Spectral theory of automorphic forms

and analytic continuation of Eisenstein

series

Bart MICHELS

Advisor: F. Brumley

Dissertation submitted in partial fulfillment of the requirements for the degree of Master de
Sciences, Technologies, Santé, Mention Mathématiques et Applications, Parcours Type
Mathématiques Fondamentales

Université Pierre et Marie Curie – Sorbonne Université

September 2018
2
Acknowledgements
Je ne peux que remercier mon encadrant, Farrell Brumley, pour les nombreux conseils et divers faits
ainsi que l’accès à des références cruciales parfois non publiées. Sans qui ce projet, une expérience de
découverte dans l’analyse fonctionnelle sur les surfaces hyperboliques, n’aurait pas été possible. J’ai
eu le plaisir d’étudier et de raffiner des démonstrations existantes sous sa direction.

Bart Michels
Paris, septembre 2018

3
4
Contents

1 Introduction 8

2 The hyperbolic plane 9

2.1 Isometries and geometry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Group actions and fundamental domains . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.4 Fourier expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

3 Operators on symmetric spaces 17

3.1 Weakly symmetric spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 Point-pair invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.3 Radially symmetric functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.4 Isotropic spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
3.5 Integral operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.6 The algebra of invariant differential operators . . . . . . . . . . . . . . . . . . . . . . . . 25
3.7 Selberg’s eigenfunction principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

4 Functions on the quotient Γ\H 29

4.1 Eisenstein series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Automorphic kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.2.1 Fourier expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.2.2 Truncated kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
4.3 Maass forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3.1 Cusp forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3.2 Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

5 Analytic continuation of Eisenstein series 47

5.1 Elementary proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.1.1 Proof by Poisson summation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.1.2 Proof by Fourier expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2 Proof via Fredholm-theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2.1 A truncated Eisenstein series. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2.2 A Fredholm equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.2.3 Truncated kernels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.2.4 Uniqueness principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.3 Bernstein’s continuation principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.3.1 Systems of equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
5.3.2 Systems of finite type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
5.3.3 Criteria for finiteness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
5.3.4 Eisenstein series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.4 Further analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

A Functional analysis 63
A.1 Bounded operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
A.2 The adjoint of an operator, C ∗ -algebras . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
A.3 Banach-Alaoglu and the Gelfand-transform . . . . . . . . . . . . . . . . . . . . . . . . . 65
A.4 Continuous functional calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
A.5 Positive operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
A.6 Compact operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
A.7 Diagonalizable operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
A.8 Spectral theory of compact normal operators . . . . . . . . . . . . . . . . . . . . . . . . 70
A.9 Trace class and Hilbert–Schmidt operators . . . . . . . . . . . . . . . . . . . . . . . . . . 71

5
B Functional calculus 73
B.1 Differentiability and holomorphy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
B.2 Weak holomorphy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
B.3 Three notions of integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
B.4 Power series and meromorphy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
B.5 Integration in function spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
B.6 Holomorphy in function spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
B.7 Meromorphy in function spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

C Fredholm integral equations 87

C.1 Regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
C.2 The Fredholm equation for bounded operators . . . . . . . . . . . . . . . . . . . . . . . . 88
C.3 Fredholm theory for compact operators . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

D Riemannian geometry 92
D.1 Riemannian manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
D.2 Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
D.3 Geodesics and parallel transport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
D.4 Geodesics on Riemannian manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
D.4.1 The exponential map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
D.4.2 Geodesics and distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
D.4.3 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
D.5 Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
D.6 The Laplace–Beltrami operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
D.7 Isometry groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
D.8 Stabilizers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

E Symmetries of manifolds 101

E.1 Isotropic manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
E.2 Homogeneous spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
E.3 Symmetric spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

F Differential operators 103

F.1 Grading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
F.1.1 Global grading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
F.1.2 Grading at a point . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
F.1.3 Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
F.2 Elliptic regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
F.3 Invariant differential operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
F.4 Differential operators on Lie groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

G Spectral theory of the Laplacian 112

G.1 Unbounded operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
G.2 The Laplacian as a symmetric operator . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
G.3 Extensions and essential self-adjointness . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
G.4 Operators with compact resolvent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
G.5 The spectrum of the Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

H Whittaker functions 118

H.1 Asymptotic expansion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
H.2 Second solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

6
7
1 Introduction
Real analytic Eisenstein series for PSL2 (R) in their general form and their meromorphic continuation,
were first studied by Selberg, with the aim of determining the decomposition of the spectrum of the
Laplace operator ∆ on a finite-volume hyperbolic surface: it measures the default of diagonalizability
of ∆. His ultimate objective was to prove a trace formula, with applications to representation theory
of PSL2 (R).
Selberg’s work was later generalized to other groups by Langlands, and various proofs were published
by Bernstein, Selberg, Colin de Verdière and others. We will limit our attention to Eisenstein series
on PSL2 (R) whose associated character of the maximal compact subgroup PSO2 (R) is trivial. That
is, they are functions on the hyperbolic plane H ∼ = PSL2 (R)/ PSO2 (R), which are invariant under the
action of a lattice Γ ⊆ PSL2 (R). The Eisenstein series becomes a function E(w, s) on H × C, defined
for suitable s.
Throughout the text, we will assume that Γ has only one cusp located at ∞ (for definitions, see the
next section). Without this assumption, the notations become heavier and the proofs slightly more
technical. This allows us to focus more on the methods used to prove meromorphic continuation, and
the fundamental difficulties that arise.
One question we need to ask ourselves, is what it means to analytically (or meromorphically) continue
the Eisenstein series: it is a function of two variables, so one can for example interpret holomorphy
to mean that E(w, s) is holomorphic for fixed w, or the stronger property that it is holomorphic as a
function on U with values in the Fréchet space of smooth functions C ∞ (H). That is: that the limit

E(w, s) − E(w, s0 )
lim
s→s0 s − s0
exists as an element of C ∞ (H) (Definition B.4). One can also wonder about additional regularity
conditions. We will show:
Theorem 1.1. The real analytic Eisenstein series as defined in (4.1) has a C ∞ -meromorphic contin-
uation to C, which is jointly smooth away from poles, as a function on H × C. It satisfies a functional
equation of the form
E(w, 1 − s) = φ(s)E(w, s)
for some C-valued meromorphic function φ satisfying φ(s)φ(1 − s) = 1.
Here, C ∞ -meromorphy can equivalently be formulated in terms of Laurent-expansions (Appendix B.4)
or as C ∞ -holomorphy up to a C-valued meromorphic factor (5.29). We will give two proofs: one is due
to Selberg and uses the theory of Fredholm integral equations. The other uses Bernstein’s continuation
principle, although in the second proof we only show the existence of a meromorphic continuation for
a weaker topology on C ∞ (H), the L2loc topology.
Minor refinements aside, my personal contribution to the study of Eisenstein series consists of a
Fredholm theorem for noncompact integral operators (C.5), a study of holomorphic and meromorphic
functions with values in function spaces (Appendix B), as well as a proof of a notorious finite type
condition in an application of Bernstein’s continuation principle (§5.3.4). It is possible that these
results are not new, and are simply hard to find in the literature.

8
2 The hyperbolic plane
2.1 Isometries and geometry
Denote by H = H2 = {z ∈ C : =m z > 0} the upper half-plane. We will often write =m z = y and
talk about “the function y” when what we mean is the imaginary part. We recall some classical facts
about the geometry of H.
Proposition 2.1. The group G = PSL2 (R) acts transitively and faithfully on H by homogra-
phies/Möbius-transformations:  
a b az + b
· z :=
c d cz + d
Moreover, it is the full group of holomorphic automorphisms.

Proof. That this defines a group action on P1 (C) is because the usual action SL2 (R) y A2 (C) is linear,
so it descends to P1 (C), and it is trivial on the subgroup of diagonal matrices {± I}. That H is stable
follows from
 
=m z a b
(2.2) =m(γz) = , γ =
|cz + d|2 c d

To see that the action is transitive, from the above formula we see that a fixed z can be sent to a point
with any imaginary part, after which a horizontal translation can take care of the real part. That it is
the full group of holomorphic automorphisms, follows from Schwarz’s lemma in complex analysis.
Thus, the action PSL2 (R) y P1 (C) decomposes into three orbits:

P1 (C) = H t P1 (R) t −H

Proposition 2.3. The stabilizer of i is

   
cos θ sin θ
K := PSO2 (R) = ± : θ ∈ R /{± I}
− sin θ cos θ

Proof. This can of course be done by a direct computation (or, indeed, using Schwarz’s lemma). We
give a conceptual argument, which explains what’s so special about the point i that we obtain the
rotation group, and not some arbitrary conjugate of it. For γ ∈ SL2 (R), we have γi = γ if and only if
there is λ ∈ C× with    
i i
γ =λ
1 1
that is,       
1 0 i
γ ·i+ ·1 =λ
0 1 1
 


0 1
We would like to substitute 1 = 1 0 and i = 0 1 so that the LHS becomes γ ·
, and
1 0
interpret the RHS by looking at the action of λ ∈ C on C ∼
× 2
= R as an R-linear map. Let’s make this
formal. We have natural isomorphisms of R-vector spaces

C2 ∼
= R2 ⊗ R C ∼
= R 2 ⊗ R R2 ∼
= M2 (R)

where the last is obtained by looking at the dual (R2 ⊗R R2 )∗ as a space of bilinear forms. It sends

t
t
t

a b ⊗ c d to the matrix a b c d . Under these isomorphisms, the action of a matrix
A ∈ SL2 (R) on C2 translates as

LA ↔ LA ⊗ id ↔ LA ⊗ id ↔ LA ◦ RI t

9
where LA denotes left multiplication and RB right multiplication. The action of a complex number
λ ∈ C× is given by
LλI ↔ id ⊗Lλ ↔ id ⊗B ↔ LI ◦ RB t
where B ∈ M2 (R) is the matrix obtained by considering multiplication by λ on C ∼
= R2 as an R-linear
map. Thus B lies in the subgroup of GL2 (R) generated by rotations and homotheties, including those
with negative ratio.
Back to our problem. We obtain, now formally, the equivalent condition that
   
0 1 0 1
γ = B
1 0 1 0
 
0 1
for some B ∈ ± SO2 (R). Taking determinants shows that ± B has to be a rotation. Because
1 0
is a rotation, we finally have that γi = i is equivalent to γ ∈ ± SO2 (R).
By (E.7) we obtain an isomorphism of homogeneous G-spaces
∼
G/K −
→H

We now have (at least) three ways to give H a Riemannian structure:

1. G being a Lie group and K a compact subgroup, it admits a G-left-invariant and K-right-
invariant metric, which is determined by the choice of a positive definite quadratic form q on
Te G: Extend q to a G-left-invariant metric by pushing it forward:

hg := (Lg )∗ q

and integrate over K to obtain a K-right-invariant one:

Z
hg =
e (Rk )∗ hgk dk
K

Then ge defines a G-invariant metric on H. There are many q we could have started with, so we
look for other ways to choose a metric on H.
2. H inherits the Riemannian metric dzdz from the complex structure, which we rescale as
dzdz
y2
Here, dz = dx − idy and dz = dx + idy.
3. We consider H as the 2-dimensional hyperbolic space as defined in (D.7)(c), with metric

dx2 + dy 2
y2

Conformal maps are holomorphic, hence we have a natural inclusion of the orientation preserving
isometries:
Isom+ (H) ⊆ G
It is an equality, as can be seen from the second definition of the metric, the formula (2.2) for the
transformation of imaginary parts, and the formula
 
d(γz) 1 a b
(2.4) = , γ =
dz (cz + d)2 c d

That is, every smooth bijection which preserves angles automatically preserves lengths! Since the
usual topology on G coincides with the compact-open topology, this is an equality of Lie groups, with
Isom(H) given the Lie group structure from Myers–Steenrod (D.47).

10
The Riemannian measure of H becomes
dxdy
dµ(x, y) =
y2

as computed in more generality in (D.43). We state some further results that will be useful in the
sequel.
Proposition 2.5. The geodesics on H are given by half-circles that are orthogonal to the real line,
including vertical lines (which can be seen as degenerate circles). Consequently, through every two
points z = (x1 , y1 ) and w = (x2 , y2 ) there is a unique geodesic segment, the length of which is given
by the hyperbolic distance

(x2 − x1 )2 + (y2 − y1 )2
 
(2.6) d(z, w) = arcosh 1 +
2y1 y2

In particular, the distance is a smooth function of the rational function

(x2 − x1 )2 + (y2 − y1 )2
(2.7) u(z, w) =
2y1 y2
Proposition 2.8. We have the bounds

(2.9) d(z, w) > | log(y1 /y2 )|

(2.10) (x1 − x2 )2 6 (exp(d(z, w)) − 1) · 2y1 y2

Proof. 1. Note that arcosh is increasing, so we can ignore the contribution from the real parts, and
we have
  
1 y1 y2
d(z, w) > arcosh + = | log(y1 /y2 )|
2 y2 y1

Alternatively, let γ = (γx , γy ) : [0, T ] → H be the smooth geodesic segment joining z and w. On
an interval [t0 , t1 ] where γy is monotone, we have for the length:
v
u 2  2
∂γ
u ∂γ
Z t1 u
∂t
x
+ ∂ty
L(γ|[t0 ,t1 ] ) = t
2 dt
t0 γy (t)
 
Z t1  ∂γy 
∂t
> dt
t0 γy (t)
Z γ(t1 )
dy
=
γ(t0 ) y

and we conclude by summing over those intervals. In fact, because γ describes an arc of a circle,
there exists t ∈ [0, T ] for which γy is monotone on [0, t] and [t, T ].
2. The second inequality follows from
p
arcosh(t) = log(t + t2 − 1) > log t

2.2 Group actions and fundamental domains

Proposition 2.11. A subgroup Γ ⊆ G is Fuchsian if the following equivalent conditions hold:
1. Γ is discrete.

11
 2. Γ acts properly discontinuously on H.
Note that in for general continuous group actions, we only have the implication 2 =⇒ 1. The
nontrivial implication follows for example from the fact that G y H is a proper group action. This in
turn follows from the general result (D.49) on isometry groups of metric spaces. It is possible to give
an elementary proof for this particular action Γ y H. See e.g. [Clark, 2018, Theorem 5].
Definition 2.12. A lattice Γ ⊂ G is a Fuchsian group of finite covolume vol(Γ\H). This is in particular
the case when Γ is cocompact.
From now on, Γ will denote a lattice.

Definition 2.13. An open (closed) fundamental domain for Γ is a connected regular open (closed)
set F ⊂ H for which the projection H → Γ\H is:
1. injective when restricted to the interior F ◦ ,
2. surjective when restricted to the closure of F ,

and whose boundary ∂F is of measure zero. An open (closed) fundamental polygon is an open (closed)
fundamental domain which is convex (for the hyperbolic metric) and whose boundary (in P 1 (C))
consists of a finite number of geodesic segments together with a finite number of points in P 1 (R). In
addition, we require the polygon to be locally finite, meaning that every compact set intersects only
finitely many translates γF , γ ∈ Γ.

A convex open set is automatically regular, as is a convex closed set with nonempty interior. One can
show that local finiteness of the fundamental domain is equivalent to requiring that the continuous
bijjection F /G  H/G is a homeomorphism [Beardon, 1983, Theorem 9.2.4].
The precise definition of a fundamental domain is of little importance, the important property is that F
with the hyperbolic measure has the same volume as the quotient Γ\H, thanks to the last requirement.
We present two nice constructions of fundamental polygons:

Proposition 2.14. Let w ∈ H have trivial stabilizer. Then the Dirichlet polygon

D = {z ∈ H : d(z, w) < d(z, γw) ∀γ ∈ Γ − {1}}

is an open fundamental polygon for Γ. Nontrivial fixed points for the action Γ y P 1 (C) lying on
the boundary ∂D are called vertices, and one has that ∂D has an even number of vertices, joined by
geodesic segments.
Let A 6 G be the subgroup of diagonal matrices of determinant 1 and N be the unipotent upper-
triangular group. Recall the Iwasawa decomposition

(2.15) G = N AK

We call an element parabolic (hyperbolic, elliptic) if it is not the identity and conjugate to an element
of N (A, K). They can be characterized in terms of their fixed points or trace:
• parabolic elements have one fixed point, which lies on P 1 (R), and trace equal to 2.
• hyperbolic elements have two fixed points, which lie on P 1 (R), and trace larger than 2.

• elliptic elements have two complex conjugate fixed points, and trace less than 2.
This gives a partition of G − {1} into three sets. Their fixed points are called parabolic, elliptic or
hyperbolic accordingly. A cusp for Γ is an orbit of parabolic fixed points.
One shows by direct calculation that two elements of G commute if and only if they have the same
fixed points, and that they must lie in the same conjugate of N , A or K. Because discrete subgroups
of N , A and K are cyclic, this shows that stabilizers are cyclic. We see that:

12
Proposition 2.16 (Elliptic fixed points). There is a finite number of elliptic orbits under Γ. The
stabilizer of an elliptic point is a finite cyclic group consisting of elliptic elements. There exist elliptic
fixed points if and only if Γ has no elliptic elements, which is the case if and only if the projection
H  Γ\H is a covering map.
From the theory of Riemann surfaces, it follows that:
Corollary 2.17. Γ\H is a Riemann surface and the projection onto it is a ramified holomorphic
covering map, whose ramification points are precisely the elliptic fixed points of Γ.
Because G acts by orientation-preserving isometries, we see that Γ\H is an oriented Riemannian
manifold, and the metric and volume form on H descend to the quotient. In particular, geodesics
descend to the quotient, and by Hopf–Rinow (D.39) it is still a complete Riemannian manifold.
One can show that:
Proposition 2.18 (Cusps). There is a finite number of cusps for Γ. The stabilizer of a parabolic
point is an infinite cyclic group consisting of parabolic elements. There are no cusps if and only of Γ
has no parabolic elements, which is the case if and only if Γ is cocompact.
Let a be a cusp for Γ. There exists σa ∈ G with σa ∞ = a, so that ∞ is a cusp for σa−1 Γσa , and its
stabilizer Γ∞ is a cyclic group generated by a parabolic element of the form
 
1 t
, t 6= 0
0 1

Because −1 
s−1
  −1
s2 t
   
0 1 t s 0 1
=
0 s 0 1 0 s 0 1
we may assume t = 1, so that Γ∞ is generated by
 
1 1
T :=
0 1

Proposition 2.19. [Ford, 1929] Suppose ∞ is a cusp for Γ, with stabilizer generated by T . Let β ∈ R
arbitrary and let F∞ = {z ∈ H : <e z ∈ (β, β + 1)}. Then the set

(2.20) F = {z ∈ F∞ : |cz + d| > 1 ∀γ ∈ Γ − Γ∞ }

is an open fundamental polygon. That is, it consists if the points in the vertical strip F∞ that are
exterior to the (Euclidean) half-circles with center −d/c and radius |c|−1 , for
 
a b
γ= ∈ Γ − Γ∞
c d

We will refer to such a domain (with, say, β = 0) as the standard fundamental domain. One shows
that two translates of F share either an elliptic fixed point or one side (which may contain a fixed
point), and that the transformation γ ∈ Γ sending F to an adjacent fundamental domain, fixes the
point or fixes (setwise) the side they share.
Corollary 2.21. Suppose ∞ is a cusp for Γ, with stabilizer generated by T . Then
   
a b
c∞ := inf |c| : ∃ ∈ Γ − Γ∞ > 0
c d

Proof. An element of the set Γ∞ \Γ is determined by its bottom row. Denote it by (c, d). If (c, d) ∈
Γ∞ \Γ, then so is (c, d + nc) for n ∈ Z. Take any β ∈ R, say β = 0, and consider the fundamental
domain F from the previous proposition. Replacing d by d + nc if necessary, we know that there is a
half-disk of radius |c|−1 centered at a point on [0, 1], which is disjoint from F . If c could be arbitrarily
small, then F would be empty.

13
If Γ ⊆ SL2 (Z), then trivially c∞ > 1.
Corollary 2.22. Let z = x + iy ∈ H with y > 0 and γ ∈ Γ − Γ∞ . Then, with c∞ as above:
1. =m(γz) 6 1/(c2∞ y).

2. d(z, γz) > 2 log(c∞ y) for c∞ y > 1.

 
a b
Proof. Let γ = .
c d

1. We have =m(γz) = y/|cz + d|2 and |cz + d| > |c|y > c∞ y because c 6= 0.

2. By (2.9),  
 y  > log(c2∞ y 2 )

d(z, γz) > log 
=m(γz) 

Note how the first inequality implies the following: there exists a neighborhood V of the cusp ∞, such
that unless γ ∈ Γ∞ , we have γV ∩ V = ∅. It is almost like saying that the action of Γ extends properly
discontinuously to the cusp.

2.3 Examples
The prototypical example of a lattice is Γ = PSL2 (Z). If Γ has coefficients contained in a field K, then
so are its cusps. Thus because the action PSL2 (Z) y P 1 (Q) is transitive, there is precisely one cusp,
which we call ∞. There are two elliptic orbits: one containing i, with stabilizer of order 2, and one
containing j = exp(2πi/3), with a stabilizer of order 3. The fundamental domain from (2.19) becomes,
with β = −1/2:   
−1 1
z = x + iy ∈ H : |z| > 1, x ∈ ,
2 2
A subgroup Γ0 of finite index in a lattice Γ is again a lattice: it has a fundamental domain which is
a finite union of fundamental domains for Γ. In particular one can consider congruence subgroups of
level N ∈ N, they are subgroups of PSL2 (Z) which contain the kernel Γ(N ) of the surjective reduction
homomorphism
PSL2 (Z)  PSL2 (Z/N Z)
The construction of the lattice PSL2 (Z) ⊂ PSL2 (Q) generalizes as follows:
Definition 2.23. A lattice Γ ⊂ PSL2 (R) is arithmetic if, when L denotes the set of traces of elements
of Γ (which are well-defined up to sign), then
(a) Q(L) is a finite extension of Q and L ⊆ OK .

(b) If φ : K → C is an embedding such that φ(L) is unbounded, then φ(t) = ± t for all t ∈ L.
All arithmetic lattices can be constructed out of quaternion algebras over number fields. The above
characterization is due to Takeuchi [Takeuchi, 1975]. In particular, the set of arithmetic lattices is
countable, while the set of all lattices is uncountable.

2.4 Fourier expansions

Let Γ ⊂ G be a lattice, with a cusp at ∞ and with stabilizer Γ∞ generated by T . Let f : H → C be
a smooth Γ-invariant function. Then in particular f (x + iy) is Γ∞ -invariant for fixed y, and we can
Fourier-expand it:

14
Proposition 2.24. There exist functions fbn (y) for n ∈ Z, with
X
(2.25) f (z) = fbn (y)e(xn)
n∈Z

where we denote e(u) = exp(2πiu) for brevity. They are given by

Z 1
(2.26) fbn (y) = f (x + iy)e(−nx)dx
0

We call fb0 (y) the constant term, and we denote it also by Cf . Perhaps misleadingly, it is not a constant
function.
When Γ has a cusp a 6= ∞, we can give a similar Fourier expansion: take σa ∈ G with σa ∞ = a, so
that ∞ is a cusp for σa−1 Γσa . The Fourier expansion at ∞ w.r.t. this lattice will serve as the Fourier
expansion at the cusp a. In order to avoid using heavy notation, we will almost always assume that Γ
has a cusp at ∞.
We see from (2.26) that the fbn (y) are smooth. Now suppose that f is an eigenfunction of the Laplacian
−∆ with eigenvalue λ. By Fourier expanding f , we have separated the variables x and y, so that we
expect to obtain differential equations in the fbn (y). Indeed, with the formula for −∆ (D.46), we find:
Proposition 2.27. For all n ∈ Z, fbn (y) is a solution of the differential equation
y 2 F 00 + (λ − 4π 2 n2 y 2 )F = 0
Proof. By repeated partial integration, we have that the Fourier coefficients are rapidly decreasing, in
a locally uniform way when y varies:
Z 1
|fbn (y)| 6 |2πn|−p |f (p) (x + iy)|dx ∀p > 0, n 6= 0
0

(k)
and similarly for the derivatives We may thus apply −∆ termwise to the Fourier expansion,
fbn (y).
and we obtain X
−∆f (z) = − y 2 (fbn(2) (y) − 4π 2 n2 fbn (y))e(xn)
n∈Z

The claim follows by comparing the Fourier coefficients (which are unique) with those of
X
λf (z) = λfbn (y)e(xn)
n∈Z

For the constant term, we find:

Proposition 2.28. Let λ = s(1 − s) with s ∈ C. Then the constant term Cf is a linear combination
of (
y s and y 1−s : s 6= 21
y s and y s log y : s = 21
Proof. By inspection, those are linearly independent solutions to the differential equation F 00 =
−λy −2 F .
Now consider n 6= 0. Substituting u = 2π|n|y gives the following equation for G(u) = fbn (u/(2π|n|)):
G00 (u) + (λu−2 − 1)G(u) = 0
This differential equation is studied in (H): we have G(u) = W (2u) for
some Witthaker function W .
More precisely, let λ = s(1 − s) so that λ = 14 − m2 with m = ± s − 21 . Suppose <e s > 0. Then

fbn (y) = c1 W0,s−1/2 (4π|n|y) + c2 W0,s−1/2 (−4π|n|y)

is a linear combination of the Witthaker functions from (H.6). Using a more careful analysis of
Whittaker functions, one can generalize this to all s, with the condition <e s > 0.

15
Proposition 2.29. Suppose in addition that f (z) = o(e2πy ) as y → ∞, uniformly in x. Then the
same asymptotic holds for the fbn (y), and we have

fbn (y) = an W0,s−1/2 (4π|n|y)

for some an ∈ C.
Proof. That the fbn (y) satisfy the same asymptotic relation, follows from their definition (2.26). Now,
by (H.6), the Whittaker function W0,s−1/2 (−4π|n|y) is asymptotically equivalent to e2π|n|y as y → ∞,
hence fbn (y) is a linear combination of W0,s−1/2 (4π|n|y) alone.
Corollary 2.30. When f (z) = o(e2πy ) as y → ∞, uniformly in x, is an eigenfunction of −∆ with
eigenvalue s(1 − s), then f (z) − Cf = O(e−2πy ) as y → ∞, uniformly in x.

Proof. We have fbn (y) = On (e−2π|n|y ) for all n 6= 0, because it is a scalar multiple of a decaying
Whittaker function. We want to get rid of the dependence on n, by estimating the coefficients an .
Recall that Z 1
|fbn (y)| 6 |f (x + iy)|dx
0
Evaluating in small y gives
|an |  e|n|
so that f (z) − Cf is bounded by the geometric series
X
e|n| e−2π|n|y  e−2πy ,  − 2πy < 0
n6=0

We conclude by noting that any  < 2π works for all y > 1.

16
3 Operators on symmetric spaces
When M is a smooth manifold, one wants to understand linear operators on C ∞ (M ). Differential
operators in particular, but not unrelated is the notion of convolution operators: It is a particularly
useful tool because it gives a way to regularize non-smooth functions, and because of the existence of
approximations of the identity: Define a compactly supported smooth function ρ on Rn by
  
exp − 1
: |x| < 1

(3.1) ρ(x) = c · 1 − |x|2

0 :x>1
R
where c > 0 is used to normalize the function so that Rn ρ = 1. Define

(3.2) ρδ (x) = δ −n ρ(x/δ)

ρδ = 1. For f ∈ L1 (Rn ) resp. f ∈ C(Rn , C) the convolution ρδ ∗ f
R
We then have for all δ > 0 that Rn
is defined and
ρδ ∗ f → f
for L -convergence resp. locally uniform convergence. From now on let f, g ∈ C ∞ (Rn ) be smooth and
1

observe that the identity

Z Z
(f ∗ g)(x) = f (x − y)g(y)dy = f (y)g(x − y)dy
Rn Rn

(valid whenever any of the integrals converges) does not only show commutativity of the convolution,
but also tells us something about the action of translation-invariant differential operators (as defined
in (F.27)): Let D = Dα be a monomial in the ∂/∂xi (for the notation, see (F.1)), then
Z Z
D(f ∗ g) = (Df )(x − y)g(y)dy = f (y)(Dg)(x − y)dy
Rn Rn

That is:
(3.3) D(f ∗ g) = (Df ) ∗ g = f ∗ (Dg)
whenever we can switch the order of integration and differentiation, which is for example the case
when either f or g has compact support or when they are both in the Schwartz space
S(Rn ) = f ∈ C ∞ (Rn ) : |Dα f (x)|  |x|−M ∀α ∈ Nn , M > 0


The existence of approximations of the identity then allows an unusual proof of the following
Proposition 3.4. Let f : U → R have continuous mixed partial derivatives of order 2 on an open set
U ⊆ R2 . Then
∂2f ∂2f
=
∂x1 ∂x2 ∂x2 ∂x1
on U .
Proof. The usual proof presented in a multivariable calculus course goes along the following lines: Fix
(a, b) ∈ U , let I = [a−, a+]×[b−, b+] and use Fubini’s theorem to choose the order of integration
in
∂2f ∂2f
ZZ ZZ
dx2 dx1 and dx1 dx2
I ∂x1 ∂x2 I ∂x2 ∂x1
Use the fundamental theorem of calculus to show that the integrals are equal. Then use a mean value
theorem to conclude that, as  → 0:
 2   2 
∂ f ∂ f
(2)2 (a, b) + o(1) = (2)2 (a, b) + o(1)
∂x1 ∂x2 ∂x2 ∂x1
Divide by (2)2 and let  → 0.

