Заголовок
Источник: http://picxxx.info
Ссылка на PDF: http://picxxx.info/pml.php?action=GETCONTENT&md5=fb5d66223b35d5055548fe84faa0b985
Конец заголовка
Unipotent Automorphic Representations:
Global Motivation
JAMES ARTHUR
Contents
§1. Introduction
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
§2. Endoscopic data .................. .
§3. The discrete part of the trace formula ..........
§4. The conjectural multiplicity formula ..... .....
§5. The expansion of Idisc,t(f) ...............
§6. The sign characters e, and r..............
§7. The expansion of Edis,t(f) ...............
§8. A combinatorial formula for Weyl groups .........
§9. Concluding remarks ..................
1
5
14
20
27
37
42
54
67
§1. INTRODUCTION
In the paper [3], we gave a conjectural description of the discrete
spectrum attached to the automorphic forms on a general reductive
group. The main qualitative feature of this description was a Jordan
decomposition into semisimple and unipotent constituents. This is in
keeping with the dual nature of conjugacy classes and characters, and
in fact, with a general parallelism between geometric objects and spectral objects that is observed in many mathematical contexts. Such
a decomposition for automorphic representations would of course be
parallel to the Jordan decomposition for rational conjugacy classes.
It would also be analogous to the Jordan decomposition that is an
essential part of the representation theory of finite algebraic groups.
The decomposition should actually apply uniformly to the automorphic representations in certain families. The families or "packets" are
indexed by certain parameters which are the source of the decomposition. The quantitative side of the conjectures in [3] is a formula for the
multiplicity with which a representation in any packet occurs in the
Supported in part by NSERC
Automorphic Forms, Shimura
Varieties, and LFunctions
Operating Grant A3483.
Copyright e 1990 by Academic Press, Inc.
All rights of reproduction in any form reserved.
ISBN 0121766519
2
JAMES ARTHUR
spectrum. It is a generalization of the formula for tempered
representations which is implicit in the examples in [15]. In terms of
the Jordan decomposition, tempered automorphic representations are
semisimple. The multiplicity formula for nontempered automorphic
representations contains some new signs. These are constructed out
of the root numbers of certain Lfunctions, attached to the semisimple
part of the given automorphic representation.
In this paper, we shall try to give some motivation for the conjectures. Some version of the conjectures, at least for many classical
groups, ought to follow from the stable trace formula. This is certainly so in the few cases where the stable trace formula has been
established [15], [22]. In general, one would need to combine the
theory of endoscopy with the ordinary (or twisted) trace formula to
obtain a stable trace formula. There are still a number of problems to
be solved, but one can guess what the final answer will be. The purpose of this paper is to show that it is compatible with the conjectures
discrete
of [3].
For purposes of introduction, let G be a connected, simply connected group over a number field F. We shall be interested in the
spectral side of the trace formula. The essential ingredient we shall
study is a certain distribution
f E C (G(A)),
Idisc, (f) ,
which is discrete in the parameters which describe the representations
of G(A). It is given by an explicit formula (3.1), one term of which
involves the trace of f on the discrete spectrum. When the stable
trace formula has been established, the payoff will be an identity
(1.1)
Idisct(f) = (G, H)SIdsc(fH)
H
in which H ranges over elliptic endoscopic groups, t(G, H) is a certain
is a pullback to G of a stable distribuconstant, and f +
tion on H(A). (Recall that the endoscopic groups are a natural family
of quasisplit groups attached to G. Recall too that a stable distribution is a special case of an invariant distribution, which arises as a
natural consequence of the difference between rational conjugacy and
geometric conjugacy. We refer the reader to [3, §3] for a brief discussion of these notions and of the LanglandsShelstad transfer mapping
on U\~j) Id'i s (fH) is not generally stable.
fH.) As~3a &distribution
U13~111JU
C~IVl V~l
SIdic(fH)
G!(A), SI',H
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
3
However, the trace of f on the discrete spectrum is also usually not
stable. Endoscopic groups were actually invented by Langlands with
the aim of measuring this lack of stability.
The endoscopic groups on the right hand side of (1.1) should all
contribute to the multiplicity formula for representations in the discrete spectrum. However, the trace of f on the discrete spectrum is
only one of several terms in the explicit formula for Idisc,t(f). The
other terms are the surviving remnants of Eisenstein series, and are
parametrized by (conjugacy classes of) proper Levi subgroups of G.
Each such term is a linear combination of distributions, which are obtained by taking the trace of a product of two operators, one being the
action of f on the induced discrete spectrum, and the other being an
intertwining operator that comes from Eisenstein series. These additional terms have one important function. They account for that part
of the discrete spectrum of a given H which under functoriality maps
into the continuous spectrum of G. However, the additional terms
also contribute irrelevant information, which complicates the study
of (1.1). The attempt to separate the extraneous information from
the contribution of the discrete spectrum leads to combinatorial difficulties. The main point of this paper is to solve these combinatorial
problems.
The results are given in §5§8. In §5 we expand Idisc,t(f) into a
linear combination of irreducible characters. This hinges on the conjectures of [3]. However, we have only the modest goal of showing that
the conjectures are compatible with (1.1), so we are free to assume
them. Each coefficient in the expansion contains a certain quotient of
Lfunctions, which comes from the global intertwining operators. If
the irreducible character is tempered, this quotient should equal the
parity of the pole of the Lfunction at s = 1. If the irreducible character is nontempered, however, it will have a unipotent part. When the
corresponding unipotent element is not even, the quotient must also
be expressed in terms of the order of the Lfunction at the center of
the critical strip. The exact relation is given by Proposition 5.1, which
we prove in §6. Together with Lemma 7.1, it provides the justification for the sign characters which appear in the general multiplicity
formula.
In §7 we establish a parallel expansion of the right hand side of (1.1)
into irreducible characters. This requires various properties from endoscopy, some known and others which are expected to hold, which
we discuss in §2 and §3. The endoscopic groups H consist of the
4
JAMES ARTHUR
quasisplit form of G, together with groups of smaller dimension. By
reasons of induction, then, the stable distributions SIdHj are uniquely
determined by (1.1). However, we must derive the expansion in §7
without reference to the left hand side of (1.1). The coefficients in the
expansion have to be given as certain undetermined constants, which
can be regarded as "stable multiplicities", and which only later are
tied precisely to the sign characters discussed above. For a parameter
which contributes to the tempered discrete spectrum, the corresponding coefficient will be familiar from [15, §6,7] and [12, §12]. It is then
just equal to 1, divided by the order of a certain finite group.
Our aim is to show that with the assumption of the conjectures of
[3], the left and right hand sides of (1.1) are equal. We would thus
like to establish a term by term identification of the two parallel expansions. However, this is not immediately obvious. What remains to
be proved at the end of §7 is a sort of analogue for Weyl groups of the
endoscopy identity (1.1). The expansion of Idisc,,(f) contains certain
constants i(x), which are defined if x is any connected component of
a complex reductive group. The expansion for the right hand side
of (1.1) is identical, except that i(x) is replaced by another constant
e(x). In the first case, i(x) is given by a finite sum over elements
in the Weyl set of x. It is the analogue for Weyl groups of the left
hand side of (1.1). The second constant e(x) is the analogue of the
right hand side of (1.1), and is given as a finite sum over the isolated
conjugacy classes in x. In §8 we prove that i(x) equals e(x) for every
component x. This establishes the term by term identification of the
expansions of each side of (1.1).
At the end of §8 the reader might be wondering whether the paper
has provided the global motivation claimed in the title. It is true that
the identity (1.1) is weaker than the conjectural multiplicity formula
(and the local conjectures on which it is based). However, the identity
can still provide significant information about the discrete spectrum,
for either G or its endoscopic groups. This is especially so if for
one of the groups, the conjectures are known to hold. The group
GL(n) is such an example, thanks to recent work of Moeglin and
Waldspurger [21]. The twisted version of (1.1), applied to GL(n),
will relate the discrete spectrum of many classical groups to that of
GL(n). In particular, it should yield some version of the multiplicity
formula for the quasisplit orthogonal and symplectic groups. We shall
finish the paper in §9 with an informal discussion of these questions.
Throughout the paper we shall adopt the following notational con
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
5
ventions. Suppose that E is a set on which a group r acts. We shall
denote the set of orbits of r on S by either Orb(r, S) or S/r. In
general, if A and B are subsets of a group r, we shall write
Cent(A,B)
=
{b E B: blab = a, for allaEA}
for the pointwise centralizer of A in B, and
Norm(A,B)
= {b E B: bAb = A}
for the normalizer of A in B. Next, suppose that C is a finite union of
connected components in a (nonconnected) algebraic group. Then C+
denotes the algebraic group generated by C, and Co is the connected
component of 1 in C+. If s is any element in C, we set
C,
=
Then C8 is also an algebraic
C =
Cent(s,C°).
group, with identity component
(C
=
Cent(s,C0).
(This differs from the notation of [2] and some other papers, in which
the symbol C, was reserved for the identity component of the centralizer.) We shall also write
Z(C) = Cent(C,C°).
This group is the intersection of Co with the center of C+, and is
contained in Z(C°). Finally, if X is any topological space, 7ro(X)
denotes the set of connected
components of X.
§2. ENDOSCOPIC
DATA
Suppose that G is a connected component of a reductive algebraic
group over a number field F. Then G+ stands for the group generated
by G, and GO is the connected component of 1 in G+. We shall assume
that G(F) is not empty. As in [3, §6], we shall also assume that G is
an inner twist of a component in a quasisplit group. More precisely,
we assume that there is a map
r
:
G

G*,
6
JAMES ARTHUR
where G* is a component such that (G*)° is quasisplit, and such that
G*(F) contains an element which preserves some Fsplitting of (G*)O
under conjugation. It is required that rv extend to an isomorphism of
G+ with (G*)+ such that for any a E Gal(F/F), the map
ra(r/1):
G*
.
G*
is an inner automorphism by an element in (G*)°.
The standard situation is when G+ = G°. By allowing G to be
a more general component, we are providing for applications of the
twisted trace formula [5]. Associated to the connected component
G° we have the Lgroup
LGO
=
G
WF.
It is a semidirect product of a complex connected group G° with the
Weil group WF of F. (As in [3], we follow the notation of Kottwitz
[12], so that G° stands for the identity coset of the Lgroup. The
symbol LGO can then be reserved for the full Lgroup of GO.) We
have not assumed that G+ is a semidirect product of Go with a finite
cyclic group, but this does not seem to be a serious concern. In
particular, it is reasonable to define the Lgroup LG+ of G+ simply as
a semidirect product of LGO by the cyclic group 7ro(G+) of connected
components in G+. The action of ro(G+) on G° is dual to its action
by outer autmorphisms on G°. The action of 7ro(G+) on WF could
be defined by some map of 7ro(G+) into H1
However,
for simplicity we shall assume that ro(G+) and WF (as subgroups of
LG+) commute. Associated to the component G we have an "Lcoset"
(WF, Z(G°)).
LG
= G
WF,
in which G is a coset of G° in a group G+ such that
LG+
=
G+ X WF .
Z(G)
=
Cent(G, G)
Notice that
is in general a proper subgroup of the center
Z(G°)
=
Cent(Go,G°)
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
7
of G0. We must always be careful to distinguish between these two
groups. The Galois group r = rF of F over F acts on both Z(G)
and Z(G°). The subgroups of Finvariant elements are given by
Z(G)r
=
Cent(LG, O)
z(Go)r
=
Cent(LGO, G).
and
These too are not
generally equal.
A==
Observe that
(Z(G)r)°
is the maximal rinvariant torus in the center of G+. It is of course
not the dual group of the maximal split torus AG in the center of G+.
It is associated, rather, to the dual of the real vector space
aG
=
Hom(X*(G)F,R)
.
(X*(G)F denotes the module of Frational characters on G+.) More
precisely,
X*(G)F  X,(A)
so that the complex dual space a* c = X*(G)F 0 C is the Lie algebra
of A0. We shall write
KG =
(A0eo)
=
Z(G)rn (z(GO)r)
for the group of fixed points of G in Ado. It is a closed subgroup of
Ado whose identity component equals AG.
The theory of endoscopy for nonconnected groups is the subject of
work in progress by Kottwitz and Shelstad. As in [3, §6], we shall
guess at the ultimate form of some of this theory by extrapolating from
the connected case. Thus, an endoscopic datum (H, 7H,s, ) should
consist of a connected quasisplit group H over F, an extension
I
a
H

7'
WF

1,
coset s in G/Z(G°), and an Lembedding of 7 into
The definition is similar to the one given in [3, §3,§6] except
semisimple
LGO.

8
JAMES ARTHUR
that s is now a coset of Z(G°) instead of a single element in G. It is
required that
((H) = Cent(s,G°)° ,
oG of any element in the coset s, and that
the connected centralizer in
h eH,
s=(h)s1= a(wh)(h),
where wh is the image of h in WF and a(.) represents a locally
trivial element in H (WF, Z(G)). In other words, a() belongs to
(2.1)
ker1 (WF, Z(G°)), the kernel of the map
H1(WF,Z(G0))
H

1(WF, Z(GO)),
v
in which v runs over the valuations of F. It is further required that the
two extensions 1 and LH define the same map of WF into Out(H),
the group of outer automorphisms of H.
Recall that an endoscopic datum is said to be elliptic if the set ((7/)s
is not contained in any proper parabolic subset of LG. Equivalently,
the datum is elliptic if and only if the group
(Z(H )r)/r(z(WHr) n z(G)r
is finite, or again, if and only if ((AH) equals Ad. Finally, two elliptic
endoscopic data (H, , s,$) and (H', C', s', ') are equivalent if there
exist dual isomorphisms a : H  H' and /3: 7'iX, together with
an element g E G° such that
g(/3(h'))g1
'(h'),
=
h' E 1,
and
gsg1
=
s'.
Suppose that (H, ', s, g) is an elliptic endoscopic datum. We shall
write Aut(H) for the group of elements g in G° such that gsg1 = s,
and g~(H)gl =('). Then Aut(H) is a reductive subgroup of G°.
Notice that g(H)Z(G°)r is a closed subgroup of Aut(H). We shall
need to know later that it is of finite index. Equivalently, we must
establish
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
LEMMA 2.1. The identity component of Aut(H)
(/)(Z(°6)r)
° =
=
{g E G
:
slgsjlg
E
equals
(fH)A(Ao)
Let sl be a fixed element in the coset
Cs
9
s, and write
Z(G°)}
=
{g E G: gsg1 = s}
and
Cs = {g E G :
slgsg1 = 1}.
Then
g1
sl9gsl
9
is an injective map from Cs/Cs onto a closed subgroup Z(s) of Z(G°).
LEMMA 2.2. The subgroup
Z'(s) = {slzsz1:
is of finite index in Z(s).
z
E Z(G0)}
Suppose that g belongs to Cs. We can write g = glz, where
glThenbelongs to the derived subgroup Gder of G and z belongs to Z(G°).
sgs1 91 = ssiglsl
91 r1 ·szsl
sgsgx
g1 1i z 1
In particular, both gl, and z belong to Cs. But the element
s1gl1 g1 lies in Gder. The lemma follows from the fact that Gder
has finite center.
PROOF OF LEMMA 2.1: According to the first condition in its definition, Aut(H) is contained in Cs. Let Aut'(H) be the subgroup of
elements g E Aut(H) such that slgsllg1 belongs to Z'(s). The last
PROOF:
lemma tells us that Aut'(H) is of finite index in Aut(H).
Let g be an element in Aut'(H). Then we can write
1 E Cs, z1 E
9 = 911 ,
(G°).
Suppose also that h is an element in g(H). The second condition in
the definition of Aut(H) implies that ghg1 equals hllh, for some
element hi E (H). We can write this as
hzihlzl1
=
(hgih1)lhg
.
10
JAMES ARTHUR
hi and g, commute with sl. It follows easily from (2.1) that
hglh1 also commutes with sl. Therefore hzlh'z1 commutes with
s1, and belongs to the subgroup Z(G) of Z(G°). Now
Both
hzlhlz1
=
a(Zl)z11 ,
where a is the projection of h onto F = Gal(F/F). The action of F
on Z(G°) factors through a finite quotient Gal(E/F), and this action
preserves the subgroup Z(G). We obtain a homomorphism
g +
a(Z1i)Z1
from Aut'(H) to the finite group H' (Gal(E/F),
g lies in the kernel of this map. Then
a(zi)z71
=
Z(G)). Suppose that
a(z)z1,
a
E r,
for some element z E Z(G). In other words, there is a decomposition
Z1 = zz2, for elements z in Z(G) and z4 in Z(G°)r. We can therefore
write g = gz4, where the element gj = g z lies in the centralizer C,.
In other words, g belongs to the subgroup C8Z(G°)r.
It remains only to observe that ~(H) is the identity component of
C,. We obtain an embedded chain
(H)Z(G°)r
C
CCZ(G°)r
of normal subgroups of finite index.
nite index in Aut(H), and the two
component.
Aut'(H) C Aut(H)
Therefore ~(H)Z(G°)r is of fi
C
groups have
the same
identity
O
Let (H, 7, s,
5) be a fixed endoscopic datum. One is interested in
the Lhomomorphisms of WF into LG whose image is contained in
~(H). (Recall that an Lhomomorphism between two extensions of
WF is a homomorphism which commutes with the projection onto
WF.) One might like to be able to identify such objects with Lhomomorphisms of WF into the Lgroup LH of H. However, this is
not always possible. The Lgroup is a semidirect product H x WF
relative to an Laction of WF on H [20, 1.4]. (The action of WF of
course factors through the quotient F of WF.) But the two extensions
1t and LH of WF by H need not be isomorphic. In other words, there
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
11
might not be an Lembedding of LH into LG which coincides with
the image of C. Fortunately the problem is not serious. In the case
that G = Go, the question can be resolved by taking a zextension of
G, as has been explained in [20, (4.4)]. In the general case, Shelstad
has pointed out that it is necessary to work directly with extensions
of the endoscopic groups H. Suppose, then, that
(2.2)
1
)
Z1
)
' H
H1
1
is a central extension of quasisplit groups over F. We shall review
the question of whether there exists an Lembedding

