and the fact that ZpGq t1u in the case here that G is simple, we obtain ` 1 p r r rpN, Gq1 1 |ZpGq |1 |OutN pGq|1 2 |OutN pGq|. 2 ` 1 r1 can then be related to the nite group The resulting power 2 p S pGq S S S {ZpGq attached to G and . One nds that r |S |1 |OutN pG, q| ` 1 r1
2

234

4. THE STANDARD MODEL

r r with OutN pG, q being stabilizer of in OutN pGq, after rst noting that the product p r |ZpGq | |OutN pGq| p equals 1, 2 or 4 in the respective cases that G equals SOpN, Cq (N odd), SppN, Cq or SOpN, Cq (N even). We then note that the integer m in the putative multiplicity formula satises r r m |OutN pGq| |OutN pG, q|1 , according to the denition in 1.5. Finally, by the local Theorem 2.2.1, which we are taking for granted, we can write r r fN pq f G pq. It follows that r r pG Sdisc, pf G q m |S |1 G pq f G pq, This in turn can be written as (4.7.6)
G Sdisc, pf q m |S |1 G pq f G pq,

r f P HpN q.

r f P HpGq,

r r r r since the mapping f f G takes HpN q onto SpGq. The identity (4.7.6) is the stable multiplicity formula of Theorem 4.1.2 r for the parameter P 2 pGq, since the constant pS 0 q equals 1 for the trivial group S 0 . We can therefore apply its analogues from Corollary 4.1.3 to the data G1 in the endoscopic expansion r p1 Idisc, pf q f P HpGq, pG, G1 q Sdisc, pf 1 q,
G1 PEell pGq

for G. The right hand side becomes a double sum (4.7.7) pG, G1 q |S1 |1 1 p 1 q f 1 p 1 q,
G1 1

over G1 P Eell pGq and 1 P pG1 , q. The analysis of this expression then follows the general discussion in 4.4 that led to Lemma 4.4.2. It will be useful to revisit briey the argument in the simpler case here, especially r since the case P 2 pGq was formally ruled out of the earlier discussion. (See [K2, 12] and [K3, p. 191].) p The outer sum in (4.7.7) can be taken over G-orbits of elliptic endoscopic 1 , in which the corresponding element s1 P G is treated as a ZpGq p p data G p coset. The inner sum is over the set of G1 -orbits of homomorphisms 1 L G1 that map to . We can replace it by the coarser set of from A to p orbits under the stabilizer AutG pG1 q of G1 in G, so long as we multiply the summand by the quotient |OutG pG1 q| |OutG pG1 , 1 q|1 .

4.7. ON THE GLOBAL THEOREMS

235

p Again, OutG pG1 , 1 q is the stabilizer of 1 (as a G1 -orbit) in the nite group p OutG pG1 q AutG pG1 q{G1 . The rst coecient in (4.7.7) equals p pG, G1 q |ZpG1 q |1 |OutG pG1 q|1 . We can therefore write (4.7.7) as a double sum p (4.7.8) |S1 |1 |ZpG1 q |1 |OutG pG1 , 1 q|1 1 p 1 q f 1 p 1 q
pG1 , 1 q

p over G-orbits of pairs. The main step is to apply the correspondence (1.4.11) to the indices pG1 , 1 q. We thereby transform the sum in (4.7.8) to a double sum over p G-orbits of pairs pG , sG q, where G is an L-homomorphism from A to L G that maps to , and s is an element in the centralizer S G G . The set p of G-orbits of G is the nite set pG, q of order m . We can therefore remove the sum over G , identifying each G with , if we multiply the summand by m . The sum over sG then becomes a sum over elements s in the underlying nite abelian group S S . It remains to express the summand in (4.7.8) in terms of and s. It follows easily from the denitions that
` ` p OutG pG1 , 1 q S,s {S,s X G1

p S ` {S ` X G1 , ,s ,s where ( ` S,s g P S : gsg 1 s ,

` p S ` S,s {ZpGq Centps, S q, ,s

s P S,

and p p p G1 G1 {ZpGq . Since S is abelian, we also have S ` S S , ,s and p p S ` X G1 S pG1 q{ZpGq . ,s The quotient of this second group by the subgroup p p p ZpG1 q ZpG1 q {ZpGq equals p S pG1 q{ZpG1 q S pG1 q S pG1 q S1 .

236

4. THE STANDARD MODEL

We thus obtain a reduction p |S1 |1 |ZpG1 q |1 |OutG pG1 , 1 q|1 |S |1 for the product of the rst three coecients in (4.7.8). Applying Lemma 4.4.1 to the fourth coecient, we obtain 1 p 1 q G1 ps1 q G ps sq ps sq. Finally, we can write the fth term as f 1 p 1 q xs s, y fG pq,
r P
1

by applying Theorem 2.2.1 to each completion v of . We have now expressed all ve factors in the summand of (4.7.8) in terms of and s. We can therefore replace the double sum over G1 and s1 by the simple sum over s P S described above. We obtain Idisc, pf q m |S |1 ps sq xs s, y fG pq
sPS P r

m |S |1
r P

xPS

pxq xx, y fG pq

m pq fG pq,

where # (4.7.9) m pq m , 0, if x , y 1 , otherwise.

On the other hand, we can write ` G Idisc, pf q tr Rdisc, pf q ,

r f P HpGq.

This is because does not contribute to any of the terms with M G in the expansion (4.1.1) of Idisc, pf q. We have thus established the multiplicity formula ` r (4.7.10) tr RG pf q m pq fG pq, f P HpGq,
disc, r P

postulated by Theorem 1.5.2, under the heuristic assumptions we have made. In so doing, we have shown that the sign character pxq is forced on us by the known dependence of global intertwining operators on L-functions for GLpN q. It has perhaps been dicult to follow the stream of informal remarks of this section. Let us recapitulate what we have done. After taking on some plausible assumptions, we have sketched a proof of four main global assertions the triviality of orthogonal -factors (Theorem 1.5.3(b)), the

4.8. REMARKS ON GENERAL GROUPS

237

existence of poles of L-functions (Theorem 1.5.3(a)), the stable multiplicr ity formula for parameters P 2 pGq (Theorem 4.1.2), and the spectral r multiplicity formula for parameters P 2 pGq (Theorem 1.5.2). The stable multiplicity formula is at the heart of things. Had we taken it as an assumption (albeit a less plausible one), we could easily have established the other three assertions. In particular, under this condition, our justication of the spectral multiplicity formula above can be regarded as a formal proof. We shall state this as a lemma for future use. Lemma 4.7.1. Assume that the stable multiplicity formula (4.7.6) holds r r for any N 1, G P Esim pN q and P 2 pGq. Then the spectral multiplicity formula (4.7.10) also holds for any N , G and . Our proof of the lemma is represented by the elementary argument above, beginning in the paragraph following (4.7.6). The condition on the stable multiplicity formula is to be interpreted broadly. It implicitly includes r Theorems 1.4.1, 1.4.2 and 2.2.1 (in the denition of the set 2 pGq and the G pq), and also includes the condition that there be no conlinear form f r tribution to the discrete spectrum of G from parameters not in 2 pGq (by deduction from (4.7.6)). This last global condition is among the most dicult we will have to prove. It is one of the reasons that the simple arguments of this section have to be replaced by the rigorous proofs that will occupy the rest of the volume. We will appeal to Lemma 6.7.1 several times later, in each case at the end of an extended argument that establishes the other conditions. 4.8. Remarks on general groups We shall conclude with a brief description of how some of the earlier discussion of the chapter might be extended. A general heuristic comparison of spectral and endoscopic terms was the topic of [A9]. This followed the special case of the generic discrete spectrum treated in [K1, 1012]. The article [A9] is not so easy to read, for among other reasons, the fact that it was written without the benet of the volume [KS] of Kottwitz and Shelstad. We are in a stronger position here. In addition to [KS], we now have the results of the earlier sections of this chapter on which to base our general remarks. The special cases treated in these sections, namely the groups r r G P Esim pN q and the twisted group G GpN q that are the main focus of the volume, are clearly simpler. For a start, they come with a well dened substitute L for the global Langlands group. In addition, the locally trivial 1-cocycles that complicate the general theory are not present. There is also the further simplication that the groups S are all abelian. However, we can still sketch how some of our arguments extend to the general case, referring as necessary to the relevant sections of [A9]. For this discussion, we take G to be a general triplet pG0 , , q over the global eld F . Following [KS, 2.1] (but with the notation of 3.2), we form

238

4. THE STANDARD MODEL

the dual set p p p p G G0 G0 L , p which is a bitorsor under the dual group G0 . We also form the L-set
L

p G G WF

G0 L ,

which is an L-bitorsor under the L-group L G0 . Recall that L L is an p p extension of the automorphism from G0 to L G0 , which depends on . It p must be used in G when we form its semidirect product with WF . Recall also that an endoscopic datum G1 for G represents a 4-tuple pG1 , G 1 , s1 , 1 q that satises (analogues of) the conditions (2.1.1)(2.1.4) of [KS]. We are p taking s1 here to be a semisimple element in G. We observe that in the notation here, the condition (2.1.4a) of [KS] asserts that s1 1 pg 1 q a1 pw1 q 1 pg 1 q s1 , g1 P G 1,

where w1 is the image of g 1 in WF , and a1 pwq is a locally trivial 1-cocycle p from WF to ZpG0 q. We must of course assume that we have a suitable substitute for the Langlands group LF . We take it to be an object in the category of locally compact topological groups. We suppose that we have at our disposal an extension L WF , F
with compact connected kernel KF , that takes the place of LF . We could of is the actual Langlands group L . In practice, course assume simply that LF F L will be a larger group, or rather a group that contains some quotient of F LF over WF that is relevant to the endoscopic study of G. For example, we will suggest later in 8.5 how to replace all of our complex groups L with one locally compact group L . F We suppose that along with L , we are also given a distinguished family F pGq of L-homomorphisms bdd

: L F

G0 .

p These objects should have bounded image in G0 , and be determined up to the general equivalence relation introduced in [K1]. Namely, two Lhomomorphisms and 1 are equivalent if (4.8.1) 1 puq zpuq g puq g 1 , u P L , F

p where g belongs to G0 , and z is the pullback of a locally trivial 1-cocycle p 0 q. For any P pGq, we set S pG0 q equal to the group from WF to ZpG bdd p of elements s P G0 such that (4.8.2) spuq zpuq puq s, u P L , F

where z is again the pullback of a locally trivial 1-cocycle. Our guiding intuition is to be based on the following expectation: there should be a

4.8. REMARKS ON GENERAL GROUPS

239

canonical L-homomorphism LF L F from the actual Langlands group to L , determined up to conjugacy by the F kernel KF , such that the restriction mapping from pGq to bdd pGq is a bdd bijection that preserves the centralizers S pG0 q. Given the set pG0 q, we dene pGq to be the family of equivalence bdd classes of L-homomorphisms : A F
L

G0 ,

A L SU p2q, F F

whose restriction to L lies in pG0 q, and for which the centralizer set F bdd S S pGq is not empty. Equivalence is dened here by the evident analogue of (4.8.1). The set S pGq is dened as in (4.8.2), but with in place of , and s ranging p p over elements in G rather than G0 . The elements in pGq are then to be treated as the global parameters for G. We would actually need to assume more broadly that L comes also F with a distinguished family pHq of special parameters for any group H bdd that arises in the endoscopic study of G. For example, H could be a Levi subgroup M of G0 , or a twisted endoscopic group G1 for G, or perhaps an r auxiliary extension G1 of G1 . In fact, we would typically permit G itself to vary over some suitable family that allows for induction arguments. This of course is exactly what we have done in the special cases we are treating in this volume. In general, we assume that the sets pHq are functorial bdd with respect to endoscopic embeddings among the associated L-groups. We then dene the general families of parameters pHq and the centralizers S pHq for H, as we did above for G. Finally, we suppose that L comes with local homomorphisms F LFv L F over the corresponding Weil groups, which are dened as usual up to conjugation. We then obtain localization mappings v , P pGq,

from pGq to the local parameter sets ` pGv q. We can of course form the subset pG, q of global parameters in pGq that are attached to a given central character datum pXG , q for G. The localizations then provide a mapping ` tpq, cpq , P pG, q, as in the special case of 3.3. We recall that tpq is the nonnegative real number given by the norm of the imaginary part of the archimedean innitesimal character of , and that c P CA pG, q is the equivalence class of families of conjugacy classes in L G attached to unramied places v of F .

240

4. THE STANDARD MODEL

For any such pair pt, cq, we dene pG, t, cq to be the set of P pG, q such that ` tpq, cpq pt, cq. r If c belongs instead to some quotient CA pG, q of CA pG, q, as in 3.3, we can of course dene pG, t, cq to be the union of sets associated to the points in the bre of c in CA pG, q. We can also dene the set r r pG1 , t, cq pG1 , 1 , t, cq, r r for any element G1 P EpGq with auxiliary datum pG1 , 1 q, as the union of r 1 , 1 q attached to points in the bre of c in CA pG1 , 1 q. r r subsets of pG r The general goal is of course to understand the discrete part of the trace formula, and the spectral information it contains. In 3.3, we observed that it has a canonical decomposition Idisc,t pf q Idisc,t,c pf q, f P HpGq,
r cPCA pG,q

into c-components. It is therefore enough to understand any such component, and specically, to compare the c-analogues of the spectral and endoscopic expansions (4.1.1) and (4.1.2). In our special case of groups r G P Esim pN q, where we can appeal to strong multiplicity 1 for GLpN q and its generalization in Theorem 1.3.2, we indexed c-components in (3.3.12) r and (3.3.13) directly in terms of parameters P pN q. This is not possible in general. Instead, one must somehow decompose the linear forms ` (4.8.3) tr MP,t,c pw, q IP,t,c p, f q , f P HpGq, and (4.8.4) p1 Sdisc,t,c pf 1 q, f P HpGq,

in the general c-analogues of (4.1.1) and (4.1.2) into contributions from parameters P pG, t, cq. The question is embedded in the larger problem of establishing general global intertwining relations for (4.8.3) and stable multiplicity formulas for (4.8.4). There is no reason to expect that we could establish global intertwining relations in advance. They will no doubt have to be proved by local-global methods, which ultimately rely on the identity of (4.1.1) and (4.1.2). At r least this is how we will be dealing with the groups G P Esim pN q in this volume. On the other hand, we might often expect to have stable multiplicity formulas already in hand before beginning the comparison. One reason is that they apply only to connected quasisplit groups, while the general object G is much broader. We will not see this distinction for much of the present r volume, since the groups G P Esim pN q of interest in the rst eight chapters are in fact quasisplit. Our proof of the stable multiplicity formula for these groups will have to come out of the comparisons of the expansions (4.1.1)

4.8. REMARKS ON GENERAL GROUPS

241

and (4.1.2), and will be one of our most dicult tasks. Once we do have a proof, however, it will be much easier to deal with inner forms of G. We shall therefore assume that the linear form (4.8.4) satises the stable multiplicity formula of 4.1, or rather the adaptation of (4.1.12) to this general object. We thus assume that for our general triplet G, we can write p1 p1 Sdisc,t,c pf 1 q Sdisc,1 pf 1 q,
1 PpG1 ,t,cq

for any G1 P Eell pGq, where (4.8.5) p1 Sdisc,1 pf 1 q |S1 |1 pS 0 1 q 1 p 1 q f 1 p 1 q.

The coecients |S1 |1 and pS 0 1 q in (4.8.5) have already been dened. The 1 sign character G1 in the coecient 1 p 1 q G1 ps1 q is dened exactly as in (1.5.6). The linear form f 1 p 1 q is supposed to be a product 1 1 1 fv , fv pv q, f1 f 1 p 1 q
v v 1 r of stable characters on the groups G1 pFv q, attached to the localizations v 1 . Part of our assumption here is that we have already dened these of local objects. Given the stable multiplicity formula (4.8.5), the general c-analogue of the endoscopic expansion (4.1.2) will be more accessible than that of the spectral expansion (4.1.1). We shall sketch what extensions of the analysis of 4.4 are needed to deal with it. We do not have to be concerned with the induction hypotheses that complicated the earlier discussion. We do have to address the main complication of the general case here, the fact that the set S can be larger than the centralizer of the image of . For example, p p S contains the full center ZpG0 q of G0 , rather than just its subgroup of -invariants. We dene the associated two quotients of S by
1

(4.8.6) and

p S S {ZpG0 q
0 p S 0 pS q S {S ZpG0 q.

The sets S S pGq, S S pGq and S S pGq are then bitorsors under the respective complex groups S S pG0 q, S S pG0 q and S pG0 q. In our special cases G P E r q and G G r S sim pN q, G GpN

p (and so long as G SOp2, Cq), S does equal the centralizer, and this notation matches what we have been using. In the general case, we write C C pG0 q for the actual centralizer p of the image of in G0 , following Kottwitz [K1, 10]. We then dene the

242

4. THE STANDARD MODEL

associated two quotients C C pG0 q and C C pG0 q of C by p C C {ZpG0 q and

0 p C 0 pC q C {C ZpG0 q . These are normal subgroups of nite index in the corresponding quotients S and S , since p p ZpG0 q X C ZpG0 q

and
0 0 p p C ZpG0 q S ZpG0 q.

In particular, the connected components C 0 and S 0 are equal. We thus obtain a canonical injection ` p S {C S {C ker1 F, ZpG0 q . We note again that in the special cases above, we have C S and C S , and therefore no need for this supplementary notation. In the general case, the set S retains its central role. For example, the chain (4.1.13) of subsets of pGq is dened in terms of S , as before. We can of course dene a similar chain of subsets of pG, t, cq, as well as an analogous chain of subsets of pG1 , t, cq for any endoscopic datum G1 r r with auxiliary datum pG1 , 1 q. It is the connected group S 0 attached to S , or rather its analogue for a parameter 1 , that determines the coecient pS 0 1 q in (4.8.5). (We were actually anticipating the subsequent denition (4.8.6) of the quotient S 1 when we stated (4.8.5).) Since this coecient vanishes if the center of S 0 1 is innite, we can take the sum in the formula p1 for Sdisc,t,c pf 1 q above over the subset ( disc pG1 , t, cq 1 P pG1 , t, cq : |ZpS 1 q| 8 of pG1 , t, cq , or better, the smaller subset (4.8.7) ( s-disc pG1 , t, cq 1 P pG1 , t, cq : |ZpS 0 1 q| 8 of stable-discrete parameters. For general perspective, we note that if we were to form the set s-ell pG1 , t, cq ell pG1 , t, cq X s-disc pG1 , t, cq of stable-elliptic parameters in ell pG1 , t, cq, it would reduce simply to the subset 2 pG1 , t, cq. r We are assuming that c is any class in some given quotient CA pG, q. The c-analogue of the endoscopic expansion is the formula p pG, G1 q S 1 pf 1 q
disc,t,c G1 PEell pGq

for the c-component Idisc,t,c pf q, f P HpGq,

4.8. REMARKS ON GENERAL GROUPS

243

of Idisc,t pf q given by (3.3.10). We therefore have an expansion p1 (4.8.8) pG, G1 q Sdisc,1 pf 1 q

G1 1

for Idisc,t,c pf q, where G1 is summed over Eell pGq, and 1 is summed over disc pG1 , t, cq. The coecient pG, G1 q is given in (3.2.4) as (4.8.9) ` ` p p p 0 pG q1 ker1 F, ZpG0 q 1 ker1 F, ZpG1 q ZpG1 q 1 OutG pG1 q1 , p1 while Sdisc,1 pf 1 q is dened by (4.8.5). The heart of 4.4 was a transformation of the double sum over pG1 , 1 q to a double sum over p, sq, according to the correspondence (1.4.11). The general process here is a little more complicated. It will be instructive for us to describe the steps in symbolic form, although this should not disguise the fact that we are basically repeating discussion from 4.4. The general correspondence (1.4.11) can be regarded as an equivariant bijection XpGq Y pGq p between left two G0 -sets. The domain XpGq is the set of pairs x pG1 , 1 q, 1 belongs to the set EpGq of endoscopic data for G, taken up to the where G image 1 pG 1 q of G 1 in L G1 and up to translation of the associated semisimple p p element s1 P G by ZpG0 q, while 1 is an actual L-homomorphism from A to r r the group L G1 , which factors through the L-embedding 1 . The codomain is the set of pairs y p, sq, where belongs to the set F pGq of actual L-homomorphisms from A to L G0 , and s is a semisimple element in S . The projections of the two kinds of pairs onto their rst components yield a larger diagram XpGq Y pGq EpGq F pGq p of G0 -equivariant mappings. We note that the correspondence restricts to a bijection between the p 0 -invariant subsets G ( Xdisc,t,c pGq x pG1 , 1 q : G1 P Eell pGq, 1 P Fs-disc pG1 , t, cq and ( Ydisc,t,c pGq y p, sq : P Fdisc pG, t, cq, s P S ,ell q of XpGq and Y pGq. We are writing Eell pGq, Fdisc pG, t, cq and Fs-disc pG1 , t, cq here for the respective preimages of Eell pGq, disc pG, t, cq, and s-disc pG1 , t, cq r r in sets EpGq, F pGq and F pG1 , 1 q. This bijection also ts into a diagram, as above, but with EpGq and F pGq replaced by Eell pGq and Fdisc pG, t, cq.

244

4. THE STANDARD MODEL

p The mappings of G0 -sets XpGq, Xdisc,t,c pGq, etc. all descend to mapp p pings among the corresponding sets G0 zzXpGq, G0 zzXdisc,t,c pGq, etc., of p p p G0 -orbits. The set G0 zzEell pGq of G0 -orbits in Eell pGq equals Eell pGq, by 0 zzF p p0 denition. The set G disc pG, t, cq of G -orbits in Fdisc pG, t, cq projects onto disc pG, t, cq, but this mapping is not generally bijective. However, it ` p follows from the denitions that the group ker1 F, ZpG0 q acts transitively on its bres, and that the stabilizer of any equals the injective image of ` p p S {C in ker1 F, ZpG0 q . The bre of in G0 zzFdisc pG, t, cq therefore has order ` p ker1 F, ZpG0 q S 1 C , (4.8.10)
since |S | |S |. That said, we see that the mappings all t together into a general diagram

p p G0 zzXdisc,t,c pGq > G0 zzYdisc,t,c pGq (4.8.11) p G0 zzEell pGq } Eell pGq p G0 zzFdisc pG, t, cq disc pG, t, cq.

The diagram (4.8.11) is the focal point for our transformation of the endoscopic expression (4.8.8). Speaking in symbolic terms, we need to follow the path that leads from the lower left hand corner of the diagram to the lower right hand corner. The rst step will be to change the double sum in p (4.8.8) to a sum over the set of G0 -orbits in Xdisc,t,c pGq. This requires two 1. changes in the inner sum over The variable 1 in (4.8.8) is supposed to be summed over the set r r s-disc pG1 , t, cq s-disc pG1 , 1 , t, cq. We can instead sum it over the larger set p r r r p G1 zzFs-disc pG1 , 1 , t, cq G1 zzFs-disc pG1 , t, cq, if we multiply the summand by the product ` p ker1 F, ZpG1 q 1 S1 C1 1 . (4.8.12) r Indeed, this product equals the inverse of the analogue for G1 of (4.8.10), r 1 G1 imply that the mapping since the conditions on the extension G p p r ker1 pF, G1 q ker1 pF, G1 q

4.8. REMARKS ON GENERAL GROUPS

245

is an isomorphism. The variable G1 in (4.8.8) is supposed to be summed over the set p Eell pGq G0 zzEell pGq in the lower left hand corner of the diagram. This is what we want. However, p p p the stabilizer of G1 in G0 is the extension AutG pG1 q of G1 , rather than G1 . 1 over the set of Aut pG1 qThis means that we should really be summing G 1 , t, cq, rather than G1 -orbits. We can do this, so long as we p orbits in Fdisc pG multiply the summand by the number (4.8.13) |OutG pG1 q| |OutG pG1 , 1 q|1

p of G0 -orbits in the given AutG pG1 q-orbit of 1 . As earlier, OutG pG1 , 1 q is p the stabilizer of 1 , regarded as a G0 -orbit, in OutG pG1 q. We have shown how to express (4.8.8) as an iterated sum over the base p set G0 zzEell pGq and its pointwise bres in (4.8.11). That is, we can write (4.8.8) as a sum over orbits x pG1 , 1 q in the upper left hand set in (4.8.11), so long as the summand is rescaled by the product of (4.8.12) and (4.8.13). We write this in turn as a sum over elements y p, sq in the upper right hand set in (4.8.11), with the expectation of being able to express the summand in terms of the bijective image p, sq of pG1 , 1 q. In other words, p we express (4.8.8) as an iterated sum over the base set G0 zzFdisc pG, t, cq and its pointwise bres in (4.8.11). The second step is to write this as a double sum over and s in more familiar sets. As it presently stands, the expression is an iterated sum over orbits of p L-homomorphisms P Fdisc pG, t, cq under G0 and orbits of elliptic elements p 0 . We can take the rst sum over s P S ,ell under the stabilizer of in G in the quotient disc pG, t, cq, provided that we multiply the summand by p the order (4.8.10) of its bre. The stabilizer of in G0 is the centralizer C . The second sum is over the set of orbits in S ,ell under C , or equivalently, the set of orbits under the quotient C of C . However, we would prefer to take a sum over the set S 0 zzS ,ell EpS ,ell q E,ell of orbits in S ,ell under the subgroup C 0 S 0 of C . The C -orbit of s is bijective with the quotient of C by the subgroup C ` Centps, C q. ,s The C 0 -orbit of s is bijective with the quotient of C 0 by the subgroup C ,s Centps, C 0 q. We can therefore take the second sum over orbits s in E,ell , if we multiply the summand by the quotient |C ` {C ,s | |C {C 0 |1 , ,s

246

4. THE STANDARD MODEL

which is to say the quotient (4.8.14) |C ` {C ,s | |C |1 . ,s

We have now established that the double sum over G1 and 1 in the endoscopic expression (4.8.8) can be replaced by a double sum over P disc pG, t, cq and s P E,ell , provided that the summand is multiplied by the product of the four factors (4.8.10), (4.8.12), (4.8.13) and (4.8.14). The summand itself becomes the product of these four factors with (4.8.9) and the right hand side of (4.8.5). Observe that four of these six factors are given in terms of the pair x pG1 , 1 q, rather than its image y p, sq. As in 4.4, some of the components of these factors will cancel from the product. Others can be expressed directly in terms of y. For example, the denominator in (4.8.13) equals
` ` p p |OutG pG1 , 1 q| |C,s {C,s X G1 ZpG0 q |,

where
` ` C,s tg P C : gsg 1 su Cy

is the preimage of C ` in C . In particular, (4.8.13) specializes to the ,s factor (4.4.5) in 4.4. We conclude that (4.8.8) can be written as the sum over P disc pG, t, cq and s P E,ell of the product of two expressions (4.8.15) and (4.8.16) |0 pG q|1 |S |1 pS 0 1 q 1 p 1 q f 1 p 1 q, in which y p, sq is the image of x pG1 , 1 q under the bijection in (4.8.11). The expressions (4.8.15) and (4.8.16) are the two factors (7.9) and (7.10) in [A9]. They also reduce to the corresponding two expressions of 4.4, where the group C is equal to S . (The multiplicity m in the second expression in 4.4 does not appear here, since now represents an element r in a set pGq rather than an equivalence class in pGq.) As in the special cases of 4.4, the product of the two expressions simplies. We shall outline the steps. The product of the rst two factors in (4.8.15) has a reduction
` ` p p p |C,s {C,s X G1 ZpG0 q |1 |C1 |1 |C ` {C 0 ZpG1 q |1 , ,s ,s ` ` p p p |C,s {C,s X G1 ZpG0 q |1 |C1 |1 |ZpG1 q |1 |C ` {C ,s |1 ,s

as in the discussion in 4.4. (See also [A9, p. 4849].) The entire expression (4.8.15) can then be written as p p |C ,s {C 0 ZpG1 q |1 |ZpG1 q |1 ,s p |0 pC ,s q|1 |C 0 X ZpG1 q |1 ,s p |0 pS ,s q|1 |S 0 X ZpG1 q |1 , ,s

4.8. REMARKS ON GENERAL GROUPS

247

since C 0 S 0 . In the expression (4.8.16), the coecient ,s ,s ` `` p 0 r p pS 0 1 q S1 {ZpGq S 0 {S 0 X ZpG1 q ,s ,s satises an identity p pS 0 1 q pS 0 q |S 0 X ZpG1 q | , ,s ,s by (4.1.9). Lemma 4.4.1, whose proof in 4.6 carries over to the general case, allows us to write the coecient 1 p 1 q in (4.8.16) as 1 p 1 q G1 ps1 q G ps xs q. Finally, we can again write (4.8.17)
1 f 1 p 1 q fG p, sq, r1

to denote the dependence of the linear form in (4.8.16) on and s. All of the components in (4.8.15) and (4.8.16) have now been expressed in terms of and s. With no further need to refer to the original pair x pG1 , 1 q, we can allow x to denote an element in the group S of connected components of S . We can then write the endoscopic expression (4.8.8) as a triple sum 1 (4.8.18) |0 pG q|1 |S |1 |0 pS ,s q|1 pS 0 q G ps xq fG p, sq ,s
x s

over P disc pG, t, cq, x P S and s P E,ell pxq. The transformation (4.8.18) of the endoscopic expression (4.8.8) is the general analogue of the formula of Lemma 4.4.2 (with allowance for our slightly dierent use of the symbol ). We would again expect the linear form 1 1 fG p, xq fG p, sq, x P S , s P S ,ss pxq, to depend only on the component x of s. If this is so, we can compress the sum over s in (4.8.18) to the coecient epxq e pxq of Proposition 4.1.1. After a change of variables in the sum over x, we will then be able to write (4.8.18) as a double sum 1 (4.8.19) |0 pG q|1 |S |1 e pxq G pxq fG p, s xq
x

over P disc pG, t, cq and x P S . This is the general analogue of the expression (4.4.11) of Corollary 4.4.3. The c-discrete part Idisc,t,c pf q of the trace formula is thus equal to the elementary endoscopic expression (4.8.19). The other half of the problem would be to extend the analysis of 4.3 to the spectral expansion of Idisc,t,c pf q. The formal aspect of this process can be carried out. For example, the commutative diagram (4.2.3) makes sense for any parameter in the general set pGq. However, one is immediately forced to deal with the linear form

248

4. THE STANDARD MODEL

(4.8.3). One would need to decompose (4.8.3) explicitly into suitable linear forms fG p, uq, P disc pG, t, cq, u P N , dened explicitly in terms of local normalized intertwining operators. This would be the general analogue of Corollary 4.2.4. We would again expect the linear form fG p, xq fG p, uq, x P S , u P N pxq, to depend only on the image x of u in S . If this is so, the methods in 4.3 would lead to an elementary spectral expression for Idisc,t,c pf q as a double sum i pxq G pxq fG p, xq (4.8.20) |0 pG q|1 |S |1
x

over P disc pG, t, cq and x P S . We will say no more about the putative expression (4.8.20), except to point out that it is the general analogue of the expression (4.3.9) of Corollary 4.3.3. In particular, the coecient ipxq i pxq in the expression equals the coecient epxq e pxq in (4.8.19), by Proposition 4.1.1. Moreover, the linear form fG p, xq in (4.8.20) ought to be equal to the linear form 1 fG p, s xq in (4.8.19), by some general version of the global intertwining relation. This would establish a term by term identication of the two expressions (4.8.19) and (4.8.20). We have noted that the chain of subsets (4.1.13) of pGq also makes sense in this general setting. Of most signicance is the subset 2 pGq, which ought to govern the discrete spectrum. To be more specic, we could say that the comparison of (4.8.19) and (4.8.20) has two purposes. One is to prove that parameters in the complement of 2 pGq do not contribute to the discrete spectrum. The other is to compute the contribution of parameters that do lie in 2 pGq. The expected contribution of will be a multiplicity formula for representations in a global packet , in terms of associated characters xx, y on the nite group S . This would be the natural generalization of the multiplicity formula (4.7.9) of Theorem 1.5.2, with allowance made for the larger group S , the fact that S could be nonabelian, and the possibility of G being a twisted group. It could also be regarded as the degenerate case of the proposed global intertwining relation. Given the stable multiplicity formula (4.8.5) for P pGq, one would try to deduce the actual spectral multiplicity formula for representations by a comparison of (4.8.19) with (4.8.20), as in our special case discussed at the end of the last section. We add one nal comment on the title of this chapter. It seems apt, despite being a possible case of cultural appropriation! The contents of the chapter do represent a model of sorts. The model is standard, in the sense that it governs the general endoscopic comparison of trace formulas. However, we are still a very long way from being able to apply it to arbitrary r groups. We can hope that the results we obtain for the groups G P Esim pN q,

4.8. REMARKS ON GENERAL GROUPS

249

combined with similar results for other classical groups, will lead to an endoscopic classication of a much broader class of groups, which share only the property of being endoscopically related to general linear groups. What about more general groups? Langlands has outlined a remarkable, though still speculative, strategy for applying the trace formula to cases of functoriality that are not tied to endoscopy [L13], [L14], [FLN]. His proposal is probably best suited to the stable trace formula. If it works, it might ultimately yield a general stable multiplicity formula as a byproduct. One could then imagine applying this formula in the standard model, as above, to a general endoscopic classication of representations.

CHAPTER 5

A Study of Critical Cases

5.1. On the case of square integrable In the last chapter, we studied the contribution to the trace formula of many global parameters . We did so by applying induction arguments to what we called the standard model. The culmination of these eorts was Proposition 4.5.1, which established the stable multiplicity formula in a number of cases. It also yielded the multiplicity formula of Theorem 1.5.2 for these cases, with its interpretation as a vanishing condition. We recall that the stable multiplicity formula is the assertion of Theorem 4.1.2. It is at the heart of all of our global theorems. We have now established r r it for global parameters P pN q that do not lie in ell pN q, or in any of the r ell pGq of pN q attached to elements G P Esim pN q. In this chapter r r subsets we shall begin the study of the cases that remain. These cases are of course the most important. They are also the most dicult. The questions become increasingly subtle as we consider successive subsets of parameters in the chain (5.1.1) r r r sim pGq 2 pGq ell pGq, r G P Esim pN q,

attached to a given G. The answers will draw upon new techniques, which we will introduce in this and future chapters. These will be founded on localglobal methods, in which local and global techniques are used to reinforce each other. We begin by setting up a general family of global parameters, on which we will later impose local constraints. Throughout the chapter, we will let (5.1.2) r F
8 N 1

r FpN q

denote a xed family of global parameters in the general set r

8 N 1 1 1

r r

r pN q.

r We assume that F is closed under the projections i ,

251

1 i r,

252

5. A STUDY OF THE CRITICAL CASES

and under direct sums p 1 , 2 q 1 2 , r 1 , 2 P F.

r In other words, F is the graded semigroup generated by its simple components r b , P Fsim , with the grading deg N, r P pN q, r inherited from . We have obviously written r r r Fsim F X sim r r r here. We shall also write Fell , F2 pN q, Fsim pGq, etc., for the intersection of r r F with the corresponding subset of . Throughout this chapter, N will again be xed. We shall assume inr ductively that all of the local and global theorems hold for any P F with degpq N . This is of course the blanket induction hypothesis we have r carried up until now. We are introducing it formally within the family F, since it is in the limited context of such families that we will eventually establish the general local theorems. Its interpretation here will be for the most part obvious. r There is one point in the induction hypothesis for F that does call for an explanation. It concerns the assertion of Theorem 1.5.3(b). Consider a pair i i b i , i 1, 2, r of distinct self-dual, simple parameters in F. Motivated by the remark following the proof of Lemma 4.3.1 in 4.6, we shall call p1 , 2 q an -pair, for want of a better term, if the following condition holds: the generic pair p1 , 2 q is orthogonal, in that it satises the condition imposed on the pair p1 , 2 q in Theorem 1.5.3(b), and the tensor product 1 b 2 is the sum of an odd number of irreducible representations of SLp2, Cq of even dimension. The corresponding sum (5.1.3) 1 2 ,

which we will call an -parameter, is a self-dual, elliptic parameter for a general linear group. If w is the associated twisted Weyl element for the relevant maximal Levi subset, it follows from the analysis of (4.6.9) and (4.6.10) in the proof of Lemma 4.3.1 that the corresponding global normalizing factor satises (5.1.4)
1 r pwq p 2 , 1 2 q.

Theorem 1.5.3(b) asserts that this number equals 1. Our induction hypothr esis for F will be that the assertion is valid for any -pair p1 , 2 q with degp1 q ` degp2 q N.

5.1. ON THE CASE OF SQUARE INTEGRABLE

253

It then follows from the remark in 4.6 mentioned above that Lemma 4.3.1 r is valid for any P FpN q that is not an -parameter. To make further progress, we shall also take on the following temporary r hypothesis on the elements in Fell pN q. r r Assumption 5.1.1. Suppose that G P Eell pN q, and that belongs F2 pGq. Then there is a unique stable linear form r f f G pq, f P HpGq, r on HpGq with the general property r r f G pq fN pq, together with a secondary property that f G pq f S pS qf O pO q, in case G G S GO , S O , and fG fS fO are composite. The reader will recognize in this hypothesis a global analogue of the local assertion of Theorem 2.2.1(a). It would obviously follow from the local assertion, but this of course has yet to be established. We note that r r if lies in the complement of F2 pGq in FpGq, the global assertion can be reduced to the corresponding assertion for a Levi subgroup of GLpN q, which then follows from our induction hypothesis. In other words, the assertion of r Assumption 5.1.1 is valid more generally for any P FpGq. (If G GS GO , r r S q and FpGO q, rather r we have naturally to dene FpGq as the product of FpG r than as a subset of FpN q.) This is the way Theorem 2.2.1(a) was stated. We need to pause here for some further discussion of the condition P r F2 pGq in Assumption 5.1.1. For it raises a second point concerning our r induction hypothesis for F. If is not simple generic, which is to say r that it does not lie in the set sim pN q of self dual cuspidal automorphic r representations of GLpN q, the datum G P Eell pN q such that belongs to r F2 pGq is given by the induction hypothesis and the original constructions of 1.4. However, we will obviously have to deal also with parameters that are simple generic. According to the global theorems we have stated, there will be three r r equivalent ways to characterize the subset sim pGq of sim pN q attached to a r given G P Esim pN q. Theorem 1.4.1, which we have used for the formal denir tion, characterizes sim pGq in terms of the discrete spectrum of G. Theorem r f P HpGq, r r f P HpN q,

254

5. A STUDY OF THE CRITICAL CASES

r 1.5.3(a), together with the fact that G for any P pGq, describes r sim pGq in terms of the poles of L-functions. The third characterization is provided by the stable multiplicity formula of Theorem 4.1.2. It tells us r r that sim pGq is the subset of elements P sim pN q such that the linear G r form Sdisc, on HpGq is nonzero. We have of course not proved any of these r theorems, regarded as conditions for the generic elements in Fsim pN q. But we could ultimately still take the assertion of any one of them as a denition r r of the subset Fsim pGq of Fsim pN q. The third characterization is the least elementary. However, with its roots in harmonic analysis, it would be the most natural one to work with as we try to extend the global part of our induction hypothesis here to N . We shall therefore take it as the basis for a temporary denition, which we will later show is equivalent to the one in 1.4. r Suppose that F is generic, in the sense that it is contained in the subset r r of generic global parameters in . If G belongs to Esim pN q, we could simply r dene Fsim pGq by the third condition above, namely as the subset of eleG r ments P Fsim pN q with Sdisc, 0. For we know from (3.3.14) that if r r belongs to Fsim pN q, there is a G P Esim pN q with this property. Indeed, the left hand side of (3.3.14) is nonzero, while it follows from Proposition 3.4.1 and Theorem 1.3.2 that the contribution to the right hand side of (3.3.14) r r of any datum in the complement of Esim pN q in Eell pN q vanishes. We expect rsim pN q to be uniquely determined by . We know this to be so if G P E 1 or N is odd. But if 1 and N is even, we would not be able to rule out the possibility at present that could lie in two dierent sets r Fsim pGq. A more serious problem with this denition is the possibility of a conict with Assumption 5.1.1 in some of our later constructions. We shall therefore adopt a slightly more exible (though ultimately equivalent) convention. r For the generic family F, we simply assume that we have attached global subsets r r (5.1.5) Fsim pGq t P Fsim pN q : S G 0u
disc,

r r of Fsim pN q to the simple endoscopic data G P Esim pN q, so that r r (5.1.6) FpN q Fsim pGq.
r GPEsim pN q

Assumption 5.1.1, as it applies to simple pairs pG, q, r r G P Esim pN q, P Fsim pGq,

represents an axiom that is imposed afterwards. Once it is granted, we could r then make our temporary denition precise by taking Fsim pGq to be the set rsim pN q such that S G of all parameters P F disc, 0, and such that the linear r pq transfers to G. Assumption 5.1.1 for simple pairs pG, q then form fN

5.1. ON THE CASE OF SQUARE INTEGRABLE

255

becomes the assertion (5.1.6). In working with this temporary denition, we must keep in mind the hypothetical possibility that the inclusion (5.1.5) might be proper, or the possibility that of the union (5.1.6) might not be disjoint. But in any case, we have now agreed on an ad hoc convention for r the simple parameters in FpN q, thereby introducing a new denition of the r sets F2 pGq to which Assumption 5.1.1 is supposed to apply. We shall resolve r it in terms of the original denition (for a particular family F) at the end of the chapter. r If F is nongeneric, in the sense that it contains elements in the compler r ment of in , we will assume that the relevant global theorems hold for r the generic part of any simple parameter P Fsim . That is, we assume r that the three ways of dening sim pGq are equivalent for such parameters. r2 pGq and FpGq of F are then all well dened. r r The subsets F We have thus rened the global induction hypothesis, as it applies to r parameters in a given family F. We observe that the two adjustments concern the two parts of Theorem 1.5.3. Part (a) of the theorem represents an induction hypothesis for generic families that is assumed to have been resolved in the nongeneric case. Part (b) represents an induction hypothesis in the nongeneric case that is irrelevant for generic families. We will ultir mately study the two kinds of families F separately, the generic case being the topic of Chapter 6, and the nongeneric case the topic of Chapter 7. In this chapter, however, we will generally be able to treat the two r cases together. Having assumed inductively that for a given family F the r relevant theorems hold for elements P FpN q with N N , we must r try to show that they hold also for parameters P FpN q. This is a long term proposition. It will not be settled in general until the later chapters, r after we have constructed families F with specied local conditions. In the meantime, we shall see what can be deduced from the induction hypotheses, when used in combination with Assumption 5.1.1. r In this section, we shall look at the basic case that lies in Fell pN q. With rell pN q is the union our convention above for the simple generic elements, F r r over G P Eell pN q of the sets F2 pGq. However, this temporary convention still G leaves open the question of how to assign a meaning to the centralizers S r attached to simple pairs pG, q, with a generic element in Fsim pGq. In situations where pG, q is understood to have been xed, we shall implicitly agree that # 1, if G G, G (5.1.7) S S 0, otherwise, r for any G P Esim pN q. This is what we expect, since G should be the unique r datum with P Fsim pGq. In any case, the dierence
0 Sdisc, pf q G 0 Sdisc, pf q,

r f P HpG q,

256

5. A STUDY OF THE CRITICAL CASES

between Sdisc, pf q and its expected value will then be dened as in (4.4.9) (with G in place of G1 ). We will of course also be interested in the case G that is not simple generic. In this case, S is given by induction, and the linear form 0 Sdisc, pf q is then dened by Assumption 5.1.1. We x r r r P Fell pN q, and take G to be a xed datum in Eell pN q with P F2 pGq. We remind ourselves that G is uniquely determined by , except possibly in the case that is simple generic. The centralizer set ` r r S pN q S GpN q

r for GpN q is a connected abelian bi-torsor. It follows that the global inr tertwining relation for GpN q and reduces simply to the assertion of Assumption 5.1.1. In particular, the stated conditions of Corollaries 4.3.3 and r 4.4.3 (with GpN q in place of G) both hold. In the case of Corollary 4.3.3, there is also an implicit condition, namely that the proof of Lemma 4.3.1 is valid under the induction hypotheses now in force. If we assume that is not an -parameter (5.1.3), this condition holds, and both corollaries are r valid for GpN q and . We then see from the identity of Proposition 4.1.1, which is elementary in this case, that the corresponding expressions (4.3.9) r r and (4.4.11) are equal. It follows that for f P HpN q, the dierence between rq and the linear form Idisc, pf
0 N r rdisc, pf q r

r 0 rdisc, pf q

r GpN q

r equals the dierence between Idisc, pf q and the linear form

0 N r r sdisc, pf q r GpN q r 0 sdisc, pf q.

r r rN Since GpN q0 GLpN q, we know that 0 rdisc, pf q 0. It follows from the r denition (4.4.10) (again with GpN q in place of G) that r p rpN, G q 0 Sdisc, pf q 0. (5.1.8)
r G PEsim pN q

r r If G belongs to the complement of Esim pN q in Eell pN q, we know from 0SG G our induction hypothesis that disc, pf q, the dierence between Sdisc, pf q and its expected value, vanishes. One of our long term goals is to show r that the same is true if G belongs to Esim pN q, or indeed if G is replaced r by any element G P Esim pN q. In other words, we would like to establish the stable multiplicity formula of Theorem 4.1.2 for the given , and any r G P Esim pN q. The next lemma gives some preliminary information. Lemma 5.1.2. Assume that for the given pair pG, q, r r G P Eell pN q, P F2 pGq,

5.1. ON THE CASE OF SQUARE INTEGRABLE

257

is not an -parameter (5.1.3). Then the stable multiplicity formula (4.1.11) holds for G and if and only if r S pf q 0 S pf q 0, f P HpG q,
disc, disc,

for every

r P Esim pN q with G G.

Proof. One direction is clear. If 0 Sdisc, 0 for every G G in r r r r pG Esim pN q, (5.1.8) tells us that 0 Sdisc, pf G q vanishes for any f P HpN q. The required assertion in this direction already being known if G is composite, we r r r r can assume that G P Esim pN q. The correspondence f f G from HpN q to 0SG r SpGq is then surjective, by Corollary 2.1.2. It follows that disc, vanishes, r as an OutN pGq-symmetric stable linear form. In other words, the stable multiplicity formula for r f P HpGq, S G pf q, disc,

is valid. G Conversely, suppose that 0 Sdisc, 0. Then (5.1.8) reduces to r p rpN, G q 0 Sdisc, pf q 0.
G G

r But if G P Esim pN q is distinct from G, (4.4.12) tells us that ` 0 r f P HpG q, Sdisc, pf q Sdisc, pf q tr Rdisc, pf q , r since does not lie in pG q unless is generic and simple, in which case the convention (5.1.7) leads to the same outcome. Applying Proposition 3.5.1 to the resulting formula ` rpN, G q tr Rdisc, pf q 0,
G G

r in which tf u is replaced by any compatible family of functions tf u, we deduce that ` G G, tr Rdisc, pf q Sdisc, pf q 0, as required. Corollary 5.1.3. Suppose that G and are as in the lemma, and that one of the further conditions (i) N is odd, (ii) 1, or r (iii) G R Esim pN q, holds. Then r (5.1.9) S G pf q m |S |1 G pq f G pq, f P HpGq,
disc,

r while Sdisc, 0 for every G P Esim pN q distinct from G. In other words, r the stable multiplicity formula holds for and any G P Esim pN q.

258

5. A STUDY OF THE CRITICAL CASES

G Proof. If (iii) holds, the stable multiplicity formula for Sdisc, pf q is a consequence of our induction hypotheses, as we noted above. It then follows G from the lemma that Sdisc, 0 for all G G, which is to say, every G . G r We can therefore assume that G P Esim pN q. Recall that Sdisc, vanishes unless the characters and G on F attached to G and are equal. We know that G determines G uniquely as a simple twisted endoscopic datum, G if it is nontrivial or if N is odd. In these cases then, Sdisc, 0 for every r G P Esim pN q distinct from G. The required stable multiplicity formula (5.1.9) for G then follows from the lemma.

r The remaining case is for N even, 1, and G P Esim pN q. It is r not so easily resolved. There is now a second group G_ P Esim pN q with r G_ G , and even though does not belong to F2 pG_ q (except G_ possibly when is generic and simple), we cannot say a priori that 0 Sdisc, vanishes. In this case, G and G_ are split groups, whose dual groups are ` SppN, Cq and SOpN, Cq. We write L for the group, isomorphic to GL 1 N , 2 that represents a Siegel maximal Levi subgroup of both G and G_ . For convenience, we shall also write r r opG q |OutN pG q|, r G P Esim pN q,

an integer equal to either 1 or 2. There are two transfer mappings f f L fL , and

_ f _ f _,L fL ,

r f P HpGq r f _ P HpG_ q,

r r from HpGq and HpG_ q respectively to the space SpLq IpLq. It is not hard to see from the denitions that they have a common image in SpLq, r which we denote by S 0 pLq. Lemma 5.1.4. Suppose that G and are as in Lemma 5.1.2, and also r that N is even, 1, and G P Esim pN q. Then there is a linear form hL hL pq, r on S 0 pLq such that
G r Sdisc, pf q m |S |1 G pq f G pq opGq f L pq,

r hL P S 0 pLq,

r f P HpGq,

and
G r Sdisc, pf _ q opG_ q f _,L pq,
_

r f _ P HpG_ q, r f P HpG q,

while
Sdisc, pf q 0 Sdisc, pf q 0,

r for every G P Esim pN q distinct from G and G_ .

5.1. ON THE CASE OF SQUARE INTEGRABLE

259

r Proof. If G P Esim pN q is distinct from G and G_ , the character G is distinct from , and
Sdisc, 0 Sdisc, 0.

The corresponding summand in the identity (5.1.8) thus vanishes, and the left hand side of (5.1.8) reduces to a sum of two terms. It follows from (3.2.6) that rpN, Gq opGq rpN, G_ q opG_ q r r since p p |ZpGq | |ZpG_ q | 1. The identity then takes the form ` G ` G_ (5.1.10) opGq1 0 Sdisc, pf q ` opG_ q1 0 Sdisc, pf _ q 0, r r for any compatible pair of functions ( r r f, f _ : f P HpGq, f _ P HpG_ q . As members of a compatible family of functions, f and f _ are not independent. Their stable orbital integrals have the same values on Levi subgroups that are common to G and G_ . But any such Levi subgroup is conjugate to a subgroup of L. Since this is the only obstruction to the independence of f and f _ , the two summands on the left hand side of (5.1.10) r can dier only by the pullback of a linear form on S 0 pLq. In other words, we can write _ 0 G_ r r S pf _ q S G pf _ q opG_ q f _,L pq, f _ P HpG_ q,
disc, disc,

r where represents a linear form on S 0 pLq. G The required formula for Sdisc, pf q then follows from (5.1.10), and the specialization
0 G Sdisc, pf q G Sdisc, pf q m |S |1 G pq f G pq

of the denition (4.4.9). Corollary 5.1.5. With the conditions of the lemma, the linear form r r f _ opG_ q f _,L pq, f _ P HpG_ q, r is the restriction to HpG_ q of a unitary character on G_ pAq. r Proof. Since does not belong to pG_ q (except possibly when is generic and simple), (4.4.12) tells us that ` G_ 0 G_ G_ Sdisc, pf _ q Sdisc, pf _ q tr Rdisc, pf _ q .
G Since the given linear form equals Sdisc, pf _ q, the corollary follows.
_

The last two lemmas and their corollaries came with the condition that r P Eell pN q not be an -parameter (5.1.3). We shall now see what happens in this supplementary case.

260

5. A STUDY OF THE CRITICAL CASES

Lemma 5.1.6. Suppose that for the given pair pG, q, r r G P Eell pN q, P F2 pGq,

is an -parameter (5.1.3). Then ` tr Rdisc, pf q 0 0 Sdisc, pf q,

r f P HpG q,

r for every G P Esim pN q. Moreover, the generic components 1 and 2 of 1 and 2 satisfy `1 2 , 1 2 1, in accordance with Theorem 1.5.3(b). Proof. We are assuming that is a sum of an -pair p1 , 2 q of simple parameters. By assumption, the generic components 1 and 2 are either both orthogonal or both symplectic, while the unipotent components 1 and 2 of 1 and 2 are of opposite type. Therefore 1 and 2 are themselves of opposite type. This implies that the datum G G1 G2 is composite. It follows that ` ` ` r r r r r r (5.1.11) f G pq tr pf G q tr 1 pf 1 q tr 2 pf 2 q , f P HpN q, r where 1 b 2 is the automorphic representation in the packet for G obtained from Assumption 5.1.1 and our induction hypothesis. We would like to argue as in the proof of Lemma 5.1.2. However, we do not know that the identity (5.1.8) is valid. The problem is that we r do not have a proof of Lemma 4.3.1 (with GpN q in place of G) for the parameter . This alters the contribution of Corollary 4.3.3 to (5.1.8). We must describe the resulting modication of the identity. With the limited information presently at our disposal, we can make use r of Corollary 4.3.3 (with GpN q in place of G) only if the summand in (4.3.9) is multiplied by the quotient
N r pwu q N,1 puq N pxu q1 sN,0 puq1

r of the two sides of the analogue for GpN q of the unproven identity of Lemma rN 4.3.1. There is of course only one summand, since the centralizer S is r connected. Given that S N is also an abelian torsor, we deduce that

N,1 puq N pxu q N,0 puq 1. The remaining factor in the quotient satises `1 N r pwu q r pwq 2 , 1 2 , according to (5.1.4). By Assumption 5.1.1, the linear forms in (4.3.9) and (4.4.11) again satisfy the global intertwining relation
1 1 fG p, s xq fG p, xq fG p, xq,

5.1. ON THE CASE OF SQUARE INTEGRABLE

261

r r but with GpN q and f again in place of G and f . This equals the trace (5.1.11). We can now compare the modied form of (4.3.9) with the corresponding expression (4.4.11). We will then be led to an explicit, possibly r nonzero contribution of the composite group G P Eell pN q to the sum (5.1.8). In fact, by substituting the formulas for the coecients above, we nd that the left hand side of (5.1.8) equals ` ` 1 1 r (5.1.12) C p 2 , 1 2 q 1 tr pf G q ,
1 where C is the product of the positive constants C and e pxq. This is the modication of (5.1.8), an identity that could also have been obtained directly from rst principles rather than the general formulas of Corollaries 4.3.3 and 4.4.3. We have shown that the left hand side of (5.1.8) equals (5.1.12). In the r summand of G P Esim pN q in (5.1.8), we can use (4.4.12) to write ` 0 r f P HpG q, (5.1.13) Sdisc, pf q tr Rdisc, pf q ,

r since does not belong to pG q. In the identity itself, we can replace the r u by an arbitrary compatible family of functions family tf ( r r f P HpG q : G P Eell pN q , writing f f in case G G. We conclude that ` ` 2 rpN, G q tr Rdisc, pf q ` C tr pf q 0,
r G PEsim pN q

for the nonnegative coecient ` 2 1 1 C C 1 p 2 , 1 2 q . The left hand side of this expression can be written as a nonnegative linear combination of irreducible characaters. Proposition 3.5.1 then tells us that each of the coecients vanishes. In particular, the right hand side of (5.1.13) equals 0. This gives the rst assertion of the lemma. From the vanishing 2 1 of the remaining coecient C , and the fact that C is strictly positive, we see also that `1 1 2 , 1 2 0. This is the second assertion of the lemma. Lemma 5.1.6 is an important complement to the previous lemmas and their corollaries. It completes our discussion of square integrable parameters r r P F2 pGq attached to composite endoscopic data G P Eell pN q. We now know that these parameters do not contribute to the discrete spectrum of any r simple endoscopic datum G P Esim pN q. Lemma 5.1.6 also gives a reduction of our induction hypothesis, as it applies to Theorem 1.5.3(b). To be precise, we have shown that the number (5.1.4) attached to an -pair p1 , 2 q is equal to 1 whenever degp1 q ` degp2 q N.

262

5. A STUDY OF THE CRITICAL CASES

With this step, we can retire the terms -pair and -parameter from our lexicon! 5.2. On the case of elliptic In the last section, we looked at square-integrable parameters ( r r (5.2.1) P 2 pGq : G P Eell pN q . We resolved what remained to be done when G lies in the complement of r r Esim pN q in Eell pN q. We shall now add some observations for the other critical case, that of elliptic parameters ( r r (5.2.2) P ell pGq : G P Esim pGq , or more precisely, parameters in (5.2.2) that lie in the complement of (5.2.1). Recall that the sets (5.2.1) and (5.2.2) can both be identied with subsets of r pN q. Their union is precisely the set of parameters that were not treated by the general induction arguments of Chapter 4. Their intersection, the set of square integrable parameters for simple data G, will be the most dicult family to treat. Among other things, we will eventually have to show that the linear form of Lemma 5.1.4 vanishes. r The elliptic parameters P ell pGq that are not square-integrable have their own set of diculties. These stem from the global intertwining relation for G, since we are not yet in a position to prove this identity. We will G therefore not be able to establish the predicted formula for Sdisc, pf q, which reduces to 0 in this case, or to establish the expected property that the G representation Rdisc, pf q vanishes. Another consequence is that we will not yet be able to resolve the case r 1 of (4.5.11) left open from 4.5. In this section, we shall put together what can be obtained from the standard model. We will introduce formulas to which the unresolved intertwining relation contributes an obstruction term. These formulas will eventually be strengthened to yield what we want. They will also be applied slightly dierently in the next section to give more information about the square integrable parameters of 5.1. For the rest of the chapter, we will be following the global notation of the last section (and the previous two chapters). In particular, F is a global r r eld, while G P Eell pN q and P pGq are of course to be understood as global objects over F . In this section, we will be assuming that G is simple. It will then be convenient to denote the complement of an embedded set in r (5.1.1) in its successor by the appropriate superscript. Thus sim pGq is the 2 2 pGq is the complement of pGq r r r r2 complement of sim pGq in 2 pGq, while ell r ell pGq. We therefore have a decomposition in r sim r2 r r (5.2.3) ell pGq sim pGq 2 pGq ell pGq. r It is the set 2 pGq that we will be looking at in this section. ell

5.2. ON THE CASE OF ELLIPTIC

263

r We will ultimately be focused on a family F of parameters (5.1.2), on which we will later impose local constraints. Our present concern will therefore be with the subset r r r r F 2 pGq F X 2 pGq, G P Esim pN q,
ell ell

r r of F. There is actually no compelling reason to work with F in much of this section, since Assumption 5.1.1, the one axiom we have imposed beyond r2 our blanket induction hypothesis, is irrelevant to parameters in Fell pGq. r2 However, we may as well state the results for Fell pGq, since it is in this context that they will be applied. r Suppose then that N and G P Esim pN q are xed, and that is a given r2 pGq. Then is characterized by the conditions parameter in the set Fell # 21 2q q`1 r , ` ` (5.2.4) S Op2, Cqq Op1, Cqrq , q 1, along with the requirement that the Weyl group W contain an element w in the regular set W,reg . Recall that p q` denotes the kernel of a sign charac ter, which was dened in terms of the degrees Ni of the simple components i of . The condition on the existence of w is that there be an element in S whose projection onto any factor Op2, Cq belong to the nonidentity connected component. This holds either if q r, or if there is an even number of indices i with Ni odd. We may as well write ( S,ell x P S : E,ell pxq H for the subset of components in the quotient S of S that contain elliptic elements. Then S,ell is also the set of components x which contain a regular Weyl element wx P W,reg pxq. This element is unique. r We write M for a Levi subgroup of G such that belongs to M pGq. We can then identify the multiplicity free parameter M 1 r r with an element in 2 pM, q. We can also write M GLpN1 q GLpNq q G as in the local notation (2.3.4). Then if q`1 r and N Nq`1 ` ` Nr , r the group G G represents the unique element in Esim pN q such that r2 pG q. Our condition q 1 is of course because does not lies in F r lie in F2 pGq. It implies that M is proper in G, and that we can apply our induction hypotheses to M and M .

264

5. A STUDY OF THE CRITICAL CASES

The rst step is to apply the results of Chapter 4 to the component r r r GpN q, as we did in the last section. The centralizer set S pN q for GpN q r pN q contains is connected, for any parameter at all, so the quotient S r one element x. The global intertwining relation for GpN q and is still r valid in this case. It is the formula (4.5.1), with GpN q in place of G, which r we established inductively for any parameter in the complement of ell pN q, near the beginning of the proof of Proposition 4.5.1. (It was of course the r exceptional set ell pN q that we treated in the last section, where we relied on Assumption 5.1.1.) In particular, the conditions of Corollaries 4.3.3 and ` r 4.4.3 apply to GpN q, , and the corresponding terms in the two expansions (4.3.9) and (4.4.11) are equal. The expansions are themselves then equal, and the two corollaries yield the familiar identity
0 N r r rdisc, pf q

rdisc, r 0 sN pf q,

r r f P HpN q.

Since
0 N r r rdisc, pf q

` N r tr Rdisc, pf q 0

in this case, it follows from the denition (4.4.10) that r p rpN, G q 0 Sdisc, pf q 0. (5.2.5)
r G PEsim pN q

r2 This is the same as the identity (5.1.8), but for the parameter P Fell pGq r rather than the parameter in F2 pGq of 5.1. We shall use it in much the same way. The next lemma is stated in terms of a compatible family ( r r (5.2.6) f P HpG q : G P Eell pN q , or more precisely, the subset of functions in a compatible family parametrized r r by the subset Esim pN q of Eell pN q. If G G, we write f f . In case N r is even and G 1, so that there is a second group G_ P Esim pN q with _ f , if G G_ . G_ 1, we will also write f Lemma 5.2.1. Suppose that for the given pair pG, q, r r2 G P Esim pN q, P Fell pGq, the index r in (5.2.4) is greater than 1. Then the sum ` rpN, G q tr Rdisc, pf q (5.2.7)
r G PEsim pN q

equals (5.2.8) c
xPS,ell

` 1 G pxq fG p, s xq fG p, xq ,

for a positive constant c cpG, q, and any compatible family of functions (5.2.6).

5.2. ON THE CASE OF ELLIPTIC

265

Proof. The general strategy of proof is pretty clear by now. We need to establish a formula for the sum (5.2.7). We have a formula for the parallel sum on the left hand side of (5.2.5). For each index of summation r G P Esim pN q, we have then to obtain an expression for the dierence ` r (5.2.9) tr Rdisc, pf q 0 Sdisc, pf q, f P HpGq, of the corresponding two summands. We shall apply the standard model, specically Corollaries 4.3.3 and 4.4.3, to the pair pG , q. We observe that the two corollaries would have been easy to establish directly in the case at hand, since the centralizer S S G for (5.2.9) is considerably simpler than the general object S treated in 4.3 and 4.4. The most important case is that of G G. It is responsible for the 1 linear forms fG p, s xq and fG p, xq in (5.2.8). We recall that these objects are dened by (4.2.4) and (4.2.5). They satisfy the conditions of Corollaries 4.4.3 and 4.3.3, with pG, q being our given pair. In other words, they are well dened functions of the component x in S . For the rst function 1 fG p, s xq, this fact was established under general conditions prior to the proof of Proposition 4.5.1. For the function fG p, xq, it follows from the special nature of , specically, the fact that there is only one element wx in the Weyl set W pxq. We can therefore apply Corollaries 4.3.3 and 4.4.3 to our pair pG, q. r Together, they tell us that for any f P HpGq, the dierence ` G ` G G Idisc, pf q 0 sG pf q Idisc, pf q 0 rdisc, pf q disc,
G 0 rdisc, pf q 0 sG pf q disc,

equals C
xPS

` 1 i pxq G pxq fG p, s xq fG p, xq .

On the one side of the identity, we can write ` G 0 G G rdisc, pf q 0 sG pf q tr Rdisc, pf q 0 Sdisc, pf q. disc, This follows from the denitions, since the expected value of the trace of G Rdisc, pf q is 0. It also follows from the denitions that the coecient i pxq on the other side vanishes unless x belongs to the subset S,ell of S . In this case, we have 0 i pxq |W |1 s0 pwq | detpw 1q|1
wPW,reg pxq

1 s0 pwx q | detpwx 1q|1 ` q 11 1 . 2 The other side of the identity therefore equals ` q ` 1 C 1 G pxq fG p, s xq fG p, xq . 2
xPS,ell

266

5. A STUDY OF THE CRITICAL CASES

This gives a formula for (5.2.9) in case G G. r Suppose that G is an element in Esim pN q that is distinct from G. We r claim that the dierence (5.2.9) vanishes. If does not belong to pG q, this is just the identity (4.4.12). In fact, we already know that both sides of r (5.2.9) vanish unless G . Suppose then that does belong to pG q. Then G G . are distinct elements in E rsim pN q, this implies that G 1, Since G and G G_ , r q, and G ` q _ S S pG_ q Spp2, Cq . By assumption, r q 2. We are therefore dealing with a case that was treated by induction in 4.5. Indeed, the pair pG_ , q satises the condition dimpT q dimpT ,x q q 2,
_ _ _ _ _ x P S , _

since the subtorus T ,x of the maximal torus T of S dened in 4.5 _ actually equals T in this case. As in the discussion of the conditions (i) and (ii) in the proof of Proposition 4.5.1, we see that the global intertwining relation (4.5.1) holds for pG_ , q. The dierence (5.2.9) therefore vanishes in the remaining case G G_ , by Corollaries 4.3.3 and 4.4.3. The last step is to apply the identity (5.2.5). We can replace the funcr tions tf u in (5.2.5) by any compatible family (5.2.6). Substituting the formulas we have obtained in the cases G G and G G, we conclude that (5.2.7) equals (5.2.8), with ` 1 q c rpN, Gq C 2 . Remark. The only property of the constant c in (5.2.8) we will use is its positivity. We could of course write it more explicitly. It follows from (3.2.6) and (4.3.5) that ` 1 q c rpN, Gq C 2 ` q 1 r 2 |OutN pGq|1 m |S |1 1 2 ` 1 1 1 q 1 r 2 |OutN pG, q| |S | 2 , q`r` and one can check that this equals 1 , where 0 if N is even and 2 1 if n is odd. We have to treat the case of r q 1 in (5.2.4) separately. The underr2 lying object of course remains a parameter in the complement Fell pGq of r r F2 pGq, for some G P Esim pN q. However, we now suppose that it falls into the special case # 21 (5.2.10) S Op2, Cq

5.2. ON THE CASE OF ELLIPTIC

267

of (5.2.4). This implies that N is even and that G 1. We thus have a r pair of split groups G and G_ in Esim pN q, with
_ S S pG_ q Spp2, Cq.

In this case, the group M GLpN1 q can be identied with a Levi subgroup of both G and G_ . If G equals either G or G_ , the subset ( S,ell x P S : E,ell pxq H
of the group S S pG q consists of one element, which we denote simply by x1 . In each case, the Weyl set W,reg px1 q consists of one element w1 .

Lemma 5.2.2. Suppose that for the given pair pG, q, r r2 G P Esim pN q, P Fell pGq, the index r in (5.2.4) equals 1. Then the sum of ` rpN, G q tr Rdisc, pf q (5.2.11)
r G PEsim pN q

and (5.2.12) equals (5.2.13)

1 8 1 8

_ pf _ qM p1 q fG_ p, x1 q

1 fG p, s x1 q fG p, x1 q ,

for any compatible family of functions (5.2.6). Proof. The proof is essentially the same as that of the last lemma. r For each G P Esim pN q, we have to obtain an expression for the dierence (5.2.9) that we can substitute into the identity (5.2.5). In particular, we have to apply Corollaries 4.3.3 and 4.4.3 to the data G G and G G_ . We require separate discussion here simply because we know less about the contribution from G G_ . That is, the inductive arguments of 4.5 are not strong enough to establish the expected global intertwining relation for G_ . This will account for the extra term (5.2.12). We can say that the conditions of the two corollaries are applicable to both G and G_ . The justication follows from the remarks at the beginning of the proof of the last lemma, except in the application of Corollary 4.3.3 to G_ . In this case, we simply adopt the convention used to treat (4.5.11) in the proof of Proposition 4.5.1, which relies on the fact that the element wx P W,reg pxq introduced earlier is unique. Once again, we shall apply both corollaries together. Assume that G is one of the groups G or G_ . For any r f P HpG q, the dierence ` 0 0 rdisc, pf q 0 s disc, pf q tr Rdisc, pf q Sdisc, pf q

268

5. A STUDY OF THE CRITICAL CASES

then equals
C

xPS

` i pxq G pxq pf q1 p, s xq fG p, xq , G

where C is the constant (4.3.5) attached to G . It follows easily from Corollary 4.6.2 and the original denition (1.5.6) that the sign character G equals 1. Since i pxq vanishes unless x belongs to S,ell , a subset of S that consists of the one element x1 , the last expression therefore reduces to ` (5.2.14) C i px1 q pf q1 p, s x1 q fG p, x1 q . G

If G G, the identity component of the group S S is abelian. We have ` 1 q 1 i px1 q 2 2 , as in the proof of the last lemma. Since rpN, Gq C
1 2

1, 8

the product of rpN, Gq with (5.2.14) in this case is equal to the expression (5.2.13), with f f . If G G_ , the group S S _ is connected, and simple of rank 1. We obtain _ s0 pwq | detpw 1q|1 i px1 q |pW q0 |1
_ wPW,reg px1 q

1 2

s0 pwx q | detpwx 1q|1 1 , 4

as in the proof of the case (4.5.11) in Proposition 4.5.1. We also have the reduction pf q1 p, s x1 q pf q1 p, 1q pf qM pq G G in (5.2.14). Since _ 1 rpN, G_ q C 4 1 , 8 the product of rpN, G_ q with (5.2.14) in this case is equal to p1q times the expression (5.2.12), with f _ f . The minus sign here is critical. It is what forces us to place (5.2.12) and (5.2.13) on opposite sides of the proposed identity. We continue the argument at this stage as in the proof of the last lemma. r If G belongs to Esim pN q, but is not equal to G and G_ , does not belong r q. The dierence (5.2.9) then vanishes by (4.4.12) (or simply from to pG the fact that G ). It remains only to apply (5.2.5), with a general r compatible family (5.2.6) in place of the functions tf u. Substituting the G, G G_ , and G distinct formulas we have obtained in the cases G from G and G_ , we conclude that the sum of (5.2.11) and (5.2.12) equals (5.2.13), as required. We have so far worked exclusively with parameters of degree equal to the xed integer N of our running induction hypotheses. This will not be sucient. We will soon have to consider some elliptic parameters of degree

5.2. ON THE CASE OF ELLIPTIC

269

greater than N . For such objects, we will of course not be able to rely on the results of this section. We will need to see what can be salvaged. Suppose that ` 1 1 r r r is a general parameter in our family FpN` q, with degree N`
1 N1

` `

r Nr

greater than N . We may as well assume that ` belongs to the subset r r Fdisc pN` q of FpN` q. In other words, the simple components i of ` are self-dual. We dene `, i ,
i

odd

r an auxiliary parameter in Fell pN`, q, where N`, Ni .

i

odd

r Suppose that ` belongs to pG` q, for a given simple datum r G` P Esim pN` q. Following earlier notation, we write M` for a Levi subr group of G` such that ` belongs to M` pG` q. Then M` GLpN1 q 1 GLpNr q r G`, , where 1 r i {2s is the greatest integer in i {2, and G`, is the unique i r r element in Esim pN`, q such that `, lies in 2 pG`, q. We can identify the corresponding parameter `,M` 11 r1 `, r with an element in 2 pM` , ` q. We will naturally have an interest in the r stable linear form on HpG` q attached to ` . It can be written as the pullback r f G` p` q f M` pM` q, f P HpG` q, of a stable linear form on M` pAq. This linear form will thus be well dened r by our Assumption 5.1.1 on FpN q, so long as N`, N . There is one part of the standard model that can be established for many r ` . It is a variant of (5.2.5), in which the sum is taken over Eell pN` q instead P E pN q, the stable linear form S r rell ` of its subset Esim pN` q. For any G disc,` r q is dened by (3.3.9), (3.3.13) and Corollary 3.4.2. Its expected on HpG value is given by the analogue of Corollary 4.1.3, as a linear combination of linear forms r f p q, f P SpG q, P pG , ` q. These objects are again well dened by Assumption 5.1.1 whenever N`, N . We can therefore dene the stable linear form
0 Sdisc,` pf q,
1 1 1 1

r f P HpG q,

270

5. A STUDY OF THE CRITICAL CASES

as the dierence between Sdisc,` pf q and its expected value, so long as N`, N .

r Lemma 5.2.3. Suppose that ` P Fdisc pN` q is as above, and that (5.2.15) Then (5.2.16)
r G PEell pN` q

N1 ` ` Nr N. r p rpN` , G q 0 Sdisc,` pf q 0, r r f HpN` q.

r Proof. We are taking the sum in (5.2.16) over the full set Eell pN` q for the obvious reason that there is only limited scope for the inductive applir p cation of Theorem 4.1.2. That is, we cannot say that 0 Sdisc,` pf q vanishes r for general elements G in the complement of Esim pN` q. We do know that this linear form is well dened, since the condition (5.2.15) obviously implies that N`, N . We also know that Lemma 4.3.1 is valid for the pair ` r GpN` q, ` , which is to say that the application of Theorem 1.5.3 to the proof of Lemma 4.3.1 in 4.6 remains valid. The assertion (a) of Theorem r 1.5.3 is not actually relevant to GpN` q, since in this case it is only RankinSelberg L-functions that contribute to the factor s0 pwq in (4.6.13). The condition on the other assertion (b) still holds, by virtue of its extension in Lemma 5.1.6 and the condition (5.2.15). Lemma 4.4.1 is of course also valid ` r for GpN` q, ` , since there were no implicit conditions in its proof from 4.6. With these remarks, we can see that the discussion preceding the special r case (5.2.5) for GpN q applies here as well. In particular, the conditions of Corollaries 4.3.3 and 4.4.3 remain` in force, and the analogues of these r corollaries hold for the general case GpN` q, ` at hand. Moreover, the ` r analogues for GpN` q, ` for the expressions (4.3.9) and (4.4.11) are equal. It follows that N 0 N` r r r r r r f P HpN` q, pf q 0 s ` pf q, r
disc,` disc,`

where

0 sN` r rdisc,` pf q

equals the left hand side of (5.2.16). Since

0 N` r rdisc,` pf q r

0,

we obtain the required identity (5.2.16). 5.3. A supplementary parameter ` In this section, we shall return to the square integrable parameters r P Fell pN q of 5.1. Recall that takes the general form r (5.3.1) 1 r , i Fsim pNi q, r r r and lies in the subset F2 pGq of Fell pN q attached to some G P Eell pN q. (Recall also that G is unique, except possibly if is generic and simple. In the

5.3. A SUPPLEMENTARY PARAMETER `

271

r latter case, we simply x a G with P Fsim pGq and treat it as if it were unique, knowing that this is what we will eventually establish.) Our goal is G to establish the stable multiplicity formula for Sdisc, pf q. In 5.1, we found that the proof was rather direct under any of the conditions N odd, G 1 r r or G Esim pN q. However, the remaining case of N even and G P Esim pN q split is considerably harder. We will need to establish it for the relevant r parameters in Fell pN q in order to complete the induction argument of this chapter. In the special case that is simple generic, we will also need to deal with the rst part (a) of Theorem 1.5.3. This is the condition on the poles of Lr functions, which we recall is intimately tied up with the denition of Fsim pGq and the corresponding special case of the stable multiplicity formula. It represents a second critical global assertion that will have to be established r in the context of the family Fell pN q. The other global assertions are less pressing. They will be resolved later, in terms of the stable multiplicity formula and the local results we establish. We recall that the second part r (b) of Theorem 1.5.3 has already been resolved for FpN q. It was established as one of the assertions of Lemma 5.1.6. r To deal with these questions, we can assume that G P Esim pN q is simple. r Given the parameter P F2 pGq, we set N` N1 ` N, where the constituents i of have been ordered so that N1 Ni for each i. We then introduce the supplementary parmeter (5.3.2) ` 21 2 r

r r r in FpN` q. There is a unique element G` in Esim pN` q such that ` P pG` q, p ` contains the product and such that G p p p p G1 G G1 G, r where G1 G1 is the element in Esim pN1 q such that 1 is contained in r r r Fsim pG1 q. Then ` lies in the subset Fell pG` q of FpG` q. We will need to work with two maximal Levi subgroups of G` . The rst applies to the case ` Neven and 1, in which we have a maximal of 1 Levi subgroup L GL 2 N of G. We then have the corresponding Levi subgroup ` L` G1 L GL 1 N G1 2 of G` . It is adapted to the decomposition ` 1 1 of , and comes with the linear form f L` f L` p1 q, r f P HpG` q,

272

5. A STUDY OF THE CRITICAL CASES

where is the linear form attached to L in Lemma 5.1.4. The second is a special case of the general denition prior to Lemma 5.2.3. This is the Levi subgroup r M` GLpN1 q G`, , G`, P Esim pN`, q, r of G` such that ` belongs to M pG` q. It is adapted to the decomposition
`

` 21 `, , of , and comes with the linear form f f M` pq,

`, 2 r

r f P HpG` q,

in which is identied with the product 1 `, . One of our ultimate goals is to show that vanishes. An essential step will be to establish analogues of Lemmas 5.2.1 and 5.2.2 for the parameter ` . These will be needed for the induction argument that is to drive the proof of the local theorems over the next two chapters. They will also be used directly for the general proof of the global theorems in Chapter 8. As with their predecessors, we shall state the lemmas in terms of a compatible family of functions ( r r (5.3.3) f P HpG q : G P Eell pN` q , with the same notation f f and f _ f in the cases G G` and G G_ . We shall again deal separately with the cases r 1 and r 1. ` Suppose rst that r 1. This is the case that lies in the complement rsim r r F2 pGq of Fsim pGq in F2 pGq. The problem is to understand the linear form of Lemma 5.1.4, and to show ultimately that it vanishes. We can therefore assume that N is even and 1, as in the earlier lemma, and in particular that G is split. We then have the maximal Levi subgroups L and L` of G and G` respectively. To deal with this case, we shall also have to strengthen our induction hypothesis. We shall assume that in dealing with the square integrable parameter here, we have already been able to treat the elliptic parameters of 5.2. Lemma 5.3.1. Suppose that for the given pair pG, q, r r G P Esim pN q, P F2 pGq,

the index r in (5.3.1) is greater than 1, while N is even and 1. Assume also that the global theorems are valid for parameters in the complement of r r Fell pN q in FpN q. Then the sum ` ` rpN` , G q tr Rdisc,` pf q ` b` f L` p1 q (5.3.4)
r G PEsim pN` q

equals (5.3.5) c`
xPS` ,ell

` 1 G` ` pxq fG` p` , s` xq fG` p` , xq ,

5.3. A SUPPLEMENTARY PARAMETER `

273

for positive constants b` bpG` , ` q and c` cpG` , ` q, and any compatible family of functions (5.3.3). Proof. This is the rst of two lemmas in which we have to work with parameters of rank greater than N . The proofs require extra discussion, since they are beyond the scope of our running induction hypotheses. The supplementary arguments are usually not dicult, but they do contain a number of points to be checked. We shall include more detail here than in the next lemma. We may as well start with a general compatible family (5.3.3). We then r r r x a function f P HpN` q that for any G P Eell pN` q has the same image r q as the corresponding function f in (5.3.3). We can then work in SpG r interchangeably with the function f or the family (5.3.3). This will simplify the notation at times, particularly when G equals G. In principle, the lemma should simply be the formula for ` rpN` , G q tr Rdisc,` pf q (5.3.6)
r G PEsim pN` q

that would be the direct analogue for ` of Lemma 5.2.1. In particular, the coecient in (5.3.5) is to be the analogue c`
1 2

rpN` , G` q C`

of the coecient c in the corresponding expression (5.2.8) of Lemma 5.2.1. However, the induction assumptions hold only proper subparameters of . We will see that most proper subparameters of ` can either be ruled out by general considerations, or expressed in terms of proper subparameters of . For the ones that cannot, we will be able to use to describe the defect in the associated terms. In other words, we will keep track of what changes the existence of makes in the arguments used to derive the earlier formula. The discussion will again center around a sum r r r p rpN` , G q 0 S (5.3.7) pf q, f P HpN` q,
disc,` r G PEell pN` q

over endoscopic data. Recall that this is the left hand side of the formula (5.2.16), established for more general ` in Lemma 5.2.3. In the case at hand, the sum N1 ` N2 ` ` Nr in (5.2.15) equals N . The formula therefore holds, and (5.3.7) vanishes. The problem is to analyze the summands in (5.3.7). We shall apply what we can of the standard model for G to the stable linear form 0 r S pf q 0 S G pf q, f P HpG q,
disc,` disc,`

r r where f maps to the function f P SpG q in (5.3.7). The problem now ranges over E pN q, rather than the rell ` is complicated by the fact that G rsim pN` q of simple data whose analogue for N indexed the sum in subset E

274

5. A STUDY OF THE CRITICAL CASES

the original formula (5.2.5). We recall that 0 Sdisc,` pf q is dened as the dierence between the original stable form Sdisc,` pf q, and its expected formula (5.3.8) |S |1 pS 0 q p q f p q PF pG ,` q

given by the twisted analogue of Corollary 4.1.3. Consider a general element G G1 G 2 , r Gk P Esim pN k q, k 1, 2,

r in Eell pN` q. Thus, N` N 1 ` N 2 , and G1 and G2 represent the two data denoted GS and GO in Assumption 5.1.1. We can write Sdisc, pf q p5.3.9paqq Sdisc,` pf q

and p5.3.9pbqq
0 Sdisc,` pf q

0 Sdisc, pf q,

where the sum in each case is over parameters 1 2, with ` 1 2 . The decomposition (5.3.9(a)) follows from (3.3.9), (3.3.13) and Corollary 3.4.2. Since the expected value (5.3.8) of the left hand side of (5.3.9(a)) has a similar decomposition, with the summand of any in the compler ment of FpG , ` q taken to be 0, the dierence does satisfy (5.3.9(b)). We observe also that the summand of in (5.3.9(a)) has an obvious product decomposition (5.3.10)
1 2 Sdisc, pf q Sdisc,1 pf 1 q Sdisc,2 pf 2 q,

r k P FpN k q, k 1, 2,

if f f 1 f 2 . We can assume that the two nonnegative integers N 1 and N 2 in the partition of N` attached to G satisfy N 1 N 2 . The most important cases will then be the trivial partition (i) N 1 0, N 2 N` , r in which G lies in the subset Esim pN` q of simple data, and the partition (ii) N 1 N1 , N 2 N attached to 1 and . Before we discuss these, however, we shall rst take care of the others.

