Functoriality for the classical groups
Cogdell, J. W.; Kim, H. H.; Piatetski-Shapiro, I. I.; Shahidi, F.
2004-09-09 00:00:00
by J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, and F. SHAHIDI Functoriality is one of the most central questions in the theory of automor- phic forms and representations [1, 2, 35, 36]. Locally and globally, it is a manifesta- tion of Langlands’ formulation of a non-abelian class field theory. Now known as the Langlands correspondence, this formulation of class field theory can be viewed as giving an arithmetic parameterization of local or automorphic representations in terms of admissible homomorphisms of (an appropriate analogue) of the Weil- Deligne group into the Langlands dual group or L-group. When this conjectural parameterization is combined with natural homomorphisms of the L-groups it pre- dicts a transfer or lifting of local or automorphic representations of two reductive algebraic groups. As a purely automorphic expression of a global non-abelian class field theory, global functoriality is inherently an arithmetic process. In this paper we establish global functoriality from the split classical groups G = SO ,SO ,or Sp to an appropriate general linear group GL , associated n 2n+1 2n 2n N to the natural embedding of L-groups, for globally generic cuspidal representations π of G (A) over a number field k. We had previously presented functoriality for thecaseG = SO in [6], but were limited at that time by a lack of suit- n 2n+1 able local tools in the other cases. The present paper is by no means a simple generalization of [6]. There were serious local problems to be overcome in the development of the tools that now allow us to cover all three series of classical groups simultaneously and that will be applicable to other cases of functoriality in the future. In addition, we have completely determined the associated local im- ages of functoriality and as a result are able to present several new applications of functoriality, including both global results concerning the Ramanujan conjecture for the classical groups and various applications to the local representation theory of the classical groups. There are several approaches to the question of functoriality: the trace for- mula, the relative trace formula, and the Converse Theorem. In this work we use the Converse Theorem, which is an L-function method. The Converse Theorem itself states that if one has an irreducible admissible representation Π ⊗ Π of GL (A),then Π is in fact automorphic if sufficiently many of its twisted L-functions L(s, Π × τ),with τ cuspidal automorphic representations of smaller GL (A), are nice [7, 9]. As a vehicle for establishing functoriality from cuspidal representations π =⊗π of some G (A) to an automorphic representation of During the work on this project J. W. C. was partially supported by the NSA through grant MDA 904-03-1-0043, H. H. K. by a grant from NSERC and an AMS Centennial Fellowship, I. I. P-S. by the NSF through grant DMS-0072450, and F. S. by the NSF through grant DMS-0200325. DOI 10.1007/s10240-004-0020-z 164 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI GL (A), there are three main steps. The first is to construct a candidate lift Π =⊗ Π . This is done by locally lifting each local component representation π v v of G (k ) to Π of GL (k ) in such a way that twisted local L-and ε-factors are n v v N v matched. At the archimedean places and the finite places where π is unramified we may accomplish this local lift by using the local Langlands correspondence, i.e., the local arithmetic Langlands classification. At the remaining finite set of places where π is ramified we must finesse the lack of a local Langlands correspondence by using the stability of the local γ -factor under highly ramified twists. This highly ramified twist has the effect of “washing out” any subtle local properties of the representation and gives a matching of local representations π of G (k ) and Π v n v v of GL (k ) for which the twisted local L-and ε-factors match after this highly N v ramified twist. We used this method in [6], however the key new ingredient here is a uniform method of expressing the local γ -factor as the Mellin transform of a Bessel function in fairly wide generality which is applicable in all of our cases as well as many more [55]. With this new general result in hand, the necessary stability result then follows from the asymptotic properties of the Bessel functions as in [8]. With this, we can finally lift π locally to Π at all places and form v v a candidate lift Π =⊗ Π such that L(s, Π × τ) = L(s,π × τ) for all τ in a suit- able twisting set. The second step is to then control the analytic properties of the twisted L-functions L(s,π × τ) on the classical groups. As in our previous work, we control these L-functions through the Fourier coefficients of Eisenstein series – the Langlands-Shahidi method. Once we know that the L(s,π × τ), and hence L(s, Π × τ), are nice for a suitable twisting set of τ , we may move to the third main step, which is the application of the Converse Theorem for GL to the representation Π. This then gives global functoriality from any of the G to the appropriate GL (Theorem 1.1). Assuming the existence of global functoriality, the result which we establish here, Ginzburg, Rallis, and Soudry had previously used their descent technique to characterize the image of global functoriality for globally generic representations of the split classical groups [56]. In particular, they show the image of global functori- ality consists of isobaric sums of certain self-dual cuspidal representations of GL (A) satisfying an appropriate L-function criterion (Theorems 7.1 and 7.2). Using the rigidity of isobaric representations afforded by the strong multiplicity one theorem for isobaric representations of GL (A) [21], this implies that there is in fact no ambiguity in our global functorial lift coming from our use of the highly ramified twist and we are able to then compute explicitly the compatible local functorial lifts of the various series of generic representations of G (k ).For thecase of n v G = SO this was done in [30] and we follow that general method here, but n 2n+1 again giving a uniform treatment for all split classical groups. For generic super- cuspidal representations π we show that their lift is a local isobaric sum of certain self-dual supercuspidal representations of general linear groups, again satisfying the FUNCTORIALITY FOR THE CLASSICAL GROUPS 165 appropriate L-function condition (Theorem 7.3). With the local supercuspidal lift in hand, one can then use the classification of local generic discrete series repre- sentations for the classical groups [24, 25, 39, 40, 43] to compute the explicit form of their lifts (Proposition 7.3), and in turn work one’s way up the classification to obtain explicit lifts of generic tempered representations (Proposition 7.4) and finally of arbitrary generic representations (Proposition 7.5). Finally we are able to re- fine these local results to compute the local factors of our global functorial lift Π (Theorem 