Professional Documents
Culture Documents
Spectral Theory
of
Automorphic Forms
Henryk Iwaniec
1995
El proyecto que hace ya diez años puso en marcha la Revista
Matemática Iberoamericana, incluı́a la publicación esporádica de
monografı́as sobre temas de gran interés y actualidad en áreas cuya
actividad hiciera aconsejable una recapitulación llevada a cabo por uno
de sus artistas importantes.
Pretendemos que estas monografı́as de la Biblioteca de la Re-
vista Matemática Iberoamericana puedan servir de guı́a a aquellos
que no siendo especialistas deseen explorar territorios de matemáticas
en parte consolidados, pero vivos y con mucho por descubrir y entender.
Es nuestro propósito ofrecer verdaderas expediciones desde el confor-
table hogar de las matemáticas que todos compartimos hasta la terra
incognita en los confines del firmamento lejano, donde las ideas estan
en continua ebullición.
Para iniciar esta andadura hemos tenido la gran fortuna de poder
contar con la presente monografı́a que sobre formas modulares y su
teorı́a espectral ha escrito el profesor Henryk Iwaniec. Como directores
de la Revista Matemática Iberoamericana queremos agradecerle el
entusiasmo que desde un principio mostró en este empeño, y el cuidado
exquisito que ha puesto en su elaboración.
Ha sido una suerte poder contar con el magnı́fico trabajo de com-
posición y maquetación de Domingo Pestana. Su dedicación y buen
hacer han sido una ayuda inestimable.
Preface xiii
Introduction 1
Chapter 0
Harmonic analysis on the euclidean plane 3
Chapter 1
Harmonic analysis on the hyperbolic plane 7
1.1. The upper half plane 7
1.2. H as a homogeneous space 12
1.3. The geodesic polar coordinates 16
1.4. Bruhat decomposition 18
1.5. The classification of motions 18
1.6. The Laplace operator 20
1.7. Eigenfunctions of ∆ 21
1.8. The invariant integral operators 28
1.9. The Green function on H 35
Chapter 2
Fuchsian groups 39
2.1. Definitions 39
2.2. Fundamental domains 41
2.3. Basic examples 44
2.4. The double coset decomposition 48
2.5. Kloosterman sums 51
2.6. Basic estimates 53
IX
X Contents
Chapter 3
Automorphic forms 57
3.1. Introduction 57
3.2. The Eisenstein series 61
3.3. Cusp forms 63
3.4. Fourier expansion of the Eisenstein series 65
Chapter 4
The spectral theorem. Discrete part 69
4.1. The automorphic Laplacian 69
4.2. Invariant integral operators on C(Γ\H) 70
4.3. Spectral resolution of ∆ in C(Γ\H) 75
Chapter 5
The automorphic Green function 77
5.1. Introduction 77
5.2. The Fourier expansion 78
5.3. An estimate for the automorphic Green function 81
5.4. Evaluation of some integrals 83
Chapter 6
Analytic continuation of the Eisenstein series 87
6.1. The Fredholm equation for the Eisenstein series 87
6.2. The analytic continuation of Ea (z, s) 90
6.3. The functional equations 93
6.4. Poles and residues of the Eisenstein series 95
Chapter 7
The spectral theorem. Continuous part 103
7.1. The Eisenstein transform 104
7.2. Bessel’s inequality 107
7.3. Spectral decomposition of E(Γ\H) 110
7.4. Spectral expansion of automorphic kernels 113
Contents XI
Chapter 8
Estimates for the Fourier coefficients of Maass forms 117
8.1. Introduction 117
8.2. The Rankin-Selberg convolution 119
8.3. Bounds for linear forms 121
8.4. Spectral mean-value estimates 123
8.5. The case of congruence groups 126
Chapter 9
Spectral theory of Kloosterman sums 133
9.1. Introduction 133
9.2. Analytic continuation of Zs (m, n) 134
9.3. Bruggeman-Kuznetsov formula 138
9.4. Bruggeman-Kuznetsov formula reversed 141
9.5. Petersson’s formulas 144
Chapter 10
The trace formula 149
10.1. Introduction 149
10.2. Computing the spectral trace 154
10.3. Computing the trace for parabolic classes 157
10.4. Computing the trace for the identity motion 161
10.5. Computing the trace for hyperbolic classes 161
10.6. Computing the trace for elliptic classes 163
10.7. Trace formulas 166
10.8. The Selberg zeta-function 168
10.9. Asymptotic law for the length of closed geodesics 170
Chapter 11
The distribution of eigenvalues 173
11.1. Weyl’s law 173
11.2. The residual spectrum and the scattering matrix 179
11.3. Small eigenvalues 181
11.4. Density theorems 185
XII Contents
Chapter 12
Hyperbolic lattice-point problems 189
Chapter 13
Spectral bounds for cusp forms 195
13.1. Introduction 195
13.2. Standard bounds 196
13.3. Applying the Hecke operator 198
13.4. Constructing an amplifier 200
13.5. The unique ergodicity conjecture 202
Appendix A
Classical analysis 205
A.1. Self-adjoint operators 205
A.2. Matrix analysis 208
A.3. The Hilbert-Schmidt integral operators 209
A.4. The Fredholm integral operators 210
A.5. Green function of a differential equation 215
Appendix B
Special functions 219
B.1. The gamma function 219
B.2. The hypergeometric functions 221
B.3. The Legendre functions 223
B.4. The Bessel functions 224
B.5. Inversion formulas 228
References 233
∗
The author would like to thank the participants and the Mathematics Department
for their warm hospitality and support.
XIII
Introduction
e(z) = e2πiz .
1
2 Introduction
operators, the Hecke operators. We shall get into this territory only
briefly in Sections 8.4 and 13.3 to demonstrate its tremendous potential.
Many important topics rest beyond the scope of these lectures; for
instance, the theory of automorphic L-functions is missed entirely.
A few traditional applications are included without straining for
the best results. For more recent applications the reader is advised to
see the original sources (see the surveys [Iw 1, 2] and the book [Sa 3]).
There is no dearth of books on spectral aspects of automorphic
functions, but none covers and treats in detail as much as the expansive
volumes by Dennis Hejhal [He1]. I recommend them to anyone who is
concerned with doing reliable research. In these books one also finds a
very comprehensive bibliography. Those who wish to learn about the
theory of automorphic forms on other symmetric spaces in addition to
the hyperbolic plane should read Audrey Terras [Te]. A broad survey
with emphasis on new developments is given by A. B. Venkov [Ve].
Chapter 0
Harmonic analysis
on the euclidean plane
R2 = (x, y) : x, y ∈ R .
ds2 = dx2 + dy 2
∂2 ∂2
D= + .
∂x2 ∂y 2
are eigenfunctions of D;
3
4 Harmonic analysis on the euclidean plane
Using polar coordinates one finds that the Fourier transform is also
radially symmetric, more precisely,
Z +∞ √
k̂(u, v) = π k(r) J0 ( λr) dr ,
0
where X
K(z, w) = k(z + p, w) ,
p∈Z2
where r(`) denotes the number of ways to write ` as the sum of two
squares,
r(`) = # (m, n) ∈ Z2 : m2 + n2 = ` ,
T
+ O(T 1/3 ) .
# ϕ : λ(ϕ) ≤ T =
4π
In view of the above connection the Gauss circle problem becomes the
Weyl law for the operator D (see Chapter 12).
Chapter 1
Harmonic analysis
on the hyperbolic plane
H = z = x + iy : x ∈ R, y ∈ R+ .
|z − w| + |z − w|
(1.2) ρ(z, w) = log .
|z − w| − |z − w|
7
8 Harmonic analysis on the hyperbolic plane
where
|z − w|2
(1.4) u(z, w) = .
4 Im z Im w
To describe the geometry of H we shall use well-known properties
of the Möbius transformations
az + b
(1.5) gz = , a, b, c, d ∈ R , ad − bc = 1 .
cz + d
Observe that a Möbius transformation g determines the matrix ac db
z−w
(1.6) gz − gw = .
(cz + d)(cw + d)
In particular this shows that
|gz − gw| = |z − w| ,
if both points are on the curve
(1.7) Cg = z ∈ C : |cz + d| = 1 .
If c 6= 0 this is a circle centered at −d/c of radius |c|−1 . Since Cg is
the locus of points z such that the line element at z is not altered in
euclidean length by the motion g, g acts on Cg as an euclidean isometry.
Naturally Cg is called the isometric circle of g. By (1.6) we get
d
(1.8) gz = (cz + d)−2
dz
−2
so we call |cz + d| the deformation of g at z. In this language the
interior of Cg consists of points with deformation greater than 1 and
the exterior consists of points with deformation less than 1. Note that
g maps Cg to Cg−1 and reverses the interior of Cg onto the exterior of
Cg−1 .
The upper half-plane 9
∗ ∗
For g = c d ∈ G we introduce the function
(1.9) jg (z) = cz + d .
that the complex plane Ĉ splits into three G-invariant subspaces, name-
ly H, H (the lower half-plane) and R̂ = R ∪ {∞} (the real line, the
common boundary of H and H). Moreover we have
Figure 2. Geodesics in H .
(1.13) π−α−β−γ.
4πA − K A2 ≤ L2 ,
(1.14) 4πA + A2 ≤ L2 ;
G = N AK .
g = nak, n ∈ N , a ∈ A, k ∈ K .
To see this first use a, k to make a matrix with a given lower row,
a ∗ ∗ ∗ ∗ ∗
= ,
a−1 − sin θ cos θ γ δ
y 1/2
1 x
n = n(x) = , a = a(y) = .
1 y −1/2
14 Harmonic analysis on the hyperbolic plane
The pair (r, ϕ) is called the geodesic polar coordinates of the point z.
These are related to the rectangular coordinates by k(ϕ) e−r i = x + iy,
i.e.
The length element and the measure are expressed in the new coordi-
nates as follows:
(1.17) dµz = 4 du dϕ .
18 Harmonic analysis on the hyperbolic plane
has one or
p two fixed points in Ĉ, they are: b/(d − a) if c = 0 and
(a − d ± (a + d)2 − 4)/2c if c 6= 0. Hence there are three cases:
(i) g has one fixed point on R̂.
(ii) g has two distinct fixed points on R̂.
(iii) g has a fixed point in H and its complex conjugate in H.
Accordingly g is said to be parabolic, hyperbolic, elliptic, and naturally
this classification applies to the conjugacy classes. Every conjugacy
class {g} contains a representative in one of the groups N , A, K. The
elements of ±N , ±A, K other that ±1 are parabolic, hyperbolic and
elliptic respectively. They act on H simply as follows
(i) z 7→ z + v , v ∈ R, (translation, fixed point ∞),
+
(ii) z 7→ p z , p∈R , (dilation, fixed points 0, ∞),
(iii) z 7→ k(θ) z , θ ∈ R, (rotation, fixed point i).
A parabolic motion has infinite order, it moves points along horo-
cycles (circles in H tangent to R̂).
A hyperbolic motion has infinite order too; it moves points along
hypercycles (the segments in H of circles in Ĉ through the fixed points on
R̂). The geodesic line through the fixed points of a hyperbolic motion
g is mapped to itself, not identically. Of the two fixed points one is
repelling, the other is attracting. The dilation factor p is called the
norm of g. For any z on this geodesic | log p| is the hyperbolic distance
between z and gz.
An elliptic motion may have finite or infinite order; it moves points
along circles centered at its fixed point in H.
where ∂/∂z = (∂/∂x − i∂/∂y)/2 and ∂/∂ z̄ = (∂/∂x + i∂/∂y)/2 are the
partial complex derivatives. Check directly that ∆ is invariant.
∂2 ∂ 1 ∂2
(1.21) ∆ = u(u + 1) + (2u + 1) + .
∂u2 ∂u 16 u(u + 1) ∂ϕ2
1.7. Eigenfunctions of ∆.
(1.22) (∆ + λ)f = 0 .
and
(2πy)1/2 Is−1/2 (y) ∼ ey ,
as y → +∞, where Kν (y) and Iν (y) are the standard Bessel functions
(see Appendix B.4). Suppose that f (z) does not grow too fast, more
precisely
f (z) = o(e2πy ) ,
as y → +∞. This condition forces f (z) to be a multiple of the function
Now let us show how the Whittaker function evolves by the method
of images. For a function f (z) to be periodic in x of period 1 one needs
to verify the transformation rule
1
Z Z
(1.29) f (z) = Ws (rz) fs (r) γs (r) ds dr ,
2πi
(1/2) R
where the outer integration is taken over the vertical line Re s = 1/2,
Z
(1.30) fs (r) = f (z) Ws (rz) dµz ,
H
The proof of the next result reduces to the ordinary Fourier series
expansion for periodic functions.
(1.35) fn eε|n|
00 0 m2
u(u + 1) F (u) + (2u + 1) F (u) + s(1 − s) − F (u) = 0 .
4u(u + 1)
where Z
m
f (s) = f (z) Usm (z) dµz ,
H
1 π
Z p
(1.43) Fs (u) = (2u + 1 + 2 u(u + 1) cos θ)−s dθ
π 0
Fs (u) = F (s, 1 − s; 1, u) ,
Γ(s)2 1
Gs (u) = u−s F s, s; 2s, .
4π Γ(2s) u
Eigenfunctions of ∆ 27
1 1
(1.47) Gs (u) = log + O(1) , u → 0,
4π u
(1.48) Gs (u) = −(4πu)−1 + O(1) ,
0
u → 0,
(1.49) Gs (u) u−σ , u → +∞ .
where Z ν Z ν
−σ
u ξ σ−1 dξ 1
0 0
and Z 1 Z 1
(1 − ξ)σ−1 dξ 1 .
η η
Then using geodesic polar coordinates with the origin at z we get the
same expression for ∆w k(z, w).
and
(1.54) h−∆F, F i ≥ 0 .
(∆ + λ) F = 0 in D ,
(1.55)
F =0 on ∂D
the Hilbert space with respect to the inner product (1.51), where H is
reduced to D. Observe that the formula (1.52) remains valid if H is
replaced by D. This yields the following inequality
Z
∂F 2 ∂F 2 ∂F 2
Z
h−∆F, F i = + dx dy ≥ dx dy .
D ∂x ∂y D ∂y
∂F −1
Z Z
2 −2
F y dy = 2 F y dy ,
∂y
∂F 2
Z Z
2
F dµ ≤ 4 dx dy .
D D ∂y
1
(1.56) h−∆F, F i ≥ hF, F i
4
(we have assumed tacitly that F is real but, of course, this is not nec-
essary). Hence we conclude that if F is a non-zero solution to the
Dirichlet problem for a domain in the hyperbolic plane then its eigen-
value satisfies λ ≥ 1/4. This fact explains the absence of Whittaker
functions Ws (rz) in (1.29) and the spherical functions Usm (z) in (1.41)
beyond the line Re s = 1/2.