17
We present here two other proofs.
First proof. Write D1 = ∂/∂x1 and D2 = ∂/∂x2 . Then

D1 D2 (ρδ ∗ f ) = D1 (ρδ ∗ D2 f ) = ρδ ∗ D1 D2 f

on the one hand and

D1 D2 (ρδ ∗ f ) = D1 (D2 ρδ ∗ f ) = D2 ρδ ∗ D1 f = ρδ ∗ D2 D1 f

on the other. Taking the point-wise limit for δ → 0 gives the result!
Second proof. We can alter the argument slightly as follows: By commutativity of convolution:

(3.5) D1 ρδ1 ∗ D2 ρδ2 ∗ f = D2 ρδ2 ∗ D1 ρδ1 ∗ f

for δ1 , δ2 > 0. We want to let δ1 , δ2 → 0. This is tricky. We may assume that f has compact support,
without changing f on a sufficiently small open set, by multiplying it with a bump function. Then
taking limits is allowed:
Fix δ2 and let δ1 → 0. The LHS converges (pointwise) to D1 (D2 ρδ2 ∗ f ), there is no problem. The
RHS is a problem of interchanging limit and integral. For the RHS, we have that D1 ρδ1 ∗ f → D1 f
locally uniformly, and thus uniformly because its support is contained in B(supp f, 1). In particular,
it is uniformly bounded as δ1 → 0, and we can apply dominated convergence to justify the pointwise
convergence
D2 ρδ2 ∗ (D1 ρδ1 ∗ f ) → D2 ρδ2 ∗ D1 f (δ1 → 0)
where D2 ρδ2 ∈ L1 (Rn ) is used as a majorant. We obtain

D1 (D2 ρδ2 ∗ f ) = D2 ρδ2 ∗ D1 f

We could also have obtained this by using that convolution is continuous as a map L1 × L1 → L1 (by
Young’s inequality).
We now want to let δ2 → 0. The convergence of the RHS is no problem. For the LHS, the problem is
to interchange limit and differentiation. We see from the RHS that this converges locally uniformly as
δ2 → 0. We also have that D2 ρδ2 ∗ f converges locally uniformly, in particular, at at least one point.
This means that we interchange limit and D1 in the LHS :

D1 (D2 ρδ2 ∗ f ) → D1 (D2 f ) (δ2 → 0)

pointwise, and in fact locally uniformly.

While a lot less elementary (we use the dominated convergence theorem when interchanging integra-
tion and differentiation), the two new proofs provide a framework for possible generalizations. The
difference between them is subtle. The first proof relied on the nature of the action of differentiation
as summarized in (3.3). In the second proof we only used the observation that Dρδ ∗ f = ρδ ∗ Df , and
instead of exploiting this as in the first proof, we used the fact that convolution is commutative.

3.1 Weakly symmetric spaces

One can hope to generalize these powerful tools to Riemannian manifolds other than Rn , to study
commutativity of differential operators invariant under some group of isometries. The difficulty lies in
finding an analogue of convolution, satisfying a property similar to (3.3). When G is a Lie group with
(say) left-invariant Haar measure µ and f, g ∈ C ∞ (G, C) such that at least one of them has compact
support, we can define Z
(f ∗ g)(x) = f (xy −1 )g(y)dµ(y)
G

When D is a differential operator that is invariant under right translations, we have that D(f ∗ g) =
(Df ) ∗ g. But unless G is abelian, we do not have f ∗ g = g ∗ f . We take a second look at how (3.3) is

18
derived, for a differential operator D on Rn . One way is to make a substitution in the integral, but we
don’t expect to have access to that in general. We try to forget the fact that x − y is a subtraction,
and look at its symmetries instead. Let k(x, y) = f (x − y) : Rn × Rn → R. It is almost symmetric:
when µ : x 7→ −x denotes the inversion, then k(x, y) = k(µy, µx), and we have, denoting the argument
on which D acts with a subscript,
(Dx k)(x0 , y0 ) = (Dx k(µy, µx))(x0 , y0 )
= (Dx k(µy0 , µx))(x0 )
= ((µ∗ D)y k(µy0 , y))(µx0 )
= (µ∗ D)y k(µy0 , µx0 )
Where µ∗ D is the pushforward of D by µ, which is the differential operator defined by
(µ∗ Df )(x0 ) = D(f (µx))(µ−1 x0 )
as in Appendix F. Now of course, in this case Dx k(x0 , y0 ) is just Dx k(µy0 , µx0 ), but we look for a
conceptual way to understand this. Let G be the translation group on Rn . Note that k is invariant
for the under the diagonal action G y Rn × Rn on point pairs:
Definition 3.6. Let S be a Riemannian homogeneous G-space. A point-pair invariant on S is a
smooth function k : S × S → C with
k(σx, σy) = k(x, y) , ∀σ ∈ G
Because Dx is G-invariant by assumption, Dx k is again a point-pair invariant. Note also that for all
x0 , y0 ∈ Rn , there exists σ ∈ G such that both σx0 = µy0 and σy0 = µx0 : it suffices to take σ to be
the translation by −x0 − y0 . We conclude that
Dx k(x0 , y0 ) = Dx k(σx0 , σy0 ) = Dx k(µy0 , µx0 )
and consequently,
(Dx k)(x0 , y0 ) = ((µ∗ D)y k(x, y))(x0 , y0 )
We have managed to shift the operator from the first to the second argument, by replacing it with its
pullback. Now let D(1) and D(2) be G-invariant differential operators on Rn and note that:
Dx(1) Dx(2) k = Dx(1) (µ∗ D(2) )y k
= (µ∗ D(2) )y Dx(1) k
= Dx(2) Dx(1) k
(1)
where in the last step we used that Dx k is again a point-pair invariant. That is, the action of
translation-invariant differential operators on translation-point-pair invariants is commutative.1
We are not done yet. We still need a notion of approximations of the identity, and an analogue for the
identity Df ∗ g = f ∗ Dg. The first ingredient is not a problem:
Definition 3.7. Let S be a homogeneous Riemannian manifold with distance function d. By homo-
geneity, there exists  > 0 such that B(x, ) is a normal neighborhood of x for each x ∈ S. By (D.37),
the function d(x, y)2 is smooth on the set of point-pairs that are at distance less than . For δ < ,
define now the smooth function, analogous to (3.2):
  
1


−

exp 
2  : d(x, y) < δ

(3.8) ρδ (x, y) = cδ ·

d(x,y)

 1− δ

0 : d(x, y) > δ

R
where cδ is chosen to have S ρδ (x, y)dy = 1 for all x.
1 It is important to note that, in fact, we are cheating here: when switching the order of differentiation, we are using

the fact that the differential operators Dx and Dy0 on Rn × Rn commute, for given differential operators D, D0 on Rn ,
while that’s precisely what we’re trying to show! That is, this proof is circular. But it gives a way to reduce the statement
about more general manifolds, to the case of Rn .

19
Proposition 3.9. For f ∈ C(S, C), we have

ρδ ? f → f

locally uniformly.

Proof. The proof is the same as for convolution on Rn . Write

Z
ρδ ? f − f = ρδ (x, y)(f (y) − f (x))dy
S
6 sup |f (y) − f (x)|
y∈B(x,δ)

and we conclude using Heine’s theorem, which says that f is locally uniformly continuous.
The previous discussion gives rise to a notion of weakly symmetric spaces:

Definition 3.10. [Selberg, 1956] A weakly symmetric (Riemannian) space is a triple (S, G, µ) with S
is a Riemannian manifold, G a locally compact transitive group of isometries of S and µ an isometry
(not necessarily in G) such that:
1. µ2 ∈ G

2. µGµ−1 = G
3. for all x, y ∈ S there exists σ ∈ G with

σx = µy , σy = µx

The prime example is Rn with G the translation group and µ : x 7→ −x the reflection.
This gives rise to a large class of Riemannian manifolds on which the algebra of invariant differential
operators is commutative. Because we won’t need those general results, and in order to avoid technical
difficulties, we will limit our attention to the smaller class of symmetric spaces (E.9) without attempting
to state the results under minimal hypotheses.

3.2 Point-pair invariants

Proposition 3.11 (Invariant differential operators applied to point-pair invariants). Let S be a sym-
metric space.
1. Applying an invariant differential operator D ∈ D(S) to either argument of k yields again a
point-pair invariant.
2. We have an unambiguous action of D(S) on point-pair invariants: if we denote the action of D
on the first argument by Dx and on the second by Dy , then

Dx k(x0 , y0 ) = Dy k(x0 , y0 )

3. The action of D(S) on point-pair invariants is commutative: for invariant differential operators
D1 , D2 ∈ D(S) we have:
D1 D2 k = D2 D1 k

Proof. 1. We have

Dy k(x0 , y0 ) = Dy (k(σx0 , σy))|y=y0 = (Dy k)(σx0 , y)|y=σy0

for σ ∈ G, where the last equality follows from the invariance of D. Similarly for the first
argument.

20
 2. Because k is symmetric, we have

Dx k(x0 , y0 ) = Dx (k(y, x))|x=x0 ,y=y0 = (Dy k)(y0 , x0 )

where the last equality follows from the chain rule. Because Dy k is symmetric, this equals
Dy k(x0 , y0 ).
3. Follows from
D1 D2 k = D1,x D2,y k = D2,y D1,x k = D2 D1 k
Definition 3.12 (Point-pair invariant of compact support). Let S be a symmetric space with isometry
group G. A point pair-invariant k on S has compact support if the following equivalent conditions hold:
1. There exists y0 ∈ S such that k(·, y0 ) has compact support.
2. For all y ∈ S, the function k(·, y) has compact support.
3. For all y0 ∈ S, there exists a neighborhood U of y0 and a compact T ⊆ S such that the support
of k(·, y) is contained in T for all y ∈ U .
4. For compact V ⊆ S, the restriction k : V × S → R has compact support.
We denote by A(S) the set of point-pair invariants of compact support.
Proof of equivalence. 1 =⇒ 2: Let σ ∈ G with σy0 = y, then

supp k(·, y) = σ · supp k(·, y0 )

is compact.
2 =⇒ 3: By (E.8) with x0 = y0 there exists a compact neighborhood U of y0 and a homeomor-
phism on its image ψ : U → G such that ψ(y)y0 = y for all y ∈ U . Then

supp k(·, y) = ψ(y) supp k(·, y0 ) ⊆ ψ(U ) · supp k(·, y0 )

which is compact, because ψ(U ) and supp k(·, y0 ) are.

3 =⇒ 4: We can cover V by a finite number of relatively compact open sets Ui such that the
restriction of k to Ui × S has support contained in aScompact set U i × Ti . Then the restriction
to V × S has support contained in the compact set U i × Ti .
4 =⇒ 1: Take V = {y0 }.
Remark 3.13. In general, for a smooth function k : M × M → R these conditions are not equivalent.
That is, compact support of k(y, ·) at each point y does not imply that locally in y the support lies
in one uniform compact set. Consider for example M = R and k : R × R which has smooth bumps on
each of the rectangles  
1 1
, × [2n, 2n + 1] (n > 0)
2n+1 2n
and is 0 elsewhere. Then all k(y, ·) have compact support, k(0, ·) has empty support, but the restriction
of k to [−δ, δ] × R has unbounded support for all δ > 0.
Definition 3.14. Consider the set A(S) of point-pair invariants of compact support. We equip it with
pointwise addition and the multiplication
Z
(k1 ◦ k2 )(x, y) = k1 (x, w)k2 (w, y)dw
S

This makes A(S) a commutative algebra:

Proposition 3.15. The composition k1 ◦ k2

21
 1. is smooth.
2. is a point-pair invariant.
3. has compact support.

4. satisfies k1 ◦ k2 = k2 ◦ k1 .
Proof. 1. Because k1 , k2 have compact support, locally in x and y the integrand has compact
support, hence defines a smooth function.
2. Because G acts by isometries:
Z Z
k1 (σx, w)k2 (w, σy)dw = k1 (x, σ −1 w)k2 (σ −1 w, y)dw
S S
Z
= k1 (x, w)k2 (w, y)dw
S

3. Fix x, then the support of k1 (x, ·) is compact, call it T . Because T is compact, the support of
k2 : S × T → R (as a function, not as a point-pair invariant) is also compact, hence the integrand
is identically 0 for y outside of a compact set.
4. Because point-pair invariants are symmetric,
Z Z
k1 (x, w)k2 (w, y)dw = k1 (w, x)k2 (y, w)dw
S S

Many results about composition of point-pair invariants carry through if at least one of them has
compact support. We will at times use such results without explicitly mentioning that they the
conditions are not strictly satisfied because of issues with supports.

3.3 Radially symmetric functions

Let S be a symmetric space with isometry group G. We have a right group action of G on real or
complex-valued functions by Lg f (x) = f (gx).2

Definition 3.16. Let S be a symmetric space with isometry group G, let x0 ∈ S and f ∈ C ∞ (S)
smooth. Then f is radially symmetric about x0 if it is invariant under the stabilizer of x0 in G.
Example 3.17. If k is a point-pair invariant, then k(·, x0 ) is radially symmetric about x0 .

We can make every function on a symmetric space radially invariant about a point, as follows. The
stabilizer K of x0 is a compact Lie group by (D.50). It has a unique right-invariant Haar measure µ
such that K has volume 1. (And because compact groups are unimodular, it is in fact bi-invariant.)
Definition 3.18. The symmetrization of f about x0 is
Z
fxrad
0
(x) = f (x, x0 ) = f (rx)µ(dr)
K

Proposition 3.19 (Properties of symmetrization). Let M be a Riemannian manifold, x0 ∈ M with

stabilizer K, and f 7→ fxrad
0
denote the symmetrization map.
1. f (x, x0 ) is radially symmetric.
2. f is radially symmetric if and only if f (x) = f (x, x0 ).

3. f (x0 , x0 ) = f (x0 ).
2 Sometimes denoted Rg .

22
 4. If f is smooth, so is f (x, x0 ).
5. If D is a K-invariant differential operator, then (Df )(·) = D(f (·, x0 )).
Proof. 1. It is radially symmetric because for g ∈ K:
Z Z Z
f (gx, x0 ) = f (rgx)µ(dr) = f (rx)(Rg )∗ µ(dr) = f (rx)µ(dr)
K K K

because µ is right-invariant, where Rg : h 7→ hg denotes the right regular representation of g ∈ K.

R
2, 3. Because K dµ = 1.
4, 5. Because K is compact and f smooth, we can switch integration and differentiation.
There is a converse for (3.17). Each radially symmetric k(x, x0 ) can be extended to a point-pair
invariant on S × S, as follows. Take (x, y) ∈ S × S, and let σ ∈ G such that σy = x0 . Define

k(x, y) := k(σx, x0 )

Proposition 3.20. k is
1. well-defined, i.e. does not depend on the choice of σ.
2. a point-pair invariant.
3. smooth.
Proof. 1. It is well-defined precisely because we assume that k(x, x0 ) is radially symmetric.
2. This follows from the construction and well-definedness: let x, y ∈ S, σy = x0 and τ ∈ G. Then
(στ −1 )(τ y) = x0 and hence:

k(τ x, τ y) := k(στ −1 τ x, x0 ) = k(σx, x0 ) = k(x, y)

3. By (E.8), we can choose σ in a smooth way as a function of y, in a neighborhood of any y0 .

To summarize, we have the following correspondence:
Theorem 3.21. For each point x0 ∈ S there is a bijection between radially invariant functions g(x)
and point-pair invariants k(x, y). Under this bijection:
1. g(x) = k(x, x0 ).
2. Dg corresponds to Dk for D ∈ D(S), the action of D on k being unambiguous by (3.11).

3.4 Isotropic spaces

In the case of isotropic Riemannian manifolds, such as H, radially symmetric functions are easiest to
understand:
Proposition 3.22. Let S be an isotropic symmetric space. Then there exists δ > 0 such that:
1. Every radially symmetric function f around x0 ∈ S is locally a function of the radial distance
r to x0 : there exists g : R>0 → R with f = g ◦ r in the geodesic ball B(x0 , δ). Moreover, g is
smooth on [0, δ).
2. For every point-pair invariant k there exists a smooth g : [0, δ) → R such that for all x, y ∈ S
with 0 < d(x, y) < δ we have k(x, y) = g(d(x, y)). Moreover, g is smooth on [0, δ).
Proof. By (E.2), G acts transitively on sufficiently small geodesic spheres around x0 , say those of
radius < δ.

23
 1. Thus f is locally a function of r, on B(x0 , δ). We want that g is smooth. Note that r : B(x0 , δ) →
[0, δ) has a smooth section: Take a normal coordinate chart φ : B(x0 , δ) → Rn (§D.4.1) and define
h(r) = φ−1 (r, 0, . . . , 0). Then g = f ◦ h is smooth.
2. For fixed x, the function k(x, ·) is radially symmetric, and the above applies.
Remark 3.23. 1. One can show that g is a smooth function of r2 resp. d(x, y)2 . This follows from
the fact that a smooth even function f (x) on R is a smooth function of x2 (3.24).
2. In the case of H, every two points are joined by a unique geodesic, hence by (D.41) the exponential
maps are global diffeomorphisms, and we can take δ = ∞ in the above proposition.
Proposition 3.24. Let f : R → R be a smooth and even function. Then there exists a smooth
g : R>0 → R with f (x) = g(x2 ).
Proof. See the Mathoverflow post [Mathoverflow, 2011] for various proofs.
Proposition 3.25. Let M be a Riemannian manifold, x0 ∈ M , (xi ) normal coordinates at x0 and r
the radial distance to x0 , defined on an open neighborhood U of x0 . Let f : U − {x0 } → R be smooth
with f = h ◦ r for some h : R>0 → R. Then

−∆f = D(h) ◦ r

for some differential operator D of degree 2, whose highest degree coefficient is nonzero for t ∈ R>0
sufficiently small.
Proof. Note that h is automatically smooth by (3.22). That such D of degree 2 exists follows imme-
diately from (D.45) and the chain rule. Let’s look at the highest degree coefficient. If g is the metric
and g ij the components of the inverse of its matrix in the coordinates (xi ), then the highest degree
coefficient is
1 X ij i j
g xx
r2 i,j

Because g ij = δij at x0 , by continuity this is close to r−2 i (xi )2 = 1 in a neighborhood of x0 , hence

P
nonzero.

3.5 Integral operators

On any measure space X we can consider integral operators on L2 (X) of the form
Z
f 7→ k(z, w)f (w)dw
X

2
where k ∈ L (X × X) is a kernel. See (A.7). We will denote the action on f by k ? f and the operator
itself simply by k. If X = S is a symmetric space, k1 , k2 are compactly supported point-pair invariants
and k1 ◦ k2 their composition from (3.14), then by Fubini:

(k1 ◦ k2 ) ? f = k1 ? (k2 ? f )

Integral operators whose kernel is a point-pair, are related to radially symmetric functions as follows:
Proposition 3.26. Let S be a symmetric space, x0 ∈ S, g radially symmetric about x0 , f ∈ C ∞ (S)
and k ∈ A(S) a compactly supported point-pair invariant. Let h be the point-pair invariant extension
of g from (3.21). Then:
1. Convolution commutes with the bijection between point-pair invariants and radially symmetric
functions: k ◦ h is the point-pair invariant associated to k ? g.
2. Convolution commutes with symmetrization:

(k ? f )rad rad
x0 = k ? fx0

24
Proof. 1. Because
Z Z
(k ◦ h)(x, x0 ) = k(x, y)h(y, x0 )dy = k(x, y)g(y)dy = (k ? g)(x)
S S

we have that k ◦ h is necessarily the unique extension of k ? g.

2. Let R be the isotropy subgroup of x0 , then the RHS evaluated in x is
Z Z Z Z
k(x, y) f (ry)drdy = k(x, y)f (ry)dydr
S R
ZR ZS
= k(x, r−1 y)f (y)dydr
R S
Z Z
= k(rx, y)f (y)dydr
R S

where we were allowed to apply Fubini because by assumption on k and compactness of R, the
integrands have compact support.
Proposition 3.27 (Convolution and invariant differential operators). For k ∈ A(S) compactly sup-
ported, D ∈ D(S) invariant and f ∈ C ∞ (S):

(Dk) ? f = D(k ? f ) = k ? Df

Proof. The first equality is immediate, because we can differentiate under the integral sign:
Z Z
Dx k(x, y)f (y)dy = Dx k(x, y)f (y)dy
S S

For the second equality, fix x0 ∈ S. We may suppose f is radially symmetric about x0 ∈ S, because
D and k? commute with symmetrization ((3.19).5) and (3.26).2) and we haven’t changed the value of
any of the involved functions at x0 .
Under the bijection from (3.21), Say f corresponds to h. Then we have the correspondences:

k ? Df ↔ k ◦ Dh
Df ↔ Dh
f ↔h
k?f ↔k◦h
D(k ? f ) ↔ D(k ◦ h)

by (3.26)(1) and (3.21). Because D(k ◦ h) = k ◦ Dh, it follows that k ? Df = D(k ? f ).

3.6 The algebra of invariant differential operators

We continue the discussion in the very beginning of this section. For real Lie groups G, we showed
that the algebra of left-invariant differential operators is isomorphic to its universal enveloping algebra
U (g) (F.39) with the Poincaré–Birkhoff–Witt theorem as a corollary. One can show that the algebra of
bi-invariant differential operators is isomorphic to the center of U (g). In particular, it is commutative.
The same conclusion holds for symmetric spaces. We give two proofs that are very different from the
proof for Lie groups. They mimic the two proofs of the analogous result for Rn (3.4).
Theorem 3.28. Let S be a symmetric space. Then the algebra of invariant differential operators
D(S) is commutative.
First proof. [Selberg, 1956] Let f be a smooth function and x0 ∈ S. By (3.19) and (3.21) there
exists a point-pair invariant k such that Dk(x0 , x0 ) = Df (x0 ) for all D ∈ D(S). By (3.11), we have
D1 D2 f (x0 ) = D2 D1 f (x0 ) for all D1 , D2 ∈ D(S).

25
The ingredients for the proof were: homogeneity of S, the compactness of the isotropy groups of one
(hence every) point of S, and the fact that every two points are switched (i.e. that point-pair invariants
are symmetric.)
It is stable by the action of D(S) on point-pair invariants by differentiation with respect to the first
variable. We will assume that the action is on the first variable in the proof below.

Second proof of (3.28). For small δ > 0 we have a compactly supported point-pair invariant ρδ (x, y)
that is an ‘approximation of the identity’, and for smooth f :
Z
ρδ (x, y)f (y)dy → f (x) as δ → 0
S

where the convergence is locally uniformly (3.9). Take invariant D1 , D2 ∈ D(S) and small δ1 , δ2 > 0.
Then D1 ρδ1 and D2 ρδ2 commute. We have:
Z Z Z
(D1 ρδ1 ◦ D2 ρδ2 )(x, y)f (y)dy = D1 ρδ1 (x, w) D2 ρδ2 (w, y)f (y)dydw
S S S

by Fubini, because for fixed x the integrands have compact support. It remains to show that we
can interchange limits, differentiation and integration to show that this approaches (D1 D2 f )(x) as
δ1 , δ2 → 0. By symmetry, we then have (D1 D2 f )(x) = (D2 D1 f )(x). The validity of this interchanging
can be checked in the exact same way as we did in the second proof for differential operators on Rn
(3.4).
With more representation theory, it is possible to give a proof of the same flavor as the proof for Lie
groups. Because S is homogeneous, we can write it as a quotient of a Lie group G by a compact
Lie group K. The Harish-Chandra homomorphism provides a map from a commutative algebra con-
structed from the Lie algebras of g and h, to the algebra D(G/K) of invariant differential operators
on S. See e.g. [Helgason, 1984, Theorem 5.13].

3.7 Selberg’s eigenfunction principle

Point-pair invariant integral operators have another very useful property. Recall that they respect
radial symmetry (3.26), preserve smoothness and commute with invariant differential operators (3.27).
The idea of the Selberg eigenfunction principle is to exploit these properties by letting point-pair
invariants act on a vector space of dimension 1.
Example 3.29. On Hn+1 , for s ∈ C, the functions y s and y n−s are eigenfunctions for the Laplacian
with eigenvalue s(n − s).
Proof. From the formula from (D.46) for −∆.
Proposition 3.30 (Spherical eigenfunctions). Let S be an isotropic Riemannian manifold, z0 ∈ S
and λ ∈ C.
1. There exists a punctured open neighborhood V of z0 such that every radially symmetric eigen-
function for −∆ defined on some open subset of V extends globally to V in a unique way, and
the kernel ker(∆ + λ) has complex dimension exactly 2 on V .
2. Let S = H and λ = s(1 − s) with s ∈ C. Then there is a unique such eigenfunction ωs (z, z0 )
that extends continuously to z0 with ω(z0 , z0 ) = 1, it is y0−s times the radial symmetrization of
y s about z0 and it is defined globally.

Proof. 1. Let f be such an eigenfunction. By (3.22), f = h ◦ r on V for some smooth h : R>0 →

R. Note that r is smooth on a punctured neighborhood of z0 by (D.37). By (3.25), solving
∆f + λf = 0 is equivalent to solving Dh = 0 for some differential operator D of degree 2, with
nonzero highest degree coefficient at t > 0 sufficiently small. So by (F.18), local solutions extend
globally and we can calculate the dimension.

26
 2. Note that H is isotropic, so the result from 1. applies. By (3.29), y s is an eigenfunction with
eigenvalue s(1 − s). Let us check that its radial symmetrization is still an eigenfunction. Indeed,
by (F.32) ∆ is G-invariant so that by (3.19), taking the symmetrization commutes with ∆.
Finally, if we divide by y0s we obtain a normalized ωs (·, z0 ), with ωs (z0 , z0 ) = 1.
It remains to show that this is the unique eigenfunction that extends continuously to z0 . This
is done by explicitly writing ∆ in polar coordinates, as in done in [Bump, 1996, Proposition
2.3.4].
Applying the correspondence from (3.21) to ωs (z, z0 ), we obtain a point-pair invariant ωs of eigenvalue
s(1 − s) for −∆. We may thus reformulate the conclusion as follows: there is, up to scalar, a unique
point-pair invariant on H with eigenvalue s(1 − s). It has the property that ωs (z, z) = 1 for all z. For
z1 ∈ H arbitrary, ωs (·, z1 ) satisfies the properties from (3.30)(2) with z0 replaced by z1 . This implies
that ωs does not depend on the choice of z0 .
Proposition 3.31 (Selberg eigenfunction principle). Let s ∈ C and f ∈ C ∞ (H) be an eigenfunction
of −∆ with eigenvalue s(1 − s). Then f is an eigenfunction for any compactly supported point-pair
invariant k:
(k ? f )(z) = (k ? ωs (·, z))(z) · f (z)
Proof. Fix z ∈ H, then we have (k ? f )rad
z = k ? fzrad by (3.26)(2). The symmetrization fzrad is still
an eigenfunction with the same eigenvalue by (3.19).5, so fzrad (w) = f (z)ωs (w, z) by uniqueness of
spherical eigenfunctions. That is,

(k ? f )rad rad
z (w) = (k ? fz )(w) = f (z) · (k ? ωs (·, z))(w)

Evaluating in w = z gives the result. Since z was arbitrary, we are done.

Accordingly, we define:
Definition 3.32. Let k be a point-pair invariant on H with compact support. Fix any z0 ∈ H and
define the Selberg-transform

k(s) = (k ? ωs (·, z0 ))(z0 )

b
= (k ◦ ωs (·, ·))(z0 , z0 )
= (k ? y s )(z0 ) · y0−s

where ◦ is the convolution product from (3.14).

The Selberg eigenfunction principle then reads

k(s) · f
k?f =b

Proposition 3.33 (Properties of the Selberg transform). 1. The Selberg-transform k 7→ b

k(s) is an
algebra homomorphism A(H) → O(C) to the ring of entire functions.

2. For all k, b k(1 − s).

k(s) = b
Proof. 1. Immediate. 2. Because ω1−s has the same eigenvalue as ωs .

Proposition 3.34 (The Selberg transform of approximations of the identity). Let ρδ be a point-pair
invariant approximation of the identity (3.8) for small δ > 0. Then:
1. ρbδ (s) → 1 locally uniformly in s.

2. −∆ρ
\δ (s) → s(1 − s) locally uniformly in s.

27
Proof. 1. Fix w0 ∈ H. We have

ρbδ (s) = y0−s (ρδ ? y s )(w0 )

Z
= y0−s ρδ (w0 , w)y s dw
H

The support of the integral is contained in some small geodesic ball B(y0 , δ). Inside the integral,
we approximate y s by y0s :
 Z 
 
 −s s s
|ρbδ (s) − 1| = y0 ρδ (w0 , w)(y − y0 )dw

 B(y0 ,δ) 
Z
6 y0−s  ρδ (w0 , w)|y − y0 ||s| max(y σ−1 , y0σ−1 )dw
 
B(y0 ,δ)

where we used the mean value inequality for y s . Let Iδ ⊆ R be the interval

Iδ = =m B(y0 , δ) = [y0 e−δ , y0 eδ ]

We see that for w ∈ B(y0 , δ) the difference |y − y0 | is small, in fact O(δ) for δ → 0. Thus as long
as s stays in a compact set K, we have

|ρbδ (s) − 1|  δ · max sy σ−1 y0−s 

 
s∈K
y∈Iδ

and the convergence follows.

2. Recall that y s is an eigenfunction of −∆ with eigenvalue s(1 − s) (3.29). Thus

\δ (s)y s = (−∆ρδ ? y s )
−∆ρ
= −∆(ρδ ? y s )
= −ρbδ (s)∆(y s )
= ρbδ (s)s(1 − s)y s

Thus −∆ρ
\δ (s) = ρbδ (s)s(1 − s) → s(1 − s) locally uniformly.

