LECTURES ON THE ARTHUR–SELBERG TRACE FORMULA

arXiv:math/9505206v1 [math.RT] 10 May 1995

STEVE GELBART

MSRI, Spring 1995

These are Notes prepared for nine lectures given at the Mathematical Sciences
Research Institute, MSRI, Berkeley during the period January–March 1995. It is a
pleasant duty to record here my gratitude to MSRI, and its staff, for making possible
this 1994–95 Special Year in Automorphic Forms, and for providing such a setting
for work.
The purpose of these Notes is to describe the contents of Arthur’s earlier, foun-
dational papers on the trace formula. In keeping with the introductory nature of
the lectures, we have sometimes illustrated the ideas of Arthur’s general theory by
applying them in detail to the case of GL(2); we have also included a few lectures
on the “simple trace formula” (and its applications), and on Jacquet’s relative trace
formula.
The TEX preparation of these Notes I owe to Wendy McKay, who patiently and
professionally transformed my messy scrawl into something readable; her expertise,
and good cheer under the pressure of weekly deadlines, is something I shall not soon
forget.
I wish to thank the auditors of these lectures for their interest, and , J. Bernstein,
D. Goldberg, E. Lapid, C. Rader, S. Rallis, A. Reznikov, and D. Soudry, for their
helpful suggestions and explanations. Finally, I wish to thank J. Arthur, H. Jacquet,
and J. Rogawski for many tutorials on these and related topics over the past year; I
hope they do not mind seeing some of their comments reappear in these Notes.

Table of Contents

Lecture I. Introduction to the Trace Formula

Lecture II. Arthur’s Modified Kernels I: The Geometric Terms
Lecture III. Arthur’s Modified Kernels II: The Spectral Terms
Lecture IV. More Explicit Forms of the Trace Formula
Lecture V. Simple Forms of the Trace Formula
Lecture VI. Applications of the Trace Formula
Lecture VII. (G, M)-Families and the Spectral Jχ (f )
Lecture VIII. Jacquet’s Relative Trace Formula
Lecture IX. Some Applications of Paley–Wiener, and Concluding Remarks

1
2 STEVE GELBART

Lecture I. Introduction to the Trace Formula

The ultimate purpose of the trace formula is to better understand the structure
of automorphic forms on a general reductive group G. In particular, one wants
to understand the decomposition of the right regular representation R0 (g) of G(A)
in the space of cusp forms L20 (G(F ) \ G(A)). The original idea of the (Selberg)
trace formula—roughly
R
speaking—was to describe the trace of the integral operator
R0 (f ) = G(A) f (g)R0 (g)dg for f in Cc∞ (G(A)); this Selberg succeeded to do for certain
special G. On the other hand, Arthur’s trace formula is valid for arbitrary G, and
asserts the equality of two different sums of distributions on G, namely:
X X
(0.1) JoT (f ) = JχT (f ).
o∈O χ∈X

Here the left side of (0.1) is summed over certain equivalence classes in the group
of rational points G(F ); it is called the geometric side of the trace formula. On the
right side of (0.1), the sum is over certain “cuspidal data” χ = {(M, σ)} of G; this is
the spectral side of the trace formula (one of whose terms is trR0 (f )).
Our purpose in these lectures is to describe and develop this formula in complete
detail, but mostly in the context of GL(2). We also wish to describe some refinements
of (0.1) necessary for applications to automorphic forms. The basic references are
Arthur’s original papers [A1]–[A10]. Earlier expositions of the GL(2) trace formula,
such as [Ge] and [GJ], will be referred to occasionally; but as they mostly predate
Arthur’s general development of the trace formula, their point of view will largely be
ignored, and their arguments not reproduced here.

1. Some history. “The” trace formula was introduced by Selberg, in the context
of a semi-simple Lie group G and discrete subgroup Γ, in his famous 1956 paper [Se].
In that paper Selberg first of all described a general formula in the case of compact
quotient Γ \ G (see the next section below for a brief description). Secondly, he
treated in detail certain special non-compact quotient cases such as SL2 (Z) \ SL2 (R);
here Selberg analytically continued Eisenstein Series E(z, s) in order to handle the
continuous spectrum of L2 (Γ \ G), and derived a formula for the trace of R(f ) in
the space of cusp forms. In both the case of compact and non-compact quotient,
Selberg was motivated by applications to geometry and number theory (lengths of
geodesics in the Riemann surface in the case of compact quotient, traces of Hecke
operators in case of non-compact quotient, etc.). But despite the obvious desirability
of generalizing these results in the case of non-compact quotient, the obstacles—such
as analytically continuing the Eisenstein series in higher rank—were formidable. A
big breakthrough came in the 1960’s when Langlands developed a theory of Eisen-
stein series valid for any reductive group G, and used it to describe the continuous
spectrum of L2 (G(F ) \ G(A)) (cf. [La 1]). As an outgrowth of this theory, Langlands
 LECTURE I. INTRODUCTION TO THE TRACE FORMULA 3

was led to his amazing conjectures about automorphic forms and their L-functions,
i.e., the “Langlands program”, with its cornerstone the principle of functoriality of
automorphic forms. A typical result in this program is the following:
Theorem 1 ([JL]). Let D be a division quaternion algebra over a number field F
and G′ = D ∗ . Then to each automorphic cuspidal representation σ of G′ (i.e., an
irreducible G′ (A)-submodule of L2 (Z(A)G′ (F ) \ G′ (A)) of dimension greater than one)
there exists a corresponding automorphic cuspidal representation π = π(σ) of G(A) =
GL2 (A) (an irreducible G(A)-submodule of the space of cusp forms L20 (Z(A)G(F ) \
G(A)) with the property that for each place v in F unramified for D (i.e. where
G′ (Fv ) ≃ GL2 (Fv )), σv ≃ πv .
Interestingly enough, this correspondence (the well-known Jacquet–Langlands cor-
respondence) can be proved using any one of the three basic tools of the theory of
automorphic forms: the theory of automorphic L-functions, the theory of theta-series
liftings, or the trace formula (for both G′ and G = GL2 , essentially as developed by
Selberg). However, in this case at least, the trace formula approach gives the strongest
payoff, namely a characterization of the image of the correspondence σ → π(σ): a
cuspidal representation π of G(A) is of the form π(σ) for some σ on a given G′D
if and only if for each ramified prime v, πv is a square integrable (discrete series)
representation of Gv = GL2 (Fv ).
Subsequently, a (“twisted”) trace formula for GL2 was developed by Saito-Shintani
and Langlands to prove another important instance of functoriality, namely base
change (with spectacular applications to Artin’s conjecture for Galois representations,
and hence ultimately to Wiles’ proof of Fermat . . . ). Thus the need to generalize
the trace formula to a general reductive group G was clear, and this is precisely what
Arthur did in his papers between the early 1970’s and late 1980’s.

2. The Case of Compact Quotient. In order to motivate Arthur’s work, let us

explain how the form (2) takes in the case of compact quotient. Thus G is a semi-
simple group over (the number field) F , with G(F )\G(A) compact. In particular, G is
anisotropic, without unipotent elements. With R(g) the right regular representation
in L2 (G(F ) \ G(A)) and f in Cc∞ (G(A)), a straightforward computation gives
Z
(R(f )φ)(x) = f (y)φ(xy) dy =
G(A)
Z Z X
f (x−1 y)φ(y) dy = ( f (x−1 γy))φ(y) dy,
G(A) G(F )\G(A)
γ∈G(F )

i.e., R(f ) is an integral operator on φ in L2 (G(F ) \ G(A)), with kernel

X
Kf (x, y) = f (x−1 γy).
γ∈G(F )

Since G(F ) \ G(A) is compact, it is then a well known theorem (cf. [GGPS], Chapter
I), that R(f ) is of trace class, and that
Z
trR(f ) = Kf (x, x) dx,
G(F )\G(A)
4 STEVE GELBART

the integral of the kernel over the “diagonal subgroup” G(F ) \ G(A)∆ . In this case,
the (Selberg) trace formula amounts to a calculation of traceR(f ) using two different
expressions for this kernel. On the one hand, we can write
X
Kf (x, y) = Ko (x, y)
o∈O

where O denotes the set of conjugacy classes in G(F ), and

X X
Ko (x, y) = f (x−1 γy) = f (x−1 δ −1 γ0 δx)
γ∈o={γ0 } δ∈Gγ0 (F )\G(F )

where Gγ0 (F ) is the centralizer of γ0 in G(F ). Then

Z Z X
Jo (f ) = Kf (x, x) dx = f (x−1 δ −1 γ0 δx) dx =
G(F )\G(A) G(F )\G(A) δ∈G (F )\G(F )
γ0

Z Z
−1
f (x γ0 x) dx = meas(Gγ0 (F ) \ Gγ0 (A)) f ((x−1 γ0 x) dx,
Gγ0 (F )\G(A) Gγ0 (A)\G(A)

and thus
X X Z
trR(f ) = Jo (f ) = meas(Gγ (F ) \ Gγ (A)) f ((x−1 γx) dx.
o∈O Gγ (A)\G(A)
{γ}

On the other hand, the compactness of G(F ) \ G(A) also implies a discrete decom-
P
position R(g) = π mπ π(g) where π runs through the irreducible “automorphic”
representations of G(A), and mπ < ∞ is the multiplicity of π in R. Equivalently,
P
L2 (G(F ) \ G(A)) P
= Lπ where Lπ is the invariant subspace realizing the mπ copies
of π, and R(f ) = Rπ (f ) where each Rπ (f ) is an integral operator in Lπ with kernel
X
Kπ (x, y) = (R(f )φ)(x)φ(y).
φ an ON basis forLπ

Thus we also have

X
trR(f ) = Jχ (f ),
χ∈{(G,π)}

where Z
Jχ (f ) = Kπ (x, x) dx.
G(F )\G(A)

The power of the trace formula derives from the fact that this equality
X X
(2.1) Jo (f ) = Jχ (f ),
namely
X Z XZ X
−1
mγ f (x γx) dx = ( (R(f )φ)(x)φ(x)) dx
Gγ (A)\G(A) π G(F )\G(A) φ
{γ}

relates two completely different kinds of data, geometric and spectral (the latter of
which we rarely know much about, but very much want to!).
 LECTURE I. INTRODUCTION TO THE TRACE FORMULA 5

3. Difficulties in the Non-Compact Case. Arthur’s great contribution was to

figure out how to overcome all the obstacles that present themselves in generalizing
(2.1) from the compact to the non-compact quotient setting. Some of the principal
problems in the non-compact setting are these:
(a) Although R(f ) is still an integral operator, with kernel
X
Kf (x, y) = f (x−1 γy),
γ

this kernel is no longer integrable over the diagonal; and

(b) The regular representation R(g) no longer decomposes discretely, Eisenstein
series are required to describe the continuous spectrum, and R(f ) is of course no
longer of trace class. Nevertheless, Arthur derives a formula (0.1) which still relates
“geometric” and “spectral” distributions attached to G—and has powerful applica-
tions to number theory (even if neither side of (0.1) represents a formula for trR(f )!).
For example, in this formula,
Z
T
Jo (f ) = KoT (x, x) dx,
Z(A)G(F )\G(A)

where KoT is a modification of the kernel

X
Ko (x, x) = f (x−1 γx)
γ∈o

which can be integrated over the diagonal. To see why this modification is really
needed, and what it accomplishes, let us look at the example of G = GL(2), and o
the hyperbolic conjugacy class
 
α0
o = {δ −1 01
δ : δ ∈ M(F ) \ G(F )},
!
a 0
where M = { }. In this case,
0 b
Z
Jo (f ) = Ko (x, x) dx
Z(A)G(F )\G(A)

(where Z is the center of scalar matrices). A formal computation then shows

Z X  
α0
Jo (f ) = f (x−1 δ −1 01
δx) dx
Z(A)G(F )\G(A) δ∈M (F )\G(F )
Z  
α0
= f (x−1 01
x) dx
ZZ(A)G(F )\G(ZA)  
α0
= ( dy)f (x−1 01
x) dx,
M (A)\G(A) Z(A)M (F )\M (A)

i.e.,  
α0
Jo (f ) = m(Z(A)M(F ) \ M(A))Ff ( 01
),
  R  
with Ff ( α0 01 ) the nice “orbital integral” M (A)\G(A) f (x−1 α0 01 x) dx of a compactly
supported function f in Cc∞ (Z(A)\G(A)). Thus Jo (f ) is infinite for the simple reason
that
m(Z(A)M(F ) \ M(A)) = m(F ∗ \ A∗ ) = ∞.
6 STEVE GELBART

To compensate for this, Arthur “modifies” Ko to KoT , using a truncation parameter

T = (T1 , T2 ) in the Lie algebra of M. As we shall see, this modification operator (in
this case) “truncates” F ∗ \ A∗ to a finite interval of length T2 − T1 , but turns Ff into
a “weighted” orbital integral! In general, for an arbitrary G, it will turn out that
Ko will be integrable “over the diagonal” only when o intersects no proper parabolic
subgroup P of G. Of course for anisotropic G this always happens, but in general
one really needs to modify Ko (as in the above example), prove that the modified
kernels KoT are indeed integrable, and then evaluate them (as generalized weighted
orbital integrals).
What about the spectral side? Here it turns out that L2 (Z(A)G(F ) \ G(A)) =
P 2 2
χ∈X Lχ , where each Lχ is a G(A)-invariant (not necessary discretely decomposable)
submodule indexed by certain cuspidal data X = {(M, σ)}, where M is a Levi sub-
group of a parabolic P ⊂ G, and σ is a cuspidal automorphic representation of M(A).
P
Thus we have K(x, y) = Kχ (x, y), the spectral decomposition of Kf (x, y) corre-
sponding to (0.1). Here it also turns out that Kχ (x, y) will not be integrable over the
diagonal, unless the cuspidal data (M, σ) = (G, π) (i.e. P = G). Thus we again need
to deal with modified KχT (which are integrable . . . ), eventually arriving at the trace
formula (0.1) by integrating the equality
X X
(3.1) KoT (x, y) = KχT (x, y)
o χ

over the diagonal.

Remarks. (1) The expressions on either side of (3.1) no longer (generally) represent
P
the kernel Kf (x, y) = f (x−1 γy), and hence (0.1) is no longer a real trace formula at
all. Nevertheless, what is true is the following. Let X(G) be the subset of cuspidal data
P
χ = (M, σ) ∈ X with M = G. Then χ∈X(G) Kχ (x, y) is the kernel of R(f ) restricted
to L20 (Z(A)G(F ) \ G(A)), it is integrable over the diagonal, and (0.1) asserts that
X X X
(3.2) trR0 (f ) = Jχ (f ) = JoT (f ) − JχT (f ),
χ∈X(G) o X\X(G)

i.e., Arthur’s trace formula may be viewed as a trace formula after all.

(2) In most applications of the trace formula to automorphic forms we want to com-
pare the formula for one group with the trace formula for another. In that case the
form of (0.1) is no less useful than (3.2).

4. Possible Applications. We have already mentioned the applications to quater-

nion algebras and base change, both of which we plan to discuss (at the level of GL(2),
not GL(n) - which is the subject matter of [AC], using the more difficult GL(n) trace
formulas). We also plan to discuss (i) Jacquet’s relative trace formula, where the idea
is to integrate the (different expressions for the) kernel over interesting non-diagonal
subgroups of G × G; and (ii) applications of the simple trace formula such as to
the embedding of a (local) discrete series representation of GL2 (Fv ) into a (global)
cuspidal representation of GL2 (A). In all these applications, one needs to further re-
fine Arthur’s first form of the trace formula. This is because (as already hinted) the
“truncation” i.e., modification process used to make Ko or Kχ integrable introduces
 LECTURE I. INTRODUCTION TO THE TRACE FORMULA 7

various defects into the distributions JoT (f ) and JχT (f ). For example, they may no
longer be invariant distributions (invariant under conjugation by G(A)), or factoriz-
able as the product of local distributions over G(Fv ). Thus one needs to replace and
rearrange the sums of distributions appearing in (0.1) by still more complicated but
ultimately more practical expressions.
References
[A1] Arthur, J., “The Selberg trace formula for groups of F -rank one”, Annals of Math. 100
(1974), pp. 236–385.
[A2] Arthur, J., “A Trace formula for reductive groups I: Terms associated to classes in G(Q)”,
Duke Math. J., Vol. 45, No. 4 (1978), pp. 911–952.
[A3] Arthur, J., “A Trace formula for reductive groups II: Applications of a truncation opera-
tor”, Compos. Math. 40 (1980), pp. 87–121.
[A4] Arthur, J., “A trace formula in invariant form”, Annals of Math. 113 (1981), pp. 1–74.
[A5] Arthur, J., “On a family of distributions obtained from Eisenstein series I: Application of
the Paley–Wiener theorem”, Amer. J. of Math., Vol. 104, No. 6 (1982), pp. 1243–1288.
[A6] Arthur, J., “On a family of distributions obtained from Eisenstein series II: Explicit for-
mulas”, Amer. J. of Math., Vol. 104, No. 6 (1982), pp. 1289–1336.
[A7] Arthur, J., “The trace formula for reductive groups”, Lecture for Journées Automorphes,
Dijon, 1981.
[AC] Arthur, J. and Clozel, L., Simple Algebras, Base Change, and the Advanced Theory of the
Trace Formula, Annals of Math. Studies No. 120, Princeton Univ. Press, 1989.
[Ge] Gelbart, S., Automorphic Forms on Adele Groups, Annals of Math. Studies No. 83, Prince-
ton Univ. Press, 1975.
[GJ] Gelbart, S. and Jacquet, H., “Forms of GL(2) from the analytic point of view”, in Proc.
Symp. Pure Math., Vol. 33, Part I, pp. 213–251.
[GGPS] Gelfand, I. M., Graev, M., and Piatetski-Shapiro, I., Automorphic Functions and Repre-
sentation Theory, Saunders, 1969.
[JL] Jacquet, H., and Langlands, R. P., Automorphic Forms on GL(2), Springer Lecture Notes,
Vol. 114, 1970.
[La1] Langlands, R. P., “On the functional equations satisfied by Eisenstein series , Springer
Lecture Notes, Vol. 544, 1976.
[Se] Selberg, A., “Harmonic analysis and discontinuous groups in weakly symmetric Rieman-
nian spaces, with applications to Dirichlet series”, J. Indian Math. Soc. 20 (1956), pp.
47-87; see also Proc. Int. Cong. Math. 1962, pp. 177–189.
8 STEVE GELBART

Lecture II. Arthur’s Modified Kernels I: The Geometric Terms

Our purpose here is to develop the geometric side of Arthur’s first form of the trace
formula X X
JoT (f ) = JχT (f ),
o∈O χ∈X

for the group G = GL(2). In particular, we must introduce the modified kernels
koT (x, f ), and prove they are integrable.
N.B. The treatment of the trace formula in [Ge] and [GJ] largely ignored the
integrability of these modified kernels.

1. The Definition of Ko . In the case of compact quotient, the geometric side of

the trace formula is summed over ordinary G(F )-conjugacy classes o in G(F ) (or
in Z(F )\G(F ), where Z is the center of G). For arbitrary G, it turns out to be
best to make the following definition: γ1 and γ2 in Z(F )\G(F ) are called equivalent,
or “conjugate”, if their semi-simple components are G(F ) conjugate in the usual
sense. That is, γ1 ∼ γ2 means, with the Jordan decompositions γi = γis γiu (i = 1, 2),
that γ1s is G(F )-conjugate to γ2s . For each such equivalence class o, and for f ∈
Cc∞ (Z(A)\G(A)), set X
Ko (x, y) = f (x−1 γy),
γ∈o

and let O denote the collection of all such classes o in Z(A)\G(A).

Note that if γ is elliptic, i.e. if γ is not conjugate in G(F ) to an element of any proper
parabolic subgroup P (F ), then it is automatically semi-simple. (Its eigenvalues do
not lie in F , and hence are distinct.) Thus our new notion of conjugacy reduces to
ordinary conjugacy.
On the other hand, suppose γ is G(F )-conjugate to an element p in a proper
parabolic
n o subgroup P (F ). For GL(2), this means P is the Borel subgroup B =
ax
0b
, so modulo Z(F ),
   
α0 1y
p = ps pu with ps = 01
, and pu = 01
.

Case (a): α = 1. In this case, the class o containing γ consists of two ordinary
unipotent conjugacy classes in Z(F )\G(F ): the trivial
 class {1},and the ordinary
conjugacy class of any non-trivial unipotent element 0 1 (since α0 01 already con-
1 x0
   
1 x0 1 αx0
jugates 0 1
to 0 1
).
Case (b): α 6= 1. In this case, o is an ordinary hyperbolic conjugacy class of the
form n   o
α0
oα = δ −1 0 1 δ | δ ∈ M(F )\G(F ) .
 LECTURE II. ARTHUR’S MODIFIED KERNELS I: THE GEOMETRIC TERMS 9
 
αx
Note that M(F ) is just the centralizer G α 0  (F ), and that 0 1
also belongs to oα
01
      −1
αx 1 y −1 α 0 1y
(since 0 1
= 01 01 01
where y = x/(α − 1) ).

2. The Integrability of Elliptic Ko . As suggested in Lecture I, it is a general

phenomenon that for an elliptic class o, the kernel Ko (x, y) is integrable over the
diagonal subset Z(A)G(F )\G(A). We now prove this for G = GL(2), following the
general arguments of [A1]. The crucial step is a lemma which first requires some
preliminary notions to be recalled.
According to the Iwasawa decomposition for G, we have
(2.1) G(A) = N(A) M(A) K,
 
1y
where N is the standard unipotent subgroup { 01
}, and K is the standard maximal
Q  
a0
compact subgroup Kν of G(A). If we denote by aM the Lie algebra of M = { 0b
},
i.e. ,
= Hom(X(M), R) ≈ R2 ,
aM
then we can define the familiar function
 
a0
HM : M(A) → aM by HM 0b
= (log |a|A , log |b|A ),
and extend it to G(A) by way of the decomposition 2.1, i.e. if g = nak in G(A), then
HM (g) = HM (a).
Lemma 2.1 ( see [A1, p 362] ). Suppose Ω is a subset of G(A) which is compact
modulo Z(A). Then there exists a number dΩ > 0 with the following property: if
γ ∈ G(F ) is such that
g −1 γng ∈ Ω
for some n ∈ N(A) and g ∈ G(A) with αH(g) > dΩ , then γ ∈ B(F ).
Assuming this Lemma (for the moment), let us quickly dispose of the integrability
of Ko . For this, and for other purposes later on, we also need to recall the notion of
a Siegel domain, together with an explicit integration formula for its volume.
First we refine the Iwasawa decomposition 2.1 by writing, for each g in G(A),
(2.2) g = znht mk
where z ∈ Z(A), n ∈ N(A), k ∈ K,
 
et 0
ht = 0 e−t
,
with et an idele equal to 1 at every finite place and the real number et at each infinite
place, and  
m = a0 01 ,
with a ∈ A1 = {a ∈ A× | |a|A = 1}. The corresponding integration formula, for f on
Z(A)\G(A), is
Z Z Z Z Z ∞
(2.3) f (g) dg = f (natk)e−2t dt d× a dn dk.
Z(A)\G(A) −∞
10 STEVE GELBART

Now for each compact subset C1 of N(A), compact subset C2 of A1 , and real number
c (not necessarily positive), let Sc denote the set of points g in G(A) of the form (2.2)
with n ∈ C1 , a ∈ C2 and t > c/2. Such a set is called a Siegel set. Note that
Sc ⊂ { g ∈ G(A) | H(g) > c }.
Fact 1 (see [Go, Thm 9]). For any c1 , c2 ,
Sc 1 ∩ γ Sc2 6= ∅
for only finitely many γ in G(F ) modulo B(F ).
Fact 2 (see [Go, p 16]). If Sc is sufficiently large (i.e. C1 and C2 are sufficiently large,
and c is sufficiently small ), then
(2.4) G(A) = G(F ) · Sc .
In other words, Sc is essentially a fundamental domain for G(F ) acting on G(A);
more precisely, Fact 2 says only that certain Sc contain a fundamental domain.
Henceforth, we always assume that Sc is so chosen that 2.4 holds, and we refer to
such a Sc as a Siegel domain.
Remark. From the above, it is clear that
Z(A)G(F )\G(A)
has finite volume. In fact, by 2.3 and 2.4,
Z Z Z Z Z Z ∞ 
−2t
dg ≤ dg = e dt d× a dn dk,
Z(A)G(F )\G(A) Z(A)\Sc c

with the outer three integrals over compact domains.

Proposition 2.2. If o is elliptic, then
X
x 7→ Ko (x, x) = f (x−1 γx)
γ∈o

is absolutely integrable over Z(A)G(F )\G(A), and its integral is

Z
Jo (f ) = m(Z(A)G(F )\G(A)) f (x−1 γx) dx.
Gγ (A)\G(A)

Proof. It suffices to prove the integrability assertion, since the ensuing formula for
Jo (f ) follows just as in the case of compact quotient.
First note that for x lying in any fixed set Ω1 , which is compact modulo Z(A), the
sum defining Ko (x, x) is actually finite and hence Ko defines a nice, smooth function
on Z(A)G(F )\G(A). Indeed, f (x−1 γx) 6= 0 implies x−1 γx ∈ Ωf = support of f .
Thus γ itself must lie in a compact set modulo Z(A), namely Ω1 Ωf Ω−1 1 , and such a
set can contain only finitely many γ in Z(F )\G(F ).
On the other hand, if γ is elliptic, then f (x−1 γx) 6= 0 implies (by Lemma 2.1) that
αH(g) < dΩf . I.e., the support of
X
Kell (x, x) = f (x−1 γx)
γ elliptic
 LECTURE II. ARTHUR’S MODIFIED KERNELS I: THE GEOMETRIC TERMS 11

lies in the “doubly truncated” Siegel set

Sc d = Z(A)C1 {ht }C2 K with c < 2t < dΩf .
Equivalently, the support of Kell is compact modulo Z(A), and the integrability is
clear.
Remark. The argument above actually proves that the full elliptic contribution is
absolutely integrable, i.e. , that
Z
|Kell (x, x)| dx < ∞.
Thus, by the Lebesgue dominated convergence theorem, it follows that
X Z
|Ko (x, x)| dx < ∞,
o elliptic

and hence that the sum X

Jo (f )
o elliptic

converges absolutely.
It remains now to prove Lemma 2.1. It might be tempting to do so for GL(2)
by simply trying to multiply out some 2 × 2 matrices (as we did in [Ge, p 201]).
However, this more quickly than not leads to an incomplete proof (as happened in
fact in [Ge]), whereas the approach of [A1, p 362] works simply and generally, as we
shall show below.
Proof of Lemma 2.1. Let
ρ : G → Gℓ(V )
denote the adjoint representation of G on the space of trace zero 2 × 2 matrices over
F ; its highest weight vector is Λ = α, the unique simple root of G with respect to
the torus M. Fix a basis {e0 , e1 , e2 } of V such that
ρ(a)ej = aΛ−jα ej for a ∈ M (j = 0, 1, 2)
ρ(n)e0 = e0 for n ∈ N and
ρ(w)e0 = e2
For ξ in V (A), define Y
kξk = k ξ ν kν ,
ν
where, for finite ν, k ξν kν = maxj {|ξνj |ν } (where (ξνj ) denote the coordinates of ξν
with respect to the basis {ej }), and for infinite ν, k ξν kν is defined by the Hilbert
space structure that makes {ej } into an orthonormal basis. (To make sense out of the
infinite product, we restrict ξ to be “primitive”, i.e. , k ξν kν = 1 for all but finitely
many ν.)
To prove the Lemma, suppose that
(∗) g −1 γng ∈ Ω
with n ∈ N(A) and Ω compact modulo Z(A). Write g = n1 ak, so that (∗) implies
a−1 n−1
1 γnn1 a ∈ KΩK.
12 STEVE GELBART

Because the map g 7→ ρ(g)e0 is continuous, and ρ is trivial on Z(A), it follows that

ρ(a−1 n−1
1 γnn1 a) · e0 ≤ e
2d0

for some d0 > 0. So suppose that γ does not belong to B(F ). Then by Bruhat’s
decomposition,
γ = b0 wn0
for some b0 in B(F ) and n0 in N(F ). But for any b ∈ B(A) and g ∈ G(A), our
assumptions on ρ and e0 imply

k ρ(gb) · e0 k = eαH(b) k ρ(g) · e0 k.