This is given by
It remains to show that the above integral does not depend on z, but
this is obvious because G acts on H transitively and ω, k are point-pair
invariants.
Proof. Since Λ does not depend on the eigenfunction f (z) we take for
computation f (w) = (Im w)s and specialize (1.60) to the point z = i
giving Z +∞ Z +∞ 2
x + (y − 1)2 s−2
Λ=2 k y dx dy .
0 0 4y
√
Changing the variable x = 2 uy and next the variable y = er one
easily completes the computation getting Λ = h(t).
d
L1 f (z) = f (z + t) = fx (z) .
dt
The formula (1.72) is obtained by differentiating at t = 0 as follows:
d z d z d z
L2 f (z) = f = Re fx (z) + Im fy (z) ,
dt tz + 1 dt tz + 1 dt tz + 1
d z
= −z 2 = −x2 + y 2 − 2i xy .
dt tz + 1
The formula (1.73) is obtained by differentiating at t = 0 as follows:
d
L3 f (z) = f (e2t z) = 2xfx (z) + 2yfy (z) .
dt
Finally (1.74) is obtained by adding the following easy formulas:
∂2 ∂2 ∂ ∂
L1 L2 = (y 2 − x2 ) 2
− 2xy − 2x − 2y ,
∂x ∂x∂y ∂x ∂y
∂2 ∂2
L2 L1 = (y 2 − x2 ) 2 − 2xy ,
∂x ∂x∂y
1 ∂2 ∂2 ∂2 ∂ ∂
L3 L3 = 2x2 2 + 4xy + 2y 2 2 + 2x + 2y .
2 ∂x ∂x∂y ∂y ∂x ∂y
∂f ∂G
Z Z
G(u(z, w)) (∆ + s(1 − s))f (w) dµw = G −f d`
V ∂U ∂n ∂n
∂G 1 ∂ log |z − w| ∂H(z, w)
Z Z Z
f d` = − f (w) d` + f (w) d`
∂U ∂n 2π ∂U ∂n ∂U ∂n
1
Z
f (w) d` .
2πε ∂U
Fuchsian groups
2.1. Definitions.
39
40 Fuchsian groups
Proposition 2.3. Every Fuchsian group of the first kind has a finite
number of generators and fundamental domain of finite volume.
(2.3) Fa (Y ) = σa P (Y ) .
For Y sufficiently large the cuspidal zones are disjoint, the set
[
F (Y ) = F \ Fa (Y )
a
Fundamental domains 43
Let Γ be a finite volume group. The quotient space Γ\H (the set
of orbits) is equipped with the topology in which the natural projection
π : H −→ Γ\H is continuous. In fact Γ\H is a Hausdorff connected
space, and with properly chosen analytic charts becomes a Riemann
surface. If the group Γ contains only hyperbolic elements, besides the
identity, then Γ\H is a compact, smooth surface of genus g ≥ 2. If Γ
has elliptic elements, then Γ\H has branch points at the fixed points
of the elliptic motions. If Γ has parabolic elements, then Γ\H is not
compact; in this case one usually compactifies Γ\H by adding cusps
with suitable charts.
It is easy to think of the Riemann surface Γ\H as being constructed
from a normal polygon by glueing pairs of congruent sides at equivalent
points.
44 Fuchsian groups
X̀ 1 |F |
(2.7) 2g − 2 + 1− +h= .
j=1
mj 2π
Of all the finite volume groups the most attractive ones for number
theory are the arithmetic groups. Since any comprehensive definition is
rather involved (cf. [Kat]), we content ourselves with basic examples.
All cusps for Γ(N ) are rational points a = u/v with (u, v) = 1 (under
0 0 0
the convention that ±1/0 = ∞). Two
u0 cusps a = u/v and a = u /v are
u
equivalent if and only if ± v ≡ v0 (mod N ). There are no elliptic
elements in Γ(N ), if N ≥ 3.
∗ ∗
Γ0 (N ) = γ ∈ SL2 (Z) : γ ≡ (mod N ) ,
∗
1 ∗
Γ1 (N ) = γ ∈ SL2 (Z) : γ ≡ (mod N ) .
1
48 Fuchsian groups
There are no elliptic fixed points of either order if 4|N or 9|N , respec-
tively. Every cusp for Γ0 (N ) is equivalent to a rational point a = u/v
with v ≥ 1, v|N , (u, v) = 1. Two cusps a = u/v, a0 = u0 /v 0 of the above
form are equivalent if and only if v = v 0 and u ≡ u0 (mod (v, N/v)).
Therefore the number of inequivalent cusps for Γ0 (N ) is given by
X
(2.14) h= ϕ((v, w)) .
vw=N
are the stability groups of cusps and B denotes the group of integral
translations, i.e.
1 b
(2.15) B= : b∈Z .
1
We shall partition the set σa−1 Γσb into double cosets with respect to B.
First let us examine the subset of the upper-triangular matrices,
i.e. those having the fixed point at ∞ ,
∗ ∗ −1
Ω∞ = ∈ σa Γσb .
∗
(2.16) Ω ∞ = B ω∞ B = ω∞ B = B ω ∞
for some
1 ∗
(2.17) ω∞ = ∈ σa−1 Γσb .
1
All other elements of σa−1 Γσb fall into one of the double cosets
for some
∗ ∗
(2.19) ωd/c = ∈ σa−1 Γσb ,
c d
shows that the double coset Ωd/c determines c uniquely whereas a and
d are determined modulo integral multiples of c. In fact Ωd/c does
not depend on the upper row of ωd/c . To see this, take ω = ac d∗ ,
ω1 = ac1 d∗ , two elements of σa−1 Γσb with the same lower row. Setting
We have
α ∗
−1
σ∞ Γσ∞ = σ0−1 Γσ0 = : αδ ≡ 1 (mod γq)
γq δ
and √
−1 α q ∗
σ∞ Γσ0 = √ √ : αδq ≡ 1 (mod γ) .
γ q δ q
For a cusp a = 1/v with vw = q, (v, w) = 1 we may take
√
√ w √
σa = .
v w 1/ w
We find that
−1 α ∗
σa Γσa = : αδ ≡ 1 (mod γv), (α+γ)(δ−γ) ≡ 1 (mod w)
γq δ
Kloosterman sums 51
and
√
−1 α √w ∗
√
σ∞ Γσa = : αδ ≡ 1 (mod γv), γ ≡ δ(mod w) .
γv w δ/ w
In the above matrices α, β, γ are integers. For other pairs of cusps the
sets σa−1 Γσb have a similar structure.
d
Ωd/c (m) = e m , m ∈ Z.
c
depends only on the equivalence classes of cusps but not on their rep-
resentatives nor on the choice of scaling matrices.
Let us see closely what the above constructions yield for the mod-
ular group Γ = SL2 (Z). In this case there is only one cusp a = b = ∞
for which we obtain the classical Kloosterman sum
X dm + an
S(m, n; c) = e
c
ad≡1 (mod c)
where τ (c) is the divisor function, τ (c) cε . For c prime (the hardest
case) this bound was derived by A. Weil [We] as a consequence of the
Riemann hypothesis for curves over finite fields. The special case n = 0
is simple; the Kloosterman sum reduces to the Ramanujan sum
X* dm X c
(2.26) S(m, 0; c) = e = µ δ,
c δ
d (mod c) δ|(c,m)
where the star restricts the summation to the classes prime to the mod-
ulus; hence the generating Dirichlet series for Ramanujan sums is equal
to
∞
X X
(2.27) c−2s S(m, 0; c) = ζ(2s)−1 δ 1−2s .
c=1 δ|m
(2.31) ca < |F | .
First for any c in the set Cab we estimate the number (2.24). Surpris-
ingly there is no sharp bound available for each Sab (0, 0; c) individually.
We show the following
d0 d ca
(2.34) 0 − ≥ 0 .
c c cc
54 Fuchsian groups
Hence, the bound (2.32) is derived by applying the box principle. Sim-
ilarly, if 0 < c, c0 ≤ X, we get from (2.34) that
d0 d
(2.36) 0 − ≥ ca c−1 X −1 .
c c
Summing this inequality over c ≤ X and 0 ≤ d < c where d0 /c0 is
chosen to be the succesive point to d/c we get (2.33).
Notice that c(a, b)2 ≥ cab which one can deduce by applying
Sab (0, 0; c) ≥ 1 to (2.32) with c = c(a, b). Incidentally (2.32) follows
from (2.33).
and
X
(2.38) c−1 |Sab (m, n; c)| ≤ c−1
ab X .
c≤X
10
# γ ∈ Γa \Γ : Im σa−1 γz > Y < 1 +
(2.39) .
ca Y
γ∈Γa \Γ
Y1 <Im γz<Y2
where yΓ (z) is the invariant height defined by (3.8) and the implied
constant depends on the group alone.
Chapter 3
Automorphic forms
3.1. Introduction.
57
58 Automorphic forms
(∆ + λ)f = 0 , λ = s(1 − s)
A s B
(3.5) fa (y) = (y + y 1−s ) + (y s − y 1−s ) .
2 2s − 1
The non-zero coefficients in (3.4) are bounded by
for any ε > 0, with the implied constant depending on ε and f . There-
fore, as y → +∞, we have
|n|≤N
|s| 1 1/2
f (σa z) = fa (y) + O ( (1 + )) .
y ca y
The zero-th term exists only if 1/2 < s ≤ 1, and it takes the form
fa (y) = fˆa (0) y 1−s with the coefficient bounded by
1 1/2 −s+1/2
fˆa (0) s − ca .
2
X Z 1
2
|fa (y)| + |fˆa (n) Ws (iny)|2 = |f (σa (x + iy))|2 dx .
n6=0 0
The Einsenstein series 61
Hence,
1−2s X Z +∞
ˆ 2Y ˆ 2
|fa (0)| + |fa (n)| Ws2 (iny) y −2 dy
2s − 1 Y
n6=0
Z +∞ Z 1
10
= |f (σa z)|2 dµz ≤ 1 +
Y 0 ca Y
because every orbit {σa γz : γ ∈ Γ} has no more than the above number
of points in the strip P (Y ) by Lemma 2.10. On the other hand it follows
from Z +∞
2
Ks−1/2 (y) y −1 dy |s|−1 e−π|s|
|s|
that Z +∞
Ws2 (iny) y −2 dy |n| |s|−1 e−π|s| ,
Y
if 2π|n|Y ≤ |s|. Setting 2πN Y = |s| one infers the desired estimate
for the sum of non-zero Fourier coefficients. To estimate the zero-th
coefficient take Y = c−1 a .
To estimate f (σa z) apply Cauchy’s inequality to the Fourier ex-
pansion as follows:
X
|f (σa z) − fa (y)|2 ≤ |n| (|n| + Y )−2 |fˆa (n)|2
n6=0
X
· |n|−1 (|n| + Y )2 |Ws (nz)|2
n6=0
Y −2 (c−1
a Y + |s|) y
−1
(y −1 |s| + Y )2
Note that ψ̂(s) (|s| + 1)−A (by repeated partial integration) in the
vertical strips σ1 ≤ Re s ≤ σ2 where σ1 , σ2 and A are any constants.
Hence, it is clear that the integral (3.12) converges absolutely.
Cusp forms 63
The innermost integral is just the zero-th term fa (y) in the Fourier
expansion (3.2) of f at the cusp a. In fact the above argument is valid
for any f ∈ A(Γ\H) such that |f | is integrable over F . Hence we obtain
(3.16) e
L(Γ\H) = C(Γ\H) e
⊕ E(Γ\H)
where the tilde stands for closure in the Hilbert space L(Γ\H) (with
respect to the norm topology).
Automorphic forms in the space C(Γ\H) are called cusp forms.
Therefore, a cusp form f is an automorphic function which is an eigen-
function of the Laplace operator and which has no zero-th term in the
Fourier expansion at any cusp, i.e. (3.15) holds. We denote
at any cusp a, by Theorem 3.1. From the estimate (3.7) it follows that
f decays exponentially at every cusp; more precisely, it satisfies
f (z) e−2πy(z) .
where the first term above exists only if a, b are equal. To the innermost
sum we apply Poisson’s formula
X XZ +∞
p(ωcd (z + n)) = p(ωcd (z + t)) e(−nt) dt
n∈Z n∈Z −∞
where
ωcd (z + t) = a/c − c−2 (t + x + d/c + iy)−1 .
Changing t 7→ t − x − d/c the Fourier integral becomes
Z +∞
d a 1
e(nx + n ) p( − 2 ) e(−nt) dt .
c −∞ c c (t + iy)
and
Z +∞
(t2 + y 2 )−s e(−nt) dt
(3.19) −∞
where
By the trivial estimate (2.38) for Kloosterman sums and the crude
bound Ws (z) min{y 1−σ , e−2πy } we infer the following
Fourier expansion of the Einsenstein series 67
X a s−1/2
(3.25) ϕ(n, s) = π s Γ(s)−1 ζ(2s)−1 |n|−1/2
b
ab=|n|
3
(3.26) res E(z, s) = .
s=1 π
Note that the holomorphy of all ϕ(n, s) on the line Re s = 1/2 is equiv-
alent to the non-vanishing of ζ(s) on the line Re s = 1, the latter fact
being equivalent to the Prime Number Theorem. To obtain some sym-
metry we put
so that
where
X a t
(3.30) ηt (n) = .
b
ab=n
Show using the functional equation that E(z, 1/2) ≡ 0. Then evaluate
(3.29) at s = 1/2 to get the following expansion involving the divisor
function τ (n)
∞
∂ 1 √ √ X
E(z, ) = y log y + 4 y τ (n) K0 (2πny) cos(2πnx) .
∂s 2 n=1
Chapter 4
69
70 The spectral theorem. Discrete part
whence
so ∆ is symmetric. Moreover
Z
(4.3) h−∆f, f i = |∇f |2 dx dy ≥ 0
F
so −∆ is non-negative.
This is called the automorphic kernel. Here and below we require the
absolute convergence of all relevant series and integrals.
First we assume that k(u) is smooth and compactly supported on
R+ . After this case is worked out, we shall replace the compactness by
a weaker condition applying a suitable approximation. Clearly
L : B(Γ\H) −→ B(Γ\H) .
Put g(z) = (Lf )(z) where f ∈ B(Γ\H). Let us compute the zero-th
term of g at a cusp a (see (3.2))
Z 1 Z 1 Z
ga (y) = g(σa n(t)z) dt = k(σa n(t)z, w) f (w) dµw dt
0 0 H
Z Z 1 Z
= k(z, w) f (σa n(t)w) dt dµw = k(z, w) fa (Im w) dµw
H 0 H
L : C(Γ\H) −→ C(Γ\H) .
