Vol. 45. No.

DUKE MATHEMATICAL JOURNAL@

December 1978

A TRACE FORMULA FOR REDUCTIVE GROUPS I TERMS ASSOCIATED TO CLASSES IN G ( Q )

JAMES G. ARTHUR Preliminaries . . . . . . . . . . . The kernel Kp(x, y) . . . . . . . . A review of Eisenstein series . . . The second formula for the kernel The modified kernel identity . . . Some geometric lemmas . . . . . Integrability of KE(x, f ) . . . . . . Weighted orbital integrals . . . .

.................. .................. .................. ..................

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .......... .......... .......... ..........

9 15 919 924 928 935 938 942 947

Introduction This paper, along with a subsequent one, is an attempt to generalize the Selberg trace formula to an arbitrary reductive group G defined over the rational numbers. Our main results have been announced in the lectures [l(c)]. In his original papers [8(a), (b)], Selberg gave a novel formula for the trace of a certain operator associated with a compact quotient of a semisimple Lie group and a discrete subgroup. When the discrete subgroup is arithmetic, the situation is essentially equivalent to the case that the group G is anisotropic. Then G(A) is a locally compact topological group and G(Q) is a discrete subgroup such that G(Q)\G(A), the space of cosets (with the quotient topology), is compact. The operator is convolution on L2(G(Q)\G(A)) by a smooth, compactly supported function f on G(A). Let us recall how to derive the formula in this case. To understand the idea, it is not necessary to be an expert in algebraic groups, or even to be familiar with the notion of adeles. If (A E L W Q ) \ G ( A ) , define
Then R is a unitary representation of G(A), and the convolution operator is defined by,

(R(f)+)(x) can be written

Received April 7, 1978. Revision received June 28, 1978. Partially supported by NSF grant MCS77-0918

JAMES G. ARTHUR

by changing the variable of integration twice. Thus, R( f ) is an integral operator. Its kernel equals

where 0 is the set of conjugacy classes in the group G(Q), and

On the other hand, there is an even more formal way to write the kernel. It is not hard to show that R decomposes into a direct sum of irreducible representations of G ( A ) ,each occurring with finite multiplicity. In other words

where 3C is a set of unitary equivalence classes of irreducible representations of G ( A ) , and the restriction of the representation R to the subspace L2(G(Q)\G(A);, is equivalent to a finite number of copies of x. For each x 2, let % be a suitable orthonormal basis of L2(G(Q)\G(A));,. Then

is a second formula for the kernel of R ( f ) . The Selberg trace formula comes from integrating both formulas for the kernel over the diagonal. (The necessary convergence arguments are easily established.) Thus

where J . , ( f )is the integral over x in G(Q)\G(A) and J x ( f )is the integral of

of

A TRACE FORMULA FOR REDUCTIVE GROUPS I

913

kx(x, f ) = Kx(x, 4. If {y} stands for a set of representatives of the conjugacy classes in G(Q), and G(y) denotes the centralizer of y in G, the formula can be written

Thus, by simple formal manipulations, the trace of the operator R(f) has been expressed as a sum of integrals off over conjugacy classes. If G is not anisotropic, G(Q)\G(A) is no longer compact and everything breaks down. Selberg suggested a program for obtaining a formula and studied some special cases. Further progress has since been made by others, but the program has been carried out completely for only a small number of groups, essentially just GLz and related groups. What are the difficulties in the noncompact setting? In general R(f) is still an integral operator. However, the natural definition of 0 seems to be the equivalence classes composed of those elements in G(Q) whose semisimple components are G(Q)-conjugate. If G is anisotropic, all the elements in G(Q) are semisimple so this relation is then just conjugacy. With KAx, y) defined as above, the kernel of R(f) can again be written

One of the main complications in the noncompact setting is that R no longer decomposes discretely. R contains a continuous family of representations for every parabolic subgroup of G defined over Q. The intertwining operators are provided by Eisenstein series, whose main properties we recall in $3. We now have to define the set 3 If G = GLn and investigations of Jacquet, Shalika and Piatetskii-Shapiro proceed as expected, 3 could be defined as the classes of irreducible unitary representations of G(A) relative to a certain relation, weaker than unitary equivalence. Otherwise, 3 must be defined as in $3, in terms of cuspidal automorphic representations of Levi components of parabolic subgroups of G. At any rate we obtain an identity

in $4 by equating two different formulas for the kernel of R(f). We can now set x = y and ask whether the various functions on each side are integrable. It turns out that K(x, x) is integrable precisely when o intersects no group P(Q), with P a proper parabolic subgroup defined over Q. The more parabolic subgroups that meet o, the worse will be the divergence of the integral. The integral of Kx(x,x) behaves in the same sort of way. We are therefore forced in $5 to modify the identity of adding suitable correction terms to each side. The correction terms are indexed by the conjugacy classes of proper parabolic subgroups, P, of G, and they also depend on a point Tin a positive Weyl chamber.

914

JAMES G . ARTHUR

The construction of the correction terms proceeds as follows: one multiplies the identity (the analogue for P of (2)), by the characteristic function of a subset of P(Q)\G(A) associated to T . The product is then summed over P(Q)\G(Q) to make the functions left G(Q)-invariant. We denote the corrected functions by kzx, f ) and k;(x, f ) and obtain a new identity

When T is large the correction terms all vanish on a large compact subset of G(Q)\G(A). On this set, kr(x,f ) equals Ko(x, x). Moreover, if o does not intersect any proper parabolic subgroup, kxx, f ) equals Ko(x,x) everywhere. Similar remarks apply to the functions on the right. Our main goal will be to show that each side of the new identity (3) is integrable and that the integrals may be taken inside the sums. If Wf) and J^(f) stand for the integrals of the summands, we will then have a formula

which generalizes the Selberg trace formula for compact quotient. In this paper we will deal with the left hand side of (3). We prove the integrability in $7. The argument is partly geometric and partly combinatorial. It is a standard technique in classical reduction theory to partition the upper half plane modulo SLfl) into a compact region and a noncompact region whose topological properties are particularly simple. Imitating a construction in Langlands' manuscript on Eisenstein series, we introduce a similar partition of G(Q)\G(A). This partition and other geometric facts are discussed in $6. We begin $7 with some manipulations that yield different formulas for the functions k ( x , f). This amounts to studying the components of k(x, f ) on various subsets of the domain of integration. It turns out that all the components which are not integrable on a given subset cancel. What remains can be estimated by the Poisson summation formula as in GLv [5, XVI]and is at length shown to be integrable. The classes o which do not meet any proper parabolic subgroup contain only semisimple elements, and are actual conjugacy classes. Thus J^f ) is independent of T and can be expressed as an orbital integral as on the left hand side of (1). In $8 we study the distributions J^ for a wider collection of o, the classes we will term unramified. We will show that Wf) can still be expressed as an orbital integral of f. However, it will be the orbital integral with respect to a measure which is not invariant, but weighted by a function obtained by taking the volume of a certain convex hull. It has been shown [I(^)] that this particular weighted orbital integral can be evaluated explicitly for certain special functions f.

A TRACE FORMULA FOR REDUCTIVE GROUPS I

915

The main use of the trace formula would be to obtain information about the right hand side from knowledge of the left hand side. One hopes that it will someday shed light on Langlands' functoriality conjecture, perhaps the fundamental problem in the whole area. Three special cases of this conjecture have already been solved [5], [ 6 ( c ) ] , by using the trace formula for GL2. In each [7], case, a trace formula for one group G was compared with the trace formula for G' = GL2. An arbitrary funciton f on G ( A )was used to define a function f ' on G 1 ( A )so as to make the left hand sides of the two formulas equal. The correspondence between automorphic representations was then extracted from the resulting equality of right hand sides. The formula (4), whose proof we will complete in the next paper, is not as explicit as the trace formula for GLz [5, pg. 516-5171. For example, if o is the class consisting of unipotent elements, we cannot express J^f ) as a weighted orbital integral. Nevertheless, I believe that (4) can eventually yield the same kind of results for general G . One would characterize the distributions J7'. and Jl by the properties required by the applications rather than by their explicit formulas.

5 1. Preliminaries
Let G be a reductive algebraic group defined over Q. We shall fix, for once and for all, a minimal parabolic subgroup Po, and a Levi component, M p , of P(,, both defined over Q. In this paper we shall work only with standard parabolic subgroups of G ; that is, parabolic subgroups P , defined over Q, which contain PO.Let us agree to refer to such groups simply as "parabolic subgroups." Fix P. Let N p be the unipotent radical of P and let M p be the unique Levi component of P which contains M p . Denote the split component of the center of M p by Ap. N p , M p and A p are all defined over Q. Let X(Mp)Qbe the group of characters of M p defined over Q. Then ap = Horn ( ~ ( M P R) ~ , ) is a real vector space whose dimension equals that of An. Its dual space is a'p = X(Mp)Q(8R. We shall denote the set of simple roots of ( P , A ) by Ap. They are elements in X(Ap)Qand are canonically embedded in a$. The set A. = A p is a base for a root system. In particular, we have the co-root a^ in a p for every root a Ap. Suppose that Pi and P2 are parabolic subgroups with PI C P2. a $ comes with an embedding into (A, while ap2is a natural quotient vector space of dp,. The group M p D P I is a parabolic subgroup of M p with unipotent radical

N;:

N ~ n, M ~ ; .