Thus (with γ = b0 wn0 ) we compute that

ρ((a−1 n−1
1 b0 w)(n0 nn1 a)) · e0 = e
αH(a)
ρ(a−1 n−1
1 b0 ) · e2

= e2αH(a) k ρ(n∗ ) · e2 k
≥ e2αH(a) k e2 k = e2αH(g)

(since H(g) = H(a), ρ(a) e2 = a−α e2 , and k ρ(n) e2 k ≥ k e2 k for any n in N(A)).
Thus the Lemma is proved for dΩ > d0 .

3. The Definition of KT o (x, y). For o the (unique non-trivial) unipotent class in
Z(F )\G(F ), we saw in Lecture I that Ko (x, y) is not integrable over the diagonal.
A somewhat less trivial computation (see Lecture III) shows that the hyperbolic
Ko (x, y) are also not integrable. So let us finally introduce the required modifications
of these kernels.
For P = B, let τbB denote the characteristic function of the positive Weyl chamber

aB + = { (r1 , r2 ) ∈ aB = aM | r1 − r2 > 0 },

i.e. , the cone in R2 where α(r1 , r2 ) = r1 − r2 is positive. Then define, for any o ∈ O
and T = (T1 , T2 ) in aB + ,
X
(3.1) koT (x, f ) = Ko (x, x) − KB,o (δx, δx) τbB (H(δx) − T )
δ∈B(F )\G(F )

where
X Z
KB,o (x, y) = f (x−1 γny) dn.
γ∈o∩Z(F )\M (F ) N (A)

This is Arthur’s modified kernel koT (x, f ).

Note that if o is elliptic, then o ∩ M(F ) = ∅. In this case KB,o ≡ 0, and

koT (x, f ) = Ko (x, x).

 LECTURE II. ARTHUR’S MODIFIED KERNELS I: THE GEOMETRIC TERMS 13

In general, the term subtracted in 3.1 from Ko (x, x) is the “correction term at ∞”
which ends up making koT (x, f ) integrable. Before proving this is so, let us note the in-
ductive nature of the definition of koT (thereby making more plausible the appearance
of KB,o ). Set
X
(3.2) KB (x, y) = KB,o (x, y)
o
X Z
≡ f (x−1 γny) dn.
γ∈Z(F )\M (F ) N (A)

Then KB (x, y) is just the kernel of R(f ) acting in L2 (Z(A)N(A)B(F )\G(A)) in place
of L2 (Z(A)G(F )\G(A). Indeed for ϕ in L2 (Z(A)N(A)B(F )\G(A)),
Z
R(f )ϕ(x) = f (x−1 y)ϕ(y) dy
Z(A)\G(A)
Z Z !
X
(3.3) = f (x−1 γny) dn ϕ(y) dy,
Z(A)N (A)B(F )\G(A) N (A)
Z(F )\M (F )

as claimed. In general, for an arbitrary G, the definition of koT (x, f ) takes into account
all the proper parabolics P of G, subtracting off from Ko (x, x) not just one term (for
the minimal parabolic B), but the sum
X X
(3.4) (−1)dim AP /Z KP,o(δx, δx) τbP (H(δx) − T ),
P ⊂G δ∈P (F )\G(F )

where
X
KP (x, y) = KP,o (x, y)
o∈O

is the kernel of RP (f ) in L2 (Z(A)NP (A)MP (F )\G(A)). (Here AP is the maximal

split torus in MP , the Levi component of P = MP NP , and RP is the right regular
representation of G(A) in L2 (Z(A)NP (A)MP (F )\G(A)).)

Remark. Observe that

koT (x, f ) = Ko (x, x).
for x in a compact (modulo Z(A)) set Ω (how large depends on T). Indeed, for x
in an appropriate such set Ω, τbB (H(δx) − T ) is identically zero∗ . Thus x is indeed
a modification of Ko (x, x) only “near infinity” (and in higher rank, more than one
parabolic is needed to effect these modifications near infinity . . . ).

∗
Proof: Fix c and T such that α(T ) > c. Fact 1 of Section 2 implies that in Sc , τbB (H(δx)−T ) 6= 0
only for finitely many classes δi in B(F )\G(F ). Then for each such class, τbB (H(δi x) − T ) = 0 on
the compact set ΩTi in Sc where c < H(δi x) < α(T ). Thus τbB (H(δx) − T ) is identically zero on
ΩT = ∩ni=1 ΩTi .
14 STEVE GELBART

4. The Integrability of koT (x, f ). Henceforth, let us assume that o is unipotent or

hyperbolic. To prove that koT (x, f ) is integrable over Z(A)G(F )\G(A), it suffices to
show that it is integrable over a Siegel domain Sc . So pick any T ∈ aB + , and write
Sc = BT ∪ BeT ,

where

BT = { g ∈ Sc | αH(g) > α(T ) > 0 },

and BeT is its complement in Sc . Clearly koT (x, f ) is integrable over BeT , since
c < αH(g) ≤ α(T ).
Thus it suffices to prove that koT (x, f ) is integrable over BT . (N.B. In general, the
partition of a Siegel domain requires more non-trivial “reduction theory” than is
evident in the case of GL(2); this is what leads to the “geometric” combinatorial
lemmas in §6 of [A2].)

Lemma 4.1. There exists a constant c0 such that αH(wn) ≤ c0 for all n ∈ N(A);
moreover, if γ in G(F ) satisfies the inequality
αH(γx) > c0
for some x in G(A) with αH(x) > c0 , it must follow that γ ∈ B(F ).

As the proof is similar to that of the crucial Lemma 2.1 in §2 of this Lecture, we
leave the details to the reader (see [A1, pp 334–335]).
Now fix T in aB + such that α(T ) > c0 . Then for x in BT , Lemma 4.1 implies that
τbB (H(δx) − T ) 6= 0
(if and) only if δ ∈ B(F ). Thus (for such x),
X X X
f (x−1 γx) = f (x−1 δ −1 γδx)τbB (H(δx) − T ),
γ∈o δ∈B(F )\G(F ) γ∈o

and hence
(
T
X X
ko (x, f ) = f (x−1 δ −1 γδx)
δ∈B(F )\G(F ) γ∈o
Z )
X
−1 −1
− f (x δ µnδx) dn τbB (H(δx) − T )
N (A) µ∈M (F )∩o

Next fix T such that α(T ) > max{c0 , dΩf }, with dΩf as in Lemma 2.1. Then by
Lemma 2.1, we may replace the sum here over γ ∈ o by a sum over
B(F ) ∩ o = N(F ) · M(F ) ∩ o
(This last equality holds since the class of γ in B(F ) is determined by its semi-simple
part, i.e. its M(F )-part.) Thus we get
 LECTURE II. ARTHUR’S MODIFIED KERNELS I: THE GEOMETRIC TERMS 15

(
X X X
T
ko (x, f ) = f (x−1 δ −1 µνδx)
δ∈B(F )\G(F ) µ∈M (F )∩o ν∈N (F )
Z )
−1 −1
(4.1) − f (x δ µnδx) dn τbB (H(δx) − T )
N (A)
XX X
= Fbδx,µ (ν) τbB (H(δx) − T )
δ µ ν∈N (F )−{1}

where
Fx,µ (n) = f (x−1 µnx)
defines a Schwartz-Bruhat function on N(A) ≈ A with Fourier transform
Z
Fb x,µ (ν) = f (x−1 µnx)ψN (nν) dn
N (A)
 
(Here ψN ( 10 x1 ) = ψ(x) for some fixed non-trivial character on F \A, and we have
applied the Poisson summation formula to Fδx,µ .) Finally we are ready to prove
Proposition 4.2. For any o ∈ O, koT (x, f ) is absolutely integrable over
Z(A)G(F )\G(A).
Proof. We may assume o is not elliptic. By the above arguments, it suffices to check
the integrability over BT , with α(T ) > max{c0 , dΩf } as above. As in that domain
formula (4.1) holds, we conclude
Z
(4.2) koT (x, f ) dx ≤
Z X X
Fbx,µ (ν) τbB (H(x) − T ) dx.
Z(A)B(F )\G(A) µ∈M (F )∩o ν∈N (F )−{1}
   
α0 1w
Note that for any µ = 01
in M(F ) ∩ o, ν = 0 1
in N(F ) − {1}, and x = znht mk,
Z     
α0 1y
Fbx,µ (ν) = f x−1 01 01
x ψ(yw) dy
A

= e2t Fbk∗ ,µ (e2t ν)

 2t

1e w
with k ∗ = (h−1 2t
t nht )mk. (Here e ν = 0 1 , and we confuse N(F ) with F .) Note
also that the variable x in (4.2) runs through a Siegel set. Thus k ∗ runs through a
compact set, and we have, for any N > 0
X
Fbk∗ ,µ (e2t ν) ≤ cN (e−2t )N
ν6=1

Applying Iwasawa’s decomposition (2.3) to (4.2), we then conclude that

Z Z Z Z Z !
∞
T −2tN
ko (x, f ) dx ≤ C e dt dk d× a dn < ∞
N (F )\N (A) F × \A1 K α(T )

as required.
16 STEVE GELBART

Summing up, we have

Theorem 4.3. For α(T ) sufficiently large,
XZ
koT (x, f ) dx < ∞.
o∈O Z(A)G(F )\G(A)
P
Hence the left side o∈O JoT (f ) of (1) of Lecture I is defined and absolutely con-
vergent.
Indeed, it suffices to remark that for a given f ∈ Cc∞ (Z(A)\G(A)), koT (x, f ) is
identically zero except for finitely many hyperbolic
  o. This isclear
 from the defining
−1 −1 α 0 −1 −1 α 0
equation (3.1), and the fact that x δ 01
δx and x δ 01
nδx can belong to
the fixed compact mod Z(A) set Ωf for only finitely many α; to see this, observe
that conjugating by δx preserves eigenvalues. Thus the sum over o in O in Theorem
4.3
PR
is infinite only over the elliptic classes, where we already know the relevant sum
|Ko (x, f )| dx is finite (see the Remark after the proof of Proposition 2.2).

Concluding Remarks. (a) The terms JoT (f ) are indeed generalizations of the
orbital integrals appearing on the geometric side of Selberg’s compact quotient trace
formula. Indeed, for o elliptic (which is automatic
R
for G anisotropic), JoT (f ) reduces
to a multiple of the usual orbital integral Gγ (A)\G(A) f (x−1 γx) dx.
We have seen—for GL(2)—that the kernel Ko (x, x) is already integrable on Sc if o is
an elliptic class (as opposed to hyperbolic or unipotent). For general G, the proper
distinction is between those o which never intersect any proper parabolic subgroup
subgroup P of G and those that do. Indeed Arthur does not talk about “elliptic” o
at all, but rather defines a modified kernel
koT (x, f )
for any o, and proves that koT (x, f ) is integrable. But of course—as we have al-
ready observed—koT (x, f ) = Ko (x, x) precisely when o is elliptic in this sense (that it
intersects no proper P . . . ).
References
[A1] Arthur, J., The Selberg trace formula for groups of F-rank one, Annals of Math. 100 (1974),
pp. 236–385.
[A2] Arthur, J., A trace formula for reductive groups II: Terms associated to classes in G(Q), Duke
Math. J., 45, No. 4 (1978), pp. 911–952.
[A7] Arthur, J., The trace formula for reductive groups, Lectures for Journées Automorphes, Dijon
(1981).
[Ge] Gelbart, S., Automorphic Forms on Adele Groups, Annals of Math. Studies No. 83, Princeton
University Press, Princeton NJ, 1975.
[GJ] Gelbart, S. and Jacquet, H., Forms of GL(2) from the analytic point of view, in Proc. Symp.
Pure Math. (Corvallis) , vol. 33 Part I, (1979), pp 213–251.
[Go] Godement, R., Domaines fondamentaux des groupes arithmetiques, Seminaire Bourbaki 257,
W. A. Benjamin, Inc., New York, (1963).
 Lecture III. Arthur’s Modified Kernels II: The Spectral Terms
arXiv:math/9505206v1 [math.RT] 10 May 1995

In order to describe the right hand side of Arthur’s trace formula, we need first
to recall the spectral expansion of the kernel
X
Kf (x, y) = Kχ (x, y).
χ∈X

1. The Spectral Kernels Kχ . Henceforth, by a “Levi subgroup M of G” we

understand the Levi component of a parabolic subgroup P of G; for any such M ,
we have
M 1 (A) = {m ∈ M (A) : |λ(m)|A = 1 ∀λ ∈ X(M )F }.
In general, the decomposition of

L2 (Z(A)G(F )\G(A))

into right G(A)-invariant subspaces is determined by spectral data χ = {(M, r)} ∈

X, where the pair (M, r) consists of a Levi subgroup M of G and a cuspidal repre-
sentation r of Z(A)\M (A); the class {(M, r)} derives from the equivalence relation

(M, r) ∼ (M ′ , r)

if and only if M is conjugate to M ′ by a Weyl group element w, and r′ = rw on

Z(A)\M 1 (A). Then to each such cuspidal datum χ = {(M, r)} is associated a
subspace L2χ , such that
M
L2 (Z(A)G(F )\G(A)) = L2χ ,
χ

and the corresponding decomposition of kernels

X
Kf (x, y) = Kχ (x, y)
χ

is explicitly describable in terms of Eisenstein series.

For G = GL(2), cuspidal data comes in two possible forms:
Case (i): M = G, r = a cuspidal representation π of Z(A)\G(A). In this case,
L2χ = Lπ , the irreducible subrepresentation of L20 (Z(A)G(F )\G(A)) realizing π,
and
X
(1.1) Kχ (x, y) = Kπ (x, y) = R(f )φ(x)φ(y)
{φ}= o.n. basis of Lπ
17
18 STEVE GELBART
     
a1 0 a1 0
Case (ii): M = ,r = µ aa21 with µ a character of F x \A1
0 a2 0 a2
(identified with the pair (M, µ−1 ). In this case,

XZ ∞
(1.2) Kχ (x, y) = Kµ (x, y) = E(x, ρ(µ, it)(f )φµ , it, µ) E(y, φµ , it, µ) dt.
{φµ } −∞

where {φµ } runs through an ON basis for the induced representation space
 
G(A) a1
IndB(A) µ = ρ(µ, o),
a2

consisting of functions φ : G(A) −→ C such that

    
a1 x a1 a1 1/2
φ g =µ φ(g)
0 a2 a2 a2 A

 
a1 x R
for all in B(A), and K |φ(k)|2 dk < ∞; ρ(µ, s) denotes the induced
0 a2  
G(A) s
representation space IndB(A) µ aa21 aa21 A , and E(g, φµ , s, µ) is the Eisenstein series

X
φsµ (γg)
γ∈B(F )\G(F )

attached to the function

φsµ (g) = φµ (g)esα(H(g))
in ρ(µ, s).
Remarks. (i) The series defining E(g, φµ , s, µ) converges only for Re(s) > 1/2,
but E(g) itself has a meromorphic continuation to all of C, with a possible pole at
s = 1/2. These and other facts from the theory of Eisenstein series we shall simply
assume, referring the reader to [GJ] for more details.
(ii) If µ2 ≡ 1, then (and only then) Kχ (x, y) has an additional “discrete” term,
namely
Z
(1.3) (R(f )µ)(x)µ(y) = f (g)µ(det g) dg µ(det x)µ(det y).
Z(A)\G(A)

(this comes from the residue of the corresponding Eisenstein series at s = 1/2 . . . ).
(iii) Above, Kχ (x, y) represents the kernel of R(f ) acting in the space L2χ , which
(for any χ = {(M, µ)}) consists of all ϕ(g) which are orthogonal to the space of
cusp forms L20 (Z(A)G(F )\G(A)), and such that for almost every x in G(A), the
projection of the function
Z
ϕxN (m) = ϕ(nmx) dn,
N (F )\N (A)
 LECTURE III. ARTHUR’S MODIFIED KERNELS II 19

onto the space L2 (Z(A)M (F )\M 1 (A)) transforms under Z(A)\M 1 (A) as a sum or
representations r, where (M, r) is a pair in χ.
The beauty of this description is that it generalizes to a general reductive group
G, as we shall now describe.
For any fixed parabolic P = MP NP in a reductive G, one may describe a de-
composition
M
(1.4) L2 (Z(A)NP (A)MP (F )\G(A)) = L2 (P ) = L2 (P )χ ,
χ∈X

2 2
where L (P )χ consists of those ϕ in L (Z(A)NP (A)MP (F )\G(A)) with the fol-
lowing property: for each standard parabolic subgroup B of G, with B ⊂ P , and
almost every x in G(A), the projection of the function
Z
m −→ ϕB,x (m) = ϕ(nmx) dn
NB (F )\NB (A)

onto the space of cusp forms in L2 (Z(A)MB (F )\MB1 (A)) transforms under MB1 (A)
as a sum of representations rB , in which (MB , rB ) is a pair in χ. If there is no such
pair in χ, ϕB,x will be orthogonal to the space of cusp forms on Z(A)MB (F )\MB1 (A)
(as happens above for G = GL(2), P = G = B, and χ = {(M, µ)}).
Remark. For any fixed χ in X, let Pχ be the set of standard parabolics B, where
(MB , rB ) belongs to χ. Then from the theory of Eisenstein series it follows

L2 (Z(A)NP (A)MP (F )\G(A))χ

will be zero unless there is a group in Pχ contained in P .
Let us return now to the decomposition
M
(1.4) L2 (Z(A)NP (A)MP (F )\G(A)) = L2 (P ) = L2 (P )χ ,
χ∈X

described above. If we let KP,χ (x, y) denote the integral kernel of the restriction
of RP (f ) to L2 (P )χ , then one can still—in general—write down a formula for
KP,χ (x, y) in terms of Eisenstein series. Clearly
X X
(1.5) KP,o (x, y) = KP,χ (x, y),
o∈O χ∈X

as each side equals the integral kernel of RP (f ) (see §3 of Lecture II, where we
computed the “geometric” expression for the kernel KP , for G = GL(2) and P =
Borel).
Summing Up. We have
X
Kf (x, y) = Kχ (x, y)
χ∈x

and (for any parabolic P of G),

X
KP (χ, y) = KP,χ (x, y).
χ∈x
20 STEVE GELBART

Concluding Remarks. (a) When P = G,

KP (x, y) = KG (x, y) = K(x, y)

and
KP,χ (x, y) = Kχ (x, y).

Thus the first spectral expansion is a special case of the second.

(b) For G = GL(2), and P = B (the Borel), we have L2 (P )χ = 0 if χ =
(G, π). On the other hand, if χ = {(MP , µ)}, then L2 (P )χ just consists of ϕ
in L2 (Z(A)NP (A)MP (F )\G(A), which under left action by M 1 (A) transforms ac-
cording to a sum of µ and µ−1 . The corresponding kernel KB,χ (x, y) is given by
the formula
Z
(1.6) KB,χ (x, y) = KB (x, my)(µ(m) + µw (m)) dm.
Z(A)M(F )\M 1 (A)

This is easily checked by computing the composition of R(f ) with the projection
operator Pχ defined on L2 (Z(A)N (A)M (F )\G(A)) by
Z
(Pχ ϕ)(y) = ϕ(my)(µ(m) + µw (m)) dm.
Z(A)M(F )\M 1 (A)

Moreover, an analogue of (1.6) (without the subscript B) holds for the kernel
Kχ (x, y).
(c) The series X X
Kχ (x, y) = KG,χ (x, y)

and X
KP,χ (x, y)

converge absolutely (to K(x, y) and KP (x, y) respectively).

(d) In general, for any G and χ = {(M, r)}
X
(1.7) KG,χ (x, y) =
X Z
1 X X
E(χ, ρ(σ, λ)(f )φ, λ)E(y, φ, λ) dλ,
n(AP ) σ ia∗
P /iaG
∗
P φ

where: the sum over P is over “associated” parabolics in G, n(AP ) is the number
of chambers in aP , a∗P = X(MP ) ⊗ R is dual to aP = Hom(X(MP )F , R), ρ(σ, λ) is
G(A)
the right regular representation in IndP (A) σ ⊗ eλ(HP (p)) , φ runs through a suitable
K-finite) basis for ρ(σ, λ), and—critically, σ runs through (classes of) irreducible
G(A)
unitary representations of M (A) such that functions in IndP (A) σ belong to L2 (P )χ .
Note that for χ = {(G, π)}, (1.3) just reduces to
X
R(f )φ(x)φ(y),
φ
 LECTURE III. ARTHUR’S MODIFIED KERNELS II 21

since L2 (P )χ ≡ 0 unless P = G. Thus our original formula for Kπ (x, y) is indeed

consistent with the statement that any KP,χ (x, y) is expressible in terms of Eisen-
stein series (but now the Eisenstein series in question belong to φ in Lπ , viewed
as induced from themselves on P = G). At the same time, formula (1.7) is also
general enough to include the extra term R(f )(µ)(x)µ(y) in case G = GL(2), and
χ = {(M, µ)} with µ = µ−1 ; indeed, the parabolic P = B in (1.7) contributes the
leading term (1.2) to Kχ (x, y), while the parabolic P = G contributes the crucial
additional term (1.3).

2. Modified Kernels and Truncated Kernels. Recall the identity

X X
(1.5) KP,o (x, y) = KP,χ (x, y),
o∈O χ∈X

introduced towards the end of the last section. By complete analogy with the
geometric side (compare formulas (3.1) and (3.4) of Lecture II), Arthur defines the
modified spectral kernel functions
X X
(2.1) kχT (x, f ) = (−1)dim(AP /Z) KP,χ (δx, δx)τ̂P (HP (δx) − T ),
P ⊂G δ∈P (F )\G(F )

which in the case of GL(2) look like

X
kχT (x, f ) = Kχ (x, x) − KB,χ (δx, δx)τ̂B (HB (δx) − T ).
δ∈B(F )\G(F )

Because of the identity (1.5), one clearly has

X X
(2.2) koT (x, f ) = kχT (x, f ),
o∈O χ∈X

i.e., the modification of the geometric expression for the kernel of R(f) is equal to
the modification of the spectral expression for this kernel.
Despite its simplicity, this identity (2.2) is the starting point of Arthur’s trace
formula in the case of non-compact quotient. Of course, we emphasize (again!)
that neither side of (2.2) represents (in general) the kernel of R(f ). However, we
have already seen in Lecture II that each of the functions koT (x, f ) is (absolutely)
integrable over Z(A)G(F )\G(A), and that the sum of the resulting distributions
Z
JoT (f ) = koT (x, f ) dx
Z(A)G(F )\G(A)

also converges absolutely. Thus by (2.2), we have shown that

X Z X
(2.3) JoT (f ) = kχT (x, f ) dx,
o Z(A)G(F )\G(A) χ
22 STEVE GELBART

i.e., to obtain the first form of Arthur’s trace formula, we need “only” show that the
integral (on the right side of (2.3) may be taken inside the sum over χ; in particular,
each kχT (x, f ) is integrable, and the sum of the resulting distributions
Z
(2.4) JχT (f ) = kχT (x, f ) dx

converges absolutely to the left side of (2.3).

To prove the absolute integrability of each kχT (x, f ), and of the sum of such
kernels, it turns out to be necessary to introduce a truncation operator on
G(F )\G(A). Given T in a+ as before, the truncation of a continuous function
ϕ(x) on Z(A)G(F )\G(A) is the function
X
(2.5) ΛT ϕ(x) = ϕ(x) − ϕN (δx)τ̂B (H(δx) − T ),
δ∈B(F )\G(F )

where Z
ϕN (δx) = ϕ(nx) dn
N (F )\N (A)

denotes the constant term of ϕ. (Again, for x ∈ Sc , the sum over δ is finite, by
Fact 1 of Lecture II . . . ) In general, for arbitrary reductive G,
X X
(2.6) ΛT ϕ(x) = (−1)dim(AP /Z) ϕNP (δx)τ̂P (H(δx) − T ).
P ⊂G δ∈P (F )\G(F )

Note that if ϕ(x) is cuspidal, i.e., ϕNP = 0 for every NP , then

ΛT ϕ = ϕ.

Of course, for a general automorphic ϕ, this is no longer true; however, the whole
idea of truncation is that ΛT ϕ still equals ϕ on a large part of Sc (how large
depends on T ), and suitably modifies ϕ “near infinity” so as to be integrable over
all of Sc . In particular, for given “large” T , there is a compact (mod Z(A)) set Ω
such that
ΛT ϕ ≡ ϕ on Ω.

(Both this statement—and its proof—resembles its analogue for the modified ker-
nels koT (x, f ); see Remark 3.5 of Lecture II.)
Now for any χ in X, let ΛT2 Kχ (x, x) denote the function obtained by truncating
the function
KG,χ (x, y) = Kχ (x, y)

with respect to the second variable and then setting y = x. The strategy Arthur
follows in [A3] to prove that
Z !
X XZ
(2.7) kχT (x, f ) dx = kχT (x, f ) dx
χ χ
 LECTURE III. ARTHUR’S MODIFIED KERNELS II 23

is the following: Using properties of ΛT , he first shows that

Z X XZ
T
(2.8) Λ2 kχ (x, x) dx = ΛT2 kχ (x, x) dx
χ χ

with
XZ
ΛT2 kχ (x, x) dx < ∞
χ

and Z X
ΛT2 kχ (x, x) dx < ∞;
χ

then he shows that for sufficiently large T ,

Z X
(2.9) ΛT2 kχ (x, x) − kχT (x, f ) dx = 0.
χ
P P
From these facts it follows that o JoT (f ) converges absolutely to χ JχT (f ), where
for sufficiently large T , JχT (f ) is given by either (absolutely convergent) integral
Z Z
kχT (x, f ) dx or ΛT2 Kχ (x, x) dx.

In particular, the first form of Arthur’s trace formula is established.

Concluding Remarks. Set
X
Kcusp (x, y) = Kχ (x, y)
χ∈X(G)

where X(G) denotes the set of cuspidal data {(M, σ)} with M = G, and R0 be the
regular representation of G(A) restricted to
M
L20 = L2χ .
χ∈X(G)

Then for any T ,

(2.10) ΛT2 Kcusp (x, y) = Kcusp (x, y),

this kernel is integrable over the diagonal, and equals to R0 (f ). Thus Arthur’s first
form of the trace formula gives the more familiar trace formula
X X
(2.11) tr R0 (f ) = JoT (f ) − JχT (f ).
o∈O X\X(G)

Here only the proof of (2.10) is elementary; it results from the observation that
y −→ Kcusp (x, y) is cuspidal (see the beginning of §4 below). The remaining facts
are discussed in [A1], at least for the case of rank 1. To prove that tr R0 (f ) exists
and equals RKcusp (x, x) dx one uses arguments similar to those used in the (earlier)
proof that koT (x, f ) dx < ∞.
24 STEVE GELBART

XZ
3. Proof of (2.8) and the Absolute Convergence of ΛT2 Kχ (x, y) dx <
χ
∞. When Arthur described the spectral expansion
X
(3.1) K(x, y) = Kχ (x, y).

in [A2], he stressed its absolute convergence by way of the following:

Lemma 3.2. (see Lemma 4.4 of [A2]) There exists an N such that for any differ-
ential operators D1 and D2 in U(g),

XX Z X
n(A)−1 D1 E(x, ρ(σ, λ)(f )(φχ ), λ)D2 E(y, φχ , λ) dλ
χ P ia∗
P /iaG φχ
∗

≤ c(D1 , D2 )kxkN kykN .

Here kxk denotesQa “norm” or “height” function on G(A), which we may (for
GL(2)) take to be v kxv k, with kxv k = sup{|(xv )ij |v , |(x−1 v )ij |v }. This norm
satisfies the properties kxk = kx−1 k, kx1 x2 k ≤ kx1 kkx2 k, and kx > kc1 , for some
c1 ; moreover, it is commensurable with eα(H(x)) , at least for x in a fixed Siegel
domain.
The above Lemma is essentially part of Langlands’ theory of Eisenstein series;
from it, we get

(3.3) |D1,x D2,y Kχ (x, y)| ≤ c(D1 , D2 )kxkN kykN ,

with a similar identity for K(x, y) itself . (Here D2,y indicates that D2 is operating
only in the y variable, etc..) These identities say that K(x, y) and Kχ (x, y) are
“slowly increasing” in each variable. It is by way of these facts that the finiteness
of
XZ
|ΛT2 K(x, x)| dx
χ

is established.

Lemma 3.4. (see Lemma 1.4 of [A3]) The truncation operator transforms “slowly
increasing functions” ϕ into “rapidly decreasing” ones; more precisely, suppose ϕ
in C ∞ (Z(A)G(F )\G(A)) is such that for some N (the “degree of slow increase”),

|Dϕ(x)| ≤ CD kxkN for all x ∈ Sc

and differential operators D. Then for any M

|ΛT ϕ(x)| ≤ CM kxk−M

for x in Sc (and some CM ).