7.4) with a second application of the Converse Theorem. This explicit knowledge of the local functorial lifts is crucial to our applications. Let us note that in the case G = SO Jiang and Soudry [26, 27] were n 2n+1 able to combine our global functoriality with the local descent to Sp and then 2n the theta correspondence to prove a Local Converse Theorem for SO over a 2n+1 p-adic field. This allowed them to prove the injectivity of the local functorial lifts as we have defined them here and establish the local Langlands correspondence . Once the local descent is available in the other cases, we would expect for SO 2n+1 similar results to follow. However, for the other classical groups the Local Converse Theorem will be more subtle since the torus does not act transitively on the set of generic characters. This will lead to more than one generic representation in each local L-packet, distinguished by their character of genericity. For a clean statement one may need to pass to similitude groups. The global application we present is indeed of an arithmetic nature and concerns the Ramanujan conjecture for generic representations of the split clas- sical groups. In the late 1970’s, when the first counterexamples to the general- ized Ramanujan conjecture for reductive groups were found for Sp and U [18], 4 3 the Ramanujan conjecture for a general reductive group G(A) was refined and conjectured to hold for generic cuspidal representations of quasi-split reductive groups [18, 45, 50]. On the other hand, Langlands, in Section 3 of [35], sug- gests that the Ramanujan conjecture should hold for cuspidal representations of quasi-split groups which functorially lift to isobaric representations of GL (A) (cf. the Remark at the end of Section 10 here). This is the case for the globally generic representations of our classical groups G (A) as we have noted above. Thus, with either formulation, we would expect that if π ⊗ π is a generic cus- pidal automorphic representation of G (A) then each local component π should n v be tempered. This is widely believed to hold for GL . The best current general bounds towards Ramanujan for GL (A) over a number field are those of Luo, Rudnick, and Sarnak [37]. Via functoriality we are able to link the Ramanu- jan conjecture for globally generic representations of the split classical groups to the Ramanujan conjecture for cuspidal representations of GL (Theorem 10.1). In particular, we show that the Ramanujan conjecture for these groups, in its strong form giving temperedness at all places, would follow from the Ramanujan conjecture for GL (Corollary 10.2), at least for globally generic cuspidal rep- N 166 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI resentations, and any bounds towards Ramanujan for GL ,such asthe Luo– Rudnick–Sarnak bounds, lead to similar bounds for the classical groups (Corol- lary 10.1). We note that once our results on functoriality are extended to the case of global function fields, which is primarily a matter of understanding the theory of L-functions for the classical groups over a global function field, then the Ra- manujan conjecture for these groups over a global function field would become a theorem, thanks to Lafforgue’s proof of the Ramanujan conjecture for GL over a global function field [33]. We hope to return to this extension in future papers. Even though functoriality is inherently arithmetic, many of its applications are to the local representation theory of the groups G . These results seem difficult to establish locally on the classical groups themselves, but are rather straightforward applications of functoriality. The first local application presented in this paper is a proof of Mœglin’s conjecture on the “dimension relation” for generic discrete series representations π of p-adic split classical groups of G (k ) [38]. This rela- v n v tion essentially states that the sum of the sizes of the Jordan blocks associated to π is equal to the dimension of the natural representation of the L-group of G , v n which is itself equal to the rank N of the general linear group GL to which π functorially lifts (Theorem 8.1). Our second application is to establishing of the basic properties of the conductor of a generic representation π of G (k ).The v n v conductor is the exponent f (π ) occurring in the local ε-factor ε(s,π ,ψ ).We v v v show, as is known to be the case for general linear groups [19], that f (π ) is a non-negative integer and f (π ) = 0 iff π is unramified (Theorem 9.1). E. Lapid v v has informed us that this should in turn have applications to the relative trace for- mula. Finally, we turn to one local application which in turn is expected to have global arithmetic applications. Using our bounds towards Ramanujan we show that the local normalized intertwining operators N(s,π ×π ) with π a local component v v of a globally generic cuspidal representation π of G (A) and π a generic repre- sentation of GL (k ), are holomorphic and non-vanishing for Re(s) ≥ 0 (Theo- m v rem 11.1). For G = SO this was done in [29]. This local result is neces- n 2n+1 sary for the understanding of the global residual spectrum of the classical groups G (A) [29]. While this project has been in the works for several years, the finalization of the proof of functoriality and the formulation of most of the applications took place while three of the authors were participants in the Thematic Program on Automorphic Forms held at the Fields Institute for Research in the Mathematical Sciences in the spring of 2003. We would like to thank the Fields Institute for providing us with a wonderful working environment. We would also like to thank D. Ban, C. Jantzen, G. Muic, and M. Tadic for helpful discussions on the classi- ´ ´ fication of generic discrete series representations. Finally, we thank the referee for several pertinent comments and corrections. FUNCTORIALITY FOR THE CLASSICAL GROUPS 167 CONTENTS 1. Functoriality for classical groups ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 167 2. The Converse Theorem . . . .... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 170 × GL . .... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 171 3. L-functions for G n m 4. Stability of γ -factors for G × GL ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 173 n 1 5. The candidate lift .. ... ... .... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 187 6. Global functoriality .... ... .... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 193 7. The image of functoriality ..... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 194 8. A conjecture of Mœglin . .. .... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 213 9. The conductor of a generic representation ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 217 10. The Ramanujan conjecture . ... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 221 11. Normalized local intertwining operators .. ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 225 A. Appendix ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... ... ... ... ... ... ... ... ... ... ... .... 