28
4 Functions on the quotient Γ\H
Armed with many tools to deal with smooth functions on symmetric spaces such as H, we begin our
study of Real analytic Eisenstein series. Fix a lattice Γ ⊂ G. Unless otherwise
  stated, we will assume
1 1
that it has only one cusp, at ∞, and that its stabilizer is generated by . We will often use the
0 1
notation σ = <e(s).

4.1 Eisenstein series

Note how, for s ∈ C, y s is invariant under Γ∞ . We define the real analytic Eisenstein series by summing
the images of y s right cosets of Γ∞ :
Definition 4.1 (Real analytic Eisenstein series).
X
E(w, s) = y(γw)s
γ∈Γ∞ \Γ

The series is Γ-invariant at those s for which it converges absolutely. We will drop ‘real analytic’ and
simply refer to it as the ‘Eisenstein series’.
Proposition 4.2. The series E(w, s) converges uniformly and absolutely on compact subsets of H ×
{σ > 1}.
We first give a proof for Γ = PSL2 (Z), based on ideas from [Charollois, 2017, Proposition 1.1]. Then
by Bézout’s theorem we have a bijection
Γ∞ \Γ −→ {(c, d) ∈ Z2 : gcd(c, d) = 1}/{± 1}
(4.3)
 
a b
γ= 7−→ (c, d)
c d
2
Proof. Let K ⊆ H be compact. For each w ∈ K, |cw + d| is a positive definite quadratic form in
(c, d) ∈ R2 , hence positive on the unit ball for the sup norm on R2 . That is, |cw + d| > C · max(|c|, |d|)
with C independent of (c, d) ∈ R2 − {0}. By continuity and compactness of K, we may assume
C = C(K) independent of w.
Now for R ∈ N there are  R pairs (c, d) ∈ Z2 − {0} with max(|c|, |d|) = R. We deduce that
∞
X yσ σ
X R
2|E(w, s)| 6 2σ
 y
|cw + d| R2σ
(m,n)∈Z2 −{0} R=1

and the conclusion follows.

The proof relied crucially on the bijection (4.3). In general, we don’t have such a nice description
of Γ∞ \Γ, and we have to exploit the discretenesss of Γ in a different way. The proof below uses the
arguments from [Iwaniec, 2002, Lemma 2.10], where one can also find explicit bounds.
Proof in the general case. Let K ⊆ H be compact. There exists δ > 0 such that K lies above the
horizontal line =m w = δ. For the trivial coset γ0 ∈ Γ∞ \Γ, we have =m(γ0 w) = =m(w), for all
other cosets we have that =m(γw) 6 1/(δc2∞ ) is bounded, by (2.22). We want them to be small. Let
w = x + iy. We have γ(w) = y|cw + d|−2 , so we want |cw + d|2 to be large, as before. Let M > 1 and
consider the set
S(M ) = {γ ∈ Γ∞ \Γ − {γ0 } : |cw + d|2 6 M }
We want to bound its cardinality. For γ ∈ S(M ), the bound on the imaginary part of cw + d implies
c 6 M 1/2 y −1 6 M 1/2 δ −1 , and the bound on the real part implies |cx + d|2 6 M . Now observe that,
when γ, γ 0 ∈ S(M ) are distinct, then
 
0 −1 ∗ ∗
γ(γ ) = ∈ Γ∞ \Γ − {γ0 }
cd0 − dc0 ∗

29
so that |cd0 − dc0 | > c∞ , that is,
 d d0  c∞ d  M 1/2
   
∀γ 6= γ 0

 − > , x + 6
c c0  cc0  c c
That is, the elements of S(M ) correspond to fractions d/c that lie in a bounded interval around
x ∈ R and such that each two of them are at least a certain distance apart. Estimating naively, using
c∞ 6 c  M 1/2 , gives that #S(M )  1 + M 1/2 (M 1/2 )2  M 3/2 . Here we bounded the size of the
gaps from below by M −1 . But we can do better: take a dyadic partition of the interval [c∞ , M 1/2 δ −1 ],
say [2n c∞ , 2n+1 c∞ ] for 0 6 n 6 21 log2 M +O(1). Then in such an interval, fractions d/c are at distance
−1 n+1
at least c∞ (2 c∞ )2 and so it contains at most

M 1/2 (2n+1 c∞ )2
1+ ·  M 1/2 2n
c∞ 2n c∞
such fractions. Summing over n, we obtain

#S(M )  M 1/2 · M 1/2 = M

We saved ourselves a factor M 1/2 . We conclude by

X 1
|E(w, s)|y −σ 6 1 +
|cw + d|2σ
γ∈Γ∞ \Γ−{γ0 }
∞
X 1
61+ 2σ
#S(n + 1)
n=1
n
∞
X 1

n=1
n2σ−1

Note that with the estimate S(M )  M 3/2 we can only show convergence for σ > 54 .
We give another geometric proof, which consists of estimating each term of the Eisenstein series by
an integral over a small domain, and piecing those domains together. The argument is makes it more
intuitive why converges for σ > 1.
Second proof in the general case. [Cohen and Sarnak, 1980, Corollary 1.7]For δ > 0, let kδ be a point-
pair invariant supported on points at distance 6 δ and with values in [0, 1]. We could take an approxi-
mation of the identity, but all we care about is that it has small support and that its Selberg-transform
kδ (s) is nonzero at real arguments. By the Selberg eigenfunction principle,3
b
Z Z
dx1 dy1 dx1 dy1
(4.4) kδ (σ)y2σ =
b kδ (w, z)y1σ 2 6 y1σ
H y1 d(z,w)<δ y12

where we denote
z = x1 + iy1 , w = x2 + iy2
Let w0 ∈ H. Because Γ acts discontinuously, its orbit is discrete. Let δ0 > 0 be such that the closed
geodesic ball B(w0 , δ0 ) is disjoint from its translates γB(w0 , δ0 ) = B(γw0 , δ0 ) for γ ∈
/ Γw0 . Then
B(γw0 , δ0 ) is disjoint from B(µw0 , δ0 ) for γµ−1 ∈
/ Γw0 . Then for all w ∈ B(w0 , δ0 /2), we have the
same relation with δ0 replaced by δ = δ0 /2:

B(γw, δ) ∩ B(µw, δ) = ∅ , γµ−1 ∈

/ Γw0

Fix the standard fundamental domain F for Γ, contained in the standard fundamental domain F∞ =
[0, 1] × R>0 for Γ∞ . Let S be a set of representatives for Γ∞ \Γ. Suppose first that Γw0 is trivial. Then
3 Since we’re only interested in an inequality, we don’t actually need the Selberg eigenfunction here: we could directly

estimate y2σ  d(z,w)<δ y1σ−2 dx1 dy1 with an implicit constant depending on y2 and σ in a controlled way. That way,
R

one can avoid the use of point-pair invariants.

30
when γ runs through s, the images of the smaller balls B(γw, δ) in the quotient Γ∞ \H are disjoint.
Let  > 0 such that at least one larger ball B(γw0 , δ0 ) lies just above the line =m w = . By the
estimate (2.9), it lies below the line =m w = e2δ0 . And by (2.22), all other larger balls lie below the
line =m w = (c2∞ )−1 . Let M = max(e2δ0 , (c2∞ )−1 ). Because the images of the larger balls in the
quotient Γ∞ \H are disjoint, we have, for w ∈ B(w0 , δ):
X X
(γw)σ 6 sup =m(γw)σ
γ∈S γ∈S w∈B(w0 ,δ)
1 XZ dx1 dy1
6 y1σ
kδ (σ) γ∈S z∈B(γw0 ,δ0 )
b y12
Z
1 dx1 dy1
6 y1σ
kδ (σ) F∞ ∩{=m z6M }
b y12

which converges for σ > 1. The absolute and uniform convergence follows now from Weierstrass’s
M-test.
Suppose now that Γw0 is not trivial, i.e. that w0 is an elliptic fixed point. Then the stabilizer is still
finite, say of size m. Then S can be partitioned in at most m subsets Si such that for all i and γ, µ ∈ Si
we have γµ−1 ∈ / Γw0 . The same argument as before now shows that
X m X
X
sup =m(γw)σ 6 sup =m(γw)σ
γ∈S w∈B(w0 ,δ) i=1 γ∈Si w∈B(w0 ,δ)
Z
1 dx1 dy1
6m· y1σ
kδ (σ)
b F∞ ∩{=m z6M } y12

for some M > 0 depending on w0 and δ0 . We conclude again using the M-test.
Proposition 4.5 (Analytic properties of the Eisenstein series). 1. For fixed w ∈ H, E(w, s) is
holomorphic in s.
2. When we forget the complex structure of the second argument, E(w, s) is jointly smooth (as a
function on an open set of R4 ). In particular, for fixed s with σ > 1 it is smooth in w.
3. For fixed s with σ > 1, E(w, s) is an eigenfunction of the Laplacian −∆ with eigenvalue s(1 − s).
4. For fixed s with σ > 1, E(w, s) is real analytic in w.
5. E(w, s) is jointly real analytic.
Proof. 1. By locally uniform convergence.
2. We have to show that the series of partial derivatives converges locally uniformly, for all higher
order derivatives. We can factor out y s . By the Cauchy–Riemann equations, it suffices to consider
dk ∂ l ∂ m −s
ds ∂y ∂x (E(w, s)y ), and by Hurwitz’s theorem we can suppose k = 0. By induction,

∂l ∂m
|cw + d|−2s = s(l+m) |cw + d|−2(s+l+m) Pl,m (cx + d, cy, c)


∂y ∂x

where s(l+m) denotes the rising factorial and Pl,m is some polynomial of total degree at most
2(l + m). As long as y is bounded away from 0, we can bound

Pl,m (cx + d, cy, c)  (|cx + d| + |cy| + |c|)2(l+m)

 (|cx + d| + |cy|)2(l+m)
 |cw + d|2(l+m)

reducing the convergence to that of the Eisenstein series, where any of the above proofs can be
used.

31
 3. Because we can now differentiate term-wise. The function y s is an eigenfunction with that
eigenvalue, and hence so is each =m(γw)s because −∆ is invariant under isometries.
4. By elliptic regularity (F.25).

5. By (5.44), itself a consequence of elliptic regularity for systems of equations.

We give another proof of the smoothness and the fact that E(w, s) is a Laplacian eigenfunction, which
illustrates the power of convolution operators:
Proof. Let k be a compactly supported point-pair invariant on H. By the Selberg eigenfunction
principle, we have
k ? (=m γw)s = bk(s) =m(γw)s
Because the Eisenstein series E(w, s) is continuous in w and converges locally uniformly, dominated
convergence implies

(4.6) k ? E(·, s) = b
k(s)E(·, s)

k(s) 6= 0, we conclude that E(w, s) is smooth in w.

Now, fix s. If k is smooth, then so is the LHS. If b
To find such k, let k be an approximation of the identity, then by (3.34) b k(s) can be as close to 1 as
k(s) 6= 0.
we want. In particular, there exists k with b
Now, note that (4.6) also implies that E(w, s) is jointly smooth, because the variables w and s are
separated in the LHS.
Now let k = −∆ρδ be the Laplacian applied to an approximation of the identity. We have4 (4.6) on
the one hand, and, by (3.27) for the action of the invariant differential operator −∆,

(4.7) k ? E(·, s) = ρδ ? (−∆E(·, s))

Now let δ → 0; then (4.6) converges to s(1 − s)E(·, s) by (3.34), and (4.7) converges locally uniformly
to −∆E(·, s) by (3.9).
Lemma 4.8. For σ > 1, we have E(w, s) = y s + Os (y 2 ) as y → ∞.

Proof. We look at the last proof of (4.2), and use the same notation. For y0 = =m w0 sufficiently large,
the stabilizer Γw0 is trivial, because there are only finitely many elliptic orbits under Γ. We are thus
in the first case in that proof. By (2.22), for y0 sufficiently large the ball B(w0 , δ0 ) is disjoint from
the balls B(γw0 , δ0 ) with γ ∈ / Γ∞ (it suffices to have 2 log(c∞ y) > 2δ0 ). For the argument to work,
we also need it to be disjoint from those balls with γ ∈ Γ∞ . Therefore it suffices that the real parts of
points in B(w0 , δ0 ) take values in an interval of length less than 1. We know that the imaginary part
in this ball is at most y0 eδ0 , so by (2.10) real parts differ by at most

(exp(2δ0 ) − 1) · 2(y0 eδ0 )2

which is less than 1 for δ0 of size  y0−1 . For large y0 , and γ ∈

/ Γ∞ , the balls B(γw0 , δ0 ) lie below the
line =m w = 1, so we have, for appropriate δ0  y0−1 :
X
|E(w0 , s) − y0s | 6 =m(γw0 )σ
γ∈S−Γ∞
Z
1 X dx1 dy1
6 y1σ
kδ0 (σ) γ∈S−Γ∞ z∈B(γw0 ,δ0 )
b y12
Z
1 dx1 dy1
6 y1σ
kδ (σ) F∞ ∩{=m z61}
b
0
y12
4 Again by dominated convergence, not by the Selberg eigenfunction principle, since we don’t know yet that E(·, s)
is a Laplacian eigenfunction!

32
The second factor is bounded independently of w0 . We have to bound the first. Let kδ be the non-
normalized approximation to the identity
  
1


−

exp 
2  : d(z, w) < δ

−1
(4.9) kδ (z, w) = cδ ρδ =


 1 − d(z,w)
δ

0 : d(z, w) > δ


(compare with (3.8)). By (3.34), we have b kδ ∼ c−1

δ as δ → 0, where
 
Z
1
c−1
δ = exp − 2  dµ(w)
 

B(z,δ) 1 − d(z,w)
δ

for all z ∈ H. Fix z. We want to bound this from below. The integrand is  1 on B(z, δ/2). Note
that a ball B(z, ) contains a Euclidean rectangle I with sides of length 1/ by virtue of the formula
(2.6) for d(z, w) and the approximation
 p  √ √
arcosh(1 + x) = log 1 + x + (x + 1)2 − 1 = log(1 + O( x)) = O( x) (x > 0)

We conclude that c−1 2

δ  vol(Iδ/2 )  δ . Substituting this in our main estimate, we obtain

kδ0 (σ)−1 ∼s cδ0  δ0−2  y02

|E(w0 , s) − y0s | s b

We are ready to apply the study of Fourier expansions from Section 2.4:
Theorem 4.10. The Eisenstein series has the Fourier expansion
X
(4.11) E(w, s) = y s + φ(s)y 1−s + an (s)W0,s−1/2 (4π|n|y)e(nx)
l∈Z−{0}

valid for w ∈ H and σ > 1, for a certain holomorphic function φ and certain functions an .

Proof. First fix s. Then by the preceding lemma we have that (2.29) applies, which gives the shape of
the nonconstant terms. By (2.28), the constant term has the form b(s)y s + φ(s)y 1−s . By the preceding
lemma, b(s) = 1 for σ > 2. It remains to argue that φ, b and an are holomorphic, by uniqueness of
analytic continuation it then follows that b(s) = 1 also for 1 < σ 6 2. The constant term equals
Z 1
C(y, s) = E(x + iy, s)dx
0

which is indeed holomorphic in s for fixed y. But we want to obtain holomorphy of each of the two
terms b(s)y s and φ(s)y 1−s . Note that b and φ do not depend on y, so taking y, y 0 for which the vectors
(y s , y 1−s ) and (y 0s , y 01−s ) are linearly independent, we can solve for b(s), φ(s) and conclude that they
are holomorphic.
The general asymptotics for Laplacian eigenfunctions of polynomial growth apply, and (2.30) gives:
Corollary 4.12. We have, for w ∈ H:

E(w, s) = y s + φ(s)y 1−s + Os (e−2πy ) , (y → ∞)

Remark 4.13. While (4.8) only proves that the Eisenstein series is y s + Os (y 2 ), the study of Fourier
expansions and Witthaker functions gives us for free that it is, in fact, y s + Os (y 1−σ )!

33
4.2 Automorphic kernels
Let Γ be a lattice which we assume SL2 (Z), unless otherwise stated. We want to study functions on
the quotient Y = Γ\H, and thus, naturally, integral operators on this manifold. While a point-pair
invariant on H sends Γ-invariant functions to Γ-invariant functions, it is itself not an integral kernel
on Y × Y . We thus define:
Definition 4.14. Let Γ be a (possibly cocompact) lattice in G. Let k be a continuous point-pair
invariant of compact support on H. Its automorphization is
X
(4.15) K(z, w) = k(z, γw)
γ∈Γ

This is well defined by requiring that k has compact support: because Γ acts properly discontinuously,
the sum is in fact finite on compact subsets of H × H. Thus K is at least as smooth as k, is Γ-invariant
in both variables and symmetric.
More precisely, for fixed z, the number of terms in the definition of K(z, w) is at most the number
of closed fundamental domains that intersect the compact set supp k(z, ·). Or more generally, for a
compact set L, the support of the restriction k|L×H is compact, and the number of terms is bounded
independently of z ∈ L. In particular, when Y is compact the number of terms in the definition of K
is bounded independently of z and w.
From now on we assume Γ is not cocompact. One can wonder whether it is still the case that the
number of terms in the definition of K is bounded. The answer is no, but we do have a good upper
bound. Throughout this subsection we will denote

z = x1 + iy1 , w = x2 + iy2

Proposition 4.16. For y1 → ∞, we have

(4.17) K(z, w) Γ,k y1

uniformly in w.
That is, we gain automorphy of the kernel at the cost of having to work with an unbounded (possibly
not even square-integrable) kernel.
Proof. Because k has compact support, it is bounded, so it suffices to estimate the number of terms
in the definition of K(z, w).
Say k is supported on points at distance at most R. By the lower bound (2.9) for d(z, w), we have
that the support supp k(z, ·) is contained in the horizontal strip R × [y1 e−R , y1 eR ]. In particular, for
y1 sufficiently large, only horizontal translates of the standard fundamental domain F can intersect
the support supp k(z, ·). But how many? We have from (2.6) that d(z, w) 6 R implies (x2 − x1 )2 6
2y1 y2 (cosh(R) − 1), thus there are at most  y1 such fundamental domains.
We will recover this estimate using the Fourier expansion, in (4.31).
Remark 4.18 (Large imaginary parts). Suppose k is supported on point pairs (z, w) at distance at
most R. As we have seen in the proof above, this implies that there exists a constant c > 0 such that
k(z, w) = 0 unless y1 /c < y2 < y1 c. We show that the same holds for K as Pa function on Y × Y :
let z, w ∈ F lie in the standard fundamental domain. In the sum K(z, w) = γ∈Γ k(z, γw), only the
therms with =m(γw) > y1 /c give a nonzero contribution.
Now note that by (2.22), for every  > 0 there exists A > 0 depending on Γ such that if =m v > A,
then either γ ∈ Γ∞ or =m(γv) < . Take  sufficiently small so that F lies above the line =m v = .
Then for y1 /c > A and w ∈ F , we cannot have =m(γw) > y1 /c unless γ ∈ Γ∞ and w already has
imaginary part y2 > y1 /c.
The kernel K is symmetric, so by changing the roles of z and w, we similarly have y1 > y2 /c if
(z, w) ∈ supp K with y2 /c > A. We conclude that:

y1 k y2 , (z, w) ∈ F ∩ supp K

34
In particular, for a fixed K and z, w ∈ F , the expression

“y → ∞”

is unambiguous when it is understood that (z, w) ∈ supp K. It means that both y1 , y2 → ∞, or

equivalently, at least one of them. We will use this abuse of notation throughout this section. Finally,
we note that for y1 > Ac or y2 > Ac we have
X
K(z, w) = k(z, γw)
γ∈Γ
(4.19) X
= k(z, γw)
γ∈Γ∞

Because K is Γ-invariant in both variables, we can consider it as a kernel on Y × Y , and we have:

Proposition 4.20. For f : H → C measurable and Γ-invariant, which descends as f : Γ\H → C, we
have:

(4.21) K ?Y f = k ? f

Proof. By unfolding: let F be any fundamental domain for Γ. Fix z ∈ H, then

Z Z
K(z, w)f (w)dµ(w) = K(z, w)f (w)dµ(w)
Y
ZF X
= k(z, γw)f (w)dµ(w)
F γ∈Γ
XZ
= k(z, γw)f (γw)dµ(γw)
γ∈Γ F
XZ
= k(z, w)f (w)dµ(w)
γ∈Γ F
Z
= k(z, w)f (w)dµ(w)
H

by noting that the sum over γ ∈ Γ is secretly a finite sum.

For non-cocompact lattices Γ, the convolution operator K is not necessarily Hilbert–Schmidt. But
using the estimate (4.17), we find:
Proposition 4.22. The automorphization K defines a bounded self-adjoint convolution operator on
L2 (Y ).
Proof. Let F be the standard fundamental domain, so that y1  y2 uniformly for z, w ∈ F and
(z, w) ∈ supp K. First, for fixed z ∈ Y we have:
Z
dx2 dy2
(K ? f )(z) = K(z, w)f (w)
F y22
|f (w)|
Z
(4.23)  dx2 dy2
{y1 y2 } y2

where we restrict the domain of integration, because letting the last integral run over F may not give
a finite value. We are tempted to say that, by Cauchy-Schwarz

|f (w)|2
Z Z
|(K ? f )(z)|2  dx2 dy2 dx2 dy2
{y1 y2 } {y1 y2 } y22

35
 2
where the second factor is bounded by kf k2 , and the first factor is as small as  y1 . But we are not
2
happy, because integrating y1 kf k2 against the hyperbolic measure dxy1 2dy1 does not give a finite value.
1
2
The problem is that estimating the second integral by kf k2 is too crude. Estimating directly kK ? f k2
instead, we have, switching the order of integration:

|f (w)|
Z Z
2 dx1 dy1
kK ? f k2  y1 2 dx2 dy2
F {y1 y2 } y2 y12
|f (w)|
Z Z
dx1 dy1
= dx2 dy2
F y22 {y1 y2 } y1

The inner integral is O(1), and the boundedness follows.

If we want to make the bound more explicit, we see from the proof and that of the upper bound (4.17)
that when k is supported on point pairs at distance 6 R, then

(4.24) kKk  kkk∞ · g(R)

for some fixed continuous function g : R>0 → R>0 of exponential growth. In particular, when the
support of k does not increase, the upper bound depends only on kkk∞ .

4.2.1 Fourier expansions

Consider a smooth compactly supported point-pair invariant k and its automorphization K(z, w) on
Γ\H. For fixed z, it is in particular Γ∞ -invariant in w. Fourier-expanding the function K(z, w + x)
for fixed (z, w) and evaluating it in x = 0 gives:
Proposition 4.25. We have
X
(4.26) K(z, w) = K
b n (z, w)
n∈Z

where Z 1
K
b n (z, w) = K(z, w + x)e(−nx)dx
0

We see that the Kb n are smooth functions in (z, w). They are Γ-invariant in z and Γ∞ -invariant in w.
We call K0 the constant term.
b
As for the Fourier expansion of smooth functions R → C, we know that the Fourier coefficients K bn
are rapidly decreasing as |n| → ∞: We have the general bound
Z 1 p 
−p
 ∂ K(z, w + x) 
|Kn (w, z)| 6 |2πn|
b   dx ∀p > 0, n 6= 0
0
 ∂xp 

2 2
Note how the Poincaré metric dx y+dy
2 tends to 0 as y → ∞. Thus the function k, which depends only
on the distance, gets “spread out” more and more as its arguments approach i∞, and we expect its
derivatives to go to 0:
Proposition 4.27 (Approximating an automorphic kernel by its constant term). For y → ∞ we have

(4.28) b 0 (z, w) + ON (y −N )
K(z, w) = K

Proof. We start from the Fourier expansion (4.26). From (4.19) we have
X
K(z, w) = k(z, γw)
γ∈Γ∞

36
for y large enough. Because the sum is now restricted to γ ∈ Γ∞ we can unfold and obtain:
Z 1 X
Kn (z, w) =
b k(z, γ(w + x))e−2πinx dx
0 γ∈Γ
∞
Z
= k(z, w + x)e−2πinx dx
R

By (3.22) we have that k is a smooth function of the hyperbolic distance. Say k(z, w) = Φ(u(z, w)),
where
|z − w|2
u(z, w) =
y1 y2
as in (2.7) and Φ is smooth on R. Then

∂(|z − w − x|2 ) 1
Dx k(z, w + x) = Φ0 (u(z, w + x)) ·
∂x y1 y2

where the second factor is a degree 1 polynomial in x. We see a factor (y1 y2 )−1 appear, which we think
of as small. When we differentiate again, we don’t just get one term with a factor (y1 y2 )−2 , there
is also a term with only a factor (y1 y2 )−1 which comes from differentiating the polynomial. Keeping
track of all terms, we obtain by induction that DxN k(z, w + x) is a (finite) sum of functions of the form
1
Φ(b) (u(z, w + x)) · P (z, w, x)
(y1 y2 )b
with b 6 M , P a polynomial of certain degree a in x and at most 2a in x1 , x2 , y1 , y2 .
For every such term we have 2b − a = N : indeed, when differentiating it, we either differentiate Φ(b) ,
which increases b by 1 and increases a by 1, or we differentiate the polynomial, which keeps b constant
and decreases a by 1. In either case, we see that 2b − a increases by 1. In particular from 2b − a = N
we have 2(b − a) = N + a > N . Since we may assume x1 , x2 bounded (w.l.o.g. z and w lie in the
standard fundamental domain) each such term is bounded by

(y12 + y22 )a
N |Φ(b) (u(z, w + x))|  |Φ(b) (u(z, w + x))| · y 2a−2b
(y1 y2 )b
6 |Φ(b) (u(z, w + x))| · y −N

where we denote y for either y1 or y2 , this abuse of notation being justified by (4.18).
For the Fourier coefficients, we now have for all N > 0:
Z
1
|K
b n (z, w)| 6 |DN k(z, w + x)|dx
(2π|n|)N R x
Z
1
N |Ψ(u(z, w + x))|dx
|ny|N R

where Ψ = b6N |Φ(b) | has compact support. It remains to bound the integral. By the lemma below,
P
it is N y. We conclude that
1 X 1 1
K(z, w) − K
b 0 (z, w) N  N −1
y N −1 |n|N y
n6=0

Lemma 4.29. Let Φ : R>0 → R be a continuous compactly supported function, and z, w ∈ H. Then,
with u as in (2.7):
Z Z  
y2 y1
(4.30) Φ(u(z, w + x))dx = (y1 y2 )1/2 Φ x2 + + − 2 dx
R R y1 y2
where the integrand is bounded by an integrable function independently of z and w.

37
Proof. We have
|z − w − x|2 (x1 − x2 − x)2 + (y1 − y2 )2
u(z, w + x) = =
y1 y2 y1 y2
so that substituting t = (x + x2 − x1 )/(y1 y2 )1/2 gives (4.30). Finally, say Φ is supported on [0, A], and
say |Φ| 6 B, then we have the upper bound
Z   Z √A
y2
2 y1
Φ x + + − 2 dx 6 2 Bdx < ∞
R y1 y2 0

b 0 : as y → ∞ we
In particular, we recover the estimate (4.17) by applying this to the constant term K
have, with k = Φ ◦ u:
Z
K0 (z, w) =
b k(z, w + x)dx
R
Z  
y2 y1
(4.31) = (y1 y2 )1/2 Φ x2 + + − 2 dx
R y1 y2
where the integral is bounded by a constant.
From (4.20), automorphized kernels behave well with respect to convolution. We can refine this as
follows:
Proposition 4.32. Let k be a compactly supported point-pair invariant with automorphization K,
and f : H → C be smooth and Γ-invariant, with constant term Cf . Then

(4.33) (K
b 0 ?F f )(z) = (k ? Cf )(z) , y1 > B

where the first convolution is on the standard fundamental domain F , the second is on H, and B is a
constant depending on Γ and k.
Recall that the Fourier coefficients Kb n are not Γ-invariant in the second variable, so it does not make
sense to view Kn as a convolution operator on the quotient Γ\H. Note also how we regard Cf (z) here
b
as a function of z, while it is (by definition) constant on horizontal lines.
Proof. Let B > 0 be large enough so that (4.19) holds for y1 > B. We have, by unfolding,
Z
(K
b n ?F f )(z) = K
b n (z, w)f (w)dµ(w)
F
Z Z 1
= K(z, w + x)f (w)dxdµ(w)
F 0
Z Z 1 X
= k(z, γ(w + x))f (w)dxdµ(w)
F 0 γ∈Γ
∞
Z Z 1 X
= k(z, γw + x)f (w)dxdµ(w)
F 0 γ∈Γ
∞
Z Z 1 X
= k(z, γw + x)f (w)dxdµ(w)
F 0 γ∈Γ
Z Z 1
= k(z, w + x)f (w)dxdµ(w)
H 0
Z Z 1
= k(z, w) f (w + x)dxdµ(w)
H 0

One might wonder why we don’t prove that

(K
b n ?F f )(z) = (k ? fbn )(z)

38
for all n ∈ Z and y1 sufficiently large. The only thing that prevents us is a subtlety in the definition of
Kb n : we Fourier-expanded K(z, w + x) for fixed (z, w) and obtain a Fourier series (4.26) without the
oscillating factor e(nx). While for f , we Fourier expanded f (x + iy) for fixed y, and not f (x + iy + x0 )
for fixed (x, y), so we do get the factor e(nx). In the end, it is of little importance which way we define
the Fourier coefficients, and we will care very little about K b n for n 6= 0.