 LECTURE III. ARTHUR’S MODIFIED KERNELS II 25

Remarks. (a) Since we are not proving Lemma (3.5), let us at least explain why
it is plausible for G = GL(2). In this case, it is clear from Lemma 2 of Lecture II,
that for x “near infinity”, in particular, such that α(H(x) > α(T ) > C0 ,
ΛT ϕ(x) = ϕ(x) − ϕN (x),
Thus ΛT removes the constant term form φ, making it rapidly decreasing at ∞
(think roughly of a classical modular form for SL2 (Z) with vanishing constant term
. . . ).
(b) What Lemma 1.4 of [A3] actually asserts is that for any pair of positive integers
M and N , we can choose D1 , . . . Dn such that for any x in Sc ,
X 
T ′ ′ −N
(3.5) |Λ ϕ(x)| ≤ sup |Di ϕ(x )| · kx k · kx′ k−M
i x′

But if ϕ happens to be slowly increasing of degree N , along with its derivatives, then
the expression in parentheses is bounded, and so Lemma 3.4 immediately follows.
It is this latter form of the Lemma—that is, (3.5)—that we shall now apply.
Proposition 3.5. The truncated kernels ΛT2 K(x, x) and ΛT2 Kχ (x, x) are abso-
lutely integrable, and
Z XZ
ΛT2 K(x, x) dx = ΛT2 Kχ (x, x) dx.
χ

Proof. Let us show that for any M , there exists constants c′ and cχ such that for
x in Sc ,
ΛT2 K(x, x) ≤ c′ kxkN −M and ΛT2 Kχ (x, x) ≤ c′χ kxkN −M .
Since kxk ≍ eα(H(x)) on any Siegel domain, this will suffice, by Iwasawa’s decom-
position, to prove that Z
|ΛT2 K(x, x)| dx < ∞,

and hence by (Lebesgue’s) dominated convergence theorem, that

Z XZ
ΛT2 K(x, x) dx = ΛT2 K(x, x) dx,
χ

an absolutely convergent sum.

Applying Lemma 3.4 (in the form of identity (3.5)) to the second variable of
K(x, y) yields
|ΛT2 K(x, y)| ≤ C0 kyk−M · sup sup (|Di,y K(x, y ′ )|ky ′ k−N ),
y ′ D1 ,...Dn

which by (3.3) is dominated by

C0 kyk−M sup(C(1, Di )kxkN ky ′ kN )ky ′ k−N = C0 kyk−N sup(C(1, Di ))kxkN .
y ′ ,i i

So now setting x = y gives the desired conclusions (for K(x, x)), and for Kχ (x, x)
one argues similarly.
26 STEVE GELBART
R R
4. Proof of (2.9) Relating ΛT2 Kχ (x, x) dx to kχT (x, t) dx. Note first that
when χ = (G, π), there is nothing to prove. Indeed, if Kπ is the kernel of R(f )
restricted to any cuspidal subspace Lπ of L2 (Z(A)G(F )\G(A)), then

y −→ Kπ (x, y)

is itself cuspidal: for any ϕ∗ in (L20 )⊥ ,

Z
Kπ (x, y)ϕ∗ (y) dy = R(f ) ◦ (Projection onto Lπ ⊂ L20 ) = 0.

Thus it follows ΛT2 Kπ (x, y) = Kπ (x, y). On the other hand, as recalled in §1,
KP,χ (x, y) ≡ 0 for all proper P in G (when χ = (G, π)). Thus we also have
(from the definition (2.1) of kχT (x, f )) that kχT (x, f ) = KG,χ (x, x) = Kπ (x, x), i.e.,
ΛT2 Kπ (x, x) ≡ kχT (x, f ), and there is nothing to prove.
Henceforth,
 we shall  fixG = GL(2), and assume χ = {(M, σ)} with M =
a 0 a 0
, and σ = µ(a/b) (or rather its class, relative to the relation
0 b 0 b
µ ∼ µ−1 in (F x \A1 )∧ ). Adapting the general arguments of §2 of [A3] to this case,
we shall now (at least) prove that
Z
|ΛT2 Kχ (x, x) − kχT (x, f )| dx = 0

for sufficiently large T .

By definition,

− (ΛT2 Kχ (x, x) − kχT (x, f )) =

X Z
Kχ (x, nδx)τ̂B (HB (δx) − T ) dn
δ∈B(F )\G(F ) N (F )\N (A)
X
− KB,χ (δx, δx)τ̂B (HB (δx) − T ).
δ∈B(F )\G(F )

Note that Kχ (x, nδx) = Kχ (δx, nδx) (since E(x, φ) is left G(F )-invariant). There-
fore the above difference equals
(Z )
X
(4.1) τ̂B (HB (δx) − T ) Kχ (δx, nδx) dn − KB,χ (δx, δx) .
δ∈B(F )\G(F ) N (F )\N (A)

Next recall our expression (1.6) for KB,χ (and KG,χ ) in the Concluding Re-
mark (b) of §1. Applying it to (4.1), with y = δx, we get that the expression in
parenthesis equals
Z
Kχ (y, ny) dn − KB,χ (y, y) =
N (F )\N (A)
Z (Z )
(K(y, nmy) − KB (y, my)) dn ×
Z(A)M(F )\M 1 (A) N (F )\N (A)

µ(m) + µ−1 (m) dm.

 LECTURE III. ARTHUR’S MODIFIED KERNELS II 27

Applying the definition of K and KB , we find, in turn, that the last expression in
parentheses equals
Z X Z X
−1
f (y γ n m y) dn − f (y −1 γ n m y) dn
N (F )\N (A) γ∈Z(F )\G(F ) N (A) µ∈Z(F )\M(F )
 
Z X Z X
=  f (y −1 γ n m y) − f (y −1 γ n m y) dn
N (A) γ∈N (F )Z(F )\G(F ) N (A) µ∈Z(F )\M(F )

Now apply Bruhat’s decomposition

Z(F )N (F )\G(F ) = Z(F )\M (F ) ∪ (Z(F )\M (F ))wN (F )

to the first summation above. The result is that the entire last expression equals
Z X X
f (y −1 µ w ν n m y) dn.
N (A) ν∈N (F ) µ∈Z(F )\M(F )

Using this, it remains to show that

Z
|ΛT2 Kχ (x, x) − kχT (x, f )| dx
Z X
≤ τ̂B (H(δx) − T ) ×
Z(A)G(F )\G(A) δ∈B(F )\G(F )
Z Z XX
f (x−1 δµ w ν n mδχ)(µ(m) + µ−1 (m) dm dχ.
Z(A)M(F )\M 1 (A) N (A)

Z
= τ̂B (H(y) − T ) ×
Z(A)B(F )\G(A)
(4.2)
Z Z X X 
f (y −1 δµ w ν n m y)(µ(m) + µ−1 (m) dm dy

=0 for sufficiently large T .

So suppose y is such that this expression (4.2) is not zero. Then our hypothesis
on f implies
y −1 µ w ν n m y
belongs to the compact (mod Z(A)) subset Ωf of G(A). Writing out the Iwasawa
decomposition n1 ak for y implies

g = a−1 n−1 ′
1 ν w µ n m n 1 a ∈ Ωf

with Ω′f the compact ( mod Z(A)) set KΩf K, i.e., α(H(g)) remains bounded
(from below and above).
28 STEVE GELBART

On the other hand, write

g = a−1 n−1
1 ν w µ n m n1 a
−1 −1 −1
= (n−1
1 ν)
a
a−1 )w µ m a (nm n1 )a
= n∗ a∗ wn′

where n∗ , n′ ∈ N (A) and a∗ = a−1 aw (µm)m . Then

c < H(g) = H(a∗ ) + H(wn′ ) = −2H(a) + H(wn′ )

implies

−2H(a) > c − H(wn′ ),

i.e.
H(wn′ ) c c0 − c
(4.3) H(a) < − < = c′
2 2 2

(using Lemma 4.1 of Lecture II).

To complete the proof, take T “sufficiently large” to mean that α(T ) > c′ . Then
the condition τ̂B (H(y) − T ) = τ̂B (H(a) − T ) 6= 0 will be incompatible with (4.3),
i.e. (4.2) will be identically zero, as claimed.

5. An Alternate Strategy?. In any treatment of the trace formula prior to the

1978 paper [A2] (see [A1]), no modified kernels appear. For example, in [GJ] one
simply and (formally) truncates both sides of the identity
X X
K(x, y) = f (x−1 γy) = Kχ (x, y)
γ

to obtain a trace formula (for GL(2)) of the form

Z XZ
(5.1) ΛT2 K(x, x) dx = ΛT2 Kχ (x, x) dx.
χ

Our purpose now is to explain where the modified kernels are actually hiding in(5.1)
and why they can not be avoided in the general theory.
Using the crucial Lemma 2.1 of Lecture II, one can easily prove (see p. 235 of
[GJ]):
Lemma 5.2. For T sufficiently large (i.e. α(T ) larger than the constant dΩf
described in Lemma 2.1 of Lecture II ),
X X
ΛT2 K(x, x) = KoT (x, f ) + KoT (x, f ) ≡ K T (x, f )
o elliptic o unipotent
or
hyperbolic
 LECTURE III. ARTHUR’S MODIFIED KERNELS II 29

Note that Lemma (5.2) implies that the truncated geometric kernel used in (5.1)
is already a sum of the modified geometric kernels. Thus an alternate strategy for
obtaining the trace formula
X X
(5.3) JoT (f ) = JχT (f )
o χ

suggests itself: since we also have

X
ΛT2 K(x, x) = ΛT2 Kχ (x, x).
χ

(the spectral analogue of the geometric formula for ΛT2 K(x, x) afforded by Lemma
(5.2)) , why not just integrate each of these expressions term by term to obtain the
trace formula identity (5.3) with
Z
JχT (f ) = ΛT2 Kχ (x, x) dx?
Z(A)G(F )\G(A)

Indeed, by Lecture II we know the left hand side of (5.3) is nicely behaved, and by
the first three sections of this lecture, we know the right hand side converges too.
This strategy was in fact carried out in [GJ] for GL(2), where it turns out that
terms on either side of (5.3) can be computed completely explicitly, as (linear)
polynomials in T (see §6 of [GJ] and §§1 and 2 of the next lecture). Then when
these expressions are plugged into the formula
X X
tr R0 (f ) = JoT (f ) − JχT (f ),
o∈O χ∈X\X(G)

the terms depending on T cancel out (as they must, since the left hand side is
independent of T ). Thus we are left with our sought-after formula for tr R0 (f ).
What goes wrong with this strategy in general? Well, for an arbitrary reductive
group G, it would be hopeless to try and compute each term JoT (f ) and JχT (f )
completely explicitly; fortunately, however, it is also unnecessary. What Arthur’s
more indirect strategy (involving modified spectral as well as geometric kernels)
allows one to prove is that JχT (f ) and JoT (f ) are a priori polynomials in T (for
sufficiently large α(T )). Thus we can concern ourselves only with the constant
terms of these polynomials (since our goal is to make (5.3) explicit, and only terms
not depending on T will survive on the right-hand side . . . ). These constant terms
((Jo (f ) ≡ Jo0 (f ) and (Jχ (f ) ≡ Jχ0 (f )) are by no means trivial to compute, but they
do turn out to be feasible to treat in Arthur’s general theory, as we shall soon see.
30 STEVE GELBART

Lecture IV. More Explicit Forms of the Trace Formula

In the first few lectures, we have described the first form of Arthur’s trace for-
mula, which in its symmetric formulation reads
X X
(*) JoT (f ) = JχT (f );
χ

more suggestively,
X X
(**) tr(R0 (f )) = JoT (f ) − JχT (f ).
o χ∈X\X(G)

After developing this formula in [A2] and [A3], Arthur devoted the next ten years
(and several hundred pages of work) to rewriting this formula in a more explicit
form, suitable for applications. In particular, in order to apply (∗∗) effectively to the
theory of automorphic forms, it seems essential to know a lot about the distributions
JoT (f ) and JχT (f ). How do they depend on T ? How can they be expressed explicitly
in terms of weighted orbital integrals or weighted characters? How can they
be factored into local distributions on the groups Gv ?
Our purpose in the remainder of these lectures is to explain some of Arthur’s
answers to these questions. But to make this task easier, we proceed in stages.
One of Arthur’s first general results is that each distribution JoT (f ) or (JχT (f )) is
a polynomial in T , of degree at most dim(AP0 /Z). Another is that each “unram-
ified” JoT (f ) (or JχT (f )) has a fairly explicit representation as a weighted orbital
integral (or character). To motivate these and other general results, we shall first
recall the familiar example of GL(2).

1. Explicit Results for GL(2) (Geometric Side). As we already observed in

general in Lecture II, the elliptic terms JoT (f ) are independent of T , and simply
equal ordinary (factorizable) orbital integrals (see Proposition 2.5 of Lecture II).
On the other hand, the hyperbolic and unipotent terms are non-trivial (linear)
polynomials in T , explicitly described in the propositions below.
Proposition 1.1. If o denotes a hyperbolic class with representative γ, then for
α(T ) sufficiently large,
Z
T ∗ 1
Jo (f ) = (T1 − T2 )m(F \A ) f (x−1 γx) dx
M(A)\G(A)
Z
1 ∗ 1
− m(F \A ) f (x−1 γx)α(H(wx) + H(x)) dx
2 M(A)\G(A)
 LECTURE IV. MORE EXPLICIT FORMS OF THE TRACE FORMULA 31

Remarks. (1) The proof of this proposition is provided by the formal computa-
tions of [GJ, pp.
R 237–238] now justified by the already established a priori absolute
convergence of koT (x, f ) dx. The assumption α(T ) ≫ 0 is needed so as to be able to
apply Lemma 4.1 of Lecture II (which in the case of GL(2) just requires α(T ) > 0).
(2) In §3 of Lecture I, we saw that the unmodified hyperbolic kernel Ko was not
integrable because of the appearance of the (divergent) volume of F x \Ax ; Propo-
sition 1.1 shows that the integral of the modified kernel indeed “truncates” this
R interval of length (T1 − T2 ), which appears as a linear term in T
volume to a finite
with coefficient f (x−1 γx) dx; however, the constant term of this polynomial no
longer involves an invariant orbital integral (because of the presence of the “weight
function” α(H(wx)) + H(x)).
Proposition 1.2. For the unipotent class o, and α(T ) sufficiently large,

Jo (f ) = m(Z(A)G(F )\G(A))f (1)

Z
∗ 1
+ f. p. (ζ(F, s)) + (T1 − T2 )m(F \A ) F (y) dy
s=1 A

where Z  
1 a
F (y) = f (k −1 0 1
k) dk is in S(A),
K

and f. p.s=1 (ζ(F, s)) denotes the “finite part” at s = 1 of the (meromorphic) zeta-
function Z
ζ(F, s) = F (a)|a|s dx a
Ax

of Tate.
Remarks. (1) Iwasawa’s decomposition implies (formally) that
Z Z    
ζ(F, s) (at s = 1) = f k −1 10 a1 k dk |a| dx a
x K
ZA    
= f x−1 10 11 x dx,
Z(A)N (A)\G(A)

the non-trivial unipotent orbital integral of f . Thus the constant term of this JoT (f )
is indeed analogous to the constant term of the hyperbolic JoT (f ): one involves a
regularized (unipotent) orbital integral, and the other a weighted (hyperbolic)
orbital integral.
(2) Because the left hand side of (**) is independent of T , it must be that the linear
terms in T appearing in both theP hyperbolic and unipotent JoT (f ) cancel out with
T
the linear terms appearing in Jχ (f ). Indeed one can check (using Iwasawa’s
decomposition) that the linear terms in Propositions 1.1 and 1.2 (for all hyperbolic
o) combine into the single term

X Z Z
∗ 1
L(T ) = (T1 − T2 )m(F \A ) f (k −1 γnh) dn dh;
γ∈Z(F )\M(F ) K N (A)
32 STEVE GELBART

then
P this term is seen (later) to exactly cancel the linear term in T arising from
T
χ Jχ (f ); see §2 below, also p. 239 of [GJ]).
(3) The proof of Proposition 1.2 again results from formal computations—see p. 236
Rof [GJ]—all of which are justified now by the a priori absolute convergence of
koT (x, f ) dx. As in the hyperbolic case, the appearance of the linear term, and the
“shearing off” (i.e., regularization) of the relevant orbital integral, are explained by
the fact that the kernel Ko (x, x) was modified before being integrated.
(4) The need to modify Ko (in this unipotent case) is clarified by the exercise below.
Exercise. Show that for the unipotent class o,
Z
JχT (f ) = Ko (x, x) dx
Z(A)G(F )\G(A)
= val {ζ(F, s)} + m(Z(A)G(F )\G(A))f (1),
s=1

with ζ(F, s) as in the proposition above. In particular, Jo (f ) = ∞ when the

“principal part” of ζ(F, s), namely m(F ∗ \A1 )F̂ (0)/(s − 1), is non-zero.
Solution. Recall that
o = {1} ∪ o′ ,
where o′ consists of the elements
n   o
δ −1 10 1t δ : t ∈ F ∗ , δ ∈ B(F )\G(F )

Thus we compute
Z Z X X    
k (x, x) dx =
o′ x−1 δ −1 10 1t δx dx
Z(A)G(F )\G(A) t6=0 δ∈B(F )\G(F )
XZ    
= f x−1 10 1t x dx,
t6=0 Z(A)B(F )\G(A)

which, by Iwasawa’s decomposition, equals

Z Z Z !
X     X
−1 1 a−1 t
f k k |a|−1 dx a dk = F (at) |a| dx a
0 1
K F x \Ax t∈F x F x \Ax t∈F x
Z
= F (a)|a| dx a
A∗

with F (y) as defined in Proposition 1.2.

Concluding Remark. As we shall see in the next lecture, the explicit expressions
just obtained for JoT (f ) (hyperbolic or unipotent) already suffice to express these
distributions as (a sum of) products of local distributions Jv (fv ); this fact will turn
out to be crucial in yielding a simple form of the trace formula in Lecture V.
 LECTURE IV. MORE EXPLICIT FORMS OF THE TRACE FORMULA 33

2. Explicit Results for GL(2) (Spectral Side). As already observed in gen-

eral in Lecture III, the “elliptic” spectral terms JχT (f ), where χ = {(G, π)}, are
independent of T ; they just reduce to the distributional characters

JχT (f ) = tr π(f ).

As for the remaining “hyperbolic” and “unipotent” χ, we have the following:

Proposition 2.1. Suppose χ = {(M, µ)} is “hyperbolic” in the sense that µ 6=
µ−1 (i.e., there is no 1-dimensional residual spectrum contributing to Kχ (x, y)).
Then for T sufficiently large,
Z ∞
(2.1) JχT (f ) = (T1 − T2 ) tr(ρ(µ, it)(f )) dt
−∞
Z ∞
+ tr(M (−it)M ′ (it)ρ(µ, it)(f )) dt
−∞

with the intertwining operators M (s) (and their derivatives) to be explained below.
Because the proof of this result directly generalizes to any G, we include it.
By definition, for χ regular (µ 6= µ−1 ),
Z X Z ∞ 
(2.2) JχT (f ) = E(x, ρ(µ, it)(f )φ, it)ΛT E(x, φ, it) dt dx.
Z(A)G(F )\G(A) φ −∞

However, because we have already established the absolute convergence of this

integral, we can interchange the orders of integration and write
XZ ∞
T
(2.3) Jχ (f ) = (ΛT E(ρ(µ, it)(f )φ, it), ΛT E(φ, it)) dt.
φ −∞

Here we have used also the well-known fact that ΛT is an orthogonal projection
operator on L2 , and hence (φ1 , ΛT φ2 ) = (ΛT φ1 , ΛT φ2 ), with (·, ·) denoting the
usual inner product in L2 .
To continue with the proof, we obviously need a more explicit expression for
the inner product of these truncated Eisenstein series. This is provided by the
“Maass–Selberg relations” (generalized by Langlands to an arbitrary G in [La1]);
in the present context they imply

(2.4)
(ΛT E(φ′ , it), ΛT E(φ, it))
= (T1 − T2 )(φ′ , φ) + (M (−it)M ′ (it)φ′ , φ)
1
+ {(φ′ , M (it)φ)e2it(T1 −T2 ) − (M (it)φ′ , φ)e−2it(T1 −T2 ) }.
2it
Here M (s) = M (s, µ) is the operator
Z
φ −→ M (s)φ(g) = φ(wng) dn
N (A)
34 STEVE GELBART

intertwining ρ(µ, s) with ρ(µw , −s). (Like E(g, φ, µ, s), M (s) is initially defined
only for Re(s) > 1/2, but is analytically continuable in C, and holomorphic on
iR.) Note M (−s) = M (−s, µw ) maps ρ(µw , −s) to ρ(µ, s), and the inner product
(φ1 , φ2 ) is defined by
Z Z
(φ1 , φ2 ) = φ, (mk)φ2 (mk) dm dk.
Z(A)M(F )\M 1 (A) K

d
Finally, M ′ (s) is the derivative ds M (s, µ) : ρ(µ, s) −→ ρ(µw , −s) (defined by iden-
tifying ρ(µ, s) with ρ(µ, 0) . . . ).
Now formula (2.4) is valid not just in our case (when µ is “regular”), but also
when µ = µ−1 ; in fact we shall require (2.4) below when we treat the “unipotent”
contribution JχT (f ). However, when µ 6= µ−1 , it is a simple matter to observe that
the third (and most complicated) term on the right side of (2.4) actually vanishes.
Indeed, as M (s) = M (s, µ) maps φ into ρ(−s, µ−1 ), it follows that φ′ and M (it)φ
transform under M 1 (A) according to distinct characters (namely, µ and µ−1 ). Thus
the inner product
Z Z
(φ′ , M (it)φ) = φ′ (k)M (it)φ(h) dh( µ2 (m) dm)
Z(A)M(F )\M 1 (A)

vanishes (and so, similarly, does (M (it)φ′ , φ)). So plugging (2.4) into (2.3) simply
yields
Z ∞ Z ∞
(2.5) (T1 − T2 ) tr(ρ(µ, it)(f )) dt + tr(M (−it)M ′ (it)ρ(µ, it)(f )) dt,
−∞ −∞

as required.
We now move on to the more problematic case when µ = µ−1 . In this case,
plugging in (2.4) into (2.3) yields (in addition to the terms in (2.5)) the term

XZ ∞ 
e2it(T1 −T2 )
(2.6) E(T ) = (ρ(µ, it)(f )φ, M (it)φ)
−∞ 2it
φ

e−2it(T1 −T2 )
− (ρ(µ, it)(f )φ, M (−it)φ) dt.
2it

(Here we have used the fact that the adjoint of M (s) is M (−s).) But this still
isn’t enough to describe JχT (f ). In this singular situation, what’s missing is the
contribution from the one-dimensional residual spectrum, i.e., the contribution
Z
T,res
Jµ (f ) = µ(f ) µ(x)ΛT (µ(x) dx
Z(A)G(F )\G(A)

from the residual kernel

KµT,res (x, y) = µ(f )µ(x)µ(y).

Analyzing both these functions of T (namely E(T ) and JµT,res (f )) yields finally:
 LECTURE IV. MORE EXPLICIT FORMS OF THE TRACE FORMULA 35

Proposition 2.7. For α(T ) ≫ 0, and χ = {(M, µ)} with µ = µ−1 ,

Z ∞ Z ∞
JχT (f ) = (T1 − T2 ) tr(ρ(µ, it)(f )) dt + tr(M (−it)M ′ (it)ρ(µ, it)(f )) dt,
−∞ −∞
1
+ 4 tr(M (0)ρ(µ, 0)(f )) + µ(f )m(Z(A)G(F )\G(A))

plus n∗1 (T ), a negligible term (for T very large).

Proof. In [GJ, pp. 239–240], a Fourier analysis argument is given to show that
1
E(T ) = 4 tr(M (0)ρ(µ, 0)(f )) + r1 (T )

with r1 (T ) negligible for α(T ) large. On the other hand, it is a pleasant exercise to
check that

JµT,res (f ) −→ µ(f )m(Z(A)G(F )\G(A)) as T → ∞.

Indeed,
Z
µ(f ) µ(x)ΛT (µ(x)) dx
Z(A)G(F )\G(A)
Z  X 
= µ(f ) µ(x) µ(x) − µ(x)τ̂B (δx) − T ) dx
Z(A)G(F )\G(A) δ∈B(F )\G(F )
Z Z X
= µ(f ) dx − µ(f ) τ̂B (H(δx) − T ) dx
δ∈B(F )\G(F )
Z
= µ(f )m(Z(A)G(F )\G(A)) − µ(f ) τ̂B (H(x) − T ) dx
Z(A)B(F )\G(A)
= µ(f )m(Z(A)G(F )\G(A)) − n1 (T )

with
e−2(T1 −T2 )
n1 (T ) = µ(f )m(F ∗ \A1 ) −→ ∞ as α(T ) → ∞.
2
So setting n∗1 (T ) = r1 (T ) + n1 (T ) completes the proof. 
Concluding Remark. (a) Proposition 2.1 expresses JχT (f )
directly as a poly-
nomial in T when χ is regular. But as already suggested, Arthur has proven in
general that JχT (f ) is polynomial in T . Thus it must follow that the “negligible”
term n∗1 (T ) (appearing in the expression for JχT (f ) for singular χ in Proposition 2.7)
is actually zero identically. (Indeed, it is a polynomial in T , which goes to 0 as T
goes to infinity . . . .) On the other hand, the expression E(T ) (or equivalently
(6.28) in [GJ, p. 239] is definitely not a polynomial; rather E(T ) includes a negli-
gible term r1 (T ) which must cancel the negligible (exponentially decreasing) term
n1 (T ) appearing in the expression for JµT,res (f ).
(b) Our computation of
Z
JµT,res (f ) = µ(f ) µ(x)ΛT (µ(x)) dx
Z(A)G(F )\G(A)
36 STEVE GELBART

really amounts to the computation of the inner product of two truncated “degen-
erate Eisenstein series” (induced from µ on P = G . . . ). In general, the explicit
description of JχT,res (f ) with χ “singular” will involve an inner product

(2.8) (ΛT EP (φ′ , λ′ ), ΛT EP (φ, λ))

G(A)
where φ in IndP (A) σ belongs to L2 (P )χ , but P does not belong to Pχ . Thus the
natural generalization of Proposition 2.7 must involve Arthur’s generalization of
/ Pχ , but like the formula for JµT,res (f ) giving only
Langlands’ formula (valid for P ∈
an asymptotic expression for (2.8); see [A9] ).

3. Results for General G. Arthur’s generalization of the above results, and in

fact the GL(2) results themselves, seem much easier to digest and understand when
viewed not as a long chain of unrelated results, but rather as realizations of a few
fundamental phenomena and principles. Of course Arthur knowingly developed his
trace formula with these thoughts in mind. To begin to explain them, let us start
with the symmetry of the formula itself.

Remark 3.1. Consider the formula

X X
JoT (f ) = JχT (f ).
o∈O χ∈X

Although at first sight the indexing sets O and X seem asymmetric and unrelated,
they can actually be viewed as mirror images of one another. Indeed, if o ∈ O,
consider those (standard) parabolics B of G which are minimal with respect to the
property that o meets MB . Then o ∩ MB is a finite union of MB (F ) conjugacy
classes which are “elliptic” in the sense that they meet no proper parabolic subgroup
of MB . Moreover, if W0 denotes the restricted Weyl group of (G, A0 ), then O is
in bijective correspondence with the set of W0 -orbits of pairs (MB , cB ), where
B is a parabolic subgroup of G, and cB is an elliptic conjugacy class in MB (F ).
Thus O indeed corresponds naturally to X, itself defined to be the set of W0 -orbits
of pairs (MB , rB ), where rB is an irreducible cuspidal automorphic representation
of MB (A). This description of O has the additional benefit that it highlights the
analogy between “elliptic” classes and “cuspidal” representations. (For convenience
here, we have ignored centers . . . .)