228 1. Functoriality for classical groups Let k be a number field and let A = A be its ring of adeles. We fix a non-trivial continuous additive character ψ of A which is trivial on the principal adeles k. We will let G denote a split classical group of rank n defined over k.More specifically, we will consider the following three cases. (i) Odd orthogonal groups. In this case G = SO , the split special orthogonal n 2n+1 group in 2n + 1 variables defined over k. For definiteness, we will take G as the con- nected component of the isometry group of the form Φ = . The con- 2n+1 nected component of the Langlands dual group of G is G = Sp (C). n 2n = SO , the split special orthogonal (ii) Even orthogonal groups. In this case G n 2n group in 2n variables defined over k. We will again take G as the connected compon- ent of the isometry group of the form Φ = . The connected component 2n is G = SO (C). of the Langlands dual group of G n 2n (iii) Symplectic groups. In this case G = Sp ,the symplectic groupin 2n variables n 2n defined over k. For definiteness, we will take G as the isometry group of the alternat- 0 Φ ing form J = . The connected component of the Langlands dual group 2n −Φ 0 of G is G = SO (C). n 2n+1 In these realizations we can take the standard Borel subgroup of G to be rep- resented by upper triangular matrices. We will denote this Borel subgroup by B and its unipotent radical by U . The abelianization of U is a direct sum of copies of k n n and we may use ψ to define a non-degenerate character of U (A) which is trivial on U (k). By abuse of notation we continue to call this character ψ. n 168 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI In each of these cases there is a general linear group GL such that G em- beds naturally in GL (C) = GL .Since both G and GL are split, this embedding N n N L L completely determines an L-homomorphism ι : G → GL by extending ι to be n N the identity on the Weil group. By Langlands’ principle of functoriality [1, 2, 5], as- sociated to these L-homomorphisms there should be a transfer or lift of automorphic representations from G (A) to GL (A) as in the following table. n N 0 0 0 0 L L L L G G ι : G → GL GL GL n N n n N N SO Sp (C) Sp (C) → GL (C) GL (C) GL 2n+1 2n 2n 2n 2n 2n SO SO (C) SO (C) → GL (C) GL (C) GL 2n 2n 2n 2n 2n 2n Sp SO (C) SO (C) → GL (C) GL (C) GL 2n 2n+1 2n+1 2n+1 2n+1 2n+1 To be more precise, let π =⊗ π be an irreducible automorphic representation of G (A). For v afinite place of k where π is unramified the representation π of G (k ) is v v n v completely determined by its Satake parameter, a semi-simple conjugacy class [t ] in L 0 L G [2, 47]. [t ] then determines a semi-simple conjugacy class [ι(t )] in GL .Anun- v v n N ramified irreducible admissible representation Π of GL (k ) is called the local functorial v N v lift of π if its associated semi-simple conjugacy class in GL is [ι(t )], or equivalently, v v −s −1 L(s, Π ) = det(I − t q ) = L(s,π ). v v v If v is an archimedean place, then by the arithmetic Langlands classification π is determined by an admissible homomorphism ϕ : W −→ G where W is the local v v v Weil group of k [2, 34]. The composition ι ◦ ϕ is an admissible homomorphism of W v v v into GL and hence determines a representation Π of GL (k ) such that L(s, Π ) = v N v v L(s,π ).This isagain the local functorial lift of π . v v An irreducible automorphic representation Π =⊗ Π of GL (A) is called v N a functorial lift of π if for every archimedean place v and for almost all non-archi- medean places v for which π is unramified wehavethat Π is a local functorial lift v v of π . In particular this entails an equality of (partial) Langlands L-functions L (s, Π) = L(s, Π ) = L(s,π ) = L (s,π). (We had previously referred to this lift v v v∈ /S v∈ /S as a weak lift, but there is nothing weak about it. This definition of a functorial lift is consistent with the formulations in [1, 5, 36].) Let π be an irreducible cuspidal representation of G (A). We say that π is glob- ally generic if there is a cusp form ϕ ∈ V such that ϕ has a non-vanishing ψ-Fourier coefficient along U , i.e., such that −1 ϕ(ug)ψ (u) du = 0. U (k)\U (A) n n FUNCTORIALITY FOR THE CLASSICAL GROUPS 169 Cuspidal automorphic representations of GL are always globally generic in this sense. For cuspidal automorphic representations of the classical groups this is a condition. In general the notion of being globally generic may depend on the choice of splitting of the group. However, as is shown in the Appendix to this paper, given a π which is globally generic with respect to some splitting there is always an “outer twist” which is globally generic with respect to a fixed splitting. This outer twist provides an abstract isomorphism between globally generic cuspidal representations and will not effect the L-or ε-factors nor the notion of the functorial lift. Hence we lose no generality in considering cuspidal representations that are globally generic with respect to our fixed splitting. The principle result that we will prove in this paper is the following. Theorem 1.1. — Let k be a number field and let π be an irreducible globally generic cuspidal automorphic representation of G (A).Then π has a functorial lift to GL (A). n N The low dimensional cases of this theorem, that is, when n = 1, are already well understood. As we will need them in the later sections of this paper, let us review them briefly here. (i) Odd orthogonal groups. When n = 1 the split SO PGL . The associated 3 2 lifting from PGL to GL simply takes a representation π of PGL (A) and views it as 2 2 2 arepresentation Π of GL (A) having trivial central character. (ii) Even orthogonal groups. When n = 1 the split SO G GL . The natural 2 m 1 embedding of L-groups then embeds GL in GL as a split rank one torus. The asso- 1 2 ciated lifting then takes a character µ of A GL (A) to the appropriate constituent −1 −1 of the induced representation Ind(µ ⊗ µ ), namely the isobaric sum Π = µ µ which takes the local Langlands quotient at each place if there is reducibility [35]. Let us note that if we take a character µ of A and let π be the corresponding repre- sentation of SO (A) then the standard L-function of π is the degree two L-function 2 µ associated to the standard embedding of L-groups discussed above, so indeed we have −1 −1 L(s,π ) = L(s,µ)L(s,µ ) = L(s,µ µ ), with similar equalities locally. In what follows,wewillmakerecourseto the work of Mœglin and Tadi´ c on local discrete se- ries representations [39, 40]. In keeping with their conventions, we will adopt the con- vention that SO (k ),for a p-adic local field k , has no supercuspidal representations, 2 v v nor discrete series representations. (iii) Symplectic groups. When n = 1 then Sp SL and this functoriality is also 2 2 well understood. The map on dual groups is then PGL (C) SO (C) → GL (C), 2 3 3 which is the adjoint representation of PGL (C).Thus if π is a generic cuspidal repre- sentation of Sp (A) SL (A) then its functorial lift Π to GL (A) is the adjoint square 2 2 3 lifting of Gelbart and Jacquet [11]. Thus we will concentrate primarily on the cases where n ≥ 2. 