COMPOSITIO MATHEMATICA, Vol. 40, Fasc. 1, 1980, pag.

87-121 1980 Sijthoff & Noordhoff International Publishers - Alphen aan den Rijn Printed in the Netherlands

A TRACE FORMULA FOR REDUCTIVE GROUPS II: APPLICATIONS OF A TRUNCATION OPERATOR

James Arthur*

1. 2. 3. 4.

A truncation operator ............. ............ Integrability of k(x,) .. The operator Mp(7r) .... . Evaluation in a special case 5. Conclusion .
..........

89 98

107
113 120

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

Introduction

This paper, as promised in the introduction to [l(c)], contains an identity which is valid for any reductive group G over 0, and which generalizes the Selberg trace formula for anisotropic G. We have already shown that a certain sum of distributions on G(A)', indexed by equivalence classes in G(Q), equals the integral of the function
xEt

kT(X, f),

E G(Q)\G(A)1.

The main task of this paper is to show that the integral may be taken inside the sum over X. There does not seem to be any easy way to do this. We are forced to proceed indirectly by first defining and studying a truncation operator A on functions on G(O)\G(A)'. Recall that kT(x,f) was obtained by modifying the function Kx(x, x). We shall apply the results of 1 to the function

ATA TK (x, x),

xE

G(Q)\G(A)1,

*Partially supported by NSF Grant MCS77-0918.

0010-437X/80/01/0087-36 $00.20/0
87

88

J. Arthur

[21

obtained from K,(x, y) by truncating in each variable setting x = y. It will turn out that the function
xE

separately, and

jA Af2 K.(x, x)|

T

is

integrable. Then in 2, our main chapter, we shall show that for sufficiently regular,
XET

(O)tG(A)1

(A TA 2Kx(x, x) - kT(x, f)) dx

converges absolutely. We shall also show that for each X, the integral over G(Q)\G(A)' equals 0. If we set JT(f) equal to
JG (O)\G( A)1

A ITA TK(x, x) dx

=
G (Q)\G( A)1

kT(x f) dx,

the

identity associated to G is then

= =ZJ f). EJATE) ECH

XE

We should note that the distributions JT and J are not in general invariant. Moreover, they depend on a choice of maximal compact subgroup and minimal parabolic subgroup. However, it should be possible to modify each of the distributions so that they are invariant and independent of these choices, and so that the identity still holds. We hope to do this in a future paper. Both formulas for JT (f) are likely to be useful. The integral on the right is particularly suited to evaluating JT on the function obtained by subtracting f from a conjugate of itself by a given element in G(A)'. It can also be used to show that JT(f) is a polynomial function in T. We shall not discuss these questions here. On the other hand, the integral on the left can be calculated explicitly if the class X is unramified. We do this in 4. The result follows from a formula, announced by Langlands in [4(a)], for the inner product of two truncated Eisenstein series. It was by examining Langlands' method for truncating Eisenstein series that I was led to the definition of the operator A T

[3]

A trace formula for reductive groups II

89

1. A truncation operator

Let G be a reductive algebraic group defined over Q. We adopt the definitions and notation of [l(c)]. In particular, K is a maximal compact subgroup of G(A) and Po is a fixed minimal parabolic subgroup of G defined over Q. Again we shall use the term 'parabolic subgroup' for a parabolic subgroup P of G, defined over 0, which contains Po. We would like to prove that the terms on the right hand side of the identity given in Proposition 5.3 of [l(c)] are integrable functions of x. To this end, we shall introduce a truncation operator for functions on G(Q)\G(A)1. Recall that T is a fixed, suitably regular point in a . If 4 is a continuous function on G(Q)\G(A)1, define (A T)(x) to be the function
(- i1)dim(A/Z) 8
G IN()\(A

P8 5P(Q)\G(Q) N(Q)\N(A) E

p4(nSx) rp(H(8x) T).

-

(the sum over P is of course over all parabolic subgroups.) Note the similarity with our definitions of the functions k (x, f) and kT(x, f) in [l(c)]. If 4 is a cusp form, A T = 0. It is a consequence of [l(c), Corollary 5.2] that if +(x) is slowly increasing, in the sense that

|+(X)| ' CllxlN",

for some C and N, then so is

ATO(x).

LEMMA 1.1: Fix P1. Then for

fE C(G(Q)\G(A)'),
=0

|fv
unless

N(Q)\Ni(A)

nT(nlx) dn1 AT
E A1.

r(Ho(x) - T) < 0 for each

PROOF: For any P, let D2(ao;P) be the set of s E such that s-la >0 for each a E AP. Applying the Bruhat decomposition to P(Q)\G(O), we find that fN,(o)\N,(^) AT (nlx) dn1 equals the sum over P and s E f(ao;P) of the integral over n in N(Q)\N(A) of the

product of (-l)dim(A/Z) with

Ni(Q)\Ni(A) vEws No(OQ)ws nNoQ)\NO)

f|l()\Av'N0 f2lN(0, C(nwsvnlx). rp(H(wsvn x)

T) dnl.

90

J. Arthur

[41

Since Ni(Q)\Ni(A)= No(Q)\No(Q)Nl(A), this last

Jw' NO(Q)w, nNO()\NiQ)N,(A)

expression equals

nN ;N0o,
w

(^(nwsnx)fp(H(wsnx)- T)dnl.

Decompose

-No(Q)w, n No(Q)\No(G)Ni(A) as (w-'No(O)ws n No(Q)\w-'No(A)ws nNN(Q)N,(A)) (wNo(A)ws n N(QO)Nl(A)\No(Q)Ni(A)) = (w-'No(Q)ws n N1(Q)\w-'No(A)w~ n Nl(A))
x

x(w -No(A)ws n No(Q)Ni(A)\No(Q)N (A)).

This induces write
a

decomposition

of the measure dn1 as

n*= n s ni*,

dn, dn*.

Then

wsnn* = wsnW
and

finally, combine the integral over hn with the integral over n in N(Q)\N(A). Because s lies in 2(a0; P), Non wsNlw n M is the unipotent radical of a standard parabolic subgroup of M. It follows
that

(No n wsN, w'n M)N = Ns

is the unipotent radical of a uniquely determined parabolic subgroup Ps of G, which is contained in P. We have shown that

fN(O\NI(A) (n1x) dn 1 equals

dn (_(dAP 1)dim(A/Z) sEDR(ao; P) wsNo(A)wsnN&O)NI(A)\N(O)Ni(A) | 4 (nwsn*x)p(H(wsn*x)- T)dn.

sN(Q)\Ns(A)

change the order of summation, and consider the set of P which give rise to a fixed Ps. Fix s E Q. Define S' (resp. S1) to be the set of a Eo such that s-'a is a positive root which is orthogonal (resp. not orthogonal) to a1. If P, is one of the groups that appear in the above formula, A' will be a subset of S'. Those P which give rise to a fixed P, are exactly the groups for which AoP is the union of Ao and a subset S of S1. Thus, for fixed s with ,A C S', we will obtain an alternating sum over S C SI of the corresponding functions ip. We
We shall

[5]

A trace formula for reductive groups II

91

apply Proposition 1.1 of [l(c)]. Let Xs be the characteristic function of the set of H E oa such that for a E AO- A U Si, wa(H) >0, while mja(H) < 0 for a in SI. Here ma is the element in Ao corresponding to a. Then fN,(O)N,(A) 4(nlx) dnl is sum over s E f1 and over all subsets A4 of S, of the integral over n* in w'sNo(A)ws n NN(O)N,(A)\ NI(O)N,(A) and n in Ns(O)\Ns(A) of the product of
1(nwsn*x)Xs(Ho(wsn*x) - T)
with -1 raised to a power equal to the number of roots in Ao- S' U Sl. Suppose that for some s, Xs(Ho(wsn*v)- T) does not vanish. Then if