5.3. A SUPPLEMENTARY PARAMETER `

275

If 0 N 1 N1 , there is no index of summation in (5.3.9), since N1 Ni for each i. It follows that (5.3.11)
0 Sdisc,` pf q

0,

in this case. In the remaining case that N 1 N1 , both N 1 and N 2 are less than N . We can then apply our induction hypothesis to the factors on the right hand side of (5.3.10). This tells us that each factor equals its expected value, and so therefore does the product. It follows that the summands in (5.3.9(b)) all vanish, and that (5.3.11) holds in this case as well. Thus, the summand of G in (5.3.7) vanishes unless the corresponding partition N` N 1 ` N 2 satises (i) or (ii). Consider the case (ii). This of course does not characterize G uniquely. Moreover, there are several indices 1 2 in (5.3.9(b)) attached to the given G . Among these, it is the pair (5.3.12) pG , q pG1 G_ , 1 q, G1 G1 ,

that is the most signicant. We shall deal rst with the other pairs. Assume that pG , q is any pair associated with the partition (ii) that is not equal to (5.3.12). If the parameter 1 2 is not equal to 1 , the second component 2 contains the factor 1 with multiplicity r r two. This means that 2 lies in the complement of Fell pN q in FpN q. The given condition of the lemma then tell us that the global theorems are valid 2 for pG2 , 2 q, and in particular, that the factor Sdisc,2 pf 2 q in (5.3.10) equals its expected value. Our induction hypothesis tells us that the other factor 1 Sdisc,1 pf 1 q in (5.3.10) also equals its expected value, since N 1 N1 is less than N . It follows that the summand of in (5.3.9(b)) vanishes. Assume next that 1 , but that the datum G G1 G2 is not equal to r G1 G_ . If G1 G1 the parameter 1 1 does not lie FpG1 q. It then follows by induction that
0 1 Sdisc,1 pf 1 q 1 Sdisc,1 pf 1 q 0,

and therefore that the summand of again vanishes. On the other hand, if G1 equals G1 , the second component G2 cannot equal G. This is because p p G1 and G are both orthogonal or both symplectic (by denition, since the p p same is true of the parameters 1 and ), whereas the factors G1 and G2 of must be of opposite type. Since we p the dual twisted endoscopic group G have ruled out G2 G_ , the only other possibility is a datum that satises G2 2 1. In this case,
0 G2 Sdisc,2 pf 2 q G Sdisc,2 pf 2 q 0,
2

by (3.4.7), and so the summand of in (5.3.9(b)) vanishes once again.

276

5. A STUDY OF THE CRITICAL CASES

Suppose then that pG , q equals (5.3.12). The decomposition (5.3.10) takes the form
G1 G Sdisc, pf q Sdisc,1 pf1 q Sdisc, pf _ q,
_

G1 G1 ,

where f f1 f _ . The rst factor becomes

G G G1 Sdisc,1 pf1 1 q f1 1 p1 q,

with an inductive application of Theorem 4.1.2 to G1 . According to Lemma 5.1.4, the second factor satises
G Sdisc, pf _ q opG_ q f _,L pq. r
_

It follows that
0 Sdisc, pf q r Sdisc, pf q opG_ q f L` p1 q,

G since the expected value of Sdisc, pf q is 0. This provides the contribution to the sum (5.3.7) of G G1 G_ . In fact, from what we have shown, it represents the only contribution from composite endoscopic data G . To be precise, the sum of those terms in (5.3.7) corresponding to data G in the r r complement of Esim pN` q in Eell pN` q equals

(5.3.13) G

rpN` , G1 G_ q opG_ q f L` p1 q. r

It remains to consider the partition (i) attached to simple data r P Esim pN` q. This represents the point at which we began the proof of Lemma 5.2.1, the model we are following from 5.2. For each such G , we have to obtain an expression for the dierence ` r f P HpG q, (5.3.14) tr Rdisc,` pf q 0 Sdisc,` pf q, of the corresponding two summands in (5.3.6) and (5.3.7). The most important case is the datum G G` . It has the property that G equals the sign character G` 1 1 r of ` . It is in fact the only element in Esim pN` q with this property unless N1 is even and 1 1, in which case G` is split, and there is a second r r split datum G_ P Esim pN` q. The remaining elements G P Esim pN` q can be ` ignored, since ` 0 Sdisc,` pf q Sdisc,` pf q tr Rdisc,` pf q 0, r by (3.4.7). In particular, for the elements G P Esim pN q distinct from G` _ , the dierence (5.3.14) vanishes. and G` Suppose rst that G G_ (so in particular, N1 is even and 1 1). ` We claim that (5.3.14) vanishes in this case as well. Since ` does not lie r in pG_ q, we would expect both terms in (5.3.14) to vanish, but we are of ` course not in a position to prove this. We also cannot appeal to (4.4.12), r since it applies only to parameters in pN q.

5.3. A SUPPLEMENTARY PARAMETER `

277

To study (5.3.14), we need to apply the expansions (4.1.1) and (4.1.2) (with G_ in place of G). The proper spectral terms in (4.1.1) correspond to ` _ proper Levi subgroups M` of G_ of which ` contributes to the discrete `
` spectrum, in the sense that the representation Rdisc,` is nonzero. Upon _ with this property. Indeed, any M _ reection, we see that there is no M` ` is either attached to a partition of N` that is not compatible with ` , _ in which case we apply Corollary 3.4.3, or M` is a product of groups for which we can combine our induction hypothesis with the fact that the set _ r 2 pM` , ` q is empty. It follows from (4.1.1) that ` G_ G_ ` ` Idisc,` pf _ q tr Rdisc,` pf _ q 0.

M_

The proper endoscopic terms in (4.1.2) are indexed by composite endoscopic r data pG_ q1 in Eell pG_ q. After further reection, we see that there is no com` ` _ q1 for which the associated linear form S 1 posite pG` disc,` is nonzero. Indeed, either the two factors of pG_ q1 give a partition of N` that is incompatible ` with ` , in which case we apply Proposition 3.4.1, or they are data for which we can combine our induction hypothesis with the fact that the set ` r G_ q1 , is empty. It follows from (4.1.2) that `
` ` Idisc,` pf _ q Sdisc,` pf _ q 0.

G_

G_

The vanishing of (5.3.14) then follows from the fact that the expected value G_ ` of Sdisc,` is 0. Suppose nally that G G` . In this case we have again to apply (4.1.1) and (4.1.2) (with G` in place of G). The expansions will now contain proper terms that are nonzero, so we will want to work with the renements represented by Corollaries 4.3.3 and 4.4.3. This will lead to a formula for the value ` G` G` r f P HpG` q, (5.3.15) tr Rdisc,` pf q 0 Sdisc,` pf q,
0 of (5.3.14) at G G` . The connected group S` is just SOp2, Cq, so we are still dealing with a rather elementary form of the standard model. However, we will have to be careful in applying our induction hypotheses to derive analogues of the corollaries for ` , since the condition

N` deg ` N is more serious in this case. The proper spectral terms for G` in (4.1.1) are attached to proper Levi subgroups. We claim that these terms vanish for Levi subgroups that are not conjugate to the group M` described prior to the statement of the lemma. We recall that the orthogonal or symplectic factor G`, of M` is the unique r r datum in Esim pN`, q, such that 2 pG`, q contains the subparameter `, 2 r .

278

5. A STUDY OF THE CRITICAL CASES

The claim follows easily from the original classication Theorem 1.3.2 for GLpN q, given the structure of Levi subgroups of G` , and the nature of ` . This leaves the terms attached to conjugates of M` . Corollary 4.3.3 is designed express the sum of such terms. With a little thought, and a brief review of our later justication of Lemma 5.2.1, we see that its proof in 4.3 extends to G` , since it requires only that the degree N`, of `, be less than N . The expansion (4.1.1) therefore behaves as expected. We see that the dierence ` G G r (5.3.16) I ` pf q tr R ` pf q , f P HpG` q,
disc,` disc,`

equals the corresponding expansion (4.3.9) of Corollary 4.3.3. Arguing then as in the proof of Lemma 5.2.1, we conclude that if we take the product of (5.3.16) with p1q times the corresponding coecient rpN` , G` q in (5.3.6), we obtain the contribution of the linear forms fG` p` , xq to the given expression (5.3.5). The proper endoscopic terms for G` in (4.1.2) are attached to composite endoscopic data r G1 G 1 G1 , G1 P Esim pN 1 q, k 1, 2,
` 1 2 k k

r in Eell pG` q. Corollary 4.4.3 is designed to express the sum of such terms. However, its extension to G` would require the stable multiplicity formula for the linear forms (5.3.17)
1 1 1 1 1 Sdisc,1 pf 1 q Sdisc,1 pf1 q Sdisc,1 pf2 q,
` 1 2

attached to parameters
1 1 1 ` 1 2 , 1 r 1 k P FpNk q, k 1, 2,

such that
1 1 ` 1 2 , 1 1 r and functions f 1 f1 f2 in HpG1 q. It is possible to treat most compos` 1 , 1 q with our induction hypotheses, augmented by the given ite pairs pG` ` r r assumption on parameters in the complement of Fell pN q in FpN q. In fact, after a moments consideration, we see that the stable multiplicity formula holds in all cases, with the possible exception of the pair 1 pG1 , ` q pG1 G, 1 q. `

Moreover, Lemma 5.1.4 can be applied to this exceptional pair. It tells us in this case that the linear form (5.3.17) equals the dierence between its expected value and the linear form (5.3.18) opGq f L` p1 q r Recalling our justication of Lemma 5.2.1 as needed, we see that apart from this defect, the proof of Corollary 4.4.3 in 4.4 extends to G` . It follows that the sum of the dierence G G r (5.3.19) I ` pf q 0 S ` pf q, f P HpG` q,
disc,` disc,`

5.3. A SUPPLEMENTARY PARAMETER `

279

with (5.3.18) equals the corresponding expansion (4.4.11). Arguing then as in the proof of Lemma 5.2.1, we conclude that if we add the expression (5.3.20) rpN` , G` q opGq f L` p1 q r

to the product of (5.3.19) with rpN` , G` q, we obtain the rest of the given ex 1 pression (5.3.5), namely the contribution from the linear forms fG` p` , s` xq. We have now nished our discussion of the dierence (5.3.15). It is of course equal to the dierence between (5.3.19) and (5.3.16). Our conclusion is that the sum of (5.3.20) with the product of rpN` , G` q and (5.3.15) equals the entire expression (5.3.5). This in turn completes our discussion of the full set simple endoscopic r data G P Esim pN` q. For we have already seen that the dierence (5.3.14) vanishes if G G` . Therefore the sum of (5.3.20) with the sum over G of the product of rpN` , G q and (5.3.14) equals (5.3.5). It thus follows that the sum of (5.3.20) with the spectral sum (5.3.6) equals the sum of (5.3.5) with the contribution of the simple data G to the endoscopic sum (5.3.7). We observed earlier that the contribution of the remaining composite data G to (5.3.7) equals (5.3.13). The sum of the two terms (5.3.13) and (5.3.20) equals b` f L` p1 q, for the positive coecient r r b` rpN` , G1 G_ q opG_ q ` rpN` , G` q opGq. This represents the supplementary summand in the original given expression (5.3.4). We have now shown that (5.3.4) equals the sum of (5.3.5) with the endoscopic sum (5.3.7). We have also observed that (5.3.7) vanishes, by Lemma 5.2.3. We have therefore completed the proof of Lemma 5.3.1, at last. r Suppose now that r 1. Then 1 belongs to the subset Fsim pGq of r the F2 pGq. In this case, ` 2, so that N` 2N is even and ` 1. r In particular, we have the two split groups G` and G_ in Esim pN` q. The ` maximal Levi subgroup M` GLpN q of G` becomes also a maximal Levi subgroup of G_ . ` This case carries the added burden of the assertion of Theorem 1.5.3(a) on the poles of L-functions. We shall set 1 if is not generic, or in case is generic, the Langlands-Shahidi L-function Lps, , _ q attached to ` _ _ a maximal parabolic subgroup P` M` N` of G_ has a pole at s 1. ` This is what Theorem 1.5.3(a) predicts. We set 1 if is generic, and it is the L-function Lps, , ` q attached to a maximal parabolic subgroup P` M` N` of G` that has a pole at s 1. In other words, 1 if Theorem 1.5.3(a) does not hold for . We recall that G` is dened in terms of the group G G assigned to . We recall further that G is dened by induction unless is (simple) generic,

280

5. A STUDY OF THE CRITICAL CASES

r in which case it was taken to be some preassigned element in Esim pN q such r r that lies in the subset Fsim pGq of Fsim pN q specied by the temporary denition of 5.1. In the latter instance, there remains some possible ambiguity in the choice of G. It will be resolved by the stable multiplicity formula, or equally well, the formula 1. We will of course have to establish both formulas. If N is even and 1, we have the two split groups G and G_ in r Esim pN q, with a common maximal Levi subgroup L GLp 1 N q. In this 2 case, we must also be concerned about the possible existence of the linear form . Recall that r r S _ pf _ q opG_ qf _,L pq, f _ P HpG_ q,
disc,

and that the right hand side represents a unitary character on G_ pAq. It will be convenient also to write r r f P HpGq, S G pf q opGq f G pq,
disc,

r for a stable linear form on HpGq. The formula of Lemma 5.1.4 then reduces to G r r Sdisc, pf q opGq f G pq opGq f L pq, since the coecients in the formula satisfy r r m |S |1 G pq opGq 1 1 opGq, in the case at hand. We then obtain f G pq f G pq ` f L pq, r f P HpGq.

As in the case of r 1 above, we also have the maximal Levi subgroup ` ` L` G L G GL 1 N GL 1 N q G 2 2 of G` . It comes with the stable linear form f f L` p q, r f P HpGq.

Lemma 5.3.2. Suppose that for the given pair pG, q, r r G P Esim pN q, P F2 pGq, the index r in (5.3.1) equals 1, while N is even and 1. Then the sum of ` ` rpN` , G q tr Rdisc,` pf q ` 1 f L` p q (5.3.21) 2
r G PEsim pN` q

and (5.3.22) equals (5.3.23)

1 8 1 8

_ pf _ qM` pq fG_ p` , x1 q `

` 1 fG` p` , s` x1 q fG` p` , x1 q ,

5.3. A SUPPLEMENTARY PARAMETER `

281

for any compatible family of functions (5.3.3), and elements x1 P S` ,ell as in Lemma 5.2.2.

Proof. The lemma will again be a modication of its counterpart from 5.2. We have thus to see what can be salvaged of the proof of Lemma 5.2.2. This time there will be modications introduced by the sign , as well as from the possible existence of . As in the proof of the last lemma, we shall work interchangeably with a general compatible family (5.3.3) and a xed r r r function f P HpN` q, which for any G P Eell pN` q have the same image in r q. SpG The sign is not hard to keep track of. The formula of Corollary 4.3.3 is based on the inductive assumption that the sign obeys Theorem 1.5.3(a). Its direct analogues for G` and G_ would hold only if equals 1. If we ` look to see what changes makes if the proof of Lemma 5.2.2 is applied to ` , we observe that we need only insert a coecient in the negative summands in the expressions (5.2.12) and (5.2.13) (or rather their analogues for G_ and G` ). The expressions so obtained are just (5.3.22) and (5.3.23). ` Our task then is to describe what change the possible presence of makes in the analogue ` rpN` , G q tr Rdisc,` pf q (5.3.24)
r G PEsim pN` q

of (5.2.11). It amounts to a re-examination of the relevant parts of the proof of Lemma 5.3.1, specically the ner analysis of the summands in r p rpN` , G q 0 Sdisc,` pf q.
r G PEell pN` q

The sum is the left hand side of (5.2.16). It vanishes, since the condition of Lemma 5.2.3 remains valid, so we can write (5.3.24) as the sum of ` p rpN` , G q trpRdisc,` pf qq 0 Sdisc,` pf q (5.3.25)
r G PEsim pN` q

with (5.3.26)
G

p rpN` , G q 0 Sdisc,` pf q,

r r G P Eell pN` q Esim pN` q.

This was of course also the foundation of the proof of Lemma 5.3.1. Having outlined the argument for this last lemma in considerable detail, we can aord to be briefer here. The eect of will be seen in the endoscopic contributions to (5.3.26) r r or (5.3.25) of proper products G1 G2 in any of the sets Eell pN` q, Eell pG` q rell pG_ q. (Products from the sets Eell pN` q and Eell pG` q were denoted r r or E ` G1 G2 and G1 G1 q in the proof of Lemma 5.3.1, while products from 1 2 r the set Eell pG_ q were seen to be innocuous.) The presence of will be a `

282

5. A STUDY OF THE CRITICAL CASES

consideration only if both G1 and G2 are taken from the set tG, G_ u. We have then to describe the change makes in the expected formula for (5.3.27)
G1 G2 1 2 Sdisc,` pf1 f2 q Sdisc, pf1 q Sdisc, pf2 q,

r fk P HpGk q,

r where G1 G2 is the group G GG_ in Eell pN` q, the group G1 GG ` ` _ q1 G_ G_ in E pG_ q. in Eell pG` q, and the group pG` ell ` Now the formula for (5.3.24) that would be given by the analogue of Lemma 5.2.2 for ` (modied by the signs as above) is just the sum of the expressions (5.3.22) and (5.3.23). However, this formula is predicated on the compound linear form Spf1 f2 q in each of the three cases of (5.3.27) being equal to its expected value. Suppose that the discrepancy 0 Spf1 f2 q is in fact nonzero. In the rst case, it contributes a nonzero summand to (5.3.26), which must then be added to (5.3.24) as a correction term for the formula to continue to hold. In the second and third cases, it alters the proper p endoscopic component of the corresponding term 0 Sdisc, pf q in (5.3.25). The change in the associated summand of (5.3.25) must then be subtracted from (5.3.24) as a correction term, for the formula to remain valid. These remarks are of course with the understanding that the function f1 f2 in r each case has the same image in SpG1 G2 q as the chosen compatible family. r o In the rst case, (5.3.27) equals the product of opGqrpG_ q with the linear form G L f1 pq f2 pq f GL p q f L` p q. The expected value of this linear form is 0. We have then to add a term equal to the product of the coecient rpN` , G q opGq opG_ q r ` r with f L` p q to (5.3.24). r In the second case, (5.3.27) equals the product of opGq2 with the linear form G G f1 pq f2 pq f GG p q. The expected value of this linear form equals f GG p q ` G ` G G G L L f1 pq f2 pq f1 pq ` f1 pq f2 pq ` f2 pq f GG p q ` f LG p q ` f GL p q ` f LL p q. We note that f LG p q f GL p q f L` p q, since f GG acquires a symmetry from the automorphism in AutpG` , G1 q ` that interchanges the two factors of G1 GG. We must therefore subtract `
1 4

5.3. A SUPPLEMENTARY PARAMETER `

283

the product of the coecient rpN` , G` q rpG` , G Gq opGq2 r with the dierence ` f GG p q f GG p q 2f L` p q ` f LL p q from (5.3.24). r In the third case, (5.3.27) equals the product of opG_ q2 with the linear form L L f1 pq f2 pq f LL p q. The expected value of this linear form is 0. We must therefore subtract the product of the coecient rpN` , G_ q pG_ , G_ G_ q opG_ q2 r ` ` with f LL p q from (5.3.24). We note that it is the coecient 1 rather than 1 that appears 8 4 in the latter two cases, since the associated two endoscopic data G G and G_ G_ each come with an extra automorphism that interchanges the two factors. Combining the three cases, we see that the sum total of what must be added to (5.3.24) is the linear form ` 1 L` p q ` 1 2f L` p q ` f LL p q 1 f LL p q 4f 8 8
1 2 f L` p q. 1 8 1 8

Its sum with (5.3.24) is the given expression (5.3.21). The formula of Lemma 5.2.2, adjusted for the lack of information we have about ` , thus tells us that (5.3.21) equals the sum of (5.3.22) and (5.3.23). The proof is complete. From the last proof, we can extract formulas for the stable distributions attached to pG` , ` q and pG_ , ` q. We record them here for use later, ` when we come to the general global classication in Chapter 8. Corollary 5.3.3. Under the conditions of the lemma, we have ` G` 1 ` G` r Sdisc,` pf q tr Rdisc,` pf q ` 4 opG` q fG` p` , x1 q f GG p q and ` G_ ` G_ ` ` _ r ` Sdisc,` pf _ q tr Rdisc,` pf _ q 1 opG_ q fG_ p` , x1 q ` f _,LL p q . 4 ` Proof. These formulas are implicit in the proof of the lemma. If G is equal to either G` or G_ , it would not be hard to compute the dierence ` ` Sdisc,` pf q tr Rdisc,` pf q directly from the general expansions (4.1.1) and (4.1.2). There will be one spectral term, corresponding to the Levi subgroup M` of G` or G_ , and one `

284

5. A STUDY OF THE CRITICAL CASES

endoscopic term, corresponding to the proper endoscopic datum G1 of G` ` or pG_ q1 of G_ . The point is that we have already done the calculations, ` ` and can in principle read these terms o from the lemma. The spectral terms are in the second summands of (5.3.23) and (5.3.22) respectively, the contributions of the terms with coecient p q. These summands must be taken over to the left hand side of the relevant part of the formula given by the lemma, and then multiplied by the inverse of the coecient pN` , G q that occurs in (5.3.21). Since
1 8

rpN` , G q1

1 4

r opG q,

we see that the spectral term in the resulting formula for (5.3.28) will be
1 4 Sdisc,` pf q,

G P tG` , G_ u, `

r opG` q fG` p` , x1 q

if G G` , and
_ 1 r ` 4 opG_ q fG_ p` , x1 q `

if G G_ . ` The endoscopic terms do not come from the rst summands in (5.3.23) and (5.3.22). For these summands represent only expected values. Moreover in the case of (5.3.22), the summand corresponds to the simple endoscopic datum G_ for G_ rather than pG1 q_ . However, the two endoscopic terms ` ` ` are easily extracted from the proof of the lemma. They are essentially the formulas for (5.3.27), in the cases that G1 G2 equals G1 and pG_ q1 . We ` ` of course have again to place each of these ingredients on the left hand side of the relevant part of the formula of the lemma. Accounting for the coecients, and inserting the extra sign p1q in the case of (5.3.22), we then see that the endoscopic term in the formula for (5.3.28) will be
1 r 4 opG` q f GG p q

if G G` , and 1 opG_ q f LL p q 4 r ` if G_ G_ . The stated formulas follow. ` Remark. If we change the conditions N even and 1 in the statement of the lemma, the formula simplies. For without these conditions, does not exist, and the last summand in (5.3.21) can be taken to be 0. In other words, (5.3.21) reduces to the sum (5.3.24) over G . In the particular case that N is odd, the term (5.3.23) can also be taken to be 0, since the set S` ,ell is then empty. It follows from a simplied form of the proof of the lemma that the sum of (5.3.24) with (5.3.22) vanishes if N is odd, and equals (5.3.23) if N is even and 1.

5.4. GENERIC PARAMETERS WITH LOCAL CONSTRAINTS

285

5.4. Generic parameters with local constraints Our aim is to apply the discussion of the last three sections to the proof of the local theorems. We shall do so in due course, by constructing particular r families F that among other things, satisfy the conditions of 5.1. We begin here by imposing some ad hoc local conditions on the members of a general r family F, with a view to strengthening the results obtained so far. Suppose that r r r r F FpN q, FpN q pN q,
N

r is a family of parameters in the general set . We assume as in 5.1 that r is the graded semigroup generated by its subset Fsim of simple elements. r F However, we shall also take on a temporary hypothesis of a dierent sort. r It consists of three local constraints on elements in F, which we state as r introduced in 5.1 and a positive integer follows in terms of notation for F N. r Assumption 5.4.1. There is a nite, nonempty set V V pFq of archimedean valuations of F for which the following three conditions hold. r r (5.4.1)(a) Suppose that P Fsim pGq, for some G P Esim pN q, and that r v P V . Then v P 2 pGv q. r rsim (5.4.1)(b) Suppose that P F2 pGq, for some G P Eell pN q. Then there is r ` pG q for any G P Esim pN q r a valuation v P V such that v does not lie in v v p p over v with G G. v r r2 (5.4.1)(c) Suppose that P Fdisc pGq for some G P Eell pN q. Then there is a valuation v P V such that the kernel of the composition of mappings S Sv Rv , S S pGq,

contains no element whose image in the global R-group R R pGq is regular. The intent of the rst condition (5.4.1)(a) is that Theorem 1.4.2 should r hold for simple generic parameters P Fsim pN q and valuations v P V . The condition in fact will allow us to bypass the possible ambiguity of the provir r sional denition of Fsim pGq, adopted for families F in 5.1. The remaining two conditions contain technical assertions that are less clear. They will make better sense once we see how to apply them later in this section, and after we have constructed families for which they hold in 6.3. r We x a family F that satises Assumption 5.4.1. We may as well r assume at this point that the elements in F are all generic, since it is to the

286

5. A STUDY OF THE CRITICAL CASES

local classication of tempered representations in Chapter 6 that they will be applied. However, we will continue to denote them by (rather than ). This preserves the notation of the last three sections, and will also ease the transition to the family that will be applied to nontempered representations in Chapter 7. r We do not assume a priori that the elements in F satisfy Assumption 5.1.1. We instead x the positive integer N , and as in 5.1, assume inr ductively that all the local and global theorems hold for any P F with degpq N . We shall combine this inductive property with Assumption 5.4.1 to establish the original conditions in Assumption 5.1.1. r Lemma 5.4.2. Suppose that G P Eell pN q, and that belongs to F2 pGq. Then the conditions of Assumption 5.1.1 hold for the pair pG, q. Proof. The assertion of course includes the case that is simple. Expanding on the remark concerning (5.4.1)(a) above, we recall that r Fsim pGq was dened provisionally in (5.1.5) and (5.1.6) as a set of paramG r eters P Fsim pN q for which the corresponding stable linear form Sdisc, is nonzero. The problem is that for a given , we do not know yet that this r characterizes the datum G P Esim pN q uniquely. We do know from our rer marks in 3.4 that if G P Esim pN q is another such datum, and if G is distinct from the quadratic character G , then Sdisc, vanishes. Morer over, (5.4.1)(a) asserts that for any v P V , v lies in ell,v pN q, and that Gv is r r the unique datum in Ev pN q such that pGv q contains v . This condition will , unless G equals G , which in turn implies that fail if G is replaced by G v v p G. Since the properties G G and G G imply that G G, p p p G r the condition (5.4.1(a)) does indeed characterize the set Fsim pGq. r If is not simple, we recall that the condition P F2 pGq (or the more r general condition P FpGq) is dened inductively. In this case, our inr duction hypotheses imply that v belongs to pGv q, for any valuation v of F. Given and G, we write the set of valuations of F as a disjoint union V U Vun , where V is given by Assumption 5.4.1 and U is a nite set, and where pG, q is assumed to be unramied at every v in the remaining set Vun . We will apply the twisted endoscopic decomposition (3.3.14) of Corollary 3.3.2 to a function r r r r r r f fV fU fun , f P HpN q, that is compatible with this decomposition. In so doing, we will rst x the functions r r r r fV fv , fv P Hv pN q,
vPV

5.4. GENERIC PARAMETERS WITH LOCAL CONSTRAINTS

287

and r fun
vPVun

r fv ,

r r r and then allow fU P HU pN q to vary. We are writing Hv pN q here for the r rU pN q for the C-tensor product over v P U Hecke module on GpN, Fv q, and H r of Hv pN q. r We are assuming that belongs to F2 pN q. We shall use (5.4.1)(b) to r dene fV . This condition, together with the stronger condition (5.4.1)(a) in r r case lies in the subset Fsim pN q of F2 pN q, tells us that there is a valuation r q, for any datum G in Esim pN q r w P V such that w does not lie in pGw w G. We then choose a function f P H pN q such that r p p rw over w with Gw w f Gw of f r r r in SpG q r fw,N pw q is nonzero, but so that the image fw w w w . The existence of f follows easily from Proposition r vanishes for any such Gw w 2.1.1, and the trace Paley-Wiener theorems [CD] and [DM] at w. At the r other places v P V , we choose fv subject only to the requirement that r r fv,N pv q not vanish. The function fV then has the property that its value r r fV,N pV q fv,N pv q
v

is nonzero, but its transfer r fV

vPV

rG fv v

r p p vanishes for any G P Esim pN q with G G. r The other function to be xed is fun . We take it simply to be any decomposable function r r fun fv
vPVun

` r r in the unramied Hecke algebra H N, Kun pN q such that the value r r fun,N pun q fv,N pv q,
vPVun

r is nonzero. Of course, Kun pN q stands here for the product over v P Vun of r the standard maximal compact subgroups Kv pN q of GLpN, Fv q. We have thus to apply the identity N r r p rpN, G q Sdisc, pf q (5.4.2) Idisc, pf q
r G PEell pN q

to the variable function (5.4.3) r r r r f fV fU fun , r r fU P HU pN q.

288

5. A STUDY OF THE CRITICAL CASES

We have not asked that the given datum G be simple. However, if r G P Eell pN q is a datum that is not simple, and distinct from G, we can say that G Sdisc, Sdisc, 0. This follows from the application of our induction hypothesis to the comr posite factors of G , since as an element in 2 pGq, cannot also belong to q. Suppose next that G P E r rsim pN q, but is again distinct from G. If pG p G, we obtain p G r r r r p p S pf q S pf f f q 0,
disc, disc, V U un

r p p by our choice of fV . If G G, we have G G ,

r given that lies in pGq. This implies that Sdisc, 0, as we agreed in G therefore all vanish, and we 3.4. The summands in (5.4.2) with G obtain r r p (5.4.4) I N pf q rpN, Gq S G pf G q. disc, disc,

We shall compare the last formula with the spectral expansion of N r r Idisc, pf q. This expansion is the analogue of (4.1.1) for GpN q, described rst at the end of 3.3. It is a sum of terms parametrized by Levi subgroups M 0 GLpN1 q GLpNr q of GLpN q. From the classication of automorphic representations of GLpN q (Theorem 1.3.3), we know that there is at most M (taken up to conjugacy) for which the corresponding term is nonzero. It is then a consequence of the denitions (or if one prefers, the twisted global intertwining relation discussed at the end of 4.2) that this term is a nonzero multiple of the linear form r r r r fN pq fV,N pV q fU,N pU q fun,N pun q r cpV , Vun q fU,N pU q, where cpV , Vun q is a nonzero constant. Combined with (5.4.4), this tells r us that fU,N pU q is a nonzero multiple of the linear form r rG G rG pG pG Sdisc, pf G q Sdisc, pfV fU fun q, rG r and hence depends only on fU . Therefore, fU,N pU q is the pullback of a r U q. linear form on SpG r We are now free to take fV to be a variable function. It follows from the conditions (5.4.1)(a) and (5.4.1)(b), and the discussion of the specic r r function fV above, that the corresponding linear form fV,N pV q is also the r pullback of a linear form on SpGV q. Enlarging the nite set U , if necessary, r is the characteristic function of Kun pN q. Having r we can assume that fun

5.4. GENERIC PARAMETERS WITH LOCAL CONSTRAINTS

289

r r treated the function fU , we conclude that fN pq is indeed the pullback of a r linear form on SpGV q. In other words, we can write r r r r (5.4.5) fN pq f G pq, f P HpN q, r where the right hand side represents the restriction to the image of HpN q r of some linear form on SpGq. r If G is simple, the linear form on SpGq is uniquely determined by (5.4.5), since Corollary 2.1.2 tells us that the twisted transfer mapping is surjective. r In particular, it can be identied with a unique stable linear form on HpGq. If G GS GO is composite, and S O is the corresponding r decomposition of , we dene the linear form on SpGq as the tensor product r r of the linear forms on SpGS q and SpGO q given by S and O . We then have to check that this denition is compatible with (5.4.5). In other words, we need to show that r r r r (5.4.6) f G pq f G pS O q, f P HpN q, where the left hand side is dened by (5.4.5), and the right hand side is dened by the linear forms attached to S and O . To establish (5.4.6), we return briey to the discussion surrounding the N r r function (5.4.3) above. For this particular function f , we know that Idisc, pf q is on the one hand equal to r p rpN, Gq S G pf G q,
disc,

r and on the other, is a constant multiple of the value at f of the left hand side of (5.4.5). This constant is easy to compute, as an elementary special case of the general remarks in 4.7 for example. The reader can check that its product with rpN, Gq1 equals the product of mpq with |S|1 . In other words, the stable multiplicity formula r r p S G pf G q mpq |S |1 f G pq
disc,

r holds for the given function f . We are using the fact that is generic here, so there is no -factor in the computation, and no sign G pq. Suppose again that G GS GO and S O . Our general induction hypothesis then tells us that p p p S G pf G q S S pf S q S 0 pf O q
disc, disc,S

mpS q |SS |

disc,O 1 S

f pS q mpO q |SO |1 f O pO q,

p for any decomposable function f G f S f O in SpGq. We thus have two r pG formulas for Sdisc, . Replacing f G by the image f G of the function (5.4.3), and noting that mpq |S |1 mpS q |SS |1 mpO q |SO |1 , r we see that (5.4.6) holds if f is the function (5.4.3). Now the local analogue r is replaced by an arbitrary function fV P HpGV q (this r r of (5.4.6) holds if f

290

5. A STUDY OF THE CRITICAL CASES

is Theorem 2.2.1(a) for the archimedean valuations v P V , which is part of r what we are taking for granted from [Me] and [S7]), or by the unit fun in the G p q and f G p q r r Hecke algebra (by denition). Since the components fV V un un r and fun in (5.4.3), the local analogue of r are nonzero for the functions fV r r r (5.4.6) holds also if f is replaced by an arbitrary function fU P HpGU q. With r this fact in hand, we replace the xed component fV in (5.4.3) by a general r V q. The required identity (5.4.6) then follows for a general function in HpG r r r r decomposable function f P HpN q, since it holds for its local factors fV , fU r . This completes the last step in the proof of the lemma. and fun Remark. The assertion of the lemma holds more generally if is any r element in FpGq. For as we remarked after stating Assumption 5.1.1, the r assertion for parameters in the complement of F2 pGq follows from the corresponding assertion for a Levi subgroup of GLpN q. r We shall now establish some global identities for FpN q. Keep in mind r q is composed of generic global parameters with rather serious local that FpN constraints. For this reason, the identities we obtain will not be general enough to have much interest in their own right. Their role is intended rather to be local. They will drive the local classication we are going to establish in the next chapter. We will return to global questions later in Chapter 8, where having armed ourselves with the required local results, we will be able to establish the global theorems in general. We shall prove four lemmas, which represent partial resolutions of the four main lemmas from 5.2 and 5.3, and apply to parameters in the family r FpN q. We continue to denote these generic parameters by rather than , in order to match the notation from the two earlier sections. Lemma 5.4.3. Suppose that pG, q, r r2 G P Esim pN q, P Fell pGq, r is as in Lemma 5.2.1, but with F being our family of generic parameters that satisfy Assumption 5.4.1. Then r R pf q 0 0 S pf q, f P HpG q,
disc, disc,

for every r f P HpGq.

r P Esim pN q, while the expression (5.2.8) vanishes for any

r Proof. Given the pair pG, q, we choose a place v P V pFq that satises (5.4.1)(c). We then consider the expression (5.2.8) in Lemma 5.2.1, for a decomposable function ` r r f fv f v , fv P HpGv q, f v P H GpAv q , r in HpGq, and a corresponding decomposition v v

5.4. GENERIC PARAMETERS WITH LOCAL CONSTRAINTS

291

of . The main terms in (5.2.8) become

v fG p, xq fv,G pv , xv q fG p v , xv q

and
1 1 fG p, xq fv,G pv , xv q pf v q1 p v , xv q, G

by (4.2.4) and (4.2.5). Since is generic, we also have reductions s 1 and pxq 1. The expression (5.2.8) therefore equals ` 1 v pf v q1 p v , xv q fv,G pv , xv q fG p v , xv q fv,G pv , xv q . (5.4.7) c G
xPS,ell

We will derive the required properties from Corollary 3.5.3. This entails identifying (5.4.7) with the supplementary expression (3.5.12) of the corollary. We have therefore to expand the linear forms in fv from (5.4.7) in terms of the basis T pGv q described at the beginning of the proof of Proposition 3.5.1. The element xv in (5.4.7) is the image of x in Sv . We shall write rv pxq for its image in the local R-group Rv . We shall also identify v with a xed r local preimage in the set 2 pMv , v q, where Mv is a xed Levi subgroup over Gv such that this set is not empty. (We are thus following the local notation (2.3.4) rather than letting Mv denote the localization of the global Levi subgroup M attached to , a group that could properly contain Mv .) The linear form fv,G pv , xv q in (5.4.7) then equals ` ` x r r (5.4.8) xrv , v y tr RPv rv pxq, v , v IPv pv , fv q ,
r v Pv pMv q

in the notation (2.4.5). The other term in (5.4.7) that depends on fv is the linear form 1 fv,G pv , xv q. We are not assuming the local intertwining relation of Theorem 2.4.1, which asserts that it equals fv,G pv , xv q. However, our assumption that v is archimedean and generic, and hence that Theorem 2.2.1 holds, leads to a weaker relation between the two linear forms. For if the assertion (b) of this theorem is combined with the theory of the R-groups Rpv q, atr tached to representations v P v pMv q and reviewed in 3.5, one obtains an endoscopic characterization of the irreducible constituents of the induced representation IPv pv q. They correspond to characters on the group Sv pGq whose restriction to the subgroup Sv pMv q equals the character attached to 1 v . It is then not hard to see that fv,G pv , xv q has an expansion ` 1 ` r (5.4.9) xrv , v y tr RPv rv pxq, v , v IPv pv , fv q , x r
r v Pv pMv q

where
1 r r RPv prv , v , v q 1 v prv q RPv prv , v , v q,

v P v pMv q, rv P Rv ,

292

5. A STUDY OF THE CRITICAL CASES

for a (sign) character rv v prv q v prv q1 on the 2-group Rv . This is included in Shelstads endoscopic classication of representations for real groups, as we will recall (6.1.5) in the next section. The local intertwining relation would tell us that the character v is trivial for each v . As it applies to the parameter v , Theorem 2.2.1(b) can thus be summarized as the existence of a natural isomorphism from the endoscopic R-group Rv onto any of the associated representation theoretic R-groups Rpv q. A triplet ` v Mv , v , rv pxq , v P v pMv q,
G or rather its orbit under the local Weyl group W0 v , can therefore be identied with an element in the basis T pGv q. It follows from the denition (3.5.3) that we can write (5.4.8) and (5.4.9) respectively as ` fv,G pv , xv q apv , xv q fv,G Mv , v , rv pxq r v Pv pMv q

and
1 fv,G pv , xv q

r v Pv pMv q

` a1 pv , xv q fv,G Mv , v , rv pxq ,

` for coecients apv , xv q and a1 pv , xv q v rv pxq apv , xv q. The expression (5.4.7) can therefore be written as (5.4.10) dpv , f v q fv,G pv q,
v PT pGv q

where dpv , f v q equals the sum ` v c a1 pv , xv q pf v q1 p v , xv q apv , xv q fG p v , xv q G

x v

r over elements x P S,ell and v P v pMv q such that the triplet ` Gv Mv , v , rv pxq belongs to the W0 -orbit represented by v . Our expression (5.4.10) for (5.2.8) matches the supplementary expression (3.5.12) of Corollary 3.5.3 (with the group G1 in (3.5.12) being the group G here). Observe also that any x P S,ell maps to a regular element in the global R-group R . It follows from the condition (5.4.1(c)) on v that the image rv pxq of x in the local R-group Rv in nontrivial. In other words, the coecient dpv , f v q vanishes if v is represented by a triplet of the form pMv , v , 1q. We can therefore apply Corollary 3.5.3 to the formula given by Lemma 5.2.1. We nd that the coecents dpv , f v q all vanish, as do the multiplicities of any irreducible constituents of representations Rdisc, in (5.2.7). It follows that the expression (5.2.8) vanishes , as required, and

5.4. GENERIC PARAMETERS WITH LOCAL CONSTRAINTS

293

that the representations Rdisc, are all zero. Finally, as we observed in the proof of Lemma 5.2.1, the dierence r r R pf q 0 Sdisc, pf q, G P Esim pN q, f P HpG q, disc,

vanishes unless G G , in which case it equals (5.2.8). Having just shown that (5.2.8) itself vanishes, we can say that the dierence vanishes also in the case G G. We conclude that r r R pf q 0 0 S pf q, G P Esim pN q, f P HpG q,
disc, disc,

as required. This completes the proof of the lemma. Lemma 5.4.4. Suppose that pG, q, r r2 G P Esim pN q, P Fell pGq, r is as in Lemma 5.2.2, with F again being our family of generic parameters that satisfy Assumption 5.4.1. Then r R pf q 0 0 S pf q, f P HpG q,
disc, disc

r for every P Esim pN q, while the expressions (5.2.13) and (5.2.12) vanish r r for any functions f P HpGq and f _ P HpG_ q respectively. Proof. This is the case of a parameter (5.2.4) with r 1. Apart from the analysis of the new term (5.2.12), which is attached to G_ and M , the proof is similar to that of the last lemma. For the given pair pG, q, r we choose v P V pFq so that the condition (5.4.1)(c) holds. We then write the expression (5.2.13) in a form (5.4.10) that matches the supplementary expression (3.5.12) of Corollary 3.5.3. To deal with (5.2.12), we can argue as in the treatment of the parameter (4.5.11) from the proof of Proposition 4.5.1. In the simple case here, the r restricted parameter M P 2 pM, q for the Levi subgroup M GLpN q is just the simple component 1 of . We can therefore identify M with the automorphic representation 1 of GLpN q. It follows from the denitions of its two terms that (5.2.12) equals ` _ 1 _ _ r (5.4.11) 8 tr p1 RP pw1 , 1 , 1 qq IP p1 , f q . This is essentially the analogue for G_ of the general expression (4.5.14) obtained in the proof of Proposition 4.5.1. As in the earlier discussion of (4.5.14), we observe that (5.4.11) is a nonnegative integral combination of irreducible characters. Its sum with the rst term (5.2.11) in the formula given by Lemma 5.2.2 is therefore of the general form (3.5.1). Since the other expression (5.2.13) is of the form (3.5.12), we can apply Corollary 3.5.3 to the formula. We conclude that the expressions (5.2.13) and (5.2.12) both vanish, as required, and that the representations Rdisc, are all zero. We then recall that as in the proof of Lemma 5.2.2, the dierence r r R pf q 0 Sdisc, pf q, G P Esim pN q, f P HpG q,
disc,

294

5. A STUDY OF THE CRITICAL CASES

equals a positive multiple of either (5.2.13) or (5.2.12) if G is equal to the associated group G or G_ , and vanishes otherwise. The dierence therefore vanishes in all cases, and
Rdisc, pf q 0 0 Sdisc, pf q,

r r G P Esim pN q, f P HpGq,

as required. Lemma 5.4.5. Suppose that pG, q, r r G P Esim pN q, P F2 pGq,

r is as in Lemma 5.3.1, for our family F of generic parameters with local constraints. Then the linear form in (5.3.4) vanishes. Proof. Recall that was dened in Lemma 5.1.4 as a linear form on r S 0 pLq. It represents the defect in the stable multiplicity formula for each of the pairs pG, q and pG_ , q. The lemma thus asserts that the stable multiplicity formula is valid for either of these pairs. We observe also that the premise of Lemma 5.3.1 holds here. It amounts to the rst vanishing assertion of Lemma 5.4.3 for parameters in the comr r plement of Fell pN q in FpN q. (It was this global theorem that was required in the proof of Lemma 5.3.1.) Therefore Lemma 5.3.1 is valid in the case at hand. The supplementary term (5.4.12) b` f L` p1 q, r L` G1 L, f P HpG` q,

in (5.3.4) is clearly the main point. If this linear form were a nonnegative linear combination of characters in f , the entire expression (5.3.4) would be of the general form (3.5.1), and we could argue as in the proof of Lemma 5.4.3. According to Corollary 5.1.5, the pullback of from LpAq to G_ pAq is a unitary character. However, it is not clear that the pullback of to GpAq has the same property. To sidestep this question, we introduce the composite r endoscopic datum G_ G1 G_ in Eell pN` q. The group L` , regarded itself 1 r as an endoscopic datum in EpN` q, has a global Levi embedding into both _ . If f P HpG_ q is the associated function in the compatible r G` and G1 1 1 L family (5.3.3), it follows from the denitions that f1 ` equals f L` . The supplementary term in (5.3.4) therefore equals (5.4.13) b` f1 ` p1 q,
L

r f1 P HpG_ q, G_ G1 G_ . 1 1

Since 1 is assumed to be a proper subparameter of , its degree N1 is less than N . It follows from our induction hypothesis that the stable linear form r on HpG1 q attached to 1 is a nonzero unitary character. Corollary 5.1.5 L then tells us that f1 ` p1 q represents a positive multiple of a unitary character (possibly 0) in f1 . It follows that if we replace the supplementary term (5.4.12) by (5.4.13), the expression (5.3.4) will indeed be of the general form (3.5.1).

5.4. GENERIC PARAMETERS WITH LOCAL CONSTRAINTS

295

r With this modication, we choose v P V pFq so that the condition (5.4.1)(c) holds for the pair pG` , ` q. Following the proof of Lemma 5.4.3, we then write the expression (5.3.5) in Lemma 5.3.1 as a sum (3.5.12) for which the required condition of Corollary 3.5.3 holds. We can therefore apply Corollary 3.5.3 to the formula provided by Lemma 5.3.1. It then follows that the coecients in the irreducible decomposition of the modied form of (5.3.4) vanish, as does the entire expression (5.3.5). In particular, the sumr mand (5.4.13) vanishes. Since the stable linear form on HpG1 q attached to 1 is nonzero, this can only happen if vanishes, as claimed. Lemma 5.4.6. Suppose that pG, q, r r G P Esim pN q, P Fsim pGq,

r is as in Lemma 5.3.2, for our family F of generic parameters with local constraints. Then the linear form in (5.3.21) vanishes, while the sign in (5.3.22) and (5.3.23) equals 1. Proof. We will have to draw on the techniques of all of the last three lemmas, since all of the corresponding unknown quantities occur here. We will also have to include something more to take care of the sign . However, the basic idea is the same. We shall apply Corollary 3.5.3 to the formula provided by Lemma 5.3.2. r We choose v P V pFq so that (5.4.1)(c) holds for the pair pG` , ` q. As in the proof of Lemma 5.4.3, we then write the expression (5.3.23) on the right hand side of the formula of Lemma 5.3.2 in the general form (3.5.12). The fact that there is now an unknown sign in the expression is of no consequence. For the condition on the coecients in (3.5.12) required for Corollary 3.5.3 remains a consequence of our given condition (4.5.1)(c). The left hand side of the formula is the sum of two expressions (5.3.21) and (5.3.22). The rst of these contains the supplementary summand (5.4.14)
1 2

f L` p q,

r L` G L, f P HpG` q.

Following the proof of the last lemma, we introduce the composite endor scopic datum G_ G G_ in Eell pN` q, with a corresponding function 1 r _ q in the compatible family (5.3.3). Since L` again has a global f1 P HpG1 Levi embedding into both G` and G_ , the supplementary summand (5.4.14) 1 equals (5.4.15)
1 2

f1 ` p q,

r f1 P HpG_ q, G_ G G_ . 1 1
L

It then follows from Corollary 5.1.5 and the denition of that f1 ` p q represents a positive multiple of a unitary character (possibly 0) in f1 . The expression (5.3.21), written with (5.4.15) in place of (5.4.14), is consequently of the general form (3.5.1). As for the second expression (5.3.22), we note that the parameter `,M` for M` GLpN q can be identied with the

296

5. A STUDY OF THE CRITICAL CASES

automorphic representation of GLpN q. Appealing to the denitions, as in the proof of Lemma 5.4.4, we can write the expression in the form ` _ _ 1 r _ _ r (5.4.16) f _ P HpG_ q. ` 8 tr p1 RP` pw , , qq IP` p , f q , Since equals either p`1q or p1q, (5.4.14) is a nonnegative linear combination of characters on the group G_ pAq. It follows that the sum of (5.3.21) ` and (5.3.22) can be written as an expression of the general form (3.5.1). We can thus apply Corollary 3.5.3 to the formula given by Lemma 5.3.2. We then see that, among other things, the expressions (5.4.14) and (5.4.16) each vanish. From the vanishing of (5.4.14), we deduce that either or equals zero. Suppose that is the factor that vanishes. The linear form r f G pq f G pq ` f L pq f L pq, f P HpGq, r on HpGq is then induced from the Levi subgroup LpAq. This contradicts the local condition (5.4.1(a)) for the simple parameter . It is therefore that vanishes, as claimed. The operator _ RP` pw_ , , q
_ in (5.4.16) is unitary, and commutes with IP` p , f _ q. The vanishing of (5.4.16) then implies that this operator equals 1. In particular, our normalized intertwining operator here is a scalar, as we expect from the fact _ that the centralizer group S` is isomorphic to SLp2, Cq. To show that it actually equals 1, we appeal to the relation between normalized intertwining operators and Whittaker models. Since the parameter is generic, case (iii) of Corollary 2.5.2 tells us that _ RP` pw_ , ,v , v q 1

for any v. Therefore

_ RP` pw_ , , q

_ RP` pw_ , ,v , v q 1.

We conclude that the sign equals 1, and therefore that the second assertion of the lemma also holds. In the last lemma, N is even and equals 1. If one of these conditions is not satised, the formula of Lemma 5.3.2 becomes simpler, as we noted at the end of the last section. In particular, is not dened, and does not contribute a summand to (5.3.21). If N is odd, the summands in (5.3.23) are also not dened, and the term (5.3.23) does not exist. However, the formula of Lemma 5.3.2 is otherwise valid, and the sign is invariably present in the term (5.3.22). The simplied analogue Lemma 5.4.6 asserts that 1. It is proved as above. r We can now clarify the temporary denition of the sets Fsim pGq adopted r in 5.1. We shall show that the three possible ways to characterize Fsim pGq are all equivalent.

5.4. GENERIC PARAMETERS WITH LOCAL CONSTRAINTS

297

r Corollary 5.4.7. Suppose that belongs to the subset Fsim pN q of our r family F of generic parameters with local constraints, and that G belongs to r Esim pN q. Then the following conditions on pG, q are equivalent. G r (i) The linear form Sdisc, on HpGq does not vanish. (ii) Theorem 1.4.1 holds for , with G G. (iii) The global quadratic characters G and are equal, and the global L-function condition 1 holds. Proof. The rst assertion 0 of Lemma 5.4.6 tells us that condition r (i) characterizes G uniquely in terms of . If G is any element in Esim pN q, we know that ` r Sdisc, pf q tr Rdisc, pf q , f P HpG q, since the proper terms in the expansions (4.1.1) and (4.1.2) all vanish by our induction hypotheses. It follows that the condition of Theorem 1.4.1 characterizes the postulated datum G uniquely, and that G G. The conditions (i) and (ii) are therefore equivalent. The third assertion 1 of Corollary 5.4.7 (together with (3.4.7), and the remark above in case N is odd or 1) tells us that (i) implies (iii). But we know that for the given , there is at most one G that satises (iii). The conditions (i) and (iii) are therefore also equivalent. The condition (i) was the basis for our temporary denition in 5.1. It has been the pivotal condition for our arguments in this chapter. The condition (ii) was used for the original denition from Chapter 1. Knowing r now that the three conditions are equivalent for the family F of this section, rsim pGq to be the set of P Fsim pN q such that any one of the r we can dene F r conditions holds for pG, q. It is clear from the discussion that Fsim pN q is rsim pN q of the sets Fsim pGq. r a disjoint union over G P E We shall add a comment to what we have established in this section. The preceding Lemmas 5.4.3, 5.4.4, 5.4.5 and 5.4.6 apply to the respective formulas of Lemmas 5.2.1, 5.2.2, 5.3.1 and 5.3.2. We recall that the four formulas became increasingly complex as each one took on another unknown quantity. The four unknowns were the global intertwining relation for G (or G` ), the global intertwining relation for G_ (or G_ ), the linear form , ` and the sign . By the time we reached the last Lemma 5.3.2, all four of these unknowns were present as terms in the associated formula. Their resolution, under the local constraints of this section, rests exclusively on Corollary 3.5.3 of Proposition 3.5.1. The basic principle behind Proposition 3.5.1 can be described as follows: if a linear combination of reducible characters vanishes, and if the coecients are all positive, then the characters are themselves equal to zero. This is an obvious simplication, but it serves to illustrate the main idea for the proofs of the last four lemmas. The critical point is that the unknown quantities, or rather the ones that do not go into the supplementary expression (3.5.12)

298

5. A STUDY OF THE CRITICAL CASES

of Corollary 3.5.3, must all occur with the appropriate signs. They do most fortunately for us! This phenomenon seems quite remarkable. We have seen it before, in the treatment of the parameters (4.5.11) and (4.5.12) of Proposition 4.5.1 and in Lemma 5.1.6, and we will see it again, most vividly perhaps in our completion of the global classication in 8.2. I do not have an explanation for it.

CHAPTER 6

The Local Classication

6.1. Local parameters The next two chapters will be devoted to the proof of the local theorems. Our methods will be global. Given the appropriate local data, we shall construct a family of global parameters to which we can apply the results of Chapter 5. Since we are aiming for local results, we take F to be a local eld in 9 Chapters 6 and 7. We will then use the symbol F to denote the auxiliary global eld over which the global arguments take place. This follows a general convention of representing global objects by placing a dot over symbols that stand for the given local objects. To obtain a general idea of the strategy, consider a local parameter
1 1

r r ,

r r i P sim pGi q, Gi P Esim pNi q,

r r where the sets sim pGi q and Esim pNi q are understood to be over the local eld 9 r F . We are proposing to construct suitable endoscopic data Gi P Esim pNi q 9 r 9 9 and parameters i P sim pGi q over a global eld F such that 9 9 9 pF, Gi , i q pFu , Gi,u , i,u q, 9 at some place u of F . This will lead to a family ( 9 9 9 9 r r 9 F Fpi , . . . , r q 91 1 9r r : 9i 0 of global parameters to which we can try to apply the methods of Chapter 5. In this chapter we will treat the case that is generic. We will 9 r show that the resulting global family F has local constraints (5.4.1), where 9 r 9 V V pFq is a set of archimedean places of F . In the next chapter, we will deal with general parameters. There we will establish local constraints that 9 r are roughly parallel to (5.4.1), but with V V pFq being a set of p-adic places. We shall prepare for this analysis with some remarks on local Langlands 9 parameters, which we will apply presently to the localizations v of global 9 9 r parameters P F. These fall into three categories. We will have separate 9 9 observations for the completions Fv at F at v u, at nonarchimedean places v u, or at archimedean places v u.
299

1 i r,

300

6. THE LOCAL CLASSIFICATION

The rst remarks are aimed at general Langlands parameters over 9 F Fu . They pertain thus to the objects for which we are trying to establish the local theorems. Since the main questions are nonarchimedean, we assume for the time being that F is p-adic. Our purpose is to introduce a temporary modication of one of our denitions. It will be a local p-adic analogue of the temporary global denition we adopted after Assumption 5.1.1, and later resolved in terms of the original denition in Corollary 5.4.7. At issue is the denition of the local r parameter sets pGq. We begin by stating one more theorem, which can be regarded as a local analogue of Theorem 1.4.1. This last of our collection of stated theorems amounts to a special case of Theorem 2.2.1, just as the initial Theorem 1.4.1 was a special case of Theorem 1.5.2. We will resolve it, along with the generic case of all the other local theorems, in 6.8. r Theorem 6.1.1. Suppose that P sim pN q is a simple generic local rell pN q such that parameter. Then there is a unique G P E r r r r fN pq f G pq, f P HpN q, r r for a linear form f G pq on SpGq. Moreover, G is simple. r We consider now the denition of the set pGq attached to a given r simple datum G P Esim pN q. As in the global case, there will be three equivr alent ways to characterize the subset sim pGq of simple local parameters r r in sim pN q. Our original denition gives sim pGq as the subset of irreducible N -dimensional representations of LF that factor through the image of L G in GLpN, Cq. The second characterization is in terms of poles of local L-functions, and will be postponed until 6.8. The third is provided by r Theorem 6.1.1. It characterizes sim pGq as the subset of local parameters r P sim pN q for which the datum G of the theorem equals G. The rst characterization is the simplest. However, the local Langlands parameters for G are quite removed from the local harmonic analysis that governs the classication of Theorem 2.2.1. The third is more technical, but is formulated purely in terms of local harmonic analysis. For this reason, we shall temporarily replace the original denition by one that is based on Theorem 6.1.1. We shall make it parallel to the temporary global denition from the beginning of 5.1. r r r Let Hcusp pN q be the subspace of functions f P HpN q such that for each r rell pN q, f G lies in the subspace Scusp pGq of SpGq. It follows from r r G P E Proposition 2.1.1 that the mapping r r r r f f G, f P Hcusp pN q,
r GPEell pN q

r r r descends to an isomorphism from the image Icusp pN q of Hcusp pN q in IpN q rell pN q of the spaces Scusp pGq. If is any r onto the direct sum over G P E

6.1. LOCAL PARAMETERS

301

r element in pN q, we can then write r r (6.1.1) fN pq f G pq,

r GPEell pN q

r r f P Hcusp pN q,

r r for uniquely determined linear forms f G pq on each of the spaces Scusp pGq. The formula (6.1.1) is a rough local analogue of the global decomposition (3.3.14). It tells us that we can attach local subsets (6.1.2) r r r sim pGq t P Fsim pN q : f G pq 0 for some f P Hcusp pGqu

r r of sim pN q to the simple endoscopic data G P Esim pN q so that r r (6.1.3) sim pN q sim pGq.
r GPEsim pN q

To be denite, and in minor contrast to the global convention of 5.1, we r dene sim pGq by taking the inclusion in (6.1.2) to be equality. That is, we r r dene sim pGq to be the set of all parameters P sim pN q such that the G pq on H rcusp pGq does not vanish, as on the right hand side linear form f of (6.1.2). This temporary denition will be in force throughout Chapter 6. During this period, the local theorems will be interpreted accordingly. For example, Theorem 6.1.1 will now include the assertion that the union (6.1.3) is actually disjoint. If we assume inductively that the local theorems r (including Theorem 6.1.1) are all valid for parameters in pN1 q with N1 r r N , we can dene the full set pGq in terms of the fundamental sets sim pG1 q rsim pN1 q, exactly as we did for the global case in 1.4. This in with G1 P E turn allows us to describe the centralizers S by the global prescription of 1.4 and the convention (5.1.7), and then to introduce the subsets r r r r sim pGq 2 pGq ell pGq pGq, according to the later global denitions of 4.1. We will thus temporarily forget about general Langlands parameters for r the p-adic group G (though we will still be able to treat elements in pN q as representations of LF ). We will come back to them in 6.8, after proving the local theorems in their interpretation above. We will then be able to r resolve the temporary denition of pGq in terms of the original one. This will yield the general local theorems for generic in their original form. The next remarks are very elementary. They concern special elements r in pGq, which will still be identied with local Langlands parameters, and 9 9 which will apply to localizations v at p-adic places v of F distinct from u. The parameters we have in mind are spherical, in the sense that they are attached to the p-adic spherical functions of [Ma]. In particular, they are r unramied if the given group G P Esim pN q over F is unramied. The group G is quasisplit over F . It is unramied, by denition, if it splits over an unramied extension of F . This is the case unless N is even, p G SOpN, Cq, and the quadratic character G is ramied. The group

302

6. THE LOCAL CLASSIFICATION

p itself is split over F unless N is even, G SOpN, Cq and G 1. Let K be our xed maximal compact subgroup of GpF q. Then K is special in general, and hyperspecial if G is unramied. An irreducible representation of GpF q is said to be K-spherical (or simply spherical) if its restriction to K contains the trivial representation. A local Langlands parameter will r be called spherical if its L-packet contains a K-spherical representation. We recall the description of these familiar objects. Let M0 be the standard minimal Levi subgroup of G. Then M0 is a x maximally split, maximal torus over F , whose dual M0 is the group of p Spherical parameters factor through the image of diagonal matrices in G. L M in L G, and in particular descend to the quotient W of L . Let 0 F F L x0 WF 0 : WF M0 M be the canonical splitting of L M0 , which is to say the identity mapping of WF to the second factor of L M0 . If is a point in the complex vector space a 0 ,C XpM qF b C, M we write |w| for the point in the complex torus x AM0 pM0 q0 XpM0 qF b C x attached to and the absolute value |w| of an element w P WF . The mapping pwq 0 pwq |w| , w P WF , is then a Langlands parameter for M0 , whose image in pGq we continue to denote by . The elements in the family t : P a 0 ,C u M are the spherical parameters for G. Under the Langlands correspondence for the torus M0 , the spherical parameter P pM0 q maps to the spherical character ptq epHM0 ptqq , t P M0 pF q, on M0 pF q. If is purely imaginary, is unitary and has bounded image x in M0 . Treating as an element in pGq, we write in this case for the set of irreducible constituents of the induced representation IP0 p q, where P0 P PpM0 q is the standard Borel subgroup of G. This is the L-packet of . For any , it contains a unique K-spherical representation. Our nal remarks concern archimedean parameters. These will be ap9 9 plied to localizations v at the archimedean places v of F distinct from u. The complex case presents little diculty, and we will in any case be choos9 ing the global eld F in 6.2 to be totally real. We therefore assume for the r rest of this section that F R, with G P Esim pN q continuing to be a simple endoscopic datum over F . We shall review a few points in Shelstads general classication of the representations of GpF q in terms of local Langlands r parameters P pGq.

6.1. LOCAL PARAMETERS

303

9 The archimedean valuations v u of F will in fact play a central role in the global arguments of this chapter. To prepare ourselves, it will be helpful r to review the archimedean parameters P 2 pGq attached to discrete series representations of GpF q. This amounts to an explicit specialization of some of the observations in 1.2. The local Langlands group for F is just the real Weil group WF WR . 2 It is generated by the group C and an element F , with relations F 1 and 1 z P C . F z F z, A parameter : WF L G, when composed with our mapping of L G into GLpN, Cq, becomes a selfdual, N -dimensional representation of WF . We are interested in the case r that lies in 2 pGq, which is to say that the group ` p S S pGq Cent pWF q, G is nite. As we noted in the general discussion of 1.2, this is equivalent to the condition that the irreducible constituents of the representation of WF attached to all be distinct and self-dual. Recall [T2] that the irreducible representations of WF are all of one or two dimensions. There are two self-dual one-dimensional representations, the trivial representation and the sign character of R pWF qab . The irreducible self-dual representations of WF of dimension two are parametrized 1 by positive half integers P 2 N, and are given by 1 0 pzz q , z P C , z 0 pzz 1 q 0 1 F . p1q2 0 The image of any such representation lies in Op2, Cq or Spp2, Cq, according to whether is an integer or not. To construct all the parameters in 2 pGq, we simply select families of distinct irreducible representations of WF that p are of the same type (either symplectic or orthogonal) as G. To be more precise, let p rN {2s be the greatest integer in N {2. If p SppN, Cq, any P 2 pGq is a direct sum of p distinct two-dimensional r G representations of symplectic type. It can therefore be identied with a set p1 , . . . , p q, i P 1 N N, 2 and

p of p distinct, positive, proper half integers. If G SOpN, Cq, with N odd, r 2 pGq is a direct sum of p distinct two-dimensional repa parameter P resentations of orthogonal type and a uniquely determined one-dimensional representation. It can therefore be identied with a set p1 , . . . , p q, i P N,

304

6. THE LOCAL CLASSIFICATION

of p distinct positive integers. The one-dimensional representation is trivial p if p is even, and is the sign character if p is odd. If G SOpN, Cq, with r 2 pGq to be nonempty. We therefore N even, we of course want the set assume that G is split if p is even, and nonsplit if p is odd. There remain two possibilities in this case, which we characterize by setting p1 equal either r to p or p 1. A parameter P 2 pGq can then be identied with a set p1 , . . . , p1 q, i P N,

of p1 distinct positive integers. It corresponds to a direct sum of p1 distinct two-dimensional representations of orthogonal type, augmented by the sum of the trivial and sign characters in case p1 p 1. The notation was used before. In the proof of Lemma 2.2.2, it was ` N that represented the innitesimal character of used for a vector in 1 Z 2 ` 1 r ell pN q. In this section, is a vector in 1 N p (where as an element in 2 r p1 p in the rst two cases above) attached to P 2 pGq. Having long ago r r agreed to identify 2 pGq with a subset of ell pN q, we must now be prepared to identify with the vector ` 2p1 ` 1 N p q P 1 Z 2Z 2 ` N in 1 Z . In any case, in this section represents the innitesimal char2 r acter of as an element in 2 pGq. We shall say that is in general position if its innitesimal character is highly regular, in the sense that the positive half-integers ( (6.1.4) i , |i j | : 1 i j p1 are all large. p We should also note that if G SOpN, Cq, for N even or odd, GpF q has a central subgroup of order 2. In this case, we can speak of the central character of , with the understanding that it refers to the representations r in the corresponding L-packet . One shows that it is trivial if and only if the integer c i
i

is even. It would not have been dicult to describe more general Langlands parameters in explicit terms. (See [L11, 3] for the case of a general group over R.) We have treated the square integrable (archimedean) parameters r P 2 pGq because they are the objects we will use in the next section to construct automorphic representations. We recall in any case that Theorem r 2.2.1 is valid for general archimedean parameters P pGq (assuming the r r two special cases for the twisted groups GpN q and G of the work in progress by Mezo and Shelstad, which would also give Theorem 2.2.4).