170 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI The preparations for and proof of Theorem 1.1 when n ≥ 2 will take place over the next five sections. Note that the case of G = SO is completely contained in n 2n+1 our previous paper [6], but we include it here to provide a uniform treatment of all classical groups. 2. The Converse Theorem In order to effect the functorial lifting from G to GL we will use the Converse n N Theorem for GL as we didin[6].Wegivethe formulation here. Letusfixa finite set S of finite places of k.For each integer m,let A (m) ={τ | τ is a cuspidal representation of GL (A)} 0 m A (m) ={τ ∈ A (m) | τ is unramified for all v ∈ S}. 0 v We set N−1 N−1 S S T (N − 1) = A (m) and T (N − 1) = A (m). m=1 m=1 × × If η is a continuous character of k \A ,let us set S S T (S; η) = T (N − 1) ⊗ η ={τ = τ ⊗ η : τ ∈ T (N − 1)}. Theorem 2.1 (Converse Theorem). — Let Π =⊗ Π be an irreducible admissible rep- resentation of GL (A) whose central character ω is invariant under k and whose L-function N Π L(s, Π) = L(s, Π ) is absolutely convergent in some right half plane. Let S be a finite set of × × finite places of k and let η be a continuous character of k \A . Suppose that for every τ ∈ T (S; η) the L-function L(s, Π × τ) is nice, that is, satisfies 1. L(s, Π × τ) and L(s, Π × τ) extend to entire functions of s ∈ C, 2. L(s, Π × τ) and L(s, Π × τ) are bounded in vertical strips, and 3. L(s, Π × τ) satisfies the functional equation L(s, Π × τ) = ε(s, Π × τ)L(1 − s, Π × τ) ˜ . Then there exists an automorphic representation Π of GL (A) such that Π Π for almost N v all v.More precisely, Π Π for all v ∈ / S. In the statement of the theorem, the twisted L-and ε-factors are defined by the products L(s, Π × τ) = L(s, Π × τ)ε(s, Π × τ) = ε(s, Π × τ ,ψ ) v v v v v v v of local factors as in [7, 6]. FUNCTORIALITY FOR THE CLASSICAL GROUPS 171 To motivate the next few sections, let us describe how we will apply this theo- remto the problemof Langlands lifting from G to GL . We begin with our globally n N generic cuspidal automorphic representation π =⊗ π of G (A).For each place v we v n need to associate to π an irreducible admissible representation Π of GL (k ) such v v N v that for every τ ∈ T (S; η) we have L(s,π × τ ) = L(s, Π × τ ) v v v v ε(s,π × τ ,ψ ) = ε(s, Π × τ ,ψ ). v v v v v v For archimedean places v and those non-archimedean v where π is unramified, we take Π to be the local functorial lift of π described above. For those places v where π v v v is ramified, we will take for Π an essentially arbitrary irreducible admissible generic representation of GL (k ) having trivial central character. However, we must choose N v our finite set of places S of k such that S contains the places where π is ramified and × × choose our character η of k \A such that η is sufficiently highly ramified so that L(s,π × η ), L(s, Π × η ), ε(s,π × η ,ψ ),and ε(s, Π × η ,ψ ) are all standard. This v v v v v v v v v v will be possible by the result on the stability of the local γ -factors that we establish in Section 4. Now consider the restricted tensor product Π =⊗ Π . This is an irreducible representation of GL (A). With the choices above we have L(s,π × τ) = L(s, Π × τ) ε(s,π × τ) = ε(s, Π × τ) for Re(s) 0 and all τ ∈ T (S; η). This is our candidate lift. The theory of L-functions for G × GL , which we address in the next sec- n m tion, will then guarantee that the twisted L-functions L(s,π × τ) arenicefor all τ ∈ T (S; η). Then the L(s, Π × τ) will also be nice and Π satisfies the hypotheses of the Converse Theorem. Hence there exists an irreducible automorphic representation Π of GL (A) such that Π Π for all archimedean v and almost all finite v where π N v v is unramified. Hence Π is a functorial lift of π . 3. L-functions for G × GL n m Let π be a globally generic cuspidal representation of G (A).For τ a cuspidal representation of GL (A) we will let L(s,π × τ) be the completed L-function as de- fined in [51] via the theory of Eisenstein series. The local factors are then defined via the arithmetic Langlands classification for the archimedean places, through the Satake parameters for the finite unramified places, as given by the poles of the associated γ -factors (or local coefficients) if π and τ are tempered, by analytic extension if π v v v 172 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI and τ are quasi-tempered, and via the representation theoretic Langlands classifica- tion otherwise. The global theory of these twisted L-functions is now quite well understood. Theorem 3.1. — Let S be a non-empty set of finite places of k.Let η be a character of × × 2 k \A such that, for some v ∈ S,the square η is ramified. Then for all τ ∈ T (S; η) the L-function L(s,π × τ) is nice, that is, 1. L(s,π × τ) is an entire function of s, 2. L(s,π × τ) is bounded in vertical strips of finite width, and 3. we have the functional equation L(s,π × τ) = ε(s,π × τ)L(1 − s, π × τ). ˜ ˜ Proof. —(1) In thecase ofG = SO we explicitly established this in [6]. In all n 2n+1 cases this now follows from the more general Proposition 2.1 of [32]. Note that in view of the results of Mui´ c [44] and of [4], the necessary result on normalized intertwining operators, Assumption 1.1 of [32], usually referred to as Assumption A [28], is valid in all cases as proved in [28, 31]. Note that this is the only part of the theorem where the twisting by η is needed. (2) The boundedness in vertical strips of these L-functions is known in wide gen- erality, which includes the cases of interest to us. It follows from Corollary 4.5 of [12] and is valid for all τ ∈ T (N − 1), provided one removes neighborhoods of the finite number of possible poles of the L-function. (3) The functional equation is also known in wide generality and is a conse- quence of Theorem 7.7 of [51]. It is again valid for all τ ∈ T (N − 1). In order to mediate between the result as stated and the references for its proof, let us recall how these twisted L-functions are obtained from the theory of Eisenstein series. Given our classical group G and a general linear group GL with m ≥ 1 let n m G be the classical group of the same type as G ,but of rank m + n.Then ifwe let m+n n M = GL × G then M is a Levi subgroup of a standard maximal parabolic subgroup m n P = P ⊂ G .Let d = m + n and let N = N be the unipotent radical of P. m,n m+n m,n ×,1 × Let A denote the group of ideles of norm 1. Fix a subgroup A ⊂ A such × × ×,1 that A R and A = A × A . It suffices to assume that τ is unitary and its + + × × central character is a character of k \A which is trivial on A .Any cuspidal rep- resentation τ of GL (A) can be written as τ τ ⊗| det | ,where τ is unitary with central character trivial on A ,and then L(s,π × τ) = L(s + s ,π × τ ).Note that if τ ∈ T (S; η),then so is τ . Hence we may assume that τ is unitary. With π and τ as in the theorem, then σ = τ ⊗ π is a unitary globally generic representation of M(A). As such, we can form the induced representation G (A) I(s,σ) = Ind (| det | τ˜ ⊗ π). P(A) FUNCTORIALITY FOR THE CLASSICAL GROUPS 173 If α is the simple root associated to the maximal parabolic subgroup P and we let, as usual, α = ρ /