4.2.2 Truncated kernels

We have seen that automorphic kernels increase sufficiently slowly at the cusp so that they define
a bounded convolution operator. In various situations it is desirable to have a compact operator
acting on L2 (Y ). From (4.28) we see that it is the constant term Kb 0 that prevents K from being
square-integrable. We want to subtract that constant term. Our criteria are:
1. We want a kernel that is automorphic in both variables.
2. It has to be smooth.
Inspired by K − K b 0 , we construct a smooth automorphic kernel that looks like it. Now, K b 0 is only
Γ∞ -invariant in the second variable, so that K0 f need not be Γ-invariant when f is. In general, there
b
are three options to make a non-automorphic f automorphic:
1. The ‘method of images’: Take the sum of its images under the action of Γ, as we did when
defining the Eisenstein series: X
f (γz)
γ∈Γ

This preserves smoothness, but has the disadvantage that convergence can be painful to show.
2. Restrict f to the standard fundamental domain F , and translate it to other fundamental domains:
X
[f ]F (z) := f |F (γ −1 z) , z ∈ γF
γ∈Γ

This is well-defined except on a measure 0 set (the boundary ∂F and its translates). We can
make it smooth and everywhere defined by letting a compactly supported point-pair invariant
act on it; the result will still be Γ-invariant (4.20).
3. A variation of the second method: if we assume that f (z) is invariant under Γ∞ , then so is
α(y)f (z) where we take α ∈ C ∞ (R) of the form
(
0 :y6A
α(y) =
1 :y >A+1

Now α(y)f (z) is supported on points with large imaginary part. Taking A sufficiently large so
that there are no elliptic fixed points with imaginary part > A, then

[α(y)f ]F

is everywhere defined and smooth: there is no need to let a point-pair invariant act on it. Indeed:
it is smooth in a neighborhood of F , because the transformation that maps F to an adjacent
fundamental domain F 0 , fixes (setwise) the intersection F ∩ F 0 , so that [α(y)f ]F is zero in a
neighborhood of F ∩ F 0 .
We thus define, following [Iwaniec, 2002, §4.2] resp. [Brumley, 2015, §5.8] resp. [Cohen and Sarnak,
1980, p. 23]:

39
Definition 4.34. For a compactly supported point-pair invariant k and smooth Γ-invariant f , define
 
XZ 1
L1 f = K ?Y f −  k(·, γ · + t)dt ?Y f
γ∈Γ 0
(4.35)
L2 f = k ? [k ? f − k ? Cf ]F
L3 f = [k ? f − α(y1 ) · k ? Cf ]F

None of these operators has a reason to be self-adjoint. It would have been more natural to define
X
L01 f = K ?Y f − b 0 (·, γ ·) ?Y f
K
γ∈Γ∞ \Γ

but that sum need not converge, essentially because it gives a double sum over γ ∈ Γ∞ \Γ, another
one coming from the definition of K. It would also have been more natural to define

L02 f = k ? [(K − K
b 0 ) ?F f ]F

which is almost the same, by (4.33), and we will make the comparison precise. We stay with the former
definition of L2 simply to respect the definition from the source.
We check that each of these defines a compact convolution operator on L2 (Y ).
1. For L1 we have to check that the sum converges locally uniformly. Indeed, k has compact support,
and the argument is exactly the same as for the convergence and smoothness of (4.15). For y1
large, we have
XZ 1
H1 (z, w) := k(z, γw + t)dt
γ∈Γ 0

X Z 1
= k(z, γw + t)dt
γ∈Γ∞ 0

X Z 1
= k(z, γ(w + t))dt
γ∈Γ∞ 0

XZ 1
= k(z, γ(w + t))dt = K
b 0 (z, w)
γ∈Γ 0

so that L1 is a compact convolution operator by the estimate (4.28).

2. For L2 , we note that by (4.33), for z ∈ F with y1 > C = C(k, Γ) sufficiently large,

(k ? f − k ? Cf )(z) = ((K − K
b 0 ) ?F f )(z)

where the second convolution is on F . Thus f 7→ k ? f − k ? Cf is the convolution operator on

F defined by the kernel

χ{y1 >C} (K(z, w) − K

b 0 (z, w))
 
Z 1 X
b 0 (z, w) −
+ χ{y1 6C} K(z, w) + K k(z, γw + t)dt
0 γ∈Γ

=: K(z, w) − H2 (z, w)

We know from the bound (4.28) that the term supported on y1 > C is bounded. The term
supported on y1 6 C is compactly supported on F × F . This bounded kernel makes K − H2 a
compact convolution operator on L2 (F ). Hence so is L2 = K ◦ (K − H2 ).

40
 3. For L3 , we find, for z ∈ F :
XZ 1
(k ? f − α(y1 ) · k ? Cf )(z) = K ?Y f − α(y1 ) k(z, γw + t)f (w)dt
γ∈Γ 0

=: K ?Y f − H3 ?Y f

That is, we see that H1 and H3 coincide for large values of y1 , hence their kernels differ by a
compactly supported, bounded function on Y × Y . We conclude that L3 is compact.
Note that the kernel H3 need not be smooth in the second variable on Y . We can fix this by
choosing a larger value of A in the definition of α, depending on the support of k.
Each of the integral operators Li has a kernel that is rapidly decreasing. Thus not only do they define
compact operators, they also send functions of polynomial growth to L2 functions.
Since we have only modified our kernels a little bit, we expect the Selberg eigenfunction principle to
hold approximately:
Proposition 4.36. Let f be a Laplacian eigenfunction on Γ\H with eigenvalue s(1 − s). Then

(Lj − b
k(s))f = −Hj ? f , j ∈ {1, 3}
(4.37)
k(s)2 )f = −K ? Hj ? f , j = 2
(Lj − b

where for z ∈ F with y1 sufficiently large,

Hj ? f = K ? Cf , j ∈ {1, 3}
(4.38)
K ? Hj ? f = K ? K ? Cf , j=2

Proof. From the compatibility relation (4.20) we have that K ?f = b

k(s)f , and the first identity follows.
For the second, we note that k ◦ k has Selberg-transform bk(s)2 .

4.3 Maass forms

Let Γ ⊂ G be any lattice, and denote Y = Γ\H for the quotient, as before. We do not assume that Γ
has only one cusp, in order to illustrate some subtle points.
Definition 4.39. When ∞ is a cusp for Γ and f is Γ-invariant, we say f
1. is of polynomial growth at ∞ if there exist N > 0 such that

f (x + iy)  y N (y → ∞)

uniformly in x.
2. vanishes at ∞ if
lim f (x + iy) = 0
y→∞

uniformly in x.
When a is any cusp for Γ, and σa ∈ G is such that σa ∞ = a, then ∞ is a cusp for σa−1 Γσa . When f
is Γ-invariant, we say f
1. is of polynomial growth at a if the σa−1 Γσa -invariant function σa−1 ◦ f ◦ σa is of polynomial growth
at ∞.
2. vanishes at a if σa−1 ◦ f ◦ σa vanishes at ∞.
Definition 4.40. A Maass form (of weight 0) for Γ is a real analytic f ∈ C ∞ (H, C) that is:
1. an eigenfunction for −∆.

41
 2. Γ-invariant
3. of polynomial growth at all cusps.
We denote by H(Γ, λ) the space of Maass forms with eigenvalue λ, and by H c (Γ, λ) the subspace of
Maass cusp forms. 5
Note that, by elliptic regularity, if f is a C 2 eigenfunction for −∆, it is automatically real analytic
(F.25). If Γ is cocompact, there are no cusps and H(Γ, λ) = H c (Γ, λ) is a full eigenspace of ∆.
Example 4.41. The Eisenstein series E(w, s) lives in H(Γ, s(1 − s)) for all σ > 1.

4.3.1 Cusp forms

Recall that Maass forms with eigenvalue s(1 − s) admit a Fourier expansion w.r.t. every cusp, whose
constant term Cf is a linear combination of y s and y 1−s (2.28).
Proposition 4.42. Let λ ∈ C and f ∈ H(Γ, λ), consider the following statements:
(a) Cfa = 0 in the Fourier expansion at every cusp a.

(b) f ∈ H c
(c) f ∈ L2 .
We always have (a) =⇒ (b) =⇒ (c). For λ 6= 0 the three are equivalent, and for λ = 0 we have

H c (Γ, λ) = {f : Cfa = 0, ∀a} ⊆ H(Γ, λ) ∩ L2 (Y )

where the inclusion is strict when there are cusps.

Proof. We try to work at every cusp separately: is f is square-integrable in a neighborhood of a cusp
a iff it vanishes at a? W.l.o.g. suppose a = ∞. Let λ = s(1 − s) with σ > 21 . We know f has a Fourier
expansion of the form
f (x + iy) = Cf (y) + R(x + iy)
where R is exponentially decreasing at ∞, uniformly in x, and Cf = c1 g1 + c2 g2 is a linear combination
of (
g1 (y) = y s and g2 (y) = y 1−s : s 6= 21
g1 (y) = y s and g2 (y) = y s log y : s = 21

In particular, R is L2 in a neighborhood of ∞ and vanishes at ∞.

(a) =⇒ (b),(c): Immediate, from the Fourier expansion f = Cf + R.

(b) =⇒ (c): R vanishes at ∞, so if f vanishes at ∞ then so does Cf , so that c1 = 0 and also c2 = 0
if σ 6 1. In particular, the constant term contains only terms of the form y ζ or y ζ log y with
ζ < 23 . This implies that Cf is square-integrable at ∞ against the hyperbolic measure dxdy
y2 .

(c) =⇒ (a): Suppose f is L2 in a neighborhood of ∞. Unless we know something about s, we cannot

simply conclude from the Fourier-expansion that f vanishes at ∞: if for example 21 < σ 6 1, there
could be a term y 1−s which is L2 but does not vanish at ∞. We want to invoke spectral theory
to conclude that λ ∈ R. We could study the restriction of the Laplacian to smooth functions
on a neighborhood of ∞, but nothing guarantees that it has only nonnegative eigenvalues: the
Laplacian is positive on complete manifolds (G.22).
Hence if we suppose that f is globally L2 , we know that λ is real. Then σ = 21 or s = 1. If s = 12 ,
we deduce that c1 = c2 = 0, because g1 and g2 have different growth and do not vanish at ∞.
5 This
is probably not the standard notation, if there exists one. Idt is inspired by a notation for the space of Harmonic
Maass forms (hence the ‘H’).

42
 If s = 1, we have c1 = 0, and we see that f is a constant function plus the function R which
vanishes at ∞. If ∞ is the only cusp, we can conclude that

H(Γ, λ) ∩ L2 (Y ) = H c (Γ, λ) ⊕ C

If there are multiple cusps, the situation is more mysterious. We observe that constant functions
are in L2 without being cusp forms, but there might be other exceptions. We leave the case
λ = 0 at that.
If s 6= 21 with σ = 12 , there is a subtlety: we have to verify that the oscillations of y s and
y 1−s do not resonate in such a way that a nontrivial linear combination of them can be L2 .
This could only possibly happen if |c1 | = |c2 |: otherwise, one term dominates the other. Let
−c1 /c2 = exp(iθ) and s = 21 + it. Using the bound exp(iT ) − 1  d(T, 2πZ) we have
 −1  
c Cf (y) = y 1/2 ei(θ+t log y) − 1

2
   
π 3π
 y 1/2 χ y : θ + t log y ∈ , + 2πZ
2 2
2
so that kCf kL2 is at least the measure of the set appearing in the characteristic function. But
this set contains infinitely many pairwise disjoint intervals of length 1 (say), hence Cf ∈/ L2 .
Contradiction.
Note how we only needed ‘local’ information near a cusp to conclude that if f vanishes at a cusp, then
it is L2 near the cusp. For the other implication, we needed global information: the proof does not
exclude Maass forms that are L2 at some cusp and are of nontrivial growth at that cusp.
In the proof we saw that, by spectral theory of the Laplacian, there can only be cusp forms if λ ∈ R>0 :

H c (Γ, λ) = 0 , λ∈
/ R>0

This implies that for λ ∈/ R>0 a Maass form is determined by its constant term, and even by the
asymptotics of its constant term:
Proposition 4.43. If Γ has cusps, f ∈ H(Γ, λ) and g ∈ H(Γ, µ) with f ∼ g  1 as y → ∞ at every
cusp, then λ = µ. If λ = µ ∈
/ R>0 , then f = g.

Proof. Say ∞ is a cusp, and let λ = s(1 − s) and µ = ζ(1 − ζ) with σ = <e s, <e ζ > 21 . By assumption,
f is not a cusp form so that by the Fourier expansion there exists c, d ∈ C such that f ∼ c · y s for
σ > 12 , f ∼ cy 1/2 log y or f ∼ cy 1/2 for s = 21 and f ∼ (cy it + dy −it )y 1/2 if s 6= σ = 21 . Thus λ is
determined by the asymptotics at ∞, and the first statement follows.
Moreover, we see that the constant terms of f and g at every cusp are equal up to possibly a term
y 1−s for σ > 12 , y 1/2 for s = 12 and 0 if s 6= σ = 12 . In any case, f and g are equal up to an L2 Maass
form. Hence if λ ∈ / R>0 , it follows that f − g = 0.

In particular we obtain:
Proposition 4.44 (Uniqueness principle). Suppose Γ has one cusp, ∞. Then for σ > 1, the Eisenstein
series E(w, s) is the unique function f such that
1. f ∈ H(Γ, s(1 − s))

2. f ∼ y s as y → ∞.
Equivalently, it is the unique Maass form with
1. (∆ + s(1 − s))f = 0

2. Cf = y s + φ(s)y 1−s for some φ(s) ∈ C.

43
4.3.2 Dimensions
We discuss the dimension of H(Γ, λ) and H c (Γ, λ). As a start, we can refine the result from (4.42) by
bounding the codimensions of the various spaces.

Proposition 4.45. If N is the number of cusps of Γ, then:

1. For λ = s(1 − s) with σ > 1:
dimC H(Γ, λ) 6 N

2. For λ ∈ R>0 (i.e. σ = 12 ):

H(Γ, λ) H(Γ, λ)
dimC 2
= dimC c 6 2N
H(Γ, λ) ∩ L (Y ) H (Γ, λ)

3. For λ = 0 (i.e. s = 1):

H(Γ, λ) H(Γ, λ) H(Γ, λ)

dimC 6 dimC c 6 dimC + N 6 2N
H(Γ, λ) ∩ L2 (Y ) H (Γ, λ) H(Γ, λ) ∩ L2 (Y )

Proof. As in the proof of (4.42), we have from the Fourier expansion that a Maass form is determined
modulo L2 Maass forms by the coefficients of the two functions g1 , g2 appearing in its constant term.
For σ > 1, at each cusp there is only one such coefficient that vanishes precisely on (the nonexistent)
L2 Maass forms: the coefficient of y s . That is, we have an injective linear map

H(Γ, λ)
H(Γ, λ) = ,→ CN
H(Γ, λ) ∩ L2 (Y )

by sending a Maass form to the coefficients of y s in its N Fourier expansions. The first inequality
follows.
For λ ∈ R>0 , the coefficients of both g1 and g2 vanish precisely on cusp forms, and the second statement
follows similarly. For λ = 0, there are the coefficients of y 1 that vanish precisely on L2 forms, which
gives the rightmost inequality. An L2 harmonic Maass form is determined modulo cusp forms by the
coefficients of y 0 in its constant terms, hence

H(Γ, λ) ∩ L2 (Y )
dimC 6N
H c (Γ, λ)

and the middle inequality follows. The leftmost inequality follows from the inclusion H c ⊆ H ∩ L2 .
Theorem 4.46. All the spaces H(Γ, λ) are finite-dimensional.

We give various proofs. Note that by (4.45) it is equivalent to show that H c or H ∩ L2 is finite
dimensional.
Proof 1, compact case. When Γ is cocompact, we know by a general result on Riemannian manifolds,
that the eigenspaces of the Laplacian are finite-dimensional (G.28). That is, all H(Γ, λ) have finite
dimension.

Proof 2, general case. When Γ is noncompact, one can show, by analyzing the resolvent of the hyper-
bolic Laplacian, that the L2 -eigenvalues of −∆ still go to infinity, counting multiplicities (G.29). In
particular,
L2 (Y ) ∩ H(Γ, λ)
has finite dimension. The conclusion follows.
We give a proof that does not use the fact that the Laplacian has compact resolvent.

44
Proof 3, compact case. Assume again that there are no cusps. Fix λ = s(1 − s). Let k be a point-
pair invariant on H and K its automorphization (4.15). By the Selberg eigenfunction principle and
(4.20), each f ∈ H(Γ, λ) is an eigenfunction for K with eigenvalue b k(s). Because Y is compact, K is
automatically a Hilbert–Schmidt integral operator on L2 (Y ), and thus compact. Its eigenspaces for
nonzero eigenvalues are finite-dimensional, either by the general spectral theory of compact operators
k(s) 6= 0
(A.35) or by the spectral theorem for compact self-adjoint operators (A.49). In particular, if b
then
dimC H(Γ, λ) < ∞
k(s) 6= 0. This is possible: we can let k be an approximation
Hence given s it suffices to find k with b
of the identity (3.34).
By modifying the kernel K, we can make the argument work for noncompact quotients as well:
Proof 4, general case. Take k and K as before. When Y is noncompact, K has no reason to be a
compact operator. Suppose first that there is only one cusp, which we may assume is ∞. We carefully
select one of the compact truncated kernels from (4.35). Suppose a truncation L coincides with K
on a linear subspace V of L2 (Y ). Because L2 -eigenspaces of L corresponding to nonzero eigenvalues
µ are finite-dimensional, it follows that the µ-eigenspace of K restricted to V is finite-dimensional.
Take L = L3 , so that L coincides with K on the space of Maass forms f with Cf = 0. From the
proof of (4.42), we see that this includes in particular the H c (Γ, λ). Fix such a λ = s(1 − s) ∈ R
and take k an approximation of the identity, so that µ = b k(s) 6= 0. The µ-eigenspace of K restricted
to V = {f : Cf = 0} has finite-dimension; in particular H c (Γ, λ) has finite dimension. We might as
well have taken L = L2 , which coincides with the compact operator K ◦ K on cusp forms, and whose
k(s)2 .
Selberg transform is b
In the case of multiple cusps, the proof is similar. The only complication is that one has to truncate
the kernel K at all cusps.
We give another proof, which does not use spectral theory, except for relying on the fact that cusp
forms have constant term of their Fourier expansions equal to 0, for which we have used the positivity
of the Laplacian. It relies on the Baire category theorem via a mysterious lemma below.
Proof 5, general case. [Borel, 1997, Theorem 8.5] Take λ > 0. We show that H c (Γ, λ) is closed in
L2 (Y ), so that its finite-dimensionality follows from the general lemma below. Take a sequence (fn )
of cusp forms in this eigenspace, which converges in L2 to f . We show that f is a cusp form and a
Laplacian eigenfunction with the same eigenvalue. Because Y has finite volume, it converges in L1 ,
and thus in distribution: for compactly supported φ ∈ C0∞ (Y ), by dominated convergence:
Z Z
fn φ → fφ
Y Y

This shows that f is, in the distributional sense, an eigenfunction of −∆ with eigenvalue λ:
Z
f · (∆ + λ)φ = 0 , φ ∈ C0∞ (Y )
Y

By elliptic regularity for weak solutions (F.26), it follows that f is smooth. Consequently, f is a
Laplacian eigenfunction in the strong sense.
It remains to show that f is a cusp form. Because λ ∈ R>0 , this is equivalent to f having no constant
term in the Fourier expansion w.r.t. every cusp. Theargument
 is from [Borel, 1997, Proposition 8.2].
1 1
Fix a cusp, w.l.o.g. ∞, with stabilizer generated by . Let Cf denote the constant term in the
0 1
Fourier expansion of f . Suppose Cf (y0 ) 6= 0 for some y0 > 0. take A > 0 such that [0, 1] × [A, ∞] is
contained in the standard fundamental domain, so that the projection H →R Y is injective on it. We
∞
may suppose y0 > A. Take a compactly supported φ ∈ C ∞ (R>0 ) for which A Cf φ 6= 0. That is,
Z 1 Z ∞
f (x + iy)φ(y)dydx 6= 0
0 A

45
At the same time, by assumption:
Z 1Z ∞
fn (x + iy)φ(y)dydx = 0 ∀n
0 A

By the choice of A, this integral expression is L1 -continuous in f : we have

Z 1 Z ∞ 
g(x + iy)φ(y)dydx 6 kgk1 kφk∞ , g ∈ L1 (Y )
 


0 A

But fn → f in L1 , a contradiction.
Lemma 4.47. Let Z be a locally compact Hausdorff space with a positive finite measure µ. Let V
be a closed subspace of L2 (Z) contained in L∞ (Z). Then V is finite-dimensional.
The proof is so elegant that we cannot omit it. Note: we are not just requiring that V is closed in
L2 (Z) ∩ L∞ (Z).
Proof. [Borel, 1997, Lemma 8.3] Because Z has finite volume, kf k2 6 µ(Z) kf k∞ for all measurable
f , so we have a continuous inclusion L∞ (Z) ,→ L2 (Z), which restricts to a continuous bijection
∼
→ (V, k·k2 ). Because V is closed in L2 (Z), it is closed in L∞ (Z), and this is a continuous
j : (V, k·k∞ ) −
bijection between Banach spaces. By the open mapping theorem (A.13), the inverse of j is continuous.6
Let c > 0 such that
kf k∞ 6 c · kf k2 , (f ∈ V )
Suppose v1 , . . . , vn are pairwise orthonormal functions in V . For all a1 , . . . , an ∈ C we have
X  X X 1/2
ai vi (z) 6 c · ai vi = c · |ai |2 , (a.e. z)
 

2

Taking ai = vi (z) we have, for all z ∈ Z:

X
|vi (z)|2 6 c2 , (a.e. z)

Integrating over z gives

n 6 c2 µ(Z)
Hence dim V 6 c2 µ(Z).

6 This is the most mysterious part: it relies on Baire’s theorem, whose proof for general complete metric spaces uses

the axiom of choice.

46
5 Analytic continuation of Eisenstein series
Before embarking on various proofs of analytic (better: meromorphic) continuation of real analytic
Eisenstein series in the sense of Theorem 1.1, we recall that there is a plenitude of notions of holomorphy
and meromorphy: We will use the terminology from Appendix B rather freely. As in the previous
 we assume that Γ has one cusp, at ∞, whose stabilizer is generated by the parabolic element
section,

1 1
.
0 1

5.1 Elementary proofs

There are a few methods which apply (as far as is known) only to specific lattices. We do not present
them in full detail here; the reader is invited to consult the references.

5.1.1 Proof by Poisson summation

Take Γ = PSL2 (Z). Recall the bijection

Γ∞ \Γ −→ {(c, d) ∈ Z2 : gcd(c, d) = 1}/{± 1}

(5.1)
 
a b
γ= 7−→ (c, d)
c d

due to Bézout’s theorem. Summing over all pairs (c, d) ∈ Z2 we obtain, for σ = <e(s) > 1:
−s
|mw + n|2
X 
2ζ(2s)E(w, s) =
y
(m,n)∈Z2

By grouping the pairs (m, n) by the value of (m, n)/ gcd(m, n). This looks a lot like the Riemann zeta
function: X
2ζ(2s) = (n2 )−s
n∈Z−{0}

We know how to meromorphically continue the ζ-function, using Poisson summation for the Jacobi
theta function: Define X 2
θ(t) := e−tπn (t > 0)
n∈Z

Poisson summation gives the functional equation

√
θ(1/t) = t · θ(t)

One then expresses the ζ-function in terms of θ by

Z ∞ 
θ(t) − 1 s dt
π −s γ(s)ζ(2s) = t
0 2 t

The functional equation for θ then provides both the meromorphic continuation of ζ, the locations and
order of its poles as well as the functional equation.
More generally, one can attach a Jacobi theta function to a quadratic form, such as

|mw + n|2
(m, n) 7→
y
Poisson summation gives a functional equation for this theta function, and similarly to the proof for
ζ, we deduce the meromorphic continuation of E(w, s). For details, see e.g. [Garrett, 2011].

47
5.1.2 Proof by Fourier expansion
Recall that Laplacian eigenfunctions of polynomial growth admit a Fourier expansion that is well-
understood, and this applies in particular to the Eisenstein series (4.10). Take again Γ = SL2 (Z).
Thanks to the fact that this group has very nice structure, we can compute all terms of the Fourier
expansion of E(w, s) explicitly. One finds that the constant term equals
ξ(2s − 1) 1−s
CE(w,s) = y s + y
ξ(2s)
where ξ(s) = π −s/2 Γ(s/2)ζ(s). The nonconstant terms are explicitly expressible in terms of Bessel
functions. Assuming the meromorphic continuation of ζ, together with the meromorphic continuation
of Bessel functions one can conclude from here. Moreover, the functional equation for ζ together with
the functional equation for Bessel functions, provides the functional equation for E(w, s). For details,
see e.g. [Charollois, 2017, Théorème 6.1.1], [Brumley, 2015, §4.2].
One may wonder why we are not satisfied with these proofs. The main reason is that they don’t
generalize well to arbitrary lattices Γ, and more generally Eisenstein series for other groups, such as
SLn (R): The problem with the proof by Poisson summation is that general Γ have no nice arithmetic
structure. And even when it has, Poisson summation is not always possible. For example, one can give
a proof using Poisson summation for a specific class of Eisenstein series for SLn (R) with the lattice
Γ = PSLn (Z); minimal-parabolic Eisenstein series. This was observed by Langlands; the argument can
also be found in [Garrett, 2012b].
The issue with the proof by Fourier expansion is that we have a priori no way to meromorphically con-
tinue the holomorphic function φ(s) that appears in the constant term (4.10). Instead, we will deduce
the meromorphic continuation (and functional equation) of φ(s) from the meromorphic continuation
of E(w, s).

5.2 Proof via Fredholm-theory

The following proof of meromorphic continuation, due to Selberg, is largely based on the lecture
notes [Cohen and Sarnak, 1980]. The idea is to use the Selberg eigenfunction principle to write the
Eisenstein series, on the half-plane {σ > 1} where it is defined, as the solution to a Fredholm equation.
We then prove that this Fredholm equation has a unique solution for s in a larger domain containing
the half-plane {σ > 1}, and that the solution depends analytically on s.

5.2.1 A truncated Eisenstein series.

Fredholm theory is about L2 functions. But E(w, s) is not an L2 -function, so we want to ‘truncate
it’, by subtracting its constant term. Its constant term is y s − φ(s)y 1−s . We know very little about
φ(s), so we prefer not to subtract it, and leave it untouched together with the nonconstant terms:
they constitute the part of E(w, s) that is nontrivial to analytically continue. Luckily, that mysterious
part of the constant term is in L2 . The first part, y s , is not. We want to subtract it while preserving
automorphy, so we proceed similarly to how we constructed the third truncated kernel L3 : Let A > 0
and define for <e s > 1 and w ∈ F in the standard fundamental domain:
(5.2) E(w,
e s) = E(w, s) − α(y)y s
where α ∈ C ∞ (R) is such that (
0 :y6A
α(y) =
1 :y >A+1
and A is chosen large enough so that there are no elliptic fixed points with imaginary part > A, that
is, so that
[E(w,
e s)]F
is smooth and Γ-invariant. It is still analytic in s for fixed y (and indeed, still jointly differentiable).
Finding a meromorphic continuation of E is equivalent to finding one of E. e Note also that E e is in L2
for all s, as follows from (4.12).

48
5.2.2 A Fredholm equation
Now let k be a compactly supported point-pair invariant on H and K its automorphization. By the
Selberg eigenfunction principle and (4.20), E is an eigenfunction on Γ\H of the integral operator
defined by K, with eigenvalue b
k(s). We expect this to be approximately true for E.
e We have:

(5.3) (K − b k(s))(α(y)y s )
e = −(K − b
k(s))E

Now K is a kernel supported on point pairs at distance at most (say) R, so that (K ? f )(z) depends
only on the values f (w) for d(z, w) < R. Because d(z, w) > | log(y1 /y2 )| by (2.9), we have

(K ? (α(y)y s ))(z) = (K ? y s ))(z) =: b

k(s)y s (z)

for y1 > (A + 1)eR , that is, for y1 large enough. We conclude that the RHS in (5.3) is compactly
supported, with support bounded independently of s.
We are ready to apply Fredholm theory:

Theorem 5.4. E(w, e s), and thus the Eisenstein series E(w, s), has an analytic continuation to {<e s >
1/2} − [ 21 , 1]. More precisely, it is pointwise holomorphic and jointly smooth. Moreover, E(w, e s) is
square-integrable for all such s.

Proof. Let k be a compactly supported point-pair invariant on H and K its automorphization. We

e by inverting the operator K − b
want to solve (5.3) for E k(s). First, fix s. Recall that K is a bounded
2
self-adjoint operator on the Hilbert space L (Γ\H) (4.22). When b k(s) is not in the spectrum of K, we
can solve the equation and obtain7

E(w,
e k(s))−1 (K − b
s) = (K − b k(s))(α(y)y s ) ∈ L2 (Γ\H)

A regularity theorem for Fredholm equations (C.3) implies that it is smooth in w.

