The Trace Formula in Invariant Form: Annals of 114 (1981), 1-74
The Trace Formula in Invariant Form: Annals of 114 (1981), 1-74
The trace formula for GL2 has yielded a number of deep results on automorphic forms. The same results ought to hold for general groups, but so far, little progress has been made. One of the reasons has been the lack of a suitable trace formula. In [l(d)] and [l(e)] we presented a formula
G is a reductive group defined over Q, and f is any function in C;(G(A)l). The left hand side of (I*) is the trace of the convolution operator off on the space of cusp forms on G(Q) \ G(A)'. It is a distribution which is of great importance in the study of automorphic representations. One would hope to study it through o E O} and { 1 x E X \ X(G)}. Unfortunately, these ; : the distributions distributions depend on a number of unpleasant things. There is the parameter T , as well as a choice of maximal compact subgroup of G(A)' and a choice of minimal parabolic subgroup. What is worse, they are not invariant; their values change when f is replaced by a conjugate of itself. In any generalization of the applications of the trace formula for GL2, we would not be handed the function f. We could only expect to be given a function such as
+(f1: 77
trdf
1 7
whose values are invariant in f. Here TT ranges over the irreducible tempered representations of G(A)'. The decomposition of trRcusp( ) into the right hand f side of (l*) would then be of uncertain value, for the individual terms actually depend on f and not just +( f ).
0003-486X/81/01141/0001/074$03.70/1 a 1981 by Princeton University (Mathematics Department) For copying information, see inside back cover.
JAMES ARTHUR
The purpose of this paper is to modify the terms in (1) so that they are invariant. Under certain assumptions on the local groups G(Q_),we will obtain a formula
in which the individual terms are invariant distributions. The definitions will be such that I = 1 if x belongs to X(G). We will therefore also have :
the analogue of (I*). The main assumptions on the local groups G ( Q ) are set forth in Section 5. One expects them to hold for all groups, but they are a little beyond the present state of harmonic analysis. They are, essentially, that any invariant distribution, I, on G(A)' can be identified with a distribution, f, on the space
This will apply in particular to the invariant distributions I. and I . In Section 13 we shall show that fo and f are natural objects on G. They are independent of any choice of maximal compact subgroup, maximal split torus, or even Haar measure. In this sense they are similar to the terms in the trace formula for compact quotient. A formula akin to (2*) is proved for G = GL, in [10(b), 581. (See also [9].) The main step is the Poisson summation formula on the group of idkles. At the right moment a sum over the multiplicative group of the field is replaced by a sum over Grossencharakters. Likewise, our main step is to apply the trace formula to the Levi components, M, of proper parabolic subgroups of G. To do this, we need to derive a function in CX(M(Al1) from f. Therein lies the difficulty. We can always assume inductively that (2) is valid on M, and use it instead of (1). Then we need only produce a function in ^(M(A)l). However, this is difficult enough. It will not be done completely until Section 12. The main step is a splitting formula for some tempered distributions, I M , proved in Section 11. In the applications of the trace formula for GLn it is important to show that on certain functions, 6,many of the distributions vanish. In Section 14 we shall study this phenomenon on a group of higher rank. The group will be G L . + will be the function in $(G(A)l) one expects to associate to a function on the general linear group of a division algebra. Our main tool is again the splitting formula of Section 11. The reduction to this formula is an elementary exercise.
The formula (2) actually follows rather formally from the existence of some auxiliary data. We present the formal manipulations in Section 4, along with an attempt to motivate our definitions. It is the proof of Theorem 4.2 that contains the crucial application of (2) to Levi subgroups. The best way to first read this paper is to look at (2.5) and the statement of Theorem 3.2, and then go directly to Sections 4 and 5. Section 4 is in fact intended as a second, more technical introduction to the paper. After Section 5 the reader might return to the earlier sections. Section 6 contains some lemmas which are used frequently throughout the rest of the paper. They are best motivated by the calculations of Sections 2 and 3. We shall conclude this introduction by illustrating how our methods apply to the trace formula of G = GL,. Let I be the group of iddes (on Q) and let I 1 be the subgroup of idkles of norm 1. We have subgroups
and
K = O,(R) X ~ G L , ( z . , ) =
P
UK,.
v
Let lo(f ) and &( f ) be the values of lJ'( f ) and Then the trace formula for GL, is
J^( f ), respectively, at T = 0.
Since we are in the special case of G = GL,, the distributions on the right can be evaluated explicitly (see [7], [4]). We shall copy them from [7], with minor modifications to fit our setting. Associated to each o E (9 is a semisimple conjugacy class { y } in G(Q). Let 6 ( G ) be the set of o for which the eigenvalues of y are not rational. For 0 e (3(G), lo(/) equals
where G(A, y ) denotes the centralizer of y in G(A). If o belongs to 8 \ (2(G) we can take y =
(n
JAMES ARTHUR
{(a, fS),(fS, a)} (consisting of one or two elements) of ordered pairs of nonzero rational numbers. If o contains two elements, lo(f ) equals
where if x = a
(1 "k,
for a
M(A), k
and rs(u) is a certain smooth function on {u = II ,,-c u,,: uu e Qz - { l}} with the property that the expression in the brackets in (c) extends to a continuous function of u E I1,,esQz. An irreducible unitary representation of M(A) consists of an ordered pair
of characters on I. There corresponds a representation J(P, v) = induced from P(A) to G(A). If s e C, write P,(u) = ii(u)IulS, and fu),
Note that the restriction, ji, of p to I1 can be identified with the orbit {p,: s iR}. Let R(p,v,s) = @R(^",vu,s)
be the normalized intertwining operator between I(p, v, s) and I(v, p, - s) defined as in [7, p. 5211. (Here, as in the rest of the paper, we shall agree that I(p, v, s ) acts on a space of functions on K . The space is independent of s, so the derivative R'(p, v, s ) makes sense.) If
is the operator which arises in the functional equation of Eisenstein series. Suppose that x e 9C \ 9C(G). It corresponds to a Weyl group orbit of automorphic representations of M(A)\ or as we prefer, a set {(ji, F),(v,^)} (containing \ one or two elements) of ordered pairs of characters on QX I l . Then Jx(f ) equals the sum of (el
-
i e x t r ( ~ ( ^ , 0 ) I ( w 3 0 , )), p f
and
where ex = 1 if jix = F and is 0 if they are distinct. If we sum over o e 6 and x 9C\9C(G), the contributions to the trace formula from the expressions (a), (b), (c), (d), (e), (f), and (g) are the respective analogues of terms (ii), (iv), (v), (i), (vi), (vii) and (viii) in [7]. As distributions in f,the expressions (a), (d), (e) and (f) are invariant. They remain unchanged iff is replaced by the function
This is not true of the other terms. If o belongs to (9 \ (9(G), J,,( f vanish, and can in fact be calculated explicitly. It equals
f ) does not
where
6
and
JAMES ARTHUR
Both fp and fp, are smooth functions on I 1 X J 1 . Similarly, one can calculate the resit of replacing f by f Y - f i n (g). It equals
Suppose that /is any function in C P ( G ( A ) )whose restriction to G ( A ) l is f . If p and v are characters on I set
Then
is a Schwartz function on f X f . It is the Fourier transform of a Schwartz function on I X I. Let &( f ) denote the restriction of this latter function to I 1 X 1'. It depends only on f , and not f. The expression ( g ) equals
This is just the sum over (jii, F) e x of the values of the Fourier transform of <^u{ f ) at We would like to apply the Poisson summation formula on 1' X J 1 to ^(f). In this case it is enough to know that & ( f ) is a Schwartz function; however if there is to be any hope for the general case it will be essential to show that Kv( f ) is compactly supported. The proof of this fact for general G and M, as we have already noted, is the main goal of this paper. The proof for GLo is actually rather trivial. At first glance, it might appear hopeless, for the function
(w).
f) certainly has poles. Remember, however, that <i>h(is not a function on I X I , but the restriction of a function to 1 X 1'. This is what saves the day. We can assume that
Almost all the terms in this sum over v equal 0. For any w , ( ~ ~ . vt r~ ( ) w , v w , f J ) I u is the Fourier transform on Qz X Qz of the function
+ -
is the Fourier transform of a Schwartz function on Q: X QE. It need not have and compact support. However, if CWis any compact subset of lI1^,,,..+,% C = Q: X Cw,the intersection of C X C with I1X Z1 is compact. The support of &( f ) is certainly contained in a finite union of such sets and is therefore compact. It depends only on the support off. Now apply the Poisson summation formula on I' X I1. The sum over x 9C\ %(G) of (g) equals
It follows from (b*) and (g*) that so that Zo is an invariant distribution. If x e 9C\ 9C(G), define ZJf) to be the sum of the expressions (e) and (f). Then Z is clearly an invariant distribution. Finally, set I = Jo if o 0(G). This of course is also an invariant distribution. Then we have
This is our trace formula in invariant form for GLo. The reader familiar with the invariant trace formula for GL2 in [ l q b ) ] will observe that it is different from the formula we have just given. For in [lqb)], Poisson summation was applied to
JAMES ARTHUR
the contribution from terms (b) and (c) above, whereas we have applied it to the contribution from (g). While being more immediately suited for the applications, the formula in [lqb)] is harder to prove for GLo, and may be impossible to establish directly for arbitrary G.
Contents
1. A review of the trace formula
2. The distributions J and Jx.
3. Noninvariance 4. The main problem: Discussion and motivation 5. Invariant harmonic analysis 6. Convex sets and some related functions 7. Some examples 8. The distributions J^ and }^ 9. The map C#IL 10. The invariant distributions lM, 11. A splitting property 12. Compact support 13. The invariant distributions lo and Ix 14. An example
1. A review of the trace formula Suppose that G is a reductive algebraic group defined over a field F of characteristic 0. More than anything else this paper concerns Levi components of parabolic subgroups of G defined over F , or as we shall call them, Levi subgroups of G. If M is a Levi subgroup, let A = AMbe the split component of the center of M; set
a = a ^ = Hom(X(MIF,R),
where X(M)Fis the group of rational characters of M defined over F. Now a is a real vector space whose dimension equals that of the split torus A. Suppose that L is a Levi subgroup of G which contains M. Then L is also a reductive group defined over F, and M is a Levi subgroup of L. We shall denote the set of Levi subgroups of L which contain M by P ( M ) . Let us also write T L ( ~for the set ) of parabolic subgroups of L, defined over F , which contain M, and let 6fL(M) denote the set of groups in (~L{M) which M is a Levi component. Each of for these three sets is finite. If L = G, we shall usually denote the sets by (M) F(M) and 9 ( M ) . (In general, if a superscript L is used to denote the depen-
dence of some object in this paper on a Levi subgroup, we shall often omit the superscript when L = G.) We shall try to reserve the letters L and M for Levi subgroups of G, and to use the letters P, Q and R for parabolic subgroups. If Mo C L are Levi subgroups of G, and P e FL(Mo), there is a unique Levi component Mp of P which contains Mn. It is defined over F. The unipotent radical, Np, of P is also defined over F. We shall write Ap and a ? for A M and a ^ . Suppose that M C Ml C L and there is a unique are Levi subgroups of G. If Q l C P L ( ~ ^ ) R e C P M l ( ~ ) , group, Q(R), in v ( M ) which is contained in Q and whose intersection with M, is R. Notice that there is a natural map from a Mto a^. We shall denote its kernel by a b Suppose for the moment that F is a local field and that Mo is a Levi subgroup of G. We will want to work with particular maximal compact subgroups of G(F), which we will call admissible relative to Mn. If F is Archimedean we will take this to mean that the Lie algebras of A M and K are orthogonal with respect to the Killing form of G. If F is non-Archimedean the vertex of K in the Bruhat-Tits building of G must be special and must belong to the apartment associated to a maximal split torus of Mn. Any K which is admissible relative to Mo has the following properties. (i) G = P(F)K for any P e CP(Mo). (ii) Any coset in G(F)/Mo(F) which normalizes Mo(F) has a representative in K. (iii) K = (Np(F) n K)-(Mp(F)n K) for any P e F(Mo). If L is a group in (Mo) n L ( F ) is a maximal compact subgroup of L ( F ) which is admissiK ble relative to Mo. For the rest of this paper G will be a fixed reductive group defined over the field Q of rational numbers. We fix a minimal Levi subgroup Mo of G. Then A. = AM,is a maximal Q-split torus of G. We shall say that a maximal compact subgroup
of G(A) is admissible relative to Mn if for each valuation v on Q, K is a maximal compact subgroup of G ( Q ) which is admissible relative to Mn, and if for any embedding of G into G L , defined over Q, KL= GL,,(o,. n G(QJ for almost all v . Fix such a K. Then K satisfies the conditions of [l(d)]. For any P e ?T( Mo) we can define the function
10
JAMES ARTHUR
from G(A) to a p as in [l(d), $11. Let Q be the Weyl group of (G, Ao). For any s e Q let w be a fixed representative of s in G(Q). w, is determined only modulo Hp(wS1) is uniquely determined. In [l(c)], Mo(Q), but for any P e thinking of the standard maximal compact subgroup of GL,,(A), we mistakenly stated that ws could also be chosen in K. However, we can chose an element Ws E K such that
1WSaWsp - wSaws1
so)*
To =
2
a=Al,"
ha&
The lemma will be proved inductively on the length of s. Suppose that and that the length of s1 is greater by one than the length of s. If we write w s l = msWsl, for m, E Mo(A), we see that ~ p ~ ( w - '= Hpo(ms)+ S - ' H ~ ~ ( W ; ~ ) = H,,(W;~) ) It follows by induction that Hp(wsl)
-
+s-l~~,(w;~).