In light of the symmetry between O and X (expressed by this Remark 3.1), it is

not surprising that most of the general results Arthur proves for the distributions
JoT (f ) have an immediate analogue for JχT (f ). For example, if o = {(MB , cB )} is
such that JoT (f ) is a polynomial of degree at most dim(AB /Z), then so is JχT (f ),
for χ = {(MB , rB )}. Or, if o = {(MB , cB )} is suitably unramified that JoT (f )
is expressible as a weighted orbital integral, then the corresponding JχT (f ) should
be expressible as a weighted average of characters. Let us explain these matters in
some more detail now, and along the way bring into play some other basic principles
as well.
 LECTURE IV. MORE EXPLICIT FORMS OF THE TRACE FORMULA 37

Proposition 3.1 (see Proposition 2.3 of [A4]). For each fixed f , JoT (f ) is a
polynomial in T of degree at most dim(aP0 /aG ) (with P0 the minimal parabolic).
The proof has two main ingredients. First, for any parabolic P ⊃ P0 , and point
X in a0 , a certain characteristic function
X
Γ′P (H, X) = (−1)dim(AR /Z) τPR (H)τ̂R (H − X)
{R : R⊃P }

is defined on a0 , with τPR the characteristic function on a0 of the set

{H ∈ a0 : α(H) > 0, α ∈ ∆R
P}

and τ̂R of the set

ˆ R }.
{H ∈ a0 : ω̃(H) > 0, ω̃ ∈ ∆
(Here ∆ ˆ R is the dual basis in a∗ corresponding to the basis of coroots {α∨ : α ∈ ∆R }
R
in aR .)
For example, for GL(2), and X = (X0 , −X0 ) in a+ ′
0 , ΓB (H, X) = τB (H)− τ̂B (H −
G
X); as a function on the one-dimensional subspace a0 = {(T, −T )} ≈ {T }, this is
just the characteristic function of the interval [0, X0 ].
On the other hand, for G = GL(3) and (X1 , X2 , −(X1 + X2 )) = X in a+ 0 , the
function Γ′P0 ( · , X) (viewed on the two-dimensional space {(r1 , r2 , −(r1 + r2 ))}) is
just the characteristic function of the shaded region below:

Figure 1
In general, the crucial property of Γ′P (H, X) is that it has compact support, and
that its integral Z
Γ′P (H, X) dH
aG
P

is a polynomial in X of degree q = dim(AP /Z). (Here aG P is the subspace of aP

annihilated by a∗G , so in particular, aP = aG
P ⊕ a G .)
M ,T
Second, certain geometric distributions Jo Q (analogous to JoT ) are defined for
each Levi subgroup M of G. Namely, for any fixed suitably regular point T1 in
a+ ∞
0 , class o in O and fQ in Cc (Z(A)\MQ (A)), define

n
X
MQ ,T1 M ,T1
Jo (fQ ) = Joi Q ,
i=1
38 STEVE GELBART

where o ∩ MQ (F ) is the union (possibly empty) of “classes” o1 , . . . on in MQ . (If

M ,T
o ∩MQ (F ) = ∅, then Jo Q 1 ≡ 0; on the other hand, if M = G, then JoM,T1 = JoT1 .)
To prove Proposition 3.1, Arthur derives the formula
X MQ ,T1 Z
(3.2) JoT (f ) = Jo (fQ ) Γ′Q (H, T − T1 ) dH,
Q⊃P0 aG
Q

where fQ in Cc∞ (Z(A)\MQ (A))Ris explicitly determined by f (and T varies freely

through T1 + a+ 0 ). Since each Γ′Q (H, T − T1 ) dH is a polynomial in T − T1 of
degree at most dim(AP0 /Z), it follows that the same is true of JoT (f ), as claimed.
Note that JoT (f ) will be independent of T , i.e., be of degree zero, precisely when o
is elliptic, i.e., has empty intersection with every proper parabolic Q appearing in
(3.2) above.
Remark 3.3. As expected, there is an analogue of Proposition 3.1 for JχT (f ),
proved in just the same way (in keeping with the principle that the geometric and
spectral terms are different facets of the same type of object). For χ in X, and M a
Levi subgroup of G, one similarly defines distributions JχM,T on M (A) and proves
(analogously) that
X Z
(3.4) JχT (f ) = JχMQ ,T1 Γ′Q (H, T − T1 ) dH
Q⊃P0 aG
Q

is (again) a polynomial in T of degree ≤ dim(A0 /Z). Interestingly, one also obtains

the equality
X X
(3.5) JoM,T (f ) = JχM,T (f ),
0∈O χ∈X

which may be viewed as (a first form of) Arthur’s trace formula “relativized” to an
arbitrary Levi subgroup MP of G! This relativization is reminiscent of Langlands’
use of the spectral decomposition of L2 (P ) (P any parabolic of G) in order to
describe L2 (G) itself, and it constitutes another of the unifying themes in Arthur’s
work.
To turn next to explicit formulas for JoT (f ) (and JχT (f )), we need first to recall
Arthur’s notion of an unramified class of datum. For χ = {(MB , rB )} this just
means that for any pair (MB , rB ) in χ, the only element of Ω(aP , aP ) fixing rB is
the trivial element. On the other hand, for o = {(MB , cB )}, the notion is a little
less obvious.
Namely, fix a class o, and choose a parabolic subgroup P and a semisimple
element γ in o such that γ belongs to M (F P) but not to the Levi subgroup of
parabolic properly contained in P . Then let (γ) denote the (possibly empty) set
P P
of roots α in = (P, AP ) such that the centralizer of γ in yα (the root subspace
of the Lie algebra of NP belonging to α) is not zero, and let A′ be the intersection
of the kernels of these roots α (regarded as characters of A). Now assume that
every element in o is semi-simple, and choose a parabolic subgroup P1 , and an
element w ∈ Ω(a, a1 ) such that AP1 = wA′ w−1 . Setting γ1 equal to wγw−1 , we call
o unramified if the centralizer G(γ1 ) is contained in M1 .
 LECTURE IV. MORE EXPLICIT FORMS OF THE TRACE FORMULA 39

Note that  
  α 0 0
α 0
γ= or γ =  0 β 0 
0 1
0 0 1
determines a semi-simple class in GL(2) (or GL(3)) which is unramified, with P
the Borel subgroup, A′ = AP , P1 = P , and γ1 = γ. On the other hand, any class
in GL(3) containing a non-semisimple element, like say
 
α 0 0
γ =  0 1 x,
0 0 1
is automatically ramified, and its corresponding distribution JoT (f ) will not be of
the nice type described in the proposition below.
Proposition 3.6. Suppose o is an unramified class in O, and γ1 is a (semi-
simple) element in o. Then for sufficiently large T ,
Z
JoT (f ) = m(Gγ1 (F )\G1γ1 (A)) f (x−1 γ1 x)v(x, T ) dx,
Gγ (A)\G(A)
with v(x, T ) (the weight function) equal to the volume of the convex hull of the
points [
{w−1 T − w−1 HP0 (wx) : w ∈ Ω(a1 , a2 )}
P2
projected onto a1 /aG.
Let us describe v(x, T ) graphically for G = GL(2) or GL(3), with γ1 a diagonal
element (with distinct eigenvalues). We shall also sketch the proof of the proposition
for GL(2), making clear how the volume v(x, T ) naturally arises.
N.B. We should really write vo (x, T ) for v(x, T ), since this function indeed depends
on the nature of o.
For GL(2) or GL(3), and γ1 as just specified, v(x, T ) is precisely the volume of
the convex hull of the projection of
{w−1 T − w−1 HP0 (wx) : w ∈ Ω(aP0 , aP0 )}
onto aP0 /aG . For GL(2), this is just the length of the line segment determined
by the points (T1 , T2 ) − H(x) and (T2 , T1 ) − wH(wx) in aP0 projected onto the
one-dimensional space aP0 /aG, i.e.,
v(x, T ) = 2(T1 − T2 ) − (α(H(x)) + α(H(wx))).
Thus Proposition 3.5 indeed reduces to Proposition 1.1.
On the other hand, for GL(3), v(x, T ) (at least for T = 0) reduces to the area
of the convex set
H pw(x)

H p (x)
40 STEVE GELBART

Figure 2

whose extreme points are indexed by the Weyl group elements {w1 = e, w2 , . . . , w6 }.
Note here that
w−1 HP0 (wx) = HPw (x),

where Pw denotes the (not-necessarily standard) parabolic subgroup w−1 P0 w, and

HPw (x) is defined analogously to HP0 using the decomposition G = Pw K. (Indeed
wx = nm0 k implies w−1 HP0 (wx) = w−1 HP0 (m0 ), and x = w−1 nm0 k implies
HPw (x) = HP0 (w−1 mw) = w−1 HP0 (m0 ), i.e., HPw (x) = w−1 HP0 (wx).) Thus we
could have alternately described v(x) = v(x, 0) in terms of the points

{HP (x) : P ∈ P(M0 )},

where P(M0 ) denotes the set of parabolics (not necessarily standard) in G whose
Levi part equals M0 .

Remark 3.7. Weighted orbital integrals of the above type were first systemat-
ically studied by Arthur in the context of real groups. In [A8], Arthur studied
general properties of these real weighted orbital integrals, motivated by a sugges-
tion of Langlands that such integrals would arise in any projected general treatment
of the trace formula. In particular,
R for matrix coefficients of discrete series repre-
sentations, Arthur related f (x−1 γx)v(x) dx to the characters of the discrete series
(whence the title of [A8]). On the other hand, Arthur’s weight functions v(x) were
also interpreted in [A8] in terms of what would eventually be known as (G, M)
families. We shall return to this point in a later lecture, where (G, M ) families
are discussed in earnest.

On the Proof of Proposition 3.6 for GL(2). Precisely because o is unramified

(in this case, a hyperbolic class with semi-simple element γ1 ), the centralizer G(γ1 )
is contained in M (in this case equals M ) and we can express the modified kernel
as

koT (x, T )
X
= f (x−1 δ −1 γ1 δx)
δ∈M(F )\G(F )
X Z
− f (x−1 δ −1 γ1 nδx) dn τ̂B (H(δx) − T )
δ∈B(F )\G(F ) N (A)

X  X
= f (x−1 δ −1 η −1 γ1 ηδx)

B(F )\G(F ) η∈N (F )
Z )
−1 −1
− f (x δ γ1 nδx) dn τ̂B (H(δx) − T ) .
N (A)
 LECTURE IV. MORE EXPLICIT FORMS OF THE TRACE FORMULA 41

Thus

JoT (x, f )
Z
= koT (x, T ) dx
Z(A)G(F )\G(A)

Z  X
= f (x−1 η −1 γ1 nx)
Z(A)B(F )\G(A)  η∈N (F )
Z )
−1 −1
− f (x n γ1 nx) dn τ̂B (H(x) − T ) dx
N (A)

(here we made the change of variables n −→ (γ1−1 n−1 γ1 )n in the second integral)
Z
= f (x−1 γ1 x)(1 − τ̂B (H(x) − T )) dx,
Z(A)M(F )G(A)

(here we used the decomposition

Z(A)B(F )\G(A) ≈ N (A)Z(A)M (F )\G(A) · N (F )\N (A),

carrying out first the integration over N (F )\N (A)). But now observe the following:
because we know koT (x, T ) is absolutely integrable, we also know that the last
integral Z
A= f (x−1 γ1 x)(1 − τ̂B (H(x) − T )) dx
Z(A)M(F )\G(A)

converges absolutely, and equals (from the change of variables x → wx)

Z
B= f (x−1 γ1 x)(1 − τ̂B (H(wx) − T )) dx,

i.e.,

JoT (x, f )
= A = 12 (A + B)
Z
1
= 2 f (x−1 γ1 x)(1 − τ̂B (H(x) − T ) − τ̂B (H(wx) − T )) dx
Z(A)M(F )\G(A)
Z
= f (x−1 γ1 x)v ∗ (x, T ) dx,
M(A)\G(A)

with
Z
∗
v (x, T ) = (1 − τ̂B (H(mx) − T )) − τ̂B (H(wmx) − T )) dm,
Z(A)M(F )\G(A)

and it remains only to prove:

42 STEVE GELBART

Claim. Set vT∗ (x) = 1 − τ̂B (H(x) − T ) − τ̂B (H(wx) − T ). Then

Z
v ∗ (x, T ) = vT∗ (mx) dm = Arthur’s weight factor v(x, T ).
Z(A)M(F )\M(A)

Proof. Note that vT∗ (x) is identically zero unless

(a) Both τ̂B (H(x) − T ) and τ̂B (H(wx) − T ) are 1 (in which case vT∗ (x) = −1);
or
(b) Both are zero (in which case vT∗ (x) = 1). But the first possibility for
vT∗ (mx) 6= 0 implies that
α(H(mx)) > α(T ) and α(H(wmx)) > α(T ),
which (for α(T ) sufficiently large . . . ) implies (by Lemma 4.1 of Lecture
II) that w ∈ B(F ), an obvious contradiction. On the other hand, it is
straightforward to check that the second possibility for vT∗ (mx) 6= 0 implies
α(H(mx)) − (T1 − T2 ) < α(H(m)) < (T1 − T2 ) − α(H(x)).
Thus the decomposition
    t 
a 0 e 0
Z(A)M (F )\M (A) = : a ∈ F x \A1 ·
0 1 0 1
Z Z Z (T1 −T2 )−α(H(x))
vT∗ (mx) dx = dx a dt
F x \A1 α(H(wx))−(T1 −T2 )

= m(F \A1 )v(x, T ).

x


Remark 3.8. The analogue of Proposition 3.6 for a spectral unramified dis-
tribution JχT (f ) is also a “straightforward” generalization of the GL(2) situation
(Proposition 2.1), namely, we have:
Proposition 3.9. For χ = {(M, r)} unramified,
X Z
1
JχT (f ) = tr(MPT (σ)χ ρ(σ, λ)(f )) dλ,
n(Ap ) ia∗P /ia∗G
P ∈Pχ

with MPT (σ)χ the operator on IndP (σ, λ) defined by

X X ewζ(T ) M (w, 0, σ)−1 M (w, ζ, σ)
cP lim Y
ζ→0
P2 ∈Pχ w∈Ω(aP ,aP ) (wζ)(α∨ )
1 2
α∈∆P2

and M (w, ζ, σ) the operator intertwining IndG

P1 σ| |λ+ζ with IndG w
P2 σ | |w(λ+ζ) .
As suggested in §2, the crucial point of the proof is Langlands’ formula for the
inner product of truncated Eisenstein series, and since this is valid for general
G (for Eisenstein series EP with P ∈ Pχ ), the proof of Proposition 2.1 generalizes
directly. The interesting point here is that the integrals appearing in the description
of JχT (f ) above really should be viewed as weighted character sums. In fact, using
the notion of (G, M ) families, Arthur eventually views both MPT (σ) and the weight
functions v(x) as special examples of functions constructed out of similar (G, M )
families; we shall return to this point in Lecture VII.
 LECTURE IV. MORE EXPLICIT FORMS OF THE TRACE FORMULA 43

(New) References
[A8] Arthur, J., The characters of discrete series as orbital integrals, Invent. Math. 32 (1976),
205–261.
[A9] , On the inner product of truncated Eisenstein series, Duke. Math. J. 49 (1982),
no. 1, 35–70.
44 STEVE GELBART

Lecture V. Simple Forms of the Trace Formula

In this lecture and the next, we explain (at least for GL(2)) how the trace formula
can be molded into a simpler form, and then used to obtain dramatic results in
representation theory, number theory, and the theory of automorphic forms. In
addition to providing a welcome respite from the general theory, this detour will
provide impetus and direction for the further discussion of Arthur’s development
of a general trace formula.
Roughly speaking, the trace formula takes on a simpler form when more re-
strictive hypotheses are put on the test functions f . The more restrictive the
hypotheses, the easier it is to establish the corresponding “simple trace formula”;
but the less restrictive the hypotheses, the more effective it is to apply the resulting
trace formula to the theory of automorphic forms.
For example, the “very simple” trace formula of Deligne-Kazhdan (described
in §3 below) has nice applications to local representation theory, but cannot give
complete results on the functorial lifting of automorphic forms, and cannot give
applications to the computation of Tamagawa numbers. For such applications one
seems to need the “simple trace formula of Arthur”, described in §2 below (in the
context of GL(2)).
We start by showing (in §1) that each term on the right side of the formula
X X
tr(R0 (f )) = JoT (f ) − JχT (f )
o χ∈X(G)
/

may be expressed as a finite sum of products of local distributions on each Gv .

Q
Then we show (in §2) that for specially chosen f = v fv , sufficiently many of
these local distributions vanish to ensure that the resulting trace formula reduces
to the simple formula
X
tr Rd (f ) = m(Z(A)G(F )\G(A))f (1) + Jo (f )
o
elliptic

(as in the case of compact quotient).

Henceforth, we assume Y
f= fv
v

is such that for all v, fv ∈ Cc∞ (Zv \Gv ), and for almost all v, fv is the characteristic
function of Zv Kv .
 LECTURE V. SIMPLE FORMS OF THE TRACE FORMULA 45

1. Factorization into Local Distributions. We start with the geometric terms

Jo (f ). Fix f = Πfv , and let Sf denote the finite set of places of F outside of which
fv is the characteristic function of Zv Kv .
Proposition 1.1. Suppose γ is hyperbolic and belongs to o. Then Jo (f ) (the
constant term of the polynomial JoT (f )) is expressible as the sum (over v in Sf ) of
the products
Y Z Z
c fw (g −1 γg) dg fv (g −1 γg)vv (g) dg.
w6=v Mw \Gw Mv \Gv

Here vv (g) = α(Hv (g) + Hv (wg)) is the local weight function.

Proof. It is easy to check that for g in G(A),
X
v(g) = vv (gv ),
v

with vv (gv ) ≡ 0 on Zv Kv . Thus, Proposition 1.1 of Lecture IV immediately implies

XZ
Jo (f ) = c f (g −1 γg)vv (g) dg
v M(A)\G(A)
X Y Z Z
−1
= c fw (g γg) dg fv (g −1 γg)vv (g) dg.
v w6=v Mw \Gw Mv \Gv

But if v ∈
/ Sf , the second integrand is identically zero. Thus the proposition is
clear. 
The factorization of the unipotent term is a bit more complicated. Recall from
Proposition 1.2 of Lecture IV that the constant term of the (non-trivial part of the)
unipotent term JoT (f ) is computed by subtracting off from
Z Z  
1a
ζ(F, s) = f (k −1 0 1
k) dk|a|s dx a
Zx K

its principal part at s = 1, and then setting s = 1.

So first rewrite this last integral as L(s, 1F )θ(s), where
ZZ  
1 1a
θ(s) = f (k −1 0 1
k) dk|a|s dx a.
L(s, 1F )

By Tate’s theory, θ(s) is holomorphic at s = 1, whereas

λ−1
L(s, 1F ) = + λ0 + · · · .
s−1

Therefore, the constant term Jo (f ) we need to compute is λ−1 θ′ (1) + λ0 θ(1), and
arguing as in [GJ, pp. 242–243], we get:
46 STEVE GELBART

Proposition 1.2. With Sf as before, and o unipotent,

Y Z     
1
Jo (f ) = λ0 fv g −1 10 11 g dg
v
L(1, 1v ) Zv Nv \Gv
X Y Z     
1 d
+ λ−1 g −1 10 11 g dgv ×
L(1, 1v ) Zv Nv \Gv ds s=1
u∈Sf v6=u
ZZ    
1
fu k −1 10 a1 k dk |a|s dx a
L(s, 1u )
+ m(Z(A)G(F )\G(A)f (1).

Remarks. (1) Each of the local unipotent orbital integrals

Z  
fv (g −1 10 11 g) dg = ζ(Fv , 1)
Zv Nv \Gv

converges, and for almost every v equals (1 − Nv−1 )−1 . It is the appearance of
the “convergence factors” L(1, 1v )−1 which makes possible the convergence of the
product of these local integrals in Proposition 1.2. (The divergence of the global
unipotent orbital integral
Z     YZ    
f g −1 10 11 g dg = fv g −1 10 11 g dg
v

was of course the reason for introducing the modified JoT (f ) in the first place . . . ).
(2) If u ∈
/ Sf , then
ZZ    
1 1
fu k −1 10 a1 k dk|a|s dx a = · ζ(1Ov , s) = 1.
L(s, 1u ) L(s, 1u )

and its derivative is zero; that’s why we need only sum over u ∈ / SF in Proposi-
tion 1.2.
We turn now to the spectral contributions. Here we need to recall some crucial
facts about normalized intertwining operators.
Recall the operator
Z
(M (s)φ)(g) = φ(wng) dn (Re(s) > 21 )
N (A)

intertwining ρ(µ, s)with ρ(µw , −s). This is the operator—or rather its analytic
continuation to iR—which appears in our description of the spectral “constant
term” Jχ (f )(χ = (M, µ)). The problem is that in analytically continuing M (s) we
lose its Euler product factorization (just as we lose the Euler product for ζ(s) in
analytically continuing it to the left of Re(s) = 1). It is to restore this factorization
that one needs the normalized intertwining
Q operators R(s) recalled below.
For a fixed additive character ψ = v ψv of F \A, and v any place of F , set

L(s, µ2v )
mv (s, µv , ψv ) = .
L(s + 1, µ2v )ε(s, µ2v , ψv )
 LECTURE V. SIMPLE FORMS OF THE TRACE FORMULA 47

Then define local normalized intertwining operators

R(s, µv ) : ρ(s, µv ) −→ ρ(−s, µw

v)

by the equality
M (s, µv ) = m(s, µv , ψv )R(s, µv )
R
(where M (s, µv )φ is just defined by the local integral Nv φ(wng) dn, convergent
for Re(s) ≫ 0). The crucial property of this normalized operator R(s, µv ) is that
it takes the (normalized) Kv -fixed vector φ0v in ρ(s, µv ) (for µv unramified) to the
unique normalized Kv -fixed vector in ρ(−s, µ−1 v ); moreover, as an operator valued
function
Q of s, R(s, µv ) is holomorphic in Re(s) > − 21 . Thus we can define, for
φ = φv in ρ(s, µ), Y
R(s, µ)φ = R(sv , µv )φv .
v

Since R(s, µv )(φv ) = φ0v

almost everywhere, this product makes perfect sense; more-
over, R(s, µ) is holomorphic in Re(s) > − 12 . But clearly (at least for Re(s) ≫ 0),

M (s, µ) = m(s, µ)R(s, µ),

where Y
m(s, µ) = mv (s, µv , ψ)v).
v

So since m(s, µ) is meromorphic, we indeed obtain the analytic continuation of

M (s, µ) to iR with its “Euler product” factorization intact.
N.B. By the (global) functional equation of Hecke L-functions,

L(s, µ2 ) L(1 − s, µ−2 )

m(s, µ) = =
L(s + 1, µ2 )ε(s, µ2 , ψ) L(s + 1, µ2 )

Proposition 1.3. For χ “unramified”, the (constant term of the) spectral contri-
bution Jχ (f ) equals
Z ∞ ′
m (it, µ)
tr(ρ(it, µ)(f )) dt
−∞ m(it, µ)
X ∞ YZ
+ ( tr(ρ(it, µv )(fv )) tr(Ru (it, µu )−1 Ru′ (it, µu )ρ(µ, it)(fu )) dt.
u∈Sf −∞ v6=u

In case χ is “ramified” (µ2 ≡ 1), there are the additional terms

(1.4) tr(M (0)ρ(µ, 0)(f ))

and
µ(f )m(Z(A)G(F )\G(A)).

Proof. Straightforward (see [GJ, p. 243]). 

48 STEVE GELBART

2. Simplifications of the Formula. A first kind of simplification comes form

making the relatively mild assumption on
Y
f= fv

that its local hyperbolic orbital integrals

Z    
(*) f g −1 a0v 10 g d∗ g ≡ 0
Mv \Gv

for at least two place v = v1 , v2 .

Proposition 2.1. With f satisfying the above assumption (*), we have
X X
tr R0 (f ) = m(Z(A)\G(A))f (1) + Jo (f ) − µ(f )m(Z(A)G(F )\G(A))
o elliptic µ2 =1

Proof. First consider the spectral contributions Jχ (f ), with χ = {(M, µ)}. If

v1 is a place where (*) holds, then tr ρ(it, µv1 )(fv ) = 0 (by a well known com-
putation of the trace of induced representations). Thus the first term appear-
ing in Proposition 1.3 immediately disappears. But if (*) also holds at a second
place v2 , then it is clear that the second term there must also always vanish. On
the other hand, if µ2 = 1, then M (0, µ) intertwines the irreducible representa-
tion ρ(0, µ) with itself; hence M (0, µ) must be a scalar operator. In particular,
tr(M (0)ρ(0, µ)(f )) = λ tr(ρ(0, µ)(f )) = 0 by our assumption on f , and we conclude
P
from Proposition 1.3 that the full spectral contribution χ = (M, µ)Jχ (f ) reduces
P
to ( µ2 =1 µ(f ))m((Z(A)G(F )\G(A)).
Now we consider the possible geometric contributions Jo (f ), following Propo-
sitions 1.1 and 1.2. For the hyperbolic contributions, it is again obvious that if
assumption (*) holds at two places, then each Jo (f ) is zero (just as for the unram-
ified spectral forms). On the other hand, for the unipotent contribution, it suffices
to make the following observation:
For any local fv in Cc∞ (Zv \Gv ),
Z     Z    
fv g −1 10 11 g dg = lim |1 − a−1 | fv g −1 a0 10 g dg.
Zv Nv \Gv a→1 Mv \Gv

This is a simple yet important identity, which we encourage the reader to verify
(using Iwasawa’s decomposition and an appropriate change of variables in N ). It
implies that the local unipotent orbital integral of fv vanishes as soon as its hyper-
bolic orbital integrals vanish. In particular, as soon as (*) holds for two places, it
follows that all of the unipotent term Jo (f ) vanishes, except for the term involving
f (1). 
In the next lecture, we shall explain how this simplification of the trace for-
mula is used to prove the Jacquet–Langlands correspondence between automorphic
representations between a quaternion algebra and GL(2), and to show that the
corresponding Tamagawa numbers are equal.
 LECTURE V. SIMPLE FORMS OF THE TRACE FORMULA 49

3. The “Simple” Trace Formula. In the early 1980’s, Deligne and Kazhdan
introduced a remarkably simple trick into the trace formula repertoire, the so-called
“simple trace formula” (see [BDKV]). As we shall see in the next lecture, it yields
powerful local results on functorial lifting, with surprisingly little work.
Theorem 3.1 (The Simple Trace Formula of Deligne-Kazhdan). Suppose f =
Q ∞
v fv in Cc (Z(A)\G(A)) satisfies the following two properties:
(i) At one place v = v1 , fv1 is the matrix coefficient of a supercuspidal repre-
sentation of Gv1 ; and
(ii) at a second place v = v2 , fv2 is supported on the set of regular elliptic
elements of Gv2 .
Then
(a) R(f ) has its image in L20 (Z(F )G(F )\G(A)), hence is of trace class, with
X
tr R(f ) = tr π(f ); and
π cuspidal

(b) we also have

X Z
tr R(f ) = m(Z(A)Gγ (F )\Gγ (A) f (x−1 γx) dx
Gγ \G(A)
{γ}
regular
elliptic

Remarks. (1) The condition we really need in (i) is that fv1 is a supercusp form
in the sense of Harish-Chandra, i.e., (like the matrix coefficient of any supercuspidal
representation),
Z
fv1 (gnh) dn = 0 for all g, h, in Gv .
N v1

(2) Both the theorem, and its proof below, hold more generally for any G; see [Ro1],
§1, for example.
(3) This simple trace formula is indeed less subtle than the simple form of Arthur’s
trace formula discussed in the last section. In particular, note that the distribution
f → m(Z(F )G(F )\G(A))f (1) disappears entirely from this simpler form of the
trace formula (and hence, for example, no application to Tamagawa numbers is
possible . . . ). It is just this crudeness, however, that makes the formula much
easier to prove!
Proof. For (a), it suffices to show that R(f ) has image in L20 (Z(F )G(F )\G(A) for
then the theorem of Gelfand and Piatetski-Shapiro (see [GPS] and [Go2]) implies
R(f ) is of trace class). To check that R(f )φ is a cusp form for any φ we simply
compute:
Z Z Z X 
R(f )φ(nx) dx = f (x−1 nvg)φ(g) dg dn
N (F )\N (A) N (F )\N (A) Z(A)\G(A) v∈N (F )
Z Z 
= f (x−1 ng) dn φ(g) dg = 0
Z(A)N (F )\G(A) N (A)
50 STEVE GELBART

since Z YZ
f (x−1 ng) dn = fv (x−1
v ngv ) dn = 0
N (A) v Nv

by our assumption (i) on fv1 .

To prove (b), note that a γ in G(F ) is regular elliptic as soon as it is regular
elliptic in Gv = G(Fv ) for some place v. But f (x−1 γx) 6= 0 implies {supp(f )} ∩
{Orbit of γ} 6= 0, which in turn implies {supp(fv2 )} ∩ {Orbit of γ} 6= 0. So by
hypothesis (ii), γ is regular elliptic in Gv2 , i.e.,
X
(3.2) Kf (x, x) = f (x−1 γx),
{γ}
regular
elliptic

and conclusion (b) follows. 