6.1. LOCAL PARAMETERS

305

The results for generic archimedean parameters (including their pror r jected analogues for GpN q and G) do not include the local intertwining relation of Theorem 2.4.1 (or of Theorem 2.4.4). They do, however, implicitly contain a weaker form of the identity. It is a consequence of Shelstads proof of the generic case r f 1 p, xq f 1 p1 q xx, y fG pq, P bdd pGq, x P S ,
G r P

of the formula (2.2.6) of Theorem 2.2.1(b), specically the fact that the pairing xx, y in this formula is compatible with the short exact sequence (2.4.9). In particular, if M P 2 pM, q is a square integrable parameter that maps to , is a disjoint union over M P M of sets M pGq on which the p dual group R of R acts simply transitively. As the irreducible constituents of induced representations IP pM q, these sets are in turn compatible with the self-intertwining operators RP pw, M , M q, in the sense that follows from r the work of Harish-Chandra [Ha4, Theorem 38.1] and Knapp and Stein [KnS, Theorem 13.4]. (See [A10, (2.3)], for example, and the general review in Chapter 3 prior to Lemma 3.5.2.) It then follows from [A10, (2.3)], the 1 formula above for fG p, xq, and the denition (2.4.5) of fG p, xq (which includes the denition of the factor xr, xy on the right hand side of (2.4.5)) u r that (6.1.5) fG p, xq M pxq xx, y fG pq, x P S ,
r P

r where M is the projection of onto M , and M is the pullback to S of a character on R . The as yet unknown characters M pxq of course are what weaken (6.1.5). They free the identity from its dependence on the ner normalization of intertwining operators in 2.32.4, and the particular pairing xx, y. 1 Shelstads proof of the formula for fG p, xq is contained in [S2] and [S5]. It follows from her proof of [S2, Corollary 5.3.16 and Theorem 5.4.27], the discussion of R-groups in [S2, 5.3], and her conversion of the spectral transfer factors of [S4] to a pairing xx, y represented by the specialization of [S5, Theorem 7.5] to connected quasisplit groups. It is clear that the resulting identity (6.1.5) is indeed a weaker version of the local intertwining identity. It reduces Theorem 2.4.1, in the case of archimedean F and generic p , to a question of base points for the simply transitive action of R on r each of the sets M pGq. We recall, however, that Shelstad has also shown r that if is the unique pB, q-generic representation in , then x , y 1 [S5, Theorem 11.5]. It then follows from Theorem 2.5.1(b) that M p q 1 in this case. The formula (6.1.5) will be our starting point for the general proof of Theorem 2.4.1. Recall that the mapping r p S

306

6. THE LOCAL CLASSIFICATION

for G is generally a proper injection. In using the archimedean parameters with our global arguments, we will need the following lemma. r Lemma 6.1.2. Suppose that G P Esim pN q is a simple endoscopic datum r over our archimedean eld F , and that P pGq is a corresponding generic r p p parameter. Then the image of in S generates S . Proof. We will prove the lemma with G being any connected quasisplit group over F . We will also take the opportunity to recall the initial foundations of real endoscopy, on which the results of this chapter ultimately depend. We rst choose a parameter M P 2 pM, q, for some Levi subgroup M r of G. As we noted in the case G P Esim pN q above, is a disjoint union over p M P M of sets M pGq on which the group R acts simply transitively. The set of characters x , y on S attached to representations P M pGq p is thus a coset R , where is the character x , M y on SM . From the short exact sequence (2.4.9) for G, it is clear that we need only show that p the injective image of M generates SM . We may therefore assume that M G, or in other words, that lies in 2 pGq. Since 2 pGq is assumed not to be empty, G has a discrete series, and therefore a maximal torus T over F that is anisotropic modulo ZpGq. As in r the examples for G P Esim pN q described explicitly above, factors through L T under an admissible L-embedding into L G. Moreover, as the image of a general consequence of the fact that belongs to 2 pGq, the centralizer p p p of its image in G is contained in T . This implies that S T and S p {ZpGq . It is then not hard to see from the denitions that there is a p T p canonical isomorphism from the dual S of S onto the nite 2-group (6.1.6) _ pG, T q{_ pG, T q X t : P X pT q, P F u, where _ pG, T q is the set of co-roots for pG, T q. (This assertion relies on the fact that T {ZpGq is anisotropic, and hence that the norm of any element in _ pG, T q vanishes.) Following Langlands and Shelstad, we write ` DpT q W pG, T q{W GpRq, T pRq , ` where W GpRq, T pRq is the subgroup of the full Weyl group induced by elements in GpRq. This set is bijective with . More precisely, one sees from Harish-Chandras classication of discrete series, and the resulting natural grouping of these objects into packets according to their innitesimal characters, that if a base point 0 P is xed, there is a canonical bijection invp0 , q, P ,

from to DpT q. We also write ` EpT q im H 1 pF, Tsc q H 1 pF, T q ,

6.2. CONSTRUCTION OF GLOBAL REPRESENTATIONS 9

307

where Tsc is the preimage of T in the simply connected cover of the derived p group of G. This group is canonically isomorphic with S . For by the remarks in [K2, 1.1] based on Tate-Nakayama duality, one can write ` p p EpT q im 0 pTsc q^ 0 pT q^ ` p p ^ p T {ZpGq S . But the mapping g g pgq1 , P F , in which g P G normalizes T , descends to a canonical bijection ` DpT q ker H 1 pF, T q H 1 pF, Gq from DpT q onto a set that is in fact contained in EpT q. (See [L9, p. 702].) If 0 P is xed, the mapping of DpT q into EpT q thus gives an embedding p of into S . We take 0 to be the unique pB, q-generic representation in . It then p follows from [S5, Theorem 11.5] that the embedding of into S gives the pairing xx, y, x P S , P , of Theorem 2.2.1(b). (See also [S4, 13].) We are trying to show that the p p image of the set in S generates the group S . It follows from Shelstads initial observation [S1, Lemma 2.1] on the set DpT q that the image in (6.1.6) of the coroot _ of any noncompact root of pG, T q is contained in the image of . We have only to show that such coroots generate the group (6.1.6). Having recalled the background, it will now be very easy for us to prove the lemma. We choose a base for the roots pG, T q of pG, T q by taking the chamber in the Lie algebra of T that contains a Harish-Chandra parameter of 0 . It then follows from [V1, Theorem 6.2(f)] that the simple roots are all noncompact. The simple coroots _ P _ pG, T q therefore map into the image of in (6.1.6). Any coroot _ P _ pG, T q is of course an integral combination of simple coroots. The group (6.1.6) is therefore generated by the images of simple coroots, and hence also by the image of . We p conclude that S is generated by the image of , as claimed. 9 6.2. Construction of global representations We now begin the construction of the global objects needed to establish the local theorems. The methods of this section will be representation theoretic. They will also be more elementary than in earlier chapters, to the extent that they do not depend on the running induction hypotheses we have been carrying since 4.3. We shall construct global representations from local ones as an exercise in the invariant trace formula. r We x a simple endoscopic datum G P Esim pN q over our local eld F . The main goal of this section will be to inate any square-integrable

308

6. THE LOCAL CLASSIFICATION

representation of GpF q to an automorphic representation. The squareintegrability condition rules out the case that F C, so we assume that F is either real or p-adic. 9 We have rst to identify F with a localization of some global eld F . The construction of suitable global elds from local elds is well known, but we may as well formulate what we need as a lemma, for the convenience of the reader. Lemma 6.2.1. Assume that the local eld F is either real or p-adic, and 9 that r0 is a positive integer. Then there is a totally real number eld F with the following properties. 9 9 (i) Fu F for some valuation u of F . 9 (ii) There are at least r0 Archimedean valuations on F . Proof. This is a simple exercise in Galois theory. The main point is 9 to construct a totally real eld F so that (i) holds. There is no problem if 9 F R, since we can then set F Q. We can therefore take F to be a nite extension of the p-adic eld Qp . Then F Qp pq, where is a root of an irreducible monic polynomial q over Qp of degree n. We identify the space of monic polynomials of degree n over a eld E with E n . Since F is open in the union of the nite set of extensions of Qp of degree n, we can replace q by any polynomial in some open neighborhood Up of q in Qn . Consider also the set of monic polynomials of degree n over p R. The subspace of such polynomials with distinct real roots is an open subset U8 of Rn . Since the diagonal image of Qn in Rn Qn is dense, the p intersection Qn X pU8 Up q is nonempty. We can therefore assume that the coecients of q lie in Q, and 9 that the roots of q in C are distinct and real. We then see that F Qpq is 9 9 a totally real eld, and that F Fu for some place u of F over p. To construct a totally real eld with many Archimedean places, we re? 9 9 ? place F by a eld F p q 1 , . . . , q k q, where q1 , . . . , qk are distinct (positive) prime numbers. These primes can be arbitrary if F R, but must be ? chosen so that Qp p q i q Qp if F is p-adic as above. There are innitely many such primes, as one sees readily from Dirichlets theorem on primes in arithmetic progression and the law of quadratic reciprocity (or its second ? 9 ? supplement in case p 2). The eld F p q1 , . . . , qk q then satises the required conditions. We can now construct the automorphic representation. We are assuming r that F is real or p-adic, and that G is the xed datum in Esim pN q. We take 9 9 F to be a totally real global eld with Fu F , as in the lemma, for which the set S8 of archimedean places is large. We then set S8 puq S8 Y tuu

6.2. CONSTRUCTION OF GLOBAL REPRESENTATIONS 9

309

and
u S8 S8 tuu.

To compactify the notation, it will be convenient to denote other objects 9 9 9 associated with F by a dot. For example, we have the adeles A over F , the `9 `9 9 9 r r r r 9 9 endoscopic sets EpN q E GpN q and E v pN q E Gv pN q over F and Fv , `9 9 9 r r for any valuation v on F , as well as the Hecke modules HpN q H GpN q `9 9 9 9 9 r r r r 9 r and Hv pN q H Gv pN q on GpN, Aq and GpN, Fv q. We can write EpN q 9 9 r r r E u pN q and HpN q Hu pN q as usual for corresponding objects attached to 9 F Fu . Lemma 6.2.2. Suppose that P 2 pGq is a square integrable represen9 9 r tation of GpF q. We can then nd a simple endoscopic datum G P E sim pN q, 9 9 9 together with an automorphic representation of GpAq that occurs discretely ` 9 9 9 9 in L2 GpF qzGpAq , with the following properties: 9 9 (i) pGu , u q pG, q. 9 (ii) For any valuation v R S8 puq, v is spherical. u 9 9 9 (iii) For any v P S8 , v is a square integrable representation of GpFv q 9 whose Langlands parameter v P 2 pGv q is in general position. 9 9 Proof. We have rst to dene the endoscopic datum G over F . The p 9 p complex dual group G must be equal to G. This determines G uniquely p SppN, Cq, and up to a quadratic id`le class character 9 if 9 if G e G p SOpN, Cq. There are two kinds of local constraints on . The comG 9 9 9 ponent u of at u must equal the quadratic character G on F determined by the given local endoscopic datum. The other constraints are at u the archimedean places v P S8 , and are imposed by the requirement that 9 Fv q have a discrete series. The question is relevant here only if N is even 9 Gp p and G SOpN, Cq. As we noted in 6.1, the requirement is that v p , 9 v
9 where p 1 N , and v is the sign character on Fv R . We take G 9 2 to be any quadratic id`le class character with the given nite set of local e constraints. (We take to be equal to 1 if all of the constrained local 9 9 9 9 characters v equal 1.) This determines an endoscopic datum G over F such 9 u G, and such that at each v P S8 , the real group Gv pFv q has a 9 9 that G discrete series. 9 To construct the automorphic representation , we turn to a natural generalization of an argument applied by Langlands to the group GLp2q [L9, p. 229]. We shall apply the trace formula to a suitable function f9 on 9 GpAq. (See also [CC] for a similar application of the twisted trace formula for GLpN q.)

310

6. THE LOCAL CLASSIFICATION

Let 9 K 8,u

vRS8 puq

9 Kv

9 9 be a standard maximal compact subgroup of GpA8,u q, the group of points 9 9 G with values in the nite ad`les of F that are 0 at u. We take f98,u to be e 9 8,u on GpA8,u q. At u, we take f9u P HpGq 9 9 the characteristic function of K to be a pseudocoecient f of . In other words, f9u has the property that for any irreducible, tempered representation of GpF q, the relation # ` 1, if , (6.2.1) f9u,G p q tr pf q 0, otherwise,
u holds. At the remaining set of places S8 , we will take a function u f98 pxu q f9v pxv q 8
u vPS8

on the group 9 9u GpF8 q

u vPS8

9 9 GpFv q

obtained by letting each f9v be a stable sum of pseudocoecients corre9 sponding to a Langlands parameter v P 2 pGv q in general position. In other words, f9v fv ,
v Pv

where fv is a pseudocoecient of the square integrable representation v 9 of GpFv q. The existence of pseudocoecients follows from [BDK] in the p-adic case, and [CD] in the archimedean case. 9 9 9 9 Let Z 8,u be the intersection of K 8,u with (the diagonal image in GpA8,u q 9 F q. This actually equals the center of GpF q, a group of 9 9 9 of) the center of Gp p p order 1 or 2, except in the abelian case G SOp2, Cq. In case G does equal 9 Fv q is compact 9 to SOp2, Cq, the existence of discrete series implies that Gp u 9 if v either belongs to S8 or equals u, and therefore that Z 8,u is a nite 9u f9u on GpF u q GpF q be constant 9 9 group. We require that the function f8 8 9 on (the diagonal image of) Z 8,u . We are also requiring that v be in genu eral position for each v P S8 . Finally, it will be convenient to x a place u , and then require that regularity of the innitesimal character of v P S8 v dominate that of the other parameters v1 . We can meet these conditions by rst choosing the parameters v1 and the corresponding function f9u,v f9v1
8
u,v v 1 PS8

u,v 9 so that the product f98 f9u is constant on Z 8,u . At the place v, we then dene v in terms of its innitesimal character by setting

v nv

6.2. CONSTRUCTION OF GLOBAL REPRESENTATIONS 9

311

where v is xed, and n is a large integer such that v represents an element 9 9 in 2 pGv q whose central character on Z 8,u is trivial. The resulting function
u u,v f98 f9u f98 f9v f9u

then has the required properties. We shall apply the trace formula to the function
u f9pxq f98 pxu q f9u pxu q f98,u px8,u q, 8

9 9 x P GpAq,

9 in HpGq. Since f9 is cuspidal at more than one place, we can use the simple 9 version [A5, Theorem 7.1] of the invariant trace formula for G. Since f9 is both cuspidal and stable at some archimedean place, the invariant orbital integrals on the geometric side vanish on conjugacy classes that are not semisimple. (See [A6, Theorem 5.1].) The geometric side of the trace formula therefore reduces to a nite sum ` 9 9 (6.2.2) vol G pF qzG pAq f9G pq 9

of invariant orbital integrals f9G pq 9 f9px1 xq dx,

9 9 9 9 G pAqzGpAq

9 9 taken over the semisimple conjugacy classes in GpF q that are R-elliptic at each place in S8 . u Consider the place v P S8 chosen above. The innitesimal character of v , or more precisely, of any of the representations in the packet of v that dene f9v as a sum of pseudo-coecients, depends linearly on n. It follows from fundamental results of Harish-Chandra [Ha1, Lemma 23], [Ha2, Lemmas 17.4 and 17.5] that (6.2.3)
9 f9G pq cpq nqpG q ,

where cpq is independent of n, and ` 1 9 9 9 qpG q 2 dimpG q rankpG q . The constant cpq vanishes for all but a nite set of , but is nonzero at 9 9 9 1. The exponent qpG q is maximal when the centralizer G of in G 9 which is to say that is central. In this case, equals G, f9G pq f9G p1q, 9 9
u by our condition on f98 . The dominant term in (6.2.1) therefore corresponds to 1. It follows that (6.2.2) is nonzero whenever n is suciently large. The geometric expansion (6.2.2) equals the spectral side of the trace formula. The parabolic terms in [A5, Theorem 7.11(a)], namely those corresponding to proper Levi subgroups M in the general expansion (3.1.1),

312

6. THE LOCAL CLASSIFICATION

vanish on functions that are pseudocoecients of discrete series. (See [A6, p. 268].) The spectral side therefore reduces simply to the trace ` G 9 9 9 mdisc pq f9G pq tr Rdisc pf9q
9

of f9 (operating by right convolution) on the automorphic discrete spectrum 9 of G. It follows that for n suciently large, there is an irreducible represen9 tation of GpAq with positive multiplicity mpq, and with 9 9 ` 9 f9G pq tr pf9q 0. 9 9 It remains to establish that the automorphic representation 9 9u 9 9 9 v 8 b u b 8,u
v

has the properties (i) (iii). At rst glance, these properties appear to be consequences of the denition of f9, and the fact that ` u u ` ` ` 9 (6.2.4) tr f98 p8 q tr f9u pu q tr f98,u p 8,u q tr pf9q 0. 9 9 9 For example, (ii) follows immediately from the denition of f98,u as the 9 characteristic function of K 8,u . However, something more is needed for (i) and (iii). The problem is that the denition of a pseudocoecient (6.2.1) 9 (in the case of (i)) does not rule out the possibility that the localization u could be a nontempered representation whose character on the elliptic set matches that of . We can treat (iii) by an argument based on the innitesimal character. 9 9 As a constituent of the automorphic discrete spectrum of G, is unitary, u , has the same 9 and so therefore are its localizations v . If v belongs to S8 9 v 9 innitesimal character as v , since the trace of v pf9v q is nonzero when f9v is the stable sum of pseudocoecients for v above. It follows from basic estimates for unitary representations, and the fact that v is in general 9 position, that v must be tempered. Appealing again to the denition of f9v , we conclude that v itself belongs to the L-packet v , for the Langlands 9 9 parameter v P 2 pGv q in general position we have chosen. This is the condition (iii). Our proof of (iii) does not apply to (i). For the innitesimal character of 9 u need not be in general position if F R, and is also too weak to help us if F is p-adic. We shall turn instead to a separate global argument, which would also give a dierent proof of (iii). While the argument is somewhat more sophisticated, it makes use only of the results of 3.4, and therefore does not rely on any induction hypothesis. Corollary 3.4.3 tells us that occurs in the subspace 9 ` ` 9 9 9 9 9 9 9 9 GpF qzGpAq L2 GpF qzGpAq L2
9 disc, disc

6.2. CONSTRUCTION OF GLOBAL REPRESENTATIONS 9

313

of the discrete spectrum attached to a (uniquely determined) global param9 9 r 9 eter P pN q. Given , we then use the formula (3.3.14) to write 9 N r Idisc, pf q 9
9 r G PE ell pN q

9 r p rpN, G q Sdisc, pf q, 9

9 9 r r f P HpN q.

9 The quadratic character attached to satises 9 G , 9 9 9

according to the denitions. It follows that the stable linear form Sdisc, 9 9 r attached to any G P E ell pN q vanishes unless G equals G . This yields 9 the usual simplication in the right hand side of the last formula. To reduce 9 r it further, we have to specialize the function f in the global Hecke module 9 r HpN q. We will take decomposable functions

(6.2.5) and (6.2.6)

9 9 9 9 r 9 r r r r f f v f u f u f 8,u 8
v

f9

u f9v f98 f9u f98,u

9 9 r r r 9 in HpN q and HpGq respectively so that for any v, f v and f9v have the same 9 r r 9 images in SpGv q. We impose no further conditions on the components f u and f9u at u. However, we will constrain the other components of f9 as above. We take the component f98,u off9 away from S8 puq to be the characteristic ` u 9 r 9 9 function of K 8,u in H GpA8, q . If v belongs to S8 , we take f9v to be the r 9 image in the symmetric Hecke algebra HpGv q of the stable sum of pseudo9 9 r r r coecients for v . In addition, we choose f v so that its image f in SpG q v v 9 9 PE r r ell,v pN q distinct from Gv . The existence of f v in this 9 vanishes for any Gv case is a consequence of Proposition 2.1.1, as we noted at a similar juncture in the proof of Lemma 5.4.2. 9 u r The condition on a function f v at v P S8 implies that the global transfer 9 9 9 p r r r p 9 f of f vanishes for any G P E ell pN q with G G. Combined with the property of G above, it tells us that 9 r p Sdisc, pf q 0, 9

314

6. THE LOCAL CLASSIFICATION

9 r 9 for any G P E ell pN q distinct from G. The formula given by (3.3.14) therefore reduces to 9 9 9 N r r9 9 pG Idisc, pf q rpN, Gq Sdisc, pf G q 9 9
9 9 G rpN, Gq Sdisc, pf9q. 9

9 We shall combine this with the expansions (4.1.1) and (4.1.2) for G. 1 P E pGq is a global elliptic endoscopic datum for G that is 9 9 Suppose G ell 9 Since the dual group G1 is then composite, the localization p distinct from G. u 1 9 G1 of G1 at any v is distinct from Gv . If v belongs to S8 , the transfer f9v of v r r 9 f9v to SpG1 q therefore vanishes, since f9v is the image in HpGv q of the stable v 9 sum of pseudocoecients for the square integrable parameter v P 2 pGv q. It follows that 1 f91 f9G 0. 9 The proper terms in the endoscopic expansion (4.1.2) of I 9 pf q therefore
disc,

vanish. Similarly, the proper terms in the spectral expansion (4.1.1) of Idisc, pf9q also vanish, as we noted earlier in the proof. It follows that 9 ` 9 G tr Rdisc, pf9q Idisc, pf9q Sdisc, pf9q. 9 9 9 Combining this with the reduction above, we conclude that ` 9 N r 9 (6.2.7) tr Rdisc, pf9q rpN, Gq1 Idisc, pf q. 9 9 We can regard the two sides of (6.2.7) as linear forms in the correspond9 r 9 ing pair of variable functions f9u and f u . From (6.2.4) and the fact that is 9u is a pseudocoecient of , the a constituent of Rdisc, , we know that if f 9 left hand side of (6.2.7) is nonzero. Indeed, it is a nontrivial sum of irre9 9 ducible characters on the group GpAq, taken at the positive denite function f9. The same is therefore true of the right hand side. It then follows from the u denition at any v P S8 of f9v in terms of v , together with the relation of 9 r 9 the corresponding function f v with v , that the innitesimal characters of 9 v and v correspond. At this point we are taking for granted a property of archimedean innitesimal characters, namely that their twisted transfer is compatible with the Kottwitz-Shelstad transfer of functions. The property can be regarded as a part of the work [Me], [S7] in progress. It could also be treated as a more elementary but nonetheless interesting exercise in twisted transfer factors, following its untwisted analogue [A13, (2.6)], for example. 9 The innitesimal character of v is by denition that of the Langlands 9 parameter v . It cannot be in general position if v is not generic. On 9 the other hand, we have agreed that it equals the innitesimal character of 9 r the original parameter v , regarded as a generic element in v pN q. Since 9 v was chosen to be in general position, v must therefore be generic. The

6.2. CONSTRUCTION OF GLOBAL REPRESENTATIONS 9

315

nongeneric component of any global parameter is of course unaected by 9 9 localization. It follows that the global parameter is generic. Its 9 u u is consequently also generic. 9 localization 9 9 We have shown that as a self-dual, local parameter for GLpN q, u u is generic. We will need to know further that it is tempered. We have therefore to show that as an N -dimensional representation of the local Langlands 9 group LF LFv , u is unitary. 9 Consider the expansion (4.1.1) for the term 9 9 N N r r Idisc, pf q Idisc, pf q 9 9 on the right hand side of (6.2.7). Appealing to the general discussion of 4.3, or simply to the local denition (2.2.1) and the properties in 2.5 of local intertwining operators for GLpN q, we see that it is a scalar multiple of 9 9 9 r 9 r r r f N pq. As a function of the variable component f u P HpN q of f , the right hand side of (6.2.7) therefore takes the form 9 r 9 cu P C. (6.2.8) cu f u,N pu q, Since the left hand side of (6.2.7) does not vanish if the corresponding funcr tion f9u P HpGq is a pseudocoecient of , the coecient cu is nonzero. In 9 r r addition, there is a (twisted) cuspidal function f u P Hcusp pN q such that 9 r 9 r f u,N pu q is nonzero. (Recall that Hcusp pN q is the subspace of functions r r r r r f P HpN q such that f belongs to Scusp pG q for every G P Eell pN q, or r r equivalently, such that fM 0 for any proper Levi subset M of GpN q. 9 r r That we can take f u above to be in Hcusp pN q follows from Proposition 9 u lies in the subset ell pN q of pN q. Indeed, the r r 2.1.1.) We conclude that r linear form on HpN q attached to a local parameter in the complement of r ell pN q vanishes on Hcusp pN q. We have been making free use here of the r local Langlands correspondence for GLpN q. In particular, the original interr pretation of ell pN q in 1.2 tells us that as an N -dimensional representation 9 of LF , u decomposes into irreducible constituents that are self-dual and 9 multiplicity free. It follows that u is unitary. In other words, the linear 9 9 r 9 r r form f u,N pu q in f u P HpN q is tempered. We have shown that the right hand side (6.2.8) of (6.2.7) is tempered in 9 r r the variable component f of f u . We want to conclude that the same is then true of the left hand side of (6.2.7). At this point, it would be convenient to be able to say that the local twisted transfer mapping 9 r r r f f u f G r r from HpN q onto SpGq is represented by a correspondence that extends to the underlying Schwartz spaces. A proof of this property would no doubt be accessible. It is perhaps already known, but I do not have a reference. We

316

6. THE LOCAL CLASSIFICATION

need only the special case represented by the twisted tempered character in (6.2.8), and this is not hard to obtain directly. Let me give a very brief sketch of the argument. The twisted tempered character in (6.2.8) has exponents. They are r attached to the chamber a` of any given cuspidal parabolic subset P of r P r GpN q, and are linear forms on the complex vector space a r . If we
P ,C

r evaluate the character at functions in the subspace Creg pN q of Schwartz r r functions on GpN, F q with strongly regular support, we nd that the P exponents satisfy the inequality ` Re pHq 0, H P a` . r
P

r On the other hand, it is easy to describe the image of Creg pN q under the r r transfer mapping f f G . Given Corollary 2.1.2 and the explicit nature of ` r r the twisted transfer factors for GpN q, G , we see that the image of Creg pN q is the space of functions of whose restriction to any maximal torus T pF q r GpF q lies in the subspace of CN -reg pT q of Schwartz functions of strongly r r GpN q-regular support. The variable here represents an OutN pGq-orbit of r strongly GpN q-regular, stable conjugacy classes in GpF q. It is then not hard to show that the P -exponents implicit in the left hand side of (6.2.7) satisfy the analogue of the inequality above. This in turn implies that the left hand side of (6.2.7) does indeed represent a tempered linear form in the variable componet f f9u of f9. 9 Since the global representation is a constituent of the representation 9 Rdisc, on the left hand side of (6.2.7), its local u-component u must then be 9 tempered. It remains only to apply (6.2.4) one more time. If f9u is specialized to a pseudocoecient of the original representation of GpF q, we see that ` 9 9 f9u,G pu q tr u pf9u q 0, 9 for the irreducible representation u of GpF q we now know is tempered. It 9 then follows from the denition of a pseudocoecient that u , as required. We have therefore established the remaining condition (i) of Lemma 6.2.2. Our proof of the lemma is at last complete. 9 9 9 Corollary 6.2.3. Suppose that pF, G, q and pF , G, q are as in the 9 9 r statement of the lemma, and that P pN q is the generic global parameter 9 obtained from in the proof of the lemma. Then for any valuation v u 9 r 9 9 of F , the (OutN pGv q-orbit of the) localization v lies in the local packet 9 9 . In particular, the archimedean parameters xed at the beginning of the v 9 construction represent localizations of . 9 r Proof. We choose a pair pf , f9q of functions (6.2.5) and (6.2.6) as in the proof of lemma, but with two minor dierences. We take f9u to be the xed r 9 pseudocoecient f rather than a variable function in HpGu q, and we take

6.2. CONSTRUCTION OF GLOBAL REPRESENTATIONS 9

317

9 f98,u to be a variable function in the K 8,u -spherical Hecke algebra, rather than the unit. With the two functions being otherwise exactly as before, the earlier conclusions remain valid. In particular, the identity (6.2.7) holds. 9 r 9 Moreover, the right hand side of (6.2.7) is a nonzero multiple of f N pq. It follows that ` 9 r 9 9 9 (6.2.9) f pq cpq tr R 9 pf q ,
N disc,

9 for a nonzero constant cpq. The right hand side of (6.2.9) is nonzero as a linear form in f98,u , since f9G pq is nonzero if f98,u is a unit. Since f98,u is allowed to vary, we see 9 9 r9 9 from the left hand side that for any v R S8 puq, v does belong to v . If v u , we combine (6.2.9) with the twisted archimedean character belongs to S8 9 9 G r 9 relations for GLpN q and the fact that f v transfers to f9v . It follows that v r 9 in this case as well. belongs to
v

Recall that we have been carrying an unproven lemma from 2.3. It is Lemma 2.3.2, which we used to bring global methods to bear on the normalization of local intertwining operators. We can now see that Lemma 2.3.2 is a special case of the lemma we have just established (with the factor G of 2.3 in place of the group G here). Its two assertions (i) and (ii) are just the corresponding assertions of Lemma 6.2.2. Lemma 2.3.2 has therefore now been proved. Returning to Lemma 6.2.2 itself, we note from the early part of the 9 9 construction that the global pair pG, F q depends only on the given local pair pG, F q. In particular, it is independent of the local representation . Bearing this in mind, we formulate an important corollary that is based on a separate hypothesis. 9 9 Corollary 6.2.4. Suppose that F , G, F and G are as in Lemma 6.2.2, and that the local theorems (interpreted as in 6.1) are valid for generic r parameters 1 P pN1 q over F with N1 N . Then for any simple local 9 r 9 r parameter P sim pGq, there is a simple global parameter P sim pGq with the following properties. 9 9 9 (i) pFu , Gu , u q pF, G, q. 9 (ii) For any valuation v R S8 puq, the localization v is spherical. u , is a local parameter in general position in (iii) For any v P S8 9 v r 2 pGv q. 9 r Proof. Given , we obtain a representation P 2 pGq with ` 9 9 r r tr pf9u q f G pq f u,N pq, u 9 r r r r for any functions f9u P HpGq and f u P HpN q with the same image in SpGq. This follows from our assumption that the local theorems hold for G. We 9 9 9 can then construct the automorphic representation of GpAq of the lemma.

318

6. THE LOCAL CLASSIFICATION

9 9 9 r In proving Lemma 6.2.2, we obtained a global parameter in pN q from Corollary 3.4.3, which we then showed was generic. We will see in 9 9 a moment that is simple. Since is a constituent of the representation 9 G 9 9 Rdisc, attached to , this will tell us that G satises the condition of the 9 global datum G of Theorem 1.4.1. 9 The condition (i) applies to the valuation v u. In the proof of the theorem, we showed that 9 r pu q 0, 9 f
u,N

9 r r if f u and f9u have the same image in SpGq, and f9u is a pseudocoecient of . The local theorems we are assuming tell us that the generic parameter 9 u is uniquely determined by this condition. Since the condition also holds for , we conclude that 9 u . 9 This completes the proof of (i). It also tells us that the local parameter u 9 is also simple, as an element in is simple. Therefore the global parameter 9 r pN q. 9 We have shown that G satises the condition of the global datum G 9 9 P E pN q is another 9 r ell imposed on the simple datum by Theorem 1.4.1. If G 9 global datum that satises this condition, one can show that G equals Gu . u 9 9 r r 9 r One applies (3.3.14) to any function f P HpN q with f u,N pu q 0, taking account of Propositions 2.1.1 and 3.4.1. If we grant this property, we can then combine the resulting identity p p p p 9 G G G G u of dual groups with the identity G G 9 9 9 of quadratic characters to deduce that G G. This is the uniqueness 9 P sim pGq, as claimed. r 9 assertion of Theorem 1.4.1. It implies that 9 The conditions (ii) and (iii) apply to valuations v of F that are distinct from u. They follow from Corollary 6.2.3. Remarks. 1. We have not yet made any induction assumption on the 9 r 9 global theorems. Our claim that belongs to sim pGq is based on the original denition provided by conditions of Theorem 1.4.1. We did not fully justify the uniqueness assertion of this theorem in the proof above. There is really no call to do so. For we will impose a global induction hypothesis in 9 r 6.4, in the context of a generic family of global parameters F of the kind considered in 5.4, which will include whatever is needed. We will construct 9 r the family F in the next section. In the process, we will apply Corollary 9 9 6.2.4 as stated, knowing that does determine G by the given conditions

9 6.3. CONSTRUCTION OF GLOBAL PARAMETERS

319

p 9 p G G and G , and that this will give the uniqueness assertion of 9 9 Theorem 1.4.1 once the global induction hypothesis is imposed. 2. In the proof of Lemma 6.2.2, we chose a variable archimedean parameter v in order to work with the simple formula (6.2.3). A reader familar with the identities of Harish-Chandra on which (6.2.3) is based will observe that our condition of general position is sucient. Suppose that r 9 v P 2 pGv q, v P Su ,
8

are archimedean parameters in general position, and that the product of p their corresponding central characters (on Z 8,u if G SOp2, Cq, or simply 9 ZpF q otherwise) with that of is trivial. Then we can choose the global u 9 9 parameter in Corollary 6.2.4 so that v v for each v in S8 . The point here is that the formula (6.2.3) is just a special case of a more general formula, in which the variable n is replaced by a tensor product over v of vectors composed of the half integers (6.1.4) attached to the innitesimal characters of the parameters v . Our condition on the general position of each v means that these half integers are all larger than some preassigned constant. This forces the terms with central to dominate the others, and the sum (6.2.2) to be nonzero. 3. There are other variants of Lemma 6.2.2, which could be proved in the same way. For example, we could construct the automorphic representation 9 9 so that v equals a given square-integrable representation for every v in 9 some nite set V of p-adic places, and so that v is as in (ii) for each v not 9 in S8 puq Y V . Similarly, we could arrange for the global parameter of r 9 Corollary 6.2.4 to be equal to a prescribed element in 2 pGv q at each v P V , and to be as in (ii) at each v outside S8 puq Y V . 9 6.3. Construction of global parameters Our goal for Chapter 6 will be to establish the local theorems for Langlands parameters in the general set r r pN q,
N

taken over the xed local eld F . To this end, we now take on the natural induction hypothesis that will carry us through the chapter. We x N , and assume inductively that the local theorems (interpreted as in 6.1) all hold r for generic parameters P with degpq N . In this section, F is assumed to be real or p-adic, as in 6.2. We x an r elliptic endoscopic datum G P Eell pN q over F , which in general need not be rsim pN q of simple data. We then x a local parameter in the subset E (6.3.1)
1 1

r r

r in pGq, with simple components r i P sim pGi q, r Gi P Esim pNi q.

320

6. THE LOCAL CLASSIFICATION

(Our requirement here that the simple components be self-dual means that lies in the subset r r disc pGq t P pGq : |ZpS q| 8u r of pGq. We observe that this is the local, generic analogue of corresponding global set in the chain (4.1.13).) We assume that is not simple, or equivalently, that each Ni is less than N . The induction hypothesis above r then tells us that the data Gi P Esim pNi q are determined by the uniqueness assertion of Theorem 6.1.1. For clarity we also assume that G is in fact r simple if does not belong to ell pN q, or in other words, if i 1 for some r r i. This insures that pGq injects as a subset of pN q, which is implicit in r the assertion that the compound parameter in (6.3.1) lies in pGq. According to the induction hypothesis, we can apply Corollary 6.2.4 to 9 the pairs pGi , i q. If F is the totally real eld of Lemma 6.2.1, we will obtain global pairs 9 9 9 9 9 r r 9 pGi , i q, Gi P E sim pNi q, i P sim pGi q, that satisfy the conditions (i)(iii) of the lemma. These will determine a 9 9 r global endoscopic datum G P E ell pN q, which will be simple if some i 1, such that the resulting global parameter (6.3.2) 9
1 1

r r

9 r 9 belongs to pGq. Our aim is to establish some ner properties of , for 9 suitable choices of its simple components i . The case that each i 1 in (6.3.1) is at the heart of things. However, we will also have to be concerned with higher multiplicities in dealing with the local intertwining relation. We x a Levi subgroup M of G such that the r 9 set 2 pM, q is nonempty. For a given choice of the global datum G, we will 9 such that 2 pM , q is nonempty. The r 9 9 also choose a global Levi subgroup M general notation, rst introduced for the global parameter ` of Lemma 5.2.3, is relevant here. For example, if N Ni
i

odd

and we have G , r where 1 r i {2s, and G is the unique element in Eell pN q such that i r 2 pG q. We can then identify the local parameter lies in in (6.3.3) M 11 rr
1 1 1 1

i odd

i ,
1 r

M GLpN1 q

GLpNr q

9 6.3. CONSTRUCTION OF GLOBAL PARAMETERS

321

r with an element in 2 pM, q. Similarly, we can identify the corresponding global parameter (6.3.4) 9 9 9 91 M 11 rr
1

r 9 9 with an element in 2 pM , q. The properties we will use are summarized in the multiple conditions of the statement of the following proposition. Proposition 6.3.1. Given the local objects G, , M and M over F as 9 9 9 in (6.3.1) and (6.3.3), we can choose corresponding global objects G, , M 9 M over F as in (6.3.2) and (6.3.4) such that the following conditions 9 and are satised. 9 (i) There is a valuation u of F such that 9 9 9 9 9 pFu , Gu , u , Mu , M,u q pF, G, , M, M q, and such that the canonical maps S 9 and S S 9 are isomorphisms. (ii) For any valuation v outside the set S8 puq, the local Langlands parameter 9 9 9 v 1 1,v r r,v 9 is a direct sum of quasicharacters of Fv , while the corresponding decomposition of the subparameter 9 9 1,v r,v contains at most one ramied quasicharacter. u 9 (iii)(a) Set V S8 . Then for any v P V , the parameters i,v lie in r 2 pGi,v q, and are in relative general position. Moreover, the canonical map9 ping p9 p M,V SM SM , 9 obtained from the combined places v P V , is surjective. (iii)(b) Suppose that each i equals 1. Then there is a v P V with the 9 9 r r property that if v lies in pG q for some G P E sim,v pN q, the dual group v v p p G equals G. v (iii)(c) Suppose that some i is greater than 1. Then there is a v P V such that the kernel of the composition of mappings S Sv Rv 9 9 9 9 contains no element whose image in the global R-group R R pGq belongs 9 9 to R,reg . 9
9 M

SM

322

6. THE LOCAL CLASSIFICATION

Remarks. 1. The statement of the proposition does contain a rather large number of conditions. I have tried to organize them in a way that makes sense. For example, they represent renements of the conditions (i) (iii) of Corollary 6.2.4. In addition, the three parts (a)(c) of (iii) are parallel to the three conditions (5.4.1)(a)(5.4.1)(c) of Assumption 5.4.1. 2. The rst assertion of (iii)(a) means that the integers (6.3.5)
r i1

9 pi,v q ti,v,k : 1 k pi rNi {2s, 1 i ru 9

are all large, and that their dierences are large in absolute value. The second assertion of (iii)(a) is that for any character on the abelian group SM , there is a representation 9 r9 9 9 9 v , v P M,v , V
vPV

r9 in the product M,V of local archimedean packets such that 9 psq xs, v y, s P SM . 9
vPV

3. The conditions (b) and (c) of (iii) are perhaps more palatable here than in the original forms (5.4.1)(b) and (5.4.1)(c), since we have now seen their application to the lemmas of 5.4. 9 Proof. We choose the totally real eld F according to Lemma 6.2.1. 9u F for a xed place u, and that F have 9 The only requirements are that F suciently many real places. We will then apply Corollary 6.2.4 inductively to each of the local pairs pGi , i q, as agreed above. This will yield global 9 9 9 r r 9 endoscopic data Gi P E sim pNi q and generic global parameters i P sim pGi q. 9 i q determine a global endoscopic datum 9 Once chosen, the global pairs pGi , 9 p 9 r 9 9 r 9 p G P E ell pN q with G G, together with a generic global parameter P pGq 9 of G with 9 as in (6.3.2). They also determine a global Levi subgroup M x 9 9 x r 9 9 M M , together with a generic global parameter M P pM , q as in (6.3.4). 9 9 9 9 The conditions Gu G, u , Mu M and M,u M in (i) follow directly from the corresponding conditions of Corollary 6.2.4 for the pairs 9 9 pGi , i q. The isomorphisms of centralizers in (i) are also consequences of the construction. They follow easily from the mappings
LG i

9 > L Gi

9 > and the denitions of the local and global centralizers.

9 6.3. CONSTRUCTION OF GLOBAL PARAMETERS

323

Before we consider the remaining conditions (ii) and (iii)(a)(iii)(c), we 9 9 recall that the construction of each i , and hence of , entails a number of choices. These include a choice of quadratic id`le class character i Gi for e 9 9 p i SOpNi , Cq. If Ni is even, i determines the outer twisting each i with G 9 9 9 9 of Gi . In this case, the local values i,v of i are predetermined at each v in the set S8 puq. If Ni is odd, i determines the L-embedding of L Gi into 9 GLpNi , Cq. Its local value is xed at u, but is otherwise arbitrary. We impose two further local conditions on the family ti u. The rst 9 9 is a constraint on the characters i with Ni odd, at the supplementary archimedean places. We x a set of distinct places V o tvi : Ni oddu
u in the set V S8 . We then require that for any v P V , # pv qpi `1 , if v vi , 9 (6.3.6) i,v pv qpi , if v vi ,

9 where v is the nontrivial quadratic character on Fv R , and 1 pi 2 pNi 1q. This of course requires that there be at least as many places in V as there are indices i with Ni odd, one of our reasons for insisting that 9 F have suciently many real places. The second constraint applies to all 9 characters i , at the places v R S8 puq. We require that for any such v, 9 there be at most one i such that i,v ramies, and for good measure, that the ramication be tame. This is certainly possible. It is well known that one can choose a quadratic extension of any number eld with arbitrarily prescribed localizations at nitely many places. If we choose the characters i successively with increasing i, we can insist that i,v be unramied at any 9 9 place v R S8 puq which divides 2, or at which j,v ramies for some j i. 9 We can now establish (ii). Suppose that v lies in the complement of 9 S8 puq. Corollary 6.2.4 tells us that for any i, the parameter i,v in the 9 r 9 r subset pGi,v q of v pNi q is spherical. According to the remarks in 6.1, this 9 means that it is a direct sum of pNi 1q unramied quasicharacters of Fv , together with another quasicharacter that is unramied if and only if i,v is 9 9 unramied. In particular, the sum v is a direct sum of quasicharacters, as 9 claimed. Our condition above on i,v implies further that there is at most 9 one i such that i,v contains a ramied quasicharacter. This gives the second claim of (ii). It remains to establish the property (iii), with its three parts (a)(c), for u the valuations v in V S8 . We rst recall that for any v P V , we have a decomposition 9 i,v
pi k0

9 i,v,k

324

6. THE LOCAL CLASSIFICATION

for any index i, where ( 9 i,v,k : 1 k pi rNi {2s 9 are irreducible, self-dual, two-dimensional representations of WFv , and i,v,0 9 is a one-dimensional quadratic character that occurs only when Ni is odd. 9 In case Ni is odd, the two-dimensional representations i,v,k are orthogonal, and hence have a nontrivial determinant. It follows that pi 9 9 9 i,v,0 detpi,v q detpi,v,k q1
k1

# 9 i,v pv qpi

v , 1,

if v vi , otherwise,

where vi P V is the valuation xed above. For any i, the two dimensional representations ( 9 i,v,k : 1 k pi , 1 i r are all mutually inequivalent, as we recall from their origins in the proof of Lemma 6.2.2. Apart from a harmless condition on the central character, the only constraint was a regularity condition on the innitesimal characters of these parameters. For a given v, we choose the parameters successively with 9 increasing i. We can then insist that the innitesimal character of i,v be highly regular, in a sense that dominates the condition already imposed on 9 the parameters i,v with j i. The rst assertion of (iii)(a) is then valid. The second assertion of (iii)(a) requires more discussion. We are cer9 9 tainly free to replace M by the subgroup G , since the general linear 9 are connected. In fact, we may as well assume simply that factors of M M G G, to save notation. In other words, we shall assume that the group S S is nite. We have then to show that the global objects can 9 be chosen so that the mapping p p9 9 (6.3.7) V S , S S {ZpGq , 9 9 9 is surjective. Given Lemma 6.1.2, the main step is to verify that the dual mapping p 9 (6.3.8) S pS {ZpGq SV Sv 9 9 9 9
vPV

is injective. The center p p 9 9 ZpGq ZpGqv , v P V,

plays no role here, so it will be enough to verify that the mapping ` S 0 pSV q 0 pSv q 9 9 9
vPV

is injective. We have therefore to show that if s P S lies in the identity 9 0 component S for each v P V , then s 1. 9
v

9 6.3. CONSTRUCTION OF GLOBAL PARAMETERS

325

9 We are regarding v as an N -dimensional representation of WFv WR . 9 Since its two dimensional constituents are distinct, they each contribute a discrete factor Op1, Cq to the centralizer Sv . The one-dimensional con9 stituents are parametrized by the subset I o of indices i with Ni odd. They of course need not be distinct. In fact, for any v P V , they are all trivial unless v vi for some i P I o , in which case the representation corresponding 9 to i is the sign character v of Fv R . It is these repeated factors that dene the identity component of Sv . In particular, in the case v vi , we 9 see that 0 S SOp|I o | 1, Cq. 9
v

0 The intersection of S with S is given by the factors of S attached to 9 9 9

the subset I 1 of indices i P I o with Ni 1, since the factors corresponding to other indices i P I o map to nontrivial components of Sv attached to 9 two-dimensional representations of WFv . If v vi1 , for some i1 P I 1 , the 9 0 intersection of S with S takes the form 9 9
v

tiPI 1 : ii1 u

Op1, Cq

`
9 v

Op1, Cqk ,

k |I o | 2,

where the subscript on the left follows the notation dened prior to (1.4.8), and the exponent on the right is understood to be 0 if |I 1 | 2. The intersection of these groups, as i1 ranges over I 1 , is trivial. In particular, 0 any element in the intersection of S with all of the groups S is trivial. 9 9v We have shown that the mapping (6.3.8) is injective. The dual mapping of abelian character groups p9 p9 SV S is therefore surjective. Recall that we have an injection p9 p9 V pv q SV pSv q. 9 9
vPV vPV

9 It follows from Lemma 6.1.2, applied to each eld Fv , that the image of the p9 set V generates the nite group SV . Its image under the mapping (6.3.7) 9 p therefore generates S 9 . To insure that the image of (6.3.7) actually equals

p9 9 S , we simply enlarge F . That is, we carry out the entire argument anew, 9 9 9 9 with a totally real extension F 1 in place of F , and with 1 1 v for each v 1 of F 1 over v. (In allowing dierent completions 1 to be equal, we 9 1 9 place v v are actually applying the variant of Corollary 6.2.4 described in Remark 2 9 9 following its proof.) If F 1 is large enough, the global parameter 1 obtained as above gives rise to a composition 1 9
V1

p SV S S1 9 9

326

6. THE LOCAL CLASSIFICATION

9 9 9 of surjective mappings. Denoting F 1 and 1 by the original symbols F and 9 , we obtain global data for which the mapping (6.3.7) is indeed surjective. This completes the proof of the second assertion of (iii)(a). Consider next the condition (iii)(b). Then each i equals 1, and the r group S is nite. This is the case in which the datum G P Eell pN q need not be simple. If has any one-dimensional constituents, we take v v1 vi1 , for any xed index i1 P I 1 with Ni1 1. Otherwise, we take v to be 9 r any xed valuation in V . Suppose that v lies in pG q, for some simple v P E equals SOpN, Cq, the two-dimensional 9 p datum Gv pN q over Fv . If Gv sim 9 constituents of v are all orthogonal, since they are mutually inequivalent. By our local induction hypothesis, the localization assertion of Theorem 9 1.4.2 is valid for each Gi . (See also Remark 1 after the proof of Corollary p p 9 6.2.4.) This implies that the group Gi Gi equals SOpNi , Cq, for any i with Ni 1. The same being trivially true if Ni 1, it follows that p p G SOpN, Cq G ,
v

as required. The other possibility is that SppN, Cq. Then all of 9 v are symplectic, which implies that the two-dimensional constituents of p Gi equals SppNi , Cq for any i with Ni 1. If there were any remaining 9 one-dimensional constituents of v , one of them, namely 9 i ,v,0 i ,v , 9
1 1

p G v

would be equal to v , while all of the others would be trivial. Under such 9 circumstances, the image of v cannot be contained in SppN, Cq. Therefore, p there are no one-dimensional representations, and Gi equals SppNi , Cq for all i. It follows that p 9 p G SppN, Cq G , 9 9 r and in particular, that G P E sim pN q is simple. We have established (iii)(b). Consider nally the condition (iii)(c). Then some i is greater than 1, and the group S is innite. In this case, G is required to be simple. We choose v as above, namely to be vi1 if there is an index i i1 with Ni 1, and to be any xed valuation in V otherwise. Let x be an element in S 9 whose image in the global R-group R lies in R,reg . We have to show that 9 9 its image in the local R-group Rv is nontrivial. The fact that R,reg is 9 9 nonempty implies that S is of the form 9 Op2, Cqq Op1, Cqr . In particular, i 2 for each i. If there is an i with Ni 1 and i 2, we choose the index i1 above so that i1 2. In this case, i1 contributes a factor Op2, Cq to the local group Sv . The image of x in R then projects onto 9 9 the nonidentity component of this factor, and is nontrivial as required. If i1 cannot be chosen in this way, there is an i with Ni 1 and i 2. In this case, i contributes a product of several factors Op2, Cq to Sv . The image 9
1

9 6.3. CONSTRUCTION OF GLOBAL PARAMETERS

327

of x in Rv then projects onto the product of nonidentity components, and 9 again is nontrivial as required. This establishes the last of the conditions (iii)(c), completing the proof of the lemma. The construction of Proposition 6.3.1 will allow us to apply global methods to the proof of the local theorems. The essential case is that of p-adic F . However, we will still have work to do in the case F R. This will be based on a variant of the construction, which entails a more restrictive 9 choice of the global eld F . Lemma 6.3.2. Suppose that the given local objects F , G, , M and M are such that F R, and such that the two dimensional constituents (6.3.9) ti : Ni 2u 9 of are in relative general position. Then we can choose global objects F , 9 9 9 9 , M and M that satisfy the conditions of Proposition 6.3.1, and so G, that 9 (6.3.10) i,v i , 1 i r, v P S u .
8

9 Proof. The rst step is to choose the global eld F . Since F R, we 9 can start with the rational eld Q. We will then take F to be the totally ? ? real Galois extension Qp q1 , . . . , qk q of Q, for distinct (positive) prime numbers q1 , . . . , qk , as in the proof of Lemma 6.2.1. We need to take k to 9 be large, in order that the set S8 of archimedean valuations of F be large. Once k is xed, we choose the primes qi to be large relative to the degree 9 2k of F over Q, and so that qi 1pmod 4q. 9 9 Then the discriminant of F equals q1 qk , so that F {Q is unramied at any k . Taking u to be any xed archimedean prime p that is not large relative to 2 9 valuation of F , we obtain 9 Fu F R, as required. 9 9 r The next step is to choose global endoscopic data Gi P E ell pNi q with 9 Gi,u Gi . This amounts to the choice of quadratic id`le class characters i e 9 9 p for F , which are 1 by denition unless Gi SOpNi , Cq. The degrees Ni here are always equal to 1 or 2, since F R. The archimedean local conditions (6.3.11) 9 i,v i ,
u v P S8 , 1 i r,

9 are forced on us by (6.3.10). As in the proof of Proposition 6.3.1, we chose i successively with increasing i so that i,v is unramied at any place v R S8 such that j,v ramies for some j i. In addition to this condition, we can 9 9 also insist that i,v be unramied for any v R S8 whose norm is not large k . We choose each subject to these conditions, and thereby 9i relative to 2 9 9 r obtain the required global endoscopic data Gi P EpNi q.

328

6. THE LOCAL CLASSIFICATION

p Consider an index i such that Gi SOp2, Cq. The archimedean local9 izations (6.3.11) of i then each equal the nontrivial quadratic character R . It follows that the quadratic extension E of F attached to by class 9i 9 9i of R 9 eld theory is totally imaginary. The group Gi is of course the correspond9 9 ing anisotropic form of SOp2q. Let Zi8 Z 8 pGi q be the (central) subgroup of 9 9 9 GpFv q, G8
vPS8

9 dened as in the proof of Lemma 6.2.2. It is the diagonal image in G8 of 8 of GpA8 q. 9 9 9 9 9 the intersection of GpF q with the maximal compact subgroup K 9 We shall show that Zi8 equals Z{2Z. 9 9 We shall rst observe that Zi8 is isomorphic to the cyclic subgroup Ci 9 . Obviously of roots of unity in Ei 9 9 9 Gi pF q tx P Ei : pxq x1 u, 9 9 9 where generates the Galois group Ei {F . Therefore Zi8 is the group of 1 . The Dirichlet unit theorem asserts that 9 units u of Ei such that puq u 9 the group of all units is the direct product of Ci with a free abelian group of 8 is nite, it can contain no unit of 9 rank r [Cassels, 18]. Since the group Zi 9 9 innite order, and is therefore contained in Ci . Since Ei is totally complex, 8 . Therefore Z 8 does equal C 9 9 9i acts by complex conjugation on Zi i 9 i is generated by a primitive Let m 2 be the positive integer such that C mth root of unity 2i m e m . 9 Then Qrm s is the maximal cyclotomic subeld of Ei . If m is greater than 2, there is a prime number p m that ramies in Qrm s, and hence also in 9 9 9 9 Ei . On the other hand, we know from the choice of F and i that Ei {Q can ramify at p only if p is very large. Indeed, it follows from the construction 9 that any such p must be large relative to the degree 2k`1 of Ei over Q. But k`1 , and approaches innity as m the degree pmq of Qrm s{Q both divides 2 9 becomes large. This contradicts the existence of p. Therefore m 2, and Ei contains only the square root p1q of unity. We have therefore established that 9 9 Zi8 Ci Z{2Z. We note in passing that our ramication condition on the primes qi is more stringent than necessary. The argument we have just given would apply to 9 any totally real extension F {Q that ramies only at primes that are large relative its degree. 9 9 r 9 We now use Corollary 6.2.4 to construct global parameters i P sim pGi q. k is even, the product of |S |-copies of the central character Since |S8 | 2 8 9 of i on the group Zi8 Z{2Z equals 1. According to Remark 2 follow9 ing Corollary 6.2.4, we can then choose the simple parameters i so that

9 6.3. CONSTRUCTION OF GLOBAL PARAMETERS

329

(6.3.10) holds. This construction is not the same as that of the proof of Lemma 6.3.2. For example, the earlier property (6.3.6) is incompatible with the requirement (6.3.10) here. However, the conditions (i), (ii) and (iii) of 9 the proposition remain valid for the resulting compound parameters and 9 M here. Indeed, the assertions (iii) that relied on (6.3.6) in the earlier construction are now trivial consequences of (6.3.10). The lemma follows. Remark. In following parts of the proof of Proposition 6.3.1, we have made the argument above a little more elaborate than necessary. It would have been sucient to take 9 9 9 Ei F Qi , 9 for a suitable imaginary quadratic extension Qi of Q. However, the argument we have given could conceivably be of interest in its own right. The primary objects in both Proposition 6.3.1 and Lemma 6.3.2 are the simple global parameters 9 i , 1 i r.

They are in turn given by Corollary 6.2.4, which relies on the application of the simple trace formula in Lemma 6.2.2 that is the source of the underlying global construction. Once chosen, they determine the required global pair 9 9 pG, q from the given local pair pG, q. We took the secondary objects G and as the given data in the proposition for the obvious reason that they are the objects to which the local theorems apply. Let us put them aside for a moment. Suppose instead that we have been given only the simple constituents i of , and that we have 9 constructed corresponding simple global parameters i by Corollary 6.2.4. We can then form the associated family (6.3.12) ( 9 9 9 9 r r 9 F Fp1 , . . . , r q 91 1 9r r : 9i 0

of compound global parameters. The supplementary constraints we imposed on the parameters during the proofs of Proposition 6.3.1 and Lemma 6.3.2 9 are independent of G and G. More precisely, the primary objects can be chosen independently of any pair 9 9 pG, q, 9 9 9 9 9 9 9 r P FpGq, G P E ell pN q, N 0,

9 9 r 9 (with G being simple if R ell pN q). The conditions (iii)(a)(iii)(c) 9 r of Proposition 6.3.1 then imply that F satises Assumption 5.4.1, with 9 u r V V pFq being the set of archimedean places S8 . We will therefore be 9 r able to apply the global results of 5.4 to F.

330

6. THE LOCAL CLASSIFICATION

6.4. The local intertwining relation for We turn now to our main task, the proof of the local theorems. We are treating local generic (Langlands) parameters in this chapter. Our general goal is to apply the inductive procedure of Chapter 5 to the global parameters constructed in the last section. In this section, we shall establish the local intertwining identity of Theorem 2.4.1. We are taking F to be a local eld. If F is the complex eld C, the case (ii) of Corollary 2.5.2 tells us that the relevant intertwining operator equals 1. Theorem 2.4.1 then follows from the denitions, and the fact that any centralizer S is connected in this case. We can therefore assume that F is real or p-adic as in the last two sections. r We x a datum G P Esim pN q over the local eld F and a local Langlands parameter 1 1 r r r in pGq, as in (6.3.1). In this section we will assume that some i is greater r than 1, or in other words, that does not lie in 2 pGq. Given G and , we 9 9 , G, , M and M that satisfy the conditions of 9 9 9 choose the global objects F 9 9 Proposition 6.3.1. Having xed the simple local constituents 1 , . . . , r of 9 we then form the family , 9 9 r r 9 F Fp1 , . . . , r q of global parameters (6.3.12). As we have noted, the conditions of Assump9 u r tion 5.4.1 follow for F and V S8 from the conditions (iii) of Proposition 6.3.1. The construction of the global objects relies on the induction hypothesis from 6.3 that the local theorems are valid for all parameters of degree less than N . To this, we now add the formal global induction hypothesis promised in Remark 1 following Corollary 6.2.4. We assume that the global 9 r theorems hold for global parameters of degree less than N in the set F. The induction arguments of Chapter 5 therefore apply to the parameters in 9 r F. In particular, they will allow us to extend our global hypothesis to the 9 9 r compound parameter P pN q of this section. The rst step in the global induction argument, the case of simple parameters, will be resolved in 6.7. We studied the global intertwining relation in 4.5. We established it directly, by induction arguments that can clearly be carried out in the context 9 r of global parameters in the set F, in the cases outlined in Corollary 4.5.2. As we remarked after this corollary, the reduction also carries over without change to the local intertwining relation of Theorem 2.4.1. In other words, the local intertwining relation holds for our generic parameter , unless it r r belongs to ell pG q for some G P Esim pN q, or unless it belongs to the local analogue for of one of the two exceptional classes (4.5.11) or (4.5.12). In each of these cases, it suces to consider elements s and u that map to

6.4. THE LOCAL INTERTWINING RELATION FOR

331

a point x in the subset S,ell of S . (We understand S,ell to be the local analogue of the global set S,ell introduced near the beginning of 5.2.) We 9 observe that Proposition 4.5.1 also yields the other global theorems for , unless falls into the rst of the three cases above. We thus have three exceptional cases to deal with. We shall study them together. In each case, we shall apply the results of Chapter 5 (and the 9 assertion of Proposition 4.5.1) to the global parameter . Keep in mind 9 9 that since is generic, the point s and the sign character G are both 9 9 trivial. r The case of P ell pGq is characterized by the local generic form of (5.2.4). The conditions are # 21 2q q`1 r , (6.4.1) ` ` S Op2, Cqq Op1, Cqrq , q 1, with the requirement that the Weyl group W contain an element w in the regular set W,reg . The other two cases are represented by (4.5.11) and (4.5.12) (which is to say the local generic forms of (4.5.11) and (4.5.12) with and in place of and ). In all three cases, the associated global pair 9 9 pG, q satises an identity ` 1 9 9 9 r 9 9 (6.4.2) f9G p, xq f9G p, xq 0, f9 P HpGq, 9 9
xPS,ell

9 where x is the isomorphic image of x in S . This follows from Lemmas 9 5.4.3 and 5.4.4 in the case (6.4.1), and Corollary 4.5.2 (together with the assertion of Lemma 5.4.4) in the cases (4.5.11) and (4.5.12). Recall that r if q r 1 in (6.4.1), there is a second group G_ P Esim pN q such that _ q, and such that S _ Spp2, Cq. The corresponding global pair r P pG _ , q actually belongs to the second case (4.5.11) (with pG, q equal to 9 pG 9 9 9 pG, _ q). It represents that part of the specialization to r 1 of (4.5.11) that was excluded by the hypotheses of Proposition 4.5.1. Since this was later covered by Lemma 5.4.4, in the assertion that (5.2.13) vanishes, we can include it now in our discussion of (4.5.11). In the cases (4.5.11) and (4.5.12), we set q 1. We assume that f9 f9v
v

is a decomposable function. The summand in (6.4.2) can then be decomposed into a dierence of products 1 9 9 9 9 f9v,G pv , xv q f9v,G pv , xv q 9 9
v v

over all valuations v. In order to remove the places v outside S8 puq from this equation, we have rst to establish a very special case of the local intertwining relation.

332

6. THE LOCAL CLASSIFICATION

Lemma 6.4.1. Suppose that F is nonarchimedean. Then the local intertwining relation holds in case r 3 and Ni 1, 1 i r.

Proof. The constraints we have imposed here are obviously pretty stringent. Applying them to the cases (6.4.1), (4.5.11) and (4.5.12), we see that the group M is particularly simple. In fact, M is abelian except in the case (4.5.12), with r 3, where we have G Mder Spp2q. Notice that this last case is the setting of Corollary 2.5.2(v). One could perhaps use the earlier corollary to give a direct local proof of the lemma, but there would still be a number of supplementary points that I have not tried to verify. We shall instead give a less direct proof based on the application of Lemma 2.5.5. We must show that the objects G, M , and M over F satisfy the conditions (i)(iii) of Lemma 2.5.5. The condition (ii) follows from [KS, 5.3]. (We assume here that the stable distribution on M pF q attached to M in [LL] is the same as the linear form given by Theorem 2.2.1(a), an induction hypothesis we will resolve in Lemma 6.6.2.) Before we consider the condition (i) on char F , we shall establish the nal condition (iii) of Lemma 2.5.5. The global identity (6.4.2) in the restricted case here is based on the id`le class characters e 9 9 i i , 1 i r, of order 1 or 2, which are obtained in Proposition 6.3.1 from the given u characters i i on F . Suppose that v lies in the set S8 S8 of 9 archimedean places. Since a completion v at v must be equal to p`1q or r 9 are all induced from a minimal parabolic p1q, the representations in v subgroup. It then follows from the remark on [S5, Lemma 11.5] following (6.1.5) that the characters M ,v p q in (6.1.5) are all trivial. Therefore 9
1 9 9 9 9 f9v,G pv , xv q f9v,G pv , xv q, 9

x P S .

The identity (6.4.2) becomes 1 9 9 9 9 9 9 f9v,G pv , xv q f9v,G pv , xv q f98 p8 , x8 q, 9 9

xPS,ell vRS8 vRS8

where 9 9 f98 p8 , x8 q
vPS8

9 9 f9v,G pv , xv q. 9

It is a direct consequence of the condition (iii)(a) of Proposition 6.3.1 that r 9 the linear mapping from functions f98 in HpG8 q to functions 9 9 x f98 p8 , x8 q, x P S ,

6.4. THE LOCAL INTERTWINING RELATION FOR

333

on S is surjective. This implies that 1 9 9 9 9 f9v,G pv , xv q, f9v,G pv , xv q (6.4.3) 9 9

vRS8 vRS8

for any x P S,ell . 9 It follows from the fundamental lemma and Theorem 2.5.1(b) that almost all of the factors on each side of (6.4.3) equals 1. From the denitions 9 and the local correspondence for Mv (which we are assuming by induction), it follows that for any v R S8 puq and x P S,ell , we can choose f9v so that the corresponding factor on the right hand side of (6.4.3) is nonzero. This 9 allows us to reduce (6.4.3) to an identity in f f9u and u . We obtain (6.4.4)
1 fG p, xq epxq fG p, xq,

r x P S,ell , f P HpGq,

for a complex number epxq eu pxu q 9

vRS8 puq

1 9 9 9 9 f9v,G pv , xv q f9v,G pv , xv q1 9 9

that is independent of f . The identity (6.4.4) would seem to be a little weaker than the condition (iii) of Lemma 2.5.5. However, with the elliptic elements x P S,ell acounted for, the condition follows easily for a general pair ps, uq by induction from the usual arguments of descent. If we impose the condition (i) that charpF q 2, will then satisfy the three requirements (i), (ii) and (iii) of Lemma 2.5.5. The lemma then tells us that the coecient epxq equals 1, and therefore that the local intertwining relation is valid if charpF q 2. It remains to deal with the case that charpF q 2. From (6.4.3), (6.4.4), and what we have just established, we see that 9 ev pxv q 1,
vPS2

9 9 where S2 is the set of valuations of F with charpFv q 2. The totally 9 constructed in the proof of Lemma 6.2.1 has the property that real eld F 9 Fv F for each v P S2 . A slightly more sophisticated construction, which 9 we leave to the reader, would give another totally real eld F with the 9 F for every v P S , but such that property Fv 2
|S2 | |S2 | ` 1 2k ` 1.

(The point is that any totally real eld has a totally real extension of any given degree, in which a given prime splits completely.) In applying Propo9 9 9 sition 6.3.1 to both F and F , we can choose the quadratic characters i so that P S not equal 9 9 i,v i,v for any valuations v P S2 and v 9 and i 2 to the distinguished valuations u P S2 and u P S2 . It follows that 9 9 ev pxv q ev pxv q eu pxq, x P S,ell ,

334

6. THE LOCAL CLASSIFICATION

for any such v and v , where eu pxq 0 is a xed complex number. We conclude that 1 9v 9 eu pxq ev px q ev pxv q 1,
v PS2

vPS2

and therefore that epxq epxq

vu

9 ev pxv q 1. eu pxq
vPS2

The local intertwining relation for charpF q 2 then follows again from (6.4.4). We now return to the general case. We are trying to extract the local intertwining relation from the global identity (6.4.2), which we recall applies 9 to the global parameter attached to by Proposition 6.3.1. We shall exploit the dierent factors of the terms of (6.4.2) attached to a decomposable function f9 f9v . The rst step is to remove the factors at places v R S8 puq. We claim that they satisfy the local intertwining relation (6.4.5)
1 9 9 9 9 f9v,G pv , xv q f9v,G pv , xv q, 9 9

v R S8 puq.

9 9 To see this, we recall that v is a direct sum of quasicharacters of Fv by 9 Proposition 6.3.1(ii). If one of these constituents is not self-dual, v does not r 9 factor through an element in 2 pMv q. It is then easy to see that xv does not belong to Sv ,ell , or equivalently, that the Weyl element wx does not belong 9 to Wv ,reg . The identity (6.4.5) then follows inductively by reductions similar 9 9 to those in 4.5 prior to Proposition 4.5.1. If all of the constituents of v are self-dual, it follows from the second assertion of Proposition 6.3.1(ii) that 9 9 there are at most three distinct constituents. The local pair pGv , v q then satises the condition of Lemma 6.4.1, and so (6.4.5) again follows. Since we can choose the functions f9v in (6.4.5) so that each side is nonzero, we can remove these factors from the global identity (6.4.2). The identity becomes 1 9 9 9 9 (6.4.6) f9v,G pv , xv q f9v,G pv , xv q 0. 9 9
xPS,ell vPS8 puq vPS8 puq

For what remains to be proved of the local intertwining relation postulated in Theorem 2.4.1, the main step turns out to be the case of certain archimedean parameters . There ought to be a local proof in this case, perhaps following from the methods of [KnZ2] and [S3][S5]. However, this still seems illusive, to me at least, so we shall continue to apply the global methods above. We do know that there is nothing further to prove in case F C. We shall therefore assume for the present that F R. The degrees of the irreducible constituents i of are then equal to 1 and 2.

6.4. THE LOCAL INTERTWINING RELATION FOR

335

r We are writing M as usual for a xed element in 2 pM, q, and we identify SM with a subgroup of S . The action of SM on S by translation stabilizes the subset S,ell , and gives a simply transitive action x x xM , x P S,ell , xM P SM ,

of SM on S,ell . In the case F R under present consideration, we already have much information about the associated L-packets. According to the r results of Shelstad reviewed in 6.1, we have the archimedean packets r of irreducible representations, and a surjective mapping M and M r r r from to M . We also have a canonical pairing x , y on S such that 1 xx, y fG pq, x P S , (6.4.7) fG p, xq
r P

and (6.4.8) fG p, xq
r P

pM q xx, y fG pq,

x P S,ell ,

r for any function f P HpGq in the associated archimedean Hecke algebra. In the second equation, x is restricted to the subset S,ell of S , and pM q M pxq 1 is the corresponding restriction of the character M in (6.1.5), which der pends only on the image M of in M . The local intertwining relation asserts that pM q equals 1. Lemma 6.4.2. Assume that F R, and that is in general position. Then there is an element 1 P SM such that
1 fG p, xq fG p, x 1 q,

r f P HpGq, x P S,ell .

Proof. The lemma asserts simply that the sign takes the special form pM q x1 , M y, r M P M .

In other words, pM q can be identied with the restriction to the image r p p of M in SM of a character on SM . Given that it treats only a pretty severe specialization of both F and , the lemma would appear to be a pretty minor step. In point of fact, it represents essential preparation for the general argument. Its proof amounts to an exercise in the Fourier analysis on a nite abelian group. 9 9 9 9 9 9 We choose global data pF , G, , M , M q as in Lemma 6.3.2. Then F is a totally real eld, whose set S8 of archimedean places is large. The valuation u belongs to S8 , and we have 9 9 9 9 9 pFv , Gv , v , Mv , M,v q pF, G, , M, M q,

336

6. THE LOCAL CLASSIFICATION

for each v P S8 . The identity (6.4.6) becomes 1 f9v,G p, xq 0, f9v,G p, xq (6.4.9)

xPS,ell vPS8 vPS8

for a test function f98

vPS8

f9v ,

r f9v P HpGq,

r which is now a nite product of functions from the same space HpGq. r If f P HpGq is xed, we can regard the function (6.4.7) of x P S as an r inverse Fourier transform of the function fG pq of P . In particular, 1 p, xq as a Fourier transform on S . It will we can recover fG pq from fG be convenient to take a partial Fourier transform, relative to the variable represented by the SM -torsor S,ell in S . Bearing in mind the general lack of a canonical splitting for the sequence 1 SM S R 1, we dene
1 fG p, q |SM |1

xPS,ell

1 fG p, xq pxq1 ,

p for any character P S . The function

1 apq fG p, q,

p P S ,

obviously satises the equivariance relation (6.4.10) ap R q R prell q1 apq, p R P R ,

p p under translation by the subgroup R of S , where rell rpxq is the image in R of (any) x P S,ell . We can therefore write
1 fG p, xq a_ pxq,

x P S,ell , x P S,ell .

for the inverse transform (6.4.11) a_ pxq

p PSM

apq pxq,

p The last summand is obviously invariant under translation of P S by R , p p p and therefore does represent a function of in the quotient SM S {R . We dene the partial Fourier transform fG p, q of fG p, xq in the same way. It is then easy to see that
1 fG p, q pq fG p, q,

p P S ,

for the function pq # r pM q, if x , M y, for some M P M , 0, otherwise,

6.4. THE LOCAL INTERTWINING RELATION FOR

337

p p p on SM S {R . (As above, we are making no distinction between a p p p function on SM and an R -invariant function on S .) The original function is therefore the inverse transform fG p, xq paq_ pxq of the product
1 paqpq pq apq pq fG p, q. These denitions are of course examples of formal operations in nite abelian p Fourier analysis. In general, any function a on S that satises (6.4.10) has an inverse transform (6.4.11) on S,ell . It satises the inversion formula apq |SM |1 a_ pxq pxq1 , xPS,ell

which specializes to
xPS,ell

a_ pxq |SM | ap1q.

Any two functions have a convolution pa1 a2 qpq

p 1 PSM

1 a1 p1 q a2 p1 q,

which satises pa1 a2 q_ a_ a_ . 1 2 We shall use these operations to transform (6.4.9) to the assertion of the lemma. We order the valuations in S8 as v1 , . . . , vn , and set p ai pq av pq f91 p, q, P S ,
i

vi ,G

for 1 i n. The left hand side of (6.4.9) can then be written n n pai q_ pxq a_ pxq i
xPS,ell i1 i1

`
x

` _ pa1 an q_ pxq pa1 q pan q pxq

` ` |SM | pa1 an qp1q pa1 q pan q p1q ` |SM | a1 p1 q an pn q p1 qa1 p1 q pn qapn q

1 ,...,n

|SM |
1 ,...,n

1 p1 q pn q a1 p1 q an pn q,

where the last two sums are taken over the set ( p p1 , . . . , n q P pSM qn : 1 n 1 . Suppose that p1 , . . . , n q is a multiple index such for each i, i belongs to r p r the image of M in SM . We can then choose the functions f9vi P HpGq

338

6. THE LOCAL CLASSIFICATION

p so that for any P S , ai pq is nonzero if lies in the coset of i , and is zero otherwise. This follows from the results of Shelstad, by which R is r identied with the representation theoretic R-group RM of any M P M . The functions f9vi serve to isolate the summand of p1 , . . . , n q as the only nonvanishing term in our expression for the left hand side of (6.4.9). Since the right hand side of (6.4.9) vanishes, we conclude that (6.4.12) p1 q pn q 1, if 1 n 1.

p r The character 1 in SM belongs to the image of M . In fact, it follows from [S5, Theorem 11.5] that corresponds to the pBM , M qr generic representation M P M . Applying the identity (2.5.4) of Theorem 2.5.1(b) for the pB, q-generic constituent of IP pM q to the denition of the sign pM q, we see that p1q 1. We are assuming that the integer n |S8 | in (6.4.12) is xed and large. Having now established that p1q 1, we are free to replace n by any integer m n, since we can always set i 1 for i m. We can therefore assume also that the mapping preserves any relation 1 m 1, m 1, r in S . Since this image generp among elements in the image of M M p ates SM , by condition (iii)(a) of Proposition 6.3.1, extends uniquely to a p character on the group S . Therefore
M

pq p1 q,

p P SM ,

for a unique element 1 P SM . It follows from (6.4.7) and (6.4.8) that fG p, xq pM q xx, y fG pq
r P

r P

x1 , M y xx, y fG pq
1 xx 1 , y fG pq fG p, x 1 q,

r P

as required. Lemma 6.4.3. The formula of the last lemma holds with 1 1. Proof. We are taking to be a parameter (6.4.1), (4.5.11) or (4.5.12) r in pGq over F R. Our assumption is that is in general position, in the sense that its simple constituents i with Ni 2 are in relative general position. In the proof of the last lemma, we applied the global construction of Lemma 6.3.2. In particular, we worked with identical archimedean parameters (6.4.13) 9 i,v i , 1 i r, v P S8 ,

6.4. THE LOCAL INTERTWINING RELATION FOR

339

9 attached to the global eld F . In the proof of this lemma, we shall use the p-adic form of the construction in Proposition 6.3.1. This will give us supplementary information about . The logic of the construction will actually be slightly convoluted, since we will choose the objects of the proposition in a dierent order. The irreducible degrees Ni of equal 1 or 2. We introduce new degrees by setting # Nj , if j q, 7 Nj Nq`1 ` ` Nr , if j q ` 1, where j ranges from 1 to pq`1q. In principle, we will be applying Proposition 6.3.1 to a local parameter 7 with simple components r 7 P sim pG7 q, j j
7 r G7 P Esim pNj q, j 1, . . . , q 7 1 1

7 q q

7 , q`1

over a p-adic eld F 7 , and multiplicities

that match those of . 9 7 9 r However, we will need to choose endoscopic data G7 P E sim pNj q over a j 9 global eld F 7 before we introduce 7 . 9 9 We may as well let F 7 be the global eld F chosen in the proof of the 7 9 last lemma. For each j q, we then take Gj to be the global endoscopic 9 9 datum Gj over F that was constructed also in the proof of the last lemma. 9 The remaining datum G7 is determined by the requirements that q`1 p p p 9 9 9 p p G7 Gq`1 Gr Gq`1 Gr q`1 and q`1 G7 97 9
r

iq`1

q`1

9 i .

p 9 (By convention, i 1 in case Gi SppNi , Cq.) Once these objects are 9 9 chosen, we x a nonarchimedean place u7 of F such that for each j, the local endoscopic datum 9 G7 G 7 7
j j,u

9 is elliptic, and hence simple, over the local eld F 7 Fu7 . (This last condip7 9 tion is needed only in case Gj SOp2, Cq.) We then choose local parameters r 7 P sim pN 7 q over the p-adic eld F 7 whose central character on the nite
j j

98 9 group Zj Z 8 pGj q is trivial. The existence of 7 follows easily from our j r induction hypothesis that the local theorems hold for parameters in pN 7 q
j

over F 7 .

340

6. THE LOCAL CLASSIFICATION

9 Once we have the local parameters 7 and the global data G7 , we choose j j 7 7 9 r 9 9 global parameters j P sim pGj q over F according to Corollary 6.2.4. In fact, by Remark 2 following Corollary 6.2.4, we can choose them so that # j , if j q, 7 9 j,v q`1 r , if j q ` 1, for any v P S8 . This is because the product over v of the associated central 98 9 9 characters with that of 7 will be trivial on Zj . The global pairs pG7 , 7 q j j j 9 7 PE 9 then determine a global endoscopic datum G r sim pN q, as at the beginning of 6.3, together with a global parameter 9 9 9 9 (6.4.14) 7 1 7 r 7 7
1 q q`1

r 9 in pG7 q such that 9 7 , v P S8 . v 7 and 7 are not those of the proof of Proposition 9 9 The global objects G 9 6.3.1, since the archimedean components of 7 are all equal. They nonethe9 9 less satisfy the conditions of the proposition (with M 7 and 7 obtained M 9 9 from G7 and 7 as at the beginning of 6.3). The conditions are in fact obvious consequences of the restricted nature of the multiplicities i in (6.4.14). For example, in the cases (6.4.1) or (4.5.11) that the multiplicities are all 9 9 9 2, the datum G7 is just the factor G7 of M 7 that is classical, namely an q`1 9 9 orthogonal or symplectic group. Since the associated parameter 7 7
q`1

is then simple, the group S7 S7 in condition (iii)(a) actually equals 1, 9 M M making the condition trivial. The other conditions are either equally clear or as in the proof of the proposition. It therefore follows that the relation 9 (6.4.6) holds for 7 . We shall apply it to the case that the group S7 is M trivial. With the assumption that S7 t1u, the sum over x P S7 ,ell in (6.4.6) M reduces to a single element x1 . We identify x1 with the base point in the 97 larger set S,ell . Since Mv M for any v P S8 , the identity reduces in this case to (6.4.15) f9v,G p, x1 1 q f 7 p7 , x1 q f9v,G p, x1 q pf 7 q1 p7 , x1 q,
G G v v

for any function

v

f9v f 7 ,

r r 9 f 7 P HpG7 q, f9v P HpGv q, v P S8 .

Suppose that the point 1 is not equal to 1. Then the linear forms r fG p, x1 q and fG p, x1 1 q in f P HpGq are linearly independent. This follows from the denition (2.4.5) of fG p, xq, the disjointness of the irreducible constituents of induced tempered representations, and the fact that the image of r p r r M generates SM . Let f fv1 be a variable function in HpGq HpGv1 q

6.4. THE LOCAL INTERTWINING RELATION FOR

341

at some xed place v1 P S8 . At the other places v P S8 , and also at the place u7 , we x functions so that the contributions to the left hand side of (6.4.15) do not vanish. It follows that fG p, x1 q c1 fG p, x1 1 q, r f P HpGq,

for a constant c1 that is independent of f . This is a contradiction. We have established the required condition that 1 1, for parameters such that the group S7 equals t1u. This includes the basic elliptic case 9
M

(6.4.1), and the rst supplementary case (4.5.11). The remaining supplementary case (4.5.12), in which q 1 and 1 3, is slightly dierent. In 9 9 9 9 this case, the classical factor G7 of M 7 is larger than the group G7 G7 . q`1 2 The associated global parameter equals 9 9 9 9 9 7 7 7 7 7 , q 2 1 q`1
7 and is not simple. If it satises the further two conditions that N1 N1 1 7 and N2 is odd, the rank 7 7 N 3N1 ` N2 ` ` Nr 2 ` N1 ` N2 ` is even, and the character `7 on the group Op1, CqOp1, Cq is nontrivial. 9 M

We then see from (1.4.9) that S7 is isomorphic to Z{2Z, and also that there 9 M p is a nontrivial central element in the dual group G. The quotient S 9 7 of S7 is therefore trivial, and we still have 1 1. If these two conditions are 9 M not met, however, S7 has order 2, and the argument fails. 9
M

It therefore remains to establish the lemma in the case (4.5.12), under 7 7 the further restriction that either N1 2 or N2 is even. Since there are only two self-dual, one-dimensional characters on F R_ , this restriction 7 implies that if Ni N1 1, then Ni 2 for every i 1. There are several ways to proceed. The simplest perhaps is to introduce a new parameter 5 21 5 2 r , 1 r 5 P sim pN1 q, 1

r in pGq over F R. That is, we replace one of the three copies of 1 in the original parameter with a simple parameter 5 that is dierent from all 1 the others. If N1 1, for example, the last restriction allows us to set 5 1 equal to 1 R . If we apply Proposition 6.3.1 to 5 , we obtain a parameter 9 9 9 5 over the global eld F . In fact, one nds that 5 can be chosen so that # , if v v0 , 9 5 v 5 , if v v0 , for any v P S8 , where v0 is a xed valuation in S8 , and so that the conditions 9 9 of Proposition 6.3.1 remain valid for 5 . Then for any v v0 , 5 is an v archimedean parameter for which we have already established the lemma.

