Introducrion To The Spectral Theory

Introduction to the
Spectral Theory
of
Automorphic Forms
Henryk Iwaniec
Revista Matemática Iberoamericana
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.
Antonio Córdoba José L. Fernández

CONTENTS
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
5.2. 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
8.2. 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
9.2. 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
10.2. 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
13.2. 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
Subject Index 239
Notation Index 245

Preface
I was captivated by a group of enthusiastic Spanish mathematicians

whose desire for cultivating modern number theory I enjoyed recently
during two memorable events, the first at the summer school in San-
tander, 1992, and the second while visiting the Universidad Autónoma
in Madrid in June 1993. These notes are an expanded version of a series
of eleven lectures I delivered in Madrid∗ . They are more than a survey
of favorite topics since proofs are given for all important results. How-
ever, there is a lot of basic material which should have been included
for completeness, but was not because of time and space limitations.
Instead, to make a comprehensive exposition we focus on issues closely
related to the the spectral aspects of automorphic forms (as opposed
to the arithmetical aspects to which I intend to return on another oc-
casion).
Primarily, the lectures are addressed to advanced graduate stu-

dents. I hope the student will get inspiration for his own adventures in
the field. This is a goal which Professor Antonio Córdoba has a vision
of pursuing throughout the new volumes to be published by the Revista
Matemática Iberoamericana. I am pleased to contribute to part of his
plan.
Many people helped me prepare these notes for publication. In

particular I am grateful to Fernando Chamizo, José Luis Fernández,
Charles Mozzochi, Antonio Sánchez-Calle and Nigel Pitt for reading
and correcting an early draft. I also acknowledge the substantial work
in the technical preparation of this text by Domingo Pestana and Marı́a
Victoria Melián without which this project would not exist.
New Brunswick, October 1994 Henryk Iwaniec
∗
The author would like to thank the participants and the Mathematics Department
for their warm hospitality and support.
XIII
Introduction
The concept of an automorphic function is a natural generalization

of that of a periodic function. Furthermore an automorphic form is a
generalization of the exponential function
e(z) = e2πiz .
To define an automorphic function in an abstract setting one needs

a group Γ acting discontinuously on a locally compact space X ; the
functions on X which are invariant under the group action are called
automorphic functions (the name was given by F. Klein in 1890). A
typical case is the homogeneous space X = G/K of a Lie group G
where K is a closed subgroup. In this case the differential calculus is
available since X is a riemannian manifold. The automorphic functions
which are eigenfunctions of all invariant differential operators (these
include the Laplace operator) are called automorphic forms. The main
goal of harmonic analysis on the quotient space Γ\X is to decompose
every automorphic function satisfying suitable growth conditions into
automorphic forms. In these lectures we shall present the basic theory
for Fuchsian groups acting on the hyperbolic plane.
When the group Γ is arithmetic, there are interesting consequences
for number theory. What makes a group arithmetic is the existence of
a large family (commutative algebra) of certain invariant, self-adjoint
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
We begin by presenting the familiar case of the euclidean plane
R2 = (x, y) : x, y ∈ R .

The group G = R2 acts on itself as translations, and it makes R2 a

homogeneous space. The euclidean plane carries the metric
ds2 = dx2 + dy 2
of curvature K = 0, and the Laplace-Beltrami operator associated with

this metric is given by
∂2 ∂2
D= + .
∂x2 ∂y 2
Clearly the exponential functions
ϕ(x, y) = e(ux + vy) , (u, v) ∈ R2 ,
are eigenfunctions of D;
(D + λ)ϕ = 0 , λ = λ(ϕ) = 4π 2 (u2 + v 2 ) .
3
4 Harmonic analysis on the euclidean plane
The well known Fourier inversion

ZZ
fˆ(u, v) = f (x, y) e(ux + vy) dx dy ,
ZZ
f (x, y) = fˆ(u, v) e(−ux − vy) du dv ,
is just the spectral resolution of D on functions satisfying proper decay

conditions.
Another view of this matter is by invariant integral operators
Z
(Lf )(z) = k(z, w) f (w) dw .
R2
For L to be G-invariant it is necessary and sufficient that the kernel

function, k(z, w), depends only on the difference z − w, i.e. k(z, w) =
k(z − w). Such an L acts by convolution: Lf = k ∗ f . One shows
that the invariant integral operators mutually commute and that they
commute with the Laplace operator as well. Therefore the spectral
resolution of D can be derived from that for a sufficiently large family
of invariant integral operators. By direct computation one shows that
the exponential function ϕ(x, y) = e(ux + vy) is an eigenfunction of L
with eigenvalue λ(ϕ) = k̂(u, v), the Fourier transform of k(z).
Of particular interest will be the radially symmetric kernels:
k(x, y) = k(x2 + y 2 ) , k(r) ∈ C0∞ (R+ ) .
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 λ = 4π 2 (u2 + v 2 ) and J0 (z) is the Bessel function given by

π
1
Z
J0 (z) = cos(z cos α) dα .
π 0
Classical analytic number theory benefits a lot from harmonic anal-

ysis on the torus Z2 \R2 (which is derived from that on the free space
R2 by the unfolding technique), as it exploits properties of periodic
Harmonic analysis on the euclidean plane 5
functions. Restricting the domain of the invariant integral operator L

to periodic functions we can write
Z
(Lf )(z) = K(z, w) f (w) dw
Z2 \R2
where X
K(z, w) = k(z + p, w) ,
p∈Z2
by folding the integral. Hence the trace of L on the torus is equal to

Z X X
Trace L = K(w, w) dw = k(p) = k(m, n) .
p∈Z2 m,n∈Z
Z2 \R2
On the other hand by the spectral decomposition (classical Fourier se-

ries expansion) X
K(z, w) = λ(ϕ) ϕ(z) ϕ(w) ;
ϕ
the trace is given by

X X
Trace L = λ(ϕ) = k̂(u, v) .
ϕ u,v∈Z
Comparing both results we get the trace formula

X X
k(m, n) = k̂(u, v) ,
m,n∈Z u,v∈Z
which is better known as the Poisson summation formula. By a linear

change of variables this formula can be modified for sums over general
lattices Λ ⊂ R2 . On both sides of the trace formula on the torus Λ\R2
the terms are of the same type because the geometric and the spectral
points range over dual lattices. However, one looses the self-duality
on negatively curved surfaces yet the relevant trace formula is no less
elegant.
In particular for a radially symmetric function the Poisson sum-

mation becomes
6 Harmonic analysis on the euclidean plane
Theorem (Hardy-Landau, Voronoi). If k ∈ C0∞ (R), then

∞
X ∞
X
r(`) k(`) = r(`) k̃(`) ,
`=0 `=0
where r(`) denotes the number of ways to write ` as the sum of two
squares,
r(`) = # (m, n) ∈ Z2 : m2 + n2 = ` ,

and k̃ is the Hankel type transform of k given by

Z +∞ √
k̃(`) = π k(t) J0 (2π `t) dt .
0
Note that the lowest eigenvalue λ(1) = 4π 2 ` = 0 for the constant

eigenfunction ϕ = 1 contributes
Z +∞
k̃(0) = π k(t) dt ,
0
which usually constitutes the main term. Taking a suitable kernel (a

smooth approximation to a step function) and using standard estimates
for Bessel’s function we derive the formula
X
r(`) = πx + O(x1/3 ) ,
`≤x
which was originally established by Voronoi and Sierpinski by different

√
means. The left side counts integral points in the circle of radius x.
This is also equal to the number of eigenvalues λ(ϕ) ≤ 4π 2 x (counted
with multiplicity), so we have
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
1.1. The upper half-plane.
As a model of the hyperbolic plane we shall use the upper half of

the plane C of complex numbers
H = z = x + iy : x ∈ R, y ∈ R+ .

H is a riemannian manifold with the metric derived from the Poincaré

differential
(1.1) ds2 = y −2 (dx2 + dy 2 ) .
The distance function on H is given explicitly by
|z − w| + |z − w|
(1.2) ρ(z, w) = log .
|z − w| − |z − w|
However, a more practical relation is
(1.3) cosh ρ(z, w) = 1 + 2 u(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

up to sign. In particular both matrices 1 = 1 1 and −1 = −1 −1

give the identity transformation. We shall always make this distinction

when necessary but often without mentioning it.
Throughout we denote G = SL2 (R), the group of real matrices of
determinant 1. The group P SL2 (R) = G/(±1) of all Möbius transfor-
mations acts on the whole compactified complex plane Ĉ = C ∪ {∞}
(the Riemann sphere) as conformal mappings. A Möbius transforma-
tion maps an euclidean circle onto a circle subject to the convention
that the euclidean lines in Ĉ are also circles. Of course, the centers
may not be preserved since g is not an euclidean isometry, save for
g = ± 1 ∗1 .
If g = ∗c d∗ ∈ G, then

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
Figure 1. Isometric circles.
∗ ∗

For g = c d ∈ G we introduce the function
(1.9) jg (z) = cz + d .
The j-function satisfies the chain rule of differentiation,
(1.10) jgh (z) = jg (hz) jh (z) .
It follows from the formula
(1.11) |jg (z)|2 Im gz = Im z
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
(Im gz)−1 |dgz| = (Im z)−1 |dz| ,
which shows that the differential form (1.1) on H is G-invariant. This

implies that the Möbius transformations are isometries of the hyperbolic
plane. In addition to these isometries we have the reflection in the
imaginary line, z 7→ −z̄, which reverses the orientation. One can show
using the above properties
Theorem 1.1. The whole group of isometries of H is generated by the

Möbius transformations and the reflection z 7→ −z̄ .
Theorem 1.2. The hyperbolic lines (geodesics in H) are represented

by the euclidean semi-circles and half-lines orthogonal to R .
Figure 2. Geodesics in H .
The hyperbolic circles (loci of points in a fixed distance from a given

point in H) are represented by the euclidean circles in H (of course, not
with the same centers).
There are various interesting relations in the hyperbolic plane. For
instance the trigonometry for a triangle asserts that
sin α sin β sin γ
= =
sinh a sinh b sinh c
and
sin α sin β cosh c = cos α cos β + cos γ ,
where α, β, γ are the interior angles from which the sides of length a, b, c
are seen, respectively. The latter relation reveals that the length of sides
depends only on the interior angles.
More counter-intuitive features occur with the area. To define
area one needs a measure. The riemannian measure derived from the
Poincaré differential ds = y −1 |dz| on H is expressed in terms of the
Lebesgue measure simply by
(1.12) dµz = y −2 dx dy .
It is easy to show directly that the above measure is G-invariant.
The upper half-plane 11
Theorem 1.3 (Gauss defect). The area of a hyperbolic triangle with

interior angles α, β, γ is equal to
(1.13) π−α−β−γ.
There is a universal inequality between the area and the boundary

length of a domain in a riemannian surface called the isoperimetric
inequality; it asserts that
4πA − K A2 ≤ L2 ,
where A is the area, L is the length of the boundary, and K is the

curvature (assumed to be constant). The isoperimetric inequality is
sharp since the equality is attained for discs. In the euclidean plane
4πA ≤ L2 . In the hyperbolic plane we obtain
(1.14) 4πA + A2 ≤ L2 ;
hence A ≤ L so the area is comparable to the boundary length. This

observation should explain why the analysis on H is more subtle than
that on R2 in various aspects.
For example the lattice point problem on H is more sophisticated
than that on R2 (see Chapter 12). To illustrate the rules look at the
hyperbolic disc of radius r centered at i (the origin of H, so to speak).
Figure 3. A hyperbolic disc.

The hyperbolic area is 4π(sinh(r/2))2 , and the circumference is

2π sinh r. On the other hand the euclidean center is at i cosh r, and
the radius is sinh r so the area is π(sinh r)2 , and the circumference is
2π sinh r. Although the circumferences are the same, the euclidean area
is much larger than the hyperbolic one if r is large (still approximately
the same area for small r). Most of the hyperbolic area concentrates
along a lower segment of the boundary.
1.2. H as a homogeneous space.
Occasionally it will be convenient to work with the homogeneous

space model of the hyperbolic plane rather than the Poincaré upper
half-plane. Here we describe that model.
The group G = SL2 (R) acts on H transitively so H is obtained as
the orbit of a point

H = Gz = gz : g ∈ G .
The point z = i is special; its stability group is the orthogonal group

cos θ sin θ
n o
K = k ∈ G : ki = i = k(θ) = : θ∈R .
− sin θ cos θ
The element k(θ) ∈ K acts on H as the rotation at i of angle 2θ.
Figure 4. The action of K .

H as a homogeneous space 13
The upper half-plane H can be identified with the quotient G/K

(the space of orbits) so that a point z ∈ H corresponds to the coset gK
of all motions which send i to z. In such a realization of H the group
G acts on itself by matrix multiplication.
In order to be able to use both models alternatively and consis-

tently we need an explicit connection between the rectangular coordi-
nates of points in H and the matrix entries of group elements in G.
This is given through the Iwasawa decomposition
G = N AK .
Here A and N are the following subgroups:

n
a
o
A= : a ∈ R+ ,
a−1
n
1 x
o
N= : x∈R .
1
The Iwasawa decomposition asserts that any g ∈ G has the unique

factorization
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
i.e. take a = (γ 2 + δ 2 )−1/2 and θ such that − sin θ = γ a, cos θ = δa.

Then apply on the left a suitable translation 1 x1 to arrive at the
desired upper row without altering the lower row. Since the above pro-
cedure is unique in each step, this proves the Iwasawa decomposition.
It is plain that a point z = x + iy in H corresponds to the coset

gK in G/K for g = n a k with
y 1/2

1 x
n = n(x) = , a = a(y) = .
1 y −1/2
Figure 5. The Iwasawa coordinates.
The group A operates on H by dilations, and the group N op-

erates by translations. The upper half-plane is identified through the
Iwasawa decomposition with the group of upper triangular matrices of
determinant 1: n
∗ ∗
o
P = ∈G .
∗
We have H = G/K = N A = A N = P . Notice that A, N are abelian
whereas P is not yet the following commutativity relation holds
a(y) n(x) = n(xy) a(y) .
To perform integration on P one needs a measure. We proceed by

recalling a few facts about topological groups.
Suppose G is a locally compact group. Then G has a left-invariant

measure dg, say, which means that
Z Z
f (hg) dg = f (g) dg
for any test function integrable on G. The left-invariant measure is

unique up to a positive constant multiple; therefore,
Z Z
f (gh) dg = δ(h) f (g) dg ,
where δ(h) > 0 depends only on h because dgh−1 is another left-

invariant measure. The factor δ(h) is called the modular function of
G. Clearly δ : G −→ R+ is a group homomorphism, and one also shows
that δ is continuous. Similarly G has a right-invariant measure; it is
H as a homogeneous space 15
equal to δ(g) dg up to a constant factor. If δ(g) = 1 then G is called

unimodular. Abelian groups are obviously unimodular, and compact
groups are also unimodular because the multiplicative group R + con-
tains no compact subgroups other than the trivial one.
Now we return to the group G = SL2 (R) and its decomposition

G = N A K. Each factor, being abelian, is a unimodular group. The
invariant measures on N , A, K are given by

1 x
dn = dx if n = n(x) = ,
1
√
y
da = y −1 dy if a = a(y) = √ ,
1/ y

cos θ sin θ
dk = (2π)−1 dθ if k = k(θ) = ,
− sin θ cos θ
where dx, dy, dθ are the Lebesgue measures.

R Since K is compact, we
could normalize the measure on K to have K dk = 1.
Let us define a measure dp on P = A N by requiring that

Z Z Z
f (p) dp = f (an) da dn ;
P A N
i.e. dp = y −1 dx dy in rectangular coordinates. We shall show that

dp is left-invariant. First we need a multiplication rule on P = A N .
Given h = a(u) n(v) and p = a(y) n(x) one has hp = a(uy) n(x + vy −1 ).
Hence,
dx dy
Z Z Z
f (hp) dp = f (a(uy) n(x + vy −1 ))
P A N y
dx dy
Z Z Z
= f (a(y) n(x)) = f (p) dp ,
A N y P
which shows that dp is left-invariant.

Furthermore we derive by Fubini’s theorem the following relation
dx dy dx dy
Z Z Z Z
f (a(y) n(x)) = f (n(xy) a(y))
A N y N A y
dx dy
Z Z
= f (n(x) a(y)) 2 .
N A y
This shows that the modular function of P is equal to δ(p) = y −1

if p = a(y) n(x). Hence the right-invariant measure on P is equal to
δ(p) dp = y −2 dx dy, which is just the riemannian measure on H.
Remark. The whole group G = N A K = SL2 (R) is unimodular in

spite of being non-abelian and not compact. One can show that the
measure dg defined by
Z Z Z Z
f (g) dg = f (a n k) da dn dk
G A N K
is the invariant measure on G.
1.3. The geodesic polar coordinates.
We shall often encounter functions on H which depend only on

the hyperbolic distance. Naturally, it is more convenient to work with
such functions in geodesic polar coordinates rather than in rectangular
coordinates z = x+iy. The geodesic polar coordinates are derived from
Cartan’s decomposition
G = K AK .
This asserts that any g ∈ G can be brought to a diagonal matrix by mul-
tiplying on both sides with orthogonal matrices. To see this first mul-
tiply on the left by k1 ∈ K to bring g to a symmetric matrix g1 = k1 g.
Then by conjugation in K the symmetric matrix g1 can be brought to a
diagonal matrix a = k g1 k −1 (this is the spectral theorem for symmetric
matrices). Hence, we have Cartan’s decomposition g = k1−1 k −1 a k.
We shall write any g ∈ P SL2 (R) as g = k(ϕ) a(e−r ) k(θ), with

cos ϕ sin ϕ
k(ϕ) = , 0 ≤ ϕ < π,
− sin ϕ cos ϕ
e−r/2

−r
a(e )= , r ≥ 0,
er/2
and k(θ) as in the Iwasawa decomposition. Of course, a(e−r ) is dif-

ferent from the one in the rectangular coordinates. We have ρ(gi, i) =
ρ(k(ϕ) e−r i, i) = ρ(e−r i, i) = r; therefore r is the hyperbolic distance
from i to gi = z = x + iy.
The geodesic polar coordinates 17
Since k(ϕ) acts by rotation at i of angle 2ϕ, it follows that when

ϕ ranges over [0, π) and r over (0, +∞), the upper half-plane is covered
once except for the origin z = i. The level curves of r and ϕ are
orthogonal circles.
Figure 6. The geodesic polar coordinates.
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.
y = (cosh r + sinh r cos 2ϕ)−1 ,

x = y sinh r sin 2ϕ .
The length element and the measure are expressed in the new coordi-
nates as follows:
(1.15) ds2 = dr2 + (2 sinh r)2 dϕ2 ,

(1.16) dµz = (2 sinh r) dr dϕ .
In the (u, ϕ) coordinates, where cosh r = 1 + 2u as in (1.3), we have
(1.17) dµz = 4 du dϕ .
1.4. Bruhat decomposition.
We have just introduced two distinct decompositions of G=SL2 (R),

namely G = N A K (Iwasawa) and G = K A K (Cartan). There is yet
a third important decomposition that we wish to consider, namely
G = N AN ∪ N ωAN
where
−1
ω=
1
is the involution. Here the first part N A N = A N = N A consists of
the upper triangullar matrices. The second part N ω A N asserts that
a ∗

every g = c d with c 6= 0 factors uniquely into

1 a/c −1/c 1 d/c
(1.18) g= .
1 c 1
All the previous decompositions can be unified in terms of the Lie
algebra of G. Briefly, a decomposition of type G = P1 Q P2 , where Pj
are 1-parameter subgroups (Pj is necessarily conjugate to {exp (tXj ) :
t ∈ R} for some Xj of zero trace) and Q is a 1-dimensional subspace
yields a parametrization of G. Such a parametrization makes it possible
to reduce the analysis on G to a simpler one on the components in the
following fashion: put a suitable test function on the central component
and use characters on the outer components.
The three decompositions we have singled out for these lectures are
destined to specific tasks. In particular we shall employ the Iwasawa
decomposition (the rectangular coordinates) for the Fourier develop-
ment of automorphic forms, the Cartan decomposition (the geodesic
polar coordinates) to study the Green function and the Bruhat type
decomposition for creating Kloosterman sums.
1.5. The classification of motions.
The Möbius transformations are rigid motions of the hyperbolic

plane and they move points in distinct ways. We shall give a charac-
terization by various means. First notice that conjugate motions act
similarly, therefore the classification should be invariant under conju-
gation. Given g ∈ P SL2 (R) we denote its conjugacy class by
{g} = τ gτ −1 : τ ∈ P SL2 (R) .

The classification of motions 19
The identity motion forms a class of itself; there is nothing to examine

in this class.
An important geometric invariant of the conjugation is the number

of fixed points (and their configuration). Any

a b
g= 6= ±1
c d
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.
An important algebraic invariant of conjugation is the trace, more

precisely, its absolute value because g determines the matrix ac db up

to sign. In terms of trace the above classes are characterized as follows

(i) g is parabolic if and only if |a + d| = 2 ,

(ii) g is hyperbolic if and only if |a + d| > 2 ,
(iii) g is elliptic if and only if |a + d| < 2 .
1.6. The Laplace operator.
Denote by Tg the following operator
(Tg f )(z) = f (gz) .
A linear operator L acting on functions f : H −→ C is said to be

invariant if it commutes with all Tg , i.e.
L(f (gz)) = (Lf )(gz) , for all g ∈ G .
The invariant differential operators are particularly important; among

these, the Laplace-Beltrami operator is special. In general, on a rie-
mannian manifold, the Laplace-Beltrami operator ∆ is characterized
by the property that a diffeomorphism is an isometry if and only if it
leaves ∆ invariant. On the hyperbolic plane H the Laplace operator
derived from the differential ds2 = y −2 (dx2 + dy 2 ) is given by
∂2 ∂2 ∂ ∂
(1.19) ∆ = y2 2
+ 2
= −(z − z̄)2 ,
∂x ∂y ∂z ∂ z̄
where ∂/∂z = (∂/∂x − i∂/∂y)/2 and ∂/∂ z̄ = (∂/∂x + i∂/∂y)/2 are the
partial complex derivatives. Check directly that ∆ is invariant.
In geodesic polar coordinates (r, ϕ) the Laplace operator takes the

form
∂2 1 ∂ 1 ∂2
(1.20) ∆= + + .
∂r2 tanh r ∂r (2 sinh r)2 ∂ϕ2
Changing r into u, where cosh r = 2u + 1 (see (1.3) and (1.4)), we get
∂2 ∂ 1 ∂2
(1.21) ∆ = u(u + 1) + (2u + 1) + .
∂u2 ∂u 16 u(u + 1) ∂ϕ2
Any differential operator on H which is G-invariant is a polynomial

in ∆ with constant coefficientes, i.e. the algebra of invariant differential
Eigenfunctions of ∆ 21
operators is generated by ∆ . A great deal of harmonic analysis on H

concerns decomposition of functions f : H −→ C into eigenpackets of
∆ (an analogue of Fourier inversion).
1.7. Eigenfunctions of ∆.
A function f : H −→ C with continuous partial derivatives of order

2 is an eigenfunction of ∆ with eigenvalue λ ∈ C if
(1.22) (∆ + λ)f = 0 .
Since ∆ is an elliptic operator with real-analytic coefficients, it forces its

eigenfunctions to be real-analytic. The eigenfunctions with eigenvalue
λ = 0 are harmonic functions; among them are holomorphic functions,
i.e. those annihilated by the operator ∂/∂ z̄ (the Cauchy-Riemann equa-
tions).
There are various ways of constructing eigenfunctions of ∆ for a
given eigenvalue λ. The standard method uses separation of variables.
However a more prolific one is the method of images; it generates a lot
of eigenfunctions out of a fixed f (z) by shifting to f (gz) and still more
by averaging f (gz) over selected g in G. In this way one may search for
eigenfunctions which satisfy desirable transformation rules. Either way
the result depends on the type of coordinates in which the construction
is performed.
We first work in the rectangular coordinates z = x + iy. If one

wants f (z) to be a function in y only, i.e. constant in x, one has the
obvious choice of two linearly independent solutions to (1.22), namely
1 s 1
(1.23) (y + y 1−s ) and (y s − y 1−s ) ,
2 2s − 1
where s(1 − s) = λ. Note that s 7→ λ is a double cover of C, save for
s = 1/2, λ = 1/4. For s = 1/2 the above eigenfunctions become
(1.24) y 1/2 and y 1/2 log y ,
respectively. If s 6= 1/2 we shall often take a simpler pair y s , y 1−s .

If one wants f (z) to be periodic in x of period 1, try f (z) =
e(x) F (2πy) and find that F satisfies the ordinary differential equation
(1.25) F 00 (y) + (λy −2 − 1) F (y) = 0 .

There are two linearly independent solutions, namely
(2π −1 y)1/2 Ks−1/2 (y) ∼ e−y
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
(1.26) Ws (z) = 2 y 1/2 Ks−1/2 (2πy) e(x) ,
which is named the Whittaker function. It will be convenient to extend

Ws (z) to the lower half-plane H by imposing the symmetry
(1.27) Ws (z) = Ws (z̄) .
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
(1.28) f (nz) = χ(n) f (z) , for all n ∈ N ,
where χ : N −→ C is the character given by

1 x
χ(n) = e(x) , if n = .
1
To construct such an f we begin with the obvious eigenfunction y s by

means of which we produce χ̄(n)(Im ωnz)s and integrate these over the
group N getting
Z +∞
f (z) = χ̄(n(x)) (Im ωn(x)z)s dx
−∞
Z +∞
−1 s

= e(x) Im dx
−∞ z−x
Z +∞
1−s
= e(x) y (1 + t2 )−s e(ty) dt = π s Γ(s)−1 Ws (z) .
−∞
Here the involution ω = 1 −1 was inserted to buy the absolute con-

vergence, at least if Re s > 1/2. The resulting function has an analytic

continuation to the whole complex s-plane where it extends to an eigen-
function of ∆, periodic in x.
Changing the character into χ(n) = e(rx), where r is a fixed real
number different from 0, we obtain by the above method an eigenfunc-
tion which is a multiple of Ws (rz). The Whittaker functions are basic
for harmonic analysis on H as the following proposition clearly assures.
Proposition 1.4. Any f ∈ C0∞ (H) has the integral representation
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
and γs (r) = (2π|r|)−1 t sinh πt for s = 1/2 + it.
Therefore loosely speaking the Whittaker functions Ws (rz) with

Re s = 1/2 and r 6= 0 real form a complete eigenpacket on H. The proof
of Proposition 1.4 is obtained by application of the classical Fourier
inversion in (r, x) variables and the following Kontorovitch-Lebedev in-
version in (t, y) variables (see [Ko-Le] and [Le, p.131]):
Z +∞ Z +∞
−1
(1.31) g(w) = Kit (w) Kit (y) g(y) y dy π −2 t sinh(πt) dt .
0 0
The proof of the next result reduces to the ordinary Fourier series
expansion for periodic functions.
Proposition 1.5. Let f (z) be an eigenfunction of ∆ with eigenvalue

λ = s(1 − s) which satisfies the transformation rule
(1.32) f (z + m) = f (z) , for all m ∈ Z
and the growth condition
(1.33) f (z) = o(e2πy ) , as y → +∞ .

Then f (z) has the following expansion

X
(1.34) f (z) = f0 (y) + fn Ws (nz) ,
n6=0
where the zero-term f0 (y) is a linear combination of the functions in

(1.23) and (1.24). The series converges absolutely and uniformly on
compacta. Hence
(1.35) fn eε|n|
for any ε > 0 with the implied constant depending on f and ε.
No less important than Ws (z) is the eigenfunction of ∆ associated

with the second solution to the differential equation (1.25) given by
(1.36) Vs (z) = 2π y 1/2 Is−1/2 (2πy) e(x) .
We extend Vs (z) to the lower half-plane H by requiring the same sym-

metry as that for Ws (z). Note that Ws (z) and Vs (z) have distinct
behaviour at infinity, namely
(1.37) Ws (z) ∼ e(x) e−2πy ,

(1.38) Vs (z) ∼ e(x) e2πy ,
as y → +∞; therefore, they are linearly independent. They both will

appear in the Fourier expansion of the automorphic Green function.
Next we shall perform the harmonic analysis on H in geodesic polar

coordinates (r, ϕ). Recall the connection z = k(ϕ) e−r i . We seek an
eigenfunction of ∆ with eigenvalue λ = s(1 − s) which transforms as
(1.39) f (kz) = χ(k) f (z) , for all k ∈ K ,
where χ : K −→ C is the character given by (for m ∈ Z)

2imθ cos θ sin θ
χ(k) = e , if k = .
− sin θ cos θ
To produce such an eigenfunction we integrate χ̄(k)(Im kz)s over the

group K getting
1 π
Z
f (z) = (Im k(−θ) k(ϕ) e−r i)s e2imθ dθ
π 0
π
1
Z
= (cosh r + sinh r cos 2θ)−s e2im(θ+ϕ) dθ
π 0
Γ(1 − s)
= P m (cosh r) e2imϕ ,
Γ(1 − s + m) −s
m
where P−s (v) is the Legendre function (the gamma factors have ap-
peared because of unfortunate normalization in the literature on special
functions, see Appendix B.3).
The same eigenfunction can be obtained by the method of separa-
tion of variables. Indeed, writing f (z) = F (u) e2imϕ , where 2u + 1 =
cosh r, we find by (1.21) that F (u) solves the ordinary differential equa-
tion
00 0 m2
u(u + 1) F (u) + (2u + 1) F (u) + s(1 − s) − F (u) = 0 .
4u(u + 1)
Then we verify by partial integration that F (u) given by

π
1
Z p
F (u) = (2u + 1 + 2 u(u + 1) cos θ)−s cos(mθ) dθ
π 0
Γ(1 − s)
= P m (2u + 1)
Γ(1 − s + m) −s
is a solution to this equation (see (B.21) and (B.24)).

By either method we have obtained the classical spherical functions
(1.40) Usm (z) = P−s

m
(2u + 1) e2imϕ .
They form a complete system on H in the following sense.
Proposition 1.6. Any f ∈ C0∞ (H) has the expansion

X (−1)m Z
(1.41) f (z) = Usm (z) f m (s) δ(s) ds ,
2πi
m∈Z
(1/2)
where Z
m
f (s) = f (z) Usm (z) dµz ,
H
and δ(s) = t tanh πt for s = 1/2 + it.

The above expansion can be derived by applications of the Fourier

series representation of a periodic function together with the following
inversion formula due to F. G. Mehler [Me] and V. A. Fock [Foc]:
Z +∞ Z +∞
(1.42) g(u) = P−1/2+it (u) P−1/2+it (v) g(v) dv t tanh (πt) dt .
0 1
Here Ps (v) = Ps0 (v) denotes the Legendre function of order m = 0.

The spherical functions of order zero are special, they depend only
on the hyperbolic distance, namely Us0 (z) = P−s (2u + 1) = Fs (u), say.
Note that
1 π
Z p
(1.43) Fs (u) = (2u + 1 + 2 u(u + 1) cos θ)−s dθ
π 0
is also given by the hypergeometric function (see (B.23))
Fs (u) = F (s, 1 − s; 1, u) ,
and it satisfies the differential equation
(1.44) u(u + 1) F 00 (u) + (2u + 1) F 0 (u) + s(1 − s) F (u) = 0 .
Recall that the differential equation (1.44) is equivalent to (see (1.21))
(1.45) (∆ + s(1 − s)) F = 0
for functions depending only on the distance variable u.

There is another solution to (1.44) linearly independent of Fs (u)
given by
1
1
Z
(1.46) Gs (u) = (ξ(1 − ξ))s−1 (ξ + u)−s dξ .
4π 0
This also can be expressed by the hypergeometric function (see (B.16))
Γ(s)2 1
Gs (u) = u−s F s, s; 2s, .
4π Γ(2s) u
Lemma 1.7. The integral (1.46) converges absolutely for Re s = σ > 0.

It gives a function Gs (u) on R+ which satisfies equations (1.44) and
(1.45). Moreover, Gs (u) satisfies the following bounds
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 → +∞ .
Proof. That Gs (u) satisfies (1.45) follows by partial integration from

the identity
d s
(∆ + s(1 − s)) ξ s−1 (1 − ξ)s−1 ξ −s = s ξ (1 − ξ)s (ξ + u)s−1 .
dξ
To prove (1.47) we put ν = (|s| + 1)u, η = (|s| + 1)−1 and split

Z 1 Z ν Z η Z 1
ξ(1 − ξ) s−1 dξ
4πGs (u) = = + +
0 ξ+u ξ+u 0 ν η
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
For the remaining integral we shall use the approximation

ξ(1 − ξ) s−1 u + ξ 2 s−1 u + ξ2
= 1− =1+O
ξ+u u+ξ u+ξ
and obtain
η η η
dξ u + ξ2
Z Z Z
= +O dξ
ν ν ξ+u ν (u + ξ)2
u+η 1
= log + O(1) = log + O(1) .
u+ν u
This completes the proof of (1.47). The proof of (1.48) is similar and
(1.49) is obvious.
1.8. The invariant integral operators.
An integral operator is defined by

Z
(Lf )(z) = k(z, w) f (w) dµw ,
H
where dµ is the riemannian measure and k : H × H −→ C is a given

function called the kernel of L. In what follows we always assume
without mention that the kernel k(z, w) and the test function f (w) are
such that the integral converges absolutely. This assumption does not
exclude the possibility that k(z, w) is singular; as a matter of fact the
important kernels are singular on the diagonal z = w.
For L to be invariant it is necessary and sufficient that
k(gz, gw) = k(z, w) , for all g ∈ G .
A function with this property is called point-pair invariant; it depends

solely on the hyperbolic distance between the points. Consequently, we
can set
k(z, w) = k(u(z, w)) ,
where k(u) is a function in one variable u ≥ 0 and u(z, w) is given by
(1.4). Therefore, an invariant integral operator is a convolution.
The invariant integral operators will be used to develop the spectral
resolution of the Laplace operator. The key point is that the resolvent
of ∆ (the inverse to ∆ + s(1 − s) acting on functions satisfying suitable
growth conditions) is an integral operator with kernel Gs (u) given by
(1.46). On the other hand every invariant integral operator is a function
of ∆ in a spectral sense. We shall give a proof of this important fact first
because it requires several independent results, which will be employed
elsewhere.
Lemma 1.8. Let k(z, w) be a smooth point-pair invariant on H × H.

We have
(1.50) ∆z k(z, w) = ∆w k(z, w) .
Proof. Using geodesic polar coordinates with the origin at w (send i

to w) we get
∆z k(z, w) = u(u + 1) k 00 (u) + (2u + 1) k 0 (u) .

The invariant integral operators 29
Then using geodesic polar coordinates with the origin at z we get the
same expression for ∆w k(z, w).
For two functions F , G such that |F G| is integrable on H with

respect to the measure dµ we define the inner product by
Z
(1.51) hF, Gi = F (z) G(z) dµz .
H
If F, G ∈ C0∞ (H) then by partial integration

Z
(1.52) h−∆F, Gi = ∇F · ∇G dx dy ,
H
where ∇F = [∂F/∂x, ∂F/∂y] is the gradient of F . Hence we infer that
(1.53) h−∆F, Gi = hF, −∆Gi
and
(1.54) h−∆F, F i ≥ 0 .
Therefore, −∆ is a symmetric and non-negative operator in the space

C0∞ (H).
Theorem 1.9. The invariant integral operators commute with the

Laplace operator.
Proof. By (1.50) and (1.53) we argue as follows

Z Z
∆L f (z) = ∆z k(z, w) f (w) dµw = ∆w k(z, w) f (w) dµw ,
R
but k(z, w)∆w f (w) dµw = L ∆ f (z) .
Remarks. The lower bound (1.54) can be improved and generalized a

bit. Consider the Dirichlet problem
(∆ + λ) F = 0 in D ,

(1.55)
F =0 on ∂D
for a domain D ⊂ H with a piecewise continuous boundary ∂D, where

F is smooth in D and continuous in ∂D ∪ D. The solutions are in
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
On the other hand we have by partial integration that for each x
∂F −1
Z Z
2 −2
F y dy = 2 F y dy ,
∂y
and integrating in x we infer, by the Cauchy-Schwarz inequality, that
∂F 2
Z Z
2
F dµ ≤ 4 dx dy .
D D ∂y
Combining both estimates we obtain
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.
We return to the study of point-pair invariants. A function f (z, w)

is said to be radial at w if as a function of z it depends only on the
distance of z to w, i.e. it can be written as F (u(z, w), w). A function
f (z, w) can be radial at some point w, but not necessarily at other
points; clearly a point-pair invariant is radial at all points. Given any
f : H −→ C one can produce a radial function at w ∈ H by averaging
over the stability group Gw = g ∈ G : gw = w . One gets
π
1
Z Z
fw (z) = f (gz) dg = f (σ k(θ) σ −1 ) dθ ,
Gw π 0
where σ ∈ G is any motion which brings i to w so that Gw = σKσ −1 .