X Z +∞
(4.5) Ha (z, w) = k(z, σa n(t)σa−1 γw) dt .
γ∈Γa \Γ −∞
72 The spectral theorem. Discrete part
(4.6) Ha (σa z, w) 1 + Im z .
X Z +∞
Ha (σa z, σa w) = k(z, t + τ w) dt .
−∞
τ ∈B\σa
−1
Γσa
Since k(u) has compact support, the ranges of integration and summa-
tion are restricted by |z − t − τ w|2 Im z Im τ w. This shows that
Im z Im τ w, and the integral is bounded by O(Im z). By Lemma 2.10
we conclude that
1
Ha (σa z, σa w) 1 + Im z = 1 + Im z .
Im z
Similarly, using Lemma 4.2, one shows that all terms in (4.5) give a
uniformly bounded contribution except for γ = 1 so that
Z +∞
Ha (z, w) = k(z, σa n(t)σa−1 w) dt + O(1) .
−∞
74 The spectral theorem. Discrete part
for a > s ≥ 2. This has a kernel k(u) = Ga (u) − Gs (u) which satisfies
the conditions (4.11) (see Lemma 1.7). Recall that Rs = (∆+s(1−s))−1
(see Theorem 1.17); hence Rs has dense range in L(Γ\H), and so does L
by the Hilbert formula. The modified operator L̂ is bounded on L2 (F ).
Although L̂ annihilates many functions, the range of L̂ is dense in the
subspace C(Γ\H) ⊂ L2 (F ). Indeed, for f ∈ D(Γ\H) we create
Theorem 4.7 The automorphic Laplace operator ∆ has pure point spec-
trum in C(Γ\H), i.e. C(Γ\H) is spanned by cusp forms. The eigenspaces
have finite dimension. For any complete orthonormal system of cusp
forms {uj } every f ∈ C(Γ\H) has the expansion
X
(4.15) f (z) = hf, uj i uj (z)
j
5.1. Introduction.
m
(5.2) Gs (z/z 0 ) = − log |z − z 0 | + O(1) , as z 0 → z
2π
77
78 The automorphic Green function
The equation (5.5) holds in Bµ (Γ\H) if Re s > µ+1. Note that Bµ (Γ\H)
is not a subspace of L(Γ\H) if µ ≥ 1/2.
For several applications of the automorphic Green function we need
to control its growth. A delicate situation occurs near the diagonal
z ≡ z 0 (mod Γ); this will be manifested in a double Fourier expansion.
according to the double coset decomposition of the set σa−1 Γσb given in
Theorem 2.7. We obtain
X
(5.7) K(σa z, σb z 0 ) = δab K0 (z, z 0 ) + Kc (z, z 0 ) ,
c∈Cab
say, where Z ∞
0
Pn (y, y ) = e(ξn) k(iy + ξ, iy 0 ) dξ .
−∞
+∞
−1
ZZ
0
Pn,m (y, y ) = e(ξn + ηm) k(iy + ξ, ) dξ dη .
iy 0 − η
−∞
Lemma 5.1. Let k(u) = Gs (u) with Re s > 1. Suppose that y 0 > y.
Then we have
and
Lemma 5.2. Let k(u) = Gs (u) with Re s > 1. Suppose that y 0 y > 1.
Then we have
and
√
J2s−1 (4π mn)
0 1 −1/2 0
(5.13) Pn,m (y, y ) = |mn| Ws (iny) Ws (imy ) · p
2 I2s−1 (4π |mn|)
The above formulas are valid in limited ranges; they are applicable
for all terms in (5.7) in the domain
Dab = (z, z 0 ) ∈ H × H : y 0 > y , y 0 y > c(a, b)−2
(5.14)
where c(a, b) denotes the smallest element of Cab (see (2.22)). If (z, z 0 ) ∈
Dab then σa 6≡ σb z 0 (mod Γ) so Gs (σa z/σb z 0 ) is defined. By Lemmas
5.1 and 5.2 we obtain after changing the order of summation (the series
converges absolutely) the following Fourier expansion.
Gs (σa z/σb z 0 ) = (2s − 1)−1 y s (y 0 )1−s δab + ϕab (s) (yy 0 )1−s
X
+ (4π|n|)−1 Ws (nz 0 ) V s (nz) δab
n6=0
X
+ (2s − 1)−1 y 1−s ϕab (m, s) Ws (mz 0 )
(5.15) m6=0
X
+ (2s − 1)−1 (y 0 )1−s ϕab (n, s) W s (nz)
n6=0
X
+ Zs (m, n) Ws (mz 0 ) W s (nz)
mn6=0
where ϕab (s), ϕab (n, s) are the Fourier coefficientes of the Eisenstein
series of Ea (σb z, s) (see Theorem 3.4) and Zs (m, n) is the Kloosterman
sums zeta-function defined by
4π √
J2s−1
mn
c
p X
(5.16) 2 |mn| Zs (m, n) = c−1 Sab (m, n; c) ·
I2s−1 4π |mn|
p
c
c
getting
X
(4π|n|)−1 exp (2πin(x0 − x) + 2π|n|(y − y 0 )(1 + O(|n|−1 ))
n6=0
∞
X 0
= Re (2πn)−1 e(n(z − z 0 )) + O(e−2π(y −y) )
n=1
1 0
= − log |1 − e(z − z 0 )| + O(e−2π(y −y) ) .
π
The third line is estimated by
∞
X 0 0
1−σ
y mσ−1 e−2πmy y 1−σ e−2πy .
1
For the last line we need a bound on Zs (m, n). To this end we appeal
to the following crude estimates for Bessel functions
Now employing the trivial bounds for Kloosterman sums we infer that
4π p|mn|
Zs (m, n) exp ,
c(a, b)
and hence the last line on the right side of (5.15) is estimated by
X 4π p|mn| 0
exp − 2π|m|y − 2π|n|y e−2π(y +y)
0
c(a, b)
mn6=0
provided y 0 y > δ > c(a, b)−2 . Collecting the above estimates we obtain
Lemma 5.4. Let Re s > 1 and δ > c(a, b)−2 . Then for z, z 0 with
y 0 > y and y 0 y > δ we have
Summing we obtain
Z +∞
2 00 2 2
y (P − 4π n P ) = e(ξn) ∆z k(iy + ξ, iy 0 ) dξ = −s(1 − s)P
−∞
W = I 0 K − IK 0 = 4 π |n| .
whence we get (5.10) and (5.11) as in the Fourier expansion for the
Eisenstein series (see Theorem 3.4 and the integrals preceeding its
proof).
For n 6= 0 we have
Z +∞ n/y 0
0 −1 0
Pn,m (y, y ) = (4π|n|) Ws (iny) y e(ηmy 0 ) Vs dη .
−∞ η+i
Lemma 5.5. Let Re s > 1/2 and a 6= 0, b be real numbers. Then the
integral
Z +∞ a
(5.19) e(ηb) Vs dη
−∞ η+i
is equal to
where V (z) = Vs (z). For z = a(η +i)−1 = aη(η 2 +1)−1 −ia(η 2 +1)−1 =
axη − iayη , say, this gives
Hence,
∂2 a
a2 V = a2 x2η Vxx + 2a2 xη yη Vxy + a2 yη2 Vyy
∂η 2 η+i
= a2 (x2η − yη2 ) Vxx + 2a2 xη yη Vxy − s(1 − s)V
= 4π 2 a2 x0η V − 2πia2 yη0 Vy − s(1 − s)V
because x0η = yη2 − x2η , yη0 = −2 xη yη , Vxx = (2πi)2 V and Vxy = 2πi Vy .
Since
∂ a
V = 2πiax0η V + ayη0 Vy ,
∂η η+i
we obtain
∂2 a a ∂ a
a2 V + s(1 − s) V + 2πia V = 0.
∂η 2 η+i η+i ∂η η+i
Let v(a) denote the Fourier integral (5.19) as a function of a. Integrat-
ing by parts and using the above relation we find that v(a) satisfies the
second order differential equation
which yields
Z +∞
s+1/2 −1 s
(5.25) v(a) ∼ 2π Γ(s + 1/2) |a| (1 + η 2 )−s dη ,
−∞
if ab > 0, and
p p
v(a) = α|a|1/2 I2s−1 (4π |ab|) + β|a|1/2 K2s−1 (4π |ab|) ,
and similar asymptotics hold true for Y2s−1 and K2s−1 but with s re-
placed by 1 − s. On the other hand we infer by (5.25), (3.19) and (5.26)
that
Z +∞
s+1/2 −1 s
v(a) ∼ 2π Γ(s + 1/2) |a| e(ηb) (1 + η 2 )−s dη
−∞
2s −1 s s−1/2
= 2 (2π) Γ(2s) |a| |b| Ks−1/2 (2π|b|) .
Analytic continuation
of the Eisenstein series
for any f ∈ Ba−1 (Γ\H) where Ga (z/z 0 ) is the automorphic Green func-
tion. Let Ea (z, s) be the Eisenstein series for the cusp a. From the
Fourier expansion
87
88 Analytic continuation of the Eisenstein series
giving
Z
Ga (z/z 0 ) Ea (z 0 , s) dµz 0 .
(6.2) −Ea (z, s) = a(1 − a) − s(1 − s)
F
This is a homogeneous Fredholm equation for Ea (z, s), but the classical
Fredholm theory cannot be employed for several reasons.
The first obstacle is that the kernel Ga (z/z 0 ) is singular on the di-
agonal z = z 0 . This is a minor problem. The singularities will disappear
if we take the difference
for fixed a > b > 2. From (6.2) we obtain the homogeneous equation
Z
(6.3) −νab (s) Ea (z, s) = Gab (z/z 0 ) Ea (z 0 , s) dµz 0
F
where
−1 −1
νab (s) = a(1 − a) − s(1 − s) − b(1 − b) − s(1 − s) .
Later we shall put λab (s) = −νab (s)−1 on the other side of (6.3). Note
that λab (s) is a polynomial in s of degree four,
GYab (z/z 0 ) = Gab (z/z 0 ) − (2a − 1)−1 (Im σb−1 z 0 )1−a Eb (z, a)
+ (2b − 1)−1 (Im σb−1 z 0 )1−b Eb (z, b) ,
Replacing Gab (z/z 0 ) in (6.3) by GYab (z/z 0 ) we must bring back the
integrals of subtracted quantities over cuspidal zones. They yield
Z
(Im σb−1 z 0 )1−a Ea (z 0 , s) dµz 0
Fb (Y )
Z 1 Z +∞
= y −1−a (δab y s + ϕab (s) y 1−s + · · · ) dx dy
0 Y
Y s−a
Y 1−a−s
= δab + ϕab (s) ,
a−s a+s−1
for every b, and similar terms must be added with b in place of a. In
this way we obtain the inhomogeneous Fredholm equation
Z
−νab (s) Ea (z, s) = GYab (z/z 0 ) Ea (z 0 , s) dµz 0
F
Y s−a
+ Ea (z, a)
(2a − 1)(a − s)
Y s−b
(6.6) − Ea (z, b)
(2b − 1)(b − s)
90 Analytic continuation of the Eisenstein series
Y 1−a−s X
+ ϕab (s) Eb (z, a)
(2a − 1)(a + s − 1)
b
1−b−s
Y X
− ϕab (s) Eb (z, b) .
(2b − 1)(b + s − 1)
b
A new obstacle has emerged from the terms involving the scattering
matrix elements ϕab (s) whose analytic continuation to Re s ≤ 1 has not
yet been established. We shall kill these terms by making a suitable
linear combination of (6.6) for three values Y , 2Y , 4Y (one could also
accomplish the same goal by integrating in Y against a suitable test
function such that its Mellin transform vanishes at the points −a and
−b). We find the following equation
Z
(6.7) h(z) = f (z) + λ H(z, z 0 ) h(z 0 ) dµz 0
F
22s−1−a+b 22s−1−a+b
f (z) = + Y s−b Ea (z, b) − Y s−a Ea (z, a) ,
(2b − 1)(b − s) (2a − 1)(a − s)
f (σa z) y a
The analytic continuation of Ea (z,s) 91
and
0
H(σa z, σb z 0 ) y a e−2π max{y −y,0} ,
if y, y 0 ≥ 4Y by (6.5). To handle the problem we multiply (6.7) through-
out by η(z) = e−ηy(z) where η is a small positive constant, 0 < η < 2π.
Then we borrow from the exponential decay in the z 0 variable to kill
the polynomial growth in the z variable. Accordingly, we re-write (6.7)
as follows
Z
(6.8) η(z) h(z) = η(z) f (z)+λ η(z) η(z 0 )−1 H(z, z 0 ) η(z 0 ) h(z 0 ) dµz 0 .
F
Here η(z) f (z) is bounded in F , and η(z) η(z 0 )−1 H(z, z 0 ) is bounded in
F × F . By the Fredholm theory the kernel η(z) η(z 0 )−1 H(z, z 0 ) has a
resolvent of type
whence
λ
Z
(6.10) h(z) = f (z) + η(z)−1 η(z 0 ) Dλ (z, z 0 ) f (z 0 ) dµz 0 .
D(λ) F
and
It may surprise anyone that the functional equations for the scat-
tering matrix and the Eisenstein series come as consequences of the
apparently remote facts that ∆ is a symmetric and non-negative opera-
tor in L(Γ\H). We shall appeal to these facts to establish the following
with 0 < ε < 2π. If Re s > 1, then f (z) is a linear combination of the
Eisenstein series Ea (z, s).
where the error term is shown to be bounded using the growth condition
(6.21). Substracting the Eisenstein series we kill the leading terms α a y s
and get
X
g(z) = f (z) − αa Ea (z, s) 1
a
94 Analytic continuation of the Eisenstein series
Denote Φa (s) = [. . . , ϕab (s), . . . ] the a-th row vector of the scat-
tering matrix Φ(s) and its `2 -norm by
X
kΦa (s)k2 = |ϕab (s)|2 .
b
Since Φ(s) is unitary on the critical line (see (6.24)), it follows that Φ(s)
is holomorphic on this line and
(6.29) EaY (z, s) = Ea (z, s) − δab (Im σb−1 z)s − ϕab (s) (Im σb−1 z)1−s
96 Analytic continuation of the Eisenstein series
where in the last term we have the scalar product of two row vectors of
the scattering matrix.
s = sj is a pole of ϕab (s), then it is also a pole of ϕaa (s). The residue
of ϕaa (s) at s = sj > 1/2 is real and positive.
with
ρb = lim (s − sj )m ϕab (s) .
s→sj
Note that the first part of the zero-th term in the Fourier expansion
of Ea (σa z, s) is killed in the limit. If Re sj > 1/2 then u(z) is square-
integrable; thus its eigenvalue must be real, non-negative so sj must
lie in the segment (1/2, 1]. Moreover, if sj was not a pole of Φa (s) or
if sj had order m > 1, then ρb = 0 for any b showing that u(z) is a
cusp form. This, however, is impossible because the Eisenstein series
Ea (z, s) with s 6= sj is orthogonal to cusp forms; hence the limit u(z)
would be orthogonal to itself. We conclude the above analysis by
Theorem 6.10. The poles of Ea (z, s) in Re s > 1/2 are among the
poles of ϕaa (s) and they are simple. The residues are Maass forms;
they are square-integrable on F and orthogonal to cusp forms.