(To - s l T o ) equals
By the definition of To, this equals 0. The lemma follows. We will eventually end up with objects which are independent of any choice of Haar measures, as well as our choices of K and Mo. In the meantime, however, f we had best fix some measures. Suppose that v is a valuation on Q. I M,, is any
11
Levi subgroup of G defined over Q , for which Kc is admissible, we assign M l Kc f ( Q ) the Haar measure for which the total volume is one. Suppose that and . L e (Me P E T L c ( M ) If u is discrete, we take Haar measures on Np(Qc) and M P ( Q ) such that the intersection of each group with Kc has volume one. Then if f E Cc(L(Q,)),
Here, ap is the modular function of the group P ( Q ) . If u is Archimedean, simply fix Haar measures on all groups {Np(Q),M p ( Q ) } given as above, so that (1.1) holds, and so that groups which are conjugate under K have compatible Haar measures. Now suppose that S is a set, possibly infinite, of valuations on Q. . Suppose that L e (Mnand P e W M ~ ) We take the restricted product Mp(Q,;) and KcsNp(Q,;).Then the analogue of measures on all the groups IIeEs (1.1) holds for functions f on k E s L ( Q u ) .In this way we obtain Haar measures on the groups Np(A) and Mp(A). By further restricting our choice of measure on Np(Qu),v Archimedean, we can assume that for each P, the volume of Np(Q) \ Np(A) is one. Then our measures on adkle groups satisfy the conditions of [l(d)]. We take the Haar measure on a O= aMoassociated to some Euclidean metric which is invariant under the Weyl group Q. The metric also gives us a measure MO),i a $ is isomorphic to the group of unitary on any subspace of a 0 . If P e 9( characters on a p. We take the Haar measure on i a $ which is dual to that on a p. The measures on M(A) and a n yield a measure on M(A)l, the kernel of the map
defined in [l(d)]. M(A) is the direct product of M(A)l and A(R)O, so we also obtain a Haar measure on A(R)O, the identity component of A(R). In the first three sections of the paper we shall examine the trace formula presented in [l(d), (e)]. In these sections we will try to use the notation of [l(d), (e)], so any undefined symbols will have the meaning assigned there. In particular, if P and Q are groups in %( MO),with P C Q, a ^ is the subspace a of ap. To the set, A$, of simple roots of ( P n My, Ap) there was associated a basis {a": a e A$} of a 9; A$ was defined to be the corresponding dual basis of ( a f ) *. Then A$ and A$ are naturally embedded subsets of a : Remember also that rf and Tf . denote the characteristic functions of { H E an: a( H ) > 0, a e A$} and { H e a ^: G(H) > 0, G e A$}. When there is an obvious meaning, we shall allow notation established for parabolic subgroups of G to carry over to parabolic subgroups of a Levi subgroup of G. For example, if R Z) Q, and Q, and P, are the intersections of Q and P with M R , A% = A$, r f l = r,' a yl = a y, etc. $
z;
12
JAMES ARTHUR
As it is given in [l(d), (e)], the trace formula depends on a fixed minimal parabolic subgroup Po O(Mo). Until we remove this dependence at the end of Section 2, Po will be fixed, and the letters P, Q or R will denote groups in O(Mo) which contain Po. The terms in the formula are indexed by sets 0 and X . O can be defined as the set of semisimple conjugacy classes in G(Q). The elements in X are Weyl orbits of irreducible cuspidal automorphic representations of Levi subgroups of G ( Q clearly acts on the set of pairs {(M, p)}, M a group in (Mo and p a cuspidal representation of M(A)'). The trace formula is an identity
associated to functions f in C?(G(A)'). T is any suitably regular point (depending on the support of f ) in a the positive chamber in a n defined by Po.: and : , / I,!' are distributions whose definitions we will recall in the next section. Many of our arguments will be inductive, so we will need to keep track of distributions on Levi components L = Mp of parabolic subgroups P, Po C P. We and 36 associated to L. If o can certainly define the sets of equivalence classes is a class in (9, o fl L(Q) is a union, possibly empty, of classes o1,. . . , o n in O L . Po fl L is a fixed minimal parabolic subgroup of L and T remains a point in the ' : associated positive chamber of a o. We therefore have the distributions 1 on CT( L(A)'). Define
aL
Similarly, suppose that x belongs to X . Then x is a G-Weyl orbit of irreducible cuspidal automorphic representations on Levi subgroups. This decomposes into a finite union, again possibly empty, of L-Weyl orbits x i , . . . , xn in 36 L . Again define
for all f E C^L(A)l), 2. The distributions lo JX and In this section we shall show that J^( f ) and I:( f ) are polynomials in T; that is, as functions of T they belong to the symmetric algebra on a;5,c. We will also take the opportunity to recall the definitions of the distributions. Fix f e
C F ( G ( A ) l )and o
where Z = AG,and
Suppose that T1 is a fixed suitably regular point in a n . We shall let T vary freely in Tl a:, and try to relate J r with the distributions JF,Tl. It evidently will be a question of expressing T p ( H p ( 8 x ) T ) in terms of the functions T F ( H p ( 8 x )- T,), Q ranging over parabolic subgroups that contain P. This suggests making the following inductive definition:
If X is a point in a n , define functions on a n , indexed by parabolic subgroups Q 3 Po, by demanding that for all Q Po, Ty(H-X)= ( - 1 ) dim(AR/Z) T [ ( H ) r ; ( H , X ) .
'
{ R :R ^ Q ]
The definition is indeed inductive; if r n ( H , X ) has been defined for all R f Q , then I;(H, X ) is specified uniquely by the formula. K ( H , X ) depends only on the projections of H and X onto a Q and it is invariant under a G . To express it in another way, consider the sum
R l , the sum over R vanishes (see the remark It is easy to verify that if following [l(d), Corollary 6.21). It follows that r b ( H , X ) equals (2.1). If G = GLg, a p/ a c is two dimensional. If X is in a : rp(- , X ) is the , characteristic function of the shaded region.
QZ
14
JAMES ARTHUR
It is the algebraic sum of the characteristic functions of the chambers at each of the four vertices. In general we have
d p /ag,
the support of
is contained in a fixed compact set, which is independent of X. Proof. If Q Z P, set ^^( H ) equal to the characteristic function of )
(H: S(H) > 0, s E i p \ i o ] .
Mimicking the construction of T^{H, X), we define functions ?r/ H, X) inductively by demanding that for all Q 3 Po, d'm(AR/ Z ) F / Q ( H ) ? ~ ( Hx ) . , fy(H-X)= (-1)
{ R :R 3 Q ]
Then
The values of these functions are easily seen from inspection. Modulo sign, rQ(., X) is just the characteristic function in a Q/ag of a parallelepiped with opposite vertices 0 and X. In particular, FQ(H,X) is compactly supported as a function of H e a /ag. The lemma will be proved by induction on dimG. We have
(-l)dim(Ap/Z)
rp ( H ) f y ( H ) = Q
0.
Therefore the outer sum may be taken over only those R not equal to G. For a : given R # G, and H E a, put
H= H*
Then where
+ H.,
H* E a:,
H* e a:.
15
is a linear map from a: to a h h i c h is independent of Q. If the summand corresponding to R does not vanish, H* will lie in a fixed compact set. So, therefore, will L ( H * ) . Applying the induction assumption to the group M n , we see that H* must lie in a fixed compact subset of a:. It follows that H is contained in a fixed compact subset of a:. The Fourier transform of Ti(., X ) will be an entire function on a;,c. It is easy to calculate. Let A be a point in a ,; whose real part belongs to - ( a "p . Then
+
: With this change of variables we must multiply by the volume of a modulo L$ X L y , where L$ and Ls are the lattices generated by ( 3 : A?} and { a v : a As} respectively. The result is the sum over Q 2) P of the product of
Define
and
if R 2) Q. (As suggested in $1, we sometimes write 06 as OS if R = G.) We have proved LEMMA The Fourier transform of the function 2.2.
H ^ T p ( H , X),
H fza p / a r .
16
JAMES ARTHUR
To evaluate the integral over IC /aG of Tp(. , X ) , replace A by t A , t > 0, in the formula and let t approach 0. The resulting limit must exist and be independent of A. Since @ ( A ) l O p ( A ) l is homogeneous of degree 9 dim( A p / Z ) , the result is
-
2 (-1) 91 { Q : Q 3 P )
dirn(AP/AQ)^
^ ,(
)"A
)-Ie
-1 QO .
It is a polynomial in X which is homogeneous of degree 9. Now we can return to our discussion of JAf ). In the expression for J3 ) , f make the substitution
Take the sum over Q outside the sum over P , and write the integral over ( x , 8 ) in (G(Q)\G(A)l)X(P(Q)\G(Q))as an integral over (Q(Q)\G(A)l) (P(Q)\Q(Q)). x Then J ^ f ) the sum over {Q:Q Z Po}, and the integral over x in Q(Q)\ G(A)l, is ) of
Decompose the integral over x into an integral over n in AQ(Q)\%(A), m E MQ(Q)\ MQ(A)l, E AQ(R)' f l G(A)l and k K . Since a MQ(Q), JAf) equals the sum over Q Z)Po of the product of
for 8
f~
with
2
{ P :Po C P C Q )
(-l)di~n(A~/AQ)
If u l , u2 belong to Mp(A)l,
equals
where
a smooth compactly supported function on Mp(A)l. The sum over P can be regarded as a sum over standard parabolic subgroups of Mp. It follows that (2.2) equals J p , T l ( We therefore have fp).
In particular, JAf ) is a polynomial in T. Next, take x e X . Then J$ f ) is the integral over x in G ( Q )\ G(A)l of
K p ( x , y ) is the kernel of the restriction of the operator Rp(f ) to the invariant Mp(Q)\ G ( A ) l ) x L2(Np(A) of Mp(Q)\ G ( A ) l ) .(See [ l ( d ) ,$3, subspace L2(Nr,(A) $41.) It can be obtained by projecting
Mp(Q)\ G ( A ) l ) .Then regarded either as a function of x or y, onto L2(Np(A) K p x ( x , y ) , regarded as a function of either x or y, is smooth. The analogue of (2.3) is established as above. The argument for J^f ) follows that for IT( f ) identically until we come to the integral
However, if we allow u , , u2 to belong to G(A)l,this integral is just the kernel of the restriction of the operator
18
JAMES ARTHUR
is a representation ~ 3of Mn(A)l~ the Hilbert space L~(N,(A)M,(Q) n . on M^A)l\MdA)l). Associated to x , we have a subspace L~(N,(A)M,(Q) n Mp(Al1 M/^A)')^ which is invariant under the operator ~ f n ' ~ ( , ( If we take \ fQ). ul and u2 to be elements in Mp(A)l,we obtain the kernel of the restriction of ~ ' % . ~ { fto this subspace. We therefore have o)
JbT and
PROPOSITION Suppose L = L for Q 3 Po, and that f E C W ) l ) , 2.3. ?( f )are polynomial functions of T . 0 and x E X . Then Jb f ) and '(
Jt,
These polynomials can be defined for all T. We shall denote the values f ) respecassumed at To, the vector defined by Lemma 1.1, by J b ( f ) and Theorem 7.11 and [ l ( e ) Theorem 2.11 that the series , tively. It follows from [ l ( d ) , 2 1 f ) \ and X X 1 f ) 1 converge. We obtain the identity
Ja
Jt(
It(
from (1.2). Suppose that Pl is another minimal parabolic subgroup in 9 ( M o ) .There is a d unique element s in Qi such that P = w l P p . If P is a parabolic subgroup that contains Po, P' = w l P w is a parabolic subgroup containing Pi. Suppose that
y = nmk, n
Ap(A), m e Mp(A), k
K,
= s-^Hp(y) H,(w;~).
Therefore
It follows that
=
{ P I :
(-1)
dim(Ap./Z)
P'~PO}
SF'(Q)\G(
Now
d We have shown that if Po and T are replaced by P and Hp( w s l ) definition of J^f ) , the result is the same. By Lemma 1.1,
+ s l T in the
^ It follows that J (f ) = Jo( f ) is independent of Po. The same argument applies to J : ( f ) and also to the corresponding distributions on Levi subgroups. Thus, depends on Mo and K, but not on a minimal each of the distributions J f and parabolic subgroup of L. They are defined for any L in Mo). Suppose that L' and L belong to ( M oand that L = w s l L w sfor s S2. ) Suppose that f C L(A)l),and x
Jk
Then from the argument above we see that and for all o and
x.