The mapping f 7→ fw will be called the mean-value operator.
Lemma 1.10. The mean-value fw (z) is radial at w. Moreover we have

(1.57) fz (z) = f (z) .
Proof. Suppose z, z1 are at the same distance from w. Then there

exists g1 ∈ Gw which sends z1 to z. Applying g1 we derive that
Z Z Z
fw (z1 ) = f (gz1 ) dg = f (gg1 z) dg = f (gz) dg = fw (z)
Gw Gw Gw
thus proving the first assertion. The second assertion is straightforward

Z Z
fz (z) = f (gz) dg = f (z) dg = f (z) .
Gz Gw
Lemma 1.11. An invariant integral operator L is not altered by the

mean-value operator, i.e. we have
(Lf )(z) = (Lfz )(z) .
Proof. Let k(z, w) be a kernel of L which is point-pair invariant. We

argue as follows
Z Z Z
(Lfz )(z) = k(z, w) fz (w) dµw = k(z, w) f (gw) dg dµw
H Gz
Z Z ZH Z
= k(z, w) f (gw) dµw dg = k(gz, w) f (w) dµw dg
Gz H Gz H
Z Z
= dg k(z, w) f (w) dµw = (Lf )(z) .
Gz H
A function f (z, w) which is radial at every w ∈ H may not neces-

sarily be a point-pair invariant, but if in addition f (z, w) is an eigen-
function of ∆ in z for any w with eigenvalue independent of w, then
f (z, w) is a point-pair invariant. Such a function is unique up to a
constant factor.
Lemma 1.12. Let λ ∈ C and w ∈ H. There exists a unique function

ω(z, w) in z which is radial at w such that
ω(w, w) = 1 ,
(∆z + λ) ω(z, w) = 0 .
This is given by
(1.58) ω(z, w) = Fs (u(z, w)) ,
where Fs (u) is the Gauss hypergeometric function F (s, 1 − s; 1; u).
Proof. Setting ω(z, w) = F (u) with u = u(z, w) we find that F (u)

satisfies the differential equation (1.44); thus it is a linear combination of
Fs (u) and Gs (u), but the normalization condition F (0) = ω(w, w) = 1
determines (1.58).
It follows immediately from Lemmas 1.10 and 1.12 that
Corollary 1.13. If f (z) is an eigenfunction of ∆ with eigenvalue

λ = s(1 − s), then
(1.59) fw (z) = ω(z, w) f (w) .
As a consequence notice that if an eigenfunction f vanishes at a

point w then fw ≡ 0. Now we are ready to prove the following basic
result
Theorem 1.14. Any eigenfunction of ∆ is also an eigenfunction of

all invariant integral operators. More precisely, if (∆ + λ)f = 0 and
L is an integral operator whose kernel k(u) is smooth and compactly
supported in R+ , then there exists Λ = Λ(λ, k) ∈ C depending on λ and
k but not on f such that Lf = Λ · f , i.e.
Z
(1.60) k(z, w) f (w) dµw = Λ f (z) .
H
Proof. By Lemma 1.11 and Corollary 1.13 we obtain (1.60) with

Z
(1.61) Λ= k(z, w) ω(z, w) dµw .
H
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.
The converse to Theorem 1.14 is also true. It asserts the following
Theorem 1.15. If f is an eigenfunction of all invariant integral opera-

tors whose kernel functions are in C0∞ (R+ ), then f is an eigenfunction
of ∆.
Proof. Let k(z, w) be a point-pair invariant such that (1.60) holds

true with Λ 6= 0 (if Λ = 0 for all k, then f ≡ 0 and the assertion is
obvious). Applying ∆ to both sides we get
Z
∆z k(z, w) f (w) dµw = Λ(∆f )(z) .
H
But ∆z k(z, w) is another point-pair invariant so by the hypothesis the

above integral equals Λ0 f (z) for some Λ0 ∈ C. By combining both
relations we get (∆ + λ)f = 0 with λ = −Λ0 Λ−1 .
There is a striking resemblance of Cauchy’s formula for holomor-

phic functions to the integral representation (1.60) for eigenfunctions of
∆. The latter is particularly helpful for testing the convergence of se-
quences of eigenfunctions as well as for estimating at individual points.
To elaborate further we shall establish an explicit expression for Λ
in terms of the eigenvalue λ and the kernel function k(u). This is given
by the Selberg/Harish-Chandra transform in the following three steps:
Z +∞
q(v) = k(u) (u − v)−1/2 du ,
v
r 2
(1.62) g(r) = 2 q sinh ,
2
Z +∞
h(t) = eirt g(r) dr .
−∞
Theorem 1.16. If k ∈ C0∞ (R+ ) and if f is an eigenfunction of ∆ with

eigenvalue λ = s(1 − s), where s = 1/2 + it, t ∈ C, then (1.60) holds
with Λ = h(t).
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).
Theorem 1.16 says that an invariant integral operator is a function

of the Laplace operator (in the spectral sense) given by the
Selberg/Harish-Chandra transform (1.62). The assumption in Theor-
ems 1.14-1.16 that the kernel k(u) is compactly supported is not essen-
tial though a certain control over the growth is required. It is simpler
to express the sufficient conditions in terms of h(t) rather than k(u).
These conditions are:
h(t) is even ,
1
(1.63) h(t) is holomorphic in the strip |Im t| ≤ + ε ,
2
−2−ε
h(t) (|t| + 1) in the strip.
For any h having the above properties one finds the inverse of the
Selberg/Harish-Chandra transform in the following three steps:
Z +∞
1
g(r) = eirt h(t) dt ,
2π −∞
1 √ √
(1.64) q(v) = g(2 log( v + 1 + v)) ,
2
1 +∞
Z
k(u) = − (v − u)−1/2 dq(v) .
π u
We shall rarely apply the relations (1.62) and (1.64) since they are
quite knotty. Instead, it is often easier to assess h(t) from the integral
representation (1.60) by testing it against a suitable eigenfunction. For
f (w) = ω(z, w) = Fs (u(z, w)) we get by (1.61) (using polar coordinates
(u, ϕ) and (1.17)) that
Z +∞
(1.62’) h(t) = 4π Fs (u) k(u) du ,
0
The Green function on H 35
where s = 1/2 + it and Fs (u) is the Gauss hypergeometric function (see

(1.43) and (1.44)). The inverse is given by (applying (1.42))
Z +∞
1
(1.64’) k(u) = Fs (u) h(t) tanh(πt) t dt .
4π −∞
1.9. The Green function on H .
Let s ∈ C with Re s > 1 and let −Rs be the integral operator on

H whose kernel function is given by (1.46), i.e.
Z
(1.65) −(Rs f )(z) = Gs (u(z, w)) f (w) dµw .
H
Theorem 1.17. If f is smooth and bounded on H, then
(1.66) (∆ + s(1 − s))Rs f = f .
In other words Rs is the right inverse to ∆ + s(1 − s) so that

Gs (u(z, w)) is the Green function on the free space H. Recall that
(1.67) (∆ + s(1 − s))Gs = 0 .
Before proving Theorem 1.17 let us make a few remarks. First we

emphasize that Gs is singular on the diagonal z = w. More precisely
we have
−1
(1.68) Gs (u(z, w)) = log |z − w| + Hs (z, w) ,
2π
say, where Hs is smooth and has bounded derivatives on H × H. The
logarithmic singularity of Gs is the critical property for (1.66) to be
true; indeed, ignoring this property one could guess wrongly that the
operator ∆ + s(1 − s) annihilates Rs since it annihilates Gs .
The proof of (1.66) depends on the following formula.
Lemma 1.18. If f is smooth and bounded on H, then

Z
(1.69) −(∆+s(1−s))Rs f (z) = Gs (u(z, w)) (∆+s(1−s))f (w) dµw .
H
A formal argument using the symmetry of ∆ + s(1 − s) (which is

not justified for singular kernels) seems to yield the result immediately,
but a rigorous proof is by no means easy. For a clear proof we use some
differential operators derived from the Lie algebra of G = SL2 (R). Let
us recall a few basic facts (cf. [La]). The Lie algebra g of the group G
over R consists of all 2 × 2 matrices X such that
∞
X (tX)n
exp (tX) = ∈ G, for all t ∈ R .
n=0
n!
One can show that g consists of trace zero matrices and that

0 1 0 0 1 0
X1 = , X2 = , X3 =
0 0 1 0 0 −1
form a basis of g over R. Note that
for the basis matrices we have

1 t
exp (tX1 ) = 1 + tX1 = ,
1

1
exp (tX2 ) = 1 + tX2 = ,
t 1
t2 t3
t
e
exp (tX3 ) = 1 + tX3 + + X3 + · · · = .
2 6 e−t
If X ∈ g then {exp (tX) : t ∈ R} is a one-parameter subgroup of G, and
the map t 7→ exp (tX) is a group homomorphism which is real-analytic
in a neigbourhood of t = 0. Thus we can define a linear operator
LX : C ∞ (G) −→ C ∞ (G) by
d
(1.70) LX f (z) = f (exp (tX)z) .

dt t=0
Clearly LX satisfies the Leibnitz rule LX (f g) = f LX g + gLX f so LX

is a differentiation (the Lie derivative).
Let L1 , L2 , L3 denote the differential operators derived from the
basis matrices X1 , X2 , X3 respectively. We shall show that
∂
(1.71) L1 = ,
∂x
∂ ∂
(1.72) L2 = (y 2 − x2 ) − 2xy ,
∂x ∂y
∂ ∂
(1.73) L3 = 2x + 2y ,
∂x ∂y
1
(1.74) 2∆ = L1 L2 + L2 L1 + L3 L3 .
2
The Green function on H 37
The formula (1.71) is obtained by differentiating at t = 0 as follows:
d
L1 f (z) = f (z + t) = fx (z) .
dt
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
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
Now we are ready to give a rigorous proof of (1.69). Let gt =

exp (tXj ). Since Rs is a G-invariant operator, we have
Z
−Rs f (gt z) = Gs (u(z, w)) f (gt w) dµw .
H
Differentiating in t and then putting t = 0 we get (1.69) for each of the

three operators Lj in place of ∆ + s(1 − s); hence (1.69) is derived for
∆ + s(1 − s) by (1.74).
For the proof of Theorem 1.17 we split H = U ∪ V , where U is the
disc (euclidean) centered at z of radius ε > 0 and V is the area outside
the disc; accordingly we split the integral (1.69). One sees clearly that
the integral over the disc U vanishes as ε tends to 0. To evaluate the
complementary integral over V we shall use Green’s formula
∂f ∂g
Z Z
(1.75) (g Df − f Dg) dx dy = g −f d` ,
V ∂V ∂n ∂n
where D = ∂ 2 /∂x2 + ∂ 2 /∂y 2 is the Laplace operator on R2 , ∂/∂n is the

outer normal derivative and d` is the euclidean length element. We get
∂f ∂G
Z Z
G(u(z, w)) (∆ + s(1 − s))f (w) dµw = G −f d`
V ∂U ∂n ∂n
by (1.67). Here on the right-hand side the integral of G ∂f /∂n vanishes

as ε → 0 so we are left with
∂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
by (1.68). The last integral vanishes as ε → 0, and the preceding one

is equal to (using euclidean polar coordinates)
1
Z
f (w) d` .
2πε ∂U
This tends to f (z) as ε → 0, thus completing the proof of (1.66).

Chapter 2
Fuchsian groups
In this chapter we give basic facts about groups of motions acting

discontinuously on the hyperbolic plane.
2.1. Definitions.
The group M2 (R) of 2× 2 real matrices is a vector space with inner

product defined by
hg, hi = Trace (ght ) = Tr (ght ) .
One easily checks that kgk = hg, gi1/2 is a norm in M2 (R) and that

2 2 2 2 2 a b
kgk = a + b + c + d , for g = .
c d
Besides its usual properties this norm satisfies kghk ≤ kgk khk.
The embedding G = SL2 (R) −→ M2 (R) induces a metric topology
in G. A subgroup Γ ⊂ G is discrete if the induced topology in Γ is
discrete, i.e. the set
γ ∈ Γ : kγk < r
is finite for any r > 0. Observe that a discrete group is countable.
39
40 Fuchsian groups
Let X be a topological space (Hausdorff) and Γ be a group of

homeomorphisms of X acting on X. We say that Γ acts on X dis-
continuously if the orbit Γx of any x ∈ X has no limit point in X.
Equivalently, any compact subset Y ⊂ X is disjoint with γY for all but
a finite number of γ ∈ Γ. Observe that the stability group Γx of a point
x ∈ X is finite.
Proposition 2.1 (Poincaré). A subgroup of SL2 (R) is discrete if and

only if when considered as a subgroup of P SL2 (R) it acts discontinu-
ously on H.
A subgroup Γ ⊂ P SL2 (R) acting on H discontinuously is called a

Fuchsian group. We shall allow a Fuchsian group to act on Ĉ, of course,
not necessarily discontinuously.
Proposition 2.2. Let Γ be a Fuchsian group and z ∈ Ĉ. Then the

stability group
Γz = γ ∈ Γ : γz = z
is cyclic (not necessarily finite if z ∈ R̂).
An element γ0 of a Fuchsian group is called primitive if γ0 generates

the stability group of its fixed points, and in case γ0 is elliptic it has
the smallest angle of rotation. Any γ other than the identity motion is
a power of a unique primitive element, γ = γ0n , n ∈ Z.
There is a multitude of Fuchsian groups. Surprising is the following

result of J. Nielsen [Ni]: if Γ ⊂ P SL2 (R) is non-abelian and hyperbolic
(it contains only hyperbolic elements besides the identity), then Γ acts
discontinuously on H (an elegant proof was given by C. L. Siegel [Si1]).
A Fuchsian group Γ is said to be of the first kind if every point on

the boundary ∂H = R̂ is a limit (in the Ĉ-topology) of the orbit Γz for
some z ∈ H. Clearly, any subgroup of finite index of a Fuchsian group
of the first kind is a Fuchsian group of the first kind. But a Fuchsian
group of the first kind cannot be too small. Obviously, it cannot be
cyclic, and a fortiori the stability group Γz of a point z ∈ Ĉ is not of
the first kind.
Fundamental domains 41
2.2. Fundamental domains.
A Fuchsian group can be visualized by its fundamental domain.

Two points z, w ∈ Ĉ are said to be equivalent if w ∈ Γz; we then write
z ≡ w (mod Γ). A set F ⊂ H is called a fundamental domain for Γ if
i) F is a domain in H,
ii) distinct points in F are not equivalent,
iii) any orbit of Γ contains a point in F (the closure of F in the
Ĉ-topology).
Any Fuchsian group has a fundamental domain not, of course,
unique. However, all fundamental domains have the same positive vol-
ume (possibly infinite) Z
|F | = dµz .
F
A fundamental domain of a Fuchsian group Γ of the first kind can be

chosen as a convex polygon. Specifically, suppose w ∈ H is not fixed by
any γ ∈ Γ other than the identity motion, then the set

D(w) = z ∈ H : ρ(z, w) < ρ(z, γw) for all γ ∈ Γ, γ 6= 1
is a fundamental domain of Γ; it is called a normal polygon (due to

Dirichlet). One can show that D(w) is a polygon with an even number
or sides (subject to the convention that if a side contains a fixed point
of an elliptic motion of order 2 from Γ, then this point is considered as
a vertex and the side divided into two sides). The sides of D(w) can be
arranged in pairs of equivalent sides so that the side-pairing motions
generate the group Γ.
From the above properties of D(w) follows (for a complete proof
see C. L. Siegel [Si2])
Proposition 2.3. Every Fuchsian group of the first kind has a finite
number of generators and fundamental domain of finite volume.
A Fuchsian group of the first kind will be called more briefly a

finite volume group. The finite volume groups split further into two
categories according to whether the fundamental polygon is compact
(after closure in Ĉ) or not. In the first case we call Γ a co-compact
group.
42 Fuchsian groups
Suppose that the polygon F is not compact. Then F must have a

vertex on R̂, and since the two sides of F which meet at such a vertex
are tangent (because they are orthogonal to R̂), they form a cusp.
Proposition 2.4. A fundamental domain of a finite volume group can

be chosen as a polygon all of whose cuspidal vertices are inequivalent.
For a fundamental domain of Γ chosen as in Proposition 2.4 the

two sides joined at a cuspidal vertex are equivalent so the side-pairing
motion fixes the vertex, is a parabolic motion, and generates the stabil-
ity group of the vertex. For this reason a cuspidal vertex is also called
a parabolic vertex. Conversely, cusps for Γ are exactly the fixed points
of parabolic motions of Γ. Hence, we have
Proposition 2.5. A finite volume group is co-compact if and only if it

has no parabolic elements.
A fundamental polygon all of whose cuspidal vertices are distinct

mod Γ will be convenient for various computations. Throughout we
denote cusps by gothic characters a, b, c, . . . The stability group of a
cusp a is an infinite cyclic group generated by a parabolic motion,

Γa = γ ∈ Γ : γa = a = hγa i ,
say. There exists σa ∈ G such that

1 1
(2.1) σa ∞ = a , σa−1 γa σa = .
1
We shall call σa a scaling matrix of the cusp a; it is determinated up to

composition with a translation from the right side. The semi-strip

(2.2) P (Y ) = z = x + iy : 0 < x < 1, y ≥ Y
is mapped into the cuspidal zone
(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
is compact (after closure) and adjacent to each Fa (Y ) along the horo-

cycles

(2.4) σa L(Y ) , L(Y ) = z = x + iY : 0 < x < 1 .
In this way the fundamental polygon F is partitioned into the central

part F (Y ) and the cuspidal zones Fa (Y ) so that
[
(2.5) F = F (Y ) ∪ Fa (Y ) .
a
Figure 7. Cuspidal zones and central part.
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
There exists a quite explicit construction (in terms of matrix entries

of the group elements) of a fundamental domain that is more practical
for us than the normal polygon. Suppose Γ is not co-compact. By
conjugation we may require that a = ∞ is a cusp whose stability group
1 1
Γ∞ is generated by 1 so a fundamental domain of Γ∞ is any vertical
strip of width 1, say

F∞ = z ∈ H : β < x < β + 1 .
Define F to be the subset of F∞ which is exterior to all the isometric

circles Cγ with γ ∈ Γ, γ ∈
/ Γ∞ (see Section 1.1), i.e.

(2.6) F = z ∈ F∞ : Im z > Im γz for all γ ∈ Γ, γ ∈
/ Γ∞ .
Thus, F consists of points of deformation less than 1 inside the strip

F∞ . One can show that the polygon (2.6) is a fundamental domain
of Γ. We shall call it the standard polygon (it was first introduced by
L. R. Ford [For]). This polygon will be used effectively in Section 2.6
to establish various estimates which are uniform with respect to the
group.
2.3. Basic examples.
There are various ways to construct a finite volume group. One

may start by drawing a convex hyperbolic polygon F ⊂ H of an even
number of sides and finite volume. However, not every such F is a
fundamental domain of a group Γ ⊂ P SL2 (R). The polygon F must
satisfy various conditions. For example, since the action of Γ on F tes-
selates H, the sum of interior angles of F at equivalent vertices is of type
2πm−1 where m is the order of the stability groups for these vertices.
Poincaré has given a complete characterization of fundamental poly-
gons of discrete groups which is quite appealing (the angle conditions
are barely insufficient).
A subgroup Γ ⊂ P SL2 (R) is called a triangular group of type
(α, β, γ) if it is generated by the reflections on the sides of some tri-
angle with interior angles α, β, γ (note that it always takes an even
number of reflections to make a group element, an analytic automor-
phism of Ĉ). Since triangles with the same angles are congruent, all
groups of the same type (regardless of the ordering of angles) are con-
jugate in P SL2 (R). A triangular group is discrete if and only if it is of
Basic examples 45
type (π/p, π/q, π/r), where p, q, r are integers with

1 1 1
2 ≤ p, q, r ≤ +∞ , 0< + + < 1.
p q r
An example of a triangular group is the Hecke group Γq where q is
an integer greater than 2 which is generated by the involution z 7→ −1/z
and the translation z 7→ z + 2 cos(π/q).
Therefore, a fundamental

domain of Γq is the triangle F = z ∈ H : |x| < λ/2 , |z| > 1 of
volume |F | = π(1 − 2/q). Moreover, i is an elliptic vertex of order 2,
ζq = e(1/2q) is an elliptic vertex of order q, ∞ is the cusp, and g = 0
is the genus of Γq \H. One can show that Γq is maximal, i.e. is not
contained in any smaller volume group.
Figure 8. Fundamental domain for Γq .
There is a general group-theoretical recipe for finite volume groups,

but it does not reveal much geometry (see R. Fricke and F. Klein [Fr-
Kl]).
Proposition 2.6. Any finite volume subgroup of P SL2 (R) is gener-

ated by primitive motions A1 , . . . , Ag , B1 , . . . , Bg , E1 , . . . , E` , P1 , . . . , Ph
satisfying the relations
mj
[A1 , B1 ] · · · [Ag , Bg ]E1 · · · E` P1 · · · Ph = 1 , Ej = 1,
where Aj , Bj are hyperbolic motions, [Aj , Bj ] stands for the commuta-
tor, g is the genus of Γ\H, Ej are elliptic motions of order mj ≥ 2, Pj
are parabolic motions and h is the number of inequivalent cusps.
46 Fuchsian groups
The symbol (g; m1 , . . . , m` ; h) is group invariant, and is called the

signature of Γ; it satisfies the Gauss-Bonnet formula
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.
First we introduce the quaternion group

√ √ √
a+b n (c + d n) p
Γ(n, p) = √ √ √ :
(c − d n) p a−b n
(2.8)
2 2 2 2
a, b, c, d ∈ Z , a − b n − c p + d np = 1 .
Here n is a positive integer, and p is an odd prime number such that

(n/p) = −1, i.e. n is not a quadratic residue modulo p. Using this
property one can show that every element different from ±1 has trace
of absolute value greater than 2; whence it is hyperbolic. Therefore
Γ(n, p) is discrete by a general result of Nielsen (see Section 2.1) and
co-compact by Proposition 2.5.
Our next example is the familiar modular group

a b
(2.9) SL2 (Z) = : a, b, c, d ∈ Z , ad − bc = 1
c d
with its fundamental domain F = {z = x + iy : |x| < 1/2, |z| > 1}

which is the normal polygon D(iv) with v > 1 as well as a standard
√
polygon. Moreover, i is an elliptic vertex of order 2, ζ = (1 + i 3)/2 is
an elliptic vertex of order 3, ∞ is the cusp, and g = 0 is the genus of
SL2 (Z)\H.
Let N be a positive integer. The principal congruence group of level

N is the subgroup Γ(N ) of the modular group consisting of matrices
congruent to the identity modulo N , i.e.

1
(2.10) Γ(N ) = γ ∈ SL2 (Z) : γ ≡ (mod N ) .
1
Basic examples 47
Γ(N ) is a normal subgroup of Γ(1) = SL2 (Z) of index

Y
µ = [Γ(1) : Γ(N )] = N 3 (1 − p−2 ) .
p|N
The number of inequivalent cusps is

Y
h = µ N −1 = N 2 (1 − p−2 ) .
p|N
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.
Figure 9. A fundamental domain for the modular group.
Any subgroup of the modular group containing Γ(N ) is called a

congruence group of level N . Two basic examples are

∗ ∗
Γ0 (N ) = γ ∈ SL2 (Z) : γ ≡ (mod N ) ,
∗

1 ∗
Γ1 (N ) = γ ∈ SL2 (Z) : γ ≡ (mod N ) .
1
48 Fuchsian groups
The group Γ0 (N ) is called the Hecke congruence group of level N , and

has index
Y 1
(2.11) µ = [Γ0 (1) : Γ0 (N )] = N 1+ .
p
p|N
The number of inequivalent elliptic fixed points of order 2 is

Y −1
(2.12) ν2 = 1+ if 4 - N ,
p
p|N
and the number of those of order 3 is

Y −3
(2.13) ν3 = 1+ if 9 - N .
p
p|N
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
In particular, if N is prime there are two inequivalent cusps for Γ 0 (N )

at ∞ and 0; they are equivalent to 1/N and 1, respectively. All of the
above properties of congruence groups can be found in [Sh].
2.4. The double coset decomposition.
Let Γ be a group of finite volume but not co-compact. For such a

group the Fourier expansion at cusps is available to help examine auto-
morphic forms through the coefficients. The Fourier expansion will be
derived from a decomposition of Γ into double cosets with respect to
the stability groups of cusps. Choose two cusps a, b for Γ (not necessar-
ily distinct) and the corresponding scaling matrices σa , σb (determined
up to translations from the right side, see (2.1)). Let us recall that
σa ∞ = a, σa−1 Γa σa = B and σb ∞ = b, σb−1 Γb σb = B, where Γa , Γb
The double coset decomposition 49
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 .
∗
Suppose Ω∞ is not empty. Take ω∞ = σa−1 γσb ∈ Ω∞ with γ ∈ Γ

and evaluate at b getting γb = σa ω∞ ∞ = σa ∞ = a. This shows that
the cusps a, b are equivalent, the stability groups are conjugate, and
ω∞ = σa−1 γσb is a translation. Now suppose ω1 = σa−1 γ1 σb is another
element of Ω∞ . We obtain
γ γ1−1 a = σa ω∞ ω1−1 σa−1 a = σa ω∞ ω1−1 ∞ = σa ∞ = a
showing that γγ1−1 ∈ Γa ; whence ω∞ ω1−1 = σa−1 γγ1−1 σa ∈ σa−1 Γa σa =

B. Combining the results we conclude that the subset Ω∞ is not empty
if and only if the cusps a, b are equivalent, in which case
(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
(2.18) Ωd/c = B ωd/c B
for some

∗ ∗
(2.19) ωd/c = ∈ σa−1 Γσb ,
c d
with c > 0. The relation

1 m a ∗ 1 n a + cm ∗
(2.20) =
1 c d 1 c d + cn
50 Fuchsian groups
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

γ = σa ωσb−1 ∈ Γ, γ1 = σa ω1 σb−1 ∈ Γ, we obtain γγ1−1 = σa ω ω1−1 σa−1 =

σa ∗ ∗∗ σa−1 . Evaluating at a we infer that γγ1−1 ∈ Γa ; whence ω ω1−1 ∈

B which shows that a ≡ a1 (mod c) as claimed.
By the above investigation we have established the following
Theorem 2.7. Let a, b be cusps for Γ. We then have a disjoint union

[ [
(2.21) σa−1 Γσb = δab Ω∞ ∪ Ωd/c
c>0 d(mod c)
where δab = 1 if a, b are equivalent, otherwise it vanishes. Moreover

−1 ∗ ∗

c, d run over numbers such that σa Γσb contains c d .
As an example take the Hecke congruence group Γ = Γ0 (q) and

the cusps ∞, 0. The scaling matrices are
√
1 −1/ q
σ∞ = , σ0 = √ .
1 q
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.
2.5. Kloosterman sums.
The double coset decomposition (2.21) is a tool for working with

the group Γ by means of additive characters. Specifically, to a double
coset Ωd/c = B ωd/c B we shall attach the character
d
Ωd/c (m) = e m , m ∈ Z.
c
Accordingly, to the inverse coset Ω−1 −1

d/c = Bωd/c B = Bω−a/c B = Ω−a/c
is attached the character
a
Ω−a/c (n) = e − n , n ∈ Z.
c
Given m, n ∈ Z and c in the set

∗ ∗
(2.22) Cab = c>0: ∈ σa−1 Γσb
c ∗
the Kloosterman sum is created by convolving of the above characters

as follows:
X d a
(2.23) Sab (m, n; c) = e m +n .
a ∗
c c
−1
c d ∈B\σa Γσb /B
We shall refer to c as the modulus and to m, n as frequencies. Note the

symmetries Cab = Cba and Sab (m, n; c) = Sba (n, m; c). Observe that
the Kloosterman sum depends on the choice of the scaling matrices in
the following simple fashion
Sa0 b0 (m, n; c) = e(αm + βn) Sab (m, n; c) ,

52 Fuchsian groups
if a0 = τa a, σa0 = τa σa n(α) and b0 = τb b, σb0 = τb σb n(β) for some

τa , τb in Γ and n(α), n(β) in A. In particular this shows that

∗ ∗
(2.24) Sab (0, 0; c) = # d (mod c) : ∈ σa−1 Γσb
c d
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)
defined for all positive integer moduli c. We have a deep bound
(2.25) |S(m, n; c)| ≤ (m, n, c)1/2 c1/2 τ (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
For m = n = 0 we get the Euler function S(0, 0; c) = ϕ(c) and

∞
X ζ(2s − 1)
(2.28) c−2s S(0, 0; c) = .
c=1
ζ(2s)
Basic estimates 53
2.6. Basic estimates.
Let us return to the general case of a finite volume group Γ which

is not co-compact. In applications we shall need some control over the
number of cosets in the decomposition (2.21). Let c(a, b) denote the
smallest element of the set Cab . Put ca = c(a, a), i.e.

∗ ∗
(2.30) ca = min c > 0 : ∈ σa−1 Γσa .
c ∗
That ca exists is seen from the construction of the standard polygon

for the group σa−1 Γσa ; c−1
a is the radius of the largest isometric circle.
Since the polygon contains the semi-strip P (c−1
a ) of volume ca , we have
(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
Proposition 2.8. For any c ∈ Cab we have
(2.32) Sab (0, 0; c) ≤ c−1

ab c
2
where cab = max{ca , cb }. Furthermore, we have on average

X
(2.33) c−1 Sab (0, 0; c) ≤ c−1
ab X .
c≤X
Proof. By symmetry we can assume without loss of generality that

∗ ∗ ∗0 ∗0
ca ≥ cb . If γ = c d and γ = c d , where c, c0 > 0 are both in
0

σa−1 Γσb , then

00 0 −1 ∗ ∗
γ =γγ = ∈ σa−1 Γσa
c00 ∗
where c00 = c0 d − cd0 . If c00 = 0 then the cusps a, b are equivalent,

γ 00 = 1 ∗1 , c0 = c and d0 = d. If c00 6= 0 then |c00 | ≥ ca ; whence

d0 d ca
(2.34) 0 − ≥ 0 .

c c cc
54 Fuchsian groups
In particular for c0 = c this yields
(2.35) |d0 − d| ≥ ca c−1 .
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).
Corollary 2.9. The Kloosterman sums satisfy the following trivial

bounds
(2.37) |Sab (m, n; c)| ≤ c−1

ab c
2
and
X
(2.38) c−1 |Sab (m, n; c)| ≤ c−1
ab X .
c≤X
Lemma 2.10. Let a be a cusp for Γ, z ∈ H and Y > 0. We have
10
# γ ∈ Γa \Γ : Im σa−1 γz > Y < 1 +

(2.39) .
ca Y
Proof. Conjugating the group we can assume that a = ∞, σa = 1

and Γa = B. Then the fundamental domain of Γa is the strip
P = {z = x + iy : 0 < x < 1, y > 0} .
Let F be the standard polygon of Γ so F consists of points in P of

deformation less than 1. For the proof we may assume that z ∈ F .
Then for any γ = ∗c d∗ ∈ Γ with c > 0 the point γz lies on the
isometric circle |cz + d| = 1 or in its interior. In any case we have
Basic estimates 55
|cz + d| ≥ 1. Since Im γz = y|cz + d|−2 > Y , this implies y > Y ,

c < y −1/2 Y −1/2 , |cx + d| < y 1/2 Y −1/2 . By the last inequality and the
spacing property (2.34) we estimate the number of pairs {c, d} with
C ≤ c < 2 C by 1 + 8 c−1 a Cy
1/2 −1/2
Y ≤ 10 c−1
a Cy
1/2 −1/2
Y . Adding
−n −1/2 −1/2 −1 −1
these bounds for C = 2 y Y , n ≥ 1, we get 10 ca Y . This
is an estimate for the number of relevant γ’s not in Γa . Finally adding
1 to account for Γa we obtain what is claimed.
Lemma 2.11. Let a be a cusp of Γ, z, w ∈ H and δ > 0. We have
#{γ ∈ σa−1 Γσa : u(γz, w) < δ}

(2.40) p
δ(δ + 1)(Im w + c−1 a ) + (δ + 1)(ca Im w)
−1
+1
where the implied constant is absolute.
Proof. Without loss of generality we can assume that a = ∞, σa = 1

and Γa = B. The condition u(γz, w) < δ is equivalent to
(2.41) |γz − w| < 2 (δ Im w Im γz)1/2 .
Taking the imaginary part we infer that Y1 < Im γz < Y2 , where Y1 =

Im w/4(δ + 1) and Y2 = 4(δ + 1)Im w. Looking at the real part we find
that the number of elements γ 0 ∈ Γa such that γ 0 γ satisfies (2.41) does
not exceed 1 + 4 (δ Im w Im γz)1/2 . Therefore, the total number of γ’s
satisfying (2.41) does not exceed
X
1 + 4 (δ Im w Im γz)1/2 .

γ∈Γa \Γ
Y1 <Im γz<Y2
Applying Lemma 2.10, by partial summation, this is bounded by

−1 1/2 −1/2
1 + c−1
a Y1 + (δ Im w)1/2 (Y2 + c−1
a Y1 ),
which yields (2.40).
Corollary 2.12. Let z ∈ H and δ > 0. We have

p
(2.42) #{γ ∈ Γ : u(γz, z) < δ} δ(δ + 1) yΓ (z) + δ + 1
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.
Let Γ be a finite volume group. A function f : H −→ C is said to

be automorphic with respect to Γ if it satisfies the periodicity condition
f (γz) = f (z) , for all γ ∈ Γ .
Therefore, f lives on the Riemann surface Γ\H. We denote the space

of such functions by A(Γ\H).
Some automorphic functions can be constructed by the method of

images. Take a function p(z) of sufficiently rapid decay on H. Then
X
f (z) = p(γz) ∈ A(Γ\H) .
γ∈Γ
Very important automorphic functions are given by a series over the

cosets of an infinite subgroup of Γ rather than over the whole group.
For such construction, of course, the generating function p(z) must be
invariant with respect to this subgroup. If Γ is not co-compact, we take
57
58 Automorphic forms
cosets with respect to the stability group of a cusp. In this way we

obtain the Poincaré series
X
(3.1) Ea (z|p) = p(σa−1 γz) ∈ A(Γ\H)
γ∈Γa \Γ
where p(z) is any function on H which is B-invariant (periodic in x of

period 1) and satisfies a suitable growth condition.
From now on we consider a non co-compact group Γ. Let a be a

cusp for Γ and σa be a scaling matrix. Any f ∈ A(Γ\H) satisfies the
transformation rule

1 m
f (σa z) = f (σa z) ,
1
for all m ∈ Z; therefore, it makes sense to write the Fourier expansion

X
(3.2) f (σa z) = fan (y) e(nx)
n
where the coefficients are given by

Z 1
fan (y) = f (σa z) e(−nx) dx .
0
If f is smooth then the series (3.2) converges absolutely and uniformly

on compacta.
An automorphic function f ∈ A(Γ\H) which is an eigenfunction of

the Laplace operator
(∆ + λ)f = 0 , λ = s(1 − s)
is called an automorphic form (of Maass [Ma]). We denote by As (Γ\H)

the space of automorphic forms with respect to Γ for the eigenvalue
λ = s(1 − s). Thus As (Γ\H) = A1−s (Γ\H) ⊂ A(Γ\H).
For an automorphic form the Fourier expansion (3.2) can be made

more explicit. In this case Proposition 1.5 yields
Introduction 59
Theorem 3.1. Any f ∈ As (Γ\H) satisfying the growth condition
(3.3) f (σa z) = o(e2πy ) , as y → +∞
has the expansion

X
(3.4) f (σa z) = fa (y) + fâ (n) Ws (nz)
n6=0
where the zero-th term fa (y) is a linear combination of the functions

(1.23), i.e.
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
(3.6) fâ (n) eε|n| ,
for any ε > 0, with the implied constant depending on ε and f . There-
fore, as y → +∞, we have
(3.7) f (σa z) = fa (y) + O(e−2πy ) .
We shall often examine the behaviour of an automorphic function

in various cuspidal zones (see (2.3) and Figure 7). When it comes to
controling the growth in the a-zone, we shall express the relevant bounds
in terms of Im σa−1 z. In order to work in all zones simultaneously we
introduce the invariant height of z by
(3.8) yΓ (z) = max max {Im σa−1 γz} .

a γ∈Γ
Thus, we have yΓ (σa z) = y if y is sufficiently large. If the group is

understood from the context, we drop the subscript Γ. Observe that
yΓ (z) is bounded below by a positive constant depending only on Γ,
say,
yΓ = min yΓ (z) > 0 .
z∈H
For example if Γ = Γq is the Hecke triangle group, then yΓ = sin(π/q)

(see Figure 8).
Our objective is to expand an automorphic function into automor-

phic forms subject to suitable growth conditions. The main results hold
in the Hilbert space
L(Γ\H) = {f ∈ A(Γ\H) : kf k < ∞}
with the inner product

Z
hf, gi = f (z) g(z) dµz .
F
Observe that bounded automorphic functions are in L(Γ\H) because

F has finite volume. The inner product is a powerful tool for analytic
as well as arithmetic studies of automorphic forms. Be aware of the
positivity of the norm kf k = hf, f i1/2 . Contemplate the arguments
exploiting this obvious fact throughtout the lectures.
When f is square integrable we shall improve the estimates of
Theorem 3.1 substantially.
Theorem 3.2. If f ∈ As (Γ\H) ∩ L(Γ\H), then it has bounded Fourier

coefficients. In fact assuming kf k = 1 and Re s ≥ 1/2 (normalization
conditions) we have
X
|n| |fâ (n)|2 (c−1
a N + |s|) e
π|s|
|n|≤N
where the implied constant is absolute. Hence
|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) = fâ (0) y 1−s with the coefficient bounded by
1 1/2 −s+1/2
fâ (0) s − ca .
2
Proof. We get by Parseval’s identity that
X Z 1
2
|fa (y)| + |fâ (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 |fâ (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
which yields our claim upon taking Y = |s|/2πy.
3.2. The Eisenstein series.
We begin with the Poincaré series (3.1) evolved from
p(z) = ψ(y) e(mz)
where m is a non-negative integer and ψ is a smooth function on R+ .