The set, A&, of simple roots of (Mp2n P I , A p l ) is a subset of Apl. As is well known, P2 -* A;: is a bijection between the set of parabolic subgroups Ps which contain Pi, and the collection of subsets of Apt.Identify a p with the subspace

{ H ap, : a ( H ) = 0 , a Afe}.

916 If

JAMES G . ARTHUR

c& is defined to be the subspace of a P annihilated by a h , then

The subspace, (a%)*,of a$, spanned by A$: is in natural duality with afe and we have a, :
=

(a$:)* @ a:,.

The space afe also embeds into a^, and Any root a A& is the restriction to afe of a unique root f3 Afe. Define c^ to be the projection onto a& of the vector fS^ in afc. Then is a basis of afe. Let
A& = {ma : a E A&} be the corresponding dual basis for (a&)*. If m A ; ! , aA& a(mV)=m(&),

define m afe by V

Then { ^ and {a} is another pair of dual bases for afe and (afe)*. v} If Pi is a parabolic subgroup, and Pi appears in our notation as a subscript or a superscript, we shall often use only i, instead of Pi, for the subscript or superscript. For example, If the letter P alone is used, we shall often omit it altogether as a subscript. Thus, P = NM, A is the set of simple roots of (P, A), and so on. G itself is a parabolic subgroup. We shall often write Z for Ac, and g for do. The following proposition is trivial to prove. However, it will be the ultimate justification for much of this paper so we had best draw attention to it. PROPOSITION Suppose that Pi and Pz are parabolic subgroups, with 1.1.

PiC Pi. Then

{P : P,

Yc c
P

(- l)(dim(A/A,) =
P,)

if Pi = Pa otherwise.

(The sum is, of course, over parabolic subgroups P ; A = Ap is uniquely determined by P.) Proof. The sum can be regarded as the sum over all subsets of A?. The result follows from the binomial theorem. Let A (resp. Ac) be the ring of adeles (resp. finite adeles) of Q. Then

A TRACE FORMULA FOR REDUCTIVE GROUPS I

917

is the restricted direct product over all valuations v, of the groups G(Qv).Fix, for once and for all, a maximal compact subgroup

of G(A) so that the following properties hold: (i) For any embedding of G into GLn, defined over Q , K,, = GLJo,,) n G(Qv) for almost all finite places v. (ii) For every finite v, Kv is a special maximal compact subgroup. (iii) The Lie algebras of Ky and Ao(R) are orthogonal with respect to the Killing form. Suppose that P is a parabolic subgroup. If m = V\mv lies in M(A), define a v vector H d m ) in ap by

Hu is a homomorphism of M(A) to the additive group up. Let M(A)l be its kernal. Then M(A) is the direct product of M(A)l and A(R)O, the component of 1 in A(R). By our conditions on K, G(A) = P(A)K. Any x E G(A) can be written as nmak, n E MA), m E M(A)l, a A(R)O, k K. We define Hp(x) = H(x) to be the vector Hu(ma) = Hu(a) in an. Notice that if PI C P2, a& is the image of Mp2(A)lunder Hpl. We shall denote the restricted Weyl group of (G, An)by fl. f l acts on a. and a%in the usual way. For every s fl we shall fix a representative w s in the . intersection of G(Q) with the normalizer of An. w is determined modulo Mo(Q). If Pi and P2 are parabolic subgroups, let fl(al, a2) denote the set of distinct isomorphisms from a, onto a, obtained by restricting elements in H to dl. Pi and P2 are said to be associated if fl(ai, a,) is not empty. If s1 belongs to fl(al, a2),there is a unique element s in f l whose restriction to al is sl and such that s l a is a positive root (that is to say, a root of (Po,An))for every a E A$i. Thus, H(al, az)can be regarded as a subset of fl; in particular, ws. an element is in G(Q) for every sl E fl(al, a2). We shall need to adopt some conventions for choices of Haar measures. For any connected subgroup, V, of No, defined over Q , we take the Haar measure on V(A) which assigns V(Q)\V(A) the volume one. Similarly, we take the Haar measure of K to be one. Fix Haar measures on each of the vector spaces an. On the spaces a$ we take the dual Haar measures. We then utilize the isomorphisms to define Haar measures on the groups Ap(R)O. Finally, fix a Haar measure on G(A). For any P , let

918 and

JAMES G . ARTHUR

There is a vector pp in a$ such the modular function ^ Here rip stands for the Lie algebra of Np. There are on P(A) equals e. unique Haar measures on M(A) and M(A)l such that for any function h E Cp(G(A)),

h(nmak) e - 2 P ~ ( H ~ ( a ) ) dndadmdk.

We should recall some properties of height functions associated to rational representations of G. Let V be a vector space defined over Q . Suppose that {v,, . ., vn}is a basis of V(Q).If & V(Qt;), and
+

define if v is finite, and

if v

R. An element

f =

fl ti,in
u

V(A) is said to be primitive if ll&lli'

for almost all v, in which case we set

I is called the height function associated to the basis {vl, that

. ., vn}.Suppose

is a homomorphism defined over Q . Let K A be the group of elements k E K such that llA(k)vII = Hull for any primitive v E V(A).It is possible to choose the basis { v l , ., vn}such that K A is of finite index in K, and also so that for each a Ac, the operator A(a) is diagonal. We shall always assume that for a given A, the basis has been chosen to satisfy these two conditions. From our basis on V(Q)we obtain a basis for the vector space of endomorphisms of V(Q).Every

A TRACE FORMULA FOR REDUCTIVE GROUPS I

919

element in G(A) is primitive with respect to the corresponding height function and for every primitive v E V(A), and every x G(A), If t > 0, define
Gt = {x G(A) : IIA(x)ll :s t}.

Suppose that A has the further property that G, is compact for every t. It is known that there are constants C and N such that for any t, the volume of Gt (with respect to our Haar measure) is bounded by CtN.For the rest of this paper we shall simply assume that some A, satisfying this additional property, has been fixed, and we shall write llxll for IIA(x)ll. This "norm" function on G(A) satisfies llxll 2 1,
\\k,xk211= llxll, llxyll :s llxll-llyll,

and for constants C and N, elements x, y G(A), and ki, k2belonging to a subgroup of finite index in K. Once 1 1 1 1 has been fixed, we shall want to consider different rational representations A of G. In particular, suppose that the highest weight of A is A, for some element A in a:. Then there are constants cl and cg such that for all x G(A). By varying the linear functional A, we can then show that for any Euclidean norm 1 1 - 1 1 on we can choose a constant c so that
\\Ho(x)\\:s c(l

+ log llxll),

x G(A).

Suppose that P is a parabolic subgroup. Recall that there is a finite number of disjoint open subsets of a , called the chambers of a . Their union is the complement in a of the set of hyperplanes which are orthogonal to the roots of (P, A). a+ is one chamber. According to Lemma 2.13 of [6(b)] set of chamthe bers is precisely the collection, indexed by all P' and s fl(a, a'), of open subsets s l ( a ' ) + .We shall write n(A) for the number of chambers. More generally, if Pi C P , let np(Al)be the number of connected components in the orthogonal complement in al of the set of hyperplanes which are orthogonal to the roots of (Pi fl M, Al).
$2. The kernel Kp(x, y)

Let R be the regular representation of G(A)l on L2(G(Q)\G(A)'). The map which sends f to the operator

JAMES G . ARTHUR

gives us a representation of any reasonable convolution algebra of functions f on G(A)l. For example, we could take Cc(G(A)l),which is defined as the topological direct limit over all compact subsets T of G(A)l, of the spaces of continuous functions on G(A)l supported on T. We shall be more interested in the smooth functions on G(A)l. For any place v, let G(QV)ldenote the intersection of G(Qwith G(A)l. Then G(R)l contains the connected component of 1 in G(A)l. Notice that for any v, Kvis contained in G(Qv)l. Suppose that KOis an open compact subgroup of G(AfY. Then the double coset space Ko\G(A)l/K0 is a discrete union of countably many copies of G(R)l. In particular it is a differentiable manifold. Suppose that T is a compact subset of G(A)l such that Let Cm(G(A)l, KO) the algebra of smooth functions on Ko\T/Ko which are F, be supported on T. %(g(R)l @ C), the universal enveloping algebra of the complexification of the Lie algebra of G(R)l, acts on this space on both the left and the right. We shall denote these actions by the convolution symbol. The seminorms

indexed by elements X, Y %(g(R)l@ define a topology on Cm(G(A)l, KO). C), F, Let C%G(A)l) be the topological direct limit over all pairs (T, KO),of the spaces Cm(G(A)l, KO).If r is any positive integer we can define C m ) the F, same way, except that we of course take only those seminorms for which the sum of the degrees of X and Y is no greater than r. Finally, any subgroup L of G(A)l acts on C;(G(A)l) by We shall write CXG(A)l)L for that matter, we will write XL,if X is any set on (or which L acts) for the set of L-invariant elements. Suppose that f E C':(G(A)l). R(f) is an integral operator on G(Q)\G(A)l with kernel

