Analysis on arithmetic quotients Introduction
Bill Casselman
University of British Columbia [email protected]
This course will be interested in the description, in analytic terms, of various spaces of functions on Γ \G where G is the group of real points on a reductive group defined over Q, and Γ is an arithmetic subgroup.
What do I mean by ‘analytic terms’? Well, the ultimate goal in the theory of automorphic forms is to derive applications to number theory, particularly those relating to a wide range of L-functions. There are many applications of this kind that depend ultimately on sophisticated analysis, and indeed can often be formulated only given a number of results in analysis. I shall occasionally mention these, but for the most part I shall not say much about them, even when the topic at hand is relevant. You can already discern my bias if you compare my account with most others in the recent literature, which are largely concerned not with arithmetic quotients Γ \G but with adelic quotients G(Q)\G(A), where G(Q) is the Q-rational group implicitly determining Γ, and A is the ring of rational adeles.
This is no loss, at least as far as my purposes are concerned, since all the hard analysis on adelic quotients reduces ultimately to hard analysis on arithmetic quotients. Further, it seesm to me that working directly on adele groups often obscures the essential nature of arguments. Restricting to arithmetic quotients should make the basic problems of analysis, as well as their solutions, clearer.
The course will come in two parts. In the first, I’ll explain what happens for SL
2, where almost all important phenomena already appear, but in much simplified fashion. Here I’ll give complete proofs.
In the second part, I’ll discuss groups of higher rank, but with necessarily less complete coverage.
Necessarily, because the approach I propose is largely still conjectural.
In the body of this introduction, I’ll sketch without details—and in particular without proofs—what happens for SL
2, remarking from time to time on what happens more generally.
One topic I’ll not deal with, at least not on the first pass, is the Arthur-Selberg trace formula.
Contents
1. Geometry
2. Compact quotients 3. SL(2,Z)
4. The Fourier expansion
5. The topological vector spaces of interest 6. Cusp forms
7. A model continuous spectrum 8. The construction of Eisenstein series 9. Truncation
10. Wave packets
11. Fourier-Eisenstein transform 12. The volume formula
13. Automorphic forms and de Rham cohomology
14. References
1. Geometry Let
G = SL
2(R) K = SO
2H = {z |
IM(z) > 0}
Γ = an arithmetic subgroup
• The group SL
2(C) acts on P
1(C) = C ∪ {∞} by M¨obius transformations. If
g =
a b c d
then
a b c d
: z 7−→ az + b cz + d . If g is real, then
IM
g(z)
=
IM(z)
|cz + d|
2,
so G takes H to itself. The isotropy group of i is K, so H may be identified with G/K.
• The metric
ds
2= dx
2+ dy
2y
2is G-invariant. The corresponding invariant area is
dx dy y
2.
The corresponding invariant Laplacian, acting on functions on H, is
∆ = ∆
H= y
2∂
2∂x
2+ ∂
2∂y
2.
• Functions on Γ\H may be identified with functions on Γ\G invariant on the right by K.
Since
dg(z)
dz = dz
(cz + d)
2g =
a b c d
the group G acts on holomorphic forms f (z) dz
n/2according to the formula (in Shimura’s notation) f (z) 7−→ f | [g
−1]
nwhere
f | [g]
n(z) = f g(z)
j(g, z)
nj(g, z) = cz + d .
Since j(gh, z) = j(g, h(z))j(h, z) these forms lift covariantly to functions F : G → C such that
F (gk) = ε(k)
−nF (g) .
where
ε: k =
c −s
s c
7−→ ε(k) = j(k, i) = c + is . Explicitly, the correspondance sets
F (g) = f | [g]
n(i) .
• Elements of the universal enveloping algebra U(g) act on functions on G/K, f 7→ R
Xf . The centre Z(g) of the enveloping algebra is a polynomial algebra in the Casimir element C. The right action of C on functions on G fixed by K is the same as the left action of ∆ on functions on H.
A basis of sl
2is
α =
1 0 0 −1
κ =
0 −1
1 0
σ =
0 1
1 0
. Since
C = α
24 + σ
22 − κ
22 , the element Ω = C
g+ κ
2is an elliptic operator in U (g).