Concluding Remark. Implicit in the proof above is the fact that Kf (x, x) is an
absolutely integrable smooth function on Z(A)G(F )\G(A), and hence tr(R(f ))
(= tr(R0 (f ))) is given by its integral. In fact (3.2) implies Kf (x, x) is compactly
supported on Z(A)G(F )\G(A) (as we saw in Lecture II). Interestingly enough, if
we drop the hypothesis (ii), so that (3.2) need no longer hold, it still follows (from
(i) above) that RKf (x, x) is absolutely integrable (in fact, rapidly decreasing), and
that tr R(f ) = Kf (x, x) dx; for a proof, see [Ro1, §1].

(New) References
[BDKV] Bernstein, J., Deligne, P., Kazhdan, D, and Vigneras, M.-F., Representations des
groupes sur un corps local, Hermann, Paris, 1984.
[Clo] Clozel, L., Invariant harmonic analysis on the Schwartz space of a reductive p-adic
group, in Harmonic Analysis on Reductive Groups, edited by W. Barker and P. Sally,
Progress in Mathematics, vol. 101, Birkhäuser, Boston, 1991, pp. 101–121.
[GPS] Gelfand, I. M., and Piatetski-Shapiro, I. I., Automorphic functions and representation
theory, Trudy Moskov. Mat. Obš č. 12 (1963), 389–412; Trans. Moscow Math. Soc. 12
(1963), 438–464.
[Go2] Godement, R., The Spectral Decomposition of Cusp Forms, Proc. Sympos. Pure Math.,
vol. IX, American Mathematical Society, Providence, RI, 1966, pp. 225–234.
[Ro1] Rogawski, J., Representations of GL(n) and Division Algebras on a p-adic Field, Duke.
Math. J. 50 (1983), no. 1, 161–196.
 LECTURE VI. APPLICATIONS OF THE TRACE FORMULA 51

Lecture VI. Applications of the Trace Formula

Let G = GL(2), and let G′ be the multiplicative group of a division quaternion

algebra D over F .
In this lecture, we shall explain how the simple forms of the trace formula are
used to prove the following three results:
(1) The Tamagawa number of G′ equals the Tamagawa number of G;
(2) There is a bijection (the “Jacquet–Langlands correspondence”) between the
automorphic π ′ on G′ (which are not one-dimensional) and the automorphic
cuspidal π of G (such that πv is square integrable for each place v of F
ramified in D); and
(3) Given any finite set S of finite places v of F , and square-integrable πv on
Gv for v ∈ S, there exists a cuspidal π on G such that (π)v ∼ = πv for all
v ∈ S.

1. Tamagawa Numbers. At the end of the 1950’s, Weil conjectured that the
Tamagawa number of any simply connected semi-simple group G equals one. At
the end of the 1960’s Jacquet–Langlands proposed a two step program for proving
this:
(1) Prove first that
τ (G0 ) = 1
for G0 the quasi-split inner form of G, using the Eisenstein series method
introduced by Langlands for split groups (see [La3] and [Lai]); then
(2) Use the trace formula to prove that

τ (G) = τ (G0 ).

This second step was first carried out in §16 of [JL] (for GL2 in place of
SL2 ), and then recently generalized to arbitrary G by Kottwitz to prove
Weil’s conjecture in general (see [Kot]).
We should stress that—in deriving a simple trace formula for an arbitrary semi-
simple quasi-split G—Kottwitz had to appeal not only to all of Arthur’s work on
the trace formula through 1988, but also to his own earlier works on the stable
form of the trace formula. By specializing Kottwitz’s argument below to the case
of GL(2) (in place of SL(2)), we manage to avoid both these bodies of work.
′
We start by recalling the notion of
QTamagawa measures for our G and G .
Fix a non-trivial character ψ = ψv of F \A. For each place v of F , ψv de-
termines a Fourier transform on Av = M2 (Fv ) or Dv , hence also a self-dual Haar
52 STEVE GELBART

measure dy on Av . The measure dy then determines a Haar measure dy/kyk on

Axv , where kyk is the module homomorphism from Av to Rx+ . Hence the choice of
ψ simultaneously determines Haar measures on each Gv and G′v ; similarly, Haar
measures are determined on Fv = Zv ≃ Zv′ . Such measures are called Tamagawa
measures, locally and globally. The resulting measure of Z(A)G(F )\G(A) is inde-
pendent of ψ and called the Tamagawa number τ (G) of G, i.e.,

τ (G) = m(Z(A)G(F )\G(A)).

Similarly,
τ (G′ ) = m(Z ′ (A)G′ (F )\G′ (A)).

The center Z ′ may of course be identified with the center Z ≃ F x of GL(2).

Theorem 1.1. The Tamagawa numbers of G and G′ are equal.

We start the proof with a definition.

Suppose v in F is ramified in D, i.e., Dv = D ⊗ Fv is a division quaternion
algebra over Fv . Then each quadratic extension Lv of Fv can be regarded simulta-
neously as a Cartan subgroup Tv′ = TL′ v of G′v and as an (elliptic) Cartan subgroup
Tv = TLv of Gv = GL2 (Fv ).

Definition 1.2. We say a function fv in Cc∞ (Zv \Gv ) matches fv′ in Cc∞ (Zv′ \G′v )
(and write fv ∼ fv′ ) if:
(i) fv (1) = fv′ (1);
(ii) the regular hyperbolic orbital integrals
Z  
a0
fv (g −1 0 1
g) dg
Mv \Gv

vanish identically; and

(iii) for corresponding tori Tv and Tv′ and t ∼ t′ ,
Z Z
fv (g −1 tg) dg = fv′ (g −1 t′ g) d′ g.
Tv \Gv Tv′ \G′v

N.B. The Haar measures on Gv and G′v are (simultaneously) normalized by way
of the Tamagawa measures recalled above; the measures on Tv and Tv′ are fixed by
a choice of Haar measure on Lv .

Fact 1 (Local Harmonic Analysis on G). Given any fv′ in Cc∞ (Zv′ \G′v ), there exist
(infinitely many . . . ) fv in Cc∞ (Zv \Gv ) which match fv′ in the above sense.
This crucial fact can be proved using a characterization of the orbital integrals
on GL2 à la “Shalika germs”; see §6 of [La2]

Using Fact 1, we can relate Tamagawa numbers to traces, and then prove The-
orem 1.1.
 LECTURE VI. APPLICATIONS OF THE TRACE FORMULA 53

Proposition
Q ′ 1.3. Let SD denote the set Q of v in ∞F ramified in D. Given f ′ =
∞ ′ ′
v fv in Cc (Z (A)\G (A)), suppose f = fv in Cc (Z(A)\G(A)) is such that for
all v in SD , fv ∼ fv′ (in the sense of Definition 1.2 ), and for all v ∈
/ SD , fv = fv′
′
(via the natural isomorphism of Gv and Gv ). (In this case, we say f “globally
matches” f ′ , and write f ∼ f ′ ). Then
 X 
(1.4) tr R0 (f ) − tr R0′ (f ′ ) = {τ (g) − τ (G′ )} f (1) + µ(f ) ,
µ2 =1

with R0′ the representation of G′ (A) in the subspace of L2 (Z(A)G′ (F )\G′ (A)) or-
thogonal to all one-dimensional invariant subspaces.

Assuming this proposition (for the moment), we give the:

Proof of Theorem 1.1. If we rewrite (1.4) as

X
(1.5) tr R0 (f ) − tr R0′ (f ′ ) − {τ (G) − τ (G′ )} µ(f ) = {τ (G) − τ (G′ )}(f (1),
µ2 =1

then it clearly suffices to prove that both sides of this equation—valid for all f ∼ f ′
as above—equal zero. For this, we follow an argument first explicitly introduced
by Langlands in [La2].
Fix a place v0 outside SD , and consider the Hecke algebra Hv0 (Gv0 , Kv0 ). If we
fix all components of
Y
f′ = fv′
v

except the v0 -component, which we let vary through Hv0 , then (1.5) reads
X X
(1.6) cj tr(πj )v0 (fv0 ) − cµ µv0 (fv0 ) = c0 {τ (G) − τ (G′ )}fv0 (1),
j µ2 =1

where the πjv0 (and µv0 ) are unramified unitary representations of Gv0 . But the
distribution fv0 → fv0 (1) is given in terms of tr πv0 (fv0 ) (for all unramified tem-
pered πv0 ) by integration against Plancherel’s measure, which is continuous in the
obvious sense. Similarly, the left-hand side of (1.6) defines a discrete measure (on
this same unramified dual). Thus it follows from (the uniqueness part of) the Riesz
representation theorem for measures that both sides of (1.6) must be zero. 

Corollary (of Proposition 1.3 and the Proof of Theorem 1.1). For match-
ing f and f ′ on G(A) and G′ (A),

tr R0 (f ) = tr R0′ (f ′ ),

where R0′ is the representation of G′ (A) in the subspace of L2 (Z(A)G′ (F )\G′ (A))
orthogonal to all one-dimensional invariant subspaces.

It remains now to complete the:

54 STEVE GELBART

Proof of Proposition 1.3. Because |SD | ≥ 2, our assumption on f (that it matches

f ′ ) implies that fv has vanishing (regular) hyperbolic orbital integrals for at least
two places v1 and v2 . Thus by Proposition 2.1 of the last lecture (Lecture V)

(1.7) tr R0 (f ) = m(Z(A)G(F )\G(A))f (1)

X Z
+ m(Z(A)Gγ (F )\Gγ (A) f (x−1 γx) dx
Gγ (A)\Gγ (A)
γ regular
elliptic
X
− µ(f )m(Z(A)G(F )\G(A)).
µ2 =1

On the other hand, for G′ , the trace formula for compact quotient in Lecture I
implies

tr R0′ (f ′ ) = m(Z(A)G′ (F )\G(A))f ′ (1)

X Z
+ m(Z(A)G′γ (F )\G′γ (A) f ′ (x−1 γx) dx
{γ}inZ ′ (F )\G′ (F ) G′γ (A)\G′ (A)
X
− µ(f ′ )m(Z(A)G′ (F )\G′ (A))
µ2 =1

So to prove the proposition, it remains to check that: (i) the regular elliptic orbital
integral terms match up on G and G′ ; and (ii) µ(f ) = µ(f ′ ) for each character µ.
As for (i), let {L} run through a set of representatives for the classes of quadratic
extensions L of F which don’t split at any v ∈ SD . Then each integral
Z
m(Z(A)Gγ (F )\G(A)) f (x−1 γx) dx
Gγ (A)\G(A)

in (1.7) is of the form

Z
(*) m(Ax Lx \Lx (A)) f (g −1 γg) dg
Lx (A)\G(A)

for some γ 6= 1 in F x \Lx . (Note that if γ belongs to a quadratic extension L which

splits at some v ∈ SD , then our hypothesis on f implies
Z
f (gv−1 γgv ) dgv = 0,
TLv \Gv

i.e., the orbital integrals for such γ don’t appear in the expression for tr R0 (f ).) On
the other hand, each regular orbital integral
Z
′ ′
m(Z(A)Gγ (A)\Gγ (A)) f ′ (x−1 γx) dx
G′γ (A)\G′ (A)

appearing in the expression for tr R0′ (f ′ ) is also of the form (*), since quadratic L
which don’t split at v ∈ SD are precisely the quadratic extensions of F embeddable
in D. Thus the regular elliptic orbital integrals indeed match up term by term.
 LECTURE VI. APPLICATIONS OF THE TRACE FORMULA 55

As for (ii), let us verify that

µv (fv ) = µv (fv′ )

for each place v of F . For v ∈

/ SD this is a tautology. So let us assume v ∈ SD . By
Weyl’s integration formula for Zv′ \G′v we compute
Z
µv (fv′ ) ∼
= fv′ (g)µv (det g) dg
Zv \G′v
X Z Z 
= 1
2 δ(t)µ(det t) fv′ (g −1 tg) dg dtv
{Lv } Zv \TL
′
v
′ \G′
TL v v

where {Lv } indexes the Cartan subgroup TL′ v , and

(a1 − a2 )2
δ(t) =
a1 a2

if t has eigenvalues a1 and a2 . But because fv matches fv′ , this last sum equals
X Z Z  Z
1
2 δ(t)µ(det t) fv (g −1 tg) dg dtv = fv (g)µv (g) dg
{Lv } Zv \TLv TLv \Gv Zv \Gv

(since the orbital integrals of fv vanish off the elliptic Cartans). Thus µv (fv′ ) =
µ(fv ), as required, and the proof of Proposition 1.3 (and hence Theorem 1.1) is
complete. 

2. The Jacquet–Langlands Correspondence. Let A0 (G) denote the collection

of irreducible invariant subspaces Vπ of R0 (g) in L20 (Z(A)G(F )\G(A)). According
to the “multiplicity one” result of [JL], proved only using the theory of Whittaker
models, each irreducible unitary representation

π = ⊗πv

of G(A) is realizable at most once in L20 . Thus the (distinct) spaces Vπ in A0 (G)
may be confused with the set of (classes of) irreducible automorphic cuspidal repre-
sentations of GL(2). Moreover, “strong multiplicity one” (proved using “only” the
theory of L-functions) implies that Vπ = Vπ′ as soon as πv and πv′ are equivalent
for almost all v.
Now consider the group G′ = Dx , and the collection A0 (G′ ) of irreducibly invari-
ant subspaces of R0′ (g) acting in the subspace of L2 (Z(A)G′ (F )\G(A)) orthogonal
to all the one-dimensional invariant subspaces. Then the trace formula implies
that “multiplicity one” and “strong multiplicity one” hold in A0 (G′ ) as well (even
though Whittaker models disappear in this setting, and the theory of L-functions
cannot be developed à la Hecke . . . ). Indeed, this is a corollary of the following
Jacquet–Langlands correspondence between G′ and G:
56 STEVE GELBART

Theorem 2.1. There exists a 1–1 mapping

Vπ′ −→ Vπ

from A0 (G′ ) to A0 (G), with the property that πv ≡ πv′ for all v ∈
/ SD ; the image
consists of all those π in A0 (G) with the property that πv is square-integrable
(mod Z) for all v ∈ SD .
Remark. A weaker form of this correspondence asserts the existence of a map
between the representations π ′ and π above; i.e., to each π ′ on G′ realizable in
some G′ on A0 (G) there exists a π in A0 (G) such that πv′ ≡ πv for all v ∈ / SD . As
noted in Lecture I, this can be proved using L-functions but the argument does not
yield a characterization of the image, nor the multiplicity one results for G′ .
The trace formula proof of Theorem 2.1 starts from the basic identity

tr R0 (f ) = tr R0′ (f ′ )

valid for matching f and f ′ ; see the Corollary to the proof of Theorem 1.1 in the
preceding section. Equivalently, and more suggestively,
X X
(2.2) tr π(f ) = tr π ′ (f ′ ).
π∈A0 (G) π ′ ∈A′0 (G)

To extract the desired bijection between Vπ′ and π from the identity (2.2) we need
to refine this identity to one of the following type:
Fix any finite set S of places of F containing SD and the archimedean places,
and for each v ∈ / S, fix a given representation πv0 which is unramified; then
XY XY
(2.3) tr πv (fv ) = tr πv′ (fv′ ),
Vπ v∈S Vπ′ v∈S

with fv ∼ fv′ , and the sums taken over all Vπ in A0 (G) (resp. Vπ′ in A0 (G)) such
that πv (resp. πv′ ) is isomorphic to πv0 for all v ∈ / S. Note that by strong multiplicity
one for GL(2), the left-hand side of (2.3) contains at most one term.
Assuming the truth of (2.3), let us explain the slick fashion in which Theorem 2.1
can be proved. Suppose first that no Vπ in A0 (G) corresponds to a given Vπ′ in
A′0 (G′ ). Then (2.3) (with S the set of places outside of which D and πv′ are p-adic
and unramified, and πv0 = πv′ for v ∈ / S) implies that the left-hand side of (2.3) is
zero. Thus we also have
X
(2.4) tr πS′ (fS′ ) = 0,

where Y
πS′ = ⊗ πv′ , fS′ = fv′
v∈S
v∈S

is arbitrary inCc∞ (ZS (A)\G′S (A)),and the sum is over all elements of A′0 (G′ ) (like,
for example, Vπ′ ) such that (π)v ∼
= πv0 for v ∈
/ S. But (2.4) contradicts a well-known
result of local harmonic analysis known as “linear independence of characters” (see
Lemma 16.1.1 of [JL] or Lemma 5.11 of [Ro1]).
 LECTURE VI. APPLICATIONS OF THE TRACE FORMULA 57

Similarly, we can begin to characterize the image of this correspondence Vπ′ → π.

Indeed, suppose π = ⊗πv is in the image of this correspondence, but πv is not
square-integrable for some v0 in SD . Then it follows
tr πv0 (fv0 ) = 0,
since πv0 would then be induced, and fvo ∼ fv′ 0 implies that fv0 has vanishing
hyperbolic integrals. Thus the left-hand side of (2.3) would vanish, leading to the
same contradiction as above.
To complete the characterization of the image, we need to recall some facts
about the square-integrable representations of Gv and G′v . In particular, let h , ie
denote the inner product for the space of class functions on the set of regular elliptic
elements of Zv \Gv or Zv \G′v defined by
X Z
1
hf1 , f2 ie = 2 δ(t)f1 (t)f2 (t) dt,
Tv Zv \Tvreg

the sum extending over the conjugacy classes of compact tori of Gv (conveniently
confused with those of G′v ). For G′v , the Peter–Weyl theorem and the Weyl integra-
tion formula imply that the characters of the irreducible representations of Zv \G′v
comprise a complete orthonormal set with respect to h , ie . On the other hand,
for Gv the so-called “orthogonality relations for square integrable representations”
imply that the characters of the square-integrable irreducible representations of
Zv \Gv comprise at least an orthonormal set with respect to h , ie Thus, given
any square-integrable representation πv on Gv , v ∈ SD , we can determine an irre-
ducible πv′′ on G′v by the condition
hχπv , χπv′′ ie = aπv 6= 0;
we also fix fv′ in Cc∞ (Zv \G′v )
such that

′ ′ 1 if πv′ ∼
= πv′′ ;
tr πv (fv ) =
0 otherwise.
Namely, we take fv′ = χπv′′ .
So suppose now that π in A0 (G) is such that πv is square-integrable for each
v ∈ SD . Then the trace formula identity (2.3) implies that
Y X
(2.5) aπ v = 1,
v∈S Vπ′

with the sum taken over Vπ′ in A′0 (G) such that πv′ ≃ πv′′ for v in S, and πv′ ≃ πv
for v ∈
/ S. Indeed, for v ∈ S,
Z
tr πv (fv ) = f (g)χπv (g) dg
Zv \Gv
X Z
1
= 2 δ(t)χπv (t)Φt (f ) dt
{Tv } Zv \Tvreg

X Z
= 1
2 δ(t)χπv (t)Φt (fv′ ) dt

= hχπv , χπv′′ ie = aπv ,

with Z Z
Φt (fv′ ) = fv (g −1 tg) dg = fv′ (g −1 tg) dg = Φt (fv′ ).
Tv \Gv Tv \G′v
58 STEVE GELBART

Claim. The sum on the right side of (2.5) has exactly one term in it.
Indeed, if there were none, the left-hand side of (2.5) would be zero (which is
impossible by our assumption in aπv , whereas as if there were two or more terms,
we would have Y
|aπv | ≥ 2
(again impossible, since each |aπv | = |hχπv , χπv′′ ie | ≤ kχπv ke kχπv′′ ke ≤ 1). Thus we
conclude there must be exactly one Vπ′ in A′0 (G′ ) corresponding to π in A0 (G),
and the theorem is finally proved.

2. Embedding of Local Discrete Series in Cusp Forms.

Theorem 2.1. Let V be any finite number of finite places of F , and πv a dis-
crete series representation of Zv \Gv for each v in V . Then there exists a cuspidal
automorphic representation π of G(A) such that πv = (π)v for each v in V .
Proof. We can assume that for at least one place v1 ∈ V , πv1 is actually supercusp-
idal (by adding such a pair (v1 , πv1 ) into the collection {v, πv } if necessary). Before
applying the simple trace formula of Deligne-Kazhdan to this situation, we need to
recall two important facts about local harmonic analysis on Gv .
For Gv non-archimedean, a function fv in Cc∞ (Zv \Gv ) is called a pseudo-
matrix coefficient for the discrete series representation πv of Zv \Gv if tr πv (fv ) =
1, but tr τv (fv ) = 0 for any tempered irreducible admissible representation τv not
equivalent to πv . That such pseudo-matrix coefficients exist in the generality of a
general reductive G is a theorem of Bernstein, Deligne and Kazhdan (see [BDK],
[Clo] [BDKV]). Moreover, one knows that for such functions fv and γ (regular)
elliptic in Gv , Z
Φγ (fv ) = fv (x−1 γx) dx = cχπv (γ)
Tv \Gv

for a non-zero constant c. In particular, the orbital integral Φγ (fv ) is not identically
zero (since χπv 6≡ 0 on the regular elliptic elements; see [Clo]).
Remark. Suppose πv is supercuspidal, and we denote one of its matrix coefficients
suitably normalized by fv (g). Then fv belongs to Cc∞ (Zv \Gv ),
tr πv (fv ) = 1, and
tr τv (fv ) = 0
for any irreducible admissible τv not equivalent to πv . Thus the terminology pseudo-
matrix coefficient is indeed apt.
Now given V and {πv } as in the hypothesis of Theorem 2.1, fix
Y
f= fv in Cc∞ (Z(A)\G(A))
v

such that (a) for each v ∈ V , fv is a pseudo-matrix coefficient; and (b) for a fixed
finite place v2 outside V , fv2 is supported on the regular elliptic set of Gv2 . Then
by the simple trace formula of Deligne-Kazhdan, for such f ,
X X
(2.2) tr π(f ) = mγ Φγ (f ),
π γ
 LECTURE VI. APPLICATIONS OF THE TRACE FORMULA 59

where the sum on the left is over the cuspidal automorphic π on Z(A)\G(A) such
that (π)v ∼= πv for each v ∈ V , and the right-hand side is summed over the regular
elliptic conjugacy classes in Z(F )\G(F ). Thus to prove Theorem 2.1, it suffices to
prove that this right-hand side (and hence the left-hand side) does not vanish for a
particular special choice of such f .
Equivalently, write Y
Φγ (f ) = Φγ (fv ),
v

a product of local orbital integrals. It will suffice then to choose f such that
Φγ (f ) 6= 0 for a particular elliptic regular γ0 , but is zero for all other γ.
To this end, first choose γ0 and fv2 (still assumed to have support in the regular
elliptic set) such that

Φγ0 (fv ) 6= 0 for each v in V ∪ {v2 }.

(Recall that for each pseudo-matrix coefficient, Φγ (fv ) = cχπv (γ) is not identically
zero on the elliptic regular set of Gv . . . .) This particular choice of γ0 will intersect
the maximal compact Kv of Gv for almost all finite places outside V ∪ {v2 }. Let us
denote this (infinite) set of places by V ∗ , and for each v in V ∗ set fv equal to the
unit element of Hv = H(Gv , Kv ). Finally, choose fv at the remaining finite places
(outside V ∪ {v2 } ∪ V ∗ ) to be suchQthat Φγ0 (fv ) 6= 0.
With this refined choice of f = fv (still arbitrary at the infinite places!) it is
clear that

(2.3) Φγ (f ) 6= 0

for some elliptic regular γ only if the coefficients of the characteristic polynomial
of γ are v-integral for each v in V ∗ , i.e. are rational with uniformly bounded
denominator. But such rational numbers lie in a lattice of F . Thus we can choose
fv at the archimedean places to have support so small near γ0 as to intersect the
Q
conjugacy class of no other γ satisfying (2.3). With this choice of f = fv ,
Φγ (f ) 6= 0 if and only if γ = γ0 , and the proof is complete. 
Concluding Remark. Crucial to our argument was the fact that the pseudo-
matrix coefficients fv , v ∈ V , eliminated from (2.2) all cuspidal τ = ⊗τv with
τv not equivalent to πv for v in V . This worked in our case since the only non-
tempered τv which might satisfy tr τv (fv ) 6= 0 is the trivial representation (when
πv is the Steinberg representation), and such a representation can never occur as
a local component of a cusp form on GL(2). In general, an additional argument
is necessary, involving some kind of limit-multiplicity argument (due to de George-
Wallach); see [Ro1] or Appendix 3 of [BDKV].

(New) References
[BDK] Bernstein, J., Deligne, P., and Kazhdan, D, Trace Paley–Weiner theorem for reductive
p-adic groups, J. Anal. Math. 47 (1986), 180–192.
[Clo] Clozel, L., Invariant harmonic analysis on the Schwartz space of a reductive p-adic
group, in Harmonic Analysis on Reductive Groups, edited by W. Barker and P. Sally,
Progress in Mathematics, vol. 101, Birkhäuser, Boston, 1991, pp. 101–121.
60 STEVE GELBART

[Kot] Kottwitz, R., Tamagawa Numbers, Ann. of Math. 127 (1988), 629–646.
[Lai] Lai, K. F., Tamagawa Numbers of Reductive Algebraic Groups, Compositio Math.,
vol. 41, Kluwer Acad. Publ., Dordrecht, 1980, pp. 153–188.
[La2] Base Change for GL(2), Ann. of Math. Stud., vol. 96, Princeton University Press,
Princeton, NJ, 1980.
[La3] , The volume of the fundamental domain for some arithmetical subgroups of
Chevalley groups, Proc. Sympos. Pure Math., vol. 9, Amer. Math. Soc., Providence, RI,
1966, pp. 143–148.
 LECTURE VII. (G, M )-FAMILIES AND THE SPECTRAL Jχ (f ) 61

Lecture VII. (G, M )-Families and the Spectral Jχ (f )

The transition from Arthur’s first (“coarse”) form of the trace formula to a more
explicit (“fine”) expansion results from formal operations on collections of functions
called (G, M)-families, operations which pervade the “fine o-expansion” as well as
the “fine χ-expansion”.
The purpose of this lecture is to describe some basic examples and properties
of (G, M )-families, and to explain the pivotal role they play, especially in making
explicit the spectral ramified terms Jχ (f ).

1. Definitions and Basic Examples. Fix a Levi subgroup M of G, and let P(M )
denote the set of not necessarily standard parabolic subgroups of G for which M is
the Levi component.
Definition 1.1.
Suppose that for each P in P(M ), cP (Λ) is a smooth function on ia∗M . Then the
collection
{cP (Λ) : P ∈ P(M )}
is called a (G, M)-family if the following condition holds: if P and P ′ are adjacent
groups in P(M ), and Λ belongs to the hyperplane spanned by the common wall of
the chambers of P and P ′ , then

(*) cP (Λ) = cP ′ (Λ).

Remark 1.2. The compatibility condition (*) here is equivalent to the property
that whenever P and P ′ are elements of P(M ) contained in a given parabolic
subgroup Q, and Λ belongs to ia∗Q , then

cP (Λ) = cP ′ (Λ).