If we had the Selberg trace formula for compact quotient in mind, we would be inclined to decompose the formula for K(x, x) into terms corresponding to conjugacy classes in G(Q). It turns out, however, that an equivalence relation in G(Q), weaker than conjugacy, is more appropriate to the non-compact setting. Any y G(Q) can be uniquely written as ysyu,where y, is semi-simple, yu is unipotent, and the two elements commute. We shall say that elements y and y'

A TRACE FORMULA FOR REDUCTIVE GROUPS I

92 1

in G(Q) are equivalent if ys and y; are G(Q)-conjugate. Let 0 be the set of equivalence classes. Every class in 0 contains one and only one conjugacy class of semi-simple elements in G(Q). If y is in G(Q), and H is a subgroup of G, defined over Q, let H(y) be the centralizer of y in H. Both H(y) and its identity component, H(y)O, are defined over Q. If R is any ring containing Q, let H(R, y) be the centralizer of y in H(R). Now suppose that o is a class in 0. is clearly possible to choose a parabolic It subgroup P and a semisimple element y in o such that y belongs to M(Q), but such that no M(Q)-conjugate of y lies in Pl(Q), for Pl a parabolic subgroup of G, Pl 5 P. In other words, y lies is no proper parabolic subgroup of M, defined over Q. Elements in M(Q) with this property are called elliptic elements. M(y)O is a reductive group, defined over Q, which is anisotropic modulo A. The group P i s not uniquely determined by o. However if (MI, y') is another pair, associated to o as above, then y' = w y w l for some element w G(Q). Since A' is the split component of the center of G(yl)O,it equals wAwP1. Therefore w = wsq, for s n ( a , a'), and 7 M(Q). It follows that there is a bijection from 0 onto the set of equivalence classes of pairs (M, c), where P = NM is a parabolic subgroup and c is a conjugacy class in M(Q) of elliptic elements, two pairs (M, c) and (MI, cl) being defined to be equivalent if c' = wscw;l for some s in O(a, a'). Suppose that o is a class in 0and that P and y are as above. Let S denote the set of roots of (P, A). S determines a decomposition

of the Lie algebra of N. Let S(y) be the set of roots a in S such that the centralizer of y in na is not zero. The elements in S(y) are, of course, characters on A. Let A' be the intersection of their kernels. We can choose parabolic subgroups Pl C P2, and an element s E fl(a, dl), such that A2 = wSA'w;l. Set y, = wsywsl. It is an elliptic element in Mi@). Az is the split component of the center of G(y1)O,and P1(yl)Ois a minimal parabolic subgroup of G(y1)O. Notice that Pl and P, are equal if and only if every element in the class o is semisimple. In general, any element in o is conjugate to ylvl, where v l is a unipotent element in P1(y1)O.vl must lie in the unipotent radical of P1(yl)O, so in particular it belongs to Nl(Q).
2.1. Suppose that P = NM is a parabolic subgroup defined over LEMMA Q. Suppose that p is in M(Q). Then if (A Cc(N(A)),

Proof. Neither side of the putative formula changes if p is replaced by an M(Q)-conjugate of itself. After noting that the previous discussion can be ap-

922

JAMES G . ARTHUR

plied to classes in M(Q) as well as G ( Q ) , we can assume that there is a parabolic subgroup P i , with Pi C P , such that y, E M i @ ) , and yu M(Q, 7 , ) H N l ( Q ) . Now the Lie algebra of N can be decomposed into eigenspaces under the action of A l . It follows that there exists a sequence of normal yS-stable subgroups of N which are defined over Q and satisfy the following properties: is abelian for each k. (i) Nk + l\Nk (ii) If 8 N k , and T ) either belongs to N or equals y,,, T ) - ~ ~ " ' T ) belongs to S Nk + 1 . We shall show that for all k , 0 5 k 5 r ,

equals (2.2)
8

N(Q,y,)\N(Q)

y
T)

^(y-'8-^8).

6 N(Q.y.)

The assertion of the lemma is the case that k = 0. The equality is immediate if k = r. B y decreasing induction on k , we assume that (2.2) equals

82 NI.I.Q.Y.W~,(Q)\Nk(Q) +

1( ~ - ~ 8 ; ~ 8 ; ~ ~ ~ ) 8 ~ 8 1 ) . 4
7

For fixed

E Nk(Q), we change variables in the sum over

T).

We find that

where

A TRACE FORMULA FOR REDUCTIVE GROUPS I

is a compactly supported function on the discrete set

Nk(Q,
YS)

Nk

+ i(Q)\Nk(Q).
Y

The map y
-^

Nk(')'s)Nk + 1 y'sly^ysy,

Nk(~s)Nk +

is an isomorphism from Nk(y,,)Nk lN + \k Therefore

onto itself which is defined over Q.

It follows from the lemma that if o E 0, and if y E o l"l M(Q), then yy belongs to o for each y in N(Q). In other words A similar remark holds for the intersection of o with any parabolic subgroup of M.

LEMMA 2.2 Under the hypotheses of Lemma 2.1,

Proof. The proof can be transcribed from the proof of Lemma 2.1 by replacing each sum over a set of rational points by the integral over the corresponding set of A-valued points. 0 If o E 0, define

Then K(x, Y )
=

S KO(x,Y ) .
0

More generally, if P is a parabolic subgroup, define

Then

JAMES G . ARTHUR

equals

This is just the kernel of Rp(f), where Rp is the regular representation of G(A)l on L2(N(A)M(Q)\G(A)l).

53. A review of Eisenstein series In this section we shall recall those results on Eisenstein series which are needed for the trace formula. They are due to Langlands; the main ideas are in the article [6(a)] while the details appear in [6(b)]. Suppose that P is a parabolic subgroup. Recall that the space of cusp forms, LEusp (M(Q)\M(A)I), on M(A)I is the space of functions 4 in L2(M(Q)\M(A)l) such that for any parabolic subgroup Pi, with Pl S P ,

where p ranges over all irreducible unitary representations of M(A)l, and each Vp is an M(A)I-invariant subspace of L& (M(Q)\M(A)'), isomorphic under the action of M(A)l to a finite number of copies of p. An irreducible unitary representation p of M(A)l is said to be cuspidal if V p 0. Suppose that P' is another parabolic subgroup, and that pr is an irreducible unitary cuspidal representation of Mt(A)l. We shall say that the pairs (M, p ) and (MI, p') are equivalent if there is an s E Sl(a, a t ) such that the representation

is unitarily equivalent to p t . Let ^? be the set of equivalence classes of pairs. Then corresponding to any x E ^?we have a class, Px,of associated parabolic subgroups. If P is any parabolic subgroup and x E ^?,set
LEusp (M(Q)\M(AI1)x =

{ P : ( M , P ) X)

(D

VP.

This is a closed M(A)l-invariant subspace of LLso (M(Q)\M(A))I), which is empty if P does not belong to Px.We have

A TRACE FORMULA FOR REDUCTIVE GROUPS I

925

Again suppose that P is fixed and that x SK. Suppose first of all that there is a group Pi in % which is contained in P. Let \li be a smooth function on Nl(A)Ml(Q)\G(A) such that vanishes for a outside a compact subset of Al(Q)\Al(A), transforms under K R according to an irreducible representation W, and as a function of m , belongs to L&(M1(Q)\Ml(A)l). One of the first results in the theory of Eisenstein series is that the function

is square integrable on M(Q)\M(A)l, ([6(a), Lemma I]). We define LYM(Q)\M(A)l),/ to be the closed span of all functions of the form @, where PI runs through those groups in Sftv which are contained in P , and W is allowed to vary over all irreducible representations of K R .If there does not exist a group Pi g?,/which is contained in P , define L2(M(Q)\M(A)l),/ to be {O}. It follows from [6(a),Lemma 21, that L2(M(Q)\M(A)l) is the orthogonal direct sum over all x E of the spaces L2(M(Q)\M(A)l),/. For any P , let II(M) denote the set of equivalence classes of irreducible unitary representations of M(A). If t, a; and 'n II(M), let 'n{,be the product of ' with the quasicharacter n If t, belongs to /a*, IT{ is unitary, so we obtain a free action of the group ia* on II(M). II(M) becomes a differentiable manifold whose connected components are the orbits of ia*. We can also transfer our Haar measure on ia* to each of the orbits in II(M);this allows us to define a measure d'n on II(M). If Pz 3 P, let I I p 2 ( M ) be the space of orbits of iai on II(M). IPqM) inherits a measure from our measures on II(M) and i a i . For 'n II(M), let Xj? be the space of smooth functions which satisfy the following conditions: (i) 4 is right K-finite. (ii) For every x E G ( A ) ,the function is a matrix coefficient of 'n.