342

6. THE LOCAL CLASSIFICATION

We leave the reader to check that the resulting application of (6.4.6) yields 9 the required property 1 1 for the remaining parameter 5 0 . v The last two lemmas represent the main step. They yield the local intertwining relation of Theorem 2.4.1 in the special case that F R and is in general position. With this information, it will be possible to deduce the general case. Proposition 6.4.4. Theorem 2.4.1 is valid for generic parameters P r pGq over the general local eld F . Proof. As we agreed at the beginning of the section, the theorem holds unless belongs to one of the three exceptional cases (6.4.1), (4.5.11) and (4.5.12), and the given elements s and u map to a point x in the set S,ell . We must allow F to be either real or p-adic, since the parameters over R treated above were required to be in general position. We therefore take r P pGq to be an arbitrary generic parameter over the given eld F , of the form (6.4.1), (4.5.11) or (4.5.12). We shall apply Proposition 6.3.1 directly to . From the objects G, , 9 9 9 9 M and M over F attached to , we obtain global objects G, , M and M 9 that satisfy the conditions of the proposition. We then obtain the over F identity (6.4.2) from the results of 5.4 and Corollary 4.5.2, and its reduction (6.4.6) implied by Lemma 6.4.1. It follows that the sum 91 pv , xv q f 1 p, xq 9 9 9v,G pv , xv q fG p, xq 9 9 fv,G f G
xPS,ell
u vPS8 u vPS8

vanishes, for any decomposable function u r 9 r f9u f f98 f, f9v P HpGv q, f P HpGq.
u vPS8

u 9 r 9 For any v P S8 , the parameter v P pGv q is in general position. It follows from Lemmas 6.4.2 and 6.4.3 that 1 9 9 9 9 f9v,G pv , xv q f9v,G pv , xv q,

x P S,ell .

The last identity then reduces to a linear relation ` 1 u f98,G p, xq fG p, xq fG p, xq 0 9

xPS,ell

in f , with coecients
u f98,G p, xq 9

u vPS8

9 9 f9v,G pv , xv q. 9

u As linear forms in f98 , the coecients are linearly independent. This follows from the denition (2.4.5), the disjointness of constituents of induced tempered representations, and the condition (iii)(a) of Proposition 6.3.1. We

6.5. ORTHOGONALITY RELATIONS FOR

343

therefore come to the conclusion that

1 fG p, xq fG p, xq,

r x P S,ell , f P HpGq.

This is the identity to which we had reduced Theorem 2.4.1. We have now established it simultaneously for each of the three critical cases (6.4.1), (4.5.11) and (4.5.12). Therefore Theorem 2.4.1 holds for each of the three cases, and hence for any generic parameter. Corollary 6.4.5. The canonical self-intertwining operator attached any 0 r r P pGq, w P W and M P M satises (6.4.16) RP pw, M , q 1. r Proof. As we noted near the end of 2.5, the extension M of M r is canonical in this case, and so therefore is the intertwining operator in (6.4.16). To establish (6.4.16), it suces to consider the two special cases represented by the local analogues of (4.5.11) and (4.5.12). We shall sketch a proof for these cases that combines what we have just established with another global observation. From the local objects G, , M and M over F , we obtain global objects 9 9 9 9 9 G, , M and M over F . It is not hard to see from the global identity (4.5.14) (which holds for both (4.5.11) and (4.5.12)) and the condition (iii)(a) of Proposition 6.3.1 that RP pw, M , q epwq 1, r for a complex number epwq. The right hand side of the local intertwining relation (2.4.7) then satises fG p, uq fG p, wuq epwq fG p, uq, since u and wu have the same image in S . It follows that epwq 1. 6.5. Orthogonality relations for We are working towards the local classication of tempered representations. We will continue to use global methods, and in particular, the constructions of global data in 6.2 and 6.3. However, the global methods will have to be supplemented with some new local information. In this section, we shall stabilize the orthogonality relations given by the local trace formula, and as well as their twisted analogues. In particular, we shall use the local intertwining relation we have now established to rene the general results of [A11, 13]. We suppose henceforth that our local eld F is nonarchimedean. We are assuming inductively that the local theorems are valid for generic parameters of degree less than our xed positive integer N . The theorems now have the interpretation of 6.1. However, we recall that this entails essentially no change in their application. Let us be a little more explicit about the last point. Suppose that lies r r in the complement of sim pN q and pN q. We can then use our induction

344

6. THE LOCAL CLASSIFICATION

hypothesis to dene a substitute of the local Langlands group LF , exactly as in the global case treated in 1.4. It is a complex reductive group L F over F , which comes with a GLpN, Cq-orbit of L-embeddings into L GLpN q. r r r If G belongs to Esim pN q, lies in the subset pGq of pN q if and only if one of these embeddings factors through the L-image of L G in L GLpN q. The local centralizer sets S , S and S are then given as usual in terms of the p image of L in L G. We can also form the local set pG, q of G-orbits of L G that map to . This set has order L-homomorphisms from L to mpq |pG, q| equal to 1 or 2, like its global analogue. r Our concern in this section will be the subset ell pGq of elliptic paramr r eters in pGq. These are the elements P pGq such that the subset S ,reg,ell ts P S ,ss : |S ,s | 8u of S is nonempty. We have taken the liberty here of writing S ,reg,ell S ,reg X S ,ell , where S ,reg ts P S ,ss : S 0 is a torusu ,s conforms to the usual denition of regular semisimple. It is then clear that the two characterizations of S ,reg,ell are equivalent. It is also easy to see that if S ,reg,ell is nonempty, it actually equals S ,ell . Observe that is elliptic if and only if there is an elliptic pair r pG1 , 1 q, G1 P Eell pGq, 1 P 2 pG1 q, r such that the image of 1 in pGq equals . In fact, it follows from the denitions that our basic correspondence pG1 , 1 q p, sq denes a bijection between local, generic, elliptic variants Xell pGq tpG1 , 1 q : G1 P Eell pGq, 1 P F2 pG1 qu and Yell pGq tp, sq : P Fell pGq, s P S ,ell u of the general sets introduced in 4.8. This descends to a bijection between their quotients r Xell pGq tpG1 , 1 q : G1 P Eell pGq, 1 P 2 pG1 qu and r Yell pGq tp, xq : P ell pGq, x P S,ell u. Our primary goal is to describe the linear form r (6.5.1) f f 1 p1 q, f P HpGq, pG1 , 1 q P Xell pGq,

6.5. ORTHOGONALITY RELATIONS FOR

345

in terms of p, sq. This is the content of Theorem 2.2.1(b), the central assertion of the local theorems. The local intertwining relation provides a r r solution of the problem when lies in the complement of 2 pGq in ell pGq, as we saw in Proposition 2.4.3. Part of the remaining problem, in which r lies in 2 pGq, will be to complete the inductive denition of f 1 p1 q by establishing Theorem 2.2.1(a). This includes Theorem 6.1.1, which will be r r needed to rene our temporary denition of the subset sim pGq of 2 pGq r r and complete the inductive denition of the sets 2 pGq and ell pGq. We will deal with these matters in 6.6 and 6.7. For the orthogonality relations of this section, it is best to work with the set Tell pGq that was part of the proof of Proposition 3.5.1. Recall that Ttemp pGq is the set of GpF q-orbits of triplets pM, , rq, where M is a Levi subgroup of G, is a unitary, square-integrable representation of M pF q, and r is an element in the R-group Rpq of . The subset Tell pGq consists of (orbits of) triplets such that r belongs to the subset Rreg pq of elements in Rpq for which the determinant dp q detpr 1qaM is nonzero. Extending the convention we are using in this volume (in contrast r r r to that of [A10]), we write Ttemp pGq and Tell pGq for the sets of OutN pGqorbits of elements in Ttemp pGq and Tell pGq respectively. We recall that there is a linear form fG p q attached to any pM, , rq. For general groups G, fG p q is determined only up to a scalar multiple, which depends on various choices related to the normalization of intertwinr ing operators. Having made such choices for G P Esim pN q in Chapter 2, we may now be more precise, even though this is not really necessary for orthogonality relations. The case that M G is not at issue, since fG p q is then just equal to fG pq. If M is proper in G, we apply our induction r hypothesis to the tempered representation P 2 pM q. We obtain a unique r bounded, generic parameter M P 2 pM q, whose packet contains the reprer sentation M , and whose image in pGq we denote by . We can then set ` r r (6.5.2) fG p q tr RP pr, M , q IP pM , f q , f P HpGq, in the notation (2.4.2) (but with there written as M here). We note that the R-group RpM q of M dened in [A10] is a priori dierent from the R-group R of dened in 2.4. We have of course agreed to identify the stabilizer W of M with a subgroup of W pM q, the Weyl group that contains the stabilizer W pM q of M . It is a consequence of the disjointness of tempered L-packets for M that W contains W pM q. On the other hand, it follows from the form (2.3.4) of M , the corresponding form of the Weyl group W pM q, and our induction assumption that Theorem 2.2.4(b) holds for the factor G of M , that any element in W stabilizes M .

346

6. THE LOCAL CLASSIFICATION

Therefore W equals W pM q, in the generic case at hand. Moreover, we 0 now know that elements in the subgroup W of W give scalar intertwining 0 operators for the induced representation IP pM q. It follows that W is contained in the subgroup W 0 pM q of W pM q. We thus have a surjective mapping (6.5.3)
0 R W {W W pM q{W 0 pM q RpM q,

r M P M ,

r which makes sense for any parameter P pGq. At the beginning of the next section, we shall observe that this mapping is an isomorphism, and hence that RpM q equals R . The essential part of the study of (6.5.1) turns out to be the special case r r in which f lies in the subspace Hcusp pGq of cuspidal functions in HpGq. It follows from [Ka, Theorem A] that the mapping that sends fG P Icusp pGq to the function fG p q, P Tell pGq, on Tell pGq is a linear isomorphism from Icusp pGq onto the space of complex valued functions of nite support on Tell pGq. (This result is also a special case of the orthogonality relations in [A10, 6].) A similar assertion applies r r to the space Icusp pGq of OutpGq-invariant functions in Icusp pGq, if we replace r Tell pGq by Tell pGq. Suppose that pG1 , 1 q and p, sq are corresponding pairs in the elliptic sets Xell pGq and Yell pGq. Observe that G1 equals G if and only r r if s equals 1, in which case maps to the subset 2 pGq of ell pGq. When this is so, we suppose that is such that the linear form f G pq is dened, according to the prescription of Theorem 2.2.1(a), since this is the case not covered by our induction hypothesis. In all cases, we then have an expansion r (6.5.4) f 1 p1 q c,x p q fG p q, f P Hcusp pGq,
r PTell pGq

for uniquely determined complex numbers c,x p q that depend only on the image x of s in S,ell . The expansion (6.5.4) was a central object of study in [A11], but with the dierent notation p1 , q c,x p q. The article [A11] was aimed at general groups, in which the family of functionals f G f G pq, P 2 pGq, could be dened only as an abstract basis of the space of stable linear forms on Icusp pGq, rather than by local Langlands parameters. In particular, there were no groups S in [A11] associated to the linear forms , and hence no pairs p, sq. This accounts for the earlier notation, and the cruder results of that paper. The local intertwining relation gives a description of the coecients in r r (6.5.4) if lies in the complement of 2 pGq in ell pGq. In this case, and

6.5. ORTHOGONALITY RELATIONS FOR

347

S are given by (6.4.1). We can then take M 1 q r r to be the parameter in 2 pM, q that was part of the discussion of the last section. Since F is nonarchimedean, our induction hypothesis gives a perfect r pairing x , y on SM M . We form the subset ( r r T,ell pGq pM, M , rq : M P M , r P Rreg pM q r of Tell pGq attached to , a set that could a priori be empty. The coecients in (6.5.4) can then be expressed in terms of the pairing xx, y xx, M y, r pM, M , rreg q,

r on S,ell T,ell , where as in 2.4, x , M y is an extension of the character r x , M y on SM to the SM -torsor S,ell . To see this, we consider the local r intertwining relation (2.4.7), as f varies over the space Hcusp pGq. After expanding the two sides of (2.4.7), according to the denitions (2.4.6), (6.5.4), (2.4.5) and (6.5.2), we need only compare the coecients of fG p q. We see that # r xx, y, if P T,ell pGq, (6.5.5) c,x p q 0, otherwise, r r for any x P S,ell and P Tell pGq, and for the given parameter P 2 pGq. ell We note that the two sides of (6.5.5) have a parallel dependence on the r extension M of the component M of , a choice we have not built into our notation for . We turn now to orthogonality relations. The starting point is the canonical Hermitian inner product p Ipf, gq IpfG , gG q fG pq gG pq d
ell pGq

r on Icusp pGq. Recall that ell pGq denotes the set of elliptic conjugacy classes in GpF q, and d is a canonical measure that is supported on a set of strongly regularly classes in ell pGq. According to the local trace formula, this inner product has a spectral expansion (6.5.6) Ipf, gq mp q |dp q|1 |Rp q|1 fG p q gG p q,
r PTell pGq

where mp q denotes the number of elements in Tell pGq that map to , as an r equivalence class tpM, , rqu in Tell pGq, and Rp q is the R-group Rpq. The formula (6.5.6) is a specialization of [A10, Corollary 3.2], in which we have written gG p q pgqG p _ q.

348

6. THE LOCAL CLASSIFICATION

The R-groups are all abelian here, and the 2-cocycles introduced in [A10, p. 86] all split. The supplementary coecient mp q occurs in (6.5.6) because r we have summed over Tell pGq instead of Tell pGq. r A similar identity holds for the twisted component GpN q. There is a canonical twisted Hermitian inner product r, g q IpfN , gN q r p r r r Ipf r fN pq gN pq d
r ell pN q

` r rr r r on the space Icusp pN q Icusp GpN q , where fN fGpN q is the usual r r r image of a given function f P Hcusp pN q in the space Icusp pN q, while ` r ell pN q ell GpN q is the set of elliptic G0 pN, F q-orbits in GpN, F q, and r r r r d is a canonical measure on ell pN q that is supported on the strongly regr ular classes in ell pN q. This inner product also has`a spectral expansion. It r r r is a sum of terms over the analogue Tell pN q Tell GpN q for GpN q of the rell pGq. By the special properties of GpN q0 GLpN q, Tell pN q is our set r r set T ` r r ell pN q ell GpN q of multiplicity free elements 1 r r in pN q. The twisted spectral expansion is then rr r r (6.5.7) Ipf , g q |dpq|1 fN pq gN pq,
r Pell pN q

where |dpq| 2r . Twisted orthogonality relations for GLpN q, which include the cases treated in [W4], are a part of work in progress by Waldspurger r on the twisted local trace formula for GpN q. (See also [Li].) Like its global r q in place of G), the discrete part of the twisted analogue (3.1.1) (with GpN local trace formula contains a sum over elements w P W pM qreg . In the case r at hand, M `GpN q0 is the Levi subgroup such that belongs to the set r r M pN q M GpN q , and the sum over W pM qreg can be restricted to the elements wreg that act on the space aM Rr , and for which | detpwreg 1qaM | |dpq| 2r . r For any G P Esim pN q, there is also a stable Hermitian inner product p Spf, gq Spf G , g G q f G pq g G pq d
reg pGq

r on the space Scusp pGq. We recall that ell pGq denotes the set of elliptic, stable conjugacy classes in GpF q, and d is a canonical measure supported on the strongly regular classes in ell pGq. The denition obviously extends r to products of groups, such as the compound elements in Eell pN q and Eell pGq. rr It can be used to stabilize the inner products Ipf, gq an Ipf , g q. One obtains

6.5. ORTHOGONALITY RELATIONS FOR

349

two expansions (6.5.8) and (6.5.9) rr Ipf , g q

r GPEell pN q

Ipf, gq

G1 PEell pGq

p pG, G1 q Spf 1 , g 1 q,

r f, g P Hcusp pGq,

p r r rpN, Gq Spf G , g G q,

rr r f , g P Hcusp pN q,

for local coecients pG, G1 q |ZpG1 q |1 |OutG pG1 q|1 and ` p r r rpN, Gq GpN q, G 1 |ZpGq |1 |OutN pGq|1 2 analogous to their global counterparts described at the end of 3.2. The linear mappings r r (6.5.10) Icusp pGq Scusp pG1 q
G1 PEell pGq

and (6.5.11) given by fG

G1

r Icusp pN q

r GPEell pN q

r Scusp pGq,

rG r f are thus isometric isomorphisms, f 1 and fN

G

relative to the natural inner products on the right hand spaces. For the ordinary expansiom (6.5.8), we refer the reader to [A11, 2], which is the restriction to cuspidal functions of the general stabilization of [A16, 10]. The treatment of the twisted expansion (6.5.9) will be similar. We need to derive an explicit spectral expansion for Spf, gq. We shall use the spectral expansion (6.5.7) of the left hand side of (6.5.9) to deduce information about the right hand side. However, in order to accommodate the limited understanding we have at this point, we need to adopt some temporary terminology. r We shall say an element P ell pN q is a cuspidal lift if there is a r G P Eell pN q such that (i) r r fN pq f G pq, r r f P Hcusp pN q,

r r where f G pq is the linear form on Scusp pGq dened by (6.1.1), and (ii) r P 2 pGq.

It is a consequence of (6.1.1) that G is uniquely determined by (i). If r r satises both (i) and (ii), and lies in the subset sim pN q of ell pN q, then r r G lies in the subset Esim pN q of Eell pN q. This is a formal consequence of r r our temporary denition of the subset sim pGq of 2 pGq. Conversely, if

350

6. THE LOCAL CLASSIFICATION

r r P sim pN q satises (i), and the corresponding datum G lies in Esim pN q, (ii) itself follows from the denition. r Theorem 2.2.1(a) asserts that every P ell pN q is a cuspidal lift, but we have yet to establish this. If is not simple, there is a unique G with r P 2 pGq by our induction assumption, but it remains to establish the rst condition (i). This amounts to the secondary assertion (2.2.4) of Theorem 2.2.1(a). In any case, we can certainly assume inductively that for each r N 1 N , every element 1 P ell pN 1 q is a cuspidal lift. Suppose that we are given a subset r c pN q ell pN q ell that consists entirely of cuspidal lifts. Then c pN q is a disjoint union over ell r G P Eell pN q of the subsets r r r c pGq c pN q X 2 pGq. 2 ell As in the untwisted case, one can show that the mapping r r fN fN pq, r r r P ell pN q, fN P Icusp pN q,

r is a linear isomorphism from Icusp pN q onto the space of complex valued funcr r rc tions of nite support on the set ell pN q Tell pN q. Let Icusp pN q denote the r r subspace of elements in Icusp pN q that as functions on ell pN q are supported r r c pN q. Then for any G P Eell pN q, the mapping (6.5.11) sends on the subset ell rc r rc Icusp pN q onto a subspace of Scusp pGq, which we shall denote by Scusp pGq. rc We note that Scusp pGq can be identied with the space of functions of nite r support on c pGq. 2 r Proposition 6.5.1. Suppose that G P Eell pN q, and that f and g are rcusp pGq such that f G and g G lie in S c pGq. Then r functions in H cusp (6.5.12) Spf, gq mpq |S |1 f G pq g G pq.
r Pc pGq 2

Moreover, the formula (6.5.8) can be written as (6.5.13) Ipf, gq mpq |S |1

r Pell pGq xPS,ell

f 1 p1 q g 1 p1 q,

where in the last summand, pG1 , 1 q corresponds to p, xq. r r r r r Proof. Let f and g be functions in Hcusp pN q such that fN and gN G and g G under the mapping (6.5.11). Then the only are the preimages of f nonvanishing term on the right hand side of (6.5.9) is that of the given group G. Using (6.5.7) to rewrite the left hand side of (6.5.9), we obtain r r |dpq|1 fN pq gN pq rpN, Gq Spf, gq.
r Pc pN q ell

6.5. ORTHOGONALITY RELATIONS FOR

351

By assumption, each in the sum on the left is the pullback of a linear form r on Scusp pGq, which we are also denoting by . That is r r fN pq gN pq f G pq g G pq. From the explicit formula for rpN, Gq above, it is easy to check that |dpq|1 rpN, Gq1 mpq |S |1 , r for any P c pGq. The identity (6.5.12) follows. Observe that we can 2 r r enlarge the index of summation in this formula from c pGq to 2 pGq, since 2 c r r elements in the subspace Scusp pGq of Scusp pGq can be regarded as functions r r on 2 pGq that vanish on the complement of c pGq. In fact, we are free to 2 r 2 pGq by the subset 2 pGq of G-orbits of local parameters (rather p replace r than OutpGq-orbits), so long as we remove the coecient mpq. To establish the second assertion, we apply the formula we have just obtained to any G1 P Eell pGq. If G1 G, we can assume inductively that r every element in 2 pG1 q is a cuspidal lift. In other words, f 1 and g 1 are images rcusp pG1 q that satisfy the condition of the proposition, with of functions in H r r c pG1 q 2 pG1 q. We can therefore write (6.5.8) as 2 Ipf, gq pG, G1 q mp1 q |S1 |1 f 1 p1 q g 1 p1 q.
G1 PEell pGq r 1 P2 pG1 q

The process of converting this to the formula (6.5.13) is an elementary local analogue of the discussion of the global multiplicity formula in 4.7, itself a very simple case of the standard model of 4.34.4. Indeed, the last double sum is over pairs pG1 , 1 q in the local set Xell pGq dened at the beginning of the section. We have only to rewrite it in terms of the bijective image p, xq of pG1 , 1 q in the second local set Yell pGq. From the explicit formula for pG, G1 q, it is easy to see that mpq |S |1 pG, G1 q mp1 q |S1 |1 . The identity (4.5.13) follows. We shall actually use (6.5.13) in its interpretation as a set of orthogonality relations for the matrix coecients tc,x p qu in (6.5.4). We of course have to require that the linear form (6.5.4) attached to a given pair p, xq in Yell pGq be dened, which is to say that for the corresponding pair pG1 , 1 q r in Xell pGq, the element 1 P 2 pG1 q is a cuspidal lift. By induction, this is r only at issue if x 1, which in turn implies that lies in the subset 2 pGq r of ell pGq. To describe the orthogonality relations, we write ap, xq mpq1 |S | and bp q mp q1 |dp q| |Rp q|, r P Tell pGq.

352

6. THE LOCAL CLASSIFICATION

Corollary 6.5.2. Suppose that yi pi , xi q, i 1, 2, are two pairs in Yell pGq such that the associated, linear forms (6.5.4) are dened. Then # ap1 , x1 q, if Y1 Y2 , (6.5.14) bp q c1 ,x1 p q c2 ,x2 p q 0, otherwise. r
PTell pGq

r Proof. We identify Icusp pGq with the space of nitely supported functions of t u, and we identify the right hand side of (6.5.10) with the space of nitely supported functions of an index set tyu that includes all pairs p, xq P Yell pGq for which (6.5.4) is dened. Then the isomorphism (6.5.10) is given by a matrix C tcpy, qu such that cpy, q c,x p q if y p, xq. r We can assume that the basis of SpGq dened by the appropriate subset of tyu is orthogonal. We can then write the two expressions (6.5.6) and (6.5.13) for Ipf, gq as an orthogonality relation apyq1 cpy, 1 q cpy, 2 q bp1 q1 1 ,2
y

for the matrix C, where A tapyqu and B tbp qu are the invertible diagonal matrices that represent the two inner products. The matrix form of this identity is C A1 C B 1 , which becomes CBC A upon inversion. The required formula follows. As a preliminary application of the orthogonality relations of the corollary, we shall show that the expansion (6.5.4) simplies in the case not r covered by (6.5.5), namely that P 2 pGq. Lemma 6.5.3. (a) Suppose that p, xq is a pair in Yell pGq such that r r lies in the subset 2 pGq of ell pGq, and such that the associated linear form (6.5.4) is dened. Then r (6.5.15) f 1 p1 q c,x pq fG pq, f P Hcusp pGq.
r P2 pGq

(b) Suppose that yi pi , xi q, i 1, 2, are two pairs in Yell pGq that both satisfy the conditions of (a). Then # mp1 q1 |S1 |, if y1 y2 , 1 (6.5.16) mpq c1 ,x1 pq c2 ,x2 pq 0, otherwise. r
P2 pGq

6.6. LOCAL PACKETS FOR COMPOSITE

353

Proof. The rst part (a) asserts that the denition (6.5.4) simplies r in the case that belongs to 2 pGq. We have to show c,x p1 q 0 for any r r xed element 1 in the complement of 2 pGq in Tell pGq. r ,ell , for a unique parameter 1 in We know that 1 lies in a packet T 1 r r the complement of 2 pGq in ell pGq. We shall apply the formula (6.5.5) for the coecients r c ,x p q, x1 P S ,ell , P Tell pGq,
1 1 1

or rather the inversion |S1 ,ell |

1

x1 PS1 ,ell

# 1, if 1 , xx1 , 1 y c1 ,x1 p q 0, otherwise,

of this formula that is a consequence of the denition of the pairing x , y. It allows us to write bp1 q c,x p1 q bp q c,x p q |S1 ,ell |1
r PTell pGq

x1 PS1 ,ell

xx1 , 1 y c1 ,x1 p q

|S1 ,ell |1

x1

xx1 , 1 y

bp q c,x p q c1 ,x1 p q.

Since 1 , Corollary 6.5.2 tells us that the last sum over vanishes. Since bp1 q 0, we then see that c,x p1 q vanishes, as required. The denition (6.5.4) thus reduces to the required formula (6.5.15) of (a). Having established (a), we substitute (6.5.15) into the formula (6.5.14) of Corollary 6.5.2. The required formula (6.5.16) of (b) follows. 6.6. Local packets for composite We are now ready to start proving the theorems that give the local classication. They consist of the ve local theorems we stated in Chapters 1 and 2, specialized to generic local parameters . We have already established the local intertwining relation Theorem 2.4.1 for pairs pG, q. r What remains is the classication of general L-packets described by Theorems 1.5.1 and 2.2.1. We will also have to deal with the supplementary r Theorems 2.2.4 and 2.4.4 for pairs pG, q. These will be treated afterwards in 6.8, as will the questions in 6.1 related to Theorem 6.1.1. We note again that in the archimedean case, the local classication has been established. With the exception of our formulation of Theorem 2.4.1, the theorems for archimedean parameters are included in the general results of Shelstad for real groups (and work in progress by Mezo for twisted real groups). We recall that our proof of Theorem 2.4.1 for generic parameters, completed by global means in 6.4, applies to all local elds F . We thus continue to assume that the local eld F is nonarchimedean, as in the last section. We also maintain our induction hypothesis that the

354

6. THE LOCAL CLASSIFICATION

r local theorems hold for generic parameters in of degree less than the xed r integer N . Suppose that G is a xed datum in Esim pN q. We are trying to classify the irreducible tempered representations of GpF q. In particular, we r r want to attach a packet of (OutN pGq-orbits of) irreducible representar bdd pGq, which satises the endoscopic character identity tions to any P (2.2.6) of Theorem 2.2.1. r Suppose for the moment that lies in the complement of 2 pGq in r r bdd pGq. Then is the image of a parameter M P 2 pM, q, for a proper Levi subgroup M of G. The assertion of Theorem 2.2.1(a) actually applies to r data G in the larger set Eell pN q. However, as we have noted before, the usual argument of descent on any such G reduces the statement to a corresponding assertion for M , which then follows from our induction hypothesis. For Theorem 2.2.1(b), we can appeal to Proposition 2.4.3, since we have established the local intertwining relation of Theorem 2.4.1 for generic parameters. We r obtain a packet of (orbits of) of irreducible tempered representations r of GpF q, and a pairing on S that satises the endoscopic character identity (2.2.6). Therefore, Theorem 2.2.1 holds for . So does Theorem 1.5.1(a), since apart from the assertion for unramied pairs pG, q that we veried in 6.1, it represents only a less precise form of Theorem 2.2.1. As for this remaining assertion, Corollary 2.5.2(iv) tells us that the relevant intertwining operators are trivial. The local intertwining relation then tells us that the associated unramied character x , y is trivial. However, since we are dealing with tempered representations, we have r more to establish. We must show that the packet also satises the supplementary conditions of Theorem 1.5.1(b). These are tied to the question on the relation between the endoscopic and representation theoretic R-groups R and RpM q, raised in our discussion of (6.5.3) from the last section. The question is now easy to resolve. We can identify (6.5.10) as a mapr ping from the space of nitely supported functions of P Tell pGq to the space of nitely supported functions of p, xq P Yell pGq. It follows from (6.5.4) that the mapping factors through the projection (6.5.3), or rather the restriction of (6.5.3) to the subset R,reg of R . But we also know that the mapping (6.5.10) is an isomorphism, and so in particular, is surjective. It follows that the projection (6.5.3) takes R,reg onto Rreg pM q, for any r M P M . This statement amounts to the assertion that the set T,ell is r not empty if belongs to ell pGq. It might appear to be insignicant, since r R,reg consists of only one element if lies in ell pGq, and is in fact empty otherwise. Nonetheless, its application to Levi subgroups L M gives us what we want. Indeed, any element in R belongs to the subset RL ,reg r attached to some L and some parameter L P ell pLq that maps to . It follows that the projection (6.5.3) is an isomorphism, and hence that RpM q r equals R for any M P M .

6.6. LOCAL PACKETS FOR COMPOSITE

355

In the proof of Proposition 3.5.1, we described the general classication of temp pGq by harmonic analysis. This characterizes temp pGq as the image of a bijection ` ( L P L, P 2,temp pLq, P Rpq , pL, , q t u,
G where the left hand side represents a W0 -orbit of triplets, and the right hand side is given by the irreducible constituent p b q of the character ` pr, f q tr RP pr, M , q IP pM , f q r

of Rpq HpGq. Our goal here is to establish the ner endoscopic classication of Theorem 1.5.1. In the proof of Proposition 2.4.3, we constructed r r the packet of over temp pGq. It is a priori a multiset, equipped with a p given by the irreducible constituents p b q of the character mapping to S ` px, f q fG p, xq xx, M y tr RP pwx , M , q IP pM , f q r r
r M P M

r of S HpGq. We have just seen that we can identify RpM q with the R1 group R S {S of . It follows that the two kinds of classication are compatible. In particular, the packet structure of the second imposes an endoscopic, character theoretic interpretation on the rst. The original classication, for its part, yields ner properties of the packets. This is the implication we need to exploit here. It tells us that the r r elements in are multiplicity free. Combined with the properties of M that are part of our induction hypothesis, it also conrms that the mapping r p r from to S is a bijection. Finally, it tells us that the complement of 2 pGq r in temp pGq is a disjoint union over local parameters in the complement of r 2 pGq in bdd pGq of the packets . These are the conditions of Theorem r r 1.5.1(b), for the p-adic case at hand. We have established Proposition 6.6.1. Theorems 1.5.1 and 2.2.1 hold for generic paramr r eters in the complement of 2 pGq in bdd pGq. Having dealt with the complementary set, we assume from this point on r that lies in 2 pGq. Then r r (6.6.1) 1 r , i P sim pGi q, Gi P Esim pNi q, for distinct simple parameters i . In this section, we shall treat the case that r r r 1, or equivalently, that lies in the complement sim pGq of sim pGq. 2 9 , G, q that 9 9 From Proposition 6.3.1 we obtain corresponding global objects pF satisfy the given list of conditions. In particular, 9 9 9 1 r r 9 is a global parameter which lies in 2 pGq. Moreover, the set ( 9 9 9 9 r r 9 (6.6.2) F Fp1 , . . . , r q 91 1 9r r : 9i 0

356

6. THE LOCAL CLASSIFICATION

is a family of global parameters that satises the conditions of Assumption 5.4.1. We can therefore apply the relevant lemmas of Chapter 5 to the pair 9 9 pG, q. In order to extract local information for pG, q from the global properties 9 9 of pG, q, we need to be able to remove valuations v R S8 puq. The role of the following lemma will be similar to that of Lemma 6.4.1. Lemma 6.6.2. Theorem 2.2.1 holds for if r 3 and Ni 1, 1 i r. r Proof. We are assuming that belongs to sim pN q. It is therefore a 2 direct sum N inequivalent characters of order 1 or 2, where 1 N r 3. p Therefore G is orthogonal, and G is simple. The secondary condition of Theorem 2.2.1(a) consequently does not apply here. The primary condition of Theorem 2.2.1(a) follows from the global property 9 9 r 9 r9 9 f N pq f G pq, 9 9 r r f P HpN q,

9 9 r r established in Lemma 5.4.2. For we can x the component f u of f away from 9 9 r 9 r r 9 u so that f N pq is a nonzero multiple of its component fN pq f u,N pu q r r at u. In other words, the linear form fN pq in f is the pullback of a stable r 9 linear form f G pq in f P HpGq. For the main assertion (b) of Theorem 2.2.1, we shall be content to treat p the case N 3. Then G SOp3, Cq and G Spp2q. We may as well also assume that the quadratic id`le class character e 9 9 9 1 2 3 9 9 equals 1, since we could otherwise replace each i by its product with 1 9 corresponds to a homorphism of W 9 to G. We shall p 9 . Then 9 F reduce the assertion to results of [LL], using the properties of Whittaker models implicit in Lemma 2.5.5. We do not know a priori that the stable distribution attached to in [LL] is the same as the linear form f G pq. We therefore set G f pq fG pq, f P HpGq,
P

where is the packet attached to the local parameter in [LL]. Since 9 9 is a local factor u of the global parameter , the global results of [LL] 9 G pq as a local factor of the -component S G 9 9 implicitly exhibit f 9 pf q of disc, 9 the stable trace formula for G. In particular, if f9 f9u f9u f f9u ,

6.6. LOCAL PACKETS FOR COMPOSITE 9

357

G G we can x f9u so that Sdisc, pf9q is a nonzero multiple of f pq. On the other 9 p hand, it follows easily from the fact that G SOp3, Cq that the formula (3.3.14) reduces to an identity 9 9 9 N r r9 9 pG Idisc, pf q rpN, Gq Sdisc, pf G q, 9 9

9 9 r r f P HpN q,

9 r 9 of which the left hand side is a nonzero multiple of f N pq. We choose the 9 9 r r r r variable component f f u of f at u so that it has the same image in SpGq 9 r as f . We can then x the complementary component f u so that it has the 9 r 9 r 9 same image in SpGu q as f9u , and so that f N pq is a nonzero multiple of f G pq. It follows that
G f pq e pq f G pq,

f P HpGq,

for a nonzero constant e pq. The last equation is the analogue of the identity (6.4.4) from the proof of Lemma 6.4.1. If charpF q 2, we can combine it with a corresponding analogue of Lemma 2.5.5, which is very simple in this case and will appear in [A27]. It follows that e pq 1. If charpF q 2, we can vary the global 9 eld F as in the proof of Lemma 6.4.1. It follows that e pq 1 in this case as well. We see therefore that
G f pq f G pq,

f P HpGq,

in all cases. The possible ambiguity between [LL] and Theorem 2.2.1(a) thus resolved, we conclude that the packet of [LL] satsies the required identity of Theorem 2.2.1(b). We shall now apply the general results from Chapter5. Recall that Lemma 5.4.2 arms the validity of Assumption 5.1.1 for the parameter 9 r 9 P F2 pGq. As in the special case established in the last lemma, we need to convert this global property to the local assertion of Theorem 2.2.1(a). To r do so, we temporarily allow G to be an element in the larger set Eell pN q. In other words, 9 9 pG, q pGu , u q is the local component of pair 9 9 pG, q, 9 9 9 r 9 r 9 G P E ell pN q, P F ell pGq. 2

This is compatible with our understanding that the basic objects in the con9 9 struction of 6.3 are really the simple parameters t1 , . . . , r u and t1 , . . . , r u 9 9 q. rather than the pairs pG, q and pG, r Lemma 6.6.3. The linear form on HpN q attached to the parameter r sim pGq is the pullback of a stable linear form P 2 f f G pq, r f P HpGq,

358

6. THE LOCAL CLASSIFICATION

r on HpGq. If G G1 G2 is not simple, and r 1 2 , i P 2 pGi q, i 1, 2, then

G G f G pq f1 p1 q f2 p2 q,
1 2

f f1 f2 .

Proof. This is the local form of Lemma 5.4.2, which we shall apply 9 9 9 to the global parameter . Recall that u , for a valuation u of F . If 9v will be v is an archimedean valuation, the analogue of the lemma for F assumed as part of the theory of twisted endoscopy in [Me]. Suppose that v u is not archimedean. If the corresponding centralizer S v is innite, 9 9v follows by induction. In view of condition the analogue of the lemma for F (ii) of Proposition 6.3.1, this leaves only the case that the integer N is less 9 than 4, and v is a sum of N -inequivalent characters of order at most 2. In p these cases, G must be orthogonal, and the analogue of the lemma follows 9 from Lemma 6.4.1. The assertions of the lemma for F Fu thus follow from the global assertion of Lemma 5.4.2, and their analogues for v u. The product formula from the lemma allows us to apply our induction r hypotheses to any datum G P Eell pN q that is composite. We can therefore 9 r 9 r assume henceforth that G P Esim pN q and G P E sim pN q, as before. Observe that the lemma includes the assertion is cuspidal lift, which will allow us to make use of the orthogonality relations of the last section. We are now ready to exploit the essential global properties of Chapter 5. Corollary 5.1.3, together with the application to Lemma 5.1.4 of its supplement Lemma 5.4.5 in case N is even and 1, tells us that the stable 9 9 9 multiplicity formula of Theorem 4.1.2 is valid for G and . We thus have (6.6.3)
9 9 9 9 G 9 9 Sdisc, pf9q |pG, q| |S |1 f G pq, 9 9

r 9 f9 P HpGq,

9 since G pq 1 in the generic case at hand. We shall apply this to the 9 -component 9 G 9 9 p1 (6.6.4) Idisc, pf9q pG, G1 q Sdisc, pf91 q 9 9
9 9 G1 PEell pGq

of the stabilized trace formula given by (3.3.15). More precisely, we apply 9 the corollary of Theorem 4.1.2 to the groups G1 that index the sum in (6.6.4). It thus follows from Corollary 4.1.3 that the right hand side of (6.6.4) equals 9 9 9 pG, G1 q |S1 |1 f 1 p1 q. 9
9 9 G1 PEell pGq 9 9 1 PpG1 ,q

9 This is the expression (4.7.7) (with ) from the beginning of the elementary proof of Lemma 4.7.1, which we recall was the conditional justication in 4.7 of the spectral multiplicity formula. The conditions from 4.7 included the local theorems, which of course are what we are now trying to

6.6. LOCAL PACKETS FOR COMPOSITE

359

establish (for generic parameters). However, the simple arguments from the proof of Lemma 4.7.1 are still available to us here. We shall revisit them very briey, in order to obtain a modied global formula that will help us prove the local results. Arguing as in the proof of Lemma 4.7.1, we convert the last expression 9 to the -analogue
9 9 pG1 ,1 q

p 9 9 9 9 |S1 |1 |ZpG1 q |1 |OutG pG1 , 1 q|1 f91 p1 q 9 9

p 9 9 9 of (4.7.8). The double sum is over the set of G-orbits of pairs pG1 , 1 q in the 9 9 9 of 4.8 such that 1 maps to . It can be replaced by a double family XpGq p 9 9 9 9 sum over the set of G-orbits of pairs pG , xG q in the associated family Y pGq 9 9 such that G maps to . The expression becomes 9 |SG |1 f91 p1 q, 9
9 9 pG ,xG q

since we can write p 9 9 9 |S1 |1 |ZpG1 q |1 |OutG pG1 , 1 q|1 |SG |1 , 9 9 as again in the proof of Lemma 4.7.1. This summand depends only on the 9 9 image x of xG in S . The right hand side of (6.6.4) therefore equals 9 9 9 9 |pG, q| |S |1 f91 p1 q, 9
xPS 9 9

9 9 9 9 9 9 where pG1 , 1 q maps to the pair p, xq, and where we recall that pG, q is 9 9 the preimage of in pGq, a set of order 1 or 2. 9 Consider next the formula (4.1.1) (with ) for the left hand side 9 of (6.6.4). We claim that cannot contribute to the discrete spectrum of 9 9 any proper Levi subgroup M of G. For as we have argued in the past, 9 9 r 9 our induction hypothesis on the factor G P EpN q of M prevents any 9 9 r 9 parameter P pN q that does contribute to the discrete spectrum of M 9 9 r from lying in the subset ell pN q that contains . Such assertions are of 9 r course implicit in the denition of the set pN q, which rests ultimately on the general classication of Theorems 1.3.2 and 1.3.3. There is consequently 9 no contribution from M to (4.1.1). The left hand side of (6.6.4) then equals ` G 9 9 G Idisc, pf9q tr Rdisc, pf q n pG q f9G pG q, 9 9 9 9 9 9
9 G

9 9 where G ranges over the set pGq of irreducible unitary representations of 9 9 GpAq, and n pG q are nonnegative integers. We are assuming that f9 lies in 9 9

360

6. THE LOCAL CLASSIFICATION

r 9 9 9 HpGq. This means that fG pG q depends only on the image of G in the 9 9 r Gq of orbits in pGq under the restricted direct product 9 9 set p r 9 r 9 OutN pGq OutN pGv q.
v

We can therefore write the left hand side of (6.6.4) as 9 9 n pq f9G pq, |pG, q| 9 9 9 9
9 r 9 PpGq

for a modied coecient (6.6.5) 9 9 n pq |pG, q|1 9 9

9 9 G PpG,q 9

n pG q 9 9

9 9 9 in which pG, q is the preimage of in pGq. 9 We have converted (6.6.4) to an identity 9 (6.6.6) f91 p1 q, n pq f9G pq |S |1 9 9 9 9 9
PpGq 9 r 9 xPS 9 9

r 9 f9 P HpGq,

9 9 9 9 where pG1 , 1 q maps to p, xq. We shall apply this identity when f9 equals a product f9 f98 f9u f98,u , relative to the decomposition 9 9 9 9 9 9 9 9 9 9 9 GpAq G8 Gu G8,u GpF8 q GpFu q GpA8,u q 9 9 of GpAq. 9 Suppose that v is a valuation of F in the complement of S8 puq. The r 9 is then dened, and satises the conditions of Theorem 2.2.1. packet v 9 r 9 For if S 9 is innite, v lies in the complement of 2 pGv q, and therefore
v

represents the case we have already established. If Sv is nite, Proposition 9 6.3.1 tells us that N 3, in which case the local classication follows from Lemma 6.6.2. We can therefore write ` 9 (6.6.7) pf98,u q1 p8,u q1 xx, 8,u ypf98,u q 9 p 8,u q, 9 9 9
G 8,u 9

9 where 8,u ranges over elements in the packet ! ) 8,u r 9 , x , v y 1 for almost all v 9 9 v : v P v 9
vRS8 puq

of 9 8,u and 9 9 xx, 8,u y

vRS8 puq

vRS8 puq

9 v ,

9 9 xxv , v y

6.6. LOCAL PACKETS FOR COMPOSITE

361

is the corresponding product of local pairings. If v is archimedean, we have 1 9 xxv , v y f9v,G pv q, 9 9 pf9v qp1 q 9 9 v
v Pv 9 r9

by the results of Shelstad. We can therefore write 9 9 9 9 xx, 8 y pf98 q 9 p8 q, (6.6.8) pf91 qp1 q
8 8 G 9 8

9 where 8 ranges over representations in the packet ) ! r9 9 9 9 v : v P v 8

vPS8

of 9 8 and xx, 8 y 9 9
vPS8

vPS8

9 v ,

xxv , v y 9 9

is the associated product of local pairings. 9 9 9 Consider the remaining valuation v u of F . Then Fu F , Gu G 9 u . At this point we assume that the function f9u lies in the subspace and r r Hcusp pGq of HpGq. We can then write 1 (6.6.9) f9u p1 q c,x pq pf9u,G qpq, 9 u
r P2 pGq

in the notation of Lemma 6.5.3. We shall use the global identity (6.6.6) to extract information about the coecients c,x pq. More precisely, we shall exploit the identity obtained by substituting the formulas (6.6.7), (6.6.8) and (6.6.9) for the three factors of the summand ` 9 9 9 9 f91 p1 q f91 p1 q f91 p1 q pf98,u q1 p8,u q1
8 8 u u

in (6.6.6). The right hand side of (6.6.6) becomes |S |1 xx, 8 y c,x pq xx, 8,u y f9G pq, 9 9 9 9 9 9 9
xPS 9

where the inner sum is over products 8 b b 8,u 9 9 9 9 r of representations 8 in the packet of 8 , in 2 pGq and 8,u in the packet 9 9 8,u , and x is the isomorphic image of x in S . Suppose that P S is 9 p 9 of 9 a character on the 2-group S S . It follows from the condition (iii)(a) 9 9 of Proposition 6.3.1 that there is a representation 8, in the packet of 8 9 such that 9 9 xx, 8, y pxq1 , x P S .

362

6. THE LOCAL CLASSIFICATION

For the places outside S8 puq, we take the representation 8,u p1q 9 v p1q
vRS8 puq

9 9 in the packet of 8,u such that for each v, the character x , v p1qy on Sv 9 98 and f98,u , the formula is 1. With an appropriate choice of the functions f (6.6.6) then reduces to an identity c,x pq pxq1 fG pq (6.6.10) n p, q fG pq |S |1
xPS

r r for any f P Hcusp pGq, where is summed over 2 pGq on each side, and ` 9 (6.6.11) n p, q n 8 pq b b 8,u p1q 9 9 r Since the characters of representations P 2 pGq are linearly independent rcusp pGq, we conclude that on H c,x pq pxq1 , (6.6.12) n p, q |S |1
xPS

r for any P 2 pGq. In particular, the number (6.6.11) does depend only on the local data p, , q at u, as the notation suggests. r Lemma 6.6.4. (a) For any P 2 pGq, n p, q is a nonnegative integer. (b) If mpq and mpq are the orders of the preimages of and in 2 pGq and 2 pGq respectively, the product n p, q mpq1 mpq n p, q r is also a nonnegative integer. Proof. By denition, n p, q is the number n pq given by (6.6.5), 9 9 with 9 9 9 8, b b 8,u p1q. The summands n pG q on the right hand side of (6.6.5) are nonnegative 9 9 integers. To prove (a), we need to show that the sum itself is divisible by 9 9 p the positive integer |pG, q|. We can therefore assume that G SOpN, Cq, r 9 9 9 with N even, since the group OutN pGq under which pG, q is an orbit would r 9 otherwise be trivial. Then OutN pGq is a group of order 2, the nontrivial element being represented by a point in OpN, Cq with determinant equal to 9 9 p1q. We can also assume that the rank Ni of each component i of is 9 r 9 even, since the stabilizer of would otherwise be the group OutN pGq itself. r9 With this condition, we consider the archimedean component v P v of 9 9 9 at any v P S8 . The irreducible components of v are distinct and of degree 2, by the condition (iii)(a) of Proposition 6.3.1. It follows easily that r 9 9 9 9 as an OutN pGq-orbit of irreducible representations of GpFv q, v contains r N pGq acts freely on the indices of 9 9 two elements v,G . This implies that Out

6.6. LOCAL PACKETS FOR COMPOSITE

363

summation G in (6.6.5). Since the original multiplicity n pG q is invariant 9 9 9 9 9 under the action of any F -automorphism of G, the sum in (6.6.5) is an even 9 9 integer, and is hence divisible by the order 2 of pG, q. This establishes (a). The proof of (b) is simpler. Since 9 9 9 mpq mpq |pG, q|, we need only show that the sum in (6.6.5) is in all cases divisible by the r number of elements G in the OutN pGq-orbit of . This number equals r 9 either 1 or 2. If it equals 2, OutN pGq acts freely on the indices of summation G in (6.6.5). It follows that the product 9 r n p, q mpq1 n pG q, 9 9
G 9

is also a nonnegative integer, as required. We can now establish the essential properties of composite parameters r for our group G P Esim pN q over the local p-adic eld F . r Proposition 6.6.5. (a) For every P sim pGq, there is a nite subset 2 r r of 2 pGq, together with a bijection x , y, r p from onto S , such that (6.6.13) f 1 p1 q xx, y fG pq,
r P

r P ,

r f P Hcusp pGq, x P S ,

where pG1 , 1 q is the pair (in Xell pGq) that maps to p, xq. r r r (b) As ranges over sim pGq, the subsets of 2 pGq are disjoint. 2 Proof. The key ingredient will be the orthogonality relations of Lemma 6.5.3(b). We shall combine them with the formula (6.6.12) for the nonnegative integer n p, q. p p Suppose that P S and 1 P S1 are characters of the nite groups 1 in sim pGq. Applying (6.6.12) to each r attached to two parameters and 2 1 1 p , q, we write of the integers n p, q and n r n p, q n1 p 1 , q
r P2 pGq

mpq1 mpq n p, q n1 p 1 , q
x,x1

mpq |S |1 |S1 |1

pxq 1 px1 q1

mpq1 c,x pq c1 ,x1 pq ,

where x and x1 are summed over S and S1 respectively. The orthogonality relation (6.5.16) then tells us that the last sum over vanishes unless p, xq

364

6. THE LOCAL CLASSIFICATION

equals p1 , x1 q, in which case it equals mpq1 |S |. The original sum over therefore vanishes unless 1 , in which case it equals pxq1 1 pxq. |S |1
xPS

This in turn vanishes unless 1 , in which case it equals 1. We have established an identity # 1, if p1 , 1 q p, q, r n p, q n1 p 1 , q (6.6.14) 0, otherwise. r
P2 pGq

Suppose rst that p1 , 1 q p, q. Then the nonnegative integers n p, q r and n p, q are either both zero or both nonzero. Since the sum over of r their product equals 1, we see that there is a unique element pq in 2 pGq such that # 1, if pq, n p, q n p, q r (6.6.15) 0, otherwise, r for any P 2 pGq. Then taking the case that p1 , 1 q p, q, we see from (6.6.14) that the mapping p, q pq is injective. In other words, if we dene r p tpq : P S u, r P sim pGq, 2

r r the subsets t u of 2 pGq are mutually disjoint. This is the assertion (b) of the proposition. To establish the remaining identity (6.6.13), we have only to invert (6.6.12). It follows from (6.6.15) that c,x pq equals pxq if pq for p some P S , and that c,x pq 0 otherwise. The required formula (6.6.13) then follows from (6.5.15), if we set xx, pqy pxq, p x P S , P S .

This completes the proof of the proposition. r Corollary 6.6.6. If belongs to , then mpq mpq. r Proof. The corollary follows from the equality (6.6.15) of n p, q with n p, q.

6.7. LOCAL PACKETS FOR SIMPLE

365

6.7. Local packets for simple We continue with the discussion of 6.6. In this section we shall complete the local classication of tempered representations. We are taking r G P sim pN q to be a simple datum over the local p-adic eld F , for the xed positive integer N . It remains to treat the simple generic parameters r P sim pGq. However, to exploit the global construction of 6.2, we will have to start with a representation instead of a parameter . r r We shall write sim pGq for the set of elements in 2 pGq that do not sim pGq of the disjoint sets we constructed r r lie in the union over P 2 r in the last section. Let us also write Isim pGq for the subspace of functions rcusp pGq such that fG p q 0 for every in the complement of sim pGq r fG P I r r r in Tell pGq, and Ssim pGq for the subspace of functions f G P Scusp pGq such G pq 0 for every in sim pGq. that f 2 r Lemma 6.7.1. The isomorphism (6.5.10) maps Isim pGq isomorphically rsim pGq. onto S r Proof. In the statement, Ssim pGq is to be understood as a subspace of 1 collections of functions tf 1 f G u in r Scusp pG1 q
G1 PEell pGq

such that f 1 0 for G1 G. We observe from (6.5.6) that the orthogor r nal complement of Isim pGq in Icusp pGq may be identied with the space of rell pGq that are supported on the complement of sim pGq. By r functions on T r (6.5.13), the orthogonal complement of Ssim pGq can be identied with the 1 u such that f G is supported on sim pGq. It follows r space of collections tf 2 easily from the formulas (6.5.5) and (6.6.13) that the mapping (6.5.10) sends the rst orthogonal complement onto the second. Since the mapping is an r r isometry, it does indeed send Isim pGq isomorphically onto Ssim pGq. r Remark. Suppose that f P Hcusp pGq is a pseudocoecient of a reprer sim pGq. This means that sentation P # 1, if , f,G p q 0, otherwise, r r for any P Tell pGq. Then f lies in Isim pGq, according to the denition above. It follows that 1 G1 f f 0, r for any datum G1 P Eell pGq distinct from G. We will now turn to the global arguments. We cannot use the discussion of 6.3, since it was based on the inductive application of Corollary 6.2.4. Corollary 6.2.4 is not available because it requires that the local theorems

366

6. THE LOCAL CLASSIFICATION

r be valid for parameters P pN q. We have therefore to go back to the original Lemma 6.2.2, and the arguments from its proof. r We x a representation P sim pGq. Having chosen this object, we apply 9 9 9 Lemmas 6.2.1 and 6.2.2. We obtain global objects pF , G, q that satisfy the 9 9 9 conditions of Lemma 6.2.2. In particular, pF, G, q equals pFu , Gu , u q, for a 9 9 place u of F . If v belongs to S8 , v belongs to the L-packet of a Langlands parameter v in general position. If v lies in the complement of S8 puq, v is a 9 spherical representation, corresponding to a spherical Langlands parameter v . The key property of is of course that it occurs in the automorphic 9 9 discrete spectrum of G. This fact was established in the rst half of the proof of Lemma 6.2.2. It then led naturally to the main point of the second half of the proof, the application of Corollary 3.4.3 to . This in turn allowed us 9 9 9 P pN q to , which we then showed was in r to attach a global parameter 9 9 r the subset of pN q of generic parameters. In Corollary 6.2.3, we saw that for any valuation v u, the Langlands 9 9 parameter of the localization v of is the localization v of . Our task 9 9 here will be to establish properties of the localization 9 u 9 r of at u. These will be consequences of the fact that lies in sim pGq, together with the results we have established in sections following Lemma 6.2.2. r We did see in proof of Lemma 6.2.2 that lies in ell pN q. (See the discussion following (6.2.8).) Moreover, it is a consequence of (6.2.9) (in the proof of Corollary 6.2.3) that there is a constant cpq 0 such that r f,N pq cpq, 9 G r r rG for any function f f u in Hcusp pN q with f f . It then follows r r that the linear form f fN pq on Hcusp pN q vanishes on the kernel of the r r transfer mapping from Hcusp pN q to Scusp pGq. This in turn implies that r r r r for any f P Hcusp pN q at all, the component f pq of G P Eell pN q in the is distinct from G. Therefore, satises decomposition (6.1.1) vanishes if G the condition (i) of the denition of a cuspidal lift in 6.5. We do not know a priori that is simple. Suppose however that it lies r r in the complement sim pN q of sim pN q. Then Lemma 6.6.3 implies that ell the conditions (i) and (ii) in the denition of a cuspidal lift are equivalent. r More precisely, if we apply this lemma to the unique datum G P Eell pN q r sim pG q (the existence of which is guaranteed by our induction with P s hypotheses), we see that G G. In other words, lies in the subset r r sim pGq of sim pN q. It then follows from the formula (6.6.13) of Proposition 2 ell 6.6.5 that G f pq f,G p q.
r P

6.7. LOCAL PACKETS FOR SIMPLE

367

r Since f is a pseudocoecient of , a representation in a set sim pGq that r is disjoint from , the sum on the right vanishes. On the other hand, since satises the condition (i), we have
G r f pq f,N pq cpq 0.

This is a contradiction. r r We have shown that lies in the subset sim pN q of ell pN q. In this case, the rst condition (i) for to be a cuspidal lift implies the second condition r r (ii). In other words, lies in the subset sim pGq of sim pN q by virtue of the fact that the denition in 6.1 is implied by the condition we have already r established. The local parameter P sim pN q is therefore a cuspidal lift. 9 Consider the global parameter . It represents a generic automorphic representation of GLpN q whose component at u lies in the discrete series 9 9 of GLpN, F q, since the corresponding component u of belongs to 9 r sim pN q. The automorphic representation is therefore cuspidal, and con9 r sequently lies in the set sim pN q of simple global parameters. This argument is familiar from the proof of Corollary 6.2.4, where we were working with 9 the benet of the stronger hypothesis. As in the earlier case, also satises the condition ` 9 9 cpq 9 cpq , 9 P A2 pGq, 9

of Theorem 1.4.1. It therefore meets the formal requirement for belonging 9 r 9 r to the subset sim pGq of sim pN q, according to the original denition of 1.4. In the case here, however, we will have to work for a moment with the provisional denition of 5.1. 9 9 r Having shown that lies in sim pN q, we are now free to form the associated family (6.7.1) 9 9 r r 9 F Fpq t u

of global parameters. To make use of the results of Chapter 5, we need to know that the conditions of Assumption 5.4.1 hold, with V again being the set S8 of archimedean valuations. We cannot appeal to Proposition 6.3.1, as we did for the earlier compound parameter (6.6.1). However, the properties 9 we need are already in the earlier Lemma 6.2.2. We know that v equals the original archimedean parameter v . Conditions (5.4.1(a)) and (5.4.1(c)) 9 then follow from the constraints we placed on v in the construction of in the rst half of the proof of Lemma 6.2.2. Since the condition (5.4.1(b)) is 9 r not relevant to this case, Assumption 5.4.1 is valid for F. 9 r We can therefore apply the results of Chapter 5 to the family F. We note that the local and global induction hypotheses of Chapter 5 are vacuous 9 here, since the basic simple parameters and have degree N . In particular, r 9 we have to rely on the provisional denition of 5.1 for the set Fsim pGq. This

368

6. THE LOCAL CLASSIFICATION

poses no diculty. For we established earlier that ` 9 9 S 9 pf q tr R 9 pf q ,

disc, disc,

r 9 for a function f9 P HpGq, chosen prior in (6.2.7) in the proof of Lemma 6.2.2, with the property that the right hand side is nonzero. Therefore, the left 9 r 9 hand side is not identically 0. Therefore belongs to F2 pGq, according to any of the three equivalent conditions of Corollary 5.4.7. Among the results of Chapter 5 we can use is Lemma 5.4.2, which arms that Assumption 5.1.1 is valid. Then, following the proof of Lemma 6.6.3, we can establish the local assertion that r r r r fN pq f G pq, f P HpN q, r for a stable linear form f G pq on HpGq. We note that we have already r r proved the weaker assertion for cuspidal functions f P Hcusp pN q, namely that is a cuspidal lift. We had to do so independently, in order to show 9 that the parameters and are simple. The main result from Chapter 5 is the stable multiplicity formula (6.7.2)
9 9 9 G 9 9 Sdisc, pf9q |pG, q| f9G pq, 9

r 9 f9 P HpGq,

9 for . It follows from Corollary 5.1.3, and the application to Lemma 5.1.4 of its supplement Lemma 5.4.6, in case N is even and 1. 9 We can now follow the arguments of 6.6, but with one essential dierence. In 6.6, we started with a local parameter . Through a correspondence 9 9 r based on our global construction, we then obtained a local packet of r 2 pGq. In this section, the correspondence has had to representations in go in the opposite direction. Starting with a representation in the subset r r sim pGq of 2 pGq, the global construction has led us through a correspondence 9 9 to a local parameter . Since has been xed from the beginning, we shall have to write r for the variable representation in 2 pGq that accompanies our attempt to reverse the correspondence. We note that the group S is now trivial. The sums over x P S that were a part of the discussion of the last section will therefore not occur here. In particular, the analogue of the formula (6.6.6) is just 9 r 9 9 9 (6.7.3) n 9 p q f9 9 p q f9pq, f9 P HpGq,
G r 9 9 PpGq

where 9 9 n p q |pG, q|1 9 9

9 9 G PpG, q 9

n pG q 9 9

6.7. LOCAL PACKETS FOR SIMPLE

369

as in (6.6.5). It follows directly from the stable multiplicity formula above. We know that is a cuspidal lift. We can therefore consider the expansion (6.5.15) of r f G pq, f P Hcusp pGq, in terms of elliptic values of discrete series characters fG p q, r r f P Hcusp pGq, P 2 pGq. r P 2 pGq.

We would like to describe the coecients c p q c,1 p q, Retracing the steps from 6.6, we arrive at a reduction (6.7.4) n p q c p q of the formula (6.6.12), for nonnegative numbers n p q n p1, q, as in (6.6.11). Thus n p q n p q 9 9 is dened by the analogue of (6.6.5) above, in terms of the global representation 9 8 p1q b b 8,u p1q 9 9 ` r 9 9 9 in GpAq , and the global multiplicities n 9 p q in the automorphic discrete
G

r P 2 pGq,

9 spectrum of G. The arguments 1 are of no signicance in c,1 p q and q; they are simply atavistic references to the character on the trivial n p1, group S . We observe that the global representation 9 9 9 9 8 b b 8,u attached to the original representation is of the required form. Since it 9 occurs in the discrete spectrum of G, the number n pq attached to is strictly positive. Finally, the analogue of Lemma 6.6.4 is valid for any . It tells us that both n p q and the product r n p q mp q1 mpq n p q are actually nonnegative integers. The next step is to apply the orthogonality relations (6.5.16), as at the beginning of the proof of Proposition 6.6.5. The properties of the coecients c p q in Lemma 6.5.3 were stated with the understanding that is r a cuspidal lift. Suppose that is an arbitrary parameter in sim pGq that is also a cuspidal lift. The corresponding expansion (6.5.15) supplies with its own set of coecients c p q. It then follows from (6.5.16) that the sum (6.7.5) mp q1 c p q c p q
r P2 pGq

370

6. THE LOCAL CLASSIFICATION

vanishes unless equals , in which case it equals mpq1 . Setting , and making the substitution (6.7.4), we obtain n p q n p q 1. r

It then follows from the fact that n p q and n p q are nonnegative integers r that # 1, if , (6.7.6) n p q n p q r 0, otherwise, r for any representation P 2 pGq. It follows from (6.7.4) and (6.7.6) that # 1, , (6.7.7) c p q 0, otherwise, r for any P 2 pGq. The vanishing assertion in the last formula leads in turn to a collapse in the sum (6.7.5). The formula for (6.7.5) reduces simply to the relation mpq1 c pq c pq 0, in case . It follows that # 1, if , (6.7.8) c pq 0, otherwise, r for any cuspidal lift P sim pGq. In particular, the parameter is uniquely determined by the representation . We see now that the indirect global process by which we associated the parameter to gives a canonical local construction. For it follows from (6.7.8) that the correspondence is a well dened mapping from r r sim pGq to sim pGq. We have next to show that this mapping is a bijection. We shall formulate the assertion in terms of the inverse mapping , as a complement to the correspondence of Proposition 6.6.5. r Proposition 6.7.2. (a) Every parameter in the subset sim pGq of r sim pN q is a cuspidal lift. (b) There is a canonical bijection , r r from sim pGq onto sim pGq such that (6.7.9) f G pq fG p q, r r f P Hcusp pGq, P sim pGq. r P sim pGq,

Proof. The relation (6.7.9) is simply the general expansion (6.5.15) of Lemma 6.5.3, together with the identity (6.7.7) for the coecients. To establish the proposition, however, we must prove the prior assertion that r r the original mapping is a bijection from sim pGq onto sim pGq. The

6.7. LOCAL PACKETS FOR SIMPLE

371

r condition (6.7.8) implies that the mapping is injective. Let us write c pGq sim r r for its image. Our task is then to show that c pGq equals sim pGq. The sim r problem is that we do not know that every element P sim pGq satises the rst condition (i) of a cuspidal lift. It is essentially the question of showing that the union (6.1.2) is disjoint. r r r r Consider the subspaces Isim pGq and Ssim pGq of Icusp pGq and Scusp pGq introduced at the beginning of the section. Lemma 6.7.1(a) asserts that the linear mapping r fG f G , f P Icusp pGq, r r takes Isim pGq isomorphically onto Ssim pGq. It is clear from the original denition that the invariant pseudocoecients f,G , r P sim pGq,
G f f ,

r r are a basis of the complex vector space Isim pGq. For any P sim pGq, set r where P c pGq is the image of . Then f can be regarded as a stable sim pseudocoecient of , in the sense that # 1, if , f p q 0, otherwise, r for any P c pGq. The stable pseudocoecients sim f , r P c pGq, sim

r are then a basis of the complex vector space Ssim pGq. We can augment this basis with objects from the last section. The r parameters in sim pGq are all cuspidal lifts. They also provide stable 2 pseudocoecients f , which can be expressed in terms of the corresponding packets by f |S |1 f ,
r P

r r and which span the orthogonal complement of Ssim pGq in Scusp pGq. The expanded family of pseudocoecients f , parametrized by the set rc r r c pGq sim pGq sim pGq, 2 2 r then forms a basis of Scusp pGq. r Suppose that is any element in sim pGq. By denition, is a par r r rameter in sim pN q such that the linear form f G p q on Scusp pGq dened by r c pGq that is distinct (6.1.1) is nonzero. Suppose that is any parameter in 2 r P c pGq, 2

372

6. THE LOCAL CLASSIFICATION

r r from . By Proposition 2.1.1, we can choose a function f P Hcusp pN q that is a preimage of , in the sense that # f , if G G, r pf q 0, otherwise, r r for any G P Eell pN q. Since is a cuspidal lift, we can regard f as a r twisted pseudocoecient of . In particular, pfN qp q vanishes, since and r represent distinct elements in ell pN q. It follows from (6.1.1) that r r pf qp q pf qG p q fN p q 0. On the other hand, there must be a such that f p q is nonzero, since r Scusp pGq is spanned by the functions f . There must therefore be a paramr r r eter P c pGq that equals . Since P sim pGq, must lie in c pGq. 2 sim c pGq. This gives the assertion r r We have shown that sim pGq equals sim r (a) of the proposition, since the elements in c pGq are cuspidal lifts. It sim r r also tells us that the original mapping from sim pGq to sim pGq is surjective, and hence a bijection. The inverse mapping therefore exists, as asserted in (b). The proof of the proposition is complete. Corollary 6.7.3. If equals , then mpq mpq. Proof. The corollary follows from the equality of n pq with n pq in r (6.7.6). r We dene the L-packet of any simple parameter P sim pGq to be the singleton r t u. r Propositions 6.6.5 and 6.7.2 together thus attach an L-packet to any r square integrable Langlands parameter P 2 pGq. However, there is still something more to be said about the corresponding character identities. We shall state it as a joint corollary of the two propositions. r Corollary 6.7.4. Suppose that is any parameter in 2 pGq. Then the character identity (6.7.10) f 1 p1 q xx, y fG pq, x P S ,
r P

established in Propositions 6.6.5 and 6.7.2 for a cuspidal function f , remains r valid if f is an arbitrary function in HpGq. Proof. At rst glance, the assertion might seem to be immediate. In the case that is not simple, for example, the global formula (6.6.6) from which we derived the cuspidal version (6.6.13) of the required character

6.7. LOCAL PACKETS FOR SIMPLE

373

r 9 identity is valid for any function f9 P HpGq. Likewise, the local transfer formula r r (6.7.11) fN pq f G pq r of Lemma 6.6.3 holds for any function f P HpGq. However, we need to be careful. For we do not know a priori that the right hand side of (6.7.11) equals the value at x 1 of the right hand side of (6.6.13), if the function f r belongs to the complement of Hcusp pGq. Nor can we expect an immediate rescue from (6.6.6). For we cannot a priori rule out the possible existence 9 9 of global representations on the left hand side for which u is not elliptic. The simplest way to establish the lemma is to appeal to the two general theorems of [A11], and their twisted analogue for GLpN q (which we discussed briey in the proof of Proposition 2.1.1). The proof of these theorems is not elementary, and has not yet been written down in the twisted case. However, we have already used them in the proof of Proposition 2.1.1. We may as well use the theorems again here, since it is exactly for this sort of problem that they were intended. In particular, they serve as a substitute for the fundamental lemma for the full spherical Hecke algebra [Hal]. The coecients on the right hand side of (6.7.10), regarded as a sum over r r the full set temp pGq whose summands are supported on the subset , are already determined from the case that f is cuspidal. This is because the characters of discrete series are determined by their values on the strongly r elliptic set. Suppose then that f is a general function in HpGq. Suppose G pq alternatively as the value at x 1 of the right also that we dene f hand side of (6.7.10), rather than the left hand side of (6.7.11). Theorem 6.1 of [A11] then asserts that the resulting linear form in f is stable. Theorem 6.2 of [A11] asserts in addition that (6.6.9), with the left hand side being the analogue for G1 of the linear form just dened, is valid for f and for any element x in S . Lastly, the twisted analogue for GLpN q of Theorem 6.2 of [A11] asserts that (6.7.11) holds for f . In other words, the alternative denition for f G pq just given matches the one we have always used. We conclude that the character identity (6.7.10) does hold for the general function f . We have established the general character identity of Theorem 2.2.1(b), r for parameters P 2 pGq. Since Lemma 6.6.3 takes care of its rst assertion (a), Theorem 2.2.1 itself is valid for any such parameter. We note that the character identity (6.7.10) also allows us to complete some of the global 9 r induction hypothesis for the family F. Applied to the multiplicity formulas 9 (6.6.8) and (6.7.3) we have deduced, it gives the contribution of to the discrete spectrum postulated in Theorem 1.5.2. Observe that Corollary 6.6.6 of Proposition 6.6.5 can be regarded as a proof of the assertion (b) of Theorem 2.2.4. Together with Corollary 6.7.3 of Proposition 6.7.2, it resolves an important induction hypothesis from Proposition 6.6.1. We are speaking of the hypothesis that for any parameter

374

6. THE LOCAL CLASSIFICATION

r r in the complement of 2 pGq in bdd pGq, the group W pM q equals W , in the notation of (6.5.3). This is what allowed us to conclude that the mapping (6.5.3) was surjective, which we recall was a starting point for Proposition 6.6.1. Proposition 6.6.1 itself tells us that Theorem 2.2.1 holds for any such . Theorem 2.2.1 is thus valid for any generic, bounded parameter r P bdd pGq. Recall that this includes the assertion (a) of Theorem 1.5.1. We shall state what remains of Theorem 1.5.1(b) as a joint corollary of the three Propositions 6.6.1, 6.6.5 and 6.7.2. r r Corollary 6.7.5. The set temp pGq of irreducible, tempered (OutN pGqr bdd pGq of orbits of ) representations of GpF q is a disjoint union over P r the packets . Proof. Proposition 6.6.1 includes the assertion that the complement of r 2 pGq in temp pGq is a disjoint union of packets , in which ranges over r r r r the complement of 2 pGq in bdd pGq. Propositions 6.6.5(b) and 6.7.2(b), r r combined with the denition of the set sim pGq, tell us that 2 pGq is the r r disjoint union over P 2 pGq of the remaining packets . The corollary follows. We have now proved both of the local Theorems 1.5.1 and 2.2.1 for r parameters P bdd pGq. This gives the local Langlands correspondence r (up to the action of the group OutN pGq) for any quasisplit orthogonal or r symplective group G P Esim pN q over the p-adic eld F . Actually, we should really say that the local classication will be complete once we have resolved the induction assumptions on which the arguments of the chapter have been based. We shall do so now. 6.8. Resolution We have established Theorems 1.5.1, 2.2.1 and 2.4.1 for generic local parameters . These are the theorems that characterize the tempered repr resentations of our classical group G P Esim pN q over the p-adic eld F . We are not quite done, however. As we have just noted, we have still to resolve the induction hypotheses on which the proof of the three theorems has been based. There are actually three kinds of hypotheses. First and foremost, we have the supplementary local Theorems 2.2.4 and 2.4.4 to prove. These theorems are of interest in their own right, since they give new information about characters and local harmonic analysis. But they were also forced on us for the statement of Theorem 2.4.1. We must prove them for the r relevant parameters in pN q in order to complete the local hypothesis of this chapter. Secondly, we must also complete the global induction hypothesis of the chapter. That is, we must convince ourselves that the global theorems 9 r hold for parameters in FpN q. Lastly, we need to resolve the temporary

6.8. RESOLUTION

375

r denition of the local set sim pGq in 6.1 in terms of the natural denition provided by local Langlands parameters. This will take care of Theorem 6.1.1, the last of the local assertions. Consider then the supplementary local Theorems 2.2.4 and 2.4.4. r They concern the case of a datum G P Esim pN q, where N is even and p SOpN, Cq. As we recall, the standard outer automorphism of G gives a G r bitorsor G under G. The theorems apply in this case to a generic parameter r r in the subset of pGq of pGq. The proofs are similar to those of Theorems 2.2.1 and 2.4.1. In particular, they rely on the global methods that originate with the construction of 6.3. However, the arguments now are considerably easier. For one thing, the parameter is never simple, so we have no need of the general discussion of 6.7. For another, we will be able to use a number of properties that were established ab initio in the earlier arguments. Suppose for a moment that
1 1

r r

p r r r is a general local parameter in pGq. Recall that the centralizer S in G of the image of is an S -torsor. It has a decomposition r (6.8.1) S Op i , C Spp i , Cq GLp j , Cq
` iPI pGq

iPI pGq

jPJ

that is parallel to (1.4.8). We have written p q here for the preimage of ` p1q under the sign character (noting that this minus sign is not related to the superscript in IG pGq). Similar remarks hold if the local pair pG, q 9 9 9 is replaced by a pair pG, q over a global eld F . In the global case, the 9 9 r denitions of 3.3 apply to the G-bitorsor G. In particular, we can form the 9 -component 9 9 9 9 9 G r r r r r Idisc, pf q Idisc, pf q, f P HpGq, 9 9 9 r of the discrete part of the twisted trace formula for G, following (3.3.12). It satises the analogue 9 9 9 9 r r r r r r p pG, G1 q S 1 pf 1 q (6.8.2) Idisc, pf q 9 disc, 9
9 9 G1 PEell pGq

9 of (3.3.15). We can also form the -component

9 9 9 9 G r r r Rdisc, pf q Rdisc, pf q, 9 9

9 9 r r f P HpGq,

9 G 9 9 of the canonical extension of the representation Rdisc, of GpAq. 9

376

6. THE LOCAL CLASSIFICATION

Returning to the focus of this chapter, we consider a generic local parameter r r 1 1 r r , i P sim pGi q, Gi P Esim pNi q, r in disc pGq. We use Proposition 6.3.1 to construct simple global parameters 9 r 9 i P sim pGi q, and the associated family 9 9 r r 9 F Fp1 , . . . , r q 9 9 of compound global parameters. If pG, q is the global pair attached to 9 9 r r r9 pG, q, then lies in pGq. The corresponding centralizers S and S are isomorphic. They are given by a product (6.8.1), but without the general linear factors, since the simple constituents i of are self dual. The other conditions of Proposition 6.3.1, which were used to validate the results of Chapter 5, will not be needed here. Having described the underlying global machinery, we can be brief in r our treatment of the two local theorems for G. We will be content simply to sketch a proof of the main cases, those of square integrable parameters r r P 2 pGq for Theorem 2.2.4, and elliptic parameters P ell pGq for Theorem 2.4.4. The various arguments of descent needed to deal then with the general cases are similar to those of the corresponding Theorems 2.2.1 and 2.4.1, and will be left to the reader. r We consider Theorem 2.2.4 rst, for a parameter in 2 pGq. In the r as we have GpN q, as an object that is part of the r case F R, we treat G work in progress by Shelstad and Mezo. We shall therefore assume that F is a p-adic eld, as in 6.6. We then have # 1 r , ` (6.8.3) r S Op1, Cqr q , r r for i P sim pGi q and Gi P Esim pNi q, with the requirement that some Ni be odd. The proof of Theorem 2.2.4 will be a straightforward application of the discussion of 6.6. 9 9 r r We rst apply (6.8.2) and Corollary 4.1.3 to a function f P HpGq. As in the remarks preceding (6.6.6), we obtain 9 9 p 9 9 r r r r r 9 r Idisc, pf q pG, G1 q S 1 pf 1 q disc, 9
9 G1

r9 9 9 |S |1 |pG, q|

9 r 9 f 1 p1 q,
r r xPS

9 9 9 9 r 9 r 9 r where pG1 , 1 q maps to the pair p, xq. The condition that lies in 2 pGq is equivalent to the property 9 9 |pG, q| 1.