J^n=w
3. Noninvariance
G(A)l,define
20
JAMES ARTHUR
A distribution J on G(A)l is said to be invariant if J( f y ) = J( f ) for all f and y. Our distributions Jo and Jx are definitely not invariant. In this section we shall evaluate them on functions of the form f y - f. The calculation is similar to that of the last section. Fix a minimal parabolic subgroup Po in 9'(Mn). Given f and y, fix a suitably regular point T in ci $. In the formula for (I f Y),K p ,,(ax, 8x) will be replaced by :
2
=MP(Q)
r}
f ~ ( x - l 8 - l ~ n 8 x ) d = K , , ( S X ~ -Sxy-l). n ~,
o NP(A)
Thus, JA f ) equals
which, after a change of variables, may be written as the integral over x in G (Q) \ G(A) of
Let Kp(8x) be any element in K such that 8 x - K p ( 8 x ) lbelongs to P(A). Then fP(HP(kd
-
T ) = fp(Hp(8x)
T+H~[KP(~~)Y))
and
21
smooth compactly supported function on M,(A)l. The net result of the calculation is the formula
f Next, take x X . The analogous formula holds for .I.&?( y). It is proved the same way. The only additional point is that
R P ( f Y = MY )r1RP(f ) R P ( Y
1.
Modifying the discussion above (in the way we obtained (2.4) in the last section), we come to the formula
Both sides of (3.1) and (3.2) are polynomials in T. If we take the values at T = To of each side, the resulting distributions are all independent of Po. However, the sums are still only over those Q which contain Po. Suppose that Q' is any parabolic subgroup in ^( Mo). There is a unique Q containing Po such that Q r = wS-'Qws for some s in a. Then uk,(G,-'k, y) equals
It follows that
Jp f,,, (
) = J?(
fn
) and
( fy-,)f= J yn (^ ,
) for all
and
22
JAMES ARTHUR
x. Thus, the sums from (2.1) and (2.2) may be taken over all Q e ^ ( M n ) . The
number of Q which are conjugate to a given parabolic subgroup containing Po equals the order of S2 divided by the order of the Weyl group of the given Levi l/l S2 component. The corresponding summands must then be multiplied by 1 B A ~ Q 1. We summarize the results we have just established as a theorem. We want to leave room for future induction arguments, so we shall state the results for distributions on L(A)l rather than on G(A)l. THEOREM Suppose that L 3.2. y E L(A)l. Then
Mn), that f
CZ L ( A ) ~ )and that ,
and
forallo e 6 a n d x e 36.
Here f p is defined in the obvious way. That is, fp, y ( m )equals
where
This theorem is the basis of all that follows. We shall explain its role in the next section.
23
Suppose that for every L E (LV(,) U(L) is some vector space of functions with common domain a subset of L(Qs). We assume that U(L) is complete with respect to some topology. We assume in addition that for any y E L(QS)l= L(Qs) n L(A)\ the map
f-fY,
fEU(L),
f+f~,yj
given by the formula (3.3), sends U(L) continuously to U(Mp). By a distribution on U(L) we mean an element in U(LIr, the dual topological vector space of U(L). We suppose, finally, that we have been given a family of distributions { JL E U(L)'} and a family of nonzero complex numbers {c(L)}, each indexed by (M(,) such that for any L, y E L(Qs), and f E U(L),
Our primary example of such a scheme comes from letting U(L) = C7(L(Qs)') and c ( L ) = I^). If S contains the Archimedean place, we can take { JL} to be one of the families {: I } or {Jk}. For if f is a function in CT( L(Qs)I), the product off with the characteristic function of
is a function in CZL(A)'). Conversely, any function in CZL(A)l) can be J obtained in this way, for a large enough set S. It is in this sense that : and J i are regarded as distributions on CT(L(Qs)l). We would like to be able to associate a natural family of invariant distributions to each family {JL}. This will be possible if we are given some additional data. Suppose that for every M E Mo), V( M ) is a second complete topological vector space. Suppose that for every pair LVC L, we are given a continuous map
We shall sometimes write + for +$. In this case, (4.2) says that +( f y ) = +( f ) for each y and f. It follows that for every i in V(M)', the distribution
24
JAMES ARTHUR
The first statement of (4.3) implies that $I' is injective; the second states that any invariant distribution on U ( M ) is of the form ($I)'(i).If I is any invariant distribution on U ( M ) we shall let f be the unique element i in V ( M)' such that $ I f ( i )= I. PROPOSITION Suppose that {&} satisfies (4.2) and (4.3). Then for 4.1. every family { J L } of distributions satisfying (4.1) there is a unique family { I LE U( L ) ' } of invariant distributions such that for every f E U( L ) ,
Proof. Fix { J L } .Assume inductively that I ^ has been defined for all groups M E M o ) such that dim M < dim L . Define
for f E U ( L ) . We want to evaluate IL( f y - f ) , for y E L ( Q s ) . The function f equals f L . Therefore JL(f - f ) equals the sum on the right hand side of (4.1), but taken only over those Q # L . The same observation gives a formula for &( f Y - f ). It follows that IL( f Y - f ) equals the difference between
and
Now g L ( M ) is a subset of P ( M o ) . A group Q E P ( M o ) belongs to ( ~ L { Mif) and only if M C M y . Therefore I L ( f y - f ) is the sum over all Q E T L ( M o ) , Q # L , of the product of c( M& L ) with
This last expression is 0 by our induction assumption. Thus, I L is an invariant 0 distribution, as required.
25
Suppose that U ( L ) = C^(L(Qs)') and that maps <^, satisfying the hypotheses of the proposition, have been defined. If S contains the Archimedean valuation, we can regard J^ and as distributions on C ^ ( L ( Q s ) l ) ,as we noted above. Then by Theorem 3.2 we obtain two families { l f : o E Q} and (1;: x e X } of invariant distributions on C;(L(QS)l). Our invariant trace formula is a formal consequence of the definitions.
Ji
THEOREM Suppose that f (E C^(L(Qs)l).Then the series Z O l ^ f ) and 4.2. Zy1;( f ) converge absolutely, and
2 I f ( f = 2 q f 11 x
0
Proof. Assume inductively that the theorem holds if L is replaced by any with dim M < dim L. The series group M E ( M p
2 l )I W
0
equals
2
0
and
The first term is finite (see the remark following Proposition 2.3). By the assumption (4.3) <t)L( f ) can be regarded as the image under & of a function in CX( M(Qs)').The second term is then finite by our induction assumption. Thus, Z 0 1 I f ( f)l is finite. The same argument shows that
S m a r l y SJ 1 3 f ) ! is finite and
The required identity now follows from (2.5)and our induction assumption.
JAMES ARTHUR
with U(M) = CT(L(Qs)l). It is a task which will consume the rest of the paper. We conclude the section with an attempt to motivate how we will make the definition. The first step is obviously to define the spaces V(M) and the maps <^ = .@ : There are two apparent possibilities. We could try taking <^( ) to be what is f sometimes called the Harish-Chandra transform of f , obtained by taking orbital integrals of f. Then V(M) would be a space of functions on the regular semisimple conjugacy classes of M(QS)l.This was the approach taken in [5(a)]. Alternatively, we might take V(M) to be a space of complex valued functions on the irreducible tempered characters of M(QS)l.Then <^( ) would be defined by f the character values of f. It is this second alternative that we will choose. The solution that it eventually yields seems quite natural. We will discuss our candidates V(M) and <^ in Section 5, and the extent to which the hypothesis (4.3) is known. Once V(M) has been chosen we will define the maps &,. This amounts to associating distributions
on C')c(L(QS)') to irreducible tempered representations TT of M, which vanish if L does not contain M, and for which (4.2) holds. The only distributions that we know at present which satisfy (4.2) are the families { Jk} and { 1:). But x can be represented by a cuspidal automorphic representation on a Levi subgroup of G (which is a Levi subgroup of L if J i does not vanish). For certain x (those we called unramified in [l(e)]) Jx(f ) can be expressed explicitly in terms of these corresponding cuspidal representations [l(e), $41. We shall simply define /L , by , the appropriate analogue of this formula. At the same time, we define distributions
for regular elements y in L(Qs). They are obtained from an appropriate analogue of a formula [l(d), (8.7)] proved for Jo(f ) for unramified classes o. In Section 8 we shall show that these distributions satisfy (4.2). ( We can then define the value of <^ at TT to be J^, , f ). It is still necessary to show that <^k( ) belongs to V(M). This can be regarded as the main problem of f our paper. We will first solve the analogous problem for C(L(Qs)),the Schwartz space on L(Qs). For if U(L) is taken to be (?(L(Qs)),the spaces V(M) and maps <^: be defined by the prescription outlined above. In Section 9 we shall can also
27
show that <^ maps C?(L(Qs)) continuously to V(M). The distributions { J & , y } will all be tempered and satisfy (4.2), so Proposition 4.1 will provide us with a family { I -^ , } of invariant distributions on C?(L(Qi,)).In Section 11 we show how each I ^ , decomposes into distributions on the local groups L(Q, ). This allows maps CP(L(Qs)l) continuously into the us in Section 12 to prove finally that (^L associated space V( M).
Condition (4.3) becomes a question in local harmonic analysis, which has not yet been answered in complete generality. It is possible that an affirmative answer is not too far distant. At any rate, we shall simply assume what is needed. If H is any locally compact group, let II(H) denote the set of equivalence classes of irreducible (continuous) unitary representations of H. If the notion of a tempered representation is defined for H, we will let I I t ( H ) stand for those classes that are tempered. Suppose that v is a valuation on Q. If v is discrete there corresponds a rational prime p,. If v is real, set p, = e. Suppose that M is a Levi subgroup of G defined over Q, . Harish-Chandra defines the map H I * from ML(QL) C , = Hom(X(M, ) ^ , R) by setting to I
Suppose that
I(ML(Qc)). I If
is a vector in a *, c , we set
Suppose that LVLis another Levi subgroup of G defined over Q , M{ C M and ( M; ), that a E n Mi(QL)).If P E PC(we can lift aLto the parabolic subgroup PL(Q,), and then induce up to M ( Q ) . The class of the resulting representation ~ It is convenient to define, : of M ( Q ) is independent of P . We denote it by a. in a noncanonical way, a "norm" function on I I ( M ( Q ) ) . If v is discrete and 7 7 - E II(MJQL)), set 11 11 = 0. However, if v is Archimedean, let A be a fixed left invariant differential operator on M ( Q ) of order two. We assume that A is positive definite and that it commutes with right translations on M ( Q ) by K f?LV(Q,). For any T , E II(M,(Q,)) we obtain an operator 77-(A) on the space on which 7 7 - acts. Let I1 77-, II be its smallest eigenvalue. It is a positive number. We can follow the same prescription to define 11 W 11 for any W in n(Ki).
28
JAMES ARTHUR
Now suppose that S is a finite set of valuations on Q. Suppose that for each v e S, Mc is a Levi subgroup of G defined over Q . We shall refer to 91L = II-s Me as a Levi S-subgroup of G , and we write 91Ls = IIwsMu(Qu). Any 7r II(xs) is a unique tensor product 8 > m s ~ uof irreducible representations of the groups Mu(Qu)[3(a)]. If [ = @u.s[u is a vector in C y a:,c we shall put
Then we Most of the time we will take each Mu equal to a fixed M e (Mn) shall write uM for 0%. In this situation, we shall sometimes want to embed a a E, by vector A e a diagonally into
Our first candidate for U(M) will be the Schwartz space on M(Qs). Actually, only the Schwartz spaces <3(M(Qu)),v e S, appear in the literature. However, Harish Chandra's definition ([5(a)], [5(b)]) extends easily to M(Qs). Indeed, if v e S let Z r and of be t h e functions on M(Qu) used in [5(a)] and 1 in [5(b)] in the definition of <3( M(QJ). Given m = 1 uEsmu M(Qs), set
and
If the Archimedean valuation v belongs to S, let X, and XR be operators on Cm(M(Qs))which act, through Cw(M(Qu)),as left and right invariant differential operators. If the Archimedean valuation does not belong to S, set X, = Xn = 1. For any n 2 0 and f E Cw(M(Qs)),put
11 f 11 xL,x R ,
,I
=
ffi
sup
eM(Qs )
(1(xLxR ) ( m ) l ~ ~ ( m )+ l~ l ' ( r n ) ) - ~ } . / (
K,
let Q^(M(Qs)) be the space of smooth, KOft M(Qs) bi-invariant functions f on M(Qs) such that
< 00 for all XL, R and n. The seminorms 11 - 11 ^ , ^ , induce a topology on G(M(Qs)). X
XI., n X,,,
1 f1
The Schwartz space Q(M(Qs)) can then be defined as the topological direct limit, over all KO,of the spaces Q^(M(Qs)). When U( M) = Q(M(Qs)), we will take V( M) to be a space of functions on litemp(M(Qs)). V is any real vector space, let Diff(V) denote the space of If differential operators with constant coefficients on V. Now, suppose that d) is a complex valued function on lIwp(M(Qs)). If 7r is a finite sum of representations = Suppose that 911= k E S M u a is {7ri} in lItemp(M(Qs)), <f)(7r) Xi^'). put Levi S-subgroup of M and that D e Diff(@uEsiaz).If o e IItemP(911;,)and [ e @uEsia ( u ~is)a ~ finite sum of classes in TemP(M(Qs)). Then d)((ol)^) is defined. If it is a smooth function of L we shall denote its derivative with respect to D at [ = 0 by D&(o ^). Otherwise, put D&(o M , = 00. Suppose that KOis an open compact subgroup of Kg. Let G(M(Qs)) be the space of complex valued functions d) on l I t P ( M ( Q s ) )such that (i) $(7r) = 0 unless 7r has a (KOf l M(Qs))-fixed vector. (ii) For any 911, D the n 2 0, the seminorm
z,
\\d)HD,p
sup
(1 + I I o I ~ ) " ~ D & ( u ~ ) ~
a e ~ h p C x s )
is finite. We topologize $( M(Qs)) with the seminorms 11 . II p, Define 3( M(Qs)) to be the union over all KOof the spaces &(M(Qs)), equipped with the direct limit topology. 5(M(Qs)) is our first candidate for the space V(M). Given M and f e Q(M(Qs)), let + ( f ) be the function on IItemp(M(Qs)) whose value, <f)( f, v), at v is the trace of the operator '(fl ={f(x)v(x)dx.