ρ ,α then as in [51] P P G (A) d
sα, ˜ H I(s,σ) = Ind (e σ). P(A) L L L Since the adjoint action r of Monthe Lie algebra n of N has two irreducible constituents in general, that is, r = r ⊕ r , then the L-functions which naturally arise 1 2 in the theory of intertwining operators and Eisenstein series for these representations will be a product L(s,σ, r )L(2s,σ, r ) where 1 2 L(s,σ, r ) = L(s,π × τ), the L-function of interest, and a second L-function, namely L(2s,σ, r ) = L(2s,τ, Sym ) if G = SO , 2 n 2n+1 and if m ≥ 2 and G = Sp or G = SO ,then n 2n n 2n L(2s,σ, r ) = L(2s,τ, ∧ ). In these later two cases, if m = 1,then r = r is irreducible. 4. Stability of γ -factors for G × GL n 1 This section is devoted to the formulation and proof of the stability of the local γ -factors for generic representations of the split classical groups. This result is neces- sary for defining a suitable local lift at the non-archimedean places where we do not have the local Langlands conjecture at our disposal. Following the ideas of [8] our method will be first to express the γ -factor as the Mellin transform of a certain par- tial Bessel function. This has been done in our cases as well as others in [55]. Then we will analyze the asymptotics of the Bessel functions as in [8] to obtain the stability. A more complete exposition and extensions to quasi-split groups will soon be available in [10]. For this section, let k denote a non-archimedean local field of characteristic zero. Let π be a generic irreducible admissible representation of G (k) and let η be a con- tinuous character of GL (k) k .Let ψ be a fixed non-trivial additive character of k. Let γ(s,π × η, ψ) be the associated γ -factor as defined in Theorem 3.5 of [51]. These are defined inductively through the local coefficients of the local induced representa- tions analogous to those given above. They are related to the local L-and ε-factors by −1 ε(s,π × η, ψ)L(1 − s, π ˜ × η ) γ(s,π × η, ψ) = . L(s,π × η) 174 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI The mainresultof thissection is thefollowing. Theorem 4.1. — Let π and π be two irreducible admissible generic representations of 1 2 G (k). Then for every sufficiently highly ramified character η of k we have γ(s,π × η, ψ) = γ(s,π × η, ψ). 1 2 For the case of G = SO this is [8]. n 2n+1 4.1. Preliminaries on Bessel functions. — Let us review the basic definitions from Section 3 of [8]. Note that, as their proofs show, the results in Section 3 of [8] are valid for any Chevalley group over k,not just SO . In this paper we specialize them 2n+1 to the split classical groups. Fix G = G and recall that B = B is the standard upper triangular Borel sub- n n group of G, T = T the standard maximal split torus of B, i.e., the diagonal matrices in G ,and U = U is its unipotent radical. Let Φ be the set of positive roots defin- n n ing U and let ∆ denote the associated simple roots. Let W be the Weyl group of G. Then W N(T)/Tand for each w ∈ W we choose a representative in N(T),which by abuseofnotation wewill continueto call w. To be specific, for what follows it will be necessary to choose the representatives as in Section 2 of [55] (see Section 4.2 be- low). For α ∈ Φ let U denote the one parameter root subgroup corresponding to α [57]. For any w ∈ W letusset − + −1 U = U and U = w Bw ∩ U w w α>0 wα<0 + − so that U = U U . w w Recall that we say that w ∈ W supports a Bessel function if for every α ∈ ∆ such that wα >0 we have that wα ∈ ∆ .Ifwelet w denote the long Weyl element of W then this is equivalent to w w being the long Weyl element of the Levi subgroup M 0 w of some standard parabolic subgroup P ⊃ B. In this case, U is the unipotent radical of P .Let A denote the center of M .Then w w w A ={t ∈ T | wα(t) = 1 for all α ∈ ∆ with wα >0}. Suppose that w ∈ W is such that w supports a Bessel function and the only w ∈ W with w ≤ w in the Bruhat order which support a Bessel function are w itself and the identity e. This is equivalent to P being a maximal parabolic subgroup. Let ∨ × α = α be the simple root associated to P . There is an injection α from k into w w ∨ × 0 ∨ × A such that α(α (t)) = t for all t ∈ k and, setting A = α (k ),we havethe w FUNCTORIALITY FOR THE CLASSICAL GROUPS 175 decomposition A = ZA ,where Z = Z is the (finite) center of G. (See, for example, w G the remarks after Assumption 5.1 in [55].) Now let π be an irreducible admissible generic representation of G(k).Let v ∈ V be such that the associated Whittaker function W ∈ W (π, ψ) satisfies π v W (e) = 1.Then if w ∈ W supports a Bessel function, and is a minimal non-trivial such with respect to the Bruhat order, we may formally define the associated Bessel 0 × function as the function on A k defined by −1 J (a) = W (awu)ψ (u) du. π,w v U (k) Since the arguments of Section 4 of [8] again only depended on G being a Chevalley group, then by the Corollary to Proposition 4.2 we know that J exists and is inde- π,w pendent of the choice of v ∈ V used to define it. This function is hard to work with. As a substitute, for every compact open subgroup Y ⊂ U (k) we define the partial Bessel function j (a) by π,w,v,Y −1 j (a) = W (awy)ψ ( y) dy. π,w,v,Y v In thecasewhere π and w are fixed, we will simply write j = j . π,w,v,Y v,Y 4.2. An integral representation for γ(s,π × η, ψ). — Our proof of the stability of the γ -factor is based upon expressing the γ -factor as the Mellin transform of one of our Bessel functions. Proposition 4.1. — Let π be a generic representation of G (k) and η a non-trivial character × 2 of k such that η is ramified. Then for each classical group G there exists a Weyl element w which supports a Bessel function and is minimal, non-trivial with this property, an elementary factor g(s,η), and a rational number δ such that for every sufficiently large open compact subset Y ⊂ U (k), setting j = j , we have v,Y π,w,v,Y −1 s−n+δ × γ(s,π × η, ψ) = g(s,η) j (a)η(a)|a| d a. v,Y The data for each classical group is as follows. (i) If G = SO , then the Weyl element is w = −I . The elementary n 2n+1 2n−1 n+1 factor is simply g(s,η) = η(−1) and δ = 1/2. 176 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI (ii) If G = SO , then the Weyl element is w = −K where n 2n 2n−2 . 01 K = . 2n−2 . n+1 2 −1 The elementary factor is g(s,η) = η(−1) γ(2s,η ,ψ) and δ = 1. −1 (iii) If G = Sp , then the Weyl element is w = −I . The elemen- n 2n 2n−2 n+1 2 −1 tary factor is again g(s,η) = η(−1) γ(2s,η ,ψ) and now δ = 0. −1 0 In all cases, a = diag(a, 1,..., 1, a ) ∈ A . This proposition is essentially Corollary 1.2 of [55]. To obtain it in this form, we must relate the Bessel functions of [55] to the ones we have defined here. While this is essentially an exercise, it will be useful to have it written down. 4.2.1. Corrections to [55]. — We begin with some minor corrections to [55]. In that paper the relevant Weyl elements w were miscalculated. This results in the following corrections (in the notation of that paper). (i) In the case GL × SO ⊂ SO the relevant Weyl element w given in 1 2n+1 2n+3 0 (4.19) is replaced by (−1) w = −I . 