Theorem 6.11. The Eisenstein series Ea (z, s) has no poles on the line
Re s = 1/2.
98 Analytic continuation of the Eisenstein series
Proposition 6.12. For s 6= 1/2 the Eisenstein series Ea (z, s) does not
vanish identically.
Hence,
lim (σ − 1)2 kEaY (σ)k2 = α + O(Y −1 )
σ→1
where α is the residue of ϕaa (s) at s = 1. On the other hand the residue
of Ea (z, s) at s = 1 is an eigenfunction of ∆ with eigenvalue zero so it
is a harmonic function in L(Γ\H); hence, it is constant. By the Fourier
expansion this constant is equal to the residue of ϕaa (s) whereas the
other coefficients must be regular at s = 1. Therefore,
lim (s − 1) Ea (z, s) = α .
s→1
where λ1 = s1 (1 − s1 ) and λ2 = s2 (1 − s2 ).
we get Z X
0
(Y ) − fn0 (Y ) g−n (Y )
= fn (Y ) g−n
L(Y ) n
d 0
(y) − fn0 (y) g−n (y) = (λ1 − λ2 ) y −2 fn (y) g−n (y) .
fn (y) g−n
dy
Remarks. If the zero-th terms are of type fa (y) = fa+ y s + fa− y 1−s ,
then (6.34) becomes
X
hf Y , g Y i = (s1 − s2 )−1 fa+ ga− Y s1 −s2 − fa− ga+ Y s1 −s2
a
X
+ (s1 + s2 − 1)−1 fa+ ga+ Y s1 +s2 −1 − fa− ga− Y 1−s1 −s2
a
where s = 1/2 + iv (note that for the last approximation one needs the
functional equation (6.23)). Hence, upon taking the limit σ → 1/2 we
derive
hE Y (·, s), t E Y (·, s)i = (2s − 1)−1 Φ(1 − s) Y 2s−1 − Φ(s) Y 1−2s
with ψ ∈ C0∞ (R+ ) will emerge from the contour integral (see (3.12))
1
Z
(7.2) Ea (z|ψ) = ψ̂(s) Ea (z, s) ds
2πi
(σ)
103
104 The spectral theorem. Continuous part
This approach requires some control over the growth of Ea (z, s) in the
s variable. So far our knowledge is rather poor in this aspect, namely
that Ea (z, s) is a meromorphic function in s of order ≤ 8. We need a
polynomial growth on average over segments of the line Re s = 1/2.
Consider the subspace C0∞ (R+ ) of the Hilbert space L2 (R+ ) with
the inner product
+∞
1
Z
(7.4) hf, gi = f (r) g(r) dr .
2π 0
defined by
+∞
1
Z
(7.5) (Ea f )(z) = f (r) Ea (z, 1/2 + ir) dr .
4π 0
The estimate (6.20) shows that the Eisenstein series Ea (z, 1/2 + ir)
barely fails to be square-integrable on F . However by partial integration
in r we get a slightly better bound for the Eisenstein transform, namely
at any cusp b. The gain of the logarithmic factor is small, yet sufficient
to see that the Eisenstein transform is in L(Γ\H), i.e.
∆ Ea = E a M
where
1
(M f )(r) = − r2 + f (r) .
4
There are various orthogonalities worthy of record. By Proposition
7.1 the Eisenstein spaces Ea (Γ\H) for distinct cusps are orthogonal.
Also every Ea (Γ\H) is orthogonal to the space C(Γ\H) spanned by cusp
forms, by Theorem 6.7. Finally Ea (Γ\H) is orthogonal to the residues
of any Eisenstein series Eb (z, s) at poles s = sj in 1/2 < sj ≤ 1 because
the eigenvalues satisfy 0 ≤ λj = sj (1 − sj ) < 1/4 ≤ r 2 + 1/4 for r ∈ R.
Therefore, arguing by means of orthogonality we conclude that
(to be precise use the above relations for a finite collection of functions
fj and then drop this condition in (7.8) by positivity).
Now all the conditions for Bessel’s inequality are satisfied; hence
X X 1 Z +∞
2
|h(tj ) uj (w)| + |h(r) Ea (w, 1/2 + ir)|2 dr
j a
4π −∞
(7.9) Z
≤ |K(z, w)|2 dµz .
F
kernel, but if you feel unconfortable with the discontinuity think of k(u)
as a compactly supported smooth approximation to this characteristic
function).
In order to estimate the Selberg/Harish-Chandra transform h(t),
rather than computing explicitly, we appeal to the integral representa-
tion Z
h(t) = k(i, z) y s dµz
H
as in the proof of Theorem 1.16. For s = 0 this gives
i
Z
h = k(i, z) dµz = 4πδ ,
2 H
which
√ is just the hyperbolic area of a disc of radius
√ r given by sinh(r/2)
= δ. Since u(i, √ z) < δ implies |y − 1| < 2δ, we have |y s − 1| ≤
|s| |y − 1| ≤ |s| 2δ ; whence
i √ i
|h(t) − h | ≤ |s| 2δ h .
2 2
This yields 2πδ < |h(t)| < 6πδ, if |s| ≤ (8δ)−1/2 .
Here we have u(z, w) ≤ δ and u(z, γw) ≤ δ; whence u(ω, γw) ≤ 4δ(δ+1)
by the triangle inequality for the hyperbolic distance. Setting
Nδ (w) = # γ ∈ Γ : u(w, γw) ≤ 4δ(δ + 1)
we obtain
i
Z Z
2
|K(z, w| dµz ≤ Nδ (w) k(i, z)2 dµz = Nδ (w) h .
F H 2
Inserting the above estimates into (7.9) we get
X0 X 1 Z 0
2
|uj (z)| + |Ea (z, 1/2 + ir)|2 dr < (πδ)−1 Nδ (z)
j a
4π
110 The spectral theorem. Continuous part
X XZ T
2
(7.10) |uj (z)| + |Ea (z, 1/2 + it)|2 dt T 2 + T y(z)
|tj |<T a −T
One can derive from (7.10) many valuable estimates for the spectra
and the eigenfunctions. For example one can show quickly that
NΓ (T ) = # j : |tj | < T T 2 .
(7.11)
1
X Z
(7.12) Ea (z|ψ) = ψ̂(sj ) uaj (z) + ψ̂(s) Ea (z, s) ds
2πi
1/2<sj ≤1 (1/2)
Spectral decomposition of E(Γ\H) 111
because the uaj (z) are mutually orthogonal as well as being orthogonal
to each of the Eisenstein series Ea (z, s) on the line Re s = 1/2.
The above argument, however, does not apply to ψ̂(s) on the line
Re s = 1/2 because the inner product hEa (·, s), Ea (·, s0 )i diverges. The-
refore the expansion (7.12) cannot be regarded as a spectral decompo-
sition (in the sense of a continuous eigenpacket) since the coefficient
ψ̂(s) in the integral does not agree with the projection of Ea (·|ψ) on
Ea (·, s). To get the proper representation we rearrange this integral by
an appeal to the functional equation
X
Ea (z, 1 − s) = ϕab (1 − s) Eb (z, s)
b
1
Z
ψ̂(s) Ea (z, s) ds
2πi
(1/2)
(7.14) X 1 Z
= hEa (·|ψ), Eb (·, s)i Eb (z, s) ds .
4πi
b
(1/2)
This is the desired expression for the projection on the Eisenstein series.
Note that it takes all the Eisenstein series Eb (z, 1/2 + ir) to perform
the spectral decomposition of one Ea (z|ψ).
112 The spectral theorem. Continuous part
Combining Theorems 4.7 and 7.3 one gets the spectral decompo-
sition of the whole space L(Γ\H). Any f ∈ D(Γ\H) has the spectral
expansion obtained in synthesis of (4.15) and (7.15).
Hence, we obtain
Notice that our initial domain was Re s > 1; however, the spectral
expansion (7.18) is valid in Re s > 1/2 without modification by analytic
continuation. In order to extend the result to the complementary half-
plane we consider the integral
1
Z
Iα (s) = χsa (v) Ea (z, v) Ea (w, 1 − v) dv
2πi
(α)
where α > 1/2 is sufficiently close to 1/2 so that all Eisenstein series
are holomorphic in the strip 1/2 < Re s < α. By Cauchy’s theorem
1 X
(7.19) Gs (z/w) − G1−s (z/w) = − Ea (z, s) Ea (w, 1 − s)
2s − 1 a
since the discrete spectrum series and the integral Iα (s) are invariant
under the change s 7→ 1 − s within the strip 1 − α < Re s < α, and also
X X
(7.20) Ea (z, s) Ea (w, 1 − s) = Ea (z, 1 − s) Ea (w, s)
a a
by (6.22’).
1 X
res Gs (z/w) = − uk (z) uk (w) .
s=sj 2s − 1 s =s
k j
1 X
Gs (z/w) = − uj (z) uj (w)
(s − 1/2)2
sj =1/2
1 X
− Ea (z, 1/2) E a (w, 1/2) + · · · .
4(s − 1/2) a
In the half-plane Re s < 1/2 the Green function inherits a lot of simple
poles from the Eisenstein series besides the finite number of poles at
1 − sj from the discrete spectrum (see (7.19)).
Chapter 8
8.1. Introduction.
117
118 Estimates for the Fourier coefficients of Maass forms
The required properties are inherited from those of the Eisenstein series
Ea (z, s) through the following integral representation
Z
(8.10) Θj (s) Laj (s) = 8 |uj (z)|2 Ea (z, s) dµz
Γ\H
where
s s s
(8.11) Θj (s) = π −1−s Γ(s)−1 Γ( )2 Γ( + itj ) Γ( − itj ) cosh πtj .
2 2 2
In the half-plane Re s > 1, where the series (8.9) converges absolutely
by virtue of (8.7), this integral representation is derived by unfolding
the integral as follows.
X Z
|uj (z)|2 (Im σa−1 γz)s dµz
γ∈Γa \Γ Γ\H
Z X Z +∞
2 s 2
= |uj (σa z)| y dµz = |ρaj (n)| |Wsj (ny)|2 y s−2 dy
n6=0 0
B\H
+∞
1
Z
= cosh(πtj ) Laj (s) (2π)−s |Kitj (y)|2 y s−1 dy .
π 0
satisfies
with some complex a = (an ). However, one cannot often take advantage
of having a special combination; therefore, we might as well consider
linear forms in general and seek estimates in terms of the `2 -norm
N
X
2
kak = |an |2 .
1
Yet, for special linear forms, one should be able to improve upon
the individual bound (8.19) by exploiting the variation in sign of ν aj (n).
For example, if the an are given by additive characters (observe that
the Fourier coefficients of a Maass form are determined only up to the
twist by a fixed additive character because the scaling matrix σa can be
altered by a translation from the right side), we shall prove the following
Proof. We have
Z 1
ρaj (n) Wsj (iny) = uj (σa z) e(−nx) dx .
0
122 Estimates for the Fourier coefficients of Maass forms
where
Z +∞
(8.22) ϕaj (x) = uj (σa z) y −1 dy |sj |1/2
0
by (8.3) (note that uj (σa z) is a cusp form for the group σa−1 Γσa ) with
the implied constant depending on the group. Summing over n we get
Z 1
−1/2 sj 1 − s j X
π Γ Γ ρaj (n) = FN (x) ϕaj (x) dx
2 2 0
|n|≤N
Lots of other results stem from that simple idea of writing ρaj (n)
as the Fourier transform of uj (σa z). Here is a cute one (its proof follows
instantly by (8.21), (8.22) and Parseval’s identity; this was pointed out
to me by W. Duke).
First, however, let us show what can be infered from Bessel’s in-
equality (7.10) which does not require the spectral theorem in its full
force. Indeed, integrating over a horocycle segment in the a cuspidal
zone at height y one gets
X X
|ρaj (n) Wsj (nz)|2
|tj |<T n6=0
(8.24) X 1 Z T X
+ |ϕac (n, s) Ws (nz)|2 dt T 2 + yT ,
4π −T
c n6=0
where X
S(x) = an e(−nx) .
n
The Parseval formula asserts (we assume that kak < +∞)
Z 1 X
|S(x)|2 dx = |an |2 = kak2 .
0 n
Here and thereafter the three dots stand for the corresponding con-
tribution of the continuous spectrum. Furthermore, this is bounded
by Z Z 1 1 Z 1
|S(x)|2 |K(z, z 0 )| dx dx0 = |S(x)|2 H(z, y 0 ) dx ,
0 0 0
say. Suppose that k(u) is real, non-negative so that we can drop the
absolute value in the automorphic kernel K(z, z 0 ). Then the integral
Z 1
0
H(z, y ) = K(z, z 0 ) dx0
0
Therefore,
0 0 1/2 y 0 10 i
H(z, y ) ≤ (y y) g log + g(0) + h .
y ca 2
For the Fourier pair h(t) = exp(−t2 /4T 2 ), g(r) = π −1/2 T exp(−r 2 T 2 )
this yields
Theorem 8.2. For T ≥ 1 and any complex a = (an ) with kak < +∞
we have
X X 1 Z T 1
2
|a ⊗ uj (σa z)| + |a ⊗ Ec (σa z, + it)|2 dt
(8.25) 4π −T 2
|tj |<T c
< T (y + 22 c−1 2
a ) kak .
Now take a = (an ) in which all but one entry vanishes getting
X
|ρaj (n) Wsj (nz)|2
|tj |<T
(8.26) X 1 Z T
+ |ϕac (n, s) Ws (nz)|2 dt < T (y + 22 c−1
a )
c
4π −T
X X 1 Z T
2
(8.27) |νaj (n)| + |ηac (n, t)|2 dt T 2 + c−1
a |n| T
4π −T
|tj |<T c
The general bounds so far established are not bad if one considers
the achieved degree of uniformity in present parameters. However, for
congruence groups some estimates can be improved. What makes this
possible is the existence of a special basis in L(Γ\H) which diagonalizes
the Hecke operators. For the classical automorphic forms this is the core
The case of congruence groups 127
1 X
(8.29) (Tn f )(z) = √ f (τ z) .
n
τ ∈Γ1 \Γn
1 X X az + b
(8.30) (Tn f )(z) = √ f ;
n d
ad=n b(mod d)
(the other operators are not even normal). Therefore, in the space
of cusp forms C(Γ0 (N )\H) an orthogonal basis {uj (z)} can be chosen
which consists of simultaneous eigenfunctions for all Tn , i.e.