M(Qs )
(For the existence of the integral and of trace class see [5(a)] and [12].) It follows fairly readily from the definition of Q(M(Qs)) that
<=<f):f+<?>(f),
f e e(M(Qs )),
maps Q(M(Qs)) continuously into 5(M(Qs)), Then (4.3) is the following assumption, which we take for granted from now on.
30
JAMES ARTHUR
ASSUMPTION For every M e (Mn) maps 6(M(Qs)) onto 5(M(Qs)). 5.1. (b The image of the transpose, (b', is the space of all tempered invariant distributions on M(Qs). The assumption will hold for S if it holds for each v in S. If v is Archimedean, it can be established from the results of [l(a)] and [8]. If v is discrete, the first statement of the assumption can probably be proved with the results of [5(c)], but the second statement is not known. However in the case that G = G L , the induced representations uA4are all irreducible (see [2] and [6]), and the second statement of the assumption can presumably be proved from this fact. Important examples of invariant tempered distributions on M are the orbital integrals. Let M ( Q S ) be the set of regular semisimple elements in M(Qs). An element y belongs to M ( Q s ) if and only if M ( Q S ) the centralizer of y in M(Qs), is of the form = IIGsTo(Qo),where each To is a maximal torus M defined over Q . Let
be the coefficient of degree equal to the rank of M in the characteristic polynomial of 1 - Ad(m). If m = L s m , ; belongs to M(Qs), set
identify I v t h a uniquely determined linear function of f" on g(M(Qs)). Suppose that L E and that P e ^( M). If f e 6 ( L(Qs)), M)
is a Schwartz function on M(Qs). For TT e I I t e ( M ( Q S ) ) , is a simple exercise it to show that (b( f, r L ) , the character of the induced representation TT^ evaluated fp, In fp) at f, equals <>( TT). particular, the element <^( in $(M(Qs)) depends only on M and not on P. We denote it by fM. As a function off, f M ( r ) is an invariant distribution. Assumption 5.1 then implies that the map f -> fM factors through a map (b -> GM from 5(L(Qs)) to 5(M(Qs)). It satisfies the formula
31
The map behaves well with respect to orbital integrals. If y belongs to L ( Q S )n M, it is easy to show that for all +. As we suggested in Section 4, our second and main candidate for U(M) is the space C?(M(QS)l). It could be defined as the space of compactly supported functions on M(QS)l which are restrictions to M(QS)l of functions in (3(M(Qs)). The orbital integrals of a compactly supported function should differ from those of an arbitrary Schwartz function only by being of bounded support in the variables y. This suggests a definition for our corresponding candidate for V(M). ). It will be a space of functions on ~ e m p ( M ( Q s ) lNow I I t P ( M ( Q s ) l )is the set under the action of orbits of ia * in IItemp(M(Qs)) Let Lat(S) = Lat(M, S) be the stabilizer in ia* of any 77. If T is as above, put = Ts n M(Qs)l and = n M(Qs)reg. We can project any <f> e g(M(Qs)) onto a function
+
(779
A)
7 7 ~ 7' 3
e ntemp(M(Qs)),
A e ia*.
G8
on IItemp(M(Qs)l). y belongs to s l , 1 3 ) depends only on the function (5.3). If We can define the notion of the support of a function in 5(M(Qs)), or as we prefer, of the function (5.3) on lltemp(M(Qs)l). <^ is the function (5.3), let If supp(d>l)be the set of pairs ('T,~ u p p ~ ( + where Tis as above and supper ~)), (+I) is the closure in Tsl of the support of the function Suppose that S = ( 9 , S ( 9 ) ) is any collection of pairs such that S(T) is a of compact subset of Tsl for each 5 . Let gS(M(Qs)l) be the set of functions <^>I the form (5.3) such that supp(<f>l) S;that is, such that suppT(<f>l) contained C is in S ( 9 ) for every 9 . The inverse image of gS(M(Qs)l)under the map (5.3) is a closed subset of g(M(Qs)). We give ^(M(QS)l) the topology induced by this map. We then define gC(M(Qs)l)to be the union over all such collections S of the spaces gS(M(Qs)l), equipped with the direct limit topology. Now if <f> = +( f ) , for f e <3(M(Qs)), the function (5.3) depends only on the restriction of f to we obtain a function on M(QS)l. In other words, for every f e C:(M(Qs)l) f). IItemp(M(Qs)l), which we continue to denote by <?)( It follows from standard properties of orbital integrals that
32
JAMES ARTHUR
is a continuous map from C;(M(QS)l) to 3,,(M(QS)l). Then with U(M) = C:(M(Qs)l), V(M) = $(M(Q_)l) and (fi equal to this map, (4.3) becomes the following assumption, which we also take for granted. onto ASSUMPTION For every M e (Mo) (fi maps C:(M(Qs)l) 5.2. ( M ( Q S ) l ) . The image of the transpose, <}>', the space of all invariant is distributions on M(QS)l. This assumption, too, will hold for S if it holds for each v in S. For archimedean v, it is essentially the characterization of orbital integrals of smooth functions of compact support, a well known problem. It has been solved for G = GLn in [10(b)] but there has apparently been nothing established for other groups. However, it will certainly be needed in any of the applications of the trace formula, so there seems no harm in assuming it at this point. If v is discrete the assumption does not amount to anything new. The orbital integrals of compactly supported functions can be characterized in terms of their Shalika germs. Moreover, Harish-Chandra has shown that the linear span of the orbital integrals is dense in the space of all invariant distributions on a p-adic group. His unpublished argument also uses Shalika germs. In summary, Assumptions 5.1 and 5.2 each contain two assertions; the statements in each case apply separately to real and p-adic groups, so there are eight assertions in all. The two assertions of Assumption 5.1 are known for real groups and unknown for p-adic groups, although probably within reach of present methods. The two assertions of Assumption 5.2 are known for p-adic groups and unknown for real groups.
could be extended to a smooth function on in*. We proved this geometrically, by exhibiting the function as the Fourier transform of a compactly supported function. We could have proved the result directly by transcribing the proof of Lemma 2.1. We shall in fact do this. We will obtain a more general statement, in which eA(x) replaced by an arbitrary function of A. is In the discussion of the functions r$ ., X) we used the fact that if P$ R , y - l ) d l m ~ ~ v T/ p~ (H)TR(H)= 0. i <? ~ )
{ Q :P C Q C R )
33
If the real part of A belongs to - a : we can integrate each summand against , e A ( H(see the proof of Lemma 2.2). We obtain the formula ) dim(Ap/AQ)^?(),l Q " ( \ 11 = 0 p (6.1) (-1)
{ Q :P C Q C R )
By analytic continuation, it is valid for all A a gcNow, suppose that c p ( A )is any smooth function on ia & = f a g . If Q 3 P, define cn(A) = cP(AQ), where, as before, AQ is the projection of A onto ia*_,. Copying our construction in Section 2, we define functions inductively by demanding that for all Q 3 P,
They are each defined on the complement of a finite set of hyperplanes in i a ; . It follows from (6.1) and (6.2) that if Q 3 P,
6.1. c;(A) LEMMA If c p ( A )is a smooth function on q, extends to a smooth function on i a g . Proof. Any element A in ia 5 can be written
c$+Ay
s-^
where each cz is a complex number and Ay belongs to subset of Ap. Let An,p denote the element
a^,^. Q 3 P, A,, If
is a
2
in a ?,,c.If Go(\) is a function on i
c-6
+ Ax
S E ~ ,
,, set
We then have the function cylp and also the function for R 3 Q. Suppose $ that a is a root in A p , which does not belong to AT. If c denotes the projection of a" onto a y ,
TIP,
34
Since
JAMES ARTHUR
whenever R 2) Q 2) P. Similarly, 6,$jp(A) = O&p(A)Qp(A) if Rl 2) R 2) Q 2) P. Mimicking the construction of cp(A), we define functions $,p(A) on E inductively by demanding that for all Q 3 P,
=
-
2
{ R I :R I ' Q )
~ " ; ~)"C-,~,~(A ) (
2
{ R :R 1 3 R 3 Q }
(-
l)"m(A~'Afl)
EQ/P(A).
Now k p ( A ) ' is a product of linear forms defined by those roots a in Ap\ A$. Fix such an a. Then the parabolic subgroups R , with R 2) Q, occur in pairs ( R , R'); if R is such that a does not vanish on a R , we define R' by setting
= cR',/P(A).
Since
dim(^^/^,.) = dim(^^/^,) + 1 ,
(6.4) vanishes whenever A(av) = 0. It follows from Taylor's theorem that (6.4) is divisible as a smooth function by the linear form A(av).Therefore, EQIp(A) is a smooth function of A. The lemma will now be proved by induction on dimG. Notice that if RDQ2)P,
and
Therefore,
Suppose that R = G. Then p) equals cWp(\), and is independent of Q. In view of (6.1), the sum over Q will vanish. It follows that cg(A) equals the sum over {R: P C R$L G} of (?R,p)~,(A function defined by (6.3),but with the (G, cp(A),Q) replaced by (MR,Glp(A), P n MR). Since %^(A) is a smooth function on F, our lemma follows from the induction assumption. To motivate the next lemmas we turn again to the example
.: where X is any point in a Then c'p(A) is the Fourier transform of the function r n ( - ,X). Consider first the function r;(-, -X). If there is no Q 2 P with $(H)fQ(H X) = 1, then Fp(H, -X) equals 0. Otherwise, let R be the largest group with this property. (R is defined by letting A 3 e the union over all such Q 3 of the sets A?.) As X is in the positive chamber, 7 H X)TR(H X) equals 1. This implies T^(H X)TR(H X) = 1, from which one can verify that Tp(H X) = 1 (see [l(b), Lemma 2.2 and the ensuing discussion]). Thus, S ( H X) > 0 for all (5 An. It follows that
which is ( - l)c1i1niAp/Z) if R = P , and 0 otherwise. Thus, rp(-, -X) is the and product of ( - l)dim(Ap/z) the characteristic function of
^ X) It follows from Lemma 2.2 that ( - D ' ^ ~ P / ~ ) I '-H, - X ) and Tp(H, have the same Fourier transforms. Therefore, modulo a set of measure 0, T i ( - ,X) is the characteristic function of
In other words, the figure drawn for GL3 in Section 2 is valid in general.
36
JAMES ARTHUR
In [l(b)] we studied families of points {Xn: P e 9( M)} which we called A*,,-orthogonal. This means that for every pair ( P , P') of adjacent groups in 3'(M) - Xp, is a multiple of the co-root associated to the unique root in Xn A n f' (-An-). Suppose that this multiple is always positive, and that, in addition, each point Xn lies in (a:)'. It follows from [l(b), Lemma 3.21 and what we have just shown, that the characteristic function of the convex hull of {Xn: P e 9 ( M ) } equals the function
In [l(b)] we calculated the Fourier transform of the characteristic function of the convex hull. It equals
In particular, this function of A extends to a smooth function on id?,. Suppose that for each P E 9(M), cp(A) is a smooth function on id?,. We shall call the collection
a (G, M)-family if the following condition holds: suppose that P and P' are adjacent groups in 9 ( M ) , and that A belongs to the hyperplane spanned by the common wall of the chambers of P and P'. Then c f i ) = cp,(A). This condition is equivalent to the property that whenever P and P' are elements in 9 ( M ) which are contained in a given parabolic subgroup Q, and A belongs to i a 5 , then +(A) = cp.(A). In particular, there is a well defined function, @A),
37
on la?,. The collection {e^^} is a (G, M) family if and only if {Xp} is AM-orthogonal,in the sense defined above.