For the absolute convergence it is sufficient that (see (2.39))
(3.9) ψ(y) y (log y)−2 , as y → 0 .

This yields a kind of weighted Poincaré series, which was considered by

A. Selberg [Se1],
X
Eam (z|ψ) = ψ(Im σa−1 γz) e(mσa−1 γz) .
γ∈Γa \Γ
For m = 0 this becomes a kind of weighted Eisenstein series,

X
(3.10) Ea (z|ψ) = ψ(Im σa−1 γz) .
γ∈Γa \Γ
For ψ(y) = y s with Re s > 1 we obtain the Eisenstein series

X
(3.11) Ea (z, s) = (Im σa−1 γz)s .
γ∈Γa \Γ
Since p(z) = y s is an eigenfunction of ∆ with eigenvalue λ = s(1 − s),

so is the Eisenstein series, i.e. Ea (z, s) ∈ As (Γ\H) if Re s > 1. But
Ea (z, s) is not square integrable over F .
If ψ is compactly supported on R+ , we call Ea (z|ψ) an incomplete

Eisenstein series. In this case Ea (z|ψ) is a bounded automorphic func-
tion on H and hence clearly square integrable over F . The incomplete
Eisenstein series is not an automorphic form because it fails to be an
eigenfunction of ∆. However, by Mellin’s inversion one can represent
the incomplete Eisenstein series as a contour integral of the Eisenstein
series
Z
1
(3.12) Ea (z|ψ) = Ea (z, s) ψ̂(s) ds ,
2πi (σ)
where σ > 1 and

Z +∞
(3.13) ψ̂(s) = ψ(y) y −s−1 dy .
0
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
To pursue the above analysis we select two linear subspaces
B(Γ\H) the space of smooth and bounded automorphic functions,

E(Γ\H) the space of incomplete Eisenstein series.
We have the inclusions E(Γ\H) ⊂ B(Γ\H) ⊂ L(Γ\H) ⊂ A(Γ\H). The

space B(Γ\H) is dense in L(Γ\H) but E(Γ\H) need not be.
3.3. Cusp forms.
Let us examine the orthogonal complement to E(Γ\H) in B(Γ\H).

Take f ∈ B(Γ\H), Ea (∗|ψ) ∈ E(Γ\H) and compute the inner product
Z X
hf, Ea (∗|ψ)i = f (z) ψ(Im σa−1 γz) dµz .
F γ∈Γa \Γ
Interchange the summation with the integration, make the substitution

z 7→ γ −1 σa z and use the automorphy of f to get
X Z
hf, Ea (∗|ψ)i = f (σa z) ψ(y) dµz .
γ∈Γa \Γ σa
−1
γF
Here, as γ runs over Γa \Γ the sets σa−1 γF cover the strip P = {z ∈ H :

0 < x < 1} once (for an appropriate choice of representatives) giving
Z Z +∞ Z 1
f (σa z) ψ(y) dµz = f (σa z) dx ψ(y) y −2 dy .
P 0 0
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
Lemma 3.3. Let f (z) be an automorphic function absolutely integrable

over F . Let Ea (z|ψ) be an incomplete Eisenstein series associated with
the cusp a and a test function ψ ∈ C0∞ (R+ ). Then we have
Z +∞
(3.14) hf, Ea (∗|ψ)i = fa (y) ψ(y) y −2 dy
0
where fa (y) is the zero-th term in the Fourier expansion of f at a .

Now suppose f ∈ B(Γ\H) is orthogonal to the space E(Γ\H). Then

the integral (3.14) vanishes for all ψ ∈ C0∞ (R+ ). This is equivalent with
the condition
(3.15) fa (y) ≡ 0 , for any cusp a.
Denote by C(Γ\H) the space of smooth, bounded automorphic

functions whose zero-th terms at all cusps vanish. Therefore, we have
just proved the orthogonal decomposition
(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
Cs (Γ\H) = C(Γ\H) ∩ As (Γ\H)
the space of cusp forms with eigenvalue λ = s(1−s). Every f ∈ Cs (Γ\H)

has the expansion
X
f (σa z) = fâ (n) Ws (nz)
n6=0
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) .
In particular this shows that a cusp form is bounded on H. Also by

Theorem 3.2 the Fourier coefficients fâ (n) of a cusp form are bounded.
They will be the subject of intensive study in forthcoming chapters.
Clearly, ∆ : C(Γ\H) −→ C(Γ\H) and ∆ : E(Γ\H) −→ E(Γ\H). It
will be shown that ∆ has a pure point spectrum in C(Γ\H), i.e. the
space C(Γ\H) is spanned by cusp forms. This will be accomplished by
means of compact integral operators. On the other hand, in the space
E(Γ\H) the spectrum will turn out to be continuous except for a finite
Fourier expansion of the Einsenstein series 65
dimensional subspace of point spectrum. Here the analytic continua-

tion of the Eisenstein series is the key issue. After this is established,
the spectral resolution of ∆ in E(Γ\H) will evolve from (3.12) at once
by contour integration. The eigenpacket of the continuous spectrum
consists of the Eisenstein series Ea (z, s) on the line Re s = 1/2 (analyt-
ically continued), and the point spectrum subspace is spanned by the
residues of Ea (z, s) at poles on the segment 1/2 < s ≤ 1.
3.4. Fourier expansion of the Eisenstein series.
We begin by expanding a general Poincaré series Ea (z|p) associated

with a cusp a and a test function p, see (3.1). Let b be another cusp not
necessarily different from a. Applying the double coset decomposition
(2.21) we infer that
X X
Ea (σb z|p) = p(σa−1 γσb z) = p(τ z)
γ∈Γa \Γ τ ∈B\σa
−1
Γσb
X X X
= δab p(z) + p(ωcd (z + n)) ,
c>0 d(mod c) n∈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)
In the special case p(z) = ψ(y) e(mz) where m is a non-negative integer

we get
Z +∞
d a y c−2 −m c−2
e(nx + n + m ) ψ e − nt dt .
c c −∞ t2 + y 2 t + iy
Summing over the coset representatives c, d we encounter the Kloost-

erman sum Sab (m, n; c) and obtain the Fourier expansion
Eam (σb z|ψ) = δab e(mz) ψ(y)

X X Z +∞
(3.17) y c−2 −m c−2
+ e(nx) Sab (m, n; c) ψ 2 2 e − nt dt
c>0 −∞ t +y t+iy
n∈Z
where δab is the diagonal symbol of Kronecker.

It remains to compute the integral in (3.17). We give results only
for m = 0 and ψ(y) = y s which is the case of the Eisenstein series. We
have
Z +∞
Γ(s − 1/2) 1−2s
(3.18) (t2 + y 2 )−s dt = π 1/2 y
−∞ Γ(s)
and
Z +∞
(t2 + y 2 )−s e(−nt) dt
(3.19) −∞
= 2π s Γ(s)−1 |n|s−1/2 y −s+1/2 Ks−1/2 (2π|n|y)
for n 6= 0 (see Appendix B). Substituting these evaluations into (3.17)

we arrive at the explicit Fourier expansion of the Eisenstein series.
Theorem 3.4. Let a, b be cusps for Γ and Re s > 1. We have

X
(3.20) Ea (σb z, s) = δab y s + ϕab (s) y 1−s + ϕab (n, s) Ws (n z)
n6=0
where
Γ(s − 1/2) X −2s

(3.21) ϕab (s) = π 1/2 c Sab (0, 0; c) ,
Γ(s) c
X
(3.22) ϕab (n, s) = π s Γ(s)−1 |n|s−1 c−2s Sab (0, n; c) ,
c
and Ws (z) is the Whittaker function given by (1.26).
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
Corollary 3.5. For s on the line Re s = σ > 1 we have

(3.23) Ea (σb z, s) = δab y s + ϕab (s) y 1−s + O (1 + y −σ ) e−2πy
uniformly in z ∈ H, the implied constant depending on s and the group.
We have mentioned in the conclusion of the previous section that

the analytic continuation of Ea (z, s) to Re s ≥ 1/2 would be required
for the spectral decomposition of the space E(Γ\H). For some groups
the continuation can be deduced from the Fourier expansion; we shall
show it for the modular group. By (3.21) and (2.28) we get
Γ(s − 1/2) ζ(2s − 1)

(3.24) ϕ(s) = π 1/2
Γ(s) ζ(2s)
where ζ(s) is the Riemann zeta-function. By (3.22) and (2.27) we get
X a s−1/2
(3.25) ϕ(n, s) = π s Γ(s)−1 ζ(2s)−1 |n|−1/2
b
ab=|n|
(since there is only one cusp a = b = ∞ we have dropped the subscript

ab for simplicity).
Since the Fourier coefficients ϕ(s), ϕ(n, s) are meromorphic in the
whole complex s-plane and the Whittaker function Ws (z) is entire in
the s-variable, the Fourier expansion (3.20) furnishes the meromorphic
continuation of E(z, s) to all s ∈ C. In the half-plane Re s ≥ 1/2 there
is only one simple pole at s = 1 with constant residue
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
(3.27) θ(s) = π −s Γ(s) ζ(2s)
so that
(3.28) ϕ(s) = θ(1 − s) θ(s)−1

by the functional equation for the Riemann zeta-function. Now we can

write the Fourier expansion (3.20) in the elegant fashion
θ(s) E(z, s) = θ(s) y s + θ(1 − s) y 1−s

∞
(3.29) √ X
+4 y ηs−1/2 (n) Ks−1/2 (2πny) cos(2πnx)
n=1
where
X a t
(3.30) ηt (n) = .
b
ab=n
Since the right side of (3.29) is invariant under the change s 7→ 1 − s,

it yields the following functional equation
(3.31) θ(s) E(z, s) = θ(1 − s) E(z, 1 − s) .
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
The spectral theorem.

Discrete part
4.1. The automorphic Laplacian.
The Laplace operator ∆ acts on all smooth automorphic functions.

However, for the purpose of the spectral decomposition of L(Γ\H) we
choose the initial domain of ∆ to be

D(Γ\H) = f ∈ B(Γ\H) : ∆f ∈ B(Γ\H) ,
which is dense in L(Γ\H). We shall show that ∆ is symmetric and
non-negative. Therefore, by Theorem A.3, ∆ has a unique self-adjoint
extension to L(Γ\H).
By Stokes’ theorem we have
∂f
Z Z Z
∆f g dµz = − ∇f ∇g dx dy + g d`
F F ∂F ∂n
where F is a bounded domain in R2 with a continuous and piecewise

smooth boundary ∂F , f, g are smooth functions, ∇f = [∂f /∂x, ∂f /∂y]
is the gradient of f , ∂f /∂n is the outer normal derivative and d` is
the euclidean length element. For F ⊂ H the boundary integral can be
given a hyperbolic invariant form
∂f
Z
g d`
∂F ∂n
69
70 The spectral theorem. Discrete part
where ∂/∂n = y∂/∂n and d` = y −1 d` are G-invariant. In this form

the Stokes’ formula remains valid for any polygon F ⊂ H of finite area
provided f, g, ∆f, ∆g are all bounded. Letting F be a fundamental
polygon for a group Γ we find that the boundary integral vanishes
because the integrals along equivalent sides cancel out. Therefore we
obtain
Lemma 4.1. For f, g ∈ D(Γ\H) we have

Z
(4.1) h−∆f, gi = ∇f ∇g dx dy ;
F
whence
(4.2) h∆f, gi = hf, ∆gi
so ∆ is symmetric. Moreover
Z
(4.3) h−∆f, f i = |∇f |2 dx dy ≥ 0
F
so −∆ is non-negative.
Observe that the quantity y 2 ∇f ∇g = y 2 (fx gx + fy gy ) is G-invari-

ant so the integral (4.1) does not depend on the choice of a fundamental
domain.
From Lemma 4.1 it follows that the eigenvalue λ = s(1 − s) of an
eigenfunction f ∈ D(Γ\H) is real and non-negative. Therefore, either
s = 1/2 + it with t ∈ R or 0 ≤ s ≤ 1.
4.2. Invariant integral operators on C(Γ\H).
The spectral resolution of ∆ in C(Γ\H) will be performed by means

of invariant integral operators. Recall that such an operator is given by
a point-pair invariant kernel k(z, w) = k(u(z, w)) which yields
Z
(Lf )(z) = k(z, w) f (w) dµw
H
for f : H −→ C. Restricting the domain of L to automorphic functions

we can write Z
(Lf )(z) = K(z, w) f (w) dµw
F
Invariant integral operators on C(Γ\H) 71
for f ∈ A(Γ\H), where F is a fixed (once and for all) fundamental

domain of Γ and the new kernel is given by the series
X
(4.4) K(z, w) = k(z, γw) .
γ∈Γ
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
where fa (y) is the zero-th term in the Fourier expansion of f at a.

Hence, if fa (y) is identically zero then so is ga (y). This proves
Proposition 4.1. An invariant integral operator L maps the subspace

C(Γ\H) of B(Γ\H) into itself,
L : C(Γ\H) −→ C(Γ\H) .
Next we examine the automorphic kernel K(z, w). Unfortunately

K(z, w) is not bounded on F × F no matter how small you make the
support of k(u). The reason is that when z, w approach the same cusp,
the number of terms which count in (4.4) grows to infinity. In order
to get a bounded kernel we shall substract from K(z, w) the so called
“principal parts”
X Z +∞
(4.5) Ha (z, w) = k(z, σa n(t)σa−1 γw) dt .
γ∈Γa \Γ −∞
Clearly Ha (z, w) is a well defined automorphic function in the second

variable. An important fact is that the principal parts do not alter the
action of L on C(Γ\H) (see Corollary 4.4 and the remarks after it).
Lemma 4.2. For z, w ∈ H we have uniformly
(4.6) Ha (σa z, w) 1 + Im z .
Proof. Changing w into σa w we need to estimate
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
Lemma 4.2 shows that Ha (z, w) is bounded in the second variable
(4.7) Ha (z, ·) ∈ B(Γ\H) .
Proposition 4.3. Given z ∈ H the principal part Ha (z, w) as a func-

tion in w is orthogonal to the space C(Γ\H), i.e.
(4.8) hHa (z, ·), f i = 0 if f ∈ C(Γ\H) .

Invariant integral operators on C(Γ\H) 73
Proof. Changing z into σa z we obtain by unfolding the involved

integral over the fundamental domain that
Z +∞ Z 1 Z +∞
hHa (σa z, ·), f i = k(z, n(t)w) dt f (σa w) dµw
0 0 −∞
Z +∞ Z +∞ Z 1
= k(z, t + iv) dt f (σa w) du v −2 dv
0 −∞ 0
where w = u + iv. The last integral is equal to fa (v) so it vanishes,

proving (4.8).
We define the total “principal part” of the kernel K(z, w) by adding

all Ha (z, w) over inequivalent cusps
X
(4.9) H(z, w) = Ha (z, w) .
a
Then we substract H(z, w) from K(z, w) and call the difference
(4.10) K̂(z, w) = K(z, w) − H(z, w)
the “compact part” of K(z, w). This becomes a kernel on F × F of

an integral operator L̂, say, acting on functions f : F −→ C. From
Proposition 4.3 we obtain
Corollary 4.4. For f ∈ C(Γ\H) we have Lf = L̂f .
Proposition 4.5. Let F be a fundamental polygon for Γ whose cuspidal

vertices are all distinct mod Γ. Then the kernel K̂(z, w) is bounded on
F × F.
Proof. As γ ranges over non-parabolic motions, the points z, γw are

separated by an arbitrarily large distance for almost all γ uniformly in
z, w ∈ F . Therefore, since k(u) is compactly supported, we have
X
K(z, w) = k(z, γ w) + O(1) .
γ parabolic
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) .
−∞
Combining both estimates we can write

X
K̂(z, w) = Ja (z, w) + O(1) ,
a
where Ja (z, w) is defined by

X Z +∞
Ja (z, w) = k(z, γw) − k(z, σa n(t)σa−1 w) dt .
γ∈Γa −∞
It remains to show that Ja (z, w) is bounded in F ×F . This is the crucial

part of the proof. We apply the Euler-MacLaurin formula
X Z Z
F (b) = F (t) dt + ψ(t) dF (t) ,
b∈Z
where ψ(t) = t − [t] − 1/2, getting

X Z +∞
Ja (σa z, σa w) = k(z, w + b) − k(z, w + t) dt
b∈Z −∞
Z +∞ Z +∞
= ψ(t) dk(z, w + t) |k 0 (u)| du 1 .
−∞ 0
Remarks. The basic results established in this section continue to hold

true for kernels k(u) which are not necessarily compactly supported but
decay fast, a sufficient condition being that
(4.11) k(u) , k 0 (u) (u + 1)−2 .
Such a generalization can be derived from the compact case by a suitable

approximation or by refining the above estimates. One should also
realize that Ha (z, w) is an incomplete Eisenstein series in the second
variable. Indeed we have
Z +∞
y0
k(z, n(t)z 0 ) dt = y 0 y g(log )
p
(4.12)
−∞ y
with g(r) given by (1.62); whence

X
Ha (σa z, w) = ψ(Im σa−1 γw) = Ea (w|ψ)
γ∈Γa \Γ
Spectral resolution of ∆ in C(Γ\H) 75
√
where ψ(v) = vy g(log(v/y)). Although ψ(v) might not be compactly
supported, it decays quite rapidly. Therefore, Proposition 4.3 comes
straight from the definition of the space C(Γ\H).
4.3. Spectral resolution of ∆ in C(Γ\H).
To this end we shall employ a proper invariant integral operator

L. By Proposition 4.5 the modified operator L̂ is of Hilbert-Schmidt
type on L2 (F ); in fact L̂ has a bounded kernel. Therefore, the Hilbert-
Schmidt theorem applies to L̂ (see Appendix A.3). Any function from
the range of L̂ has the series expansion
X
(4.13) f= hf, uj i uj (z) .
j≥0
Here {uj }j≥0 is any maximal orthonormal system of eigenfunctions of

L̂ in the space L2 (F ). But the range of L̂ is definitely not dense in
L2 (F ) so the spectral expansion (4.13) does not hold for all f ∈ L2 (F ).
It may happen that L̂ is the trivial operator giving nothing but the zero
function to expand.
To find a good L consider the resolvent operator

Z
−(Rs f )(z) = Gs (u(z, w)) f (w) dµw
H
whose kernel Gs (u) is the Green function, singular at u = 0. In order

to kill the singularity take the difference (Hilbert’s formula for iterated
resolvent)
(4.14) L = Rs − Ra = (s(1 − s) − a(1 − a)) Rs Ra
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
g = (s(1 − s) − a(1 − a))−1 (∆ + a(1 − a))(∆ + s(1 − s))f ∈ D(Γ\H)

such that Lg = f . Moreover, if f ∈ C(Γ\H) then g ∈ C(Γ\H) so by

Corollary 4.4 we also get L̂g = f . Therefore, the subspace C(Γ\H) ∩
D(Γ\H) is in the range of L̂ and it is dense in C(Γ\H). This, together
with the Hilbert-Schmidt theorem and Corollary 4.4, proves the follow-
ing
Proposition 4.6. Let L : D(Γ\H) −→ D(Γ\H) be the integral oper-

ator given by (4.14). Then L maps the subspace C(Γ\H) densely into
itself where it has pure point spectrum. Let {uj } be a complete or-
thonormal system of eigenfunctions of L in C(Γ\H). Then any f ∈
C(Γ\H) ∩ D(Γ\H) has the expansion (4.13), which converges absolutely
and uniformly on compacta.
Since L, ∆ commute and they are symmetric operators, it follows

from Corollary A.9 for the space H = C(Γ\H) that L has a complete
orthonormal system of eigenfunctions in C(Γ\H) which are cusp forms.
Applying Proposition 4.6 for this system we conclude the spectral res-
olution of ∆ in C(Γ\H).
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
converging in the norm topology. If f ∈ C(Γ\H) ∩ D(Γ\H) then the

series converges absolutely and uniformly on compacta.
Remark. The space C(Γ\H) is perhaps trivial for generic groups as

conjectured by Phillips and Sarnak [Ph-Sa]. For the Hecke triangle
groups Γq the recent numerical computations by Winkler [Wi] and
Hejhal [He2] provide some evidence that the even cusp forms prob-
ably do not exist already for q = 7 though the odd ones appear in
abundance.
Chapter 5
The automorphic Green function
5.1. Introduction.
Recall that Gs (z, z 0 ) = Gs (u(z, z 0 )) is the Green function on H (on

a free space, so to speak, see Section 1.9). We now consider a Green
function in the context of Γ\H where Γ is a finite volume group. This
is constructed by the method of images giving
X
(5.1) Gs (z/z 0 ) = Gs (z, γz 0 ) , if z 6≡ z 0 (mod Γ) .
γ∈Γ
Gs (z/z 0 ) is not defined for z ≡ z 0 (mod Γ). We suppose that Re s =

σ > 1 so the series converges absolutely by virtue of Gs (u) u−σ and
(2.41). To simplify notation we have not displayed the dependence of
the automorphic Green function on the group since it is fixed through-
out the analysis. However, in order to avoid confusion, we write a slash
between points of the automorphic Green function in contrast with a
comma in the case of the free space Green function.
Given z ∈ H we have (see (1.47))
m
(5.2) Gs (z/z 0 ) = − log |z − z 0 | + O(1) , as z 0 → z
2π
77
78 The automorphic Green function
where m is the order of the stability group Γz (m = 1 except for the

elliptic fixed points).
The automorphic Green function is an automorphic form in each
variable; it has the properties: Gs (z/z 0 ) = Gs (z 0 /z) = Gs (γz/γ 0 z 0 ) for
γ, γ 0 ∈ Γ, z 6≡ z 0 (mod Γ) and
(5.3) (∆z + s(1 − s))Gs (z/z 0 ) = 0 .
The resolvent operator −Rs restricted to automorphic functions is

given by the kernel −Gs (z/z 0 )
Z
(5.4) −(Rs f )(z) = Gs (z/z 0 ) f (z 0 ) dµz 0 .
F
By Theorem 1.17 Rs is the inverse to ∆ + s(1 − s) on the space B(Γ\H).

In fact one can show by examining the arguments in Section 1.9 that
the equation
(5.5) (∆ + s(1 − s))Rs f = f
is valid in the bigger space Bµ (Γ\H) of smooth automorphic functions

satisfying the following growth condition
(5.6) f (σa z) y(z)µ .
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.
5.2. The Fourier expansion.
Let us begin by expanding a general automorphic kernel

X
K(z, z 0 ) = k(z, γz 0 )
γ∈Γ
where k(z, z 0 ) = k(u(z, z 0 )) is a point-pair invariant. We allow a log-

arithmic singularity of k(u) at u = 0 and assume the series converges
The Fourier expansion 79
absolutely whenever z 6≡ z 0 (mod Γ) as in the case of the Green func-

tion. Suppose a, b are cusps for Γ (not necessarily distinct). We split
the series X
K(σa z, σb z 0 ) = k(z, τ z 0 )
−1
τ ∈σa Γσb
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
where K0 exists only if a = b in which case

X
K0 (z, z 0 ) = k(z, z 0 + n) .
n
By Poisson’s summation we obtain

X
K0 (z, z 0 ) = e(−nx + nx0 ) Pn (y, y 0 ) ,
n
say, where Z ∞
0
Pn (y, y ) = e(ξn) k(iy + ξ, iy 0 ) dξ .
−∞
For c ∈ Cab the computation of Kc (z, z 0 ) is similar. We have

X X
Kc (z, z 0 ) = k(z + n, ωcd (z 0 − m))
d(mod c) m n
+∞
X X ZZ
= e(ξn + ηm) k(z + ξ, ωcd (z 0 − η)) dξ dη
d(mod c) m n −∞
X X a d
= e(n(−x + ) + m(x0 + ))
mn
c c
d(mod c)
+∞
−c−2
ZZ
· e(ξn + ηm) k(iy + ξ, ) dξ dη
iy 0 − η
−∞
X
−2
=c Sab (m, n; c) e(−nx + mx0 ) Pn,mc−2 (y, c2 y 0 )
mn
where Sab (m, n; c) is the Kloosterman sum (see (2.23)) and
+∞
−1
ZZ
0
Pn,m (y, y ) = e(ξn + ηm) k(iy + ξ, ) dξ dη .
iy 0 − η
−∞
It remains to compute the integrals Pn (y, y 0 ) and Pn,m (y, y 0 ) ex-

plicitly in terms of k(u). In full generality this seems to be a hard task;
however, the special case k(u) = Gs (u) is all we need. There is a nat-
ural and elegant approach through the theory of Green function of an
ordinary differential equation. But it is not short; therefore, we now
state and use the results before proving them at the end of this section.
Lemma 5.1. Let k(u) = Gs (u) with Re s > 1. Suppose that y 0 > y.
Then we have
(5.8) P0 (y, y 0 ) = (2s − 1)−1 y s (y 0 )1−s
and
(5.9) Pn (y, y 0 ) = (4π|n|)−1 Vs (iny) Ws (iny 0 ) , n 6= 0
where Ws (z), Vs (z) are defined by (1.26) and (1.36), respectively.
Lemma 5.2. Let k(u) = Gs (u) with Re s > 1. Suppose that y 0 y > 1.
Then we have
π 1/2 Γ(s − 1/2)

(5.10) P0,0 (y, y 0 ) = (yy 0 )1−s ,
2s − 1 Γ(s)
πs y 1−s
(5.11) P0,m (y, y 0 ) = Ws (imy 0 ) , m 6= 0 ,
(2s − 1)Γ(s) |m|
πs y 0 1−s
(5.12) Pn,0 (y, y 0 ) = Ws (iny) , n 6= 0 ,
(2s − 1)Γ(s) |n|
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|)
according to whether mn > 0 or mn < 0.

An estimate for the automorphic Green function 81
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.
Theorem 5.3. Let Re s > 1 and (z, z 0 ) ∈ Dab . We have
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
5.3. An estimate for the automorphic Green function.
We shall use the Fourier expansion (5.15) to estimate Gs (σa z/σb z 0 )

in cuspidal zones. From the first and the fourth lines on the right side
of (5.15) we assemble (2s − 1)−1 (y 0 )1−s Eb (σa z, s) (see Theorem 3.4).
To the second line we apply the asymptotics
Ws (mz 0 ) = exp (2πimx0 − 2π|m|y 0 ) (1 + O(|m|−1 )) ,

Vs (nz) = exp (2πinx + 2π|n|y) (1 + O(|n|−1 ))
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
I2s−1 (y) min{y 2σ−1 , y −1/2 } ey ,

J2s−1 (y) min{y 2σ−1 , y −1/2 } .
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
Gs (σa z/σb z 0 ) = (2s − 1)−1 (y 0 )1−s Eb (σa z, s)

(5.17) 1 0
− log |1 − e(z − z 0 )| + O(e−2π(y −y) ) .
π
Evaluation of some integrals 83
5.4. Evaluation of some integrals.
As promised, we now give proofs of Lemmas 5.1 and 5.2. We first

evaluate Pn (y, y 0 ) by an appeal to the theory of the Green function (see
Appendix A.5), and then we apply the result to Pm,n (y, y 0 ) (different
computations can be found in [Fa]).
The Fourier integral
Z +∞
0
Pn (y, y ) = e(ξn) k(iy + ξ, iy 0 ) dξ
−∞
has the singular kernel k(z, z 0 ) = Gs (u(z, z 0 )) yet Pn (y, y 0 ) is continuous

in R+ × R+ including the diagonal. Clearly Pn (y, y 0 ) is symmetric.
Differentiating in y twice we get
Z +∞
00
P = e(ξn) kyy (iy + ξ, iy 0 ) dξ ,
−∞
and integrating by parts in ξ two times we get

Z +∞
2
(2πin) P = e(ξn) kxx (iy + ξ, iy 0 ) dξ .
−∞
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
−∞
by (∆z + s(1 − s))k(z, z 0 ) = 0 (see (1.68)). This means that Pn (y, y 0 )

as a function in y satisfies the Bessel differential equation
(5.18) P 00 (y) + (s(1 − s)y −2 − 4π 2 n2 ) P (y) = 0 .
Next, applying (1.66) to functions of type f (z) = e(nx) g(y) we

infer that Z +∞
T Pn (y, y 0 ) g(y 0 ) dy 0 = g(y)
−∞
for any g smooth and bounded in R+ , where T is the second order

differential operator associated with the equation (5.18). Therefore,
Pn (y, y 0 ) is a Green function for T .
There are two linearly independent solutions to (5.18), namely

I(y) = Vs (iny) and K(y) = Ws (iny) (see (1.36) and (1.26), respec-
tively) for which the Wronskian is equal to
W = I 0 K − IK 0 = 4 π |n| .
This follows by the asymptotic formulas (see (1.38) and (1.37))
I(y) ∼ e2π|n|y , I 0 (y) ∼ 2π |n| e2π|n|y ,

K(y) ∼ e−2π|n|y , K 0 (y) ∼ −2π |n| e−2π|n|y ,
as y → +∞, and by the fact that the Wronskian is constant. If n = 0

we have a pair I(y) = y s , K(y) = y 1−s for which the Wronskian is
equal to 2s − 1.
Finally, by the theory of the Green function for the equation (5.18)
it follows that Pn (y, y 0 ) is given as in Lemma 5.1.
For the proof of Lemma 5.2 we express Pn,m in terms of Pn as the

integral
+∞
−ηn y0
Z
0
Pn,m (y, y ) = e − ηm P n y, dη .
−∞ η 2 + (y 0 )2 η 2 + (y 0 )2
Since y > y 0 (η 2 + (y 0 )2 )−1 for any η ∈ R, we can apply Lemma 5.1 in

the whole range of the above integration.
First consider n = 0. In this case
Z +∞
0 −1 1−s 0 s
P0,m (y, y ) = (2s − 1) y (y ) e(−ηm) (η 2 + (y 0 )2 )−s dη ;
−∞
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
Hence, we shall get (5.12) and (5.13) by the following

Evaluation of some integrals 85
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
(5.20) 4π(2s − 1)−1 Γ(s)−1 (π|a|)s if b = 0 ,

1/2
√
(5.21) 4π|a| Ks−1/2 (2π|b|) J2s−1 (4π ab) if ab > 0 ,
p
(5.22) 4π|a|1/2 Ks−1/2 (2π|b|) I2s−1 (4π |ab|) if ab < 0 .
Proof. We appeal to the equation
(∆ + s(1 − s))V (z) = 0
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
a2 yη2 (Vxx + Vyy ) + s(1 − s)V = 0 .
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
(5.23) a2 v 00 (a) + (s(1 − s) + 4π 2 ab) v(a) = 0 .