The main profit from splitting the space of cusp forms into new-
forms is the multiplicativity of the Fourier coefficients. Precisely, if u(z)
is a newform on Γ0 (N ), then its first Fourier coefficient ν(1) does not
vanish (it is customary to normalize u(z) by setting ν(1) = 1 which
we reject in favor of the L2 -normalization to avoid confusion), and
λ(n) = ν(n)/ν(1) satisfy the following rules of multiplication
X mn
λ(m) λ(n) = λ 2 if (n, N ) = 1 ,
(8.39) d
d|(m,n)
for all x ≥ 1 and any ε > 0, the implied constant depending only on ε.
Moreover, we have
Remarks. The upper bound of (8.42) follows easily from the Ramanu-
jan-Petersson conjecture (8.4), but this profound conjecture is beyond
the reach of current knowledge. By (8.7) we have
X
(8.44) |ν(n)|2 x N −1 + λ1/2
0<n≤x
Proof. The lower bound of (8.42) is easy to prove, just observe that if
p - N then either |λ(p)| ≥ 1/2 or |λ(p2 )| ≥ 1/2 since λ(p)2 = λ(p2 ) + 1
by (8.39). For the proof of the upper bound of (8.42) we consider the
Rankin-Selberg L-function
∞
X
L(s) = |λ(n)|2 n−s
1
The case of congruence groups 131
since τ (`)2 `ε for any 0 < ε < s − 1, the implied constant depending
on ε alone. Iterating the obtained inequality we get
k
L2 (s) L(s − εk) (λN )10
for any k ≥ 1 and 0 < εk < s − 1, the implied constant depending only
on ε, s and k. Taking the root of degree 2k with k large we get the
bound
(8.45) L(s) (λ N )ε
for any ε > 0 and s > 1, the implied constant depending on ε and
s. This bound implies (8.42) with the extra factor xε which can be
removed by means of (8.44).
The lower bound for |ν(1)|2 in (8.43) follows immediately by com-
paring the upper bound of (8.42) with (8.17) for x = (λ N )10 . The
upper bound for |ν(1)|2 is obtained by combining (8.44) with the lower
bound of (8.42) for x = λ1/2 N .
(8.46) |ν(1)|2 N −1 (λ N )ε .
132 Estimates for the Fourier coefficients of Maass forms
Spectral theory of
Kloosterman sums
9.1. Introduction.
133
134 Spectral theory of Kloosterman sums
4π p|mn|
0
Ws (imy ) Ws (iny) J2s−1
c
Γ(s − 1/2)2
(c2 y 0 y)1/2−s c−3 s−3/2 ,
Γ(2s)
Iν (y) Kν (y 0 ) , if y 0 ≥ y ,
0
(9.4) 2 Dν (y, y ) =
Iν (y 0 ) Kν (y) , if y 0 ≤ y .
We obtain
136 Spectral theory of Kloosterman sums
1
δab δmn Ds−1/2 (y 0 , y) + Zs (m, n) Ks−1/2 (y 0 ) Ks−1/2 (y)
2|n|
X
= (s − sj )−1 (1 − s − sj )−1 ρ̄aj (m) ρbj (n) Kitj (y 0 ) Kitj (y)
(9.5) j
X 1 Z
+ (s − v)−1 (1 − s − v)−1 ϕ̄ac (m, v) ϕbc (n, v)
c
4πi (1/2)
· Kv−1/2 (y 0 ) Kv−1/2 (y) dv .
1 X
−(s − )−2 ρ̄ak (m) ρbk (n) K0 (y 0 ) K0 (y) .
2
sk =1/2
Let h(t) be a test function which satisfies the conditions (1.63). Put
f (s) = 4π(s − 1/2)(sin πs)−1 h(i(s − 1/2)). Thus f (s) is holomorphic in
the strip ε ≤ Re s ≤ 1 − ε and bounded by (recall that s = 1/2 + it)
f (s) (|t| + 1)−1−ε | cosh πt|−1 .
Also f (s) = −f (1 − s). Multiply (9.8) through by f (s) and integrate
over the vertical line Re s = 1 − ε getting
1 1
Z Z
Zs (m, n) f (s) ds + Zs (m, n) f (s) ds
2πi 2πi
(1−ε) (ε)
1 1 1
Z
= δab δmn sin π s − f (s) ds
2π|n| 2πi 2
(1−ε)
1 f (s)
Z X
− ϕac (m, 1 − s) ϕbc (n, s) ds .
2πi c
2s − 1
(1−ε)
On the right side the diagonal part becomes −|n|−1 δab δmn h0 , where
1 +∞
Z
(9.9) h0 = t tanh(πt) h(t) dt ,
π −∞
and the continuous spectrum integrals yield
Z +∞ X
1 1 h(t)
ϕac m, − it ϕbc n, + it dt
−∞ c
2 2 cosh πt
after moving to the critical line. On the left side the second integral is
brought to the first one by moving to the line Re s = 1−ε. By Cauchy’s
theorem, the poles at s = sj and s = 1 − sj contribute
X h(tj )
−4π ρ̄aj (m) ρbj (n)
j
cosh πtj
Bruggeman-Kuznetsov formula 139
2
Z
Zs (m, n) f (s) ds
2πi
(1−ε)
X 4π p|mn|
−1/2 −1 ±
= |mn| c Sab (m, n; c) h
c
c
where
1
Z
+
h (x) = J2s−1 (x) f (s) ds
2πi
(1−ε)
if mn > 0, and h− (x) is the same integral with I2s−1 in place of J2s−1
if mn < 0. Moving to the critical line we get
+∞
h(t) t
Z
+
(9.10) h (x) = 2i J2it (x) dt
−∞ cosh πt
+∞
4
Z
−
(9.11) h (x) = K2it (x) sinh(πt) h(t) t dt .
π −∞
where
+∞
4
Z
(9.15) Kf (t) = cosh(πt) K2it (x) f (x) x−1 dx .
π 0
Theorem 9.4. Let mn < 0. For any f (x) satisfying (9.14) we have
X 4π p|mn|
−1
c Sab (m, n; c)f
c
c
X
(9.16) = Kf (tj ) ν̄aj (m) νbj (n)
j
X 1 Z ∞
+ Kf (t) η̄ac (m, t) ηbc (n, t) dt
c
4π −∞
where
π −1
(9.17) Bν (x) = 2 sin ν J−ν (x) − Jν (x) .
2
Note that B2it (x) ∈ L2 (R+ , x−1 dx). Therefore, the image of the trans-
form h 7→ h+ falls into the subspace spanned by the functions B2it (x),
t ∈ R. It is a very large subspace but not dense in L2 (R+ , x−1 dx). In-
deed the Bessel functions J` (x) of order ` ≥ 1, ` ≡ 1 (mod 2) are missed
because they are orthogonal to B2it (x) (see Appendix B.5). Given
f ∈ L2 (R+ , x−1 dx) put
Z +∞
(9.18) Tf (t) = B2it (x) f (x) x−1 dx
0
so Tf (t) B2it (x) is the projection of f (x) onto B2it (x). Then define the
continuous superposition of these projections by
Z +∞
∞
(9.19) f (x) = B2it (x) Tf (t) tanh(πt) t dt .
0
Bruggeman-Kuznetsov formula reversed 143
We also define
∞
2
Z
∞
(9.20) f = Tf (t) tanh(πt) t dt .
π 0
Incidentally, by (B.37) and (9.18) you can derive another formula for
f ∞ which involves f directly, namely
∞
1
Z
∞
(9.21) f = J0 (x) f (x) dx .
2π 0
X 4π √mn
∞ −1 ∞
δab δmn f + c Sab (m, n; c) f
c
c
X
(9.22) = Tf (tj ) ν̄aj (m) νbj (n)
j
X 1 Z ∞
+ Tf (t) η̄ac (m, t) ηbc (n, t) dt .
c
4π −∞
One can prove under the conditions (9.13) that (see Appendix B.5)
Z +∞
∞
(9.25) f (x) = xy J0 (xy) Hf (y) dy ;
1
144 Spectral theory of Kloosterman sums
where
Z +∞
(9.28) Nf (`) = J` (x) f (x) x−1 dx .
0
Note that J` (x) = i`+1 B` (x); hence Nf (`) = i`+1 Tf (i`/2). By virtue of
the Neumann series expansion for f 0 (x) our problem reduces to finding
an analogue of (9.12) with J` (x) as the test function attached to the
Kloosterman sums. The latter will bring us to the classical (holomor-
phic) cusp forms of weight k = ` + 1.
These are called automorphic forms of weight k with respect to the group
Γ. Throughout we asume that k is even and positive. The space Mk (Γ)
has finite dimension, precisely
X̀ 1 k hk
(9.30) dim Mk (Γ) = (k − 1)(g − 1) + 1− +
j=1
mj 2 2
Petersson’s formulas 145
Observe that y k f (z) ḡ(z) ∈ A(Γ) so it does not matter what fundamen-
tal domain is taken.
As in the proof of Theorem 3.2 one shows that the Fourier coeffi-
cients of a cusp form f normalized by hf, f ik = 1 satisfy the bound
X (k − 1)! ˆ
(9.34) |fa (n)|2 c−1
a N +k
(4πn)k−1
1≤n≤N
(k − 2)! X ˆ
(9.37) Pam (z) = fajk (m) fjk (z) ,
(4πm)k−1 j
where fˆajk (m) denotes the m-th Fourier coefficient of fjk in cusp a. On
the other hand we have the Fourier expansion of Pam (z) in cusp b due
to Petersson, Rankin and Selberg
∞
−k
X n (k−1)/2
(9.38) jσb (z) Pam (σb z) = P̂ab (m, n) e(nz) ,
n=1
m
say, with
X 4π √mn
k −1
(9.39) P̂ab (m, n) = δab δmn + 2πi c Sab (m, n; c) Jk−1
c
c
(for a proof apply the double coset decomposition (2.21) as in the case
of Eam (σb z|ψ) in Section 3.4). Comparing this with the n-th coefficient
on the right side of (9.37) we arrive at the following
Remarks. For k = 2 the Poincaré series (9.35) does not converge abso-
lutely (a proper definition in this case was given by Hecke); nevertheless,
(9.36)-(9.40) are true. The series of Kloosterman sums converges abso-
lutely if k ≥ 4 and at least conditionally if k = 2 by the spectral theory
of Kloosterman sums, which we have already established.
10.1. Introduction.
A truly beautiful formula has been derived from the spectral theo-
rem by A. Selberg (see [Se2]). The Selberg trace formula establishes a
quantitative connection between the spectrum and the geometry of the
Riemann surface Γ\H.
is called the trace. Suppose for a moment that Γ\H is a compact quo-
tient and K(z, w) is given by a smooth compactly supported function
k(u). Then K is of trace class. Integrating the spectral decomposition
X
K(z, z) = h(tj ) |uj (z)|2
j
149
150 The trace formula
we get
X
Tr K = h(tj )
j
XZ
Tr K = k(z, γz) dµz .
γ∈Γ F
[γ] = τ −1 γτ : τ ∈ Γ
where
XZ
Tr KC = k(z, γz) dµz .
γ∈C F
and
X X 1 Z +∞ 1 2
2
(10.8) h(tj ) |uj (z)| + h(r) Ea (z, + ir) dr
j a
4π −∞ 2
After the trace formula is established for the iterated resolvent and
the required convergence is not a problem, we shall relax the condition
a > s > 1 by analytic continuation. Then we recommend to the reader
to generalize the formula by contour integration. Of course, the result-
ing integrals will be the same as those previously developed however
receiving them through another channel is an attractive exercise.
Notice that we have changed the Eisenstein series into the truncated
ones because they match inside F (Y ).
Extending the integration to the whole of F we immediately get
an upper bound. For each point tj in the discrete spectrum we get 1
and for each r in the continuous spectrum we get, by (6.35),
(10.9) 1 1
= Tr (2ir)−1 Φ( − ir) Y 2ir − Φ( + ir) Y −2ir
2 2
ϕ0 1
+ 2h log Y − ( + ir) .
ϕ 2
Here Φ(s) is the scattering matrix, h is its rank, i.e. the number of
inequivalent cusps, ϕ(s) = det Φ(s) and
ϕ0
(s) = Tr Φ0 (s) Φ−1 (s) .
ϕ
Computing the spectral trace 155
For a proof of the last equation employ the eigenvalues of Φ(s) and a
unitary diagonalization. It follows from (10.9) that −(ϕ 0 /ϕ)(1/2 + ir)
is real and bounded below by a constant depending on the group. A
good upper bound is not known, but see (10.13).
Accordingly we need to evaluate the integral
+∞
1 h(r) 1 1
Z
Φ( − ir) Y 2ir − Φ( + ir) Y −2ir dr .
I(Y ) =
4π −∞ 2ir 2 2
We write
+∞
1 1 1
Z
I(Y ) = r−1 h(r) Φ( − ir) Y 2ir − Φ( ) dr
4πi −∞ 2 2
by giving and taking back Φ(1/2) and then exploiting the symmetry
h(r) = h(−r). We now move the integration upwards to Im r = ε
getting
1 1
Z
I(Y ) = −Φ( ) r−1 h(r) dr + O(Y −2ε )
2 4πi
Im r=ε
1 1
Z
(10.10) r−1 h(r) dr = − h(0) .
2πi 2
Im r=ε
1 1
I(Y ) = Φ( ) h(0) + O(Y −2ε ) .
4 2
With the truncated Eisenstein series EcY (z, s) the argument is more
involved. We use the inequality
Z +∞ Z +∞
2 −2
|Ws (iy)| y dy |s| |Ws (iy)|2 y −3 dy
Y Y /2
to estimate as follows
Z X Z +∞
Y 2 2
|Ec (z, s)| dµz = |ϕac (n, s)| |Ws (iny)|2 y −2 dy
n6=0 Y
Fa (Y )
Z +∞ X
|s| |ϕac (n, s) Ws (iny)|2 y −3 dy .
Y /2 n6=0
Computing the trace for parabolic classes 157
We have shown above that the truncated traces over cuspidal zones are
absorbed by the error term in (10.11) so the inequality (10.11) turns
into equatility.
where H(Y ) is the region of the upper half plane with the cuspidal zones
removed. Conjugating by the scaling matrix σa we get
Z
Y
Tr KC = k(z, z + `) dµz .