Hc(wsn*x)-T)= a2 t, t E R, EA0
ta is positive for a in A0- A U S, and is not positive for a E Sl. If
_w

E=l,
(s-'(Ho(wsn*x) - T))
=
> a EAO

ta(S(s-la1)

EAo\S1

E t.m(s-la),
a

where s-la is orthogonal to a' if less than or equal to 0. Now

E S'. This last number is

clearly

s-l(Ho(wn*x) - T) = Ho(x) - T + s-Ho(wsvwsl) + (T- s-IT),

for some element

E Ao, it is well known that m(s-'Ho(wsvw-1) is nonnegative and w (T - s-i) is strictly positive. D Therefore m (Ho(x) - T) is negative for any w E al. From the definition of AT we obtain
v

E No(A). If

COROLLARY 1.2: ATAT =AT.

LEMMA 1.3: Suppose that 41 and 42 are continuous functions on G(O)\G(A)'. Assume that 41 is slowly increasing, and that 42 is rapidly decreasing, in the sense that for any N, the function IlxIN I42(x)l is bounded on any Siegel set. Then
'

(A Ti, 42) = (41, A T2).

92

J. Arthur

[61

PROOF: The inner product (ATb1, )2) is defined by an convergent integral. It equals
2 |fG(O)\()IP (-l)dim(A/Z) 8EP(Q)\G(Q) (Q)\G(A)1

absolutely

f1NNA l(nxx)'rp(H(Sx)- T)b2(x) dn dx

N(Q)\N(A)

2 (- 1)dim(A/Z)

fP

l(nx)4)2(x)ip(H(x)- T) dx dn
-

=2
P

f (_ 1)dim() fN(Q)\N(A) JP(Q)\G(A)1^) l(x)2(nx)fp(H(x)

(l1, A T02).

T) dx dn.
[

This last expression reduces to

REMARK: It can be shown that AT extends to an orthogonal projection on L2(G(Q)\G(A)'). We would like to show that under suitable conditions, AT+P(x) is rapidly decreasing at infinity. The argument begins the same way as the proofs of Theorems 7.1 and 8.1 of [l(c)]. Suppose <4 is a continuous function on G(Q)\G(A)1. Apply Lemma 6.4 as in the beginning of the proof of Theorem 7.1 of [l(c)]. We find that AT4(x) is the sum over {P1, P2: Po C P1 C P2} and 8 E Pi(Q)\G(Q), of

F'(Sx, T)2l(Ho(Sx)- T)0pp2(Sx),

where

pP2(y)= {P: PICPCPI (-1)dim(AZ) 2

N(Q)\N(I)

(ny) dn.

For the moment, fix 8 and x. We regard 8 as an element in G(Q) which we are free to left multiply by an element in Pi(O). We can therefore assume, as in [l(c), 7] that 8x = n*n*mak,

where k E K, n*, n*, and m belong to fixed compact subsets of N2(A), N2(A) and MI(A)' respectively, and a is an element in Ai(R)

[7]

A trace formula for reductive

groups II

93

with oa(Ho(a)- T) O. Therefore

PIP2(8X) 4p,p,2(n*n*mak) = P ,P2(n*mak)

=

=
=

ppp2(aa-ln*mak) (>pP2(ac),

where c belongs to a fixed compact subset of G(A)1 which depends only on G. The function 4P,,p2 resembles the function estimated in the corollary of [3, Lemma 10]. We want a slightly different statement of the estimate, however, so we had best re-examine the proof. If a E A2, let 4 Pa, Pi C Pa C P2, be the parabolic subgroup such that A ? = A is the complement of a in A4. For each a, let {Yai,... Y,n} be a basis of n2(Q), the Lie algebra of N2(Q). We shall assume that the basis is compatible with the action of Al, so that each Ya,i is a root vector corresponding to the root Pa,i of (M2 n P1, Al). We shall also assume that if i j, the height of Pa,i is not less than the height of 3,j. Define naj, 0 j < n,, to be the direct sum of {Yai,..., Y,,j} with the Lie algebra of N2, and let Na,j = exp na,j. Then Naij is a normal subgroup of Ni which is defined over 0. If V is any subgroup of N1, defined over 0, let 7r(V) be the operator which sends b to
-

tf
Then

V(o)\V(A)

4b(ny) dn, y E G(A).

pp,2 is the transform of 4 by the product over a E A2 of the

nta

operators

7T(N2) - 1T(Na) = 2 i=1

T(N.,i-) T(Nai).
-

If Ko is an open compact subgroup of G(Af), G(Q)\G(A)l/Ko is differentiable manifold. We assume from now on that 4 is a function on this space which is differentiable of sufficiently high order. Suppose that I is a collection of indices

a {i: a Al, l ia na}.

Then

NI = Na,i-i

94

J. Arthur

[81

and

N.
are normal

H NCi,
ot

subgroups of N1. Let n' be the span of {Y,,i} and let

=

n'(O)' be the set of elements

aYai,, r

rae * C

Then if n is

any positive integer,

n

_ (r

is a nonzero real number. By the Fourier inversion formula for the group A/Q, p,,p,(y) is the sum over all I of

f ?E tnlO)
4

\l(Q A)

dx

du
JN ~~(Q)\N~I(^)

(ue(X)y)4((X, 4)).
product defined

Here e and $i are as in [l(c), 7] and (,) is the inner by our basis on n'. If n is a positive integer,
n