Let Xp('n)be the completion of Y@('n). It is a Hilbert space. If 4 t, a;, define

3Vp('n),

and

926

JAMES G . ARTHUR

and

Then Zp(7rr) is a representation of G(A) which is unitary if ( E ia*. Notice that Xp(7r) = {O} unless there is a subrepresentation of the regular representation of M(A)lon L^M(Q)\M(A)l) which is equivalent to the restriction of 7r to M(A)l. Given x 3let Xp(7r)),be the closed subspace of Xp(7r)consisting of those 4 such that for all x the function belongs to LWQ^M(A)l)),. Then

(I do not know whether Xp(7r)xcan be nonzero for more than one x.) Suppose that KOis an open compact subgroup of G(Ac)and that W is an equivalence class of irreducible representations of Ka. Let %p(7r)xxo the subspace of functions be in Xp(,TT\which are invariant under KO n K , and let SVpWxXojw the space of be functions in Xp(7r)xa which transform under KR according to W . It is a consequence of the decomposition of the spaces L2(M(Q)\M(A)l)),, established by Langlands in $7 of [6(b)],that each of the spaces %p(7r)x80,w finite dimensionis al. We shall need to have ortho-normal bases of the spaces SVpWx. Let us fix such a basis, SQp(7r)),, each 7r and x such that for We and such that every 4 9 p W Xbelongs to one of the spaces Xp(7r)xxo,w. shall need these bases in the next section, when we give a second formula for the kernal K(x, y) in terms of Eisenstein series. Suppose that 7r II(M), E Pp(7r), t, E a;,= and s n ( a , a'). If Re i, belongs to pp + (a$)+,the Eisenstein series and global intertwining operators are defined by
1

E(x, 4 ~= ) and (Mb, ~r)<br)(x) =

V<i)i(8x)
6 P(Q)\G(Q)

. 8~(8xy,
1

(b^wynx) - 8p(w2nx)2
1

N'(A) r l w,N(A)w;'\N'(A)

aP(x) dn. The properties that we will need are all contained, at least implicitly, in [6(b)] (see especially Appendix 11), and have been summarized in [l(c)]. For the con-

A TRACE FORMULA FOR REDUCTIVE GROUPS I

927

venience of the reader, we shall recall them again here. E(x, <bf)and M(s, ~ r ~ )can ~ continued as meromorphic functions in t, to d ) be a;. If t, ia*, E(x, A) is a smooth function of x, and M(s, TT{) is a unitary operator from Xp(irg)to Xpi(sirr). There is an integer N such that for any 4 , sup (ll~ll-~lE(x, +)I
x G(A)

is a locally bounded function on the set of t, a at which E(x, c$~) regular. If : is h CJG(A))" and t lI(ar, a"), the following functional equations hold:

(ii) E(x, M(s, -n-)d)) E(x, 4). = (hi) M(ts, ir) = M(t, sir) M(s, Tr). Let Sft be a class of associated parabolic subgroups. Let L9 be the set of collections F = {Fp : P Sfi} of functions such that (i) Fp,(sir) = M(s, i r ) F p ( 4 , (ii) IIF1I2 =
PI- 9

s 0 ( a , a'),

n(A)-I

IIFp(ir)l12dir< w .

Then the map which sends F to the function

extends to a unitary map from onto defined for F i n a dense subspace of Ly, a closed, G(A)-invariant subspace Lj,(G(Q)\G(A)) of L2(G(Q)\G(A)). Moreover, there is an orthogonal decomposition L2(G(Q)\G(A)) =

@
9

&(G(Q)\G(A)).

We could equally well have defined subspaces L;(G(Q)\G(A)l) of L?(G(Q)\G(A)l). The only change in the formulation would be to integrate over WAf) instead of H M . ( ) This would allow us to decompose the representation R into a direct integral of the representations Ip(ir) (Let us agree to denote the restriction of the representation Zp(ir) to G(A)l by Zp(7r) as well.) If x 2, we could replace L9 by a space of collections F = {Fp : P E Sft} such that Fp(ir) belongs to ^ p ( ~ for each ir. We would obtain a decomposition of the space )~ L^Q^G(A)l)^. More generally, if P C P2, and if t, is a suitable point in a;, define

928

JAMES G . ARTHUR

The discussion above holds if the functions E(x) are replaced by Ep2(x). then It amounts to a description of the decomposition of the representation Rp, into a direct integral of { Z p ( 7 r ) : P c P2}. We shall end this section with some simple remarks on representations of the 0 universal enveloping algebra. %(go' C ) acts, through the representation Ip(7r),on the vector space ^ { 7 There are two involutions of %(g(R)l0 C ) . r) The first,

x + x,
is just conjugation with respect to the real form anti-involution),
g(R)l.

The second, (actually an

x +x*,
is the adjoint map. For any h e C(G(A)l and (We also have where h*(x) = h(xP1).) shall denote the left invariant differential operator We associated to X by R(X). If we wish to emphasize the fact that we are differentiating with respect to a variable x, we shall write Rx(X).Then It follows that for every E ^(-n) and X E % ( g ( R ) l ^i C ) that there a locally bounded function c ( { ) , defined on the set of points t, E a* at which E(x, <^) is regular, such that
TT

W M ) , X, Y E %(g(R)l 0 C ) , and

4 . The second formula for the kernel The theory of Eisenstein series yields another formula for K ( x , y). The equality of this formula with that of $2 is what will eventually lead to our trace formula. The following result is essentially due to Duflo and Labesse [2].

LEMMA 4.1. For any m 2 1 we can choose elements gg E I Z ~ ( G ( A ) ~ ) ~ R , E Q(G(R)KRand Z E %g(R)l 63 CKRsuch that Z * & + g; is the Dirac distribution at the identity in G(R)l.
g;

Proof. Let A be any elliptic element in %(g(R)l0 ClKR. For example, A could be obtained as a linear combination of the Casimir operators of G(R)land

A TRACE FORMULA FOR REDUCTIVE GROUPS I

929

KR.We can assume that A is KR-invariant.Let f l be a small open KR-invariant neighborhood of 1 in G(R)l. By the existence of a fundamental solution [4, Pg. 1741 we can find, for any n > 0, a function h on f l such that An * h is the Dirac distribution. By the elliptic regularity theorem, we can choose n so large that h belongs to Cm(fl).Since A is KR-invariant,we can assume that h Cm(fl)^. Moreover, h is infinitely differentiable away from the identity. Let $I be any function in C;(fl)^ which equals 1 in a neighborhood of the identity. Then A * h$I equals the Dirac distribution in a neighborhood of the identity, and is smooth away from the identity. Thus gi = h$I is in C^fl)^, and g;, the difference between An * gi and the Dirac distribution, is actually in C:(fl)^. The lemma is established with Z = An. COROLLARY 4.2. Suppose ro > deg Z. Then any h C':(G(A)lequals

where, in the notation of the lemma, hl = h * Z, h2 = h , and gi is the product of gi with a multiple of the characteristic function of an open compact subgroup of G(Af). Proof. Let KObe any open compact subgroup of G(Af)under which h is biinvariant. Define gi(xR xi), xR G(R)l,xf G(Af)l, i = 1, 2, to be gi(xR) divided by the volume of KOif xf KO,and 0 otherwise. The corol0 lary then follows from the lemma. LEMMA 4.3. There is a constant N a n d a continuous semi-norm II l l o on Cc(G(A)l)such that for any f E CC(G(A)l) and x, y G(A)l,

This lemma is well known, at least in the F\G(R)l setting ([3, Lemma 91). The extension to adele groups is easy. Recall that a function f on G(A)l is said to be K-finite if the space spanned by the left and right K-translates off is finite dimensional. LEMMA 4.4. There is an ro and a continuous seminorm \\-\\ro C?>(G(A)l) on such that ifX, Y %(g(R)l (g> C), r 2 ro + deg X + deg Y, and f is a K-finite function in C:(G(A)l),

930
is bounded by

JAMES G. ARTHUR

llX * f

Yllro - llxllN - llyllN. ) is finite. Let E(x, ir) = E(x, T T be~

Here N is as in Lemma 4.3.