B\σa H(Y )
Notice that the set B\σa H(Y ) is contained in the box {z : 0 < x ≤ 1,
0 < y ≤ Y } and it contains the box {z : 0 < x ≤ 1, Y 0 < y ≤ Y }
where Y 0 Y = c−2
a . Therefore,
Z 1 Z Y Z 1 Z Y
Y
k(z, z + `) dµz ≤ Tr KC ≤ k(z, z + `) dµz .
0 Y0 0 0
Here we have
1 Y Y
` 2 −2
Z Z Z
k(z, z + `) dµz = k ( ) y dy
0 0 0 2y
Z +∞
−1
= |`| k(u) u−1/2 du .
(`/2Y )2
where Z +∞ √
k(u) u−1/2 log 2Y
L(Y ) = 2 u + γ du .
0
+∞
1
Z
k(u) u−1/2 du = q(0) = g(0) .
0 2
+∞
−1 +∞ v log u
Z Z Z
−1/2
k(u) u log u du = p du dq(v)
0 π 0 0 u(v − u)
−1 +∞ 1 log uv
Z Z
= p du dq(v)
π 0 0 u(1 − u)
Z 1
1 1
Z +∞
1 log u du
Z
= q(0) p du − p log v dq(v)
π 0 u(1 − u) π 0 u(1 − u) 0
(do not try to integrate by parts !). In the last line the first integral is
−2π log 2, the second integral is π and the third one is
+∞ +∞
r
Z Z
log v dq(v) = log(sinh ) dg(r)
0 0 2
+∞
r 1
Z
log(sinh ) dg(r) = g(0)(γ + log 2) − h(0)
0 2 4
(10.17) Z +∞
1
+ h(t) ψ(1 + it) dt
2π −∞
160 The trace formula
where
∞
Γ0 X 1 1
(10.18) ψ(s) = (s) = −γ − − .
Γ n=0
n + s n + 1
1
Z
0
g (r) = − eirt h(t) t dt
2πi
Im t=ε
and employ the formula for the Laplace transform of log(sinh r/2) (see
[Gr-Ry, 4.331.1, p. 573, and 4.342.3, p. 575])
+∞
r 1
Z
(10.19) log(sinh ) de−νr = γ + log 2 − + ψ(1 + ν) .
0 2 2ν
1
L(Y ) = (2s − 1)−1 (log Y − ψ(s + ) − log 2) + (2s − 1)−2
(10.21) 2
− (the same for s = a) .
+∞
1
Z
(10.23) k(0) = r tanh(πr) h(r) dr .
4π −∞
Tr P = N P 1/2 + N P −1/2 .
In particular for k(u) given by (10.1) we get from (10.3) and (10.25)
that the trace for the class C = P ` is
Summing over ` 6= 0 we get the total trace from all hyperbolic classes
having equivalent fixed points
X ∞
X
−1
Tr KC = 2 (2s − 1) (ps+k − 1)−1 log p
(10.27) C=P ` k=0
− (the same for s = a) .
1
Z Z
Tr KC = k(z, k(θ`)z) dµz = k(z, k(θ`)z) dµz
S m H
1 +∞ g(r) cosh(r/2)
Z
(10.29) Tr KC = dr .
m 0 cosh r − cos(2π`/m)
If one prefers to have an expression in terms of h rather than g we state
another formula (a proof is cumbersome)
valid for Re µ > −1 and t 6= 2πn (see [Gr-Ry, 3.543.2, p. 357]). Hence,
Z +∞ ∞
r 1 X sin(2k + 1)α
e−µr cosh (cosh r − cos 2α)−1 dr = .
0 2 sin α k + µ + 1/2
k=0
Computing the trace for elliptic classes 165
This yields
∞
π` −1 X π`
Tr KC = (2s − 1) m sin (s + k)−1 sin(2k + 1)
m m
k=0
− (the same for s = a) .
By the same token we can borrow the Euler constant to get (see (10.18))
X −1 X s + k π`
= ψ sin(2k + 1) .
m m m
k 0≤k<m
Hence, we arrive at
−1 X s + k sin(2k + 1)π`/m
Tr KC = ψ
(10.31) (2s − 1) m2 m sin π`/m
0≤k<m
that
X sin(2k + 1)π`/m
= m − 2k − 1.
sin π`/m
0<`<m
Therefore,
X 1 X 2k + 1 s + k
(10.32) Tr KC = −1 ψ .
(2s − 1) m m m
C=R` 0≤k<m
166 The trace formula
All parts from which to build the trace formula have now been
manufactured. Let us first assemble these for the particular pair h, g
given by (10.2) and (10.3).
The series over the discrete spectrum and the integral accounting
for the continuous spectrum in the Resolvent Trace Formula converge
absolutely due to (7.11) and (10.13). Therefore the Dirichlet series over
the hyperbolic classes also converges absolutely in Re s > 1; in fact one
gets quickly from the lowest eigenvalue λ0 = 0 that
X 1
(10.36) p−s log p ∼ , as s → 1+ .
s−1
P
1 ∞ −ϕ0 1
X Z
h(rj ) + h(r) +ir dr
j
4π −∞ ϕ 2
|F | ∞
Z
= h(r) r tanh(πr) dr
4π −∞
(10.37)
∞
XX
+2 (p`/2 −p−`/2 )−1 g(` log p) log p
P `=1
Z∞
X X π` −1 cosh π(1−2`/m)r
+ 2 m sin h(r) dr
m cosh πr
R 0<`<m −∞
h(0) 1
+ Tr I − Φ − h g(0) log 2
4Z 2
∞
h
− h(r) ψ(1 + ir) dr .
2π −∞
168 The trace formula
Remarks. The above terms come from (10.11), (10.20), (10.23), (10.25)
and (10.30). The series and integrals converge absolutely. For alterna-
tive expressions see (10.16) and (10.29).
where the outer product ranges over the primitive hyperbolic conjugacy
classes in Γ of norm p = N P > 1. The infinite product converges
absolutely; therefore it does not vanish in Re s > 1. Differentiating in
s it gives
∞
Z0 X X log p
(10.39) (s) = ;
Z ps+k − 1
P k=0
therefore 0 0
−1 Z −1 Z
(2s − 1) (s) − (2a − 1) (a)
Z Z
is exactly the contribution of the hyperbolic motions to the Resolvent
Trace Formula (10.35). This formula yields the analytic continuation
of (Z 0 /Z)(s) to the whole complex s-plane; the key point is that all
poles of (Z 0 /Z)(s) are simple and have integral residues. This is clear
in every term of (10.35) except for the contributions from the elliptic
classes and the identity motion which have to be combined together
into
X |F | X 1
−1 −1
m log m + (2s − 1) Rm (s) − ψ(s) +
m
2π m
m
where the integration goes along any curve which joins a with s avoiding
poles. Since Z(s) = Z(a) F (s) this proves
The remaining zeros and poles of Z(s) in the half plane Re s < 1/2
can be likewise determined by examining the Resolvent Trace Formula
(in order to interpret the continuous spectrum integral in (10.35) use the
expansion (11.9) of −ϕ0 (s)/ϕ(s) into simple fractions and the residue
theorem). Besides one can also derive a functional equation of type
(10.40) Z(s) = Ψ(s) Z(1 − s)
where Ψ(s) is a certain meromorphic function of order 2 which can
be written explicitely in terms of elementary functions, the Euler Γ-
function and the Barnes G-function,
∞
s/2 −s(s+1)/2−γs2 /2
Y s n −s+s2 /2n
G(s + 1) = (2π) e 1+ e .
n=1
n
As a matter of fact one can attach to Z(s) a finite number of local factors
corresponding to the identity, the parabolic and the elliptic classes so
that the complete zeta-function satisfies a simpler equation Z ∗ (s) =
Z ∗ (1 − s) (cf. [Vig1] and [Fi]).
p + 1 −δ (p+p−1 ) 1 −δp
e log p = 1 + O e log p .
p−1 p
Estimating the error term trivially we are left with the following
The elliptic and parabolic classes contribute no more than the above
bound. Gathering these estimates we arrive at
X X
(10.42) q(log p) log p = s−1
j X
sj
+ O(Y + X 1/2 T ) .
P 1/2<sj ≤1
Hence, we can drop in (10.42) the excess over p ≤ X whithin the error
term already present. Then we choose Y = X 3/4 to minimize the error
term and conclude with the following asymptotic expression
172 The trace formula
173
174 The distribution of eigenvalues
By (B.11) we get
|F | 2
(11.2) NΓ (T ) + MΓ (T ) ∼ T .
4π
This asymptotic is called Weyl’s law (to pay tribute for numerous
results he established in similar contexts). The strongest form of Weyl’s
law ever established for a general surface Γ\H is
|F | 2 h
(11.3) NΓ (T ) + MΓ (T ) = T − T log T + cΓ T + O(T (log T )−1 ) ,
4π π
where cΓ is a constant (cf. [He1, Theorem 2.28] and [Ve, Theorem 7.3]).
176 The distribution of eigenvalues
(11.4) MΓ (T ) T log T ;
whence,
|F | 2
(11.5) NΓ (T ) = T + O(T log T ) .
4π
In this case the situation is clear. The determinant of the scattering
matrix ϕ(s) is shown to be a product of Dirichlet L-functions so it is
meromorphic of order 1 thus leading to (11.4). An immediate conse-
quence of the resulting asymptotic (11.5) is that the congruence groups
have infinitely many linearly independent cusp forms. This is a rather
indirect argument. Sadly enough, not a single cusp form has yet been
constructed for the modular group.
λ1/2
(11.6) m(λ) .
log(λ + 3)
This easily obtained bound is the best available today, even for the
modular group for which presumably m(λ) = 1. Any improvement
upon (11.6) in the order of magnitude would be welcomed.
Weyl’s law 177
with a1 6= 0 and 0 < b1 < b2 < · · · < bn → +∞. The series converges
absolutely in Re s ≥ 1 + ε so it is bounded, and therefore, by Stirling’s
formula
b2s
1 ϕ(s) |s|
−h/2
.
Also notice that ϕ(s) does not vanish in a half plane Re s > σ0 with σ0
sufficiently large. In the half-plane Re s ≥ 1/2 it has a finite number of
poles, all in the segment 1/2 < sj ≤ 1. Following Selberg we put
Y s − sj
ϕ∗ (s) = b2s−1
1 ϕ(s) .
s − 1 + sj
1/2<sj ≤1
ϕ∗ (s) |s|−h/2 in Re s ≥ 1 + ε ,
ϕ∗ (s) ϕ∗ (1 − s) = 1 in the whole s-plane ,
|ϕ∗ (s)| = 1 on the line Re s = 1/2 ,
|ϕ∗ (s)| ≤ 1 in the half plane Re s ≥ 1/2 .
We appeal to Jensen’s formula
Z 1 X
log |f (e2πiθ )| dθ = − log |zj |
0 j
where the summation ranges over the zeros of ϕ∗ (s) in Re s > 1/2
different from 1. But a zero zj corresponds to a pole sj = 1 − zj by the
functional equation; therefore, the above inequality can be written as
X0 1 − s
j
log < +∞
j
sj
where sj ranges over the poles of ϕ∗ (s) in Re s < 1/2 different from 0.
Putting sj = βj + iγj we rewrite this inequality again as
X0
(11.7) (1 − 2βj ) |sj |−2 < +∞ .
βj <1/2
ϕ∗ 0 X 1 1
(11.8) − (s) = − .
ϕ∗ s − sj s − 1 + s̄j
βj <1/2
ϕ0 X 1 1
(11.9) − (s) = − + 2 log b1
ϕ j
s − s j s − 1 + s̄ j
where sj ranges over all poles of ϕ(s) with proper multiplicity (notice
that all poles of ϕ(s) are in the strip 1 − σ0 ≤ Re s ≤ 1).
On the critical line s = 1/2 + it the terms of (11.8) are positive
−1
(1 − 2βj ) (1/2 − βj )2 + (t − γj )2 > 0,
ϕ0
(11.10) − (s) > 2 log b1 + O(|s|−2 ) .
ϕ
The residual spectrum and the scattering matrix 179
The poles of Φ(s) in the segment 1/2 < sj ≤ 1 yield the so-called
residual eigenvalues λj = sj (1−sj ). There is always one at s0 = 1 which
corresponds to the lowest eigenvalue λ0 = 0 with constant eigenfunction
u0 (z) = |F |−1/2 .
There are groups having many residual points arbitrarily close to 1, but
they are not congruence groups.
whence ϕ∞∞ (s) has no poles in Re s ≥ 1/2 except for a simple one at
s = 1. The same can be shown for other cusps.
How small can the first positive eigenvalue λ1 in the discrete spec-
trum of Γ\H possibly be ? Undoubtedly it is an important question and
an intricate one too. In particular, we wish to know if λj ≥ 1/4 which
would mean that the positive discrete spectrum lies on the continuous
one (if Γ has cusps). We shall call λj with 0 < λj < 1/4 exceptional,
equivalently λj = sj (1 − sj ) with 1/2 < sj < 1, emphasising they are
not welcomed. In practice the exceptional eigenvalues distort rather
than simplify results. That exceptional eigenvalues exist, both the cus-
pidal and the residual ones, for some groups was known to A. Selberg,
proved by B. Randol and constructed by M. N. Huxley (just to name
a few from a long list of investigators).
g+1
(11.14) λ1 ≤ 2 ≤ 6,
g−1
Some of the methods used to establish the above results adapt well
to general quotients F = Γ\H of finite volume. Thus P. G. Zograf [Zo]
has generalized (11.14) to show that
(11.15) λ1 ≤ 8π (g + 1) |F |−1
for any F which is not compact and has volume |F | ≥ 32π (g + 1);
in particular such a surface has exceptional eigenvalues. Suppose the
corresponding group Γ is a subgroup of P SL2 (Z) of finite index n ≥ 1
and that F = Γ\H has exactly one cusp (such groups have been studied
by H. Petersson [Pe2]). One shows that the Eisenstein series for Γ and
P SL2 (Z) differ by factor n−s ; therefore Γ has only λ0 = 0 in the residual
spectrum. Consequently, λ1 is cuspidal, and using (11.15) λ1 can be
made very small; just take a subgroup of signature (0; 2, . . . , 2; 1) with
a large number of elliptic classes of order 2.