H (_-

ai)l
Op,,,(y) equals

can be regarded as an element in I(q(R)' C). Then the sum over I and over E n'(Q)' of

~~~(1.1) (f~ | l(O(1.1) dX|

A)
f

du

N(Q)\N7(A)

Ry(Ad(y-l)Yn )4(ue(X)y)((X, )).

Now, we set
y = 8x = ac,
above. Since o2(Ho(a)- T) 0, a belongs to a fixed Siegel set in M2(A). It follows that the integrand in (1.1), as a function of X, is invariant by an open compact subgroup of n'(Af) which is inas

[9]

A trace formula for reductive groups II

95

dependent of a and c. Consequently, (1.1) vanishes unless ~ belongs to a fixed lattice, L'(Ko), in n'(R). But for n sufficiently large
fEn (o)'nL (Ko)

is finite for all I. Let c,(Ko) be the supremum over all I of these numbers. Then IP4,,p2(ac)l is bounded by

Cn(Ko) 2 f \ I
Let

N(Q)\Nj(A)

(R(Ad(c)- Ad(a)-' Y?)4)(uac)\ du.

pB

= Za 13a,i. Then

PB

is a positive sum of roots in A . For any n,

Y= Ad(a -l)Y e-"n(Ho(a)) Y? = e-nI(Ho(Sx))y.

We can choose a finite set of elements {Xi} in t(q(R)' ing only on n and Ko, such that for any PI, P2, I and c,

QC), depend-

Cn(Ko) Ad(c)-l Y
is a linear combination of {X,}. Since c lies in a compact set, we may assume that each of the coefficients has absolute value less than 1. We have thus far shown that IA TO(x) is bounded by the sum over all Pi, P2 and 8 E Pi(Q)\G(Q) of the product of

F'(Sx, T)ro-(Ho(Sx)- T)
with

(1.2)

2 I

,
i

Ni(Q)\N()

j R(Xj)(u8x)I du*

e-""nI(O))

LEMMA 1.4: Let 6 be a Siegel set in G(A)'. For any pair of positive integers N' and N, and any open compact subgroup Ko of G(Af), we can choose a finite subset {X5} of tI(g(R)l (0 C) and a positive integer r which satisfy the following property: Suppose that (S, do) is a measure space and that f(oa, x) is a measurable function from S to Cr(G(Q)\G(A)I/Ko). Then for any x E 6,

1A T(crO,x)l do-

96

J. Arthur

[10]

is bounded

by
E yGc(A)1 sup

(f

R(Xi)(o, y) da * Iyll-N) * Ix.

b in

PROOF: Substitute result is

o(a) for
S

(1.2) and integrate over o. The

(1.3)

2 I

( N(0)\N(A)

iR(Xi)4(o, u8x)l do du e-n"(Ho(x". above,

If 5x = ac, with a and c as

118xll-< 11all -11

l.

We are assuming that a2(Ho(a) - T) $ 0. Since fi3 is a positive sum of roots in A2 we conclude from [l(c), Corollary 6.2] that I|all is bounded by a fixed power of eI(Ho(a)) = e I(Ho0(x))

It follows that for any positive integers N and N1 we so that (1.3) is bounded by a constant multiple of
sup E yeG(A)lY i

may choose

IR(Xi)(o-, )ly da- ix K) IXI-NI. '

[2]) that there is a constant cl such that for any

It is well known (see y E G(Q) and x E ,

|iiYXi-N
The

.< C111XII-N,

only thing left to estimate is

5EP1(0)\G(Q)

F x(Sx, T)oa(He(Sx)- T).

The summand is the characteristic function, evaluated at 8x, of a certain subset of

{y E G(A)': w(Ho(y) - T) > 0, r E al.

The sum is bounded

by
P 6EPl(O)\G(O)

,(Ho(Sx)- T).

[11]

A trace formula for reductive

groups II

97

It follows from [l(c), Lemma 5.1] that we can find constants C2 and N2 such that for all P1 this last expression is bounded by C2xllIN2. Set N1 = N' + N2. N1 dictates our choice of n, from which we obtain the differential operators {Xi}. The theorem follows with any r greater than all the degrees of the operators Xi. o
In the next section we will need to have analogues of the operators AT for different parabolic subgroups of G. If PI is a parabolic subgroup, and 4 is a continuous function on Pl(Q)\G(A)', define
ATPI(x)=
{R:P0CRCP1}

Z (_ l)dim(AR/IA) 5ER(O)\PI(O)

x 1I

4(nSx) dn TR(Ho(x) - T).

NR(O)\NR(A^)

LEMMA 1.5: Suppose that P is a parabolic continuous function on P(Q)\G(A)'. Then

{Pj: PoCPiCP} 5EPi(Q)\P(O)

subgroup and

is a

0(Sx)rv (Ho(Sx)(nx) dx.

T)

equals
(1.4)

|fN(Q)\N(A)

PROOF: We need to prove that (1.4) is the sum over {R: Po C R C P} and 8 E R(Q)\P(Q) of the product of

N) NR(Q)\NR(A) .$)(nSx)
with

dn

(1.5)

{PI: RCPICP}

T(o( (- 1)dim(ARIA1) I(Ho(Sx) T) *(Ho(SX)-

Consider Lemma 6.3 of [l(c)], with A a point in -(a*)+. The sum given in that lemma then reduces to (1.5). It follows from [l(c), Prop. 1.1] that (1.5) vanishes if R. P and equals 1 if R = P. This establishes
Lemma 1.5. 0

98

J. Arthur

112]

2.

Integrability of kT(x, f)

We take r to be a sufficiently large integer, and continue to let T be suitably regular point in ao. In [l(c)] we associated to every f E C'(G(A)') a function, k(x, f), on G(Q)\G(A)1.

THEOREM 2.1: For sufficiently

regular T,

ve^ G(O)\G(A)1

E|Yl Ik T(x, f)l dx

is finite.
We will not prove the theorem directly. Rather, we shall relate kT(x,f) to the truncation operators whose asymptotic properties we have just studied. We shall operate on Kp,(x, y), which of course is a function of two variables. If PI C P2, we shall write A T'Pl (resp. A r.PI) for the operator A TP,, acting on the first (resp. second) variable. LEMMA 2.2: For any X E X, k T(x, f)
E E

equals

{PI,P2: PoCP1CPZ) 5EPi(O)\G(Q)

X

acr(Ho(x) - T)

l {P: PCPCP2} (-1)dim(AZ)A TP Kp,x(SX, SX)}. ,

PROOF: The given expression is the sum over all chains P1 C P C P2 C P3 and over 8 E Pi(QC)\G(Q), of

(-1)dA3A2)3(HO()
expression equals

T)3 (Ho (Sx)

T)(-1)dim('AZ)A

P'Kp.X(x, 8x).

As we have done many times, we appeal to [l(c), Prop. 1.1]. We see that the sum over P2 equals 0 unless P = P3. Therefore the given

{Pi,P: PoCP1CP}
* P

(- l)dim(A/z)

5EPi(O)\G(O)

p (Ho(Sx)- T) p

rp(Ho(Sx)- T) A'TPKp, (x, Sx). 21K.(X,~)

[13]

A trace formula for reductive

groups II

99

Apply Lemma 1.5 to the sum over Pi. We obtain

{P: PoCP}

(_ I)dim(A/Z) 8EP(O)\G(O) rp(Ho(SX)- T) 2

* |fNoN) Kp,(Sx, nSx) dn. N(0)\N(A)

Since

Kp,(8x, n8x) = Kpx(Sx, 8x),

this last expression equals Fix Pi C P2. Motivated

E N(A),

kx(x, f), as required.

by the last lemma,
we shall examine the

expression
{P: PiCPCP2},

2 (_(-1)dim(A/A:2)

N(Q)\NI(A)

Kp(x, ny) dn.

It

equals
{P: PiCPCP2}

(- l)dim(A/A2)
=
P

7EPi(O)\P(O) iN(A) TEMi(O)

2 f(x-'y- y) dn

(_ l)dim(A/A2) yEPl(Q)\P(O) Kp,(yX, y).

F(Pi, P2) be the set of elements in P1(O)\P2(Q) which do not belong to Pi(Q)\P(Q) for any P, with P1 C P CP2. By [1(c), Prop. 1.1] the above expression equals
Let
yEF(PI, P2)