L.
Proof. The only possible singularities are along hyperplanes Va")= 0 where a is a reduced root of (G, AM). Such a singularity occurs only in the terms corresponding to those P for which either a or - a is a simple root. But such groups in 9 ( M ) occur in pairs (P, P') where P and P' are adjacent, and have a and - a respectively as a simple root. If we multiply the corresponding pair of terms by \(av), and then take \ to be a point in general position on the hyperplane \(a") = 0, the result is 0, since cp(A) = co,(A). It follows from Taylor's theorem that cMdoes not have a singularity on the hyperplane. Fix a (G, M) family {cp(\)}. We shall often denote the value of cM(\) at \ = 0 simply by cM.To calculate it, set
A = tA,
and let t approach 0. If p = dim(AM/AG)we obtain
tR,Ac=
In particular, this expression is independent of A. Likewise, if Q contains some M), group in 9( we shall write cp for ~$0). It equals
where 9 = dim(Ap/AG). Now, fix a group L in (M)If Q E 9 ( L ) , P E 9 ( M ) , and P C Q, the function cp(\), \(= ia;,
+
is a (G, L ) family. Suppose that Q 9 ( L ) is fixed. If R p ( M ) , Q(R) is the unique group in 9 ( M ) such that Q(R) C Q and Q(R) f l L = R. Define a function c$' on i a t, by
38
JAMES ARTHUR
Then {c$(\): R e W M ) } ( L ,M ) family. In particular, we have functions is an c$(\) and (cf)'(\), P c P ( M ) , and their values c$ and ( c f ) ' at \ = 0. In general, c$ depends on Q, and not just on L. If it is independent of Q, we shall sometimes denote it by c;. If each of the functions
depends only on L, and not on Q, we shall denote it by c^(\), or even c,,(\), since R determines L uniquely. ) Suppose that {dp(\)} is a second ( G , M ) family. Then { ( ~ d ) ~ ( \= cp(\)dp(\)} is also a ( G , M ) family. There is a very simple formula for (cd)^{\). For geometric intuition, consider the case that G = GL3, M = Mn, and cp(\)dp(A) = eA('p)eW).
The volume of the shaded region equals c,%,. The volumes of the six hatched regions add up to d M . Each of the other six regions has volume equal to c&, for a maximal parabolic subgroup Q. LEMMA We have, in general, 6.3.
Since {dp(\)} is a ( G , M ) family, each function d$\) is well defined, in that it depends only on Q and not on the group P C Q. Interchanging the order of
Proof. Set
for each P ^ ( M ) . In view of our earlier remarks on convex hulls, c $ ( A ) will M ( M vanish if Q belongs to the complement of 9( ) in 9 M ) ; if Q belongs to 9( ) , c $ ( A ) trivially equals 1. The corollary follows from the lemma.
(M)
is independent o f Q. Then
Proof. This follows from Lemma 6.3 and Corollary 6.4. We remark that all of the results of this section are valid if the functions c p ( A ) take values in a complete topological vector space, instead of just C. Of course for Lemma 6.3 we would need to assume the space was also an algebra.
7. Some examples
Examples of ( G , M ) families occur naturally in harmonic analysis. One elementary example is obtained from the Weyl group translates of a point. Take M = M o , fix Po 9 ( M o ) and let T be a point in u o . Any P e ^ M y ) equals w 'Pows for a unique element s B. Define X p = s ' T . Suppose that P' = ( W ~ , ) P ~ Wis , adjacent to P. Then s' equals sas, sa the simple reflection , corresponding to a e Ape. The point X p - X p , equals s l ( T - s i l T ) , which is a multiple of s l a " . But s l a v = /Iv, where /I is the unique root in An ft ( A p , ) . Thus
is a ( G , M o ) family
40
JAMES ARTHUR
There is another elementary example, which we will use later, in Section 11. Fix M e (Mn) ft is any reduced root of (G, A), we can form the co-root f f . If For any P 9 ( M ) , set Xg = /?" if ft is a root of ( P , A), and let ~ f=, 0 otherwise. Let {ro} be a set of real numbers, and define X p to be the sum over all reduced roots /? of %Xg. Suppose that P and P' are adjacent. If ft is a root of both ( P , A) and (P', A), Xg equals X;,. If it is a root of neither, both vectors are zero. The only reduced root of (P, A) which is not a root of (P', A) is the unique root ft in , A n n (-Ap-). It follows that Xp - X p , is always a multiple of /?". Thus, is a (G, M) family. Suppose that L e (M)and that Q <E 9(L). We have the ( L , M) family
Write where Xp is the sum of r$& over those reduced roots /? which do not vanish namely, the reduced roots of on or, and XR is the sum over the remaining /?, (L, AM). It is clear from the definitions that XR is independent of Q and that Xp is independent of R. Therefore,
That is, c? depends only on L, and L, and not on Q. We denote it by $1. For our next example, fix a finite set S of valuations on Q, and fix Now G(Qc) is a subgroup of G(A), so for any x e G(Qc) and M <E (MO) P e 9(M), we have the vector Hp{x) in a . Suppose that P and P' are adjacent and that a is the unique root in An Fl (-Ap,). Then -Hp(x) Hp,(x) is a multiple of a". For the case that S = { R } this is [l(b), Lemma 3.61. The proof for general S is identical. Therefore,
ways to define
on(\, x ) ,
There is the ( G , L ) family { e m } there is the ( G , L) family derived as in or Section 6 from the ( G , M ) family {up(\, x ) : P O ( M ) } .The two are obviously the same. Notice also that if Q e O ( L ) and x belongs to L ( Q s ) ,the functions
x depend only on R and not on Q. We denote them by u^A, x ) or v R ( A , ) . Sometimes, however, we will take x to be a general point in G ( Q c )and use the function
Our final three examples, which are all basically the same, are derived from the intertwining operators between induced representations. Let v be a valuation on Q , and suppose that My is a Levi subgroup of G defined over Q . (From now on, unless we state otherwise, we will only consider groups My for which K is and a vector admissible.) Take a representation rye I I ( M y ( Q y ) ) If P e 9(M u ) , we can lift the representation I T ( to P ( Q v ) , and then induce to G ( Q y ) . This gives a representation Z p ( ~ , ( ) of G ( Q y ) on a Hilbert space '}Cp(ry). take '^Cp(%) to be the space of square integrable functions <^> from We K to the space on which I T acts such that
Then T p ( ~ y ) depends only on I T and not on lo. Suppose that v is real. Then there are canonically normalized intertwining operators
R P ~ ~ xP ( ( ) ~ %~ , ( d 7 1: p ^ ^ p p , p' %Mu),
42
JAMES ARTHUR
whenever P, P' and Pf' belong to ^(Me). Suppose that $o is a function in Kp(77-,,)such that &(k) = &(I) for all k K O . (In particular, the representation T,, is of class 1.) Then the normalizations have the property that for all P'. Suppose that L,, e ( M u and R e P~(M,,). Consider the induced representation
^:k ^ ^ k , k ) 7
k , e K,, n L,,(Q,,),
Finally, suppose that W e II(K,,). The space, ' } C p ( ~ ) w , vectors in Xp(77-,,) of is which transform under KOaccording to W, finite dimensional. Let Rp,lp(77-1;)w be the restriction of Rp,lp(77-,,) this space. Then for any D E Diff(iaz) there are to constants C and N such that
for all 77-,, e I I ( M ( Q ) ) , and W I I ( K ) . These properties were established in [l(f)l. Now take v to be a discrete valuation. There should also be a canonical way to normalize the intertwining operators between induced representations (see [ l q a ) , p. 2821). However to check the required properties, or even make the definitions in full generality would require a better understanding of harmonic analysis than is now available. Shahidi [ l l ] has introduced a normalization which applies to a large number of cases and is presumably the one of [lqa)]. At any rate, given Harish Chandra's work on the unnormalized intertwining operators (see [12]), it should be possible to define ad hoc normalizations with the right properties. From now on, we shall just assume the existence of operators
for which properties (7.1)-(7.6) hold. Suppose that S is a finite set of valuations. Let M be a group in ( M oand let 77- = @ v e s ~ cbe a representation in I I ( X S ) . For P e T( M) we define the
induced representation
which satisfy the analogues of (7.1)- (7.6). Suppose that Po e 9 ( M ) is fixed. Then, with
IT;,
as defined in Section 5,
is a family of operator valued functions of A e i a *. We can interpret each operator as a direct sum of operators on finite dimensional spaces. In fact, let K O be an open compact subgroup of Kc, and let W be an irreducible unitary w which are representation of KR. Let ' 5 C p ( ~ ) K , be the space of vectors in Xp(77-) invariant under KO and which transform under Kn according to W . Then '}Cp(7r)Ko,w is finite dimensional, and is an invariant subspace of each operator R p ( A ,I T , Po). Suppose that Q <E ^ ( M ) , and that groups P, P' e 9 ( M ) are both contained in Q. Then P = Q ( R ) and Pf = Q ( R f ) for uniquely determined It groups R , R f q M ' - ' ( M ) . follows from (7.5) that if A ia;,
RP,\P('7r\ = RP'lP(r1.
Consequently for any such A,
= R p ( A , r , Po).
Therefore {Rp(\,7T,Po): P e M ) } is a ( G , M ) family. More generally, if L e (M and Po P(M) we have the ( L , M ) family
composed of the intertwining operators acting on %$IT).We can form the operators R , * po) and R ~ I T , Q ^ M ) , Po), (the values of A = 0 of R L ( A , I T , Po) and Rb(A, I T , Po).) In general, each of these new operators is unbounded, but it can still be regarded as a direct sum of operators on finite dimensional spaces. It follows from the definition that if P, is a second fixed group in ^ ( M I ,
44
For any L f e C?(L(Qs)),IT
(E (E
JAMES ARTHUR
This is certainly an ( L ,M)-family. That each function is smooth in A can be obtained from (7.6) and the differentiation variant of the dominated convergence theorem. In particular, any differentiation with respect to A can be interchanged with the trace operation. It follows that for any Q e g L ( M ) the number + b ( f , I T , Po), obtained a priori from the family {+,(A, f , I T , Po): P (E P ( M ) } , also f equals tr(ZpJI T , & ( I T , Po)). Similarly,
Combining this last formula with (7.7), we find that ^(f , I T , Po) is independent of Po. We shall denote it simply by &( f , I T ) .The similarity to our notation of Section 4 is of course intentional. Incidentally, the same reasoning establishes ) more generally that for any Q e T L ( ~the,number ^(f , I T , Po) is independent of the group Po (E ( S I L ( ~ ) . It is obvious that each function +:(A, f , I T , Po) depends only on the unitary Fix equivalence class of IT. Suppose that M , (E P ( M ) . Qo e 6 p L ( ~ , ) and Po e ^ ( M I , with PoC QO- Then Po= Q o ( R o ) for R o e 6 p M 1 ( M ) . Given I T E I I t e m p ( M ( Q s ) )let r1be the induced representation ^ ( I T ) of M l ( Q s ) . If Q is , , any other group in 6 p L ( ~ , )and A e i a Ll,
for f e (3( L ( Q s ) ) . In other words, the ( L , M,) family associated to ' n M 1 and Qo is the same as the one derived from the ( L , M ) family { & ( A , f , I T ,Po): P (E 6 p L ( M ) } . The ( M , , M ) families derived from ( L ,M ) families are also related to intrinsically defined ( M I ,M ) families:
Proof. Let R n = Po n M,. Then Po = Q( R o ) . We need to evaluate the trace of an operator on %;(IT). Now %$IT)can be regarded as the space of square
45
integrable, K ("I Q(Qs) equivariant functions from K ("I L(Qc) to the Hilbert space ^ C ~ ; ( T ) .Then I&( f ) is an integral operator with kernel K ( k l ,k 2 )equal to
^P^M),
This establishes the lemma. ) Finally, suppose that f belongs to C y ( L ( Q s ) and that Po E 6 f L ( ~ and consider the functions )
TT E
ri(M(Qs)). Fix
That they are smooth in A follows again from (7.6). Notice that
depends only on the restriction of f to L(QS)land the orbit of denote this last function of A by
TT-
under i n *. We
where now f is taken to be a function in C,c(L(QS)' TT- is a representation in and n ( M ( Q s ) l ) .Then {<f,p(A,, T T , Po): P E 6 f L ( M ) }is an ( L ,M ) family which f satisfies properties analogous to those described above. In particular, we can associate a number
JAMES ARTHUR
is finite. (Since &(x) is left M-invariant, the integrand is L(Qs)-invariant.) and Since Proof. Write the variable x as I1 L-EsxL- let y = I1 %syL-.