Proof. Let u be a cusp form for Γ with kuk = 1. By Lemma 4.1 the
eigenvalue of u is equal to the energy integral (Dirichlet)
Z
λ= |y ∇u(z)|2 dµz
F
Small eigenvalues 183
1
F 0 = z : |x| < , |z| > 1
2
and γν ranges over the coset representatives. Consider also another fun-
damental polygon for the modular group obtained from F 0 by applying
the involution ω = 1 −1 . Observe that F 00 = F 0 ∪ ωF 0 covers the
√
strip {z : |x| < 1/2, y > 3/2}, which contains F 0 . Take the unions
F1 = ∪(F 0 + b) and F2 = ∪(F 00 + b) over 0 ≤ b < B. We obtain
X Z
2λB = |y ∇u(z)|2 dµz
ν
γν F 2
XZ
= |(y ∇)u(γν z)|2 dµz
ν F2
XZ +∞ Z B
≥ √ |(y ∇)u(γν z)|2 dµz .
ν 3/2 0
Here for each ν we apply the Fourier expansion of u(γν z) in the cusp
aν = γν ∞, X nx
u(γν z) = cνn (y) e
qν
n6=0
π 2 Z +∞ Z B
=3 √ |u(γν z)|2 dµz
q 3/2 0
π 2 Z
≥3 |u(z)|2 dµz .
q
γν F 1
Proof. Apply Theorem 9.2 and Weil’s bound (2.25) for Kloosterman
sums. It shows that the series Zs (m, n) converges absolutely in Re s >
3/4 so its poles at sj have Re sj ≤ 3/4; whence, λj = sj (1 − sj ) ≥ 3/16.
is most desired for applications. We shall call any such result with expo-
nent c > 0 a density theorem. A density theorem with sufficiently large
exponent completely substitutes for the Selberg eigenvalue conjecture
in the same fashion as the density theorems for zeros of L-functions
serve in the absence of the Riemann hypothesis.
Let us first deal with the simplest case of a hyperbolic group. Our
approach is an excercise with the trace formula (10.37). Take the test
function
sin tL 4
(11.21) h(t) =
tL
so that the Fourier transform g(x) is supported in the segment
[−4L, 4L]. Choose 4L to be the length of the shortest closed geodesic
on Γ\H or any smaller number. Then there is no contribution from
the hyperbolic classes to the trace formula (the parabolic and elliptic
classes do not exist by assumption) and therefore we are left with the
cute identity
X sin tj L 4 |F | +∞ sin tL 4
Z
= tanh(πt) t dt .
j
tj L 2π 0 tL
On the right side apply the bound tanh(πt) ≤ πt getting π|F |(2L)−3 .
On the left side discard all but the exceptional points for which use the
bound sinh x ≥ x(2x + 1)−1 ex to get
Inserting these estimates into the above identity we establish the fol-
lowing density theorem
X
(11.22) e4(sj −1/2)L ≤ π |F | (2L)−3 (L + 1)4 .
1/2<sj ≤1
1 2s − 1
Z
+
h (x) = −i J2s−1 (x) h i(s − ) ds
2 sin πs
(σ)
√
with 1/2 ≤ σ < 1. By the trivial estimate |Jν (x)| ≤ π|xν Γ(ν +1/2)−1 |
this yields h+ (x) (xe2L )2σ−1 . Furthermore, by Weil’s bound for
Kloosterman sums (2.25) we get
X 3 −2
c−2σ |S(n, m; c)| τ (nq) (n, q)1/2 q −2σ+1/2 σ −
4
c≡0 (mod q)
189
190 Hyperbolic lattice-point problems
In both asymptotics the error term has been sharpened to O(xθ+ε ) with
various θ many times by methods of exponential sums. One expects
that the best exponent should be θ = 1/4 whereas the best known one
is θ = 23/73 due to M. N. Huxley [Hu3].
where H(t) is any decreasing majorant of |h(t)|, and the implied con-
stant depends on the group Γ and the points z, w. It remains to choose
the function k(u) and survey its transform h(t).
For the hyperbolic circle problem the aim is to estimate the number
n o
P (X) = # γ ∈ Γ : 4u(γz, w) + 2 ≤ X .
Hyperbolic lattice-point problems 191
Γ(s − 1/2) s
(12.8) h(t) = π 1/2 X + O(Y + X 1/2 ) ,
Γ(s + 1)
This yields an upper bound for P (X) through (12.7). A lower bound of
the same type is obtained by applying the above result with X replaced
by X − Y . Now clearly, the optimal choice is Y = X 2/3 giving (12.6).
(12.10) ad − bc = 1
(12.11) a 2 + b2 + c 2 + d2 ≤ X .
(12.16) ad − bc = n
√
By (12.18) it follows that Kn (z, w) = n Tn K(z, w). On the other
hand applying Tn to the spectral decomposition (7.17) we infer that
(asumming that uj are eigenfunctions of Tn , see Section 8.6)
X
n−1/2 Kn (z, w) = λj (n) h(tj ) uj (z) ūj (w)
j
+∞
1 1 1
Z
+ ηt (n) h(t) E(z, + it) E(w, + it) dt .
4π −∞ 2 2
Now insert the trivial bound |λj (n)| ≤ λ0 (n) = σ(n) n−1/2 . The rest
of the proof proceeds as in the case n = 1 but with k(u) replaced by
k(nu). It gives
√ X X
2/3
Pn (X) = 3 λ0 (n) n +O
n n
thus completing the proof of Theorem 12.4.
If one applies all these results to the group (12.13) and transform
the points as in the proof of Theorem 12.3, one will end up with
Remark. The error term O(X 2/3 ) in the above results has never been
improved even slightly for any group. Recently F. Chamizo [Ch] has
established numerous sharper results on average while R. Phillips and Z.
Rudnick [Ph-Ru] gave insightful analysis of the limit for various error
terms relevant to the hyperbolic circle problem. Both works suggest
that the exponent 2/3 might be lowered to any number greater than
1/2.
Chapter 13
13.1. Introduction.
195
196 Spectral bounds for cusp forms
T
1 1
X Z
2
(13.1) |uj (z)| + |E(z, + it)|2 dt T 2 + T y
2π 0 2
0<tj <T
for any ε > 0 and z ∈ H; the implied constant depends on ε alone. This
estimate, if true, is very deep since it has the Lindelöf hypothesis for
certain L-functions as a consequence. From (13.1) we can only deduce
(13.2) with ε = 1/2 while we already know that it holds with ε = 1/4.
Standard bounds 197
We choose
cosh(πt/2) cosh πT /2
(13.4) h(t) = 4π 2
cosh πt + cosh πT
so that h(t) > 0 everywhere and h(t) 1 if t = T + O(1). The Fourier
transform of h(t) is equal to
cos xT
(13.4) g(x) = 2π .
cosh x
Hence one can show that the Selberg/Harish-Chandra transform satis-
fies
Hence,
X
(13.8) |uj (z)|2 T .
T <tj <T +1
This bound is as sharp as one can get by playing with test functions
alone, save for a factor (log λj )−1 .
The same bound can be established for eigenstates on any compact
Riemann surface X. In such generality the exponent 1/4 is best possible
as the example of the sphere X = S 2 shows. If, however, X is negatively
curved, then a stronger bound is probably true.
where
νn = # γ ∈ Γn : γz = z .
We are looking for estimates which are uniform in n. To simplify we
consider only the special point z = i. In this case
νn = # a2 + b2 = n = r(n) .
Applying the Hecke operator 199
a b
For γ = c d ∈ Γn with γi 6= i we have
4 n u(γi, i) = a2 + b2 + c2 + d2 − 2n
(13.10)
= (a − d)2 + (b + c)2 ≥ 1 .
where X mn
r(m, n) = dr .
d2
d|(m,n)
where kak1 denotes the `1 -norm. Next apply r(m, n) (m, n) (mn)ε
and Cauchy’s inequality to the leading term getting
X X 2
an λj (n) |uj (z)|2 T kak22 + T 1/2 N 3/2 kak21 N ε .
h(tj )
j n≤N
On the left side the terms are non-negative, so we can drop all but one
getting
1/2
L2j |uj (z)|2 tj kak22 + tj N 3/2 kak21 N ε
(13.12)
where X
Lj = an λj (n) .
n≤N
1/2 ε
|uj (z)|2 tj N −1 + tj N 3/2 N tj
1/5
by (13.12). Choosing N = tj we conclude that
1/5+ε
(13.14) |uj (z)| λj
X N 1/2
λj (p)2 − λj (p2 ) ∼
Lj = .
√ 2 log N
p≤ N
Moreover, we have
X
kak22 = λj (p)2 + 1 tεj N 1/2
√
p≤ N
1/4
Choosing N = tj we obtain the following unconditional result
Remarks. Using more refined estimates one can get (13.15) with the
exponent 5/24 in place of 7/32. Also the result holds true for any z ∈ H
so it yields a bound for the L∞ -norm
5/24+ε
(13.16) kuj k∞ λj .
The same estimate has been established for the eigenfunctions with
respect to the quaternion group [Iw-Sa].
Here uj is normalized so that dµj has total mass equal to 1. The total
mass of dνt is infinity since the Eisenstein series cannot be normalized.
In the language of quantum mechanics dµj represents the probability
density of a particle being in the state uj . Z. Rudnick and P. Sarnak
[Ru-Sa] suggest that (the unique ergodicity conjecture)
(13.18) µj → µ , as j → ∞ .
This has been proved for subsequences of full density (the quantum
ergodicity theorem) by A. I. Shnirelman, S. Zelditch and Y. Colin de
Constructing an amplifier 203
1/2
for any fixed h 6= 0 and N λj (these are building blocks for shifted
convolution L-functions). Therefore, the classical problem concerning
bounds for such sums is given a new face. A usual heuristic about
oscillatory sums led Sarnak to the stronger quantitative form of the
conjecture, namely that
Z Z
−1/4+ε
(13.20) f (z) dµj (z) = f (z) dµ(z) + O(λj )
X X
for any f ∈ C0∞ (X). Then, together with W. Luo [Lu-Sa] they showed
that the above approximation holds true on average with respect to the
spectrum. More precisely,
They also gave estimates for discrepancy showing that the expo-
nent −1/4 in (13.20) is best possible.
205
206 Appendix A. Classical analysis
If λ ∈
/ R this shows that the operator T −λ has zero kernel, the resolvent
is defined on (T − λ) D(T ) and it is bounded by |Im λ|−1 . It remains
to show that (T − λ) D(T ) = H. Suppose g ∈ H is orthogonal to
(T − λ) D(T ) so
(A.2) Rλ − Rγ = (λ − γ)Rλ Rγ .
where the last inner product is bounded by |Im λ|−2 |Im γ|−1 kf k kgk.
Hence, it is plain that the limit exists as γ → λ, and it is equal to
d
hf, Rλ gi = hf, Rλ Rλ gi = hRλ f, Rλ gi
dλ
because Rλ∗ = Rλ̄ .
λ1 0
U −1 AU = .. , for every A ∈ F .
.
0 λn
The Hilbert-Schmidt integral operators 209
Λ= ..
.
0 λn
and a unitary matrix U such that A = U ΛU −1 . Moreover, if A is real,
then U can be chosen real.
X
(A.3) f (z) = hf, uj i uj (z) .
j≥0
(A.5) (I − λK)g = 0
which converges absolutely in the disc |λ| < kKk−1 . For λ in this disc
the Neumann series yields an L2 -solution. This solution is unique up
to a function vanishing almost everywhere because the homogeneous
equation (A.5) implies
is majorized by
∞
X
|K(x, y)| + |λ| A(x) B(y) |λ|j kKkj
j=0
vol D = V < +∞ ,
(A.12)
|K(x, y)| ≤ K < +∞ .
By Hadamard’s inequality
YX
|det (aij )|2 ≤ |aij |2
j i
so we get
√
ξ1 , . . . , ξ m
(A.14) K ≤ ( m K)m .
η1 , . . . , η m
Let us denote
Z Z
ξ1 , . . . , ξ m
Cm = · · · K dξ1 · · · dξm ,
ξ1 , . . . , ξ m
Z Z
x, ξ1 , . . . , ξm
Cm (x, y) = · · · K dξ1 · · · dξm .
y, ξ1 , . . . , ξm
By (A.14) we get
√
|Cm | ≤ ( m K V )m ,
√
|Cm (x, y)| ≤ ( m + 1 K)m+1 V m .
√ 2
is majorized by (use the inequality ( m x)m < m! e2x )
∞ √
X m |λ| K V )m 2
1+ < 2 e8(|λ|KV )
1
m!
Cm (x, y)
Z Z
ξ1 , . . . , ξ m
= ··· K(x, y) K
ξ1 , . . . , ξ m
m
X
` ξ1 , ξ 2 , . . . . . . . . . , ξ m
+ (−1) K(x, ξ` ) K dξ1 · · · dξm
y, ξ1 , . . . , ξˆ` , . . . , ξm
`=1
= Cm K(x, y)
m Z
ξ` , ξ1 , . . . , ξˆ` , . . . , ξm
X Z Z
− K(x, ξ` ) · · · K dξ1 · · · dξm
y, ξ1 , . . . , ξˆ` , . . . , ξm
`=1
m Z
X
= Cm K(x, y) − K(x, ξ` ) Cm−1 (ξ` , y) dξ` .
`=1
which is the same one as for the resolvent Rλ (x, y) (see (A.11)). Since
for small |λ| the Fredholm equation has unique solution, we conclude
that (A.13) is true for all λ ∈ C by analytic continuation.
Finally, from (A.10) and (A.13) we conclude that for any λ with
D(λ) 6= 0 the unique solution to the Fredholm equation is given by
λ
Z
(A.19) g(x) = f (x) + Dλ (x, y) f (y) dy .
D(λ) D
(A.20) Tg = f
(A.24) T G(y, y 0 ) = 0 if y 6= y 0 .
we obtain
+∞ y
∂ ∂
Z Z
0 0 0
G(y, y ) g(y ) dy = G(y, y 0 ) g(y 0 ) dy 0
∂y 0 0 ∂y
Z +∞
∂
+ G(y, y 0 ) g(y 0 ) dy 0
y ∂y
because the terms G(y, y − 0) g(y) and −G(y, y + 0) g(y) cancel out by
continuity. Differentiating again we get
+∞ +∞
∂2 ∂2
Z Z
0 0 0
G(y, y ) g(y ) dy = G(y, y 0 ) g(y 0 ) dy 0
∂y 2 0 0 ∂y 2
∂ ∂
+ G(y, y − 0) − G(y, y + 0) g(y) .
∂y ∂y
+∞ ∂ ∂
Z
0 0 0
T G(y, y ) g(y ) dy = G(y, y − 0) − G(y, y + 0) g(y) .
0 ∂y ∂y
∂ ∂
(A.25) G(y, y − 0) − G(y, y + 0) = 1 .
∂y ∂y
Green function of a differential equation 217
Z +∞
(B.1) Γ(s) = e−x xs−1 dx .
0
219
220 Appendix B. Special functions
+∞ ∞
(−1)n
Z X
(B.2) Γ(s) = e−x xs−1 dx + (s + n)−1 .
1 n=0
n!