Kp,(yx, y). implicit in the

In this last formula we have affected the cancellation

alternating sum over P. In order to exploit the equation we have just derived, we interrupt with a lemma.
Suppose that for each i, 1_ i < n, we are given a parabolic subgroup Qi D P1, points xi, y, E G(A) and a number ci such
LEMMA 2.3:

that

2 ci
i=1

Nl(O)\Ni(A)

KQ.(Xi, nmyi) dn

100

J. Arthur

114)

vanishes for all m E Mi(Q)\M,(A)'. Then for

n

any X E f,

z hX(m) Ci=C) Nl(O)\Nl(A) KQ,.X(xi, nmyv) dn i=1

also vanishes for all m E M1(Q)\M,(A)'.

PROOF: Suppose that for a given X' E A, there is a group R in P~ which is contained in P1. We would like to prove that for any function x, E L2(MR(Q)\MR(A)1)X,, the integral

(2.1)

I MR(O)\MR(A) fNOM)\N hx(nm)+,'(m) dn dm

vanishes for X. X'. Suppose that X. X', and that E QO(7r)v for some Q C Qj, and some ir E H(MQ). The construction of Eisenstein series is such that if the function
mNl(O)\Nl(A)

EQ,(nmy, <) dn,

m E Mi(Q)\Mi(A)',

is substituted for h. in (2.1), the result is 0. It follows from the estimates of [l(c), 4] that (2.1) itself is 0. The same estimates yield constants c and N such that

h,x(m)l < c||m|rN,

mE

M,(Q)\M,(A)'.

By assumption, Exh(m) equals 0. Consequently (2.1) is zero even when = X'. The function hx is continuous. Because (2.1) vanishes for all X, hX satisfies the hypotheses of [4(b), Lemma 3.7]. h, is
therefore zero.

To return to the proof of the theorem, we look for conditions imposed on x, y and y by the nonvanishing of

(2.2)
Set
y=

Kp,(yx, my),

m E Mi(Q)\Mi(A)'.

y1k,

yi E Pi(A) n G(A),' k E K.

There is a compact subset of

G(A)', depending only on the support of

[15]

A trace formula for reductive groups II 101

f, which contains some point

x-'^y-lnrmyx,

n E N(A), rl E Mi(Q),

whenever (2.2) does not vanish. Fix CE l1 and let A be a rational representation of G with highest weight de, d > 0. Choose a height function 1 1 as in [l(c), 1]. If v is a highest weight vector, we can choose a constant c, such that

IlA(x-'yn-rmy,)vll

<

cl

whenever x-'ynrlmyl lies in the given compact subset of left side of this inequality equals

G(A)1. The

edw(HoYAl))JA(x-'l-')vUJ = edw(Hoy))'IA(x- y-')vll,

which is no less than a constant

multiple of

edm(HO(y)) e -d(Ho(yx))
In other words, w (Ho(yx) - Ho(y)) is no less than a fixed constant. It follows from this observation that we may choose a point ToE ao, depending only on the support of f, such that

(2.3)

Ti(Ho(yx)- Ho(y)- To) = 1

whenever (2.2) does not vanish identically in m. We conclude from Lemma 2.3 that if (2.3) fails to hold for a given x, y and y, then

(2.4)

Kp,,x(yx, my),

m E Mi(Q)\Mi(A)',

vanishes for all X and m. Combining [l(c), Lemma 5.1] with what we have conclude that for fixed x and y,

just shown,

we

Kp,,x(yx, y),

yE

F(Pi, P2),

vanishes unless y belongs to a finite subset of F(Pi, P2), of X. Therefore the sums in
{P: PiCPCP2}

independent

(- 1)dim(A/A2)

N(Q)\NI(A)

Kp(x, nmy) dn

yEF(Pi,P2)

~ Kp((yx, my)

102

J. Arthur

[161

are finite. Since the expression vanishes for all m in Mi(Q)\Mi(A)', we can apply Lemma 2.3. We obtain an equality of functions of y for each X. We are certainly at liberty to apply our truncation operator to those functions. It follows that for any X E S,

{P: PICPCP2

(-l)dim(A/Z)

TPKpoX(X, y)
,

equals

(_l)dim(Az) yEF(PI,P) A?' IKp,,X(yX y).

We have thus far shown that
| 2f (Q)\G(^A) IkT(x, f)l dx
X

is bounded

by the sum over {P1, P2: Po C Pi C P2} of

JPl(Q)\G(A)1

2 yEF(PI,P2) cr(Ho(x) - T)A PV'Kp,x(yx, x)l dx. 2 X

Let 6 be a fixed Siegel set in Mi(A)' with Mi(Q)6 = Mi(A)', and let F be a compact subset of NI(A) with NI(O)F = N1(A). Then the last Po(A), a E integral is bounded by the integral over n E F, m E t nl A,(R) n G(A)1, and k E K, of

e-2pp i(a))o2(HO(a) - T) 2
x Y

IA'PKp,,(ynmak, mak)I. 2

Suppose that for n, m, a and k as above, and for some rh E Ml(A)l' y EF(Pi, P2) and X E

Kp,,,(ynmak, rhak) # 0.
Write y = vwrr, for v E N0(Q), ir E PO(Q) and s E QM2, the Weyl group of (M2, Ao). It follows from Lemma 2.3 that there is a fixed compact subset of G(A)' which contains
a- m-lnl wspla,

for points

nl E No(A) and pi E M1(A)'NI(A). Fix w E d, and let A and

[17]
v

A trace formula for reductive groups II

103

be as above. A(ws)v is a weight vector, with weight si. The vector

A(a-lm-lnlwspla)v ed('-s"XHo(a)) e-dsw(Ho(m))v

-

can be written as a sum of weight vectors, with sw. By the construction of our height function,

weights higher than

ed(w-swxHo(a)) e-ds(Hi(m))

vI l IIA(a-lm-'n1 wspla)vl.
depending only on the

It follows that there are constants c' and c, support of f, such that
< <

K(w - sm)Ho(a))l c'Ism(Ho(m))| c(l + log|lmll).

Since s fixes a2 pointwise, the inequality

I(w -s-)(Ho(a))| c c(l + logllmll)

holds for the projection of w onto a2. In other words, we may take X to be an element in A2. For each such A, w -sm is a nonnegative integral sum of roots in A2. We claim that the coefficient of the element a in A2, such that w = wa, is not zero. Otherwise we would have (a - s )(m ') = 0, or equivalently, sx = . This would force s to belong to QM, for some parabolic subgroup P, P C P CP2. This contradicts the assumption that y = vwsr belongs to F(P1, P2), so the coefficient of a is indeed positive. We can assume that a has the additional property that

I2(Ho(a) T) $ 0.
-

| I on a0 there is a constant c such that

(2.5)
<

It follows from Corollary 6.2 of

[l(c)] that for any Euclidean norm

Il(Ho(a)Il c(1 + logirlII).

We have shown that if a E A1(R) n G(A)1 is such that for some X, y, n, m, mi and k,

(2.6)

fol2(Ho(a)- T)Kp,,x(ynma, rhak)l

compact subgroup

does not vanish, then the inequality (2.5) holds. Suppose that f is right invariant under an open

104

J. Arthur

[181

Ko of G(Af). Then if Ip,(r,f)4o 0 for some rT and f E Ap,(7r)X, the function E(y, d) is right Ko-invariant in y. Therefore for any x, y and X, Kp,,(yx, y) is right Ko-invariant in y. It follows that (2.6) is right invariant in mi under the open compact subgroup
kCK

n (kIkok-1) n M,(Af)l