In particular, let
F (M )
denote the parabolic subgroups of G whose Levi components contain M ; then for
each Q in F (M ) one may define a smooth function cQ (λ) on ia∗Q ⊂ ia∗P through
the formula
cQ (Λ) = cP (Λ).
for any parabolic P in P(M ) contained in Q.
62 STEVE GELBART

Example 1.3. Let {XP : P ∈ P(M )} be a collection of points in aM with the

property that for each pair (P, P ′ ) of adjacent groups in P(M ), XP − XP ′ is per-
pendicular to the hyperplane spanned by the common wall of the chambers of P
and P ′ ; pictorially:

X p'

Xp

Figure 3
Then the resulting family of functions
cP (Λ) = eΛ(XP )
on ia∗M is a (G, M )-family. Indeed for Λ on such a hyperplane,
cP (Λ)
= eΛ(XP −XP ′ ) = 1.
cP ′ (Λ)

Remark 1.4. The similarity here with Figure 2 of Lecture IV is not coincidental.
Indeed, for any fixed x in G(A), the points
{−HP (x) : P ∈ P(M )}
comprise an “orthogonal family” of points as depicted in the figure above; this
is proved in Lemma 3.6 of [A8], where such “AM -orthogonal” families of points
were first introduced. (N.B. A collection of points {XP } in aM comprises an AM -
orthogonal family if an only if XP − XP ′ is a multiple of the coroot associated to
the unique root in ∆P ∩ −∆P ′ ).)
Remark 1.5. If L is any Levi subgroup of G containing M , the more general
notion of an (L, M )-family is defined exactly as above: for each P in P L (M ) (the
parabolic subgroups of L with Levi component M ), cP (λ) is smooth on ia∗M , and
if P ′ in P L (M ) is adjacent to P , then cP and cP ′ agree on the wall between the
corresponding adjacent chambers.
For the next Lemma, we need to recall a function θP (Λ) defined on ia∗M by
Y
θP (Λ) = (aP )−1 Λ(α∨ ),
α∈∆P

where aP is the covolume of a certain coroot lattice. (Again, there is the more
general notion of θPQ for any parabolic Q ⊃ P . . . .)
 LECTURE VII. (G, M )-FAMILIES AND THE SPECTRAL Jχ (f ) 63

Lemma 1.6. For any (G, M )-family {cP (Λ) : P ∈ P(M )}, the function
X
cM (Λ) = cP (Λ)θP (Λ)−1 ,
P ∈P(M)

initially defined away from the hyperplanes Λ(α∨ ) = 0, extends to a smooth function
on ia∗M .
Proof. The only possible singularities along Λ(α∨ ) = 0 occur in terms correspond-
ing to those P for which α or −α is a simple root. But such groups occur in pairs
of adjacent P , P ′ , where cP and cP ′ agree. I.e., the contribution of these P ’s to
cM (Λ) is
cP (Λ) cP ′ (Λ)
+ ,
θP (Λ) θP ′ (Λ)
where the numerators are equal, and the denominators appear as negatives of one
another; thus the singularities cancel. 
Remark. The functions cM (Λ) arise naturally as the weight functions of the
weighted orbital integrals described in Lecture IV. Indeed, consider the (G, M )-
family
{cP (Λ) = eΛ(XP ) ; XP = −HP (x)}
recalled above. In [A8] it was shown that the function
X
cM (Λ) = eΛ(XP ) θP (Λ)−1
P ∈P(M)

equals the Fourier transform of the characteristic function of the convex hull of the
points XP (pictured in Figure 3 above). Thus it follows (without appealing to the
more general Lemma 1.6 above) that cM (Λ) is a smooth function on ia∗M (being
the Fourier transform of a compactly supported function). In particular, cM (0)
is defined, and equal to the volume of this convex hull, namely the weight
function vM (x).
It turns out that (G, M )-families pervade even more extensively the spectral
terms Jχ (f ). Here, as we shall explain below, (G, M ) families arise which are
products of families of the above (geometric) type, with (G, M )-families defined
in terms of intertwining operators.

2. Motivation and Examples Arising from the Spectral Distributions

JχT (f ). Recall (from Lecture III) that
Z
JχT (f ) = ΛT2 Kχ (x, x) dx,
Z(A)G(F )\G(A)

with
X Z
1 X X
Kχ (x, y) = E(x, ρ(σ, λ)(f )φ, λ)E(y, φ, λ) dλ.
n(AP ) σ ia∗
P /iaG
∗
P φ
64 STEVE GELBART

Here P = MP NP runs over the “associated” parabolics of G, φ runs thru a K-finite

basis for ρ(σ, λ), and σ runs through (classes of) irreducible unitary representations
G(A)
of MP (A) such that functions in IndP (A) σ belong to L2 (P )χ .
Formally then,
XXZ
T
Jχ (f ) = ΨTσ (λ, f, P ) dλ,
P σ ia∗
P /iaG
∗

where
1
ΨTσ (λ, f, P ) = tr(ΩTχ,σ (P, λ)ρ(σ, λ)(f )),
n(AP )
and Z
(ΩTχ,σ (P, λ)φ′ , φ) = ΛT E(x, φ′ , λ)ΛT E(x, φ, λ) dx.
Z(A)G(F )\G(A)

But according to the inner product formula for truncated Eisenstein series, this
operator ΩTχ,σ (P, λ) equals the value at λ′ = λ of

X X ′ ′
e(t λ −tλ)(T )
(2.1) MP1 |P (t, λ)−1 MP1 |P (t′ , λ′ ) .
θP1 (t′ λ′ − tλ)
P1 t,t′ in
W (aP ,aP1 )

Here MP1 |P (t, λ) is the intertwining operator defined on Ind(σ, λ) by

Z
φ(wt−1 nx) dx
NP1 (A)∩wt NP (A)wt−1 \NP1 (A)

(initially for λ with large real part, then for any λ by analytic continuation). Denote
T
by ωχ,σ (P, λ) the value of the operator defined in (2.1) at λ′ = λ. Actually, for
P ∈/ Pχ , the operators ΩTχ,σ (P, λ) and ωχ,σ T
(P, λ) are only asymptotically equal
(for large T ); this is the thrust of [A9].
Claim.
T
(2.2) tr(ωχ,σ (P, λ)ρ(σ, λ)(f ))

may be evaluated by setting λ′ = λ in the sum over s ∈ W (aP , aP ) of

X
tr( cQ (Λ)dQ (Λ)θQ (Λ)−1 ρ(σ, λ)(f )),
Q∈P(M)

where Λ = Λs = sλ′ − λ, and cQ and dQ are two (G, M ) families to be described

below.
To see this, rewrite the expression (2.1) as

X X X ′
e(t(sλ −λ))(T )
MP1 |P (t, λ)−1 MP1 |P (ts, λ′ ) .
θP1 (t(sλ′ − λ))
s∈WP = P1 t∈W (aP ,aP1 )
W (aP ,aP )
 LECTURE VII. (G, M )-FAMILIES AND THE SPECTRAL Jχ (f ) 65

For any P1 and t ∈ W (aP , aP1 ), set Q = wt−1 P1 wt for any representative wt of t in
G(F ). Then
(P1 , t) ↔ Q
is a bijection between pairs which occurs in the sum above and groups Q ∈ P(M ).
It can also be checked that θP1 (t(sλ′ − λ)) = θQ (sλ′ − λ); moreover,
′ ′
MP1 |P (t, λ)−1 MP1 |P (ts, λ′ )e(t(sλ −λ))(T ) = MQ|P (λ)−1 MQ|P (s, λ′ )e(sλ −λ)YQ (T ) ,

where MQ|P (λ) = MQ|P (1, λ), and

YQ (T ) = t−1 T

(or rather the projection of this point onto aM ). Thus the expression (2.1) indeed
equals the value at λ′ = λ of the sum over s in W (aP , aP ) of
X
cQ (Λ)dQ (Λ)θQ (Λ)−1
Q∈P(M)

with

(2.3) cQ (Λ) = eΛ(YQ (T )) ,

Λ = sλ′ − λ,
and

(2.4) dQ (Λ) = (MQ|P (λ)−1 MQ|P (s, λ′ )).

Remark 2.5. Below we shall see that {cQ (Λ)} and {dQ (Λ)} both define (G, M )-
families. Thus the product

{e(Λ) = c(Λ)d(A)}

is also a (G, M )-family, and the corresponding function

X
eM (Λ) = cQ (Λ)dQ (Λ)θQ (Λ)−1
Q∈P(M)

is smooth on ia∗M . In particular, the expression (2.2) may indeed be evaluated at

λ′ = λ (which corresponds to evaluating eM (Λ) at Λ = sλ − λ).
Remark 2.6. In (2.3) and (2.4), set λ′ = ζ + λ in ia∗M , and s = the identity.
Then the value of eM (Λ) at λ′ = λ corresponds to the (operator) expression

X MQ|P (λ)−1 MQ|P (λ′ )

val eΛ(YQ (T ))
λ =λ
′ θQ (ζ)
Q∈P(M)
X X M (w, λ)−1 M (w, λ + ζ)
= lim ewζ(T ) Q ∨
ζ→0
P1 w∈W (aP ,aP1 ) α∈∆P (wζ)(α )
1
66 STEVE GELBART

This is precisely the operator we denoted by MPT (σ)χ in Proposition 3.9 of Lec-
ture IV. Thus the “weight function” MPT (σ)χ appearing in the weighted character
expression for the unramified spectral term JχT (f ) in Proposition 3.9 is indeed
realizable as a special value of eM (Λ) for an appropriate (G, M )-family eP .
N.B. As we observed in the case of GL(2) (in §2 of Lecture IV), the expression
for JχT (f ) simplifies considerably for unramified spectral data χ, since the sum
over t, t′ in W (aP , aP1 ) in (2.1) need only be taken over t = t′ , i.e., only the term
corresponding to s = 1 really appears in the expression for (2.2), and we indeed
have
X X 1 Z
JχT (f ) = tr(MPT (σ)χ ρ(σ, λ)) dλ
σ
n(A P ) ia ∗ /ia∗
P ∈Pχ P G

as asserted in Proposition 3.9 of Lecture IV.

Proposition 2.7. Both cQ (Λ) and dQ (Λ) define (G, M )-families.
Proof. We have to show that if Q and Q′ are adjacent in P(M ), then

YQ (T ) − YQ′ (T )

is perpendicular to the hyperplane in ia∗M spanned by the wall common to the

chambers of Q and Q′ . Recall that for any Q in P(M ), YQ (T ) = t−1 T if Q =
wt−1 P1 wt for some fixed P1 ∈ P(M ). But if Q′ is adjacent to Q, then Q′ = wt−1
′ P1 wt′
′
with t = sα t, and sα the simple reflection corresponding to an α ∈ ∆P0 . Therefore
YQ (T ) − YQ′ (T ) = t−1 (T − s−1α T ), which is a multiple of t
−1 ∨
α (see footnote1
−1 ∨ ∨ ∨
below). But s α = β , where β is the unique root in ∆Q ∩ (−∆Q′ ). Thus
YQ (T ) − YQ′ (T ) is perpendicular to the hyperplane between Q and Q′ . 
As for dQ (Λ), we recall the standard functional equation

MQ′ |P (s, λ) = MQ′ |Q (1, sλ)MQ′ |P (s, λ),

valid for s ∈ W (aM , aM ) and Q ∈ P(M ) (see equation (1.2) of [A8]). Using it gives

dQ′ (Λ) = MQ′ |P (λ)−1 MQ′ |P (s, λ′ )

= MQ|P (λ)−1 MQ′ |Q (λ)−1 MQ′ |Q (sλ′ )MQ|P (s, λ′ ).

So it remains to show that for Λ belonging to the hyperplane common to Q′ and Q,

MQ′ |Q (λ)−1 MQ′ |Q (sλ′ ) = I.

For this, see §§1 and 2 of [A8].

1 Formally, α∨ = 2α/(α, α), and s−1α = T − 2(T, α)/(α, α) α. Therefore, for any β, (β, T −
s−1
α (T )) = (β, T )−(β, T )+2(T, α)/(α, α) (β, α) = (T, α)(β, α∨ ), i.e., T −s−1
α T is indeed a multiple
of α∨ .
 LECTURE VII. (G, M )-FAMILIES AND THE SPECTRAL Jχ (f ) 67

3. A Splitting Formula for Products of (G, M )-families. In the last section,

we explained how JχT (f ) is expressed in terms of integrals of special values of the
functions eM (Λ), where eP = cP (Λ)dP (Λ) is the product of two (G, M )-families.
The result below is crucial in completing the evaluation of these integrals.
Proposition 3.1 (see Lemma 6.3 of [A4]). In general,
X
(cd)M (Λ) = cQ ′
M (Λ)dQ (Λ),
Q∈F (M)

where F (M ) denotes the set of parabolics of G whose Levi component contains (but
does not necessarily equal ) M .
We will not prove this proposition, since the reader can find it in [A4]. But
we shall at least define the terms cQ ′
M and dQ appearing here, and explain their
significance.
Given any (G, M )-family {cP (Λ)}, and a parabolic Q containing some P in
P(M ), how first do we define the function cQ ?
Since ia∗Q ⊂ ia∗P , it is tempting to simply set

cQ (Λ) = cP (ΛQ )

with ΛQ the projection of Λ onto ia∗Q . But is this well-defined? If P ′ is another

element of P(M ) contained in Q, we must check that

cP ′ (Λ) = cP (Λ)

as soon as λ ∈ ia∗Q . But we have already remarked that the compatibility condition
on the (G, M )-family {cP } implies this is the case (Exercise!), and thus cQ is well-
defined on ia∗Q .
Now for any pair of parabolics Q ⊂ R, define
Y
R
θ̂Q (Λ) = (âR
Q)
−1
Λ(w̃∨ ),
ˆR
w̃∈∆ Q

with w̃∨ in aR
Q defined by

α(w̃∨ ) = w̃(α∨ ), α ∈ ∆R
Q

(and the constant âR

Q the covolume of a certain lattice . . . ). Then the functions

c′Q (Λ), Q⊃P

are defined (initially on the complement of a finite set of hyperplanes in ia∗P ) by

X
c′Q (Λ) = R
(−1)dim(AQ |AR ) θ̂Q (Λ)−1 cR (Λ)θR (Λ)−1 .
{R : R⊃Q}

In Lemma 6.1 of [A4], it is shown that these functions extend to smooth functions
on all of ia∗P .
68 STEVE GELBART

Now that the functions d′Q in Proposition 3.1 are defined, there remains the
simpler task of defining the cQM (Λ); these are just the cM functions (1.5) derived
from the following (L, M )-families of functions {cQ
R (Λ)}.
Fix Q arbitrary in F (M ), and let L = LQ denote its Levi component. By P L (M )
we denote the set of parabolic subgroups of L whose Levi component is M , and
for each R ∈ P L (M ), we set Q(R) equal to the unique group in P(M ) which is
contained in Q and such that Q(R) ∩ L = R. Then the resulting family of functions

{cQ L
R (Λ) = cQ(R) (Λ) : R ∈ P (M )}

comprises an (L, M )-family, and the functions cQ

M (Λ) appearing in Proposition 3.1
are the corresponding cM -functions
X
cQ
M (Λ) = cQ Q
R (Λ)θR (Λ)
−1
.
R∈P L (M)

4. GL(2) Revisited. Although we already discussed (in Lecture IV) the evaluation
of JχT (f ) in this case, we wish to return to this calculation now from the more general
standpoint of Arthur’s theory.
According to the discussion in Section 2 above, JχT (f ) is given by the polynomial

XZ X X n val o
(4.1) λ = λ′ tr(eM (Λ)ρ(σ, λ)(f )) dλ,
P ia∗
P \iaG
∗
σ s∈W (aP )

where e(Λ) is the (G, M )-family given by the product of the family of functions
(2.3) and (2.4). When P = B (the Borel subgroup), we have
 it −it 
λ= ,
2 2
and  it′ −it′ 
λ′ = , ,
2 2
with t′ = t + ζ and Λ = sλ′ − λ.
For χ = {(M, µ)}, the contribution from the term s = 1 reduces to
(4.2)  
Z n eiζ(T1 −T2 )/2 e−iζ(T1 −T2 )/2 M (w, it)−1 M (w, i(t + ζ)) o
tr lim + ρ(µ, it)(f ) dt
R ζ→0 iζ −iζ

Indeed, the two elements of P(M ) are just B and B̄ (the subgroup wBw−1 opposite
to B), and so

eM (Λ) = eB (Λ)θB (Λ)−1 + eB̄ (Λ)θB (Λ)−1

cB (Λ)dB (Λ) cB̄ (Λ)dB̄ (Λ)
= + .
iζ −iζ
 LECTURE VII. (G, M )-FAMILIES AND THE SPECTRAL Jχ (f ) 69

But (2.3) clearly implies that

cB (Λ) = eΛ(T ) = eiζ(T1 −T2 )/2

and
cB̄ (Λ) = e−Λ(T ) = e−iζ(T1 −T2 )/2 .
Also MB|B (1, λ) = I, so dB (Λ) = I, and it remains only to compute dB̄ (Λ). For
this, let Tw denote the linear transformation φ(x) −→ φ(wx) taking IndB̄ σ w | |wλ
to IndB σ | |λ . Then it is straightforward to check that

dB̄ (Λ) = MB̄|B (λ)−1 MB̄|B (1, λ′ )

= (Tw M (w, λ))−1 (Tw M (w, λ′ ))
= M (w, it)−1 M (w, i(t + t′ )),

and thus (4.2) holds. To continue, there is no need to appeal to the fact that the
expression in brackets in (4.2) comes from the (G, M )-family {eP }. Instead, one
can just compute the limit in question to be

(T1 − T2 ) + M (−it)M ′ (it),

and hence the contribution to JχT (f ) (from s = 1) to be

Z ∞ Z ∞
(T1 − T2 ) tr(ρ(µ, it)(f ) dt + tr(M (−it)M ′ (it)ρ(µ, it)(f ) dt.
=∞ =∞

This of course agrees with formula (2.5) of Lecture IV, and for χ unramified, it
accounts for all of JχT (f ).
Now let us concentrate on the more interesting contribution to (4.1), namely the
one corresponding to s = w. In this case, Λ = wλ′ − λ, and this equals (−it, it)
when λ′ = λ. So as above, we compute

cB (Λ) = eΛ(T ) = e−2it(T1 −T2 )

cB̄ (Λ) = e−Λ(T ) = e2it(T1 −T2 )
dB (Λ) = M (w, λ′ )

and

dB̄ (Λ) = M (w, λ)−1 .

In other words, for s = w

( )
val
′
λ = λ tr(eM (Λ)ρ(σ, λ)(f )

e2it(T1 −T2 ) e−2it(T1 −T2 )

= tr(M (w, −it)(ρ(µ, it)(f )) − tr(M (w, it)(ρ(µ, it)(f )) .
2it 2it
This is exactly the expression (6.28) in [GJ]. But instead of evaluating its integral
over R using the simple manipulations of [GJ], let us instead appeal to the splitting
formula for eM (Λ).
70 STEVE GELBART

Lemma. For s = w

eΛ(T ) − e−Λ(T )
eM (Λ) = (cd)M = dB (0)
−2it
dB (Λ) − dB (0) Λ(T )
+ ·e
−2it
d (Λ) − dB (0) −Λ(T )
+ B̄ ·e
2it

with e±Λ(T ) = e∓2it(T1 −T2 ) .

Proof. According to Proposition 3.1, we need to compute

cQ ′
M (Λ)dQ (Λ)

for each Q ∈ F(M ). So first fix Q = G. By definition, dG (Λ) = dB (0) = dB̄ (0),
and then d′G (Λ) = dG (Λ). On the other hand, we compute (from the definition of
cQ G
M ) that cM (Λ) = cM (Λ). Thus the leading term of (cd)M (Λ) is indeed

eΛ(T ) − e−Λ(T )
dB (0) ,
−2it
as claimed.
Now fix Q = B. Then

dB (Λ) − dB (0)
d′B (0) = ,
−2it

while cB L
M (Λ) = cB (Λ). Indeed, in this case L = M and P (M ) = M ! Similarly

dB̄ (Λ) − dB̄ (0)

cB̄ ′
M (Λ)dB̄ (Λ) = · cB̄ (Λ),
−2it
and we’re done. 
Proposition. For T large, the contribution to JχT (f ) from (the Borel subgroup B
and ) s = w is just

tr(dB (0)ρ(0, µ)(f )) = tr(M (0, µ)ρ(0, µ)(f ));

in particular, only the leading term in (cd)M (Λ) contributes to JχT (f ).

Proof.
Let us suppose for the moment that JχT (f ) can be calculated by taking a com-
pactly supported Schwartz function B(t) with B(0) = 1, and then computing
Z ∞
JχT (f ) = lim lim tr(ωχT (t)ρ(µ, it)(f ))B(εt) dt.
ε→0 α(T )→∞ −∞

By the last lemma, the contribution from s = w (and P = B) to

tr(ωχT (t)ρ(µ, it)(f ))

 LECTURE VII. (G, M )-FAMILIES AND THE SPECTRAL Jχ (f ) 71

is
 
e2it(T1 −T2 ) − e−2it(T1 −T2 )
tr(dB (0)ρ(µ, it)(f ))
2it
tr((dB (Λ) − dB (0))ρ(µ, it)(f ))
+ e−2it(T1 −T2 )
−2it
tr((dB̄ (Λ) − dB (0))ρ(µ, it)(f ))
+ e2it(T1 −T2 ) .
−2it
But for any real T ,
Z T
eiyT e−iyT
+ = eiyx dx
iy −iy −T
Z ∞
= χT (x)e−iyx dx,
∞

with χT the characteristic function of [−T, T ]. Thus we can write the contribution
of the leading term as
Z ∞ Z ∞ 
lim lim χα(T ) (x) e2itx tr(dB (0)ρ(µ, it)(f ))B(εt) dt dx
ε→0 α(T )→∞ −∞ −∞
Z ∞ 
= lim lim χα(T ) (x)F̂ε (x) dx
ε→0 α(T )→∞ −∞
Z ∞ 
= lim F̂ε (x) dx = lim Fε (0),
ε→0 −∞ ε→0

where

Fε (t) = tr(dB (0)ρ(µ, it)(f ))B(εt)

= tr(M (0, µ)ρ(µ, 0)(f )) at t = 0,

since B(0) = 1. On the other hand, for the second (or third) term, the contribution
is Z ∞
lim lim e∓2it(T1 −T2 ) Fε (t) dt
ε→0 α(T )→∞ −∞

where for each ε,

 
dB (t) − dB (0)ρ(µ, it)(f )
Fε (t) = tr B(εt)
±2it

is a smooth Schwartz function on R, whose Fourier transform vanishes at infinity.

Thus the second and third terms contribute nothing to JχT (f ). 
Concluding Remarks. (1) In treating the GL(2) calculation of JχT (f ) from the
point of view of Arthur’s general theory, we have made a big mountain out of
a small mole hill. However, already for GL(3), there are compelling reasons for
appealing to the general theory—not just for using the splitting formula for products
of (G, M )-families, but also for exploiting the compactly supported function B in
72 STEVE GELBART

the calculation of Jχ (f ). Indeed, without using this function B, there is the need to
directly establish the integrability of a function like tr(M (w, λ)ρ(σ, λ)(f )) (in order
to justify applying a Riemann-Lebesgue Lemma as T → ∞). In the case of GL(2),
this can be checked directly, but already for GL(3) the required estimates on the
growth of M (w, λ) are non-trivial (see Section 3 of [Ja1]), and in general they are
simply not available.
(2) It remains to prove that the above computation (with the insertion of the
“smoothing function” B(t)) is justified. This is precisely the goal of Arthur’s paper
[A7] and—we hope—part of the subject matter of Lecture IX.

(New) Reference
[Ja1] Jacquet, H., The continuous spectrum of the relative trace formula for GL(3) over a
quadratic extension, Israel Journal of Math. 89, 1–59.
 LECTURE VIII. JACQUET’S RELATIVE TRACE FORMULA 73

Lecture VIII. Jacquet’s Relative Trace Formula

In deriving Arthur’s trace formula

X X
(*) JoT (f ) = JχT (f ),
o χ

we started by restricting the kernel functions

X X
Kf (x, y) = Ko (x, y) = Kχ (x, y)
o χ

to the diagonal subgroup.

Z(A)G(F )\G(A) × Z(A)G(F )\G(A).

Then after modifying these restricted kernels, we were able to integrate them and
obtain the formula (*). The term “trace formula” seemed apt because—at least for
cuspidal data χ = {(M, r)} with M = G, JχT (f ) actually represents the trace of
R(f ) restricted to L2χ .
The idea of Jacquet’s relative “trace formula” is to restrict the kernel functions
Kf (x, y) to interesting subgroups of the diagonal, and then integrate them against
possibly non-trivial automorphic forms on these subgroups. Although the re-
sulting integrals no longer represent the “trace” of anything, there results from this
new approach a dramatic array of interesting possible applications to automorphic
forms (See [JLR] and [Ja1] for a general discussion, and further references.)
In this lecture we shall concentrate on a particular example (cf. [Ye]) which
already shows the power of this method, and indicates how the further development
of the method is inextricably linked to the general theory developed by Arthur.
Even in our exposition of this simplest example, we are influenced by Jacquet’s
more general ideas (see [Ja1]); moreover, the approach to the computations in §4
is adapted from joint work with him and Rogawski on similar computations for
GL(3).

1. Description of a Base Change Theorem. Fix a quadratic extension E of

the number field F , and let ωE/F denote the corresponding quadratic character of
F ∗ \AF . Let us call a cuspidal automorphic representation π of GL2 (AF ) stable if
it is not of the form π(µ′ ) for any grossencharacter µ′ of E.
74 STEVE GELBART

Theorem (cf. [Ye]). Every stable cuspidal π on GL2 (AF ), with central charac-
ter ωE/F , has a base change lift to a cuspidal representation πE of GL2 (AE );
moreover, the resulting map
π −→ πE

is a bijection onto the cuspidal representations Π of GL2 (AE ) which (have trivial
central character and ) are distinguished with respect to GL2 (F ), i.e., the “period”
Z
(1.2) φ′ (h) dh
Z(A) GL2 (F )\ GL2 (AF )

is non-zero for some φ′ in the space of Π.

Remarks. (i) In [La2], Langlands established the existence of a base change lift
for GL2 for an arbitrary cyclic extension E of F , and in[AC] these results were
dramatically generalized to GLn . Both these works characterize the image of base
change in terms of the Galois invariance of the representations “upstairs”; this is
not surprising, since the proofs proceed from a comparison of the (“invariant”)
trace formula over F with the (Galois) “twisted” (“invariant”) trace formula over
E.
(ii) In [HLR] it is shown that an automorphic cuspidal representation Π of GL2 (E)
(trivial on its center) is the base change lift of an automorphic cuspidal representa-
tion π of GL2 (AF ) (with central character ωE/F ) if and only if Π is distinguished.
The argument there combines the base change result of [La2] with properties of
the so-called Asai L-function L(s, Π, Asai) (which has a pole at s = 1 if and only
if π is distinguished . . . ). (By contrast, the relative trace formula proof of Theo-
rem 1.1 (see below) makes no appeal to [La2].) The main thrust of [HLR] was to
study algebraic cycles on certain Shimura varieties attached to GL2 over E, whose
Hasse–Weil zeta functions are computable in terms of

L(s, Π, Asai).

It was precisely the desire to extend these results to the context of quaternion
algebras that led to the idea of the relative trace formula introduced in [JLai].
(iii) In [JLai], Jacquet and Lai used a comparison of relative trace formulas on
GL2 /E and the multiplicative group G′ of a division quaternion algebra over E (the
relative subgroups of the diagonal being the F -diagonal subgroups) to prove that Π′
on G′ /E is G′ (F ) distinguished if and only if its Jacquet–Langlands correspondent
Π on GL2 /E is GL2 (F ) distinguished. This was applied in [Lai2] to extend the
results of [HLR] to the case of certain compact Shimura varieties.

2. Outline of the Proof of Theorem 1.1. It is convenient to use the notation H

for the group GL2 over F , and G for the group ResE
F GL2 ( so that H(A) = GL2 (AF )
and G(A) = GL2 (AE )). By
Y
f= fv
 LECTURE VIII. JACQUET’S RELATIVE TRACE FORMULA 75

we denote a test function in Cc∞ (H(A), ωE/F ), where for almost every place v, fv
is the identity element of the local Hecke algebra H(Hv , KvH , ωv ); similarly,
Y
f′ = fv′

belongs to Cc∞ (Z(AE \G(AF )), with fv′ almost everywhere the identity in

H(ZGv \Gv , KvG ).

Corresponding to f there is the convolution operator

Z
R(f ) = f (h)R(h) dh
Z(A)\H(A)

in the space L2 (H(F )\H(A), ω); it is an integral operator with kernel

X
Kf (x, y) = f (x−1 γy).
γ∈Z(F )\H(F )

Similarly f ′ determines an integral operator R′ (f ′ ) in L2 (ZG (A)G(F )\G(A)) with

kernel X
Kf ′ (g1 , g2 ) = f ′ (g1−1 γg2 ).
γ∈ZG (F )\G(F )

The proof of Theorem 1.1 will come from comparing certain “relative traces” asso-
ciated to these kernels, as we shall now explain.
Step I. Prove that for “matching” f and f ′ ,
ZZ
(2.1) Kf (n1 , n2 )ψN (n−1
1 n2 ) dn1 dn2
[N (F )\N (AF )]2
Z Z
= Kf ′ (n′ , h)ψN
′ ′ ′
′ (n ) dn dh.
Z(A)H(F )\H(A) N ′ (F )\N ′ (AF )

Here N (resp. N ′ ) is the standard maximal unipotent of H (resp. G), and ψN

′
(resp. ψN ) the corresponding character of N (F )\N (AF ) (resp. N ′ (F )\N ′ (AF ))
determined by a fixed additive character ψ (of F \A) (with ψ ′ = ψ ◦ trE/F ).
Remarks. (i) A special feature of this particular relative trace formula approach is
the absolute convergence of the integrals on either side of the identity (2.1). On
the left hand side, this is clear from the compactness of the domains of integration;
for the right side, this follows from the fact that for sufficiently large T ,
Z
ΛT2 Kf ′ (n′ , h)ψN
′ ′
′ (n ) dn
′
N ′ (F )\N ′ (A)

equals (as opposed to just approximates)

Z
Kf ′ (n′ , h)ψN
′ ′ ′
′ (n ) dn .
76 STEVE GELBART

For a proof, see Proposition 2.1 of [Ja1].