9 = dim(AM/AL), we can rewrite the above integral as a sum of products of integrals over L(QL-),,o\ L(QL-).If v is Archimedean, the convergence of the resulting integral can be proved as in [l(b), Lemma 7.21. If v is discrete one knows that any finite dimensional L(Q,,) module over Q has a basis of eigenvectors for any given split torus in L ( Q ) , such that the corresponding lattice is stabilized by K O . This fact, together with [5(b), Lemma 131 (see also the proofs of Corollary 4.7.3 and 4.8.4 of [12]), allows us to transcribe the Archimedean proof. The argument is the same, so we need not present the 0 details. If y e L(Qs)reg l M and f e 6( f L(Qs )) define
By the last lemma, the integral converges absolutely, and each Jh, is a tempered distribution. If Q e gL(M), we have
defined in
belongs to
We see from the discussion of Section 6 that ub(x,y ) is same as the function defined in Section 2. As a function of x , ub(x,y ) is left Q ( Q s ) invariant. We write
48
JAMES ARTHUR
This equals
Our second family of distributions has actually already been defined. We introduce new notation only to point out the analogy with the distributions we have just defined. If 77- e l I t e m p ( M ( Q s )and f e ( 3 ( L ( Q s ) ) ,define )
JM,Af}=W^\
= tr(i^77-.
f )~^(77-. I ) , pi,
(8.2)
@.(fQ)
This for any Po e F ( M ) . formula follows from Lemma 7.1 along with the fact, noted in Section 7, that the right hand side is independent of Po.
f y ) equals
Substituting for
In this expression we can replace IpJ77-, y ) by IpJr A ,y ) the final limit. We obtain the trace of the operator
-'
49
where
One sees immediately from the definition of induced representation that for (fi E X p ( r ) = X ~ T Tand k K f l L(Qs), ),
Substitute this expression in (8.3). Take the sum over P inside the sum over Q. We shall show that the limit in A can be taken inside the sum over Q. That is, that
has a limit as A approaches 0. Suppose that Po C Q. If we can show that the limit exists in this case, then the limit will exist for an arbitrary Pi in TL(M).In fact, it will just be the conjugate by R p d l p ( ~ ) the limit for Po. When we evaluate the of trace, the two limits will be equal. Therefore, we may assume Po C Q. Now it is clear that if m MQ(A), It follows from this fact, (8.4), and (7.5) that if P E ^( P C Q, M),
6.)Therefore
Our notation here is confusing. The operator Up(\, TT, on the left hand side y) TT, has been obtained, via (6.3), from Up(AQ, y), while on the right hand side, Ub(A, TT,y) is obtained from UPJAQ,TT,y), and is independent of P. Now the existence of the required limit follows from Lemma 6.2. It follows that ~ ( T f y , , equals the sum over Q E T L ( M ) of the trace of the operator T
where Po can be taken to be any group in T L ( M ) with Po C Q. ). This operator acts on X P o ( ~= X ~ ( T TBut X p ( r ) can be interpreted as ) l the space of square integrable, (K f Q(Qs))-equivariant functions from K n L(Qs) to the Hilbert space X g ? , y ( ~ ) .Now (7.5) and (8.4) tell us how to
50
JAMES ARTHUR
interpret the operators R$(T, Po) and U ~ ( Ty) in this picture. Then (8.5) , becomes an integral operator with kernel K(k,, k2) equal to
R$(T, P ,
Therefore
nM~).
equals
Given groups M C L in (Mo) function f i n C?(L(Qs))and a class a ntenJ )I, we defined M(Qs in Section 7. Let <^( f ) be the map that sends According to Lemma 8.3, tlLs(fY)= for any y
E
Q&M)
in
2 Wfv,,)
is a continuous map from e{L(Qs)) to $(M(Qs)). This will establish all the hypotheses of Section 4 (modulo Assumption 5.1, of course) for the case that U(L) = C?( L(Qs)), V( M ) = ^( M(Qs)) and <^ is as just defined. The proof of continuity is essentially a result in local harmonic analysis. In order not to stray too far afield, we shall be brief. The reader familiar with Harish-Chandra's work on the harmonic analysis on the Schwartz space will have no trouble with the details. (See [5(a)] and 1121.) Fix Po E ?PL(M).Given f e t\L(Qs)) and Q e TL(M),let +h(f, Po) be the map that sends T e I I t ( M ( Q s ) ) to f c ( f , T , Po) = tr(lP"(T, f ) R ~PO)). T
f + @gf?po)
is a continuous map from (3( L(Qs)) to 5( M(Qs)). Proof. Let 9TL = k c s M . , be a Levi S-subgroup of M. We need to show that for any n and any D Diff(@.,iaE),
is a continuous seminorm on 6(L(Qs)). Now <>b( aM,Po) is the value at t = 0 f, of a linear combination of functions
w a,,,
LIE.5
0,
nternp(M.,(QJ)
for groups P., and Pi in OL(M,,).We must show that for D e Diff(iaz) and q and n nonnegative integers, the value at t = 0 of
is a continuous seminorm on (3(L(Q,,)). The case of v discrete poses no problem. For then, f is bi-invariant under an open compact subgroup of K,,. This means that the operator
is of finite rank. Moreover, the operator vanishes unless a,, belongs to a subset of IItemp(M,,(Q.,)) which is compact (in the obvious sense). These facts are easy consequences of a result of Harish-Chandra [5(c), Lemma 31. The continuity of our seminorms follows from the definition of (3(L(Q,,)). Now suppose that v is Archimedean. For simplicity assume that L = G. If T is an operator on Xp(o.,) let 11 T 11 be the trace of its positive semidefinite square root. For W in Tl(K,,) let Pw be the projection of %p(a,,) onto ';)Cp(a,,)w, and
52 define
JAMES ARTHUR
For any n and any D e Diff(ia*), there is a continuous seminorm 6 ( G ( Q , ) ) such that
11 - 1 1 on
( ( 11 D"!PJ(J,> A )w1,w21 1 1 < l l f ll(1 + I I ~ ~ l l ) "+ l I ~ ~ l I ) - " +l 1IuJI)-", for all f 6 ( G ( Q , ) ) , 0, e n t e m p ( M , ( Q u ) )and W17 e I I ( K , ) . This can be W2 established fairly readily from the definitions. It is essentially the easy half of the theorem stated in [ l ( a ) ] .The estimate we require then follows from this inequality, (7.6), and the fact that
COROLLARY 9.2.
( ) Now suppose Ml p M ) and that Pl is a group in 6 f L ( ~ ^ which contains a given Po e ^ ( M I . Suppose also that Q e p ( M l ) . Then for any f e C ( L ( Q s ) ) , <f>y( f , Pi) is a function in 5 ( M l ( Q s ) ) .Then <^( f,PllM is a function in g ( M ( Q s ) ) . Its value at 77 e l l t e m p ( M ( Q s ) )is < } > y ( f , ~Pi), which by (7.8) equals ^, % ( f , 7 ~ Po). Thus ,
(gel)
<f'b(f^l ).M
W , 1'
'0
The map < is particularly simple on spherical functions. I f f is bi-invariant ;^ under the maximal compact subgroup K n L(Qs)of L ( Q s ) , it follows from (7.3) that <^( f ) vanishes if M # L. Finally, suppose that S is a disjoint union of S and { v } . Let f be a function in 6 ( L ( Q s ) ) ,and a Schwartz function on L,(Q,) which is bi-invariant under K O n L J Q , ). Then if / = /A
53
satisfy the assumptions of Section 4. Suppose that y is an element in L ( Q s )n M. We defined the distribution J ^ ,in Section 8. If MI e G L ( ~ o and does not ) 2 contain M, define 1 1 to be 0. Then by Lemma 8.2,
for all y in L(Qs). It follows from Proposition 4.1 that there are invariant tempered distributions ~eL(Qs\eg^M>',Me^(Mo), on L(Qs)such that
Jl,,(f)=
M.
2
Ee
%,(&(f))
L(~")
for all L and M. Observe that I^, will vanish unless M C L. Although they are defined by a simple formula, these distributions are in some ways rather complicated. For example,
G , , ( f ) L,,(fL =
e(G(QJ),
depends only on the image of f in 5(G(Qs)).However, as a distribution on $(G(Qs)), ,has no simple formula. I^, ,} Notice that we could have used the distributions { J;, instead of { J&, } to obtain invariant distributions { I ^ ,,}. However, as the distributions J ^ , , were used to define the maps 'fi'y, this leads to nothing new. In fact, 1 5 , vanishes if M L, and
^(f
= J$,,,(f)
= ' f i ( f 7 7 7 ) = t r 7 r ( f ).
The rest of this section will be devoted to proving a useful property of the and that y distributions. Suppose that M C MI C L , are groups in ( M o ) belongs to Ll(Qdreg M. We shall show that n ,(+) can be expressed as a linear combination of
fk,
~ 1
(f&,,(h):
'fi
$(L,(Q~)).
First, we shall prove two lemmas. Suppose that M e ( M o fixed, and is that {cp(\): P E 6?(M)}is a ( G , M) family. We assume in addition that if M C L C L l , and Q e 6?(Ll), number c? is independent of Q. We denote it the by cf-1.
54
JAMES ARTHUR
for fixed Po e ^(M). Since cL = cg, the value of the left hand side of the required formula at TT is
by Lemmas 6.3 and 7.1. This is just the value at TT of the right hand side.
Proof. Let f be any function in C2(Ll(Qs)) such that <>( ) = <>. f -The left hand side of the required formula equals
We shall prove the lemma by induction on dim(L,/M). Apply the induction hypothesis to the sum over L in (10.2), and then add the result to the right hand side of the required formula. We obtain
( L ,M , : M C L C M , C L l }
by Lemma 6.3 and Corollary 6.4. It follows from (8.1) that the expression (10.1) equals
Recalling the inductive definition of the distributions equals (10.3). We shall apply the last lemma with
with
as in Section 7. The constants rn are still to be chosen. Fix Ml E ( M )Choose positive numbers lp for each of the reduced roots of ( M I ,A ) such that for every the sum R E V ( M ) , over all reduced roots /? of ( R , A ) of ln/3" belongs to the chamber in a associated to R. Fix t e R. For any reduced root /3 of (G, A ) let rn = tla if /? vanishes on a ^ , and let rn = 0 otherwise. Now suppose that L e ( M )We claim that c^ vanishes unless L C Ml. For as we saw in Section 7 ,
It is the volume in a h of the convex hull of the points { X R : R e ^ ( M I } defined in Section 7. Each X R is orthogonal to a ^ , and by our choice of {rn}, it is also orthogonal to a ^ . Therefore, if L is not contained in Ml, the points { X R } all lie in a proper subspace of a h . The convex hull then has volume 0; c\, then does vanish. On the other hand, if L = Ml, each point X R lies in the chamber of ah corresponding to R. The volume of the convex hull does not vanish by the results of Section 6. Therefore c,^ # 0. Now fix L 1 ( M i ) <f> E 3(Ll(Qs)), For
is a polynomial in t , whose highest term is c 3 & , ^ ( @ )The right hand side of . the identity in Lemma 10.2 is also a polynomial in t . Equating the highest terms,
56
we find that = dim( a El), of
JAMES ARTHUR
For any L in this last sum, there are natural maps The composition gives us a map from a31 to a i l . Suppose that it is not an isomorphism. Then by our choice of {ro}, the images in a i l of vectors {Xn: Q e '5""(M)} span a proper subspace of a i l . It follows that c p = 0. Thus we may include only those L in the sum for which the map is an isomorphism, or what is the same thing, for which a 3 = a$@ a$. We have established
LEMMA 10.3. Suppose that M C Ml c Ll are groups i n (M(,)Then for every L e p l (M) there is a constant d( L), which vanishes unless
a L l = a^,@ aE1, .M such that for all <f> e 3(Ll(Qs)) and y e L 1 ( Q s ) fl M,
fL1 Y .MI,
2
LEE^'(M)
d(~)fi,,(<f>L).
11. A splitting theorem It is important to be able to express the distributions I^, on L(Qs) in terms of distributions on the local groups L ( Q ) . As before, M C L are groups in (M(, and S is a finite set of valuations on Q.