of Mi(Af)1. We apply Lemma 1.4 with the group G replaced by M1. For any positive integers N1 and N' we can choose a finite set {Xi} of elements in l1(mI(R)'lC), the universal enveloping algebra of the complexification of the Lie algebra of Mi(R)', such that for all n E F, m E in Po(A), r E 6, a E Ai(R) n G(A)1 and k E K,

(2.7)
is bounded

yEF(PiP2) X

lA P'KP, ,(ynmak, mak)l

by
sup

(2.8)

C
i

ue Ml()

E)(

(Xi)Kp1(ynmak, uak)|l \IuL-N) I,mh-.

qU(q(R)' C) such that

We can choose elements {YJ} in

Ad(ak)-Xi = Ad(k)-'X, =
where

c,(k) Yj,

cij(k) are continuous functions on K. Recall that Kp,,x(x, y) is ultimately defined in terms of f. The function Ry(Yj)Kp,,(x, y) is defined the same way, but with f replaced by f* Y*. The support of f* Y* is contained in the support of f, so we can assume that (2.3) is valid whenever Ry( Y)Kp,,(yx, y) does not vanish. By Corollary 4.6 of
[l(c)],

E|Ry(Yj)Kp,,,(x, y)J x
is bounded by a constant multiple of a power of from Corollary 5.2 of [l(c)] that the expression
yEF(PI,P2)
=

|xll| Ilyll. It follows

IRRy( Yj)Kp,.x(yx, y)l

*

yEF(PI,P2) x

IR2y(Yi)Kp,x((yx, y)i C

i(H(Ho(y(x)- To) Hoy)-

[19]

A trace formula for reductive groups II

105

is also bounded by a constant multiple of a power of l||xl IlyIl. By taking N1 to be large enough we obtain constants C2 and N2 such (2.8), and therefore (2.7), is bounded by

m C2Im IINjIaN21211r- N .
Set mh = m in (2.7). Integrate the resulting expression over n E F, m E nP(A), k EK and a in the subset of elements in Al(R)n G(A)' which satisfy (2.5). There are constants C3 and N3 such that the result is bounded by

C3

lmllN3-N

dm.

If we set N; = N3, this is finite. The proof of Theorem 2.1 is complete.

LEMMA 2.4: For T sufficiently
G(O)\G(A)1

regular, and r sufficiently large,

G(O)\G(A)

kfCT(X, ) dx =
.

A K (x, x) dx,

for all f E Cc(G(A)1) and X E

PROOF: It follows from the proof of Theorem 2.1 that the integral of kT(x,f) is the sum over all P C P2 of the product of (-1)dim(A2/Z) with

I(o)tG(Ab)1 yEF(PI,P2) o2(Ho(x) T) 2 Pl(Q)\G(A)

A T'P1Kp1,x(yx, x) dx.

As a double integral over x and y this converges absolutely. If P1 = P2 # G, the integrand is zero. If Pi = P2 = G, the result is the integral of A2Kx(x,x). We have only to show that if PI CP2, the result is zero. Let f2(P1, P2) be the set of elements s in QM2 such that sa and s-'a are positive roots for each a E A' and such that s does not belong to any f2M, with PI C P CP2. Then the above integral equals the sum over all s E f2(P1, P2) of

fL\GA)' yePi(o)nwfXP,(o)wH\Pi) Pl(O)\G(A)1 yEP (O)fw?(O)ws\Pi(O)O'(Ho(x)

Since

T) A TP'Kp1,x(wsyx, x) dx.

oa(Ho(x) T) * A2P'Kp1,X(WyX, x)
-

a2(Ho(yx) - T) * A PKp,x(wsyx, x)

106

J. Arthur

[20]

for

any y E PI(Q), this equals

(2.9)

f(

(Pi(O)nws IPl(O)ws)\G(A)1

a2(Ho(x) - T)

TP'Kp,,X(Wsx, x) dx.

If s E f(P, P2), ws-Pows nM, is the standard minimal parabolic subgroup of M1, since s-'a >0 for a E A'. Therefore MI n w -'P w, equals M1i P,, for a unique parabolic subgroup P, of G, with PoC Ps C P. Write the integral in (2.9) as a double integral over 1 Ms(Q)Ns(A)\G(A)1 x (Pi(Q) n wS-'PI(O)ws)\Ms,()Ns(A). Pi n w-P, w is the semi-direct product of Mln wT'Piws and N n w;'Pw,, and Mi nwT1plw decomposes further as the semidirect product of M,(Q) and N'(Q). Therefore, (2.9) equals the integral over x in M,(Q)N,(A)\G(A)l of the product of a2(Ho(x)- T) and

(2.10)
This last

dn
JN(Q)\N[(A)
N(A)nw- I Pi(A)ws \N(A)

dnl A rTPKpx(wsnlnx, nlnx).

expression equals
f

J A2P'Kp,x(Wsninx, nx) dnl dn

= fi

A'PKp,,x(wsnlx, nx) dn, dn.

We apply Lemma 1.1 to the parabolic subgroup Mi n P, of Ml. Then this expression vanishes unless w (Ho(x)- T) is negative for each C SA. On the other hand, we can assume that (2.3) holds, with y, x, E and y replaced by Ws, nix, and nx respectively. In other words,
w

(Ho(wnlx)) -> (Ho(x)) + w (To)

for each

E A But it is well known that Ai.

s

(Ho(w,snlx)) < (sHo(x)),

so there is a constant C,

depending only on the support of f, such that

((x)sHo(xx)s )) < C

for every w in Al. These two conditions on Ho(x), we repeat, are based on the assumption that (2.10) does not vanish. We obtain a third

[21]

A trace formula for reductive groups II 107

condition by demanding that L(Ho(x) - T) not vanish. We shall show o that these three conditions are incompatible if T is sufficiently

regular.
We write the

projection of Ho(x)- T on a, as
-

CIC,ca

E + E> c,,w

c, and c,. Now the Levi component of Ps equals Mi ws'Mlws. Therefore sao is orthogonal to al. Then for wo E 1,

The first and third conditions on Ho(x) translate to the positivity of each

0 o(Ho(x) - sHo(x))
equals
weo(T-sT)+
Now w - s
v

Y caro0(sa')+

E cmo(w --stw).

is a nonnegative sum of co-roots, so the sum over w is Moreover we can replace each a in the sum over A by nonnegative. the corresponding root in AO\As. Since s maps the roots in this latter set to positive roots, the sum over a is also nonnegative. Finally, for any wo, wo(T - sT) can be made arbitrarily large for T sufficiently regular. We thus contradict the second condition on Ho(x). Therefore (2.10) is always zero so the integral of k(x, f) equals that of

A2K,(x, x).
3. The operator MP(tr) For

any x E Y, set

JTf) = I G(0)\G(A)1 k

(x, f) dx.

In this section we shall give another formula, which reveals a different set of properties of the distributions JT. We shall build on Lemma 2.4, which is a partial step in this direction. Fix P, r E II(M) and X E 3. Suppose that A is a linear operator on or ep(ir) under which one of the spaces 'P(w)x, is invariant. Here Ko is an open compact subgroup of Xp(1r)x,Ko,w

Xp(Wr)x,K0

108

J. Arthur

[22]

G(Af) and

W is an equivalence class of irreducible

representations of

KR. We shall write AX, AXKO or AX.KOW for the restriction of A to the

subspace in question. Suppose that t is a Siegel set in G(A)'. It is a consequence of Lemma 1.4 and [l(c), (3.1)] that given any integer N' and a vector b E 9p(rr)x, we can choose a locally bounded function c(g) on the set of E ap,c at which E(x, 4b) is regular, such that