• Any function in L
2(Γ\G) may be considered a distribution. The elliptic operator R
Ωwith domain C
c∞(Γ\G) is esentially self-adjoint. The domain of its closure is the same as the F in L
2(Γ\G) such that R
ΩF , considered as a distribution, is again in L
2(Γ\G). The domain of any power Ω
nis the same as the intersection of the domains of all X in U
2n(g).
• On G define the norm
kgk = sup δ(a), δ
−1(a) if
g = k
1ak
2. Since kgk = sup
kvk=1kg(v)k
kghk ≤ kgk khk .
2. Compact quotients
Suppose for the moment that Γ \G is compact. For example, Γ could be the group of units of norm 1 in a quaternion algebra over Q, or more generally in a quaternion algebra B over a totally real number field, such that B ⊗
FR ∼ = M
2(R) for exactly one real embedding of F into R.
There are two related results of interest: (1) The unitary representation of G on L
2(Γ\G) decomposes into an orthogonal direct sum of irreducible representations of G, each with finite multiplicity. (2) On the subspace of L
2(Γ\G) of functions transforming with respect to the right action of K by ε
n, the operator Ω decomposes into an orthogonal direct sum of eigenspaces, with a discrete set of eigenvalues passing off to ∞.
There is more that can be said, for example about the asymptotic distribution of eigenvalues, but it will
not be said here.
3. SL
2( Z )
More interesting is the case in which Γ \G is not compact. As I have already said, I’ll look here at only one example, but one for which most interesting phenomena already appear. Let
Γ = SL
2(Z)
P = group of upper triangular matrices A = group of diagonal matrices
|A| = connected component of A
N = group of upper triangular unipotent matrices δ = the modulus character of P : p 7→
detAd
n(p) H
Y= {z = x + iy | y ≥ Y }
Γ
P= Γ ∩ P Thus
δ:
t 0 0 1/t
7−→ t
2.
The region
D = {z ∈ H | |z| ≥ 1, |
RE(z)| ≤ 1/2}
is a fundamental domain for Γ.
i
1 /2 + i √ 3 /2
This region has finite area, less than Z
∞√3/2
dx dy y
2. In this region
kgk ≍ inf
γ∈Γ
kγgk ≍ y .
• For Y ≥ 1 the canonical map
Γ
P\H
Y→ Γ\H
is an embedding. Its image is called a parabolic domain of Γ \H. In other words, the quotient Γ\H is the
union of a parabolic domain, with simple geometry, and a compact part which is more complicated.
4. The Fourier expansion
If F is any continuous function on Γ \H then it is invariant under horizontal translatiosn z 7→ z + n. We may define its Fourier series
X
Z
F
n(z)e
2πinxwhere
F
n(z) = Z
10
F (x + z)e
−2πinxdx .
Since H modulo horizontal translations may be identified with iR
≥0, this may be identified with a function F
n(y) on (0, ∞). If F is smooth enough, this series will converge.
For us, the most important term will be the constant term F
0, although the other terms are not without interest. For example, if F (z) is a classical automorphic form one can express it as a converging analytic series
F (z) = c
0+ c
1e
2πz+ c
2e
4πz+ c
3e
6πz+ · · · = X
c
ne
−2πnye
2πnix, with F
n(y) = c
ne
−2πny, and the coefficients c
nare of great interest.
If F is a continous function on Γ \G one may define its ‘constant term’
F
0(g) = Z
Γ∩N\N
F (ng) dn
which will be a function on N Γ
P\G. The Iwasawa decomposition G = N |A| K tells us that this is in bijection with |A| × ({±I}\K).
In fact, the constant term can be defined for any distribution on Γ\G. If f is a function on Γ
PN \G, we can define formally the ‘Eisenstein series’
E
f(g) = X
ΓP\Γ
f (γg) .
If this converges, it will be invariant Γ. In particular, if f is fixed by K it may be identified with a function f (y) on N \H ∼ = {iy | y > 0}, and then
E
f(z) = X
ΓP\Γ
f
IM(γz) .
There are many criteria for the convergence of this series. The basic feature in all of them is that f (nak) = f (ak) must grow small as δ(a) → 0 (or, in the special case, that f(y) grow small as y does).