If b = 0 the solutions are v(a) = α |a|s + β |a|1−s where α, β are

constants. We shall determine these constants from the asymptotic
formula
(5.24) V (z) ∼ 2π s+1/2 Γ(s + 1/2)−1 |y|s , as y → 0
which yields
Z +∞
s+1/2 −1 s
(5.25) v(a) ∼ 2π Γ(s + 1/2) |a| (1 + η 2 )−s dη ,
−∞
as a → 0. Hence, by (3.18) and the duplication formula for the gamma

function (see (B.6) in the Appendix B) we get
4π(π|a|)s
v(a) ∼ ,
(2s − 1)Γ(s)
and so one determines that β = 0 and then that v(a) is exactly equal
to (5.20).
If b 6= 0 the solutions to (5.23) are given by Bessel’s functions
√ √
v(a) = α|a|1/2 J2s−1 (4π ab) + β|a|1/2 Y2s−1 (4π ab) ,
if ab > 0, and
p p
v(a) = α|a|1/2 I2s−1 (4π |ab|) + β|a|1/2 K2s−1 (4π |ab|) ,
if ab < 0. From the power series expansion for Bessel’s functions it

follows that
√ √
J2s−1 (4π ab) ∼ Γ(2s)−1 (2π ab)2s−1 ,
p p
I2s−1 (4π |ab|) ∼ Γ(2s)−1 (2π |ab|)2s−1 ,
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|) .
From this, one determines the constants β = 0 and α = 4π Ks−1/2 (4π|b|)

so v(a) is given exactly by (5.21) or (5.22) according to the sign of ab.
This completes the proof of Lemma 5.5 and that of Lemma 5.2.
Chapter 6
Analytic continuation
of the Eisenstein series
The analytic continuation of the Eisenstein series Ea (z, s) is fun-

damental for the spectral resolution of ∆ in the space E(Γ\H) of incom-
plete Eisenstein series. There are many ways to perform the analytic
continuation (Selberg, Faddeev, Colin de Verdière, Langlands, Bern-
stein, . . . ). We shall present one of Selberg’s methods which uses the
Fredholm theory of integral equations (see [Se2]).
6.1. The Fredholm equation for the Eisenstein series.
To get started we fix a number a > 2 and consider the resolvent

Ra . We have
Z
−1
(6.1) −(∆ + a(1 − a)) f (z) = Ga (z/z 0 ) f (z 0 ) dµz 0 ,
F
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
Ea (σb z, s) = δab y s + ϕab (s) y 1−s + Ea∗ (σb z, s)
87
88 Analytic continuation of the Eisenstein series
it is apparent that Ea (z, s) belongs to Bσ (Γ\H) with σ = Re s. Suppose

that 1 < σ ≤ a − 1 so (6.1) applies to

f (z) = (∆ + a(1 − a))Ea (z, s) = a(1 − a) − s(1 − s) Ea (z, s)
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
Gab (z/z 0 ) = Ga (z/z 0 ) − Gb (z/z 0 )
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,
(a − s)(a + s − 1)(b − s)(b + s − 1)

(6.4) λab (s) = .
(b − a)(a + b − 1)
The new kernel Gab (z/z 0 ) is continuous in (z, z 0 ) ∈ H × H since the

leading term in the asymptotic (5.2) for Gs (z/z 0 ) does not depend on
s, so it cancels.
The second obstacle is that Gab (z/z 0 ) is not bounded. We handle
this problem in the z 0 variable by substracting suitable contributions
when z 0 is in cuspidal zones. To this end we fix a fundamental polygon
F having inequivalent cuspidal vertices and partition it into
[
F = F (Y ) ∪ Fb (Y )
b
The Fredholm equation for the Eisenstein series 89
where Y is a large parameter, Fb (Y ) are the cuspidal zones of height Y

and F (Y ) is the central part (see (2.1)-(2.5)). We define the truncated
kernel on H × F (not on H × H) by setting GYab (z/z 0 ) = Gab (z/z 0 ), if
z 0 ∈ F (Y ) and
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) ,
if z 0 ∈ Fb (Y ). Note that GYab (z/z 0 ) is automorphic in z but not in z 0

(the second variable is confined to the fixed fundamental domain; its
range could be extended over all H by Γ-periodicity, but there is no
reason to do so). In the z 0 variable in F the truncated kernel GYab (z/z 0 )
is continuous except for jumps on the horocycles Lb (Y ). As z 0 ap-
proaches a cusp, GYab (z/z 0 ) decays exponentially by Lemma 5.4, but in
the z variable the kernel GYab (z/z 0 ) is not bounded; it has polynomial
growth at cusps which is inherited from the Eisenstein series Eb (z, a)
and Eb (z, b). More precisely, we infer from the approximation (5.17)
the following bound
0
(6.5) GYab (σa z/σb z 0 ) y a e−2π max{y −y,0} , if y, y 0 > Y .
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)
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
where λ = λab (s) is given by (6.4),
h(z) = (22s−1 − 1)−1 (2s−1+a − 1) (2s−1+b − 1) νab (s) Ea (z, s) ,
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)
H(z, z 0 ) = (2s−1+a − 1)−1 (2s−1+b − 1)−1

· GYab − 2s−1 (2a + 2b )G2Y 2s−2+a+b 4Y
Gab (z/z 0 ) .

ab + 2
For notational simplicity we did not display the dependence of h(z),

f (z), H(z, z 0 ) on the complex parameter s nor on the fixed numbers a,
b.
6.2. The analytic continuation of Ea (z, s).
We are almost ready to apply the Fredholm theory to the equa-

tion (6.7). There are still minor problems with the growth of f (z)
and H(z, z 0 ) in the z variable. These functions are not bounded (as
required in our version of the Fredholm theory), but they have polyno-
mial growth at the cusps. More precisely, given a > b > c + 1 we have,
uniformly, for s in the strip 1 − c ≤ Re s ≤ c that
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
(6.9) Rλ (z, z 0 ) = D(λ)−1 Dλ (z, z 0 )
where D(λ) 6≡ 0 and Dλ (z, z 0 ) are holomorphic in λ of order ≤ 2;

therefore of order ≤ 8 in s in the strip 1 − c ≤ Re s ≤ c.
For any λ with D(λ) 6= 0 we have a unique solution to (6.8) given
by Z
η(z) h(z) = η(z) f (z) + λ Rλ (z, z 0 ) η(z 0 ) f (z 0 ) dµz 0 ;
F
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
The function Dλ (z, z 0 ) has a power series expansion in λ whose coef-

ficients are bounded in F × F . Therefore, (6.10) yields the analytic
continuation of Ea (z, s). Putting
Aa (s) = (2s+a−1 − 1) (2s+b−1 − 1) D(λ)
where λ = λab (s) is given by (6.4) and
Aa (z, s) = (22s−1 − 1) λ D(λ) h(z)
where h(z) is given by (6.10), we conclude


Proposition 6.1. Given c > 1 denote S = s ∈ C : 1 − c ≤ Re s ≤ c .
There are functions Aa (s) 6≡ 0 on S and Aa (z, s) on H × S with the
following properties:
(6.11) Aa (s) is holomorphic in s of order ≤ 8 ,

(6.12) Aa (z, s) is holomorphic in s of order ≤ 8 ,
(6.13) Aa (z, s) is real-analytic in (z, s) ,
(6.14) Aa (z, s) ∈ As (Γ\H) ,
(6.15) Aa (z, s) = Aa (s) Ea (z, s) if 1 < Re s ≤ c ,
(6.16) Aa (z, s) eε y(z) ,
the implied constant depending on ε, s and Γ.
From (6.15) we draw the analytic continuation of Ea (z, s) to the

strip S, and since c is arbitrary we get the meromorphic continuation
to the whole s-plane. Furthermore, by (6.16) we retain certain control
of growth, namely
Corollary 6.2. Suppose s is not a zero of Aa (s). For any ε > 0 we

have
(6.17) Ea (z, s) eε y(z) ,
the implied constant depending on ε, s and Γ.
This is not a very strong bound, nevertheless it helps us to proceed

further. From (6.17) we infer the validity of the Fourier expansion
X
(6.18) Ea (σb z, s) = δab y s + ϕab (s) y 1−s + ϕab (n, s) Ws (nz) ,
n6=0
for all s with Aa (s) 6= 0. We also obtain the meromorphic continuation

of the coefficientes ϕab (s), ϕab (n, s). After this is known, the estimate
(6.17) improves itself via the Fourier expansion (6.18). Manipulating
skillfully with the exponentical decay of the Whittaker function one
shows that
(6.19) ϕab (n, s) |n|σ + |n|1−σ
and
(6.20) Ea (σb z, s) = δab y s + ϕab (s) y 1−s + O(e−2πy )

The functional equations 93
as y → +∞, for any s with Aa (s) 6= 0. However, the implied constants

in (6.19) and (6.20) may depend badly on s.
We conclude with the following obvious, yet basic observation
Theorem 6.3. The meromorphically continued Eisenstein series are

orthogonal to cusp forms.
Proof. The inner product of a cusp form against an Eisenstein series

exists because the former has exponential decay at cusps and the latter
has at most polynomial growth. For Re s > 1 the orthogonality follows
by the unfolding method, and for the regular points with Re s < 1 it
follows by analytic continuation (the inner product converges absolutely
and uniformly on compacta in s).
6.3. The functional equations.
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
Lemma 6.4. Suppose f ∈ As (Γ\H) satisfies the growth condition
(6.21) f (z) eε y(z) ,
with 0 < ε < 2π. If Re s > 1, then f (z) is a linear combination of the
Eisenstein series Ea (z, s).
Proof. Since f ∈ As (Γ\H), it has the Fourier expansion
f (σa z) = αa y s + βa y 1−s + O(1)
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
in H. Hence g ∈ As (Γ\H) ∩ L(Γ\H) which implies g = 0 because ∆ has

only non-negative eigenvalues in L(Γ\H). Therefore,
X
f (z) = αa Ea (z, s) .
a
Let E(z, s) denote the column vector of the Eisenstein series

Ea (z, s) where a ranges over all inequivalent cusps. Recall the Fourier
expansions (6.18). The first coefficients of the zero-th term form the
identity matrix
I = δab ,
and the second coefficients form the scattering matrix

Φ(s) = ϕab (s) .
Theorem 6.5. The column-vector Eisenstein series satisfies the func-

tional equation
(6.22) E(z, s) = Φ(s) E(z, 1 − s) .
Proof. Suppose Re s > 1 and Aa (1 − s) 6= 0 so the Eisenstein series

Ea (z, 1 − s) is defined by meromorphic continuation
Ea (z, 1 − s) ∈ A1−s (Γ\H) = As (Γ\H) .
Moreover, Ea (z, 1 − s) satisfies the growth condition (6.21) by virtue of
Corollary 6.2; therefore, by Lemma 6.4 it follows that
X
Ea (z, 1 − s) = ϕab (1 − s) Eb (z, s) .
b
This relation extends to all s ∈ C by analytic continuation so changing
s into 1 − s we get (6.22).
From (6.22) by the symmetry of Φ(s) one gets another functional
equation
t
(6.22’) E(z, s) E(w, 1 − s) = t E(z, 1 − s) E(w, s) .
Theorem 6.6. The scattering matrix satisfies the functional equation

(6.23) Φ(s) Φ(1 − s) = I .
For s with Re s = 1/2 the scattering matrix is unitary,
(6.24) Φ(s) t Φ(s) = I .
For s real the scattering matris is hermitian.
Poles and residues of the Eisenstein series 95
Proof. The functional equation (6.23) follows by (6.22). Next, we see

the symmetry
(6.25) ϕab (s) = ϕba (s)
from the Dirichlet series representation (3.21) if Re s > 1, and it extends

to all s by analytic continuation. In matrix notation (6.25) takes the
form
(6.26) Φ(s) = t Φ(s) .
We also read from the Dirichlet series representation that
(6.27) Φ(s) = Φ(s) ,
for all s by analytic continuation. Since s = 1−s on the line Re s = 1/2,

it follows by combining (6.26) with (6.27) that Φ(1 − s) = t Φ(s). This
and the functional equation (6.23) yield (6.24). Finally, it follows from
(6.26) and (6.27) that Φ(s) is hermitian for s real.
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
(6.28) kΦa (s)k2 = 1 , if Re s = 1/2 .
6.4. Poles and residues of the Eisenstein series.
We shall infer some information about poles of Φ(s) and E(z, s) in

Re s > 1/2 from a certain formula of Maass and Selberg for the inner
product of truncated Eisenstein series (the whole series Ea (z, s) is not
in L(Γ\H) because of polynomial growth at cusps). We set
(6.29) EaY (z, s) = Ea (z, s) − δab (Im σb−1 z)s − ϕab (s) (Im σb−1 z)1−s
if z is in the zone Fb (Y ) and we substract nothing if z is in the central

part F (Y ). The truncated Eisenstein series satisfies the bound
EaY (z, s) e−2π y(z) ,
for z ∈ F , the implied constant depending on s and Y (see (6.20)).
Proposition 6.8 (Maass-Selberg). If s1 , s2 are regular points of the

Eisenstein series Ea (z, s) and Eb (z, s), respectively, and s1 6= s2 , s1 +
s2 6= 1, then
hEaY (·, s1 ), EbY (·, s2 )i = (s1 − s2 )−1 ϕab (s2 ) Y s1 −s2

+ (s2 − s1 )−1 ϕab (s1 ) Y s2 −s1
(6.30) + (s1 + s2 − 1)−1 δab Y s1 +s2 −1
− (s1 + s2 − 1)−1 Φa (s1 ) Φb (s2 ) Y 1−s1 −s2
where in the last term we have the scalar product of two row vectors of
the scattering matrix.
This relation is derived by application of Green’s formula to the

central part F (Y ) of a fundamental polygon. The resulting boundary
terms on equivalent sides segments cancel out, and the remaining inte-
grals along horocycles of height Y for each cusp are computed using the
Fourier expansions (6.18). Only the zero-th terms survive the integra-
tion and they make up the right-hand side of (6.30). A similar relation
holds true for general Maass forms (see the end of this section for more
details).
We shall need (6.30) for a = b and s1 = σ + iv, s2 = σ − iv with

σ > 1/2 and v =6 0. In this case we obtain
X
kEaY (·, σ + iv)k2 + (2σ − 1)−1 Y 1−2σ |ϕab (σ + iv)|2
(6.31) b
−1 2σ−1
= (2σ − 1) Y − v Im ϕaa (σ + iv) Y −2iv
−1
provided s = σ + iv is a regular point of Φa (s). Hence, by examining

the growth of individual terms and using the positivity of the left side
one derives immediately
Theorem 6.9. The functions ϕab (s) are holomorphic in Re s ≥ 1/2

except for a finite number of simple poles in the segment (1/2, 1]. If
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.
Now we are ready to examine poles and residues of the Eisenstein

series Ea (z, s) in Re s > 1/2. Suppose sj is a pole of order m ≥ 1. Then
the function
u(z) = lim (s − sj )m Ea (z, s)
s→sj
does not vanish identically, and it belongs to Asj (Γ\H). Moreover, it

has the Fourier expansion at any cusp of type
X
u(σb z) = ρb y 1−sj + ρb (n) Wsj (nz)
n6=0
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.
Next we determine what happens on the line Re s = 1/2. We let

σ → 1/2 in (6.31) showing that
(6.32) kEaY (·, σ + iv)k 1 ,
for any fixed v ∈ R, including v = 0 because Φ(s) is holomorphic and

unitary on Re s = 1/2 (use (6.28)) and real on R.
Theorem 6.11. The Eisenstein series Ea (z, s) has no poles on the line
Re s = 1/2.
Proof. Suppose s0 = 1/2 + iv is a pole of Ea (z, s) of order m ≥ 1,

say. Since Φ(s) is regular at s = s0 , we have
u(z) = lim (s − s0 )m Ea (z, s) = lim (s − s0 )m EaY (z, s) .

s→s0 s→s0
Hence, it follows by (6.32) that kuk = 0 so u(z) ≡ 0 because u(z) is

continuous (in fact real-analytic).
Proposition 6.12. For s 6= 1/2 the Eisenstein series Ea (z, s) does not
vanish identically.
Proof. The zero-th term of Ea (σa z, s) is equal to y s +ϕaa (s) y 1−s 6≡ 0.
Show that the Eisenstein series Ea (z, 1/2) vanishes identically if

and only if ϕaa (1/2) = −1.
Proposition 6.13. The point s = 1 is a pole of Ea (z, s) with residue
(6.33) res Ea (z, s) = |F |−1 .

s=1
Proof. Suppose ϕaa (s) is regular at s = σ > 1/2. Letting v → 0 in

(6.31) we get
X
kEaY (σ)k2 + (2σ − 1)−1 Y 1−2σ |ϕab (σ)|2
b
−1 2σ−1
= (2σ − 1) Y + 2 ϕaa (σ) log Y − ϕ0aa (σ) .
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
Comparing both limits we infer that α2 |F | = α + O(Y −1 ). Letting

Y → +∞ we obtain α = |F |−1 as claimed.
Now we provide a proof of the Maass-Selberg relations for arbitrary

Maass forms which do not grow exponentially at the cusps. Such forms
have the Fourier expansion (3.4). As with the Eisenstein series we
truncate f (z) by substracting the zero-th terms in cuspidal zones, i.e.
we put
f (z) − fa (Im σa−1 z) if z ∈ Fa (Y ) ,

Y
f (z) =
f (z) if z ∈ F (Y ) .
By (3.7) the truncated form has exponential decay at cusps, i.e.
f Y (z) e−2π y(z) , for z ∈ F .
Theorem 6.14. Let f ∈ As1 (Γ\H) and g ∈ As2 (Γ\H). Suppose f, g

satisfy (3.3) at any cusp. Then, for Y sufficiently large we have
X
(λ1 − λ2 ) hf Y , g Y i = fa (Y ) ga0 (Y ) − fa0 (Y ) ga (Y )

(6.34)
a
where λ1 = s1 (1 − s1 ) and λ2 = s2 (1 − s2 ).
Proof. We begin by applying Green’s formula (the hyperbolic version)

Z Z
(λ1 − λ2 ) f g dµ = (f ∆g − g ∆f ) dµ
F (Y ) F (Y )
Z ∂g ∂f
= f −g d` .
∂F (Y ) ∂n ∂n
The boundary ∂F (Y ) consists of segments of sides of F and the horo-
cycles σa L(Y ), where L(Y ) = {z = x + iY : 0 < x < 1} (the beginning
of cuspidal zones, see (2.4)). Since the integrals along the segments of
equivalent sides cancel out, we are left with
Z
(λ1 − λ2 ) f g dµ
F (Y )
XZ ∂ ∂
= f (σa z) g(σa z) − g(σa z) f (σa z) dx
a F (Y ) ∂y ∂y
after the change of variable z 7→ σa z. Next, by the Fourier expansion

X
f (σa z) = fn (y) e(nx) with f0 (y) = fa (y) ,
n
X
g(σa z) = gn (y) e(nx) with g0 (y) = ga (y) ,
n
we get Z X
0
(Y ) − fn0 (Y ) g−n (Y )

= fn (Y ) g−n
L(Y ) n
after integration in 0 < x < 1. Furthermore, by the Whittaker differ-

ential equation (1.25) for the Fourier coefficients,
d 0
(y) − fn0 (y) g−n (y) = (λ1 − λ2 ) y −2 fn (y) g−n (y) .

fn (y) g−n
dy
If n 6= 0 this has an exponential decay as y → +∞ so we can integrate

in y > Y getting
Z +∞
0
fn (Y ) g−n (Y ) − fn0 (Y ) g−n (Y ) = −(λ1 − λ2 ) fn (y) g−n (y) y −2 dy .
Y
Summing over n 6= 0 we infer that

Z
= fa (Y ) ga0 (Y ) − fa0 (Y ) ga (Y )
L(Y )
Z +∞ Z 1
− (λ1 − λ2 ) f Y (σa z) g Y (σa z) dµz .
Y 0
Here the last integral is equal to (after the change z 7→ σa−1 z)

Z
f Y (z) g Y (z) dµz .
Fa (Y )
Finally, summing over the cusps we arrive at (6.34) by collecting these

integrals together with the one we began with.
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
upon dividing by λ1 − λ2 = (s1 − s2 )(1 − s1 − s2 ) which requires the

condition λ1 6= λ2 . In particular, for the Eisenstein series this reduces
to (6.30).
Now that we know that the Eisenstein series are holomorphic on

the critical line Re s = 1/2 we can extend (6.30) by examining carefully
what happens at s1 = s2 = σ + iv as σ → 1/2. In matrix notation all
the relations (6.30) read simultaneously as the following one
hE Y (·, s), t E Y (·, s)i = (2iv)−1 Φ(s) Y 2iv − Φ(s) Y −2iv

+ (2σ − 1)−1 Y 2σ−1 − Φ(s) Φ(s)Y 1−2σ

where s = σ + iv, v 6= 0. Here we apply the following approximations
Y 2σ−1 = 1 + (2σ − 1) log Y + · · · ,

Y 1−2σ = 1 − (2σ − 1) log Y + · · · ,
Φ(σ + iv) = Φ(s) + (σ − 1/2) Φ0 (s) + · · · ,
Φ(σ + iv) Φ(σ − iv) = 1 + (2σ − 1) Φ0 (s) Φ(s)−1 + · · · ,
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

(6.35) + 2 log Y − Φ0 (s) Φ(s)−1 ,
for Re s = 1/2, s 6= 1/2. Furthermore, at the center of the critical strip
(6.36) hE Y (·, 1/2), t E Y (·, 1/2)i = (2 log Y − Φ0 (1/2)) (1 + Φ(1/2)) .

Chapter 7
The spectral theorem.

Continuous part
To complete the decomposition of the space L(Γ\H) into ∆-inva-

riant subspaces it remains to do it in the subspace E(Γ\H) spanned by
the incomplete Eisenstein series Ea (z|ψ) (the orthogonal complement
C(Γ\H) was already shown to be spanned by Maass cusp forms, see
Chapter 4). The spectral expansion for the incomplete Eisenstein series
X
(7.1) Ea (z|ψ) = ψ(Im σa−1 γz) ,
γ∈Γa \Γ
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)
after moving the integration from Re s = σ > 1 to the line σ = 1/2

followed by an application of the functional equation for the Eisenstein
series. Recall the bound (which is uniform in vertical strips)
Z +∞
(7.3) ψ̂(s) = ψ(y) y −s−1 dy (|s| + 1)−A .
0
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.
7.1. The Eisenstein transform.
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
To a cusp a we associate the Eisenstein transform
Ea : C0∞ (R+ ) −→ A(Γ\H)
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
(7.6) (Ea f )(σb z) y 1/2 (log y)−1 , as y → +∞
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 : C0∞ (R+ ) −→ L(Γ\H) .
Proposition 7.1. For f, g ∈ C0∞ (R) we have
(7.7) hEa f, Eb gi = δab hf, gi .

The Eisenstein transform 105
Proof. For the proof we consider the truncated Eisenstein transform

Z +∞
Y 1
(Ea f )(z) = f (r) EaY (z, 1/2 + ir) dr
4π 0
where EaY (z, s) is the truncated Eisenstein series (see (6.29)). We get
the approximation
y(z)1/2
(EaY f )(z) = (Ea f )(z) + O
log y(z)
on integrating by parts in r, wherein the equality holds if z ∈ F (Y ).
Hence, we infer that
k(Ea − EaY )f k (log Y )−1/2 ,
and by the Cauchy-Schwarz inequality this gives the approximation
hEa f, Eb gi = hEaY f, EbY gi + O (log Y )−1/2 .

Next we compute the inner product

hEaY f, EbY gi
Z +∞Z +∞
1
= 2
f (r0 ) g(r) hEaY (·, 1/2 + ir 0 ), EbY (·1/2 + ir)i dr dr 0
(4π) 0 0
by an appeal to the Maass-Selberg formula (see Proposition 6.8)

hEaY (·,1/2 + ir 0 ), EbY (·, 1/2 + ir)i
i 0
= 0 ϕab (1/2 + ir) Y i(r +r)
r +r
i 0
− 0 ϕab (1/2 + ir 0 ) Y −i(r +r)
r +r
i 0
i(r−r0 )
+ δ ab − Φ a (1/2 + ir ) Φ b (1/2 + ir) Y
r − r0
i 0 0
+ 0
δab Y i(r −r) − Y i(r−r ) .
r−r
Since all terms are continuous in (r, r 0 ) ∈ R+ × R+ (recall that Φ(s) is
unitary on Re s = 1/2) we gain the factor log Y by partial integration
in r for all but the last term. We obtain
hEaY f, EbY gi
Z +∞Z +∞ 0 0
δab 0 Y i(r −r) −Y i(r−r ) 0 −1

= f (r ) g(r) dr dr + O (log Y ) .
(4π)2 0 0 i(r0 − r)
Since r is bounded below by a positive constant, the innermost integral

in r 0 approximates to
+∞
du
Z
f (r) 2 sin(u log Y ) = 2π f (r)
−∞ u
up to an error term O((log Y )−1 ), which is estimated by partial in-

tegration. Collecting the above results and letting Y → +∞ we get
(7.7).
Corollary 7.2. The Eisenstein transform Ea maps isometrically

C0∞ (R+ ) into L(Γ\H).
Remark. One can extend the Eisenstein transform Ea to an isome-

try of L2 (R+ ) into L(Γ\H), of course not surjectively. This is a close
analogue of the Plancherel theorem for the Fourier transform.
The image Ea (Γ\H) of the Eisenstein transform Ea is called the

space of the Eisenstein series Ea (z, s). Clearly Ea (Γ\H) is an invariant
subspace for the Laplace operator; more precisely, ∆ acts on Ea (Γ\H)
through multiplication, 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
R(Γ\H) ⊕ Ea (Γ\H) ⊂ E(Γ\H)

a
where R(Γ\H) is the space spanned by residues of all the Eisenstein

series in the segment (1/2, 1]. The spectral theorem will show that
R(Γ\H) together with Ea (Γ\H) fill densely the space E(Γ\H).
Bessel’s inequality 107
7.2. Bessel’s inequality.
Suppose fj are mutually orthogonal in a Hilbert space. Then for

any f in that space we have
X X X
kf − fj k2 = kf k2 − 2 Re hf, fj i + kfj k2 .
j j j
In particular, if fj are chosen so that hf, fj i = kfj k2 this gives

X X
kf − fj k2 = kf k2 − kfj k2 ,
j j
and hence, by the positivity of the norm k · k we obtain the Bessel

inequality
X
(7.8) kfj k2 ≤ kf k2 .
j
(to be precise use the above relations for a finite collection of functions
fj and then drop this condition in (7.8) by positivity).
We employ Bessel’s inequality in the space L(Γ\H) for an auto-

morphic kernel X
f (z) = K(z, w) = k(z, γw)
γ∈Γ
where z is the variable and w is fixed. Suppose k(u) is smooth and

compactly supported on R+ ; then f (z) is also compactly supported on
H so it belongs to L(Γ\H). Anticipating the spectral expansion for
K(z, w) (see Theorem 8.1) we choose the functions
fj (z) = h(tj ) uj (z) uj (w) ,

Z B
1
fa (z) = h(r) Ea (z, 1/2 + ir) Ea (w, 1/2 + ir) dr
2π A
to approximate f (z) so that the inequality (7.8) is quite sharp. Here

uj (z) range over an orthogonal system in the space of discrete spec-
trum (Maass cusp forms and the residues of the Eisenstein series in
the segment (1/2, 1]), Ea (z, s) are the Eisenstein series and h(r) is the
Selberg/Harish-Chandra transform of k(u). The integral is cut off at
fixed heights A, B with 0 < A < B < +∞ because we do not know

yet if the full integral converges. Observe that fa (z) is the Eisenstein
transform
fa (z) = (Ea g)(z)
for g given by
g(r) = 2 h(r) E a (w, 1/2 + ir) ,
if A ≤ r ≤ B and g(r) = 0 elsewhere. This is not smooth at the
end points; nevertheless Proposition 7.1 remains valid by a suitable
approximation or by essentially the same proof.
From the discussion concluding the previous section we have learn-

ed that all fj (z), fa (z) are mutually orthogonal. To compute the pro-
jections of fj on the kernel f (z) = K(z, w) we appeal to Theorem 1.16.
We get
hf, fj i = h(tj ) uj (w) hf, uj i

Z
= h(tj ) uj (w) k(z, w) uj (z) dµz
H
= |h(tj ) uj (w)| = kfj k2 .
2
Similarly, by Proposition 7.1, we get

B
1
Z
hf, fa i = |h(r) Ea (w, 1/2 + ir)|2 dr
2π A
B
1
Z
= |g(r)|2 dr = kfa k2 .
2π A
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
Here we have dropped the restriction A ≤ r ≤ B by positivity and we

have added integrals over negative r by symmetry.
The upshot will come out of (7.9) for the kernel k(u) which is the
characteristic function of the segment 0 ≤ u ≤ δ (this is an admissible
Bessel’s inequality 109
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 .
Next we estimate the L2 -norm of K(z, w). We begin by

Z X Z
2
|K(z, w)| dµz = k(γ 0 z, w) k(γ 0 z, γw) dµz
F γ,γ 0 ∈Γ F
XZ
= k(z, w) k(z, γw) dµz .
γ∈Γ H
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
where 0 restricts the summation and the integration to points sj =

1/2 + itj and s =√1/2 + ir in the disc |s| ≤ (8δ)−1/2 . By Corollary 2.12
we get Nδ (z) δ y(z) + 1. Choosing δ = (4T )−2 we obtain
Proposition 7.2. Let T ≥ 1 and z ∈ H. We have
X XZ T
2
(7.10) |uj (z)| + |Ea (z, 1/2 + it)|2 dt T 2 + T y(z)
|tj |<T a −T
where the implied constant depends on the group Γ alone.
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)
To this end integrate (7.10) over the central part F (Y ) ⊂ F with Y T

and use Theorem 3.2 to extend the integral of |uj (z)|2 over the whole
fundamental domain at the cost of a small error term. Ignoring the
integrals of |Ea (z, 1/2 + it)|2 one obtains (7.11). For the continuous
spectrum analogue of this result see (10.13).
7.3. Spectral decomposition of E (Γ\H).
Since E(Γ\H) is spanned by the incomplete Eisenstein series

Ea (z|ψ), it suffices to decompose Ea (z|ψ) for any ψ ∈ C0∞ (R+ ). The
Mellin transform of ψ is entire, and it satisfies the bound
ψ̂(s) (|s| + 1)−A
in any vertical strip, where A is an arbitrary positive number. Thus

the integral (7.2) converges absolutely if σ > 1. Moving to the line
Re s = 1/2 (which is justified because Ea (z, s) has at most polyno-
mial growth in s on average due to Proposition 7.2 and the Lindelöf-
Phragmen convexity principle) we pass a finite number of simple poles
in the segment (1/2, 1] and get
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
where uaj (z) is the residue of Ea (z, s) at s = sj . Here the coefficients

ψ̂(sj ) are given by the inner product
(7.13) ψ̂(sj ) = hEa (·|ψ), uaj i kuaj k−2
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
(see (6.23)). We also use the formula (see Lemma 3.2)

Z +∞
δab y 1−s + ϕab (1 − s) y s ψ(y) y −2 dy .

hEa (·|ψ), Eb (·, s)i =
0
Multiplying this by Eb (z, s) and summing over b ,

X
hEa (·|ψ), Eb (·, s)i Eb (z, s) = ψ̂(s) Ea (z, s) + ψ̂(1 − s) Ea (z, 1 − s)
b
by the functional equation. Finally integrating this in s on the line

Re s = 1/2 we obtain
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).
The expansions (7.12-7.14) extend to all functions f ∈ E(Γ\H) by

linearity. Some of the residues uaj (z) of the Eisenstein series Ea (z, s)
associated with distinct cusps can be linearly dependent, for instance
the residues at s = 1 since they are all equal to a constant. Moreover,
the residues need be normalized so as to give the L2 -norm equal to one.
We let Rsj (Γ\H) be the space spanned by the residues of all Eisenstein
series at s = sj ; thus
dim Rsj (Γ\H) ≤ h = the number of cusps.
Then we let R(Γ\H) be the space spanned by all residues of all Eisen-
stein series in the segment (1/2, 1], which have the orthogonal decom-
position
R(Γ\H) = ⊕ Rsj (Γ\H) .
1/2<sj ≤1
In each space Rsj (Γ\H) we choose an orthonormal basis out of which

we assemble the basis {uj (z)} of R(Γ\H).
With regard to the integrals of Eisenstein series on the line s =
1/2 + ir neither further rearrangement nor any normalization is de-
sired. The collection of these Eisenstein series is called the eigenpacket.
From the above considerations we conclude the following spectral de-
composition of E(Γ\H).
Theorem 7.3. The space E(Γ\H) of incomplete Eisenstein series splits

orthogonally into ∆-invariant subspaces,
E(Γ\H) = R(Γ\H) ⊕ Ea (Γ\H) .
a
The spectrum of ∆ in R(Γ\H) is discrete; it consists of a finite number

of points λj with 0 ≤ λj < 1/4. The spectrum of ∆ in Ea (Γ\H) is
absolutely continuous; it covers the segment [1/4, ∞) uniformly with
multiplicity 1. Every f ∈ E(Γ\H) has the expansion
X
f (z) = hf, uj i uj (z)
j
(7.15) X 1 Z +∞
+ hf, Ea (·, 1/2 + ir)i Ea (z, 1/2 + ir) dr
a
4π −∞
which converges in the norm topology, and if f belongs to the initial
domain

D(Γ\H) = f ∈ A(Γ\H) : f, ∆f smooth and bounded ,
it converges point-wise absolutely and uniformly on compacta.
Spectral decomposition of automorphic kernels 113
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).
7.4. Spectral expansion of automorphic kernels.
The spectral theorem is a powerful tool for analytic studies in au-

tomorphic forms. Particularly handy is the spectral series expansion
for the automorphic kernel
X
K(z, w) = k(z, γw)
γ∈Γ
as well as for the automorphic Green function

X
(7.16) Gs (z/w) = Gs (z, γw) .
γ∈Γ
First suppose k(u) ∈ C0∞ (R+ ) so as a function of z for w on com-

pacta this K(z, w) has the absolutely and uniformly convergent spectral
expansion given by (4.15) and (7.15). The projections of K(z, w) on
the eigenfunctions are computed in Theorem 1.16; they are
hK(·, w), uj i = h(tj ) uj (w) ,

hK(·, w), Eb (·, 1/2 + ir)i = h(r) E b (w, 1/2 + ir) .
Hence, we obtain
Theorem 7.4. Let K(z, w) be an automorphic kernel given by a

point-pair invariant k(z, w) = k(u(z, w)) whose Selberg/Harish-Chandra
transform h(r) satisfies the conditions (1.63). Then it has the spectral
expansion
X
K(z, w) = h(tj ) uj (z) uj (w)
j
(7.17)
X 1 Z +∞
+ h(r) Ea (z, 1/2 + ir) E a (w, 1/2 + ir) dr
a
4π −∞
which converges absolutely and uniformly on compacta.

Remark. Our initial assumption that k(u) is compactly supported was

replaced by the weaker conditions (1.63) using a suitable approximation.
Next we develop an expansion for the automorphic Green function.

Formally speaking it is a special case of an automorphic kernel for
k(u) = Gs (u), but the above result does not apply directly because
Gs (u) is singular at u = 0. We annihilate the singularity by considering
the difference k(u) = Gs (u) − Ga (u) with Re s > 1 and a > 1. As
a function of w the difference Gs (z/w) − Ga (z/w) has the spectral
expansion
X
Gs (z/w) − Ga (z/w) = gj (z) uj (w) + Eisenstein part.
j
To compute the coefficients gj (z) we save work by appealing to prop-

erties of the resolvent. We get
Z

gj (z) = Gs (z/w) − Ga (z/w) uj (w) dµw
F
= (∆ + s(1 − s))−1 uj (z) − ((∆ + a(1 − a))−1 uj (z)
= (s − sj )−1 (1 − s − sj )−1 − (a − sj )−1 (1 − a − sj )−1 uj (z)

= χsa (sj ) uj (z) ,
say. Furthermore we perform the same computations for projections on

the Eisenstein series. We obtain
Theorem 7.5. Let a > 1 and Re s > 1. Then

X
Gs (z/w) − Ga (z/w) = χsa (sj ) uj (z) uj (w)
j
X 1 Z
(7.18) + χsa (v) Ea (z, v) E a (w, v) dv
a
4πi
(1/2)
where the series and integrals converge absolutely and uniformly on

compacta.
Spectral decomposition of automorphic kernels 115
Remark. The spectral theory for the resolvent operator is treated in

greater generality by J. Elstrodt [El].
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
I1/2 (s) = Iα (s) − (2s − 1)−1 Ea (z, s) Ea (w, 1 − s)
for s in the strip. This furnishes the spectral expansion of Gs (z/w) −

Ga (z/w) in Re s < α through the analytic continuation of Iα (s). More-
over, it shows the following functional equation
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’).
From the spectral expansion (7.18) it is plain that the point

s = sj > 1/2 from the discrete spectrum gives a simple pole of
Gs (z/w) with residue
1 X
res Gs (z/w) = − uk (z) uk (w) .
s=sj 2s − 1 s =s
k j
If s = 1/2 belongs to the discrete spectrum, then it gives a pole of order

2, but the residue comes from the Eisenstein series only. More precisely,
the Laurent expansion at s = 1/2 begins as follows
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
Estimates for the Fourier

coefficients of Maass forms
The main goal of this chapter is to establish auxiliary estimates

which are needed later when looking after the convergence of various
series. Crude results would often do. However since it is effortless to
be general, sharp and explicit in some cases, we go beyond the primary
objective.
8.1. Introduction.
Let {uj (z) : j ≥ 0} be a complete orthonormal system of Maass

forms for the discrete spectrum together with the eigenpacket {Ec (z, s) :
s = 1/2 + it, t ∈ R} of Eisenstein series for the continuous spectrum for
L(Γ\H). These have Fourier expansions of the type
X
(8.1) uj (σa z) = ρaj (0) y 1−sj + ρaj (n) Wsj (nz) ,
n6=0
X
s 1−s
(8.2) Ec (σa z, s) = δac y + ϕac (s) y + ϕac (n, s) Ws (nz)
n6=0
where ρaj (0) = 0 if uj (z) is a cusp form or otherwise uj (z) is a linear

combination of residues of Eisenstein series at s = sj > 1/2. The
117
118 Estimates for the Fourier coefficients of Maass forms
Whittaker function satisfies Wsj (nz) ∼ e(nz) as n → +∞; whence the

tail of (8.1) looks like an expansion into exponentials. However the
terms opening the series have rather peculiar shape; they look like a
power series as long as the argument is much smaller than |tj |. Then
the transition occurs somewhere near 2π|n|y ∼ |tj | if |tj | is large.
The coefficients ρaj (n) can be exponentially large in |tj | since they
must diminish the exponential decay of Wsj (nz) to give kuj k = 1.
Actually, uj (z) is quite small everywhere. If uj is a cusp form we
obtain by Theorem 3.2 that
|sj | 1
(8.3) uj (σa z)2 1+
y ca y
where the implied constant is absolute. From this
1/4
(8.3’) uj (z) λj (y + y −1 )−1/2
where the implied constant depends on the group. Similarly, if uj is a
residual form, then
(8.4) uj (z) y(z)1−sj .
For visual aesthetics we scale down the Fourier coefficients to
4π|n| 1/2
(8.5) νaj (n) = ρaj (n) ,
cosh πtj
4π|n| 1/2
(8.6) ηac (n, t) = ϕac (n, 1/2 + it) ,
cosh πt
if n 6= 0. Note that π(cosh πtj )−1 = Γ(sj ) Γ(1 − sj ) for sj = 1/2 + itj .
Since of every two points sj , 1−sj only one counts, in subsequent writing
we make a unique selection requiring either tj ≥ 0 or 1/2 < sj ≤ 1.
We shall see that the coefficients so scaled are bounded on average in
various ways with respect to n and tj .
By Theorem 3.2 we infer that the Fourier coefficients of a cusp
form uj (z) satisfy
X X
(8.7) |νaj (n)|2 (1 + λ−1
j ) |s j | + ,
ca
0<|n|≤X
where the implied constant is absolute. Hence

(8.8) νaj (n) |tj |1/2 + |n|1/2 ,
where the implied constant depends on the group.
The Rankin-Selberg convolution 119
8.2. The Rankin-Selberg convolution.
A more precise estimate than (8.7), at least when X is sufficiently

large, can be established using analytic properties of the series
X
(8.9) Laj (s) = |νaj (n)|2 |n|−s .
n6=0
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
Here the Mellin transform of |Kitj (y)|2 is given by the product of

gamma functions (see Appendix B.4) which leads us to (8.10).
From (8.10) we deduce that the Rankin-Selberg function Laj (s)
has an analytic continuation over the whole s-plane. In the half-plane
Re s ≥ 1/2 the poles of Laj (s) are among those of Ea (z, s) so they are
simple and lie in the segment (1/2, 1]. At s = 1 the residue is
(8.12) res Laj (s) = 8 |F |−1

s=1
by Proposition 6.13. Note that Θj (1) = 1 and kuj k = 1.