For which s can we do this? Because K is self-adjoint, it has real spectrum. Let k = −∆ρδ , the
Laplacian applied to an approximation of the identity. By (3.34), b
k(s) converges locally uniformly to
s(1 − s) as δ → 0. Write s = σ + it and note that the imaginary part

=m(s(1 − s)) = t(1 − 2σ)

is strictly negative as long as σ > 21 and s is not real. So if we fix s0 ∈ {σ > 12 } − R and take δ
sufficiently small, then K − b
k(s) is invertible for all s in an open neighborhood U of s0 . We are ready
to apply Fredholm theory: the RHS of (5.3) has compact support in w, independently of s. By (C.5),
the solution Ee we obtain, is analytic in s ∈ U and jointly smooth.
Now note that the construction of E(w,e s) for each s depends on the choice of k. We have to argue
that we obtain a unique solution, i.e. that the solutions obtained on neighborhoods U of each s0 glue
together. There are two ways to see this:
1. For k = −∆ρδ , we have bk(s) → s(1 − s) locally uniformly as δ → 0. We can write {σ > 21 } − R
as an increasing union of relatively compact open sets (Un ), the smallest of which intersects the
half-plane {σ > 1}. Letting δ → 0 as n → ∞, we obtain solutions on each of the Un , which glue
together by uniqueness of analytic continuation for fixed w.
2. We can first construct an analytic continuation on neighborhoods of s0 , for <e s0 = 1. Those
neighborhoods overlap with the half-plane {σ > 1}, hence either by uniqueness of the solution
to the Fredholm equation or by uniqueness of analytic continuation, that solution must coincide
with the already defined E(w,
e s) for σ > 1. We want to proceed in this fashion, extending our
solution E(w, s) bit by bit until we cover all of {σ > 21 } − R. But care must be taken when gluing
e
those solutions, because uniqueness of analytic continuation does not apply to such successive
7 Note that we cannot simply say that the operator K − bk(s) cancels with its inverse: the RHS of (5.3) must not be
read as the bounded operator K − bk(s) applied to the function α(y)y s . Indeed, the latter is not in L2 !

49
 extensions. (Unlike in the previous argument, where in each step the larger domain contains all
the smaller ones.) Instead, we seek to apply uniqueness of the solution to the Fredholm equation.
Suppose we have analytically continued E e to an open connected set V containing the half-plane
{σ > 1}, and that there is a point s0 on the boundary of V which is contained in the open set
{σ > 12 }−R. Fredholm theory gives us a point-pair invariant k, an open connected neighborhood
U of s0 and a solution E(w, s) of (5.3) on U . We want to show that it coincides with the already
found Ee on U ∩ V . Therefore, we need that E e satisfies (5.3) for the point-pair invariant k (and
not just some other point-pair invariant).
We know that E(w, s) is jointly smooth, and the solution to our Fredholm equations, as well.
Hence so is Ee + α(y)y s . By uniqueness of analytic continuation and joint smoothness, it is still
a Laplacian eigenfunction with eigenvalue s(1 − s). Thus the Selberg eigenfunction principle
applies for s ∈ U ∩ V . In particular, E
e satisfies (5.3) for the point-pair invariant k, on U ∩ V .
Uniqueness of the solution of the Fredholm equation implies that E coincides with E. e
Finally, we must argue that we can cover all of {σ > 12 } − R in this fashion. We can consider a
maximal connected open set of {σ > 12 } − R on which E(w,
e s) is defined and apply Zorn’s lemma.
Or more elementary, take an arbitrary point s0 in that domain, a compact segment linking it to
the half-plane {σ > 1}, and use compactness of the segment to show that we can reach the point
s0 in a finite number of steps.
Remark 5.5. 1. What is preventing us from analytically continuing beyond the line σ = 12 , is the
spectrum of the operators K: they are not necessarily compact, hence might not have discrete
spectrum. This makes that, for each s, we have to construct a point-pair invariant k whose
Selberg-transform bk(s) avoids the real line (in a neighborhood of s). If the spectrum of K
were discrete (with the exception of 0), it would suffice that bk(s) avoids 0, and we obtain a
meromorphic continuation. Not surprisingly, this will be our strategy.
2. While the operators K are not compact, they are still bounded, and their spectrum is contained
in the closed disk B(0, kKk). One can wonder whether, for s0 fixed, we can find K such that
k(s) avoids just the spectrum of K in a neighborhood of s0 (and not the entire real line). If, for
b
example, the kδ = −∆ρδ from the proof give us K with kKk 6 C bounded independently of δ,
we would have an analytic continuation for all s such that s(1 − s) is not in the disk B(0, C).
Unfortunately8 this is not the case: the best bound we have for kKk is (4.24): While the support
of the kδ is controlled, their sup norm is not.

5.2.3 Truncated kernels

We want to meromorphically continue beyond the line σ = 21 . As remarked above, we want a compact
e to an L2 function whose analytic continuation is known, in order to
integral operator that sends E
apply Fredholm theory. Consider any of the truncated kernels Li defined in (4.35). Let us fix one, say
L3 . We want to view the equation (4.37), with f = E(w, s), as a Fredholm equation:

(5.6) k(s))E(w, s) = −H3 ? (y s + φ(s)y 1−s )

(L3 − b
k(s)(y s + φ(s)y 1−s )
= −α(y)b
There are two issues:
1. The RHS is not in L2 (as before).
2. The RHS involves the function φ, whose meromorphic continuation is not known.
We know how to deal with the first problem: we can modify the Eisenstein series and let E(w,
e s) =
E(w, s) − α(y)y s , just as before. The main problem is the function φ(s), which we know virtually
nothing about. The strategy is as follows:
8 And not surprisingly: otherwise the meromorphic continuation of the Eisenstein series would have only finitely many

poles.

50
 1. Solve (5.6) for all s ∈ C, without the term involving φ in the RHS. Call the solution E ∗ .
2. Conclude, simply by taking linear combinations, that

E ∗ (w, s) + φ(s)E ∗ (1 − s)

solves (5.6) with the full RHS, for σ > 1.

3. Conclude that E ∗ (w, s) + φ(s)E ∗ (1 − s) must coincide with E(w, s) for σ > 1.
4. Exploit the fact that E is a Laplacian eigenfunction to derive a formula for φ in terms of E ∗ ,
which we use to meromorphically continue it.
The last step is the most mysterious one; it relies on spectral properties of the Laplacian. Note that,
while E
e satisfies

(5.7) (L3 − b
k(s))E(w,
e k(s)(y s + φ(s)y 1−s ) + (L3 − b
s) = −α(y)b k(s))(α(y)y s )

we never solve this equation directly. The problem is that it does not have enough symmetry: when we
solve (5.7) without the term involving φ and call the solution Ee ∗ , we cannot conclude that E
e ∗ (w, s) +
∗
φ(s)E (w, 1 − s) solves (5.7) with the full RHS. Instead, we define the more symmetric truncation
e
E(w, s) = E(w, s) − α(y)(y s + φ(s)y 1−s ) and solve

(5.8) (L3 − b k(s)(y s + φ(s)y 1−s ) + (L3 − b

k(s))E(w, s) = −α(y)b k(s))(α(y)(y s + φ(s)y 1−s ))

first without the terms involving φ, call the solution E ∗∗ and then conclude that

E ∗∗ (w, s) + φ(s)E ∗∗ (w, 1 − s)

solves (5.8) with the full RHS.

Lemma 5.9 (Partial solution to the Fredholm equation). For any compactly supported point-pair
invariant for which b
k(s) is not identically zero, the equation

k(s))E ∗∗ (w, s) = −α(y)b

(L3 − b k(s)y s + (L3 − b
k(s))(α(y)y s )

has a L2 -meromorphic jointly smooth (away from poles) solution E ∗ (w, s) for s ∈ C.
Proof. The RHS is simply L3 (α(y)y s ), and equals L3 y s = 0 for large values of y. Thus the RHS
is compactly supported with support bounded independently of s. We can apply Fredholm theory.
Because L3 is a convolution operator with bounded kernel, there are now two options:
1. Fredholm theory for integral operators with bounded kernel (C.5), (C.8).
2. The general Fredholm theorem for bounded operators (C.15).
The conclusion follows.
Corollary 5.10. Let E ∗∗ (w, s) be as in (5.9). Then for σ > 1,

E(w, s) = E ∗∗ (w, s) + φ(s)E ∗∗ (w, 1 − s)

That is, with E ∗ (w, s) := E ∗∗ (w, s) + α(y)y s we have

(5.11) E(w, s) = E ∗ (w, s) + φ(s)E ∗ (w, 1 − s)

Proof. The first identity follows because both sides are L2 -solutions of (5.8), which has a unique
solution for s in a suitable open set. The second follows by definition of E ∗ and E.
Because E ∗∗ (w, s) is jointly smooth away from poles and L2 -meromorphic, it is C ∞ -meromorphic
(B.31), and hence so is E ∗ (w, s). But it is worth remembering that E ∗ (w, s) equals α(y)y s plus an
L2 ∩ C ∞ -valued meromorphic function.

51
5.2.4 Uniqueness principle
We recall that the Laplacian −∆ has a unique self-adjoint unbounded extension to L2 (Y ), whose
domain contains the smooth L2 -functions, and which is still positive (G.22). This gave rise to the
uniqueness principle (4.44): a smooth L2 Laplacian eigenfunction must have a nonnegative real eigen-
value. The idea is to use this to extract φ(s) from (5.11).
Lemma 5.12. There exists a discrete set S ⊂ C with the following property: for each s0 ∈ {σ >
1} − R − S that is not a pole of E ∗ (w, s) or E ∗ (w, 1 − s), φ(s0 ) is the only value of λ(s0 ) ∈ C for which
E ∗ (w, s0 ) + λ(s0 )E ∗ (w, 1 − s0 ) is an eigenfunction of −∆ with eigenvalue s0 (1 − s0 ).
First proof. Let s0 ∈ {σ > 1} − R. That λ(s0 ) = φ(s0 ) works follows from (5.11). If there are two
distinct such λ, then E ∗ (w, 1 − s0 ) is an eigenfunction. Being the sum of the L2 functions α(y)y 1−s0
and E ∗∗ (w, 1 − s0 ), it is in L2 . Because s0 (1 − s0 ) ∈
/ R, this implies E ∗ (w, 1 − s) = 0. Now should
the set of such s0 be not discrete, then it has an accumulation point, and by uniqueness of analytic
continuation we would have E ∗ (w, s) = 0 for all s ∈ C. In particular, E(w, s) = 0. Contradiction.9
Second proof. As before, we know that λ(s0 ) = φ(s0 ) is a solution, and we want to show that it is the
only one. Expressing that E ∗ (w, s0 ) + λ(s0 )E ∗ (w, 1 − s0 ) is an eigenfunction gives:
(5.13) (s0 (1 − s0 ) − ∆)E ∗ (w, s0 ) + λ(s0 )(s0 (1 − s0 ) − ∆)E ∗ (w, 1 − s0 ) = 0
If the coefficient of λ(s0 ) is nonzero for some w, then λ(s0 ) is uniquely determined. If it were zero for
all w, then E ∗ (w, 1 − s0 ) is an eigenfunction with eigenvalue s0 (1 − s0 ), hence it is identically 0. As
before, we conclude that the set of such exceptional s0 is discrete.
We want to use (5.11) to meromorphically continue E(w, s). We already have a meromorphic contin-
uation of E ∗ (w, s), so it remains to continue φ(s). For this, we want to use (5.13).
Proposition 5.14. There is a meromorphic continuation of φ(s) to the entire complex plane.
Proof. Care must be taken, because it is not obvious that when λ(s0 ) solves (5.13) for a fixed w, it is
a solution for all w. We give an argument that avoids this problem.
We claim that there exists w0 such that the meromorphic function (s(1 − s) − ∆)E ∗ (w0 , 1 − s) is not
identically zero. Indeed: if it were identically zero for all w0 , then E ∗ (w, 1 − s) would be a Laplacian
eigenfunction, in particular for σ > 1, where it is L2 . Then E ∗ is identically zero, a contradiction.
Now, we don’t fix s, but we fix w0 for which (s(1 − s) − ∆)E ∗ (w0 , 1 − s) is a nontrivial meromorphic
function. Then (5.13) defines a meromorphic function λ(s). It coincides with φ(s) in the half-plane
σ > 1, because we have seen that there is a dense open subset of that half-plane where there is a
unique solution. We conclude that
(s(1 − s) − ∆)E ∗ (w0 , s)
φ(s) =
(s(1 − s) − ∆)E ∗ (w0 , 1 − s)
is the desired meromorphic continuation. Note that, a posteriori, φ(s) does solve (5.13) for all other
w1 . Indeed, if (s(1 − s) − ∆)E ∗ (w1 , 1 − s) is not identically zero, then solving (5.13) at w = w1 gives a
meromorphic continuation of φ, which must be the same as the one we found by evaluating at w0 . If
it is identically zero, the equation is trivial. We can also see this as follows: if φ(s) is constructed by
evaluating at some w0 , then E(w, s) = E ∗ (w, s) + φ(s)E ∗ (w, 1 − s) defines a meromorphic continuation
of the Eisenstein series. By joint differentiability, it is still a Laplacian eigenfunction, which implies
that (5.13) holds for all w.
Finally, we conclude, combining (5.10) and (5.14):
Theorem 5.15. The Eisenstein series E(w, s) has a jointly smooth C ∞ -meromorphic continuation to
the entire complex plane.
More precisely, we have shown that E(w, s) equals α(y)y s + φ(s)α(y)y 1−s plus an L2 ∩ C ∞ -valued
meromorphic function.
9 The Eisenstein series is not identically 0, because it has a nonzero constant term.

52
5.3 Bernstein’s continuation principle
Roughly speaking, Selberg’s proof of meromorphic continuation consists of constructing an equation,
depending holomorphically on s in a certain sense, satisfied by the Eisenstein series for σ > 1 and
which has a unique solution for all s. Using Fredholm theory, we showed that the unique solution
depends meromorphically on s. The proof depends crucially on the specific form of the Fredholm
equation. A more general framework is given by Bernstein’s continuation principle:
Continuation principle. Consider a topological C-vector space V , an open con-
nected subset S ⊆ C and for each s ∈ S a system of non-homogeneous linear
(5.16) equations for elements v ∈ V , depending holomorphically on s ∈ S. If the system
has a unique solution v(s) for all s in some open subset of S, then it has a unique
solution for almost all s, which depends meromorphically on s.
We always assume our topological vector spaces to be Hausdorff. We will make the statement more
precise in (5.30). We do not claim that the principle holds always, but we will give a sufficient condition,
and apply it to the Eisenstein series.
A continuation principle can also be formulated when S is a higher-dimensional complex manifold.
See for example [Garrett, 2001], on which much of the formalization below is based.

5.3.1 Systems of equations

Definition 5.17 (Systems of equations). Let V be a C-vector space. A (non-homogeneous) linear
equation on V is a triple (T, W, w) where W is a complex vector space, T : V → W a linear map
and w ∈ W . A system of linear equations indexed by a set I is a family Ξ = (Ti , Wi , wi )i∈I of linear
equations, and a solution of the system is a v ∈ V such that Ti v = wi for all i ∈ I. We denote the set
of solutions by Sol(Ξ).
Contrary to [Garrett, 2001], we do not assume V and W to be topological vector spaces and the maps
T : V → W to be continuous. While we will, later, assume that V and W are topological in order to be
able to talk about holomorphy, the continuity of the T : V → W is not necessary for the continuation
principle to hold. Although in our application to Eisenstein series, these linear maps will be differential
operators on a Fréchet space of smooth functions, and thus indeed continuous.
Remark 5.18. 1. (A system as a single equation) Given a system Ξ, we can always replace it by
a single equation without changing the solution set, by taking direct products:
!
Y
Sol((Ti , Wi , wi )i∈I ) = Sol (Ti )i∈I , Wi , (wi )i∈I
i∈I

2. (A system of linear forms) Similarly, we can replace it by a system where all the target spaces have
dimension 1. Indeed: take one equation (T, W, w). If (ej )j∈J is a basis for W and (µj ) are their
dual linear forms (in general not a basis of W ∗ ), then Sol(T, W, w) = Sol((µj ◦ T, C, µj (w))j∈J ).
We can then do this for every equation (T, W, w) of the system. More generally, in this argument
we can replace the (µj ) by any family of linear functionals on W that separates points.
To make sense of what it means for a family of systems of equations to be holomorphic, we have to
clarify what it means for a function with values in Hom(V, W ) to be holomorphic. Let U ⊆ C be open.
One option is to say that A : U → Hom(V, W ) is holomorphic iff the C-valued function s 7→ µ(As v) is
holomorphic for all v ∈ V and µ ∈ W ∗ . But this definition is too restrictive: when for example V = C,
so that Hom(V, W ) ≡ W , W is a topological vector space and A : U → W is holomorphic in the sense
of (B.4), then µ ◦ A is holomorphic when µ is continuous, but usually not fore general µ. It becomes
clear that the correct notion of holomorphy must depend on the topology of W .
Definition 5.19. Let V be a C-vector space and W a topological C-vector space. The weak operator
topology on Hom(V, W ) is the initial topology with respect to the maps φv,µ for v ∈ V and µ ∈ W ∗
(the continuous dual), defined by
φv,µ : Hom(V, W ) → C
(5.20)
T 7→ µ(T (v))

53
If for some reason one does not want to restrict to continuous µ, one can always recover that non-
topological definition by giving W the discrete topology, so that its algebraic dual and continuous dual
coincide.
That it makes Hom(V, W ) into a topological vector space, follows from the general result below applied
to X = Hom(V, W ) and Y = C. Alternatively, by noting that this initial topology coincides with
the topology induced by the seminorms pv,µ = |φv,µ |. In particular, Hom(V, W ) is a locally convex
topological vector space. It is Hausdorff iff the φv,µ separate points of Hom(V, W ). Equivalently, iff
W ∗ separates points.
Proposition 5.21 (Induced topologies on algebraic structures). 1. Let G be a group, H a topo-
logical group and Φ ⊆ Hom(G, H) a set of homomorphisms. Then G is a topological group for
the initial topology with respect to Φ.
2. Let K be a topological field, X a K-vector space, Y a topological K-vector space and Φ ⊆
Hom(X, Y ) a set of linear maps. Then X is a topological K-vector space for the initial topology
with respect to Φ.

Proof. 1. We show that inversion is continuous on G. It suffices to show that if U ⊆ G is of the

form φ−1 (V ) with V ⊆ H open, then U −1 is open. Because the diagram

φ
G H
−1
(·) ( · )−1
φ
G H

commutes, we have that U −1 = φ−1 (V −1 ) is open. We show that the multiplication G × G → G

is continuous. Let U and V be as before. Because φ : G → H is continuous, the map φ × φ :
G × G → H × H is continuous. Because the diagram
φ×φ
G×G H ×H
• •
φ
G H

commutes, we have that •−1 (U ) = (φ × φ)−1 (•−1 (V )) is open.

2. Similarly.

Note that the weak operator topology is usually only defined on the subspace L(V, W ) ⊆ Hom(V, W )
of continuous linear maps, when V is also assumed to be a topological vector space.
Definition 5.22 (Holomorphic operator-valued functions). Let U ⊆ C be open, V and W be C-vector
spaces with W topological. We call A : U → Hom(V, W ) weakly holomorphic if the following equivalent
conditions hold:

1. A is weakly holomorphic (B.7) for the weak operator topology.

2. All C-valued functions φv,µ ◦ A (5.20) are holomorphic.
Proof of equivalence. 1 =⇒ 2: Immediate. 2 =⇒ 1: By (5.23), every continuous linear functional on
Hom(V, W ) is a finite linear combination of the φv,µ .

Proposition 5.23 (The dual of a weak topology). Let V be a C-vector space, whose topology is
induced by a family of linear maps φi : V → C. Then the continuous dual V ∗ consists of finite linear
combinations of the φi .

54
Proof. The argument is inspired by the proof of [Bade, 1954, Lemma 3.3]. Let θ : V → C be continuous.
By continuity of θ, there exists δ > 0 and a finite set J ⊆ I such that, for all v ∈ V , |φj (v)| < δ for all
j ∈ J implies |θ(v)| < 1. By linearity, |φj (v)| < δ for all j implies |θ(v)| < . In particular, each θ(v)
is determined by the values φj (v). Define Φ : V → C|J| by
Φ(v) = (φj (v))j∈J
Then Φ is injective and continuous, and θ factors through Φ. Write θ = f0 ◦ Φ for some f0 defined
on the image of Φ. We can extend f0 to a linear map f : C|J| → C. It follows that θ is a linear
combination of the φj .
Of the plenitude of reasonable topologies10 on Hom(V, W ), the weak operator topology is the weakest
one, so that the corresponding notion of holomorphy is the least restrictive.
Example 5.24. Let V and W be finite-dimensional, with bases (ej ) and (fk ). A family of linear
operators Ts : V → W is holomorphic iff the entries of its matrix representation in these bases are
holomorphic.
Definition 5.25 (Holomorphic families of systems of equations). Let V be a C-vector space and S ⊆ C
open.
1. A holomorphic family of linear equations on V is a family (Ts , W, ws )s∈S where W is a topo-
logical C-vector space, T : S → Hom(V, W ) is weakly holomorphic and w : S → W is weakly
holomorphic.
2. A holomorphic family of systems of equations is a system of holomorphic families of equations:
Ξ(s) = ((Ts(i) , W (i) , ws(i) )s∈S )i∈I

For convenience, we will call it simply a holomorphic system. The solution set becomes a function of
s:
Sol(Ξ(s)) = Sol((Ts(i) , Ws(i) , ws(i) )i∈I )
Remark 5.26 (Holomorphic systems of linear forms). Suppose that the continuous dual W ∗ separates
points, which is for example the case when W is Hausdorff and locally convex, by the Hahn–Banach
separation theorem. Let (µi ) be any generating set of W ∗ . By Remark 5.18, an equation (T, W, w) is
equivalent to the system (µ(i) ◦ T, C, µ(i) (w))i∈I . We also have that a holomorphic family of equations
(Ts , W, ws ) is holomorphic iff (µ(i) ◦ Ts , C, µ(i) (ws ))i is holomorphic: for the inhomogeneous terms ws
this follows from the definition of weak holomorphy. For the linear maps, similarly.
Proposition 5.27 (Composition of holomorphic families). Let U, V, W be C-vector spaces with V
and W topological. Let S ⊆ C be open and A : S → Hom(U, V ) and B : S → L(V, W ) weakly
holomorphic.
1. Then B ◦ A : s 7→ Bs ◦ As is holomorphic.
2. If v : S → V is weakly holomorphic, then Bv : s 7→ Bs vs is weakly holomorphic.
Proof. [Garrett, 2001]
1. This is a corollary of Hartog’s theorem on separate analyticity. Let µ ∈ W ∗ and u ∈ U . By
definition of weak holomorphy and continuity of the Bt , the C-valued function
S×S →C
(s, t) 7→ µ(Bt (As (u))
is separately analytic. By Hartog’s theorem, it is jointly analytic. In particular, the diagonal
function
s 7→ (s, s) 7→ µ(Bs (As (u))
is analytic.
10 strong, weak, ultrastrong, σ-weak, . . .

55
 2. With the identification V ≡ Hom(C, V ), a weakly holomorphic vs corresponds to a weakly
holomorphic operator-valued function. Thus this reduces to the first statement.
Remark 5.28. In the setting of (5.27), note that the map S → Hom(U, V )⊕Hom(V, W ) : s 7→ (As , Bs )
is weakly holomorphic. One could try to give an elementary argument of the first statement by showing
that composition of linear maps Hom(U, V ) ⊕ Hom(V, W ) → Hom(U, W ) is holomorphic (in the sense
of (B.4)) for the weak operator topologies, and it will follow that the composition

S Hom(U, V ) ⊕ Hom(V, W ) Hom(U, W )

s 7−→ (As , Bs ) 7−→ B s ◦ As

is holomorphic. But composition of linear operators is in general not holomorphic for the weak operator
topology; it need not even be continuous.
Finally, we clarify what it means for a V -valued function to be meromorphic.
Definition 5.29. Let V be a topological C-vector space and U ⊆ C open. A function f : U → V is
(weakly) meromorphic if it is locally of the form Pg with g : U → V (weakly) holomorphic (B.4) and
P a C-valued polynomial. Note that in the case of Fréchet spaces (or whenever Laurent expansions
exist), this corresponds to the usual definition in terms of Laurent coefficients (Appendix B.4).
We are ready to formulate a precise continuation principle:
Continuation principle. Let V be a topological C-vector space, S ⊆ C be open
and connected, and Ξ(s)s∈S a holomorphic system of equations on V , in the sense
of (5.25). Suppose there is an open subset U ⊆ S such that for s ∈ U , the system
(5.30)
has a unique solution v(s). Then Ξ(s) has a unique solution for all s ∈ S except
possibly a closed discrete set, and the solution v(s) is weakly meromorphic in s.
In particular, v(s) is meromorphic in U .

5.3.2 Systems of finite type

The prime example of holomorphic systems for which the continuation principle holds, are systems
which “essentially” consist of only finitely many equations.
Definition 5.31 (Systems of finite type). Let V be a topological C-vector space and Ξ(s)s∈S be a
holomorphic system of equations on V .
1. We say Ξ(s) is of finite type if the following equivalent conditions hold:

(a) There is a holomorphic family of vector subspaces of bounded dimension containing Sol(Ξ(s))
for all s. Formally: there exists a finite-dimensional vector space L and a holomorphic fam-
ily of linear maps λ : S → Hom(L, V ) such that Sol(Ξ(s)) ⊆ λs (L) for all s ∈ S. We call
(L, (λs )s∈S ) a finite holomorphic envelope.
(b) There is a holomorphic family of affine subspaces of bounded dimension containing Sol(Ξ(s))
for all s. Formally: there exists a finite-dimensional vector space L, a holomorphic family of
linear maps µ : S → Hom(L, V ) and a holomorphic map v : S → V such that Sol(Ξ(s)) ⊆
µs (L) + vs for all s ∈ S.
2. We say Ξ(s) is locally of finite type if every s ∈ S has an open neighborhood on which Ξ is of
finite type.

Proof of equivalence. If (a) holds, then (b) follows with µ = λ and vs = 0 for all s. If (b) holds, then
(a) follows with L0 = L ⊕ C and λs (l, z) = µs (l) + z · vs .
Example 5.32. 1. If V is finite-dimensional, every system is of finite type.

56
 2. If Ξ(s) has a unique solution v(s) for all s which depends holomorphically on s, then Ξ is of finite
type: We can take L = {0} in condition (b).
Recall that a finite-dimensional C-vector space L has a canonical topology: it is the topology induced
by any norm. It comes with a canonical notion of L-valued holomorphic functions. Recall also that a
holomorphic system can often be reduced to a holomorphic system whose equations have target spaces
that are one-dimensional (5.26).
Theorem 5.33 (Continuation principle for finite type systems). With the notations from the contin-
uation principle (5.30): if Ξ(s) is of finite type and consists of equations whose target spaces Wi are
one-dimensional, then the continuation principle holds.
If moreover V is locally convex and quasi-complete, so that V -valued (weakly) meromorphic functions
have Laurent-expansions (Appendix B.4), then v(s) is holomorphic in U .
The idea is that the unique solvability of a finite system of equations can be expressed in terms of the
nonvanishsing of a determinant. If the determinant is a holomorphic function that does not vanish in
some open set, then it vanishes globally at at most a discrete set of points. There is a subtlety: the
determinant will be the one of the linear system Ts ◦ λs on L, but this system has zero determinant if
λs is not injective:
Proof. Let λs : L → V be a finite holomorphic envelope of Ξ(s).
Injective envelope. Suppose first that λs is injective for all s. We will later remove this hypothesis.
Because Sol(Ξ(s)) ⊂ λs (L) we have the equality Sol(Ξ(s)) = λs (Sol(Ξ(s) ◦ λs )). If Σ(s) = Ξ(s) ◦ λs
has a unique solution u(s) for all s, then Ξ(s) has the unique solution v(s) = λs (u(s)). And if u(s) is
weakly meromorphic, then so is v(s) (5.27).
Reduction to a finite system. Hence, replacing, Ξ(s) by Ξ(s) ◦ λs , which is still a holomorphic
system by (5.27), we may suppose that V = L is finite-dimensional and that λs = idV for all s. We
may also suppose V = CN . Let (ej ) be the standard basis of V , and denote by (xj ) the coordinates
of x ∈ V . By assumption, Ξ consists of equations of the form
N
X
aij (s)xj = wi (s) (i ∈ I)
j=1

where the aij are holomorphic by (5.24).

Now let U ⊆ S be open such that Ξ(s) has a unique solution v(s) for s ∈ U . Take any s0 ∈ U .
Select N equations of Ξ(s) that determine the solution for s = s0 , say the equations corresponding to
i1 , . . . , iN . Then the determinant det(aik j (s0 )) 6= 0. This determinant is a holomorphic function of s,
hence det(aik j (s)) 6= 0 for all s ∈ S except possibly for a closed discrete set P . For such s, by Cramer’s
rule the unique solution of the subsystem is given in terms of the adjugate matrix by
 
wi1 (s)
1
u(s) = adj(aik j )  ... 
 
det(aik j (s))
wiN (s)

In particular, u is a weakly meromorphic V -valued function in the sense of (5.29). Because u(s) is the
unique solution of a subsystem, for s ∈ U − P it must coincide with the solution v(s) of the full system,
which is unique by assumption. In particular, v(s) is weakly holomorphic on U − P . By repeating the
argument with s0 a point of P ∩ U , we conclude that v(s) is weakly holomorphic on U .
Solvability of the full system. We did not show that the meromorphic continuation u(s) is a
solution of the full system for s ∈ S − P . It is true: u(s) being a solution of a linear equation of Ξ(s)
is a holomorphic condition in s. It is satisfied in U , hence by uniqueness of analytic continuation for
C-valued functions, it is satisfied in S − P .
A solution of Ξ(s) for s ∈/ S − P (which exists, as we now know) is automatically unique: already the
solution of a finite subsystem is unique.
Reduction to injective λs . Finally, we get rid of the assumption that the λs are injective. The
dimension of the range of λs is bounded by dim L. Take s0 ∈ S for which the dimension is maximal,

57
and take a subspace L0 ⊆ L of minimal dimension such that λs0 (L0 ) = λs0 (L). Then in particular
λs0 |L0 is injective. Because λs is holomorphic, λs |L0 is injective for s ∈ S − R, for some closed and
discrete R ⊆ S. In particular, λs (L0 ) = λs (L) for such s.
We now apply the case where “λs = idV ” to the system Ξ(s) ◦ λs . We know that this has a unique
solution for s ∈ U − R, say v(s), it is weakly holomorphic in U − R and it has a weakly meromorphic
continuation to S. (It may have poles outside R.) Let w(s) be the unique solution of Ξ(s) for s ∈ U .
By uniqueness, we must have w(s) = λs (v(s)) for s ∈ U − R, and we conclude that λs (v(s)) is a weakly
meromorphic continuation of w(s) in S, but its values may differ from those of w at points of U ∩ R.
This completes the proof of the continuation principle.
Holomorphy in U . We would like to show that:11

• λs (v(s)) has removable singularities in U .