6.8. RESOLUTION

377

The twisted analogue of (6.6.6) then becomes 9 9 9 r 9 r9 r f 1 p1 q, (6.8.4) n pq f p q |S |1 9 9 9

G 9 PpGq 9 9 r r x P S G 9 r where n pq is the multiplicity of in Rdisc, , and is the canonical exten9 9 9 9 9 r 9 9 sion (4.2.7) of to GpAq dened whenever n pq is nonzero. Recall that 9 9 in the proof of the local results in 6.6 and 6.7, we evaluated n pq. We 9 9 showed that it satises the formula of Theorem 1.5.2,with m being equal to 9 1, at least if the component 8,u of outside S8 puq equals 8,u p1q. With 9 9 9 9 8,u , n 9 pq vanishes unless the representation 8 lies in r 9 this condition on 9 9 9 , the representation u 9 the product 9 of the archimedean packets 8 v

9 lies in the packet u at F Fu , and the character 9 9 9 xx, 8 y xx, y, x P S ,

is trivial, in which case n pq 1. We have only to investigate the form 9 9 taken by (6.8.4) for a suitable product 9 9 9 r r r f f 8 f f 8,u , 9 r f f u.

Let be a xed character on S . By the condition (iii)(a) of Proposition 9 6.3.1, we can choose a representation 8, in 8 that maps to the character 9 1 in S . The archimedean form of Theorem 2.2.4, which we are taking 9 9 r r 9 for granted, implies that any extension 8, of 8 to G8 determines an 9 9 r r r r extension of to S . We choose the rst factor f 8 P HpG8 q so that # 9 9 1, if 8 8, , 9 9 r 9 r f p 8 q 8,G 0, otherwise, 9 for any representation 8 in the packet 8 . We choose the third factor 9 9 9 r r f 8,u P HpG8,u q so that # 1, if 8,u 8,u p1q, 9 9 98,u 9 8,u r r f 9 p q G 0, otherwise, for any representation 8,u in the packet 8,u . We recall that 8,u p1q is the 9 9 9 9 9 product over v R S8 puq of those representations v P v with x , v y 1. 9 9 9 r 9 r It therefore has a canonical extension 8,u p1q to GpA8,u q. At the remaining 9 r r place u, we take f f9u to be any function in HpGq HpGu q. This represents a minor departure from this stage of the discussion in 6.6, where f was chosen to be cuspidal.

378

6. THE LOCAL CLASSIFICATION

Consider rst the right hand side 9 9 ` 9 r 9 r r 9 r 9 f 1 p1 q |S |1 f 1 p1 q f 1 p1 q pf 8,u q1 p8,u q1 |S |1 8 8

x r x r

of (6.8.4). Ruling out the trivial case that N 2, we shall assume that N 4, since N is even. It follows from the condition (ii) of Proposition 9 9 r 6.3.1 that v does not lie in 2 pGv q for any v R S8 puq. It then follows from Theorem 2.4.4 (which we will discuss in a moment), and the appropriate descent argument that ` 9 9 9 9 9 r r 9 r r r xx, 8,u y f 8,u p 8,u q 1. pf 8,u q1 p8,u q1 9
8,u P8,u 9 9 G

We see also from the denitions that 9 9 9 9 r 9 r r r f 1 p1 q xx, 8 y f 8 8

8 P8 9 9

r 9 p q 8 8,G

rx prq1 .

The right hand side of (6.8.4) becomes rx r |S |1 prq1 f 1 p1 q,

r x

r r where pG1 , 1 q is the preimage of p, xq. To describe the other side, we write pq as in 6.6 for the representation of GpF q in the packet with r r x , y , and we take pq pq to be the extension of pq determined by the canonical extension (4.2.7) of 8, b pq b 8,u p1q, 9 9 9 9 the canonical extension of 8,u p1q, and the extension we have chosen for 9 8, . The left hand side of (6.8.4) becomes 9 9 r9 r n pq f p q 9 9
G 9 PpGq 9

9 r n pq f 9 9

9 9 9 r r r8,u r8,u q 9 p q f r pr q f 9 p 8 G 8,G G

` rr r fG pq .

The formula (6.8.4) therefore reduces to ` rr r rx r (6.8.5) fG pq |S |1 prq1 f 1 p1 q,

xPS r r

p for the given character P S . is any xed point in S , the product of pr q with either side of rx r r If x r (6.8.5) is independent of the extension of . We sum each of the two

6.8. RESOLUTION

379

r products over P S . Since # |S |

1

p PS

rx rx pr q prq

1, 0,

r r if x x , otherwise,

the right hand simplies. We conclude that ` r r f 1 p1 q xr, y fG pq . x r rr

p PS

r r The extension of to S is arbitrary. It determines the extension r r pq pq of pq to GpF q on the right hand side. The formula becomes r r r r x r rr xr, y fG prq, f 1 p1 q f P HpGq.
P

This is the assertion (a) of Theorem 2.2.4. We recall that the assertion (b) of the theorem was Corollary 6.6.6. We have therefore established Theorem r 2.2.4, the rst supplementary local theorem for parameters P 2 pGq. Consider next the assertion of Theorem 2.4.4, for a parameter in the r r complement of 2 pGq in ell pGq. The conditions on are $ & 21 2q q`1 r , (6.8.6) ` %S Op2, Cqq Op1, Cqrq , r q 1, with the requirement that there be a regular element w P W,reg in the r r Weyl set W for pG, q. This is the twisted analogue for G of the main case (6.4.1) treated in 6.4. There is no need to be concerned with the other two r exceptional cases (4.5.11) and (4.5.12). Their local analogues for G follow from Corollary 6.4.5, together with a natural descent argument. We are r also taking for granted the analogue for G of Lemma 2.4.2, which represents a reduction of M by descent. In other words, we shall assume as in 6.4 that M is minimal with respect to , in the sense that there is a parameter r M P 2 pM, q. Theorem 2.4.4 can be proved in the same way as Theorem 2.4.1. Given 9 9 9 the local objects G, , M and M , we again choose global objects G, , M 9 M according to Proposition 6.3.1. Then lies in the subset ell pGq of 9 r 9 and 9 9 r 9 r ell pGq over the global eld F . The rst step is to establish the G-analogue `9 9 9 9 9 r 9 r r9 9 r r r (6.8.7) f 1 9 p, xq fp, xq 0, f P HpGq,
G G r r xPS,ell

of the global identity (6.4.2). Recall that (6.4.2) was one of the assertions of the renement Lemma 5.4.3 of Lemma 5.2.1. The original lemma was given by a very simple case of the standard model, which applies equally well to 9 9 r G. Another consequence of Lemma 5.4.3 is that does not contribute to

380

6. THE LOCAL CLASSIFICATION

9 r the discrete spectrum of any datum G P E sim pN q. This implies that the 9 r analogue for G of the sum (5.2.7) in Lemma 5.2.1 vanishes. The formula 9 r (6.8.7) then follows from the analogue for G of the lemma itself. The next step is to remove the contributions to (6.8.7) of the valuations r v in the complement of S8 puq. This reduces to the analogue for G of Lemma 6.4.1, which amounts to the assertion of Theorem 2.4.4 in case Ni 1, 1 i r 3.

Since the Levi subgroup M of G can have no symplectic factors, we see from the rst observation in the proof of Lemma 6.4.1 that M is actually abelian. It is then easy to see that G is a quasisplit form of either SOp4q or SOp2q. We leave to the reader the exercise of establishing Theorem 2.4.4 in either r of these cases. Once we have the G-analogue of the lemma, we can use the descent argument following its proof to treat the contributions to (6.8.7) of 9 r valuations v R S8 puq. This reduces (6.8.7) to the G analogue 9 9 9 9 r 9 pv , xv q 0 r1 pv , xv q 9 r 9 r f f (6.8.8) 9
r r xPS,ell vPS8 puq v,G vPS8 puq v,G

9 9 r r of (6.4.6), for functions f v P HpGv q. The assertion of Theorem 2.4.4 for is the local identity (6.8.9) r1 fG p, xq fG p, xq, r r r r r r x P S,ell , f P HpGq. r r

As in the proof of Theorem 2.4.1, the main point is to treat the special case that F equals R, and is in general position. To do so, we need to persuade r ourselves that Lemmas 6.4.2 and 6.4.3 remain valid if G is replaced by G. For Lemma 6.4.2, we use the simply transitive action x x xM , r r x P S,ell , xM P SM , r r

r of the group SM on S,ell in place of its action on S,ell . The proof for r G then carries over directly to G. For Lemma 6.4.3, we again introduce 7 supplementary degrees Nj . The original degrees Ni of of course equal 1 or 2, since F is archimedean. If Ni 2 for each i q, we set # Nj , if i q, 7 Nj Nq`1 ` ` Nr , if i q ` 1, as in the original proof. If Ni 1 for some to be q ` 1, we set $ Nj , & 7 Nj 1, % Nq`1 ` ` Nr , index i q, which we can take if i q, if i q ` 1, if i q ` 2.

6.8. RESOLUTION

381

We leave the reader to verify that with this modication, the proof of Lemma r 6.4.3 for G carries over directly to G. Once we have established (6.8.9) for F R and in general position, we can apply the arguments Proposition 6.4.4. These arguments carry over r directly from G to G. They yield the formula (6.8.9) for any F and . This completes our discussion of Theorem 2.4.4, the remaining supplementary local theorem. 9 r The second question to resolve is the global induction hypothesis for F. The global assertions are Theorems 1.5.2, 1.5.3, 4.1.2 and 4.2.2 (as well as Theorems 1.4.1 and 1.4.2, taken as implicit foundations for Theorem 1.5.2). 9 9 r We must check that they are valid for parameters in the set FpN q. As 9 r we noted at the end of 6.3 and in 6.7, the family F satises Assumption 5.4.1. It consequently also satises Assumption 5.1.1, by Lemma 5.4.2. We 9 9 r can therefore apply any of the results of Chapter 5 to parameters P F. The central global assertion is the stable multiplicity formula of Theorem 4.1.2. The general global induction argument used in Proposition 4.5.1 can 9 r be restricted to the family F. It yields the stable multiplicity formula for 9 9 r parameters P FpN q that do not fall into the critical cases studied in Chapter 5. For these remaining cases, the formula was established in Lemmas 9 9 r 5.4.35.4.6. Theorem 4.1.2 therefore holds for any parameter P FpN q. Theorem 1.5.2 is of course the actual multiplicity formula. It describes 9 9 r representations in the discrete spectrum of a group G P E sim pN q. For the 9 r 9 packet attached to a parameter P F2 pGq, the formula follows from the local classication and the stable multiplicity formula, as we noted after Corollary 6.7.4, or as we can see from a direct appeal to the obvious variant 9 9 r r 9 of Lemma 6.7.1. If lies in the complement of F2 pGq in FpN q, Theorem 1.5.2 is a vanishing assertion. It follows from (4.4.12), (4.5.5) and the stable multiplicity formula. Theorem 1.5.3 has two parts. The assertion (a) follows from Lemma 5.4.6. The assertion (b) is not part of the induction argument for generic parameters, as we can see from the remarks at the beginning of 5.1. It need not concern us here. For Theorem 4.2.2, the essential assertion (b) applies 9 9 r to parameters P F2 pGq. It is a natural biproduct of the discussion of the local Theorem 2.2.4 above. Finally, Theorems 1.4.1 and 1.4.2 follow from r 9 our resolution of the denition of the global set Fsim pGq in Corollary 5.4.7. r The last question concerns the denition of the local sets sim pGq. We need to resolve the temporary denition in 6.1 in terms of the natural denition in terms of local Langlands parameters. To do so, we introduce a third condition in terms of local L-functions. We assume for the rest of the section that the local eld F is p-adic, although the discussion would also be valid in the archimedean case. Suppose

382

6. THE LOCAL CLASSIFICATION

r that P sim pN q is a simple datum. The question is whether lies in the r r subset sim pGq of sim pN q. Following the global construction from Lemma r 5.3.2, we introduce the local parameter ` in the set pN` q. Then N` 2N is even and ` 1, and there are two split groups G` and G_ ` r in Esim pN` q over F . They share a maximal Levi subgroup M` GLpN q. As in the global case, G` is distinguished from G_ by the condition that ` p p p G` contain the product G G. To emphasize the analogy with the global resolution of Corollary 5.4.7, let us write 1 if the local Langlands-Shahidi L-function Lps, , _ q at` _ _ tached to a maximal parabolic subgroup P` M` N` of G_ has a pole at ` p s 0. This is what we expect if maps LF into L G. For if G is orthogonal, _ is the symmetric square representation of M . The restriction of _ to x` ` ` L G in M then contains the trivial representation. Similarly, x` the image of p x if G is symplectic, _ is the skew-symmetric square representation of M` , ` L G again contains the trivial represenwhose restriction to the image of tation. We can also write 1 in case it is the L-function Lps, , ` q attached to a maximal parabolic subgroup P` M` N` that has a pole at s 0. This of course is what we have to rule out. r Corollary 6.8.1. Suppose that P sim pN q and that G belongs to rsim pN q. Then the following conditions on the pair pG, q over our local E p-adic eld F are equivalent. r r (i) The linear form f G pq on Scusp pGq in (6.1.1) does not vanish. (ii) As a local Langlands parameter for GLpN q, factors through the image of L G in GLpN, Cq. (iii) The local quadratic characters G and are equal, and the local L-function condition 1 holds. Proof. The conditions (i) and (ii) represent our two possible denitions r sim pGq. They will be linked by the last condition (iii). of We recall that Henniart [He2] has established the identities Lps, , S 2 q Lps, S 2 q and Lps, , 2 q Lps, 2 q of local L-functions. The left hand sides are Langlands-Shahidi L-functions attached to the square integrable representation of GLpN, F q, while the right hand sides are the L-functions attached to representations of LF . With these relations, the corollary becomes a consequence of the isomorphism between the representation theoretic and endoscopic R-groups of the parameter r ` P pG` q. For it follows from the remarks preceding the statement of the corollary that (ii) is equivalent to (iii), with L-functions interpreted as on the right. It follows from the local intertwining relation that (i) is also equivalent to (iii). More precisely, (i) is valid if and only if the induced representation IP` p q is reducible, since the left hand side of the analogue of

6.8. RESOLUTION

383

(2.4.7) for pG` , ` q then represents the expected nonzero linear form r on Scusp pG Gq. By the denition of the representation theoretic R-group Rp q, IP` p q is reducible if and only if the Plancherel density P` p, q is nonzero at 0, or equivalently, the normalizing factor rP` |P ` p, q rP` |P ` p q has no pole at 0. The equivalence of (i) with (iii) then follows from the denition (2.3.3) of the normalizing factors in terms of L-functions. Corollary 6.8.1 is clearly parallel to its global counterpart Corollary 5.4.7, from Chapter 5. Together, they resolve the temporary denitions of the r local and global sets sim pGq from 7.1 and 6.1 in terms of the original denitions. In each case, it is the condition (i) that gives the temporary denition in terms of harmonic analysis, and condition (ii) that gives the original denition. Notice, however, that while the global conditions (i) and (ii) of Corollary 5.4.7 are closely related, the local conditions (ii) and (iii) of Corollary 6.8.1 are the ones that are most closely related to each other. There is also another dierence. Corollary 5.4.7 was proved, for the r general family F with local constraints in 5.4, as part of a running induction 9 r argument on N . Corollary 6.8.1 was proved, with the help of the family F constructed in this chapter, only after the induction hypothesis on N had been resolved. With the proof of Corollary 6.8.1, we have reconciled the temporary defr inition of sim pGq in terms of harmonic analysis with the natural denition in terms of local Langlands parameters. In particular, we have established Theorem 6.1.1. This completes our proof of the local Langlands corresponr r dence for any group G P Esim pN q (up to the action of the group OutN pGq of order 2, in case G is of type Dn ).

CHAPTER 7

Local Nontempered Representations

7.1. Local parameters and duality We come now to the local theorems for nongeneric parameters . Our general strategy will be similar to that used in the last section to treat generic parameters . Given local objects pF, G, q, we shall construct cor9 9 9 9 r responding global objects pF , G, q, and a family of global parameters F to which we can apply the results of Chapter 5. This section is parallel to 6.1. Its purpose is to examine local parameters 9 that occur as completions v of what will be our special global parameters 9 . These completions will again fall into dierent categories. First of all, we 9 will have the place v u at which v equals the given local parameter . 9 u , we will again need to control the behaviour To obtain information about 9 of completions v at a set of places v P V distinct from u. In contrast to the last chapter, however, V will be a nite set of p-adic places rather than u the set S8 of complementary archimedean places. We are assuming in this chapter that F is a local eld. Suppose r that P pGq is a general local parameter for a simple endoscopic datum r G P Esim pN q over F . The main local result to establish is Theorem 2.2.1. The rst assertion (a) of this theorem is now known. It was reduced in Lemma 2.2.2 to the generic case, which we resolved in Lemma 6.6.3. In r particular, the OutN pGq-symmetric linear form (7.1.1) f 1 p 1 q, r f P HpGq,

of the second assertion (b) of the theorem is well dened. The main point is to show that it has an expansion (2.2.6). We shall introduce a formal expansion of (7.1.1). Recall that the pair 1 , 1 q is the preimage of p, sq, for a semisimple element s in S . We rst pG observe that f 1 p 1 q depends only on the image x of s in the quotient S of S . This would of course be a consequence of the local intertwining relation for , which we have yet to prove. However, the claim is easy to establish directly. It follows from the local form of the discussion of the left hand side of (4.5.1), given prior to the statement of Proposition 4.5.1. Knowing that (7.1.1) depends only on x, we can expand it formally as a linear combination of irreducible characters on S . Building a translation of r S by s into the denition, we obtain a general set of OutN pGq-symmetric
385

386

7. LOCAL NONTEMPERED REPRESENTATIONS

linear forms f fG pq, r r P , f P HpGq, x P S , which is in bijection with the set of irreducible characters x xx, y, on S , such that (7.1.2) f 1 p 1 q
r P

xs x, y fG pq,

r for any x P S and f P HpGq. The assertion (b) of Theorem 2.2.1 is that each r r P is a nonnegative, integral linear combination of (OutN pGq-orbits of) irreducible unitary characters on GpF q. r We shall take the liberty of calling the packet of , in anticipation r r of what we expect to prove. For and will then represent the same object, diering only in their interpretations as packets over the respective p r sets S and unit pGq. 9 9 We will use the expansion (7.1.2) to study the localization u of we want to understand. It will be necessary to allow the valuation v to be archimedean as well as p-adic. This is because the archimedean packets of [ABV], which exist for general groups, are not dened by twisted transfer to GLpN q. In fact, it will be valuations in the union U S8 puq S8 Y tuu 9 of S8 and tuu that parametrize completions v for which we have to establish Theorem 2.2.1. The set U will in some sense be the analogue of the singleton tuu from Chapter 6. The rest of the section will be devoted to parameters that are to be 9 localizations v at places v in complement of U . Such valuations are of course nonarchimedean. We will break them into a disjoint union of a nite 9 set V , and the complement S V,U of V and U . The localizations v at V,U will play the role of the spherical parameters described in 6.1. As vPS we have already said, the remaining set V will be the analogue of the set u S8 in Chapter 6. We will take it to be any large nite set of valuations v of 9 9 F for which the degrees qv of the corresponding residue elds are all large. We assume for the rest of the section that the local eld F is nonarchimedean. Suppose for the moment that G is a general connected reductive group over F . An irreducible representation of GpF q determines a connected component in the Bernstein center of G. This is given by a pair pM , r q, where M is a Levi subgroup of G, and r is a supercuspidal representation of the group ( M pF q1 x P M pF q : HM pxq 0

7.1. LOCAL PARAMETERS AND DUALITY

387

such that is a subquotient of an induced representation IP pr, q, where r, , P a ,C , M is the a ,C -orbit of supercuspidal representations of M pF q attached to r . M We write pq p1qdimpAM0 {AM q , where M0 is a minimal Levi subgroup contained in M . Since pM , r q is determined up to conjugacy, the sign pq is an invariant of . It plays an important role in the duality operator of Aubert-Schneider-Stuhler. The duality operator D DG is an involution on the Grothendieck group KpGq of the category of GpF q-modules of nite length. It is dened [Au, 1.5] by G (7.1.3) DG p1qdimpAP0 {AP q iG rP , P
P P0

P P PpM q,

where P M NP ranges over standard parabolic subgroups of G, and iG and P G rP are the functors of induction and restriction between KpGq and KpM q [BZ, 2.3]. A key property of DG is that it preserves irreducibility up to sign. More precisely, suppose that is an irreducible representation of GpF q, with image rs in KpGq. Then (7.1.4) Drs pq rps, p for an irreducible representation of GpF q. (See [Au, Corollaire 3.9 and Erratum] and [ScS, Proposition IV.5.1].) Duality also has the property of being compatible with endoscopy. Suppose that G1 is an endoscopic datum for G, equipped with auxiliary datum r r pG1 , 1 q and corresponding transfer factors, as in 2.1. The transpose of the 1 associated transfer mapping f f 1 of functions is a transfer S 1 SG of disr r tributions. Here S 1 is any stable, p 1 q1 -equivariant distribution on G1 pF q, 1 is the invariant, 1 -equivariant distribution on GpF q dened by and SG
1 p SG pf q S 1 pf 1 q,

f P HpG, q.

Now the representation theoretic character provides an isomorphism from the complex vector space KpGqC KpGq bZ C onto the space of invariant distributions on GpF q that are nite, in the sense that they are nite linear combinations of irreducible characters. The duality mapping DG can therefore be regarded as an involution on the space of nite invariant distributions on GpF q. The compatibility property is a formula (7.1.5)
1 pD1 S 1 qG pG, G1 q pDG SG q,

388

7. LOCAL NONTEMPERED REPRESENTATIONS

r in which S 1 is assumed to be nite, D1 is the duality involution for G1 , and pG, G1 q p1q
dimpAM0 {AM 1 q
0

It comes with the implicit assertion that D1 maps the subspace of stable distributions to itself. (See [Hi], [A25].) More generally, suppose that G is a twisted triplet pG0 , , q, as in 2.1. One can dene the complex vector space KpGq b C as the span of objects (2.1.1). The duality involution (7.1.3) makes sense in this generality, provided that the indices of summation P are treated as standard parabolic subsets of G [A4, 1]. It acts on the Grothendieck group KpGq of the subcategory of G` pF q-modules of nite length whose irreducible constituents are of the form (2.1.1). The compatibility formula (7.1.5) then extends to this setting, with the appropriate denition of the sign pG, G1 q. (For a proof of this and other properties discussed below, we refer the reader to the paper [A25] in preparation.) We note that we still have local parameters P pGq, if G represents a general triplet pG0 , , q. For any such , we write p pw, u1 , u2 q pw, u2 , u1 q, w P WF , u1 , u2 P SU p2q, for the dual parameter that interchanges the two SU p2q-factors in the group LF SU p2q WF SU p2q SU p2q. We also form the sign pq p1q
dimpAM0 {AM q

where M0 is a minimal Levi subset of G, and M is a minimal Levi subset 0 for which the L-group L M contains the image of . r Assume now that G belongs to Esim pN q. At this point, we are thinking 9 of particular parameters that are to be localizations v at places v outside p r of U . They will be of the form , for generic parameters P pGq. p corresponds to a r Any pair of local parameters P pGq and pair of irreducible, self-dual representations and of GLpN, F q. It follows from the conjecture of Zelevinsky [Z, 9.17], proved by Aubert [Au, Thor`me 2.3] and Schneider and Stuhler [ScS], that equals . However, e e p our real interest is in the twisted characters ` r r r fG pq tr pf q and ` r r r fG pq tr pf q r r r in f P HpN q. These are nite invariant distributions on GpN, F q, which we r r can denote by and . They behave in a simple way under the duality r for GpN q. For it can be seen from twisted extensions of the r operator D arguments of [Au], applied to the very special case at hand, that r r rr D pq .

7.1. LOCAL PARAMETERS AND DUALITY

389

p (See the general construction of [MW4, 3].) The parameters and also provide nite, stable distributions on GpF q. We may as well denote r r them again by and , since they are the unique OutN pGq-preimages of r and under the twisted transfer mapping of distributions r S S SGpN q . r To describe their behaviour under the duality operator D DG , we apply r rr the analogue for GpN q of (7.1.5) and the formula for D above. We obtain r r r rr r pDq pG, Gq D pG, Gq pq . r r The mapping S S is an injection if S is required to be OutN pGq-invariant. Since r r pG, Gq pq pq in this case, we conclude that (7.1.6) D pq .

p We shall apply these remarks to the case of the general expansion (7.1.2). Since the groups S and S are the same, we can dene a canonical r r bijection from and , with matching characters xx, y xx, y, r x P S , P .

p r r Lemma 7.1.1. Suppose that , for P pGq and G P Esim pN q. Then (7.1.7) r for any P . Proof. The formula (2.2.6) is valid for the generic parameter . It can be written 1 xx, y , G
r P

xs , y pq pq , p

r where pG1 , 1 q corresponds to p, xq, and is regarded as an OutN pGqinvariant distribution on the right hand side. We shall apply the operator D DG to each side of this identity. It follows from (7.1.5) and the analogue for G1 of (7.1.6) that DG 1 pG, G1 q pD1 1 qG pG, G1 q p1 q p 1 qG . G Since pG, G1 q p1 q pq, we obtain
1 DG 1 pqp 1 qG pq G . G

390

7. LOCAL NONTEMPERED REPRESENTATIONS

The original identity becomes

1 G

r P

xx, y pq D

` p xx, y pq pq .

On the other hand, the expansion (7.1.2) can be written 1 xs x, y , G

r P

r where is again understood as an OutN pGq-invariant distribution on the right hand side. This identity then becomes ` 1 G xx, y xs , y
r P

r P

` xx, y xs , y ,

since s can be regarded as an element in S . It gives us a second expansion 1 of G , as a function of x into characters xx, y. Identifying the coecients of xx, y, we obtain the required formula (7.1.7). r Theorem 2.2.1 asserts that the distribution attached to any P r is a character (up to the action of OutN pGq). Combined with Lemma 7.1.1, it implies the stronger assertion that is actually equal to the irreducible character . It also implies the identity of signs p (7.1.8) pq pq xs , y, for every representation P . Conversely, suppose that the identity p (7.1.8) is for all . Then equals , so the assertion of Theorem 2.2.1 holds for in this stronger form. r r Given the generic parameter P pGq, we write G for the subset of r representations P such that the identity (7.1.8) holds. We then dene the corresponding subset ( r r p (7.1.9) G : P G r r r r r of . The subset G G of the packet thus consists of p (OutN pGq-orbits of) irreducible representations . The assertion of p amounts to the equality Theorem 2.2.1 for the dual parameter (7.1.10) r r G

of the two sets. 1 Lemma 7.1.1 also has implications for the function fG p, q on the left hand side of the nongeneric local intertwining relation (2.4.7) of Theorem

7.1. LOCAL PARAMETERS AND DUALITY

391

2.4.1. To describe it, we rst transfer the duality operator to the function r r spaces IpGq and SpGq. r r Recall that IpGq is a space of functions on temp pGq. Any such function extends by analytic continuation to a function on the corresponding set of r standard representations, and hence by linearity, to a function on pGq and r on the complex Grothendieck group KpGqC . We can therefore regard IpGq as a space of linear functions on the vector space of OutN pGq-symmetric r invariant distributions on GpF q that are nite. Recall also that SpGq is r a space of functions on bdd pGq. These objects can also be extended by analytic continuation and linearity. The generic form of Theorem 2.2.1, established in Chapter 6, can be used to characterize the space of stable r distributions within the larger space of OutN pGq-symmetric, invariant, nite r distributions. It allows us to regard SpGq as a space of linear functions on r N pGq-symmetric stable distributions of GpF q that are nite. the space of Out r r These descriptions of IpGq and SpGq are compatible with the convention r above of treating a parameter P pGq as a nite stable distribution (as in r (7.1.6)), and a representation P pGq as a nite invariant distribution (as in the proof of Lemma 7.1.1). With this understanding, we dene duality r r operators D DG on both IpGq and SpGq by the adjoint actions pDaqpq apDq, and pDbqpq bpDq, r r b P SpGq, P bdd pGq. r r This gives continuous involutions on both IpGq and SpGq. p Suppose that G, and are as in Lemma 7.1.1. It is easy to deduce from the lemma that (7.1.11)
1 fG p, s xq pq pDfG q1 p, s xq,

r r a P IpGq, P temp pGq,

r f P HpGq,

for any element x in the group S S . To do so, we rst observe that pDfG q1 p, xq xx, y pDfG qpq, x P S ,
r P

according to the denition (2.4.6) and the generic form of (2.2.6) we have established. We can therefore write pDfG q1 p, xq xx, y fG pDq

pq since the lemma tells us that

r P

xs x, y fG p q,

p D pq pq xs , y .

392

7. LOCAL NONTEMPERED REPRESENTATIONS

It then follows from (2.4.6) and (7.1.2) that 1 xs x, y fG pq pq fG p, xq. pDfG q1 p, xq pq

r P

Replacing x by s x, we obtain (7.1.11). Now the generic form of Theorem 2.4.1 we have established implies that pq pDfG q1 p, s xq pq pDfG q p, s uq, where u is any element in N whose image xu in S equals x. Given the reduction of Lemma 2.4.2, the assertion of Theorem 2.4.1 for the dual pap rameter is consequently equivalent to the relation r (7.1.12) pq pDfG q p, s uq fG p, uq, u P N , f P HpGq. This is a statement about the behaviour under duality of intertwining operators, equipped with the normalization (2.4.4) of Chapter 2. To be more precise, we x a pair r pM, M q, M P 2 pM, q, p as in the preamble to Proposition 2.4.3. Then M equals M , for a xed r 2 pM, q. We assume that M is proper in G, and that the parameter M P analogue of Theorem 2.2.1 holds inductively for pM, M q. According to the r r denition (2.4.5), fG p, uq is the sum, over M in the set M M , of the linear forms ` (7.1.13) fG p, u, M q xr, M y tr RP pwu , M , q IP pM , f q . u r r With this notation, it is not hard to see that Theorem 2.4.1 implies the identity of linear forms r p (7.1.14) pq pDfG q p, s u, M q fG p, u, M q, u P N , f P HpGq, r for every representation M P M . This is a straightforward consequence r of the fact that either side of (7.1.12) equals the sum over M P M of the linear forms on the corresponding side of (7.1.14). Conversely, it is clear that if (7.1.14) is valid for all M , then Theorem 2.4.1 holds for . It is clear from the denition (7.1.13) that the putative identity (7.1.14) includes a condition on the representation theoretic R-groups of and . r D. Ban [Ban] has shown that for the groups G P Esim pN q in question, the two R-groups are indeed the same. We recall that the endoscopic R-groups of and are equal by denition. These are in turn isomorphic to any of the representation theoretic R-groups of , by the results of Chapter 6. Theorem 8.1 of [Ban], which relies on the results of [G], implies that the same is true of the representation theoretic R-groups of . More precisely, r the normalized intertwining operators for representations M P M and r correspond, up to a scalar multiple, under the action of D. p M P M Stated in terms of the notation (7.1.13), this amounts to a weaker form r p (7.1.15) fG p, u, M q puqpDfG qp, u, M q, u P N , f P HpGq,
M

7.1. LOCAL PARAMETERS AND DUALITY

393

of (7.1.14), for a sign M puq. The formula (7.1.15) is reminiscent of (6.1.4), the weaker form of the local intertwining relation for archimedean F and generic that follows from Shelstads work. In fact, the two formulas have almost identical roles. They 9 apply to elements v in the set V of special valuations of F (archimedean in Chapter 6, p-adic here) by which the local intertwining relation (for generic parameters in Chapter 6, and general parameters here) is deduced by global means. In each case, M represents a character on the subgroup W of W pM q. Its inclusion in the formula has the eect of making the left hand side insensitive to the specic normalizations introduced in Chapter 2. The formulas would both hold (with dierent choices of characters M ) for any normalizations Rpw, M q of the intertwining operators that are multiplicative in w. They follow directly from the property that the operators 0 Rpw, M q are scalars for w P W , and form a basis of the space of all intertwining operators of the induced representation IP pM q as w ranges over R . It is this property that follows from [Ban]. We note that Theorem p 8.1 of [Ban] requires that M be unitary. This condition is included in our induction assumption that Theorem 2.2.1 is valid for M . (The results in [Ban] and [G] were actually stated for G split, but with our knowledge of r the groups RpM q from 6.6, they extend easily to the case that G P Esim pN q is quasisplit.) r r Given the generic parameter P pGq, we write GM for the subset of r representations M P M such that the identity (7.1.14) holds. We then dene the corresponding subset ( r r p (7.1.16) GM M : M P GM p r of M . The assertion of Theorem 2.4.1 for the dual parameter amounts to the equality (7.1.17) r r GM M

of the two sets. Our discussion of duality has been designed for application to local com9 pletions v at places v R U . The role of such completions will be similar to that of their analogues from Chapter 6. To emphasize this point, we can specialize the local parameters and further. A nite dimensional representation of WF is said to be tamely ramied if its kernel contains the wildly ramied inertia group. In other words, the representation is trivial on the pro-p part of the inertia subgroup IF of WF , p being the residual characteristic of F . We shall say that the representation is quadratic if its irreducible constituents have dimension at most two. We thus observe a further symmetry with the discussion of 6.1. For it is clear r that the tamely ramied, quadratic representations of WF that lie in 2 pGq are in some sense analogues of the archimedean parameters described prior to Lemma 6.1.2.

394

7. LOCAL NONTEMPERED REPRESENTATIONS

p We shall say that the local parameters and are tamely ramied and quadratic if their restrictions to WF have these properties. p r Lemma 7.1.2. Suppose that and P pGq are tamely ramied, r quadratic parameters for our simple endoscopic datum G P Esim pN q over the p-adic eld F . r p p (a) The image of G in the dual group S generates S , and contains the trivial character 1. r p p (b) The image of GM in the dual group SM generates SM , and con tains the trivial character 1. Lemma 7.1.2 is intended as an analogue of Lemma 6.1.2. It provides the limited local information that will be used to deduce the general local theorems by global means. The two assertions of the lemma are obviously equivalent to their analogues for . It is really the rst assertion (i) (taken for ) that is parallel to Lemma 6.1.2. The second assertion (ii) is actually a stronger analogue of what was available in 6.1. It represents a renement (for the p-adic eld F here) of the formula (6.1.5), in which the characters M pxq are all trivial. As we shall see in 7.3, this renement will greatly simplify our proof of the local intertwining relation. We shall give a proof of Lemma 7.1.2 in [A25] that is based on the representations of Hecke algebras, r as well as a proof of a natural variant of the lemma for the bitorsor G rsim pN q. attached to an even orthogonal group G P E We should note that analogues of Lemma 7.1.2(i) have been established in complete generality by Moeglin [M1] [M2] [M3]. Using a remarkable family of partial duality involutions, she has established results that among r r r other things imply that G equals , for any generic parameter P pGq. Her results are partially global, in that they depend on the generic form of Theorem 2.2.1, which was established by global means in Chapter 6. 9 7.2. Construction of global parameters We will now introduce the families of global parameters needed to establish the general nontempered local theorems. The construction is essentially that of 6.26.3. However, we shall have to impose local constraints at a nite set of nonarchimedean places. These will be used to apply the duality properties described in the last section. In particular, they will be used in the next section to modify the discussion of 5.4. r Suppose that G P Esim pN q is a xed simple twisted endoscopic datum r for GpN q over the given local eld F . Consider a nongeneric parameter (7.2.1)
1 1

r r ,

r in pGq, with simple components r i P sim pGi q, r Gi P Esim pNi q.

9 7.2. CONSTRUCTION OF GLOBAL PARAMETERS

395

Given the local pairs pGi , i q over F , we will want to choose global pairs 9 9 9 9 9 r r 9 pGi , i q, Gi P E sim pNi q, i P sim pGi q, 9 over a suitable global eld F such that 9 9 9 pFu , Gi,u , i,u q pF, Gi , i q, 1 i r, 9 for some valuation u of F . These will determine a global endoscopic datum 9 9 r G P E sim pN q such that the global parameter 9 9 9 (7.2.2) 1 1 r r r 9 belongs to pGq. The process will be similar to that of Proposition 6.3.1, except that supplementary constraints on the global data will have to be slightly dierent. There is also the minor dierence that G was taken from r the larger set Eell pN q in Proposition 6.3.1, rather than the set of simple data r Esim pN q we have specied here. As in the generic case in 6.3, we x a Levi subgroup M of G such that r 9 the set 2 pM, q is nonempty. Given G, we will also choose a global Levi 9 9 9 r subgroup M of G such that 2 pM, q is nonempty. The analogues (7.2.3) and (7.2.4) 9 1 9 9 9 M 11 rr
1

M 11 rr

of (6.3.3) and (6.3.4) can then be identied with xed elements in the rer r 9 9 spective sets 2 pM, q and 2 pM , q. As before, we will have a special interest in the case that each i equals 1. This of course is the case that 9 9 9 9 9 M G and M , with M G and M . Recall that there are decompositions i i b i , Ni mi ni , 1 i r, r r where i P sim pHi q is a simple generic parameter for a datum Hi P Esim pmi q over F , and i is an irreducible representation of SU p2q of dimension ni . Set I m,n ti : mi m, ni nu, for any positive integers m and n. We then have another decomposition 8 n (7.2.5) p b n q,
n1

where n

p i i q,

m1 iPI m,n

and n is the irreducible representation of SU p2q of dimension n. The global 9 constituents of will be of the form 9 9 9 i i b i , 1 i r,

396

7. LOCAL NONTEMPERED REPRESENTATIONS

9 r 9 9 r where i P sim pHi q is a simple generic parameter for some Hi P E sim pmi q 9 9 9 over F , and i is the extension of i to the group SLp2, Cq. As we noted 9 9 r 9 back in 1.4, these objects determine the simple data Gi with i P sim pGi q. For the construction, we will apply a variant of Proposition 6.3.1 directly to the local generic parameter 8 r n p q p i i q.
n1 i1

9 There will be four possibilities for the global eld F . The simplest case is 9 when the local eld F equals C. In this case we take F to be any imaginary quadratic extension of Q. If F R, and if the two dimensional generic constituents ti : mi 2, 1 i ru 9 of are in relative general position, we simply take F Q. If F R, but these two dimensional generic constituents are not in relative general 9 position, we take F to be a totally real eld with several archimedean places. 9 Finally, if F is p-adic, we take F to be any totally real eld of which F 9 equals some completion. In all cases, we x a valuation u of F such that 9u F . As agreed in the last section, we also x a large nite set V of F nonarchimedean places, which does not contain u, and for which qv is large 9 for any v P V . This set will assume the role played in Proposition 6.3.1 by the complementary set u S8 S8 tuu of archimedean places. According to our remarks in 7.1, it will be the set U S8 puq S8 Y tuu that often takes the place of the individual valuation u in Chapter 6, while it will be the union S8 pu, V q SpU, V q U Y V that assumes the earlier role played by S8 puq. Proposition 7.2.1. Given the local objects G, , M and M over F as 9 9 9 in (7.2.1) and (7.2.3), we can choose corresponding global objects G, , M 9 M over F as in (7.2.2) and (7.2.4) such that the following conditions 9 and are satised. 9 (i) There is a valuation u of F such that 9 9 9 9 9 pFu , Gu , u , Mu , M,u q pF, G, , M, M q, and such that the canonical maps SM SM 9 and S S 9 are isomorphisms.

9 7.2. CONSTRUCTION OF GLOBAL PARAMETERS

397

(ii) For any valuation v outside the set S8 pu, V q, the local Langlands parameter v 1 1,v r r,v , 9 9 9 9 is a direct sum of tamely ramied quasicharacters of Fv , while the corresponding decomposition of the subparameter 9 9 1,v r,v contains at most one ramied quasicharacter. r 9 9 (iii)(a) For any v P V , the parameters i,v lie in 2 pHi,v q and are each a direct sum of tamely ramied representations of WFv of dimension one 9 9 v equals the dual v of a generic parameter p and two, so in particular, r 9 v P pGv q. Moreover, the canonical mappings p p9 r r pM q SM SM M 9 9
M,V

vPV

M,v

and r G 9
M,V

vPV

r p G 9

M,v

p9 p q SM SM

are surjective. (iii)(b) Suppose that each i equals 1. Then there is a v P V with the 9 9 r r property that if v lies in pG q for some G P E sim,v pN q, the dual group v v p p G equals G. v (iii)(c) Suppose that some i is greater than 1. Then there is a v P V such that the kernel of the composition of mappings S Sv Rv 9 9 9 9 contains no element whose image in the global R-group R R pGq belongs 9 9 to R,reg . 9 Proof. The statement is clearly very close to that of Proposition 6.3.1. Notice that the condition (iii)(a) here applies to the two sets of p-adic representations from Lemma 7.1.2 (but with G replaced by M in the rst set). The condition (iii)(a) of Proposition 6.3.1 applies to the L-packets over R from Lemma 6.1.2. We have included the possibility here that F C. In this case, however, each constituent is one-dimensional. Since it is also self-contragredient, i 9 must then be the trivial character on C . We simply take i to be the 9 trivial id`le class character for F . The various conditions then follow without e diculty. 9 We can therefore assume that F C. Then F is a totally real eld, chosen according to the description prior to the statement of the proposition. There will be some supercial dierences from the setting of Proposition 9 6.3.1. For example, we are allowing a case F Q in which there is only one real valuation. Moreover, the constituents i of need not be distinct, since they could be matched with dierent representations i of SU p2q. Finally,

398

7. LOCAL NONTEMPERED REPRESENTATIONS

we will be applying the construction of Proposition 6.3.1 as it applies to the larger set of valuations u u S8 pV q S8 Y V
u in place of S8 . However, the general structure of the proof will be similar. The rst step is to construct the primary global pairs

9 9 pHi , i q,

9 9 r r 9 9 Hi P E sim pmi q, i P sim pHi q,

9 over F from the given local pairs pHi , i q, r r Hi P Esim pmi q, i P sim pHi q,

over F . We apply Corollary 6.2.4, supplemented by Remark 3 following its proof. For every i, the global pair then has the properties that 9 9 9 pFu , Hi,u , i,u q pF, Hi , i q, 9 9 that i,v is spherical for every v R S8 pu, V q, and that i,v is an archimedean u parameter in general position for every v P S8 . The new condition applies to the nite set V of nonarchimedean places. According to the supplementary remark, we can arrange that for each v P V , the local Langlands parameter r 9 9 i,v belongs to 2 pHi,v q. In fact, we can assume that i,v is a direct sum 9 of distinct, irreducible, self-dual, tamely ramied representations of WFv of 9 dimension two with unramied determinant, together with an unramied character of WFv of order 1 or 2, in case mi is odd. Our hypothesis that the 9 residual characteristics of valuations v P V are large insures that there are enough such two-dimensional representations of WFv . 9 9 9 As we noted above, the simple generic pairs pHi , i q can then be com9 bined with the representations i of SLp2, Cq to give nongeneric pairs 9 9 pGi , i q, 9 9 9 r 9 9 Gi P E sim pNi q, i P Esim pGi q.

9 9 r These objects in turn determine a global endoscopic datum G P E sim pN q p 9 r 9 9 p with G G, together with a global parameter P pGq as in (7.2.2). They x x 9 9 9 also determine a global Levi subgroup M of G with M M , together with 9 r 9 9 a global parameter M P pM , q as in (7.2.4). The condition (i) follows. To obtain the remaining conditions of the proposition, we need to impose 9 9 more local constraints on the basic global pairs pHi , i q. We can by and large follow the prescription from the proof of Proposition 6.3.1, with V now being u the chosen set of nonarchimedean valuations instead of S8 . In particular, we will choose the global pairs successively with increasing i. Consider the quadratic id`le class character e 9 9 i Hi i detpi q 9 9 9 9 p attached to any Hi with Hi SOpmi , Cq. Its local values i,v at places v P V are predetermined if mi is even. In this case, the conditions on i,v 9

9 7.2. CONSTRUCTION OF GLOBAL PARAMETERS

399

imply that 9 i,v pv qqi , qi

1 2

mi ,

9 where v is the nontrivial unramied quadratic character on Fv . If mi is odd, we apply the criterion of Proposition 6.3.1 to V . Namely, we x a set of distinct places V o tvi : mi oddu, and then require that # pv qqi `1 , if v vi , 9 (7.2.6) i,v pv qqi , if v vi ,

qi

1 2

pmi 1q,

9 for any v P V and i with mi odd. Recall that the local values i,v are also determined at places v P S8 puq if mi is even, and at v u in general. To u 9 be denite, we arbitrarily x i,v in the remaining case of v P S8 and mi odd. (We cannot use the criterion of Proposition 6.3.1, which we have just u applied to our p-adic set V here, since we are not assuming that S8 is large. There is no need to do so in any case, since V is now taking the earlier role u of S8 .) 9 Having xed the local values i,v at any v P S8 pu, V q and i with p Hi SOpmi , Cq, we choose the global quadratic characters as in the proof of Proposition 6.3.1. We require that for any v R S8 pu, V q there be at most 9 one i such that i,v ramies, and that the ramication be tame. Recall that 9 this is done by insisting that i,v be unramied at any place v R S8 pu, V q, 9 which divides 2, or at which j,v ramies for some j i. Once we have 9 chosen the quadratic character i , we obtain the global endoscopic datum x 9 9 p Hi from the dual group H i Hi . We then construct the global parameters r 9 9 i P sim pHi q according to Lemma 6.2.2 and Corollary 6.2.4 (modied by u the accompanying Remark 3). At the places v in either S8 or V , we rst 9 i,v in terms of their predecessors j,v , with j i. specify the localizations 9 This is how we arrange that as i varies, the constituents of i,v are mutually v. disjoint, and in fact, in relative general position if v P S8 The required property (ii) follows just as in the proof of Proposition 6.3.1. The same is also largely true of the remaining assertions in (iii)(a) (iii)(c). They are for the most part trivial unless the subset I 1 of indices i P I o with mi 1 is not empty. As in the earlier proof, it is really only for this possibility that the property (7.2.6) was needed. The rst assertion of (iii)(a) is an immediate consequence of the construction. In the other assertion of (iii)(a), the rst mapping is analogous to that of Proposition 6.3.1(iii)(a). As before, it suces to treat the case that M G. Working with the valuations vi as i varies over I 1 , we again show that the mapping p p S 9 S 9
V

400

7. LOCAL NONTEMPERED REPRESENTATIONS

is surjective. We then establish this part of the assertion, using Lemma 7.1.2(a) in place of Lemma 6.1.2, from the fact that the set V is large. The second mapping in (iii)(a) has no counterpart in Proposition 6.3.1. However, the required surjectivity follows in the same way from Lemma 7.1.2(b). For the conditions (b) and (c) of (iii), we can again take v to be any valuation in V if I 1 is empty. Otherwise, we must take v v1 vi1 , for any xed index i1 in I 1 . It might appear that (iii)(b) is complicated here by the existence of simple components i i b i with i P I 1 and ni ni1 . Indeed, the condition would fail on this account if we were allowing G to r be a general element in Eell pN q, as in Proposition 6.3.1. However, we have rsim pN q to be simple here. The condition (iii)(b) deliberately taken G P E then follows easily. The proof of the last condition (iii)(c) is identical to that of Proposition 6.3.1. The discussion of this section has been parallel to that of 6.3. Having no need of a nontempered analogue of Lemma 6.3.2, we are now at the point we reached at the end of 6.3. Given simple local parameters i i b i 9 9 9 over F , we have constructed global parameters i i b i over F . We can 9 then form the associated family (7.2.7) 9 9 9 9 r r 9 F Fp1 , . . . , r q t 91 1 9r r : 9i 0u

of compound global parameters. This represents another family of the sort treated in Chapter 5. In contrast with the earlier generic case, Assumption 5.1.1 requires no further eort to justify here. It follows from Lemma 2.2.2 and what we have established in the generic case. We recall that this assumption was the basic premise of Chapter 5. 9 r We now formally adopt the induction hypotheses for F that were in force in Chapter 5. We x the positive integer N , and assume that the r local theorems all hold for general parameters P with degpq N . To this, we add the induction assumption that the global theorems all hold for 9 9 9 r parameters in the given family F with degpq N . The global hypothesis is understood to include the conditions on Theorem 1.5.3, described near the beginning of 5.1. The assertion of Theorem 1.5.3(b) thus represents an induction assumption for the nontempered fam9 r ily F, which is resolved in Lemma 5.1.6. The assertion of Theorem 1.5.3(a) 9 represents a condition on the simple generic constituents i of parameters 9 9 r P F, which has also been treated. It is part of the resolution from Corollary 5.4.7, which among other things, is the foundation for the denition of 9 r 9 r the subsets FpGq of F. We recall that the global theorems for simple generic parameters with local constraints were established in Lemma 5.4.6. The local constraints for the simple generic parameters i here are admittedly at 9 u the nonarchimedean places V , rather than the archimedean places S8 of 5.4. But with the local theorems in place now for all of the constituents of

7.3. THE LOCAL INTERTWINING RELATION FOR

401

9 i , the proof of Lemma 5.4.6 carries over in a straightforward way. Alternatively, we observe that the general proof of the global theorems, which we will complete in 8.2, can be carried out for generic parameters with no 9 reference to the local results for nongeneric parameters of this chapter. We would like to apply the results of Chapter 5 to the nongeneric family (7.2.7) with local constraints. Assumption 5.1.1 is easy to verify in this case. It follows from the reduction in Lemma 2.2.2 of the local assertion of Theorem 2.2.1(a) and our induction assumption on the generic constituents i of parameters in (7.2.7). (The argument here is similar to that of the early 9 part of the proof of Lemma 8.1.1, at the beginning of the next chapter. This places the family (7.2.7) in the general framework of 5.1. In particular, the general lemmas of 5.15.3 are valid (as applicable) for its parameters. However, their stronger renements in 5.4, which we used to establish the local theorems of Chapter 6, hold only for the generic family of 5.4. We are going to have to extend them to the nongeneric family (7.2.7). To deal with Theorem 2.4.1, we will need nongeneric renements of 9 r2 9 Lemmas 5.2.1 and 5.2.2. These apply to parameters in the subset Fell pGq 9 9 9 r 9 r of F attached to a simple datum G P E sim pN q. Any such takes the form # 9 9 9 9 9 21 2q q`1 r , ` ` (7.2.8) q 1, S Op2, Cqq Op1, Cqrq , 9 9 as in (5.2.4). In contrast to the earlier generic renements of 5.4, we will not need a separate argument for the case that the index r equals 1. To deal with Theorem 2.2.1, we will need nongeneric renements of Lemmas 5.3.1 and 5.3.2. These apply to parameters (7.2.9) 9 9 9 1 r , 9 r i P F sim pNi q,

9 r 9 r in the subset F2 pGq of F. We will again not require a separate argument for the case that r equals 1. However, we will otherwise follow the general strategy of Chapter 5. In particular, we shall introduce the simple datum 9 9 r G` P E sim pN` q, for N` N1 ` N , and the parameter 9 9 9 ` 1 , as in 5.3. We shall establish the necessary renements in the next section. The argument will be similar to that of 5.4, except that the exceptional place v will be taken from the set of p-adic valuations V . In order to reduce the problem to Corollary 3.5.3, which was the underlying tool in 5.4, we shall 9 9 apply the duality theory of the last section to the parameters v and `,v . 7.3. The local intertwining relation for The nongeneric form of Theorem 2.4.1 will be easier than the generic case treated in 6.4. This is because we will have some special cases to work

402

7. LOCAL NONTEMPERED REPRESENTATIONS

with. They consist of a limited number of local intertwining relations that can be obtained by duality from the generic form of the theorem already established. We will be able to combine them with a simplied form of the global argument from 6.4. First, however, we will have to deal with the question raised at the end of the last section. We shall establish a pair of lemmas that extend the global results of 5.4. The rst lemma gives the nongeneric renements of Lemmas 5.2.1 and 9 r2 9 r 5.2.2, for parameters (7.2.8) in the subset Fell pGq of F. It will be used later in the section to derive the general local intertwining relation. Lemma 7.3.1. Suppose that (7.3.1) 9 9 pG, q, 9 9 r2 9 9 r G P E sim pN q, P Fell pGq,

9 r is an analogue for the nongeneric family F of the pair given in either Lemmas 5.2.1 or 5.2.2. Then the expressions ` 1 9 9 99 9 9 c 0, (7.3.2) c G pxq f9G p, s xq f9G p, xq , 9 9 9 9
xPS,ell

and (7.3.3)
1 8

9 9 _ 9 9 pf9_ qM p1 q f9G_ p, x1 q , 9

if r 1,

r 9 r 9 from these lemmas vanish, for any functions f9 P HpGq and f9_ P HpG_ q. 9 r Proof. This lemma represents an analogue for F of Lemmas 5.4.3 and 5.4.4, which we recall were the renements of Lemmas 5.2.1 and 5.2.2 for generic parameters. In particular, (7.3.2) corresponds to one of the two expressions (5.2.8) or (5.2.13) (according to whether r 1 and r 1), while (7.3.3) corresponds to (5.2.12). As in the two renements for generic parameters, we actually have to establish several vanishing assertions. In stating this lemma, we have included only those assertions that will be used directly in the coming proof of the nontempered local intertwining relation. Following the argument from the two lemmas of 5.4, we shall reduce the proof to the assertion of Corollary 3.5.3. The main step is to write the expression (7.3.2) in the form (5.4.10) obtained in the proof of Lemma 5.4.3. We choose v P V according to the condition (iii)(c) of Proposition 9 p r 9 7.2.1. In particular, v is the dual v of a generic parameter v P pGv q, as prescribed by the condition (iii)(a) of the proposition. We take f9 f9v f9v , r 9 r 9 f9v P HpGv q, f9v P HpGv q,

to be decomposible. Then (7.3.2) equals ` 1 9 v 9 9 9 9 9 9 c G pxq f9v,G pv , s xv qpf9v q1 9 p v , s xv q f9v,G pv , xv q f9G p v , xv q , 9 9 9 9 9 9 9 9 9 G

xPS,ell

7.3. THE LOCAL INTERTWINING RELATION FOR

403

in the notation (5.4.7). There are two linear forms in f9v in each summand of this expression. To write them in terms of the generic parameter v , we 9 shall apply the duality operator Dv for the group Gv . 91 pv , s 9 xv q. It equals 9 The rst linear form is fv,G 9 9 pv qpDv f9v,G q1 pv , s xv q pv qpDv f9v,G qpv , s xv q, 9 9 9 9 9 9 by (7.1.11) and the local intertwining relation for the generic parameter v . This in turn can be written as pDv f9v,G q pv , s xv , v q pv q 9 9 9
r v Pv pMv q

in the notation (7.1.13), or what is the same thing, the product of pv q with the analogue for Dv f9v,G , v and s xv of the expression (5.4.8). (We are 9 9 9 r r pMv q for the Mv -packet here, as in (5.4.8), rather than writing v v,Mv as in (7.1.13). The Levi subgroup Mv is chosen in terms of v , and is not 9 generally equal to the localization Mv of Proposition 7.2.1.) As in the earlier 1 9 discussion of (5.4.8), we can then express f9v,G pv , s xv q as a general sum 9 9 9 ` 9 1 pv , xv q pDv f9v,G q Mv , v , rv pxq , 9
r v Pv pMv q

9 following the denition (3.5.3). It is understood that 1 pv , xv q is a complex coecient, that rv pxq rv ps xq ` is the image of x in the local R-group Rv , and that Mv , v , rv pxq repre9 sents an element in the set T pGv q, as in the proof of Lemma 5.4.3. 9 9 The other linear form is f9v,G pv , xv q. It can be written as 9 9 9 p f9v,G pv , xv , v q 9
r v Pv pMv q

9 in the notation (7.1.13), or equivalently, the analogue for f9v.G , v and xv of 9 9 (5.4.8). Our expectation (7.1.14) is that these expressions are equal to their analogues above. However, we do not need this ner property here. We can instead use the weaker formula (7.1.15) obtained from [Ban]. It asserts that 9 9 p 9 f9v,G pv , xv , v q v pxq pDv f9v,G q pv , xv , v q, 9 9 r for any v P v pMv q with associated sign v pxq. We may therefore express 9 9 f9v,G pv , xv q as a general sum 9 ` pv , xv q pDv f9v,G q Mv , v , rv pxq , 9 9
r v Pv pMv q

according again to the denition (3.5.3), and for another complex coecient 9 pv , xv q.

404

7. LOCAL NONTEMPERED REPRESENTATIONS

We can now write the original expression (7.3.2) in the desired form. For we see from the formulas above that (7.3.2) is equal to pv , f9v qpDv f9v,G qpv q, (7.3.4) 9
9 v PT pGv q

where pv , f9v q equals the sum 9 ` v 9 9 9 9 9 c G pxq 1 pv , xv qpf9v q1 9 p v , s xv q pv , xv q f9G p v , xv q , 9 9 9 9 9 G

x v

` r over elements x P S,ell and v P v pMv q such that the triplet Mv , v , rv pxq
G belongs to the W0 v -orbit represented by v . We now apply Lemmas 5.2.1 and 5.2.2. Together, they assert that if tf u is any global compatible family of functions (5.2.6), the sum ` rpN, G q tr Rdisc, pf q , (7.3.5) 9 9 r G PE sim pN q 9

added to (7.3.3) in case r 1, equals (7.3.2). We are following the earlier 9 9 convention of writing f9 or f9_ in place of f in case G equals G or G_ . The expression (7.3.5) is of course a nonnegative linear combination of irreducible 9 characters on the groups G pAq. In the case r 1, the expression (7.3.3) equals the analogue `` 1 9 r 9 (7.3.6) tr 1 RP _ pw1 , 9 , 1 q IP _ p 9 , f9_ q
8 1 1

9 r for F of the expression (5.4.11) obtained in the proof of Lemma 5.4.4. It too is a nonnegative linear combination of irreducible characters, in this case on 9 9 the group G_ pAq. We can therefore write the sum of (7.3.5) with (7.3.3) as 9 r r cG p q fG p q, G P E sim pN q, P pG q,
G

for nonnegative coecients cG p q. 9 r For any G P E ell pN q, let Dv be the duality operator on G . At this v point, we will actually be concerned with the absolute value p p D p q p q D p q , P pG q, of where dual (7.1.4) of the representation This extends to an involution p p p D p q D p q b ,v b ,v , b ,v ,
v v v v v Dv , v v is the v p v v v v v v v .

on

r pG q,

and an adjoint involution p p pD f qp q f pD q f pp b ,v q

v G G v G v

r on IpG q. Our sum of (7.3.5) and (7.3.3) can then be written as 9 p r r pG p q pDv fG qp q, c (7.3.7) G P E sim pN q, P pG q,
G

7.3. THE LOCAL INTERTWINING RELATION FOR

405

for nonnegative coecients p pG p q cG pDv q. c The earlier expression (7.3.4) takes the form p p pv , f9v qpDv f9v,G qpv q, (7.3.8) 9
9 v PT pGv q

where p pv , f9v q pv q pv , f9v q and pv q pv q, ` v Mv , v , rv pxq . The assertions of Lemmas 5.2.1 and 5.2.2 may thus be characterized as an identity between the expressions (7.3.7) and (7.3.8), for any compatible family of functions tf u. But it is easy to see from the denitions in 2.1 p and 3.4 that the family of operators tDv u denes a correspondence of comp patible families. To be precise, suppose that tf u and tf u are two families r q such that of functions on the spaces HpG p p Dv fG fG , 9 r G P E sim pN q.

p Then if tf u is a compatible family, the same is true of tf u. The identity between (7.3.7) and (7.3.8) therefore remains valid, with the functions p p p Dv fG replaced by fG , for any compatible family tf u. We can now apply Corollary 3.5.3. It tells us that the coecients pG p q c p v , f9v q all vanish. Since (7.3.2) equals (7.3.8), it vanishes, as asserted. and p 9 Corollary 3.5.3 likewise tells us that (7.3.6) vanishes, as do any of the summands in (7.3.5). Since (7.3.3) equals (7.3.6), it also vanishes. This is the second assertion of the lemma. The second lemma gives the nongeneric renements of Lemmas 5.3.1 9 r 9 r and 5.3.2, for parameters (7.2.9) in the subset F2 pGq of F. It will be used in the next section to establish what is left of Theorem 2.2.1. Lemma 7.3.2. Suppose that 9 9 pG, q, 9 9 r 9 9 r G P E sim pN q, P F2 pGq,

9 r is an analogue for the nongeneric family F of the pair given in either Lemmas 5.3.1 or 5.3.2. Then 9 9 9 9 9 9 9 (7.3.9) S G pf9q |pG, q| |S 9 |1 G pq f G pq,
9 disc,

r 9 for any function f9 P HpGq. 9 r Proof. This lemma is an analogue for F of Lemmas 5.4.5 and 5.4.6. It is stated with the implicit assumption of the condition of Lemma 5.3.1,

406

7. LOCAL NONTEMPERED REPRESENTATIONS

namely that the global theorems are valid for parameters in the complement 9 9 r of Fell pN q in FpN q. 9 The required identity (7.3.9) is the stable multiplicity formula for . We can assume that N is even and 1, since the formula is otherwise valid 9 by Corollary 5.1.3. It then follows from Lemma 5.1.4 that the left hand side minus the right hand side is a multiple of the linear form 9 f9 f9L pq, r 9 f9 P HpGq,

9 r 9 9 9 on HpGq obtained from the maximal Levi subgroup LpAq of GpAq. (We 9 9 are writing L and in place of L and here, to be consistent with the 9 surrounding notation.) We must show that vanishes. As we agreed at the end of the last section, we introduce the simple 9 9 r datum G` P E sim pN` q, for N` N1 ` N , and the parameter 9 9 9 ` 1 9 r 9 in F ell pG` q. We have then to modify the proofs of Lemmas 5.4.5 and 5.4.6 in the manner of the last lemma. This entails working with compatible families 9 r of functions (5.3.3) on the groups G P E sim pN` q, and the correspondence p on such objects under the local duality operators D`,v . We shall be brief. 9 9 9 We rst introduce the composite endoscopic datum G_ G1 G_ in 1 9 r r 9 E sim pN` q, with corresponding function f91 P HpG_ q in the compatible family 1 9 9 (5.3.3). We are including the case r 1 here, in which G1 G. It follows from the denitions that the supplementary summands in the expressions (5.3.4) and (5.3.21) of Lemmas 5.3.1 and 5.3.2 can be written 9 9 b` f91 ` p1 q and
1 2 9 L 9 9 f91 ` p q 9 L

respectively. We then apply Corollary 3.5.3 as in Lemmas 5.4.5 and 5.4.6, p in combination with the operators Dv . This leads to the conclusion that 9 9 9 9 the linear forms 1 and , attached to the maximal Levi subgroup 9 ` G1 L of G_ in the respective cases r 1 and r 1, vanish. 9 9 9 L 1 9 9 If r 1, the linear form 1 attached to G1 is nonzero, according to 9 vanishes in this case, as required. our induction hypothesis. Therefore 9 Suppose then that r 1. If vanishes, the linear form
9 9 9 9 9 9 9 f9pq f9G pq ` f9L pq f9L pq,

r 9 f9 P HpGq,

9 is induced from the Levi subgroup LpAq. This contradicts the rst assertion of Proposition 7.2.1(iii)(a), which is the analogue of the local generic condip tion (5.4.1)(a), as one sees easily from the properties of Dv . Therefore, it is 9 again the linear form that vanishes, as required.

7.3. THE LOCAL INTERTWINING RELATION FOR

407

We shall now establish Theorem 2.4.1 for nongeneric parameters r P pGq. To recall the general strategy, we go back to the discussion for generic parameters at the beginning of 6.4. As we noted there, the reduction of the global intertwining relation in 4.5 can be made in the context 9 r of the family F, dened now by (7.2.7). The reduction also carries over to the local intertwining relation of Theorem 2.4.1. In other words, the local r intertwining relation holds for a parameter P pN q, unless belongs to q for some G P E r ell pG sim pN q, or belongs to the local analogue for of one of the two exceptional cases (4.5.11) or (4.5.12). As before, we prove the three exceptional cases together. Again, it is sucient to establish the assertion of Theorem 2.4.1 for elements s and u that map to a point x in the set S,ell . r r In the rst case, we x G P Esim pN q with P ell pGq. The conditions are # (7.3.10) 21 2q q`1 r , ` ` S Op2, Cqq Op1, Cqrq , q 1,

with the requirement that the Weyl group W contain an element w in 9 r W,reg . Proposition 7.2.1 then gives us the global family F, together with 9 r2 9 9 9 a global pair pG, q such that P Fell pGq is of the form (7.2.8). The global pair satises the condition of Lemma 7.3.1. This lemma then gives us an identity (7.3.11)
xPS,ell

` 1 9 9 99 9 9 G pxq f9G p, s xq f9G p, xq 0, 9 9 9 9

9 r r 9 f P HpGq.

r The second and third cases apply to local parameters P pGq, for r G P Esim pN q, that are of the local forms of (4.5.11) and (4.5.12). In each 9 r of these two cases, Proposition 7.2.1 gives a global family F, together with 9 r 9 9 9 a global pair pG, q such that P FpGq has the corresponding global form (4.5.11) or (4.5.12). In both cases, Corollary 4.5.2 again tells us that (7.3.11) is valid as stated. Recall that if q r 1 in the original case (7.3.10), there r r is a second group G_ P Esim pN q such that P pG_ q. The corresponding 9 9 global pair pG_ , q then belongs to the second exceptional case (4.5.11), and the global formula (7.3.11) in this case amounts to the vanishing of (7.3.3) established in Lemma 7.3.1. We can therefore include it in our discussion of (4.5.11), as we did for the generic parameters treated in 6.4. We have thus to extract the remaining local intertwining relation
1 fG p, s xq fG p, xq,

r x P S,ell , f P HpGq,

408

7. LOCAL NONTEMPERED REPRESENTATIONS

from the global identity (7.3.11), when represents one of the three exceptional cases above. As in 6.4, we take f9 f9v
v

to be a decomposable function. The summand of x in (7.3.11) then decomposes into a dierence of products 9 1 9 9 9 (7.3.12) G pxq 9 f9v,G pv , s xv q f9v,G pv , xv q 9 9 9 9 9
v v

over all valuations v. The rst step is to remove the contributions to these products from valuations in the complement of the nite set SpU, V q. We shall establish a local lemma that can be applied to the places v P V , as well as those in the complement of SpU, V q. To simplify the notation, we shall formulate it in terms of the basic local pair pG, q over F , specialized by p r the requirement that for some P pGq. This of course implies that r G for the subset of representations F is p-adic. For any such , we write ,M r r P that lie in pM q, the set of irreducible constituents of the induced r representation IP pM q, for some M in the set G dened prior to (7.1.16).
M

p Lemma 7.3.3. Assume that F is p-adic, and that for some P r r r pGq. Then if f P HpGq is such that the function fG ppq of P is r G , we have supported on the subset ,M
1 fG p, s xq fG p, xq,

x P S .

Moreover, if is tamely ramied and quadratic, and f satises the stronger condition that # 1, if x , y 1, (7.3.13) fG ppq 0, otherwise, r for any P , then
1 fG p, s xq fG p, xq 1,

x P S .

Proof. Using (7.1.11) and the generic form of Theorem 2.4.1, we write
1 fG p, s xq pq pDfG q1 p, s xq

pq pDfG q p, s xq, as in 7.1. Our condition on f implies that the function pDfG q pq fG pDq pq fG ppq, r P ,

r r is supported on the subset G of . It then follows from the formula ,M r (7.1.13), together with the denition of G , that the function ,M pDfG q p, s u, M q, r M P M ,

7.3. THE LOCAL INTERTWINING RELATION FOR

409

r r is supported on the subset GM of M . (We recall that u is any element in the group N N whose image in S S equals x.) We can therefore write pq pDfG q p, s uq pDfG q p, s u, M q pq
r M PM

pq
r M PG
M

r M PG
M

pDG q p, s u, M q

fG p, u, M q p p fG p, u, M q

r M PM

fG p, uq, r since the identity (7.1.14) holds by denition for representations M P GM . We have established that
1 fG p, s xq fG p, uq fG p, xq,

as required. Suppose now that is tamely ramied and quadratic. It then follows from Lemma 7.1.2(b) that (7.3.13) is indeed stronger than the rst condition 1 on f . In particular, the functions fG p, s xq and fG p, xq are equal. To see that they are equal to 1, we have only to observe that 1 p fG p, s xq xx, y fG ppq 1,
r P

by (7.1.2), Lemma 7.1.1 and the fact that (7.1.8) holds by our induction hypothesis on M . This completes the proof of the lemma. p r Remark. Suppose that for an arbitrary parameter P pGq, as p in the lemma. Then for any complex valued function a on S , there is a r function f P HpGq such that ` r a x , y fG ppq, P . This follows by duality from the assertion (b) of Theorem 1.5.1 (which we have now proved). In particular, we can always choose f so that the condition (7.3.13) holds. We will need another lemma for the unramied places v, which will again be stated in terms of pG, q. Its proof does not rely on the induction hypothesis that M G.

410

7. LOCAL NONTEMPERED REPRESENTATIONS

Lemma 7.3.4. Suppose that pG, q is unramied over F , and that f is the characteristic function of a hyperspecial maximal compact subgroup of GpF q. Then f satises the condition (7.3.13) of Lemma 7.3.3. Proof. The condition that be unramied means that the restriction of to LF is an unramied unitary representation of WF (and in particular, is trivial on the SU p2q factor of LF ). This implies that is the dual of a r tamely ramied, quadratic parameter in bdd pGq. We shall rst establish that (7.3.14)
1 fG p, xq 1,

x P S .

1 We recall that fG p, xq equals f 1 p 1 q, for any pair pG1 , 1 q that maps to p, xq. The fundamental lemma for pG, G1 q asserts that f 1 is the image in r SpG1 q of the characteristic function of a hyperspecial maximal compact subgroup of G1 pF q. Since we can assume inductively that f 1 p 1 q 1 whenever x 1, it will be enough to show that 1 fG p, 1q f G pq 1.

r r r r Let f f 0 , where f 0 is the characteristic function of the standard maxi` r mal compact subgroup of GLpN, F q. The fundamental lemma for GpN q, G r r asserts that f G equals the image f G of f in HpGq. It follows that r r f G pq f G pq fN pq. But it is easy to see from the denition (2.2.1), and the properties of Whitr taker functionals for GLpN, F q discussed in 2.5, that fN pq equals 1. The identity (7.3.14) follows. We need to be a little careful here, as I was reminded by the referee. The proof of the two fundamental lemmas by Ngo and Waldspurger applies only to the case that the residual characteristic of F is large. However, it is easily extended to the general case by using the generic forms of the local theorems we have now proved (results which of course would have been unavailable without the original two fundamental lemmas). We shall sketch a separate argument, even though it would also be possible to appeal directly to the unramied assertion of Theorem 1.5.1(a) we established at the beginning of 6.6. If we combine the identity (2.2.6) of Theorem 2.2.1 with the last assertion of Theorem 1.5.1(a), keeping in mind the remarks in 2.5 and 6.1 on spherical functions, we see that # 1, if is unramied, G f pq 0, otherwise,

7.3. THE LOCAL INTERTWINING RELATION FOR

411

r r for any P bdd pGq. This characterizes the image of F in SpGq as a function r on bdd pGq. We also see more generally that # 1, if is unramied, f 1 p1 q 0, otherwise, if pG1 , 1 q is a preimage of a given pair p, xq. Since the pair pG1 , 1 q is unramied if and only if the same is true of , we obtain the formula # 1, if 1 is unramied, f 1 p1 q 0, otherwise, r for G1 P EpGq unramied and any 1 P bdd pG1 q. Therefore, f 1 is indeed r 1 q of the characteristic function of a hyperspecial subgroup the image in SpG of G1 pF q, as asserted by the fundamental lemma for pG, G1 q. To treat the twisted fundamental lemma, we use the formula # 1, if is unramied, r fN pq 0, otherwise, r which holds for any P bdd pN q. This is not hard to establish from the r unramied case of (i) from Corollary 2.5.2. It tells us that f G `equals the G of f in SpGq, as asserted by the fundamental lemma for GpN q, G . r r image f The two fundamental lemmas thus hold in general, and the identity (7.3.14) is therefore valid with no restriction on the residual characteristic of F . To nish the proof of the lemma, we take to be the tamely ramied, p r quadratic parameter in bdd pGq such that . We must show that f satises (7.3.13). To this end, we write f 1 p, xq f 1 p 1 q xs x, y fG p q,
r P

by (7.1.2). It then follows from (7.3.14) that # 1, if x , y 1, fG p q 0, otherwise, r for any P . The required condition (7.3.13) is thus a consequence of Lemma 7.1.2(a), which tells us that if x , y 1. p We return to our proof of the local intertwining relation, in which r P pGq falls into one of the three exceptional cases (7.3.10), (4.5.11) or (4.5.12). The global identity (7.3.11) asserts that the sum over x P S,ell of (7.3.12) vanishes. We will use this identity to show that the two factors for v u in the two products in (7.3.12) are equal. The process will be considerably easier than its counterpart from 6.4, thanks to Lemma 7.3.3. 9 If v is a valuation of F in the complement of U , Proposition 7.2.1 as9 v v , for a tamely ramied, quadratic, generic parameter serts that

412

7. LOCAL NONTEMPERED REPRESENTATIONS

r 9 v P pGv q. Suppose rst that v lies in the complement of both U and r 9 V . In this case, we choose the function f9v P HpGv q so that it satises the 9 analogue for v of the condition (7.3.13). In the special case that v is 9v is the characteristic function of a unramied, we can also assume that f 9 9 hyperspecial maximal compact subgroup of GpFv q, by Lemma 7.3.4. It then follows from Lemma 7.3.3 that the corresponding two factors in (7.3.12) are each equal to 1. We may therefore remove them from the products. We have established that the two products in (7.3.12) can each be taken over the nite set of valuations v P SpU, V q. If v belongs to V , we choose f9v so that it satises the rst condition of Lemma 7.3.3, namely that the r r function f9v,G ppv q of v P v is supported on the subset Gv ,M . It then 9 follows from the lemma that 9 9 9 9 x P S,ell . f91 pv , s 9 xv q f9 9 pv , xv q,
9 v,G v,G

The identity (7.3.11) reduces to ` 1 1 9 9 9 9 9 G pxq f9U,G pU , s xq f9U,G pU , xq f9V pV , s xq, 9 9 9 9 9 9 9 9

xPS,ell 9

where
1 9 f9V pV , s xq 9 9

vPV

1 9 f9v,G pv , s xv q, 9 9 9

9 9 and f9U,G pU , xq are dened by corresponding products 9 over v P U . It is a consequence of Proposition 7.2.1(iii)(a) that as f9V varies under the given constraints, the functions 9 9 x f91 pV , s 9 xq, x P S,ell , while
V

1 9 f9U,G pU , s xq 9 9 9

span the space of all functions on S . It follows that (7.3.15)

1 9 9 9 f9U,G pU , s xq fU,G pU , xq, 9 9 9 9

x P S,ell ,

r 9 for any function f9U P HpGU q. It remains to separate the contribution of u to each side of (7.3.15) from u that of its complement S8 in U . To do so, we need only recall how we chose 9 the global eld F prior to Proposition 7.2.1. 9 is any imaginary quadratic eld. Then U consists of the If F C, F one valuation u, and (7.3.15) is simply the required identity (2.4.7) for the 9 local parameter u . If F R, and the simple generic constituents ( (7.3.16) i : mi 2, 1 i r 9 of are in relative general position, F equals Q. Again U consists of the 9 one valuation u, and (7.3.15) is the required identity for u . If F R, but the constituents (7.3.16) are not in relative general position, or if F 9 is p-adic, we can take F to be a totally real eld with several archimedean 9 places. In these two cases, the corresponding global parameter was chosen

7.4. LOCAL PACKETS FOR COMPOSITE AND SIMPLE

413

v 9 in the proof of Proposition 7.2.1 so that for any v P S8 , the analogues for v of the generic constituents (7.3.16) are in general position. It follows from the case treated above that 1 9 9 9 f9v pv , s xv q f9v pv , xv q, 9 9

r 9 f9v P HpGv q,

u for any v P S8 and x P S,ell . These linear forms are all nonzero, as one can 9 see by relating the left hand side to GLpN, Fv q, according to the denitions (2.4.6), (2.2.3) and (2.2.1). We may therefore remove their contributions to (7.3.15). This leaves only the contribution of the complementary valuation 9 v u. The formula (7.3.15) thus reduces to the required identity for u in these remaining two cases. We have now established the local intertwining relation for parameters r P pGq in any of the three outstanding cases (7.3.10), (4.5.11) or (4.5.12). r This completes the proof of Theorem 2.4.1 for any parameter P pGq over F. We also obtain the following corollary, which is proved in the same way as its generic analogue Corollary 6.4.5.