The Rankin-Selberg L-function inherits a functional equation from

that for the Eisenstein series. Precisely, by (8.10) and (6.22) the column-
vector
(8.13) Lj (s) = [. . . , Laj (s), . . . ]t
satisfies
(8.14) Θj (s) Lj (s) = Φ(s) Θj (1 − s) Lj (1 − s)
where Φ(s) is the scattering matrix for the group Γ.
On the vertical lines Re s = σ > 1 we get estimates for Lj (s)

by (8.7). Hence the functional equation will give us control of the
growth of Lj (s) in the critical strip by means of the Phragmen-Lindelöf
convexity principle provided we can control the growth of Φ(s). The
latter is an open problem for general groups. When this problem is
resolved in a number of cases, it proves that (s − 1) Laj (s) |s|A in
Re s ≥ 1 − ε. Hence, one infers by a standard complex integration the
following asymptotic formula
X
(8.15) |νaj (n)|2 ∼ 8 |F |−1 X ,
|n|≤X
as X → +∞. In particular this says that |νaj (n)| is about 2 |F |−1/2 on

average.
In the case of Γ = Γ0 (N ) one can be more precise because the

scattering matrix is computed explicitly (see [He1], [Hu1]). One can
prove that (s − 1) Laj (s) is holomorphic in Re s ≥ 1/2 and
(8.16) (s − 1) Laj (s) |s sj N |3 .
Hence one derives by a contour integration together with (8.7) that

X
(8.17) |νaj (n)|2 = 8 |F |−1 X + O(|sj | X 7/8 ) ,
|n|≤X

Bounds for linear forms 121
8.3. Bounds for linear forms.
In numerous applications one needs bounds not necessarily for the

individual νaj (n) but for mean values of some kind with respect to n
as well as the spectrum. Sometimes even the group Γ can vary.
In practice one meets linear forms
N
X
(8.18) Laj (a) = an νaj (n)
1
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
Immediately, from (8.7) and Cauchy’s inequality we derive that

N
(8.19) L2aj (a) (1 + λ−1
j ) |s j | + kak2 .
ca
This estimate is best possible apart of the implied constant.
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
Theorem 8.1. If uj is a cusp form, then

X
(8.20) e(αn) νaj (n) (λj N )1/2 log 2N
|n|≤N
where the implied constant depends only on the group.
Proof. We have
Z 1
ρaj (n) Wsj (iny) = uj (σa z) e(−nx) dx .
0
Integrating with respect to the measure y −1 dy we get

1
sj 1 − s j
Z
−1/2
(8.21) π Γ Γ ρaj (n) = ϕaj (x) e(−nx) dx
2 2 0
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
with the kernel

X sin π(2N + 1)x
FN (x) = e(nx) = .
sin πx
|n|≤N
The L1 -norm of FN (x) is small, namely

Z 1
|FN (x)| dx log 2N ,
0
and applying Stirling’s approximation to the gamma factors we infer

that X
νaj (n) |n|−1/2 |sj |1/2 |1 − sj |1/2 log 2N .
|n|≤N
Finally relax |n|−1/2 by partial summation and introduce the character

e(αn) by changing σa to σa n(α) to complete the proof.
Remarks. The estimate (8.22) is quite crude, probably ϕaj |sj |ε .

The exponent 1/2 in (8.20) cannot be reduced for all α in view of Par-
seval’s identity. Nevertheless for special α and for a properly chosen
scaling matrix σa a considerable improvement is possible, if Γ is a con-
gruence group at any rate.
Spectral mean-value estimates 123
The uniformity in α allows us to draw interesting consequences.

For example, by way of additive characters to modulus q, we can stick
to an arithmetic progression getting the same bound as (8.20), i.e.
X
(8.23) νaj (n) (λj N )1/2 log 2N .
|n|≤N
n≡a(mod q)
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).
Theorem. Let uj be a cusp form. For any a = (am ), b = (bn ), we

have
X X 1/2 π
am bn ρaj (m − n) λj cosh( tj ) kak kbk ,
2
|m|≤M |n|≤N
where the implied constant depends on the group.
A vast amount of cancellation between terms of the above bilinear

form indicates strongly that the Fourier coefficients of a cusp form are
in no way near to being an additive character. They rather tend to
be multiplicative after an adequate diagonalization, if Γ is a congru-
ence group. A multiplicative analogue of the Theorem in this case is
obviously false.
8.4. Spectral mean-value estimates.
Next we establish estimates for the Fourier coefficients of Maass

forms on average with respect to the spectrum. Many useful estimates
on average with respect to the spectrum for the congruence group Γ0 (q)
have been established by J.-M. Deshouillers and H. Iwaniec [De-Iw].
Here is a sample
X
|Laj (a)|2 (T 2 + q −1 N log 2N ) kak2
|tj |<T
for any N, T ≥ 1 (originally we had N ε in place of log 2N ). In these

lectures we prove a result of similar type (see (8.25)), which is somewhat
weaker, but it holds for any group and is no less valuable.
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
for T ≥ 1 and z ∈ H, where the implied constant depends on the group.
For fixed n we shall do better by an appeal to the complete spectral

decomposition of a particular automorphic kernel (7.16). We begin by
estimating the twisted Maass forms which are obtained from the Fourier
expansion (8.1) by multiplying its coefficients with a sequence a = (an ).
Denote these by
X
a ⊗ uj (σa z) = an ρaj (n) Wsj (nz)
n
if uj is a cusp form. Similarly, we twist the Eisenstein series and their

residues. For notational simplicity and without loss of generality we
are going to consider a = ∞, σa = 1 (change the group to σa−1 Γσa ) and
z = iy (change an to an e(−nx)). In this case we also drop the subscript
a in relevant places and set Aj (y) = a ⊗ uj (σa z). We have
Z 1
Aj (y) = S(x) uj (z) dx
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
By the spectral decomposition (7.17), we get

X Z 1Z 1
0
h(tj ) Aj (y) Aj (y ) + · · · = S(x) S(x0 ) K(z, z 0 ) dx dx0 .
j 0 0
Spectral mean-value estimates 125
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
is easily recognized as an incomplete Eisenstein series (see the remarks

concluding Section 4.2)
X
H(z, y 0 ) = ψ(Im γz)
γ∈Γa \Γ
(remember that a = ∞) where ψ(y) = (y 0 y)1/2 g(log y 0 /y) and g(r) is

the Fourier transform of h(t). The identity motion yields ψ(y). The
other motions, by Lemma 2.10, using partial summation contribute all
together at most
+∞ +∞
10 10 1
Z Z
−1 0
y |ψ (y)| dy = er/2 g(r) − g 0 (r) dr .

ca 0 ca −∞ 2
To simplify, assume that g(r) is positive and decreasing on R+ (this hy-

pothesis implies our former condition that k(u) is non-negative). Then
the last integral is bounded by
+∞
r i
Z
(g(r) − 2 g 0 (r)) cosh dr = g(0) + h .
0 2 2
Therefore,
0 0 1/2 y 0 10 i
H(z, y ) ≤ (y y) g log + g(0) + h .
y ca 2
Taking y 0 = y we obtain the following inequality

X 10 10 i
h(tj ) |Aaj (y)|2 + · · · ≤ y+ g(0) + h kak2 .
j
ca 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
for T ≥ 1 and z ∈ H. Compare this with (8.24).

Next clear (8.26) of the Whittaker functions by integrating over
the dyadic interval Y < y < 2Y with Y = c T /|n| (as in the proof
of Theorem 3.2 the integration is necessary because one cannot find a
universal value of y for which all Wsj (iny) have the expected order of
magnitude). One gets the following estimate
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
for any T ≥ 1 and n 6= 0 where the implied constant is absolute. A more

precise asymptotic formula (9.13) will be derived from the Bruggeman-
Kuznetsov formula (9.12).
8.5. The case of congruence groups.
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
of the Atkin-Lehner theory of newforms [At-Le]. The case of Maass

forms is identical except for verbal differences. We do not wish to
develop this theory from scratch here but rather only transcribe briefly
the main concepts and results for the group Γ0 (N ).
For n ≥ 1 denote the set

a b
(8.28) Γn = : a, b, c, d ∈ Z, ad − bc = n .
c d
In particular Γ1 is the modular group. Naturally Γ1 acts on Γn . The

Hecke operator Tn : A(Γ1 \H) −→ A(Γ1 \H) is defined by
1 X
(8.29) (Tn f )(z) = √ f (τ z) .
n
τ ∈Γ1 \Γn
Picking up specific representatives of Γ1 \Γn we can also write
1 X X az + b
(8.30) (Tn f )(z) = √ f ;
n d
ad=n b(mod d)
clearly, the sum is finite, the number of terms being

X
[Γn : Γ1 ] = d = σ(n) ,
d|n
and therefore, Tn is bounded on L(Γ1 \H) by σ(n) n−1/2 . One shows

the following multiplication rule
X
(8.31) T m Tn = Tmnd−2
d|(m,n)
so that in particular Tm , Tn commute. Also Tn commutes with the

Laplace operator because it is defined by the group operations.
Consider the Hecke congruence group Γ0 (N ) of level N ≥ 1. Since

A(Γ0 (N )\H) ⊂ A(Γ1 \H), every operator Tn acts on A(Γ0 (N )\H). Nev-
ertheless, only Tn with (n, N ) = 1 are interesting. First of all Tn is
self-adjoint in L(Γ0 (N )\H), i.e.
hTn f, gi = hf, Tn gi , if (n, N ) = 1 .

(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.
(8.32) Tn uj (z) = λj (n) uj (z) , if (n, N ) = 1 .
The Eisenstein series E∞ (z, 1/2 + it) is shown to be an eigenfunction

of all the Hecke operators Tn , (n, N ) = 1, with eigenvalue
X a it
(8.33) ηt (n) = .
d
ad=n
It is conjectured (Ramanujan-Petersson) that
(8.34) |λj (n)| ≤ τ (n) , (n, N ) = 1 .
By (8.33) the conjecture is obviously true in the space of continuous

spectrum. In the cuspidal space the best known bound so far is
λj (n) ≤ τ (n) n5/28
which is due to D. Bump, W. Duke, J. Hoffstein, H. Iwaniec [Bu-Du-

Ho-Iw]. For the constant eigenfunction the Hecke eigenvalue is much
larger; we have exactly
(8.35) λ0 (n) = σ(n) n−1/2 .
By virtue of (8.32) the Hecke operator Tn acts on the Fourier co-

efficients of uj (z) in the cusp a = ∞ simply by
(8.36) νj (n) = νj (1) λj (n) , if (n, N ) = 1 .
In fact it is plain from (8.30)-(8.32) that

X mn
(8.37) νj (m) λj (n) = νj , if (n, N ) = 1 .
d2
d|(m,n)
From now on we drop the subscript j for notational simplicity. The

relation (8.36) says that the Fourier coefficients ν(n) are proportional
to the Hecke eigenvalues λ(n) provided ν(1) 6= 0, but unfortunately
ν(1) may vanish for some forms. For example take a cusp form v(z) on
an overgroup Γ0 (M ) ⊃ Γ0 (N ) with M |N . Then u(z) = v(Dz) where
DM |N , is a cusp form on Γ0 (N ) all of whose coefficients ν(n) vanish,

save for n ≡ 0 (mod D). If M < N such v(Dz) is seen as an oldform.
Atkin and Lehner have shown how to split the space of cusp forms into
newforms. Let us write
C(Γ0 (N )\H) = Cold (Γ0 (N )\H) ⊕ Cnew (Γ0 (N )\H) .
Here Cold (Γ0 (N )\H) is the linear subspace of C(Γ\H) spanned by forms
of type v(Dz), where v(z) is a Maass cusp form on Γ0 (M ) with DM |N ,
M < N , and by definition Cnew (Γ0 (N )\H) is the orthogonal comple-
ment. Clearly Tn with (n, N ) = 1 maps Cold (Γ0 (N )\H) into itself be-
cause it commutes with the operator f (z) 7→ f (Dz) for every D|N .
Consequently, Tn maps Cnew (Γ0 (N )\H) into itself because Tn is hermi-
tian. Therefore, there exists a basis in Cnew (Γ0 (N )\H) of Maass cusp
forms which are common eigenfunctions of all Tn with (n, N ) = 1 (one
can work it out in each spectral eigenspace separately). These cusp
forms are called newforms of level N . The newforms are the GL2 ana-
logue of the primitive Dirichlet characters χ(mod N ).
Let us return to the space Cold (Γ0 (N )\H). A function u(z) in
Cold (Γ0 (N )\H) is called an oldform if u(z) = v(Dz), where v(z) is a
newform on some overgroup Γ0 (M ) with DM |N , M < N . In this
case we say that u(z) is an oldform of level M and divisor D. It turns
out, but it is not automatically a fact, that the space Cold (Γ0 (N )\H) is
spanned by oldforms.
Another pleasant fact is that a newform of level N , besides being

a common eigenfunction of all Tn with (n, N ) = 1, is automatically an
eigenfunction of all the operators Up , p|N , defined by
1 X z + b
(8.38) (Up f )(z) = √ f .
p p
b(mod p)
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)
λ(m) λ(p) = λ(mp) if p|N .

Observe that for all m, n ≥ 1 one has

X mn
(8.40) |λ(m) λ(n)| ≤ λ 2 .
d
d|(m,n)
Trivially |ν(n)| = |ν(1) λ(n)| ≤ |ν(1)| λ0 (n); whence by (8.17) we

get a crude lower bound
(8.41) |ν(1)|2 (λ N )−9 .
Now we are ready to use the power of multiplication to establish

the following
Theorem 8.3. Let u(z) be a newform on Γ0 (N ) with eigenvalue λ and

Fourier coefficients ν(n) = ν(1) λ(n). Then we have
X
(8.42) x1/2 (log 2N x)−1 |λ(n)|2 x (λ N )ε
0<n≤x
for all x ≥ 1 and any ε > 0, the implied constant depending only on ε.
Moreover, we have
(8.43) N −1 (λ N )−ε |ν(1)|2 λ1/4 N −1/2 log 2N .
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
where the implied constant is absolute (assuming kuk = 1). Therefore,

the novelty of (8.42) rests in the short range of x < λ1/2 N which is
crucial for applications.
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
for s > 1. First, by (8.44) and (8.41), we derive a crude estimate
L(s) |ν(1)|−2 λ1/2 (s − 1)−1 (λN )10 (s − 1)−1 ,
where the implied constant is absolute. Then by (8.40) and Cauchy’s

inequality we infer that
X X mn 2
2
L (s) ≤ λ 2 (mn)−s
m,n d|(m,n)
d
X X mn 2
≤ τ ((m, n)) λ 2 (mn)−s
m,n
d
d|(m,n)
X X
= |λ(`)|2 `−s τ ((m, n) d) d−2s
` d mn
mn=`
X
≤ ζ 2 (2s) |τ (`) λ(`)|2 `−s L(s − ε) ,
`
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 .
Presumably, the lower bound (8.42) should be x (λ N )−ε uniformly

in x ≥ 1, but it resists a proof. Such a bound is desired for numerous
applications (see for example Section 13.4). It would give us, among
other things, the following estimate
(8.46) |ν(1)|2 N −1 (λ N )ε .
Very recently J. Hoffstein and P. Lockhart [Ho-Lo] have established

(8.46) unconditionally using quite advanced results from the theory
of automorphic L-functions and ideas of C. L. Siegel concerning the
exceptional zero (see also the appendix to [Ho-Lo] by D. Goldfeld, J.
Hoffstein and D. Lieman). The estimates (8.43) and (8.46) combined
determine the true order of magnitude of ν(1).
Chapter 9
Spectral theory of
Kloosterman sums
9.1. Introduction.
Kloosterman sums were invented to refine the circle method of

Hardy and Ramanujan [Ha-Ra]. Originally, Kloosterman [Kl1] ap-
plied his refinement to counting representations by a quadratic form in
four variables. Shortly afterwards, Kloosterman [Kl2] and Rademacher
[Rad] used the idea to estimate the Fourier coefficients of classical mod-
ular forms. These coefficients were later expressed effectively as sums
of Kloosterman sums without appealing to the circle method by H.
Petterson [Pe1], R. Rankin [Ran] and A. Selberg [Se3] independently.
Therefore, a connection between Kloosterman sums and modular forms
was established right away. Next, algebraic geometry became associ-
ated with modular forms via Weil’s estimate for Kloosterman sums (a
special case of the Riemann hypothesis for curves). Consequently, the
results became deeper but still not the best possible. Only recently a
complete picture has emerged from the spectral theory of automorphic
forms. Its essence is captured in the analytic continuation of the series
X
(9.1) Ls (m, n) = c−2s Sab (m, n; c) .
c>0
due to A. Selberg [Se1]. There is an elegant treatment of this series
133
134 Spectral theory of Kloosterman sums
by D. Goldfeld and P. Sarnak [Go-Sa]. In these lectures we catch a

glimpse of that profound theory. Instead of (9.1) we shall examine
the series (5.16), which occurs on our way more naturally. Of course
both are related via the power series expansion for the Bessel functions
J2s−1 (x), I2s−1 (x); precisely we have
∞
2 s−1
X (4π 2 mn)k
(9.2) Zs (m, n) = π(4π |mn|) Ls+k (m, n) .
k! Γ(k + 2s)
k=0
Recall that the series Ls (m, n) and Zs (m, n) converge absolutely in

Re s > 1.
9.2. Analytic continuation of Zs (m, n).
We shall appeal to the properties of the automorphic Green func-

tion Gs (z/z 0 ) already established (originally Selberg has employed
Poincaré series). The idea is straightforward: on one hand the zeta-
function Zs (m, n) turns up in the Fourier expansion (5.15); on the
other hand we have the spectral decomposition (7.18) for the differ-
ence Gs (z/z 0 ) − Ga (z/z 0 ) with a > 1, Re s > 1 in terms of Maass forms.
These forms have Fourier expansion too (see (8.1) and (8.2)). Com-
paring the (m, n)-th Fourier coefficients for mn 6= 0 of both expansions
leads to the following identity
δab δmn (4π|n|)−1 Ws (iny 0 ) Vs (iny) + Zs (m, n) Ws (imy 0 ) Ws (iny)

− the same expresion for s = a
(9.3)
χsa (sj ) ρ̄aj (m) ρbj (n) Wsj (imy 0 ) Wsj (iny)
P
= j
+ the countinuous spectrum integrals
where recall that χsa (v) = (s−v)−1 (1−s−v)−1 −(a−v)−1 (1−a−v)−1 .

To be precise this step requires the absolute convergence of the relevant
hybrid Fourier/spectral series. The upper bound (8.24) is just adequate
to verify the convergence in question so (9.3) is established thoroughly.
Another point is that the Fourier expansion (5.15) is valid only for
(z, z 0 ) in the set Dab ; therefore, (9.3) requires y 0 ≥ y, y 0 y ≥ c(a, b)−2 ,
and this condition is indispensable.
Analytic continuation of Zs (m,n) 135
The relation (9.3) will simplify considerably taking the limit as

a → +∞ since the corresponding terms vanish. To see that we can take
the limit on the left hand side we apply the trivial bound O(c2 ) for the
Kloosterman sums Sab (m, n; c), and for the involved Bessel functions
we apply the asymptotics
x ν 2 ν
Jν (x) ∼ Γ(ν + 1)−1 and Kν (x) ∼ Γ(ν)
2 x
as ν → ∞ uniformly in 0 < x 1 (see the power series expansions).

Hence, we infer that
Γ(s − 1/2) 1 −1

Ws (iny 0 ) Vs (iny) = s− ,
Γ(s + 1/2) 2
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)
if y 0 y ≥ c−2 and s ≥ 2. These estimates also hold true with J2s−1

replaced by I2s−1 . Therefore, the contribution to the left side of (9.3)
of terms for s = a is bounded by O(a−1 ). On the right side we split the
series of spectral terms in accordance with χsa (v) and show that the
resulting series for s = a vanish as a → +∞. To be correct one ought
to verify the uniform convergence. For this purpose the bound (8.26)
is more than sufficient while (8.24) barely misses.
This said, we drop in (9.3) all terms with s = a. Furthermore for

simplicity we change 2π|n|y → y, 2π|m|y 0 → y 0 and to obtain symmetry
we set
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
Proposition 9.1. Suppose y 0 y ≥ 4π 2 c(a, b)−2 |mn|. For any s with

Re s > 1 we have
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 .
The sum over the discrete spectrum in (9.5) extends to all of C by

analytic continuation and gives a function invariant under the change
s 7→ 1 − s. It has simple poles at s = sj and at s = 1 − sj with the
same residue
X
−(2sj − 1)−1 ρ̄ak (m) ρbk (n) Kitj (y 0 ) Kitj (y)
sk =sj
provided sj 6= 1/2. If sj = 1/2 (so λj = 1/4 belongs to the discrete

spectrum which is necessarily cuspidal), then it contributes
1 X
−(s − )−2 ρ̄ak (m) ρbk (n) K0 (y 0 ) K0 (y) .
2
sk =1/2
This has a double pole at s = 1/2 with residue zero.
The continuous spectrum integrals in (9.5) give functions holomor-

phic in Re s > 1/2. To extend these to Re s ≤ 1/2 we repeat the
arguments given for the Green function in Section 7.4. Accordingly we
shall move the integration from Re v = 1/2 to Re v = α with α larger
but close enough to 1/2 so that all ϕac (m, v), ϕbc (n, v) are holomorphic
in the strip 1/2 < Re v < α. Assuming for a moment that s is in this
strip we pass the pole at v = s with residue
−(2s − 1)−1 ϕac (m, 1 − s) ϕbc (n, s) Ks−1/2 (y 0 ) Ks−1/2 (y) .
The resulting integral over the vertical line Re v = α gives a holomor-

phic function in the strip 1−α < Re s < α which is also invariant under
Analytic continuation of Zs (m,n) 137
the change s 7→ 1 − s. Also invariant is the above residue except for

sign change since we have
X X
(9.6) ϕac (m, 1 − s) ϕbc (n, s) = ϕac (m, s) ϕbc (n, 1 − s)
c c
by (7.20). Finally notice that

1
Dν (y 0 , y) − D−ν (y 0 , y) = − sin(πν) Kν (y 0 ) Kν (y)
π
since
π
(9.7) Kν (y) = I−ν (y) − Iν (y) .
2 sin πν
Having made the above transformations we take the difference of
(9.5) at s and 1 − s. We find that in every term left the Bessel functions
Kit (y 0 ), Kit (y) match and wash away leaving us with the clean
Theorem 9.2. The series Zs (m, n) has an analytic continuation in s

to all C, and it satisfies the functional equation
1 1
Zs (m, n) − Z1−s (m, n) = δab δmn sin π s −
2π|n| 2
(9.8) 1 X
− ϕac (m, 1 − s) ϕbc (n, s) .
2s − 1 c
Zs (m, n) has simple poles at s = sj and s = 1 − sj with residue

1 X
− ρ̄ak (m) ρbk (n)
2sj − 1 s =s
k j
provided sj 6= 1/2. At s = 1/2 it has a pole with Taylor expansion

1 X
Zs (m, n) = − ρ̄ak (m) ρbk (n)
(s − 1/2)2
sk =1/2
1 X
− ϕ̄ac (m, 1/2) ϕbc (n, 1/2) + · · ·
4(s − 1/2) c
where the first sum is void if sk = 1/2 does not exist.
Remarks. Since the poles of Zs and Z1−s at s 6= 1/2 cancel out, it

follows that (s − 1/2)(Zs − Z1−s ) is entire and so is the sum (9.6).
9.3. Bruggeman-Kuznetsov formula.
For practical purposes it is advantageous to quantify the analytic

properties of the Kloosterman sums zeta-function Zs (m, n), i.e. to
develop a kind of Poisson summation formula with test functions as
flexible as possible. In analytic number theory this is routine for the
Dirichlet series, which satisfy standard functional equations. Our case
of Zs (m, n) is not much different.
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
where of each two points sj , 1 − sj only one counts (check separately

the residue at sj = 1/2). When on the line Re s = 1 − ε we expand
Zs (m, n) into series of Kloosterman sums (see (5.16)) and integrate
termwise getting
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
and (use (9.7))
+∞
4
Z
−
(9.11) h (x) = K2it (x) sinh(πt) h(t) t dt .
π −∞
Remarks. The above presentation requires the absolute convergence

of the series for Zs (m, n) on Re s = 1 − ε which is often the case.
Nevertheless, this hypothesis can be avoided by moving to the line
Re s = 1 + ε where the series does converge absolutely. It will produce
an additional term 4 h(i/2) Z1 (m, n) from the pole of f (s) at s = 1.
The same term will reappear from evaluation of h± (x) after moving
back to Re s = 1 − ε (and further down to Re s = 1/2). Therefore
this term contributes nothing. At this point one needs the convergence
of Z1 (m, n) (not necessarily absolute) which can be established by a
routine estimate using (9.5) and (8.26) (one needs the holomorphy of
Zs (m, n) at s = 1 and some control of growth).
From the above parts we assemble the following formula
Theorem 9.3. Let a, b be cusps of Γ, mn 6= 0 and νaj (m), νbj (n),

ηac (m, t), ηbc (n, t) be the Fourier coefficients of a complete orthonor-
mal system of Maass forms and the eigenpacket of Eisenstein series in
L(Γ\H). Then for any h(t) satisfying (1.63) we have
X
h(tj ) ν̄aj (m) νbj (n)
j
X 1 Z +∞
(9.12) + h(t) η̄ac (m, t) ηbc (n, t) dt
c
4π −∞
X 4π p|mn|
−1 ±
= δab δmn h0 + c Sab (m, n; c) h
c
c
where ± is the sign of mn and h0 , h+ , h− are the integral transforms
of h given by (9.9)-(9.11).
Remarks. One could derive (9.12) by integrating (9.3) instead of (9.8).

This formula was established first for the modular group by N. V.
Kuznetsov and in a slightly less refined form by R. W. Bruggeman
[Br1]. They and many others have used the Selberg-Poincaré series
Eam (z, s) rather than the Green function. N. V. Proskurin [Pr] con-
sidered the general case of finite volume groups and forms of arbitrary
weight. Motivated by numerous applications J.-M. Deshouillers and
H. Iwaniec [De-Iw] worked out the formula with mixed cusps for the
group Γ0 (q). J. Cogdell and I. Piatetski-Shapiro [Co-Pi] have given a
conceptual proof for general finite volume groups in the framework of
representation theory. A far reaching generalization to Γ ⊂ G where
G is a real rank 1 semisimple Lie group and Γ its discrete subgroup
was established by R. Miatello and N. Wallach [Mi-Wa]. A. Good pur-
sued the study in the direction of Kloosterman sums associated with
arbitrary one-parameter subgroups of a Fuchsian group acting on H in
place of the stability groups of cusps. His work [Go] deserves deeper
penetration.
To catch a glimpse of applications we recommend one to derive

from Theorem 9.3 the following asymptotic formula (use the trivial
bound (2.38) for sums of Kloosterman sums)
X X 1 Z T 2
2
(9.13) |νaj (n)| + |ηac (n, t)|2 dt = T 2 +O(|n| T )
c
4π −T π
|tj |<T
for n 6= 0 and T ≥ 1 where the implied constant depends on the group.

Bruggeman-Kuznetsov formula reversed 141
9.4. Bruggeman-Kuznetsov formula reversed.
The formula (9.12) serves dual purposes. First it gives a way to

study the spectrum as well as the Fourier coefficients of Maass forms
on Γ\H by means of Kloosterman sums. Through these sums one can
employ a variety of methods beyond soft analysis; and when the group
is arithmetic, one can use deep facts from algebraic geometry (Weil’s
bound for Kloosterman sums). The obtained result is often sharper
than what is possible for general groups. Next, assuming the spectral
aspects are well understood, one can reverse the first approach to study
sums of Kloosterman sums. Such studies are of primary interest for
analytic number theory. At first glance this switch may seem to be a
zero value game; however, vast research in the early eighties has made
this interplay very productive (see the surveys [Iw1], [Iw2]).
When applying it to Kloosterman sums one desires a general test

function on that side of (9.12) rather than on its spectral side. This
leads us to the problem of reversing the transforms h 7→ h− and h 7→ h+ .
If mn < 0 it is possible to solve this problem entirely by means of the
Kontorovich-Lebedev inversion (1.31). Indeed, any function of C 2 class
on [0, +∞) satisfying the following conditions
(9.14) f (0) = 0 , f (j) (x) (x + 1)−2−ε , j = 0, 1, 2 ,
is realized in the image of the transform h 7→ h− ; precisely we have

+∞
4
Z
f (x) = K2it (x) Kf (t) sinh(πt) t dt
π −∞
where
+∞
4
Z
(9.15) Kf (t) = cosh(πt) K2it (x) f (x) x−1 dx .
π 0
Then Theorem 9.3 for h(t) = Kf (t) becomes

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 Kf is the integral transform of f given by (9.15).
If mn > 0 we need to invert h 7→ h+ . Unfortunately the image of

this transform is not dense in the space of functions of basic interest,
specifically in L2 (R+ , x−1 dx). For a closer examination of this subspace
let us write (9.10) as follows
Z +∞
+
h (x) = 4 B2it (x) tanh(πt) h(t) t dt
0
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
Then for h(t) = Tf (t) Theorem 9.3 becomes
Theorem 9.5. Let mn > 0. For any f satisfying (9.13) we have
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 −∞
Although the functions of type f ∞ (x) span most of the space we

often need when working with Kloosterman sums, they are never nice
to deal with in practice; therefore it is important to determine the
complementary function to f ∞ (x). This is best described in terms of
the Hankel transform
Z +∞
(9.23) Hf (x) = J0 (xy) f (y) dy .
0
To recover f use the Hankel inversion (see Appendix B.5)

Z +∞
(9.24) f (x) = xy J0 (xy) Hf (y) dy .
0
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
hence the complementary function f 0 (x) = f (x) − f ∞ (x) is given by

Z 1
0
(9.26) f (x) = xy J0 (xy) Hf (y) dy .
0
This said, we wish to have a formula complementary to (9.22)

with f 0 (x) in place of f ∞ (x). To find it first recognize that f 0 (x) is
the projection of f (x) onto the subspace of L2 (R+ , x−1 dx) spanned by
the Bessel functions J` (x) of order ` ≥ 1, ` ≡ 1 (mod 2), i.e. f 0 (x) is
given by the Neumann series (see Appendix B.5)
X
(9.27) f 0 (x) = 2` Nf (`) J` (x) ,
1≤`≡1 (mod 2)
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.
9.5. Petersson’s formulas.
Let Mk (Γ) denote the linear space of holomorphic functions

f : H −→ C satisfying the following transformation rule
(9.29) jγ (z)−k f (γz) = f (z) , γ ∈ Γ.
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
if k > 2, where (g; m1 , . . . , m` ; h) is the signature of Γ (for k = 2 the

formula is slighty different). Hence
|F |
dim Mk (Γ) ≤ k+1
4π
by the Gauss-Bonnet formula (2.7).
Any f ∈ Mk (Γ) has the Fourier expansion in cusps of type

∞
fâ (n) e(nz)
X
−k
(9.31) jσa (z) f (σa z) =
n=0
which converges absolutely and uniformly on compacta. If for every

cusp
(9.32) fâ (0) = 0 ,
then f is called a cusp form. A cusp form has exponential decay at

cusps. In particular y k/2 f (z) is bounded on H. The subspace of cusp
forms, say Sk (Γ), is equipped with the Petersson inner product
Z
(9.33) hf, gik = y k f (z) ḡ(z) dµz .
F
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
The space Sk (Γ) is spanned by the Poincaré series

X
(9.35) Pam (z) = jσa−1 γ (z)−k e(m σa−1 γz)
γ∈Γa \Γ
with m ≥ 1. This is obvious by the following formula of Petersson

(k − 2)! ˆ
(9.36) hf, Pam ik = fa (m)
(4πm)k−1
which is derived by the unfolding technique. Indeed, the subspace

spanned by the Poincaré series is closed and any function orthogonal
to this subspace is zero by (9.36). Let us choose an orthonormal basis
of Sk (Γ), say {fjk }, and expand Pam into this basis. By (9.36) we get
(k − 2)! X ˆ
(9.37) Pam (z) = fajk (m) fjk (z) ,
(4πm)k−1 j
where fâjk (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
Theorem 9.6. Let m, n be positive integers and let k be a positive

even integer. Then
(k − 2)! X ˆ
√ fajk (m) fbjk (n)
(4π mn)k−1 j
(9.40) 4π √mn
X
k −1
= δab δmn + 2πi c Sab (m, n; c) Jk−1 .
c
c
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.
Finally we generalize (9.40) by summing in accordance with (9.27)

(use (9.34) to validate the convergence). We obtain
Petersson’s formulas 147
Theorem 9.7. Let m, n > 0. For any f satisfying (9.13) we have

√
0 4π mn
X
∞ −1
− δab δmn f + c Sab (m, n; c) f
c
c
(9.41) X
= ik Nf (k − 1) ψ̄ajk (m) ψbjk (n)
2≤k≡0 (mod 2)
where we have scaled the Fourier coefficients of an orthonormal basis

{fjk } of Sk (Γ) down to
π −k Γ(k) 1/2
(9.42) ψajk (m) = fâjk (m) .
(4m)k−1
All terms in (9.41) are obtained straightforwardly except for −f ∞ ,

which needs a few words of explanation. This term comes out as follows
X
(2πik )−1 2` Nf (`)
1≤`≡1 (mod 2)
+∞ ∞
1
Z X
= (−1) (2r − 1) J2r−1 (x) f (x) x−1 dx
r
π 0 r=1
+∞
−1
Z
= J0 (x) f (x) dx = −f ∞
2π 0
by applying the recurrence relation Jn−1 (x) + Jn+1 (x) = 2ny −1 Jn (x).
Theorem 9.7 constitutes the exact complement to Theorem 9.5.