0 2n+1 (−1) This change only effects the elementary factor g(s,η) in a minor way. It will change formula (7.12) to −1 n+1 2 −1 C(s,η ⊗ σ) = η(−1) γ(2s,η ,ψ) × h 1 s−n+1/2 × × j I −I η(h)|h| d h N 2n−1 2n−1 v ˜, −1 h 1 n+1 with a similar change of η(−1) to η(−1) in formulas (1.5) and (1.6). FUNCTORIALITY FOR THE CLASSICAL GROUPS 177 (ii) In thecaseGL × SO ⊂ SO the relevant Weyl element w given in 1 2n 2n+2 0 (4.43) is replaced by (−1) w = −K . 0 2n (−1) The source of the error is an incorrect multiplication in (4.43). This change only ef- fects the elementary factor g(s,η) in a minor way. It will change formula (7.13) to −1 n+1 2 −1 C(s,η ⊗ σ) = η(−1) γ(2s,η ,ψ) × h 1 s−n+1 × × j I −K η(h)|h| d h 2n−2 2n−2 v ˜,N −1 h 1 n+1 with a similar change of η(−1) to η(−1) in formulas (1.5) and (1.7). (iii) In the case GL ×Sp ⊂ Sp the relevant Weyl element w given in (4.56) 1 2n 2n+2 0 is replaced by n+1 (−1) w = −I . 0 2n (−1) The source of the error is an incorrect multiplication in (4.56). This change only ef- fects the elementary factor g(s,η) in a minor way. It will change formula (7.26) to −1 n+1 2 −1 C(s,η ⊗ σ) = η(−1) γ(2s,η ,ψ) × h −1 s−n × × j I −I η(h)|h| d h 2n−2 2n−2 v,N ˜ 0 −1 h 1 n+1 with a similar change of η(−1) to η(−1) in formulas (1.5) and (1.7). 4.2.2. A comparison of Bessel functions. — For this section, let us use j to de- note any of the Bessel functions occurring in [55]. Our goal is to express the Bessel functions j (m ˙ ) occurring in Corollary 1.2 of [55] in terms of those we have defined v,N ˜ 0 here. (i) Let G = SO with m = 2n + 1 or 2n.In G consider the standard (upper n m n+1 triangular) maximal parabolic subgroup P = M N with Levi subgroup M = n+1 n+1 n+1 n+1 M GL × G . In our geometric model, this would be the stabilizer of the isotropic 1 n line through (0,..., 0, 1). The unipotent radical then takes the form 1t −
t, t ∗ m N (k) = n(t) = I −t t ∈ k n+1 m t ∗ where
t, t= tΦ t and t is the adjoint of t with respect to this form. m 178 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI Let π be our generic representation of G (k). Then in the expression for j (m ˙ ) n v,N ˜ 0 from Corollary 1.2 of [55] we have that m ˙ = I w = hw ∈ G ⊂ M m−2 n n+1 −1 where w is as in our integral representation and v ˜ ∈ W (π, ψ) with W (e) = 1.Here v ˜ N ⊂ N(k) is a (suitable) open compact subgroup of the opposite unipotent subgroup Nto N = N . In fact, the formulas hold for any such choice of v ˜ and sufficiently n+1 large N (see Theorem 6.2 of [55]). We nowturnto the Besselfunctionitself as giveninTheorem 6.2of [55].First × 2 ˜ ˜ j (m ˙ ) = j (m ˙ , y ) with y ∈ k satisfying ord ( y ) =−cond(ψ) − cond(η ).Then 0 0 k 0 v ˜,N v ˜,N 0 0 the Bessel function is given by (6.26) of [55], which we can write as j (m ˙ , y ) = v,N 0 ˜ 0 −1 ∨ −1 ∨ ∨ −1 ∨ −1 W (mu ˙ )ϕ(uα ( y ) α (x ˙ )n ˙α (x ˙ ) α ( y )u )ψ(u) du. v ˜ 0 α α 0 U \U M,n M ∨ × × Here α : k → Z \Z , x ∈ k a specified choice, and ϕ is the characteristic func- G M α tion of N . Throughout, n is a specific Z U -orbit representative in N and m and n ˙ ˙ ˙ 0 M −1 are related by w n = mn n ∈ M N N . ˙ ˙ ˙ ˙ n+1 n+1 n+1 Let us first consider the domain of integration. By Proposition 4.4 or Proposition 4.8 of [55] we have that −1 −1 U = U ={u ∈ U | mum ∈ U and ψ(mum ) = ψ(u)}. ˙ ˙ ˙ ˙ M,n M M M,m ˙ −1 −1 −1 In our situation, U = U ⊂ G and m ˙ = hw.Then mu ˙ m ˙ ∈ U iff wuw ∈ h U h = M n n n n + + + U ,thatis, u ∈ U = U .Since h acts trivially on U we see that U = U so n −1 M,n ˙ w w w + − that we can take U \U U \U to be U , which we note depends only upon w. M,n ˙ M n w w Next we turn to the effect of the cutoff characteristic function ϕ.Taking u ∈ U we see that the actual domain of integration is determined by the condition ∨ −1 ∨ ∨ −1 ∨ −1 uα ( y ) α (x )nα (x ) α ( y )u ∈ N . ˙ ˙ ˙ 0 α α 0 0 A priori, this condition depends on n ˙ which is related to m ˙ and hence h. In fact, as we shallsee, thisisnot the case. Firstnotethat thiscondition is equivalent to ∨ ∨ −1 −1 ∨ ∨ −1 uα (x ˙ )n ˙α (x ˙ ) u ∈ α ( y )N α ( y ) . α α 0 0 0 ∨ ∨ −1 But α ( y )N α ( y ) is another compact open subgroup of the same type, so we 0 0 0 may ignore this in our situation. As in (7.1) of [55] we write −1 w n(t) = mn n( y) where n( y) = yI ˙ ˙ ˙ ˙ ˙ ∗ ∗ − y y −y 1 2 FUNCTORIALITY FOR THE CLASSICAL GROUPS 179 with y ∈ k written as a column vector. Now according to section 7 of [55] in our situ- −1 ∗ −1 −1 ation we have t = (1, 0,..., 0, h) and x = h .Then y = 2
t, t t = (h , 0, ..., 0, 1) and ∨ ∨ −1 ∨ −1 ∨ α (x ˙ )n ˙( y)α (x ˙ ) = α (h )n ˙( y)α (h) = n ˙( y ) α α where y = (−h, 0, ..., 0, −1). So the condition on the cutoff of our domain of inte- gration is that t −1 un( (−h, 0, ..., 0, −1))u ∈ N . For certitude, let us take N to be defined as −M yI N = n( y) = y ∈ p 0 m i 1 ∗ ∗ − y y −y 1 for some sufficiently large integer vector M = (M , ..., M ).As M increases, these 1 m exhaust N. Now recall that u ∈ U and this means that we can write 1 ∗ 1x − x x ∗ m−2 u = u(x) = I −x with x ∈ k m−2 −1 which we view as embedded in M via u ∈ U U ⊂ M. Then in general un( y)u = n M m t t 1 ∗ n(uy) with uy ∈ k . In our situation y = (−h, 0, ..., 0, −1) and so u(x)y = ( x x − h, t ∗ − x , −1). Thus our domain of integration is over Y ⊂ U (k) defined by the conditions −M ∗ −M m−i 1 Y = u = u(x) | x ∈ p with h ≡ x x (mod p ) . N ,which we are To rid ourselves of the remaining dependence on h we enlarge allowed to do. By choosing M sufficiently large, which may depend on h and M , 1 2 ..., M , we obtain a domain of integration m−1 −M m−i Y = u = u(x) | x ∈ p , 1 ≤ i ≤ m − 2 which is now independent of h and with this choice of Y and M we have j (hw) = j (h) = j (h). π,v ˜,w,Y v ˜,Y v ˜,N (ii) In the symplectic case G = Sp we must use the Bessel function j (m) = n 2n v ˜,N ˜ ˜ j (m, y ) = j (mH, y ) as in (7.24) and (7.25) of [55]. Here H is the matrix ˙ ˙ 0 0 v ˜,N v ˜,N 10 h H = I 0 ∈ U U . 2n−2 M n 1 180 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI −1 Its effect in computing the Bessel function is to replace . But by (7.27) n ˙ by HnH ˙ of [55] this matrix is represented by −1 HnH = y I 1 2n Y y 1 t −1 with y = (−h , 0, ..., 0, 1). Comparing this with our formula for n ˙( y) above in the orthogonal case we see that the same analysis will go through. So in this case we also have j (hw) = j (h) for any sufficiently large open compact Y ⊂ U (k). v,Y v,N ˜ 0 4.2.3. Proof of Proposition 4.1. — With the identifications above, the fact that we have the integral representation of Proposition 4.1 is simply a restatement of Corol- lary 1.2 of [55]. To have the Proposition as stated, we must check that each Weyl element w that occurs both supports a Bessel function and is minimal non-trivial with this property. This is easy enough to check using the criterion in terms of parabolic subgroups from Proposition 3.2 of [8] mentioned above. Note that in the case of G = SO this integral representation is the same as n 2n+1 that of Proposition 4.1 of [8] which was derived from Soudry’s integral representation. 