Proof. Since f is K-finite, the sum over the vector in the algebraic direct product

equals

Apply Corollary 4.2 with h = X * f can be arbitrary.) Then h

=

* Y. (The integer m, upon which Z depends, 1 hi * gi. We can choose gi to be in

i= 1

COT(G(A)l)K. Since each function hi is K-finite, we can assume that gi is K-finite

as well. The absolute value of (4.2) equals

which is bounded by

Consider the set

9 = {(x,P, ir) : x

E, ir

IIG(M)}.

Regarded as a disjoint union of copies of IIG(M),it comes equipped with a topology. The integral (4.1) defines a measure on Y, Suppose that S is a com-

A TRACE FORMULA FOR REDUCTIVE GROUPS I

93 1

pact subset of if and that k is a K-finite function in Cc(G(A)l).It follows from the results on Eisenstein series summarized in $3 that
r

is the kernel of the restriction of the operator R(k * k*) to an invariant subspace. Since the operator is positive semidefinite, the value at x = y of this expression is bounded by

By Lemma 4.3 this is in turn bounded by

Since this last expression is independent of S it remains a bound if the original integral is taken over all of if. It follows from Schwartz' inequality that (4.1) is bounded by lxIlNllyllN . Since hl
=

x
i

Ilg*,* gil102

llhi * h?llo2.

* 2 , h2 = h and ro > deg 2 , the map which sends h to

0 is a continuous seminorm on Cp(G(A)l). The lemma follows. The K-finite functions are not dense in C;(G(A)l). However, there is a positive integer lo, depending only on G, such that for any r > lo, C^G(A)l) is contained in the closure in C - 'o(G(A)l) of the K-finite functions in C - 'o(G(A)l). If w = (Wl, W2) is a pair of equivalence classes of irreducible representations of KR,define

fJx)

deg Wl . deg W2

f
KR
X

d k o f (kf~kf)chw,(k2)dkldk2,
KR

if chw, is the character of Wi. It follows easily from the representation theory of a compact Lie group that lo may be chosen so that if r > lo and 11-11 is any continuous seminorm on C; - 'o(G(A)l),

is a continuous seminorm on C;(G(A)l), and fa converges absolutely in C: - lo(G(A)l) to f. Suppose that the measure space if is defined as in the proof of the last lemma. If S is a measurable subset of if and f is a K-finite function in C;o(G(A)l), define Z(S, f, x, y) to be
0

JAMES G . ARTHUR

(x./'.w) 6

ifi

V
@ P ( ~
+

E(x, Ip(r, f

I+)

~(,y,<i))dr.

For fixed x and y this defines a continuous linear functional on a subspace of C?(G(A)l) closure of which contains Cgo lo(G(A)l). can therefore dethe We fine Z(S,f , x , y) for any f E C> lo(G(AI1).
+

LEMMA 4.5. Zf f E C':(G(A)l), f, x, y) is continuously differentiable in Z(S, C either xor y of order r - ro - lo.IfX,Y E %(g(R)l@ )and r 2 ro + lo + deg X + deg Y ,

f R^X)R(Y*)Z(S , x, Y )
equals Z(S, X

* f * Y , x , y).

Proof. Suppose first of all that f is K-finite. If S is relatively compact the result follows from (3.1) and the proof of Lemma 4.4. In general S can be written as a disjoint union of relatively compact sets Sk. For any n ,

In the notation of the proof of Lemma 4.4, this is bounded by

As n approaches m this expression approaches 0 uniformly for y in compact subsets. It follows that the function

is differentiable in y and its derivative with respect to Y* equals

The differentiability in x follows the same way. Now suppose that f is an arbitrary function in C:(G(A)l). fn be a sequence Let of K-finite functions that converges to f in C':" lo(G(A)l.If ni > n z ,
R y ( Y * )I(S, X
=
5

* fn

x , Y ) - Ru(Y*)I(s,X

* fn;,

x, Y)]

\I(S,X

* fn, *

Y-X

* fn; * Y , X , y)l

ll~ll~llyll~ * ( f fn2)* Y\lrO, \\X -

A TRACE FORMULA FOR REDUCTIVE GROUPS I

by Lemma 4.4. Therefore the sequence Ru(Y*) Z(5, X

* fn,

X,

Y)

converges uniformly for x and y in compact sets. In particular,

m,X * f , x , Y )
is differentiable y , and

n+m

lim Z(S, X * fn

Y , x, y)

since X * fn

* Y approaches X * f * Y in CP>(G(A)I).

The differentiability in x follows the same way.

COROLLARY 4.6. Suppose that X , Y E '%(@)I C and r = rn + lo + deg X + deg Y. Then there is a continuous seminorm II II on C;(G(A)l) that if {Sk} such

is any sequence of disjoint subsets of Sl, and f E C;(G(A)l),

Proof. By the lemma, we may assume that X = Y = 1. The corollary then 0 follows from Lemma 4.4 and the remarks immediately following its proof.
COROLLARY 4.7.

Under the assumptions of Corollary 4.6,

where S is the union of the sets Sk.

Proof. Again by Lemma 4.5, we need only consider the case that = Y = 1. According to Lemma 4.4 and the remark following its proof,

But
a

1lz(5k,
k

fa,

x, Y ) \

934

JAMES G. ARTHUR

is bounded by the product of IlxllN llyllNand

x
0

llfwllro.

This last expression represents a continuous seminorm on CEO 'o(G(A)l), and is, in particular, finite. Therefore the double series above converges absolutely. It equals
+

as required.

It follows from the discussion of $3 that to any closed subset S of 9 there corresponds a closed invariant subspace of L2(G(Q)\G(A)l). Let Ps be the orthogonal projection onto this subspace. Set LEMMA 4.8. Z(S, f, x, y) is the integral kernel of Rs(f). Proof. Suppose, first of all, that f is a K-finite function in CNG(A)l). If 5 is compact, the lemma follows from the results on Eisenstein series summarized in $3. In general, S can be expressed as a disjoint union of sets Sk such that for each n,

sn= U
k = l

Sk

is compact. If

4 and ifi are functions in C;(G(Q)\G(A)l),

It follows from Lemma 4.4 and dominated convergence that this equals

JJ f(x) l{S, f,x, Y)4(Y'ldY dx

Now suppose that f is an arbitrary function in CEO+ ^^(A)). Let {fbe a sequence of K-finite functions that converges in Cp'(G(A)) to f. By the dominated convergence theorem,

A TRACE FORMULA FOR REDUCTIVE GROUPS I

Lemma 4.3 allows us to use, once again, the dominated convergence theorem. The above limit equals
( R s ( f)A,
$)a

We have proved that Z(S ,f , x, y) is the kernel of Rc(f ) . Suppose that x is an element in 3 and that S is the set { ( x ,P, T T ) } obtained by taking all P and IT. Then R s ( f ) is the projection of R ( f ) onto L2(G(Q)\GA)l),/. We shall write R x ( f )and Kx(x, y) for R s ( f ) and Z(S, f, x, y). It follows from Lemma 4.8 and Corollary 4.7 that

equals the kernel of R ( f ) . It therefore must equal K(x, y) almost everywhere. However the difference between these two functions is continuous in x and y (separately). It follows easily that

for all x and y. Suppose that P is a parabolic subgroup. If K ( x , y) is replaced by Kp(x,y) and E(x, 4 ) is replaced by Ep(x, d>), we can obtain obvious analogues of the definitions of this section, as well as of Lemmas 4.3 through 4.8. Then K p X ( x y) is , defined to be

iff is K-finite. Iff is an arbitrary function in C": lo(G(A)l),

+

for all x and y.

85. The modified kernel identity

After comparing the formulas of 42 and $4 for Kp(x, y), we note that
(5.1)

^ K p o ( x ,Y ) = ^ K p , x ( x , Y )
0

for all x and y. In this section we shall modify each of the functions K ( x , x) and K x ( x ,x) so that the sum over o remains equal to the sum over x. We shall later

936

JAMES G . ARTHUR

find that all of the modified functions are integrable over G(Q)\G(A)l. First, we shall need a lemma. If PI C P2, let
TP2
Pi

and

f :

= f:

be the characteristic functions on a. of {H E a. : a(H) > 0, a E A?} and

{ H E a,,: w(H) > 0, w E A?}.

We shall denote T$ and f$5 simply by r p and fp. LEMMA5.1. Suppose that we are given a parabolic subgroup P, and a Euclidean norm I I I I on up. Then there are constants c and N such that for all x E G(A)l and X up,

2
8

fp(H(8x) - X)

~(Ilxlle'l~~~)~.

E P(Q)\G(Q)

In particular, the sum is finite. Proof. Suppose that w E Ao. Let A be a rational representation of G on the vector space V, with highest weight dm, d > 0. Choose a height function relative to a basis on V(Q) as in $1. We can assume that the basis contains a highest weight vector v. According to the Bruhat decomposition, any element 8 E G(Q) can be written TT wgv,for TT E Po@), v No@) and s 0. It follows that llA(8)-1vII
2

1.