THEOREM Let <f> e 3(L(Qs)) a d y e L(Qs)reg l M. Suppose that S is 11.1. f a disjoint union of two subsets S' and Sf' and that <f> and y decompose relative to the product L(Qs) = L(Qs,)L(Qsas <f> = +'+" and y = y'y" respectively. Then
57
is any function such that @ ( f )= @' and @(/") = @". Then ( difference between J;, f ) and
Zb, ( @ )
is the
where f^yl9 f-''2M I , Y" stands for the invariant distribution on $ M d Q s - ) 9 9( M~(Q,;,,)) that maps a function if,' 9 if," to f$;,,.(if,;>)f$;, ,,,,(if,"). Therefore, ( 11.1) can be written as the difference between
and
This last expression is just what we want. We will therefore be done if we can (f). show that (11.2) equals The value of <^( f ) at TT e q e ( M 2 ( Q s ) )is obtained from the ( L ,M y ) family PEP ~ ( M o p ( L, r , Py), f
J;,
~),
TT",
Now @ b ( fP2IM is an element in $ ( M l ( Q s , ) ) .As we observed in Section 9, , for any group P, E 6 f L ( ~ , ) , contained in Po. Also, Lemma 7.1 allows us to write %(/", P2) as f;). Consequently, (f,'?;\V1@ y , ) ( ^ , ( f )) equals
a(
c;,
( 11.3)
2
QE'S'^MZ)
&,(@(,(/'.) ) & pi
y-(*(f{))-
lW^f,
Pi)) is
58
JAMES ARTHUR
we obtain
According to (8.1), the expression in the brackets equals the integral over y in L(Qs,,)^,,,\ L(Qst') of the product of
1~
with
(11.4)
2 f~~.,,(@(.(f'.Pi')~~~(Y 1.
QES-~(M~)
TT'
The value at
Since v n ( y ) is independent of M l , (11.2) equals the integral over y in L(Qs,,).,,,\ L(Qs,,)of the sum over all Q E P ( M )the product of of
with
/a,(
Moreover,
IDL(-y')]112.1 DL(y")] ' I 2 = 1 D L ( y )I ' I 2 ,
and
f'(x --\^z)f"{ -\^"Y)
= f ( b y )-ly(xy
1).
0
It follows that (11.2)equals J ^ ( f ) . This is what we were required to prove. If we combine the theorem with Lemma 10.3 we obtain COROLLARY For every group 11.2.
and
we have
Proof. Let v be a valuation in S, and let S' be the complement of v in S. Then we can decompose <f> = <f>'<^, y = y'yr relative to the product L ( Q g )= and L ( Q s )- L(Q,,).By the theorem, f&, ,,(+I equals
a 2 0 a^.
60
JAMES ARTHUR
We will need a slight generalization of this corollary. Suppose that v is a valuation. We have been studying the distributions
on 6(L(Q^,)).It has always been understood that the Levi subgroups M C L were defined over Q. This is clearly not necessary. For any pair M C L of Levi subgroups of G defined over Q, for which K.,, is admissible (a condition we will assume for the rest of the paper), we could just as easily have defined distributions Y, E Lu(Q, )reg n M,, 1% yi..'
L C
on LJQ,). Lemma 10.3 would certainly continue to be valid. We will still, however, retain the fixed Levi subgroups M C L defined over Q. Then there is a surjective map
for any m in M(Qs). Suppose that = 11, L is a Levi S-subgroup of L. There is a )~, = Hom(X( L)Q,R). We therenatural map from a = H o ~ n ( x ( L ^R) onto fore have a surjective map
hL: @ c^-^a/,.
vcs
1 3
61
defined above. Let us write EL(%, S ) for the set of levi S-subgroups = 1 " Lo 1 of L such that Mu C L for each v. If E is any group in P(^Tl, S), we have a surjective map h: @ a M - + @ a L L .
v
All the maps are surjective. The kernels of h and ha respectively. So we have a map
(11.5)
@ a$&+ cs-s
a^. L
We shall be most concerned with those E ^(u% S ) for which this map is an isomorphism. When this happens, we also have kerh fl kerhM= 0, and the equivalent property that ker He^ = ker Xg fl ker H>f.
COROLLABY For every = k,s in EL(%, S ) there is a constant 11.3. L,, d(E), which equals 0 unless the map (11.5) is an isomorphism, such that for all
we have
62
JAMES ARTHUR
4%') e s v
n 'sy{h.)
1.
Now, apply Lemma 10.3 to each of the distributions 1^!(&, yL). We obtain the sum over Levi subgroups L in EM^(^), defined over Q , of ~ ( L " ) ' ; , : . JLL ~ @ The constant d ( L ) will be 0 unless (11.6)
a"L L = a:&@ M aL
abL.
a M' MM"
is an isomorphism. Let = I I L . In view of (11.6) 9TL' is uniquely determined by Define d ( = c ( 9 R ' ) The map (11.5) is the composition of
@
ues ues
d(~,).
@
ues
a z - a ML.
If the first map is not an isomorphism, d ( L " ) = 0 for some v . If the second is not an isomorphism, c ( ( % ) ) = 0. Therefore d ( is 0 unless the map (11.5) is an
isonvo~hism.
and
'in.,('fc
'
ues
n,!.'
Yb,.
LC
,(<('I
eeeL(as) ,
d(&)
tl^,y(@e).
&:e ( L ( Q s ) )
!^(M(Qs))?
63
we also defined a number ^.{ f , TT) f e C F ( L ( Q ~and ) for ) ~ TT will take TT to be a tempered representation. Then
TI(M(Qc)l). We
)l) is a complex valued function on l l t e m p ( M ( ~ s Our. goal is to show that it From Corollary 9.2 and its very definition, we know that belongs to M(Qs)l). &( f ) is obtained from a function in 5(M(Qs)) the projection (5.3). We have by only to show that the orbital integrals of <^w( f ) are compactly supported.
THEOREM12.1. If M C L, < ^ h a p s C z L ( Q S ) l ) continuously to WQsI1). Proof. We shall prove the theorem by induction on dim(L/M). Suppose that T H O S T L , where for each v , T is a maximal torus of M defined over Q L . = M(Qs)l. We must show that for every Set Ts = ll LEsTL(QJ, and n compact subset C of L(QS)lthere is a bounded subset D of q1 M ( Q s )such that whenever f is a function in C:(L(Qs)l) which is supported on C, the function Y Y % n M ( Q s )reg? is supported on D.
+
'fi<^W)7
LEMMA12.2. For each <^ E SC(L(Qs)l) there is a compact subset C of [if,( < ^ ) = 0 i f y does not belong to C.
Proof. For v in S define A to be the split component of the torus Tc. Let M y be the centralizer of A c in M . It is a Levi subgroup of M defined over Q O so , yii = II,,^Mc is a Levi S-subgroup of M , and Tsl f l L(Qs)reg contained in is L ( Q s )n Therefore, Corollary 11.3 tells us that for any 7 17L(Qs)reg and <^ g(L(Qs)), ' ,(+) = d(I" ( < ! i E ) . y e<=el-(a, s)
as.
s1
Now, 9,' n L(Qs)reg also contained in L(Q,)'. This means that is depends only on the function
^,,
,(<^)
on ritemp(L(Qs)l). therefore may identify <^ with a function on L(QS)l, We which we assume belongs to L(Qs)l). Fix e ^ OX,Then He maps the center of ES suqectively onto S). aL t . With this fact it is easy to show that the function
64
vanishes unless H&) belongs to a compact subset of e E s a L which depends only on supp^). On the other hand, we can assume that d ( # 0, so by Corollary 11.3 the map (11.5) is an isomorphism. As we observed in the preamble to the corollary, this implies that ker Hy^ = ker Ht n ker HlV. Now & is a subgroup of ORs. The kernel of He^ in (& is compact. Ysl is by definition just the kernel of Hi, in ^Ts. Therefore, the restriction of He to is a proper map. Thus the map
is supported on a bounded set, which depends only on supp ^). This proves the lemma.
LEMMA 12.3. For every compact subset C of L ( Q S ) l there is a bounded subset D of Ysln L(Qs)reg such that for any function f i n C ; ( L ( Q S ) l ) supported on C, the function
is supported on D. f ) , the invariant orbital integral of f . The If L = M, JM, ( f ) equals , . lemma in this case is a well known result of Harish-Chandra. The proof for arbitrary L is no different. We can now finish the proof of the theorem. /"(&( f ) ) equals
It follows from the last two lemmas and our induction hypothesis that if f is supported on C , f^&( f ) ) vanishes for y outside a fixed bounded subset of n L(Qs)reg.But n L ( Q s ) ^ is dense in f l M(Qs)reg.The theorem follows. 13. The invariant distributions l oand I x With the completion of the proof of Theorem 12.1 we have reached our goal. We have shown that for every S the maps
are continuous and satisfy (4.2). Assume that S contains the Archimedean valuation. Then there are invariant distributions { I t : o e a} and {I:: x X } on
JAMES ARTHUR
for any f e C?(L(Qs)l). As we observed in Section 4, we can think of I f and I: as invariant distributions on C?(L(A)l). However, we had better check that the final distributions are independent of S. Suppose that S' is a larger set of valuations, the disjoint union of S and S , . There is a natural injection of C 3 L ( Q S ) l )into C T L ( Q s , ) l ) The image, f', of a function f i n C x L ( Q s ) l )is the product of f . with the characteristic function of IIoEsKu. must verify that I ^ f ) equals We I3f'). By definition, J ^ f ) J*). = On the other hand we can map any ) whose value at function (f> in U L ( Q S ) l to the function in gc(L(QSr)l) is + ( T T ) if 7r1 is of class one and is 0 otherwise. It follows from (9.2) that for any M P ( M " ) image of &(f) in ~ , ( M ( Q ~ ,equals W). the )~) Now
It follows by induction on dim L that I t ( f ) = I t ( f ' ) . Similarly I:( f ) = I:(f'). Thus, the distributions I f and I: are independent of S. We can therefore regard and as distributions on $ ( L ( A ) l ) ,the direct . limit over all S of the spaces g ( L ( Q s ) l ) Notice that C& extends to a continuous map from C"L(A)l) to g,(M(A)l). In fact, Xp(7r), Ip{Tr, f ) , Rp^('n) and <^( f ) can all be defined directly for f CT(L(A)) and TT IItemp(M(A)). is Iff is the the restriction off to L(A)l, the value of <^( f l ) at a class in 13temp(M(A)1) integral over all TT in the associated orbit in IItelnp(M(A)) of
As they are defined, 1: and {L appear to depend on all the arbitrary choices we ), made in Section 1. We shall show that, as distributions on $ ( L ( A ) ~ they do not. We shall first consider changing the maximal compact subgroup K. Fix f C,(L(A) the moment, write ^(f , K ) for &( f ). We wish to study the For dependence on K . Given P E F ( M ) and TT E IItelnp( M(A)),define
66
JAMES ARTHUR
for 4 'XP(r), n Ap(A), m M(A) and k e K n L(A). Then y p ( r ) maps 'Xp(r) isomorphically onto a Hilbert space k p ( r ) of functions on L(A). k p ( r ) is what is usually taken for the underlying space of the induced representation. It is independent of K. Now
and
so the operator ip{r, f ) is independent of K . It does, however, depend on our choice of Haar measure on L(A). For suitable A and $I, (/tplp(^)(.)(x) is the product of convergent intertwining integrals
with some scalar normalizing factors. The integrals clearly do not depend on K . We will assume from now on that each scalar factor is also independent of K and that the normalized operators are independent of any choice of Haar measure. Then the only terms in our expression for $Ip(\,f , r , Po) that depend on K are the operators Notice that
+ -
YPi")YPirA)-: ')CPirA)
'XP(4.
for any (. e 'XP(rA). Now suppose that K is replaced by another admissible maximal compact subgroup K*. The Haar measures fixed in Section 1 were tied, via (1.1), to our choice of K. We must therefore take a different set of Haar measures on all our groups, subject only to the restrictions of Section 1. If the new Haar measure on L(A) differs from the old one by a factor p(L), the operator ip(r, ) will have to f
67
be replaced by i i . [ ~ ) i ~ { w. , For P E ^ ( M ) and x f) H*(x) a M ,associated to the decomposition in We also have the Hilbert spaces ~
L ( A ) = ^,(A)M(A)(K* f l L ( A ) ) .
T T
If
+ belongs to ~ ( T T ) ,
where
TT,
We can now calculate the analogue for K* of GP(A,f , T TPo). We need only , replace Y ~ ( T T ) ~ ~ byTthe ~right hand side of (13.1) in the formula above for ( T ) ~ GF(A,f , T T , Po), and then multiply by p ( L ) . The result is
To obtain & ( T , f , K*), we multiply this expression by O p ( A ) l , sum over P v ( M ) , and let A approach 0. Now Up(A, T TK * ) is independent of P and its , value at A = 0 is 1. Therefore & ( T , f , K * ) equals the trace of the operator
We can now argue exactly as in the proof of Lemma 8.3. Formulas (13.2) and (13.3) take the place of (8.4) and (8.3). Notice that {up(\, x, K*): P T L ( M ) }is an ( L ,M ) family. If Q e P ( M ) m E M p ( A ) , and
X,
K*).