Adding (9.41) to (9.22) one gets a complete spectral decomposition for
sums of Kloosterman sums. Notice that the diagonal terms δab δmn f ∞
cancel out whereas the test functions attached to the Kloosterman sums
make up the original function f (x) = f 0 (x) + f ∞ (x). On the spec-
tral sides of (9.41) and (9.42) the Fourier coefficients of basic automor-
phic forms are having attached the integral transforms ik Nf (k − 1) =
Tf (i(k − 1)/2) and Tf (tj ) respectively (see Appendix B.5 where we call
these the Neumann and the Titchmarsh coefficients of f respectively).
The fact that the holomorphic automorphic forms participate in the
decomposition is nicely explained in a language of representations by
R. W. Bruggeman [Br2]. The partition f (x) = f 0 (x) + f ∞ (x) together
with the series (9.27) and the integral (9.20) constitutes an inversion
formula first established rigorously by D. B. Sears and E. C. Titchmarsh
(see (4.6) of [Se-Ti, p.172]). A full characterization of functions which
can be represented exclusively by the Neumann series (9.27) was given
by G. H. Hardy and E. C. Titchmarsh [Ha-Ti].
Chapter 10
The trace formula
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.
A function K : F × F −→ C and the integral operator having K as

its kernel are said to be of trace class if K(z, w) is absolutely integrable
on the diagonal z = w in which case the integral
Z
Tr K = K(z, z) dµz
F
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
(the spectral trace of K, so to speak). On the other hand from the

series X
K(z, z) = k(z, γz)
γ∈Γ
(the geometric side of K, so to speak), we get
XZ
Tr K = k(z, γz) dµz .
γ∈Γ F
Comparing both numbers one obtains the pre-trace formula. It is a

quite useful expression yet it does not reveal the geometry of the surface
Γ\H.
Following Selberg we partition the group into conjugacy classes
[γ] = τ −1 γτ : τ ∈ Γ

before computing the geometrical trace. Given a conjugacy class C in

Γ let KC denote the partial kernel restricted to elements in C
X
KC (z, z) = k(z, γz) .
γ∈C
Thus K and its trace splits into

X X
K= KC , Tr K = Tr KC
C C
where
XZ
Tr KC = k(z, γz) dµz .
γ∈C F
Two elements τ, τ 0 ∈ Γ yield the same conjugate of γ if and only if

τ 0 τ −1 belongs to the centralizer

Z(γ) = ρ ∈ Γ : ργ = γρ .
Introduction 151
Therefore, we can write (choose γ ∈ C)

X Z
Tr KC = k(z, τ −1 γτ z) dµz
τ ∈Z(γ)\Γ F
Z
= k(z, γz) dµz
Z(γ)\H
where Z(γ)\H is a fundamental domain of the centralizer, the point

being that it is a relatively simple domain. For computation it is con-
venient to observe that the above integral really depends only on the
conjugacy class of γ in the group G = SL2 (R). Precisely, if γ 0 = g −1 γg
with g ∈ G, then
Z
Tr KC = k(z, g −1 γgz) dµz = Tr KC 0
g −1 Z(γ)g\H
where g −1 Z(γ)g is the centralizer of γ 0 in Γ0 = g −1 Γg and C 0 = g −1 Cg.

In particular, after conjugating Γ with suitable g we can find a repre-
sentative of the class in one of the groups ±N , ±A, K according to
whether the class is parabolic, hyperbolic or elliptic (remember that
every element of Γ has two representations in G = SL2 (R)). This
transformation will simplify further computations, but above all it illu-
minates the geometric side of the trace.
Recall the classification of motions described in Section 1.5. To

identify the conjugation representative of a motion in one of the groups
±N , ±A, K leads us to examine its fixed points. Looking from the
surface Γ\H we shall speak of equivalent classes of fixed points modulo
Γ to refer to the fixed points of a whole conjugacy class rather than to
a solitary motion. It does happen that two distinct conjugacy classes
in Γ have the same fixed points modulo Γ.
The set of fixed points of a given γ 6= ±1 (the identity motion
γ = ±1 having all H as fixed points is an exception; it has to be
considered differently) determines the primitive class C0 , say, and every
other class C which has the same fixed points modulo Γ is a unique
power of C0 , C = C0` , say, with ` ∈ Z, ` 6= 0 subject to 1 ≤ ` < m if C0 is
elliptic of order m. If γ = γ0` , where γ0 is primitive, then Z(γ) = Z(γ0 )
is cyclic generated by γ0 , therefore the fundamental domain Z(γ)\H is
as simple as a vertical strip, a horizontal strip or a sector in H, if γ is
brought to ±N , ±A, K, respectively.
Figure 10. Fundamental domains of centralizers.
The above demostration was oversimplified as the convergence

questions were temporarily ignored. A considerable difficulty occurs
when the surface Γ\H is not compact; thus, it has cusps which pro-
duce the continuous spectrum. In this case K is not of trace class on
F for two parallel reasons: on the spectral side the Eisenstein series
Ea (z, 1/2 + it) are not square integrable whereas on the geometric side
the partial kernels KC (z, z) for parabolic classes are not absolutely in-
tegrable over cuspidal zones. Selberg has dealt with the problem by
computing asymptotically the trace on the central part F (Y ) ⊂ F with
Y tending to +∞. Let Tr Y K stand for such a truncated trace
Z
Y
Tr K = K(z, z) dµz .
F (Y )
From the spectral side we obtain that Tr Y K ∼ A1 log Y + T1 whereas

from the geometric side Tr Y K ∼ A2 log Y + T2 , where A1 , A2 , T1 , T2
are constants which can be explicitely expressed in terms of Γ. Hence
one infers that T1 = T2 which is the celebrated trace formula (not a
tautology like the relation A1 = A2 ).
In these lectures we apply Selberg’s ideas to the iterated resolvent

(4.14) given by the kernel
K(z, w) = Gs (z/w) − Ga (z/w)
where Gs (z/w) is the automorphic Green function. Therefore the gen-
erating function
(10.1) k(u) = Gs (u) − Ga (u)
Introduction 153
is smooth and bounded but not compactly supported in R+ in con-

trast to our previous practice. In this context let us record that the
Selberg/Harish-Chandra transform of k(u) is (use Theorems 1.14-1.17)
1 2 −1 1 2 −1
2 2
(10.2) h(r) = s− +r − a− +r
2 2
of which the Fourier transform is
(10.3) g(x) = (2s − 1)−1 e−|x|(s−1/2) − (2a − 1)−1 e−|x|(a−1/2) .
As a matter of fact we shall carry our computations of particular

components of the kernel for any k, h, g which satisfy adequate growth
conditions. Yet, to be precise we cannot utilize these general computa-
tions until the convergence of many series and integrals involving the
spectrum and the length of closed geodesics is established. In order to
proceed without extra work we are going to stick in the beginning to
the particular functions given above with a > s > 1 so that k, h, g are
positive and small,
(10.4) 0 < k(u) (u + 1)−s ,

(10.5) 0 < h(r) (|r| + 1)−4 ,
(10.6) 0 < g(x) e−|x|/2 .
Hence both series for K(z, z), namely

X
(10.7) k(u(z, γz))
γ∈Γ
and
X X 1 Z +∞ 1 2
2

(10.8) h(tj ) |uj (z)| + h(r) Ea (z, + ir) dr
j a
4π −∞ 2
converge absolutely and uniformly on compacta.
As a by-product of working with these particular test functions on

the side we shall release sufficient estimates to validate the convergence
in the general case. From this point on we no longer assume the special
case. Nevertheless, since it is handled without extra effort, we do not
quit the resolvent kernel because it will be fundamental in the theory
of Selberg’s zeta-function anyway.
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.
Governed by different goals in some cases we shall elaborate more

than one expression for the same thing. We would like to have only the
Fourier pair g, h present in the final formulation of the trace formula.
The original kernel function k must go into hiding, but, of course, it
can always be found in the Selberg/Harish-Chandra transform.
10.2. Computing the spectral trace.
Integrating (10.4) over F (Y ) termwise we get

X Z
Y
Tr K = h(tj ) |uj (z)|2 dµz
j F (Y )
+∞
1 Ec (z, 1 + ir)2 dµz dr .
Z XZ Y
+ h(r)
4π −∞ c F (Y ) 2
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),
EcY (z, 1 + ir)2 dµz = Tr hE Y (·, 1 + ir), t E Y (·, 1 + ir)i

XZ
c F 2 2 2
(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
because Φ(s) is bounded in 1/2 ≤ Re s ≤ 1/2 + ε for a small ε. Here
1 1
Z
(10.10) r−1 h(r) dr = − h(0) .
2πi 2
Im r=ε
To see this we move the integration downwards to Im r = −ε passing a

simple pole at r = 0 of residue h(0). By the symmetry h(r) = h(−r)
the lower horizontal line integral is equal to minus the upper one which
therefore is equal to half of the residue. We end up with
1 1
I(Y ) = Φ( ) h(0) + O(Y −2ε ) .
4 2
Finally, summing h(tj ) over the point spectrum as well as integrating

other parts against h(r) we conclude the following inequality for the
truncated trace
Z +∞
Y
X 1 −ϕ0 1
Tr K < h(tj ) + ( + ir) h(r) dr
4π −∞ ϕ 2
(10.11) j
1 1
+ h(0) Tr Φ( ) + g(0) h log Y + O(Y −ε ) .
4 2
Here for the particular k(u) given by (10.1) we have explicitly
g(0) = (2s − 1)−1 − (2a − 1)−1 ,

1
h(0) = (2s − 1)−2 − (2a − 1)−2 ,
4
1 1 1 1 1 1
h(tj ) = − − − .
2s − 1 s − sj 1 − s − sj 2a − 1 a − sj 1 − a − sj
In order to get a lower bound for Tr Y K we must estimate and sub-

stract from the upper bound (10.11) the truncated traces over cuspidal
zones Fa (Y ). If uj is a cusp form we infer from (8.3) that
Z
|uj (z)|2 dµz |sj | Y −2 .
Fa (Y )
If uj is the residue of an Eisenstein series at sj with 1/2 < sj ≤ 1 we

infer from (8.4) that
Z
|uj (z)|2 dµz Y 1−2sj .
Fa (Y )
From both estimates we conclude using (7.11) that

X Z
h(tj ) |uj (z)|2 dµz Y −2ε .
j
F \F (Y )
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
Hence, we infer from (8.14) that

R
1
Z Z
(10.12) |EcY (z, + ir)|2 dµz dr R3 Y −1 .
−R 2
Fa (Y )
Then combining with (10.5) we conclude that

+∞
1 1
Z X Z
h(r) |EcY (z, + ir)|2 dµz dr Y −1 .
4π −∞ c
2
F \F (Y )
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.
As a by-product of the work done in this section we state the

following estimate
R
1 −ϕ0 1
Z
(10.13) MΓ (R) = ( + ir) dr R2 .
4π −R ϕ 2
To see this we integrate (7.10) over F (Y ) with Y ≈ R and apply (10.9)

together with (10.12).
10.3. Computing the trace for parabolic classes.
As indicated in the introduction we shall compute the geometric

traces for each conjugacy class separately. We begin with the parabolic
motions since they require special care. There are h primitive parabolic
conjugacy classes, one for each class of equivalent cusps. The primitive
class Ca , say, attached to cusps equivalent with a consists of generators
of the stability groups of these cusps. Every parabolic conjugacy class
is obtained as a power C = Ca` for some a and ` 6= 0. Let γ = γa` , where
γa is the generator of Γa so the centralizer Z(γ) = Z(γa ) = Γa is the
stability group. By the unfolding technique the truncated trace of the
class C evolves into
Z
Y
Tr KC = k(z, γz) dµz
Z(γ)\H(Y )
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
To continue the computation we first sum over `, getting

Z +∞ X
2 k(u) u−1/2 `−1 du
(2Y )−2 √
1≤`<2Y u
Z +∞ √
k(u) u−1/2 log 2Y u + γ + O(u−1/2 Y −1 ) du

=2
(2Y )−2
= L(Y ) + O(Y −1 log Y )
where Z +∞ √
k(u) u−1/2 log 2Y

L(Y ) = 2 u + γ du .
0
And with Y replaced by Y 0 we only get O(Y 0 ). Hence we conclude that

all parabolic motions having equivalent fixed points yield
X
(10.14) Tr Y KC = L(Y ) + O(Y −1 log Y ) .
`
C=Ca
Computing the trace for parabolic classes 159
It remains to evaluate L(Y ). We split L(Y ) into

Z +∞
(10.15) L(Y ) = g(0) (log 2Y + γ) + k(u) u−1/2 log u du
0
where the first term is obtained by (see (1.62))
+∞
1
Z
k(u) u−1/2 du = q(0) = g(0) .
0 2
To the second term we apply (1.64) getting
+∞
−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
upon changing v into (sinh(r/2))2 . Collecting the above evaluations we

arrive at
+∞
r
Z
(10.16) L(Y ) = g(0)(log Y + γ) − log(sinh ) dg(r) .
0 2
If one prefers to have an expression in terms of h rather than g, we

supply the formula
+∞
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 −∞
where
∞
Γ0 X 1 1
(10.18) ψ(s) = (s) = −γ − − .
Γ n=0
n + s n + 1
For the proof we write
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
With these ingredients one derives (10.17) straightforwardly except for

the term −h(0)/4 which comes out from (10.10). Combining (10.17)
and (10.16) we get
+∞
Y 1 1
Z
(10.20) L(Y ) = g(0) log + h(0) − h(t) ψ(1 + it) dt .
2 4 2π −∞
For the particular g(r) given by (10.3) we get immediately from

(10.16) and (10.19) that
1
L(Y ) = (2s − 1)−1 (log Y − ψ(s + ) − log 2) + (2s − 1)−2
(10.21) 2
− (the same for s = a) .
Before proceeding to the non-parabolic conjugacy classes let us

observe that the leading terms in (10.14) and (10.11) coincide. After
substracting g(0) h log Y what is left on the spectral side converges
to a constant as Y → ∞. This proves, by the positivity of k(u) on
the geometric side, that the remaining partial kernels associated with
the non-parabolic classes are of trace type individually and in total.
Therefore we can simplify the job by computing the actual traces which
are the limits of the truncated ones.
Computing the trace for hyperbolic classes 161
10.4. Computing the trace for the identity motion.
The identity motion forms a class by itself, C = {1}. Thus KC (z, w)

= k(z, w) and
Z
(10.22) Tr KC = k(z, z) dµz = k(0) |F |
F
where |F | is the area of a fundamental domain. By (1.64’) we have
+∞
1
Z
(10.23) k(0) = r tanh(πr) h(r) dr .
4π −∞
For the particular h(r) given in (10.2) we compute directly that

k(0) = lim Gs (u) − Ga (u)
u→0
Z 1
1 ξ(1 − ξ) s−1 ξ(1 − ξ) a−1 dξ
= lim −
u→0 4π 0 ξ+u ξ+u ξ+u
(10.24)
Z 1
1
(1 − ξ)s−1 − (1 − ξ)a−1 ξ −1 dξ

=
4π 0
∞
1 X 1 1 1
= − = ψ(a) − ψ(s) .
4π n=0 n + s n + a 4π
10.5. Computing the trace for hyperbolic classes.
The primitive hyperbolic conjugacy classes in Γ are the most fasci-

nating of all. Following Selberg we denote such a class by P displaying
its resemblance to prime ideals in number fields. Let C = P ` . Choose
γP ∈ P and γ = γP` ∈ C. Then Z(γ) = Z(γP ) and
Z
Tr KC = k(z, γz) dµz .
Z(γP )\H
By conjugation in G we send γP to ±A. The resulting motion acts

simply as a dilation z 7→ pz by a positive factor p 6= 1. To fix nota-
tion suppose that p > 1 or else change P into P −1 (this change only
reverses the counting of the classes C = P ` ). Then log p is the hy-

perbolic distance of i to pi, thus also the distance of z to γP z for any
z on the geodesic joining the fixed points of γP . Since γP maps the
geodesic joining its two fixed points in R̂ into itself (not identically),
this geodesic closes on the surface Γ\H on which the points z, γP z are
the same. Thus log p is just the length of the closed geodesic multiplied
by the winding number (the geodesic segment joining z with γP z in
the free space may wind itself on Γ\H a finite number of times). We
shall denote p = N P and call it the norm of P (it does not depend on
representatives in the conjugacy class). The norm can be expressed in
terms of the trace
Tr P = N P 1/2 + N P −1/2 .
After conjugation in G the centralizer becomes a cyclic group gen-

erated by the dilation z 7→ pz so its fundamental domain is the hori-
zontal strip 1 < y < p. Hence we obtain
Z p Z +∞
Tr KC = k(z, p` z) dµz .
1 −∞
Furthermore, putting 2d = |p`/2 − p−`/2 | we continue the computation

as follows
p +∞
d|z| 2 −2
Z Z
Tr KC = ) y dx dy k (
1 −∞ y
Z p Z +∞
−1
= y dy k(d2 (x2 + 1)) dx
1 −∞
(10.25) +∞
log p k(u) log p
Z
= √ du = q(d2 )
d d2 u−d 2 d
log p p log p
= g 2 log( d2 + 1 + d) = g(` log p)
2d 2d
= |p`/2 − p−`/2 |−1 g(` log p) log p .
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
Tr KC = (2s − 1)−1 (1 − p−|`| )−1 p−|`|s log p

(10.26)
Computing the trace for elliptic classes 163
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
10.6. Computing the trace for elliptic classes.
The idea is the same but computations are somewhat harder. We

denote by R a primitive elliptic conjugacy class in Γ. There is only a
finite number of these. Let m = mR > 1 be the order of R. Any elliptic
class having the same fixed points as R is C = R` with 0 < ` < m.
Conjugating R in G one can assume the representative to be k(θ) where
θ = θR = πm−1 ; this acts as a rotation of angle 2θ at i. Since it
generates the centralizer, the fundamental domain of that centralizer is
a hyperbolic sector of angle 2θ at i, say S. Therefore, we have
1
Z Z
Tr KC = k(z, k(θ`)z) dµz = k(z, k(θ`)z) dµz
S m H
since it takes m images of S to cover H exactly (except for a zero

measure set). We shall continue computation in the geodesic polar
coordinates z = k(ϕ) e−r i where ϕ ranges over [0, π) and r over [0, +∞)
(see Section 1.3). Since k(θ`) commutes with k(ϕ), we get
+∞
π
Z
Tr KC = k(e−r i, k(θ`) e−r i) (2 sinh r) dr .
m 0
By u(z, k(θ)z) = (2y)−2 |z 2 + 1|2 sin2 θ this yields formulas in terms of

k(u)
+∞
π
Z
k (sinh r sin θ`)2 (2 sinh r) dr

Tr KC =
m 0
Z +∞
π k(u) du
(10.28) = p
m sin θ` 0 u + sin2 θ`
π +∞ π`
Z
= k u sin2 (u + 1)−1/2 du .
m 0 m
These are nice and practical expressions; nevertheless, we continue com-

puting since our strategy is to remove k from the scene. Applying (1.64),
by partial integration, we get (for a > 0)
Z +∞
k(u) (u + a2 )−1/2 du
0
1 +∞ 0
Z Z v
−1/2
=− q (v) (v − u)(u + a2 ) du dv
π 0 0
Z v/(v+a2 )
1 +∞ 0
Z
−1/2
=− q (v) u(1 − u) du dv
π 0 0
a +∞
Z
= q(v) (v + a2 )−1 v −1/2 dv
π 0
Z +∞
a g(r) cosh(r/2)
= dr
2π 0 sinh2 (r/2) + a2
by changing v = sinh2 (r/2). For a = sin α > 0 this yields

Z +∞
sin α +∞ g(r) cosh(r/2)
Z
2 −1/2
k(u) (u + sin α) du = dr ,
0 π 0 cosh r − cos 2α
and taking α = π`m−1 we conclude that
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)
π` −1 +∞ cosh π(1 − 2`/m)r

Z
(10.30) Tr KC = 2 m sin h(r) dr .
m −∞ cosh πr
Now let us apply (10.29) for g from (10.3). We begin by an appeal

to the following formula
Z +∞ ∞
−µx −1 2 X sin kt
e (cosh x − cos t) dx =
0 sin t µ+k
k=1
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
Next we exploit the periodicity to break the summation into residue

classes modulo m as follows
∞
X X π` X
(s + k + mn)−1 − (1 + k + mn)−1 .

= sin(2k + 1)
m n=0
k 0≤k<m
Here we have borrowed terms for s = 1 to produce convergence at no

cost because
X π`
sin(2k + 1) = 0.
m
0≤k<m
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
− the same for s = a .

Finally we sum over 0 < ` < m to compute the total trace of all elliptic
classes C = R` which have common fixed points mod Γ. It follows from
the identity
X `n sin(2k + 1)π`/m
e =
m sin π`/m
|n|≤k
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
Another interesting transformation is offered by the identity

1 X s + k
ψ = ψ(s) − log m
m m
0≤k<m
(see [Gr-Ry, 8.365.6, p. 945]). It leads to
X
Tr KC = m−1 log m − ψ(s) + (2s − 1)−1 Rm (s) ,

(10.33)
C=R`
where
X s + k
(10.34) Rm (s) = m−2 (2s + 2k − m) ψ .
m
0≤k<m
The key point in the last arrangement is that Rm (s) is meromorphic
in the whole complex s-plane with only simple poles at non-positive
integers −d of residue 2[d/m] + 1 which is a positive integer. These
properties are vital for constructing Selberg’s zeta-function (the residue
is also linked to the dimension of a certain space of differential forms
on Γ\H, but we do not dwell on this here).
10.7. Trace formulas.
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).
Theorem 10.1 (Resolvent Trace Formula). Let a > 1 and Re s > 1.

We have
X 1 1
−
j
(s−1/2)2 +t2j (a−1/2)2 +t2j
Z ∞
−ϕ0 1

1 1 1
+ − +ir) dr
4π −∞ (s−1/2)2+r2 (a−1/2)2+r2 ϕ 2
(10.35)
1 1 h 1
= h−Tr Φ − ψ s+ +log 2
(2s−1)2 2 2s−1 2
∞
|F | 2 X X log p
− ψ(s) +
2π 2s−1 ps+k −1
P k=0
1 X X 2k+1−m s+k
+ ψ
2s−1 m2 m
R 0≤k<m

Trace formulas 167
Remarks. In the above formula each term of the discrete spectrum

is counted with the multiplicity of the eigenvalue λj = t2j + 1/4, h is
the number of primitive parabolic classes (= the number of inequivalent
cusps), P ranges over primitive hyperbolic classes of norm p = N P > 1
and R ranges over primitive elliptic classes of order m = mR > 1.
These terms come from (10.11), (10.21), (10.24), (10.27) and (10.32),
respectively. A more general Resolvent Trace Formula is given by J.
Fischer [Fi].
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
The above observations permit us to construct the trace formula for a

general pair h, g.
Theorem 10.2 (Selberg’s Trace Formula). Suppose h satisfies the

conditions (1.63) and let g be the Fourier transform of h. Then
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 −∞
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).
10.8. The Selberg zeta-function.
In connection with the trace formula A. Selberg (see [Se2]) has

introduced a zeta-function which in many ways mimics the L-functions
of algebraic number fields. As in classical cases the zeta-function is
built with various local factors. We define
∞
YY
(10.38) ZΓ (s) = (1 − p−s−k ) if Re s > 1
P k=0
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
by an appeal to (10.33). It has been already observed after (10.34) that

Rm (s) has integral residues. The second part has poles at non-positive
integers (see (10.18)) with residue
|F | X 1
+ = 2g − 2 + h + ` ∈ Z
2π m
m
The Selberg zeta-function 169
by the Gauss-Bonnet formula (2.7) where g is the genus of Γ\H, h is the

number of parabolic generators (cusps) and ` that of the elliptic ones.
By virtue of the above properties we can define with no ambiguity a
meromorphic function
Z s Z0
F (s) = exp (u) du
a Z
where the integration goes along any curve which joins a with s avoiding
poles. Since Z(s) = Z(a) F (s) this proves
Theorem 10.3. The Selberg zeta-function Z(s) defined for Re s > 1 by

(10.38) has a meromorphic continuation to the whole complex s-plane.
In the half plane Re s ≥ 1/2 it is holomorphic and has zeros at the
points sj and sj of order equal to the dimension of the λj -eigenspace
except for s = 1/2, where Z(s) has a zero or pole of order equal to twice
the dimension of the (1/4)-eigenspace minus the number of inequivalent
cusps.
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]).
If you wish, the Selberg zeta-function satisfies an analogue of the

Riemann hypothesis. However, the analogy with the Riemann zeta-
function is superficial. First of all it fails badly when it comes to de-
velopment into Dirichlet’s series. Furthermore, the functional equation
(10.40) resists any decent interpretation as a kind of Poisson’s summa-

tion principle. Nevertheless, modern studies of Z(s) have caused a lot
of excitement in mathematical physics (see [Sa1]). Now it seems that
the dream of Hilbert and Pólya of connecting the zeros of the Riemann
zeta-function with eigenvalues of a self-adjoint operator is a reality in
the context of Z(s).
10.9. Asymptotic law for the length of closed geodesics.
Perhaps the most appealing application of the Selberg Trace For-

mula is the evaluation of the length of closed geodesics in the Riemann
surface Γ\H. Let us begin with a simple test function
x
g(x) = 2 (cosh ) e−2δ cosh x
2
where 0 < δ ≤ 1. Its Fourier transform is equal to

Z +∞
cosh(sx) + cosh(1 − s)x e−2δ cosh x dx

h(t) = 2
0
= 2 Ks (2δ) + 2 K1−s (2δ) = Γ(s)δ −s + O(δ −1/2 |Γ(s)|) ,
if 1/2 ≤ Re s ≤ 1; clearly, h(t) satisfies the conditions (1.63). All

terms in the trace formula (10.37) contribute at most O(δ −1/2 ) except
for the points of the discrete spectrum with 1/2 < sj ≤ 1 and the
primitive hyperbolic classes. A primitive hyperbolic class P of norm
p > 1 contributes
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
Theorem 10.4. For any δ > 0,

X X
(10.41) e−δp log p = Γ(sj ) δ −sj + O(δ −1/2 ) ,
P 1/2<sj ≤1
the implied constant depending on the group Γ.

Asymptotic law for the length of closed geodesics 171
Next, let us try another test function g(x) = 2 (cosh(x/2)) q(x),

where q(x) is even, smooth, supported on |x| ≤ log(X + Y ) such that
0 ≤ q(x) ≤ 1 and q(x) = 1 if |x| ≤ log X. The parameters X ≥ Y ≥ 1
will be chosen later. For s = 1/2 + it in the segment 1/2 < s ≤ 1 we
have
Z +∞
esx + e(1−s)x q(x) dx = s−1 X s + O(Y + X 1/2 ) ,

h(t) =
−∞
and for s on the line Re s = 1/2 we get by partial integration that
h(t) |s|−1 X 1/2 min{1, |s|−2 T 2 }
where T = XY −1 . Hence the discrete spectrum contributes

X X
h(tj ) = s−1
j X
sj
+ O(Y + X 1/2 T ) ,
j 1/2<sj ≤1
and the continuous spectrum contributes to the error term above. On

the geometric side the identity motion contributes
+∞
|F |
Z
h(t) tanh(πt) t dt X 1/2 T .
4π −∞
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
We shall clean this formula by exploiting the positivity of terms. First

substracting (10.42) from that for X + Y in place of X we deduce
X
log p Y + X 1/2 T .
X<p<X+Y
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
Theorem 10.5 (Selberg). For X ≥ 1,

X X
(10.43) log p = s−1
j X
sj
+ O(X 3/4 ) ,
p≤X 1/2<sj ≤1
where p = N P denotes the norm of primitive hyperbolic classes, i.e.

log p is the length of a closed geodesic in Γ\H counted with multiplicity.
The implied constant depends on the group.
An alternative approach to prove (10.43) makes use of analytic

properties of the Selberg zeta-function ZΓ (s). Since ZΓ (s) satisfies the
Riemann hypothesis (all zeros and poles other than 1/2 < sj ≤ 1 are on
the line Re s = 1/2 or to the left) one should ask if (10.43) holds true
with smaller error term. Here it is not clear if O(x1/2+ε ) is possible as
it is for the Prime Number Theorem assuming the classical Riemann
hypothesis. In the case of the modular group W. Luo and P. Sarnak [Lu-
Sa] have recently established by refining the arguments of H. Iwaniec
[Iw3] that
X
(10.44) log p = X + O(X 7/10+ε ) .
p≤X
Any improvement upon the exponent 3/4 is meaningful; it amounts

to showing a uniform distribution of the zeros sj on the critical line.
In the proof of (10.44) Luo and Sarnak appeal to the Weil bound for
Kloosterman sums (therefore, indirectly to the Riemann hypothesis for
curves) and to the recent estimate (8.46) of Hoffstein and Lockhart.
The result has connection with indefinite binary quadratic forms since
the length of closed geodesics on SL2 (Z)\H are given by 2 log εd with
multiplicity h(d), where εd is the fundamental unit and h(d) is the class
number (cf. [Sa4]).
Chapter 11
The distribution of eigenvalues
In this chapter we consider some fundamental issues of the Rie-

mann surface Γ\H concerning the eigenvalues of the Laplace operator.
They are the subject of study in various fields of mathematics and
physics. Hence, there are diverse techniques ranging from topology,
differential geometry, partial differential equations through automor-
phic forms and number theory. Among these the Selberg trace formula
is a conventional tool for establishing asymptotics in the spectrum. We
shall apply the trace formula to prove Weyl’s law and then tackle the
big problem of small eigenvalues.
11.1. Weyl’s law.
The interplay between geometric invariants of the Riemann surface

Γ\H and the spectrum of the Laplacian is a wonderful gift of nature. Its
content has been phrased by M. Kac [Kac] in the question: “Can one
hear the shape of a drum ?”. M.-F. Vignéras [Vig2] has constructed
strictly hyperbolic groups Γ1 , Γ2 for which the surfaces Γ1 \H, Γ2 \H
are isospectral but not isometric (more precisely they have the same
spectrum with multiplicities, but the groups are not conjugate), thus
173
174 The distribution of eigenvalues
the answer to the question is not always affirmative. Whatever the

answer for a given group, it is easy to argue that reconstructing the
surface out of the spectrum is not a very practical goal since the set
of eigenvalues is impossible to examine with the required precision.
What is possible are mainly statistical results. In practice, estimates
and asymptotics involving the spectrum in suitable segments are of
considerable value.
First we wish to evaluate the counting function of the eigenvalues

λj = 1/4 + i t2j in the discrete spectrum
NΓ (T ) = #{j : |tj | ≤ T } .
By analogy in the trace formula, the continuous spectrum is accounted
for by the integral
Z T
1 −ϕ0 1
MΓ (T ) = + it dt .
4π −T ϕ 2
We shall shed more light upon MΓ (T ) later. So far we have only shown
that each part of the spectrum cannot be immensely large, in the sense
that
NΓ (T ) T 2 ,
MΓ (T ) T 2 ,
for any T ≥ 1. These bounds are rather cheap by-products of Bessel’s
inequality (see (7.11), (10.13)) yet T 2 is the right order of magnitude.
In order to get a lower bound or an asymptotic formula we need the
complete spectral decomposition of an automorphic kernel, which in-
evitably forces us to treat the quantities NΓ (T ), MΓ (T ) together.
2
Let us apply the trace formula for the Fourier pair h(t) = e−δt
2
and g(x) = (4πδ)−1/2 e−x /4δ , where δ is a small positive parameter.
The identity motion contributes
|F | +∞ −δt2 |F |
Z
e tanh(πt) t dt = + O(1) .
4π −∞ 4πδ
The hyperbolic and the elliptic motions contribute a bounded quantity.
The parabolic motions contribute
Z +∞
−1/2 h 2 1 1
−(4πδ) h log 2 − e−δt ψ(1 + it) dt + Tr I − Φ .
2π −∞ 4 2
Weyl’s law 175
By (B.11) we get
ψ(1 + it) + ψ(1 − it) = log(1 + t2 ) + O((1 + t2 )−1 ) ;
whence the above integral is equal to (up to a bounded term)

+∞ Z +∞
1 t −1/2
Z
−δt2
2 e log t dt = √ e−t log t dt
0 2 δ 0 δ
1 1 1
= √ Γ0 −Γ log δ
2 δ 2 2
√
π
= √ − γ − log 4δ .
2 δ
From the above computations we conclude the following
Theorem 11.1. For any δ > 0 we have

+∞
1 −ϕ0 1
Z
−δt2j 2
X
+ it e−δt dt

e +
j
4π −∞ ϕ 2
(11.1)
|F | h log δ γh
= + √ − √ + O(1) ,
4πδ 4 πδ 4 πδ
where O(1) is bounded by a constant depending on the group.
By a Tauberian argument one deduces
Corollary 11.2. As T → +∞ we have
|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]).
It is plain from (11.2) that the total spectrum is quite large. As

we cannot count separately the point and the continuous spectra, it
becomes an open question which one is larger (if either) in order of
magnitude. The only answers we can get so far are for special groups
by showing that MΓ (T ) is much smaller than NΓ (T ). For congruence
groups Selberg has proved that
(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.
Nothing like (11.4) is known in general. Recent intensive studies,

initiated by R. S. Phillips and P. Sarnak [Ph-Sa], then expanded by S.
Wolpert [Wo], all indicate that the opposite situation is more likely to
be true for generic groups, i.e. the cuspidal spectrum should be small.
W. Luo [Lu] has supplied very convincing and concrete results. Also
the numerical computations by D. Hejhal [He2] tend to support such a
claim. It would be interesting to isolate a finite volume quotient Γ\H
having a cusp but no cusp forms. The Phillips-Sarnak theory is still
not complete. Had one showed that the multiplicities of eigenvalues
(the dimension of eigenspaces) for the group Γ0 (q) were bounded, the
result in [Lu] would have reached the main objective of the theory.
Unfortunately very little is known. By (11.3) one infers the following
bound for the multiplicity of λ:
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
Before devoting greater attention to congruence groups let us reveal

what the quantity MΓ (T ) really stands for in general. First we show
some estimates for the scattering matrix Φ(s). For Re s > 1 the entries
are given by Dirichlet’s series (3.21). Hence, the determinant is also
given by a Dirichlet’s series (do not underestimate the significance of
this fact !)
√ Γ(s − 1/2) h X ∞
ϕ(s) = π an b−2s
n
Γ(s) n=1
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
The local factors are repeated with multiplicities of each sj so that

ϕ∗ (s) is holomorphic in the half-plane Re s ≥ 1/2. We have
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 f (z) is a holomorphic function in |z| ≤ 1 with f (0) = 1, has no

zeros on |z| = 1, and zj ranges over the zeros in |z| < 1 repeated with
multiplicity. By changing z 7→ (s − 1)/s we map the unit disk |z| ≤ 1
onto the half-plane Re s ≥ 1/2. Then after removing a possible zero of
ϕ∗ (s) at s = 1 and normalizing to 1 we conclude that
X0 z − 1
j
− log < +∞

j
z 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
Now we can consider the Hadamard canonical product for ϕ∗ (s)

Y s − 1 + s̄j
ϕ∗ (s) = eg(s)
j
s − sj
where g(s) is a polynomial. Here a pole sj is matched with the zero

1 − s̄j so the product converges by virtue of (11.7). Since ϕ∗ (s) is
bounded in Re s ≥ 1 + ε, g(s) is constant. Therefore
ϕ∗ 0 X 1 1
(11.8) − (s) = − .
ϕ∗ s − sj s − 1 + s̄j
βj <1/2
Adding the poles in 1/2 < sj ≤ 1 we conclude that
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,
whence, −ϕ∗ 0 (s)/ϕ∗ (s) ≥ 0 and
ϕ0
(11.10) − (s) > 2 log b1 + O(|s|−2 ) .
ϕ
The residual spectrum and the scattering matrix 179
Finally, by Cauchy’s Theorem, (11.9) reveals that the integral M Γ (T )

is approximately equal to the number of poles of ϕ(s) on the left of
the critical line of height less than T up to an error term O(T ). If
the number of such poles is of order of magnitude T 2 , then most of
them must concentrate along the critical line Re s = 1/2 because of
(11.7). Selberg [Se2] has established several stronger results of this
kind. For the modular group, ϕ(s) has poles at the complex zeros of
ζ(2s); therefore, according to the Riemann hypothesis these poles are
on the line Re s = 1/4 .
11.2. The residual spectrum and the scattering matrix.
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.
Theorem 11.3. The congruence groups have no residual spectrum

besides the obvious point s0 = 1.
This result can be proved in a number of ways. Clearly it is enough

to consider the principal congruence group Γ(N ). In this case one can
compute the scattering matrix Φ(s) explicitly in terms of the Riemann
zeta-function and Dirichlet L-series and from this the claim is deduced.
A complete computation is quite involved. However, since all poles
appear on the diagonal of Φ(s) (see Theorem 6.10), one can save work
by computing only the diagonal entries
√ Γ(s − 1/2) X −2s
ϕaa (s) = π c Saa (0, 0; c) .
Γ(s) c
For the cusp a = ∞ it gives