However, there are constants c1, c2 and Nl such that for any x E G(A)l,

~~llxll~~lIA(8x)-~ vll

It follows that there is a constant c such that

(5.2)

w(Ho(8x))5 c(l

+ log 11x11)

for all w E Ao, x E G(A)l and 8 E G(Q). For each x, let F(x) be a fixed set of representatives of P(Q)\G(Q) in G(Q) such that for any 8 r(x), 8x belongs to wSA(R)OK, where w is a fixed compact subset of N(A) and S is a fixed Siege1 set in M(A)l. Then there is a compact subset wo of N,,(A) Mo(A)land a point To in a. such that for any x, and any 8 F(x), 8x belongs to woAo(R)OK, and in addition,

A TRACE FORMULA FOR REDUCTIVE GROUPS

937

for every a E A:. We are interested in those 8 such that Tp(Ho(8x) X} = 1 ; that is, such that for every w Ap. The set of points Ho(8x)in a? which satisfy (5.2), (5.3) and (5.4) is compact. In fact, it follows from our discussion that if x G(A)l, 8 F(x) and Tp(H(8x)- X} = 1 , then 181 is bounded by a constant multiple of 1x1 a power of 1x1 e l l x l Since 11 l. for some c and N , 181 too is bounded by a constant multiple of a power of 11 1x1 e'^". Because G(Q) is a discrete subgroup of G(A)l,the lemma follows 11 from the fact that the volume in G(A)l of the set is bounded by a constant multiple of a power of t. 0 COROLLARY 5.2. Suppose that T an and N 2 0. Then we can find constants c' and N'such that for any function d>onP(Q)\G(A)l, and x , y G(A)l,

c'llxIlN llyIlN Proof.

sup (ld>(u)l llullrN).

u G(A)'

The expression (5.5) is bounded by the product of sup (I+(u)\. llull-^)

u G(A)'

and

We proved in the lemma that when Tp(H(8x)- H(y) - T ) was equal to 1, ll&cll was bounded by a constant multiple of a power of l l ~ l l e ~ ~ ~ p ( " ) corollary The therefore follows from the lemma itself.
+

Suppose that T is a fixed point in G. We shall say that T is suitably regular if a0 is sufficiently large for all a E Ao. We shall make this assumption whenever it is convenient, often without further comment. For the rest of this paper, f will be a function in Cg(G(A)l), where r 2 ro + lo. We shall also assume that r is as large as necessary at any given time, again, without further comment. Suppose that x G(A)l. For o 0 and x E, define

JAMES G. ARTHUR

kz(x,f ) = and

1 (P

l)dim(A~/Z)

8 e Pm\G(Q)

Kp,o(8X, X ) . Tp(H(8x)8

kz(x,f ) =

1(- l)dim(Ap/n
P
0

8 6 P(Q)\G(Q)

v p , x ( 8 x ,8x) - Tp(H(8x)

PROPOSITION 5.3. Proof.

2 kz(x,f ) = 1k%x,f ) . x
6 P(Q)\G(Q)

The left hand side equals

(- l)dim(Ap/a
P

x x
o

Kp,o(8x, 8x)Tp(H(8x)- T ) ,

since the sums over P and 8 are finite. B y (5.1) this equals

which is just

2 k:(x, f ) . x

46. Some geometric lemmas

We want to show that each of the terms in the identity of Proposition 5.2 is integrable over G(Q)\G(A)j. In this section we shall collect some geometric lemmas which will be needed in estimating the integrals. If PI C P,, set

for H ao. We hope that the reader will understand the motivation for this definition, as well as for the next lemma, after reading $7. LEMMA 6.1. If P, 3 Pi, o"? is the characteristic function of the set of H a, such that (i) a H > 0 for all a AL () (ii) a H 5 0 for all w A\ L () ,A and (iii) W ( H )> Ofor all w E A,. Proof. Fix H E a,. Consider the subset of those VJ in A, for which w ( H ) > 0. This subset is of the form Ac, for a unique parabolic subgroup R 3 P.,. Then

Suppose that T^H} = 1 for a given Pa 3 R. Then f i H ) = 1 for all smaller P3. It follows from Proposition 1.1 that the above sum vanishes unless the original Py equals R. Thus,

A TRACE FORMULA FOR REDUCTIVE GROUPS I

is the characteristic function of

(6.1)

{X E a, : a(X) > 0, a E d)?; a(X)

0, a E Ai\Af}

If R = P u?(H) = 1 if and only if H belongs to the set (6.1), as required. We must show that if R is strictly larger than P2, cr\(H) = 0. Suppose not. Then H belongs to the set (6.1). In particular, the projection of H onto a g i e s in the positive chamber, which is contained in set of positive linear combinations of roots in Af. Thus w(H) > 0 for all w E A?. By the definition of R, w(H) > 0 for all w AR. From this, we shall show that if

each co,is positive. Suppose that w E AR c A,, and that awis the element in A, which is paired with w. Then ca- = w(H) is positive. Therefore the projection of

, onto of is in the negative chamber, so that if v A?, v(Ha) is negative. If a ,is the root in A? corresponding to v, ca,, = v(H) - ~ H R )

is positive. Thus each ca is positive. Therefore w(H) is positive for each w E A,, and in particular, for each w in A,. Therefore R = P2 so we have a 0 contradiction.
COROLLARY Fix T a For any H a?, let /:/?bethe projection of H 6.2. . : onto a\. Then if u:(H - T ) # 0, a(/:/?) is positive for each a E A\, and

for any Euclidean norm I I I I on an, and some constant c. Proof. The first condition follows directly from the lemma and the fact that T belongs to a;. To prove the second one, write The value at Hz of any root in A2 equals a H for some root a in Al\Af. But (^ by the lemma. Since for each w A2, H2belongs to a compact set. In fact the norm of Hz is bound0 ed by a constant multiple of 1 + 11/:/?11 as required.

940

JAMES G . ARTHUR

This corollary will not be used until we begin to estimate the integrals of the functions defined in $5. A second consequence of Lemma 6.1 is the special case that Pv = PI, and Pi # G. Since every functional in the set & is negative on -a'\, c$i(H) is the characteristic function of the empty set. In other words,

for all H E a, ,. Suppose that Q and P are parabolic subgroups with Q C P. Fix A a?,. Let ?;(A) be (- 1) raised to a power equal to the number of roots a E AS such that A(&) 5 0. Let be the characteristic function of the set of H a. such that for any a A& wm(H)> 0 if A(av) <s 0, and wa(H) s 0 if A(a) > 0. In the special case that none of the numbers A(@) or wJH) is zero, these definitions give functions which occur in a combinatorial lemma of Langlands [l(b)]. It is desirable to have this lemma for general H and A, so we shall give a different proof, based only on Proposition 1.1.

equals 0 if A(av) 5 0 for some a E AS, and equals 1 otherwise. Proof. If R # P , our remark following Corollary 6.2 implies that
(- l)d"n(Al/ALJ)^(^P(^)
{R,:R C PI C PI

vanishes for all H . Therefore

is the difference between and

e;(A)G(A, H)r&H) (- l)dim(AllAp)^ff). P,<-f} We shall prove the lemma by induction on dim(Ao/Ap). Define A&to be the set of roots a E AS such that a(A) > 0. Associated to A& we have a parabolic subgroup PA,with Q C PAC P. By our induction assumption, the sum over R in (6.3) vanishes unless PI C PA, in which case it equals (- l)dim(A-^p)<(H). Thus, (6.3) equals
{R,Pi:Q C R C

(6.3)

1 :

A TRACE FORMULA FOR REDUCTIVE GROUPS I

94 1

if PA# P , and equals (6.4) minus 1 if PA= P. We need only show that (6.2) equals (6.4). This is a consequence of Proposition 1.1 and the definitions. Our final aim for this section is to derive a partition of G(Q)\G(A) into disjoint subsets, one for each (standard) parabolic subgroup. The partition is similar to a construction from [6(b)], in which disjoint subsets of G(Q)\G(A) are associated to maximal parabolic subgroups. More generally, we shall partition N(A)M(Q)\G(A), where P = NM is a parabolic subgroup. The result is little more than a restatement of a basic lemma from reduction theory, which we and would do well to recall. Suppose that w is a compact subset of No(A)M0(A)l that To -(To. For any parabolic subgroup Pi, let ^(To, w) be the set of pak, p w, a Ao(R)O,

K,

such that a(Ho(a) - To) is positive for each a E Ah. By the properties of Siege1 sets we can fix w and To so that for any Pi, G(A) = Pl(Q)WT0,a).The result from reduction theory is contained in the proof of [6(b), Lemma 2.12.1. Namely, any suitably regular point Tin a$ has the following property: suppose that Pi C P are parabolic subgroups, and that x and 8x belong to @l(T0, w) for points x G(A) and 8 P(Q). Then if a(Ho(x) - T ) > 0 for all a in A$\A$l, 8 belongs to Pl(Q). Suppose that Pl is given. Let ^(To, T, w) be the set of x in Pi(To,w) such that m(H0(x) - T ) ss 0 for each w A&.Let be the characteristic function of the set of x G(A) such that 8x belongs to ^(To, T, w) for some 8 P,(Q). F1(x, T) is left Al(R)ONl(A)Ml(Q) invariant, and can be regarded as the characteristic function of the projection of ^(TO, T, w) onto Al(R)ON1(A)Ml(Q)\G(A), a compact subset of Al(R)ONl(A)Ml(Q)\G(A). LEMMA 6.4. Fix P, and let Tbe any suitably regularpoint in To + at,. Then

equals one for all x in G(A), Proof. Fix x E G(A). Choose 8 E P(Q) such that 8x belongs to ^(To, w). Apply Lemma 6.3, with Q = Po, A ( a v , and H = Ho(8x)- T. Then there is a parabolic subgroup Pl C P such that w(Ho(&f)- T ) 0 for all VJ A; and a(Ho(8x) - T ) > 0 for a AT. Therefore so the given sum is at least one.