(ii) At first sight, the equality in (2.1) seems surprising, since the double cosets

N \H/N and N ′ \G/H

which arise on each side do not match up. However, it seems to be a general feature
of the relative trace formula that the “relevant” orbits—those which survive the rel-
ative integrations—do indeed match up. In the case at hand, it is the integration
against ψ which allows for the matching up of relevant orbits. Once this is estab-
lished, the assertion that (2.1) holds for matching f and f ′ reduces quickly to a
collection of purely local assertions: two local function fv and fv′ are “matching”
if certain local relative orbital integrals (indexed by identical sets) are equal.
(iii) It is also crucial to prove a “fundamental lemma” asserting that this local
matching is compatible with the base change map BC ∗ between the Hecke algebras
of Kv and Gv . This map is dual to the natural base change morphism

BC : (g, σ) −→ (g, g, σ)

from L H = GL2 (C) × Gal(E/F ) to L G = (GL2 (C) × GL2 (C)) ⋊ Gal(E/F ). In

particular, when fv′ belongs to the Hecke algebra of Gv , the function fv in the
Hecke algebra of Hv defined by

(fv )∨ (t) = (fv′ )∨ (BC(t))

will be a matching function to fv′ . For proofs of these matching results, see [Ye]. For
more general results, and discussion of “relative fundamental lemmas”, see [Ja2],
[JaYe], [Mao], and [JLR].
Step II. Prove (again for these matching f and f ′ ) that the relative traces of the
continuous parts of the kernels Kf and Kf ′ cancel each other out.
To explain what this means, let us first write out the spectral expansion of Kf
in crude form:

Kf (x, y) = Kf, cont + Kf,sp + Kf,0 ,

where Kf,cont denotes the kernel of R(f ) restricted to the continuous spectrum of
L2 (H(F )\H(A), ωE/F ), and Kf,sp the kernel for the discrete non-cuspidal spec-
trum. Similarly,
Kf ′ (g1 , g2 ) = Kf ′ ,cont + Kf ′ ,sp + Kf ′ ,0 .
A simple but key observation is that the sp-part of the kernel contributes nothing
to the relative trace. Indeed,
Z
Kf,sp (n1 , n2 )ψ(n1−1 n2 ) dn1 dn2
[N (F )\N (AF )]2
Z
=0= Kf ′ ,sp (n′ , h)ψN
′ ′ ′
′ (n ) dn ,
N ′ (F )\N ′ (A)

since these parts of the spectrum are non-generic.

 LECTURE VIII. JACQUET’S RELATIVE TRACE FORMULA 77

Thus, if we can show that the continuous kernels contribute roughly equal relative
traces, we might then conclude
ZZ ZZ
−1
(2.2) Kf,0 (n1 , n2 )ψN (n1 n2 ) dn1 dn2 = Kf ′ ,0 (n′ , h)ψN ′ (n′ ) dn′ dh

Henceforth, we denote the maximal compact of Hv by Kv and the maximal

compact of Gv by Kv′ .
Step III. Use the equality of the relative cuspidal traces on H and G to establish
the asserted base change bijection between the cuspidal (stable) π on H and the
cuspidal (distinguished) Π on G.
To see why this is relatively (!) straightforward, let us write out the explicit
spectral expansions for Kf,0 and Kf ′ ,0 , namely,
X X
Kf,0 (x, y) = π(f )φ(x)φ(y),
π φ∈Vπ
cuspidal

and X X
Kf ′ ,0 (g1 , g2 ) = Π(f )φ(g1 )φ′ (g2 ).
Π φ′ ∈VΠ
cuspidal
Q ′S
If we assume f and f ′ are right K S = v∈S / Kv (resp. K ) invariant, where S is
a finite set of places (including the infinite ones), then these identities read
X X Y
(2.3) Kf,0 (x, y) = (f S )∨ (t(π S )) πv (fv )φ(x)φ(y)
π K S 6={0} {φπ } v∈S

and
X S
X Y
(2.3′ ) Kf ′ ,0 (g1 , g2 ) = (f ′ )∨ (t(ΠS )) Πv (fv′ )φ′ (g1 )φ′ (g2 ).
Π K ′S 6={0} {φ′Π } v∈S

If we further assume that for v ∈ / S (where both fv′ and fv are spherical), fv
′
corresponds to fv via the natural base change map between their Hecke
algebras, then,
S
(f S )∨ (t(π S )) = (f ′ )∨ (BC(t(π S ))).
So plugging these expressions into the relative trace formula identity (2.2) yields
the more explicit identity
X S
X S
(2.4) A(π, fS )(f ′ )∨ (BC(t(π S )) = B(Π, fS′ )(f ′ )∨ (t(ΠS ))
π Π

with
X
A(π, fS ) = WπψS (fS )φ (e)Wφψ (e),
{φπ }
X ′
B(Π, fS′ ) = WΠψS (f ′ )φ′ (e)D(φ′ ),
S
{φ′Π }
Z
Wφψ (e) =
′
φ(n)ψN (n) dn, (with a similar expression for W ψ )
N (F )\N (A)
78 STEVE GELBART

and
Z
D(φ′ ) = φ′ (h) dh.
Z(A)H(F )\H(A)

Note that (2.4) may be viewed as a (absolutelyQ convergent) linear combination

S
of irreducible
Q characters of the group (G) = v ∈S
/ Gv . Also, note that a particular
cuspidal on G(A) will appear on the right side of (2.4) only if it is distinguished,
i.e., D(φ′ ) 6= 0 for some φ′ . Thus it is not surprising that a “linear independence of
characters argument” yields the desired bijection of Theorem 1.1. For more details,
we refer the reader to [Ye], and §4 below.
Since the most subtle part of the proof outlined above is Step II, we focus on
this step in some detail in the sections below.

3. Step II: Subtracting off the Continuous Spectrum. Once we know that
the relative trace formulas on H and G agree for matching f and f ′ , it follows
(keeping in mind the previous discussion) that
ZZ ZZ
(3.1) Kf,0 (n1 , n2 )ψN (n−1 n2 ) dn1 dn2 − Kf ′ ,0 (n′ , h)ψN ′ (n′ ) dn′ dh
ZZ ZZ
−1
= Kf,cont (n1 , n2 )ψN (n n2 ) dn1 dn2 − Kf ′ ,cont (n′ , h)ψN ′ (n′ ) dn′ dh.

“Subtracting off the continuous spectrum” from the initial relative trace formula
identity thus amounts to being able to prove that both sides of this last identity
are identically zero.
According to (2.3) and (2.4), the left hand side of (3.1) reduces to the discrete
expression
X S
X S
A(π, fS )(f ′ )∨ (BC(t(π S )) − B(Π, fS′ )(f ′ )∨ (t(ΠS )).
π Π

So the natural strategy should be to show that the right hand side of (3.1) can be
written as the difference of two absolutely convergent expressions of the form
Z
S
(3.2) Φ(σ)(f ′ )∨ (σ) dσ

with Φ(σ) an integrable function on the unramified unitary dual of GS . In fact it

turns out that the right-hand side of (3.1) contributes a discrete as well as con-
tinuous distribution (i.e., the continuous spectrums cancel each other only modulo
an “unstable” piece of the cuspidal spectrum). Nevertheless, it is possible (after
rearranging terms) to apply the trick of Langlands from [La2, p. 211]: since (3.1)
amounts to an equality between an “atomic” and a “continuous” measure on the
unramified unitary dual of GS , both sides must be identically zero.
Consider for example the expression
ZZ
(3.3) Kf,cont (n1 , n2 )ψN (n−1
1 n2 ) dn1 dn2 ,
[N (F )\N (A)]2
 LECTURE VIII. JACQUET’S RELATIVE TRACE FORMULA 79

where
X
Kf,cont (x, y) = Kχ (x, y)
χ={(M,µ)}
XXZ ∞
= E(x, ρ(µ, it)(f )φ, it, µ)E(y, φ, it, µ) dt.
µ {φµ } −∞

Because the integration is over a compact domain, it is easy to compute that (3.3)
equals
X Z ∞
(3.4) (f S )∨ (µ, it)c(fS , µ, it) dt.
µ −∞
unramified
outside S

where (µ, it) denotes the conjugacy class in L GS attached to Ind µS kit ,
X
c(fS , µ, it) = W (ρ(µS , it)(fS )φ, ψ, µ, it)W (φ, ψ, µ, it),
{φµ }

and Z
W (φ, ψ, µ, it) = E(n, φ, it, µ)ψN (n) dn.
N (F )\N (A)
S
Moreover, because f S “matches” f ′ , we have
S
(3.5) (f S )∨ (µ, it) = f ′ )∨ (BC(µ, it),

with (BC)(µ, it) = (µ′ , it) and µ′ (z) = µ(z z̄). Thus the contribution from Kf,cont
to the right hand side of (3.1) is indeed of the form (3.2).
Now consider the relative trace formula expression
Z Z
Kfcont
′ (n′ , h)ψN ′ (n′ ) dn′ dh
Z(A)H(F )\H(A) N ′ (F )\N ′ (A)

corresponding to the quadratic extension. As we already hinted (and shall explain

in the next section), this turns out to produce discrete as well as continuous sums
S
of traces (f ′ )∨ (σ).

4. Step II Continued: Periods of Truncated Eisenstein Series, and the

Continuous relative trace over G. Our purpose here is to describe some of the
difficulties involved in proving the following result:
Theorem 4.1. The contribution of the continuous spectrum to the relative trace
formula for G is
X Z ∞ X
S S
(f ′ )∨ (µ′ , it)c1 (fS′ , µ′ , it) dt + (f ′ )∨ (µ′ , 0)c2 (fS′ , µ′ , 0).
−∞
µ′ µ′
′
µ =µ◦NE/F µ′ 6=µ◦NE/F
80 STEVE GELBART

Here both expressions are absolutely convergent,

X XZ ∞
|c2 (fS′ , µ′ , 0)| < ∞, and |c1 (fS′ , µ′ , it)| dt < ∞.
µ′ µ′ −∞

As we already explained in the last section, this theorem is exactly what is

needed to complete the proof of Theorem 1.1. Indeed, if we subtract the integral
expression here from
ZZ
Kf,cont (n1 , n2 )ψN (n−1
1 n2 ) dn1 dn2

in (3.1) we obtain (using(3.4) and (3.5)) the expression

XZ ∞ S
(4.2) (f ′ )(µ′ , it)c∗ (fS′ , µ′ , it) dt.
µ′ −∞

On the other hand, bringing the discrete sum in (4.1) over to the left-hand side
of (3.1) yields the expression
X S
(4.3) A(π, fS )(f ′ )(BC(t(π S ))
π
cuspidal
X S
B(Π, fS′ )(f ′ )∨ (t(ΠS ))
Π cuspidal
X S
− (f ′ )∨ (µ′ , 0)c2 (fS′ , µ′ , 0).
µ6=µ◦NE/F

S
Hence we indeed can conclude (by the arbitrariness of (f ′ )) that both (4.2) and
(4.3) are identically zero.
Remark. For any given grossencharacter µ′ of E which does not factor through the
norm map, let π(µ′ ) denote the cuspidal representation of H(A) whose Langlands
parameter tπv almost everywhere base change lifts to the Langlands parameter of
IndG ′ ′
B ′ µv . These π(µ ) are the cusp forms constructed by Hecke and Maass (using
either the theory of theta-functions or L-functions). The fact that (4.3) equals zero
gives (by “linear independence of characters”) an alternate proof of the existence
of such forms, as well as establishing the bijection asserted in Theorem 1.1.
Now let us return to Theorem 4.1, giving a concrete expression for the relative
trace of Rcont (f ′ ). As we have already noted—for α(T ) sufficiently large,
Z Z
Kfcont
′ (n′ , h)ψN ′ (n′ ) dn′ dh
Z(A)H(F )\H(A) N ′ (F )\N ′ (A)
ZZ
= ΛT2 Kf ′ ,cont (n′ , h)ψN
′
(n′ ) dn′ dh.

Moreover, it follows from Propositions 2.3 and 2.5 of [Ja1] that this last expression
equals ZZ
X
ΛT2 Kχ (n′ , h) ψN ′ (n′ ) dn′ dh
χ={(M ′ ,µ′ )}
 LECTURE VIII. JACQUET’S RELATIVE TRACE FORMULA 81

and, ultimately, that

Z Z
(4.4) Kfcont
′ (n′ , h)ψN ′ (n′ ) dn′ dh
Z(A)H(F )\H(A) N ′ (F )\N ′ (A)
XZ ∞ X Z 
= ΛT E(h, ρ(µ′ , it)(f ′ )φ′ , it) dh
µ′ −∞ {φ′ } Z(A)H(F )\H(A)
µ′

× W (φ′ , ψ, µ′ , it)(e) dt.

Thus our task is to obtain an explicit expression for the right-hand side of this last
equation, as T → ∞ (i.e., for T sufficiently large . . . ); in this we follow [Ja1].
Remark. Introducing a truncation operation facilitates the explicit calculation
of the relative continuous spectrum of G but, as with the ordinary trace formula,
also produces “discrete traces” from within the continuous spectrum, as we shall
now explain.
In (4.4), the sum is over the (classes of) unitary characters µ′ of E ∗ \AxE inducing
the Eisenstein series
X
E(g, φ′ , µ′ , s) = φµ (γg),
γ∈B ′ (F )\G(F )

with φµ ∈ Ind µ′ ks , and B ′ = ResE ′ ′

F B = M N . In computing the periods
Z
P(φ′ , µ′ , it) = ΛT E(h, ρ(µ, it)(f )φ′ , µ′ , it) dh,
Z(A)H(F )\H(A)

one appeals to the Bruhat-type decomposition

[
(*) G(F ) = B ′ (F )H(F ) B ′ (F )ηH(F ),
 √  √
where η in G(F ) satisfies η η̄ −1 = w. (One can take η = −√ττ 11 if τ generates E.)
The proposition below (see [JLai] and [Ja1]) is to be viewed as a “relative analogue”
of the Langlands formula for the inner product of two truncated Eisenstein series.
Proposition 4.5. For each unitary character µ′ of E ∗ \A∗E , set

m(F x \A1F ) if µ′ |F x \A1F ≡ 1;
δ1 (µ′ ) =
0 otherwise,
and 
m(A∗F E x \A∗E ) if µ(a) ≡ µ(ā);
δ2 (µ′ ) =
0 otherwise,
Then for sufficiently large α(T ) ,
P(φ′ , µ′ , it)
n eit(T1 −T2 ) Z e−it(T1 −T2 )
Z o
= δ1 (µ′ ) φ′ (k) dk − M (w, it)φ′ (k) dk
it K it K
Z
′
+ δ2 (µ ) φ′ (ηh) dh
T η(A)\H(AF )

with \
Tη = η −1 B ′ (F )η H(F ).
82 STEVE GELBART

Remark. The δ1 -term here corresponds to the contribution from the “small-cell”
part of (*), whereas the δ2 -term corresponds to the “big cell” (w 6= e). Note that
the second term is independent of the truncation parameter T . Also, all terms
must be understood in the sense of meromorphic continuation from s in some right
half-plane to s = it.
Sketch of the Proof. From the definition of ΛT , and the formula for the constant
term of E, one computes that
X
(4.6) ΛT E(g, φ′ , µ′ , s) = φ′ (γg)(1 − τ̂B (H(γg) − T )
γ∈B ′ (F )\G(F )

− M (w, s)φ′ (γg)τ̂B (H)(γg) − T ).

(This is a special case of Arthur’s “second formula” for ΛT E; see Lemma 4.1 of
[A3].) The next step is to check that (for s in the domain of convergence of the
Eisenstein series), the series
X
|φ′ (γg)| |1 − τ̂B (H(γg) − T )| + |M (w, s)φ′ (γg)|τ̂B (H(γg) − T )
γ∈B ′ (F )\G(F )

is integrable over Z(A)H(F )\H(A). This justifies the term by term integration of
(4.6) over Z(A)H(F )\H(A); see [Ja1] §3 for the case of GL(3). Then the decompo-
sition (*), together with Iwasawa’s decomposition for H(A), ultimately yields the
Proposition. Implicit here is the fact that for α(T ) sufficiently large,

τ̂B (H(ηh) − T ) ≡ 0 for all h;

this requires some proof (see Proposition 7.2 of [Ja1] for a more general setting),
and explains why the contribution from η is indeed independent of T .
In applying Proposition 4.5 to the calculation of
ZZ
(4.1) Kf ′ ,cont (n′ , h)ψN ′ (n′ ) dn′ dh,

let us first consider the contributions from characters µ′ with δ2 (µ′ ) 6= 0. These are
precisely the grossencharacters of E which factor thru the norm map from F . If in
addition δ1 (µ′ ) = 0, then Proposition 4.5 implies that the contribution of such µ′
to (4.1) is
XZ ∞ X
(4.7) θf (µ′ , it, φ′ )W (φ′ , ψ, µ′ , it) dt
µ′ −∞ {φ′ }
µ′

where Z
′ ′ ′
θf (µ , it, φ ) = δ2 (µ ) ρ(µ′ , it)(f )φ′ (ηh) dh.
Tη (AF )\H(AF )

Note we can rewrite (4.7) as

XZ ∞
S
(f ′ )∨ (µ′ , it)CfS′ (µ′ , it) dt
µ′ −∞
 LECTURE VIII. JACQUET’S RELATIVE TRACE FORMULA 83

with X
CfS′ (µ′ , it) dt = W (φ′ , ψ, µ′ , it)θfS′ (µ′ , it, φ′ ).
φ′

Now suppose δ2 (µ′ ) = 0, i.e., µ′ does not factor thru the norm map from F .
Then Proposition 4.5 implies that the contribution of each such µ′ to the relative
trace (4.1) is
Z X
(4.8) W (φ′ , ψ, µ′ , it)·
φ′
 Z Z 
′ eit(T1 −T2 ) ′ e−it(T1 −T2 ) ′
δ1 (µ ) f φ (h) dh − M (w, it)f φ (k) dk dt,
it K it K

with
f φ′ = ρ(µ′ , it)(f )φ′ .
Notice the similarity here with the formula encountered for the ordinary trace
formula term JχT (f ) in §4 of Lecture VII. There we computed the limit as T → ∞
by applying Arthur’s splitting formula for products of (G, M )-families. Here we
can do the same thing, taking (as before)

cB (it) = e2it(T1 −T2 ) and cB̄ (it) = e−2it(T1 −T2 ) .

But now we set Z

dB (it) = W (φ′ , ψ, µ′ , it) f φ′ (k) dk
K
and Z
dB̄ (it) = W (φ′ , ψ, µ′ , it) M (w, it)(f φ′ )(k) dk.
K
It must of course be checked that this d(Λ) is also a (G, M )-family (not obvious!).
The main task, though, is to compute
X
lim lim (cd)M (it)B(εt) dt,
ε→0 T →∞
φ′

with B(t) a compactly supported Schwartz function as in Lecture VII. Computing

as before, we find that
Z
d′G (it) = dB (0)(= dB̄ (0)) = W (φ′ , ψ, µ′ , 0) ρ(µ′ , 0)(f )φ′ (k) dk
K

with φ′ in Ind µ′ , and—ultimately—that the above limit equals

X Z
W (φ , ψ, 0, µ ) ρ(0, µ′ )(f ′ )φ′ (k) dk.
′ ′

Thus we find that the total contribution (to the relative trace formula) from char-
acters µ′ of this form is
X S
(f ′ )∨ (0, µ′ )C(f ′ S , µ′ )
µ′
µ′ 6=N ◦µ
84 STEVE GELBART

with Z
X
C(ψ, f ′ S , µ′ ) = W (φ′ , ψ, 0, µ′ ) ρ(0, µ′S )(fS′ )φ′ (k) dk.
φ K

It remains to observe that when δ2 (µ′ ) 6= 0 and δ1 (µ′ ) 6= 0, the contribution to

the relative trace formula is again just (4.5). There are apparent contributions
Z Z
e2it(T1 −T2 ) ′ e−2it(T1 −T2 )
f φ (k) dk − M (w, it)φ′ (k) dk
it it

from P(φ′ , µ′ , t), but they integrate to zero over R by way of a Riemann–Lebesgue
Lemma (just like the terms in cdM (t) not belonging to the leading term Q = G in
F (M ) at the end of the last Lecture). We spare the reader the details.
Concluding Remarks. A host of natural problems present themselves in connec-
tion with any general development of the relative trace formula, especially as far
as applications are concerned. In addition to the obvious problem of developing
general local “matching” theorems (and related “fundamental lemmas”), there are
these problems from the “spectral side”:
(1) Develop a general formula for the period of a truncated Eisenstein series,
analogous to Langlands’ formula for the inner product of truncated Eisen-
stein series;
(2) Apply a general theory of (G, M )-families and smoothing functions B(λ)
to the calculation of the contribution of the “continuous spectrum” to the
relative trace formula, and
(3) Prove in general that the resulting spectral expressions are sufficiently con-
vergent to justify subtracting them off from the full relative trace formula.
For the case of GL(2), this can done by direct estimates on the intertwining
operators, but already for GL(3) such a direct approach is problematic.

References
[Ja1] Jacquet, H., The continuous spectrum of the relative trace formula for GL(3) over a
quadratic extension, Israel Journal of Math. 89, 1–59.
[Ja2] , Relative Kloosterman integrals for GL(3), II, Canad. J. Math. 44 (1992), 1220–
1240.
[JLai] Jacquet, H. , and Lai, J., A relative trace formula, Comp. Math. 54 (1985), 243–310.
[JLR] Jacquet, H. , Lai, K., and Rallis, S., A trace formula for symmetric spaces, Duke Math.
J. 70 (1993), no. 2, 305–372.
[JaYe] Jacquet, H. , and Ye, Y., Une remarque sur le changement de base quadratic, C.R.A.S.
Paris 311 (1990), 671–676.
[Mao] Mao, Z., Relative Kloosterman integrals for GL(3), III, Canad. J. Math. 45 (1993),
no. 6, 1211–12430.
[Ye] Ye, Y., Kloosterman integrals and base change for GL(2), J. Reine Angew. Math. 400
(1989), 57–121.
 LECTURE IX. APPLICATIONS OF PALEY–WIENER; CONCLUDING REMARKS 85

Lecture IX. Some Applications of

Paley–Wiener, and Concluding Remarks

In Lecture VII, we explained how (G, M )-families arose in Arthur’s explicit calcu-
lation of the general spectral terms Jχ (f ). We also alluded to why a Paley–Wiener
theorem was needed to complete this calculation, and promised to return to it. In
fact, this Paley–Wiener theorem is also crucial in handling other convergence prob-
lems related to the trace formula. So our purpose now is to finally touch on these
matters, albeit briefly. This being the last lecture, we shall also conclude with a
few remarks orienting the reader to some papers of Arthur’s that we so far haven’t
even mentioned.

1. The Paley–Wiener Theorem. The Fourier transform of a compactly sup-

ported smooth function f on R extends to an entire function F (Λ) on C, and
satisfies a well-known growth condition, namely that there exists a constant N
(depending on the support of f ) such that

sup |F (Λ)|e−N | Re(Λ)| (1 + | Im(Λ)|)n < ∞

for any integer n. According to the classical Paley–Wiener Theorem, these prop-
erties characterize the image of the Fourier transform on the space of such test
functions f .
For a real reductive group G, let Cc∞ (G, K) denote the “Hecke algebra” of K-
finite compactly supported smooth functions. Then for any f in Cc∞ (G, K) and
irreducible admissible representation π of G (on a Banach space Uπ ), the Fourier
transform of f at π is defined to be the operator
Z
π(f ) = f (x)π(x) dx.
G

The resulting function

π → π(f )
has as its domain the set of irreducible (admissible) representations of G, and for
any (π, Uπ ) takes values in the space of operators on Uπ . The purpose of the
(operator) Paley–Wiener Theorem of [A10] is to characterize which functions

π 7→ F (π)

are of the form π 7→ π(f ) for some f .

86 STEVE GELBART

More precisely, let us (following Arthur) fix a minimal parabolic subgroup B of

G with Langlands decomposition N0 A0 M01 , and consider the nonunitary principal
series of induced representations ρB (σ, Λ), indexed by quasi-characters Λ of A0 and
irreducible representations σ of M01 . By a well-known theorem of Harish-Chandra,
any π is equivalent to a subquotient of some ρB (σ, Λ). This means that π(f ) will
be completely determined by the Fourier transform

fˆ: (σ, Λ) → fˆB (σ, Λ) = ρB (σ, Λ)(f ).

What should the image of this Fourier transform be?

Arthur defines P W (G, K) to be the space of functions

F : (σ, Λ) 7→ FB (σ, Λ),

which are entire, K-finite (in a sense I shall not explain here), and satisfy the growth
condition

(1.1) sup kF (σ, Λ)ke−N k Re(Λ)k (1 + k Im(Λ)k)n < ∞,
σ,Λ

for some N (and all n). Moreover, there is another, more complicated condition for
F (σ, Λ) which comes from the various intertwining maps between principal series.
Namely, if there is a relation
m
X
Dk (ρB (σk , Λk )(g)uk , vk ) = 0
k=1

valid for all g in G, where Dk is a differential operator (acting through the variable
Λ, and uk , vk are vectors in the space of ρ(σk , Λk ), then F (σ, Λ) must also satisfy
the relation
m
X
(1.2) Dk (F (σk , Λk )uk , vk ) = 0.
k=1

(Such relations exist commonly, but are difficult to characterize explicitly; indeed,
a listing of all such relations would be tantamount to a complete knowledge of the
irreducible subquotients of the principal series.)
The main result of [A10] is that

f → fˆ(σ, Λ)

provides an algebra isomorphism from Cc∞ (G, K) onto PW(G, K). The conse-
quence of this result which is actually needed for the trace formula concerns the
notion of Paley–Wiener multipliers.
By a “multiplier” of Cc∞ (G, K) is meant a linear operator C on Cc∞ (G, K) such
that
C(f ∗ g) = C(f ) ∗ g = f ∗ C(g)
 LECTURE IX. APPLICATIONS OF PALEY–WIENER; CONCLUDING REMARKS 87

for all f , g in Cc∞ (G, K) . (The algebra of all such “multipliers” coincides with
EndU(G) (Cc∞ (G, K)). The Paley–Wiener multipliers are constructed as follows.
Set

(1.3) h = ihK ⊕ h0 ,

where hK is a fixed Cartan subalgebra of the Lie algebra of K ∩ M0 , and h0 is

the Lie algebra of the split component of M0 . (Here M0 is the Levi component
of a fixed minimal parabolic subgroup of G.) If G is the Lie algebra of G, then
hC is a Cartan subalgebra of GC , and h ⊂ hC is invariant under the Weyl group
W of (GC , hC ). Let E(h)W be the algebra of compactly supported distributions on
h which are invariant under W . Then for any γ in E(h)W , the Fourier–Laplace
transform
γ̂(ν), ν ∈ h∗C
is an entire function on h∗C which is W -invariant and satisfies a growth estimate of
the form


(1.4) sup |γ̂(ν)|e−Nγ k Re(ν)k (1 + k Im νk−nγ < M
ν∈h∗
C

for some integers Nγ and nγ . According to the result below, these γ̂ provide us
with multipliers Cγ (called Paley–Wiener multipliers).
Theorem 1.5 (see Thm. 4.2 of [A10]). For each distribution γ in E(h)W , and f
in Cc∞ (G, K), there is a unique fγ in Cc∞ (G, K) such that

(1.5) π(fγ ) = γ̂(νπ )π(f )

for any irreducible admissible representation π. (Here {νπ } is the W -orbit in h∗C
associated to the infinitesimal character of π.) In particular, if we define

Cγ (f ) = fγ ,

then the map

γ → Cγ , γ ∈ E(h)W ,
is a homomorphism from the algebra E(h)W to the algebra of multipliers

EndU(G) (Cc∞ (G, K)).