• It coincides with w at points of U ∩ R.
Call y(s) = λs (v(s)), we only look at it for s ∈ U − R. Take s0 ∈ R ∩ U . By assumption, Ξ(s) is of
finite type in S, and because we now know that it has a meromorphic solution, we can show that it
is of finite type with injective envelope in a neighborhood of s0 . Indeed: suppose s0 is a pole of order
N > 0 of y(s) and take a small neighborhood Ω of s0 in which there are no other points of R. Let
L00 = C2 and define µs : L00 → V for s ∈ Ω by

µs (a, b) = y(s)(s − s0 )N a + w(s0 )b

Where we extend y(s)(s − s0 )N holomorphically to s0 ; call its value y0 . Suppose y0 and w(s0 ) are
linearly independent. Then the above µs provides a finite holomorphic envelope of Ξ(s) for s ∈ Ω, and
moreover µs0 is injective. By the injective case, we deduce that y(s) is holomorphic in a neighborhood
of s0 , in particular, at s0 . This contradicts N > 0. Now suppose y0 and w(s0 ) are linearly dependent.
Then
κs (a) = y(s)(s − s0 )N a
defines a one-dimensional holomorphic envelope of Ξ(s) for s around s0 . Suppose V has Laurent-
expansions. Then κs0 is injective: y0 = 0 would imply that s0 is a pole of y(s) of smaller order.
By the injective case, we conclude again that y is weakly holomorphic at s0 . (A contradiction with
N > 0.)
Holomorphy is a local condition, so not surprisingly, we obtain:
Theorem 5.34 (Continuation principle for locally finite type systems). If Ξ(s) is locally of finite type
and consists of equations whose target spaces Wi are one-dimensional, then the continuation principle
(5.30) holds for Ξ(s).

Proof. Write S as an increasing union (Un )n>1 of open relatively compact (in S) open sets. On each
of them, there is a weakly meromorphic continuation vn (s) by the locally finite case. We show that
they glue together to a weakly meromorphic v(s). Let P be the set of points at which Ξ(s) does not
have a unique solution. By the finite type case, P is closed and discrete in each of the Un . Hence it is
closed and discrete in their union, S. By uniqueness of the solution, the vn are successive extensions
on Un ∩ (S − P ), and it follows that when we glue them together to a weakly holomorphic v on S − P ,
this v(s) is weakly meromorphic.

5.3.3 Criteria for finiteness

We need some practical criteria to assure that a holomorphic system is locally of finite type.
Proposition 5.35 (Dominance). (Called inference by Bernstein) [Garrett, 2014a, Proposition 2.0.6]
Let V and V 0 be topological C-vector spaces, U ⊆ C open and Ξ and Ξ0 holomorphic systems on V
11 This is automatic if we know on beforehand that w(s) is holomorphic in U , which will be the case when we apply
the uniqueness principle to the Eisenstein series.

58
and V 0 , respectively. We say that X 0 dominates X if there exists a weakly holomorphic family (for
the weak operator topology) of continuous linear maps hs : V 0 → V such that

(5.36) Sol Ξ(s) ⊆ hs (Sol Ξ0 (s)) (∀s ∈ U )

If Ξ0 is (locally) finite, then Ξ is (locally) finite.

Proof. If (L, λs ) is a finite holomorphic envelope for Ξ0 in some open subset of U , then hs ◦ λs is one
for Ξ. It is holomorphic by (5.27).
Note that if Ξ0 (s) = Ξ(s) ◦ hs , then the reverse inclusion in (5.36) holds.
Proposition 5.37 (Compact operator criterion). [Garrett, 2014a, Proposition 2.0.7] Let V be a C-
Hilbert space, U ⊆ C open and Ξ be the system on V determined by a holomorphic family of bounded
operators Ts : V → V , holomorphic for the norm topology on L(V, V ). Suppose that for some s0 ∈ U ,
Ts0 is of the form λ + K with λ ∈ C − {0} and K compact. Then Ξ is of finite type in a neighborhood
of s0 .
Proof. The kernel of Ts0 is closed because Ts0 is bounded, and its range is closed because K is compact:
this is part of the spectral theory of compact operators (usually only formulated for operators from a
Banach space to itself). Call V0 its kernel and V1 its range, and let pri be the orthogonal projection
on Vi . The system pr1 ◦Ξ dominates Ξ. We how that it is of finite type around s0 . Define

φs = pr0 ⊕(pr1 ◦Ts ) : V → V0 ⊕ V1

and note that Sol(pr1 ◦Ξ(s)) = φ−1 s (V0 ⊕ 0). By spectral theory of compact operators, V0 is finite-
dimensional, so it suffices to show that φs is invertible and that its inverse is holomorphic.
By construction, φs0 is a continuous linear bijection, and it is holomorphic in s. By the open mapping
theorem (A.13), it is an isomorphism. Because φs is holomorphic for the norm topology, it is in
particular continuous, so that φs is invertible for s in a neighborhood of s0 , and the inverse is also
holomorphic (B.6). We conclude that pr1 ◦Ξ, and thus Ξ, is of finite type around s0 .
With minor modifications, a similar criterion can be proven more generally for Fredholm operators
(bounded operators that are invertible modulo compact operators) between Banach spaces, but we
will not need it. See again [Garrett, 2014a, Proposition 2.0.7]. Instead, we will be interested in the
following generalization which is suitable for non-homogeneous systems:
Proposition 5.38 (Compact operator criterion, inhomogeneous version). Let V, U, Ts , λ, K, Ts0 be as
before. Let L be finite-dimensional and let ws : L → V be bounded operator-valued, holomorphic for
the operator norm. Then there exists a finite holomorphic envelope (L0 , µs ) in a neighborhood Ω of s0
with
Ts−1 (ws (L)) ⊆ µs (L0 ) ∀s ∈ Ω
Proof. The argument is the same: we prove the stronger result that Ts−1 pr−1
1 (pr1 (ws (L))) has a finite
holomorphic envelope around s0 , by observing that

Ts−1 pr−1 −1
1 (pr1 (ws (L))) = φs ((0, ws (L)) + (V0 ⊕ 0))

So we can take L0 = L ⊕ V0 .

5.3.4 Eisenstein series

We want to apply the continuation principle to the Eisenstein series, so we’re looking for a holomorphic
system of locally finite type that characterizes the Eisenstein series. By the uniqueness principle (4.44),
we have that for σ > 1 the Eisenstein series f = E(·, s) is characterized by
(
(∆ + s(1 − s))f = 0
Cf = y s + φ(s)y 1−s for some φ(s) ∈ C

59
Equivalently, by the system of linear equations
(
(5.39)  + s(1 − s))f = 0
(∆
∂
y ∂y − (1 − s) (Cf − y s ) = 0

Consider the Fréchet space of smooth functions V = C ∞ (Γ\H) and the vector subspace W of func-
tions that are of polynomial growth. The system (5.39) is linear (inhomogeneous) in f ∈ W , and is
holomorphic (using (B.23)). With the aim of applying the continuation principle, we want this system
to be locally of finite type, and the continuation principle (5.34) will apply. Indeed: The equations of
the system are W -valued, and because V is Fréchet, continuous linear functionals of V (hence of W )
separate points, so that the system can be reduced to an equivalent holomorphic system with equations
whose target spaces are one-dimensional (5.26).
Note that the set of solutions of (5.39) of polynomial growth is finite for all s. Indeed, already
without the second equation this is true: this is precisely saying that spaces of Maass forms are
finite-dimensional (4.46).
In unpublished lecture notes [Bernstein, 1984, Lecture III, §5], one finds the informal remark that, if
one can prove that a holomorphic system has finitely many solutions for all s, usually one can also
prove that it is locally of finite type. A more detailed argument is supposedly given in an appendix
to these lecture notes, but we haven’t found this appendix, and we give an alternative to this rather
imposing remark. It should be noted that verifying local finiteness is something that has often been
neglected in existing proofs of meromorphic continuation that use the continuation principle [Garrett,
2012a].
Proposition 5.40. The holomorphic system on W defined by the equation (∆ + s(1 − s))f = 0 is
locally of finite type for the L2loc -topology on W .
It suffices to prove the following strengthening of finite-dimensionality of the spaces of Maass forms
H(Γ, λ):
Proposition 5.41. For every s0 ∈ C, there exists an open neighborhood U of s0 , a finite-dimensional
vector space L and a strongly holomorphic family of linear maps λs : L → W (s ∈ U ) such that

λs (L) ⊇ H(Γ, s(1 − s)) (∀s ∈ U )

Here W is equipped with the L2loc -topology.

Proof. We take a second look at proof 4 of (4.46). We know that Maass forms in H(Γ, s(1 − s)) are
eigenfunctions of an automorphic kernel K, with eigenvalue b k(s). We want to reduce to the compact
operator criterion (5.38) and work with L2 -functions, so we truncate our Maass forms: we know that
each f ∈ H(Γ, s(1 − s)) has a constant term of the form ay s + by 1−s . Let A > 0 be large enough such
that there are no elliptic points with imaginary part > A, so that there exists α ∈ C ∞ (R) of the form
(
0 :y6A
α(y) =
1 :y >A+1

which defines a smooth function on Γ\H, as in the section about truncated kernels. Then f ∈
C · α(y)y s + C · α(y)y 1−s + L2 (Γ\H). Define trunc f := f − α(y)Cf . The functions α(y)y s and
α(y)y 1−s define a C ∞ -holomorphic envelope (L, λs ) for f − trunc f , with L = C2 . It remains to find
an envelope for trunc f . Applying the approximate Selberg eigenfunction principle to the truncated
kernel L3 (4.37) gives
(L3 − b
k(s)) trunc f = L3 (α(y)Cf )
(Compare with (5.8).) That is, the RHS lies in the finite-dimensional subspace of L2 spanned by
L3 (α(y)y s ) and L3 (α(y)y 1−s ). These functions are (w, s)-continuous and pointwise holomorphic with
compact support bounded uniformly in s, so they are L2 -holomorphic (B.27). They give a finite
envelope for (L3 − b k(s0 ) → 1
k(s)) trunc f . Take k to be an approximation of the identity, so that b

60
(3.34). Using the compact operator criterion, we conclude that trunc f takes values in some L2 -
holomorphic finite envelope (L0 , µs ), locally around s0 .
We want to combine both envelopes. But λs is holomorphic for the Fréchet topology, while µs is
holomorphic for the L2 -topology. So all we can say for now is that the local envelope (L ⊕ L0 , λs ⊕ µs )
for H(Γ, s(1 − s)) is holomorphic for the L2loc topology.
Corollary 5.42. The Eisenstein series has a W -valued weakly L2loc -meromorphic continuation to C.
Proof. This follows now from the continuation principle (5.34). Note that L2loc -continuous functionals
of W still separate points: one can take the convolution of L2loc functions with compactly supported
‘test functions’.
Note that while the L2loc -topology is weaker than the C ∞ -Fréchet topology, the L2loc -topology is still
Fréchet; its topology is induced by a countable family of seminorms. In particular, the weak-to-strong
principle for holomorphy (hence meromorphy) holds for L2loc (B.8):
Corollary 5.43. The Eisenstein series has a W -valued strongly L2loc -meromorphic continuation to C.
The L2loc -topology on W is so weak that we cannot even conclude that E(w, s) is meromorphic for fixed
w: evaluation maps are continuous for the usual Fréchet-topology, but not for the L2loc -topology. So
proving L2loc -meromorphy hardly counts as proving a meromorphic continuation. We try to upgrade
to C ∞ -meromorphy.
We want to use an L2 -to-C ∞ result for vector-valued meromorphic functions (B.31). The only issue
is joint continuity of E(w, s) and its partial derivatives w.r.t. w, away from poles. All we know is that
they are separately continuous and L2loc -continuous. But they are harmonic for fixed w and a Laplacian
eigenfunction for fixed s, so we are in a setting similar to Hartog’s theorem on separate holomorphy:
Proposition 5.44 (Separate eigenfunctions are jointly smooth). Let f : H × U → C be (jointly)
locally integrable, separately smooth, harmonic for fixed w ∈ H and annihilated by ∆ + s(1 − s) for
fixed s ∈ U . Then f is jointly smooth, after changing its values on a set of measure 0.
Proof. This follows from elliptic regularity for overdetermined systems of differential equations (F.26).
If ∆H and ∆U denote the respective Laplace operators, then f is annihilated in the distributional
sense by ∆H,w − s(1 − s) and ∆U,s : For every test function φ ∈ Cc∞ (H × U ),
Z Z Z Z
f (w, s)∆U,s φ(w, s)dsdw = ∆U,s f (w, s)φ(w, s)dsdw
H U
ZH U
= 0
H
=0

and similarly for ∆H . These differential operators have principal symbols y 2 (ξ12 + ξ22 ) resp. ξ32 + ξ42 ,
which do not vanish simultaneously for (ξi ) ∈ R4 − {0}, so they constitute an overdetermined elliptic
system.
One can show that a separately continuous function Ra × Rb → Rc is always Lebesgue-measurable
[Johnson, 1969].
We are almost in a position to prove that E(w, s) has a W -valued strongly C ∞ -meromorphic contin-
uation. By the above discussion, it remains to show that E(w, s) is jointly smooth and jointly locally
integrable. We know that it is a continuous L2loc -valued function, so in particular its local L2 -norms
are locally bounded when s varies. By Fubini, it follows that E(w, s) is jointly L2loc .
Joint smoothness is more subtle: what (5.44) tells us is that E(w, s) is jointly smooth after changing its
values on a set of measure 0. Call F (w, s) the smooth modification of E(w, s). We want the difference
F (w, s) − E(w, s) to be 0. All way know is:
• It is almost everywhere 0.
• It is separately smooth.

61
While it is true that continuous functions which are zero almost everywhere are zero everywhere, it is
a priori unclear whether the same conclusion holds for separately continuous functions. If this is true,
then it will follow that F (w, s) − E(w, s) is everywhere 0, that E(w, s) is smooth and finally, using
(B.31), that its L2loc -meromorphy implies C ∞ -meromorphy. We don’t know how to fill this gap.

5.4 Further analysis

So far we have proved a C ∞ -meromorphic continuation of E(w, s), jointly smooth away from poles.
Much more can be said about the Eisenstein series:

Theorem 5.45. We have the functional equations

E(w, 1 − s) = φ(s)E(w, s)

φ(s)φ(1 − s) = 1
Proof. Using the uniqueness principle (4.44), this is almost immediate: both sides are Maass forms
whose constant terms are of the form y 1−s + φ(1 − s)y s resp. φ(s)y s + φ(s)φ(1 − s)y 1−s . For σ > 1,
the uniqueness principle implies the first functional equation. It extends to general s by uniqueness
of meromorphic continuation. Comparing their constant terms, we obtain the functional equation for
φ.
Note: it is not obvious why (whether) uniqueness of meromorphic continuation holds for vector-
valued meromorphic functions. But because the C ∞ -Fréchet topology on V = C ∞ (H) separates
points (already evaluations separate points) uniqueness of meromorphic continuation on V reduces to
uniqueness of meromorphic continuation for C-valued functions.
In (5.4), we showed that the meromorphic continuation of E(w, s) has no poles in the set {σ > 12 }−[0, 1].
Using the so-called Maass–Selberg relations for L2 -inner products of truncated Eisenstein series, one
can show:

Theorem 5.46. Poles can only occur in the half-open interval (0, 1], and they are simple. The residues
are square-integrable Maass forms.
Proof. See e.g. [Borel, 1997, §12.11].
Theorem 5.47. s = 1 is a pole, with residue equal to the constant function 1.

Proof. See e.g. [Borel, 1997, Proposition 12.13].

62
A Functional analysis
In this entire section, K will denote any of the two fields R and C.

Definition A.1. Let V be a normed space. A Schauder Pbasis for V is a family (si ) for which every
element can be written uniquely as a convergent series i∈I λi si (where only countable many terms
are nonzero).
A permutation
P of a Schauder basis need no longer be Schauder. In a Hilbert space, it does, and the
series i∈I λi si converges absolutely whenever it converges. (Bessel’s inequality)
Definition A.2. A Hilbert space over K is separable if the following equivalent conditions hold:
• It is separable as a topological space, i.e. has a countable dense subset.
• It has a countable Schauder basis.

• It has a countable orthonormal Schauder basis.

It implies that every Schauder basis is countable. ‘Basis’ will mean Schauder basis from now on.
Example A.3. For p > 1 the space of sequences `p (K) is Banach. For X a measure space, Lp (X) is
a Banach space (Riesz-Fischer). For p = 2 they are Hilbert.

Example A.4. For a set X and a Banach space V , the bounded functions X → V with the supremum
norm k·k∞ form a Banach space. If X has a topology and we restrict to continuous maps, the resulting
space is still complete.

A.1 Bounded operators

Definition A.5 (Bounded operator). Let X and Y be normed spaces. A linear transformation
A : X → Y is bounded if the following equivalent conditions hold:
• It is continuous.

• The preimage of the unit ball of Y contains an open ball around 0 ∈ X.

• The image of the unit ball of X is bounded.
• The image of a bounded set is bounded.
• kA(x)k  kxk uniformly for x ∈ X.

in which case we can define its norm kAk as

kAxk
sup = sup kAxk = inf {C > 0 : kAxk 6 C kxk ∀x ∈ X}
x∈X kxk kxk=1

Example A.6. In an inner product space X over R or C, for each x ∈ X the linear map hx, ·i : X → C
is bounded by Cauchy-Schwarz, with norm equal to kxk.
Example A.7 (Integral operators). Let X be a measure space and k ∈ L2 (X × X) be a C-valued
‘kernel’. Then the operator K on the normed space L2 (X) defined by
Z
f 7→ k(s, t)f (t)dt

|k|2 .
R
is bounded with norm 6 kKk2 = X2

We will see that it is compact (A.33) and even Hilbert–Schmidt (A.56).

63
Example A.8. Let X, Y be normed spaces over K. The bounded linear operators L(X, Y ) form again
a normed space for the operator norm. In particular we can consider the dual X ∗ = L(X, K).
We state some useful facts about bounded operators:
Proposition A.9. Let X, Y be Banach spaces over K. The bounded linear operators L(X, Y ) form
a Banach space.
Theorem A.10 (Hahn-Banach). Let X be a Banach space over K and Y 6 X a linear subspace.
For every bounded linear map φ0 ∈ L(Y, K) there is a bounded linear extension φ ∈ L(X, K) with the
same norm.
Corollary A.11. Let X be a Banach space over K and x ∈ X be nonzero, then there exists φ ∈
L(X, K) with φ(x) = kxk and with kφk = 1.
Proof. Apply Hahn-Banach to the linear subspace Y = Kx ⊆ X and the bounded linear map φ0 (λx) =
λ/ kxk which has norm 1.
Proposition A.12. Let A : X → Y be a linear map between separable inner product spaces. Let
(a
Pij ) be the (infinite) matrix of A in orthonormal bases. Then for A to be bounded it suffices that
2
i,j |aij | < ∞.
This condition is strong (it says that A is Hilbert–Schmidt). If the matrix is diagonal (in particular,
if X = Y and A is diagonalizable), it suffices that its entries are bounded (A.42). Thus the change of
orthonormal bases defines a bounded operator (by continuous extension of a densely defined operator,
or Hahn-Banach).
Theorem A.13 (Open mapping theorem). Let A : X → Y be a surjective bounded map between
Banach spaces. Then A is open.
Theorem A.14 (Closed graph theorem). Let A : X → Y be a linear map between Banach spaces.
Then A is bounded iff its graph G ⊂ X × Y is closed.
Proof. If A is bounded, the graph is closed by continuity. (This is true for any continuous map to a
Hausdorff space.) Conversely, if G is closed, it is a Banach subspace of X × Y . The first projection
G → X is a continuous bijection, and by the open mapping theorem it is an isomorphism. Thus
X → G : x 7→ (x, Ax) is continuous, hence so is A.

A.2 The adjoint of an operator, C ∗ -algebras

Theorem A.15 (Riesz representation theorem). On a Hilbert space, every bounded linear form is of
the form hx, ·i for a unique x.
Definition A.16. Let A : X → Y be a bounded operator between Hilbert spaces. Its adjoint is the
unique mapping A∗ : Y → X such that hAx, yi = hx, A∗ yi for all x ∈ X, y ∈ Y .
Proposition A.17 (Properties of the adjoint). We have that A∗ is linear, is bounded with kAk = kA∗ k,
(λA)∗ = λA ∗ ∗∗ ∗ ∗ ∗
2 for λ ∈ C, A = A and (AB) = B A whenever this makes sense. If X = Y then
∗
kAA k = A .

Definition A.18. A Banach-algebra over K is a complete, unital, associative normed K-algebra,
‘normed’ meaning that the norm is submultiplicative:
kxyk 6 kxk kyk ∀x, y ∈ A
∗
A C -algebra is a complex Banach-algebra with an involution (adjoint) satisfying the properties in the
above proposition.
While it can be of interest to study non-unital C ∗ -algebras, we only need to know about the algebra
of operators on a Hilbert-space, so we require our C ∗ -algebras to be unital.
Definition A.19. Let A be a C ∗ algebra and x ∈ A. Then x is normal if xx∗ = x∗ x, and self-adjoint
if x = x∗ .
On a complex Hilbert space, a bounded operator is self-adjoint iff hAx, xi ∈ R for all x.

64
A.3 Banach-Alaoglu and the Gelfand-transform
Theorem A.20 (Banach-Alaoglu). Let X be a normed space over K with dual X ∗ . As a topological
space we can view X ∗ as a subspace of the product KX . Then with the topology induced by the
product topology, the closed unit ball B ∗ of X ∗ is compact and Hausdorff.
For a normed space X, we have a natural map Φ : X → X ∗∗ by sending an x ∈ X to the evaluation
at x. Using Hahn-Banach, one shows that it is a linear isometric embedding.
Theorem A.21. Let X be a normed space. And B ∗ be as in the Banach-Alaoglu theorem. Then
Φ composed with the restriction to B ∗ is a linear isometric embedding into the normed space of
continuous functions C(B ∗ , K) equipped with the supremum norm. In particular, any normed space
can be embedded in a (C(K, K), k·k∞ ) for a compact Hausdorff space K.
Proposition A.22. Let A be a K-algebra (which we will always assume associative and unitary). An
algebra-morphism A → K (we assume it sends 1 to 1) is automatically continuous and has norm 1.
Definition A.23. For a K-algebra A, we denote Ab for the set of algebra-homomorphisms A → K. The
Gelfand-transform of A is the map Φ composed with restriction to A: b it sends a ∈ A to the evaluation
homomorphism A → K at a.
b
If A is a commutative Banach algebra over C, then A b is also called the maximal ideal space, because
algebra homomorphisms A → C correspond bijectively to maximal ideals of A, by taking their kernel.
Proposition A.24. If A is a Banachalgebra, then A b ⊆ B ∗ ⊆ A∗ is compact Hausdorff for the induced
topology by the product topology. The Gelfand transform A → C(A, b K) is a continuous algebra-
morphism with norm at most 1. If A is commutative and K = C, its kernel is the Jacobson radical of
A (the intersection of maximal ideals).
A commutative C ∗ -algebra is always semi-simple (meaning that the Jacobson radical is zero). More
generally, a nonzero normal element of a C ∗ -algebra is not contained in the Jacobson radical. One way
to show this is by introducing the spectral radius of an element of a Banach algebra, the supremum of
the absolute values of elements of its spectrum. One shows that the spectral radius does not depend on
the ambient Banach algebra, and that for normal elements of a C ∗ -algebra it equals the norm. Using
Stone-Weierstrass, one shows:
Theorem A.25. Let A be a commutative C ∗ -algebra. Then the Gelfand transform is a metric ∗ -
isomorphism (∗ meaning that it commutes with the involution) A → C(A,
b C).

A.4 Continuous functional calculus

Theorem A.26 (Functional calculus for normal elements). 1. Let A be a commutative C ∗ -algebra
and x ∈ A with spectrum spec x. Then there exists a unique metric ∗ -embedding C(spec x, C) →
A that sends the inclusion to x; the image of f is denoted f (x).
2. Let A be a C ∗ -algebra and x ∈ A normal. Then the same holds.
Sketch of proof. 1. By (A.25) there is an isomorphism Γ : A → C(A,
b C), which sends x to x
b, the
evaluation at x. We have a continuous ∗ -algebra morphism
C(spec x, C) → C(A,
b C)

by sending f to f ◦ x̂. (We use here that the image of xb is contained in spec x.) It is an isometry
(for the sup-norms) because x b : A → spec x is surjective. We then apply the ∗ -isomorphism Γ
b
∗
to obtain a metric -embedding C(spec x, C) → A. Uniqueness follows from continuity, because
every element of C(spec x, C) is a (uniform) limit of polynomials in the inclusion and its complex
conjugate, by Stone-Weierstrass.
2. By applying the first statement to the commutative C ∗ -subalgebra generated by x. (The spec-
trum of an element does not depend on the ambient Banach algebra.)
This applies in particular to normal operators on a complex Hilbert space. (In which case the bounded
operators form a C ∗ -algebra.)

65
A.5 Positive operators
Definition A.27 (Positive operator). A bounded operator A on a complex Hilbert space H is positive
or non-negative if the following equivalent conditions hold:

• hAx, xi ∈ R>0 for all x ∈ H.

• A is normal and spec A ⊆ R>0 .
• A is of the form B ∗ B.
• A is of the form BB ∗ .

• A is the square of a positive self-adjoint operator.

This defines a partial order on bounded operators. Positive implies self-adjoint. The last four conditions
are equivalent in any C ∗ -algebra. On a finite-dimensional vector space, a positive operator is precisely
a positive semidefinite operator.

Example A.28. Let K be a compact topological space and consider the C ∗ -algebra of continuous
functions C(K, C). The invertible elements are those functions with values in C× . The positive
elements are the functions with values in R>0 .
Functional calculus for normal operators provides a canonical choice for the positive square root (one
direction in the above equivalence). It is unique, because a commutative C ∗ -algebra is isomorphic (via
the Gelfand transform) to the algebra of continuous functions√ on a compact space (its maximal ideal
space). We can denote the unique positive square root by A. We also denote
√
|A| = A∗ A

A.6 Compact operators

Definition A.29 (Compact operator). Let T : X → Y be a linear map between normed spaces. Then
T is compact if the following equivalent conditions hold:
• The image of the unit ball of X is relatively compact.
• The image of every bounded set is relatively compact.
• The image of every bounded sequence contains a convergent subsequence.

And if Y is Banach:
• The image of a bounded set is totally bounded.
A relatively compact subset is bounded, so a compact operator is automatically bounded. A bounded
operator followed by a compact operator is by definition compact. A compact operator followed by
a bounded one is compact as well: we need that the continuous image of a relatively compact set U
under a bounded T is relatively compact. Indeed, we always have T (U ) ⊆ T (U ), the reverse inclusion
holds because U is compact, so its image is, and so it is a closed set containing T (U ).
Definition A.30. A finite-rank operator is one whose image has finite dimension.

Example A.31. A bounded finite-rank operator is compact, because a bounded subset of a finite-
dimensional vector space is relatively compact. The identity map on an inner product space is compact
iff the space is finite-dimensional. This holds more generally for normed spaces (Riesz).
The compact operators on a Hilbert space are complete for the operator norm. See [Bump, 1996,
Lemma 2.3.1].

Proposition A.32. For a bounded operator T between Hilbert spaces, TFAE:

66
 • T is compact.
• T ∗ is compact.
• T is the limit of a sequence of finite-rank operators (for the operator norm).

Proof. See [Conway, 1990, Theorem II.4.4].

Example A.33 (Compactness of integral operators). Let X be a measure space and k ∈ L2 (X × X).
Then Z
K : f 7→ k(·, y)f (y)dy

is compact on L2 (X) with norm at most kkk2 .

Sketch of proof. We know that it is bounded by (A.7). For compactness, we can construct a sequence
of finite rank operators that converges to K. See [Conway, 1990, Proposition II.4.7].
Such an integral operator is even Hilbert–Schmidt (A.56).
Theorem A.34 (Fredholm alternative). Let T be a compact operator on a Hilbert space and λ ∈ C× .
TFAE:

1. T − λ is injective: λ is not an eigenvalue of T .

2. T − λ is surjective.
3. T − λ is invertible.

Proof. This is part of the spectral theory of compact operators; see e.g. [Conway, 1990, pp. VII.7.9,
IVI.7.10]. On finite-dimensional spaces this is the rank-nullity theorem.
Proposition A.35 (Eigenvalues of a compact operator). Let A be a compact operator on a K-Hilbert
space H. The eigenspaces corresponding to nonzero eigenvalues have finite dimension, there are only
countably many nonzero eigenvalues and they tend to 0.