Corollary 7.3.5. The canonical self-intertwining operator attached to 0 r r any P pGq, w P W and M P M satises r RP pw, M , q 1. 7.4. Local packets for composite and simple We are at last ready to complete our proof of the local theorems. They consist of ve local theorems stated in Chapters 1 and 2. Having established everything for generic local parameters in Chapter 6 (including the generic supplement Theorem 6.1.1), we are working now with nongeneric local parameters . We dealt with the local intertwining relation of Theorem 2.4.1 in the last section. We shall sketch a proof of its supplement Theorem 2.4.4 at the end of this section. Since the nongeneric form of Theorem 1.5.1 is a special case of Theorem 2.2.1, we have then only to establish this last theorem, and its supplement Theorem 2.2.4. The general assertion (a) of Theorem 2.2.1 is now known. It was reduced in Lemma 2.2.2 to the generic case, which we resolved in Lemma 6.6.3. The assertion (b) of this theorem is of course the crux of the matter. For any pair r r pG, q, G P Esim pN q, P pGq, r over the local eld F , it postulates the existence of a packet , together r with a pairing on S that satises (2.2.6). Before we begin its proof, however, we shall derive a consequence of Theorem 2.2.1. We shall establish r a proposition that provides some elementary constituents of the packet . A special case of the proposition will then be used in the general proof of Theorem 2.2.1.

414

7. LOCAL NONTEMPERED REPRESENTATIONS

r We should rst recall that if P pGq is a general Langlands parameter r for G, the L-packet is already dened. It consists of Langlands quotients r of elements in a corresponding packet P of standard representations. r The packet P in turn is dened by analytic continuation from the case that is bounded, as is the pairing xx, y xx, y, r x P S , P P .

However, the endoscopic identity (2.2.6) will be valid for general only if r r is replaced by the packet P of standard representations. Its failure for r of irreducible representations is of course one of the reasons the packet we have had to introduce other parameters . Suppose for a moment that Theorem 2.2.1 is valid for the given pair pG, q, and that x is an element in S . We can then choose a semisimple element s P S with image x in S , such that if pG1 , 1 q is the preimage of the pair p, sq, the endoscopic datum G1 P EpGq is elliptic. The endoscopic identity (2.2.6) of Theorem 2.2.1(b) is r f 1 p 1 q xs x, y fG pq, f P HpGq.
r P

r r Recall that is a nite unit pGq-packet, which is to say that it bres over r r r the set unit pGq. Let us write pq for the bre in of any element r P unit pGq. The endoscopic identity can then be written r r xs x, y fG pq, f P HpGq. (7.4.1) f 1 p 1 q
r Punit pGq r r P pq

We shall combine it with the denition in Theorem 2.2.1(a) to give a formal r construction of the packet . 1 , 1 q, we write Given pG G1 G1 G 1 , 1 2 and
1 1 1 1 2 , 1 r i P pG1 q, i 1, 2, i 1 1 where N N1 ` N2 . We can then write f 1 p 1 q np 1 , 1 q f 1 p1 q, r 1 PpGq

r G1 P Esim pNi1 q, i

for generic parameters 1 1 1 , 1 2 r in pG1 q, and integers r r 1 1 r 1 2 np 1 , 1 q np1 , 1 q np2 , 1 q, r 1 P pG1 q, i i

7.4. LOCAL PACKETS FOR COMPOSITE AND SIMPLE

415

by applying (2.2.12) separately to G1 and G1 . For any such 1 , pG1 , 1 q maps 1 2 to a pair ` r p1 q, sp1 q , p1 q P pGq, sp1 q P Sp1 q . The analogue of (2.2.6) for standard representations is then xsp1 q, y fG pq. f 1 p1 q
r PPp1 q

Lastly, we have the decomposition fG pq

r PpGq

mp, q fG pq

r of a standard representation P Pp1 q into irreducible representations , for nonnegative integers mp, q. Combining the three decompositions, we obtain a second endoscopic identity r (7.4.2) f 1 p 1 q p 1 , q fG pq, f P HpGq,
r PpGq

where (7.4.3) p 1 , q

r np 1 , 1 q xsp1 q, y mp, q,

r P pGq.

r r 1 PpG1 q PPp1 q

It then follows from (7.4.1) that (7.4.4) xs x, y p 1 , q, r

P pq r r

r P pGq.

r We recall that there is a generic parameter P pGq attached to . It is dened by 1 |w| 2 0 , pwq w, w P LF , 1 2 0 |w| where |w| is the extension to LF of the absolute value on WF . The embedding of the centralizer S into S factors to a surjective homomorphism p from S to S . This is dual to an injection of the character group S p r into S . Our proposition asserts that the representations in take on an r endoscopic interpretation as elements in the larger packet . Proposition 7.4.1. Assume that Theorem 2.2.1 is valid for the pair r pG, q. Then there is an injection from into the set of multiplicity free

416

7. LOCAL NONTEMPERED REPRESENTATIONS

r elements in the packet such that the diagram r

X

>

r
_

p S

>

p S

r is commutative. In particular, the elements in the nontempered packet are unitary. Proof. As we observed in 2.2 and 3.5, an irreducible representation of GLpN, F q comes with a linear form in the closure pa q` of the B associated maximal dual chamber. It is a measure of the failure of to be tempered. Given parameters P pN q and P pN q for GLpN q, we have been writing and . In particular, we have linear r r forms and for parameters and in the subsets pGq and pGq of pN q and pN q. The maximal closed chamber pa 0 q` for G embeds P canonically in the chamber pa q` for GLpN q. With this understanding, it B r is clear that the linear form attached to a representation P equals r r , for the parameter P pGq such that contains . The proposition will be a consequence of the identity (7.4.4) applied to r a xed element P pGq, together with elementary properties of the factors in the summands of (7.4.3). It might be a little simpler to argue in terms of the norm on pa q` , as we did in the proof of Lemma 3.5.2, rather than B the underlying partial order. The factor np 1 , 1 q in (2.4.3) represents the r r general coecient in the expansion of a stable linear form 1 on HpG1 q in 1 . If it is nonzero, we have terms of standard stable linear forms }1 } }1 } } }, with equality if and only if 1 1 . We also have p1 q 1 r for any parameter 1 P pG1 q, and , r r P P , P ,

r for any parameter P pGq. The factor mp, q in (7.4.3) is the general coecient in the expansion of a standard representation of GpF q in terms of irreducible representations . If it is nonzero, we have } } } }, with equality if and only if is the Langlands quotient of . Taken together, these conditions place a similar restriction on the coecient p 1 , q dened by (7.4.3). Our conclusion is that if p 1 , q is nonzero, then } } }1 },

7.4. LOCAL PACKETS FOR COMPOSITE AND SIMPLE

417

r with equality if and only if lies in , for parameters p1 q and 1 1 . We note that if 1 does equal 1 , then p1 q and sp1 q s. In this case, the original pair p, sq is transformed to p , sq under the composition ` p, sq pG1 , 1 q pG1 , 1 q p1 q, sp1 q p , sq. r Suppose now that belongs to . The double sum in (7.4.3) then reduces to a sum over the single pair p1 , q p1 , q, where is the r representation in P corresponding to . As the leading coecients in their respective expansions, np 1 , 1 q and mp , q both equal 1. We see r also that xsp1 q, y xs, y xx, y. The formula (7.4.3) thus reduces to p 1 , q xx, y. It then follows from (7.4.4) that
P pq r r

r xs x, y xx, y.

The point x P S is arbitrary, and has been identied with its image in S on the right hand side of the formula. Since the image of s in S is trivial, the formula can be written xx, y xx, y, r x P S , (7.4.5)
r r P pq

after translation of x by s s1 . If we set x 1 in (7.4.5), we see that the index of summation is trivial. r r In other words, the bre pq of in consists of one element, which we simply identify with . This is the rst assertion of the proposition. The other assertions follow from the resulting simplication of (7.4.5), and the r fact that the representations in are unitary. We can now go back to our main topic, the proof of Theorem 2.2.1. Our task is to establish assertion (b) of the theorem for the given pair pG, q. r If lies in the complement of 2 pGq, the assertion follows from the local intertwining relation we have now established, or rather its embodiment in Proposition 2.4.3. We have therefore only to consider the case that r belongs to 2 pGq. We take r r (7.4.6) 1 r , i P sim pGi q, Gi P Esim pNi q, r to be a xed parameter in 2 pGq. Applying Proposition 7.2.1, we obtain 9 r 9 an endoscopic datum G P Esim pN q over the global eld F and a global 9 P F2 pGq of the corresponding form (7.2.9). The pair pG, q is r 9 9 9 parameter

418

7. LOCAL NONTEMPERED REPRESENTATIONS

as in Lemma 7.3.2. It follows from this lemma that the stable multiplicity formula (7.3.9) holds for the pair. 9 r By the global induction hypothesis on F, (7.3.9) applies more generally 9 to the groups G1 that occur in 9 9 p1 pG, G1 q Sdisc, pf91 q, (7.4.7) Idisc, pf9q 9 9
9 r 9 G1 PEell pGq

9 the endoscopic decomposition (4.1.2) of the -component of the discrete part of the trace formula. The specialization of (7.3.9) to any linear form on the right hand side was expressed in Corollary 4.1.3. Following the conditional proof of the spectral multiplicity formula in Lemma 4.7.1, as we did in the discussion of (6.6.4), we obtain a further decomposition 9 G 9 9 9 9 Idisc, pf9q |pG, q| |S |1 1 p 1 q f91 p 1 q, 9 9
xPS 9 9

9 9 9 9 where pG1 , 1 q maps to the pair p, xq. Continuing as in the relevant passage from 6.6, we turn next to the 9 spectral decomposition (4.1.1) of Idisc, pf9q. Since the parameter belongs to 9 r 9 9 9 2 pGq, it cannot factor through any proper Levi subgroup M of G. Applying 9 of M as in the discussion 9 our global induction hypothesis to the factor G 9 of 6.6, we see that the summand of M in (4.1.1) vanishes. The left hand side of (7.4.7) therefore equals ` G 9 9 G Idisc, pf9q tr Rdisc pf9q n pG q f9G pG q, 9 9 9 9 9
9 G

9 9 9 9 where G ranges over the set pGq of irreducible representations of GpAq, r 9 and n pG q are nonnegative integers. Since f9 lies in HpGq, f9G pG q depends 9 9 9 9 r Gq of orbits in pGq under the 9 9 9 9 only on the image of G in the set p restricted direct product r 9 r 9 OutN pGq OutN pGv q.
v

We can therefore write

9 G 9 9 Idisc, pf9q |pG, q| 9

9 r 9 PpGq

n pq f9G pq, 9 9 9 9

for a modied coecient 9 9 n pq |pG, q|1 9 9

9 9 G PpG,q 9

n pG q, 9 9

9 9 9 in which pG, q is the preimage of in pGq, as in (6.6.5). 9

7.4. LOCAL PACKETS FOR COMPOSITE AND SIMPLE

419

We have converted (7.4.7) to an identity 9 9 n pq f9G pq |S |1 1 p 1 q f91 p 1 q, (7.4.8) 9 9 9 9 9

PpGq 9 r 9 xPS 9 9

r 9 f9 P HpGq,

9 9 9 9 where pG1 , 1 q maps to p, xq. This formula is the analogue of the identity (6.6.6) in 6.6. We shall apply it to a product f9 f9V f9U f9V,U , relative to the decomposition 9 9 9 9 9 9 9 9 9 9 9 GpAq GV GU GV,U GpFV q GpFU q GpAV,U q. Recall that the set U S8 puq is composed of places at which we have little information, the analogue of the singleton tuu in 6.6. Recall also that the u set V of p-adic places here plays the role of the set S8 of archimedean places in 6.6. It is what we will use to extract information about the places in U . We have been trying to follow the discussion of 6.6 as closely as possible, in order to clarify the structure of the general argument. At this point we come to a minor bifurcation. Our concern is the linear form
1 91 f 1 p 1 q f9u pu q,

f f9u , x P S ,

in which pG1 , 1 q maps to p, xq. We want to express it in terms of irreducible characters. In 6.6, we proceeded directly. We decomposed the linear form (with in place of ) into a linear combination of irreducible characters in f , with formal coecients c,x pq in x that had then to be determined. It is convenient here to approach the question from the opposite direction. We shall decompose f 1 p 1 q into a linear combination of irreducible characters in x, with formal coecients fG pq in f that have yet to be understood. In other words, we shall follow the general denition (7.1.2). In fact, we will apply it to each of the valuations v of U . For each v P U , (7.1.2) gives a sum 1 91 r 9 f9v pv q xs xv , v y f9v,G pv q, 9 9 xv P Sv , f9v P HpGv q, 9 9 9 9
v 9

r9 r 9 9 over v in the nite packet v of linear forms on HpGv q. We write 1 91 (7.4.9) pf9U qpU q xs x, U y pf9U qG pU q, 9 9 9 9 9
U 9

where U ranges over the linear forms in the packet 9 ! ) r9 9 9 9 U v : v P v

vPU

of 9 U
vPU

9 v ,

420

7. LOCAL NONTEMPERED REPRESENTATIONS

and 9 9 xx, U y
vPU

9 9 xxv , v y.

We temporarily follow the same convention for the supplementary valuations v R U . We can then write ` 9 (7.4.10) pf9V,U q1 p V,U q1 xs x, V,U y pf9V,U qG p V,U q, 9 9 9 9 9
U,V 9

where V,U ranges over elements in the packet 9 ) ! r9 9 9 v : v P v , x , v y 1 for almost all v 9 9 V,U
vRV YU

of 9 V,U and 9 9 xx, V,U y We can also write (7.4.11)

1 9 pf9V qp1 q V

vRV YU

9 v ,

vRV YU

9 9 xxv , v y.

V 9

xs x, V y pf9V qG pV q, 9 9 9 9 9

9 where V ranges over elements in the packet ) ! r9 9 9 v : v P v 9 V

vPV

of 9 V and 9 9 xx, V y
vPV

vPV

9 v ,

9 9 xxv , v y.

We will presently specialize the supplementary functions f9V and f9V,U , as in the argument of 6.6. We rst substitute the formulas (7.4.9), (7.4.10) and (7.4.11) into the three factors of our linear form ` 9 9 9 9 f91 p 1 q f91 p 1 q f91 p 1 q pf9V,U q1 p V,U q1
V V U U

9 in (7.4.8). According to Lemma 4.4.1, the coecient 1 p 1 q in (7.4.8) equals 9 G on S at s x, with p, xq being the usual 9 9 the value of the character 9 9 9 9 9 9 9 image of pG1 , 1 q. The right hand side of (7.4.8) becomes |S |1 ps xq xs x, y f9G pq, 9 9 9 9 9 9 9 9 9
xPS 9

7.4. LOCAL PACKETS FOR COMPOSITE AND SIMPLE

421

an expression we write as (7.4.12) |S |1 9

xPS 9

pxq xx, y f9G pq, 9 9 9 9 9 9

where the inner sums are over products 9 V b U b V,U 9 9 9 9 9 9 of elements in the packets of V , U and V,U , and 9 9 9 xx, y f9 9 pq
G

equals 9 xx, V y f9V,G pV q xx, U y f9U,G pU q xx, V,U y f9V,U p V,U q, 9 9 9 9 9 9 9 9 9 9 9

G

9 while x is the isomorphic image of x in S , as usual. Bear in mind that 9 V , U and V,U at this point are still just linear forms. However, the two 9 9 9 supplementary indices V and V,U both range over linearly independent 9 9 sets of irreducible, signed characters. This follows from assertions (ii) and 9 (iii)(a) of Proposition 7.2.1, which tell us that the localization v at any v R U is dual to a tamely ramied, quadratic, generic parameter. We shall 9 9 now x elements V and V,U for which the sign is 1. p Suppose that P S is a xed character on the 2-group S S . 9 9 We apply condition (iii)(a) of Proposition 7.2.1 to the localization V of 9 9 . (Bear in mind that M G here, since belongs to 2 pGq.) The 9 r M condition implies that we can x an element V, in the subset 9 9 r9 r G G
9 V 9 v vPV

9 of the packet of V such that the character x , V, y on S equals the product 9 9 1 1 9 . Away from V and U , we take the element 9 9 v p1q V,U p1q
vRV YU

in the subset

r9 GV,U 9

vRV YU

r9 Gv 9
v

9 9 of the packet of V,U such that for any v, the character x , v p1qy on Sv is 9 1. We are relying here on Proposition 7.2.1(ii) and Lemma 7.1.2(a), which r9 r9 9 together imply that v p1q lies in the subset Gv of v . It follows from 9
v

the denition (7.1.9) that v, and V,U p1q are both (orbits of) irreducible 9 9 representations. We can now specialize the functions f9V and f9V,U . We have already noted that the two supplementary indices V and V,U in (7.4.12) are over 9 9 linearly independent sets of irreducible signed characters. We choose the 9 corresponding functions f9V,G pV q and f9V,U p V,U q, which represent factors of 9 9 9
G

422

7. LOCAL NONTEMPERED REPRESENTATIONS

the summands on the right hand side of (7.4.12), so that they are supported 9 9 9 9 at V V, and V,U V,U p1q. The original formula (7.4.8) then reduces to an identity 9 9 9 xx, U y pxq1 f9U,G pU q, n p, U q f9U,G pU q |S |1 9 9 9 9 9 9
U 9 U 9 xPS

9 r 9 9 9 where U is summed over pGU q, U is summed over the packet of U , and ` 9 9 9 (7.4.13) n p, U q n V, b U b V,U p1q . 9 9 This is the analogue of the formula (6.6.10) from the generic case in 6.6. 9 9 Since the groups Sv are all abelian, the function xx, U y of x P S attached 9 r 9 pq is the set of U in the packet 9 here to any U is a linear character. If 9
U

9 r9 U of U such that this character equals , we conclude that 9 (7.4.14) n p, U q f9U,G pU q f9U,G pU q, 9 9 9 9
r 9 U PpGU q 9 r9 U P pq 9
U

p r 9 for any P S and f9U P HpGU q. This is the analogue of (6.6.12) r 9 Lemma 7.4.2. For any U P pGU q, n p, U q is a nonnegative integer. 9 9 Proof. For this lemma, we need a slight embellishment of the proof of its generic counterpart Lemma 6.6.4(a). We recall from (7.4.13) that 9 9 9 n pG q, n p, U q |pG, q|1 9 9
9 9 G PpG,q 9

where 9 V, b U b U,V p1q. 9 9 9 Since the summand n pG q is a nonnegative integer, the only possible source 9 9 of diculty is the inverse of the order 9 9 |pG, q| |pG, q| p of the preimage of in pGq. We can assume that G is orthogonal and that the degrees Ni mi ni of the components i i b i of are all even, since the order |pG, q| would otherwise be 1, leaving us with nothing to prove. Suppose that one of the numbers mi is odd. Then ni is even, and the corresponding representation i of SU p2q is symplectic. Since the representation i of LF is orthogonal in this case, i would then be symplectic. r This contradicts the fact that P 2 pGq is orthogonal. We have shown that the degrees mi of the generic constituents i of i are all even. Consider the construction of the corresponding global generic constituents i from Proposition 7.2.1. It has the property that for any 9

7.4. LOCAL PACKETS FOR COMPOSITE AND SIMPLE

423

v P V , the completion i,v is a direct sum of 2-dimensional representations 9 of WFv . Moreover, for any n, the representation 9 9 9v i,v n
ti: ni nu

of LFv is multiplicity free. The completion 9 n 9 9 9 v pi,v b i q pv b n q

i n

9 p r 9 r 9 of therefore belongs to 2 pGv q, and is of the form v , for some v P 2 pGv q. 9 v is even orthogonal, there are two elements in each Since the dual group of G 9 9 9 of the sets pGv , v q and pGv , v q. It then follows from Corollaries 6.6.6 r r 9 and 6.7.3 that any element v P v , regarded as an OutN pGv q-orbit of irreducible representations, contains two representations. The same is therefore r9 true for any element v P v by duality, and hence for any component v 9 9 r N pGq acts freely 9 9 9 of the factor V, of at V . This implies that the group Out on the indices of summation G above. Since the original multiplicity npG q 9 9 9 -automorphism of G, the sum itself is 9 is invariant under the action of any F 9 9 an even integer, and is therefore divisible by the order 2 of pG, q. This completes the proof of the lemma. It remains to separate the contribution of u to each side of (7.4.14) from u that of its complement S8 in U . As in our treatment of (7.3.15) in the last 9 section, we need to recall how we chose the global eld F . If F C, or F R and the simple generic constituents (7.4.15) ti : mi 2, 1 i ru of are in relative general position, U consists of the one valuation u. In these cases, there is nothing to do. In fact, in dealing with the remaining cases, we can assume inductively that the theorem has been established for pG, q in case U tuu. This is where we will use Proposition 7.4.1. The remaining cases are when F is real but the constituents (7.4.15) of u are not in relative general position, or F is p-adic. Then S8 is not empty. 9 v of the constituents It consists of real places v such that the analogues for (7.4.15) are in general position. By our induction assumption, we can apply 9 9 Proposition 7.4.1 to the associated pairs pGv , v q. We set 8 9u v , 9
u vPS8

r 9 where for each v, v is the element in the packet such that the character 9v x , v y on Sv is trivial. The proposition asserts that v belongs to the packet 9 9 9 r9 v , and has multiplicity 1. We set u f98 f9v ,
u vPS8

424

7. LOCAL NONTEMPERED REPRESENTATIONS

r 9 where for each v, f9v is a function in HpGv q such that # 1, if v v , 91 9 1 9 f9v,G pv q 9 0, otherwise, r9 for any element v in v . We shall apply the identity (7.4.14) to the 91 product u f9U f f98 , r r 9 in which f is a variable function in the space HpGq HpGu q. There will be only one term on the right hand side of (7.4.14) at the chosen function f9U . To describe it, let pq be the element in the packet r r9 u with x , pqy . 9 It then follows from the denitions, and the fact that each v above has r9 multiplicity 1 in v , that the right hand side of (7.4.14) reduces simply to ` fG pq . The left hand side of (7.4.14) reduces to a nite sum of irreducible unitary characters in f , with nonnegative integral multiplicities 9u n p, q n p, 8 b q, r P unit pGq. r f P HpGq.

The identity therefore takes the form ` (7.4.16) n p, q fG pq fG pq ,

r Punit pGq

Observe that in the earlier cases in which U contained the one element u, (7.4.16) is actually identical to (7.4.14). Our global argument has thus led to a local identity for pG, q that holds in all cases. r Recall that pq is the element in the provisional packet attached to r the character on S . It was dened (7.1.2) only as a linear form on HpGq. r N pGqHowever, we see from (7.4.16) that it is actually a nite sum of (Out orbits of) irreducible characters. We have therefore obtained the following proposition. r Proposition 7.4.3. For any P 2 pGq, dene r r p pq, P S ,
p PS

r r where pq is the disjoint union over P unit pGq of multisets consisting of np, q-copies of . Then f 1 p 1 q ps xq fG pq xs x, y fG pq,
p r PS P pq r P

r for any f P HpGq and x P S , where xs x, y ps xq, r P pq,

7.4. LOCAL PACKETS FOR COMPOSITE AND SIMPLE

425

and as usual, pG1 , 1 q maps to the pair p, xq. We have now established the endoscopic identity (2.2.6). More prer p cisely, we have shown that the provisional packet over S determines r unit pGq-packet that satises the conditions of Theorem 2.2.1. This r the completes the proof of Theorem 2.2.1 and Theorem 1.5.1, for the pair pG, q over F . In the last two sections, we have established Theorems 2.2.1 and 2.4.1 in general. These are the essential local theorems. We also know that Theorem 1.5.1 is valid. Any of its assertions that are not subsumed in Theorem 2.2.1 follow either from Lemma 7.3.4, or Corollary 6.7.5 and the general classication of tempered reprsentations established in 6.7. This leaves only the twisted local supplements, Theorems 2.2.4 and 2.4.4. They r r r apply to parameters in the subset pGq of pGq, where G is the bitorsor rsim pN q with N even. attached to an orthogonal group G P E In 6.8, we sketched a proof of the two supplementary theorems for r r parameters in the subset bdd pGq of pGq. For any such , the dual p lies in pGq. We shall rst describe the analogues for G r r parameter of the objects in Lemma 7.1.2. r For any P pGq, there is a canonical mapping , r r P , r r where is an extension of to GpF q and is a GpF q-invariant linear form r r on HpGq, such that r r r r r r r f 1 p 1 q xs x, y f r p q, f P HpGq.
G P

p r This is the analogue for G of the case of (7.1.2) treated prior to the r statement of Lemma 7.1.1. In particular, pG1 , 1 q is the preimage of a pair r . We dene r r p, xq attached to an element x P S (7.4.17)
r p r G tp P : u, r

p r where is the preimage of in , and is an extension of the reprep p r q. It is clear that is uniquely determined by , and r sentation to GpF p r r r r hence by the extension of . This is the analogue for G of the set G r r in Lemma 7.1.2(a). If lies in the complement of 2 pGq in pGq, we can r choose a proper Levi subgroup M of G, with parameters M in 2 pM, q and r 2 pM, q as in Lemma 7.1.2(b). The analogue for G of the associated r M P r r set G is the set G of representations
M M

M , p

r M P M , r r u P N , f P HpGq, r r

r in the packet M such that the analogue (7.4.18) r r rr r r p pqpDfG qp, s u, M q fG p, u, M q,

426

7. LOCAL NONTEMPERED REPRESENTATIONS

r r of the identity (7.1.14) holds. The objects pq and D are taken with respect r which is to say that they are dened in terms of Levi subsets of G r to G, rather than Levi subgroups of G, according to the remarks prior to Lemma 7.1.1. p r Suppose that in addition to being of the form , the parameter P pGq r G in S generates S p p is tamely ramied and quadratic. Then the image of and contains the trivial character 1, while if M is proper in G, the image r p p of GM in SM generates SM and contains the trivial character 1. This r assertion is the variant of Lemma 7.1.2 for G mentioned at the end of 7.1. We shall prove it in [A25], along with the lemma itself. r Given the analogue for G of Lemma 7.1.2, we can establish the suppler mentary Theorems 2.2.4 and 2.4.4, for local parameters P pGq. In each case, we use Proposition 7.2.1 to construct a family (7.2.7) over the global 9 9 9 9 eld F . If pG, q is the global pair attached to pG, q, lies in the subset 9 r r 9 pGq of pGq. We have then simply to follow the steps of the proofs of the corresponding Theorems 2.2.1 and 2.4.1. We shall spare the details, since r r they contain no new arguments. If belongs to the subset 2 pGq of pGq, for example, the reduction of Theorem 2.2.4 to the proof of Theorem 2.2.1 in this section is essentially the same as the generic case treated at r r the beginning of 6.8. If belongs to the complement of 2 pGq in ell pGq, 9 r we rst observe that the analogue for G of the expression (5.2.8) of Lemma 5.2.1 vanishes. As in the generic case (6.8.6), this follows from the fact that 9 r the analogue for G of the other expression (5.2.7) in Lemma 5.2.1 vanishes. The proof of Theorem 2.4.4 for can then be transcribed from the proof of Theorem 2.4.1 in 7.3. With these brief remarks, we have nished our discussion of Theorems 9 r 2.2.4 and 2.4.4. We still have the global induction hypothesis for F, the nongeneric family (7.2.7). To resolve it, we must persuade ourselves that 9 9 r the relevant global theorems are valid for parameters in the subset FpN q 9 r of F. We shall again be brief. The stable multiplicity formula of Theorem 4.1.2 follows from the general reductions of Proposition 4.5.1 and the proofs of Lemmas 7.3.1 and 7.3.2. The actual multiplicity formula is the assertion 9 r 9 of Theorem 1.5.2. If lies in F2 pGq, it follows from the multiplicity formula (7.4.8) (and the local results we have now proved), or alternatively, a direct 9 appeal to the relevant variant of Lemma 6.7.1. If lies in the complement 9 r 9 r of F2 pGq in FpN q, it follows as in generic case from (4.4.12), (4.5.5), and the stable multiplicity formula. Theorem 1.5.3(a) is relevant only to generic parameters. Theorem 1.5.3(b) is also an assertion for generic parameters. However, its role in the induction argument is really for nongeneric parameters, insofar as they occur in proof of Lemma 4.3.1 from 4.6. The assertion

7.5. SOME REMARKS ON CHARACTERS

427

follows from an application of Lemma 5.1.6 to the nongeneric family (7.2.7). Finally, we have the supplementary global Theorem 4.2.2, for the relevant 9 r parameters in FpN q. Its main assertion (b) is implicit in the proof of the supplementary local Theorem 2.2.4. We passed over this proof as a variant of the generic case treated in 6.8. However, we will establish both assertions of Theorem 2.2.4, for general global parameters, in the rst section of the next chapter. We have now resolved the various induction hypotheses for general local parameters . With the end of the induction argument, we have established all of the local theorems. 7.5. Some remarks on characters We conclude this chapter by drawing attention to two character formulas for GLpN q. The rst applies to the case F R. It is the formula of Adams and Huang for a Speh character on GLpN q. The second is a renement by Tadic of results of Zelevinsky that are implicit in the discussion from 2.2 of p-adic innitesimal characters on GLpN q. In each case, the formula is an explicit expansion of an irreducible character on GLpN, F q in terms of standard characters. We will need the formulas in the next chapter. They will be used in 8.2 in the critical nal stages of the global classication. The formulas might also be of interest in their own right. They apply to untwisted analogues r of the integers np 1 , 1 q in the double sum (7.4.3), which we recall gives the denition of the coecients p 1 , q that determine the packets r , r r P pGq, G P Esim pN q.

We shall see that the two formulas are parallel. They hint at properties of r the packets that are common to the real and p-adic cases. Suppose that N mn, and that is an irreducible, tempered character on GLpm, F q. Then corresponds to an irreducible, unitary, m-dimensional representation of LF under the Langlands correspondence for GLpmq. Let us write ` ` ` r (7.5.1) n pf q tr pf q tr pf q , f P H GLpN q , b n ,

for the irreducible character on GLpN, F q corresponding to and the irreducible, n-dimensional representation n of SU p2q. We will also consider collections of pairs pmi , i q, 1 i r, where N m1 ` ` mr , and i is a virtual character on GLpmi , F q. In other words, i is a nite, integral, linear combination of irreducible characters on GLpmi , F q. Let us write (7.5.2) 1 r

428

7. LOCAL NONTEMPERED REPRESENTATIONS

for the virtual character on GLpN, F q induced from the virtual character 1 r on the Levi subgroup GLpm1 , F q GLpmr , F q. If the constituents i are irreducible, (7.5.2) is of course a standard character. The formulas in [AH] and [Tad2] express a character (7.5.1) explicitly in terms of virtual characters (7.5.2). Suppose rst that F R. Consider an irreducible representation of WF of dimension 2. It is then induced from a character z pz{|z|qk |z| z z ,
1

z P C , k P N, P C,

1 of the subgroup C of WF , where 2 pk ` q and 1 pk ` q, and 2

|z| pzzq 2 is the analytic absolute value. This representation of WF , which we denote by pk, q, parametrizes a representation in the relative discrete series of GLp2, F q, whose character we shall denote by pk, q. Following [AH, 5.10], we dene pk, q by coherent continuation for all k P Z. For xed , ( pk, q : k P Z is then a coherent family of virtual characters based at [AH, 4]. In concrete terms, pk, q can be dened by the natural continuation in k from N to Z of Harish-Chandras explicit formula for pk, q. To describe it explicitly, let p , q be the character ` x x{|x| |x| , x P R , 0, 1, P C, of R . Then p0, q p0, q p1, q, while if k 1, the sum pk, q ` pk, q equals ` ` p0, k ` q p1, k ` q ` p1, k ` q p0, k ` q ,

if k is even, and ` ` p0, k ` q p0, k ` q ` p1, k ` q p1, k ` q , if k is odd. These identities are easy consequences of Harish-Chandras formulas for the characters of representations in the relative discrete series of GLp2, Rq. Suppose that N 2n. For any k P N, we have the unitary, irreducible character pkq pk, 0q

7.5. SOME REMARKS ON CHARACTERS

429

in the relative discrete series for GLp2, Rq. The corresponding character n pkq on GLpN, F q dened by (7.5.1) is known as the Speh character attached to k and n. According to [AH, (5.11)(c)], it has an expansion ` n sgnpwq k pi wiq, pn ` 1q pi ` wiq n pkq
wPSn i1

in terms of virtual characters (7.5.2). Here, wi stands obviously for the image of i under the permutation w in the symmetric group Sn . Suppose next that F is p-adic. Let r be an irreducible unitary representation of WF of degree mr . If r pwq rpwq |w| , r pk, q r b k`1 , w P WF , k 0, is the twist of r by a real number , the tensor product

is an irreducible representation of the Langlands group LF WF SU p2q of degree m mr pk ` 1q. By the Langlands correspondence, it indexes a representation in the relative discrete series of GLpm, F q, whose character we denote by r pk, q. As in the case of R above, we need to dene r pk, q for all integers k. The value k 0 is already part of the denition here. It corresponds to the case that the character r pk, q is supercuspidal. If k 1, we follow the convention of [Tad2] of setting r pk, q equal to 1, the trivial character on the trivial group t1u. If k 1, we simply set r pk, q 0. Suppose that N mn, m mr pk ` 1q.
n The character r pkq on GLpN, F q associated to the unitary character

r pkq r pk, 0q by (7.5.1) is itself unitary. According to [Tad2, Theorem 5.4], it can be written explicitly as a signed sum ` n n r pkq sgnpwq r k pi wiq, pn ` 1q pi ` wiq
wPSn i1

of virtual characters (7.5.2). We have changed the notation of [Tad2] slightly in order to make the p-adic formula parallel to the formula for F R. In [Tad2], d equals pk ` 1q, Wn equals Sn , and the sum is taken over the subset ( pdq Wn w P Wn : wpiq ` d i, 1 i n ( w P Sn : k pi wiq 1, 1 i n of elements in Wn whose summand in the p-adic formula is nonzero. Moreover, if is the unitary supercuspidal representation corresponding to r, the character` of the representation denoted by pr, k sq on p. 342 of [Tad2] equals r k, k in our notation. 2

430

7. LOCAL NONTEMPERED REPRESENTATIONS

To write the two formulas together, we can always take r to be a unitary 1-dimensional representation of WF in case F R. The resulting character of GLp1, F q is of course the only irreducible unitary representation of a general linear group over F R that meets the formal denition of sun percuspidal. The virtual characters r pk, q and r pkq then make sense for both real and p-adic F , as do the representations r pk, q and
n r pkq r pkq b n

of LF and LF SU p2q respectively. By including r in the notation, we can then restrict P C to be real, as we did for p-adic F . The case F R will still dier slightly from that of p-adic F , in that r pk, q will remain unchanged if r is multiplied by the sign character, but this is not serious. Our interest here will generally be conned to the original case that r is trivial. At any rate, the formulas for both real and p-adic F can now be written together as n w (7.5.3) r pkq sgnpwq r pkq,
wPSn

where (7.5.4)
w r pkq n i1

` r k pi wiq, pn ` 1q pi ` wiq .

The outer automorphism x x_ of GLpN q gives an involution _ on the space of virtual characters on GLpN, F q. The virtual character in (7.5.3) satises n n r pkq_ r_ pkq,
n since the associated unitary representation r pkq of LF SU p2q has the same property. To describe the resulting mapping on the summands of (7.5.3), we write

w_ w w1 w w1 w1 w , where w : i pn ` 1q i, is the element of Sn of greatest length.

w P Sn , 1 i n,

Lemma 7.5.1. (i) The summand in (7.5.3) satises

w w r pkq_ r_ pkq,
_

w P Sn .
1

(ii) Suppose that k n, and that

w w r pkq r pkq,

for elements w, w1 P Sn . Then w w1 . Proof. (i) It is clear that r p , q_ r_ p , q,

7.5. SOME REMARKS ON CHARACTERS

431

for any P Z and P R. We shall check that this gives a bijection between the direct summands in the expansions (7.5.4) of each side of the proposed identity. Given w, we dene a permutation i i_ w wi, 1 i n, of the indices in the formal direct sum (7.5.4). Then k pi_ w_ i_ q k pw wi w w1 w w wiq k pw wi w iq `` ` k pn ` 1q wi pn ` 1q i k pi wiq. The permutation therefore stabilizes the rst argument k pi wiq of the corresponding two direct summands. Moreover, pn ` 1q pi_ ` w_ i_ q pn ` 1q pw wi ` w iq ` pn ` 1q pn ` 1q wi ` pn ` 1q i ` pn ` 1q pi ` wiq . The permutation therefore acts as p1q on the second argument pn ` 1q pi ` wiq of the direct summands. The assertion (i) follows. (ii) The rst condition k n implies that k pi wiq 1,
w r pkq

1 i n.

In particular, the linear forms in (7.5.3) are standard characters, with direct summands that lie in the mutually inequivalent family r p , q, P N, P R, of essentially square integrable characters. The second condition implies that there is a bijection i i1 of the indices such that k pi wiq k pi1 w1 i1 q and pn ` 1q pi ` wiq pn ` 1q pi1 ` w1 i1 q. These two equations imply that i1 i and w1 i1 wi, so that w1 w, as claimed. This lemma is intended only for perspective. If k n, it tells us that w is self-dual if and only if w w_ . If k n, r pkq is still self-dual _ . However, the converse question seems to be more complicated, if w w especially in case F R. The anti-involution w w_ of Sn is rather curious here. For it is the involution
w r pkq

w w w w ,

w P Sn ,

432

7. LOCAL NONTEMPERED REPRESENTATIONS

whose kernel is the Weyl group of the relevant classical group. The proper roles of these two operations are unclear to me. There are two elements w P Sn whose summands in (7.5.3) are of special w interest. The rst is the identity permutation w 1. Then r pkq equals the standard character n ` w (7.5.5) n pkq r 1 pkq r k, pn ` 1q 2i , w1 1. r
i1

This case is characterized by the condition that the direct summands indexed by i are equal, up to twists by unramied quasicharacters. The other element w is the longest permutation w w . Then r pkq equals n ` w r pkq r k ` pn ` 1q 2i, 0 .
i1

This case has the property that the factors indexed by i are often tempered. w If k n as in Lemma 7.5.1(ii), they are all tempered, and r pkq is again a w , and write standard character. With this restriction on k, we set w
w w n, pkq r pkq r pkq. r

The summand of w w in (7.5.3) is more complicated if k n. For pkq is then only a virtual character, which could be 0 if F is p-adic. If F is real, and k, n 1 are arbitrary, we can set w w as above, but we dene n ` r k ` pn ` 1q 2i, 0 , (7.5.6) n, pkq r
w r i1

where
r p , q r p| |, q. If F is p-adic, we have to dene w dierently. For any k, n 1, we set # pn ` 1q i, if i k ` 1, w i i p1 ` kq, if i k ` 2,

for any i in the interval of integers ( Nr1, ns i P N : 1 i n . Since w restricts to an order reversing bijection from Nr1, k ` 1s onto Nrn k, ns, and an order preserving bijection from Nrk ` 2, ns onto Nr1, n k 1s, it is indeed a permutation in Sn . We then set (7.5.7) n
w n, pkq r pkq r

n i1

` r k ` pn ` 1q 2i, 0 ,

where mintn, k ` 1u. The factors of n, pkq are thus (irreducible) temr pered characters in all cases. In the p-adic case, we note that the condition k n, under which we gave a simpler denition for n, pkq above, could be r relaxed to pk ` 1q n.

7.5. SOME REMARKS ON CHARACTERS

433

According to the denitions (7.5.5) and (7.5.6), n, pkq is induced from r an irreducible, tempered, essentially square integrable character of a Levi subgroup of GLpN, F q. It is therefore an irreducible tempered character on GLpN, F q. The following lemma summarizes the properties that will be needed in 8.2. Lemma 7.5.2. Suppose that k 0, n 1, r and N are as above, and that k 1 in case F is archimedean. (i) The irreducible tempered character n, pkq occurs with multiplicity r n p1q in the decomposition of the irreducible character r pkq into standard characters. (ii) The factor r pk ` n 1, 0q occurs with multiplicity 1 in the general decomposition (7.5.2) of n, pkq into essentially square integrable characters. r Proof. (i) Suppose that F R. The character n, pkq is part of the r n contribution to r pkq of the summand of w w w in (7.5.3). It repw resents the tempered part of the decomposition of r pkq into standard characters. This follows from (7.5.6), the denition (7.5.4) in case w w , and the decomposition of any direct factor r p , q into standard characters of GLp2, F q in case 0. Similarly, if w w w , the standard w characters in the decomposition of r pkq are all nontempered, since pn ` 1q pi ` wiq 0, n, pkq r 1 i n. Therefore is the only tempered character in the decomposition of n r pkq into standard characters. It actually occurs with multiplicity 1. Suppose that F is p-adic. The product
w sgnpw q n, pkq sgnpw q r pkq r

is then equal to the summand of w w in (7.5.3). Since the other summands in (7.5.3) are all standard characters (up to a sign), it suces to show that they are all nontempered. Assume that w P Sn indexes a summand in (7.5.3) that is nonzero and tempered. We consider the decomposition (7.5.4) w of r pkq. The index i of any direct summand in (7.5.4) satises either (7.5.8) (7.5.9) wi i ` k 1, pn ` 1q pi ` wiq 0, which means that the factor equals 1, or which in the absence of (7.5.8) is necessary for the factor to be tempered. The condition (7.5.8) implies that i k ` 2, since wi 1. It then follows from (7.5.9) that the restriction of w to Nr1, k ` 1s equals that of w w , and therefore maps Nr1, k ` 1s bijectively onto Nrn k, ns. This implies that w restricts to a bijection from Nrk ` 2, ns onto Nr1, n k 1s. But we must have wi i pk ` 1q, i P Nrk ` 2, ns, since the factor of i in (7.5.4) would otherwise vanish, contradicting the conw dition that r pkq is nonzero. This in turn can hold for every i P Nrk ` 2, ns

434

7. LOCAL NONTEMPERED REPRESENTATIONS

only if w maps Nrk ` 2, ns monotonically onto Nr1, n k 1s . In other words, the restriction of w to the remaining integer interval Nrk ` 2, ns also equals that of w . Therefore w equals w , as required. (ii) The general decomposition (7.5.2) for n, pkq is given explicitly by r (7.5.6) or (7.5.7), according to whether F is real or p-adic. In either case, r pk ` n 1, 0q is the factor indexed by i 1. In the case (7.5.6) of F R, it is clear from the denitions (and the fact that k 1) that the factor does occur with multiplicity 1. In the case (7.5.7) of p-adic F , the factors are all distinct, and therefore all have multiplicities 1. The lemma follows. In the course of proving (i), we established the following additional property. Corollary 7.5.3. The character n, pkq is the only tempered constituent r n in the decomposition of r pkq into standard characters. The formula (7.5.3) is an explicit expansion of the irreducible character n r pkq of GLpN, F q in terms of standard characters. Our real interest, however, is in the twisted characters for GLpN, F q. One would ultimately like a twisted analogue of (7.5.3) for any F , which is to say, an explicit description of the terms on the right hand side of the general formula (2.2.9). For p-adic F , Moeglin and Waldspurger [MW4] have established an inductive process, which is both subtle and complex, that leads to such a formula. The fact that (7.5.3) holds uniformly for any F suggests perhaps that a similar process might hold for F R. However, the twisted non-tempered characters of GLpN, Rq seem not to have been studied. In any case, it would be hard to establish an explicit twisted expansion from (7.5.3). The problem, pointed out to me by the referee, is that serious diculties accumulate from the ambiguity of extensions of self-dual representations. Fortunately, we will need only partial information, which we will in fact be able to extract from (7.5.3). Roughly speaking, we shall show that r the coecients np, q in (2.2.9) are congruent modulo 2 to their untwisted analogues. r We x a parameter P pN q over the local eld F . This was the setting of the preamble to Lemma 2.2.2, from which we take our notation. We will also use similar notation for the more elementary untwisted analogues of objects introduced in 2.2. In particular, we have the -subsets ( pN, q P pN q : p , q p , q, and P pN, q P P pN q : p , q p , q, ( of the respective sets pN q and P pN q of irreducible and standard representations of GLpN, F q. We also have untwisted analogues (7.5.10) fN pq mp, q fN pq, f P HpN q, P P pN, q,
PpN,q

7.5. SOME REMARKS ON CHARACTERS

435

and (7.5.11) fN pq
PP pN,q

np, q fN pq,

f P HpN q, P pN, q,

of the twisted expansions (2.2.7) and (2.2.8), or rather their restrictions to r r r r the subsets pN, q and P pN, q of pN q and P pN q. The expansion (7.5.3) we have just described is an explicit form of (7.5.11), in the special case that is simple and . To exploit it, we need to nd some relation between the general expansions we have just quoted with the twisted expansions of 2.2. r Lemma 7.5.4. Given P pN q, assume that prq0 and prq0 are rpN, q and P pN, q. r restrictions to GLpN, F q of representations P P r r Then the coecients in (2.2.7) and (7.5.10) and in (2.2.8) and (7.5.11) satisfy r mpr, q mp, q (modulo 2) and npr, q np, q (modulo 2). r r r Proof. According to the notation of 2.2, P pN, q and pN, q are sets rpN qu and tpN qu from (2.2.7) r of representatives of orbits in the families tP r and (2.2.8). They consist of representations of G` pN, F q that are dened by the Whittaker extensions from the beginning of 2.2. In particular, the corresponding coecients mpr, q and npr, q are indeed integers. The same r r is of course true of the general coecients mp, q and np, q in (7.5.10) and (7.5.11), whether and are self-dual (which is to say, of the given form prq0 r r and prq0 ) or not. The particular extensions and are in fact irrelevant r to the lemma. So long as they are actually representations of G` pN, F q, they are determined up to a sign, which does not aect the assertion of the lemma. The expansion (7.5.10) is just the decomposition of the standard character of GLpN, F q into irreducible characters . In particular, mp, q is the multiplicity of in the -isotypical subspace V p, q of the space on which acts (or rather the image of this space in the Grothendieck group KpGqC ). The twisted expansion (2.2.7) is the decomposition of the standard characr r r ter of G` pN, F q into irreducible characters , or rather the restriction of r r this decomposition to the subset GpN, F q of G` pN, F q. It is obtained from (7.5.10) by rst removing the summands in (7.5.10) of representations with _ , and then for the remaining , keeping track of the restrictions r of to the subspaces V p, q. For any P pN, q with _ , we thus have mpr, q m` p, q m p, q, r where m` p, q is the multiplicity of in the -isotypical subspace of V p, q, r p, q is the multiplicity of in the complementary subspace. Since and m mp, q m` p, q ` m p, q,

436

7. LOCAL NONTEMPERED REPRESENTATIONS

we see that mpr, q is congruent to mp, q modulo 2. This is the rst of r the required congruence relations. The second congruence relation will be a little more complicated. We can regard P pN, q and pN, q as partially ordered sets, dened by the partial order on the linear forms and in pa q` . The mapping from B P pN, q to pN, q is then a canonical isomorphism of partially ordered sets. To simplify the notation, we introduce an abstract partially ordered set Kpq, with corresponding isomorphisms k k , and k k onto P pN, q and pN, q. We then write mkk1 mpk , k1 q, and nkk1 npk , k1 q, k, k 1 P Kpq, for the integral coecients in the expansions (7.5.10) and (7.5.11). The set Kpq inherits an involution k k _ from the outer automorphism of GLpN q, and it follows from the expansions that
_ mk_ pk1 q_ mp_ , k1 q mpk , k1 q mkk1 k

k, k 1 P Kpq,

and
_ nk_ pk1 q_ npk , _1 q npk , k1 q nkk1 , k

for any k and k 1 in Kpq. ` Let Upq be the set of unipotent, integral Kpq Kpq -matrices a pakk1 q, such that (i) ukk 1, (ii) ukk1 0, unless k 1 k, and (iii) uk_ pk1 q_ ukk1 . It follows immediately that Upq is a group under matrix multiplication, in which our two coecient matrices pmkk1 q and pnkk1 q represent inverse elements. To exploit the third condition, we borrow our earlier notation from the slightly dierent context of 1.4. We write Kpq Ipq where ( Ipq i P Kpq : i_ i , while Jpq is a xed set of representatives in Kpq of orbits of order two under the involution and J _ pq is the complementary set of representatives. We can then introducea second group. Let UI pq be the group of unipotent, ` integral, Ipq Ipq -matrices aI paii1 q, i, i1 P Ipq,

k, k 1 P Kpq,

Jpq

J _ pq,

7.5. SOME REMARKS ON CHARACTERS

437

that satisfy the three conditions above (but with ii1 in place of kk 1 ). In this case, there are really only two conditions, since the adjoint condition (iii) is trivial. There is a natural restriction mapping a aI from Upq onto UI pq. However, this projection is not a group homomorphism. To describe the obstruction, consider the image in UI pq of a product of matrices a and b in Upq. For any indices i and i1 in Ipq, we can write pabqii1 aik bki1
kPKpq

i PIpq

aii bi i1 `

jPJ

aij bji1 `

j _ PJ _

aij _ bj _ i1

i PIpq

aii bi i1 ` 2

jPJ

aij bji1 ,

since aij _ aij and bj _ i1 bji1 . In particular, the obstruction vanishes if the matrix coecients are taken modulo 2. Let us therefore write Upq and U I pq for the two matrix groups dened in the same way as Upq and UI pq, but the coecients in the eld Z{2Z. The restriction mapping u uI is then a surjective homomorphism from Upq onto U I pq. To put the matter a little more strongly, we write x for the reduction modulo 2 of any integral matrix x. It then follows that the mapping : a aI paI q, a P Upq,

is a surjective homomorphism from Upq onto U I pq. We now complete the proof of the lemma. We can of course identify the matrices m and n in Upq with the coecient matrices of (7.5.10) and (7.5.11). In particular, n equals the inverse of m . It follows that nI pnq pm1 q pmq1 pmI q1 , since as a homomorphism, commutes with inverses. Having established the rst congruence relation, we can identify mI with the reduction modulo 2 of the coecient matrix in (2.2.7). Its inverse m1 is therefore the reduction I modulo 2 of the coecient matrix in (2.2.8). Since this is in turn equal to nI , we have also established the second congruence relation. Our interest will be in the special case of (7.5.11) that equals the Langlands quotient attached to . In this case, (7.5.11) can be written in the form (7.5.12) fN pq np, q fN pq, f P HpN q,
PpN,q

where pN, q P pN q : p , q p , q, (

438

7. LOCAL NONTEMPERED REPRESENTATIONS

and np, q np , q. This is the untwisted analogue of the twisted expansion (2.2.9). It is the abstract setting for the explicit character identity (7.5.3). The formula (7.5.3) actually applies only to case that is simple. However, it is easily extended to any parameter
1 1

r r .

When is simple, and corresponds to a pair pr, kq as in (7.5.3), we write

n pq r pkq.

for the right hand side of (7.5.3). For general , we can then set (7.5.13) pq looooooooooomooooooooooon looooooooooomooooooooooon . p1 q p1 q pr q pr q
1 r

As a character on GLpN, F q induced from an irreducible unitary character, pq is itself irreducible. It is the character of the Langlands quotient , whose value at a test function f P HpN q equals the left hand side of (7.5.12). The corresponding direct sum of the virtual characters on the right hand side of (7.5.3) is an explicit (though by now somewhat complicated) integral linear combination of standard characters on GLpN, F q. Its value at f equals the right hand side of (7.5.12). The two extremal standard characters in this explicit form of (7.5.12) are p q p1 q p1 q pr q pr q loooooooooooomoooooooooooon loooooooooooomoooooooooooon
1 r

and p q p 1 q p 1 q p r q p r q, loooooooooooomoooooooooooon loooooooooooomoooooooooooon

1 r

where if is simple, p q n pkq and p q n, pkq, in the earlier r r notation. In the notation here, it is understood and stand for the Langlands parameters of the standard characters p q and p q. We know very well that np, q 1. But from Lemma 7.5.2(i) we observe also that np, q 1. We recall that indexes the constituent of (7.5.12) that is most highly nontempered, while it follows from Corollary 7.5.3 that indexes the unique constituent of (7.5.12) that is tempered. With these remarks, we obtain the following corollary of Lemma 7.5.4.

7.5. SOME REMARKS ON CHARACTERS

439

r Corollary 7.5.5. For any P pN, q, the coecients in (2.2.9) and (7.5.12) satisfy np, q np, q (modulo 2). r In particular, r np, q 1 (modulo 2), so the parameter has nonzero multiplicity in the twisted expansion (2.2.9). It indexes the unique constituent of (2.2.9) that is tempered. r r Suppose that lies in the subset pGq of pN q attached to a simple rsim pN q. The Langlands parameter datum G P E : LF
L

of the standard character p q on GLpN, F q is dened by ` puq u, ppuq , u P LF , where 1 |u| 2 0 0 1 , |u| 2 u P LF .

ppuq

We can interpret the homomorphism p : LF SLp2, Cq as the Langlands parameter for the trivial one-dimensional representation of P GLp2, F q. As such, it has a dual parameter p : WF SLp2, Cq, which corresponds to the associated square integrable representation of P GLp2, F q. More precisely, p pw, sq s, pw, sq P LF WF SU p2q,

if F is p-adic, while if F R, we have 1 pzzq 2 0 and p pF q

0 pzzq
1 2

p pzq

z P C ,

0 1 , 1 0

440

7. LOCAL NONTEMPERED REPRESENTATIONS

in the notation of 6.1. We can therefore associate a second Langlands parameter : LF L G r in pGq to by dening ` puq u, p puq , u P LF . This is the Langlands parameter of the standard character p q on GLpN, F q. For p-adic F , it plays an important role in the work of Moeglin.

CHAPTER 8

The Global Classication

8.1. On the global theorems We are at last in a position to prove the global theorems. We have already had to treat some special cases in order to prove the local theorems. It is now time to work in the opposite direction. We shall use the local theorems we have just established to prove the global theorems in general. We will complete this argument in 8.1 and 8.2. In the remaining three sections of the chapter, we will discuss some ramications of the results. Recall that there are four global theorems. They consist of Theorems 1.5.2 and 1.5.3 (which for their statements implicitly include Theorems 1.4.1 and 1.4.2), and the global supplements Theorems 4.1.2 and 4.2.2. The global intertwining relation of Corollary 4.2.1 is actually a corollary of the associated local intertwining relation, which we have established in all cases. On the other hand, the stable multiplicity formula of Theorem 4.1.2 is really the fundamental global assertion. It represents a foundation on which the others all depend. The underlying eld F will be global throughout this chapter, unless r specied otherwise. The theorems apply to parameters in the set of selfdual representations that occur in the automorphic spectral decompositions of general linear groups over F . We shall combine the local results we have established with the global formulas of 5.15.3. To this end, we will take r r the global family F of Chapter 5 to be the entire family . To be consistent r with the conventions of 5.1, we must consider the case that F equals the r of generic parameters in , as well as that of F . However, we r r r subset r r will generally write in place of F, distinguishing when necessary between the generic and nongeneric cases. As usual, when working with nongeneric r parameters in , we will assume implicitly that we have already established the global theorems for generic parameters. We x a positive integer N , and as in 5.1, assume inductively that the r global theorems all hold for any P with degpq N . We do not need to add a local hypothesis, of course, since we have now established all the local theorems. We do however have to check that the family satises the additional hypothesis of 5.1. r Lemma 8.1.1. Suppose that G P Eell pN q and that belongs to 2 pGq. Then the conditions of Assumption 5.1.1 hold for the pair pG, q.
441

442

8. THE GLOBAL CLASSIFICATION

Proof. This is the analogue of Lemma 5.4.2, for the maximal family r r F . However, it will be a little easier for us to prove, now that we have the local theorems in hand. Suppose for example that G GS GO is composite. From our induction hypothesis, we observe that any localization r v S,v O,v belongs to the local set pGv q. The condition of Assumption 5.1.1 in this case then follows from its local version (2.2.4) in Theorem 2.2.1. We therefore assume that G is simple. If we can show that any localr ization v of still belongs to pGv q, we can appeal again to the relevant local assertion (2.2.3) of Theorem 2.2.1. Suppose that is composite. An application of our induction hypothesis to the simple constituents i of , r r combined with the inductive denition of the sets pGq and pGv q, then r v q, and hence that satises Assumption 5.1.1. conrms that v lies in pG If is simple but not generic, we can apply our induction hypothesis to r its simple factor . We see again that v belongs to pGv q, and that satises Assumption 5.1.1. We can therefore assume that is simple and generic. We have reduced the proof of the lemma to the case of a simple generic pair r r pG, q, G P Esim pN q, P sim pGq. r The problem is to show that any localization v of belongs to pGv q. The reader will recognize in this the assertion of Theorem 1.4.2, the second seed theorem from Chapter 1. However, we are now using the temporary r denition of sim pGq from 5.1, rather than the original denition from Theorem 1.4.1 in terms of which Theorem 1.4.2 was stated. The logic of the r temporary denition is actually somewhat convoluted. For we took sim pGq r sim pN q that satised the condition of to be the subset of elements P G Assumption 5.1.1, in addition to the natural global condition Sdisc, 0. We then had to reinterpret Assumption 5.1.1 for simple generic pairs as the condition (5.1.6). Our problem, thus modied, is to show that for any G r r P sim pN q, we can nd a simple datum G P Esim pN q such that Sdisc, 0 r and such that the linear form fN pq transfers to G. r sim pN q, we write For the given P ` N r r r r r rN r Idisc, pf q tr Rdisc, pf q fN pq, f P HpN q. r This is one of the various reductions of (4.1.1) (with GpN q in place of G) r with which we are now very familiar. It then follows from the GpN q-analogue (3.3.14) of (4.1.2) that r r r r pG rpN, Gq Sdisc, pf G q, fN pq f P HpN q.
r GPEsim pN q

Here we are using the fact, obtained by the usual induction argument, that G r r Sdisc, vanishes for any G in the complement of Esim pN q in Eell pN q. There

8.1. ON THE GLOBAL THEOREMS

443

r must be at least one G P Esim pN q whose summand does not vanish identically r in f . If there is exactly one, we have r r pG fN pq rpN, Gq Sdisc, pf G q, r r f P HpN q,

r r and fN pq is the pullback of a unique stable linear form on HpGq, as required. This is the case for example if N is odd or the character is nontrivial. We can therefore assume that N is even and 1. The summand of r G then vanishes unless G is one of the two split groups in Esim pN q, which _ . The identity becomes we denote as usual by G and G r r pG pG_ r _ fN pq rpN, Gq Sdisc, pf G q ` rpN, G_ q Sdisc, pf G q. r We can assume that neither summand vanishes identically in f , since we would otherwise be in the case settled above. We can also assume that r the linear form fN pq does not transfer to either G or G_ , since we would otherwise again be done. We must show that these conditions lead to a contradiction of the identity. The second condition implies that there are valuations v and v _ such r r that v and v_ do not lie in the respective local sets pGv q and pG__ q. v _ . One can then choose a function f P H pN q such r rv Suppose rst that v v v rG rG_ r that fv fv 0, but such that fv,N pv q 0. We leave the reader to check that this is possible, using the characterization of Proposition 2.1.1 r and the fact that v must still belong to the local set pG q attached to v r r r r r some G P Ev pN q. Once fv is chosen, we set f fv f v , for any function v r rv r r _ r r f v P Hv pN q such that fN pv q 0. Then f G f G 0, while fN pq 0. In other words, the right hand side of the last identity vanishes, while the left hand side is nonzero. This is a contradiction. Suppose next that v v _ . r r r r We can then choose functions fv P Hv pN q and fv_ P Hv_ pN q such that rG 0 and f G_ 0, but such that fv,N pv q and fv_ ,N pv_ q are both r_ r r fv v r r r r _ nonzero. We then set f fv fv_ f v,v , for a complementary function _ _ _ r rv,v r r _ f v,v with fN pv,v q 0. Once again, we have f G f G 0 and r fN pq 0, and therefore a violation of the last identity. We thus have obtained a contradiction in both cases. Our conclusion is that the linear r form fN pq does indeed transfer to one of the two groups. Consequently, r r there is a simple datum G P Esim pN q such that lies in the subset sim pGq r of sim pN q. This completes the proof of the lemma, in the last case of a simple generic parameter. Remark. We observed that Assumption 5.1.1, when restricted to the r case of simple generic parameters P sim pN q, is essentially the assertion of Theorem 1.4.1. The only dierence is that it is expressed in terms of r r our alternate denitions of the sets sim pN q and v pN q. In other words,

444

8. THE GLOBAL CLASSIFICATION

Assumption 5.1.1 (for simple generic ) is tied to the conditions (i) in Corollaries 5.4.7 and 6.8.1, while Theorem 1.4.2 is tied to the equivalent conditions (ii) in these corollaries. r r r r We have now placed the maximal family F (or F ) within the general framework of 5.1. In particular, the family comes with a set of induction hypotheses on the positive integer N . Our task is to complete the inductive proof by establishing the global theorems for parameters in the r r r r set FpN q pN q (or FpN q pN q). Suppose for a moment that belongs to the complement of the set r r ell pN q. Recall that this means that does not belong to 2 pGq, for any r G P Eell pN q. The global theorems reduce in this case to the two vanishing assertions of Proposition 4.5.1. However, the proposition actually imposes a r further requirement that also not belong to the larger elliptic set ell pGq, r if G P Esim pN q is simple. This second condition does not subsume the rst, r r since it permits to lie in 2 pG q for some G P Eell pN q that is not simple. And indeed, there really was something to establish in this case. We did it in Lemma 5.1.6, which was our setting for a general proof of Theorem 1.5.3(b), and where we also completed the proof of the two vanishing assertions for pairs r r pG, q, G P Esim pN q, P 2 pG q. (The corresponding two assertions do not have to be proved for the pair pG , q. They follow immediately from our induction hypothesis on N , G G and the fact that the linear forms trpRdisc,v q and Sdisc, for composite G decompose into the associated products. We note, incidentally, that the roles of G and G here have been interchanged from Lemma 5.1.6.) Proposition 4.5.1 and Lemma 5.1.6, taken together, thus yield the global theorems for any that does not belong to the elliptic set attached to any simple G. r We can therefore assume from this point on that lies in ell pGq, for r some G P Esim pN q. One sees easily from the general form (4.5.9) of S that G is uniquely determined by . r r r Suppose rst that lies in the complement 2 pGq of 2 pGq in ell pGq. ell r ell pN q. The global theorems therefore Then lies in the complement of reduce to the two vanishing assertions of Proposition 4.5.1, even though the proposition itself does not apply to this case. We shall instead apply Lemma 5.2.1 or Lemma 5.2.2, according to whether the index r in the decomposition (5.2.4) of is greater than 1 or equal to 1. The global intertwining relation (4.2.6) of Corollary 4.2.1, which we now know is valid, tells us that the given expressions (5.2.8), (5.2.12) and (5.2.13) in these lemmas all vanish. It follows that the remaining expressions (5.2.7) and (5.2.11) in the two cases also vanish. In other words, we have ` rpN, G q tr Rdisc, pf q 0,
r G PEsim pN q

8.1. ON THE GLOBAL THEOREMS

445

for any r, and any compatible family of functions (5.2.6). It then follows from Proposition 3.5.1 that ` r r (8.1.1) tr Rdisc, pf q 0, G P Esim pN q, f P HpG q. This is one of the two vanishing assertions. The second, 0 r r (8.1.2) S pf q 0, G P Esim pN q, f P HpG q,
disc,

follows from the rst. Indeed, the dierence (5.2.9) of the left hand sides of (8.1.1) and (8.1.2) was seen in the proofs of Lemmas 5.2.1 and 5.2.2 to equal one of the expressions (5.2.8), (5.2.12) and (5.2.13) we now know are all equal to 0. We have thus established the global theorems when does not lie in r 2 pGq, for any G P Esim pN q. This is a major step, even though it amounts r only to a pair of vanishing formulas. Theorem 1.5.2 asserts that the contribur tions from parameters in 2 pGq exhaust the automorphic discrete spectrum of G. The rst formula (8.1.1), applied to G G (and augmented by Corollary 3.4.3), asserts that this is indeed the case. In particular, G has no embedded eigenvalues, in the sense described at the beginning of 4.3. This result has been hard won. The proof we have just completed calls upon much of what we have done since Chapter 1. At its heart are the comparisons from the standard model in 4.3 and 4.4, in which the two vanishing formulas are treated side by side. The second formula (8.1.2) is equivalent to the stable multiplicity formula (4.1.11) for the pair pG, q. We have now established that it holds when the group S attached to pG, q is innite. 0 This of course is the case in which the coecients pS q in (4.1.11) vanish, according to the property (4.1.9). We come now to the last case, in which is square integrable. We r r assume henceforth that lies in 2 pGq, for some G P Esim pN q. In particular, 1 r is multiplicity free. For this case, there will be assertions to be established from all of the global theorems. If is compound, in the sense that r 1, the techniques we have been using will be quite sucient. The case r 1 that is simple is considerably harder. It requires something further, which we will discuss in the next section. In both cases, the stable multiplicity formula remains the key result to be proved. Lemma 5.1.4 gives a criterion for the stable multiplicity formula to be r valid for any pair pG , q, with G P Esim pN q. It tells us that we need r0 pLq vanishes. We recall that L and only show that the linear form on S are dened only if N is even and n 1, or equivalently, if there is a 1 r second group G_ P Esim pN q with nG_ 1. Then L GLp 2 N q represents _ , and is the obstruction for a maximal Levi subgroup of both G and G _ G_ the stable linear form Sdisc, Sdisc, on G_ to vanish as expected. Recall also that if N is odd or n 1, we have already established the stable multiplicity formula for each pG , q in Corollary 5.1.3.

446

8. THE GLOBAL CLASSIFICATION

Assume now that r 1. We apply Lemma 5.3.1 to the supplementary parameter ` . The global intertwining relation, which we now know holds in general, tells us that the expression (5.3.5) vanishes. It then follows from the lemma that the sum ` rpN` , G q tr Rdisc,` pf q ` b` f L` p1 q
r G PEsim pN` q

vanishes, for a positive constant b` , and any compatible family of functions (5.3.3). We can now argue exactly as in the proof of Lemma 5.4.5. That is, we replace the supplemental summand on the right by the linear form L r b` f ` p1 q, f1 P HpG_ q,
1 1

r for the composite endoscopic datum G1 G_ in Eell pN` q, with corresponding function f1 in (5.3.3). This is a nonnegative linear combination of irreducible characters in f1 , by Corollary 5.1.5 and the application of our induction hypothesis to 1 . We can therefore apply Proposition 3.5.1 to the modied sum. This allows us to say that the linear form 1 attached to the maximal Levi subgroup L` of G_ vanishes. Since the linear form 1 attached to 1 is nonzero by our induction hypothesis, we conclude that vanishes. The stable multiplicity formula thus holds for any pair pG , q, and for pG, q in particular, whenever the index r for is greater than 1. Now that we have the stable multiplicity formula for parameters r P 2 pGq with r 1, we shall establish Theorem 1.5.2 for any such . r More precisely, for any G P Esim pN q, we shall show that the contribution of to the discrete spectrum of G is the summand of on the right hand side of the analogue for G of the decomposition (1.5.5) of Theorem 1.5.2. r r If G P Esim pN q is distinct from G, does not belong to pG q, by denition. There is consequently no contribution of to the analogue for G of (1.5.5). According to the stable multiplicity formula for pG , q, the stable linear form Sdisc, attached to pG , q vanishes. It follows from (4.4.12) that ` r tr R pf q S pf q 0, f P HpG q,
disc, disc,

G_ 1

so there is also no contribution of to the discrete spectrum of G . It therefore suces to consider the case that G G. Since belongs to r 2 pGq, its contribution to (1.5.5) is a nontrivial multiplicity formula, rather than zero. But we have already seen how to prove this formula. The proof is of course the content of our conditional Lemma 4.7.1, from the end of 4.7. The conditions of this lemma (both explicit and implicit) have now all been met. The conclusion (4.7.10) of the lemma, namely that the contribution of to the discrete spectrum of G is indeed the summand of in (1.5.5), is therefore valid. Theorem 1.5.2 therefore holds for the compound parameter r P 2 pGq. The only other global theorem that is relevant to is Theorem 4.2.2. r r It concerns the bitorsor G attached to a datum G P Esim pN q, in case N is

8.1. ON THE GLOBAL THEOREMS

447

p r even, G SOpN, Cq, and lies in 2 pGq. Recall that (4.2.7) provides a canonical extension r G r Rdisc, Rdisc,
G r of the representation Rdisc, to the GpAq-bitorsor GpAq. The theorem has r r two assertions, according to whether lies in the subset 2 pGq of 2 pGq or not. The argument is similar to the special case discussed, somewhat hastily, in 6.8. We form the -component r G r r r Idisc, pf q Idisc, pf q,

r r f P HpGq,

r of the discrete part of the twisted trace formula for G, following (3.3.12). According to Hypothesis 3.2.1, it satises the analogue r1 r r r p r r Idisc, pf q pG, G1 q S 1 disc, pf q
r G1 PEell pGq

r r for G of (4.1.2). It also satises the analogue for G of (4.1.1). As we have observed many times in the case of a square integrable parameter , the terms with M G in (4.1.1) all vanish. The remaining term with M G r r is the trace of the operator Rdisc, pf q, by the general denitions at the end of 3.1. It follows that ` r1 r r r r p r (8.1.3) tr Rdisc, pf q pG, G1 q S 1 disc, pf q.
r r G1 PEell pGq

r r Suppose that lies in the complement of 2 pGq in 2 pGq. Then does r r r r not lie in pG1 q, for any twisted endoscopic datum G1 P Eell pGq. It follows 1 , or in fact the application to G1 r r from the stable multiplicity formula for G r1 of the more elementary assertion of Proposition 3.4.1, that Sdisc, vanishes. The left hand side of (8.1.3) therefore also vanishes. This is the assertion (a) of Theorem 4.2.2. r r r Now suppose that belongs to the subset 2 pGq of 2 pGq. The set S is then nonempty, and is a bitorsor over S . To establish the corresponding assertion (b) of Theorem 4.2.2, we can follow the argument from 6.8 for a generic global parameter with local constraints. Starting with (8.1.3), we obtain a generalization rr r (8.1.4) n pq fG prq |S |1 1 p 1 q f 1 p 1 q,
r PpGq r r xPS

r r of (6.8.4), in which pG1 , 1 q maps to p, xq, pGq is the subset of represenr r r tations in pGq that extend to GpAq, and n pq is the multiplicity of in G r Rdisc, . On the left hand side, is the canonical extension of dened whenever n pq 0.

448

8. THE GLOBAL CLASSIFICATION

We noted in 4.8 that the proof of Lemma 4.4.1 in 6.6 is quite general. r In particular, it carries over without change to a twisted group G. We can therefore write r r r r r xs x, y f r prq, ps xq 1 p 1 q f 1 p 1 q
G r r x P S r r xPS P

according to the general argument of 4.7. With a change of variables in the sum over x, we see that the right hand of (8.1.4) equals r x r rr xr, y fG prq. G prq |S |1 x
xPS r r P

This in turn equals

P p q

rr fG prq,

where p q is the set of representations in the global packet such r that the linear character x , y on S equals 1 G , while is the canonical extension of determined by . To be more precise, we recall r r that Theorem 2.2.4 assigns an extension of from GpAq to GpAq to any r . In the last sum, extension x , y of the linear form x , y from S to S r r r . corresponds to the canonical extension of x , y to S The formula (8.1.4) thus takes the form r r r r n pq f r prq f r prq, f P HpGq.
G G r PpGq P p q

rG r On the left hand side, represents the extension of determined by Rdisc, , while on the right hand side, it stands for the extension determined by . The formula tells us that the two extensions match, as asserted in Theorem 4.2.2(b). This completes the proof of Theorem 4.2.2, and hence our study r of global parameters P 2 pGq with r 1. r Suppose now that r 1. In other words, P sim pGq is a simple rsim pN q. As we have noted, this case will be parameter for the group G P E more of a challenge. We shall begin the discussion here, leaving the main burden of proof for the next section. There are two general problems in this case. One remains the stable multiplicity formula, which entails proving that the linear form vanishes. We shall soon see how to reduce this to the case that is not generic. The other r is the assertion of Theorem 1.5.3(a) for generic parameters P sim pGq. In 5.3, we formulated it in terms of a sign , which equals 1 or 1 according to whether the assertion is valid for or not. This of course applies only to the case that is generic. There will thus be one problem to solve if is not generic, and another if it is. We remind ourselves that we are working in the general framework of Chapter 5. In particular, we are following the convention at the beginning

8.1. ON THE GLOBAL THEOREMS

449

r r of Chapter 5 for the denition of the subset sim pGq of sim pGq. Once we have established the stable multiplicity formula in case is generic, we will r be able to say that G is uniquely determined by , and hence that sim pN q r r is a disjoint union over G P Esim pN q of the sets sim pGq. This leads directly to a proof of Theorem 1.4.1. It will therefore tell us that the temporary r denition of sim pGq is equivalent to the original denition in terms of this r theorem. We recall that these matters were treated in 5.4, when F was the family with local constraints introduced there. We are of course now taking r r r F to be either the full set or its subset of generic parameters. r We begin our investigation of the simple parameter P sim pGq as we did in the compound case above. That is, we turn to the only result available to us, the relevant lemma from 5.3. This is Lemma 5.3.2, which we apply to the supplementary parameter ` , and a compatible family of functions (5.3.3). Consider the terms in Lemma 5.3.2. We rst note that the global intertwining relation holds for the pairs pG` , ` q and pG_ , ` q. It gives ` identities
1 fG` p` , s` x1 q fG` p` , x1 q

and
_ pf _ qM` pq pf _ q1 _ p` , x1 q fG_ p` , x1 q G` `

between the terms in the two expressions (5.3.22) and (5.3.23) of the lemma. The dierence between (5.3.22) and (5.3.23) therefore equals
1 8

p1 qpf _ qM` pq 1 p1 qfG` p` , x1 q, 8

an expression we can write as

1 8

` p1 q f M` pq fG` p` , x1 q , pf _ qM` pq f M` pq.

since Now f M` pq represents a unitary character on the group G` pAq. It is induced from the Langlands quotient representation of the maximal Levi subgroup M` pAq GLpN, Aq of G` pAq. By denition, fG` p` , x1 q is a dierence of two unitary characters on G` pAq whose sum equals f M` pq. Since equals 1 or 1, the last expression is therefore a nonnegative linear combination of irreducible characters. Lemma 5.3.2 tells us that its sum with ` 1 rpN` , G q tr Rdisc,` pf q ` 2 f L` p q ,
r G PEsim pN` q

450

8. THE GLOBAL CLASSIFICATION

the expression we labelled (5.3.21) in the statement of the lemma, vanishes. Arguing as in the proof of Lemma 5.4.6, we replace the supplemental summand in (5.3.21) by the linear form
1 L` 2 f1 p

q,

r f1 P HpG_ q, 1

r for the composite endoscopic datum G_ G G_ in Esim pN` q, with cor1 responding function f1 from the family (5.3.3). By Corollary 5.1.5 and the denition of , this is a nonnegative linear combination of irreducible characters in f1 . The same is of course true of the other summands in (5.3.21). The terms in Lemma 5.3.2 thus combine to give an expression of the general form (3.5.1). The lemma itself asserts that the expression so obtained vanishes. It then follows from Proposition 3.5.1 that the corresponding coecients vanish. In particular, we see that (8.1.5) that 0, and that ` p1 q f M` pq fG` p` , x1 q 0. The second of these equations tells us that either 0, which is what we want, or 0. In other words, we have (8.1.6) 0, if 0.
Rdisc,` pf q 0,

r r G P Esim pN` q, f P HpG q,

The third equation tells us that either 1, which is what we want, or the complementary factor on the left hand side vanishes. In other words, we have (8.1.7) f M` pq fG` p` , x1 q, if 1,

r for any function f P HpG` q. What is one to make of these conditions? G Recall that Sdisc, pf q equals a scalar multiple of f G pq, according to the G denition prior to Lemma 5.3.2. If is generic, Sdisc, is nonzero, by our r denition of sim pGq. It follows that 0, and therefore that 0. We note in passing that this argument could also have been used in the proof of Lemma 5.4.6. The local condition (5.4.1)(a) we used instead is parallel to its nontempered analogue needed for the proof of Lemma 7.3.2. In any case, we have now established the stable multiplicity formula for generic . This is the result we anticipated above. It resolves the temporary denition r of sim pGq from 5.1 in terms of the original denition by Theorem 1.4.1, r and thus completes our inductive denition of the full set pGq. This leaves only the two problems mentioned above. We must show that 1 if is generic, and that 0 if is not generic. At the beginning of the next section, we shall see whether we can solve the two problems by applying the conditions (8.1.7) and (8.1.6).