Remarks. (i) This Theorem is almost an immediate consequence of the Paley–

Wiener Theorem described earlier. Indeed, the uniqueness is clear (since (1.5)
completely determines the Fourier transform of fγ ). On the other hand, the exis-
tence of fγ results from checking that the function

FB (σ, Λ) = γ̂(νσ + Λ)fˆB (σ, Λ)

belongs to PW(G, K), and hence is the Fourier transform of some fγ ; then since
π(fγ ) is given by the action of ρ(σ, Λ)(fγ ) on the appropriate subquotient, it follows
that
π(fγ ) = γ̂(νσ + Λ)π(f ) = γ̂(νπ )π(f ),
88 STEVE GELBART

as required.
(ii) That Cγ (f ) = fγ defines a “multiplier” of Cc∞ (G, K) follows formally from
(1.5): Cγ (f ∗ g) = Cγ (f ) ∗ g, since

π((f ∗ g)γ ) = γ̂(νπ )π(f ∗ g)

= γ̂(νπ )π(f )π(g)
= π(fγ )π(g).

(iii) This Paley–Wiener theorem may be viewed as an analogue for real groups of
the p-adic results of [BDK] and [Ro3].

2. Applications of Paley–Wiener to the Calculation of Jχ (f ).. Recall how

JχT (f ) was defined (for sufficiently large T ) to be the integral over the diagonal of the
truncated kernel function ΛT2 (Kχ (x, y). Using the explicit formula for ΛT2 (Kχ (x, y)
in terms of Eisenstein series, we then wrote
XXZ
(2.1) JχT (f ) = ΨTσ (λ, f ) dλ,
P σ ia∗
P /iaG
∗

where
1
ΨTσ (λ, f ) = tr(ΩTχ,σ (P, λ)ρ(σ, λ)(f )),
n(AP )
and ΩTχ,σ (P, λ) is the operator on the space of ρ(σ, λ) defined by
Z
(ΩTχ,σ (P, λ)φ′ , φ) = ΛT E(x, φ′ , λ)ΛT E(x, φ, λ) dx.
Z(A)G(F )\G(A)

We stress that the integral formula (2.1) for JχT (f ) is absolutely convergent, and
that JχT (f ) is known to be polynomial in T . The problem is to explicitly compute
JχT (f ) for a conveniently chosen value T0 of T .
At the end of Lecture IV, we used Arthur’s general theory to derive an explicit
formula for JχT (f ) in the case of G = GL(2). The strategy consisted of three steps:
(1) plug into (2.1) Langlands’ explicit formula for ΩTχ,σ (P, λ) in terms of inter-
twining operators;
(2) use the theory of (G, M )-families to rewrite ΨTσ (λ, f ) in a form still more
amenable to integration over dλ; and
(3) compute the resulting integrals over λ (and show that some of the contri-
butions are negligible as T → ∞).
It was in this last step that we resorted to a Riemann–Lebesgue Lemma whose
application is best justified by way of Paley–Wiener multipliers; indeed these mul-
tipliers allow for the insertion of a compactly supported function B(λ) into the
formula for Jχ T , as explained in Theorem 2.3 below.
For general G, there is an additional, more fundamental reason for requiring
the Paley–Wiener Theorem in the calculation of Jχ (f ). It derives from the fact
that Langlands’ explicit formula for ΩTχ,σ (P, λ) is not valid for Eisenstein series
 LECTURE IX. APPLICATIONS OF PALEY–WIENER; CONCLUDING REMARKS 89

induced from non-cuspidal σ. (Equivalently, it is not valid when P lies outside Pχ ,

the set of parabolics attached to the collection of pairs (MB , rB ) describing the
cuspidal datum χ.) In these cases, Arthur proved “only” that the “nice” formula
for ΩTχ,σ (P, λ) is an asymptotic one for large T , valid uniformly only on compact
subsets of λ.
T
More precisely, let ωχ,σ (P, λ) denote the operator on ρ(σ, λ) defined explicitly in
terms of intertwining operators as the value at λ′ = λ of
XX ′ ′
e(t λ −tλ)(T )
MP1 |P (t, λ)−1 MP1 |P (t′ , λ′ )
θP1 (t′ λ′ − tλ)
P1 t,t′

(cf. formula (2.1) of Lecture VII for complete details). Then (cf. Corollary 9.2 of
[A9]) the difference between (ΩTχ,σ (P, λ)ρ(λ, σ)(f )φ, φ) and (the “nicer” expression)
T
(ωχ,σ (P, λ)ρ(λ, σ)(f )φ, φ) is bounded in absolute value by

(2.2) r(λ)kφk2 e−ǫkT k ,

where ǫ > 0, and r(λ) is a locally bounded function on ia∗P . Thus the integrals of
these expressions over all of ia∗P /ia∗G need not approach one another as T → ∞.
T
In short, the substitution of ωχ,σ for ΩTχ,σ in (2.1) is simply not justified.
Example. For G = GL(2), we explicitly derived Arthur’s estimate(2.2) in case
χ = {(M, µ)}, with µ = µ−1 , and P = G. Namely, in the proof of Prop. 2.7 in
Lecture IV, we computed the difference of these expressions to be M e−2(T1 −T2 ) . Of
course, in this case the problem of uniformity alluded to above is a priori absent
(the variable λ being constrained to lie in the space ia∗P /ia∗G = {0} . . . ).
In general, the way out of this problem, like the problem of showing certain terms
to be “negligible” in step (3) above, is by way of Arthur’s Paley–Wiener Theorem,
as we shall now (finally) explain.
For any B in S(ih∗ /ia∗G)W , and irreducible unitary representation σ on M (A),
define
Bσ (λ) = B(iYσ + λ),
Here Yσ is the imaginary part of the orbit {νσ } in h∗C associated to the infinitesimal
character of σ∞ (and we have fixed an embedding of any a∗P in h∗ ). Recall that
{νσ + λ} is also the orbit in h∗C associated to the infinitesimal character of the
induced representation ρ(σ∞ , λ) of G∞ .
Theorem 2.3 (see Thm. 6.3 of [A5]). Suppose B in S(ih∗ /ia∗G )W is such that
B(0) = 1. For any ǫ > 0, write B ǫ for the function B ǫ (ν) = B(ǫν). Then
XZ
JχT (f ) = lim lim ΨTσ (λ, f )Bσǫ (λ) dλ.
ǫ→0 T →∞ ia∗
P /iaG
∗
P,σ

Remarks. (i) The importance of this result is that it allows us to invoke the as-
ymptotic formula for ΩTχ,σ . Indeed, if we take B in the Theorem to be compactly
T
supported, then we can indeed substitute ωχ,σ for ΩTχ,σ and write
X Z
(2.4) JχT (f ) = lim lim T
tr(ωχ,σ (P, λ)ρ(λ, σ)(f ))Bσǫ (λ) dλ
ǫ→0 T →∞ ia∗
P /iaG
∗
P,σ
90 STEVE GELBART

(since the error term (2.2), multiplied by such a compactly supported B, will indeed
approach 0 uniformly in λ).
(ii) For large T , the expression
XZ
ΨTσ (λ, f )Bσ (λ) dλ
P,σ ia∗
P /iaG
∗

is asymptotic to a polynomial P T (B) in T . So Theorem 2.3 really says

JχT (f ) = lim P T (B ǫ ),
ǫ→0

or
XZ
JχT (f ) = lim lim ΨTσ (λ, f )Bσǫ (λ) dλ,
ǫ→0 T →∞
P,σ

where the limit in T (as in the statement of the Theorem) is interpreted as the
polynomial P T (B ǫ ) which is asymptotic to the given function as T approaches ∞.
On the other hand, since

lim Bσǫ (λ) = lim B(ǫ(iYσ + λ)) = B(0) = 1,

ǫ→0 ǫ→0

the dominated convergence theorem implies

XZ
JχT (f ) = ΨTσ (λ) dλ (for T large)
P,σ
XZ
= lim ΨTσ (λ)Bσǫ (λ) dλ (T large)
ǫ→0
P,σ
XZ
= lim lim ΨTσ (λ)Bσǫ (λ) dλ
T →∞ ǫ→0
P,σ

(with limT →∞ interpreted as above). Thus all of the work of [A5] essentially boils
down to a justification of the interchange of order of two limits (over ǫ and T )!
(iii) (Concerning the Proof of Theorem 2.3). The basic idea of the proof is
explained clearly on pp. 1258–60 of [A5]. The starting point is an application of
the Paley–Wiener multiplier Theorem 1.5 to the archimedean component of any
K-finite function f in Cc∞ (G(A)). For any γ in E(h)W , this yields a function fγ in
Cc∞ (G(A)) such that

ΨTσ (λ, fλ ) = γ̂(νσ + λ)ΨTσ (λ, f ).

Using this relation, we write

XZ
JχT (fλ ) = γ̂(νσ + λ)ΨTσ (λ, f ) dλ (for T large)
P,σ ia∗
P /iaG
∗

XZ Z 
(2.5) = ΨTσ (λ, f ) γ(H)e(νσ +λ)(H) dH dλ
P,σ ia∗
P /iaG
∗ h
Z X 
= ΨTσ (H)eνσ (H) γ(H) dH
h P,σ
 LECTURE IX. APPLICATIONS OF PALEY–WIENER; CONCLUDING REMARKS 91

where Z
ΨTσ (H) = ΨTσ (λ, f )eλ(H) dλ.
ia∗
P /iaG
∗

In particular, if we fix H in h, and set

X
γ H = |W |−1 γs−1 H
s∈W

with γs−1 H the Dirac measure at s−1 H, then (2.5) implies

XX −1
(2.6) JχT (fγ H ) = |W |−1 ΨTσ (s−1 H)eνσ (s H)

s∈W P,σ

Note here that, for each value of H, the right hand side of (2.6) is a polynomial in
T (since the left hand side is). Thus we have constructed a family of polynomials in
T , call them pT (H), whose value at H = 0 is exactly JχT (f ) (since then fγ H = f ).
Now to compute pT (H) at 0, the natural thing to do is to integrate this function
against an arbitrary Schwartz R function β(H). By the Plancherel Theorem for h,
the resulting inner product pT (H)β(H) dH can be replaced by one over ih∗ /ia∗G
involving the Fourier transform
Z
B(ν) = β(H)eν(H) dH.

In particular, JχT (f ) should be obtained by having β approximate the Dirac measure

at 0, i.e., by using
βǫ (H) = ǫ− dim(h) β(ǫ−1 H)
R
with ǫ small, and β(H) dH = 1; Equivalently, this brings into play

B ǫ (ν) = B(ǫν)

with B(0) = 1, and this is roughly how one ends up with the sought-after formula
XZ
JχT (f ) = lim ΨTσ (λ, f )Bσǫ (λ) dλ.
ǫ→0 ia∗
P /iaG
∗
P,σ

The only problem is that pT (H) is not actually tempered, and hence cannot be
integrated against any β! Thus a lengthy detour is required in Sections 4–6 of [A5],
involving additional polynomials and analysis (but nothing approaching either the
Paley–Wiener theorem, or the theorem on the polynomial nature of JχT (f )).

P
3. A Final Formula for χ JχT (f ). At the end of Lecture VII, we explained
how, for GL(2), the formula
Z ∞
JχT (f ) = lim lim tr(ωχT (t)ρ(µ, it)(f )B(ǫt) dt
ǫ→0 T →∞ −∞
92 STEVE GELBART

could be computed explicitly using (G, M ) families. Now, thanks to formula 2.4,
we have a similar strategy for computing JχT (f ) in general, and the end result is a
formula directly generalizing the explicit formula
Z ∞ Z ∞
JχT (f ) = (T1 − T2 ) tr(ρ(µ, it)(f )) dt + tr(M (−it)M ′ (it)ρ(µ, it)(f ) dt
−∞ −∞
+ 14 tr(M (0)ρ(µ, 0)(f )) + µ(f )τ (G).

(Recall that the “limit in T ” is to be interpreted as the polynomial asymptotic to

the given integral as T approaches ∞.)
N.B. In Lecture IV, we were interested only in the constant term of the polynomial
JχT (f ), i.e., JχT (f ) = Jχ0 (f ). In Arthur’s general theory, one fixes on a convenient
point T0 (which in many “classical” situations is just zero; see Lemma 1.1 of [A4]),
and sets
Jχ (f ) = JχT0 (f ).
Theorem 3.2 (see Theorem 8.2 of [A6]). Suppose f ∈ Cc∞ (Z(A)\G(A)). Then
JχT (f ) equals the sum over M ∈ L(M0 ), L ∈ L(M ), σ on M (A) (compatible with
χ), and w in W L (aM )reg of the product of

|W0M | |W0 |−1 | det(w − 1)aLM )|−1

with Z X
|P(M )|−1 tr(ML (P, λ)M (P, w)ρ(σ, λ)(f )) dλ.
ia∗
L /iaG
∗
P ∈P(M)

Remarks. (i) The exact definition of each of the objects appearing here is to be
found in Sections 4 through 8 of [A6]. But the point is that they are all more
or less familiar objects related to the intertwining operators MQ|P discussed in
Lecture VII. For GL2 , there will be only three terms in the sum above: one for
M = L = M0 , with w = e and σ = µ a character of M0 (this gives the term in (3.1)
with the logarithmic derivative); one for M = M0 ,L = G, w the non-trivial Weyl
element, and σ a character of M0 (this gives the term involving tr(M (0)ρ(µ, 0)(f )));
and one with M = L = G, w = e, and µ the character µ(det g) of G(A) (this gives
the contribution µ(f )τ (G)).
(ii) The proof of Theorem 3.2 consists of two non-trivial steps. First one uses
(G, M ) families to carry out the explicit computation of
!
XXZ
T
val lim “ lim ” ωσ,χ (λ, f )Bσ (ǫλ) dλ ,
T =T0 ǫ→0 T →∞
P σ

T
where ωσ,χ (λ, f ) is a sum of special values of functions eM (Λ) for particular (G, M )-
families e(Λ) = c(Λ)d(Λ) (involving the intertwining operators MQ|P ); this is car-
ried out in Sections 3 through 5 of [A6], and the end result is a sum of terms of the
form Z X
lim tr(ML (P, λ)M (P, w)ρ(λ, σ)Bσǫ (λ)d
ǫ→0 ia∗
L /iaG P ∈P(M)
∗
 LECTURE IX. APPLICATIONS OF PALEY–WIENER; CONCLUDING REMARKS 93

with B in Cc∞ (ih∗ /ia∗G )W equal to 1 at 0. Clearly the second step must be to show
that this (limit and) test function can be made to disappear!
Equivalently, by the dominated convergence theorem, one must show that the
sum
XZ
kML (P, λ)ρ(λ, σ)(f )k dλ
σ L /iaG
ia∗ ∗

is finite. This is carried out in Sections 6 through 8 of [A6], using normalized

intertwining operators, and estimates on the scalar functions defining these nor-
malizations. (Some assumptions are made on these normalizations at p-adic places,
but at least for GL(n) these assumptions are now proved; see [Sha1] and [Sha2]).
Finally, having reached the exalted plateau which Theorem 3.2 represents, we
should mention that it is really just the beginning! In particular it is not yet known
if the explicit spectral expansion
X X X XZ X
(3.3) Jχ (f ) = tr(M∗L (P, λ, σ, w, f ) dλ
χ χ M,L σ λ∈ia∗
L /iaG
∗
P
w

converges absolutely as a multiple integral. (The problem is that the estimates on

the normalizing factors alluded to above are not uniform in χ!).
In other words, we know that
X
|Jχ (f )| < ∞,
χ

and that Z
X
kML(P, λ)ρ(λ, σ)(f )k dλ < ∞,
Q
σ∈ (M(A))χ

but not that this last sum can be summed over all χ. This lack of knowledge
causes problems in applying the trace formula, say to base change, where on needs
absolutely convergent expressions (in order to use tricks like Langlands’ to subtract
off the non-discrete spectrum). In order to get around this problem, Arthur proves
certain estimates on the convergence of
X
|Jχ (f )|
χ

which may be viewed as a weak form of the conjectured absolute convergence of

the multiple integrals (3.3). This involves the infinitesimal characters of the infinite
components σ∞ of the representations σ appearing in the spectral expansion, and an
application of the Paley–Wiener multiplier theorem. The end result is a technique
called “separation of infinitesimal characters via multipliers” which allows one to
obtain the required equality of discrete traces from the matching of the full traces;
details appear in [A12], [AC], and [La3].
94 STEVE GELBART

4. Related Works. Here are some of the topics from Arthur’s work which we
have ignored till now.
In comparing trace formulas for different groups (for example, in proving base
change for GLn ), one wants to rewrite the basic formula
X X
(4.1) Jo (f ) = Jχ (f )
o χ

in invariant form, i.e., as an identity

X X
(4.2) I0 (f ) = Iχ (f ),
o χ

where each of the terms is an invariant distribution. The reason for this is that one
needs to assert the equality of two trace formulas for “associated” (or “matching”)
functions, given only by their orbital integrals (which are invariant distributions
. . . ). Such an invariant trace formula is obtained already in [A4], where a subtle
rearrangement and modification of the terms Jo (f ) and Jχ (f ) yields the new col-
lection of invariant distribution Io (f ) and Iχ (f ) appearing in (4.2). However, this
is only theP “coarse” form
P of the invariant trace formula, analogous to the “coarse”
formula Jo (f ) = Jχ (f ). To get the “fine” expansion of the invariant trace
formula requires the papers [A11] and [A12]
Explicit formulas for Io (f ) are obtained from the analogous fine expansions of
Jo (f ) discussed in [A13] and [A14]. (In Lecture IV we described the exact form
of J0 (f )—for unramified o—in terms of weighted orbital integrals; the extreme
opposite case—of o corresponding to {1}—is the subject matter of [A13], and the
general case—a mixture of these extremes—is handled in [A14].) On the other
hand, explicit formulas for Iχ (f ) are obtained by another delicate mixture of the
techniques of Paley–Wiener
P multipliers and (G, M )-families, somewhat analogously
to the case of Jχ (f ).
Remarks. (i) In a very recent work, Labesse has explained how one can apply the
trace formula to the base change problem for GL(n) without first putting the trace
formula in invariant form. This gives an alternate, simpler approach to [AC]; see
[La3].
(ii) Recall from our discussion of Jacquet’s relative trace formula that no truncation
operator was needed to define the “relative” distributions Jo (f ) and Jχ (f ); hence
these terms are already invariant!
(iii) Outside the scope of these “introductory” Lectures are the works of Kottwitz
and others on the stable trace formula (see [Kot2] and [La5]), and the more recent
papers of Arthur on unipotent representations, A-packets, etc. (see [A15], [A16]).

(New) References
[A15] Arthur, J., On some problems suggested by the trace formula, Lie Group Representations
II,, Lecture Notes in Mathematics, vol. 1041, Springer-Verlag, 1984, pp. 1–49.
[A16] , Unipotent automorphic representations: global motivations, in Automorphic
Forms, Shimura Varieties and L-functions, vol. 1, Academic Press, 1990, pp. 1–75.
 LECTURE IX. APPLICATIONS OF PALEY–WIENER; CONCLUDING REMARKS 95

[Kot2] Kottwitz, R., Stable trace formula: elliptic singular terms, Math. Ann. 275 (1986),
365–399; Duke Math. J. 51 (1984), 611–650.
[La5] Langlands, R. P., Les débuts d’une formule des traces stable, Publ. Math. de L’Univ.
de Paris VII, vol. 13, 1983.
[Ro3] Rogawski, J., The trace Paley–Wiener Theorem in the twisted case, Trans. Amer. Math.
Soc. 309 (1988), no. 1, 215–229.
96 STEVE GELBART

References for All Lectures

[A1] Arthur, J., The Selberg trace formula for groups of F -rank one, Ann. of Math. 100
(1974), 236–385.
[A2] , A trace formula for reductive groups, I: Terms associated to classes in G(Q),
Duke Math. J. 45 (1978), no. 4, 911–952.
[A3] , A trace formula for reductive groups, II: Applications of a truncation operator,
Compositio Math., vol. 40, Kluwer Acad. Publ., Dordrecht, 1980, pp. 87–121.
[A4] , A trace formula in invariant form, Ann. of Math. 113 (1981), 1–74.
[A5] , On a family of distributions obtained from Eisenstein series, I: Application of
the Paley–Weiner theorem, Amer. J. Math. 104 (1982), no. 6, 1243–1288.
[A6] , On a family of distributions obtained from Eisenstein series, II: Explicit for-
mulas, Amer. J. Math. 104 (1982), no. 6, 1289–1336.
[A7] , The trace formula for reductive groups, Lectures for Journées Automorphes,
Dijon, 1981.
[A8] , The characters of discrete series as orbital integrals, Invent. Math. 32 (1976),
205–261.
[A9] , On the inner product of truncated Eisenstein series, Duke. Math. J. 49 (1982),
no. 1, 35–70.
[A10] , A Paley–Wiener Theorem for real reductive groups, Acta Math. 150 (1983),
1–89.
[A11] , The invariant trace formula, I: Local theory, J. Amer. Math. Soc. 1 (1988),
no. 2, 323–383.
[A12] , The invariant trace formula, II: Global theory, J. Amer. Math. Soc. 1 (1988),
no. 3, 501–554.
[A13] , A measure on the unipotent orbit, Canad. J. Math. 37 (1985), no. 6, 1237–1274.
[A14] , On a family of distributions obtained from orbits, Canad. J. Math. 38 (1986),
no. 1, 179–214.
[A15] , On some problems suggested by the trace formula, Lie Group Representations
II, Lecture Notes in Mathematics, vol. 1041, Springer-Verlag, 1984, pp. 1–49.
[A16] , Unipotent automorphic representations: global motivations, in Automorphic
Forms, Shimura Varieties and L-functions, vol. 1, Academic Press, 1990, pp. 1–75.
[AC] Arthur, J. and Clozel, L., Simple Algebras, Base Change, and the Advanced Theory of
the Trace Formula, Ann. of Math. Stud. No. 83, Princeton University Press, Princeton,
NJ, 1975.
[BDK] Bernstein, J., Deligne, P., and Kazhdan, D, Trace Paley–Weiner theorem for reductive
p-adic groups, J. Anal. Math. 47 (1986), 180–192.
[BDKV] Bernstein, J., Deligne, P., Kazhdan, D, and Vigneras, M.-F., Representations des
groupes sur un corps local, Hermann, Paris, 1984.
[Clo] Clozel, L., Invariant harmonic analysis on the Schwartz space of a reductive p-adic
group, in Harmonic Analysis on Reductive Groups, edited by W. Barker and P. Sally,
Progress in Mathematics, vol. 101, Birkhäuser, Boston, 1991, pp. 101–121.
[CLL] Clozel, L., Labesse, J.-P., and Langlands, R. P., Morning Seminar on the Trace For-
mula,, Lecture Notes, Institute for Advanced Study, Princeton (1984).
[Ge] Gelbart, S, Automorphic Forms on Adele Groups, Ann. of Math. Stud. No. 83, Princeton
University Press, Princeton, NJ, 1975.
[GJ] Gelbart, S. and Jacquet, H., Forms of GL(2) from the analytic point of view, Proc.
Sympos. Pure Math. (Corvallis), vol. 33 Part I, Amer. Math. Soc., Providence, RI,
1979, pp. 213–251.
[GGPS] Gelfand, I. M., Graev, M., and Piatetski-Shapiro, I., Automorphic Functions and Rep-
resentation Theory (1969), Saunders.
 REFERENCES FOR ALL LECTURES 97

[GPS] Gelfand, I. M., and Piatetski-Shapiro, I., Automorphic functions and representation
theory, Trudy Moskov. Mat. Obš č. 12 (1963), 389–412; Trans. Moscow Math. Soc. 12
(1963), 438–464.
[Go] Godement, R., Domaines fondamentaux des groupes arithmetiques, Seminaire Bourbaki
257, W. A. Benjamin, Inc., New York, 1963.
[Go2] , The Spectral Decomposition of Cusp Forms, Proc. Sympos. Pure Math., vol. IX,
American Mathematical Society, Providence, RI, 1966, pp. 225–234.
[Ja1] Jacquet, H., The continuous spectrum of the relative trace formula for GL(3) over a
quadratic extension, Israel Journal of Math. 89, 1–59.
[Ja2] , Relative Kloosterman integrals for GL(3), II, Canad. J. Math. 44 (1992), 1220–
1240.
[JLai] Jacquet, H., and Lai, J., A relative trace formula, Comp. Math. 54 (1985), 243–310.
[JLR] Jacquet, H., Lai, K., and Rallis, S., A trace formula for symmetric spaces, Duke Math.
J. 70 (1993), no. 2, 305–372.
[JL] Jacquet, H. and Langlands, R. P., Automorphic Forms on GL(2), Lecture Notes in
Mathematics, vol. 114, Springer-Verlag, New York, 1970.
[JaYe] Jacquet, H., and Ye, Y., Une remarque sur le changement de base quadratic, C.R.A.S.
Paris 311 (1990), 671–676.
[Kot] Kottwitz, R., Tamagawa Numbers, Ann. of Math. 127 (1988), 629–646.
[Kot2] Kottwitz, R., Stable trace formula: elliptic singular terms, Math. Ann. 275 (1986),
365–399; Duke Math. J. 51 (1984), 611–650.
[Lai] Lai, K. F., Tamagawa Numbers of Reductive Algebraic Groups, Compositio Math.,
vol. 41, Kluwer Acad. Publ., Dordrecht, 1980, pp. 153–188.
[Lab1] Labesse, J.-P., La formules des traces D’Arthur-Selberg, Seminaire Bourbaki, exposé
636, November 1984.
[Lab2] , The present state of the trace formula, Automorphic Forms, Shimura Varieties,
and L-functions (L. Clozel and J. Milne, eds.), Perspectives in Math., vol. 10, Academic
Press, Princeton, NJ, 1990, pp. 211–226.
[Lab3] , Non-invariant base change identitites, Preprint (Fall 1994).
[La1] Langlands, R. P., On the functional equations satisfied by Eisenstein series, Lecture
Notes in Mathematics, vol. 544, Springer-Verlag, New York, 1976.
[La2] , Base Change for GL(2), Ann. of Math. Stud., vol. 96, Princeton University
Press, Princeton, NJ, 1980.
[La3] , The volume of the fundamental domain for some arithmetical subgroups of
Chevalley groups, Proc. Sympos. Pure Math., vol. 9, Amer. Math. Soc., Providence, RI,
1966, pp. 143–148.
[La4] , Eisenstein series, the trace formula, and the modern theory of automorphic
forms, Number Theory, Trace Formula and Discrete Groups (Selberg Volume), Aca-
demic Press, 1989.
[La5] , Les débuts d’une formule des traces stable, Publ. Math. de L’Univ. de Paris
VII, vol. 13, 1983.
[Mao] Mao, Z., Relative Kloosterman integrals for GL(3), III, Canad. J. Math. 45 (1993),
no. 6, 1211–1230.
[MW] Moeglin, C., and Waldspurger, J.-L., Décomposition Spectrale et Séries d’Eisenstein,
Progress in Math. Series, vol. 113, Birkhäuser Verlag, 1994.
[Ro1] Rogawski, J., Representations of GL(n) and Division Algebras on a p-adic Field, Duke.
Math. J. 50 (1983), no. 1, 161–196.
[Ro2] , Automorphic Representations of Unitary Groups in Three Variables, Ann. of
Math. Stud., vol. 123, Princeton University Press, Princeton, NJ.
[Ro3] Rogawski, J., The trace Paley–Wiener Theorem in the twisted case, Trans.
Amer. Math. Soc. 309 (1988), no. 1, 215–229.
98 STEVE GELBART

[Se] Selberg, A., Harmonic analysis and discontinuous groups in weakly symmetric Rie-
mannian spaces, with applications to Dirichlet series, J. Indian Math. Soc. 20 (1956),
47–87; see also, Proc. Internat. Cong. Math. (1962), 177–189.
[Sha] Shokranian, S., The Selberg-Arthur Trace Formula, Lecture Notes in Mathematics,
vol. 1503, Springer-Verlag, New York, 1992.
[Sha1] Shahidi, F., Fourier transforms of intertwining operators and Plancherel measures for
GL(n), Amer. J. Math. 106 (1984), 67–111.
[Sha2] , Local coefficients and normalization of intertwining operators for GL(n), Com-
positio Math. 39 (1983), 271–295.
[Ye] Ye, Y., Kloosterman integrals and base change for GL(2), J. Reine Angew. Math. 400
(1989), 57–121.

Stephen Gelbart
Weizmann Institute of Science
Theoretical Math
Rehovot 76100, Israel
[email protected]