Proof. [Conway, 1990, Proposition II.4.13] We prove all statements at once. The image of any infinite
orthonormal set of eigenvectors is relatively compact and thus contains a Cauchy sequence. If there were
uncountably many eigenvalues λi 6= 0, then there would exist n ∈ N with infinitely many |λi | > 1/n.
If countably many are nonzero but they do not tend to 0, the same holds. In both cases, and also if an
eigenspace for some λi 6= 0 is infinite-dimensional, there is an infinite orthonormal set of eigenvectors
ei corresponding to |λi | > δ > 0. But

(A.36) kλi ei − λj ej k = |λi |2 + |λj |2 > δ 2

meaning that the image of the infinite bounded set of (ei ), contains no Cauchy sequence.
The same holds true more generally in a Banach space, and one can show:

Theorem A.37 (Spectral theory of compact operators). Let X be a K-Banach space and T a compact
operator on X.
1. Every nonzero element of its spectrum λ ∈ σ(T ) is an eigenvalue of T .
2. The subspaces corresponding to nonzero eigenvalues are finite-dimensional.

3. The spectrum of T is countable and can only have 0 as an accumulation point.

Sketch of proof. The essential ingredient is (the elementary) Riesz’ lemma: Given a non-dense proper
subspace Y ⊂ X and  > 0, there is x ∈ X of norm 1 and at distance at least 1 −  from Y . One can
think of this as being a substitute for the inequality (A.36). One then exploits this together with the
compactness of T ; the proof is elementary. See e.g. [Conway, 1990, VII§7].

67
A.7 Diagonalizable operators
Definition A.38. A projection on a K-Hilbert space is a self-adjoint idempotent bounded operator.
Two projections P, Q are orthogonal if P Q = 0, equivalently, if QP = 0.

Proposition A.39 (Properties of projections). We work on a K-Hilbert space H.

1. For a projection P we have ker P = ran(1 − P ) and ker(1 − P ) = ran(P ); they are closed
subspaces.
2. Projections are in 1-1-correspondence with closed linear subspaces, by sending P to its range,
and a closed subspace V to the orthogonal projection on it, which comes from the decomposition
H = V ⊕ V ⊥.
3. Under this bijection, orthogonal projections correspond to orthogonal subspaces.
4. If P is the projection on V , then 1 − P is the projection on its orthogonal complement.

5. Let (Pi ) be a family of pairwise orthogonal projections on (Vi ) and P be the projection on their
span V . Then for all x ∈ H, X
Px = Pi x
i

where the RHS is a series with countable support which converges absolutely.
Definition A.40 (Diagonalizable operator). Let H be a Hilbert space and A a bounded operator on
H. Then A is diagonalizable if the following equivalent conditions hold:
1. There exist closed pairwise orthogonal eigenspaces for A that span H, i.e. H is the orthogonal
direct sum of eigenspaces.
2. H has an orthogonal basis of eigenvectors for A.

3. There exist pairwise orthogonal projections (Pi ) and λi ∈ C such that for all x ∈ H,
X
Ax = λi Pi x
i

where the RHS is a series with countable support which converges absolutely.

4. There exist pairwise orthogonal projections (Pi ) and pairwise distinct λi ∈ C such that for all
x ∈ H, X
Ax = λi Pi x
i

where the RHS is a series with countable support.

A family of operators is simultaneously diagonalizable if the choice in each or any of these definitions
can be made independently of the operator.
P
Remark A.41. 1. This does not imply A = i λi Pi . Indeed, the Pi have norm 1 so this sum can
only converge if λn → 0. This is also sufficient; see (A.43).
2. This definition is stronger than the usual notion of diagonalizability of linear operators on Cn or
Rn .
Proposition A.42 (Properties of diagonalizable operators). Let A be a bounded diagonalizable op-
erator on a Hilbert space, then:
1. It is normal.

2. Its eigenvalues λi are bounded, and kAk = sup |λi |.

68
 3. It is compact iff only countably many eigenvalues (with multiplicites) are nonzero, and they
go to 0 (counting multiplicities). In particular, eigenspaces for nonzero eigenvalues are finite-
dimensional.
4. Positive iff its eigenvalues are real and > 0.
Sketch of proof. Let (ei ) be an orthonormal basis in which A is diagonal.
1. Because its adjoint is diagonalizable in the same basis (with the conjugate eigenvalues).
P
2. Take an element µi ei , then its image has norm at most sup |λi | times the original norm. The
bound is tight by taking eigenvectors.
3. [Conway, 1990, Proposition II.4.6] If A is compact, this follows from the more general (A.35).
Conversely, when they go to zero, we can write the operator as a limit of finite rank operators
and conclude by (A.32). Indeed, order the countably many nonzero eigenvalues (λi ) with mul-
tiplicities, and corresponding eigenvectors (ei ). If Pi denotes the orthogonal projection on hei i,
then
Xn
A − λi Pi = sup |λi |

i>n
i=1
which goes to 0 when n → ∞.
|µi |2 λi .
P P
4. Because if x = µi ei then hAx, xi =
To come back to (A.41), we have:
Proposition A.43. Let H be a Hilbert space, A a diagonalizable operator with countably many
eigenvalues λi so that there exist pairwise orthogonal projections (Pi ) such that for every x ∈ H:
X
Ax = λ i Pi x
i
P
Then λn → 0 iff A = λi Pi .
In particular, this is the case if A is compact, and it implies that A is compact if in addition the
eigenspaces for nonzero eigenvalues are finite-dimensional: then λn → 0 with multiplicities, and A is
compact by (A.42).
Proof. For a finite family of pairwise orthogonal projections (Pi ) and λi ∈ C we have
X
λi Pi = max |λi |


P
Thus using Cauchy’s criterion, λn → 0 implies that λi Pi converges, and by continuity of the
evaluation (A, x) 7→ Ax, it equals A. Conversely, kλi Pi k = |λi |, so for the series to converge we need
λn → 0.
Theorem A.44 (Simultaneous diagonalizability). Let H be a Hilbert space.
1. The restriction of a diagonalizable (bounded) operator to an invariant closed subspace is diago-
nalizable
2. Let (Ai )i∈I be a family of commuting diagonalizable (bounded) operators. Then they are simul-
taneaously diagonalizable.
Proof. 1. Because the restriction of a projection is a projection.
2. Commuting operators stabilize each other’s eigenspaces, and thus intersections thereof. If the
family is finite, we can conclude by induction: diagonalize all Ai except A1 , and then the re-
striction of A1 to the intersections of their eigenspaces. For general families, the same argument
works: we use Zorn’s lemma and show that a maximal subset J ⊆ I for which the (Aj )j∈J are
simultaneously diagonalizable, is necessarily the whole of I.

69
Proposition A.45 (Properties of simultaneously diagonalizable operators). Let H be a C-Hilbert
space. The sum and product of simultaneously diagonalizable operators A, B is diagonalizable in the
same basis.
Proof. By assumption, H has an orthonormal basis of eigenvectors for both A and B, which are visibly
eigenvectors for A + B and AB.

A.8 Spectral theory of compact normal operators

Proposition A.46 (Isolated points of the spectrum). Let H be a Hilbert space and A a normal
bounded operator. Let λ be an isolated point of its spectrum so that the characteristic function 1λ is
continuous on spec(A). Then
1. The element P = 1λ (A) given by the functional calculus of (A.26) is the orthogonal projection
on the eigenspace ker(A − λ).
2. λ is an eigenvalue of A.
Proof. [Vernaeve, 2015, Stelling 2.7.11; Kowalski, 2009, Corollary 3.8] We have that 1λ is self-adjoint
and idempotent, hence so is P : it is an orthogonal projection.
1. We want that Ax = λx iff P x = x, for x ∈ H. By functional calculus it suffices that id −λ
and 1 − 1λ divide each other in C(spec(A), C). Indeed: if f divides g in spec A then ker f (A) ⊆
ker g(A). Note that f = (1 − 1λ )/(id −λ) defines a continuous nonzero function on spec A − {λ},
and we can extend f and its reciprocal continuously at λ by assigning any value to the image of
λ. Hence id −λ and 1 − 1λ divide each other.
2. The eigenspace for λ is the range of P . Because 1λ 6= 0, we have P 6= 0, so the range is
nonzero.
Theorem A.47 (Spectral theorem for normal operators with almost discrete spectrum). Let H be a
complex Hilbert space and A a normal operator on H. Suppose that the spectrum spec(A) is discrete
or has only 0 as an accumulation point, (so that it is in particular countable). Then A is diagonalizable,
and if the Pi are the projections on the eigenspaces with eigenvalues λi , then
X
A= λ i Pi

Proof. [Vernaeve, 2015, Stelling 2.7.11] Because spec(A) is compact, we have λn → 0, at least if there
are infinitely many eigenvalues. We treat the finite case in the same breath, all statements about
convergence being trivial in that case. For nonzero λn 6= 0, we have 1λn (A) = Pn by (A.46) and the
functions 1λn and 1λm are orthogonal
P for n 6= m, so that the projections Pn and Pm are orthogonal.
Because λn → 0, we have 1 = λn 1λn in C(spec A, C) and thus
X
A= λ n Pn

by (A.26).
By (A.37), the above holds in particular for compact normal operators, and thus for compact self-
adjoint operators:
Theorem A.48 (Spectral theorem for compact normal operators). A compact normal operator on
a Hilbert space is diagonalizable, has countably many eigenvalues, and finite-dimensional eigenspaces
for nonzero eigenvalues.
Proof. The only new information is that eigenspaces for nonzero eigenvalues have finite dimension.
This follows from (A.35).
Theorem A.49 (Spectral theorem for compact self-adjoint operators). A compact self-adjoint op-
erator on a Hilbert space is diagonalizable, has countably many eigenvalues, and finite-dimensional
eigenspaces for nonzero eigenvalues.

70
Alternative proof of (A.49). One can prove diagonalizability by successively exhibiting eigenvectors,
and reducing to a compact operator with smaller norm. The key argument is that a compact self-
adjoint operator has an eigenvalue whose absolute value equals the spectral radius. See e.g. [Conway,
1990, Theorem II.5.1] or [Bump, 1996, Theorem 2.3.1].

Proof of (A.48) from (A.49). [Conway, 1990, Theorem II.7.6] Because A is normal,

A + A∗
X = <e A := and
2
A − A∗
Y = =m A :=
2i
which are always self-adjoint, commute. By (A.44), X and Y are simultaneously diagonalizable, hence
A = X + iY is diagonalizable.
As a corollary, the square root of a positive compact operator is again compact. In particular, the
absolute value |A| of a compact operator A is compact.

A.9 Trace class and Hilbert–Schmidt operators

The below can be found in [Conway, 1990, Exercise IX.2.19–20].
Definition A.50 (Trace of a positive operator). Let A be a positive operator on a complex Hilbert
space with orthonormal basis (ei ). Its trace is the sum of nonnegative terms
X
Tr A := hAei , ei i ∈ [0, +∞]
i

It does not depend on the choice of the basis.

Definition A.51. A Hilbert–Schmidt operator A between Hilbert spaces is a bounded operator for
which the positive operator A∗ A has finite trace. That is, the sum
2
X 2
X 2
kAk2 := kAei k = |hAei , fj i| < ∞
i i,j

is finite in some (or any) orthonormal bases (ei ) and (fj ), and kAk2 is called the Hilbert–Schmidt norm.

Proposition A.52 (Properties of Hilbert–Schmidt operators). On a Hilbert space, we have that:

1. Hilbert–Schmidt operators form a two-sided ideal of the algebra of bounded operators, which is
stable by taking adjoints.
2. Finite rank implies Hilbert–Schmidt implies compact. The finite rank operators are dense in the
Hilbert–Schmidt.
3. k·k2 is a norm and and kA∗ k2 = kAk2 > kAk
4. If A, B are Hilbert–Schmidt and AB positive, then AB has finite trace. If in addition BA is
positive, then Tr(AB) = Tr(BA).

|λi |2 < ∞ (counting

P
5. A diagonalizable operator with eigenvalues λi is Hilbert–Schmidt iff
multiplicities).
Definition A.53. A trace class operator A on a Hilbert space is one for which the following equivalent
conditions hold:

1. |A| is the product of Hilbert–Schmidt operators.

2. |A| has finite trace.

71
 p
3. |A| is Hilbert–Schmidt.
4. A is the product of Hilbert–Schmidt operators.
Proposition A.54 (Properties of trace class operators). On a Hilbert space with orthonormal basis
(ei ), we have that:
1. Trace class operators form a two-sided ideal of the algebra of bounded operators, which is stable
by taking adjoints.
2. Finite rank implies trace class implies Hilbert–Schmidt implies compact. The finite rank opera-
tors are dense in the trace class.

3. If A is trace class, then the trace X

Tr A := hAei , ei i
i

converges (in particular, only countably many terms are nonzero) and is independent of the basis.
4. Tr(B ∗ A) defines an inner product on Hilbert–Schmidt operators whose norm is k·k2 . The Hilbert–
Schmidt form a Hilbert space for this inner product.
5. Trace class is a Banach space for the norm k·k1 defined by kAk1 = Tr |A|.
P
6. A diagonalizable operator with eigenvalues λi is trace class iff |λi | < ∞ (counting multiplici-
ties).
Remark A.55 (The trace for non-trace-class operators). If A is an operator P on a Hilbert space H
with orthonormal bases (e i ) and (fj ), then the absolute convergence of hAe i , ei i does not imply that
(−1)j /j. Let
P
of hAfj , fj i: take H separable
√ and A diagonal in the√basis (fj )j>1 with eigenvalues
e2i+1 = (f2i+1 + f2i+2 )/ 2 and e2i+2 = (f2i+1 − f2i+2 )/ 2. Then
∞  
X 1X 1 1
hAei , ei i = −
2 n=0 2n + 2 2n + 1

converges absolutely, but

∞
X X (−1)n
hAfj , fj i =
j n=1
n

does not. Moreover, rearranging the fj can make the sum converge to anything, by Riemann’s rear-
rangement theorem. In short, the trace does not make sense for non-trace-class operators.

Example A.56 (Hilbert–Schmidt integral operators). Let X be a measure space and k ∈ L2 (X × X).
Then Z
K : f 7→ k(·, y)f (y)dy

is Hilbert–Schmidt on L2 (X) with Hilbert–Schmidt norm kKk2 equal to the L2 norm kkk2 .
Proof. [Conway, 1990, Lemma II.4.8; Bump, 1996, Theorem 2.3.2]. For kKk2 = kkk2 we use the
fact that if (ei ) is an orthonormal basis of L2 (X), then the ei (x)ej (y) are an orthonormal basis of
L2 (X × X).
Conversely, one can show that if L2 is separable, every Hilbert–Schmidt operator is of the above form.
It is self-adjoint iff k(x, y) = k(y, x) a.e.

72
B Functional calculus
We are often confronted with complex-valued functions on M × U , where M is a smooth manifold and
U ⊆ C is an open set, and we are interested in their holomorphic or meromorphic dependence on the
second variable. There are many different ways to interpret “holomorphic dependence”. For example:
Definition B.1. We call f : M ×U → C pointwise holomorphic (meromorphic) 12 if f (w, ·) is holomor-
phic (meromorphic) for all w ∈ M . We call f uniformly meromorphic 12 if there exists a closed discrete
set P ⊂ U and a map m : P → N>0 such that each f (w, ·) is meromorphic with poles contained in P ,
and with the order of p ∈ P at most m(p).
Using Weierstrass products, one can construct a holomorphic function on U (not identically zero on
each connected component), which has a zero of multiplicity m(p) at each p ∈ P . Thus uniform
meromorphy of f (w, s) is equivalent to the existence of a holomorphic g (not a zero divisor) such that
f (w, s)g(s) is pointwise holomorphic.
When f (w, s) is pointwise holomorphic and smooth for fixed s, it is reasonable to require that when
D is a differential operator on M , then Df (w, s) is still pointwise holomorphic. Because holomorphic
functions are nothing else than differentiable functions of two real variables annihilated by the Cauchy–
Riemann operator  
∂ 1 ∂ ∂
= +i
∂s 2 ∂σ ∂t
we see that Df will still be pointwise holomorphic when f is jointly smooth, seen as a function on
M × R2 :
Definition B.2. Given topological spaces X, Y, Z and a map f : X × Y → Z, we call it separately
continuous if for all x0 ∈ X and y0 ∈ Y the maps

f (x0 , ·) : Y → Z
f (·, y0 ) : X → Z

are continuous, and (jointly) continuous if f : X × Y → Z is continuous. When X, Y, Z are smooth

manifolds, we define separate smoothness and (joint) smoothness similarly.
Similarly, joint smoothness implies that the complex derivatives f (n) (w, s) are still smooth for fixed s.
It turns out that we don’t need joint smoothness for that:

Proposition B.3 (Joint regularity of complex derivatives). Let M be a smooth manifold and U ⊆ C
open. Let f : M × U → C be pointwise holomorphic. Then:
1. If f is jointly (resp. separately) continuous, then so is f 0 .
2. If f is of class C 1 then so is f 0 , in which case for every chart (xi ) of M we have that ∂f /∂xi is
pointwise holomorphic and
∂ ∂ ∂ ∂
i
f= f
∂s ∂x ∂xi ∂s
∂ |α|
3. If for every chart (xi ) the partial derivatives ∂xα f (w, s) up to order n exist and are continuous,
then f is jointly of class C n .
Proof. 1. Suppose f is (jointly) continuous. Fix (w0 , s0 ) ∈ M × U . We have, for s in a small
compact neighborhood V of s0 :
Z
0 1 f (w, ζ)
f (w, s) = dζ
2πi ∂B (ζ − s)2
12 The terminology is new, we are introducing it purely for convenience.

73
 where B is some small ball centered at s0 . If we restrict w to a compact neighborhood W of
w0 , then f (w, ζ)/(ζ − s)2 becomes uniformly continuous for (w, s, ζ) ∈ W × V × ∂B, so that the
above integral defines a continuous function of (w, s).
Similarly for separate continuity, or by noting that it follows from the case of joint continuity by
taking M = {w} of dimension zero.
2. Because f is C 1 , the integral of ∂f /∂xi (w, ·) along closed contractible contours is still 0, hence
it is holomorphic for all w. The equality follows from Cauchy’s integral formula, from which we
also see that f 0 is still C 1 .
3. We are in particular assuming that f = ∂ 0 f /∂x0 is continuous. We proceed by induction on n.
For n = 1, the partial derivatives of order 1 are indeed continuous by 1. and 2. If n > 1, we note
that
(i) each ∂f /∂xi is pointwise holomorphic and has continuous partial derivatives w.r.t. w up to
order n − 1
(ii) f 0 is pointwise holomorphic, and using Cauchy’s integral formula one shows that its partial
derivatives up to order n exist and are continuous.
and conclude by induction.
Compare this with Hartog’s lemma, which says that a separately holomorphic function is jointly
holomorphic (without any assumption of joint continuity). Such ‘separate versus joint’ properties have
been extensively studied. We will not get any deeper into that. We only remark that a separately
continuous function Ra × Rb → Rc is Lebesgue-measurable [Johnson, 1969].
When studying the spectrum of a bounded operator on a complex Banach space, or more generally,
of an element of a C ∗ -algebra, we are also confronted with functions that take values in a Banach
algebra. More generally, one can wonder about holomorphic functions with values in a topological
complex vector space. Turning back to the first question, given a smooth function f : M × U → C,
we could then take ‘holomorphic dependence’ to mean that f : U → C ∞ (M ) is holomorphic. These
notions and their properties are the subject of this section.

B.1 Differentiability and holomorphy

Let K be any of the fields R or C. All our topological vector spaces will be assumed Hausdorff.

Definition B.4. Let X, Y be topological K-vector spaces, U ⊆ X open and f : U → Y a function.

1. Let X, Y be Banach spaces. We call f (Fréchet) differentiable at x ∈ U if it is approximately
linear at x:
f (x + h) = f (x) + Ah + o(h) (h → 0)
for some bounded linear A : X → Y , the (Fréchet) differential.
2. We will be almost exclusively interested in the case where X = K, in which case the definition
generalizes to topological K-vector spaces Y : we call f differentiable at x ∈ U if the limit

f (x + h) − f (x)
lim
h→0 h
exists, in which case we call it the derivative. If K = C and f is complex differentiable in an
open set U ⊆ X, we also call it (strongly) holomorphic in U .
3. Let X and Y be general topological K-vector spaces. Suppose x = 0 ∈ U and f (0) = 0. We call
f tangent at 0 if for every neighborhood Ω of 0 ∈ Y there exists a neighborhood V of 0 ∈ X such
that V ⊆ U , and a neighborhood I of 0 ∈ K such that

(B.5) f (tV ) ∈ o(t)Ω (t ∈ I)

74
 for some o(t)-function I → K, depending on Ω and V . For general x and f (x), we call f
differentiable at x if there exists a continuous linear map A : X → Y such that h 7→ f (x + h) −
f (x) − Ah is tangent at 0.
Proof of equivalence. It is not immediately clear why the third definition generalizes the others.

3 =⇒ 2: Take a neighborhood Ω of 0 ∈ Y , so that f (x) + Ω is a neighborhood of f (x). Let A,

V and I be as in the hypothesis of definition 3. Then f (x + th) − f (x) − Ath ∈ o(t)Ω for h ∈ V ,
t ∈ I. Fix one such h 6= 0, and take t = k/h, with k ∈ K sufficiently small so that k/h ∈ I. Then

f (x + k) − f (x) o(k/h)
−A∈ Ω (k → 0)
k k
Now note that if Y is not locally convex, there is no reason to assume that, for example, B(0, 1) ·
Ω ⊆ Ω.
Take any neighborhood Ω0 of 0 ∈ Y . By continuity of scalar multiplication, there exists a
neighborhood Ω of 0 ∈ Y and a neighborhood J of 0 in K such that JΩ ⊆ Ω0 . Then o(k/h) k Ω ⊆ Ω0
f (x+k)−f (x)
for k sufficiently small, with h and o(k/h) as above. We conclude that k − A ∈ Ω0 for
0
small k. Because Ω was arbitrary,

f (x + k) − f (x)
lim =A
k→0 k

2 =⇒ 3: Take a neighborhood Ω of 0 ∈ Y . By assumption, there exists A ∈ Y (independent of

Ω) with f (x + h) − f (x) − Ah ∈ hΩ for h sufficiently small, say h ∈ B(0, δ) =: I. Let V = B(0, 1).
For h even smaller, we have f (x + h) − f (x) − Ah ∈ hΩ/2. (But we are not using that Ω/2 ⊆ Ω;
this need not be the case). And so on: we find a decreasing sequence δn ∈ R>0 with δ1 = δ such
that f (x + h) − f (x) − Ah ∈ hΩ/n for h ∈ B(0, δn ) = δn V . Hence we can define o(t) : I → K to
be t/n on the annulus B(0, δn ) − B(0, δn+1 ), and 0 elsewhere, should δn 6→ 0.
3 ⇔ 1: This is immediate, using the fact that balls form a basis of the topology.
We will rarely work with the general notion for topological vector spaces. When X and Y are K-Banach
spaces, the set of bounded linear maps L(X, Y ) is again a Banach space, so if f is K-differentiable
in U it makes sense to ask about the second order derivative, and so on. The usual facts carry
trough, with the same proofs: differentiable at x implies continuous at x, linear maps are everywhere
differentiable and equal to their differential at every point, and we have a chain rule. If X = K and Y
is a topological K-vector space, we have a canonical identification L(X, Y ) ∼= Y and we can talk about
(higher) derivatives as elements of Y . We also have a product rule and a quotient rule.

Proposition B.6. Let X be a K-Banach space, Y a K-Banach algebra, U ⊆ X open and let f, g :
U → Y be K-differentiable at x0 ∈ U .
1. Then f g is K-differentiable at x0 and we have the product rule

(f g)0 (x0 ) = f 0 (x0 )g(x0 ) + f (x0 )g 0 (x0 )

2. Suppose f (x0 ) is invertible. Then f (x) is invertible for x in a neighborhood of x0 , and the inverse
1/f is K-differentiable at x0 with derivative

(1/f )0 (x0 ) = −f (x0 )−1 f 0 (x0 )f (x0 )−1

Proof. 1. The proof is the same as for functions C → C.

2. By the chain rule, it suffices to prove that y 7→ y −1 is K-differentiable at invertible elements
of Y , with derivative h 7→ −y0−1 hy0−1 . Note that indeed, if y0 is invertible, then elements in a

75
 −1
neighborhood of y0 are invertible (when they are at distance less than y0−1 ) so that y −1 is

defined on a neighborhood of y0 . For such y we have

−1
y −1 − y0−1 = y0 (1 − y0−1 (y0 − y)) − y0−1
= (1 − y0−1 (y0 − y))−1 y0−1 − y0−1
= 1 + y0−1 (y0 − y) + o(y0 − y) y0−1 − y0−1

= y0−1 (y0 − y)y0−1 + o(y0 − y)

because we can Taylor expand (1 − x)−1 as soon as kxk < 1. The conclusion follows.
The product rule is, visibly, more generally valid when Y is a topological K-algebra. We are not
attempting to state the results in the most general possible settings.

B.2 Weak holomorphy

Everything we said about differentiability so far was elementary. The theory of holomorphic functions
in topological vector spaces is rich. Note that a function C → Cm is holomorphic iff all of its components
are. One can generalize this and define:
Definition B.7. Let X be a complex topological vector space, U ⊆ C open and f = U → X a
function. Then f is weakly holomorphic in U if for all φ ∈ X ∗ , the C-valued function φ ◦ f : U → C is
holomorphic.
It turns out that the converse holds:

Theorem B.8. Let U ⊆ C open.

1. (Dunford) If X is a Banach space, a weakly holomorphic f : U → X is strongly holomorphic.
2. More generally, this holds if X is a topological vector space, locally convex (its topology is
induced by a set of seminorms) and quasi-complete (meaning that every closed bounded subset
is complete).
Proof. 1. See [Yosida, 1980, §V.3, Theorem 1]. The proof uses a Cauchy integral formula in Banach
spaces via Bochner-integration, discussed briefly below.
2. See [Garrett, 2005]. The proof uses the more general notion of Gelfand–Pettis integration.

This implies in particular to Fréchet spaces (complete locally convex Hausdorff vector spaces whose
topology is induced by a countable family of seminorms). A Banach space is Fréchet; its topology is
induced by only one (semi)norm. We will almost exclusively work with Fréchet spaces.
Before giving more background about integration in topological vector spaces, we illustrate the power
of this equivalence. Many results about holomorphic functions U → C generalize immediately to
Banach spaces:

Corollary B.9. Let X be a complex Fréchet space, U ⊆ C open and f : U → X. TFAE:

1. f is complex differentiable (i.e. strongly holomorphic)
2. f has all higher derivatives

Corollary B.10. The locally uniform limit f of holomorphic functions fn : U → X is holomorphic

...
Corollary B.11 (Hurwitz’s theorem). . . . in which case f 0 is the locally uniform limit of fn0 .
In particular, we can differentiate a uniformly convergent series term-wise.

76
Proof. The first corollary is immediate. For the second, there’s a subtlety. Suppose X is a Banach
space, for convenience. We will sketch a proof for more general X later. Let K ⊆ U be compact.
For λ ∈ X ∗ we have fn → f uniformly on K, so λ ◦ fn → λ ◦ f uniformly on K, and by Hurwitz,
λ ◦ fn0 → λ ◦ f 0 uniformly on K. Moreover, the convergence is uniform in λ as long as kλk remains
bounded. Now suppose ∃ > 0 : ∀N ∈ N : ∃n > N ∃xn ∈ K : kfn0 (xn ) − f 0 (xn )k > . We use (A.11) to
find λn ∈ X ∗ which sends fn0 (xn ) − f 0 (xn ) to 1 and has norm at most 1/. Then kλn k is bounded, yet
the convergence is not uniform in those λn . Contradiction.
Alternatively, one can build a theory of integration, prove a Cauchy formula for f 0 and directly mimic
the proof of Hurwitz’s theorem in the general setting. We include the main elements of the approach.