IA TE(x, 0)I -Cc(4) Ilx l-",

-

for all x E . It follows that for

4, tf E p(wT)x, the integrals

dx A : IG(O)\G(A), TrE(x, O)A TE(x, ,)

and
|

~Go)( E(x, O)A TE(x, tp1) dx

converge absolutely, and define meromorphic functions in (t, j) which are regular whenever the integrands are. By Corollary 1.2 and Lemma 1.3 these meromorphic functions are equal. Thus we obtain a linear operator MT(7r) on p(ir) by defining

(Mp(rr)Xl, 4)2)= JG(Q)\G(A)1 A TE(x, 4)1) A rE(x, 2) dx = |fxi E(x, )) A TE(x, 42) dx, i G(Q)\G(A)'
for every pair A(l and ()2 in %1p(r). MT(r) depends only on the orbit of ir in HG(M). It is clear that Mp(7r)x is self-adjoint and positive definite. Notice also that

Ip(1T, k)

(iTr)

= MT (rr)X Ip (r, k)

for all k E K. It follows that for any Ko and W, M (7r)X leaves the finite dimensional space rp(7r)XKo,W invariant. Recall that in the proof of Lemma 4.1 of [1(c)], we fixed an elliptic element A in t(((R)'(C)KR. For any Ko and W, p(lr),KOW is an invariant subspace for the operator Ip(ir, A). Choose A so that for any X, ir, W and Ko, such that pp(Tr)XK0,w {0}, Ip(r, A)x,Ko,w is the

[23]

A trace formula for reductive groups II

109

product of the identity operator with a real number which is larger than 1. For example, we could take A to equal 1 + AA1l, where Al is a
suitable linear combination of the Casimir elements for G(R)' and KR. If A is any operator on a Hilbert space, IlAll1 denotes the trace class norm of A.

THEOREM 3.1: There is a positive integer n such that for any open compact subgroup Ko of G(Af),
x

I n(A)1 ffG(M) ||Mp(r)xK0

p(ir, An)XoIi d7r

is finite.

orm Assume the proof of the theorem r the moment and take r = deg An. Suppose that f is a function in C1(G(A)l), which is biinvariant under Ko. Then

IIMp(7r) Ip(rT,f)XII
=
=
<

|Mp(Tr)x,K'*Ip (r,)|lli
||Mp(7),Ko Ip(TA )-lIp(1T, *f)Ill
*

|IMT(I),,KoIp(,Tr, A n)JlP1
is of the operator Af *f) Ip(7r,
|

"

*)||.

For anyar the

norm

bounded by

cG(A)1

|(A"

f)(x)| dx.

is and in fact defines a continuous seminorm on finite, C*(G(A)'). Inf). Theorem 3.1 implies that for every f E C?(G(A)a), Thus,

J(3.1)
all
v.

2 x

n(A)-

'f G(M) IMT(M(r),

Ip(r,

dr7 f))Ji1

is finite, and in fact defines a continuous seminorm on Cr(G(A)I). In particular, the operator Mp(r)X * Ip(Tr, f)x is of trace class for almost

THEOREM 3.2: There is

an r>

rl such that

/ E C[(G(A)'), jT(f) =

for

any X and any

n(A)-''

n(M)

tr(MpT(7r), * Ip(Ir, f) dTr.

110

J. Arthur

[241

We shall prove the two theorems together, Let N and ro be the positive integers of Lemma 4.4 in [l(c)]. Choose an open compact subgroup, Ko, of G(Af) and a Siegel set 0 in G(A)'. According to Lemma 1.4 and the lemma just quoted from [l(c)], we may choose a

finite set {Yi} of elements in V(t(R)'IC) such that for xEG(A)', vE t',

r-ro+
and f a K-finite function in

deg Yi,

Cr(G(A)'IKo),
p(rr)

Z n(A)-'G(M) dE J E(M)
P

E(x, p(r, )f) ATE(y, ) d7r

is bounded

by

E,If* Yij XIIIIN IlYIl-N

When we set x = y and integrate the above GQ)\G(A)1, the result is bounded by

expression

over

vol(G(O)\G(A)') lf * Yll,0.
Suppose w = (WI, W2) is a pair of equivalence classes of irreducible representations of KR. We defined the function
f,(x) = deg W

-deg W2

Rt

chw(k) f(k lxk21)chw,(k2) dk dk2

in 4 of [l(c)]. We also defined the

positive integer to. Let

deg Yi.

r2 = r0+ eo+
As we saw in 4 of

[l(c)],
ID

fl/l, = vol(G(Q)\G(A)') X I f * Yillro, f E C?2(G(A)'),

[25]

A trace formula for reductive groups II

111

is a continuous seminorm on

C?(G(A)'). We have shown that

(3.2)

z)' E G(o)\G(A)1
X

P
*

n(A)- fJG(M4) .Ep(r) E(x, Ip(7r, f,)4)

A E(x, c) drdx

ft

is bounded by Ilfll , for every f E C2(G(A)'). Let r by any integer larger than r2 for which Lemma 2.4 is valid. Then if f E Cr(G(A)1), and X is fixed,

n(A)-1f GM) JG(M)

P

tr(M(r)x Ip(7r, f)x) d7r

*
<

n(A)-' ,(2 = CnG(M) 2 G(Q)\G(A)1 \E.ap(r)X E(x, Ip( * AE(x, 4)) dx dr

*

fe))

f|G(0)\G(A)1 X22 n(A)-' f|g(M) (2 E(x, Ip(r, f)) G

P

* ATE(x,

4)) dr

dx,

by Tonelli's theorem. The operator AT is defined in terms of sums and integrals over compact sets. If we combine Tonelli's theorem with the estimates of [l(c), 4] we find that we can take AT outside the sums over 4b, P and o, and the integral over 7r. The result is

f(0)()' A Kx(x, x) dx,

which by Lemma 2.4 equals JT(f). The proof of Theorem 3.2 will now follow from Theorem 3.1 if we take r to be larger than rl. The only remaining thing to prove is Theorem 3.1. We shall use Lemma 4.1 of [l(c)]. We can choose n, and functions g' E C(G(A))KR and g2 E- CT(G(A)')KR such that An * ga + g2 is the Dirac distribution at 1 in G(R)'. If i 1, 2, set

gi(xR Xf) = vol(Ko)-' gR(xR) ch4(xf), XR E G(R)', Xf EG(A)l,

112

J. Arthur

[26]

where chKo is the characteristic function of

Ko. Then

Ip(V7, A")XK = Ip(7, gl)x + Ip(V7, A ) p(, g2),.

Suppose that W is an irreducible KR-type and that w = (W, W). Then the trace of the restriction of MpT(Tr)X Ip(7r, An) to WP(1r)x.Ko.W is

tr(Mp(7r)x Ip(7r, gl.a),

Since the

+ MP(7r)X

Ip(, )-p(7r, g2.w)x).

than 1, this last

eigenvalues of Ip(7r, A) are all larger expression is bounded by

2 Itr(Mp(7r)x, Ip(7T, gi,)x). i=l

-

Now the trace class norm of the

operator
'

Mpr),,K, . IP(, An)-I X.K M~(Tr)~Ko

is the sum of the traces of its restriction to each of the pSp(7r)XKow. Therefore
, P

subspaces

n(A)-'

HG(M)

IP(7T,,, A)X.I |IMP(7)XK0

dr7

is bounded

by the sum over i = 1, 2 of

'

Z x

-1 n(A)-

(M)

ep(,)X G(o)\G(^'

E(x, Ip(r, gi,)Q)

.A TE(X,

) dx d7r.

This in turn is bounded by

Co | ,,

G(O)\G(A)1

Y
P

n(A)-f I Sz
JIG(A)

eXP(ff)k

E(x, Ip(7T, gi,))

*ATE(x, ) dx dr,
which is just (3.2) with f replaced Theorem 3.2, is now proved.

by g,. Theorem 3.1,

as well as

[1

[27]

A trace formula for reductive groups II

113

4. Evaluation in a special case