8.2. PROOF BY CONTRADICTION

451

8.2. Proof by contradiction The task now is to nish proving the most intractable case of the global theorems. In the last section, we xed a positive integer N , and imposed the usual induction hypothesis that the theorems all hold for global parameters r P with degpq N . We then described how our earlier results yield the r global theorems for parameters of degree N in the complement of sim pN q. r sim pN q. It remains to treat the simple global parameters P In the last section we also obtained information from the application of Lemma 5.3.2. In particular, we proved the stable multiplicity formula in the r r case that lies in the subset sim pGq of generic parameters in sim pGq. It remains only to establish the condition 1 of Theorem 1.5.3(a) in this r case. In the complementary case that P sim pGq is not generic, 1 by denition. In this case, it is only the stable multiplicity formula for that needs to be established. We thus have the two separate properties to establish, one when is generic, and the other when it is not. In each case, there is a corresponding condition (8.1.7) or (8.1.6) we can bring to bear on the proof. Are the conditions (8.1.7) and (8.1.6) we have obtained from the application of Lemma 5.3.2 sucient to establish the nal two properties? Our initial answer to this question is not likely to be encouraging. However, with patience, we will see that it eventually leads to a happy conclusion. We shall study the question for a general pair pG, q, r r G P Esim pN q, P sim pGq.

The most critical case is obviously that of N even and n 1. This is the case in which G is split, and where we have the auxiliary objects G_ , L and . We shall discuss these objects in treating the general case, with the implicit understanding that any terms in which they occur vanish if N is odd or n 1. Consider the condition (8.1.6) that either or must vanish. If it is that vanishes, as required, we see from Lemma 5.1.4 that
G r Sdisc, pf q |pG, q| |S |1 G pq f G pq opGq f L pq

r r opGq 1 1 f G pq opGq f G pq, and that

G Sdisc, opG_ q f _,L pq 0, r
_

r r for functions f P HpGq and f _ P HpG_ q that we may as well assume come from a compatible family. These are of course the stable multiplicity formulas for pG, q and pG_ , q. In this case, we also obtain the local identity f G pq f G pq ` f L pq f G pq,

452

8. THE GLOBAL CLASSIFICATION

from the formula obtained in 5.3 prior to Lemma 5.3.2. On the other hand, if it is that vanishes, we have the local identity f G pq f G pq ` f L pq f L pq. In particular, the stable distribution on GpAq attached to the original parameter transfers locally from the Levi subgroup LpAq. In this case, it follows from Lemma 5.1.4 that
G r Sdisc, pf q |pG, q| |S |1 G pq f G pq opGq f L pq

r r opGq 1 1 f L pq opGq f L pq 0, and that

G r r Sdisc, pf _ q opG_ q f _,L pq opG_ q f G pq.
_

Consider the case that is generic. Then 0, as we have seen. The problem here is to show that 1. Assume the contrary, namely that r 1. For any function f P HpG` q, we obtain identities
1 f M` pq fG` p` , x1 q fG` p` , x1 q

f GG p q f GG p q,
1 from (8.1.7), the global intertwining relation, and the denition of fG` in terms of the endoscopic transfer f GG of f to the endoscopic datum G1 G G for G` . It then follows from Corollary 5.3.3 and (8.1.5) that `
` Sdisc,` pf q 1 opG` q f M` pq. 2 r

r In addition, if f _ P HpG_ q lies in a compatible family that contains f , we ` obtain an identity

_ f M` pq fG_ p` , x1 q, `

as above. It then follows from Corollary 5.3.3, (8.1.5), and the vanishing of that
` Sdisc,` pf _ q

G_

1 4

r ` opG_ q f M` pq.

To summarize what we have obtained so far, we take two pairs of functions

_ r r tf1 P HpGq, f1 P HpG_ qu

and
_ r r tf2 P HpG` q, f2 P H` pG_ qu `

8.2. PROOF BY CONTRADICTION

453

that are compatible, in the sense that they belong to compatible families for r r Eell pN q and Eell pN` q respectively. We then have $ G G Sdisc, pf1 q opGq f1 pq, r _ G S _ & disc, pf1 q 0, S G` pf2 q 1 opG` q f M` pq, disc,` 2 2 r _ G ` M` %S 1 _ r _ disc,` pf2 q 4 opG` q f2 pq,

(8.2.1)

under the assumption that 1 (so that is generic and 0). Suppose next that is not generic. Then 1 by denition. However, in this case we will still have to show that 0. Assume the contrary, namely that 0, and consequently that 0. The rst condition implies that N is even and G is split. In particular, we can work with the p dual group G in place of L G, if we choose. The second condition implies that the invariant linear form on GpAq attached to is the pullback of the invariant linear form on the Levi subgroup LpAq. From this property, we shall establish the following consequence of the results of 7.5. Lemma 8.2.1. With our assumption that 0, the localization v of the r parameter P sim pN q at any valuation v factors through the Levi subgroup p of G. p L Proof. We are working with a simple, self-dual parameter in the r r subset sim pGq of sim pN q. It has the usual decomposition b , r P sim pmq, n ,

for positive integers m and n with N mn. The unipotent component has degree n 1, by our assumption that is not generic. The generic component has a determinant , which is an id`le class character with e 2 1, since is self dual. It will not be hard to show that the generic degree m is even, and that the character equals 1. These properties follow from a global argument we shall give at the end of the proof. In the meantime, we shall establish the lemma under the assumption that they hold. This part of the argument is purely local. We shall describe it in local notation that is compatible with the discussion of 7.5. We therefore take F to be local until near the end of the proof, at which r point we will revert to the global notation. We take G P Esim pN q to be a r split endoscopic datum over F , and P ` pGq to be a local parameter. We are assuming that N mn, where n 1 and m is even, and that is of the form p b n, P ` pmq, n ,

454

8. THE GLOBAL CLASSIFICATION

where n is the irreducible representation of SU p2q of degree n and has determinant equal to 1. The generic factor is a self-dual representation
r i1 i i

of LF , for positive integers i and distinct simple elements i P sim pmi q such that m i mi ` ` r mr . The split group G has a Siegel maximal Levi subgroup L GLpN {2q. Our primary assumption on is that the stable character r f G pq, f P HpGq, is induced from a linear form on the Levi subgroup LpF q of GpF q. We must use this to deduce that the homomorphism p : LF SU p2q G p p factors through the subgroup L of G. p into G can of course be taken to be p The embedding of L p x px, x_ q, x P L GLpN {2, Cq. Since is self-dual, there is an involution i i_ on the set of indices i such that i_ _ and i_ i . An orbit of order 2 contributes a i subrepresentation n _ n i pi b q i pi b q p of that factors through L. Similarly, for any i with _ i , the even part i of i contributes a subrepresentation 2 1 pi b n q, i
1 i

r i {2s,

p that also factors through L. If i is odd, however, there will be a supplementary subrepresentation pi b n q of whose image cannot also be embedded p in L. We have therefore to show that the set of indices I ti : _ i , i
i

oddu

is empty. Suppose rst that F C. A component i is then a quasicharacter on C , which is self dual if and only if it equals 1. If there is an i with i 1, the corresponding multiplicity i must be even, since m is even, and the other components occur in pairs. Therefore I is empty and factors p through L, as required. Suppose then that F is real or p-adic. We assume that the subset of indices I is not empty. Our task is to obtain a contradiction of the condition on the stable character f G pq. By the remarks above, it suces to show that if pi b n q
iPI

8.2. PROOF BY CONTRADICTION

455

and N

iPI

mi n,

r r and if G is the datum in Esim pN q such that belongs to pG q, then the twisted character G r f p q, f P HpG q,

is not induced from the maximal Levi subgroup L of G . The answer lies in the decomposition (2.2.12) (with pG , q in place of pG, q) of this stable character in terms of standard stable characters. Since standard stable characters on G are dened directly by the twisted transfer of standard characters from GLpN q, it suces to consider the summands in the decomposition (2.2.9) (with pN , q) in place of pN, q of the twisted irreducible character ` r r r r r f P HpN q, f f,N p q tr pf q ,

into twisted standard characters. We shall apply the character formulas of 7.5 to the direct summands ,i i b n , i P I , of . As an irreducible representation of LF of degree mi , i is self-dual and therefore unitary. If F is p-adic, LF is of course a product WF SU p2q. We write i r b k`1 , mi mr pk ` 1q, where r is an irreducible unitary representation of LF of degree mr , and k is a nonnegative integer. The irreducible character on GLpmi n, F q attached to n pi b n q then equals r pkq, in the notation of 7.5. If F R, mi equals 1 or 2. In case mi 2, i parametrizes a self-dual representation of GLp2, Rq in the relative discrete series. This in turn is parametrized by a positive integer k P N. The irreducible character on GLp2n, F q attached to pi b n q equals n pkq, in the notation of 7.5. In case mi 1, there are two possibilities for the self-dual linear character i of R , either the trivial character 1 or the sign character R . But it follows from the fact that m is even that there is a second index i1 P I such that i1 R i . The irreducible character on GLp2n, F q attached to pi i1 q b n then equals n p0q, again in the notation of 7.5. We thus obtain a decomposition n p q p,i q r pkq
iPI pr,kq

from (7.5.13) and the surjective correspondence of indices i pr, kq, i P I , k 0, where r is understood to be trivial if F R. The correspondence is almost a bijection, in that the bre of pr, kq in I consists of one element unless F R and k 0, in which case it is a pair ti, i1 u. We x an index pr1 , k1 q

456

8. THE GLOBAL CLASSIFICATION

with k1 maximal. If F R, pr1 , k1 q k1 is uniquely determined. In this case, we have k1 1. For if k1 0, we obtain an identity
r i1

pi q i

iPI

i 1 R R ,

which contradicts the condition 1. If F is p-adic, there could be several r paired with k1 . We may as well choose r1 from among these so that mr1 is maximal. We shall apply Lemma 7.5.2 and Corollary 7.5.5 to the direct summand n pk q of p q, or rather the tempered constituent n, pk q of this sumr1 1 r1 1 mand. This gives an extremal factor r1 pk1 ` n 1, 0q of r1 pk1 ` n 1, 0q, which can be compared with the other factors r p , q from the decompok sitions (7.5.3) and (7.5.4) of any direct summand r pkq of p q. It would not be hard to see directly that n pk1 ` n 1, 0q is distinct from all of these other factors. This would lead us to the required contradiction. However, it is perhaps more natural (if a little less direct) simply to restrict our attention to the tempered constituent n, r pkq p ,i q p q
iPI pr,kq

of p q. That this suces is a consequence of the parametrization of standard characters (ordinary or twisted) for GLpN q by the associated r set of Langlands parameters (either pN q or pN q, and the fact that these objects form a basis of the space of all virtual characters (ordinary or twisted) for GLpN q. For it can then be seen, using the germ expansions of twisted characters around singular points for example, that if a given r nite linear combination of twisted standard characters tu is induced from r a distribution on the maximal Levi subset ` r r L pF q GLpN {2, F q GLpN {2, F q pN q r r r of GpN , F q, the same is true of each . (The symbol pN q here denotes the standard outer automorphism of GLpN q, or rather, its restriction to r the Levi subgroup L L0 . It is not to be confused with any of the given r In other words, the Langlands parameter of each r twisted characters .) factors through the L-group of L . We are assuming that the twisted character r r r p q f,N f,N p q r is induced from such a . We know from Lemma 7.5.2(i) and Corollary r 7.5.5 that the twisted tempered character p q attached to p q oc r curs in the decomposition (2.2.9) of p q into twisted standard characters. To obtain a contradiction, we need only show that the self-dual Langlands r parameter of p q does not factor through the L-group of L0 .

8.2. PROOF BY CONTRADICTION

457

Lemma 7.5.2(ii) tells us that the factor r1 pk1 ` n 1, 0q occurs with multiplicity 1 in the direct summand n, pk1 q of p q. Here we are relying r1 on the fact that k1 1 if F R (a condition that need not hold if F is p-adic, but which is then superuous to Lemma 7.5.2(ii)). The integral argument pn ` k1 1q in this factor is maximal. It is in fact strictly larger n, that the arguments in any of the other factors of direct summands r pkq of p q, except for the maximal arguments from pairs pr, k1 q, r r1 ,

in the p-adic case. But the standard characters r pn ` k1 1, 0q and ` r1 pn`k1 1, 0q for GL mr pk1 `1q are distinct, since the same is true of the corresponding Langlands parameters. It follows that there is a self-dual factor in the decomposition of p q that occurs with multiplicity 1. In other words, there is an irreducible self-dual subrepresentation of the Langlands r parameter P pN q that occurs with multiplicity 1. This precludes r the possibility that factor through the L-group of L0 , and gives us the

required contradiction. We have shown that the existence of indices in I leads to a contradiction. The set I is therefore empty. This implies that the parameter r p P pGq factors through L, as we have seen. We have thus established the assertion of the lemma, in local notation, under the assumption that m is even and 1. This assumption was actually used only in the case that F is archimedean, but of course, we still have to justify it. We return to the global notation we have been using throughout Chapter 8. Then F , G and b are the global objects from the beginning of the proof. It remains to show that m degpq is even, and that the global id`le class character equals 1. e Let v be any valuation of F at which G and are unramied. We are free to apply the local argument above to the completion v v b n of . Since the assumption on m and v was not needed in this case, the argument tells us that the given condition on the stable character
G fv pv q,

r fv P HpGv q,

namely that it is induced from LpFv q, implies that the local parameter v p p factors through the image of L in G. This in turn implies that m is even, as required, and that the local determinant v equals 1. It then follows from Tchebotarev density theorem that the global determinant equals 1, as required. In particular, v 1 for any valuation v of F , the condition we used in the archimedean case of the local argument above. This completes the global argument, and hence the proof of the lemma.

458

8. THE GLOBAL CLASSIFICATION

The lemma implies that any localization `,v of ` factors through the p diagonal image of L in the Levi subgroup p p L L GLpN {2, Cq GLpN {2, Cq p p p of G` . The centralizer of the image of L in G` can be identied with the connected group GLp2, Cq GLp2, Cq, which contains the global centralizer Op2, Cq S` . It follows that S` maps to the identity component of S`,v . Applying this to the nontrivial point x1 in S` , we see from the global intertwining relation r that fG` p` , x1 q equals f M` pq, for any f P HpG` q. In other words, the condition (8.1.7) holds in the present case, even though 1. It then follows from Corollary 5.3.3 and (8.1.5) that
` Sdisc,` pf q

1 4

opG` q f M` pq. r

r If f _ P HpG_ q belongs to some compatible family that contains f , it again ` follows from Corollary 5.3.3 and (8.1.5), in combination with the global in_ tertwining relation and the fact that the group S` Spp2, Cq is connected, that ` _ G_ ` Sdisc,` pf _ q 1 opG_ q fG_ p` , x1 q ` f _,LL p q 4 r ` ` ` M _ 1 r 4 opG` q f ` pq ` f M` pq 1 opG_ q f M` pq. 2 r ` We thus have a second family of identities $ G Sdisc, pf1 q 0, _ G S _ G & disc, pf1 q opG_ q f1 pq, r S ` pf2 q 1 opG` q f2 ` pq, disc,` 4 r _ G` % M _ Sdisc,` pf2 q 1 opG_ q f2 ` pq, 2 r `
G M

(8.2.2)

under the assumption that 0 (so that is not generic and 1), _ _ and for functions f1 , f1 , f2 and f2 as in (8.2.1). Notice the similarity with (8.2.1). The two families (8.2.1) and (8.2.2) are completely parallel, but with G and G` interchanged with their counterparts G_ and G_ . `

8.2. PROOF BY CONTRADICTION

459

For comparison, we can also write down what the equations in (8.2.1) and (8.2.2) would be if 1 and 0, as expected. We obtain $ G G r Sdisc, pf1 q opGq f1 pq, _ G S _ & disc, pf1 q 0, (8.2.3) S G` pf2 q 0, disc,` _ G` % M _ Sdisc,` pf2 q 1 opG_ q f2 ` pq, 4 r ` under the assumption 1 and 0 (which imposes no further condi_ _ tions on ), and for functions f1 , f1 , f2 and f2 as in (8.2.1) and (8.2.2). These follow from Corollary 5.3.3 and discussion above, particularly (8.1.5) and the preceding applications of the global intertwining relation. They would also follow easily from the (as yet unproven) stable multiplicity formula. We are of course trying to establish the conditions 1 and 0. Our task is therefore to focus on the equations (8.2.1) and (8.2.2) that would follow if the conditions are not met. However, the equations in (8.2.1) and (8.2.2) are vaguely disturbing. They contain nothing that is at odds with anything we have established so far. As a matter of fact, their symmetry seems to give them an air of authority, which is reinforced by their relative simplicity. But the equations are based on premises that we must unequivocally refute! We are left with the uncomfortable feeling that our methods may have run their natural course. Given the symmetry between (8.2.1) and (8.2.2), we appear to have no choice but to look for a uniform argument that applies to both problems. We shall introduce a second supplementary parameter, which at rst sight seems to be unnatural and without promise. The surprise is that it actually works! The new parameter will be seen in fact to be quite natural, and to be exactly what is required to contradict the assumptions 1 of (8.2.1) and 0 of (8.2.2). For the given pair pG, q, we dene our second supplementary parameter by `` 3 . r Then `` belongs to pN`` q, for N`` 3N . Remember that we are assuming that N is even and 1, so that G and G_ are distinct split groups r r in Esim pN q. Let G`` and G_ be the split groups in Esim pN`` q whose dual `` p p p p groups G`` and pG_ q^ contain the respective products G G G and `` _ q^ pG_ q^ pG_ q^ . We then have the maximal Levi subgroups pG M`` M` G GLpN q G and
_ _ M`` M` G_ GLpN q G_

460

8. THE GLOBAL CLASSIFICATION

of G`` and G_ respectively. The product can then be treated `` r as a parameter for M`` whose image in pN`` q equals `` . Given an associated compatible family of functions ( r r (8.2.4) f P HpG q : G P Eell pN`` q , we shall write f f and f _ f as usual in the cases G G`` and G G_ . `` The stage is now set for the lemma that is the key to the nal classication. This lemma is founded on the techniques of 5.3, but relies also on the character formulas of 7.5 embodied in Lemma 8.2.1. Lemma 8.2.2. For the given pair pG, q, r r G P Esim pN q, P sim pN q,

assume that N is even and 1. In addition, assume that either 1 as in (8.2.1), or 0 as in (8.2.2). Then the sum ` (8.2.5) tr Rdisc,`` pf q ` 1 f M`` p q 2
r G PEsim pN`` q

vanishes for any compatible family of functions (8.2.4). Proof. Once again, we have to modify the arguments from Chapter 4. The proof will be loosely modeled on that of Lemmas 5.3.1 and 5.3.2 for ` . However, our exposition here will be slightly dierent, since we now have the local theorems at our disposal. In particular, analogues of the terms (5.3.22) and (5.3.23) need not be an explicit part of the discussion here, thanks to the local and global intertwining relations. We x a compatible family of functions (8.2.4). The arguments of Chapter 4 would lead ultimately to the vanishing of the sum ` rpN`` , G q tr Rdisc,`` pf q , (8.2.6)
r G PEsim pN`` q

if we could assume inductively that the global theorems were valid for proper subparameters of `` . We cannot do this, of course, because we have assumed that one of the contrary conditions 1 or 0 holds. However, we can use this condition, as it is reected in either (8.2.1) or (8.2.2), to modify the expected formula for (8.2.6). The modications of (8.2.6) are of three sorts. They arise from the endor scopic contributions of proper products G G1 G2 in the set Eell pN`` q, 1 G G in the sets from endoscopic contributions of proper products G 1 2 r Eell pG q attached to simple data G P Esim pN`` q, and from spectral conr tributions from proper Levi subgroups M of the groups G P Esim pN`` q. As we have seen, explicitly in the proof of Lemma 5.3.1 and implicitly in that of Lemma 5.3.2, most such contributions reduce immediately to zero. We shall have to consider only one pair of contributions of each of the three

8.2. PROOF BY CONTRADICTION

461

kinds. In each case, we will have to subtract the expected value of the contribution from the actual value. The dierence will then have to be added or subtracted, as a correction term for (8.2.6), in order that the resulting expression vanish. As in the earlier proofs, the sign of this last operation (addition or subtraction) is dictated by the role of the given contribution in the standard model. Since these signs are critical, it would be a good idea to review again the process by which they are determined. We know that we can write (8.2.6) as the sum of ` rpN`` , G q trpRdisc,`` pf q 0 Sdisc,`` pf q (8.2.7)
r G PEsim pN`` q

and (8.2.8)
G rpN`` , G q 0 Sdisc,`` pf q,

r0 G P Eell pN`` q,

r r r0 where Eell pN`` q is the complement of Esim pN`` q in Eell pN`` q, since the condition (5.2.15) of Lemma 5.2.3 holds for `` . This is the identity, with (8.2.6) on the left hand side and the sum of (8.2.7) and (8.2.8) on the right, that is to be our starting point. The rst pair of modications describes the obstruction to the expected vanishing of the summands in (8.2.8). To simplify the right hand side of the identity, we will want to move the obstructions to the left hand side. In other words, we add them to (8.2.6), since the summands come with minus signs. For the remaining terms on the right hand side, we will take what we can from Corollaries 4.3.3 and 4.4.3. We can certainly write ` r G P Esim pN`` q, pf q I 1 pf q I 1 pf q, tr Rdisc, pf q S
``

disc,``

end

spec

stand for the proper parts of the respective expansions where and (4.1.1) and (4.1.2) (namely, the sums taken only over M G and G1 G respectively, but of course with pG , `` q in place of pG, q). Corollaries 4.3.3 and 4.4.3 (and their comparison by Proposition 4.1.1 that in the case of pG, q culminated in (4.5.4)) pertain to the expected values of the terms on the two sides of this last formula. Since the expected value of the trace ` rdisc,`` pf q tr Rdisc,`` pf q is 0, they tell us that ` tr Rdisc,`` pf q 0 Sdisc,`` pf q 0 rdisc,`` pf q 0 s disc,`` pf q
1 1 0 Iend pf q 0 Ispec pf q,

1 Ispec

1 Iend

r G P Esim pN`` q,

1 1 1 1 where 0 Iend pf q and 0 Ispec pf q stand for the variance of Iend pf q and Ispec pf q respectively from their expected values. The remaining modications describe the obstruction to the expected vanishing of these last two terms. 1 The second of the three pairs applies to 0 Iend pf q. To move the resulting obstruction to the left hand side, we must subtract it from (8.2.6). The 1 third pair applies to 0 Ispec pf q. To transfer the resulting obstruction in this

462

8. THE GLOBAL CLASSIFICATION

last case, we must add it to (8.2.6). The expression (8.2.6), thus modied by the three kinds of correction, will then vanish. The arguments for the two possible conditions 1 or 0 are essentially parallel, so we shall only treat one of them in detail. We assume until near the end of the proof that the condition 0 is in force. Then is not generic, equals 1, and the equations (8.2.2) hold. The rst pair of corrections applies to the linear forms (8.2.9)
1 2 Sdisc,`` pf q Sdisc, pf1 q Sdisc,` pf2 q,

r0 G P Eell pN`` q,

r0 where G G1 G2 is either of the groups GG_ or G_ G` in Eell pN`` q, ` f f is the corresponding function in (8.2.4). If G equals and f 1 2 G G_ , the actual value of (8.2.9) vanishes, according to the rst equation ` in (8.2.2), while the expected value equals r ` G 1 opGq opG_ q f1 pq f2 4 r
M`

pq,

by the rst and fourth equations in (8.2.3). If G equals G_ G` , the actual value of (8.2.9) is
1 4 G opG_ q opG` q f1 pq f2 r r M`

pq,

by the second and third equations in (8.2.2), while the expected value vanishes by the second (or third) equation in (8.2.3). In each case, we subtract the expected value from the actual value, and then multiply the dierence by the coecient (8.2.10) rpN`` , G q
1 8 1 4

r opG q1

1 4

r r opG1 q1 opG2 q1 .

We obtain two equal contributions, whose sum (8.2.11) f M`` p q,

must be added as a correction term to (8.2.6). The second pair of corrections applies to two terms in the analogue for `` of (4.1.2). These are the linear forms (8.2.12) 1 1 2 r Sdisc,`` pf 1 q Sdisc, pf1 q Sdisc,` pf2 q, G1 P Esim pG q, G P Esim pN`` q, r where G1 G1 G2 is either of the groups G G` P Eell pG`` q or r G_ G_ P Eell pG_ q, and f 1 f1 f2 is a function in HpG1 q with the ` `` 1 q as the chosen compatible family. If G1 equals G G , r same image in SpG ` the actual value and the expected value of (8.2.12) both vanish, by the rst equation in (8.2.2) and the third equation in (8.2.3). If G1 equals G_ G_ , ` the actual value of (8.2.12) is r ` G 1 opG_ q opG_ q f1 pq f2 2 r
M`

p` q,

by the second and fourth equations in (8.2.2), while the expected value vanishes, by the second equation in (8.2.3). In each case, we subtract the

8.2. PROOF BY CONTRADICTION

463

expected value from the actual value, and then multiply the dierence by the coecient (8.2.13) 1 rpN`` , G q pG , G1 q 4 opG q1 OutpG , G1 q 1 opG1 q1 opG2 q1 . r r 4 r This time we must subtract the result from (8.2.6). The two minus signs cancel, and we obtain another copy of (8.2.11) to be added as a correction term to (8.2.6). The third pair of corrections applies to two terms in the analogues for `` of (4.1.1). These are the linear forms ` r (8.2.14) tr MP ,`` pw q IP ,`` pf q , G P Esim pN`` q, where G is either G`` or G_ , M is the corresponding Levi subgroup `` _ M`` or M`` , and w is the associated element in W`` ,reg pM q. This requires a little more discussion. Suppose rst that G equals G`` . Recall that M`` is isomorphic to GLpN q G. The actual discrete spectrum for G and vanishes under our assumption 0, since the rst equation in (8.2.2) tells us that the corresponding stable linear form vanishes. The actual value of (8.2.14) therefore vanishes. To describe the expected value, we note that the centralizer group S `` is isomorphic to SOp3, Cq. Because is not generic, the induction hypothesis underlying the proof of Lemma 4.3.1 in 4.6 is valid. By a very elementary case of the discussion in 4.6, the global normalizing factor implicit in (8.2.14) therefore equals p1q. If we then combine the rst equation in (8.2.3) with the global intertwining relation, we see that the expected value of (8.2.14) is rpGq fG`` p`` , x1 q rpGq f M`` p q, o o where x1 stands for the only element in the trivial group S`` . We must subtract the expected value from the actual value, multiply the dierence by the coecient (8.2.15) rpN`` , G`` q |W pM`` q|1 | detpw 1q|1 1 opG`` q 1 1 opGq1 , 2 r 4 8 r and add the result to (8.2.6). This gives us a third copy of (8.2.11) to be added to (8.2.6). Suppose nally that G equals G_ . In this case, we cannot quite `` apply the discussion of 4.6 to the global normalizing factor in (8.2.14), because our assumption 0 violates the earlier premises. However, we can still represent the normalizing factor as a product of quotients of two automorphic L-functions. One is the Rankin-Selberg L-function Lps, q. Indeed, it is built out of local Rankin-Selberg factors Lps, v v q attached to localizations
_ v v : LFv SU p2q pM`` q^ GLpN, Cq pG_ q^

GLpN, Cq GLpN, Cq

464

8. THE GLOBAL CLASSIFICATION

of . (We shall recall before how v can be treated as a local parameter for the group G_ .) Since is self-dual, it has a pole at s 1. The other is the Langlands-Shahidi L-function attached to the parameter ` for G_ . ` _ Since S` is isomorphic to SLp2, Cq, this L-function also has a pole at s 1. We are of course using our induction hypothesis here that Theorem 1.5.3(a) is valid for the generic component of , as we did in discussing the case G G`` above. The two L-functions each contribute a factor p1q to the product, so the global normalizing factor equals 1 in this case. (By contrast, in the case G G`` above, S` is isomorphic to Op2, Cq, and the corresponding L-function does not have a pole, and therefore does not contribute a second factor p1q.) Continuing our analysis of the case G G_ , we consider the local `` normalized intertwining operators in (8.2.14). According to Lemma 8.2.1 and our assumption 0, any localization v of factors through the p p p Levi subgroup L of G. Since L also represents a Levi subgroup of pG_ q^ , we can treat v as a local parameter for G_ as well as for G. The image in pG_ q^ of the corresponding localization `` ``,v v v v of `` is then contained in the threefold diagonal image of pG_ q^ . The centralizer of this diagonal image, taken modulo the center of pG_ q^ , is `` isomorphic to the connected group SOp3, Cq. Since w w_ represents the nontrivial element in the Weyl group of SOp3, Cq, we can choose a _ local representative of w_ in the connected centralizer pS`` ,v q0 . The local normalized intertwining operator therefore equals the identity. We have shown that if G G_ , the local and global factors of the `` operator MP ,`` pw q in (8.2.14) are all equal to 1. The operator is therefore trivial. In fact, from the global intertwining relation and the second equation in (8.2.2), we see that (8.2.14) itself equals the linear form
_ r r opG_ q fG_ p`` , x1 q opG_ q pf _ qM`` p q opG_ q f M`` p q. r ``
_

On the other hand, the expected value of (8.2.14) is equal to 0, by the second equation in (8.2.3). Subtracting the expected value from the actual value, and multiplying by the coecient (8.2.16) _ _ 1 r `` 4 rpN`` , G_ q |W pM`` q|1 | detpw`` 1q|1 2 opG_ q 1 1 opG_ q1 , `` 8 r we obtain a fourth copy of (8.2.11) to be added to (8.2.6). We have obtained four correction terms in all, each of which is equal to (8.2.6). The sum of (8.2.6) with the total correction term `1 4 8 f M`` p q 1 f M`` p q 2 therefore vanishes. Since the sum is equal to the given expression (8.2.2), we have obtained a proof of the lemma in case 0. Assume now that 1. Then is generic, 0, and the equations (8.2.1) hold. The structure of the proof is the same in this case, with the only

8.2. PROOF BY CONTRADICTION

465

minor dierences in detail resulting from our use of the equations (8.2.1) in place of (8.2.2). We can be brief. Suppose that the group G in (8.2.9) equals G G_ . Then the actual ` value of (8.2.9) equals
1 4

r r ` opGq opG_ q f M`` p q,

while the expected value is r ` 1 opGq opG_ q f M`` p q, 4 r by the rst and fourth equations in (8.2.1) and (8.2.3). If G equals G_ G` , the actual and expected values of (8.2.9) both vanish, by the second equations in (8.2.1) and (8.2.3). The product of the total dierence with the coecients (8.2.10) is equal to the linear form (8.2.11) we obtained before. Suppose that the group G1 in (8.2.12) equals G G` . Then the actual value of (8.2.12) equals r 1 opGq opG` q f M`` p q, 2 r while the expected value vanishes, by the rst and third equations in (8.2.1) and (8.2.3). If G1 equals G_ G_ , the actual and expected values of (8.2.12) ` both vanish, by the second equations in (8.2.1) and (8.2.3). The product of the total dierence with the coecient (8.2.13), when subtracted from (8.2.6), again gives a second copy of the correction term (8.2.9). Suppose that the groups G and M in (8.2.14) equal G`` and M`` . The actual discrete spectrum for M`` and is equal to what is expected, by the rst equations in (8.2.1) and (8.2.3). Since the group S`` equals Op3, Cq, the normalized intertwining operator in (8.2.14) equals 1, as expected. The global normalizing factor implicit in (8.2.14) is therefore the only quantity that diers from its expected value. The expected value equals the sign of w w in SOp3, Cq, which is p1q. The actual value equals 1, as in our discussion of G G_ for the case 0 above, this `` time using the denition of 1. Appealing again to the rst equations of (8.2.1) and (8.2.3), we see that the actual value of (8.2.12) is r opGq fG`` p`` , x1 q opGq f M`` p q, r while the expected value equals p1q times this quantity. The product of the dierence with the coecient (8.2.15) gives two more copies to (8.2.11) to be added as correction terms to (8.2.6). Finally, suppose that the groups G and M in (8.2.14) equal G_ and `` _ . The discrete spectrum for M _ and vanishes, as expected, by M`` `` the second equations in (8.2.1) and (8.2.3). The actual and expected values of (8.2.14) therefore both vanish in this case, and give no further correction terms for (8.2.6). The total correction term in this case thus equals `1 1 2 8 f M`` p q ` 4 f M`` p q 1 f M`` p q, 2

466

8. THE GLOBAL CLASSIFICATION

as it did in the earlier case. The given expression (8.2.2) therefore again vanishes. We have completed the proof of the lemma in the remaining case 1. With the proof of Lemma 8.2.2, we have nally obtained the result we want. It remains only to appeal one last time to Proposition 3.5.1. The linear form (8.2.17) f M`` p q, r f P HpG`` q,

in (8.2.5) is an induced character. In fact, it is induced from the nonzero character on the Levi subgroup M`` pAq obtained from the local constituents v v of . In particular, it is a nonnegative integral combination of irreducible characters. Since the same is true of the linear forms in f in (8.2.5), the entire expression (8.2.5) is a nonnegative linear combination of irreducible characters of the general form (3.5.1). If the expression vanishes, we can indeed apply Proposition 3.5.1. It will tell us that the corresponding coecients all vanish. Since there are nonzero coecients in (8.2.17), this cannot happen. The expression (8.2.5) therefore does not vanish. We have arrived at a contradiction to the premise of Lemma 8.2.2. In other words, we have established that 1 and 0. It follows that Theorem 1.5.3(a) holds if is generic, and that the stable multiplicity formula of Theorem 4.1.2 is valid in the remaining case that is not generic. These are the two assertions left over from the last section. r The remaining global properties for simple parameters P sim pN q are in the assertions of Theorems 1.5.2 and 4.2.2. They follow immediately. Indeed, given the stable multiplicity formula for , we appeal to (4.4.12) and Lemma 4.7.1 as above to establish the required contribution (4.7.10) to the spectral multiplicity of Theorem 1.5.2. (In case is generic, we are also implicitly relying on the fact that Lemma 8.1.1 includes a proof of Theorem 1.4.2.) If pG, q is as in Theorem 4.2.2, the simple parameter r r lies in the complement of 2 pGq of 2 pGq. The assertion (b) of this theorem is therefore irrelevant, while assertion (a) again follows from the fact that the right hand side of (8.1.3) vanishes. These are the last assertions to be r proved for the simple parameter P sim pN q. Since the simple parameters were all that remained after the last section, we have now established the r global theorems for any parameter P pN q. This completes the induction argument on N begun in the previous section, and therefore completes our proof of the global theorems. In summary, we have now resolved all of the induction hypotheses. We have consequently proved all of the theorems. These are the generic local theorems established in Chapter 6, the nongeneric local theorems treated in Chapter 7, and the global theorems proved in this chapter. In particular, r r the analogue of Corollary 5.4.7 for the full set Fsim pN q sim pN q of simple global generic parameters is valid. We shall restate it as a nal corollary of this discussion, for easy reference.

8.3. REFLECTIONS ON THE RESULTS

467

r Corollary 8.2.3. Suppose that P sim pN q is any simple, self-dual, r generic parameter of rank N , and that G belongs to Esim pN q. Then the following three conditions on the pair pG, q over our global eld F are equivalent. G r (i) The linear form Sdisc, on HpGq does not vanish. (ii) Theorem 1.4.1 holds for , with G G. (iii) The global quadratic characters G and are equal, and the global L-function condition 1 of Theorem 1.5.3(a) holds. Proof. The proof follows from the global theorems we have now established. In particular, it is identical to that of Corollary 5.4.7, the special r case for the global family F of 5.4. Corollary 8.2.3 tells us that any of the three equivalent conditions (i) r (iii) can serve as a denition for the set sim pGq of simple, generic global parameters for G. Thus ends the last ambiguity from the construction of r 1.4, which led to the complete set of global parameters pGq and the theorems they satisfy. Once again we observe the local-global symmetry, here between the global corollary we have just proved, and its local counterpart Corollary 6.8.1. We can now go on to other things. In the rest of this chapter, we shall discuss a few ramications of what we have done. We will then study the r representations of inner twists of groups G P Esim pN q in the nal Chapter 9. 8.3. Reections on the results Having nally proved the theorems, we shall pause for a few moments of reection. We will begin the section with a brief overview of the results, taken from the general perspective of characters and multiplicities. We will then pose some questions that are particular to the even orthogonal groups SOp2nq. We will complete the section with a proof that local and global L-packets contain generic representations. r We x one of the quasisplit groups G P Esim pN q over our eld F . We recall that the integer N equals 2n, 2n ` 1 or 2n, according to whether G equals a simple group SOp2n ` 1q, Spp2nq or SOp2nq from the innite p family Bn , Cn or Dn , with corresponding dual group G equal to Spp2n, Cq, SOp2n ` 1, Cq or SOp2n, Cq. In the second and third cases, G comes with an arithmetic character (a character on F if F is local or A {F if F is global) of order 1 or 2. In the third case, determines G as a quasisplit inner twist of the corresponding split group. The group G is of course split in the other two cases. The role of in the second case G Spp2nq is to specify the L-embedding of L G into the L-group of GLpN q that determines G as twisted endoscopic datum for GLpN q. Suppose for the moment that F is local. The local Langlands group LF is then dened as the split extension (1.1.1) of WF . We have classied the

468

8. THE GLOBAL CLASSIFICATION

r irreducible representations of GpF q in terms of packets indexed by the r set pGq of equivalence classes of Langlands parameters : LF
L

G.

p p Equivalence is dened by G-conjugacy as usual when G equals Spp2n, Cq r is the L-packet conjectured by Langor SOp2n ` 1, Cq. In these cases, lands for any connected reductive group over F . When G equals SOp2n, Cq, equivalence is dened by conjugacy under the exension Op2n, Cq of Z{2Z by SOp2n, Cq. In this case, the packet of is still an L-packet if the Op2n, Cqorbit of equals its SOp2n, Cq-orbit . If the Op2n, Cq-orbit contains two r SOp2n, Cq orbits, however, is a set of pairs of irreducible representations. We have thus established the full local Langlands correspondence for the split groups SOp2n ` 1q and Spp2nq, and a slightly weaker form of the correspondence for the quasisplit groups SOp2nq. It is worth emphasizing that the endoscopic classication of representar tions is by characters. That is, the elements in a packet are represented by their characters. These are dened in terms of the nite 2-group
0 p S S {S ZpGq ,

F GalpF {F q,

and twisted characters for general linear groups. Let us be more explicit. The distributional character ` fG pq tr pf q , f P HpGq, of any irreducible representation P pGq of GpF q is a locally integrable function. In other words, fG pq G p, xq f pxqdx IG p, q fG pq d,
Greg pF q reg pGq

where reg pGq G-reg pGq is the space of strongly G-regular conjugacy classes in GpF q, equipped with its natural measure, and IG p, q |Dpq| 2 G p, q,
1

P reg pGq,

is the normalized character. This assertion is a fundamental theorem of Harish-Chandra, which is at the heart of harmonic analysis on real and p-adic groups. The endoscopic classication is in terms of tempered representations, or equivalently, their (reducible) analytic continuation to standard representations. For real groups, tempered (and standard) characters have natural formulas, which are based on Harish-Chandras explicit formulas [Ha1] [Ha2] for the characters of discrete series. Characters for p-adic groups are more complex, but they have many interesting and well understood qualitative properties. It is not hard to describe endoscopic transfer in terms of normalized characters. Recall that the tempered representations of GpF q correspond to r r the subset bdd pGq of parameters P pGq of bounded image. We will need a slightly dierent convention for treating the constituents of the associated

8.3. REFLECTIONS ON THE RESULTS

469

r r packets , since they are elements in the set pGq of orbits in pGq under the group r r r OpGq OutN pGq AutN pGq{r N pGq. Int For any in this set, the sum r IG p, q IG p , q,

P pG, q,

r over the OpGq-orbit pG, q of in pGq depends only on the image of in the quotient r r reg pGq reg pGq { OpGq. r We take the quotient measure on reg pGq inherited from reg pGq. We then observe that r r (8.3.1) fG pq IG p, q fG pq d, f P HpGq,
r reg pGq

r r for any P pGq, and in particular, for in any packet . The stable anar logue of this distribution is the linear form f G pq attached to P bdd pGq. It can be expanded as an integral r r f G pq S G p, q f G pq d, f P HpGq,
r reg pGq

r r r over the space reg pGq G-reg pGq of OpGq-orbits of strongly G-regular stable conjugacy classes in GpF q, equipped with the measure inherited from r reg pGq. The normalized stable character S G p, q in the integral is obtained by transfer from GLpN q, as we shall see explicitly in a moment. r The function S G p, q of course has an analogue r r 1 S 1 p1 , 1 q S G p1 , 1 q, r 1 P G-reg pG1 q,

r for the preimage pG1 , 1 q of any given pair p, sq. For any f P HpGq, we can then write 1 1 r f p q S 1 p1 , 1 q f 1 p 1 q d 1 r G-reg pG1 q r S 1 p, 1 q p 1 , q fG pq d 1 r G-reg pG1 q r Preg pGq r S 1 p1 , 1 q p 1 , q fG pq d,
r reg pGq r 1 PG-reg pG1 q
1

by the denition of f 1 f G , and a change of variables as, for example, in [A11, Lemma 2.3]. On the other hand, if x is the image in S of the given

470

8. THE GLOBAL CLASSIFICATION

point s P S ,ss , we can also write f 1 p1 q xx, y fG pq

r P

r reg pGq

r P

r xx, y IG p, q fG pq d,

by (2.2.6) and (8.3.1). We thus obtain an identity r r S 1 p1 , 1 q p 1 , q xx, y IG p, q

1

of the two integrands. An inversion on the group S then yields the general formula r r (8.3.2) IG p, q xx, y1 S 1 p1 , 1 q p 1 , q,
xPS 1 P r G-reg pG1 q

r r in which and lie in and reg pGq respectively, and pG1 , 1 q corresponds to x according to the usual mappings pG1 , 1 q p, sq p, xq, s P S ,ss . If pG1 , 1 q equals pG, q and is the stable conjugacy class of , the second last formula reduces to r r r S G p, q IG p, q, P reg pGq,
r P

as might be expected. This is not to be regarded as a means to compute stable characters, however, since stable characters are really the primary objects. We instead use the twisted character r r r r r r (8.3.3) fN pq IN pr , q fN prq dr, f P HpN q,
r reg pN q

r attached to any self-dual parameter P bdd pN q for GLpN q, and in particr bdd pGq. The left hand side here equals the trace of pf q, ular, to any P r r r by denition (2.2.1). The domain of integration reg pN q on the right hand side is the set of strongly regular, twisted conjugacy classes in GLpN, F q. The kernel function in the integrand is the normalized twisted character
1 r r IN pr , q |Dprq| 2 N pr , q, r r

r r P reg pN q,

whose existence has been established in [Clo]. An analysis similar to that of (8.3.2), using (2.2.1) and (8.3.3) in place of (2.2.6) and (8.3.1), yields the general formula r r r r (8.3.4) S G p, q IN pr , q p, q ,
Preg pN q r r

r r for any P bdd pGq and P N -reg pGq.

8.3. REFLECTIONS ON THE RESULTS

471

r The formulas (8.3.2) and (8.3.4) classify the (OpGq-orbits of) irreducible tempered representations of GpF q in terms of twisted, tempered characters for general linear groups. Suppose for example that lies in the subset r 2 pGq of square integrable (orbits of) representations. Then the parameter r r r for the packet of lies in the subset 2 pGq of bdd pGq. The endoscopic 1 that index the terms in the formula (8.3.2) for the character of data G r consequently lie in the subset Eell pGq of EpGq, and are products G1 G1 G 1 , 1 2 r G1 P Esim pNi1 q, i

r for a partition of N into two even integers Ni1 . The stable characters S 1 p1 , 1 q in the corresponding summands decompose accordingly into products
1 r1 1 r r1 S 1 p1 , 1 q S1 p1 , 1 q S2 p1 , 2 q, 1 2

whose factors can be written in terms of twisted characters on the groups GLpNi1 , F q by the formula (8.3.4). The supplementary coecients on the right hand sides of (8.3.2) and (8.3.4) are concrete functions. The factor xx, y1 in (8.3.2) is just the value at x1 of the character on the 2-group r S attached to . The other coecients p 1 , q and p, q are transfer factors, which we know are subtle objects, but which are nonetheless dened ([LS1], [KS]) by explicit formulas. The local classication embodied in the character formulas (8.3.1)(8.3.4) was proved by global means. We recall that the proof also had a purely local component. It was the collection of orthogonality relations from 6.5, for elliptic tempered characters on GpF q, and twisted, elliptic tempered characters on GLpN, F q. r The local theorems apply also to the more general set pGq of equivalence classes of nongeneric parameters : LF SU p2q
L

G.

r These are complicated by the fact that elements in the packets are generally reducible. However, the character formulas (8.3.1)(8.3.4) remain valid for , with the point x in (8.3.2) being replaced by its translate s x. They lead to explicit character formulas for the reducible elements (which r r r we denoted by in 7.1) in (as the set over S we denoted by in 7.1), in terms of twisted, nontempered characters for general linear groups. The r nongeneric local parameters P pGq are not part of the local classication of admissible representations. They serve rather in support of a global goal, the description of nontempered automorphic representations. But since the constituents of their packets are all unitary, the parameters might still have a role in the local classication of unitary representations. We do have to account for the possible failure of the analogue of Ramanujans conjecture for GLpN q. In other words, we have to work with the larger set r r r ` pGq ` pGq X ` pN q unit unit

472

8. THE GLOBAL CLASSIFICATION

r of local parameters in place of pGq. This is the intermediate set r r r pGq ` pGq ` pGq, unit dened in 1.5 in terms of the irreducible unitary representations of GLpN, F q. r r For any P ` pGq, the elements in the packet are the induced repunit resentations (1.5.1). Their characters are obtained by analytic continuation of characters (2.2.6) that are provided by our local Theorem 2.2.1. In particular, the discussion above applies to the more general local parameters r P unit pGq. We would expect the induced representations in the packets r to be irreducible. We shall state this as a conjecture, even though its proof might be quite straightforward. r Conjecture 8.3.1. Assume that F is local, that G P Esim pN q, and that ` r is a parameter in the set unit pGq. Then the induced representations IP pM, q, r M P M , r in the packet are irreducible and unitary. Rather than attempt a proof of the assertion, let me just append a couple r of brief remarks. There is a bijection from the set ` pN q attached unit r r to GLpN q onto unit pN q. The irreducible representation P unit pN q of GLpN, F q can be regarded either as the Langlands quotient attached to the Langlands parameter , or an irreducible induced representation IP pN q pM pN q, q, for a Levi subgroup M pN q of GLpN q and the Langlands ` quotient M pN q of a parameter M pN q P M pN q . The classication of the unitary dual unit pN q in [Tad1] and [V2] is in terms of such induced representations. In particular, the irreducibility is built into the construction. The problem is to show that the irreducibility property is preserved under the transfer from GLpN q to G. One would have to show that the singularities of the relevant intertwining operators for GpF q are dominated by those for GLpN, F q. Such singularities are governed by exponents of representations [BoW, p. 113] for p-adic F , and by some variant of the conditions of [SpehV], when F is archimedean. In the case that lies in the subset r r r unit pGq ` pGq X pGq unit r of generic parameters in ` pGq, the two assertions of the conjecture are unit not hard to check. Suppose now that F is global. If the fundamental local objects are characters, the fundamental global objects are Hecke eigenfamilies c tcv u. The global theorems classify the Hecke eigenfamilies attached to automorphic representations of G in terms of those for general linear groups. But they do more. Just as the local theorems sort irreducible characters into local packets for G in relating them to GLpN q, so the global theorems characterize the automorphic representations for G attached to eigenfamilies in terms of a global packet, and hence in terms of tensor products of local

8.3. REFLECTIONS ON THE RESULTS

473

characters, even as they relate the eigenfamilies to those of GLpN q. The global process could be regarded as a description of the ramication properties of automorphic representations that are hidden in the unramied data of Hecke eigenfamilies. This is all summarized compactly in the multiplicity formula of Theorem 1.5.2. Recall that Acusp pN q denotes the set of equivalence classes of unitary, cuspidal automorphic representations of GLpN q. When we began in 1.3, we r agreed that this was the basic global object. It contains the family Acusp pN q r of self-dual such representations, which we have also denoted by sim pN q when thinking in terms of global parameters. The disjoint unions r r r Acusp pN q Acusp Acusp pN q Acusp
N N

are then the foundation for all of the global results. Theorem 1.3.2 implies that these sets map bijectively onto the corresponding sets of Hecke eigenfamilies in the disjoint unions r r (8.3.5) Csim Csim pN q Csim Csim pN q,
N N

where ( r Csim pN q c P Csim pN q : c_ c , and Csim pN q Ccusp pN q is our fundamental set of simple Hecke eigenfamilies from (1.3.11). The more concrete data in (8.3.5) can thus also serve as a foundation for the global results. Ideally, one would like to be able to formulate general global properties explicitly in terms of data from (8.3.5). Sometimes one can do so in an r elementary way. For example, Csim pN q is the subset of families c tcv u in Csim pN q that are equal to their corresponding dual families c_ tc_ u. The v same goes for the description in 1.3 of the larger sets of Hecke eigenfamilies n1 n1 ( cpq cpq b cpq cv pq cv pq qv 2 cv pq qv 2 , which are attached to elements in the set sim pN q A2 pN q that parametrizes the automorphic discrete spectrum of GLpN q. These are obtained in an elementary way from families cpq in the basic sets Csim pmq. Sometimes one must use transcendental means. This is the case for the description of the r subset Csim pGq of families c P Csim pN q attached to the given G, according to which of the partial L-functions ` s LS ps, c, S 2 q det I S 2 pcv q qv
vRS

or LS ps, c, 2 q
vRS

` s det I 2 pcv q qv

474

8. THE GLOBAL CLASSIFICATION

has a pole at s 1 (since the partial L-functions have the same behaviour at s 1 as the completed L-functions of Theorem 1.5.3). Sometimes, however, the ties are less direct. For example, it would be hard to claim that the local data at ramied places v can be obtained from the unramied data in this explicit way. Theorems 1.5.1 and 1.5.2 do reduce the question to the case of GLpN q. One might then argue that the answer is in the theory of Rankin-Selberg L-functions, specically the completion of a partial Lfunction LS ps, c1 c2 q to a full L-function Lps, 1 2 q with the appropriate functional equation. At any rate, the global theorems are ultimately statements about data from (8.3.5). In particular, Theorem 1.5.2 describes the automorphic disr crete spectrum of G in terms of parameters P 2 pGq attached to compound Hecke eigenfamilies cpq. These in turn are derived from simple r r families i P Csim pmi q, as in the special case of the subset sim pGq above. It is the multiplicity formula (1.5.3) that provides the quantitative link ber tween global data and local properties. Through the parameters P 2 pGq, r their completions v P pGv q, and the corresponding mappings S
v

Sv

of centralizers, it characterizes the automorphic discrete spectrum explicitly in terms of local characters for G. These are then classied in terms of twisted characters for general linear groups by the formulas (8.3.1)(8.3.4), and their generalizations for nongeneric parameters v . The global theorems of course have other consequences. Since some of these were discussed as they came up in the text, we will not spend more time on them here. We recall only that they include a general theory of Rankin-Selberg L-functions and -factors for pairs of representations of G that is inherited from GLpN q, the special properties of L-functions and -factors given by Theorem 1.5.3, the relation between multiplicities and symplectic -factors implicit in the sign character of Theorem 1.5.2, and the absence of embedded eigenvalues for G established in Chapter 4. Our proof of the local and global theorems is reminiscent of a fundamental theorem from the beginnings of the subject. This is the analytic classication by Hermann Weyl of the representations of compact, simply connected Lie groups [We1]. Weyls proof, which includes the Weyl character formula, is by invariant harmonic analysis. It entails a felicitous interplay of characters and multiplicities, linked by the orthogonality relations satised by irreducible characters. For a proof in the elementary case of U pN q (which we could consider the analogue of our reductive group GLpN q), see [We2, 377385]. These general observations on characters and multiplicities are all implicit in the statements of the various theorems. They have served as a pretext for us to collect our thoughts. They also lead us naturally to the

8.3. REFLECTIONS ON THE RESULTS

475

next topic of discussion, that of even orthogonal groups. This will be divided into two parts, according to whether the integer mpq attached to a parameter equals 1 or 2. We shall see that each part comes also with a broader philosophical question. We x an even orthogonal group r G P Esim pN q, p G SOpN, Cq, N 2n,

r over F . The group OpGq is of order 2 in this case, and G is either split or r r a quasisplit outer twist by the nontrivial element in OpGq. Recall that r p for any P pGq, mpq denotes the number of G-orbits in the associated r N pGq-orbit of L-homomorphisms attached to . Then pGq is the disr Aut joint union of the subset ( r r r r pGq P pGq : mpq 1 , G G , and its complement ( r r c pGq P pGq : mpq 2 , r r r r which we will usually denote by 1 pGq. We write 2 pGq, bdd pGq, 2 pGq r r r r r and 1 pGq, 1 pGq, 1 pGq, etc., for the obvious subsets pGq and 1 pGq. 2 2 bdd Consider, for example, a parameter 1 r , r r i P sim pGi q, Gi P Esim pNi q,

r r r r in the subset 2 pGq of pGq. Then lies in the subset 1 pGq of 2 pGq if 2 and only if each of the integers Ni is even. This is the most important case. r A general parameter P pGq is a self-dual sum (1.4.1) of simple parameters with higher multiplicities i . It is then easy to see that belongs to the r subset 1 pGq if and only if Ni is even for each i in the subset I of indices that correspond to self dual simple parameters. The condition can also be r expressed in terms of the general subparameter P sim pG q we have r attached to , in which the associated endoscopic datum G P Esim pN q is now an even orthogonal group. We write , r r P pG q, G P Esim pN q,

where G is also an even orthogonal group, and factors through a Levi subgroup of G that is a product of general linear groups. Then lies in r r 1 pGq if and only if and lie in the corresponding subsets 1 pG q and 1 pG q of pG q and pG q (with the convention that 1 pGq pGq in r r r2 r r 2 case N 0). The remarks here, with the accompanying notation, obviously r extend to the larger family of parameters ` pGq. r We shall limit our discussion to the subset bdd pGq of generic parameters r in pGq. (If F is global, the subscript bdd here is a little misleading. It anticipates our being able at some point to replace the complex algebraic group L by a locally compact group.) The rst part of the discussion

476

8. THE GLOBAL CLASSIFICATION

r r concerns parameters in the subset bdd pGq of bdd pGq. It is an extension of our remarks on characters and multiplicities to twisted endoscopy for G. r Suppose that F is local. For a local parameter P bdd pGq, Theorem r 2.2.4(b) tells us that the elements in are actually irreducible represenr tations, rather than nontrivial OpGq-orbits. In other words, the packet r is an actual L-packet, as we noted at the beginning of the section. The character formulas (8.3.1)(8.3.4) are of course still in force. However, there are supplementary character formulas that come from Theorem r 2.2.4(a). They apply to the G-twisted characters rr rr r r (8.3.6) fG prq IG pr, q fG prq dr, f P HpGq, r r r G-reg pGq r of extensions to GpF q of representations P . An analysis similar to r that of (8.3.2), with (2.2.17) and (8.3.5) in place of (2.2.6) and (8.3.1), yields the general formula r r r r r x r (8.3.7) IG pr, q xr, y1 S 1 p1 , 1 q p 1 , q, r r
r xPS 1 PG-reg pG1 q r r r

r r r r r r r for any P bdd pGq, P and P G-reg pGq and for , S and pG1 , 1 q as in Theorem 2.2.4(a). Given that the L-packet has already been dened by the original formula, one might ask whether (8.3.7) is superuous. It is not. Suppose for r r example that lies in the subset 2 pGq of bdd pGq. The endoscopic data r r G1 that index the terms in (8.3.7) then lie in the subset Eell pGq of EpGq, and are products r r r r r r G1 G1 G1 , G1 P Esim pNi1 q, 1 2 i r for a partition of N into two odd integers Ni1 . The stable characters in the corresponding summands decompose accordingly into products r r r r r r1 r r r1 S 1 p1 , 1 q S 1 p1 , 1 q S 1 p1 , 2 q 1 2 to which we can apply (8.3.4). The formula (8.3.6) thus expresses twisted r characters on GpF q in terms of twisted characters of general linear groups r 1 . The original formula (8.3.2) expresses ordinary characters of odd rank Ni on GpF q in terms of twisted characters of general linear groups of even rank Ni1 . Therefore, (8.3.7) represents a separate reciprocity law, which gives new local information. r Suppose now that F is global. A global parameter P 2 pGq has a global L-packet , whose contribution to the discrete spectrum is given explicitly by Theorem 1.5.2. Since its completions v lie in the corresponding r local subsets pGv q, the representations in the local packets v satisfy the supplementary twisted character formulas (8.3.7). What more is there to say? The extra piece of global information is given by Theorem 4.2.2(b). It r asserts that two canonical extensions to GpAq of any P in the discrete

8.3. REFLECTIONS ON THE RESULTS

477

spectrum, one obtained from (4.2.7), the other from Theorem 1.5.2 and r the extensions v in (8.3.7), are the same. Viewed as a signed multiplicity formula for the extension (4.2.7), it tells us that the sign in question equals 1. r Suppose again that each completion v of lies in pGv q, but that r r itself belongs to the complementary global set 1 pGq c pGq. This is the 2 2 r that contributes to the discrete case in which any representation P spectrum does so with multiplicity 2. Theorem 4.2.2(a) implies that for r G r r r any f P HpGq, the restriction of the operator Rdisc, pf q to the -isotypical 2 of the discrete spectrum has vanishing trace. The eigenspaces subspace L r G r of the operator Rdisc, pq then give a canonical decomposition (8.3.8) L2 L2 L2 , ,` , P ,

G of L2 into invariant subspaces on which the restriction Rdisc, is equivalent to . This phenomenon is related to a general philosophical question posed by A. Beilinson. Guided by his work on the geometric Langlands program, Beilinson has asked whether spaces of automorphic forms of higher multiplicity might have natural bases. Examples from [LL] lend support to his suggestion. The decomposition (8.3.8), which we have obtained from twisted endoscopy for G, can be regarded as further evidence. It would obviously be interesting to investigate this question heuristically in terms of parameters for general groups. The other half of our discussion of the even orthogonal group G is for the r r r complementary subset 1 pGq c pGq of bdd pGq. This is not implicit bdd bdd in the earlier theorems. It concerns the question of constructing genuine Lr packets from the coarser packets , and will be the setting for two new theorems in the next section. Our purpose here is to lay the groundwork for these theorems. r r Suppose that F is local. Regarding P 1 pGq as an AutN pGq-orbit bdd p of L-homomorphisms, we write pq for the corresponding pair of G-orbits r in pGq of L-homomorphisms . We also write for the preimage of r under the projection of pGq onto pGq. Then is a disjoint union of r r OpGq-torsors in pGq, which is to say, transitive OpGq-orbits (of order 2) of irreducible representations of GpF q. We would like to separate it into r two disjoint subsets , of representatives of the OpGq-orbits, which are compatible with endoscopic transfer, and are canonically parametrized by r the two elements in the OpGq-torsor pq. We will eventually realize this goal, apart from the canonical parametrization by pq. The idea is to r replace pq by a second OpGq-torsor T pq constructed directly in terms of representations. For the formal denition of T pq, consider rst a simple parameter r r P sim pGq. Notice that this automatically implies that lies in 1 pGq, bdd

478

8. THE GLOBAL CLASSIFICATION

since N is even. According to our weaker version of the local Langlands r r correspondence for G, corresponds to an OpGq-orbit P pGq of irreducible representations in unit pGq. By Corollary 6.7.3, the order mp q of r the orbit equals 2. We dene T pq in this case to be the OpGq-torsor . Consider next a general element 1 r , r i P sim pGi q, Gi P Esim pNi q, Ni even,

r in 1 pGq. In this case, we take 2 (8.3.9) ( T pq t t1 tr : ti P T pi q { to be a set of equivalence classes in the product over i of the sets T pi q. The equivalence relation is dened simply by writing t1 t if the set ( i : t1 ti i r r is even. Finally, suppose that lies in the complement of 1 pGq in 1 pGq. 2 bdd If , in the notation above, is the image of a product of local Langlands parameters for general linear groups, and there is no ambiguity in identifying the two Langlands parameters in pq with two L-packets. In this case, we simply set T pq pq. We can therefore assume that is a proper direct sum . Having already dened T p q for the parameter r P 2 pG q, we set ( (8.3.10) T pq t t t P T p q T p q { , where t1 t if the components of t1 and t are either the same or both distinct from each other. With our reference to Corollary 6.7.3 in the denition of T pq, we have implicitly taken F to be p-adic. The same denition of course holds if F is archimedean. In fact, by the general form of the Langlands correspondence r for real groups, there is a canonical OpGq-isomorphism between the torsors pq and T pq. For p-adic F , we do not have a canonical bijection. Our goal, which we will pursue in the next section, will be to construct the two L-packets for using T pq in place of pq. r Suppose that F is global, and that is a global parameter in 1 pGq. bdd r We can then dene an OpGq-torsor in a rather similar fashion. Suppose r r rst that lies in the subset sim pGq of 1 pGq. For any valuation v on F , 2 r we take v to be the representation in the packet v corresponding to the trivial character on Sv . In the case of archimedean v, where the mapping p r from v to Sv is not surjective, we appeal to Shelstads characterization r [S5, Theorem 11.5] of the elements in with Whittaker model for the v existence of v . The global product v , which is to be regarded as an orbit under the group ` r r OutN pGA q OutN pGv q ,
v

8.3. REFLECTIONS ON THE RESULTS

479

then occurs in the global discrete spectrum. More precisely, there are exactly two representations in the orbit of that occur in the discrete spectrum of G. They each occur with multiplicity 1, and together, form a transitive r r orbit under the diagonal subgroup OpGq of OutN pGA q. This follows directly from the proof of Proposition 6.7.2, or alternatively, the general assertion of r Theorem 1.5.2. We can therefore dene T pq to be the OpGq-torsor t u, in case is simple. Suppose next that is a general parameter in the subset r r 1 pGq, with simple components ti u. We then dene the global OpGq-torsor 2 T pq by the local construction (8.3.9). Finally, if lies in the complement r r of 1 pGq in 1 pGq, we follow the local prescription (8.3.10) to dene the 2 bdd global torsor T pq. r r We note that if P sim pGq is simple, and v lies in the subset pGv q of r pGv q for every v, the two representations above are actually equivalent. r But as we saw in the decomposition (8.3.8) obtained from the action of , we can indeed treat them as two separate representations in the denition. At any rate, our interest in global torsor T pq attached to P bdd pGq will be r conned to the case that some v lies in the corresponding local set 1 pGv q, since we have treated the other case above. In the next section, we will use the local torsor T pq to construct the r two L-packets attached to a local parameter P 1 pGq. We shall use the bdd global torsor T pq to establish corresponding multiplicity formulas for global r parameters P 1 pGq. This will give the rened endoscopic classication, 2 both local and global, for generic parameters for our even orthogonal group G. These results represent a dierent perspective, in at least one sense. The characters of individual representations in a local packet , cannot be dened in terms of twisted characters for general linear groups. They do satisfy internal endoscopic relations analogous to (8.3.2), but nothing akin to (8.3.4). From this point of view, the representations are less accessible r than their OpGq-orbits . Of interest also is what we do not obtain. We will not be able to dene a canonical parametrization of the two L-packets , by the two Langlands parameters P pq. The diculty seems to be deeply entrenched, at least insofar as it pertains to the methods of this volume. The obstruction can be likened to what in physics might be called a pZ, 2Zq-symmetry. The philosophical question is whether there might be any endoscopic way to resolve it. Is there any experiment within the conjectural theory of endoscopy that r would determine the predicted bijection between the two OpGq-torsors? The third topic for this section is that of Whittaker models. We shall prove a strong form of the generic packet conjecture, for any group r G P Esim pN q. To do so, we have only to combine our classication of local and global L-packets for G with the results of Cogdell, Kim, PiatetskiiShapiro and Shahidi, and Ginzburg, Rallis and Soudry on the transfer of automorphic representations with global Whittaker models. The basic idea

480

8. THE GLOBAL CLASSIFICATION

of applying the work that culminated in [CKPS2] and [GRS] to the general existence of Whittaker models is not new. I learned of it from Rallis, but the idea was also known to Shahidi, and no doubt others as well. We should recall the underlying global denition. Assume for a moment that F is global, and that G is a general quasisplit, connected group over F . Suppose also that pB, T, tXk uq is a F -stable splitting of G over F , and that F is a nontrivial additive character on A{F . The function puq F pu1 ` ` un q, u P NB pAq, where tuk u are the coordinates of u with respect to the simple root vectors tXk u, is a nondegenerate, left NB pF q-invariant character on NB pAq. It represents a global Whittaker datum pB, q for G. Given , one denes the Whittaker functional phq hpnq pnq1 dn
NB pF qzNB pAq

on the space of smooth functions h on GpF qzGpAq. Suppose that is an irreducible cuspidal automorphic representation of GpAq, acting on a closed ` invariant subspace V of L2 cusp GpF qzGpAq, G (where G denotes a central character datum). We say that p, V q is (globally) pB, q-generic if the restriction of to the space of smooth functions in V is nonzero. If this is so, the restriction v of to any local constituent v of is a local pBv , v q-Whittaker functional. We assume that F is either local or global, and that G is a xed group in r Esim pN q over F . We can then x pB, q by xing F , and taking pB, T, tXk uq to be the standard splitting. If F is global, and pGv , Fv q is a localization of pG, F q, we can take pBv , v q to be the corresponding localization of pB, q. r Proposition 8.3.2. (a) Suppose that F is local, that P bdd pGq is a r bounded generic parmeter, and that represents the OutN pGq-orbit in the r such that the linear character x, y on S is trivial. Then local packet is locally pB, q-generic. (b) Suppose that F is global, and that is a generic parameter in the r larger global set 2 pGq. Then there is a globally pB, q-generic element in r the global packet such that the linear character x, y on S is trivial. Proof. (a) If F is archimedean, the result is known for any quasisplit G, and for the genuine L-packet attached to a Langlands parameter P bdd pGq. For it follows from [V2, Theorem 6.2] and [Kos, p. 105] that there is exactly one pB, q-generic representation in the packet . It then follows from [S5, Theorem 11.5] that with the pB, q-Whittaker normalization of transfer factors we are using, is indeed the representation in r such that x , y equals 1. We can therefore assume that F is p-adic. Suppose rst that belongs r to the complement of 2 pGq. The argument in this case is purely local. We choose a local parameter M P 2 pM, q, for a proper Levi subgroup M of

8.3. REFLECTIONS ON THE RESULTS

481

r G. Let M be the representation in M such that the linear form x , M y on SM is trivial. We can assume inductively that M is pM, M q-generic, since M is proper in G. Let be the unique pB, q-generic component of IP pM q. It then follows from (2.5.3), together with the denition (2.4.17) of r the character x , y in terms of its twisted analogue x , M y for M and the self-intertwining operators RP pw, M q, that the character is trivial. This r gives the proposition for . r Suppose next that belongs to 2 pGq. In this case we use a global argument, specically the global construction of 6.3. From Proposition 6.3.1, together with the discussion at the beginning of this section in case is sim9 9 9 ple, we obtain global objects pF , G, q from the given local objects pF, G, q. 9 9 r 9 r Then is a generic parameter in the subset 2 pGq of ell pN q. It corresponds to an automorphic representation of GLpN q, which is induced from a self9 dual, cuspidal automorphic representation of a Levi subgroup. It follows by 9 induction, or by Lemma 5.4.6 in case is simple, that Theorem 1.5.3 is 9 i P sim pGi q of . This is the required 9 r 9 valid for the simple components condition for the global descent theorem of [GRS] and [So] (which depends for its proof on [CKPS2]). The theorem states that there is a globally 9 9 9 9 9 pB, q-generic cuspidal automorphic representation of GpAq whose near equivalence class (in the language of [So]) maps to that of . In other 9 9 words, satises the condition ` 9 9 cpq cp q cpq, 9 9 and thus contributes to the subspace of the automorphic discrete spectrum 9 9 of G determined by . As we noted following the proof of Corollary 6.7.4, r 9 therefore represents an element in the global packet 9 we now have at

our disposal, for which associated linear character 9 9 9 9 x P S , 9 (8.3.11) xx, y xxv , v y, 9
v

equals 1. 9 9 Since is globally pB, q-generic, each of its components v is locally 9 9 9 9 9 pBv , v q-generic. If v R S8 puq, we know that v is characterized uniquely as r9 9 the element in the local packet v such that the local linear character x , v y on Sv equals 1. This follows from the discussion based on the group Spp2q 9 at the end of 2.5, which led to the proof of Lemmas 6.4.1 and 6.6.2. We have just seen that the same condition holds for the archimedean places v P S8 . 9 This leaves only the p-adic place u of F , which we recall has the property 9u , Gu , u q equals pF, G, q. The component u of at u lies in the 9 9 that pF 9 9 r 9 . It follows from the product formula (8.3.11) that r local packet u the corresponding linear character x , u y on the group Su S equals 1, 9 9 and consequently that u equals the given representation of GpF q. But we 9

482

8. THE GLOBAL CLASSIFICATION

9 9 also know that u is pBu , u q-generic. The given representation is therefore 9 pB, q-generic, as required. (b) The existence of is proved in the same way as that of above. 9 That is to say, it follows directly from the global descent theorem of [GRS] and [So]. The assertion on the character x, y is just the condition that occurs in the discrete spectrum. Remarks. 1. We have identied the irreducible representation with r r the corresponding OutN pGq-orbit of representations in the packet . This is just our usual convention in the local case (a), and applies also to the r global case (b) if is taken to be the OutN pGq-orbit of representations in r N pGA q-orbit that actually occur in the discrete spectrum. the associated Out r The group of F -automorphisms of G with which we identied OutN pGq in Chapter 1 stabilizes pB, q. We can therefore regard a pB, q-Whittaker r functional for the representation as an OutN pGq-orbit of functionals for r N pGq-orbit of representations. the corresponding Out 2. In the local case (a), it is expected that is the unique pB, q-generic r element in the packet . In fact, we had to use the known analogue of this property for valuations v u in the proof. The property for G follows from r [JiS1] (and the existence of the packet ), in the case G SOp2n ` 1q. (See also [JiS2], for a proof of the local Langlands correspondence for generic representations of the p-adic group SOp2n ` 1q.) Extensions of [JiS1] to the other groups G were sketched in [JiS3], while the group G Spp2nq was recently treated in detail in [Liu]. Taking these for granted, we see that the automorphic representation in (b) is the (unique) element in the global r r packet whose local components are the elements v P v whose linear characters x, v y on the corresponding groups Sv are all trivial. 3. Consider again the local case (a). As we noted in 2.5, Konno [Kon] r proved that the packet , once constructed, would contain a pB, q-generic element. His methods are purely local. Their renement, based on a study of the Whittaker normalization of transfer factors in [KS, (5.3)], would presumably lead to a local proof of the proposition in case charpF q 2, as well as a proof of the uniqueness of . 8.4. Renements for even orthogonal groups We shall now complete the discussion of even orthogonal groups begun in the last section. Our goal is to construct genuine L-packets from the r cruder packets obtained in the main theorems. To see that this might be possible, we need only observe that the full force of the stabilized trace formula has yet to be exploited. We have used it so far just for test functions in the symmetric global Hecke algebra. We will now apply it as it was stated in 4.1, for arbitrary functions in the Hecke algebra. As in the last section, we x an even orthogonal group r p G P Esim pN q, G SOpN, Cq, N 2n,

8.4. REFINEMENTS FOR EVEN ORTHOGONAL GROUPS

483

over the eld F . We are interested in parameters that lie in the subset ( r r 1 pGq P pGq : mpq 2 r of pGq. For any such , the associated pair pq pG, q of Langlands parameters is a torsor under the group r r r OpGq OutN pGq AutN pGq{r N pGq Int r of order 2. In the last section, we dened another OpGq-torsor T pq in terms r of irreducible representations. The L-packets we construct from will be parametrized by T pq instead of pq. r r r If lies in the complement pGq of 1 pGq in pGq, the set pq consists of alone. A formal application of the earlier denition of T pq to this case yields a set that also contains only one element. With this understanding, r we note that for any P pGq, the set T pq behaves well under endoscopic transfer. Suppose for example that s belongs to the subset S ,ell of elliptic elements in S . The endoscopic preimage of the pair p, sq is then a pair pG1 , 1 q, where r G1 G1 G 1 , G1 P Esim pNi1 q, 1 2 i belongs to Eell pGq, and r 1 1 1 , 1 P pG1 q,
1 2 i i

belongs to

r pG1 q.