1
,
LH1
which extends the canonical embedding H < H1 of dual groups.
Consider first the kernel KF of the projection WF  F, a connected group. It would be no trouble to construct an embedding for
the preimage 7t' of KF in W7. For it follows easily from (2.1) that
equals the subgroup ~(H) x KF of LGO. In other words, there
is a splitting 0: KF * I such that
(7'1')
(2.3)
hO(k)h1
=
0(whkWh ),
h E X1, k E KF,
where wh is the image of h in WF. Now by assumption, the
WF into Out(H) defined by 71 is the same as the Laction
h

w(h),
map of
hE H , w EWF,
used to define LH. It follows that 0 can be extended to a section from
WF to 7H such that
O(w)h0(w)
=
w(h),
h E H, wE WF.
Keep in mind that it is only the restriction of 0 to KF which is a homomorphism. However, 0 is uniquely determined up to multiplication
by elements in the center Z(H) of H. Therefore
O(wli)(w2) = b(1,W2)8(w1,W2) ,
Wl,W2 E WF,
where b(wl, w2) is a 2cocycle from WF to Z(H). By (2.3), b(wl, w2)
depends only on the images of wl and w2 in rF. We shall write /
12
JAMES ARTHUR
for the image of b in H2(WF, Z(1i)), relative to the embedding of
Z(H) into Z(H1). Then / is the inflation of a class in H2 (F, Z(H1))
which is independent of 8. Suppose that 3 is trivial. That is,
Wl,W2 E WF,
Z(WI)Wl (Z(W2))Z(WlW2)1
for a function z : WF * Z(H1) which is uniquely determined up to
/3(wl,W2)
a 1cocycle.
form
=
Every element in 7
can be
represented uniquely in the
hO(w) ,
hE H, wE WF,
and the map
(1 (hO(w))
(2.4)
hz(w) X w
is then an Lembedding of 7H into LH1. Conversely, if an embedding
1 exists, the function z(w) in (2.4) will split the class /.
=
Assume that the embedding ~1 exists. Suppose also that the central
subgroup Z1 of Hi is connected. Then we can form the Lgroup
LZ! = Z, X WF, and there is a canonical projection LH1  LZ1l We
also have an exact sequence
Z(H1) Z1
Z(H)
1

,
1
of complex abelian groups. Let zl(w) be the projection of z(w) onto
Z1. Then zl is a 1cocycle from WF to Z1. In fact, if we agree not to
distinguish between a cocycle and its corresponding cohomology class,
zl is just the preimage of the class b E H2(WF, Z(H)) determined by
the long exact sequence
 H1(WF,Z(Hi1))

H1(WF,Z1)
H2(WF,Z())
H2(WF, Z(H,)) It is uniquely determined modulo the image of H'(WF, Z(Hi))


in
H1(WF, Z). The map
alI(w)
zi(w) X w,
is an Lhomomorphism of WF to LZ1.
=
wE
WF,
13
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
Suppose that LF  WF is some extension of WF. Suppose also
that b : LF  LG is an Lhomomorphism whose image is contained in
I.
H LF
~(KH). That is, V = o ObH, for some Lhomomorphism :b
Then /1== 1 o0 H is an Lhomomorphism of LF into LH1. Set

tH(t) = 7(t)9(wt),
where wt is the image of t in WF and y(t) belongs to H.
61(lH(t)) = 7(t)z(wt) x w ,
tE
LF,
Then
t E LF.
It follows that the composition of )1 with the projection LH1  LZ1
equals a1 (or rather, the pullback of al to LF.) Conversely, any Lhomomorphism 1i : LF + LH1 whose projection to LZ1 equals al
is easily seen to be of the form C1 o AH. We can summarize these
remarks in a commutative diagram
LF
al
LZ1

/SW'01/
^H
\^
LH1
'
t
LG
In conclusion, we want to associate pairs (H1, 1i) to endoscopic data,
where H1 is a central extension (2.2) and C1 is an Lembedding (2.4).
We shall call such a pair a splitting for the endoscopic datum. We shall
say that (H1, 1) is a distinguished splitting if, in addition, the map
H1(A)  H(A) between adele groups is surjective, and the central
subgroup Z1 is an induced torus. That is, Z1 is a product of tori of the
form ResE/F(Gm). In particular, Z1 is connected, as we assumed in
the discussion above. Any endoscopic datum has a distinguished splitting. For example, the cocycle b(wl,w2) E Z(H) that we described
above often splits. In this case, we can simply take (H1,, 1) = (H, Id).
In general, we can always take H1 to be a zextension of H [11, §1],
the existence of which is established in [17, pp. 721722]. The first
condition follows from [11, Lemma 1.1(3)], while the second is part
of the definition of a zextension. It is also part of the definition that
the derived group of H1 is simply connected. This in turn implies
14
JAMES ARTHUR
that Z(H1) is a complex torus. It follows from [17, Lemma 4] that
the class f E H2(WF, Z(H)) is trivial. The embedding ~1 therefore
exists, and (Hi,g1) becomes a distinguished splitting. In general, if
(H1, 1) is any distinguished splitting, one needs to know that the
canonical map
ker' (F, Z(H))
ker' (F, Z(H1))

(As before, ker1 (F, Z(H)) denotes the kernel of
H1 (Fv, Z(H)).)
H' (F, Z(H)) 
is an isomorphism.
the map
V
This follows from the proof of
use the injectivity of the map
H1(F,Z(HI))
[12, Lemma 4.3.2(a)].
,
We will also
H1(F,Z(HI1)),
which is a consequence of the long exact sequence of cohomology, and
= 7ro(Zr) is trivial.
the fact the group H°(F,
21)
§3. THE DISCRETE PART OF THE TRACE FORMULA
We are going to study a piece of the trace formula. It consists of
those distributions on the spectral side of the trace formula which
are discrete with respect to the natural measure on the relevant automorphic representations. This part of the formula contains the actual
trace on the discrete spectrum. It is thus the payload, the part which
will eventually be used to compare automorphic representations on
different groups. Of course, there are serious problems relating to the
other terms in the trace formula which will have to be overcome first.
Our intention in this paper is simply to see what can be learned once
these other problems have been solved.
Let A be the adele ring of F. We should first identify our space of
test functions on G(A), the set of Avalued points in G. Consider the
diagonalizable group Z(G) = Cent(G, GO). We shall fix a closed subgroup X of the group Z(G, A) of adele points such that X n Z(G, F)
is closed, and such that XZ(G, F)\Z(G, A) is compact. Let X be a
character on X which is trivial on X n Z(G, F). Then Co (G(AP), X)
will denote the space of smooth functions f on G(A), of compact
support modulo X, such that
f(zx)
=
x(z)lf(z),
X, z E G(A).
15
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
Let t be an arbitrary but fixed nonnegative real number. The corresponding discrete part of the trace formula is the distribution
Idisc,t(f) ,
(G(A), X) ,
f E Cc
(G(A), X) which is given by the expression
i
(3.1) Z
IWG(aM)I1
[ro(G'+)l'
{M} wEWG(aM),.r
Idet(w 1)aG ltr(M(w,O)pp,t(O,f)).
on Co

M
(See [2, §4], [4, §II.9].) We shall describe very briefly the terms in
this expression. The outer sum is over the finite set of Go (F)orbits of
Levi components M of Frational parabolic subgroups P of G°. The
inner sum is over the regular elements
WG(aM)reg
=
{W E WG(aM): det(w 1)aG
¢
0}
in the Weyl set
WG(aM) = Norm(AM,G)/M
of (G,M). As in earlier papers, we regard the Weyl elements as
operators on the real vector space
aM = Hom(X(M)F,R)
which leave invariant the kernel aG of the projection of aM onto aG.
For each M there is canonical isomorphism from
AM,O
=
AMQ(R)° ,
MQ = ResF/Q(M),
AGM denotes the preimage of aG= in AM,OO we can
extend X uniquely to a character XM on XM A, oX which is
trivial on A,0 Let LdSt(M(F)\M(A),x1) be the subspace of
L2 (M(F)\M(A), XM) which decomposes under M(A\) as direct sum
of irreducible
whose Archimedean infinitesimal characonto aM. If
a
representions
ter has norm t. Then
pp,t(O): f , PP,(O, f)
=
J\G(A\)
f(x)pp,t(O, x)dx
16
JAMES ARTHUR
stands for the corresponding representation induced from P(A) to the
group G(A)+ generated by G(A). It acts on a Hilbert space 7Hp,t of
XM equivariant functions on G(A)+. Finally,
M(w,0O): Hpt

'Hp,,
w
W
(aM)reg,
is the global intertwining operator which comes from the theory of
Eisenstein series. For a given conductor, Idisc,t(f) is a finite linear
combination of irreducible characters on G(A)+.
There are some minor discrepancies between (3.1) and the original definition [2, (4.3)]. In (3.1) we have summed over the orbits
{M} instead of all Levi components which contain a given minimal
one. This is why IWG(aM)11 appears instead of the normalizing
constant IWOMIIWoGl1 from [2]. The operator pp,t(O, f) here comes
from a representation of G(A)+ induced from a subgroup of the connected component Go(A). It is a direct sum of [ro(G+)I copies of
the corresponding operator from [2], which comes essentially from
the induced representation of GO(A). Hence the constant Iro(G+)K1
in (3.1). The difference between taking a Xequivariant function on
G(A), as we have done here, and a function defined on the subset
G(A)1 of G(A), as in [2], is purely formal. In [2], there was also
the additional assumption that f was Kfinite, but this was only for
dealing with other terms in the trace formula.
The program for comparing trace formulas on different groups, as it
is presently conceived, falls into the general framework of stabilizing
the trace formula. The basic references for this problem are [18],
[12], and [13]. The problem was solved completely for G = SL(2) in
[15]. A general solution would include: a transfer map from functions
for G to functions for endoscopic data, a stable distribution analogous
to Idisc,t for any quasisplit group, and an identity relating Idisc,t to
the corresponding stable distributions for endoscopic data. We shall
discuss the transfer first, and then describe the expected properties
of the other objects in the form of a hypothesis.
Suppose that (H,H,7s, ,) is an elliptic endoscopic datum for G.
Assume also that we have fixed a distinguished splitting (Hi, 1) for
the endoscopic datum. As we recall from §1, 1i determines an Lhomomorphism 1 : WF LZ1. Let

(1: Z1(F)\Z1(A)
(  C*
be the character associated to ac by the Langlands correspondence
for tori. Now, in the special case that G = G°, the results [20]
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
17
Langlands and Shelstad imply the existence of a canonical map
f fHi from functions f E Cc (G(A) to functions fHi(7H,) on
suitable stable conjugacy classes in H(A), with the property that
Z1 e Z1(A).
fHl(zl')H ) (C(Zl)lfH1(7H)
(See also [12], [13] and [3].) The map must be constructed as a
tensor product of the local maps fV  fH", fv E Cc (G(Fv)), which
are defined explicitly in [20]. Langlands and Shelstad expect that
fHi is the set of stable orbital integrals on H (A) of a function g in
C, (Hi(A),(1). We shall assume that this is so. In fact, we shall
of
=
,
assume that the transfer map
f/
f
fECe
,
(G(A)),
has been defined, and has this property, for general G.
We should actually modify the transfer mapping so that its domain
is the space CC (G(A, X) considered earlier. Lemma 4.4A of [20]
suggests how the functions
f(x)
=
f(x)
z
Z(GG(),
(G,), Xx
f ECe°(G(A)),
should behave under the transfer map. In general, there will be a
mapping z  z' from Z(G, F)\Z(G, A) into Z(H, F)\Z(H, A).
We also have the exact sequence
norm
1
We can then
(3.2)

Z1
)
Z(H1)