4.3. Asymptotics of Bessel functions. — In this section we investigate the asymp- totics of the Bessel functions j (a) defined above. v,Y We will follow the development presented for SO in [8]. The paper [8] was 2n+1 written for SO because that was the only case in which there was an integral rep- 2n+1 resentation for the γ -factor in terms of Bessel functions. This integral representation was presented in Section 2 of [8] and that section is specific to SO .Section 3 2n+1 and the first parts of Section 4 of [8] rely only on results about Chevalley groups from, say, Steinberg’s notes on Chevalley groups [57] and hence remain valid for any of our groups G . The remainder of Section 4 and Section 5 up through Proposi- tion 5.1 of [8] are more or less formal and rely only on standard facts about the Bruhat decomposition, the Bruhat order, and the fact that the Weyl element w oc- curring in j (a) has the property that w supports a Bessel function and that the only v,Y Weyl elements w with w ≤ w which support a Bessel function are w itself and the identity e. These facts remain true for our w as noted above, so the results of these sections of [8] remain valid in all our present cases. In particular, quoting Proposition 5.1 of [8] we have the following. Proposition 4.2. — There exists a vector v ∈ V and a compact neighborhood BK of π 1 the identity e in B\G such that if χ is the characteristic function of BK , we have that for all n 1 1 sufficiently large compact open Y ⊂ U (k) −1 j (a) = W (awy)χ (awy)ψ ( y) dy + W (a). v,Y v 1 v Y FUNCTORIALITY FOR THE CLASSICAL GROUPS 181 From this point on the situation is slightly different from that in [8] because in the cases G = SO and G = Sp the groups have non-trivial finite centers. Still n 2n n 2n following [8], for each simple root α let t → u (t) be the associated one parameter subgroup of U. For any positive integer M let U(M) =
u (t) | α ∈ ∆ ;|t|≤ q . This is a compact open subgroup of U(k) and as M grows these exhaust U. For any v ∈ V let us set −1 v = ψ (u)π(u)vdu. Vol(U(M)) U(M) Since (π, V ) is smooth this is actually a finite sum and so v ∈ V . π M π Then as noted in [8] as long as Y is sufficiently large relative to M we may choose v and K in Proposition 4.2 such that K ⊂ Stab(v ) and we have 1 1 M −1 −1 W (awy)χ (awy)ψ ( y) dy = W (awy)χ (awy)ψ ( y) dy. v 1 v 1 Y Y Consider this latter integral. If we write awy = utk with u ∈ U(k), t ∈ T(k), and k ∈ K ,sothat utk ∈ BK ,thensince K ⊂ Stab(v ) we have W (awy) = 1 1 1 1 1 M v ψ(u)W (t). As shown in Lemma 4.1 of [8] the support of W on the torus T is v v M M contained in the set T ={t ∈ T(k) | α(t) ∈ 1 +p for all simple α}. 1 M For M a positive integer, let us set T ={t ∈ T(k) | t ≡ I (mod p )}. Lemma 4.1. — For M sufficiently large, T ⊂ Z · T where Z is the center of G and M n M = M − ord(2). Proof. — Let us consider the case of G = SO . The others are handled accord- n 2n −1 ingly. With our basis, we can write an element t of the torus as t = diag(t , ..., t , t 1 n −1 ..., t ). The simple roots are then α (t) = t /t for i = 1, ..., n − 1 and α (t) = t t . i i i+1 n n−1 n M M If t ∈ T then α (t) ∈ 1 +p and α (t) ∈ 1 +p implies their ratio lies in there as M n−1 n 2 M well, that is, t ∈ 1 +p . 2 M 2 M In general, if t ∈ 1 +p then t is a unit satisfying t − 1 ≡ 0 (mod p ). Letting P(X) = X − 1 we have that ord(P (t)) = ord(2) and the roots of P(X) = 0 in O are ±1. Thus, say by Corollary 1 of Theorem 2 in ch.III, §4, no.4 of [3], we know t ≡±1 (mod p ) where M = M − ord(2). M M M Thus t ∈±1 +p .Then since α (t) = t /t ∈ 1 +p ⊂ 1 +p we have n n−1 n−1 n that t ∈±1 + p and that the sign of t and t must be the same. Continuing n−1 n n−1 with the rest of the roots in this manner, we find that ±t ∈ T and we are done since Z ={±1}. 182 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI 1 1 1 Hence if t ∈ T we can further write t = zt with z ∈ Zand t ∈ T .Itiseasy to check that for t ∈ T we have W (t) = W (t), so that if we choose M from the M v v 1 1 beginning so that T ⊂ T ∩ Stab(v) then we see that W (t) = W (t) = W (zt ) = v v v M M ω (z)W (t ) = ω (z). π v π So,in our integral,we see that W (awy)χ (awy) = 0 iff awy ∈ UT K or v 1 M 1 −1 1 y ∈ (aw) UT K . If we write this decomposition as awy = utk = u(awy)z(awy)t k , M 1 1 1 then we find −1 W (awy)χ (awy)ψ ( y) dy = v 1 −1 ψ(u(awy))ψ ( y)ω (z(awy)) dy. −1 Y∩(aw) UT K M 1 Then our previous proposition can now be written as follows. Proposition 4.3. — Fix v ∈ V such that W (e) = 1 and choose M sufficiently large so π v that T ⊂ T ∩ Stab(v). There exists a vector v ∈ V and a compact open subgroup K such π 1 M π that for Y ⊂ U (k) sufficiently large we have −1 (a) = ψ(u(awy))ψ ( y)ω (z(awy)) dy + W (a). v,Y π v −1 Y∩(aw) UT K M 1 This proposition gives us the asymptotics of j (a) in the following sense. The v,Y function W is a smooth Whittaker function and hence vanishes for a large and ex- hibits the standard asymptotics of the Whittaker function as a goes to zero. Thus the integral expression contains all asymptotics of the Bessel function as a gets large. Even though this integral is a complicated exponential sum, it only depends on π through its central character ω . 4.4. Stability of γ -functions depending on the central character. — As an immediate consequence of Proposition 4.3 we obtain the following stability result. Proposition 4.4. — Let π and π be two irreducible admissible generic representations of 1 2 G (k) having the same central character. Then for every sufficiently highly ramified character η of k we have γ(s,π × η, ψ) = γ(s,π × η, ψ). 1 2 Proof. —Let v ∈ V be chosen such that for each we have W (e) = 1. Choose i π v i i a large integer M such that T ⊂ T ∩ Stab(v ).Let K be a compact open subgroup i 0 of G such that K ⊂ Stab(v ) ∩ Stab(v ). Then in Proposition 4.3 we may take n 0 1 2 −1 K = u K u 1 0 u∈U(M) FUNCTORIALITY FOR THE CLASSICAL GROUPS 183 as in Section 6 of [8], that is, we can take K to be the same for π and π .Thenby 1 1 2 Proposition 4.3 there exist vectors v ∈ V such that −1 j (a) = ψ(u(awy))ψ ( y)ω (z(awy)) dy + W (a). v ,Y π v i i −1 Y∩(aw) UT K M 1 Since the central characters of π and π agree, we have 1 2 j (a) − j (a) = W (a) − W (a). v ,Y v ,Y v v 1 2 π π 1 2 If we now turn to Proposition 4.1 we find that as long as η is ramified we have −1 −1 γ(s,π × η, ψ) −γ(s,π × η, ψ) 1 2 s−n+δ × = g(s,η) ( j (a) − j (a))η(a)|a| d a v ,Y v ,Y 1 2 s−n+δ × = g(s,η) (W (a) − W (a))η(a)|a| d a. v v π π 1 2 But the Whittaker functions are smooth. So for Re(s) 0 and η sufficiently highly ramified we have s−n+δ × W (a)η(a)|a| d a ≡ 0. Thus for Re(s) 0 we have −1 −1 γ(s,π × η, ψ) − γ(s,π × η, ψ) ≡ 0 1 2 and then by the principle of analytic continuation this must be true for all s.Thus γ(s,π × η, ψ) = γ(s,π × η, ψ) 1 2 and we are done. 4.5. Computation of the stable forms. — To complete the proof of Theorem 4.1, as well as for application in the proof of Theorem 1.1, we will compute an explicit for- mula for the stable form of the γ -factor. In order to do this, let π be any irreducible admissible generic representation of G (k) with central character ω.Take µ ,...,µ to n 1 n be n characters of k .Then µ ⊗ ···⊗ µ defines a character of T (k) and we assume 1 n n that upon restriction to the center Z ⊂ T (k) this character agrees with the central n n G (k) character ω of π .Then ifwelet π = Ind (µ ⊗ ··· ⊗ µ ) then for an appropriate 1 2 1 n B (k) choice of the µ (in “general position”) this representation will be irreducible admis- sible generic and have central character ω. Thus for all sufficiently highly ramified η we have γ(s,π × η, ψ) = γ(s,π × η, ψ). 