For now we’ll just need this one:
Lemma 4.1. For ϕ a continuous function on Γ
PN \H with compact support on Γ\G, the series
[ephi]
E
ϕ(z) = X
ΓP\Γ
ϕ(γz)
converges absolutely to a continuous function of compact support on Γ\H. It is a right G- and g- covariant map from C
c∞(Γ
PN \G) to C
c∞(Γ\G).
The point of all this is that the constant term and Eisenstein series are closely related:
Lemma 4.2. Assume ϕ to be a continuous function on Γ
PN \G. If E
ϕconverges absolutely and uniformly
[evarphi2]
on compact subsets of H, then for F continuous on Γ\H
hF
0, ϕi
ΓP\H= hF, E
ϕi
Γ\H.
As a result, we see that it is consistent to define Φ
0for any distribution Φ on Γ \H, and in particular for Φ in L
2, as a distribution on Γ
P\H by the formula
hΦ
0, ϕi
ΓP\H= hΦ, E
ϕi
Γ\H.
5. The topological vector spaces of interest
There are several topological vector spaces of interest when one is doing analysis on Γ \G. In one group are these:
C
c∞(Γ\G) C
c−∞(Γ\G) S(Γ\G) A(Γ\G) A
mg(Γ\G) A
umg(Γ\G) A(Γ\G) and in another are these:
L
2(Γ\G) L
2,∞(Γ\G) C(Γ\G) C(Γ\G) b
The difference between the two groups is that the first are largely involved with distributions, the second just with square-integrable functions. Expressed in other terms, the first are concerned with arbitrary continuous representations of G, the second only with unitary ones.
The topologies themselves are not likely to appear in an important role, except for the unitary represen- tations on Hilbert spaces.
• C
c∞(Γ\G)
As a topological vector space, this is soemwhat complicated. If U is a relatively compact open subset of Γ\G one has on C
c∞(U ) the semi-norms
kfk
n= sup
g∈U,X∈Un(g)
R
Xf (g) .
The space C
c∞itself is the direct limit of these spaces. What is important is that a linear functional on the direct limit is continuous if and only if it is continuous on each C
c∞(U ).
• C
c−∞(Γ\G)
This is the space of distributions on Γ \G, the continuous linear functions on C
c∞(Γ\G).
• S(Γ\G)
This I call the Schwartz space of Γ \G, that of all smooth functions f such that all R
Xf vanish more rapidly at infinity than any 1/y
N. The topology is defined by semi-norms
kfk
m,n= sup
g∈Γ\G,X∈Un(g)
kgk
mR
Xf (g) .
• A(Γ\G)
This is the continuous dual of S, the space of moderate distributions .
• A
mg(Γ\G)
Functions F such that every R
XF is of moderate growth.
• A
umg(Γ\G)
Functions F of uniform moderate growth —there exists one N such that all R
XF are O(kgk
N). For a fixed N the space of F with F = O( kgk
N) is a Banach space, and this is the subspace of its smooth vectors.
• A(Γ\G)
The space of automorphic forms on Γ \G—those smooth functions of moderate growth contained in a finite-dimensional K-stable subspace and annihilated by some non-zero polynomial P (C). Classical holomorphic automorphic forms are automorphic forms, precisely since they are required to have Laurent expansions
f (z) = X
n
c
ne
2πnizwith c
n= 0 for n < 0.
• L
2(Γ\G)
Square-integrable functions—a Hilbert space.
• L
2,∞(Γ\G)
Functions F in L
2such that all R
XF , considered as distributions, are also in L
2. It is contained in A
umg(Γ\G).
• C(Γ\G)
This I call the Harish-Chandra space of Γ \G, that of all smooth functions f such that all R
Xf vanish more rapidly at infinity on D than any y
1/2/ log
Ny. These functions are just barely square-integrable.
The topology on this space is defined by semi-norms kfk
Cm,n= sup
g∈Γ\G,X∈Un(g)
log
mkgk·
R
Xf (g) .
• b C(Γ\G)
The continuous linear dual of C.