Z(H)
, 1
expect a formula
(fz)H "(H,)
=
Cl(Zl)fHl(zlHi),
where C( is an extension to Z(Hi,F)\Z(Hi, A) of the character on
Z1(F)\Z1(A), and z1 is any point in Z(H1, A) whose image in Z(H, A)
equals z'. Recall that X is a character on the closed subgroup X of
Z(G, A). We shall assume that
x(z)
z EX,
(z'),
where X' is a character on the image X' of X in Z(H, A). To define the
transfer mapping for functions in Ce°(G(A), X), we simply multiply
=
18
JAMES ARTHUR
each side of (3.2) by X(z), and integrate over z in X n Z(G,F)\X.
Let X1 be the preimage of X' in Z(H1, A), and set
Xi(Zi)
=
(l(Zl)X'(z'),
for any point z1 E X1 with image z' in X'. Then Xi is a character
on X1, and the triple (HI, X, X) satisfies the conditions we imposed
on (G,X, X). In this context our assumption is that for any function
f E C~ (G(A), X) there is a function g E CC (Hi (A), Xi) whose stable
orbital integrals are given by f H,. The function g is of course not
uniquely determined by f. However, if SI is any stable distribution
on C{(Hi(A), X1), SI(g) will be uniquely determined by f. We shall
therefore write
SI(f H) = SI(g).
The ultimate goal is to give an expansion of Idisc,t as a linear combination of stable distributions on the equivalence classes of elliptic
endoscopic data {H} for G. The coefficients will be certain constants
t(G,H), which in the case G = G° were introduced by Langlands
[18]. (Following the usual convention of metonymy, we shall often
Kottwitz
write H in place of a full endoscopic datum (H, , I).)
s,
has established a simple formula for these constants [12, Theorem
8.3.1], again when G = G0. Let
T
(G°)
=
r(G°)r(G°)1
be the relative Tamagawa number of G° [12, §5]. (T(G°) denotes the
ordinary Tamagawa number of Go, and Go is the simply connected
cover of the derived group of G°. Thus according to Weil's conjecture,
which has been established by Kottwitz [14] for groups without Es
factors, Ti(G°) simply equals T(G0).) Kottwitz' formula is then
t(G, H) = ri (G°)r (H)17ro (Aut(H)) I1
In the general case, the constants have not yet been defined. We shall
have to get by with a makeshift definition that reduces to Kottwitz'
formula when G = G°.
If we are given an equivalence class {H} of elliptic endoscopic data,
we shall usually assume implicitly that H is a representative of the
class such that / is the identity. That is, /H is an embedded subgroup
of LGO. Then Z(H)r is a subgroup of G° whose identity component
19
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
equals AG. The subgroup KG = (A&0)G of G° also has AG as its
identity component, so that KG n Z(H)r is a subgroup of finite index
in KG. For general G we shall simply define
(3.3) (G, H) = 1 (G°)r (H)l1 to (Aut(H)) I1G/KGnZ(n)r 11.
The fourth factor in the product on the right, which of course equals
1 when G = G°, is suggested by the calculations in §7.
We can now state the hypothesis. Part of it applies to any (G, X, X)
as above, and part applies to triples (G1,X1, X1) with the restriction
that G1 is a connected quasisplit group over F.
HYPOTHESIS 3.1. For any (G1, X1, X1) there is a stable distribution
SId,t on C°e (G1(A), Xi) with the property that for any (G, X, X),
the distribution
EdiSc,t(f)
(3.4)
=
d'Ct(f 1)
E (GH)Si
X)
Here f stands for any function in C° (G(A), and
H is summed over the equivalence classes of elliptic endoscopic data
5
for G.
Remarks. 1. It is understood that we have fixed a distinguished
splitting (H1, ~1) for each H. The distribution SIf st(f
H) should
then depend only on H and not on the splitting.
2. The stable distributions Sldist are uniquely determined by the
condition that Edisc,t(f) equals Idisc,t(f). For suppose that they have
been defined inductively for any group whose semisimple part has
dimension less than that of G1. Setting G = G1, one simply defines
equals Idisc,t(f).
SIsc,t(f)
=
Idisc,t(f)
E
L(G, H)Sdi"U(fHi)
In this case, the hypothesis becomes the assertion that the right hand
side is a stable distribution in f. This of course is highly nontrivial.
It is likely to be resolved only by proving a similar assertion for all
the other terms in the trace formula. There is a discussion of this
question in the paper [19].
We shall need a slightly different formula for t(G, H) in §7. For H
as
above, set
z(H)r = Z(H)rz (GGO)/Z(GO)
Since H
represents
(abelian) group.
an
H
n Z(G)r
Z(H)r/z(H)r
elliptic endoscopic datum,
Z(H)r is a finite
20
JAMES ARTHUR
LEMMA 3.2. The constant
i(G, H) equals
(3.5) Iker1 (F, Z(G°))lIro0(G)l'
Iker' (F, Z(H)) IIZ(H)r 1 IAut(H)/H Z(G0)r1
PROOF: The main point is the formula
Ti(GO)
=
(Z(Go)r) Ilker'(F, Z(G0))I1
7ro
of Sansuc and Kottwitz for the relative Tamagawa number [12,
(5.1.1)]. From this, it will be a routine matter to derive the expression (3.5) from (3.3). For Lemma 2.1 tells us that
Iro(Aut(H)) 1
=
IAut(H)/HZ(GO)rl1 IHZ(G0)r/HA~o 1
Keeping in mind that AGo is the identity component of Z(G°)r, we
deduce that
HZ(G )/HAo l=
=
IZ(G°)r/z(G°)r n (HAo)l'
(Z(GO )r n (HAoo))/AGo
o(Z(GO)r) Gl
.
Moreover,
IZ(G)r n (HAoo))/Ao I
= In Z(G0°)/H n Ao I
=
=
=
Iz(H) n z(G)r/z(H)r n KG\
n KG)I1
IrO(Z(ft)r n Z(G)r)llro(Z(Hi)rIK11G/Z(fH)r
17o(Z(fH)r) Z(fi)rll 7ro(KG)
n KG
The lemma follows from the formula above for T1 (G0) and its analogue
°
for T1(H).
§4. THE
CONJECTURAL MULTIPLICITY FORMULA
Our goal is to provide some motivation for the conjectures on nontempered automorphic representations stated in [1] and [3]. The
main global ingredient of the conjectures is a multiplicity formula for
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
21
automorphic representations in the discrete spectrum. It is a generalization of similar formula for tempered automorphic representations
which was implicit in the examples of [15] and was stated explicitly
in [12]. We shall recall the various objects from [3, §8] needed to
state the formula.
The automorphic representations which occur in the spectral decomposition should be attached to maps
(4.1)
:
LF x SL(2,C)
LGO.
such that the projection onto G° of the image LF is bounded. Here
LF is hypothetical Langlands group, which we shall assume is an
extension of the Weyl group WF by a compact connected group. The
maps themselves are subject to certain conditions. For example, b
should be globally relevant, in the sense that its image must not lie
in a parabolic subgroup of LGO unless the corresponding parabolic
subgroup of Go is defined over F. Another condition is designed
to insure that b parametrizes representations of G°(A) which lift to
G(A)+. Let
So = S (G)
be the set of elements s E G such that each point
t' E LF X SL(2, C),
s(t')S1l(t')1
belongs to Z(G°), and such that the class of the 1cocycle
t  s(t)sl1(t)' ,
tE LF,
lies in the subgroup kerl(LF, Z(G°)) of H1 (LF, Z(G)). The condition on is that
be
Recall also that two
,
S, nonempty.
parameters
01 and 1b2 are equivalent if there is an element g E oG such that
(4.2)
/2(t,u) = g11(t, u)ga(t), (t, u) E LF x SL(2, C),
where a(t) is a 1cocycle of LF in Z(G°) whose class in H1 (LF, Z(G°))
lies in kerl (LF, Z(GO)).
Let 2(G) denote the set of equivalence classes of maps (4.1) which
satisfy the required conditions [3, §8]. Let 1o(G) denote the subset
of (equivalence classes of) maps 1 E @(G) such that the set
0
S,
=
SO(G)
=
SO(G)/Z(G°)
22
JAMES ARTHUR
[4] we called these maps elliptic. They should parametrize
automorphic representations which occur in the discrete spectrum. It
will be convenient to define two other subsets of 1(G). Let us say
that b is weakly elliptic if the group Sp(G°) (obtained by replacing G
with the identity component GO) has finite center. We shall say that
/
is finite. In
is discrete if it satisfies the weaker condition that the group
3+
S+/Z(G°)
=
has finite center. (Keep in mind thatS, S,
S,(G°), S,(G°), etc., are complex reductive Lie groups which are
generally not connected.) Let o(G) and T'disc(G) denote the set of
(equivalence classes of) maps b E @(G) which are weakly elliptic and
discrete, respectively. Then we have embeddings
generated by S,
'Io(G)
V(G)
idisc(G) c T(G).
Let X be a fixed character on a subgroup X of Z(G, A) which satisfies the conditions of §3. We may as well assume that X is contained in
Z°(G, A), the adele group of the identity component of Z(G). There
is a canonical map from LGO onto the Lgroup LZO(G) of Z°(G).
The composition of any parameter b E [(G) with the map gives a
parameter in A (Z°(G)), and therefore a dual character
(: Z°(G,F)\Z°(G,A) , C*
C
C
.
I(G, X), To(G, X), etc., for the set of parameters 4I in
9(G), Io(G), etc., such that the character (, coincides with X on X.
Suppose that p E I(G, X). As in [3, §8], we can form the finite set
S, = So(G) = So/SZ(G°).
We shall write
It is a coset of
SO(G°)
=
S,(G°)/SZ(G°)
in the finite group
S,(G+)
S,(G+)/S+ Z(G°)
Now the local conjectures in [3, §6] assert that there is a set II,
representations attached
=
to 4. The elements in
of
HI, should in fact
23
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
belong to Nunit (G(A), X), the set of equivalence classes of irreducible
unitary representations of G(A)+ whose restrictions to G°(A) remain
irreducible, and whose central character on X coincides with X. There
should also be a canonical pairing
< x,7r
>,
x
E S+,
7r E nI,,
such that the functions x +< x, r > are characters of nonzero finite
dimensional representations of SI. Finally, the conjectures assert the
existence of stable distributions
(4.3)
on Cc
f~ (+),
(G1(A),Xi), for each (G1,X1,X1)
f
61E
I(G1,X1),
with G1 connected and
quasisplit.
Let us recall how the distributions (4.3) are supposed to behave with
respect to endoscopic data. Suppose that s is a semisimple element
in S3. Take H to be the connected centralizer in G° of any point in
s, and set
u = Hix(LF
SL(2,C))
There is obviously an injection H H and a surjection H
We are
that the kernel of the
assuming
and it follows that b
Therefore
1

maps
H

WF.
map LF WF is connected,
both the kernel and SL(2, C) into H.
K
H

WF
1
is a short exact sequence. We can identify H, equipped with the
canonical Laction of WF induced by K, with the dual of a well defined
quasisplit group H over F. If g is the inclusion of H into LGO, then
(H, ', s,5) is an endoscopic datum for G. It has the property that b
equals g o OH for some Lhomomorphism 4'H of LF x SL(2, C) into
7'. Now, let (H1, 1) be any distinguished splitting for the endoscopic
datum. We can construct the character X1 on a closed subgroup X1
of Z(H1,A) as in §2, and from our remarks in §2, we see that the
parameter
b1
=
(1o'H
belongs to 1(H1, X1). According to our assumptions on the transfer
map f , fHI, the distribution
f  /f H, i),
f ECO (G(A) x)
24
JAMES ARTHUR
makes sense. It should satisfy the formula
(4.4)
fH 1(1)
S, 7r> fG(7),
<
=
7rE{n,}
where s is the image of s in So,,
so, is the element
*(1(o1 1))
in
Sp(G0), and
fG(7r)
=
tr(J
x\G(A)
f(z)7r(x)dx)
As in [3], {(II} denotes the set of orbits in Il, under 7ro(G+)*, the
dual of the finite component group, which acts in the obvious way on
Unit (G(A), X) Recall that the element so was introduced in [3, §4]
to describe the signs which occurred on the right hand side of (4.4).
The objects we have just described, namely the packets HIl, the
pairings < x, r >, the stable distributions (4.3), and the formula
(4.4), are all consequences of the local conjectures [3, Conjectures
6.1 and 6.2]. The adelic versions described here are simply restricted
tensor products of the local versions in [3]. We shall assume their
existence in what follows.
We should also recall the sign character
E·: S+ ' {±1}
which occurs in the conjectural multiplicity formula.
LI
=
Set
LF X SL(2, C),
and consider the representation
r,(s,t')
of S+ x
Ad(sik(t')) ,
=
s
LF on the Lie algebra g of G. Let
To
0
(AkC
Tk =
=
k
k
k
k)
E
E
S,'Lt
,
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
25
in which Ak, Pk and Vk are irreducible
(finite dimensional) representations of S L,LF and SL(2, C) respectively. The global Lfunction L(s, Pk) will be defined as a product of
local Lfunctions. We shall assume it has analytic continuation and
satisfies the functional equation
be the
decomposition of rT
L(s, Pk)
=
e(s, Pk)L(1

,
lk)
where E(s, Pk) is a finite product of local root numbers. It follows from
the functional equation that if /k is equivalent to its contragredient
(, pk) = ±1. Let us write g for the direct sum of those
irreducible constituents Tk such that (i) Pk  pk, (ii) e(, /1k) = 1,
and (iii) dim vk is even. The sign character is then given by
x E S,
E(x) = e+(x) = n det(k(s)) ,
(4.5)
k
Ik, then
where the product is taken over those k such that Tk is contained in
, and s is any element in S which projects onto x. In other words,
(4.5')
(x)
=
det(s, EndL())
We could actually have replaced the first condition in the definition
of g by the stronger assertion (i') rk k. Indeed vk is always equal
to its contragredient, and
det(Ak(s))
=
detAk(s)1
Therefore, the contribution to (4.5) of the distinct pairs (rk, Tk) equals
1. It should also be noted that the condition (iii) above is not really
necessary. For suppose that rk satisfies (i') and (ii), but that dim(vk)
is odd. Then vk corresponds to the principal unipotent in an odd
orthogonal group. Since pk is selfcontragredient, its image must be
contained in either the orthogonal or the symplectic group. We shall
assume the generalization of the theorem of Fr6hlich and Queyrot [6]
which, in view of the sign e(, pk) = 1, implies that pk is actually
symplectic. Finally, since the representation rk is self contragredient
and preserves the Killing form, it must be orthogonal. For this to
be so, the third representation in the tensor product must actually
be symplectic. Therefore det Ak(s) = 1, and Tk contributes nothing
26
JAMES ARTHUR
to (4.5). This explains the apparent discrepancy between the present
definition (4.5) and the earlier one [3, (8.4)].
If q is any vector in the Hilbert space L2(GO(F)\GO(Af), X1), set
(R(y)q)(x)
/(lxy),
=
x
E G°(F)\G°(A),
for any points y E G(A)+ and 6E G+(F) such that (ly belongs
to G°(A). This gives an extension of the regular representation to
G(A)+. For any representation r E HIunit (G(A), X), let mo(7r) be the
multiplicity with which r occurs as a discrete summand of R. Now,
suppose that ir belongs to a packet 114,, b E [(G, X). Then we have
the nonnegative integer
(4.6)
m,(7r)
=
ISl1 E ,(x) < X,7
>
zES+.
given explicitly in terms of the pairing. The multiplicity formula
amounts to the global component of our conjecture, and will be stated
formally as a hypothesis.
HYPOTHESIS 4.1. For any representation r E HIunit(G(A),X), we
have the multiplicity formula
(4.7)
E mg,(r).
°
,Ero(G,x)
Before discussing the conjectures, we shall collect a few simple observations for our later use. Let t be a fixed map in T7(G). (We shall
sometimes not distinguish between a map and its equivalence class.)
Let Cp denote the centralizer in G° of the image of b. Then COZ(G°)
is a subgroup of Sp,(G°). The quotient
mo(r)
=
C,
Cz(G°)/Z(G°)
=
is a subgroup of S9(G°).
Now, the image of the cocycle
t E LF, s E Si(G°),
t
sk(t)s1l(t1)
in H1(LF,Z(G0)) gives a map from Sp(G°) into ker (LF,Z(G°))
whose kernel is easily seen to equal CZ(G°). We therefore obtain a
continuous injection