In this section we shall give an explicit formula for Jr(f) for a particular kind of class X E X. These special X we will call unramified; they are analogues of the unramified classes o E C for which we were able to calculate J'(f) in [l(c), 8]. The formula for JT(f) is a consequence of an inner product formula of Langlands which was announced in [4(a), 9]. Most of this section will be taken up with the proof, essentially that of Langlands, for the formula. First, however, we must demonstrate a connection between the truncation operator AT and the modified Eisenstein series defined by Langlands in [4(a)]. Fix a parabolic subgroup PI and a representation E 1I(Mi). If and iE a c, write E

0p,(ir)

Ep(x, , O)=Ep(x, 4c),

If s E Q(a1, a2), define

P D P.

M(s, ir, ) = M(s, O) by

= M(s, rt),t)-_.

M(s, 0)4

M(s, ) maps W1,(r) to ,p2(s7r). Suppose that X E X is such that P1 E Px. Then for all X E G(A)1,
O(mx),
mE

Mi(Q)\M,(A)',
the

is a cusp form in m. If P2 is a second group in P, we have following basic formula from the theory of Eisenstein series:

|J2

N2(o)\N2(A)

E(nx,4,, ) dn = E (M(s,
sEA(a1,a2)

)(x) 0)W e(s+PP2XH(x)).

A formula like this exists if P2 is replaced by an arbitrary (standard) parabolic subgroup, P. Recall that 2N(a; P) is defined to be the union over all a2 of those elements s E Q(a1, a2) such that sal = a2 contains a, and s-'a is positive for each a E A2. Then we have

(4.1)

2 f N(Q)\N(A) E(nx, (, O) dn = sEf(a1; P) Ep(x, M(s, 0)4, sO).

The verification of this formula is a simple exercise which we can leave to the reader. It can be proved directly from the series definition

114

J. Arthur

[281

of E(x, ), 4). Alternatively, one can prove it by induction on dim A, applying [4(b), Lemma 3.7] to the group M
E LEMMA 4.1: Suppose that P E P as above, that 4 E tp,(rr)X and that ; is a point in a*c whose real part OR lies in p +(a*)+. Then A TE(x, 4, O) equals

(4.2)

2 6EP2(O)\G(O) sEG(ai,a2) E2(SR)42(SR, Ho(5x)- T)) 2 2

Pi
.

(s+XHo(x))(M(ss, ())(6x).
absolutely. (The functions E2 and 42

with the sum over 8 converging are as [l(c), 8].)

PROOF: Suppose that P2 and s E Q(ai, a2) are given. In the process of verifying the equality of (8.5) and (8.6) in [l(c)], we ended up proving that for all H E ao,

E2(SR)d42(S5R, H)
was

equal to
{P: PDP2,ssE(al; P)}

(- )dim(Ap/Z) p(H). ,

Apply this to (4.2). Then decompose the sum over P2(Q)\G(Q) into a
sum over P2(Q)\P(Q) and P(Q)\G(Q). The sum over P(Q)\G(Q) will be finite by [l(c), Lemma 5.1]. If a E ad, S-la' is a nonnegative sum of elements of the form BV, for / E Al. It follows that

(SR - p2)(cX ) = (R - pl)(s-'l a + pl(S-a )-- P2(a )

is

positive. Therefore the sum

eEP2(0)\P(Q)

2 eWP(stHi6fx)) . (M(S, 00)(6Sx)

absolutely convergent, and in fact equal to Ep(8x, M(s, )4), s'). In particular, the original sum over 8 in (4.2) is absolutely convergent.
is
We find that

(4.2) equals

[29]
E
P

A trace formula for reductive groups II

115

(_l)dim(A/Z) 8EP(Q)\G(Q)

sEn(a; P)

Ep(8x, M(s, ), sO

i(H(x)

T).

If the left hand side of (4.1) is substituted into the brackets, the result is A TE(x, p, i). DTo simplify the notation, we shall assume that 7r(a) is the identity operator for all a E AI(R). This entails no loss of generality, since any wl-E H(Mi) equals rtr, for some such ir and some 71 E ia*. Given P2, define

412(X) = sER2(al,a2) E2(S4;R)2(S4R, Ho(x)- T)

.

e(S+P2xHO))(M(s,' ))(x).

If A E ia*, define

'2(A, x)

fA2(R)nG(A)l (e-(A+p2HO"(ax2(ax)) da,

for x E G(A)'. This function is not hard to compute. We have only to evaluate

fI

)W(E2(SR)42(s5R, Ho(ax) (

T) da

Since a ->H2(ax) is a measure preserving diffeomorphism from A2(R))n G(A)1 onto a2G, this last expression equals
es

es-AXH) 2(sSR)2(ssR, H T) dH.

Make a further

change of variables
H=
arE42

2 tav, t,ER.

Of course, we will have to multiply by the Jacobian of this change of measure. It is the volume of a2G modulo the lattice, L2, spanned by {a': a E A2}. The integral becomes a product of integrals of decreasing functions over half lines; it is easy to evaluate (see [l(b), Lemma

116

J. Arthur

[301

3.4]). We find that V2(A, x) equals

vol(a2/L2) sef(a1a-) H (M(s (sj Ai)(a ) a

*

(s-S-H T)

))(0).

We have been assuming that ,R is a point in pi + (a*)'. Let us suppose from now on that it is suitably regular. Then F2(A, x) can be analytically continued as a holomorphic function, for A in a tube in a*2 over a ball Bp, in a2, centered at the origin, of arbitrarily large radius. The functions

/52(A): x -* 42(A, x),

by A, span a finite dimensional subspace of For fixed A0 in Bp,, T2(A) is a square infunction from Ao + i(a )* to this finite dimensional space. tegrable Suppose that P; is another group in SY. Pick a class T' E H(M'), a vector O'E pi(7r') and a point ('E atic to satisfy the same conditions as above, and define the functions 12 and gV associated to any
indexed

L2(M2(G)\M2(A)'x K).

other group P' in ?. Then

(4.3)

G(O)\G(A)l

A TE(x,

', ,,) A TE(x, c') dx

is the sum over P2 and P2 in

Px of

2 IG(Q)\G(A)' GEP2(O)\G(Q)

q2(8x))(
(

EP(O)\G()

+2(6x)) dx.

This last inner product is given by a basic formula in the theory of Eisenstein series ([4(a), Lemma 4.6]). It equals

G O+i(a(2)* tE20(a2,a)

(M(t, A) 2(A), A(-tA)) dA,

and dA is the Haar measure where Ao is any point in Bp, n (P2 + on i(a2)* which is dual to our Haar measure on a'. Therefore, (4.3) equals the sum over P2 and s E Q(a, a2), of the integral over A, of the product of

(a*)+),

(4.4)

vol(a /L2)2
et

(sC-AXT)
-=J aE(2

[31]

A trace formula for reductive groups II

117

and

(4.5)

E
Pi

tEQ(a2,a)

e'(s''+tA) T) (4.5) E --e s'Ef(ajai)

-l
aE-A

(s''+ tA)(a

x (M(t, A)M(s,

0), M(s',I')X').

We shall show that (4.5) is a regular function of A on the tube over P2 + (a*)+. The functions M(t, A) are regular on this tube, so the only signularities are along hyperplanes

{A (s' + tA)(a) = },
for fixed s', t,

belonging

E (N(a2, C' and a E A'. Let SaThen 3 a'') be the simple reflection to a is root

(see [4(b), Pg. 35]).

= -sa

in dp5, and

{A: (sas'' + stA)(3) = 0}

hyperplane as above. Thus, the summands in (4.5) which singular along a given hyperplane occur naturally in pairs. We shall show that the two residues around the hyperplane add up to 0. Assume that (s''+ tA)(a ) =. Then (s,s'' + stA)(0))=0. The inner product from the summand of (4.5) corresponding to P'2, sas', Sat equals