The nite group r r r OpG1 q OpG1 q OpG1 q

1 2

is an extension r r 1 OutG pG1 q OpG1 q OpGq 1, which acts transitively on the product T p1 q T p1 q T p1 q. 1 2 It is clear that there is a surjective mapping T p1 q T pq, which is comr r patible with the actions of OpG1 q and OpGq. In particular, the mapping 1 , 1 q p, sq extends to a surjective mapping pG t P T pq, t1 P T p1 q. r We are really only interested in the case that lies in 1 pGq (or ever the r r smaller set 1 pGq), so that T pq has order 2 and is indeed an OpGq-torsor. bdd 1 q then typically has order 4, and is an OpG1 q-torsor. In this r The set T p 1 q T pq are Out pG1 q-torsors. case, the bres of the mapping T p G r r Suppose that F is local. If represents a subset of unit pGq, we shall write for its preimage in unit pGq. In particular, is the preimage in r r r unit pGq of the packet of OpGq-orbits in unit pGq we have attached to r 1 pGq of pGq. It follows from r any . Assume that lies in the subset bdd pG1 , 1 , t1 q p, s, tq,

484

8. THE GLOBAL CLASSIFICATION

r r the results of Chapter 6 that the OpGq-orbit of any P has order 2. The bres of the mapping r r are thus OpGq-torsors. The following theorem posits among other things a pair of canonical sections r parametrized by T pq. Their images in unit pGq represent the true L-packets attached to the two Langlands parameters in the set pq. r Theorem 8.4.1. Suppose that F is local, and that lies in 1 pGq. bdd r Then there is an OpGq-equivariant bijection t t , f f G pt q, t P T pq, f P HpGq, t P T pq, from T pq onto a pair of stable linear forms r on HpGq, and an OpGq-equivariant bijection p, tq t , r from T pq onto , such that (8.4.1) and (8.4.2) fG pq fG pt q, r f P HpGq, t P T pq, f P HpGq, t P T pq, and such that for any x P S , the identity (8.4.3) f 1 p1 1 q xx, y fG pt q, t
r P

r P , t P T pq,

f G pq f G pt q,

r f P HpGq, t P T pq,

is valid for any endoscopic preimage pG1 , 1 , t1 q of p, s, tq. Remarks. 1. The formula (8.4.1) tells us that the action t of T pq on r the stable distributions in question is compatible with the action of OpGq on corresponding parameters. The formula (8.4.2) asserts that the action t of r T pq on the relevant characters is also compatible with that of OpGq. The formula (8.4.3) is a renement (for generic parameters) of the main assertion (2.2.6) of the local Theorem 2.2.1. 2. Given (8.4.1) and (8.4.2), one sees without diculty that the proposed mappings t t and t t are uniquely determined by (8.4.3). 3. Suppose that the theorem holds for F . It then extends formally to r standard representations attached to parameters in the larger set 1 pGq. In r attached to any parameter particular, the theorem is valid for the packet in the subset r r r 1 pGq 1 pGq X ` pGq unit unit r 1 pGq. of

8.4. REFINEMENTS FOR EVEN ORTHOGONAL GROUPS

485

We shall use global methods to establish this local theorem. Before doing so, we may as well rst state the associated global theorem. The burden of proof for both theorems will then fall on the stabilized trace formula for G. r Suppose that F is global, and that P bdd pGq is a generic global parameter. We can then write p1 (8.4.4) Idisc, pf q pG, G1 q Sdisc, pf 1 q, f P HpGq,
G1 PEell pGq

in the notation (3.3.15). The essential point here is that the formula holds for general functions in HpGq, rather than just the functions in the symmetric r subalgebra HpGq. To make the argument work, we will need a nonsymmetric version of the stable multiplicity formula, at least for some , to express the terms on the right. r r Any localization v of lies in the generic subset unit pGq of ` pGq. unit Assume that Theorem 8.4.1 is valid for Fv , and that v lies in the subset r r 1 pGq of unit pGq. The assertions of Theorem 8.4.1 are then valid for v , unit as we noted above in Remark 3. This allows us to dene an isomorphism t tv between the torsors T pq and T pv q. To describe it, suppose rst that r P sim pGq is simple. According to the denition in 8.3, an element t P T pq is represented by an automorphic representation t attached r r to a certain OpGq-orbit in . We use the second assertion of the theorem to dene tv as the element in T pv q such that v,tv t,v . We extend the construction from this basic case to more general , rst in r r r 1 pGq and then in the complement of 1 pGq in 1 pGq, directly from the 2 2 bdd denitions. Suppose now that Theorem 8.4.1 holds for every completion Fv of F . The mappings t tv then allow us to globalize the two constructions of the local theorem. We dene a global OpGq-equivariant mapping t t , t P T pq,

from T pq to the space of linear forms on the global space SpGq by setting t pv,tv q.
v

r If v lies in the complement of 1 pGv q, tv of course represents the lone unit r element in the set T pv q. In this case, v,tv is understood to be the OpGv qinvariant linear form v we identied in (2.2.2) with the corresponding parameter. To describe the second construction, we extend its accompanying local notation in the natural way. Namely, we write for the preimage in

486

8. THE GLOBAL CLASSIFICATION

r unit pGq of any subset of the restricted tensor product ` r r unit pGq unit pGv q .
v

r We then dene an OpGq-equivariant mapping p, tq t , r from T pq to by setting t v,tv ,

v

r P , t P T pq,

r P , t P T pq.

This is compatible with our denition of tv in the special case that lies in r sim pGq. r Theorem 8.4.2. Suppose that F is global, and that lies in 1 pGq. 2 Then (8.4.5) Sdisc, pf q |S |1 f G pt q, f P HpGq,
tPT pq

and (8.4.6) ` tr Rdisc, pf q fG pt q, f P HpGq.

r P p1q tPT pq

Remarks. 1. The assertions are renements (for generic, square-integrable parameters) of the global Theorems 4.1.2 and 1.5.2. By not requiring f to r lie in the symmetric Hecke algebra HpGq, they place the group G of type Dn on an equal footing with our groups of type Bn and Cn . 2. We recall that the left hand side of (8.4.6) was dened in (3.4.5). The r indexing set p1q on the right hand side is the set of representations in r the global packet such that the character x, y on S is trivial. The proofs of Theorems 8.4.1 and 8.4.2 have a common core. It begins with the preliminary interpretation of the stabilized trace formula (8.4.4). To describe this, we assume inductively that both theorems hold if N is replaced by any positive integer N 1 N . Suppose that F and are as in the global Theorem 8.4.2. The analogue of (8.4.5) is then valid for the summand of any G1 G on the right hand side of (8.4.4). Making the appropriate substitution, and recalling how we treated these terms in (6.6.4) for example, we nd that the right hand side of (8.4.4) equals G (8.4.7) Sdisc, pf q ` |S |1 f 1 p1 1 q, t
xPS x1

tPT pq

where pG1 , 1 , t1 q is a preimage of p, x, tq. The left hand side of (8.4.4) reduces to the left hand side of (8.4.6). For the present, we write it simply

8.4. REFINEMENTS FOR EVEN ORTHOGONAL GROUPS

487

as (8.4.8) where pq pG, q is the preimage of in unit pGq, and n p q is the multiplicity of in the discrete spectrum. The identity of (8.4.7) and (8.4.8) will be a starting point for the proofs of both theorems. We shall establish the two theorems in succession. Proof of Theorem 8.4.1. We are assuming that F is local, and that r lies in 1 pGq. If F is archimedean, the theorem follows from the endobdd scopic Langlands classication of Shelstad. In fact, the Langlands classir cation in this case provides a canonical OpGq-bijection between the torsors pq and T pq, which is compatible with the assertions of the theorem. It is the lack of this bijection in the p-adic case that forces us to introduce T pq as a substitute for pq. We therefore assume that the local eld F is nonarchimedean. To deal with this case, we shall use the global methods of Chapter 6. In particular, we shall apply Proposition 6.3.1 to the local parameter . Having established the local theorems of Chapter 6, we do not need the induction hypothesis Ni N from Proposition 6.3.1 that allowed us to apply Corollary 6.2.4. In other words, we do not have to rule out the case that is simple. r The essential case is that of a parameter (6.6.1) in the subset 1 pGq 2 1 9 r of bdd pGq. With this assumption on , we obtain global objects F , 9 r 9 9 P Esim pN q and P 1 pGq from the local objects F , G and , which satisfy r G 2 the conditions of Proposition 6.3.1. In particular, we have 9 9 9 pFu , Gu , u q pF, G, q, 9 for a p-adic valuation u on the global eld F . We shall identify the local r r 9 r 9 group OpGq OpGu q (of order 2) with its global counterpart OpGq. From the construction of Proposition 6.3.1, together with the fact that the simple 9 9 r 9 degrees Ni degpi q degpi q are even, we see that v belongs to 1 pGv q 2 9 for each archimedean valuation v P S8 . In particular, the set pv q is r 9 r r 9 an OpGv q-torsor. This allows us to identify the group OpGq OpGq also 9 r Gv q for any v P S8 , and thereby treat pv q as an 9 with its analogue Op r OpGq-torsor. 9 9 There is a canonical OpGq-bijection t t from T pq to T pq. In the case that is simple, this follows from the fact that a global representation 9 v in T pq is uniquely determined by its component ,u at u. For 9 9 arbitrary , the bijection is then a consequence of the denition of the 9 two parallel equivalence relations for T pq and T pq. Its construction thus 9 does not require the validity of Theorem 8.4.1 for the completion u . n p q fG p q,
r PpGq Ppq

488

8. THE GLOBAL CLASSIFICATION

9 9 We do know that the theorem holds for the completion v of at any archimedean valuation v P S8 . The earlier construction then gives us an 9 9 OpGq-bijection t tv from T pq to T pv q. This in turn can be composed 9 r 9 with the isomorphism between T pv q and any of the other OpGv q-torsors 9 v . We thus obtain an OpGq-isomorphism t v,t from T pq 9 r attached to 9 r 9 9 to pv q, and an OpGq-isomorphism t v,t from T pq to T pv q, for any r9 v P v . 9 We are carrying the earlier induction hypothesis that both theorems hold if N is replaced by a positive integer N 1 N . The expressions (8.4.7) and 9 9 9 (8.4.8), obtained from the stabilized trace formula (8.4.4) but with pF , G, q in place of the earlier triplet pF, G, q, are therefore equal. To be consistent with the notation here, we need also to write f9 for the global test function 9 in HpGq that was earlier denoted by f . We x its component f98,u f9v
vRS8 puq

r 9 away from S8 puq by choosing each f9v P HpGv q to be a symmetric, spheri1 p1 q 1, where pG1 , 1 q is the v-component of the 9 9 cal function so that f9v 9 v v v 9 9 1 , 1 q of the pair attached to any x P S . We can then regard preimage pG the equality of (8.4.7) and (8.4.8) as an identity of linear forms in the complementary component 9 f98,u f98 f9u f98 f, f98 P HpG8 q, f P HpGq, of f9. The summand of x and t in (8.4.7) equals 9 9 f91 p1 1 q f91 p1 1 q f 1 p1 1 q,
t 8 8,t t

since 9 pf98,u q1 p8,u q

vRS8 puq

1 f9v p1 q 1. v

The expression (8.4.7) therefore equals 9 G 1 9 (8.4.9) Sdisc, pf9q ` |S |1 f98 p1 1 qf 1 p1 1 q. 9 8,t t
xPS x1

tPT pq

The description of (8.4.8) requires a little more discussion. p r For any character P S , we have the OpGq-orbit of representations r r9 P such that x, y . We also have the set 8 pq of representations r9 8 9 v , 9 v P v , 9
vPS8

r9 in the packet 8 such that the character 9 9 9 9 xx, 8 y xxv , v y,

vPS8

x P S ,

8.4. REFINEMENTS FOR EVEN ORTHOGONAL GROUPS

489

on S equals pxq1 . It follows from Proposition 6.3.1(iii)(a) that this set is r9 not empty. In general, the elements 8 in 8 are simply transitive orbits 9 9 in unit pG8 q under the product r 9 r 9 OpGv q. OpG8 q
vPS8

r9 9 For any 8 P 8 pq and any local representation , P p q, there is a unique global representation 8, b , b , 9 9 9 8,u where 8, belongs to the set 9 9 9 pG8 , 8 q `
vPS8

9 9 pGv , v q

9 9 8,u and P pG8,u q is spherical, such that the coecient n p q on (8.4.8) 9 9 is nonzero. This follows from Proposition 6.3.1(ii) and either the proof of Propositions 6.6.5 and 6.7.2 or simply the general assertion of Theorem 9 9 1.5.2. In particular, the representation 8, is uniquely determined by 8 and , . We can then write 9 9 9 8, 8,t8 pq v,t9v pq ,
vPS8

for a direct product 9 t8 pq t8 p8 , , q 9

vPS8

tv p8 , , q 9

9 of uniquely determined elements tv pq tv p8 , , q in T pq. Since the 9 9 nonzero coecient n p q actually equals 1, the expression (8.4.8) becomes 9 (8.4.10) f98,G p8, q fG p, q. 9 9
r9 p 9 PS 8 P8 pq , Pp q

r r Lemma 8.4.3. For any P , there is an OpGq-isomorphism t t p and 8 P 9 pq, we have r 9 from T pq onto pq such that for any P S 8 9 tv p8 , ,t q t, t P T pq, v P S8 .

Proof. We have to show that for any and , , the T pq-components 9 tv p8 , , q, v P S8 ,

of the representations 8, 8,t98 pq in (8.4.10) are independent of 8 and 9 9 9 v. The main point is to show that tv p8 , , q is independent of the rst 9 9 argument 8 . The proof of this will be based on the identity of (8.4.9) and (8.4.10) that we have obtained from the stabilized trace formula.

490

8. THE GLOBAL CLASSIFICATION

9 We rst impose a constraint on the function f98 P HpG8 q that forces the summands on the right hand side of (8.4.9) to vanish. We require that for any x 1 in S and any t P T pq, the linear form 1 9 9 9 xx, 8 y f98,G p8,t q f98 p1 1 q 9 9 8,t
r9 8 P8 9

vanishes. In other words, for any t P T pq, the sum f98,G p8,t q, f98,G p, tq 9 9 9
r9 9 8 P8 pq

p P S ,

is independent of . This forces the sum over x in (8.4.9) to vanish. We then x the indices and , in (8.4.10), and take f P HpGq to be a pseudocoecient f, of the representation , P 2 pGq. This simplies the triple r9 sum in (8.4.10) to a simple sum over 8 pq. The equality of (8.4.9) and (8.4.10) becomes 9 G (8.4.11) Sdisc, pf98 f, f98,u q f98,G p8, q. 9 9 9
r9 8 P8 pq 9

Consider the abstract spectral expansion of the left hand side of (8.4.11), as a stable linear form in f98 . Using the right hand to bound its spectral support, and keeping in mind the spectral criterion for stability in terms of archimedean Langlands parameters, we obtain a linear combination 9 9 G 9 G 9 cpt8 q f98 p8,t98 q (8.4.12) Sdisc, pf98 f, f98,u q 9
9 t8

over indices 9 t8
vPS8

9 tv ,

tv P T pq,

9 9 with complex coecients cpt8 q cpt8 , f, q. The formulas (8.4.11) and (8.4.12) were derived with the condition that f98 p, tq be independent of t. We can actually treat the two right hand sides r9 as linear forms on the space of functions in the preimage 8 pq of 8 pq 9 9 9 in pG8 q. To be precise, we claim that if the coecients cpt8 q are rescaled by the factor r r | 9 | | 9 pq|1 | 9 | | 9 pq|1 ,
8 8 8 8

the two formulas remain valid if the condition in f98 is replaced by the 9 requirement that the function f98,G on pG8 q be supported on 8 pq. 9 9 This is because the restriction mapping, from the subspace of functions 9 f98 P HpG8 q that satisfy the rst condition onto the space of functions f98,G 9 on 8 pq, is surjective. The claim follows easily from this fact, together 9

8.4. REFINEMENTS FOR EVEN ORTHOGONAL GROUPS

491

with the form of the terms on the two right hand sides. The two formulas can then be combined as an identity 9 G 9 9 cpt8 q f98 p8,t98 q, f98,G p8, q (8.4.13) 9 9
r9 9 8 P8 pq 9 t8

9 9 9 for complex coecients cpt8 q cpt8 , f, q, and any f98 P HpG8 q such that 9 the function f98,G on pG8 q is supported on 8 pq. The representation 9 9 r 8, on the left hand side is the image of 8 under the section from 9 pq 9 9
8

9 to 8 pq determined by , , while the index of summation t8 on the right 9 hand is over elements in the direct product of |S8 |-copies of T pq. The left hand side of (8.4.13) does not vanish, since we can always take f98 to be a pseudocoecient of some 8, . We can therefore choose an 9 98 so that the corresponding coecient cpt8 q on the right is nonzero. 9 index t r 9 pq is arbitrary, and that f98 is a pseudocoecient 9 Suppose that 8 P 8 9 9 of the representation 8,t98 . The left hand side then equals 1 if t8 p8 , , q 98 , and vanishes otherwise. But the right hand side reduces in this equals t 9 9 case to the nonzero coecient cpt8 q cpt8 , f, q. It follows that the product t8 p8 , , q 9 tv p8 , , q 9
vPS8

equals the index 9 t8

vPS8

9 tv .

r9 9 But 8 was taken to be any element in 8 pq. It follows that for any 9 9 9 v P S8 , tv p8 , , q equals tv , and is therefore independent of 8 . 9v tv p8 , , q is independent of v. If 9 We now have to check that t 9 consists of one element . In this case, the is simple, equals 1, and denitions give isomorphisms 9 9 9 9 t t t t,v v,t v,t r of OpGq-torsors. (We recall that for simple , T pq is just dened to be the r OpGq-torsor .) The global representation t is determined by t and a 9 r 9 9 p1q. It follows that r 9 particularly chosen factor 8 in the set
8 8

tv p8 , t q t, 9 9 by the denition of the left hand side. Having already shown that tv p8 , t q is independent of 8 , we conclude that it is also independent of v. For 9 9 9 r general , we have an isomorphism t v,t of OpGq-torsors T pq and pv q. 9 The independence of tv p8 , , q on v then follows from the denitions, and the case we have just established for the simple constituents of . r We can now construct the required isomorphism. Suppose that P . p . For any representation Then , for a unique character P S

492

8. THE GLOBAL CLASSIFICATION

, P p q, we now know that the point 9 t tv p8 , , q, r9 9 8 P 8 pq, v P S8 ,

9 in T pq is independent of 8 and v, and therefore depends only on , . We r thus obtain an isomorphism , t between the OpGq-torsors p q and T pq. Its inverse t t then satises the conditions of the lemma. We return to the proof of Theorem 8.4.1. There are two mappings to r consider. The OpGq-equivariant isomorphism p, tq t is provided by the last lemma. We dene the other mapping by setting fG pt q, t P T pq, f P HpGq. (8.4.14) f G pt q
r P

The required identities (8.4.1) and (8.4.2) then follow from the denitions. We must still show that the right hand side of (8.4.14) is stable in f in order to justify the notation on the left hand side, and of course also because it is one of the requirements of the theorem. We have also to establish (8.4.3). The mapping p, tq t is dened by the formula 9 t8 p8 , ,t q t, r9 9 8 P 8 pq, t P T pq,

according to the discussion at the end of the proof of Lemma 8.4.3. The expression (8.4.10) can then be written as 9 (8.4.15) f98,G p8,t q fG p,t q. 9
p PS tPT pq r9 8 P8 pq 9

The inner sum in brackets is what we denoted by f98,G p, tq near the begin9 ning of the proof of the lemma. As before, let us assume that f98 is such that this is independent of . The inner sum can then be written as the product of |S |1 with 9 G 9 f98,G p, tq f98,G p8,t q f98 p8,t q. 9 9 9
p PS r9 8 P8 9

Our expression (8.4.15) for (8.4.10) reduces to the simpler expression 9 G 9 f98 p8,t q fG p,t q , |S |1
tPT pq p PS

which in turn equals |S |1

tPT pq 9 G 9 f98 p8,t q f G pt q,

according to the denition (8.4.14). On the other hand, our condition on f98 implies that the summands in (8.4.9) with x 1 all vanish. The equality of

8.4. REFINEMENTS FOR EVEN ORTHOGONAL GROUPS

493

(8.4.9) and (8.4.10) becomes (8.4.16)

9 G Sdisc, pf9q |S |1 9 9

tPT pq

9 G 9 f98 p8,t q f G pt q.

The left hand side of this formula is stable in the variable function f P HpGq. 9 G 9 The coecients f98 p8,t q may be chosen arbitrarily by varying f98 . It follows that the linear form f G pt q dened by (8.4.14) is stable, as required. It remains to prove (8.4.3). The formula (8.4.16) is still valid without the condition on f98 under which it was derived. This follows from the stability 9 of each side and the fact that the subspace of functions f98 P HpG8 q that 9 satisfy the condition maps onto SpG8 q. We substitute it for the leading term on the right hand side (8.4.9) of the trace formula (8.4.4). The right hand side of (8.4.4) becomes 1 9 |S |1 f98 p1 1 q f 1 p1 1 q, 8,t t
xPS tPT pq

9 for any f98 P HpG8 q and f P HpGq, an expression we write as 9 9 9 xx, 8 y f98,G p8,t q f 1 p1 1 q. (8.4.17) |S |1 9 t
x t r9 8 P8 9

This consequently equals the expression (8.4.15) we have obtained for the left hand side (8.4.10) of (8.4.4). We choose the function f98 to be a pseudocoecient of a representation 8,t P 8 pq, for xed and t. The expres9 9 sion (8.4.15) reduces to fG p,t q. The other expression (8.4.17) becomes |S |1 pxq1 f 1 p1 1 q, t
xPS

9 9 since xx, 8 y Applying Fourier inversion on S to each side of the resulting identity, we obtain f 1 p1 1 q pxq fG p,t q xx, y fG pt q. t
p PS r P

pxq1 .

This is the required formula (8.4.3). We have completed the proof of Theorem 8.4.1 in case is a parameter r (6.6.1) in 1 pGq. It remains to consider a parameter in the complement 2 r r of 1 pGq in 1 pGq. Then is of the general form (6.3.1), and is the image 2 bdd r of a parameter M P 2 pM q, for a proper Levi subgroup M of G. We shall say only a few words about this case since it is much simpler. We dene the mappings t t and p, tq t directly from their analogues for M , guided by the requirement that they be compatible with induc1 tion. In this way, we also obtain general analogues fG pt , xq and fG pt , xq of the two sides of the local intertwining relation (2.4.7). The required identities (8.4.1) and (8.4.2) follow directly from the denitions. It remains to establish the transfer identity (8.4.3). As we saw in the proof of Proposition

494

8. THE GLOBAL CLASSIFICATION

2.4.3, this in turn will follow from the local intertwining relation. In other words, it suces to establish the following analogue of Theorem 2.4.1. Proposition 8.4.4. Suppose that F is local, and that lies in the comr r plement of 1 pGq in 1 pGq. Then 2 bdd (8.4.18)
1 fG pt , sq fG pt , uq,

f P HpGq, t P T pq,

for u and s as in Theorem 2.4.1. Proof. Recalling the discussion at the beginning of 6.4, and noting that there is nothing new to be established in the two exceptional cases (4.5.11) and (4.5.12), we see that it is enough to prove (8.4.18) for a parameter of the special form (6.4.1), with r q, and Ni even whenever q i r. We can then take u s to be an element x in the set we denoted by S,ell . We shall be brief. Suppose rst that F R. By the results of Shelstad and the multiplicativity of intertwining operators, the generalization fG pt , xq M,t pxq xx, y fG pt q, t P T pq, f P HpGq,
r P

r of (6.1.5) remains valid. If f P HpGq, this equals 1 fG p, xq fG p, xq xx, y fG pq,

r P

by the special case of (8.4.18) given by Theorem 2.4.1. We can choose r f P HpGq so that # 1, if , fG pt q fG p q 0, otherwise, r for any P . It then follows that M,t pxq 1, Therefore fG pt , xq
r P

x P S,ell , t P T pq.
1 xx, y fG pt q fG pt , xq,

for any t, x and f P HpGq, as required. Suppose now that F is p-adic. In this case, we use Proposition 6.3.1 to 9 9 9 9 9 attach global objects F , G, , M and M to F , G, , M and M , as at the beginning of 6.4. We then have the analogue ` 9 9 9 9 9 (8.4.19) f91 pt , xq f9 9 pt , xq 0, f9 P HpGq,
9 G G tPT pq xPS,ell

9 9 of (6.4.2). This follows from an application of the standard model to pG, q, as in the relevant part of the proof of Lemma 5.2.1, and the fact that 9 r 9 9G Rdisc, pf9q vanishes, as we have shown for symmetric functions f9 P HpGq 9

8.4. REFINEMENTS FOR EVEN ORTHOGONAL GROUPS

495

r 9 (and consequently any f9). We can x a symmetric function f98,u P HpG8,u q away from S8 puq so that if f9 f98 f f98,u , r 9 f98 P HpG8 q, f P HpGq, the left hand side of (8.4.19) decomposes ` 1 9 9 9 1 9 f98,G p8,t , xq fG pt , xq f98,G p8,t , xq fG pt , xq . 9 9
tPT pq xPS,ell

It then follows from the archimedean case we have proved that ` 1 9 f98,G p8,t , xq fG pt , xq fG pt , xq 0. 9 9
t x

Since the linear forms 9 f9 9 p8,t , xq, 9

8,G

9 f98 P HpG8 q, t P T pq, x P S,ell , f P HpGq, t P T pq, x P S,ell ,

are linearly independent, we conclude that

1 fG pt , xq fG pt , xq,

as required. This completes our sketch of the proof of Proposition 8.4.4, and hence also of the remaining part of the proof of Theorem 8.4.1.

We write (8.4.20) r ,t tt : P u, t P T pq, r for any P 1 pGq, as in Theorem 8.4.1. The sets ,t parametrized by bdd the two elements t P T pq are then the two L-packets attached to . If r belongs to the complementary set bdd pGq, T pq consists of one element, and the assertions of Theorem 8.4.1 hold trivially with t and t . If we extend the denition (8.4.20) to this case, ,t is simply the L-packet r we have already attached to . Corollary 8.4.5. (a) The set temp pGq of irreducible, tempered representations of GpF q is a disjoint union of the L-packets ,t , (b) The distributions t , r P bdd pGq, t P T pq, form a basis of the space of stable linear forms on HpGq. r r Proof. (a) For any P bdd pGq, the preimage of in temp pGq is the disjoint union over t P T pq of the packets ,t . The assertion follows r from its analogue for temp pGq, which was stated in Theorem 1.5.1(a) and proved in Corollary 6.7.5. (b) We can assume that F is nonarchimedean, as we have done implicitly in (a), since the assertions for archimedean F follow from the general r P bdd pGq, t P T pq.

496

8. THE GLOBAL CLASSIFICATION

endoscopic classication of Shelstad. In particular, Shelstads archimedean results imply that the spectral characterization of stability (in terms of Langlands parameters) matches the geometric characterization (in terms of stable conjugacy classes), a property we have already used in the derivation of (8.4.12). This is what must be veried for p-adic F . Theorem 8.4.1 tells us that the distributions t are stable. Part (a) above implies that they are linearly independent. What remains is to show that they span the space of all stable linear forms on HpGq. This is the r analogue for G of the assertion of Proposition 2.1.1 for GpN q, alluded to at the beginning of the proof of the proposition. It is implicit in the main results [A11] (namely, Theorems 6.1 and 6.2, together with Proposition 3.5), and will be made explicit in [A24]. Proof of Theorem 8.4.2. We assume now that F is global. We have established that Theorem 8.4.1 holds for the localizations v of the given r global parameter P 1 pGq. However, we must still retain what is left of 2 the original induction hypothesis, namely that Theorem 8.4.2 holds if N is replaced by any integer N 1 N . The expressions (8.4.7) and (8.4.8) are therefore equal. The main point is to prove the stable multiplicity formula (8.4.5). The argument will be similar to a part of the proof of Theorem 8.4.1, specically Lemma 8.4.3, so we can be brief. We shall rst take care of the other assertion, the actual multiplicity formula (8.4.6). One advantage we have here is that the right hand side of (8.4.5) is actually dened. It equals the contribution of x 1 that is missing from G (8.4.7). Following a convention from 4.4, we write 0 Sdisc, pf q for the difference between the two sides of (8.4.5). Since (8.4.7) equals the left hand G side of (8.4.6), namely the trace of Rdisc, pf q, we see that the dierence ` G G tr Rdisc, pf q 0 Sdisc, pf q, f P HpGq, equals |S |1 f 1 p1 1 q. t
xPS tPT pq

Another advantage is that we can now apply the formula (8.4.3) to the local components of the global summands f 1 p1 1 q. From a simplied variant of t the remarks at the end of 4.7, we deduce that the last expression equals fG pt q.
r P p1q tPT pq G This is just the right hand side of (8.4.6). Writing 0 rdisc, pf q for the dierence of the two sides of (8.4.6), we conclude that

(8.4.21)

0 G rdisc, pf q

G 0 Sdisc, pf q,

f P HpGq.

In particular, the stable multiplicity formula (8.4.5) implies the actual multiplicity formula (8.4.6).

8.4. REFINEMENTS FOR EVEN ORTHOGONAL GROUPS

497

It remains then to establish (8.4.5). If the localization v of lies in the r r subset unit pGv q of unit pGv q for each v, the formula (8.3.6) reduces to what r we have established for symmetric functions f P HpGq. The formula (8.4.5) is then also valid, and there is nothing further to do. We can therefore r assume that v lies in the complementary set 1 pGv q for some v. This unit property will allow us to put together a proof of (8.4.5) parallel to that of its analogue (8.4.16) from Theorem 8.4.1. Suppose that V is some nite set of valuations of F such that v lies in r r the subset 1 pGv q of unit pGv q for each v P V . We x a character on unit V in the packet S , and an element ! ) r r V v : v P v , x, v y 1 for almost all v
vRV

with the property that the character V xx, y xxv , v y,

vRV

x P S ,

V r on S equals . We also x a symmetric function f P HpGV q such that # V 1, if V , V f,G p V q 0, otherwise,

r for every V P V . We then set

V f fV f ,

fV P HpGV q,

and regard (8.4.21) as an identity in the complementary function fV . G The left hand side 0 rdisc, pf q of (8.4.21) is dened as the dierence of the two sides of the putative identity (8.4.6). The left hand side of (8.4.6) in turn equals the double sum (8.4.8). Arguing as in the derivation of (8.4.10) from (8.4.8) in the proof of Theorem 8.4.1, we see that this equals fV,G pV, q,
r V PV pq V,

r r where V pq is the subset of elements V P V such that the character 1 , and r x, V y on S equals V, ranges over some OpGq-orbit of order 2 in the set pV q. In particular, we have V, V,tV pq v,tV pq ,
vPV

for a direct product

V tV pq tV pV , q

vPV

V tv pV , q

498

8. THE GLOBAL CLASSIFICATION

r of points in T pq that is determined up to the diagonal action of OpGq. The right hand side of (8.4.6) is just equal to fV,G pV,t q,
r V PV pq tPT pq

where V,t
vPV

v,t ,

t P T pq.

The dierence of these last two expressions therefore equals the right hand G V side 0 Sdisc pfV f q of (8.4.21). Supplementing the argument used to derive (8.4.12) in the proof of Lemma 8.4.3 with the criterion for stability at places G V v P V in Corollary 8.4.5(b), we write the stable linear form 0 Sdisc pfV f q as a linear combination G cptV q fV pV,tV q
tV

over indices tV
vPSV V with complex coecients cptV q cptV , f q. The identity (8.4.21) becomes ` G f V,G pV,tV pq q f V,G pV,t q cptV q fV pV,tV q, (8.4.22) r V PV pq tV

tv ,

tv P T pq,

where f V pxq
r POpGq

` fV pxq .

r The problem is to show that for each V P V pq, the point V q is diagonal, in the sense that it is the image of a xed tV pq tV pV , point t P T pq. We shall assume that this is false, and then derive a contradiction to (8.4.22). Our assumption implies that the left hand side of (8.4.22) does not vanish. There is then a corresponding contribution to the right hand side. That is, there is a nondiagonal index tV such that the coecient cptV q is nonzero. r Suppose that V P V pq. We can then choose fV P HpGV q so that the function fV,G on V pq equals 1 at V,tV , and vanishes on the complement of V,tV . The left hand side of (8.4.22) then equals 1 if the index V tV pq tV pV , q equals tV , and vanishes otherwise. Since the right hand side of (8.4.22) equals the nonzero coecient cptV q, we deduce that tV pq equals tV . Since V is arbitrary, we conclude that tV tV pq is independent of V . Armed with this property, we obtain the contradiction from the r denition of the isomorphism of OpGq-torsors t v,t v,t , r t P T pq, v P V, v P v ,

8.4. REFINEMENTS FOR EVEN ORTHOGONAL GROUPS

499

as in the closing arguments from the proof of Lemma 8.4.3. That is, we show that tV is in fact diagonal, rst in case is simple, and then in general by an application of the denition of T pq to the simple components of . We have established the necessary contradiction. It follows that the r (OpGq-orbits of) indices tV pq and t in any summand on the left hand side of (8.4.22) are equal. The sum itself therefore vanishes. But the left hand G side of (8.4.22) equals the left hand side 0 rdisc, pf q of (8.4.21), if f V equals p the chosen function f V . Letting P S vary, we conclude that

(8.4.23)

0 G rdisc, pf q

G 0 Sdisc, pf q 0,

r for any nite set of valuations v P V with v P 1 pGv q, and any function unit (8.4.24) f fV f V , r fV P HpGV q, f V P HpGV q. It remains to establish (8.4.23) for any function f P HpGq. Assuming fv is decomposable, we x a nite without loss of generality that f
v

set S of valuations outside of which fv is unramied. We then take V to r be the subset of places v P S such that v lies in 1 pGv q. For any v in unit the complement of S, fv is the characteristic function of our xed maximal r compact subgroup Kv of GpFv q, which lies in the OpGv q-symmetric subspace r r HpGv q of HpGv q. If v lies in the complement of V in S, v lies in unit pGv q, r v q-symmetric. One sees readily from the discussion above and is itself OpG G that the left hand 0 rdisc, pf q of (8.4.22) remains unchanged if fv P HpGv q is replaced by the symmetric function 1 f v . It then follows from (8.4.22), and 2 the case of (8.4.23) we have established for functions f of the form (8.4.24), that (8.4.23) does indeed hold for any function f P HpGq. According to our denitions, the required global formulas (8.4.5) and (8.4.6) are just the two vanishing assertions of (8.4.23). Having established these assertions, we therefore completed the proof of Theorem 8.4.2. We have completed our study of local and global L-packets for the even orthogonal group G. In particular, we have obtained a rened local endoscopic classication of admissible representations, and a rened global endoscopic classication of the representations in the discrete spectrum of Ramanujan type. We have not tried to treat the packets attached to nongeneric parameters . The global methods of this section undoubtedly apply to such . However, they would have to be combined with a ner understanding of the constituents of their packets. It would be interesting to apply the p-adic results of Moeglin to this question. As we noted in 8.3, the rened endoscopic classication we have established for p-adic F does not quite imply a rened Langlands correspondence. The impediment is the internal pZ{2Zq-symmetry implicit in the lack of a canonical bijection between two sets of order 2. In principle, there will be r one copy of the group pZ{2Zq for each local parameter P 1 pGq. The bdd obstruction is actually a little more modest. Suppose that for each even N ,

500

8. THE GLOBAL CLASSIFICATION

r and each even orthogonal group G P Esim pN q over the p-adic eld F , the r pZ{2Zq-symmetry for any simple parameter P sim pGq has been broken. r In other words, the bijection between the OpGq-torsors pq and T pq predicted by the local Langlands conjecture has been dened. The same then r holds for any of the more general parameters P 1 pGq, which are atbdd tached to any N and G. This is a direct consequence of the denitions. Notice that the question here is purely local. Our makeshift global group L is too coarse at this point even to consider a global analogue of the problem.

8.5. An approximation of the Langlands group We would like to be able to streamline the denitions of 1.4. The goal would be to replace the complex groups L in terms of which we formulated the global theorems by a locally compact group that is independent of . We shall conclude Chapter 8 with a step in this direction. We shall introduce a group that characterizes most of the global parameters we have studied. The new group will be closer in spirit to the hypothetical Langlands group LF . As an approximation it is still pretty crude. It serves as a substitute for LF only so far as to describe the global endoscopic problems we have now solved. In fact, the group treats just a part of the problem, in that it pertains only to global parameters we shall call regular. Its missing complement depends on the solution of two supplementary endoscopic problems. There is no point in attempting to formulate it here, even though these problems are undoubtedly easier than the ones we solved. Our aim for now is merely to give a partial description, which illustrates what we can hope is a general construction in the simplest of situations. There are two steps. The rst will be to introduce a group that characterizes automorphic representations of general linear groups. We will then describe a subgroup that accounts also for automorphic representations of orthogonal and symplectic groups. For any N 1, we have the basic set Csim pN q, dened in 1.3 and discussed further in 8.3. We recall that it consists of simple families c tcv : v R Su of semisimple conjugacy classes in GLpN, Cq, taken up to the equivalence relation c1 c if c1 cv for almost all v. As usual, S ValpF q represents a v nite set of valuations of F that contains the set S8 of archimedean places. The special case F Cp1q Csim p1q is an abelian group under pointwise multiplication. It has an action : c c, P , c P Csim pN q,

8.5. AN APPROXIMATION OF THE LANGLANDS GROUP

501

on Csim pN q, given by pointwise scalar multiplication on families of pN N qmatrices. The complex determinant GLpN, Cq GLp1, Cq, which is of course dual to the central embedding GLp1q GLpN q, determines a second operation. Its pointwise application gives a mapping p : Csim pN q such that p p pcq N pcq. We shall use these two operations to dene an endoscopic Langlands group that treats most automorphic representations of general linear groups. We rst recall that the two operations have natural analogues for automorphic representations. We know that there is a canonical bijection cpq, P Acusp pN q, from the set of unitary cuspidal automorphic representations of GLpN q onto Csim pN q. This is the theorem of strong multiplicity one, which can be regarded as the specialization of Theorem 1.3.2 to Acusp pN q, and is the basis for our denition of the set Csim pN q. The abelian group X XF Acusp p1q of id`le class characters acts in the usual way e : b p detq, P X, P Acusp pN q, on Acusp pN q. The central character then gives a second operation, a mapping : Acusp pN q X such that pq N pq. These are the automorphic analogues of the two operations for Csim pN q. We shall write c c , c P Csim pN q, for the inverse of the bijection cpq, and also , c c , and pc q c , c P Csim pN q, for the id`le class character e c pcq . p P , for its specialization to the case N 1. We then have P , c P Csim pN q,

502

8. THE GLOBAL CLASSIFICATION

For any c P Csim pN q, we have isomorphic subgroups c t P : c cu and Xc t P X : c c u t : P c u of and X. They consist of elements of order dividing N , and are expected to be cyclic. We shall say that c and c are regular if these two subgroups are trivial, which is to say that the associated representations c b p detq, P X, are mutually inequivalent. We then write Csim,reg pN q for the set of regular elements in Csim pN q. The complement of Csim,reg pN q in Csim pN q should then be sparse in Csim pN q. We shall introduce a locally compact variant Lc of the L-group of GLpN q for every c P Csim,reg pN q. The unitary group U pN q is a compact real form of the complex dual group GLpN, Cq. For any c P Csim,reg pN q, we dene an extension (8.5.1) 1 Kc Lc WF 1 of WF by the special unitary group Kc SU pN q by setting ( Lc g w P U pN q WF : detpgq c pwq . We have identied the central character c of c here with a character on the Weil group. The group Lc is a locally compact extension of WF by the compact, simply connected group SU pN q, which need not be split. By itself, Lc depends only on c . It has the property that for any unitary, N -dimensional representation of WF with determinant equal to c , the corresponding global Langlands parameter factors through the image of the L-embedding ` Lc L GLpN q GLpN, Cq WF . We shall make Lc more rigid by attaching the local Langlands parameters associated to c. The group SU pN q is of course a compact real form of the complex special linear group SLpN, Cq. To account for the possible failure of the generalized Ramanujan conjecture for GLpN q, we will need to work with the complexication Lc,C of Lc . This is the extension (8.5.2) 1 Kc,C Lc,C WF 1 ( Lc,C g w P GLpN, Cq WF : detpgq c pwq . It has the property that any L-homomorphism : Lc
L

of WF by Kc,C SLpN, Cq dened by

G,

for a connected reductive group G over F , extends analytically to Lc,C . Our interest in Lc,C is in the local Langlands parameters ` LFv L GLpN qv GLpN, Cq WFv

8.5. AN APPROXIMATION OF THE LANGLANDS GROUP

503

attached to the local components c,v of the cuspidal automorphic representation c of c. They provide the left hand vertical arrows in the diagrams LFv WFv , Lc,C WF and are dened up to conjugacy by the subgroup Kc,C of Lc,C . We would like to treat Lc as an object that depends only on the -orbit of c. Since acts freely on Csim,reg pN q, any point c1 in the orbit of c equals c, for a unique element P . The mapping g w g pwq w, g w P Lc ,

(8.5.3)

v P ValpF q,

is then a canonical L-isomorphism from Lc to Lc1 , which is compatible with the associated diagrams (8.5.3). Before we formalize the last property in our denitions, we should build r in the group OpN q of outer automorphisms of GLpN q. This is the group of r order 2 generated by our standard automorphism pN q. Its canonical actions r r on and X give isomorphic semidirect products OpN q and X OpN q. 1 is any r The rst group OpN q acts in the obvious way on Csim pN q. If c point in the corresponding orbit of c, we can still choose a canonical element r in OpN q such that c1 c. We do need to be a little careful if the r orbit of c is self-dual, in the sense that the family c_ pN qpcq equals for r some P , since the stability group of c in OpN q then has order 2. In this case, we simply take to be the unique element in the normal subgroup r r of OpN q. The second group X OpN q acts on the L-group of GLpN q according to the automorphisms
L

1 ,

r P X, P OpN q,

dened at the beginning of 3.2. In other words, p q : g w pgq pwq w, g P GLpN, Cq, w P WF .

In particular, if c1 equals pqc, the restriction of the automorphism dened ` r by p q maps Lc to Lc1 . It follows that for any point c1 in the OpN q orbit of c, we have a canonical L-isomorphism from Lc to Lc1 , which is compatible with the associated diagrams (8.5.3). We can therefore treat c as an element in the set Csim,reg pN q of ` r OpN q -orbits in Csim,reg pN q. With this interpretation of c, we can still form a canonical extension (8.5.1), with complexication (8.5.2) and local diagrams (8.5.3). It comes with a canonical isomorphism onto the associated concrete object we have attached to any c1 in the class c, and is dened as a formal inverse limit over the points c1 , as in Kottwitzs denitions at the beginning of [K1].

504

8. THE GLOBAL CLASSIFICATION

We are now working with the family Csim,reg pN q Csim,reg

N 1

of orbits of the group `

N 1

r OpN q

in the subset Csim,reg of Csim ` Csim pN q .

N 1

`
N 1

Csim,reg pN q

We amalgamate the groups we have attached to elements of the family into a bre product (8.5.4) L pLc WF q F,reg
cPCsim,reg

over WF . We thus obtain a locally compact extension (8.5.5)

1 KF,reg L F,reg WF 1 of WF by a product KF,reg of compact, simply connected groups. The extension comes with a complexication 1 KF,reg,C L F,reg,C WF 1, where KF,reg,C is an (innite) topological product of complex special linear groups, and an associated family of conjugacy classes of local embeddings

LFv (8.5.6)

WFv , v P ValpF q.

L F,reg,C WF Consider a (continuous, semisimple) representation : L F,reg GLpN, Cq of L . In case is irreducible, we shall call it standard if it factors to an F,reg irreducible, unitary representation of a quotient Lc of L , whose restricF,reg tion to the subgroup Kc SU pN q is standard, in the extended sense that it is equivalent to either the standard embedding of SU pN q into GLpN, Cq or its dual. An irreducible standard representation thus amounts to a tensor product of a standard representation of SU pN q with a unitary character of c WF . In particular, it is trivial on the commutator subgroup WF of WF . In the general case, we shall say that is standard if it is a (nite) direct sum of irreducible standard representations. We then write reg,bdd pN q for the

8.5. AN APPROXIMATION OF THE LANGLANDS GROUP

505

set of standard representations of LF , and pN q for the set of equivalence reg classes of unitary representations : L F,reg SU p2q GLpN, Cq
whose restriction to L F,reg belongs to reg,bdd pN q. We have constructed a formal analogue L F,reg of the global Langlands group LF . It governs the basic representation theory of GLpN q, which is to say, the theory encompassed by the three theorems stated in 1.3. Of course, we have really to preface this statement with the adjective regular. Let Areg pN q be the subset of representations in ApN q, the set of automorphic representations of GLpN q introduced in 1.3 that occur in the spectral decomposition, whose cuspidal components are regular. It then follows from the denitions, together with the three theorems, that there is a canonical bijection , P pN q, reg

from pN q onto Areg pN q such that reg ,v v , v P ValpF q. Here, v P v pN q is the local parameter attached to the complexication of by (8.5.6), while ,v is obviously the local component of the automorphic representation . There is no need to be concerned that the complexication L F,reg,C of L F,reg is not locally compact. Its only role is to extend the domain of parameters P reg,bdd pN q. Any such factors to a representation of an extension of WF by a nite dimensional quotient of KF,reg , which in turn extends analytically to the extension of WF by the corresponding quotient of KF,reg,C . We recall again that the complexication L F,reg,C is needed only because we do not know the generalized Ramanujan conjecture for cuspidal automorphic representations of GLpN q. The group L F,reg is adapted to the theory of (standard) endoscopy for general linear groups, which amounts to the three theorems stated in 1.3, together with the theory of Eisenstein series for GLpN q. We shall now introduce a subgroup of L F,reg that is adapted to the more sophisticated theory of endoscopy for orthogonal and symplectic groups. Consider the subset ( r r Csim Csim pN q c P Csim pN q : c_ c
N 1 N 1

r of self-dual families in Csim . For any c P Csim , we write r r c t P : c P Csim u. Since pcq_ _ c_ 1 c, we see that r c t P : 2 P c u.

506

8. THE GLOBAL CLASSIFICATION

We shall again restrict our attention to the corresponding subset ` r r r Csim,reg Csim,reg pN q Csim pN q X Csim,reg pN q
N 1 N 1

of regular elements. For any c in this subset, the associated subset r r c t P : 2 1u of is the subgroup of elements of order dividing 2, which is of course independent of c. We form the corresponding family ` r r Csim,reg Csim,reg pN q
N 1

r r of -orbits in Csim,reg . There is then an injection

r Csim,reg Csim,reg , ` r whose image is the set of OpN q -orbits in Csim,reg that have represenr tatives in Csim,reg . It follows from Theorem 1.4.1 that ` r r Csim,reg pN q Csim,reg pGq , r GPEsim pN q

r r where Csim,reg pGq is the subset of families c P Csim,reg pN q such that c cpq, r r for some in the set A2 pGq. If N is even, the free action of on Csim,reg pN q r stabilizes the associated subsets Csim,reg pGq, as one sees from the criterion r of Theorem 1.5.3. If N is odd, it is the opposite that holds. In this case r acts simply transitively on the data G P Esim pN q. Recall however that these dierent endoscopic data all have the same endoscopic group G SppN 1q, p with dual group G SOpN, Cq. r r Suppose that c belongs to the subset Csim,reg pGq of Csim,reg pN q attached r r to G P Esim pN q, for some N . We write Kc for the standard maximal compact p r subgroup of G. Thus, Kc equals the special unitary quaternion subgroup p Spp2n, Cq if G SOp2n ` 1q, the compact orthogonal SU pn, Hq of G p subgroup SOp2n ` 1q of G SOp2n ` 1, Cq if G Spp2nq, and the compact p orthogonal subgroup SOp2nq of G SOp2n, Cq if G SOp2nq. In the rst r two cases, WF acts trivially on Kc , while in the third case, it acts according to the mapping of WF to the group of outer automorphisms dened by c . We dene an extension (8.5.7) r r 1 Kc Lc WF 1

r of WF by Kc simply by setting r r Lc Kc WF .

8.5. AN APPROXIMATION OF THE LANGLANDS GROUP

507

r We can think of Lc as a compact real form of the group words, its complexication is the extension (8.5.8) r r 1 Kc,C Lc,C WF 1

L G.

In other

p p r of WF by the complex group Kc,C G dened by the L-group Lc,C L G. Our notation here is obviously designed to be compatible with that of the groups Lc and Lc,C above. In particular, we have local diagrams r LFv WFv , r Lc,C WF in which the left hand vertical arrows are dened up to conjugacy by the r r subgroup Kc,C of Lc,C . This follows from Theorem 1.4.2, but with one r p caveat. In the case that Kc,C G SOp2n ` 1, Cq, we must take c to r be the element in its -orbit that c 1 in order that the images of the r mappings of the groups LFv actually be contained in Lc,C . r Following our convention for the group Lc above, we can treat Lc as an r r object that depends only on the image of c in the set Csim,reg pGq of -orbits r p in Csim,reg pGq. In the case G SOp2n ` 1, Cq we have just agged, this is part of the denition. In either of the other two cases, suppose that c1 c r is the element in the orbit of c attached to P . The mapping g w g pwq w, r g w P Lc ,

(8.5.9)

v P ValpF q,

r r is then a canonical L-ismorphism from Lc to Lc1 , which is compatible with r the associated diagrams (8.5.9). We can therefore regard c as a -orbit in all three cases. With this interpretation, we still have a canonical extension (8.5.7), with complexication (8.5.8) and local diagrams (8.5.9). It again comes with a canonical isomorphism onto the concrete object we have attached to any c1 in the class c. We can now dene our subgroup of L . We have assigned a group F,reg r r r r Lc with complexication Lc,C to each c P Csim,reg . Let us identify Csim,reg with its injective image in Csim,reg . For any c in its complement, we just r r set Lc Lc and Lc,C Lc,C . With these conventions, we amalgamate the groups as a bre product r r (8.5.10) L pLc WF q F,reg
cPCsim,reg

over F . We thus obtain a locally compact extension (8.5.11) r r 1 KF,reg L F,reg WF 1

508

8. THE GLOBAL CLASSIFICATION

r of WF by a product KF,reg of compact, connected groups. It comes with a complexication r r 1 KF,reg,C L F,reg,C WF 1, r where KF,reg,C is an (innite) topological product of complex dual groups, and an associated family of conjugacy classes of embeddings LFv WFv (8.5.12) , v P ValpF q.

r L F,reg,C WF
r r To embed L F,reg as a subgroup of LF,reg , consider a class c P Csim,reg . r By choosing suitable representatives of c in Csim,reg and Csim,reg , we obtain r r a canonical L-embedding of Lc into Lc from the standard embedding of Kc r into Kc . If c belongs to the complement of Csim,reg in Csim,reg , the groups r Lc,reg and Lc,reg are equal by denition. It follows from the denitions (8.5.4) and (8.5.10) that there is a canonical L-embedding

(8.5.13)

r L F,reg LF,reg ,

which is compatible with the local diagrams (8.5.6) and (8.5.12). Consider an L-homomorphism L r r r : L G, G P Esim pN q. F,reg r We shall say that is standard if it can be embedded into a commutative diagram
r L F,reg r LG

L F,reg GLpN, Cq where the left hand vertical arrow is the L-embedding we have just constructed, the right hand vertical arrow is the representation determined by G as a twisted endoscopic datum, and is a standard representation r of L F,reg dened as above. We write reg,bdd pGq for the set of standard r L-homomorphisms , taken up to the equivalence relation dened by conL G by the group Aut pGq. We dene pGq to be the set of r N r jugation of reg L-homomorphisms r r : L SU p2q L G
F,reg

r r whose restriction to L F,reg belongs to reg,bdd pN q, taken up to the same equivalence relation. r The point of course is that one could use the single group L F,reg in r reg pGq of place of the family of ad hoc groups L attached to the subset

8.5. AN APPROXIMATION OF THE LANGLANDS GROUP

509

r r regular parameters in the earlier set pGq. A parameter (1.4.1) in pGq would again be called regular if the cuspidal automorphic representations r i P Acusp pmi q in its simple constituents i i b i are regular. One sees r r from the denitions that there is a bijection from pGq onto reg pGq that reg is compatible with the local diagrams (8.5.12) and (1.4.14), and such that pL F,reg,C q L , r P pGq, reg

for suitable representatives of the corresponding parameters within their r OutN pGq-orbits. Moreover, the bijection preserves the global centralizers S , their localizations Sv , and the associated homomorphisms Sv . S
v

Therefore our theorems, insofar as they apply to regular parameters, can r indeed be formulated in terms of the group L . F,reg r We have reserved the symbols L and L for the locally compact exF F r tensions of L and L that could be expected if one included the F,reg F,reg complementary, nonregular representations in the construction. If c lies in the complement of Csim,reg pN q in Csim pN q, c should be a cyclic group, whose order d then divides N . In this case, c would be obtained by automorphic induction from the subgroup WK of WF attached to Xc by class eld theory. One would dene Lc as an extension (8.5.1) of WF by the smaller simply connected group Kc SU pkq SU pkq, loooooooooooomoooooooooooon
d

dk N,

on whose factors WF acts by permutation through its cyclic quotient WF {WK . For any point c1 c in the -orbit of c, the L-isomorphism from Lc to Lc1 would then depend only on the image of in {c , and would therefore be canonical. This would lead to a canonical group Lc that depends only on ` the image of c in the set Csim pN q of OpN q -orbits in Csim pN q. Similar r properties could be expected of any point c in the complement of Csim,reg pN q r r in Csim pN q. The would lead to a canonical group Lc that depends only on r pN q of -orbits in Csim pN q. We would then r r the image of c in the set Csim r be able to dene the larger groups L and L as bre products (8.5.4) and F F (8.5.10), but taken over the larger set Csim Csim pN q.
N

r The detailed construction of the larger groups L and L should not be F F dicult, given the global results of 8.1 and 8.2. However, it would take us beyond the methods of this volume. We add here only the remark that the two groups will come with an L-embedding r L L ,
F F

510

8. THE GLOBAL CLASSIFICATION

r and two larger families of parameters pN q and pGq. These would allow us to formulate our global results entirely in terms of the natural approxir mations L and L of the Langlands group LF . We hope to return to these F F matters at some point in the future, in the general context of quasiclassical groups.

CHAPTER 9

Inner Forms
In preparation. 9.1. Inner twists and centralizers 9.2. Statement of theorems for inner forms 9.3. Local correspondence for inner forms 9.4. Nontempered packets for inner forms 9.5. Completion of the proof

511

Bibliography
[ABV] J. Adams, D. Barbasch, and D. Vogan, The Langlands Classication and Irreducible Characters for Real Reductive Groups, Progr. Math. 104, Birkhauser, Boston, 1992. J. Adams and J.-S. Huang, Kazhdan-Patterson lifting for GLpn, Rq, Duke Math. J. 89 (1997), 423443. J. Adams and J. Johnson, Endoscopic groups and packets for non-tempered representations, Compositio Math. 64 (1987), 271309. J. Arthur, Eisenstein series and the trace formula, in Automorphic Forms, Representations and L-functions, Proc. Sympos. Pure Math. 33 (1979), Part 1, Amer. Math. Soc., 253274. , On a family of distributions obtained from Eisenstein series I: Application of the Paley-Wiener theorem, Amer. J. Math. 104 (1982), 12431288. , On a family of distributions obtained from Eisenstein series II: Explicit formulas, Amer. J. Math. 104 (1982), 12891336. , The local behaviour of weighted orbital integrals, Duke Math. J. 56 (1988), 223293. , The invariant trace formula II. Global theory, J. Amer. Math. Soc. 1 (1988), 501554. , The L2 -Lefschetz numbers of Hecke operators, Invent. Math. 97 (1989), 257290. , Intertwining operators and residues I. Weighted characters, J. Funct. Anal. 84 (1989), 1984. , Unipotent automorphic representations: conjectures, Astrisque 171 e 172 (1989), 1371. , Unipotent Automorphic Representations: Global Motivation, in Automorphic Forms, Shimura Varieties and L-functions, vol. I, Academic Press, 1990, 175. , On elliptic tempered characters, Acta. Math. 171 (1993), 73138. , On local character relations, Selecta Math. 2, No. 4 (1996), 501579. , On the transfer of distributions: weighted orbital integrals, Duke Math. J. 99 (1999), 209283. , Stabilization of a family of dierential equations, Proc. Sympos. Pure Math. 158 (2000), 7798. , A stable trace formula I. General expansions, Journal of the Inst. of Math. Jussieu 1 (2002), 175277. , A stable trace formula II. Global descent, Invent. Math. 143 (2001), 157220. , A stable trace formula III. Proof of the main theorems, Annals of Math. 158 (2003), 769873. , A note on the automorphic Langlands group, Canad. Math. Bull. 45 (2002), 466482.

[AH] [AJ] [A1]

[A2] [A3] [A4] [A5] [A6] [A7] [A8] [A9]

[A10] [A11] [A12] [A13] [A14] [A15] [A16] [A17]

513

514

BIBLIOGRAPHY

, An introduction to the trace formula, in Harmonic Analysis, the Trace Formula, and Shimura Varieties, Clay Mathematics Proceedings, vol. 4, 2005, 1263. [A19] , A note on L-packets, Pure and Applied Math Quarterly 2 (2006), 199 217. [A20] , Induced representations, intertwining operators and transfer, Contemporary Math. 449 (2008), 5167. [A21] , Problems for real groups, Contemporary Math. 472 (2008), 3962. [A22] , Report on the trace formula, to appear in Contemporary Math. [A23] , The embedded eigenvalue problem for classical groups, to appear in Clay Mathematics Proceedings. [A24] , Endoscopy and singular invariant distributions, in preparation. [A25] , Duality, Endoscopy and Hecke operators, in preparation. [A26] , A nontempered intertwining relation for GLpN q, [A27] Transfer factors and Whittaker models, in preparation. [AC] J. Arthur and L. Clozel, Simple Algebras, Base Change, and the Advanced Theory of the Trace Formula, Ann. of Math. Studies 120, Princeton Univ. Press, Princeton, N.J., 1989. [Au] A.-M. Aubert, Dualit dans le groupe de Grothendieck de la catgorie des e e reprsentations lisses de longueur nie dun groupe rductif p-adique, Trans. e e Amer. Math. Soc. 347 (1995), 21792189, and Erratum, Ibid. 348 (1996), 4687 4690. [Ba] A. Badalescu, Global Jacquet-Langlands correspondence, multiplicity one and classication of automorphic representations, Invent. Math. 172 (2008), 383 438. [Ban] D. Ban, The Aubert involution and R-groups, Ann. Sci. Ecole Norm. Sup. 35 (2002), 673693. [BV] D. Barbasch and D. Vogan, Unipotent representations of complex semisimple Lie groups, Ann. of Math. 121 (1985), 41110. [Be] J. Bernstein, P -invariant distributions on GLpnq, in Lie Group Representations II, Lecture Notes in Math., vol. 1041, Springer, New York, 1984, 50102. [BDK] J. Bernstein, P. Deligne and D. Kazhdan, Trace Paley-Wiener theorem for reductive p-adic groups, J. dAnalyse Math. 47 (1986), 180192. [BZ] J. Bernstein and A. Zelevinsky, Induced representations of reductive p-adic groups I, Ann. Sci. Ecole Norm. Sup. 10 (1977), 441472. [Bo] A. Borel, Automorphic L-functions, in Automorphic Forms, Representations and L-functions, Proc. Sympos. Pure Math. vol. 33, Part 2, Amer. Math. Soc., 1979, 2762. [BoW] A. Borel and N. Wallach, Continuous Cohomology, Discrete Subgroups, and Representations of Reductive Groups, Ann. of Math. Studies 94, Princeton Univ. Press, Princeton, N.J., 1980. [C] W. Casselman, The unramied principal series of p-adic groups I. The spherical function, Compositio Math. 40 (1980), 387406. [CS] W. Casselman and J. Shalika, The unramied principal series of p-adic groups II. The Whittaker function, Compositio Math. 41 (1980), 207231. [Cassels] J. Cassels, Global Fields, in Algebraic Number Theory, Thompson Book Company, Washington, 1967, 4284. [CL1] P.-H. Chaudouard and G. Laumon, Le lemme fondamental pondr I: construce e tions gomtriques, preprint. e e , Le lemme fondamental pondr II: Enoncs cohomologiques, preprint. e e e [CL2] [CC] G. Chenevier and L. Clozel, Corps de nombres peu ramis et formes automore phes autoduales, J. Amer. Math. Soc. 22 (2009), 467519.

[A18]

BIBLIOGRAPHY

515

[Clo] [CD]

[CLL] [CHL] [CKPS1] [CKPS2] [CH] [D1]

[D2] [DM] [F]

[FLN] [FQ] [G] [GRS]

[GKM1] [GKM2] [Hal] [Ha1] [Ha2] [Ha3] [Ha4] [Ha5]

L. Clozel, Characters of nonconnected, reductive p-adic groups, Canad. J. Math. 39 (1987), 149167. L. Clozel and P. Delorme, Le thor`me de Paley-Wiener invariant pour les e e groupes de Lie rductif II, Ann. Scient. Ec. Norm. Sup., 4e srie, 23 (1990), e e 193228. 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. R. Cluckers, T. Hales, and F. Loeser, Transfer principle for the fundamental lemma, preprint. J. Cogdell, H. Kim, I. Piatetski-Shapiro, and F. Shahidi, On lifting from classical groups to GLN , Publ. Math. Inst. Hautes Etudes Sci. 93 (2001), 530. , Functoriality for the classical groups, Publ. Math. Inst. Hautes Etudes Sci. 99 (2004, 163233. C. Cunningham and T. Hales, Good Orbital Integrals, Represent. Theory 8 (2004), 414457. P. Deligne, Les constantes des quations fonctionnelles des fonctions L, in Mode ular Forms of One Variable II, Lecture Notes in Math. 349, Springer, New York, 1973, 501597. , Les constantes locales de lquation fonctionnelle de la fonction L e dArtin dune reprsentation orthogonale, Invent. Math. 35 (1976), 299316. e P. Delorme and P. Mezo, A twisted invariant Paley-Wiener theorem for real reductive groups, Duke Math. J. 144 (2008), 341380. D. Flath, Decomposition of representations into tensor products, in Automorphic Forms, Representations and L-functions, Proc. Sympos. Pure Math. vol. 33, Part 1, Amer. Math. Soc., 1979, 179184. E. Frenkel, R. Langlands and B.C. Ngo, Formule des traces et functorialit: Le e dbut dun programme, preprint. e A. Frohlich and J. Queyrot, On the functional equations of the Artin L-function for characters of real representations, Invent. Math. 20 (1973), 125138. D. Goldberg, Reducibility of induced representations for Spp2nq and SOpnq, Amer. J. Math. 116 (1994), 11011151. D. Ginzburg, S. Rallis, and D. Soudry, Generic automorphic forms on SOp2n ` 1q: functorial lift to GLp2nq, endoscopy and base change, Internat. Math. Res. Notices 14 (2001), 729764. M. Goresky, R. Kottwitz, and R. MacPherson, Homology of ane Springer bres in the unramied case, Duke Math. J. 121 (2004), 509561. , Purity of equivalued ane Springer bres, Representation Theory 10 (2006), 130146. T. Hales, On the fundamental lemma for standard endoscopy: reduction to unit elements, Canad. J. Math. 47 (1995), 974994. Harish-Chandra, Discrete series for semisimple Lie groups. II. Explicit determination of the characters, Acta Math. 116, 1111. , Harmonic analysis on real reductive groups. I. The theory of the constant term, J. Funct. Anal. 19, 104204. , Harmonic analysis on real reductive groups II. Wave packets in the Schwartz space, Invent. Math. 36 (1976), 155. , Harmonic analysis on real reductive groups III. The Moass-Selberg relations and the Plancherel formula, Ann. of Math. 104 (1976), 117201. , The Plancherel formula for reductive p-adic groups, in Collected Papers, vol. IV, Springer-Verlag, 353367

516

BIBLIOGRAPHY

[HT]

[He1] [He2] [Hi] [JL] [JPS] [JS] [JiS1] [JiS2]

[JiS3] [Ka] [KeS] [Kn] [KnS] [KnZ1]

[KnZ2] [Kon] [Kos] [K1] [K2] [K3] [KR] [KS] [LL] [L1]

M. Harris and R. Taylor, On the Geometry and Cohomology of Some Simple Shimura Varieties, Ann. of Math. Studies 151, Princeton Univ. Press, Princeton, N.J., 2001. G. Henniart, Une preuve simple des conjectures de Langlands de GLpnq sur un corps p-adique, Invent. Math. 139 (2000), 439455. , Correspondence de Langlands et fonctions L des carrs extrieur et e e symtrique, preprint, IHES, 2003. e K. Hiraga, On functoriality of Zelevinski involutions, Compositio Math. 140 (2004), 16251656. H. Jacquet and R. Langlands, Automorphic Forms on GLp2q, Lecture Notes in Math., vol. 114, Springer, New York, 1970. H. Jacquet, I. Piatetski-Shapiro, and J. Shalika, Rankin-Selberg convolutions, Amer. J. Math. 105 (1983), 367464. H. Jacquet and J. Shalika, On Euler products and the classication of automorphic representations II, Amer. J. Math. 103 (1981), 777815. D. Jiang and D. Soudry, The local converse theorem for SOp2n ` 1q and applications, Annals of Math. 157 (2003), 743806. , Generic representations and local Langlands reciprocity law for p-adic SOp2n ` 1q, in Contributions to Automorphic Forms, Geometry and Number Theory, Johns Hopkins Univ. Press, Baltimore, MD, 2004, 457519. , On local descent from GLpnq to classical groups, preprint. D. Kazhdan, Cuspidal geometry on p-adic groups (Appendix), J. Analyse Math. 47 (1980), 136. D. Keys and F. Shahidi, Artin L-functions and normalization of intertwining operators, Ann. Scient. Ec. Norm. Sup. 4e srie, t. 21 (1988), 6789. e A. Knapp, Commutativity of intertwining operators. II, Bull. Amer. Math. Soc. 82 (1976), 271273. A. Knapp and E. Stein, Intertwining operators for semisimple groups II, Invent. Math. 60 (1980), 984. A. Knapp and G. Zuckerman, Normalizing factors and L-groups, in Automorphic Forms, Representations and L-functions, Proc. Sympos. Pure Math. vol. 33, Part 1, Amer. Math. Soc., 1979, 93106. , Classication of irreducible tempered representations of semisimple groups, Ann. of Math. 116 (1982), 389455. T. Konno, Twisted endoscopy and the generic packet conjecture, Israel J. Math. 129 (2002), 253289. B. Kostant, On Whittaker vectors and representation theory, Invent. Math. 48 (1978), 101184. R. Kottwitz, Stable trace formula: cuspidal tempered terms, Duke Math. J. 51 (1984), 611650. , Stable trace formula: elliptic singular terms, Math. Ann. 275 (1986), 365399. , Shimura Varieties and -adic Representations, in Automorphic Forms, Shimura Varieties and L-functions, vol. 1, Academic Press, 1990, 161209. R. Kottwitz and J. Rogawski, The distributions in the invariant trace formula are supported on characters, Canad. J. Math. 52 (2000), 804814. R. Kottwitz and D. Shelstad, Foundations of Twisted Endoscopy, Astrisque, e vol. 255. J.-P. Labesse and R. Langlands, L-indistinguishability for SLp2q, Canad. J. Math. 31 (1979), 726785. R. Langlands, Eisenstein series, in Algebraic Groups and Discontinuous Subgroups, Proc. Sympos. Pure Math., vol. 9, Amer. Math. Soc., 1966, 235252.

BIBLIOGRAPHY

517

[L2]

[L3] [L4] [L5] [L6]

[L7]

[L8] [L9] [L10] [L11]

[L12] [L13] [L14] [LS1] [LS2] [Lap] [Lar] [LN] [Li] [Liu] [Ma] [Me] [M1] [M2] [M3]

, Problems in the theory of automorphic forms, in Lectures in Modern Analysis and Applications, Lecture Notes in Math. 170, Springer, New York, 1970, 1861. , On Artins L-function, Rice University Studies 56 (1970), 2328. , Euler Products, Yale University Press, 1971. , On the Functional Equations Satised by Eisenstein Series, Lecture Notes in Math. 544, Springer, New York, 1976. , On the notion of an automorphic representation. A supplement to the preceding paper, in Automorphic Forms, Representations and L-functions, Proc. Sympos. Pure Math. vol. 33, Part 1, Amer. Math. Soc., 1979, 203208. , Automorphic representations, Shimura varieties, and motives. Ein Mrchen, in Automorphic Forms, Representations and L-functions, Proc. Syma pos. Pure Math. vol. 33, Part 2, Amer. Math. Soc., 1979, 205246. Stable conjugacy: denitions and lemmas, Canad. J. Math. 31 (1979), 700725. , Base Change for GLp2q, Ann. of Math. Studies 96, Princeton Univ. Press, Princeton, N.J., 1980. , Les dbuts dune formule des traces stables, Publ. Math. Univ. Paris VII e 13, 1983. , On the classication of irreducible representations of real algebraic groups, in Representation Theory and Harmonic Analysis on Semisimple Lie Groups, AMS Mathematical Surveys and Monographs, vol. 31, 1989, 101170. , Representations of abelian algebraic groups, Pacic J. Math. (1997), 231250. , Beyond endoscopy, in Contributions to Automorphic Forms, Geometry, and Number Theory, Johns Hopkins University Press, 2004, 611698. , Un nouveau point de rep`re des formes automorphes, Canad. Math. e Bull. 50 (2007), 243267. R. Langlands and D. Shelstad, On the denition of transfer factors, Math. Ann. 278 (1987), 219271. , Descent for transfer factors, The Grothendieck Festschift, Vol. II, Birkhauser, Boston, 1990, 485563. E. Lapid, On the root number of representations of orthogonal type, Compositio Math. 140 (2004), 274286. M. Larsen, On the conjugacy of element-conjugate homomorphisms, Israel J. Math. 88 (1994), 253277. G. Laumon and B.C. Ngo, Le lemme fondamental pour les groupes unitaires, Annals of Math. 168 (2008), 477573. C. Li, A Local Twisted Trace Formula and Twisted Orthogonality Relations, in preparation. B. Liu, Genericity of Representations of p-Adic Sp2n and Local Langlands Parameters, Canad. J. Math., (published electronically March 8, 2011). I.G. Macdonald, Spherical Functions on a Group of p-Adic Type, Publications of the Ramanujan Institute, Madras, 1971. P. Mezo, Spectral transfer in the twisted endoscopy of real reductive groups, preprint. C. Moeglin, Sur certaines paquets dArthur et involution dAubert-SchneiderStuhler gnralise, Represent. Theory 10 (2006), 89129. e e e , Paquets dArthur discrets pour un groupe classique p-adique, Contemp. Math. 489 (2009), 179257. , Paquets dArthur pour les groupes classiques p-adiques; point de vue combinatoire, preprint.

518

BIBLIOGRAPHY

[M4] [MW1] [MW2] [MW3] [MW4] [Mu] [N] [Paris] [Re] [Rod1] [Rod2]

[Ro1] [Ro2] [Ro3] [Sa] [ScS] [Sha1] [Sha2] [Sha3] [Sha4] [Shal] [S1] [S2] [S3] [S4]

[S5] [S6] [S7]

, Multiplicit 1 dans les paquets dArthur aux places p-adiques, to appear e in Clay Mathematics Proceedings. C. Moeglin and J.-L. Waldspurger, Mod`les de Whittaker dgnrs pour des e e e e e groupes p-adiques, Math. Z. 1996 (1987), 427452. , Le spectre rsiduel de GLpnq, Ann. Scient. Ec. Norm. Sup. 4e srie 22 e e (1989), 605674. , Pacquets stables de reprsentations temprs et rductions unipotent e e e e pour SOp2n ` 1q, Invent. Math. 152 (2003), 461623. , Sur le transfert des traces dun groupe classique p-adique a un groupe ` linaire tordu, Select Math. 12 (2006), 433515. e W. Mller, The trace class conjecture in the theory of automorphic forms, Ann. u of Math. 130 (1989), 473529. B.C. Ngo, Le lemme fondamental pour les alg`bres de Lie, preprint. e Paris Book Project, Stabilization de la formule des traces, varits de Shimura, ee et applications arithmtiques, (ed. M. Harris). e D. Renard, Endoscopy for real reductive groups, preprint. F. Rodier, Whittaker models for admissible representations of reductive p-adic split groups, Proc. Sympos. Pure Math. 26 (1974), Amer. Math. Soc., 425430. , Mod`le de Whittaker et caract`res de reprsentations, in None e e Commutative Harmonic Analysis, Lecture Notes in Math. 466, Springer, New York, 1975, 151171. J. Rogawski, The trace Paley-Wiener theorem in the twisted case, Trans. Amer. Math. Soc. 309 (1988), 215229. , Automorphic Representations of Unitary Groups in Three Variables, Ann. of Math. Studies 123, Princeton Univ. Press, Princeton, N.J., 1990. , The multiplicity formula for A-packets, in The Zeta Functions of Picard Modular Surfaces, Les Publications CRM, Montreal, 1992, 395420. I. Satake, Classication Theory of Semisimple Algebraic Groups, notes prepared by D. Schattschneider, M. Dekker, New York, 1971. P. Schneider and U. Stuhler, Representation theory and sheaves on the BruhatTits building, Pub. Math. IHES 85 (1997), 97191. F. Shahidi, On certain L-functions, Amer. J. Math. 103 (1981), 297355. , Local coecients as Artin factors for real groups, Duke Math. J. 52 (1985), 9731007. , On the Ramanujan conjecture and niteness of poles for certain Lfunctions, Annals of Math. 127 (1988), 547584. , A proof of Langlands conjecture on Plancherel measures; Complementary series for p-adic groups, Annals of Math. 132 (1990), 273330. J. Shalika, The multiplicity one theorem for GLn , Annals of Math. 100 (1974), 171193. D. Shelstad, Orbital integrals and a family of groups attached to a real reductive group, Ann. Sci. Ecole Norm. Sup. 12 (1979), 131. , L-indistinguishability for real groups, Math. Ann. 259 (1982), 385430. , Tempered endoscopy for real groups I: geometric transfer with canonical factors, Contemp. Math., 472 (2008), 215246. , Tempered endoscopy for real groups II: spectral transfer factors, to appear in Autmorphic Forms and the Langlands Program, Higher Education Press/International Press, 243283. , Tempered endoscopy for real groups III: inversion of transfer and Lpacket structure, Representation Theory 12 (2008), 369402. , On geometric transfer in real twisted endoscopy, preprint. , On spectral factors in real twisted endoscopy, preprint.

BIBLIOGRAPHY

519

A. Silberger, Introduction to Harmonic Analysis on Reductive p-adic Groups, Mathematical Notes, Princeton University Press, 1979. [Si2] , The Knapp-Stein dimension theorem for p-adic groups, Proc. Amer. Math. Soc. 76 (1978), 243246; Correction, Proc. Amer. Math. Soc. 76 (1979), 169170. [So] D. Soudry, On Langlands functoriality from classical groups to GLn , Astrisque e 298 (2005), 335390. [SpehV] B. Speh and D. Vogan, Reducibility of generalized principal series representations, Acta. Math. 145 (1980), 227299. [Sp] T. Springer, Linear Algebraic Groups, Prog. Math. 9, 1981, Birkhuser, Boston, a Basel, Stuttgart. [SpS] T. Springer and R. Steinberg, Conjugacy classes, in Seminar on Algebraic Groups and Related Finite Groups, Lecture Notes in Math., vol. 131, 1970, 167266. [Tad1] M. Tadic, Topology of unitary dual of non-archimedean GLpnq, Duke Math. J. 55 (1987), 385422. [Tad2] , On characters of irreducible unitary representations of general linear groups, Abh. Math. Sem. Univ. Hamburg 65 (1995), 341363. [T1] J. Tate, Fourier Analysis in Number Fields and Heckes Zeta Functions, in Algebraic Number Theory, Thompson, Washington, D.C., 1967, 305347. [T2] , Number theoretic background, in Automorphic Forms, Representations and L-functions, Proc. Sympos. Pure Math. vol. 33, Part 2, Amer. Math. Soc., 1979, 326. [Ti] J. Tits, Classication of Algebraic Semisimple Groups, in Algebraic Groups and Discontinuous Subgroups, Proc. Sympos. Pure Math., vol. 9, Amer. Math. Soc., 1966, 3362. [V1] D. Vogan, Gelfand-Kirillov dimension for Harish-Chandra modules, Invent. Math. 48 (1978), 7598. , The unitary dual of GLpnq over an archimedean eld, Invent. Math. 83 [V2] (1986), 449505. , The local Langlands conjecture, Contemp. Math. 145 (1993), 305379. [V3] [W1] J.-L. Waldspurger, Le lemme fondamental implique le transfer, Compositio Math. 105 (1997), 153236. [W2] , Reprsentations de rduction unipotente pour SOp2n ` 1q: quelques e e consquences dune article de Lusztig, Contributions to automorphic forms, gee ometry, and number theory, 803910, Johns Hopkins University Press, Baltimore, MD, 2004. [W3] , Endoscopie et changement de caractristique, J. Inst. Math. Jussieu 5 e (2006), 423525. [W4] , Le groupe GLN tordu, sur un corps p-adique. I, Duke Math J. 137 (2007), 185234. [W5] , Le groupe GLN tordu, sur un corps p-adique. II, Duke Math J. 137 (2007), 235336. [W6] , Lendoscopie tordue nest pas si tordue, Memoirs of AMS 908 (2008). [W7] , A propos du lemme fondamental pondr tordu, Math. Ann. 343, 103 e e 174. [We1] H. Weyl, Theorie der Darstellung kontinuierlicher halb-einfacher Gruppen durch lineare Transformationen, Gesammelte Abhandlungen, Bd. II, 68, 543647. [We2] , The Theory of Groups and Quantum Mechanics, 1931, rept. Dover Publications, 1950. [Z] A. Zelevinsky, Induced representations of reductive p-adic groups II. On irre ducible representations of GLpnq, Ann. Sci. Ecole Norm. Sup. 13 (1980), 165 210.

[Si1]