1 2 184 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI We can explicitly computethe righthand side ofthisformula. Byfirstusing themul- tiplicativity of the γ -factor [52] we obtain −1 γ(s,π × η, ψ) = γ(s,µ η, ψ)γ s,µ η, ψ 2 j j=1 if G = SO or G = SO ,while if G = Sp we obtain n 2n+1 n 2n n 2n −1 γ(s,π × η, ψ) = γ(s,η,ψ) γ(s,µ η, ψ)γ s,µ η, ψ . 2 j j=1 This computes the stable form of γ -factor in terms of abelian γ -factors. Proposition 4.5. — Let π be any irreducible admissible generic representation of G (k) with central character ω and let µ , ..., µ be any choice of characters of k in general position such that 1 n µ ⊗···⊗ µ agrees with ω upon restriction to the center. Then for every sufficiently highly ramified 1 n character η we have γ(s,π × η, ψ) −1 γ(s,µ η, ψ)γ s,µ η, ψ G = SO , SO j n 2n+1 2n j j=1 = . −1 γ(s,η,ψ) γ(s,µ η, ψ)γ s,µ η, ψ G = Sp j n 2n j=1 4.6. Proof of Theorem 4.1. — To complete the proof of Theorem 4.1 it will suffice to show that the stable form of the γ -factor computed in Proposition 4.5 is actually independent of the central character ω. There is an elementary reason for this (see the commentsatthe endofthissection), buta reason which is particularly adapted to our application is the following. First take G to be SO or SO . Then in either case the standard embedding n 2n+1 2n of the L-groups predicts a functoriality to GL with N = 2n. In either of these cases, let Π be the induced representation of GL (k) induced from these same characters, that is, GL (k) −1 −1 Π = Ind µ ⊗ ··· ⊗ µ ⊗ µ ⊗ ··· ⊗ µ . 1 n B (k) n 1 Then Π is a generic representation of GL (k) having trivial central character and by multiplicativity of the γ -factors for GL [20] we also have −1 γ(s, Π × η, ψ) = γ(s,µ η, ψ)γ s,µ η, ψ . j=1 FUNCTORIALITY FOR THE CLASSICAL GROUPS 185 Thus γ(s,π × η, ψ) = γ(s, Π × η, ψ). On the other hand, by the stability of γ -factors for GL [23] we know that the sta- ble form of the γ -factor on GL depends only on the central character. Since Π has trivial central character no matter the central character ω of π , we see that the stable form of the γ -factor for G is independent of the central character. This establishes Theorem 4.1 in these cases. The case of G = Sp is similar. Take π an irreducible admissible generic rep- n 2n resentation of Sp (k) and take characters µ , ...µ so that for sufficiently ramified η 2n 1 n we have −1 γ(s,π × η, ψ) = γ(s,η,ψ) γ(s,µ η, ψ)γ s,µ η, ψ . j=1 Now the functorial lift should be to GL ,so wetake Π to be the generic represen- 2n+1 tation of GL (k) with trivial central character given by 2n+1 GL (k) 2n+1 −1 −1 Π = Ind µ ⊗ ··· ⊗ µ ⊗ 1 ⊗ µ ⊗ ··· ⊗ µ . 1 n B (k) n 1 2n+1 Then multiplicativity of γ -factors for GL [20] gives −1 γ(s, Π × η, ψ) = γ(s,η,ψ) γ(s,µ η, ψ)γ s,µ η, ψ j=1 as well, so that γ(s,π × η, ψ) = γ(s, Π × η, ψ) for all sufficiently highly ramified η. But again the stable form of the γ factor for GL depends only on the central character of Π [23], which is trivial no matter what the central character of π . Thus the stable form of γ(s,π × η, ψ) is independent of the central character of π as well. This completes the proof of Theorem 4.1 in this case as well. We end this section with two corollaries of our stability results. The first is a corollary of Proposition 4.5 combined with Theorem 4.1 and the following obser- vations. In the notation of Proposition 4.5, for η sufficiently highly ramified, each µ η will also be highly ramified, so that L(s,µ η) ≡ 1.Then γ(s,µ η, ψ) = ε(s,µ η, ψ). j j j Similarly, by [54] as soon as η is sufficiently highly ramified we have L(s,π × η) ≡ 1, so that γ(s,π × η, ψ) = ε(s,π × η, ψ) as well. Thus we obtain the stability of local ε-factors as well as their stable form. 186 J. W. COGDELL, H. H. KIM, I. I. PIATETSKI-SHAPIRO, F. SHAHIDI Corollary 4.1. — Let π be an irreducible admissible generic representation of G (k) and let µ , ..., µ be characters of k in general position. Then for every sufficiently ramified character η 1 n we have ε(s,π × η, ψ) −1 ε(s,µ η, ψ)ε s,µ η, ψ G = SO , SO j n 2n+1 2n j j=1 = . −1 ε(s,η,ψ) ε(s,µ η, ψ)ε s,µ η, ψ G = Sp j n 2n j=1 Our second corollary combines the proof of Theorem 4.1 with the above obser- vations on the stability of the local L-factors, both for G and GL . n N Corollary 4.2. — Let π be an irreducible admissible generic representation of G (k).Let Π be any irreducible admissible representation of GL (k) with trivial central character ( N as in Theorem 1.1). Then for all sufficiently ramified characters η of k we have L(s,π × η) ≡ 1 ≡ L(s, Π × η) and ε(s,π × η, ψ) = ε(s, Π × η, ψ). As was pointed out by the referee, the formulas in Proposition 4.5 and Corol- lary 4.1 can be simplified as follows. As we noted above, for highly ramified characters, thereisno differenceinthe γ -factors and the ε-factors. The ε-factors for characters of k can then be computed via Gauss sums. As long as η is sufficiently highly ramified with respect to µ we have that there exists c such that ε(s,µη, ψ) = µ(c )ε(s,η,ψ). η η Thus under these conditions we have −1 2 ε(s,µη, ψ)ε(s,µ η, ψ) = ε(s,η,ψ) which then leads to γ(s,π × η, ψ) = γ(s,η,ψ) in Proposition 4.5 in all cases and ε(s,π × η, ψ) = ε(s,η,ψ) in Corollary 4.1 in all cases, where the natural functoriality is from G to GL .In the n N case of Proposition 4.5 this formula provides the elementary proof of the independence of the stable form from the central character of π alluded to above. We chose to leave our original proof since it then leads naturally to Corollary 4.2. These formulas for the stable form can then be obtained after the fact by using stability and then taking π to be induced from trivial characters (again as was pointed out by the referee). FUNCTORIALITY FOR THE CLASSICAL GROUPS 187 5. The candidate lift We now return to k denoting a number field. Let π =⊗ π be a globally generic cuspidal representation of G (A). In this section we will construct our candidate Π = ⊗ Π for the functorial lift of π as an irreducible admissible representation of GL (A). v N We will construct Π by constructing each local component, or local lift, Π .There will be three cases: (i) the archimedean lift, (ii) the non-archimedean unramified lift, and finally (iii) the non-archimedean ramified lift. 5.1. The archimedean lift. —Let v be an archimedean place of k. By the arith- metic Langlands classification [34, 2], π is parameterized by an admissible homo- L 0 morphism φ : W → G where W is the Weil group of k . By composing with v k k v v n v ι : G (C) → GL (C) we have an admissible homomorphism Φ = ι ◦ φ : W −→ n N v v k GL (C) and this defines an irreducible admissible representation Π of GL (k ). N v N v L L G GL n N XX EE π φ