SO(G°)/COZ(GO) r S3(GO)/C0
c
ker
(LFiZ( 0)).
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
27
According to Lemma 11.2.2 of [12], or rather its extension
hypothetical group LF, kerl(LF,Z(G°)) is isomorphic to
kerl (F, Z(G0)). In particular, ker' (LF, Z((G)) is a finite discrete
group. Therefore, the connected component S, of S, maps to the
identity element in ker (LF, Z(G0)). We obtain an identity
S0 = C
(4.8)
of connected components. In particular, if we set
to the
C,
=
CZ(G°)/CZ(G°)
= C
,
we can write the injection above as
(4.9)
S,(G°)/C
,
kerl(LF, Z(G)).
Suppose that s is a semisimple element in S,. According to our
conventions, S,,s denotes the centralizer of s in S,, and So, is the
connected component of 1 in Sa,8. We can also take the centralizer
C1,, of s in C,, and its identity component C ,. In §7 we shall use
the identities S,,s = C,, and , = C,0. These of course follow
immediately from (4.8). We shall also have occasion to consider some
slightly different centralizers. Keeping in mind that s is a coset in
G/Z(G°), we write S,,s for the centralizer in S, of any element in
the coset s. Then
So'Z(G°)/Z(G°)
subgroup of Sg,,, which by Lemma
particular, we have an equality
is a
Ss,
2.2 is of finite index. In
S,,Z(G°)/Z(G°)
of identity components. Similarly, if Cs,8 denotes the centralizer in
(4.10)
=
Co of any element in the coset s, we have
(4.11)
',s
=
C,,
C=
,,Z( ()/Z(GG°)
§5. THE EXPANSION OF
Idisc,t(f)
We have now stated two global hypotheses. As we have already
noted, Hypothesis 3.1 should be a consequence of a stable trace formula. Once this is established, one could try to combine the formulas
28
JAMES ARTHUR
(3.1) and (3.4) to deduce something approaching the multiplicity formula in Hypothesis 4.1. Our aims in this paper are more modest.
We shall simply show that the two hypotheses are compatible. We
are actually going to establish that Hypothesis 4.1, together with the
local assumptions of §3, §4 and [3, §7], implies Hypothesis 3.1. More
precisely, we shall show that the formula for Idisc,t(f) obtained by
combining Hypothesis 4.1 with (3.1) equals the formula for Edi8c,t(f)
provided by the definition (3.4). In the process we shall gain some
insight into the role of the sign characters ,.
In this section we shall derive a formula for Idisc,t from Hypothesis
4.1. By combining (4.7) with (3.1) we will obtain an expansion for
If
Idisct(f),
E
Cc
(G(A), X) ,
as a linear combination of irreducible characters. In doing this we
will need to apply a local conjecture from [3, §7] for the values of
normalized intertwining operators.
According to (3.1), Idisc,t(f) equals the sum over {M} and over
w E WG(aM)reg, of the product of
lro(G+)ll WG(aM)l1det(w  1)aG with
(5.1)
tr(M(w, 0)pP,(O,f))
Our first task is to expand (5.1) into a linear combination of irreducible characters.
For any M, and w E WG(aM), we can form the component Mw =
M * w. It satisfies the same conditions as G. Now, recall that Pp,t(0)
is the representation of G(A)+ obtained by parabolic induction from
the action of M(A) on
(5.2)
Ldisct (M(F)\M(A),XM1)
representation of M(A) has a canonical extension to the group
M,(A)+ generated by the coset Mw(A) = M(A)w. In particular,
the space (5.2) can be decomposed into a direct sum of subspaces
corresponding to irreducible representations a,, of Mu(A)+. There is
a similar decomposition
This
?iP,t
=
dHP w)
ffw
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
of the induced
space into subspaces which
29
are invariant under the
operator
(5.3)
M(w,O)pp,(0,f)
If the restriction of a, to M(A) is reducible, one sees easily that
the trace of the operator (5.3) on 1Hp(a,) vanishes. Therefore, in
computing the full trace (5.1), we need only consider representations
aw which belong to the space we have denoted by IIunit (Mw(A), XM1).
According to Hypothesis 4.1 (applied to Mw rather than G), the
multiplicity with which a representation a, E unit (Mw(A), XM1) occurs in (5.2) equals
.
E
0we o(Mw,XM,t)
^met(aw)
,
where mw (aw) is the nonnegative integer defined by (4.6). We have
written To(Mw, XM, t) to denote the set of parameters in To(Mw, XM)
whose Archimedean infinitesimal character has absolute value t. Any
pair oiw and aw, with mw (aw) O, determines a subspace of (5.2),
and also a subspace of the induced space ip,t. The restriction of
(5.3) to this latter subspace can be expressed in terms of the operators
studied in [3, §7]. It equals an expression
(5.4)
mow(aw)r(bw)(Rp(aw, bw)IZp(a, f)) ,
whose constituents we shall describe in a moment. The trace (5.1)
becomes the sum over Ow E 'o(Mw,XM,t) and aw E 1
H, of the
trace of the expression (5.4).
Given M and P, it is convenient to fix a dual parabolic subgroup
LP = P XI WF in LGO with Levi component LM = M > WF. The
choice of P and LP determines an embedding of the Lgroup LM into
LG°. It also allows us to identify WG(aM) with the dual Weyl set
WG(aM) = Norm(AM, G)/M.
Returning to (5.4), we note that Ip(a) stands for the induced representation of G(A)+ obtained from the restriction a of aw to M(A).
The operator
Rp(a,,w ,) =
Rp(affwvw,v)
V
30
JAMES ARTHUR
is a tensor product of local normalized intertwining operators defined
in [3, (7.4)]. When this operator is evaluated at a smooth vector
in 'p,t, almost all the terms in the product reduce to 1. Finally,
the scalar r(,,) in (5.4) is obtained from an infinite product of local
normalizing functions of the form [3, (7.2)]. It equals
lim
where
p,, o
(L(O, ,\,X)E(O,)l
PP,
w
o
{,,A,)~1) ,
b,,A is the twist of the global parameter
q
'k:t)W
O
t,
))
tE LF,
by the vector A in
aM, =
X(M)F®C

X,(AM)®C,
and pP,w is the contragredient of the adjoint representation of LM on
n
W1npW/w1nU pnW
.
Here nf stands for the Lie algebra of the unipotent radical of
Applying the anticipated functional equation
L(0, pP,w o° w,kA)
we
=
E(0, Pp,w o w,,A)L(1 PP,w o
LP.
,A),
write
(5.5)
r(Vw)
=
lim (L(1, pp,,,,
0\.0
o0 OP,>)L(l p, o
w,)~1)
(See [16, Appendix 2].)
Having described the terms in (5.4), we go back to the expression
we have obtained for (5.1). Recall [3, §7] that
Rp((aw,Ow) = ((Mw)Rp(aw, w),
for any character ( in
7ro(M+)*= Hom(M+/M, C*).
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
31
This allows us to write (5.1) as the sum over ,w E '0o(M,, XM, t) and
over the orbits {a,} E {lI,.} of Iro(M+)* in II,, of the expression
ml
.(aw)r(w)tr(Rp(aow, w)p(a f )) .
where
m.(ar)
=
E
CEro(M+)*
mw (C(w)C(M . )
Applying Fourier inversion on 7ro(M+) to the formula (4.6) (with G
replaced by Mw), while taking into account the property (i) of the
local Conjecture 6.1 in [3], we obtain
m(a)
=
ISwI1 uES,pw e,(u) < U, (T
>.
(5.1) equals
xi xi is i E
Therefore
;w
{ow}
uESp,.
Ew
,(u) r(w) < u,o'w> tr(Rp(crw,
w)Zp(, f))f
Suppose that b,,, belongs to I0o(Mw, XM, t). Let Vb denote the composition of i,, with our embedding LM C LGO. We claim that b
is well defined (as an equivalence class of parameters) in '(G, X, t).
Recalling (§4) the definition of equivalent parameters, we note that it
is enough to show that the map
(5.6)
ker1(F,
Z(G))

ker' (F, Z(M))
is an isomorphism. By the obvious transitivity property, we can in
fact assume that M is minimal, and hence a torus. Then Z(M)/Z(G°)
is a maximal torus in an adjoint group, on which the Galois action is
dual to a direct sum of permutation representations. The bijectivity
of (5.6) then follows from the exact sequence
7ro((Z(M)/Z(G))r )


H1(F, Z(G°))
H1(F,Z(M/))

and its analogues for the completions of F.
4.3.2(a) of [12].) This proves the claim.
H1(F,Z(M)Z(G)),
(See the proof of Lemma
32
JAMES ARTHUR
Thus, 0, maps to an element / in i(G, X, t), to which we can associate the objects S$,= S,.(G), HIT = 1,(G) and e0 = , for G. The
next step is to apply a conjectural formula [3, §7] for the trace of the
normalized intertwining operators in terms of the pairing on S,x II.
As it is stated in [3], the formula applies to the local intertwining
operators and pairings, but the product over all valuations gives a
formula for the global objects. In fact, certain constants in the local
formula (namely, c(ax, nw), Aw(IF) and c(rx, nG), in the notation of
[3, §7]) have the property that their products over all valuations equal
1. The global formula is therefore simpler. If bM denotes the parameter ,,w, but regarded as an element in I(M) rather than i(M,),
then the orbits {fa } above will be in bijective correspondence with
the representations a E Hp1M which extend to Mw(Ab)+. It follows
from Conjecture 7.1 of [3] (and also the two remarks made after the
conjecture), that
E <
u,aw, > tr(Rp(aw,, w)Zp(a, f))
equals
< x,7 > fGc() ,
w
where xu stands for the image in SO of the point
We have now obtained an expansion
u
E Sq,.
<
5 Ewo (u)r(t&W)
fG(7r),
1«IS.l1
rEII
w} E1o(Mw,XM,t)
uESp,
'
E
for the trace (5.1). We shall substitute this into our formula for
Idisc,t(f) Observe that
S
<
U,
trEIp
> fG(r)
=
lwo(G+)l 7rE{nJ1,} < xU, > fG(7)
S
Therefore, Idisc,t(f) equals the triple sum over {M}, w E WG(aM)reg
and ,6 E T'o(MW, XM, t) of the product of
IWc(aM)Ildet(w  1)aG