942

JAMES G . ARTHUR

Suppose that there are elements 8i, a2 G(Q), and parabolic subgroups Pi and P2contained in P such that F U ~ ~DT?(Ho(8ix) - T ) = F2(82x, T)T^o(8zx) X, Six W 0 , T, w),
i = 1, 2.
- T) =

1.

After left translating 8, by an element in Pi@) if necessary, we may assume that The projection of H(8,x) - T onto a$ can be written

a E A:

caa"

+ 2
w
E

cmw",

A'i

where each c,, and cWis positive. It follows that a(Ho(aix) - T ) > 0 for every a A$\A;. In particular, since T lies in To + a, 8,x belongs to ^(To, w). The : reduction theoretic result just quoted now implies that 828;' belongs to PI@) and 818a1belongs to P2(Q). In other words, 82 = (81, for some element in Pl(Q) n P2(Q). Let Q = Pl f l P,. Then HO(alx) - T and HO(8,x) - T project : . onto the same point, say Hc, on a If R equals either Pi or P2, we have w(H$) 5 0 for w and a(H$) > 0 for a E Ag. Applying Lemma 6.3, with A (at))+,we see that there is exactly one R, with Q C R C P , for which these inequalities hold. Therefore Pi = P2, and a1 and 82 belong to the same Pi@) coset in G(Q). This proves that the given sum is at most one.
$7. Integrability of kxx, f )

The primary goal of this section is to establish the integrability of each function kzx, f ) . In fact we will obtain THEOREM 7.1. For all sufficiently regular T,

is finite.

Proof. For any x, kxx, f ) equals the sum over P and over 8 E P(Q)\G(Q) of the product of and

This equals the sum over {PI, P : PoC Pi C P} and {8 E P,(Q)\G(Q)} of

For any H a. we can certainly write

A TRACE FORMULA FOR REDUCTIVE GROUPS I

{P,,P,:PC P2 c P,) by Proposition 1.1. In the notation of $6, this equals

T?(H)fp(H) =

( - l)dim(A2/A3)y3( ^ ) f 3 ( H )

We have shown that k z x , f ) is the sum over { P I ,P , Pi : Pi over 8 Pl(Q)\G(Q) of

T)Kp,o(8x, 4 . 8

c P c Pz} and

Therefore
(7.2)

is bounded by the sum over { P i ,P f l 1 C P,} and over o Cof the integral over P1(Q)\G(A)l of the product of with the absolute value of

The critical part of this expression is the alternating sum over P . In order to exploit it we shall show that the sum over y M(Q) n o in (7.4) can be taken over a smaller set. Fix P , and suppose that x P1(Q)\G(A)l. We can assume that neither (7.3) nor

vanishes. We want to show that the sum over y can be taken over the intersection of o with the parabolic subgroup Pi n M of M. We choose a representative of x in G(A)*of the form where k E K, n*, n* and m belong to fixed compact subsets of Nz(A),%(A) and Mo(A)lrespectively, and a Ao(R)On G(A)l has the property that and
(7 6 )

Here Tois as in $6. By Corollary 6.2, a(Ho(a)- T ) is positive for any a follows that the projection of Ho(a) - T onto equals

At. It

944

JAMES G . ARTHUR

where each ca and c_ is positive. Consequently, A well known lemma from reduction theory asserts that for any such a , a l n * m a belongs to a fixed compact subset of Ni(A) x M0(A)l which is independent of T. Suppose that the assertion we are trying to prove is false. Then there is a y in M(Q) n Pl(Q)\M(Q) such that

is not zero. This expression equals

Therefore there is a compact subset of G(A)l which meets alyN(A)a. Thus, a^ya belongs to a fixed compact subset of M(A)l. According to the Bruhat decomposition for M(Q), we can write y = vwsr, for v N$(Q), r M(Q) n Po@), and s flM,the Weyl group of (M, Ao), s cannot belong to the Weyl group of (Mi, Ao), so we can find an element m & not fixed by s. Let A be a rational representation of M with highest weight dm, d > 0. Let v be a highest weight vector in V(Q), the space on which G(Q) acts. Choose a height function I I relative to a basis of V(Q) as in $1. We can assume that the basis includes the vectors v and A(ws)u. The component of A(alvwsra)v in the direction of A(ws)v is e"^ - ^^oC") A(ws)v. Therefore The left side of this inequality is bounded by a number which is independent of T. On the other hand, w - sw is a nonnegative sum of roots in A$, and at least one element in A$\Ai has nonzero coefficient. It follows from (7.5) and (7.7) that the right hand side of the inequality can be made arbitrarily large by letting T be sufficiently regular. This is a contradiction. We have shown that for T sufficiently regular, (7.4) equals

According to the remark following Lemma 2.1, so the absolute value of (7.4) is bounded by the sum over y Ml(Q)

n o of

A TRACE FORMULA FOR REDUCTIVE GROUPS I

945

We would like to replace the sum over v by a sum over the rational points of n:, the Lie algebra of N?. Let be an isomorphism, defined over from the Lie algebra of No onto No which intertwines the action of Ao. The last expression equals

{RP, P c PJ c

2;

(- l)dim(Ala

1 ie n f ( Q )

/(;[-lye({ + X)x)dXl.
"(A)

Apply the Poisson summation formula to the sum over 6 . Let (-,-) be a positive definite bilinear form on no@) x no@) for which the action of Ao(Q) is selfadjoint, and let <b be a nontrivial character on A / Q . We obtain

{ E P , c P c P,}

2;

( - ~ ~ ~ 2 ( ~ /fix-lye(mx)<b((x, ~ x 1 . ~ a 0
YE

nXQ)

If n?(Q)'is the set of elements in n?(Q)which do not belong to any n?(Q),with Pi C P $ P i , this equals

6 n!(Q)

2;

f ( x - ' v e W x ) W . {))dXl,

n,(A)

by Proposition 1.1. We have shown that (7.2) is bounded by the sum over { P i , P2 : Pi C P2} of the integral over x in P,(Q)\G(A)l of

Set where k K, a A1(R)On G(A)l,and n*, n* and m lie in fixed fundamental domains in N2(A),N?(A) and Mi(A)l respectively. Of course, this change of to the variables will add a factor of e-2p~(Ho(a)) integrand. n* is absorbed in the integral over X. We need only consider points for which the integrand does not vanish. Therefore m and a-ln*a both remain in fixed compact sets. Thus (7.2)is bounded by a constant multiple of the quantity obtained by taking the sum over P I , Pa and y , the supremum as y ranges over a fixed compact subset of G(A)l, and the integral over a in A1(R)On G(A)l of the expression

The sum over y is finite. Our only remaining worry is the integral over a .

946

JAMES G . ARTHUR

Let

be the decomposition of n: into eigenspaces under the action of A!. Each A stands for a linear function on a:. Choose a basis of n:(Q) such that each basis element lies in some nA(Q).The basis gives us a Euclidean norm on n:(R). It also allows us to speak of nQ) and nA(Z). Note that

f ( Y - ~ Y ~ ( W Y ) $Ad(a)[))dX, Y n4(A), J((X, y + n,(A) is the Fourier transform of a Schwartz-Bruhat function on nt(A) which varies smoothly with y . Therefore we can reduce the integral over X in (7.8) to a finite sum of integrals over a real vector space. It follows without difficulty that for every n we can choose N such that the supremum over y of (7.8) is bounded by a constant multiple of

Every A is a unique integral linear combination of elements in A:. Suppose that S is a subset of roots A with the property that for any a A!, there is a A in S whose a coordinate is positive. Let nS(Q)' be the set of elements in n;(Q) whose projections onto nAare nonzero if A belongs to S and are zero otherwise. Then the above sum over n: all such S and over [ inns can be replaced by the double sum over

$ Z 1' . Clearly
2
c e ns(x1
2)
'

w:(Ho(a)- 2')

is bounded by

is the set of nonzero elements in nA

, ,

quotient of n by the number of roots in S. This last expression equals the product of

and

w:(Ho(a)- 2')

1
6
ng(x 2)
I '

llAd(a)[ll-n.

llAd(a)[ll-n

A TRACE FORMULA FOR REDUCTIVE GROUPS I

The first factor is finite for large enough n. The second factor equals

where each ka is a positive real number. The projection of Ho(a) - T onto a? can be written
a A:

tamy + H*,

where H* a;, and for each a A:, fa is a positive real number. If u?(Ho(a) 7') # 0, it follows from Corollary 6.2 that H* belongs to a compact set whose volume is bounded by a polynomial in the numbers { f a } . Therefore, there is an N such that the integral of (7.9) bounded by a multiple of is

This last expression is finite. The proof of Theorem 7.1 is complete.