6. Cusp forms
There is one fundamental fact about the constant term—the difference F − F
0generally decreases as y → ∞. There are various technical versions of this, but one of the useful is:
Proposition 6.1. If F lies in A
umgthen every R
XF − R
XF
0is rapidly vanishing at ∞.
[vanishingf0]
This is certainly the case for classical automorphic forms F , for which F − F
0= O(e
−2πy).
A distribution of moderate growth is called cuspidal if its constant term vanishes, or equivalently if and only if it annihilates all E
ϕ.
Let L
2cuspbe the subspace of cuspidal functions in L
2(Γ\G). By definition, we can write L
2(Γ\G) = L
2cusp⊕ L
2Eiswhere L
2Eisis the closure in L
2of the subspace spanned by the E
ϕ. This also gives us L
2,∞(Γ\G) = L
2,∞cusp⊕ L
2,∞Eis.
But L
2,∞is containd in A
umg, so
L
2,∞cusp(Γ\G) ⊆ S(Γ\G) . As a consequence we can now also write
S = S
cusp⊕ S
Eiswhere S is the closure in S of the span of the E
ϕ.
The spaces L
2cuspand S
cusplook just like the corresponding spaces for a compact quotient. I’ll say no more about them here, although in the course itself I’ll prove these statements. The complements L
2Eisand S
Eisare more interesting, from an analyst’s point of view. From a number theorist’s point of view the cusp forms are certainly more interesting, but it turns out that in order to understand them one has also to deal with the complement.
7. A model continuous spectrum
The constant functions lie in L
2Eissince they are certainly orthogonal to cusp forms. The complement of the constants in L
2Eisis a continuous integral of unitary representations. Before I explain exactly what happens, I’ll look in this section at a simple example.
For the moment, let G = R
≥0, the multiplicative group of positive real numbers, and Γ = {1}. Let D be the multiplicatively invarant derivative xd/dx. Here too we can define spaces that will be of interest:
• C(0, ∞)
This is the space of all f in C
∞(0, ∞) such that all D
mf decrease more rapidly than any 1/| log
Nx| near 0 and ∞.
The map x 7→ log x identifies the group (0, ∞) with R, and this space is isomorphic to the additive Schwartz space S(R). The Fourier transform
f 7−→ b f (s) = Z
∞0
f (x)x
−sdx x is an isomorphism of this with S(iR). The inverse map is
f (x) = 1 2πi
Z
RE(s)=0
F (s)x
sds .
• b C(0, ∞)
This the space dual to C(0, ∞), that of tempered distributions.
• S(0, ∞)
This is the space of all C
∞(0, ∞) such that all D
mf decrease more rapidly than any x
−Nat ∞ and more rapidly than any x
Nat 0. It looks more like C
c∞(0, ∞) than C(0, ∞). The Fourier transform b f is holomorphic, characterized by the condition that for all N and σ > 0 there exists C
Nsuch that
F (s)
≤ C
N(1 + |s|)
Nin the region
RE(s) ≤ σ.
• A(0, ∞)
This is the space of all functions on (0, ∞) annihilated by some polynomial P (D). It has as basis the functions x
slog
nx.
• L
2,∞(0, ∞)
One can verify that for f in C Z
∞0
f(x)
2dx = 1 2πi
Z
RE(s)=0
b f (s)
2ds
Therefore isomorphism of C(0, ∞) with C(iR) extnds to one of L
2(0, ∞) with L
2(iR).
8. The construction of Eisenstein series
Again let Γ = SL
2(Z). I want to describe roughly what the continuous spectrum of Γ\G looks like. But to simplify things slightly, I’ll look just at Γ \H.
The starting point is that if F is an eigenfunction of ∆ on H, then F
0is one of ∆ on Γ
PN \H. So we then have
y
2∂
2F
0∂y
2= D
2F − DF = λF ,
so that F
0is an automorphic form of the kind we have seen in the previous section. The solutions are of the form y
s, with λ = λ
s= s(s − 1). It seems natural to try constructing eigenfunctions of ∆ by means of Eisenstein series. Since ∆ commutes with Γ, the Eisenstein series
E
s(z) = X
ΓP\Γ
IM
(γz)
s,
if it converges, ought also to be an eigenfunction of ∆ with eigenvalue s(s − 1). It doesn’t always
converge, but it does extend meromorphically to values of s throughout C. Everyone has his own
favourite way to carry out this meromorphic extension—mine takes place in three steps described by
this diagram:
real part of s = 1/ 2
s = 1
CONVERGENCE.