with
(5.7)
ISW1J1
uESp,
e,
(u)r(w)
S
arE{nSl}
< x,,7r > fG(r).
33
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
We propose to interchange the sum over the parameters with the
sums over M and w. The outer sum will then have to be over all
parameters 1 E I(G, X, t). For any p there will be an M, unique up
to conjugacy, such that pb is the composition of a parameter mM E
To(M) with the embedding LM C LGO. The condition that OM
also belong to
for a given w E WG(aM)reg, is that the set
SM(Mw,) be nonempty. There is another way to state this. Recall
that we have identified WG(aM) with the dual Weyl set WG(aM).
Then SM (Mw) is nonempty if and only if w belongs to the subset
WV = W1(G) of elements in W0(aM) which, modulo the isomorphic
groups (5.6), centralize the image of 4p. It will be convenient for us
to regard this subset Wp as the full Weyl set associated to Sp =
S,/Z(G°). It acts on the maximal torus
0o(Mw),
T =
Afz(GO)/z(Go)
of the connected component
Sz(G)/Z(G ).
For any w E Wp, we shall write det(w  1) for the determinant of
(w  1), acting on the Lie algebra of TV. One sees easily that
Idet(w 1) Idet(w 1)GO Idet(w 1)aG lldet(w 1) ~o1
Now it is well known that Idet(w  1) o I equals the order of the kernel
S =
=
of w,
=
.
acting on the dual torus
(Z( )r) /(Z(Gf) )
(See [23, 11.1.7].)
The action of w on this torus is of course inde
pendent of w, and the kernel is just the finite group of components
in
KG
Z(G)r n (Z(G0)r).
=
Therefore
Idet(w 1).aG1
M
In particular, w belongs to
Wb,reg
=
=
Idet(w l)l17o(KG)l'
.
WG(aM)reg if and only if it lies in the set
{W E Wo : det(w  1)
6
0}
34
JAMES ARTHUR
of regular elements in W,. When this is so, the associated parameter
in o0(M,) in fact belongs to To(Mw, XM, t). We shall denote it by
iw, as above.
Actually, 0hw is not uniquely determined by t, and w. We must
decide how many parameters in To(Mw, XM, t) lie in the equivalence
class of /. Keeping in mind the isomorphism (5.6), we see that two
parameters u, map to the same b if and only if they are conjugate by
an element in WG(aM). Moreover, two such conjugates are equivalent
in 'o (Mw, XM, t) if and only if they differ by an element in W4(G°).
The number of 0hw associated to / is therefore
] G(aM)l[W((G0)[Thus,
over
1
=
IWG(aM)llIW,1
interchange of summation expresses Idisc,t(f)
E6 Ti(G, X, t) and w E WP,,reg of the product of
our
as
the sum
Iwo(KG)IlWO'l Idet(w  1)1
with
(5.7).
Suppose that b E @(G).
As in the case of a local
can define the finite set
PA4 =
PA(G)
=
parameter,
we
Norm(T,3,S)/TV
Norm(A,$S,)/AMZ(G°).
Let S$be the subgroup of elements in PA(G°) which act trivially
on Tp,. This group acts freely by translation on N^P, and the set of
orbits can be identified canonically with WV,. One sees easily from
the isomorphism (5.6) that
w W,
SO, = SM. W = S, .W,
=
for M as above. We also have the Weyl group
W, =
=
Norm(Tp,,S})/T,
Norm(AM, S})/AMZ(G°)
A/J,
of the connected component S,. This too acts freely on
and
the set of orbits can be identified canonically with S,. We obtain a
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
35
commutative diagram
1
1
I
I
it ,S R
Sc
I I
S~
~
1
1
~i 1
1
the local case [3, (7.1)]. The dotted arrows stand for splittings
of short exact sequences determined by a fixed Borel subgroup of
S,0 containing Tp. Similarly, one obtains a commutative diagram of
groups if one replaces JAp, S,,, WV, and R, by the respective finite
groups fig, S+, W and R+ they generate. We shall write u + xu
and u + wu for the projections of Af+ onto S+ and W+. Notice
that if x is any element in S,, and A/(x) is the corresponding orbit of
W, in A0, the second projection maps Ar(x) bijectively onto a subset
W(x) of W,. We shall set
as in
W(x)reg
=
W(x)
n
WO,reg
and
j^(Z)reg
=
{(
E
^(x) :
wu E W(X)reg}
We apply these observations to our formula for Idisc,t(f). According
to the horizontal exact sequence for Af, in the diagram, the double
sum over w E W,,reg and u E S$, =
can be combined into a
sum
over
the
elements
in ',. We shall write
simple
regular
S$w
(5.8)
for
set
e
(u)
=
+,(u)
any point u E A/, whose projection onto W, equals w. We also
(5.9)
r,(w)
=
r(w)
36
JAMES ARTHUR
Then
eM and r, extend to well defined characters on l+ and W,
respectively. The simple sum in its turn can be decomposed by the
corresponding vertical exact sequence into a double sum over x E S$
and u E
V(x)reg. Observe that
IWllS,,, = IWllS1
=
I1t1
=
IsiIWSIl
ISllW(x)l
It follows that Idisc,t(f) equals the sum over i in '(G, X,t) of the
product of
70 (G) I1'IS 1
=
with
E
Eiw()K1 UEn(Z)reg
xESp
ej(u)r0(wu)Idet(wu 1)11
< x, r >
fc(r)
7re{nH}
Any element w E WI operates on TV. It preserves the set ES
of roots of (S?,,T ). We shall simply write e(w) for the usual sign
attached to this permuation, namely the number (1) raised to the
power
1(E+)
where
n
(wE+)l,
S+ is the set of positive roots in E, relative to some order.
PROPOSITION 5.1. We have
ro(wU)
=
e(w)(x)
(U)
for any element u E Af.
This proposition is the motivation for the introduction of the characters eF, into the multiplicity formula of Hypothesis 4.1. We shall
prove it in the next section. In the meantime, we can combine it with
our formula for Idisc,t(f).
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
PROPOSITION 5.2. The distribution
Ilo(KG)l with
(5.10)
E
E
E.I(G,x,t) rE.{n+}
IS I
37
Idisc,/(f) equals the product of
aEST
E(X)i(X) < x, r > fG( ),
where
(5.11) i(x)
=
IW(x)l' E e(w)ldet(w 1)11 .
wE W(X),eg
§6. THE
SIGN CHARACTERS
AND ro,
ES
In this section we shall pause to study the characters e, and r,.
=
Our goal is to prove Proposition 5.1. Recall that e, E,
is the
onedimensional character (4.5) on S$= SG(G+) which comes into
the conjectural multiplicity formula. The function re is the one dimensional character ((5.5), (5.9)) on WT = W,(G+) defined by the
that
and W+ are both
global normalizing factors. We have seenS,
quotients of the finite group Af,. We can therefore identify e and re
with characters on K'+. Proposition 5.1 can be regarded as a formula
for the quotient of these two characters.
We shall begin by expressing r, in terms of the orders of certain
Lfunctions at s = 1. Let EM denote the set of roots of (GO°, A).
For each a E EM there is a representation pa of LM on the root
space g. Having already fixed the dual parabolic subgroups P and
LP = PWF, we shall write Ep C EM for the set of roots of (P, A).
Fix an element w E W, and set
EP,w = {a Ep :
Then there is a decomposition
PP,w
=
w
$
E (Ep)}
P&
&EtP,w
for the representation of LM which occurs in (5.5). Notice that the
Killing form provides an isomorphism between pa and the contragredient pa. The formula (5.5) becomes
(6.1) r4(w)
=
limn
][ L(1A(a),p&o)o)L(l+A()(),pa&o)1
rEt.p,w
38
JAMES ARTHUR
since
L(1,pa& o0,,) = L(1 + (&), pa o ) .
We are going to show that rp(w) equals the character
(lord.=1(L(sp&ook))
(6.2)
We claim that for
Ep,w such that
every root a E Ep,,,
'
Pa&o°
To this
P o1
p4,I
there is also a root ai E
.
end, observe that
Pa O ~ pwa o ad(w) o  pw o .
The first of these isomorphisms is given by the intertwining map
Ad(w): .  g ,
and the second follows from the fact that the image of w E Wq, under
the adjoint representation commutes with the image of LF x SL(2, C).
Now, consider the orbit
O,(a)
=
(w&a: j
Z}
of a under the cyclic group generated by w. The representations
(p
0o:
P E Ow(a))
equivalent to the contragredients
({ o : P E Ow(a)}.
But after a moment's thought, we see that the intersections of O(w(&)
with Sp,W and (Ep,w) contain an equal number of roots. The claim
follows. In particular, the terms in the product in (6.1) can be grouped
in such a way that pa appears in the numerator as well as the denominator. This leads directly to the formula (6.2) for re(w).
are all equivalent, and are also
Recall that
(pa o+0)(t)
=
p
(i(t, (Itin Itl¾)))
t ELF
39
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
Now there is a decomposition
(pao0 )
(Yj Vj),
=
jEJ(&)
where each pj is an irreducible unitary representation of LF and vj
is an irreducible representation of SL(2, C). Therefore, (6.2) can be
written as a product
TI IT ( 1)ord=(L(s,1j®v,))
(6.3)
&ES:p, jEJ(&)
where
L(s, pj x vj) stands for the Lfunction of the representation
t ,
j(t)vj (' It
tl
)
tE LF,
,
of LF. From the discussion above we see that the contragredient acts
as an involution pj ® Vj + ij ® Vj on the constituents of pp,,. It is,
moreover, an easy consequence of the unitarity of yj that
vj)
=
L(s, j ® vj),
ord8=1 (L(s, pj ® vj))
=
ord8=1 (L(s,
L(s,
j®
so that
ij 0® i))
In particular, the contribution to (6.3) of a distinct pair of contragredient constituents cancels. The product (6.3) need only be taken over
those constituents with
j 0 Vj
Vj
.
Since any finite dimensional representation of SL(2, C) is self contragredient, the condition is just pj Jij.
The question then is to determine the sign
(6.4)
(_l)ord.=1 L(s,p®v)
40
JAMES ARTHUR
for any irreducible representation ®0 v of LF x SL(2, C) such that p
is unitary, and i  p. Set m = deg(p) and n = deg(v). Then v maps
the matrix
( It
\t
)
tE
LF,
to the diagonal matrix
diag(tl2(n1), Itl(n3),
in
...,
t
(n1))
GL(n, C). Therefore
L(s, p v)
=
i=1
L(s + (n 2i + ), )
We must therefore describe the order of zero or pole of L(s, p) at any
real halfinteger.
Hypothesis 4.1 includes the global Langlands correspondence for
GL(m), which asserts that
L(s, p)
=
L(s, r)
for some unitary, cuspidal automorphic representation wr of GL(m, A).
(See [3, §2].) Then L(s, p) can have a real pole only if p is the trivial
one dimensional representation, in which case there is a simple pole at
s = 0 and s = 1 [7, Corollary 13.8]. Results of Jacquet and Shalika
[8, Theorem (1.3)], [9, Theorem 5.3] imply further that the only
possible zero of L(s, p) at a real half integer is at s = 1, the center of
the critical strip. The poles of L(s, p) will contribute to (6.4) if n is
odd. However, if p is trivial and n is of odd dimension greater than 1,
the poles at 0 and 1 will both contribute, and their effect will cancel.
The zeros of L(s,/p) will contribute to (6.4) if n is even. From the
functional equation
L(s,p)
=
E(s,p)L(1  s,1) ,
we see that L(s, p) has a zero at s = 2 of even or odd order,
to whether e( ,p) equals +1 or 1.
We have established
according
41
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
LEMMA 6.1. If n = deg(v) is even, the sign (6.4) equals e(, p). If
n is odd, (6.4) equals 1 unless # ®0 v is the trivial representation of
0
LF x SL(2, C), in which case (6.4) equals (1).
If we substitute the formula of Lemma 6.1 into the product (6.3),
we obtain a new expression for rp(w). To describe this in a convenient
way, we shall define a character e1/M which is closely related to the
original characters e'c and em. Let m denote the Lie algebra of M,
and let Adg/m denote the adjoint representation of LM on g/hM. The
group
Norm(AM, S)/Z(G°)
also acts by the adjoint action on g/ml, and it commutes with the
composite representation Adg/m o im of LV = LF X SL(2, C). Now,
4,
=
Norm(T t,,S)
=
we have a decomposition
Adg/,,_mo
(&
=
(9j®vj)
&E.M jEJ(&)
into irreducible representations of L'. Let us write (g/im), for the
direct sum of those irreducible constituents pj 0 vj such that (i) Jj pj, (ii) e(½,pj) = 1, and (iii) deg(vj) is even. Then (g/im) is an
invariant subspace of both L' and NI. Define
(6.5)
E,/M(u) det(,i EndL E((_g/n))) ,
=
where i is any element in
equals u. Observe that
N+
_(g/m)
whose
=
projection onto A/'+f
©
u
A+,
= N,/T,
ge,
&EEm
where
By
=
L n
(g/'m),/.
The subgroup S, of N" leaves invariant each of the subspaces g
of (g/fm)4. Since the actions of SI on g, and g
are contrais trivial on S1, and descends to a character on the
gredient,
quotient
G/M
JZT+V
Ar+/
/ cl
W+.
42
JAMES ARTHUR
Of course the main reason for defining eGIM is the formula
(6.6)
E£(u)
=
G1M (U)M(),
U
E
which follows easily from (4.5'), (6.5) and the corresponding formula
for e.
To express r,in terms of eIM, let EM,+ be the set of roots a E
SM such that the dimension of EndL' (g ) is odd. It follows from
properties of the determinant that
eGIM(w) = (1)I
,
w
E
W .
This is just the contribution from the even dimensional representations vj to the expression for r,(w) given by (6.3) and Lemma 6.1.
The contribution from the odd dimensional representations vj is simply the usual sign character e(w) attached to the group S. Thus
rg(w)
=
e(w)/GIM(W),
w
EW .
The required formula
rT(u)
=
E(Wu)Ew(U)EM(u)1,
uE
,
0
of Proposition 5.1 then follows directly from (6.6).
The formula (6.6) can be regarded as motivation for the definition
of e6. The introduction of this character might have seemed odd at
first. However, we now have a direct connection between eG and the
more familiar function rV obtained from the normalizing factors of
global intertwining operators.
§7.
THE EXPANSION
We turn now to the distribution
esis 3.1 as the sum
(7.1)
H
Edisc,t.
GH)Sdit(f)
,
OF
Edisc,t(f)
It was defined in
f E CC
Hypoth
(G(f), X),
over equivalence classes of elliptic endoscopic data. We shall convert
this into an expression which is parallel to the expansion (5.10) for
Idisc,t (f).
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
43
Hypothesis 3.1 can be regarded as a general existence assertion.
There should be a stable distribution on any quasisplit group with
the property that (7.1) equals Idisc,t(f) for any component G at all.
Our ultimate goal is to show that this assertion is compatible with
the formula (5.10) for Idisc,t(f). Since the stable distributions are
uniquely determined by the property, the problem is simply to show
that they exist. For a given quasisplit group G1, and a suitable
character X1 on a subgroup X1 of Z(G,A), we shall try to construct
the associated stable distribution SI
Sldisc,t in terms of the parameters
1 E I(G1, Xi, t). Our local assumptions in §4 attach a stable distribution
_i
on
,^
f01'
fi E CC (G (^A), X),
G1(A) to each parameter t1 E TI(G1, Xi).
(7.2)
SId ,t)=
E
l E.
(Gl ,Xl,t)
Let us therefore set
(Sf),
where
SIG (fl) = a(G,
l)f (),
for constants a(G1, 4il) to be determined. We shall assume that the
constants vanish unlessi1 belongs to (1G1, Xi, t), a countable subset
of iI(G1, Xl,t). We shall attempt to define them so that the formula
obtained by equating (7.1) with the right hand side of (5.10) is universally valid.
We fix a representative (H, H, s, ), for each equivalence class of
endoscopic data for G, such that X is a subgroup of LGO and ~ is
the inclusion mapping. We also fix a distinguished splitting (Hi, 1)
of (H, H, s, C). The character X1 is then defined on a subgroup X1 of
Z(H1, A) as in §3. We begin with the formula
(7.3)
Edisc,t(f) =
(G, H)
H
E
,E((Hl
1
,x
,t)
SI II (f H )
obtained by applying the definition (7.2) to the groups H1 in (7.1).
Our immediate goal is to convert the double sum over H and /1 to a
single sum over the orbits of G° on a certain set. In the process, we
will need to apply the formula (3.5) for the coefficients L(G, H).
44
JAMES ARTHUR
Recall that
@T(G) denotes the set of maps
': LF x SL(2, C)
LGO
satisfying certain conditions,
and taken modulo the equivalence relation (4.2). Let us write I(G) for the same set of parameters, but
without the equivalence relation, and let I(G)/G° denote the set of
G°orbits in 2(G). (We can also write 'disc(G), '(G, X, t), etc., for
the obvious subsets of QI(G).) We shall describe the order of the covering projection T(G)/G°  1(G). According to the definition (4.2),
the group ker (F, Z(G°)) acts transitively on the fibres of the projection. The isotropy subgroup is just the image of S,(G°)/Cp under
the injection (4.9). But the finite group S,(G°) is bijective with the
set Sp. Therefore, the order of each fibre in the projection equals
Iker1(FZ(G0))I ISgO1 IC0.
(7.4)
We shall apply this remark to the quasisplit groups H1 which occur
in (7.3). We can replace the sum over
l 'i E'1O(Hl, X1, t) by the sum
'
over Q(H, Al,t)/Hl1, provided that we divide by
Iker1(F.Z(H/1)) ISOl 1 ICl1,1
the analogue for Hi of the integer (7.4). Since (Hi, i) is assumed
to be a distinguished splitting, kerl(F, Z(Hi1)) equals kerl(F, Z(H)).
Combining this with the formula (3.5) for t(G,H), we are able to
a o as the sum over H and over i E o(H1 IXi, t)/H1 of
write Edisc,t()
Iker (F, Z(G0))I1I17I0(G)11Z(H)'1FIS1
(7.5)
xl Cl IIAut(H)/HZ(Go)rI1SI (a H).
Keep in mind that H really stands for the equivalence class of an
endoscopic datum (H, H, s, ). Now, suppose that we are given a
parameter 41 E ;F(H1, XI). Then 41 factors to an Lhomomorphism
from WF into ', which may then be composed with the embedding
t  LGO. In this way we obtain a parameter 'k E I(G,X).
·It :follows
from the property (2.1) of endoscopic data that the coset
s E G/Z(G°) lies in the set
S
=
S/Z(G°) = SO(G)/Z(G°).
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
45
Conversely, suppose that we are given a parameter b E I(G, X) and a
consisting of semisimple elements. Then we can define an
endoscopic datum (H, X, s, ~) as in §4. Recall that H is the quasisplit
group whose dual group is
coset s . Sg
H = Cent(s,G0)0,
equipped with the Laction induced by
'H = Htb(LF x SL(2, C)),
and g is the inclusion of ' into LGO. The parameter , then factors
through 'H. For any distinguished splitting (H1, g1) of the endoscopic
datum, we obtain the character X1 :X1 : C* as in §3, and ib then
yields a parameter il E IQ(H1, X1).
We have just established a correspondence between the pairs
(H,l1) and (,,s). We want the datum H to be elliptic and the
parameter ¢1 to be weakly elliptic. We ought to describe these conditions in terms of (t, s). Since b1lfactors through H7, and H1 equals
l(H[)Z(Hl), we have
Co Z(H1)/Z(H1) Cent(Image(p), H)Z(H)/Z(H).
=
In other words,
C, Z(H,)/Z(Hl) (Co n H)Z(H)/Z(H).
(7.6)
In particular, there is an isomorphism
ClZ(Hl)/Z(Hl) (C
v
n
H)°Z(H)/IZ(H)
of the two identity components. Notice that (C4, n H)) equals CO,s,
the connected centralizer in C, of any element in the coset s. Conse
quently
LZ(Hl)/Z(Hl)C°,sZ/c(I/Z(tH) C0,/,
z(Hf).
Thus, t1 is weakly elliptic if and only if the center of C ,,/Ci,, fn
Z(H)r is finite. Now C, n Z(H1)r is a central subgroup of CO,,
=
which contains
C
Aft
n
(Z(H)r)0. Therefore, the conditions that p1
46
JAMES ARTHUR
be weakly elliptic and H be elliptic, taken together, are equivalent to
the condition that C',, has finite center modulo AG = (Z()r)°. We
can describe this more simply in terms of the set
S3,fin = {s E SO : IZ(S,8)I
For by
<
oo}.
(4.11) we have
s,° = C,8Z(G°)/Z(G°) C°sl/C,, , n Z(G°)
=
,/c4
C.
n z(G)r.
Thus, the correspondence is between elliptic pairs (H, 1) and pairs
(b, s) such that s belongs to 9Sfin.
The foregoing discussion will enable us to interchange the order
of summation in the original double sum over H and t1. Keep in
mind that (H, H, s,,) stands for a representative of an equivalence
class of endoscopic data for which KH is a subgroup of LGO and ( is
the inclusion mapping. The equivalence classes themselves can be
identified with the G°orbits of such data. The stabilizer in 0G of
(H, 7, s, Id) is the group Aut(H) which appears in the expression
(7.5). The group Aut(H) in turn acts on the set of parameters i E
I(G, X, t) such that s belongs to S,fin. The stabilizer in Aut(H) of
a given 0 is simply the group
Cs = { ECe : csc1 = s}
of elements in C, which fix the coset s. On the other hand, we can
identify the orbits {1 } E I(H1, t)/Hl with the Horbits of {/}.
This is easily seen from the injectivity of the map
X1,
H1(r, Z(H))
H
(F, Z(H1)),
noted in §2, and the fact that H1 = Z(Hi)1(H). We can actually
take HZ(G°)rorbits of {b}, since Z(G°)r centralizes the image of
has finite index in Aut(H), by Lemma
b. But the group
of
in
the
stabilizer
and
i HZ(G°)r is the subgroup
2.1,
HZ(OG)r
C+, n (HZ(G°)r)
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
47
of finite index in C,. Therefore, we can replace the original double
sum over H and t1 by the sum over the G°orbits in the set
{(I),s):
E %l(G, X, t),
s E S,fin},
(7.5) by
(7.7) IAut(H)/HZ(G)l  C/C' n (HZ(G ))The stabilizer in G° of a given parameter ,b E 1(G, X, t) is the group
C,. We can therefore replace the sum over G°orbits
in
{(I,
s)} by a
double sum over bE F(G, x, t)/G° and over the set Orb(Cp,S,fin)
of orbits of Cp in Sfin. Obviously, SP,fin has the same set of orbits
if we multiply the summand
under C1 as under the group
Cp = CpZ(G°)/Z(G°).
The stabilizer of s in Cp equals
C,$ = ,CsZ(G°)/Z(G0) = Cent(s, C).
However, we would prefer to take the orbits in S,,fin under the connected component
Co= C Z(G°)/Z(G°).
+,
The C,orbit of s is bijective with C4/C while the C,orbit is in
bijective correspondence with the quotient of C, by the group
Cs = Cent(s, C+).
Therefore, we can indeed take the second sum over
vided that we multiply the summand by
or what is the same thing,
(7.8)
\cJ C\1
Csorbits, pro
48
JAMES ARTHUR
Finally, we can take the first sum over .E W(G,X,t) instead of
(G, X, t)/G°, if we multiply the summand by the integer (7.4). We
have shown that Edi8c,t(f) equals the sum over t E '(G,X,t) and
s E Orb(C',, S,,fin) of the expression obtained by multiplying (7.4),
(7.5), (7.7) and
the product of
(7.8) together.
We
can
write this last
(7.9) IC,s/Cns
+ '"0+1 n (HZ(G0) ILkC 1) IC+
(7.9)
expression as
IZ() I1
1
and
Iro(,G)l1 IS,11 Is I sI'l(fHI).
(7.10)
(7.9) can be simplified. We begin by writing
co,, (C, n H)Z(H)/(C, n H)°Z(H)
(Co n H)I(C,, n H)°Z(H)r
(C, n H)Z(G°)/((C n H)Oz(H)' · Z(G°)).
The first isomorphism follows from (7.6), while the second is trivial
The term
and the third is a consequence of the fact that
(C, n H) n Z(G°) = Z(G)r n Z(H)r c Z(H))r
We also observe that
C/C,, n (HZ(G°)) = C,/(C, n H)Z(G
= C,+Z(G°)/(C,
n
H)Z(G0).
This allows us to write
, /,+Sn (HZ(GO)r)l1 Ico, I
= IC,sZ(G°)/(C, n H)Z(H)rz(G0),
for the first two factors in the product (7.9). Let us divide both groups
in the quotient on the right by Z(G°). The numerator becomes
C
C,z(GO)/Z(Go) =O'sI
=
49
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
and the denominator may be written
(c, n H)°Z(H)rZ(o°)/Z(o0)
= ((C4, nH)OZ(GO)/Z(GO))(Z(f)rZ(GO)/Z(GO))
= (c, Z(o)/Z(G° ))))Z(H)r
,S2(//)r,
=
by (4.11). Therefore, (7.9) equals
+ff ^, /CL
C+/C1
H
I
= ICf,/C K(IC/)rn Z(H)(fK)r
I0C0s
/ I8'K
K,8 n Z(H)IL,.
0=
(//)r1.
We noted in §4 that