It follows from this fact, (13.2) and (7.5) that if Po and P are groups in which are both contained in Q,
P(M)
TT,
JAMES ARTHUR
where Po can be taken to be any group in ^ ( M ) with Po C Q. The rest of the proof of Lemma 8.3 carries over without further difficulty. We obtain
where fQ
Next, suppose that I L is one of the distributions 1 or 1;. Identify f with its : restriction to L ( A ) l , and write
J L ( f 7K ) = I L ( f )
to denote the dependence on K. This dependence can be studied by transcribing almost verbatim the proof of Theorem 3.2. The result is the formula
J L ( f ,K * )
P(L)
c ( M Q ) c ( ~ )J Q ( ~ ~ , K )*. ,
-1
Q&Mo)
Now let I L ( . , K ) be the invariant distribution defined by Proposition 4.1. We claim that its Fourier transform, !(-, K ) , is independent of K . Indeed, arguing as in the proof of Proposition 4.1, we find that equals the sum over all Q e 6 > f L ( ~ ) , with Q # L , of the product of p ( L ) c ( M Q ) c ( L ) " l with
Now
#(fQ,K*,
It follows that the above expression vanishes. We have shown that IL( f , K*) = ^LVL(f7 K). Now suppose that instead of fixing f we fixed an element <f> in g^,(L(A)'). Take f to be any function such that d> = d>i( K ) . The second choice of Haar f,
69
measures will necessitate replacing f by p(L) I f . Therefore fL(+, K*). We have established
I^(+,K )
equals
PROPOSITION As distributions on $(L(A)'), I": and are independent 13.2. El of K. They are also independent of the Hoar measures chosen in Section 1.
In Section 1 we also fixed MO, a minimal Levi subgroup of G. The distributions I", and I " are independent of this choice as well. Any other minimal Levi subgroup equals y ' M 0 y , for some y in G(Q). Then y ' K is a maximal compact subgroup of G which is admissible relative to y ' M Oy . If L is in (MO) y ^ L y belongs to ( ^ M O ) . We can transfer all the Haar measures chosen in y Section 1 by conjugating by y l . It is a simple exercise to check that I ( f ) = I: l L y ( f y )and I t ( / ) =)I for f E C?(L(A)l). It follows that I", and I"^ are independent of MO. 14. An example We shall conclude our paper with a look at the example of inner twistings of G L . Special cases have recently been studied by Flath [3(a)] and by Deligne and Kazdan (unpublished). For simplicity of notation we will stick to GL.,(Q) although we could just as easily work, through restriction of scalars, with an arbitrary number field. Suppose that D is a division algebra of degree d over Q. Let G and G, be the general linear groups of ranks m and n = md over D and Q respectively. The local Langlands conjecture states that for every valuation v on Q there is an injection with certain properties. This would yield an injection
For every f E C:(G(A)l) we could then define a complex valued function <^ on I I t e l n p ( ~ l ( A )by letting $ 1 ( ~ 1 ) l) equal tr(77-(f)) if is the image of TT E litelnp(G(A)l), letting it equal 0 otherwise. The conditions on (14.1) are such and that should belong to $(G,(A)l). The ultimate goal would be to establish a correspondence between the automorphic representations of G and G, by comparing the identity
0 ~ 6 ,
fol,<^l) =
xesi
2 ^(+ii.
the invariant trace formula for G,, with the trace formula for G. We will attempt something more modest here. Now (9 and ( 9 , can be identified with the semisimple conjugacy classes in G(Q) and Gl(Q) respectively.
70
JAMES ARTHUR
The theory of division algebras gives an injection It is easy to describe the image; it is also easy to say what the image of the map
f belongs to the expected image of (14.3) but o e 0,does not should be. I belong to the image of (14.2), ((<^),) ought to vanish. We will prove this when the class o is unramified, in the sense of [l(d)]. First of all let us recall some notions for the general linear group over a fixed field E. The characteristic polynomial identifies the semisimple conjugacy classes in G L ( E ) with the polynomials in E[X] of degree n and nonzero constant term. Regular semisimple conjugacy classes correspond to polynomials with distinct roots. If o is a regular semisimple class we define a partition
of n from the degrees of the irreducible factors of the characteristic polynomial. A partition can also be defined for any Levi subgroup, M, of G L defined over E. It is the unique partition p(M) = (nl,...,nr), nl> . - . s n r , of n such that M is isomorphic to IILIGLn,.Note that r is the dimension of the space a n . We can partially order the partitions of n by setting p < p whenever there are Levi subgroups Ml C My of G L such that t> = ^(MI) and p = f ( My). Then p(o) 5 N M ) , for a given o and M, if and only if o intersects M(E). Notice that if ^ ( k ) = (k, k , . . . , k) for a given divisor k of n, t)(k) 5 ( n,, . . . , n,) if and only if k divides each integer n,.
14.1. Suppose that M is a Levi subgroup of GLn and that a$ fl LEMMA a 2 = {O} for groups Mi and My in (M(defined, of course, over E). Suppose also that p(k) <: @(MI)and p(k) 5 ^(Mg) for some divisor k of n. Then O(k) 5 @(MI.
<^),
Proof. The condition on Ml and M2 is equivalent to a M= a M a v . Now for i = 1,2, let @(Mi)= (nil, . . . , n i r ). , Then k divides each n,,, and Mi = 11% Mi,, where Mij is isomorphic to GLn,,. The intersection of a and a is a space of dimension at least one. Therefore
71
Now M is the subgroup of Mi defined by a subset of the simple roots of M i with respect to some ordering on a maximal split torus. It follows that
We claim that one of the groups Mii is contained in M. Assume the contrary, and suppose that rl 5 r2. For each 1, M n M2i will be a proper subgroup of M2,. The length of p(M) will be no less than 2r2. Since this must also equal dim a M ,we obtain a contradiction. We have shown that M2, C M for some i. Therefore, M2{C Mill for some h. Let n' = n - n2,. Then k divides n' and we have a partition ~ ' ( d of n'. There is clearly a unique subgroup G' of GLn, isomorphic ) to GLn,,such that M2 is contained in M2,G'. Then M' = M i l G', Mi = Ml n G' ; : and M = M2 f l G' are Levi subgroups of G' such that a f l 0%; = {O}. The L partition p(Mi) is obtained by replacing n l h by nlll - n2,. Since k divides ) Similarly ~ ' ( k 5 @(ML). follows by induction on n ) It nlh - n2,, ~ ' ( k5 @(Mi). that ~ ' ( k 5 p(M'). Since @(MI is obtained from p(M') by adjoining n2,, ) @ ( k ) @(MI. 5 0 For each valuation v on Q we have the invariant, invo(D), of D at v . It is an element in Q/Z, and
2 invu(D) = 0.
0
Let d u be the order of invo(D). Then d is the least common multiple of the integers {do}.The image of the map (14.1) should be the set of induced cuspidal representations
where Mu is a Levi subgroup of Gl defined over Qo such that @ ( d o 5 p(Mo). ) Said another way, the image of the map (14.3) will be
where ~ F ( G ~ ( Q ~is ' ) ) the space of functions in ~ ; ( G ~ ( Q ~ )such that ') G1, % = 0 for any Levi Ssubgroup 9R = IIoEsMo Gl for which the property of fails to hold. Next, suppose that o is an unramified class in a1; in the present situation this means a regular semisimple conjugacy class in Gl(Q). For each v , o generates a regular semisimple conjugacy class in Gl(Qo) = GLn(Qo),so we I obtain a partition & Jo) of n. It follows from the theory of division algebras that o is in the image of the map (14.2) if and only if @ ( d o ) @ J o ) for all v . 5
72
JAMES ARTHUR
THEOREM 14.2. Suppose that E $;(Gl(A)l) and that fo(+l)# 0 for a given unramified c h s o E Ol. Then o belongs to the image of the map (14.2). The Proo$ Let fl be any function in C:(G1(A)l) such that +(fl) results of [l(d), $81 allow us to express Jo(fl) as a weighted orbital integral. There are parabolic subgroups Po C Pl of Gl, with Levi components Mo C Ml, a11 . defined over Q, such that Po is minimal and v(M1) = ~ ( 0 )Let yl be a point in ( fl) equals J : Ml(Q) n o. If T is a suitably regular point in
r~i,
= -HpO(ws-l) =s
-
Choose a finite set S of valuations on Q, containing the Archimedean valuation and all the places at which D does not split, such that (i) fl belongs to C:(G1(Qs)l). (ii) lD(yl)Ic= 1 for all v not in S. (iii) {x E Gl(QL):x-lYlx E KL}= Gl(QL)yl.Ku all o not in S. for Then
Moreover, the integral over G(AIYl\G(A) above can be taken over G(QS),,> G(Qs). If we identify yl with its image in Ml(Qs), we obtain Jo(fl) = V O ~ ( A . ~ ~ ( R ) \ G (G ( )Q )~ ~ ~ ~ A )JMl,yl(fl). ~ ~ It follows from the definitions of Section 4 that (14.4) ) ) Y G( ) f o b l ) z ~ O ~ ( A . ~ ~ ( R ) ~ - ~ ( Qf M~l ,~ \l ( @ l A+) ~ ~
73
We have not used any special properties of GL,, in deriving this formula. It holds for any reductive group. Our theorem will now be proved by combining the corollaries of Theorem 11.1 with Lemma 14.1. The first corollary states that
so that
and Fix v E S. For any prime p , let pr be the highest power of p which divides d o . Since the invariants of D sum to 0, there must be a w E S , distinct from v , such p that pr divides d w . Therefore @( r ) 5 @ ( L o )and p( p r ) 5 @ ( L w ) .Since a n a L c = {O}, we can apply Lemma 14.1. We see that p ( p r ) 5 @ ( M I ) .It follows MI that P ( d " ) 5 P(M1). We have identified yl with its image in M l ( Q s ) , so we shall write
2,
For each v , choose a Levi subgroup Mo of G l , defined over Q",with yo E Me C M l , and p ( M c ) = p " ( 0 ) .Let 9 , 2= M u . By Corollary 11.3, fAfl, + l ) equals the yj sum over all !2 in !2(%, S ) of This summand will be nonzero for some e. The conditions for the nonvanishing imply that of d( C ) and and for all v
E
P@")
@ ( M u )= @ " ( o )
74
JAMES ARTHUR
for all E S. Since S contains all valuations for which d o > 1, o belongs to the image of (14.2). Our theorem is proved. 0
[1] J. ARTHUR, (a) A theorem on the Schwartz space of a reductive Lie group, Proc. Nat. Acad. Sci. 72 (19751, 4718-4719. (b) The characters of discrete series as orbital integrals, Inv. Math. 32 (19761, 205-261. (c) Eisenstein series and the trace formula, in Autmmphic F m , Representations and Lfunctions, Proc. Sympos. Pure Math., Vol. 33, 1979. (d) A trace formula for reductive groups I: terms associated to classes in G ( Q ) ,Duke Math. J. 45 (19781, 911-952. (e) A trace formula for reductive groups 11: applications of a truncation operator, Comp. Math. 40 (19801, 87- 121. ( On the invariant distributions associated to weighted orbital integrals, preprint. 0 [2] I. N. BEWSTEIN A. V. ZELEVINSKY, AND Induced representations of reductive b-adic groups, I, Ann. Scient. Ec. Norm. Sup. 10 (19771, 441-472. [3] D. FLATH, (a) A comparison of the automorphic representations of GL(3) and its twisted forms, Thesis, Harvard University, 1977. (b) Decomposition of representations into tensor products, in Autmorphic F m s , Representations and Lfunctions, Proc. Sympos. Pure Math., Vol. 33, 1979. [4] S. GELBART H. JACQUET, and Forms on GL(2) from the analytic point of view, in Autmorphic F m , Representatiuns, and LFunctions, Proc. Sympos. Pure Math., Vol. 33, 1979. [5] H~ISH-CHANDRA, (a) Discrete series for semisimple Lie groups 11, Acta Math. 116 (1966), 1-111. (b) Harmonic analysis on reductive padic groups, in Harmonic Analysis on Hmtogeneous Spaces, Proc. Sympos. Pure Math., Vol. 26, 1973, (c) The Plancherel formula for reductive padic groups, preprint. [6] H. JACQUET, Generic representations, in Springer Lecture Notes, Vol. 587, 1977. [7] H. JACQUET R. P. LANGLANDS, and Automorphic Fmns on CL(2), Springer Lecture Notes, Vol. 114, 1970. [8] A. W. KNAPP,Commutativity of intertwining operators 11, Bull. A. M. S. 82 (19761, 271-273. [9] J-P. LABESSE R. P. LANGLANDS, and Lindistinguishability for SL(2), Can. J. Math. 31 (19791, 726-785. [lo] R. P. LANGLANDS, (a) On the Functional Equations Satisfied by Eisenstein Suies, Springer Lecture Notes, Vol. 544, 1976. (b) Base Change fm GL(2), Annals of Math. Studies, 1980. [ l l ] F. SHAHIDI, certain Lfunctions, preprint. On Introduction to Harmonic Analysis on Reductiue padic Groups, Mathematical [12] A. SILBERGER, Notes, Princeton University Press, 1980. (Received June 14, 1979) (Revised October 29, 1980)