The function y
s=
IM(z)
sis an eigenfunction for ∆ on H, with eigenvalue s(s − 1).
invariant under N .
Proposition 8.1. The series converges absolutely and uniformly on compact subsets of H to an eigen-
[es-convergence]
function of ∆ that satisfies the condition that
E
s− y
s= O(y
1−σ) for y ≫ 0, with σ =
RE(s).
The constant term of E
ssatisfies an ordinary differential equation that tells us it is equal to A(s)y
s+ B(s)y
1−sfor some holomorphic functions A, B. It turns out that A(s) = 1, and that in fact E
s∼ y
s+ ξ(2s − 1)
ξ(2s) y
1−swhere ξ(s) = π
−s/2Γ(s/2)ζ(s).
RESOLVANT.
For s in the striped region
RE
(s) > 1/2, s / ∈ (1/2, 1]
the value of λ
s= s(s − 1) does not lie in (−∞, 0], which includes the spectrum of the closure of ∆. The resolvant (∆ − λ
sI)
−1is therefore a bounded function on L
2in this region. We use this, roughly to find a function E
swhich is asymptotic to y
sin this region in the sense that E
s− y
sis square-integrable for y ≫ 0. It is therefore a continuation of E
s.
THE CRITICAL LINE.
For the final step, there are two competing techniques, one (from [Colin de Verdi`ere:1972] usng the Friedrichs extension, the other (from [Jacquet:1981]) using integral operators.
Neither one seems definitive, but I know no better.
In the end, one obtains not only meromorphic extension, but a functional equation E
s= c(s)E
1−s.
The function c(s) satisfies c(s)c(1 − s) = 1, and hence c(s)
= 1 on the critical line
RE(s) = 1/2.
Explain the functional equation for general E
s, using intertwining operators for principal series, talk
about Γ(s) factors.
9. Truncation
In constructing Eisenstein series we have seen many variations on the theme of cutting off a function in the neighbourhood of infinity. For Y ≥ 1 define the truncation of Φ informally by the condition that over D, with z the projection of g onto H:
Λ
YΦ(g) =
Φ(g) − Φ
0(g) if
IM(z) ≥ Y
Φ(g) otherwise.
A little more formally, define
C
YΦ(g) = E
χ[Y,∞)Φ
0(g)
(for any fixed g the series defining this has only a finite number of terms) and Λ
YΦ = Φ − C
YΦ .
If Φ lies in A
umgthen Λ
YΦ is rapidly decreasing at infinity, and in particular square-integrable. The decomposition
Φ = Λ
YΦ + C
YΦ
is orthogonal whenever that makes sense. The Maass-Selberg relation expresses hΛ
YE
s, Λ
YE
ti
in an explicit formula. (This is used to prove that E
shad no poles on the critical line. There is no circular reasoning, I promise.)
10. Wave packets
For f in C
c(1/2 + iR) the function
E
f= 1 2πi
Z
RE(s)=1/2
f (s)E
sds
lies in C, and hence may be paired with E
−t. This is proved by expressing E
sas E
s= Λ
YE
s+ C
YE
sSince it lies in L
2,∞, it is in A
umg.
11. Fourier-Eisenstein transform
For f in S(Γ\H) its Fourier-Eisenstein transform is the meromorphic function f (s) = hE b
−s, f i
Γ\H.
There are some obvious properties this function F (s) satisfies:
(a) It is meromorphic in C, holomorphic in
RE(s) ≤ 0 except for a simple pole at s = −1;
(b) it satisfies the functional equation F (s) = c( −s)F (−s);
and some more subtle ones:
(c) it is L
2on
RE(s) = −1/2, along with all s
nF (s);
(d) in a region σ
0< σ =
RE(s)h1/2,
IM(s)iε it is bounded up to constant by all 1
σ
kIM(s)
NThere exists a map back from such F (s) to S, inducing an isomorphism with S
Eis. Part of the inverse map is an integral
1 2πi
Z
RE(s)=1/2