C,,s = S,,a. In particular
100,J/00,s1 =13Is/ 81 i ro(3,,)l1.
=
The term
(7.9) can by consequence by written as
Iro(3,s)l' s1,, n Z(H)rl1.
We have now established that
Iro(KG)11 with
E I I1
E
0sEOrb(S, ,S,,nfn)
OE*(GXt)
Edic,i(f) equals
the
product
of
Io(Ss)lI
ISO,1Io,0,nz()rl sI,(f).
By assumption,
SI,/'(fH1)
Part of our local
the
image x
=
r(H1,l). f' (1).
assumption in §4 is that f H1 (1) depends only on
= s of s in the set
=
=
S= S,
Sw/Z()
Z()
/
=
o(S).
50
JAMES ARTHUR
More precisely, formula
(4.4) asserts that
fH (1) =
<
E
.rE{n,}
x,
r
>
fG(Tr)
where < *, > is the global pairing on S, x II+. We can therefore
write Edisc,t(f) as the product of Iro(nG)l1 with
(7.11)
E 7r.E{Hl}
E l IS01 xEro(S,)
E ( sEOrb(ST,xffi)
E
I
ro(
s)1r(0, s)) < 8+x,x
>
fG(7r),
where
Xfin = x n S,fin
and
T(, s) = IsVl IS, n Z(H)r l(Hi i).
The similarity with the formula in Proposition 5.2 appears promising.
We must try to define the constants a(H1, b1) so that the two formulas
for Idisc,t(f) always match.
Suppose that G1 is an arbitrary connected quasisplit group over
F, and that '1i E T(G1). Then we shall set
(7.12)
(, )u(S1)
a(G1, 1) I=S, I1e(
is the sign character (4.5) and a(3S,) is a constant,
where Fe, = ,G
to be determined, which depends only on the
the complex, connected reductive group
isomorphism class of
S,= (Sl/Z(G1))°.
We also ask that this latter constant have the property that
(7.13)
a(S1) = a(Si/Zi)iZl 1,
for any complex connected group S1, and any subgroup Z1 of the
center of S1. In particular, o(S1) is going to have to vanish if S1 has
infinite center. This implies that a(G1, i1) = 0 unless z/1 E I'(G1),
as we would expect.
51
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
We of course want to set G1 = H1. Then
S,
=
Cl Co1Z(Hi)/Z(Hi)
= (C1 n H)°Z(H)/Z(H)

= CoZ(H)/Z(H),
by the formula (7.6). Since
,, n Z(H) = C, n Z(H)rZ(G),
we obtain
C
z(HZ)/Z(H) co
H
G
H
G
a( ) = a(~,,()° s n Z(G))r (l.)
It follows from the property
(7.13) that
Therefore,
n )rl(
a(IH, il) = IS 1
Ke(5)Z)r/2(
so that
LEMMA 7.1.
(,S) =, (1) s).
For H1 and
S= Eas
(7.11), we have
,xI E ()
,
, in )S.).
(S0)
PROOF: As in
§4, let
T
(Ak Pk 0 Vk)
=
k
,
52
JAMES ARTHUR
be the decomposition of the representation
SL(2, C)  GL(g)
TV, : Sx Lx
into irreducible constituents. If I denotes the set of indices k in the
direct sum, let I' denote the subset of k such that (i) pk  pk, (ii)
e(, pk) =1, and (iii) dim(vk) is even. Then
x
E
=
det(Ak(s)),
S, 45() JI
kEI'
where s is any element in Sp which projects onto x. Notice that the
element sv, lies in both S, and SL(2, C). If k belongs to I', we obtain
Ak(S=) Vk(S) = 1,
=
since dim Vk is even. It follows that
(so)
=
H det(Ak(so))
kEPI
=
J
(_)dim(Ak)
kEI'
Now H1 is a central extension of the endoscopic group H attached
to s. The Lie algebra of H equals the centralizer of Ad(s) in g, and
the Lie algebra of H1 can be identified with the direct sum of this
algebra and a central ideal. For each k, let A) be the space of sfixed
vectors for Ak. This of course is just the intersection of the underlying
space of Ak with the Lie algebra of H. Recalling the relation between
b and li, and applying the formula for eG(gs,) to H1, we obtain
E,
(S1)=
(l)dim(\)k
kEl'
Finally, we observe that the number
EC(x) = I det(Ak(s))
kEI'
equals the product
the operators
of all the
eigenvalues, counting multiplicities,
{Ak(s): k E I'}.
of
53
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
Now the contragradient Ak  Ak defines an involution on the representations Ak with k E I'. In particular, if f is an eigenvalue, not
equal to ±1, then ~1 is also an eigenvalue with the same multiplicity.
Therefore,
(x) = (l)m(1),
where m(1) is the total multiplicity of the eigenvalue
same token,
(1). By the
Z(dim(Ak) dim(Ak)) m(1)
kEl'


is an even integer.
Consequently,
EG()E 1() = (1 1) = eG(X).
We obtain
e(Sx) (X)eG(xe) )= £G(X)E()1 =
as
( )
required.
O
The lemma allows us to write
r(1),s) =
G
)a()o )
Substituting this into (7.11), and setting
(7.14)
e(x) =
I7ro(S,) I1
(S ),
sEOrb(S 0,xain)
Edisc,t(f) equals the product of 17ro(CG)l 1 with
E £^(sx)e(x) < Sx, r> fG(r).
K
Z IS1'1
xE
E*I.(G,x,t)
we see that
rE{In,}
7ro(So)
point s,, E S,(G°) belongs to the center of S (G+). Consequently, for any point s in the component x, the group Sg, equals
S . It follows that e(x) equals e(spx). Substituting this into the
formula above, and changing variables in the sum over x E 7ro(S,),
we see that Edisc,t(f) equals the product of I[o(rcG)K1 with
The
(7.15)
E
0.*(G',X,t) "rE n }
IS 11 x.ro(Sp)E G(x)e(x) < x,
>
G(T7).
54
JAMES ARTHUR
We have now reached the stage in §7 at which we concluded §5.
Taking the two sections together, we see that Hypotheses 3.1 and
4.1 yield two parallel expansions for Idic,t(f) and Edic,t(f) into irreducible characters. Our goal is to show that these two expansions
are in fact the same. The expansions are given by (5.10) and (7.15).
They differ only in the coefficients i(x) and e(x), which are defined for
any component x E ro0(S,) by (5.11) and (7.14). We must then show
that the coefficients are equal. Recall that e(x) depends on a constant
a(Sl), which is to be defined for any complex, connected reductive
group S1 and which satisfies (7.13). We must show that a(S1) can be
defined for each S1 in such a way that i(x) and e(x) are equal for any
x. This is a property of Weyl groups which we shall establish in the
next section.
§8. A COMBINATORIAL FORMULA FOR WEYL GROUPS
Suppose that S is a union of connected components in an arbitrary
complex, reductive algebraic group. Then S+ is the reductive group
generated by S, and SO is the connected component of 1 in S+. Recall
also that we are writing Ss for the centralizer in So of any element
s E S. This group is of course not always connected. As a slight
generalization of (5.11), we set
i(S)
(8.1)
=
IW°V1
E e(w)ldet(w 1)11,
W
EWreg
where
W° =
is the
W(S°)
=
Norm(T,SO)/T
of SO relative to a fixed maximal torus
is the set of elements w in the Weyl set
Weyl group
Wreg W(S)reg
=
W =
W(S)
=
T, and
Norm(T,S)/T
such that det(w  1) 0. The determinant can be taken on the real
vector space aT = Hom(X(T), R). As in §5, e(w) = ±1 is the parity
of the number of positive roots of (S°, T) which are mapped by w to
negative roots.
As in §7, we shall write Orb(S°, E) instead of E/So for the set of
orbits under conjugation by So on an invariant subset E of S. This
will prevent any confusion of orbits with cosets. The main example is
when E equals the subset
Sfin
=
{s E S:
IZ(S,°)
<
o} ,
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
55
in which case the set Orb(S°,Sfin) is finite.
Our object is to prove
THEOREM 8.1. There are unique constants o(S1), defined for each
connected and semisimple complex group S1, such that for any S the
number
(8.2)
e(S)
ro(S)I (S0)
=
sEOrb(S° ,Sfin )
equals i(S). The constants have the further property that
(8.3)
o(S1)
=
o(S1/Z)IZ11
for any central subgroup Z1 of S1.
Remarks. 1. It is obviously enough to prove the theorem when S is
just one connected component. We shall assume this in what follows.
2. Let us agree to write a(S1) = 0 if S1 is any complex, connected
algebraic group which is not semisimple. In particular, this constant
vanishes if S1 is a reductive group with infinite center. The equation
(8.2) can then be written
e(S)
Iro(SS)1a(S0)
=
sEOrb(S° ,S)
3. Theorem 8.1 is what remains to be
proved of the comparison of
Idisc,t(f) and Edi8c,t(f) that we began in §5 and §7. It is interesting to
observe that Theorem 8.1 is actually a miniature replica of the original problem. It is a formal analogue for Weyl groups of the problem
of comparing Idisc,t(f) and Edisc,t(f), and indeed of many of the comparison problems, both local and global, that arise from endoscopy.
I do not know whether it is part of a larger theory of endoscopy for
Weyl groups, or whether results of this nature are already implicit in
the representation theory of Weyl groups and finite Chevalley groups.
We shall begin the proof of Theorem 8.1 by taking note of the
uniqueness of the constants a(S1). For a given semisimple S1, assume inductively that a(S[) has been defined for any S1 of dimension
smaller than S1. Then a(S1) is determined by the formula,
a(Si)lZ(Si)I
=
ro(Si,)L(So,8)
i(S) sEOrb( S ,S,1Z(Si))
56
JAMES ARTHUR
which follows from the required equality of e(Si) with i(Si). In other
words, the special case of (8.2) that S = So = S1 provides a definition
of the constant a(S1).
Having defined the constants a(S1) we shall next establish the property (8.3). The argument is similar to part of the discussion of §7. Suppose that S is an arbitrary component, and that Z is a finite subgroup
of Z(S°) which is invariant under conjugation by S. Then S = S/Z is
a connected component of the reductive group S+ = S+/Z, of which
the identity component S° equals S°/Z.
(i) i(S) = i(S).
(ii) e(S) =e().
(iii) If SO = S, then a(S) = a(S)ZI1.
PROOF: The property (i) follows easily from the definition (8.1). We
shall establish the other two properties together. To this end we
shall assume inductively that (ii) holds for any connected group of
LEMMA 8.2.
dimension smaller than S.
If the group
Z(S) = Cent(S,Z(S5))
is infinite, the quantities e(S), e(S), a(S0) and a(S°) all vanish, and
there is nothing to prove. We can therefore assume that Z(S) is finite.
This implies that the group Z(S+) n S is also finite. Let s be a coset
in Sfin = Sfin/Z which does not lie in Z(S+) n S. Then
Ss =
is a proper subgroup of
Cent(S,S0)
O°. Since
so= soz/z= S= /SonZ
for any element s in the coset
that
a(s)
=
g,
our induction
assumption implies
a(s)lSo n ZI.
Let Ss be the normalizer in So of the coset s. The set of orbits in
Orb(S°, S) which meet s can be identified with Orb(S, 4), a set of
cardinality
\z\\9/5s\1.
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
57
Observe that
kto(S.)l'(S°)
= IZllS /S.ll IS/S°l a(8°)ls°n zi'
= I(s,/Z)/(SOz/z)l'(o)
= Iro(&,)l'a1(S°) .
Z
sEOrb(S ,8)
Summing over all such s, we obtain
e(3)  a(S°)lZ(S+)n Sl.
If Z(S+) nS is empty, it follows immediately that e(S) equals e(S).
Suppose that Z(S+) n S is not empty. Then S acts on the group So
e(S)  a(S0°)Z(S+) n S
=
by inner automorphisms, and we have
e(S) = e(S°) = i(S°) = i(S)
from the definitions. (The equality of e(S°) and i(S°) was part of the
definition of a(S°).) Similarly e(S) = i(S). The property (i) then
implies that e(S) equals e(S) in this case as well. This is the required
property (ii). Suppose that S = S°. Then Z(S+) n S equals Z(S), a
group which of course is not empty. The property (iii) follows from
=
°
the fact that IZ(S)I IZ(S)IIZI.
The property (iii) of the lemma is the required condition (8.3) We
still have the main part of the proof of the theorem, which is to show
that e(S) equals i(S). This of course is a problem only if S is not
equal to S°.
As a warmup, let us verify the equality of e(S) and i(S) in the
special case that SO = T is a torus. Then W consists of one element
w, the adjoint operation of S on T. We can assume that this element
is regular. Recall [23, 11.1.7] that
Idet(w1)1
=
ITwI,
where TW denotes the kernel of w in T. Since e(w) =
i(S)
=
ITw"1.
1, we obtain
58
JAMES ARTHUR
On the other hand,
Cent(s,T°)
Tw =
=
S,
any element s E S. The regularity of w means that s belongs to
Sfin, and that SO = {1}. Therefore a(S°) equals 1. But the Torbit
for
of s equals the product of s with {tlw(t) : t E T}, a subtorus of
T. This subtorus has the same dimension as T, and must therefore
equal T. In other words, the orbit of s equals S, so there is only one
summand on the right hand side of (8.2). We obtain
e(S)
as
=
,ro(Ss)L1
=
ITWl1 i(S),
=
required.
Now suppose that S is arbitrary. We shall use Lemma 8.2 to effect
a simplification. First, observe that i(S) and e(S) depend only on SO
and the set of automorphisms of So induced from conjugation by S.
We may therefore assume that S+ is the semidirect product of S° with
7ro(S+). Now, let SO, be the=simplyconnected covering of the derivedof
group of S°, and let Se Z(S°)° be the connected component
the center of SO. Then
c0
=