√ Γ(s − 1/2) −1 X
ϕ∞∞ (s) = π N ϕ(c) c−2s
Γ(s)
c≡0 (mod N 2 )
√ Γ(s − 1/2) ζ(2s − 1) ϕ(N ) Y 1 −1
= π 1 − 2s ;
Γ(s) ζ(2s) N 4s p
p|N
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.
Another approach proceeds directly from the Eisenstein series

X
Ea (σa z, s) = (Im τ z)s .
−1
τ ∈Γ∞ \σa Γσa
Using a parametrization of the above cosets (one does not need to be

very explicit) the series splits into a finite number of Epstein zeta-
functions X
Q(m, n)−s
m,n
where Q is a positive definite quadratic form with rational coefficients

and m, n range over co-prime integers in an arithmetic progression. Ev-
ery such series has an analytic continuation to Re s > 1/2 by Poisson’s
summation with at most a simple pole at s = 1.
The scattering matrix Φ(s) has been computed completely in some

cases. For instance, if Γ = Γ0 (p) and p is prime we derive easily,
exploiting the computations following Theorem 2.7, that

ϕ∞∞ (s) ϕ∞0 (s)
(11.11) Φ(s) = = ϕ(s) Np (s)
ϕ0∞ (s) ϕ00 (s)
where
ps − p1−s

2s −1 p−1
Np (s) = (p − 1)
p − p1−s
s
p−1
and ϕ(s) is the scattering matrix for the modular group given by (3.24).
This can be written in the symmetric fashion (see (3.27) and (3.28))
Φ(s) = M (s)−1 M (1 − s)
where
ps

−s 1
M (s) = π Γ(s) ζ(2s) .
ps 1
The above results are consistent with these due to D. Hejhal [He1] (see
also [Hu1]). For the group Γ0 (N ) with N squarefree Hejhal provides us
with
(11.12) Φ(s) = ϕ(s) ⊗ Np (s) .
p|N
Small eigenvalues 181
11.3. Small eigenvalues.
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).
A great deal of research concerns compact, smooth Riemann sur-

faces. Suppose F is one of those having curvature −1. By the uni-
formization theorem F can be represented as a quotient Γ\H for some
hyperbolic group of signature (g; 0; 0) with g ≥ 2. In this case the
Gauss-Bonnet theorem asserts that |F | = 4π(g − 1), and Weyl’s law
implies λj → 4π|F |−1 j as j → ∞. R. Schoen, S. Wolpert and S. T.
Yau showed that λ2g−3 can be as small as one likes. In the other di-
rection P. Buser proved that λ4g−2 is never exceptional, i.e. the lower
bound
1
(11.13) λ4g−2 ≥
4
always holds true while λ4g−3 can be exceptional for arbitrary g ≥ 2.
It is remarkable that the first eigenvalue is bounded above by an
absolute constant; for example, we obtain (essentially by variational
calculus)
g+1
(11.14) λ1 ≤ 2 ≤ 6,
g−1
which is due to P. C. Yang and S. T. Yau. It is also possible to estimate

λ1 in terms of the diameter d of F alone, namely it holds that
d −2 1 4π 2
4π sinh < λ1 < + 2 ,
2 4 d
which is due to P. Buser and S.-Y. Cheng respectively.
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.
For congruence groups it is a quite different story.
Conjecture (A. Selberg [Se1]). There is no exceptional spectrum for

congruence groups, i.e. one has the bound
1
(11.16) λ1 ≥ .
4
The conjecture is known to be true for groups of small level (see

M. N. Huxley [Hu2]). In the case of the modular group H. Maass [Ma]
and W. Roelcke [Ro] have established (by slightly different methods)
somewhat sharper bounds. We apply Roelcke’s argument to prove the
following
Theorem 11.4. Let Γ be a subgroup of finite index in the modular

group. Denote by q the maximal width of cuspidal zones. Then the
cuspidal eigenvalues satisfy the lower bound
3 π 2
(11.17) λ≥ .
2 q
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
where F is any fundamental domain of Γ. Choose F = ∪γν F 0 , where

F 0 is the standard fundamental polygon for the modular group
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
where qν is the width. Choose B divisible by all qν . We obtain

Z +∞ Z B X 2πnZ +∞ 2
2
|(y ∇)u(γν z)| dµz ≥ B √ cνn (y) dy

√
3/2 n6=0 qν

3/2 0
π 2 Z +∞ X
≥3 B √ |cνn (y)|2 y −2 dy
q 3/2 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
Summing over the coset representatives γν we get 2λB ≥ 3π 2 q −2 B,

whence (11.17).
For the modular group Theorem 11.4 yields λ1 ≥ 3π 2 /2 = 14.80 . . . ,

but the true value is λ1 = 91.14 . . . according to numerical computa-
tions of D. Hejhal (see [He1, Appendix C]). If the cuspidal widths are
≤ 7, we get from (11.7) that λ1 > 3/11. In particular this proves
Corollary 11.5. There is no exceptional spectrum for congruence

groups of level ≤ 7.
The Selberg eigenvalue conjecture is also true for some non-congru-

ence groups. P. Sarnak [Sa3] has proved it for all Hecke triangle groups
using the Courant nodal line technique (see also M. N. Huxley [Hu2]).
The Roelcke technique is wasteful. It cannot work for groups of

very large level since there are a lot of eigenvalues near to 1/4. Indeed
by employing the trace formula for the group Γ0 (q) with a suitable test
function we can show that
1 1
≤ λj < + c (log q)−2 |F | (log q)−3

(11.18) # j:
4 4
where c is a large constant. The upshot of this is that the dimension
of the eigenspace λ = 1/4 satisfies m(1/4) |F |(log |F |)−3 so it is
considerably smaller than the volume.
A. Selberg [Se1] has established a remarkable lower bound.
Theorem 11.6. For any congruence group we have

3
(11.19) λ1 ≥ .
16
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.
A better bound was established in the summer of 1994 by W. Luo,

Z. Rudnick and P. Sarnak [Lu-Ru-Sa]
21
(11.19’) λ1 ≥
100
by using properties of the Rankin-Selberg convolution L-functions on
GL3 . One can show by arguments entirely within the GL2 theory a
slightly weaker bound λ1 ≥ 10/49.
Density theorems 185
11.4. Density theorems.
In practice a few exceptional eigenvalues do not cause a problem

but a larger number of these may ruin results. Therefore it is important
to investigate the distribution of the points sj in the segment 1/2 <
sj < 1 from a statistical point of view. A natural way would be to
count all points sj with certain weights such that the larger sj is, the
heavier the weight that is attached to sj . It occurs that the following
type inequality
X
(11.20) |F |c(sj −1/2) |F |1+ε
1/2<sj <1
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
sin tj L e|tj |L e|tj |L

≥ ≥ .
tj L 2|tj |L + 1 L+1
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
From the always-present point s0 = 1 we infer a bound for the length

of the shortest closed geodesic; namely, 4L ≤ 2 log(|F | log |F |). Recall
that |F | = 4π(g − 1) by the Gauss-Bonnet formula.
Next we establish a much deeper density theorem for the Hecke

congruence group Γ0 (q). It is relatively easy to derive (11.20) with ex-
ponent c = 2 from the Selberg trace formula (10.37). However our goal
is c = 4. This time it will be an excercise with Bruggeman-Kuznetsov’s
formula (9.12). Let us use the same test function (11.21) with L ≥ 4.
From (9.9), arguing as above we get h0 ≤ πL−3 . To estimate the trans-
form h+ (x) we use the contour integral
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)
uniformly in 3/4 < σ < 1. Hence, choosing σ = 3/4 + (L + log 2nq) −1

we obtain
X sin tj L 4
|νj (n)|2
(11.23) j>0
tj L
L−3 + eL n1/2 (n, q)1/2 q −1 τ (nq) (log 2nq)2
(the continuous spectrum integrals are dropped due to the positivity).
Fixing L in (11.23) and dropping all but one term we infer

Density theorems 187
Corollary 11.6. The Fourier coefficient νj (n) of a normalized cusp

form uj for the group Γ0 (q) satisfies the bound
(11.24) νj (n) λj n1/4 τ (n) log 2n

Choosing n = 1 and L = 4 + log q in (11.23) we obtain another

inequality
X
(11.25) q 4(sj −1/2) |νj (1)|2 τ (q) log6 q .
1/2<sj <1
Hence, the density theorem (11.2) with exponent c = 4 would follow if

we had the lower bound for the first coefficient
(11.26) |νj (1)|2 q −1−ε .
However, such a bound may be false for individual j. Clearly, one can
make νj (1) vanish if the λj -eigenspace has dimension > 1. Therefore,
for (11.26) to be true, a basis of cusp forms must be carefully selected.
To this end choose the Hecke basis as described in Section 8.5. Every
form in the basis is of type v(dz), where v(z) is a newform on Γ0 (r)
with dr|q (all forms are normalized with respect to the inner product
in L(Γ0 (q)\H) regardless of the level). By (8.43) the first Fourier co-
efficient of v(z) satisfies (11.26) (take into account our normalization).
The other forms v(dz) of divisor d 6= 1 do not contribute to (11.25)
yet they have a common Laplace eigenvalue already accounted for from
v(z). Since the missing factor of multiplicity does not exceed τ (q) q ε ,
(11.25) and (11.26) for newforms yield
Theorem 11.7 (Density Theorem). The exceptional eigenvalues for

Γ0 (q) satisfy
X
(11.27) q 4(sj −1/2) q 1+ε .
1/2<sj <1
Equivalently, for any σ ≥ 1/2 we have

# j > 0 : sj > σ q 3−4σ+ε .

(11.28)
The last result displays my favorite estimate for the cardinality of

the empty set, but the empty set does not always look the same !
Chapter 12
Hyperbolic lattice-point problems
The lattice-point problem is described in general as follows. Let X

be a topological space acted on discontinuously by a group Γ. Let D be
a domain in X and let z be a point in X. The problem is to estimate
the number of points of the orbit Γz = γz : γ ∈ Γ which meet
D. For a sufficiently regular domain D one expects that this number
approximates the area of D with respect to a certain measure on X
(the Haar measure if X is the homogeneous space of a Lie group).
A classical example is the Euclidean plane X = R2 acted on by the

group Γ = Z2 of integral translations. In this case we ask how many
integer points are in D ? The elementary method of packing with the
unit square works fine. If D is a disc of radius r = x1/2 centered at the
origin, then one deals with the Gauss circle problem, and the packing
method leads to
X
(12.1) r(m) = π x + O(x1/2 )
m≤x
where r(m) denotes the number of ways that m can be represented as

the sum of two squares
r(m) = # a, b ∈ Z : a2 + b2 = m .

(12.2)
189
190 Hyperbolic lattice-point problems
If D is the area under a hyperbola the problem is known as the Dirichlet

divisor problem. We obtain
X
(12.3) τ (m) = x log x − (2γ − 1) x + O(x1/2 )
m≤x
where τ (m) denotes the number of positive divisors of m

(12.4) τ (m) = # a, b ∈ N : ab = m .
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].
In this chapter we consider the lattice-point problem in the hyper-

bolic plane H. In view of the isoperimetric inequality (1.14) the packing
method is insufficient because the area of a regular domain is compa-
rable with the length of the boundary (the negative curvature causes
the problem). Even if the geometric ideas can be pushed sometimes to
produce an asymptotic, the spectral theorem will show here a lot more
power.
The lattice-point problem on H with respect to a group Γ is nothing

other than a case of evaluating the automorphic kernel K(z, w) for a
properly chosen function k(u). Recall the spectral expansion (7.17).
The points sj of the discrete spectrum in the segment 1/2 < sj ≤ 1 yield
the main term. The remaining points of the discrete and continuous
spectra lie on the critical line Re s = 1/2. The contribution of these
spectra is estimated by means of (7.10) using Cauchy’s inequality. We
obtain in general that
X Z +∞

(12.5) K(z, w) = h(tj ) uj (z) ūj (w) + O (t + 1) H(t) dt
1/2<sj ≤1 0
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 .
Recall that 4u + 2 = eρ + e−ρ where ρ is the hyperbolic distance. In

this case (12.5) will yield
Theorem 12.1. Let Γ be a finite volume group. For X ≥ 2, we have

X Γ(sj − 1/2)
(12.6) P (X) = π 1/2 uj (z) ūj (w) X sj + O(X 2/3 ) .
Γ(sj + 1)
1/2<sj ≤1
Note that the dominant term in (12.6) is attained at s0 = 1 which

corresponds to the lowest eigenvalue λ0 = 0 with constant eigenfunction
u0 (z) = u0 (w) = |F |−1/2 so it contributes π|F |−1 X.
Proof. Naturally, one would like to take k(u) = 1 if u ≤ (X − 2)/4

and k(u) = 0 elsewhere, but this kernel does not yield strong results
Rather, we take k(u) whose graph is
Figure 11. The test function k(u).
with Y to be chosen later subject to X ≥ 2Y ≥ 2. This gives only a

majorization
(12.7) P (X) ≤ K(z, w) .
We let the reader prove that the Selberg/Harish-Chandra trans-

form of k(u) satisfies
Γ(s − 1/2) s
(12.8) h(t) = π 1/2 X + O(Y + X 1/2 ) ,
Γ(s + 1)
if 1/2 < s ≤ 1 where the implied constant depends on s, and
h(t) |s|−5/2 min{|s|, T } + log X X 1/2 ,

(12.9)
if Re s = 1/2 where T = XY −1 and the implied constant is absolute.
Given these evaluations we get by (12.5) that

X Γ(sj − 1/2)
K(z, w) = π 1/2 uj (z) ūj (w) X sj +O(Y +XY −1/2 ) .
Γ(sj + 1)
1/2<sj ≤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).
For special points and groups the hyperbolic lattice-point problem

can be stated in an arithmetic fashion. Let us take z = w = i. Then
for γ = ac db we have 4 u(γi, i) + 2 = a2 + b2 + c2 + d2 . Therefore, if
Γ is the modular group, then P (X) designates the number of integer
points (a, b, c, d) on the hypersurface
(12.10) ad − bc = 1
within the ball
(12.11) a 2 + b2 + c 2 + d2 ≤ X .
In this case Theorem 12.1 together with Corollary 11.5 give
Corollary 12.2. If Γ is the modular group, then
(12.12) P (X) = 3 X + O(X 2/3 ) .
Another interesting case is for the group

a b a ≡ d (mod 2)
(12.13) Γ= ∈ SL2 (Z) : .
c d b ≡ c (mod 2)
This is a subgroup of index 3 in the modular group; in fact Γ is conjugate

−1
to Γ0 (2), namely Γ = 1 −1 Γ0 (2) 1 −1

1 1 . Theorem 12.1 gives
(12.14) P (X) = X + O(X 2/3 ) .
In this case the linear map a + d = 2k, a − d = 2`, b + c = 2m,

b − c = 2n yields integers k, `, m, n, and it transforms (12.10) into
k 2 − `2 − m2 + n2 = 1 and (12.11) into k 2 + `2 + m2 + n2 ≤ X/2. Hence,

it follows that X
P (4x + 2) = r(m) r(m + 1) .
m≤x
Combining this with (12.14) we obtain
Theorem 12.3. For x ≥ 1 we have

X
(12.15) r(m) r(m + 1) = 4 x + O(x2/3 ) .
m≤x
For the modular group the lattice-point problem can be generalized

to count integer points on the hypersurface
(12.16) ad − bc = n
where n is a fixed positive integer. Denote by Pn (x) the number of such

points within the ball (12.11).
Theorem 12.4. For X ≥ n ≥ 1 we have

X
−1 1/3 2/3
(12.17) Pn (X) = 3 d X + O(n X )
d|n
For the proof we need the Hecke operator Tn : A(Γ\H) → A(Γ\H)

defined by (see Section 8.5)
1 X
(12.18) (Tn f )(z) = √ f (τ z)
n
τ ∈Γ\Γn
where Γ is the modular group and

a b
Γn = : a, b, c, d ∈ Z , ad − bc = n .
c d
Observe that Pn (X) counts points of the orbit Γn (z) = {γz : γ ∈ Γn }

in a disc. Indeed, for γ ∈ Γn we have
1 2 X
4 u(γi, i) + 2 = (a + b2 + c2 + d2 ) ≤ .
n n
Therefore, in general the problem boils down to estimating the kernel

X
Kn (z, w) = k(γz, w) .
γ∈Γn
√
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
Theorem 12.5. If n is odd and x ≥ n ≥ 1, then

X X
−1 1/3 2/3
(12.19) r(m) r(m + n) = 4 d x + O(n x ) .
m≤x d|n
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
Spectral bounds for cusp forms
13.1. Introduction.
The eigenfunctions of the Laplace operator on a Riemannian man-

ifold are of great interest for theoretical physicists working in quantum
mechanics. The square-integrable eigenstates are particularly mean-
ingful. How do they behave on high energy levels, that is in the limit
with respect to the eigenvalues ? Do they concentrate onto specific sub-
manifolds, or sets such as closed geodesics when being on distinguished
energy levels, and if so, what is the distribution law for these levels ?
For physicists, if the individual eigenstates behave like random waves,
this is a manifestation of quantum chaos.
A simpler question, yet not easy to answer, is how large the eigen-
states can possibly be in terms of the spectrum. The case of the torus
Z2 \R2 shows that all eigenstates of the standard basis are uniformly
bounded. But this is not true on other manifolds such as the sphere S 2
on which the eigenfunctions given by the Legendre polynomials (spheri-
cal harmonics) take relatively large values at special points (see the con-
cluding remarks in Appendix B.3). However, this phenomenon seems to
be much weaker if the manifold is negatively curved such as the quotient
space Γ\H of the hyperbolic plane modulo a finite volume group.
195
196 Spectral bounds for cusp forms
In this chapter (joint work with P. Sarnak [Iw-Sa]) we demonstrate

how to break the standard upper bound for cusp forms uj on the mod-
1/4
ular group. The standard bound for uj is O(λj ), which follows easily
from the spectral decomposition for a well chosen automorphic kernel.
One can also infer the standard bound in number of other ways and for
general groups (see (8.3’)). In order to strengthen this bound in the
case of the modular group we exploit a bit of arithmetic through the
Hecke correspondence. A conjecture we make about the true size of uj
is connected with an estimate for a certain arithmetic zeta-function on
the critical line and puts the celebrated Lindelöf hypothesis for zeta-
functions into a new perspective. For us (believers in the Riemann
hypothesis, a fortiori, in the Lindelöf hypothesis) this connection insin-
uates that there is considerable chaos in the eigenstates on arithmetic
surfaces (on the modular surface at any rate).
13.2. Standard bounds.
Throughout we assume that {uj (z)} is an orthonormal system of

cusp forms for the modular group which are eigenfunctions of all the
Hecke operators. The Bessel inequality (see Proposition 7.2)
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
shows that the uj ’s are bounded on average. In the other direction

we have proved recently that the uj (z) are not bounded individually;
more precisely, at special arithmetic points (the complex multiplication
points) one has
|uj (z)| (log log λj )1/2
for infinitely many j. We conjecture that
(13.2) |uj (z)| λjε
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
A sharper estimate will follow if we restrict the spectral averaging

to a short interval. For this purpose we need the complete spectral de-
composition of an automorphic kernel (in place of the Bessel inequality)
X
K(z, z) = k(u(γz, z))
γ∈Γ
(13.3) +∞
1 1
Z
X
2
2
= h(tj ) |uj (z)| + h(t) E(z, +it) dt .
j
2π −∞ 2
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
(13.5) k(0) = T + O(1) ,

(13.6) k(u) T 1/2 u−1/4 (1 + u)−5/4 .
Moreover, by Corollary 2.12 one obtains

(13.7) # γ ∈ Γ : 4 u(γz, z) + 2 ≤ X X ,
for any X ≥ 2, the implied constant depending on z. Using these

estimates we evaluate K(z, z) as follows
X
K(z, z) = ν k(0) + k(u(γz, z)) = ν T + O(T 1/2 )
γ∈Γ
γz6=z
where ν = 1, 2, 3 is the order of the stability group of z. Combining

this with the spectral expansion we get
Z +∞
X
2 1 1
h(tj ) |uj (z)| + h(t) |E(z, + it)|2 dt = ν T + O(T 1/2 ) .
j
4π −∞ 2
Hence,
X
(13.8) |uj (z)|2 T .
T <tj <T +1
Recall that the segment (T, T + 1) contains T points tj counted with

multiplicity. This shows that the uj (z) is bounded on average over
segments of constant length. Ignoring all but one term we recover the
standard bound (somewhat differently than the derivation of (8.3’))
1/4
(13.9) |uj (z)| λj .
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.
13.3. Applying the Hecke operator.
In order to get further improvement we use the Hecke operator Tn

as in the proof of Theorem 12.4. From the spectral decomposition we
get
X
n−1/2 Kn (z, z) = h(tj ) λj (n) |uj (z)|2
j
+∞
1 1
Z
+ h(t) ηt (n) |E(z, + it)|2 dt ,
4π −∞ 2
and on the geometric side we get
X X
Kn (z, z) = k(u(γz, z)) = νn k(0) + k(u(γz, z))
γ∈Γn γ∈Γn
γz6=z
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 .
Lemma 13.1. For z = i and X ≥ 2 we have

# γ ∈ Γn : 4 u(γz, z) + 2 ≤ X τ (n) n X log X .
Proof. We are counting the number of solutions to ad − bc = n within

the ball a2 + b2 + c2 + d2 ≤ nX. Use your fingers.
From (13.10), Lemma 13.1, (13.5) and (13.6) we infer an estimate

for the geometric side of Kn (z, z); then combining this estimate with
the spectral decomposition we conclude the following
Proposition 13.2. For z = i and n ≥ 1 we have

X r(n)
(13.11) h(tj ) λj (n) |uj (z)|2 = √ T + O(T 1/2 n3/4+ε )
j
n
the implied constant depending on ε alone.
We have dropped the contribution from the Eisenstein series be-

cause it is easily absorbed by the error term (even more easily one could
drop it later after the positivity is restored).
Since 1 ≤ r(n) ≤ 4 τ (n) nε , the formula (13.11) shows that as n

gets large there exists a considerable cancellation of spectral terms due
to the variation in sign of the Hecke eigenvalue λj (n). This variation
is the key to improve (13.9). Unfortunately, for the same reason, we
cannot drop all but one term to conclude directly a good bound for the
individual cusp form.
13.4. Constructing an amplifier.
We shall overcome the lack of positivity on the spectral side of

(13.11) by means of an amplifier. First using the multiplication rule for
the eigenvalues X mn
λj (m) λj (n) = λj 2
d
d|(m,n)
we generalize (13.11) as follows:

X r(m, n)
h(tj ) λj (m) λj (n) |uj (z)|2 = √ T + O(T 1/2 (mn)3/4+ε )
j
mn
where X mn
r(m, n) = dr .
d2
d|(m,n)
Now multiply this throughout by any complex numbers am , ān and

sum over m, n ≤ N getting
X X 2 X r(m, n)
h(tj ) an λj (n) |uj (z)|2 = T am ān √

j
mn
n≤N m,n≤N
+ O(T 1/2 N 3/2+ε kak21 )
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
This linear form Lj serves to amplify the contribution of the se-

lected eigenvalue λj . We would like to make Lj large; therefore, an =
Constructing an amplifier 201
λj (n) seems to be the obvious choice (no cancellation occurs). One

expects that
X
(13.13) t−ε
j N λj (n)2 tεj N
n≤N
uniformly in j and N with the implied constant depending on ε alone.

The upper bound has been proved in Theorem 8.3, but the lower bound
is still a conjecture. If we accept this conjecture, then Lj t−ε
j N . We
2 2
also have kak2 = Lj and kak1 ≤ N Lj ; whence
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
at the special point z = i. This is a conditional result subject to the

conjectured lower bound of (13.13).
Without any conjecture we have still a good choice, namely
√
λj (p) if n = p ≤ N ,

an =
−1 if n = p2 ≤ N ,
and we put an = 0 otherwise. Since λj (p)2 − λj (p2 ) = 1, the above

choice yields
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
by the upper bound of (13.13) and

X
kak21 = |λj (p)| + 1 ≤ N 1/4 kak2 .

√
p≤ N
Inserting these estimates into (13.12) we get

1/2
|uj (z)|2 tj N −1/2 + tj N 3/2 N ε .

1/4
Choosing N = tj we obtain the following unconditional result
Theorem 13.3. For z = i we have

7/32+ε
(13.15) |uj (z)| λj ,
the implied constant depending only on ε .
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].
13.5. The unique ergodicity conjecture.
The next natural problem to consider after the conjecture (13.2) is

that of equidistribution of |uj (z)|2 and |E(z, 1/2 + it)|2 . In particular,
a basic question is what are the limits of the measures
dµj (z) = |uj (z)|2 dµ(z) ,

(13.17)
dνt (z) = |E(z, 1/2 + it)|2 dµ(z) .
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
Verdière independently in various contexts. About the continuous spec-

trum P. Sarnak has established that
48
(13.19) µt (A) ∼ µ(A) log t , as t → +∞
π
for any Jordan measurable set A ⊂ X = Γ\H. He also gave evidence for
the unique ergodicity conjecture by relating it to the Lindelöf hypothesis
in the theory of L-functions. By means of the Fourier coefficients of
uj the unique ergodicity conjecture requires essentially that there is a
considerable cancellation in the following sums
X
ν̄j (n) νj (n + h)
1≤n≤N
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,
Theorem 13.4R (Luo and Sarnak). Let f ∈ C0∞ (X) be orthogonal to

constants, i.e. X f dµ = 0 . Then
X Z 2
f dµj T 1+ε

tj ≤T X
where the implied constant depends on ε and f .
They also gave estimates for discrepancy showing that the expo-
nent −1/4 in (13.20) is best possible.
The arithmetic quantum chaos (see a stimulating article by P. Sar-

nak [Sa2]), is a new area in which modern number theory interacts with
physics more strongly than ever before.
Appendix A. Classical analysis
In the lectures we have appealed to several facts from classical

analysis. We give here a brief account of these facts.
A.1. Self-adjoint operators.
Let T : H → H be a linear operator in a Hilbert space H. Denote

by D(T ), R(T ) the domain and the range of T , respectively. Suppose
D(T ) is dense in H, then there exists the adjoint operator T ∗ defined
uniquely (Riesz theorem) by
hT f, gi = hf, T ∗ gi .
We have the properties T1 ⊂ T2 implies T2∗ ⊂ T1∗ and T ⊂ T ∗∗ .
An operator T is said to be symmetric if T ⊂ T ∗ and self-adjoint
if T = T ∗ .
Lemma A.1. The eigenvalues of a symmetric operator are real.
Proof. If T f = λf with f ∈ H, f 6= 0 and λ ∈ C, then

λhf, f i = hλf, f i = hT f, f i = hf, T ∗ f i = hf, T f i = λ̄hf, f i ;
whence λ = λ̄ as claimed.
205
206 Appendix A. Classical analysis
Lemma A.2. The eigenfunctions for distinct eigenvalues of a sym-

metric operator are orthogonal.
Proof. Let T f = λf , T g = ηg with f, g ∈ H. Then we have
λhf, gi = hT f, gi = hf, T gi = ηhf, gi ;
whence hf, gi = 0 if λ 6= η as claimed.
A symmetric operator T is said to be non-negative if
hT f, f i ≥ 0 , for all f ∈ D(T ) .
Therefore, the eigenvalues of a symmetric, non-negative operator are

non-negative
Theorem A.3 (Friedrichs). A symmetric, non-negative operator ad-

mits a self-adjoint extension.
Now let T be a self-adjoint operator in a Hilbert space H. The

operator
Rλ = (T − λ)−1
is called the resolvent. The complex number λ is said to be a regular
point for T if the resolvent Rλ is defined on the whole space H and
is a bounded operator. The remaining complex numbers comprise the
spectrum of T
σ(T ) = λ ∈ C : λ not regular .
The spectrum σ(T ) is partitioned into the point spectrum and the con-
tinuous spectrum (not a disjoint partition, in general) as follows:
- λ belongs only to the point spectrum if and only if Rλ is defined
on a set not dense in H and it is bounded.
- λ belongs only to the continuous spectrum if and only if Rλ is
defined on a dense set in H and it is unbounded.
- λ belongs to both spectra if and only if Rλ is defined on a set not
dense in H and it is unbounded.
Self-adjoint operators 207
Lemma A.4. Let T be a self-adjoint operator in H. All λ ∈ C \ R are

regular points of T . Moreover,
(A.1) kRλ k ≤ |Im λ|−1 .
Proof. For any λ ∈ C we have
k(T − λ)f k2 = k(Rλ )f k2 = k(T − Re λ)f k2 + (Im λ)2 kf k2 .
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
h(T − λ)f, gi = 0 , for any f ∈ D(T ) .
This gives hT f, gi = hf, λ̄gi. Since T = T ∗ it follows that g ∈ D(T ∗ ) =

D(T ) and λ̄g = T ∗ g = T g so λ̄ would be an eigenvalue of T if g 6=
0. Hence g = 0 which shows that (T − λ) D(T ) is dense in H. But
(T − λ) D(T ) is closed; therefore, it is equal to H.
Lemma A.5 (Hilbert formula). The resolvent of a self-adjoint operator

at regular points satisfies
(A.2) Rλ − Rγ = (λ − γ)Rλ Rγ .
Proof. This follows by applying Rλ to the obvious identity

I − (T − λ)Rγ = (T − γ) − (T − λ) Rγ = (λ − γ)Rγ .
Lemma A.6. Let T be a self-adjoint operator on H and f, g ∈ H. The

map
λ 7→ hf, Rλ gi
is holomorphic in C \ R.
Proof. Using Hilbert’s formula we infer the following relation
(γ − λ)−1 hf, Rγ gi − hf, Rλ gi = hf, Rλ Rγ gi

= hf, Rλ Rλ gi(γ − λ)hf, Rλ Rλ Rγ gi ,

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 .
A.2. Matrix analysis.
We recall a few basic facts about complex matrices (finite dimen-

sion operators) A = (aij ) ∈ Mn (C). We denote
A = (aij ) complex conjugate,

At = (aji ) transpose,
t
A∗ = A adjoint matrix.
A is non-singular if it has the inverse A−1 , or equivalently its determinat

|A| = det (aij ) does not vanish. A matrix U ∈ Mn (C) is unitary if
U ∗ U = I. It has the properties
- |det U | = 1,
- the columns of U form an orthonormal set of vectors,
- x and U x have the same length, i.e. U is an isometry.
- the eigenvalues of U are on the unit circle.
Two matrices A, B ∈ Mn (C) commute if AB = BA. A matrix
A ∈ Mn (C) is called normal if it commutes with the adjoint A∗ . Clearly
a unitary matrix is normal.
Proposition A.7 (Simultaneous diagonalization). Let F ⊂ Mn (C) be

a commuting family of normal matrices. There exists a unitary matrix
U such that
λ1 0
 
U −1 AU =  .. , for every A ∈ F .
.
0 λn
The Hilbert-Schmidt integral operators 209
A matrix is Hermitian if A = A∗ . This means that the matrix A is

normal, and it has all eigenvalues real. As a special case of Proposition
A.7 we get
Corollary A.8 (Spectral theorem for Hermitian matrices). Let A be a

Hermitian matrix. There exists a diagonal real matrix
λ1 0
 
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.
A slight generalization of this result is the following
Corollary A.9. Let L be a symmetric operator in a Hilbert space H

with eigenspaces of finite dimension and let ∆ be another symmetric
operator in H which commutes with L. Then there exists a maximal
orthonormal system of eigenfunctions of L (not necessarily a complete
system in H) which are also eigenfunctions of ∆.
Proof. Let Hλ be the eigenspace of L for eigenvalue λ. Since ∆

commutes with L it maps Hλ into itself. The unitary diagonalization
of the corresponding Hermitian matrix yields the desired system.
A.3. The Hilbert-Schmidt integral operators.
Let F be a domain in R2 . An integral operator

Z
(Lf )(z) = k(z, w) f (w) dw
F
2
with a kernel k ∈ L (F × F, dz dw) is called the Hilbert-Schmidt type
operator. Clearly L : L2 (F ) → L2 (F ), and it is a bounded operator for
ZZ
2
kLk ≤ |k(z, w)|2 dz dw .
F ×F
Moreover, if k(z, w) = k(z, w), then L is symmetric (not necessarily

self-adjoint).
Theorem A.10 (Hilbert-Schmidt). Let L 6= 0 be a symmetric integral

operator of Hilbert-Schmidt type. Then
- L has pure discrete spectrum,
- the eigenspaces of L have finite dimension,
- the eigenvalues of L can accumulate only at zero,
- L has at least one eigenvalue, the largest one being
kLf k
µ0 = sup = kLk
f 6=0 kf k
where the supremum is attained by an eigenfunction of L (the vari-

ational principle),
- the range of L in L2 (F ) is spanned by eigenfunctions of L. Let
{uj }j≥0 be any maximal orthonormal system of eigenfunctions of
L in L2 (F ), i.e.
huj , uk i = δjk , Luj = µj uj with |µ0 | ≥ |µ1 | ≥ · · · .
Then any f from the range of L has an absolutely and uniformly
convergent series representation
X
(A.3) f (z) = hf, uj i uj (z) .
j≥0
A.4. The Fredholm integral equations.
Suppose λ is a complex number, D is a domain, K : D × D −→ C

is a kernel function, and f : D −→ C is a given function. The Fredholm
equation (of the second type) is
Z
(A.4) g(x) − λ K(x, y) g(y) dy = f (x)
D
or in the operator notation (I − λK)g = f . If f ≡ 0 then the equation

is called homogeneous. We seek solutions g ∈ L2 (D); therefore, it is
natural to assume that f ∈ L2 (D) and K ∈ L2 (D × D). Denote
ZZ
2
kKk = |K(x, y)|2 dx dy < +∞ .
D×D
Clearly K : L2 (D) −→ L2 (D) is a compact operator.

The parameter λ ∈ C is said to be a characteristic number of the
kernel K(x, y) if the homogeneous equation
(A.5) (I − λK)g = 0
has a solution g ∈ L2 (D), g 6= 0. This is possible only if λ 6= 0

and then λ−1 is just the eigenvalue of the operator K whereas g is its
eigenfunction.
For small |λ| the method of successive approximations leads to

the solution. We start with g0 = f and define by induction gp =
λKgp−1 + f , i.e.
p
X
gp = λj K j f .
j=0
The infinite series

∞
X
(A.6) g= λj K j f
j=0
is called the Neumann series. The norm of g is majorized by the series

∞
X
|λ|j kKkj kf k ,
j=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
kgk ≤ |λ| kKk kgk ;
whence g = 0 almost everywhere since |λ| kKk < 1. In other words,

this shows that the inverse operator (I − λK)−1 exists and is bounded
by
k(I − λK)−1 k ≤ (1 − |λ| kKk)−1
in the disc |λ| < kKk−1 . The iterated operators K j are given by the
kernels
Z
(A.7) Kj (x, y) = K(x, z) Kj−1 (z, y) dz , j = 2, 3, . . .
D
with K1 = K. Suppose that

Z
2
A(x) = |K(x, y)|2 dy < +∞ ,
ZD
B(y)2 = |K(x, y)|2 dx < +∞ .
D
Applying the Cauchy-Schwarz inequality we show by induction that
(A.8) |Kj (x, y)| ≤ A(x) B(y) kKkj−2 , j = 2, 3, . . . .
Hence, the series

∞
X
(A.9) Rλ (x, y) = λj−1 Kj (x, y)
j=1
is majorized by
∞
X
|K(x, y)| + |λ| A(x) B(y) |λ|j kKkj
j=0
so it gives a function Rλ (x, y) in L2 (D × D), which is holomorphic in λ

in the disc |λ| < kKk−1 . One can integrate term by term showing by
(A.6) that
Z
(A.10) g(x) = f (x) + λ Rλ (x, y) f (y) dy .
D
In operator notation this asserts that (I − λK)−1 = I + λR where R is

the integral operator whose kernel is Rλ (x, y). This kernel is called the
resolvent of K, and it satisfies the Fredholm equation (use (A.9))
Z
(A.11) Rλ (x, y) = K(x, y) + λ K(x, z) Rλ (z, y) dz .
D
By the principle of analytic continuation it follows that the solution

g given by (A.10) is unique not only for λ in the disc |λ| < kKk−1 but
also in a larger domain to which Rλ has analytic continuation, and it
is in L2 (D × D).
A nice, explicit construction was given by Fredholm in the special

case of D having finite volume and K(x, y) being bounded on D × D ,
vol D = V < +∞ ,
(A.12)
|K(x, y)| ≤ K < +∞ .
In this case we shall construct two entire functions D(λ) 6≡ 0 and

Dλ (x, y) such that
(A.13) Rλ (x, y) = D(λ)−1 Dλ (x, y)
for all λ ∈ C with D(λ) 6= 0. Put

ξ1 , . . . , ξ m
K = det K(ξi , ηj ) .
η1 , . . . , η m
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 .
Hence the series

∞
X (−λ)m
(A.15) D(λ) = 1 + Cm
1
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!
showing that it converges absolutely in the whole complex λ-plane, and
(A.16) D(λ) exp(3 |λ| K V )2 .
Therefore, D(λ) is an entire function of order 2. Note that D(0) = 1 so

D(λ) does not vanish for small |λ|. Similarly, the series
∞
X (−λ)m
(A.17) Dλ (x, y) = Cm (x, y)
0
m!
where C0 (x, y) = K(x, y) is majorized by

∞ √
X ( m + 1 |λ| K V )m
eK K exp(3 |λ| K V )2
0
m!
showing that Dλ (x, y) is an entire function of order 2 in λ.