58. Weighted orbital integrals For any II 0, set

In this section we shall define another function jT(x, f).We shall show that its integral is equal to J z ( f ) . Then we will, in some cases, reduce the integral of jT(x,f ) to a weighted orbital integral off. Given o and P, define the function

It is obtained from Kp,o(x, by replacing part of the integral over N(A) by the y) corresponding sum over Q-rational points. Define

THEOREM 8.1. For all suficiently regular T ,

is .finite. Moreover, for any

0,

JAMES G . ARTHUR

Proof. The argument used to prove the first statement is essentially the same as the proof of Theorem 7.1. The integral

is bounded by the sum over { P I ,P2 : PI C P2} and o 0 of the integral over P1(Q)\G(AI1 of the product of with the absolute value of

As in the proof of Theorem 7.1 we observe that for T sufficiently regular the sum over y may be taken over Pl(Q) fl M(Q) n 0. It follows from Lemma 2.1 and Poisson summation that the expression in the brackets in (8.3) equals the sum over y in Ml(Q) f l LI of

Let n?(Q,y8)' the set, possibly empty, of elements in n?(Q,y8) be which do not belong to any n{(Q, y8), with Pl c P P2. Then (8.1) is bounded by the sum over { P I ,P2 : PI C P2} of the integral over x in MI(Q)Nl(A)\G(A)l of

The integrand, as a function of n, is left Nz(A, ys) invariant. We can therefore write n = n2nl where n2 ranges over a relatively compact fundamental set for N!(Q, y8) in N?(A, 78) and nl belongs to N l ( A , y8)\Nl(A). For any y the integrand will vanish for nl outside a compact set. Next, set
x = mak,

A TRACE FORMULA FOR REDUCTIVE GROUPS I

949

where k K, a A1(R)Of l G(A)l and m lies in a fixed fundamental domain in M1(A)l.Note that a normalizes N l ( A , y8)\Nl(A). If the integrand does not vanish m and a-ln2a will both remain in fixed compact sets. Moreover, the sum over y will be finite. It follows that (8.1) is bounded by a constant multiple of the quantity obtained by taking the sum over P I , P2 and y , the supremum as y ranges over a fixed compact subset of G(A)l and the integral over a A1(R)Oi7 G(A)l of the expression

The finitude of (8.1) now follows from the arguments of $7. Fix o. The integral of jT(x, f ) is the sum over Pl and P2 of the integral over P1(Q)\G(A)l of the product of (8.2) and (8.3). Decompose the integral over P1(Q)\G(A)l into a double integral over M1(Q)Nl(A)\G(A)l and Nl(Q)\Nl(A). Then take the integral over Nl(Q)\Nl(A) inside the sum over P and y in (8.3).The summand is then

by Lemma 2.2. The integral over Nl(Q)\Nl(A) can now be taken back outside the sum over y and P , and recombined with the integral over M1(Q)Nl(A)\G(A)l. We must of course remember that (8.2) is a left Nl(A)invariant function of x. We end up with the sum over Pl and P2 of the integral over P1(Q)\G(A)l of the product of (7.3) and (7.4).This is just the integral of k:(x,f ) . The theorem is proved. 0 Let o be a fixed equivalence class. Choose a semisimple element yl o and groups Pl C P2 as in $2. Assume that PI = P2, so that o consists entirely of semisimple elements, Any element in G ( y l )normalizes A l , since it is the split component of the center of G(y1)O. obtain a map from G ( y l )to fl(al, a l ) , We whose kernel is G ( y l ) n M I . We shall say that the class II is unram$ed if the map is trivial; in other words if G ( y l )is contained in M I , It is clear that o is unramified if and only if the only s in Q ( a 1 ,a l ) for which wSylw;l is Ml(Q)conjugate to yl is the identity,

950

JAMES G. ARTHUR

Let o be an unramified class in 0, and let yl and PI be as above. Since it consists entirely of semi-simple elements, o is an actual conjugacy class in G(Q). Suppose that P is a parabolic subgroup and that y belongs to M(Q) n 0. Then there is a parabolic subgroup P2 C P , and an element y2 M2(Q), which is M(Q)-conjugate to y , such that the split component of the center of G(y2)0is A2. Any element in G(Q) which conjugates yl to y2 will conjugate Al to A2. It follows that for some s i2(al, a2),and q M(Q), Suppose that for a parabolic P3 C P , and elements s f fl(al, a3) and q' WQ), Then there is an element Since G(y) C M , for some element t M(Q). Let i2(al; P) be the set of elements s in the union over a2of the sets fl(al, a2)such that sal = a2contains a and s-la is positive for each a A$. Then given y and P , there is a unique s in fl(al; P) such that (8.4) holds for some q M(Q). It follows from this discussion that Jp,o(y,y) equals

c G(Q, y) such that

Since the centralizer of wsyIw;l in G is contained in M , this equals the sum over 8 G(Q, yl)\G(Q) of the product of and

Suppose that A is a point in a%such that A(aV) > 0 for each a A,,. We shall show that (8.5) equals

A TRACE FORMULA FOR REDUCTIVE GROUPS I

95 1

where we have written e2 and (f>i for the functions denoted by es, and <^, in $6. To see this, write (8.5) as the sum over P,, over s Wa,, az) and over those P 3 P2such that s^a > 0 for all a A of , : For a given s, define Ps 3 P by Then the sum over P is just the sum over {P : P, C P C Pg}. Since it is an alternating sum of characteristic functions, we can apply Proposition 1 . 1 . The sum over P will vanish, for a given s, unless precisely one summand is nonzero. We have shown that for all H do,

equals the product of (- l)dim(A~/Z) the characteristic function of with This is just the function e2(5w%oA, H ) . We have shown that (8.5) equals (8.6). After substituting (8.6) for (8.5) in the expression for j%x, f ) , we must integrate over G(Q)\G(A)l. Since the integrand is left Z(R)O-invariant, we can integrate over Z(R)OG(Q)\G(A) instead. We could then combine the integral over x and the sum over 8 if we were able to prove the resulting integral absolutely convergent. But the resulting integral can be written as

where

The integral over x can be taken over a compact set. By [l(b), Corollary 3.31, the integral over a can also be taken over a compact set. It follows that Jyf ) equals (8.7). We have expressed J^f) as a weighted orbital integral off whenever o is unramified. We note that v(x, T ) , the weight factor, equals the volume of the convex hull of the projection of

onto al/g, [l(b), Corollary 3.51.

952

JAMES G. ARTHUR

1. J. ARTHUR (a) The Selberg trace formula for groups of F-rank o n e , Ann. of Math. 100(1974), 326-385. (b) The characters of discrete series as orbital integrals, Inv. Math. 32(1976), 205-261. (c) "Eisenstein series and the trace formula," to appear in Automorphic Forms, Representations and Lfunctions, Amer. Math. Soc. AND J. P. LABESSE, Sur la formule des traces de Selberg, Ann. Sci. Ecole Norm. 2. M. DUFLO SUP.4(1971), 193-284. 3. HARISH-CHANDRA, Automorphic forms on semisimple Lie groups, Springer-Verlag, 1968. Linear 4. L. H ~ R M A N D E R , Partial Differential Operators, Springer-Verlag, 1963. AND Automorphic Forms on GL(2), Springer-Verlag, 1970. 5. H.JACQUET R. LANGLANDS, 6. R. LANGLANDS (a) "Eisenstein series" in Algebraic Groups and Discontinuous Subgroups, Amer. Math. Soc., 1966. (b) On the Functional Equations Satisfied by Eisenstein series, Springer-Verlag, 1976. (c) Base charge for GLs, mimeographed notes. 7. J. P. LABESSE R. LANGLANDS, AND Lindistinguishability for SLs, preprint. 8. A. SELBERG (a) Harmonic analysis and discontinuous groups in weakly symmetric Riemannian spaces with applications to Dirichlet series, J. Indian Math. Soc. 20(1956), 47-87. (b) Discontinuous groups and harmonic analysis, Proc. Int. Cong. Math. 1962, 177-189.