0 x0S
sc
cent
is a finite covering group of S°. In particular, So equals SO/Z, where
Z is the finite central subgroup of S°. It is then readily verified that
S = S/Z, where S = Ssc x Scent is a component which normalizes Z
and such that the identity components (S (Ssc)° and (Scent)0 equal
the respective groups S°, S°c and Scent above. Applying Lemma 8.2
and the calculaion above for tori, we obtain
e(S)  i(S)
=
=
=
e(Ssc x Scent)  i(Ssc X Scent)
e(Ssc)e(Scent)  i(Ssc)i(Scent)
(e(Ssc)  i(Ssc))i(Scent).
(We have also used the fact, easily verified from the definitions, that
e and i are multiplicative on products.)
It is therefore enough to show that e(S) equals i(S) in the special
case that S° is semisimple and simply connected. We shall assume
this from now on. If s is any semisimple element in S, the group
Ss =
Cent(s, S)
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
59
is then connected, by [24, Theorem 8.1]. In this case, it is part of
our definition that e(S.) equals i(S8). If t is a semisimple element in
So, the connectedness of St implies that the set
St
=
Cent(t,S)
is either connected or empty. We can assume inductively that if
dim(St) < dim(S), then e(St) equals i(St).
LEMMA 8.3. The required equality of e(S) and i(S) is equivalent to
the formula
(8.4)
s
i(S) = E i(S) .
E
Orb(So ,S)
tEOrb(S ,S )
PROOF: If s E S and t E SO are elements that
Ss,t =
commute, we write
Cent({s,t}, S).
It is obvious that
7To((Ss)t)
=
7ro(S,,)
=
7To((S),))
The left hand side of (8.4) then equals
E i(S8)
sEOrb(S°,S)
=
sEOrb(So ,S)
e(S,)
E lrto(s,it)l'(So,t)
E=
sEOrb(S°,S) tEOrb(S,,S.)
E
{(s,t)ESx S :st=ts}/SO
E
E
Iro(Ss,,,)1(S° , ,)
tEOrb(S°,S°) sEOrb(St,St)
=
1o(s,t)'os,)l(S,)
E e(St)
tEOrb(S°,SO)
This last expression would just be the right hand side of (8.4) if e(St)
were replaced by i(St). But if t does not belong to Z(S), dim(St) is
60
JAMES ARTHUR
smaller than dim(S), and e(St) equals i(St) by our induction assumption. It t belongs to Z(S), St is just S itself. Therefore, the equality
of e(S) and i(S) is indeed equivalent to the identity (8.4).
[1
It remains for us to establish the formula (8.4), in which S is a
component such that So is semisimple and simply connected. We
shall deal with each side separately. According to [24, Theorem
7.5], any semisimple element in S normalizes some pair (T1, B1) of
groups, where T1 is a maximal torus in SO and B1 is a Borel subgroup
of So which contains T1. Let B be a fixed Borel subgroup of So which
contains our fixed maximal torus T. Then any semisimple orbit of So
in S contains an element which normalizes T and B. The normalizer
of T and B in S can be written TWB, where WB is a fixed semisimple
element in S which preserves some splitting of B. Let S' and T' denote
the centralizers of WB in So and T. Since S° is simply connected, S'
is a connected reductive group which contains T' as a maximal torus.
In particular, we can form the usual sign character a' on the Weyl
group W' of (S',T').
LEMMA 8.4. The left hand side of (8.4) equals the number
A(W',E')
(8.5)
=
IW'1
EWreg e'()
PROOF: Let N' denote the normalizer of T' in S'. Then
W = N'/T' = TN'/T.
We claim that
Norm(T', S)
(8.6)
To see
= TN'.
this, we shall consider the (open) chambers in the real vector
spaces
aT,
aT = Hom(X(T),R)
determined by the roots of S' and S°. The fact that WB preserves
=
Hom(X(T'),R)
C
a splitting in (B, T) implies that the simple roots of (B n S', T') are
just the orbits under powers of ad(wB) of the simple roots of (B, T).
The corresponding positive chambers are therefore related by a+,
aT n aT. By an argument of symmetry, any chamber in aT, becomes
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
61
the intersection of aT, with a uniquely determined chamber in aT.
Suppose that n is an element in Norm(T', SO). Then n also normalizes
T, since T is the centralizer of T' in S°. The chamber Ad(n)(aT)
contains an open subset of aT,, and therefore a chamber
n' E N',
Ad(n')(a+,)
in aT,. The map Ad(n')lAd(n) will then send the chambers aT, and
a+ to themselves. This justifies the claim (8.6).
We have agreed that any semisimple orbit of SO in S intersects
TWB. Suppose that two elements sl and s in TWB are S°conjugate.
Since T' is a maximal torus of both Ss and Ss,, s1 and s are conjugate
by an element in the group (8.6). From this it follows that there is a
canonical bijection from Orb(TN', TWB) to the semisimple elements
in Orb(S°, S). It is of course only semisimple orbits which are relevant
to (8.4). We can therefore write the left hand side of (8.4) as
,
E i(S8)
sEOrb(SO ,S)
=E
sEOrb(TN',TwB)

i(S8)
IVW(s)l1
sEOrb(TN',TwB)
5
es(w)Jdet(w  1)11
wEW(Ss)reg
sign character on the Weyl group W(Ss) of
Ss.
If s belongs to TWB, Ss need not be a subgroup of S'. However,
the elements in W(Ss) normalize T', and are induced from the group
(8.6). Therefore
where
es
stands for the
W(Ss) = Cent(s,N')/T' TCent(s,N')/T.
In particular, W(S$) is a subgroup of W'. Thus, the simple reflections
in W(S,) are also reflections in W', and since es and e' both take the
value (1) on any such reflection, we see that E. equals the restriction
of e' to W(S,). We can substitute this into the expression above.
Our characterization of W(Ss) also suggests that we should change
the sum over Orb(TN', TWB) to a sum over the smaller set
TB = Orb(T,TwB) = {tlWBtwU1: t E T} \ TWB.
'
62
JAMES ARTHUR
The expression for the left hand side of (8.4) becomes
E'(w)ldet(w 1)1
JW ' 1
sETB wEW(S.),eg
The group W'
operates on TB.
W(Ss)reg
The last
=
It is easy to check that
{W E Wreg : W(S) = } .
expression can therefore be written
\W'\1
E'(w)ldet(w 1)1
E
{(s,w)ETB
X
W's :w(s)=s}
Now T' is a finite
covering of the torus
{tlWBtWBl: t E T}\T,
and Idet(w  1)1 equals the number of fixed points of w in either torus.
In particular, this number equals the order of the fixed point set TB
of w in TB. We can therefore write our expression as
EE '()
Iw'I1
"EW'6W, sE.T
Iw1
E (W)
wEW'reg
=
A(W',e)
[
.
LEMMA 8.5. The right hand side of (8.4)
A(W,e)
(8.7)
=
IW°11
equals the number
l
e(w)
WE Wreg
PROOF: Since any semisimple conjugacy class in So meets T, we have
a bijection from Orb(W°,T) to the set of semisimple elements in
Orb(So, SO). The right hand side of (8.4) can then be written
E
tEOrb(S ,S°)
=
i(St)
E Et(w)ldet(w
E IW(St)l1 wEW(St),,g
tEOrb(W°,T)

1)1
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
63
where et stands for the sign character on the Weyl set W(St).
We need only consider elements t E T such that St is not empty. For
any such t, T is a maximal torus in the connected group (St)° = St,
and W(St) is a subset of W = W(S). We claim that et is the restriction of e to W(St). The group W(St) is generated by reflections which
lie in W(S°). Since this group acts simply transitively on W(St), it
suffices to check that e and et coincide on one element in W(St). Let
s be a semisimple element in St. Then there is a conjugate
S1
=
gE S°,
gsg1,
of s which lies in TWB. We can in fact choose g so that tz = gtg1
lies in the maximal torus T' of S. It then follows that tl equals
wi(t) for some wl E W°, and that t is fixed by the element wl' WBWl
in W(S). In other words, wl1 WBwl belongs to W(St). Since this
element normalizes the Borel subgroups wl'Bwl and wl'Bw1 n St
of So and St, we have
E(WliWBWl)
=
Et(wlWBwl)
This establishes the claim.
Since W(St) is the centralizer of t in
(8.4) becomes
IWllE
tET
The set W = W(S)
E
W°,
=
the
1.
right hand side of
6(w)ldet(w 1)11.

wEW(St),eg
operates on T, and
W(St)reg
=
{W E Wreg w(t) = t
.
The last expression can then be written
IW°11 {(t,w)ETx
=
=
Ee(w)ldet(w  1)1
Weg:w(t)=t}
IWI1 E E (w)lTWl1
wEW,,g tET"
IW 11 E (w)
sEWreg
A(W, ) ,
since Idet(w 1)1 equals the order of the fixed point set Tw of w in
=

T. The lemma is proved.
O
64
JAMES ARTHUR
A(W',') = A(W,E).
PROOF: The numbers A(W',e') and A(W, £) depend only on the
Weyl set W. They are independent of the isogeny class of the underlying component S. We shall assume inductively that the required
formula holds if S is replaced by a component of strictly smaller dimension.
We have the fixed Borel subgroups B and B' of So and S', so
we can speak of standard parabolic subgroups. Suppose that A is a
standard torus in T'. In other words, A is the split component of a
parabolic subgroup P' of S' which contains B'. Let M' be the Levi
LEMMA 8.6.
component of P' which contains T'. Then A equals AM' =
the connected component of the center of M'. Write
Z(M')°,
WA = W(M'/A)
for the Weyl group of M'/A, acting on T'/A. We can also take the
centralizer M of A in S. Then M° is the Levi component of a standard
parabolic subgroup of S°. Write
WA = W(M/A),
Weyl set of the component M, acting on T/A. The element
WB obviously embeds into WA, and WA is just the centralizer of WB
in WA = W(M°/A). If A is nontrivial, our induction hypothesis tells
for the
us that
(8.8)
A(WAEA)
=
A(WAEA),
where e' and EA are the sign characters on WA and WA.
Suppose that w is an arbitrary element in W'. The identity component of the fixed point set (T')W is a torus in T', and equals a
W'translate
w1i(A),
wl E W',
of a standard torus A in T'. The element
wlwwl1 then lies in WAreg.
It is also clear that
EJ(W)
=
EA(1WWl 1)
.
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
Now the pair (A, wl) is not uniquely determined
of such pairs actually equals
65
by w. The number
n(A)WAl ,
where n(A) is the number of chambers in Hom(X(A), R) cut out by
the hyperplanes orthogonal to the roots of the corresponding standard
parabolic subgroup. The elements in W' can be enumerated up to this
ambiguity, however, as conjugates
W1 WAW ,
WA E
WA,reg
W1 E W
We obtain
IWI1 E E(w)
wuEW
=
A
n(A)1lWA1 E eA(WA)
WAEWA
= i n(A)1A(WA,
A).
A
If W' is not equal to {1}, the sign character E' is nontrivial, and the
left hand side of the equation equals 0. Applying (8.8) to the right
hand side, we conclude that the expression
(8.9)
A(W',e) + E n(A) l(WAeA)
A.{1}
vanishes if W' : {1}.
Now suppose that w is an arbitrary element in W. The identity
component (Tw)° of the set of fixed points of w in T is a torus which
commutes with any representative in S of the Weyl element w. Copying an argument from the proof of Lemma 8.5, we see that
(TW)O
=
wl (A),
where wl belongs to W° and A is a torus in T'. In fact, we can
assume that the centralizer M° of A in S° is the Levi component of
a standard parabolic subgroup of S°. This implies that A is standard
torus in T'. The element wlwwT1 then lies in WA,reg, and
e(W)
=
A(WlWW 1) .
66
JAMES ARTHUR
given w, how many such pairs (A, wl) are there? We can certainly replace w1 by a product
For a
W'WMW1 ,
w E W', WM E W(M°),
in which w' maps A to another standard torus in T'. However, this
is the only possible ambiguity, so the number of pairs equals
n(A)IW(M°)I
=
n(A)IWAI.
We obtain
E e(w)
E
Zn(A)1'WA1 WAEWA,reg
IW01
wEW
=
A(WA)
A
=
1A(WA, A).
Zn(A)
A
If W contains more than just the one element wB, the left hand side
of the equation equals 0. Therefore, the expression
(8.10)
A(We)+ E n(A)A(WAeA)
A${1}
vanishes if W Z {WB}.
The simple reflections in W' correspond to the orbits of simple roots
of (B,T) under powers of ad(wB). It follows that W' = {1} if and
only if W = {wB}. In this case, both (8.9) and (8.10) are trivially
equal to 1. We can therefore conclude that the expressions (8.9) and
(8.10) are equal. The equality of A(W', e') and A(W, e) then follows.
We have reached the end of the lemmas that make up the proof
of Theorem 8.1. We obtain the general inequality of i(S) with e(S)
immediately by combining Lemmas 8.3, 8.4, 8.5 and 8.6. The proof
of Theorem 8.1 is now complete.
O
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
§9. CONCLUDING
67
REMARKS
Theorem 8.1 tells us that the coefficients i(x) and e(x) in (5.10) and
(7.15) are always equal. It follows that there is a term by term identification of the expansions for Idisc,t(f) and Edisc,t(f). We conclude
that Hypothesis 3.1 is a consequence of Hypothesis 4.1 (together with
the local assumptions of §3, §4 and [3, §7]). This was the task we
originally set for ourselves.
We have in fact shown that the contributions to Idisc,t(f) and
Edisc,i(f) of each parameter tp E I(G, X, t) are equal. Now there are
some parameters for which the representation theoretic hypotheses
are known. Consider the special case that G is a connected quasisplit group. Suppose that b is the image of a parameter ob0 E6(Mo)
for a minimal Levi subgroup Mo of G. Since Mo is a maximal torus
in this case, o0 is trivial on SL(2, C), and is the parameter of a unitary character on Mo(F)\Mo(A). We can take IIH to be the set of
irreducible constituents of the corresponding induced representation
of G(A). The parameter ho factors through the quotient WF of LF,
so there is no problem with the hypothetical Langlands group. In
particular, S, equals the centralizer in G of the image of WF, and
the quotient Sp is just the Rgroup R,. The pairing on S$x IIH is
then determined by the global normalized intertwining operators. In
fact, Conjecture 7.1 of [3], which we assumed in §5, is already known
in this case thanks to Keys and Shahidi [10, Theorem 5.1]. If H1
is associated to a point in a component x E SO, we could just define
the distribution f + fH1 (1) by
fH1(li)
< x,r >
=
fG(r)
.
wrEH,
Then with these interpretations, the notions that went into the discussion in §5§8 are all understood. The reader who dislikes arguments
based on unproven conjectures can regard the earlier discussion as
pertaining only to the parameters just described. It establishes that
the contribution of these parameters to
(9.1)
Edisc,t(f)