Thus Γ(s) has simple poles at non-positive integers. We have the Weier-
strass product
∞
Y s −1 1 s
s Γ(s) = 1+ 1+
n=1
n n
(B.3) ∞
−γs
Y s s/n
=e 1+ e
n=1
n
Functional equation:
Duplication formula:
1
(B.6) Γ(s) Γ(s + ) = π 1/2 21−2s Γ(2s) .
2
is valid in the angle |arg s| < π − ε with the implied constant depending
on ε. Hence,
t it
(B.8) Γ(σ + it) = (2π)1/2 tσ−1/2 e−πt/2 (1 + O(t−1 ))
e
The hypergeometric functions 221
It satisfies
1 X s + k
(B.10) ψ(s) = ψ + log m
m m
0≤k<m
1
Γ(u) Γ(v)
Z
(B.12) xu−1 (1 − x)v−1 dx = ,
0 Γ(u + v)
+∞
Γ(s − 1/2)
Z
(B.13) (x2 + 1)−s dx = π 1/2 .
−∞ Γ(s)
α F (α + 1) − β F (β + 1) = (α − β) F ,
α F (α + 1) − (γ − 1) F (γ − 1) = (α − β + 1) F ,
where only the shifted arguments are displayed; the other ones remain
unchanged.
z
(B.17) F (α, β; γ; z) = (1 − z)−α F α, β; γ; ,
1−z
1 1
(B.18) F (2α, 2β; α + β + ; z) = F (α, β; α + β + ; 4z(1 − z)) .
2 2
These formulas are meaningful for |z| < 1/2. In extended ranges the
values of F (α, β; γ, z) must be interpreted as those obtained by a suit-
able analytic continuation.
If either α or β but not γ is a non-positive integer, then the hyper-
geometric series terminates at a finite place; thus, it is a polynomial.
In particular, we obtain
1 − z
F n + 1, −n; 1; = Pn (z) ,
2
1 1 − z
F n, −n; ; = Tn (z) ,
2 2
Γ(ν + m + 1) π
Z p
m
(B.24) Pν (z) = (z + z 2 − 1 cos α)ν cos(mα) dα .
π Γ(ν + 1) 0
1 dn 2
(B.25) Pn (z) = (z − 1)n .
2n n! dz n
The generating function of the Legendre polynomials is
∞
X
(B.26) Pn (z) xn = (1 − 2zx + x2 )−1/2 .
n=0
The Legendre functions Pnm (z) vanish identically if m > n. Those with
0 ≤ m ≤ n, also called spherical harmonics, give a complete system of
eigenfunctions of the Laplacian on the sphere S 2 = {x2 + y 2 + z 2 = 1}.
The eigenvalues are λ = n(n + 1), and the λ-eigenspace has dimension
2n + 1. In the polar coordinates (x, y, z) = (sin θ cos ϕ, sin θ sin ϕ, cos θ)
the Laplacian is given by
∂2 1 ∂ 1 ∂2
+ + ,
∂θ2 tan θ ∂θ (sin θ)2 ∂ϕ2
(B.27) z 2 f 00 + zf 0 + (z 2 − ν 2 ) f = 0
The Bessel functions 225
where for ν = n the ratio is defined by taking the limit. Then Jν (z),
Yν (z) are always linearly independent, since the Wronskian is
(B.31) z 2 f 00 + zf 0 − (z 2 + ν 2 )f = 0 .
Both Iν (z), I−ν (z) are linearly independent if and only if ν is not an
integer. If ν = n is an integer, we have the relation
We set
π
(B.34) Kν (z) = (sin πν)−1 (I−ν (z) − Iν (z))
2
and obtain a pair Iν (z), Kν (z) of linearly independent solutions to
(B.31) for all ν since the Wronskian is
for x > 0 as ε tends to zero. The same discontinuity appears for Iν (z).
Each of the functions Jν , Yν , Iν , Kν is expressible in terms of
others in some way.
The Bessel functions of different order are related by many recur-
rence formulas. For the J-function we have
The same formulas hold for the Y -function. For the I-function we have
And for the K-function the above formulas hold with negative sign on
the right-hand sides.
The Bessel functions 227
2 1/2 2 1/2
J1/2 (z) = sin z , Y1/2 (z) = − cos z ,
πz πz
2 1/2 π 1/2
I1/2 (z) = sinh z , K1/2 (z) = e−z .
πz 2z
Applying the recurrence formulas one can find elementary expressions
for Bessel functions of any order ν which is half of an odd integer.
The four functions Jν (y), Yν (y), Iν (y), Kν (y) in the real positive
variable y have distinct asymptotic behaviour. For y < 1 + |ν|1/2 one
will get good approximations by the first terms in the power series. For
y > 1 + |ν|2 we have
2 1/2 π π 1 + |ν|2
Jν (y) = cos y − ν − +O ,
πy 2 4 y
(B.35) 2 1/2 1 + |ν|2
π π
Yν (y) = sin y − ν − +O ,
πy 2 4 y
and
1 + |ν|2
−1/2 y
Iν (y) = (2πy) e 1+O ,
y
(B.36) π 1/2 1 + |ν|2
Kν (y) = e−y 1 + O ,
2y y
1 −1 z ν +∞ 2
Z
1/2
Kν (z) = π Γ(ν + ) (t − 1)ν−1/2 e−tz dt
2 2 1
Z +∞
1 z
−ν
= π −1/2 Γ(ν + ) (t2 + 1)−ν−1/2 cos(tz) dt
2 2 0
π 1/2 1 −1 −z +∞ −t t ν−1/2
Z
= Γ(ν + ) e e t 1+ dt
2z 2 0 2z
228 Appendix B. Special functions
+∞
1 z 1 −ν−1
Z
= exp − t+ t dt
2 0 2 t
Z +∞
= e−z cosh t cosh(νt) dt ,
0
where Re z > 0 and Re ν > −1/2. The Mellin transforms are often
expressed as a product of gamma functions. For example we have
+∞
s+ν s−ν
Z
Kν (x) xs−1 dx = 2s−2 Γ( ) Γ( ),
0 2 2
+∞
1
Z
e−x Kν (x) xs−1 dx = 2−s π 1/2 Γ(s + )−1 Γ(s + ν) Γ(s − ν) ,
0 2
if Re s > |Re ν| and
Z +∞ Y s±µ±ν
Kµ (x) Kν (x) xs−1 dx = 2s−3 Γ(s)−1 Γ( ),
0 2
if Re s > |Re µ| + |Re ν|.
Similar formulas hold for the Mellin transform involving J-funct-
ions. One of these at s = 0 yields
Z +∞
2 sin π(µ − ν)/2
(B.37) Jµ (x) Jν (x) x−1 dx = ,
0 π (µ − ν) (µ + ν)
if Re (µ + ν) > 0. This formula shows, among other things, that the
J-functions which are distinct but have the same order modulo even
integers are orthogonal with respect to the measure x−1 dx.
Therefore Tf (t) B2it (x) is the projection of f (x) onto B2it (x). If f (x)
satisfies (B.46), we define the continuous superposition of these projec-
tions by the integral
Z +∞
∞
(B.49) f (x) = Tf (t) B2it (x) tanh(πt) t dt .
0
X Z 1
(B.51) 2` J` (x) J` (y) = xy u J0 (ux) J0 (uy) du .
0<` odd 0
Hence,
Z 1
0
(B.52) f (x) = ux J0 (ux) Hf (u) du
0
233
234 References
[Fr-Kl] R. Fricke and F. Klein, Vorlesungen über die Theorie der automorphen
Funktionen I, II. Leipzig, 1897 and 1912.
[Go] A. Good, Local Analysis of Selberg’s Trace Formula. Springer Lecture
Notes in Math. 1040, 1983.
[Go-Sa] D. Goldfeld and P. Sarnak, Sums of Kloosterman sums. Invent. Math.
71 (1983), 243-250.
[Gr-Ry] I. S. Gradshteyn and I. M. Ryzhik, Table of Integrals, Series and Prod-
ucts. Academic Press, London, 1965.
[Ha-Ra] G. H. Hardy and S. Ramanujan, Asymptotic formulae in combinatory
analysis. Proc. London Math. Soc. 17 (1918), 75-115.
[Ha-Ti] G. H. Hardy and E. C. Titchmarsh, Solutions of some integral equations
considered by Bateman, Kapteyn, Littlewood, and Milne. Proc. London
Math. Soc. 23 (1924), 1-26.
[He1] D. A. Hejhal, The Selberg Trace Formula for PSL(2,R). Springer Lec-
ture Notes in Math. 548 (1976) and 1001 (1983).
[He2] D. A. Hejhal, Eigenvalues of the Laplacian for Hecke triangle groups.
Memoirs of the Amer. Math. Soc. 469 (1992).
[Ho-Lo] J. Hoffstein and P. Lockhart, Coefficients of Maass forms and the Siegel
zero. Ann. of Math. 140 (1994), 161-176.
[Hu1] M. N. Huxley, Scattering matrices for congruence subgroups, in Modular
forms, 141-156, Ellis Horwood Series of Halsted Press, New York, 1984.
[Hu2] M. N. Huxley, Introduction to Kloostermania, in Elementary and An-
alytic Theory of Numbers, 217-306. Banach Center Publ. 17, Warsaw,
1985.
[Hu3] M. N. Huxley, Exponential sums and lattice points, II. Proc. London
Math. Soc. 66 (1993), 279-301; Corrigenda in 68 (1994), p. 264.
[Iw1] H. Iwaniec, Non-holomorphic modular forms and their applications, in
Modular Forms, 157-196. Ellis Horwood Series of Halsted Press, New
York, 1984.
[Iw2] H. Iwaniec, Spectral theory of automorphic functions and recent devel-
opments in analytic number theory. Proc. I.C.M., Berkeley, 1986.
[Iw3] H. Iwaniec, Prime geodesic theorem. J. Reine Angew. Math. 349
(1984), 136-159.
[Iw-Sa] H. Iwaniec and P. Sarnak, L∞ -norms of eigenfuntions of arithmetic sur-
faces. To appear in Ann. of Math.
[Kac] M. Kac, Can one hear the shape of a drum? Amer. Math. Monthly 73
(1966), 1-23.
References 235
Acting discontinuously, 40
amplifying linear form, 200
Atkin-Lehner theory of newforms, 127
autormophic Green function, 78
automorphic form, 58
autormophic forms of weight k, 144
automorphic function, 57
automorphic kernel, 71
Cartan’s decomposition, 16
central part F (Y ), 43
centralizer, 150
co-compact group, 41
column vector of the Eisenstein series, 94
compact part of K(z, w), 73
congruence group of level N , 47
conjugacy classes, 18, 150
continuous spectrum, 206
critical line, 95, 101
cusp, 42
cusp forms, 64, 145
cuspidal zones, 43
239
240 Subject Index
Deformation of g, 8
density theorem, 185
dilation, 19
Dirichlet divisor problem, 190
discrete subgroup, 39
distance function, 7
double cosets, 49
Eigenfunction of ∆, 21
eigenpacket, 112
Eisenstein series, 62
Eisenstein transform, 104
elliptic element, 19
energy integral, 182
Epstein zeta-functions, 180
equivalent points, 41
exceptional eigenvalue, 181
G-invariant, 4
gamma function, 219
Gauss circle problem, 6, 189
Gauss defect, 11
Gauss hypergeometric function, 35
Gauss-Bonnet formula, 46
geodesic polar coordinates, 17
geometric side of K, 150
Green function, 26, 75
Subject Index 241
Kernel of L, 28
Kloosterman sum, 51
Kloosterman sums zeta-function, 81
Kontorovich-Lebedev inversion, 23, 229
Laplace-Beltrami operator, 3, 20
Laplace operator, 20
lattice-point problem, 189
Legendre function, 25, 223
length of closed geodesics, 162
Lie algebra, 36
242 Subject Index
Lie derivative, 36
Parabolic element, 19
parabolic vertex, 42
Petersson’s formulas, 144
Petersson-Rankin-Selberg formula, 146
Poincaré area, 10
Poincaré differential, 7
Poincaré series, 58, 145
point-pair invariant, 28
point spectrum, 206
Poisson summation formula, 5
primitive motion, 40
principal congruence group of level N , 46
principal parts of an automorphic kernel, 72
Quaternion group, 46
quotient space Γ\H, 43
Subject Index 243
Radial function at w, 30
Ramanujan sum, 52
Ramanujan-Petersson conjecture, 128
Rankin-Selberg function, 119
reflection, 9
resolvent of K, 212
resolvent operator, 75, 78, 206
Resolvent Trace Formula, 166
residual eigenvalues, 179
Riemann surface, 43
riemannian measure, 10
rotation, 19
Scaling matrix, 42
scattering matrix, 94
Sears-Titchmarsh inversion, 147, 230
Selberg conjecture, 182
Selberg’s Trace Formula, 167
Selberg zeta-function, 168
Selberg/Harish-Chandra transform, 33
signature of Γ, 46
space of the Eisenstein series, 106
spectral trace of K, 150
spherical functions, 25, 223
spherical harmonics, 224
stability group, 30, 40
standard polygon, 44
Unimodular, 15
unique ergodicity conjecture, 202
upper half-plane, 7
a, b, c, 42 δab , 50
kak, 121 e(z) = e2πiz , 1
a ⊗ uj , 124 E(z, s), 94
A, 13 E(Γ\H), 63
a(y), 13 Ea (Γ\H), 106
A(Γ\H), 57 Ea , 104
As (Γ\H), 58 Ea f , 104
B, 49 Ea (z|p), 58
Bν (x), 142 Ea (z, s), 62
B(Γ\H), 63 EaY (z, s), 95
Bµ (Γ\H), 78 Ea (z|w), 62
cab , 53 Eam (z|ψ), 62
ca , 53 ηac (n, t), 118
c(a, b), 53 ηt (n), 68
Cg , 8 hf, gi, 60, 104
C(Γ\H), 64 |F |, 41
Cs (Γ\H), 64 F (α, β, γ; z), 221
Cab , 51 f 0 (x), 144
Cnew (Γ0 (N )\H), 129 f ∞, 143
Cold (Γ0 (N )\H), 129 f ∞ (x), 142
dµz, 10 fa (y), 59
dµj (z), 202 fˆa (n), 59
dνt (z), 202 fˆajk (m), 146
dνt (z), 202 fˆajk (m), 146
D, 3 f Y (z), 99
D(w), 41 Fa (Y ), 42
D(Γ\H), 69 Fs (u), 26
Dab , 81 F∞ , 44
D(λ), 91 ϕab (s), 66
Dλ (z, z 0 ), 91 ϕab (n, s), 66
∆, 20 ϕ(s) = det Φ(s), 155
245
246 Notation Index