is the same
are

(4.6)

(M(sat, A)M(s, 0)O, M(s\s', ;')r') = (M(s,, s';')*M(s,, tA) * M(t, A)M(s, ;)4, M(s', ;')4')

by the functional equations. But

M(sa, s'T')* = M(a, - s'')-l = M(sa, tA)-l,
since M(sa, tA) depends fore (4.6) equals

only on the projection of tA onto a. There-

(M(t, A)M(s, C), M(s', C")O'),

which is the inner product from the summand of (4.5) corresponding to P2, s', t. It follows that the residues of the two summands do add up to zero. Therefore (4.5) is regular at the hyperplane under consideration, and so is regular on the tube over p2 + (a*)+

118

J. Arthur

[321

Next we shall show that if s . 1, the integral in A of the product of (4.4) and (4.5) equals 0. Given such an s, choose a root a E Ad such that (s4R)(a') < 0. Change the path of integration from Re A = Ao to Re A = A0 + Nw,, where N is a positive integer which we let approach oo. We can do this by virtue of the regularity of (4.5) and the fact that the numbers

{lIM(t, A)||: Re A = A + Nm,}

are bounded

independently of N. Notice that

i le-A(r) e(ttlXT)I = e(tAo- to+N(t a-m,)X T)

is no greater than 1. Therefore, the integral over Re A = Ao+ Na approaches 0 as N approaches oo. It follows that the original integral

equals zero.
A, setting

We have only to set s = I in (4.4), multiply the result by (4.5), and then integrate over A. Make a change of variables in the integral over

A=

ZaWa, ZaEC.

With this change of measures, we must multiply the result by the volume of i(a )* modulo the lattice spanned by {,a: a E A2}. Since dA represents the measure on i(aG)* dual to that on ai, and since { a} and {a'} are dual bases, this factor equals

--i)
dim(A2/Z)

1 \ dim(A2/Z)

vol(a 2/L2)-'

The product of this factor with (4.4) then equals

2( ri2) L2)e Cvol(a \27-~/i~/

V

eZa ~a/Ed'(a )-,aT)

-~

Each a, is to be integrated over the line Ao(a v) + iR). We replace this contour with the line Ao(at) + N + iR, and let N approach oo. According to our assumptions on , CR(a ) > AO(a'), so we will pick up a residue at Za = (a '). By the arguments of the previous paragraph,

[33]
the

A trace formula for reductive groups II 119

integral of Za over the line Ao(ac)+ N + iR approaches 0 as N approaches oo. Therefore the integral of Za over Ao(at)+ iR equals the residue of the integrand at Za = ?(a ). It follows that (4.3) is the W product of vol(aG/L2) with the value of (4.5) at s = 1 and A = W.e have proved
LEMMA 4.2: (Langlands) Suppose that Pi, P; E Px, that 4 E Xlp,(r)x, 4(' E Pi(Tr')x and that t and 4' are vectors in a ,,c and a Pj,c whose real parts are suitably regular points in (a4)+ and (a )+ respectively. Then

f|IA) A E(x, -G(Q)\G(^)~

vol(a G2L2)
-

)', )A TE(x, C') dx

equals the sum over P2 E Px, s E f(al, a2), and s' E Q(a;, a2) of
e(s"+s''XT) (M(s, 00, M(s', ')). H (sC + s'')(a) aEA2
D

Both sides of the identity of the lemma are meromorphic functions in (r, c'). Therefore the identity is valid for all regular points C and r'. Recall that the elements of X are equivalence classes of pairs (MI, pi). We shall say that X is unramified if for any pair (MI, p) in X, the only element s E f((al, a,) for which sp = p is the identity. For the remainder of this section, assume that X is unramified. Suppose that = Pi = P= P and that ir = ir'. Then if 4, 4b', s and s' are as in the

lemma, (M(s,5), M(s', "')') = 0

unless s = s'. It follows that for i E ia*, (Mi(i,)), 4) equals
(sCXT)
-

lim 0

2 P2EX

sE7(4a,42)

2 vol(aQI/L2)

H-

aE42

(s)(a')

(M(s, 27 + 0), M(s', "7)<').

We can now take ir to be any class in H(M). We have shown that for P E Px and 7r E H7(M), ir, lim 2 2 M'(ir) =

vol(a/ILp) -pP2E sEnl(a,42) xe')M(s, (slma') ) n )-M(s,

aEA2

On the other hand, if P does not belong to 9x and Xr EH(M), then eP(ir)x = {0}. This fact can be extracted from the results of [4(b), 7].

120

J. Arthur

[341

We can therefore write

J(f) =
PEJ

n(A)-'

G(M)

tr(M(rr)x Ip(r, f)) d7r,

with Mp(7r)x given explicitly above in terms of the global intertwining operators. If we wanted to pursue the analogy with 8 of [l(c)], we might regard this formula as a linear combination of 'weighted characters' of f.

5. Conclusion

The results of this paper, and of [l(c)] can be summarized as an identity for the reductive group G. Namely, there is an integer r > such that for any f E Cr(G(A)') and any suitably regular point T E aO,
ree

2 J )=

xEt

JX

where

JT(f)= ,(Qo)\G(A) kT(x,f)dx

G(Q)\G(A)z

j[ (x, f)dx,

and

G (o)G()\

=
P

n(A)-

fGn(M) tr(MT(7r)x

Ip(r, f)x) drr.

be the restriction of the representation R to L2usp(G(\G)\G(A)). Let X(G) be the set of classes X E ' such that = {G}. Then Rcusp is the direct sum over all X in i(G) of the representations Rx. If X E X(G) and xr E H(G), MT(rr)X is the identity operator. It follows from the finiteness of (3.1) that if f is in C'(G(A)'), Rcusp(f) is of trace class. (This fact also follows from [3, Pg. 14] and [l(c), Corollary 4.2].) Moreover if f E Cr(G(A)'), for r as

Let

Rcusp

[35]

A trace formula for reductive groups II

121

in Theorem 3.2,
tr Rcusp(f)
=
=

xEV(G)

E tr R(f)

XE(2

( XEX(G) UG(G)
=

tr(G(7r, f)) dr

XEa(G)

2 xj(f).

Thus
tr Rcusp(f)=
fECO

2 (f)- xEx\X(G) J(f).

REFERENCES

[1] J. ARTHUR: (a) The characters of discrete series as orbital integrals. Inv. Math. 32 (1976) 205-261. (b) Eisenstein series and the trace formula, in Automorphic Forms, Representations and L-functions, A. M. S., 1979. (c) A trace formula for reductive groups I. Duke Math. J. 45 (1978) 911-952. [2] A. BOREL: Ensembles fundamentaux pour les groups arithmetiques et forms

automorphes, E. N. S., 1967. [3] HARISH-CHANDRA: Automorphic forms on semisimple Lie groups. Springer-Verlag, 1968. [4] R. LANGLANDS: (a) Eisenstein series, in Algebraic groups and discontinuous subgroups. A. M. S. 1966. (b) On the functional equations satisfied by Eisenstein, series, Springer-Verlag,
1976.

(Oblatum 18-IV-1978 & 3-XI-1978)

Department of Mathematics University of Toronto Toronto, Canada, M551A1