Idisc,t(f)
vanishes.
This paper has concerned the conjectures in [3] on unipotent (and
more general) automorphic representations. The long term goal is to
68
JAMES ARTHUR
prove them, at least in part, with the help of endoscopy and the trace
formula. A first step towards the creation of a logical structure for
the argument is to verify the compatibility of the notions involved
and to analyze the reasons for it. This has been our emphasis, and
we continue with some informal comments on the proof envisaged.
In general, Hypothesis 3.1 asserts the vanishing of the distributions
(9.1). As we mentioned earlier, one should first try to deduce this from
the trace formula. One would then use (9.1) to establish some version of the multiplicity formula (4.7). The formula could be assumed
inductively for any proper Levisubgroup. This would permit the application of the arguments in §5§8 to any parameter /I E 1(G, X, t)
which is not the image of an elliptic parameter for an elliptic endoscopic datum. The contribution to (9.1) of all such parameters could
then be shown to vanish. The only remaining contribution to (9.1)
would come from parameters k such that Sas, is finite for some element s in SV, = Sp/Z(G°). It is from this that we would hope to
deduce some form of (4.7), again using arguments of Sections 5, 6
and 7. The sign characters EG would be forced on us at this stage,
essentially because of Proposition 5.1.
Of course, it would not be feasible to apply the arguments of §5§8
in precisely the way they were presented here. The correspondence
from maps WF LG to automorphic representations is much deeper
than multiplicity formulas such as (4.7), and in any case, we would
certainly not want to assume the existence of the Langlands group LF.
We would instead have to replace the parameters 4 by the families
a = {a,: v ¢ S} of conjugacy classes in LG attached to automorphic
representations. (See [3, §1, §8].) For many G we can expect a bijection from I(G) onto the set E(G) of such families. In these cases,
the idea would be to define the centralizer S, in terms of a. This
could probably be done by considering the set of endoscopic groups
H for which a lies in the image of the map E(H) ES(G). It is of
course necessary to determine S, in order to state the multiplicity formula (4.7). By definition, a parameter 4 has a Jordan decomposition
(Oss, Ounip),
where
o'.: LF 
LG
and
: SL(2, C) , So
We would describe the Jordan decomposition in terms of a by first
determining the family aos attached to ;5,, and then describing the
nip
69
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
group So, in terms of a,,. In the case of a general G, some understanding of the fibres of the map [(G) E(G) will probably be

needed.
We have not said much about the local side of the conjectures. This
includes the definition of the stable distributions fi  fG (41), the
construction of the packets HII and the pairing < *, >, and the proof
of the local character identity (4.4). Once the stable distributions
have been defined, the packets and the pairing are determined by
(4.4) (together with the maps f  f HI). The essential part of the
local conjecture is then the assertion that for a given 4, certain linear
combinations of the distributions
f Cc (G(A),X),
f , fHl('1),
are actually characters, as opposed to more general invariant distributions. Ideally, it would be best to deduce this locally. However,
the global Hypothesis 3.1 itself carries some local information. For
it ultimately implies some version of (4.7), and any such multiplicity
formula tells us that certain distributions are in fact characters. I do
not know how far this can be pushed. It is perhaps best to wait until
Hypothesis 3.1 has actually been established.
The case in which Hypothesis 3.1 will lead to the most complete
results is the example of outer twisting of GL(n). The hypotheses
of §4 (interpreted without reliance on the parameters 4 E I(G)) are
now known for GL(n). Moeglin and Waldspurger [21] have recently
characterized the residual discrete spectrum for GL(n) in terms of the
cuspidal spectrum, and it is clear how to interpret this in terms of the
Jordan decomposition [3, §2]. On the other hand, the twisted endoscopic groups for GL(n) include all of the quasisplit classical groups
of type B, C and D (up to isogeny). One should try to deduce the conjectural properties of the spectra of these classical groups from what
is known for GL(n). We will conclude with a very brief discussion of
this example.
Set Go = GL(n). If
Jn=."i on
0 I
then
On(g)
=
J1(t1)Jn ,
gE
G°,
70
JAMES ARTHUR
is an outer automorphism of Go which leaves invariant the standard
Borel subgroup. Set
G = G° 0,.
If On denotes the same outer automorphism of G° = GL(n, C), then
G
=
G° X On
easy to describe the elliptic endoscopic data for G.
integer r, with 1 < r < n, set
It is
For each
1
1
Sr =
4X
0
On
1
regarded as a semisimple coset in G/Z(G°). Then
Cent(sr, G°) ) Sp(2r, C) x O(n  2r, C)
Define
Hr = Cent(sr,
0)°0
Sp(2r, C) x SO(n  2r, C).
Let Gr be any Lhomomorphism
Cent(sr,G )xWF C G xWF.
This determines an endoscopic datum (Hr, ',Sr,r) whose equivaWF

lence class depends only on the map
r:
Gal(F/F) , 7o(Cent(sr,,°))
Thus,
Hr _
~
SO(2r+ 1) xx SO*(n2r),
1
SO(2r + 1)
Sp(n
2r),
n
even,
n
odd,
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
71
where SO* stands for the quasisplit orthogonal group determined by
When n  2r = 2 and is trivial, Z(Hr)r is infinite, and the
endoscopic datum is not elliptic. If we rule out this exceptional case,
however, we obtain a set of representatives of elliptic endoscopic data.
Observe that 7r can be identified with LHr in each case, so there is
no need to introduce the extensions that were denoted by H1 in §2.
Suppose that
: LF SL(2,C)  LGO
*.
*
is a parameter in [(G°). We shall identify 0b with an ndimensional
representation of the group LF x SL(2, C), which can then be decomposed into a direct sum
e
+ =
©^k
k=1
of irreducible representations. The centralizer in G° of the
b is the group of intertwining operators. That is,
So,(G°)
=
So
image of
IGL(mj, C).
The parameter b belongs to [(G) if and only if it is selfcontragredient
as a representation of LF x SL(2, C). In other words, the contragredient operation acts as a permutation of order two on the irreducible
constituents Ok. Suppose that this is the case. Then S, = Sp(G) is
isomorphic to a product of components of the form
GL(m, C) )X
m
or
(GL(m, C) x GL(m, C)) x Tm ,
with
~m(1g,g2)
(9m(92), Om(gl)) ,
=
9,g2 E GL(m,C).
We are especially interested in the parameters 4 E J(G) which are
the images of elliptic parameters Or E Io(Hr), for elliptic endoscopic
data Hr. Since AG Z/2Z is finite, this means that there is an
element s E So such that SO,, is finite. The condition is equivalent to

(9.2)
S1o

(C* 01)e,
72
JAMES ARTHUR
say that the irreducible constituents /Ok of i are selfcontragredient and mutually inequivalent. Since
E C*,
01(z) Z,
which is to
=
immediately from (9.2) that there is only an orbit of So in Sp.
Therefore, b factors through only the one endoscopic datum Hr.
Fix an elliptic endoscopic datum Hr, and let tr E To0(Hr) be a fixed
elliptic parameter. The image ? of Or in @(G) then satisfies (9.2).
we see
For reasons of induction it is not necessary to consider a product of
two classical groups, so we may assume that r equals 0 or 2. Then
Hr is either an orthogonal group or a symplectic group. To study the
representations of Hr(A) attached to O r, it will be necessary to apply
Hypothesis 3.1 to both G and Hr.
A missing ingredient from the local conjectures was a canonical
definition of the stable distribution
f.  fHr(r) f,
(9.3)
r E Co(Hr(A)).
Such a definition will be provided, at least in some cases, by the
connection with G. The packet HI consists of one orbit {Trt,} C
under the group ro(G+)* Z/2Z, and we can choose 7r,
so that
II(G(A)+)
=
,r >
<
It follows from
= <
s E So.
>= 1 ,
,
(4.4) that
fHr(Or)
=
fG(TI,),
f
C
(G(A)).
A similar formula holds for the corresponding stable distributions on
the local groups G(Fv). However, this formula may not determine
(9.3) completely. The problem is that the anticipated injection
(f.Hr
fEC
(G(A))}I
{fHr:
fr E C
(H(A)) },
obtained by transfer of twisted orbital integrals, could be a strict
inclusion. This difficulty is tied up with the question of how many
local parameters
/4
=
(0
V
,,
E (HIFv),
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
73
lift to 4i. If Hr is symplectic or odd orthogonal, the only such parameter will be 0Ir itself. However, there can be a number of O' in the
even orthogonal case, and the formula then determines only a sum of
distributions (9.3).
Once the distribution (9.3) has been defined (for Hr and its endoscopic groups), the packet II^ and the pairing on Sr x II7^ will
be uniquely determined. Leaving aside the question of whether the
required local properties of these objects can be deduced from Hypothesis 3.1, let us simply assume that the local assumptions of §3, §4
and [3, §7] hold for Hr. The next problem is to determine the stable
distribution
SIH (fr) = a(Hr, r)frr (Vbr)
(9.4)
(See the notation of §7.) The distribution fHr (r) is a local object
which we are assuming is known, so it is the global constant a(Hr, 4Or)
which must be found. According to Hypothesis 3.1, we should take
the contribution of 4' to (9.1), and set it equal to 0. I have not
thought through the details, but it should just be a question of running
backwards over a couple of the more trivial arguments of §5 and §7.
The result will be a special case
OT(Hr,7r) = ISrlE Hi(Sgr)
of the general formula (7.12) we determined was compatible with Hypothesis 4.1. Observe that the sign character e,, (Sgr) appears. It
originates, through [3, Conjecture 7.1] and Proposition 5.1, from
the normalizing factors for (nontempered) intertwining operators for
GL(n).
Having determined the stable distributions (9.4) (for Hr and its
endoscopic groups), we can apply Hypothesis 3.1 to Hr. The contribution of Or to Edisc,t(fr) can be calculated as an easy special case
of the arguments in §7, or it can simply be read off from the formula
(7.15), (applied to Hr instead of G). It equals
E (is 11
eHr(x)< x, >)tr(7(fr))
(9.5)
7rE nI
r
XES
Notice that the only contribution to mo(7r) should come from the
parameter )r,. This suggests that the map Ti(H,) , E(H,) is bijective, at least if Hr is not an even orthogonal group, the case we left
ambiguous.
This discussion has been very sketchy. We have simply tried to
indicate that since the spectrum of GL(n) can be understood in terms
of a Jordan decomposition, the same should be true for the spectrum
of its endoscopic groups. The arguments of §5§8 will be essential for
this, in that they allow for the elimination of the irrelevant parameters
from the study of (9.1).
REFERENCES
[1]
J. Arthur, On some problems suggested by the trace formula, in Lie Group
Representations II, Lecture Notes in Math., vol. 1041, SpringerVerlag, 1983,
pp. 149.
[2] J. Arthur, The invariant trace formula II. Global theory, J. Amer. Math. Soc.
1 (1988), 501554.
[3] J. Arthur, Unipotent automorphic representations: Conjectures, to appear in
Ast6risque.
[4] J. Arthur and L. Clozel, Simple Algebras, Base Change and the Advanced
Princeton
Theory of the ace Formula, to appear in Annals of Math. Studies,
University Press.
[5] L. Clozel, J.P. Labesse and R.P. Langlands, Morning seminar on the trace
formula, Lecture Notes, Institute for Advanced Study, Princeton, N.J., 1984.
L[6] A. Frohlich and J. Queyrot, On the functional equations20of the Artin
function for characters of real representations, Invent. Math. (1973), 125138.
[7] R. Godement and H. Jacquet, Zeta Functions of Simple Algebras, Lecture
Notes in Math., vol. 260, SpringerVerlag, 1972.
[8] H. Jacquet and J. Shalika, A nonvanishing theorem for zeta functions of
GLn, Invent. Math. 38 (1976), 116.
[9] H. Jacquet and J. Shalika, On Euler products and the classification of automorphic representations I, Amer. J. Math. 103 (1981), 449558.
[10] D. Keys and F. Shahidi, Artin Lfunctions and normalization of intertwining
operators, Ann. Scient. Ec. Norm. Sup., 4e srie, t.21 (1988), 6789.
[11] R. Kottwitz, Rational contugacy classes in reductive groups, Duke Math. J.
49 (1982), 785806.
UNIPOTENT AUTOMORPHIC REPRESENTATIONS
75
[12] R. Kottwitz, Stable trace formula: cuspidal tempered terms, Duke Math. J.
51 (1984), 611650.
[13] R. Kottwitz, Stable trace formula: elliptic singular terms, Math. Ann. 275
(1986), 365399.
[14] R. Kottwitz, Tamagawa numbers, Ann. of Math. 127 (1988), 629646.
[16] J.P. Labesse and R. Langlands, Linditinguishabiity for SL(2), Canad. J.
Math. 31 (1979), 726785.
[16] R. Langlands, On the Functional Euations Satisfied by Eisenstein Series,
Lecture Notes in Math., vol. 544, SpringerVerlag, 1976.
[17] R. Langlands, Stable corjugacy: definitions and lemmas, Canad. J. Math.
31 (1979), 700725.
[18] R. Langlands, Les Debuts d'une Formule desu7aces Stable, Publ. Math.
Univ. Paris VII, Vol. 13, Paris (1983).
[19] R. Langlands, Eisenstein series, the trace formula, and the modem theory of
automorphic forms, to appear in Number Theory, M7ace Formulas and Discrete
Groups: Symposium in Honour of Atle Seberg, Academic Press.
[20] R. Langlands and D. Shelstad, On the definition of transfer factors, Math.
Ann. 278 (1987), 219271.
[21] C. Moeglin and J.L. Waldspurger, Le spectre residuel de GL(n), preprint.
[22] J. Rogawski, Automorphic Representations of Unitary Groups in Three Variables, to appear in Annals of Math. Studies, Princeton Univ. Press.
[23] T. Springer and R. Steinberg, Corjugacy classes, in Seminar on Algebraic
Groups and Related Finite Groups, Lecture Notes in Math., vol. 131, 1970, pp.
167266.
[24] R. Steinberg, Endomorphisms of Linear Algebraic Groups, Mem. Amer.
Math. Soc. 80, 1968.
Mathematics
Canada.
Department, University of Toronto, Toronto M55 1Al, Ontario,