Developing the determinant by the first row we obtain
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
The positions of the accented elements above are meant to be skipped.

Hence Cm (x, y) satisfies the recurrence integral formula
Z
Cm (x, y) = Cm K(x, y) − m K(x, z) Cm−1 (z, y) dz
Green function of a differential equation 215
for all m = 1, 2, . . . . Adding we find that Dλ (x, y) satisfies the Fredholm

equation
Z
(A.18) Dλ (x, y) = D(λ) K(x, y) + λ K(x, z) Dλ (z, y) dz ,
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.5. Green function of a differential equation.
Generally speaking a Green function Gλ (z, z 0 ) is the kernel of the

resolvent Rλ = (I − λT )−1 for a suitable linear operator T provided,
of course, the resolvent is an integral operator, and this, indeed, is the
case for great many operators either integral or differential. Precise
definitions of a Green function vary a bit as specific situations call for
additional properties to meet the uniqueness requirement. The Green
function depends analytically on λ in a small initial domain from which
its analytic continuation leads the way to the spectral resolution of T .
In Section A.4 we constructed the Green function for Fredholm’s

integral equation. Now we consider the ordinary differential equation
(A.20) Tg = f
where T : C ∞ (R+ ) −→ C ∞ (R+ ) is a second order differential operator

given by
(A.21) T g(y) = −g 00 (y) + p(y) g(y) , p ∈ C ∞ (R+ ) .
The Green function of the equation (A.20) (or of the operator T ) is a

function G : R+ × R+ −→ C such that
Z +∞
(A.22) T G(y, y 0 ) g(y 0 ) dy 0 = g(y) ,
0
for all g ∈ C ∞ (R+ ) with g(0) = g(+∞) = 0. Therefore G yields an

integral operator which is the right inverse to T on functions g as above.
Of course, G is not unique. We shall be looking for the Green function
which satisfies the following conditions:
(A.23) G(y, y 0 ) is continuous in R+ ×R+ and smooth everywhere except
for the diagonal y = y 0 ,
(A.24) T G(y, y 0 ) = 0 if y 6= y 0 .
Suppose a function G(y, y 0 ) satisfies the above conditions. Differ-

entiating the identity
Z +∞ Z y Z +∞
0 0 0 0 0 0
G(y, y ) g(y ) dy = G(y, y ) g(y ) dy + G(y, y 0 ) g(y 0 ) dy 0
0 0 y
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
This yields the following identity
+∞ ∂ ∂
Z
0 0 0
T G(y, y ) g(y ) dy = G(y, y − 0) − G(y, y + 0) g(y) .
0 ∂y ∂y
Hence, we conclude that the property (A.22) is equivalent to
∂ ∂
(A.25) G(y, y − 0) − G(y, y + 0) = 1 .
∂y ∂y
Green function of a differential equation 217
To construct a Green function satisfying the conditions (A.23) and

(A.24) we take two linearly independent solutions to the homogeneous
equation, say I(y) and K(y),
T I(y) = −I 00 (y) + p(y) I(y) = 0 ,

T K(y) = −K 00 (y) + p(y) K(y) = 0
and seek a solution of the type

(
0
I(y) A(y 0 ) if y > y 0 ,
G(y, y ) =
K(y) B(y 0 ) if y < y 0 ,
where A(y 0 ), B(y 0 ) are to be determined. Then (A.24) is automati-

cally satisfied. The continuity condition (A.23) and the jump condition
(A.25) yield the linear system for the unknown functions
I(y) A(y) − K(y) B(y) = 0 ,

−I (y) A(y) + K 0 (y) B(y) = 1 .
0
The determinant of this system is the Wronskian W = IK 0 − I 0 K

which is constant because W 0 = IK 00 − I 00 K = 0. Since I, K are
linearly independent, W 6= 0. Hence, we have unique solutions A(y) =
W −1 K(y) and B(y) = W −1 I(y) which give
(
W −1 I(y) K(y 0 ) if y ≥ y 0 ,
G(y, y 0 ) =
W −1 K(y) I(y 0 ) if y ≤ y 0 .
Appendix B. Special functions
Special functions have been created gradually by the demand of sci-

entists and engineers concerned with real computations. Various types
of special functions are associated with specific problems of mechanics
and physics. Due to deep intuition of mathematicians the modern ap-
proach to special functions is unified beautifully through the language
of the group representation theory. Today special functions live on
suitable symmetric spaces. Therefore, it is not surprising that some
special functions are encountered in the theory of automorphic forms.
Here we give excerpts from several sources of what is needed about spe-
cial functions for these lectures. No proofs are supplied so the reader
still has to penetrate the jungle of relevant literature for completeness.
Recommended books are [Gr-Ry], [Le], [Vi], [Wa].
B.1. The gamma function.
The gamma function of Euler is defined in Re s > 0 as the Mellin

transform of the exponential function
Z +∞
(B.1) Γ(s) = e−x xs−1 dx .
0
219
220 Appendix B. Special functions
It has a meromorphic continuation over the whole complex plane given

by
+∞ ∞
(−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
where γ = .5772156649 . . . is the Euler constant. Hence, Γ(s) does not

vanish anywhere, and so Γ(s)−1 is an entire function of order 1. Also
we have
Recursion formula:
(B.4) s Γ(s) = Γ(s + 1) .
Functional equation:
(B.5) Γ(s) Γ(1 − s) = π (sin πs)−1 .
Duplication formula:
1
(B.6) Γ(s) Γ(s + ) = π 1/2 21−2s Γ(2s) .
2
The Stirling’s asymptotic formula

2π 1/2 s s
(B.7) Γ(s) = (1 + O(|s|−1 ))
s e
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
if t > 0 with the implied constant depending on σ.
The psi function is defined by

∞
Γ0 X 1 1
(B.9) ψ(s) = (s) = −γ − − .
Γ n=0
n+s n+1
It satisfies
1 X s + k
(B.10) ψ(s) = ψ + log m
m m
0≤k<m
where m is any positive integer. We have the approximation
(B.11) ψ(s) = log s − (2 s)−1 + O(|s|−2 )
uniformly in the angle |arg s| < π − ε.
For Re u > 0, Re v > 0, Re s > 1/2 we have
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)
B.2. The hypergeometric functions.
A hypergeometric function is a solution of the differential equation
(B.14) z(1 − z) F 00 − ((α + β + 1)z − γ) F 0 − αβ F = 0
where α, β, γ are complex numbers. If γ is different from non-positive

integers, then one of the two linearly independent solutions is given by
the Gauss hypergeometric series
∞
X (α)k (β)k
(B.15) F (α, β; γ; z) = zk
(γ)k k!
k=0
with coefficients given by (s)k = Γ(s + k)/Γ(s) = s · · · (s + k − 1). The

power series converges absolutely in the unit disk |z| < 1. However, the
hypergeometric function F (α, β; γ; z) has an analytic continuation over
the plane C cut along [1, +∞). This is given by the integral represen-
tation
1
Γ(γ)
Z
(B.16) F (α, β; γ; z) = tβ−1 (1−t)γ−β−1 (1−tz)−α dt
Γ(β) Γ(γ − β) 0
when Re γ > Re β > 0 and by various recurrence formulas in the re-

maining cases. For example the following formulas will do the job
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.
The hypergeometric function satisfies many transformations rules.

Here are two examples:
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
the Legendre and Tchebyshev polynomials, respectively.

The Legendre functions 223
For |z| < 1, we have

1
(B.19) F (α, α + ; 2α + 1; 1 − z 2 ) = 4α (1 + z)−2α .
2
If Re γ > 0 and Re γ > Re (α + β), we obtain as z → 1− that
Γ(γ) Γ(γ − β − α)
(B.20) F (α, β; γ; 1) = .
Γ(γ − α) Γ(γ − β)
B.3. The Legendre functions.
Let ν, m be complex numbers. A solution to the differential equa-

tion
(B.21) (1 − z 2 ) P 00 − 2z P 0 + (ν(ν + 1) − m2 (1 − z 2 )−1 ) P = 0
is called a spherical function. This equation is encountered for boundary
value problems for domains of radial symmetry for which solutions are
naturally written in polar coordinates. We are interested exclusively in
spherical functions with m a non-negative integer, in which case they are
named Legendre functions of order m. The theory of Legendre functions
of a positive order can be reduced to that of zero order. Indeed, if Pν (z)
satisfies (B.21) with m = 0, then
dm
(B.22) Pνm (z) = (z 2 − 1)m/2 Pν (z)
dz m
satisfies (B.21) with parameter m.
One of the solutions to (B.21) with m = 0 is given by the hyper-
geometric function
1 − z
(B.23) Pν (z) = F − ν, ν + 1; 1; .
2
This function is defined and analytic in the plane C cut along (−∞, −1].
Then the Legendre function of order m derived from (B.22) is also
defined and analytic in the same domain.
There are plenty of relations between Legendre functions such as
Pν (z) = P−ν−1 (z) ,

(ν + 1) Pν+1 (z) = (2ν + 1) z Pν (z) − ν Pν−1 (z) ,
(z 2 − 1) Pν0 (z) = ν z Pν (z) − ν Pν−1 (z) .
And there are useful integral representations such as
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
If ν = n is a non-negative integer, then Pn (z) is a polynomial of

degree n given by
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
and the functions

1 (n − m)! 1/2 m
n+ Pn (cos θ) e±imϕ
2 (n + m)!
with 0 ≤ m ≤ n form a complete, orthonormal system in the eigenspace

λ = n(n + 1). Notice that these eigenfunctions are not bounded in the
spectrum. The maximum attained for m = 0 at θ = 0 is equal to
(n + 1/2)1/2 = (λ + 1/4)1/4 .
B.4. The Bessel functions.
The Bessel functions are solutions to the differential equation
(B.27) z 2 f 00 + zf 0 + (z 2 − ν 2 ) f = 0
The Bessel functions 225
where ν is a complex number. In automorphic theory this equation

arises when searching for eigenfunctions of the Laplace operator on the
hyperbolic plane by the method of separation of variables in rectangular
coordinates.
Since the equation (B.27) is singular at z = 0, there is going to be
a problem. This is resolved by cutting the complex z-plane along the
segment (−∞, 0]. Then (B.27) has two linearly independent solutions
which are holomorphic in z ∈ C \ (−∞, 0]. One of these is given by the
power series
∞
X (−1)k z ν+2k
(B.28) Jν (z) = ,
k! Γ(k + 1 + ν) 2
k=0
which converges absolutely in the whole plane. As a function of the

parameter ν, called the order of Jν (z), it is an entire function.
Changing ν into −ν does not alter the equation (B.27) so J−ν (z) is
another solution. The solutions Jν (z), J−ν (z) are linearly independent
if and only if the Wronskian W (Jν (z), J−ν (z)) = −2(πz)−1 sin πν does
not vanish identically, i.e. for ν different from an integer.
If ν = n is an integer, we have the relation
(B.29) Jn (z) = (−1)n J−n (z) .
To get a pair of linearly independent solutions which would be

suitable for all ν we choose the linear combination
(B.30) Yν (z) = (sin πν)−1 (Jν (z) cos πν − J−ν (z))
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
W (Jν (z), Yν (z)) = 2(πz)−1 .
Changing the variable z 7→ iz (rotation of angle π/2) we transform

the differential equation (B.27) into
(B.31) z 2 f 00 + zf 0 − (z 2 + ν 2 )f = 0 .
As before we seek holomorphic solutions in the complex plane cut along

the negative axis. One of these is given by the power series
∞ z ν+2k
X 1
(B.32) Iν (z) = .
k! Γ(k + 1 + ν) 2
k=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
(B.33) In (z) = I−n (z) .
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
W (Iν (z), Kν (z)) = −z −1 .
If ν = n is an integer, the functions Jn (z) and In (z) are actually

entire in z. For ν not an integer there is a discontinuity along the
negative axis, namely we have
Jν (−x + εi) − Jν (−x − εi) ∼ 2i sin(πν) Jν (x)
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
Jν−1 (z) + Jν+1 (z) = 2 ν z −1 Jν (z) ,

Jν−1 (z) − Jν+1 (z) = 2 Jν0 (z) ,
(z ν Jν (z))0 = z ν Jν−1 (z) ,
(z −ν Jν (z))0 = −z −ν Jν+1 (z) .
The same formulas hold for the Y -function. For the I-function we have
Iν−1 (z) − Iν+1 (z) = 2 ν z −1 Iν (z) ,

Iν−1 (z) + Iν+1 (z) = 2 Iν0 (z) ,
(z ν Iν (z))0 = z ν Iν−1 (z) ,
(z −ν Iν (z))0 = z −ν Iν+1 (z) .
And for the K-function the above formulas hold with negative sign on
the right-hand sides.
The Bessel functions 227
The Bessel functions of half order are elementary functions
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
where the implied constant is absolute. We don’t have clear asymptotics

for Bessel functions in the transition range y 1 + |ν|.
There are various integral representations and transforms each of

which is useful in specific situations. Here is a selection for the K-
function:
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
+∞
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.
B.5. Inversion formulas.
There are many ways of representing a function as a series and an

integral of Bessel functions. We shall present three types of expansions.
A general theory of eigenfunction expansions associated with second
order differential equations is given by E.C. Titchmarsh [Ti2].
The Hankel inversion. Suppose f is a continuous function with finite

variation on R+ such that
Z +∞
(B.38) |f (x)| x−1/2 dx < +∞ .
0
Inversion formulas 229
The Hankel transform of f of order ν with Re ν > −1/2 is defined by

Z +∞
(B.39) Hf (y) = f (x) Jν (xy) dx
0
for y > 0. It satisfies the inversion formula

Z +∞
(B.40) f (x) = Hf (y) Jν (xy) xy dy .
0
The Kontorovitch-Lebedev inversion. It is concerned with contin-

uous representations in the K-functions of imaginary order. Suppose
f (y) is a smooth function with finite variation on R+ such that
Z +∞
(B.41) |f (y)| (y −1/2 + y −1 | log y|) dy < +∞ .
0
Then the integral transform

Z +∞
(B.42) Lf (t) = Kit (y) f (y) y −1 dy
0
satisfies the inversion formula

Z +∞
(B.43) f (y) = Lf (t) Kit (y) π −2 sinh(πt) t dt .
−∞
The Neumann series. It is concerned with discrete representations

in the J-functions of integral order. Note that J` (x) for ` > 0 is square
integrable on R+ with respect to the measure x−1 dx. The Neumann
coefficients of a function f ∈ L2 (R+ , x−1 dx) are defined by
Z +∞
(B.44) Nf (`) = f (x) J` (x) x−1 dx .
0
These give the Neumann series

X
(B.45) f 0 (x) = 2` Nf (`) J` (x) .
0<` odd
If f (x) is smooth on R+ such that
(B.46) f (j) (x) x(x + 1)−4 , 0 ≤ j ≤ 2,

then the Neumann series converges absolutely and uniformly. By the

orthogonality property
Z +∞
2` J` (x) Jm (x) x−1 dx = δ`,m
0
if ` ≡ m ≡ 1 (mod 2) (see (B.37)), one shows that f 0 (x) is the projection

of f (x) on the subspace of L2 (R+ , x−1 dx) spanned by the J` (x) of odd
order. This subspace is not dense in L2 (R+ , x−1 dx) so f 0 (x) is not
always equal to f (x).
The Titchmarsh integral. It is concerned with continuous represen-

tations in the J-functions of imaginary order. Put
π −1
(B.47) Bν (x) = (2 sin ν) (J−ν (x) − Jν (x)) .
2
Note that B2it (x) ∈ L2 (R+ , x−1 dx) if t > 0 and B2it (x) is orthogonal
to all J` (x) with 0 < ` ≡ 1 (mod 2), which fact follows by (B.37). The
Titchmarsh coefficients of a function f ∈ L2 (R+ , x−1 dx) are defined by
Z +∞
(B.48) Tf (t) = f (x) B2it (x) x−1 dx .
0
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
In other words it turns out that tanh(πt) t dt is the relevant spectral

measure. The continuous packet {B2it }t>0 spans densely the orthogonal
complement to the linear subspace of the discrete collection {J1 , J3 , J5 ,
. . . } in L2 (R+ , x−1 dx). More precisely, for f (x) satisfying (B.46), the
Sears-Titchmarsh inversion holds
(B.50) f (x) = f 0 (x) + f ∞ (x) .
The above partition can be characterized nicely in terms of the

Hankel transform of order zero. Indeed, by recurrence formulas for the
J-function we obtain
2ν d
Jν (ux) Jν (uy) = u Jν−1 (ux) Jν−1 (uy) − u Jν+1 (ux) Jν+1 (uy) .
xy du
Inversion formulas 231
Integrating over 0 < u < 1 and summing over ν = 1, 3, 5, . . . , we get
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
where Hf (u) is the Hankel transform of f given by

Z +∞
(B.53) Hf (u) = f (y) J0 (uy) dy .
0
By the Hankel inversion (B.40) and the Sears-Titchmarsh inversion

(B.50) it follows that
Z +∞
∞
(B.54) f (x) = ux J0 (ux) Hf (u) du .
1
Therefore the projections f 0 (x), f ∞ (x) are obtained by truncating the

Hankel transform of f to the segments 0 ≤ u < 1 and 1 ≤ u < +∞
respectively.
References
[At-Le] A. O. L. Atkin and J. Lehner, Hecke operators on Γ0 (m). Math. Ann.

185 (1970), 134-160.
[Br1] R. W. Bruggeman, Fourier coefficients of cusp forms. Invent. Math. 45
(1978), 1-18.
[Br2] R. W. Bruggeman, Automorphic forms, in Elementary and Analytic
Theory of Numbers, 31-74, Banach Center Pub. 17, Warsaw, 1985.
[Bu-Du-Ho-Iw] D. Bump, W. Duke, J. Hoffstein and H. Iwaniec, An estimate for the
Hecke eigenvalues of Maass forms. Duke Math. J. Research Notices 4
(1992), 75-81.
[Ch] F. Chamizo, Temas de Teorı́a Analı́tica de los Números. Doctoral The-
sis, Universidad Autónoma de Madrid, 1994.
[Co-Pi] J. W. Cogdell and I. Piatetski-Shapiro, The Arithmetic and Spectral
Analysis of Poincaré Series. Perspectives in Math. 13, Academic Press,
1990.
[De-Iw] J.-M. Deshouillers and H. Iwaniec, Kloosterman sums and Fourier coef-
ficients of cusp forms. Inv. Math. 70 (1982), 219-288.
[El] J. Elstrodt, Die Resolvente zum Eigenwertproblem der automorphen
Formen in der hyperbolischen Ebene, I, II, III. Math. Ann. 203 (1973),
295-330; Math. Z. 132 (1973), 99-134; Math. Ann. 208 (1974), 99-132.
[Fa] J. D. Fay, Fourier coefficients of the resolvent for a Fuchsian group. J.
Reine Angew. Math. 293/294 (1977), 143-203.
[Fi] J. Fischer, An Approach to the Selberg Trace Formula via the Selberg
Zeta-Function, Springer Lecture Notes in Math. 1253, (1980).
[Foc] V. A. Fock, On the representation of an arbitrary function by an integral
involving Legendre’s functions with a complex index. C. R. Acad. Sci.
U.R.S.S, Doklady Akad. Nauk. U.S.S.R. 39 (1943), 253-256.
[For] L. R. Ford, Automorphic Functions. McGraw-Hill, New York, 1929.
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
[Kat] S. Katok, Fuchsian Groups. Univ. of Chicago Press, 1992.

[Kl1] H. D. Kloosterman, On the representation of numbers in the form ax2 +
by 2 + cz 2 + dt2 . Acta Math. 49 (1926), 407-464.
[Kl2] H. D. Kloosterman, Asymptotische Formel für die Fourier-Koeffizienten
ganzer Modulformen. Abh. Math. Sem. Univ. Hamburg 5 (1927),
338-352.
[Ko-Le] M. J. Kontorovitch and N. N. Lebedev, J. Exper. Theor. Phys. U.S.S.R.
8 (1938), 1192-1206, and 9 (1939), 729-741.
[Ku] N. V. Kuznetsov, Petersson’s conjecture for cusp forms of weight zero
and Linnik’s conjecture. Sums of Kloosterman sums. Math. U.S.S.R.
Sbornik 29 (1981), 299-342.
[La] S. Lang, SL2 (R). Addison-Wesley, Reading-Mass., 1973.
[Le] N. N. Lebedev, Special Functions and their Applications. Dover Publi-
cations, New York, 1972.
[Lu] W. Luo, On the non-vanishing of Ranking-Selberg L-functions, Duke
Math. J. 69 (1993), 411-427.
[Lu-Ru-Sa] W. Luo, Z. Rudnick and P. Sarnak, On Selberg’s eigenvalue conjecture.
Preprint, 1994.
[Lu-Sa] W. Luo and P. Sarnak, Quantum ergodicity of eigenfunctions on
SL2 (Z)\H. Preprint, 1994.
[Ma] H. Maas, Über eine neue Art von nichtanalytschen automorphen Funk-
tionen und die Bestimmung Dirichletscher Reihen durch Funktionalgle-
ichungen. Math. Ann. 121 (1949), 141-183.
[Me] F. G. Mehler, Über eine mit den Kugel-und Cylinderfunctionen ver-
wandte Funktion und ihre Anwendung in der Theorie der Elektricität-
svertheilung. Math. Ann. 18 (1881), 161-194.
[Mi-Wa] R. Miatello and N. R. Wallach, Kuznetsov formulas for rank one groups.
J. Funct. Anal. 93 (1990), 171-207.
[Ni] J. Nielsen, Über Gruppen linearer Transfromationen. Mitt. Math. Ges.
Hamburg 8 (1940), 82-104.
[Pe1] H. Petersson, Über die Entwicklungskoeffizienten der automorphen For-
men. Acta Math. 58 (1932), 169-215.
[Pe2] H. Petersson, Über einen einfachen Typus von Untergruppen der Mod-
ulgruppe. Archiv der Math. 4 (1953), 308-315.
[Ph-Ru] R. S. Phillips and Z. Rudnick, The circle problem in the hyperbolic
plane. To appear in J. Funct. Anal.
236 References
[Ph-Sa] R. S. Phillips and P. Sarnak, On cusp forms for cofinite subgroups of

PSL(2,R). Invent. Math. 80 (1985), 339-364.
[Pr] N. V. Proskurin, Summation formulas for general Kloosterman sums (in
Russian). Zap. Naučn. Sem. LOMI 82 (1979), 103-135.
[Rad] H. Rademacher, A convergence series for the partition function p(n).
Proc. Nat. Acad. Sci. U.S.A. 23 (1937), 78-84.
[Ran] R. A. Rankin, Modular Forms of Negative Dimensions. Dissertation.
Clave College Cambridge, 1940.
[Ro] W. Roelcke, Über die Wellengleichung be Grenzkreisgruppen erster Art.
S.-B. Heidelberger Akad. Wiss. Math. Nat. Kl., 1956, 4 Abh.
[Ru-Sa] Z. Rudnick and P. Sarnak, The behavior of eigenstates of arithmetic
hyperbolic manifolds. To appear in Comm. Math. Phys.
[Sa1] P. Sarnak, Determinants of Laplacians. Comm. Math. Phys. 110
(1987), 113-120.
[Sa2] P. Sarnak, Arithmetic Quantum Chaos. R. A. Blyth Lectures, University
of Toronto, 1993.
[Sa3] P. Sarnak, Some Applications of Modular Forms. Cambridge Tracts in
Math. 99, Cambridge University Press, 1990.
[Sa4] P. Sarnak, Class numbers of indefinite binary quadratic forms. J. Num-
ber Theory 15 (1982), 229-247.
[Se-Ti] D. B. Sears and E. C. Titchmarsh, Some eigenfunction formulae, Quart
J. Math. Oxford 1 (1950), 165-175.
[Se1] A. Selberg, On the estimation of Fourier coefficients of modular forms.
Proc. Symp. Pure Math. VII, Amer. Math. Soc. (1965), 1-15.
[Se2] A. Selberg, Collected Papers, vol. I. Springer, 1989.
[Se3] A. Selberg, Über die Fourierkoeffizienten elliptischer Modulformen neg-
ativer Dimension. C. R. Neuvième Congrès Math. Scandinaves, Hels-
ingfors (1938), 320-322. Mercatorin Kirjapaino, Helsinki, 1939.
[Sh] G. Shimura, Introduction to the Arithmetic Theory of Automorphic
Functions. Princeton Univ. Press, 1971.
[Si1] C. L. Siegel, Bemerkung zu einem Satze von Jakob Nielsen. Mat.
Tiolsskv. B (1950), 66-70.
[Si2] C. L. Siegel, Discontinuous Groups. Ann. of Math. 44 (1943), 674-689.
[Te] A. Terras, Fourier Analysis on Symmetric Spaces and Applications, I,
II. Springer-Verlag, 1985, 1988.
[Ti1] E. C. Titchmarsh, Introduction to the Theory of Fourier Integrals. Cla-
rendon Press, Oxford, 1948 (2-nd edition).
References 237
[Ti2] E. C. Titchmarsh, Eigenfunction Expansions Associated with Second-

Order Differential Equations. Clarendon Press, I, revised edition, 1962;
II, 1958.
[Ve] A. B. Venkov, Spectral Theory of Automorphic Functions and its Appli-
cations. Kluwer Academic Pub., Soviet Series 51, 1990.
[Vi] N. Ya. Vilenkin, Special Functions and Theory of Group Representa-
tions. Amer. Math. Soc. Translations of Math. Monograph. 22, 1968.
[Vig1] M.-F. Vignéras, L’équation fonctionelle de la fonction zéta de Selberg
du groupe modulaire PSL(2,Z). Asterisque 61 (1979), 235-249.
[Vig2] M.-F. Vignéras, Variétés riemanniennes isospectrales et non isométri-
ques. Ann. of Math. 112 (1980), 21-32.
[Wa] G. N. Watson, A Treatise on the Theory of Bessel Functions. Cambridge
University Press, London, 1962.
[We] A. Weil, On some exponential sums. Proc. Nat. Acad. Sci. U.S.A. 34
(1948), 204-207.
[Wi] A. Winkler, Cusp forms and Hecke groups. J. Reine Angew. Math. 386
(1988), 181-204.
[Wo] S. Wolpert, Disappearence of cusp forms in families. Ann. of Math.
139 (1994), 239-291.
[Zo] P. Zograf, Fuchsian groups and small eigenvalues of the Laplace opera-
tor. Zap. Naučn. Sem. LOMI 122 (1982), 24-29.
Subject Index
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
Barnes G-function, 169

Bessel functions, 4, 22, 224
Bessel’s inequality, 107
Bruggeman-Kuznetsov formula, 140
Bruggeman-Kuznetsov formula reversed, 141
Bruhat decomposition, 18
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
Finite volume group, 41

first kind group, 40
Fourier expansion of automorphic forms, 59, 145
Fourier expansion of Eisenstein series, 66
Fourier transform, 4
Fredholm equation, 210
free space, 35
frequencies of a Kloosterman sum, 51
Fuchsian group, 40
functional equation of Eisenstein series, 68
functional equation of the scattering matrix, 94
fundamental domain, 41
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
Green function of a differential equation, 83, 215

Green’s formula, 37
Hankel inversion, 143, 229

Hankel type transforms, 6, 143
Hecke group, 45
Hecke operator, 127
Hilbert’s formula for the iterated resolvent, 75, 207
Hilbert-Pólya dream, 170
homogeneous-space model, 12
horocycles, 19
hyperbolic circle problem, 190
hyperbolic element, 19
hyperbolic group, 40
hyperbolic plane, 7
hypercycles, 19
hypergeometric function, 26, 221
Incomplete Eisenstein series, 62

inner product, 60, 145
invariant height, 59
invariant integral operator, 4, 28
invariant linear operator, 20
inverse Selberg/Harish-Chandra transform, 34
isometric circle, 8
isoperimetric inequality, 11
iterated resolvent, 75, 152
Iwasawa decomposition, 13
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
Maass automorphic form, 58

Maass-Selberg relations, 96, 99
mean-value operator, 31
Mehler-Fock inversion formula, 26
method of averaging images, 21
Möbius transformations, 8
modular function of a locally compact group, 15
modular group, 46
modulus of a Kloosterman sum, 51
Newforms of level N , 129

Neumann coefficients, 229
Neumann series, 144, 211, 229
norm of g, 19
norm of P , 162
norm in M2 (R), 39
normal matrix, 208
normal polygon, 41
Oldform of level N , 129
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
Titchmarsh coefficients, 230

Titchmarsh integral, 230
total principal part of K(z, w), 73
trace of a conjugate class, 19
trace of a kernel, 149
trace class operator, 149
trace of L, 5
translation, 19
triangular group of type (α, β, γ), 44
truncated Eisenstein series, 96
twisted Eisenstein series, 124
twisted Maass forms, 124
244 Subject Index
Unimodular, 15
unique ergodicity conjecture, 202
upper half-plane, 7
Variational principle, 210
Weighted Eisenstein series, 62

weighted Poincaré series, 62
Weil bound for Kloosterman sums, 52
Weyl’s law, 175
Whittaker function, 22
Wronskian, 84, 217
Zero-th term, 24, 59

Notation Index
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 fâ (n), 59
dνt (z), 202 fâjk (m), 146
dνt (z), 202 fâjk (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
Φ(s), 94 Lf (t), 229

{g}, 18 Ls (m, n), 133
G = SL2 (R), 8 L(Γ\H), 60
Gs (z/z 0 ), 77 λ = s(1 − s), 21, 70
GYab (z/z 0 ), 89 λab , 88
Gs (u(z, w)), 26, 35 Λ(λ, k), 32
[γ], 150 MΓ (R), 157
Γ(s), 219 Mk (Γ) 144
Γq , 45 N, 13
Γn , 127 n(x), 13
Γz , 40 Nf (`), 144
Γ(n, p), 46 NΓ (T ), 110
Γ(N ), 46 νab , 88
Γ0 (N ), 47 νaj (n), 118
Γ1 (N ), 47 ν 2 , ν3 , 48
H, 7 ω(z, w), 32
h− (x), 139 ωd/c , 49
h+ (x), 139 ω∞ , 49
Ha (z, w), 72 Ωd/c , 49
Hf (x), 143 Ωd/c (m), 51
Hs (z, w), 35 Ω∞ , 49
Iν (y), 22, 225 p = NP, 162
jg (z), 9 P (Y ), 42
J0 (z), 4 Pam (z), 145
Jν (y), 225 P̂ab (m, n), 146
K, 12 Pn (y, y 0 ), 79
KC (z, w), 150 Pn,m (z, z 0 ), 80
Kf (t), 141 P SL2 (R), 8
k(ϕ), 12 P (X), 190
Kν (y), 22, 226 Pn (X), 193
m
K(z, w), 71 P−s (v), 25, 224
K̂(z, w), 73 ψajk (m), 147
L̂, 73 ψ(s), 160
Laj (s), 119 ψ̂(s), 62
Notation Index 247
(r, ϕ), 17 Tn f , 127

r(`), 6 Tr K, 149
Rs (f ), 35, 78 τ (c), 52
R(Γ\H), 112 u(z, w), 8
Rs (Γ\H), 112 uaj (z), 111
ρ(z, w), 7 Usm (z), 25
ρaj (n), 117 Vs (z), 24
S(m, n; c), 52 Ws (z), 22
Sab (m, n; c), 51 yΓ , 59
SL2 (R), 8 yΓ (z), 59
SL2 (Z), 46 Z(γ), 150
σa , 42 ZΓ (s), 168
Tf (t), 142 Zs (m, n), 134

Introducrion To The Spectral Theory

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Introducrion To The Spectral Theory

Uploaded by

Copyright:

Available Formats

Introduction to the

Revista Matemática Iberoamericana

Antonio Córdoba José L. Fernández

Subject Index 239

Notation Index 245

I was captivated by a group of enthusiastic Spanish mathematicians

Primarily, the lectures are addressed to advanced graduate stu-

Many people helped me prepare these notes for publication. In

New Brunswick, October 1994 Henryk Iwaniec

The concept of an automorphic function is a natural generalization

To define an automorphic function in an abstract setting one needs

We begin by presenting the familiar case of the euclidean plane

The group G = R2 acts on itself as translations, and it makes R2 a

of curvature K = 0, and the Laplace-Beltrami operator associated with

Clearly the exponential functions

ϕ(x, y) = e(ux + vy) , (u, v) ∈ R2 ,

(D + λ)ϕ = 0 , λ = λ(ϕ) = 4π 2 (u2 + v 2 ) .

The well known Fourier inversion

is just the spectral resolution of D on functions satisfying proper decay

For L to be G-invariant it is necessary and sufficient that the kernel

k(x, y) = k(x2 + y 2 ) , k(r) ∈ C0∞ (R+ ) .

where λ = 4π 2 (u2 + v 2 ) and J0 (z) is the Bessel function given by

Classical analytic number theory benefits a lot from harmonic anal-

functions. Restricting the domain of the invariant integral operator L

by folding the integral. Hence the trace of L on the torus is equal to

On the other hand by the spectral decomposition (classical Fourier se-

the trace is given by

Comparing both results we get the trace formula

which is better known as the Poisson summation formula. By a linear

In particular for a radially symmetric function the Poisson sum-

Theorem (Hardy-Landau, Voronoi). If k ∈ C0∞ (R), then

and k̃ is the Hankel type transform of k given by

Note that the lowest eigenvalue λ(1) = 4π 2 ` = 0 for the constant

which usually constitutes the main term. Taking a suitable kernel (a

which was originally established by Voronoi and Sierpinski by different

1.1. The upper half-plane.

As a model of the hyperbolic plane we shall use the upper half of

H is a riemannian manifold with the metric derived from the Poincaré

(1.1) ds2 = y −2 (dx2 + dy 2 ) .

The distance function on H is given explicitly by

However, a more practical relation is

(1.3) cosh ρ(z, w) = 1 + 2 u(z, w) ,

up to sign. In particular both matrices 1 = 1 1 and −1 = −1 −1

give the identity transformation. We shall always make this distinction

Figure 1. Isometric circles.

The j-function satisfies the chain rule of differentiation,

(1.10) jgh (z) = jg (hz) jh (z) .

It follows from the formula

(1.11) |jg (z)|2 Im gz = Im z

(Im gz)−1 |dgz| = (Im z)−1 |dz| ,

which shows that the differential form (1.1) on H is G-invariant. This

Theorem 1.1. The whole group of isometries of H is generated by the

Theorem 1.2. The hyperbolic lines (geodesics in H) are represented

The hyperbolic circles (loci of points in a fixed distance from a given

Theorem 1.3 (Gauss defect). The area of a hyperbolic triangle with

There is a universal inequality between the area and the boundary

where A is the area, L is the length of the boundary, and K is the

hence A ≤ L so the area is comparable to the boundary length. This

Figure 3. A hyperbolic disc.

The hyperbolic area is 4π(sinh(r/2))2 , and the circumference is

1.2. H as a homogeneous space.

Occasionally it will be convenient to work with the homogeneous

The point z = i is special; its stability group is the orthogonal group