B.3 Three notions of integration

There are different notions of integration in topological vector spaces. We will not use any deep results
about them and in fact all we need is some notion of integration, and a way to estimate integrals using
a triangle inequality. We present three approaches.
Given a measure space S and a topological K-vector space X, the least we can expect is that integration
commutes with continuous linear functionals. That is, for f : S → X:
Z  Z
λ f = λ◦f ∀λ ∈ X ∗
S S
R
We call f weakly integrable if such a vector S f exist, in which case we call it a weak integral or
Gelfand–Pettis integral. From the definition, we have: when RT : X → Y isR a continuous linear map
between topological vector spaces and f has a weak integral S f , then T ( S f ) is a weak integral of
T ◦ f . When X is locally convex, the Hahn–Banach separation theorem
R implies that continuous linear
functionals separate points, hence there can exist at most one such S f . One can show that:
Theorem B.12 (Existence of Gelfand–Pettis integrals). Let S be a locally compact Hausdorff topolog-
ical space with a finite positive Borel measure. Let X be a locally convex and quasi-complete complex
topological vector space. Then a compactly supported continuous f : S → X is weakly integrable.
In particular, we can take S a Lebesgue-measurable subset of some Rn , of finite measure, and we can
take X to be a Fréchet space or Banach space. Local convexity should be thought of as requiring that
convex linear combinations of small vectors are still small.
Proof. See [Garrett, 2014b, Theorem 1.0.1]. The proof is nonconstructive in that it uses compactness
of a certain set, to show that the intersection of a certain family of closed sets is nonempty. This
compactness, in its turn, relies on Tychonoff’s theorem (and thus on the axiom of choice) via [Garrett,
2014b, Proposition 4.0.1].
The above approach differs quite substantially from the theory of Riemann or Lebesgue integration.
But they do have infinite dimensional analogues.
The Riemann integral over an interval [a, b] ⊆ R is defined in the same way as for real-valued functions.
Given a topological vector space X and a function f : [a, b] → X we can define the Riemann sum for
every finite partition (subdivision) of [a, b] and for every choice of ‘tags’, which are points in the
closed subintervals defined by the partition. Partitions of [a, b] form a directed set for the relation of
being a refinement, and every choice of tags defines a net on the set of partitions, whose values are
Riemann sums. We call f Riemann-integrable if the net of Riemann sums converges to a common
Rb
vector a f ∈ X for every choice of tags. Because the Riemann-sums are linear in f , it follows that a
Riemann-integral (when it exists) is a Gelfand–Pettis integral. One can show:
Theorem B.13 (Existence of Riemann-integrals). If X is a complete and locally convex R-vector
space, then any continuous f : [a, b] → X is Riemann-integrable. Moreover, the nets converge in a
uniform way, in the following sense. Denote for a partition ∆ of [a, b] the maximum distance between
two adjacent points of the subdivision by |∆|. Then for every neighborhood U of 0 in X, there exists
 > 0 such that the Riemann sums for all partitions ∆ with |∆| <  and all choices of tags, lie in U .

77
Sketch of proof. For detailed computations, see [Nagy, 2014]. First, one uses local convexity and
uniform continuity of f to prove that, for every neighborhood U of 0, there exists  > 0 such that if
|∆| <  and Σ is a refinement of ∆, then the difference between their Riemann sums lies in V . (No
matter the choice of tags.)
Then take a sequence n → 0 and successive refinements ∆n with |∆n | < n (and arbitrary tags).
Use the first claim to conclude that the sequence of Riemann sums is Cauchy, hence convergent by
completeness.
Now take a neighborhood U of 0, take  as in the first claim and ∆ with |∆| < . Considering a
common refinement Σn of ∆ and ∆n , and using the triangle inequality, one shows that the Riemann
sums for ∆ (for any choice of tags) lie in 2U , and the conclusion follows.
R R
If f is continuous, then from the definition it follows that p◦f 6 p( f ) for every continuous seminorm
p of X.
In order to generalize the Lebesgue-integral, we need a notion of measurability. Let (S, µ) be a measure
space and X a K-Banach space. We call f : S → X weakly measurable if λ ◦ f is measurable for every
λ ∈ X ∗ , and strongly measurable if it is a.e. equal to the pointwise limit of a sequence of simple
functions. Here, simple functions are defined as for real-valued functions: f is simple if its support
has finite measure and if there exists a finite measurable partition into sets on which f is constant. If
I ⊆ R is an interval, a continuous map I → X is strongly measurable. (The argument is the same as
for real-valued functions.)
A strongly measurable function is visibly weakly measurable, and its image must be separable (have a
countable dense subset). In fact:
Theorem B.14 (Pettis). A function f : S → X is strongly measurable iff it is weakly measurable
and there exists a subset T ⊆ S whose complement has measure 0, and such that the image f (T ) is
separable.
Proof. See [Yosida, 1980, §V.4].
A strongly measurable f : S → X is called Bochner-integrable if there exists a sequence of simple
functions fn that converges a.e. to f , and such that the kf − fn k are integrable and:
Z
lim kf − fn k dµ → 0
n→∞ S

A simple function is thus Bochner-integrable, and we can define its Bochner-integral in the obvious
way, as the finite weighted sum of its values with the measure of each preimage as weight.
R For a
Bochner-integrable function, we define its Bochner-integral as the limit of the integrals S fn , with fn
as above. One can show that this limit exists and does not depend on the choice of fn . See [Yosida,
1980, §V.5]. From the definition, we see that the Bochner-integral (when it exists) is a Gelfand–Pettis
integral.
Theorem B.15 (Bochner). Let (S, µ) be a measure space, X a K-Banach space and f : S → X a
strongly measurable function.
1. f is Bochner-integrable iff kf k : S → R is µ-integrable, in which case we have the triangle
inequality: Z Z

f 6
kf k
S S

2. If Y is a K-Banach space, f : S → X is Bochner-integrable and A ∈ L(X, Y ), then A ◦ f is

Bochner-integrable (in particular, strongly measurable) and
Z Z
A◦f =A f
S S

Proof. See [Yosida, 1980, §V.5, Theorem 1, Corollary 1, Corollary 2]. Admitting all other statements,
the last equality also follows because we know Bochner-integrals are Gelfand–Pettis integrals, which
are unique in the case of Banach spaces.

78
Because the Riemann and Bochner integrals are Gelfand–Pettis integrals, we conclude that when they
both exist, they must be equal. (At least, when working with a locally convex space, so that Gelfand–
Pettis integrals are unique when they exist.) We can thus simply speak of the ‘integral’ of a function,
without ambiguity.
Remark B.16. We are not saying that when X is a Banach space, every Riemann-integrable function
is Bochner-integrable. In fact, this is no longer true in the infinite-dimensional case.13 The condition
for Bochner-integrable functions to be separably-valued up to a null-set, turns out to be quite strong.
There is an ambiguity about integrability of functions that take value in a linear subspace: Let X be
a topological vector space, Y a linear subspace and f : S → Y a function. If f is weakly integrable
as a Y -valued function, then it is (by the general result about continuous linear maps remarked in the
beginning) weakly integrable as an X-valued function, and the weak integrals coincide. What about
the converse? We make the following elementary observations:
1. If Y is a closed linear subspace and f is Riemann-integrable as an X-valued function, it is also
Riemann-integrable as a Y -valued function, simply by uniqueness of limits.
2. If X is a normed space, Y is any linear subspace and f is weakly integrable as an X-valued
function, it is also -weakly integrable as a Y valued function: by Hahn–Banach, continuous
linear functionals of Y extend to X.

B.4 Power series and meromorphy

Let X be a locally convex and quasi-complete topological C-vector space. Then it has all Gelfand–
Pettis contour integrals (B.12) and linear functionals separate points. We have seen (without proof)
that weak and strong holomorphy are equivalent in this case. Much like the equivalence of weak and
strong holomorphy reduces statements about holomorphic functions U → X to holomorphic functions
U → C, the fact that the integral commutes with linear functionals allows to reduce many statements
about vector-valued integrals to statements about integration of C-valued functions:
Proposition B.17. 1. Change of variables in integration of differentiable functions R → X. In
particular, contour integrals are well-defined.
2. Cauchy’s integral theorem: if f : U → X is holomorphic, then its integral along closed con-
tractible contours is 0.
3. Conversely, if f is continuous and the above holds, then f is holomorphic.
4. Homotopy invariance: if C1 , C2 are homotopic contours in U and f : U → X holomorphic, then
Z Z
f= f
C1 C2

5. Cauchy’s integral formula: Z

1 f (z)
f (z0 ) = dz
2πi ∂B(z0 ,δ) z − z0

6. Cauchy’s integral formula for derivatives:

Z
1 f (z)
f (n) (z0 ) = dz
2πi B(z0 ,δ) (z − z0 )n+1

Here for Cauchy’s integral representation for derivatives we also use that differentiation commutes with
linear forms X → C.
Note also how the criterion for holomorphy by contour integration allows an alternative proof for the
fact that the locally uniform limit of holomorphic functions is holomorphic. Also, Cauchy’s integral
13 Contrary to the case of real-valued functions, where every Riemann-integrable function is Lebesgue-integrable.

79
representation for the first derivative allows to generalize Hurwitz’s theorem directly: by the triangle
inequality for seminorms of Riemann-integrals in locally convex spaces, the locally uniform convergence
fn → f implies, together with Cauchy’s integral formula, the uniform convergence fn0 → f 0 . We also
have:

Proposition B.18 (Laurent expansion). Let B 0 (z0 , R) ⊆ C be a punctured disc, X as before and
f : B 0 (z0 , R) → X a mapping. Then f is holomorphic iff there exist (an )n∈Z ∈ X Z for which, pointwise,
X
f (z) = an (z − z0 )n (z ∈ B 0 (z0 , R))
n∈Z

in which case the convergence is uniform and absolute (for every continuous seminorm) in the annuli

{r1 < |z − z0 | < r2 } (0 < r1 < r2 < R)

and the Laurent-coefficients an are then uniquely determined by

Z
1 f (z)
an = dz
2πi ∂B(z0 ,δ) (z − z0 )n+1

Proof. The proof is the same as for complex-valued functions, with the absolute value replaced by
a family of seminorms. Note that, while the uniqueness of the an and the formula for them can
be deduced from the Hahn–Banach separation theorem and the complex-valued case, their existence
cannot.
We see that holomorphic functions can equivalently be defined as functions that are locally power
series. We can define a meromorphic X-valued function as one with isolated singularities and whose
Laurent-expansion has a finite singular part at every point. This can be checked on linear forms:
Proposition B.19. Let f : U → X be a function.
1. f is meromorphic iff λ ◦ f is meromorphic, for all λ ∈ X ∗ .

And if X is a topological algebra:

2. Meromorphic functions U → X form a ring.
If U is an open connected subset of C, the reciprocal of a nonzero holomorphic function U → C is
meromorphic. Consider a Banach algebra X, and a holomorphic function f : U → X which is not
identically zero. If f (s) is invertible for s in a punctured neighborhood of (say) 0, then 0 is an isolated
singularity of 1/f . But 1/f need not be meromorphic at 0. Indeed: suppose f (0) is not invertible, so
that if 1/f is not holomorphic at s0 . If it is meromorphic, then 1/f (s) ∼ (s − s0 )−N A for some N > 0
and A ∈ X − {0}. In particular:
f (0)A = Af (0) = 0
But if in addition f (0) is not a zero-divisor, this gives a contradiction. In (C.7), we give a sufficient
condition for 1/f to be meromorphic.

B.5 Integration in function spaces

We come back to the question about holomorphy of functions f : M ×U → C, where M is a Riemannian
manifold.
Proposition B.20 (Continuity of L2 -integrals). Let M be an orientable Riemannian manifold. Let
I ⊆ R be a compact interval. Let f : M × I → C be a function such that f (·, t) ∈ L2 (M ) for all t ∈ I.
Suppose that t 7→ f (·, t) is continuous, so that it is integrable:
Z
F (·) = f (·, t)dt ∈ L2 (M )
I

80
 1. Let N ⊆ M be a measurable subset. Then t 7→ f (·, t)|N ∈ L2 (N ) is still integrable and
Z
F (·)|N = f (·, t)|N dt ∈ L2 (N )
I

2. If f is (jointly) continuous, then F (·) is continuous (that is, it has a continuous representative)
and Z
F (x) = f (x, t)dt ∈ C
I
for all x ∈ M .
Proof. 1. Because integration commutes with continuous linear maps, in particular, the restriction
to N .
2. While it is true that linear operators commute with integration, the evaluation map at x is ill-
defined on L2 . We can extend it from L2 (M ) ∩ C 0 (M ) using Hahn–Banach. But C 0 (M ) cannot
in any obvious way be mapped to L2 (M ), so we still need an argument why F (·) is continuous.
Continuity is a local condition, so by the first statement we may assume M is compact. (Other-
wise, replace M by a compact neighborhood of x.) In that case, we have a continuous inclusion
map i : C 0 (M ) → L2 (M ). Here, C 0 (M ) is equipped with the L∞ norm. Its image need not be
closed. But f : I → C 0 (M ) is weakly integrable: it is continuous by uniform continuity of f .
Hence when composed with i, we obtain the (weak) integral of f as a L2 (M )-valued function.
We conclude that F (·) has a continuous representative. Finally, we can evaluate it in x by
Hahn–Banach, as we remarked already.
Proposition B.21 (Regularity of complex L2 -derivatives). Let M be an orientable Riemannian man-
ifold. Let U ⊆ C be open. Let f : M × U → C be a function such that f (·, s) ∈ L2 (M ) for all s and
f : U → L2 (M ) is holomorphic.
1. If f is (jointly) continuous, then so is f 0 (w, s).
2. If f is jointly (resp. separately) smooth, then so is f 0 (w, s).
Proof. 1. Suppose f is (jointly) continuous. Fix (w0 , s0 ) ∈ M × U . We have by Cauchy’s integral
formula (B.17): Z
0 1 f (·, ζ)
f (·, s0 ) = dζ ∈ L2 (M )
2πi ∂B (ζ − s0 )2
where B is some small ball centered at s0 . By (B.20), we can evaluate this in w. For s in a
small compact neighborhood V of s0 and w in a compact neighborhood W of w0 , we have that
f (w, ζ)/(ζ − s)2 becomes uniformly continuous on W × V × ∂B, so the integral
Z
1 f (w, ζ)
f 0 (w, s) = dζ
2πi ∂B (ζ − s)2
defines a continuous function of (w, s).
2. Similarly, from Cauchy’s integral formula.

B.6 Holomorphy in function spaces

Continuing the remarks in the introduction to this section, we make the following elementary obser-
vations:
Proposition B.22 (Joint regularity implies C k -smoothness). Let U be an open set of C = R2 and M
a σ-compact Hausdorff topological space. We have the Banach space Cb0 (M ) of bounded continuous
functions with the supremum norm, and the Fréchet space C 0 (M ) of continuous functions with the
topology of locally uniform convergence. Let f : M × U → C be jointly continuous. Denote the real
coordinates on U by (t1 , t2 ).

81
 |α|
1. If the partial derivatives ∂∂tα f of all orders exist and are jointly continuous, then f : U → C 0 (M )
is smooth. If they are also bounded, then f : U → Cb0 (M ) is smooth.

Now let M be a smooth manifold. We have the Fréchet space C ∞ (M ) with the topology of locally
uniform convergence of all partial derivatives on compact subsets of coordinate neighborhoods.
2. Let f be jointly smooth. Then f : U → C ∞ (M ) is smooth.
Proof. 1. Suppose those partial derivatives are jointly continuous, and let K ⊆ M be compact. Let
s0 ∈ U and write, for (h, k) ∈ R2 small and w ∈ K:

∂f (w, ·) ∂f (w, ·)
g(w, s0 , h, k) = f (w, s0 + (h, k)) − f (w, s0 ) − (0) · h − (0) · k
∂t1 ∂t2
For f to be differentiable at s0 , we have to show that g(w, s0 , h, k) = o(k(h, k)k) uniformly on
K. By the mean value inequality:

kg(w, s0 , h, k)k 6 k(h, k)k · ∇g(w, s0 , ·)(ξ(h,k) )

for some ξ(h,k) ∈ B(0, k(h, k)k). Because ∇g(w, s0 , 0) = 0 and ∇g(w, s0 , ξ) is continuous in ξ,
the RHS is ow (k(h, k)k). Moreover, it is jointly continuous in (w, ξ), and differentiable in ξ with
continuous derivative. We use the mean value inequality once more and by compactness of K,
we conclude that the RHS is oK (k(h, k)k), independently of w ∈ K. Thus f : U → C 0 (M ) is
differentiable. By induction, it is smooth.
If all partial derivatives of f : U → C 0 (M ) are bounded, then we can take K = U is the above
proof, and the conclusion follows.
2. This follows from 1., because smoothness of f : U → C ∞ (M ) is equivalent to smoothness of all
partial derivatives ∂x∂α f : U → C ∞ (V ) ⊆ C 0 (V ), for charts (xi ) : V → Rdim M .
Similarly one proves:
Proposition B.23 (Pointwise holomorphy iff C k -holomorphy). Let U be an open set of C = R2 and
M a smooth manifold.
1. If f : M × U → C is continuous, then it is pointwise holomorphic iff f : U → C 0 (M ) is
holomorphic.
2. If for every chart (xi ) of M the partial derivatives ∂ |α| f (w, s)/∂xα of any order exist and are
continuous, then f is pointwise holomorphic iff f : U → C ∞ (M ) is holomorphic.
Proof. That C 0 -holomorphy implies pointwise holomorphy, is because evaluations are well-defined
linear forms on C 0 . The other direction is proved as for (B.22), by using the mean value inequality.
That the regularity conditions on f can be weakened under the assumption of pointwise holomorphy,
follows from (B.3).

For convenience, we will say that f is C 0 -smooth, Cb0 -smooth, C 0 -holomorphic, C ∞ -holomorphic, etc.
We study the relation with L2 -holomorphy and pointwise holomorphy.14 From now on, let M be a
Riemannian manifold, U ⊆ C open and f : M × U → C a function.
Proposition B.24 (L2 -holomorphy implies pointwise holomorphy). Let f be (jointly) continuous and
such that f (·, s) ∈ L2 (M ) for all s. Suppose f : U → L2 (M ) is holomorphic. Then f is pointwise
holomorphic with f (w, ·)0 (s) = f 0 (·, s)(w) for all w, s.
Proof. Fix w ∈ M and s0 ∈ U . By assumption,

f (w, s) − f (w, s0 ) = f 0 (w, s0 )(s − s0 ) + (s − s0 )R(w, s) (s → s0 )

14 Most results will still hold with L2 replaced by Lp , for 1 6 p < ∞.

82
where R(w, s) = o(1) (in L2 (M )). By (B.21), f 0 (w, s0 ) is continuous in w. We want to look at this
equation for fixed w and not just as an L2 statement. We have R(w, s0 ) = 0 almost everywhere, but
that is not enough. W.l.o.g. suppose s0 = 0. We may also assume f (w, 0) = f 0 (w, 0) = 0 for all w.
For small s, by Cauchy’s integral formula in L2 :

Z
f (·, s) 1 f (·, ζ)
R(·, s) = = dζ
s 2πi ∂B ζ · (ζ − s)

for some ball B centered at 0 ∈ U . By (B.20), this is continuous for fixed s, we can evaluate it in w
and from the integral we then see that R is jointly continuous.
Now suppose f (w, s) is not complex-differentiable at s = 0. Then there exists  > 0 and a sequence
sn → 0 with |R(w, sn )| > . By joint continuity of R(w, ·), we have |R(w, 0)| >  and by joint
continuity there exists a neighborhood of (w, 0) on which |R(z, s)| > /2. But then kR(·, s)kL2 is
bounded from below as s → 0, contradiction.
Proposition B.25 (L2 -holomorphy implies C k -holomorphy). Let f be jointly continuous and such
that f (·, s) ∈ L2 (M ) for all s. Suppose that f is L2 -holomorphic.
1. Then f is C 0 -holomorphic and its C 0 -derivatives coincide with the L2 -derivatives.
2. If for every chart (xi ) the partial derivatives ∂ |α| f (w, s)/∂xα up to order n exist and are jointly
continuous, then f is C ∞ -holomorphic.
Proof. By (B.24), f is pointwise holomorphic. The two statements now follow from (B.23).
Lemma B.26. Let f : B(0, R) → C be holomorphic and 0 < r < R. Then
Z
sup |f (s)| r,R |f (z)||dz|
s∈B(0,r) ∂B(0,R)

Proof. By Cauchy’s integral formula, for s ∈ B(0, r):

|f (z)|
Z
|f (s)|  |dz|
∂B(0,R) |z − s|

We conclude using |z − s| R,r 1 for z ∈ ∂B(0, R).

Proposition B.27 (L2 -holomorphy versus pointwise holomorphy). Let f : M × U → C be continuous
and such that f (·, s) ∈ L2 (M ) for all s. Then the following are equivalent:
1. f is L2 -holomorphic.
2. f is pointwise holomorphic, and s 7→ kf (·, s)k2 is locally bounded.
3. f is pointwise holomorphic, and s 7→ f (·, s) is locally bounded by an L2 -function, independently
of s.
Remark B.28. For the extra condition in the last statement, it suffices in particular that f is con-
tinuous with support contained in T × U for some compact T ⊆ M : it is then locally bounded by a
function that is constant on T and 0 elsewhere.
Proof. 1 =⇒ 2: Pointwise holomorphy is proven in (B.24). The local boundedness of the norms
kf (·, s)k2 follows by continuity of f : U → L2 (M ).
2 =⇒ 1: As a general fact, the local boundedness of kf (·, s)k2 implies that of kf 0 (·, s)k2 : We
have
Z
1 f (w, z)
f 0 (w, s) = dz
2πi ∂B (z − s)2

83
 where B is a small ball centered at s, which we may assume of fixed radius δ > 0 as long as s
stays in a compact set. Then
Z  Z 2
2  1 f (w, z) 
kf 0 (·, s)k2 = 
2
dz  dw
M 2πi ∂B (z − s)

Z Z
δ |f (w, z)|2 |dz|dw
M ∂B
Z
2
= kf (·, z)k2 |dz|
∂B
δ 1
by Cauchy-Schwarz and Fubini. In the last step we used that z 7→ kf (·, z)k2 is locally bounded.
Thus in particular, the higher derivatives f (n) (·, s) are in L2 and their norms are also locally
bounded.
There is a converse. Suppose f 0 (·, s) has locally bounded norm, then
Z Z 2

2 0
kf (·, s) − f (·, s0 )k2 = f (w, z)dz  dw
 

M  [s0 ,s] 
Z Z
6 |s − s0 | |f 0 (w, z)|2 |dz|dw
M [s0 ,s]
Z
2
= |s − s0 | kf 0 (·, z)k2 |dz|
[s0 ,s]
2
 |s − s0 | 2
sup kf 0 (·, z)k2
z∈B(s0 ,|s−s0 |)

where [s0 , s] is a straight segment. Thus the local boundedness of kf 0 (·, z)k2 implies the conti-
nuity of s 7→ f (·, s).
Now fix s0 ∈ U . by subtracting from f the separately continuous L2 function (s − s0 )f 0 (w, s0 ),
we may assume that f 0 (w, s0 ) = 0.
From the computations above we successively have that kf 00 (·, s)k2 is locally bounded, that
kf 0 (·, s)k2 is continuous at s0 and that f (·, s) is differentiable at s0 .
3 =⇒ 2: Immediate.
2 =⇒ 3: Let s0 ∈ U , and choose 0 < r < R such that B(s0 , R) ⊂ U and that kf (·, s)k2 is
bounded on B(s0 , R). Let
g(w) = sup |f (w, s)|
s∈B(s0 ,r)
2
By (B.26) applied to all f (w, ·) :
Z Z Z
2
g(w) dw r,R |f (w, z)|2 |dz|dw
M M ∂B(s0 ,R)
Z
2
= kf (·, z)k2 |dz|
∂B(s0 ,R)

R 1
so that g is an L2 -function bounding all f (·, s) for s ∈ B(s0 , r).
3 =⇒ 1: This follows of course from 3 =⇒ 2 =⇒ 1, but we can give a direct argument.
So suppose f (·, s) is locally bounded by an L2 function. Cauchy’s integral formula shows that
the same holds for f 0 (·, s), and the mean value theorem shows that it is also true for (f (·, s) −
f (·, s0 ))/(s − s0 ). By holomorphy at fixed w, we have:
f (w, s) − f (w, s0 ) = f 0 (w, s0 )(s − s0 ) + R(w, s) (s → s0 )

84
 where R(w, s) = ow (s − s0 ). We can now apply dominated convergence to R(w, s)/(s − s0 ), and
we obtain
R(w, s)
s − s0 2 → 0 (s → s0 )

L

which implies that f : U → L2 (M ) is complex-differentiable at s0 .

Remark B.29. We do not know15 whether pointwise holomorphy, without any additional condition,
implies L2 -holomorphy, for continuous f : M × U → C such that f (·, s) ∈ L2 (M ) for all s ∈ U . We
have a partial result: Because the L2 -norm is lower-semicontinuous, using the Baire category theorem
one can show that kf (·, s)k is locally bounded in an open dense set, hence it is L2 -holomorphic in that
open dense set.
We know that the implication does not hold with ‘holomorphic’ replaced by ‘real analytic’. For
example, the function

f :R×R→R
xs
(x, s) 7→
1 + (xs)2

is jointly continuous, pointwise real analytic and in L2 (R) for fixed s, but it is not real analytic as an
L2 (R)-valued function: by a change of variables we have
1
kf (·, s)kL2  (s → 0)
|s|

So f : R → L2 (R) is not even continuous at 0.

B.7 Meromorphy in function spaces

If X is (say) a Fréchet space, by definition a meromorphic function f : U → X is locally of the form
g(s)
h(s) with g : U → X and h : U → C holomorphic, h not a zero divisor. We conclude that (with the
same notations as before):
Proposition B.30. If f : M × U → C is C 0 -meromorphic, then it is uniformly (pointwise) meromor-
phic.
A subtlety arises when we have no reason to assume that the numerator g(w, s) is jointly continuous:

Proposition B.31 (L2 -meromorphy implies C k -meromorphy). Let S ⊆ U be closed and discrete,
f : M × (U − S) → C be continuous and L2 -meromorphic. Then it is C 0 -meromorphic (and thus
uniformly meromorphic). If in addition f is smooth in w on M × (U − S) with jointly continuous
partial derivatives, then f is C ∞ -meromorphic.
Proof 1. The question is local, so we may assume S is finite and the set T of L2 -poles of f is finite.
Take a complex polynomial h(s) such that h(s)f (w, s) is L2 -meromorphic. Using what we know about
L2 -holomorphic functions (B.25), it suffices to prove that h(s)f (w, s) is continuous on M × U and not
just on M × (U − (S ∪ T )), which is done in the lemma below.
Proof 2. We have that f is C 0 -holomorphic with isolated singularities. The Laurent-coefficients of
those singularities can be expressed using integrals, and one sees that they are continuous in w. Hence
when they are zero in L2 , they are zero a.e. in w, hence zero everywhere in w.
Lemma B.32. Let S ⊆ U be closed and discrete, f : M × U → C be L2 -holomorphic and f :
M × (U − S) → C be continuous. Then f is continuous on M × U . If in addition f : M × (U − S) → C
is jointly smooth or smooth in w with jointly continuous partial derivatives, then f has that same
property on M × U .
15 That is, we do not know whether it is known.

85
Proof. Using Cauchy’s integral formula in L2 , we write f (·, s0 ) in terms of the values f (·, s) for s on
a small circle around s0 . By (B.20), we can evaluate the integral at points of M and the continuity
at (w, s0 ) follows from the continuity at (w, s) for s on that small circle. Any additional regularity
properties follow from the same integral formula.

From the second proof of (B.31) we also see that the order of the C 0 -poles equals the order of the
L2 -poles. Using the same technique, we prove a converse to (B.30):
Proposition B.33 (Uniform meromorphy implies C k -meromorphy). Let S ⊆ U be closed and discrete,
f : M × U → C uniformly meromorphic and f : M × (U − S) → C continuous.
1. Then f is C 0 -meromorphic.

2. If in addition, f : M × (U − S) → C is smooth in w with continuous partial derivatives, then it

is C ∞ -meromorphic.
Proof. By (B.23), f is C 0 - (resp. C ∞ -) holomorphic with isolated singularities. We know that the
Laurent-coefficients of isolated C 0 -singularities are continuous in w. (This time, this is simply by
definition of the space the holomorphic function takes values in.) If they are 0 at every point, then
they are 0 in C 0 (resp. C ∞ ). We conclude that the isolated C 0 -singularities are C 0 -poles, and similarly
for C ∞ .
Note that in (B.31), we haven’t said that the L2 -poles or C 0 -poles lie in S, and the two proofs we
gave, do not clarify this. It is true: using the relations between all those notions of holomorphy we
can reduce it to Riemann’s theorem on removable singularities, for functions U → C:

Proposition B.34 (Removable singularities). Let f : M × U → C be continuous. If f is L2 -

meromorphic (resp. C 0 -meromorphic) on M × U , then it is L2 -holomorphic (resp. C 0 -holomorphic).
Proof. Under the conditions from the statement, we know that L2 -meromorphy implies C 0 -meromorphy,
and that the L2 -orders of the poles are the same as their C 0 -orders. Suppose s0 is a pole. Because f
is C 0 -meromorphic, we can compose f with evaluation in each w and deduce that s0 is a pole of some
f (w, ·). Fix such a w. Then f (w, s) is meromorphic yet continuous in a neighborhood of s0 . Then s0
is a removable singularity of f (w, s), contradiction.

86
C Fredholm integral equations
In this section, we let M be an oriented Riemannian manifold equipped with its canonical measure,
and let k : M × M → C be a (measurable) kernel. We assume that k defines a bounded convolution
operator on L2 (M ) by Z
K : g(x) 7→ k(x, y)g(y)dy
M

This is the case if k ∈ L2 (M × M ) is Hilbert–Schmidt (A.56) but it can also be bounded without being
square-integrable (4.22). We denote by kKk its operator norm, and by kkk2 the L2 norm when it is
finite, which then equals the Hilbert Schmidt norm kKk2 . Recall that kKk 6 kKk2 . We study the
Fredholm equation (of the second type)

(C.1) (1 − λK)g = f

where λ ∈ C and f ∈ L2 (M ). We seek to answer the questions:

1. When does (C.1) have a unique L2 solution g?
2. Is g smooth when f is smooth?
3. Is the dependence of g on λ holomorphic? Meromorphic?
The operator 1 − λK has a bounded inverse (by definition) when λ = 0 or λ−1 is not in the spectrum
−1
of K. This answers the first question. Observe that it suffices that |λ| < kKk , in which case

(C.2) (1 − λK)−1 − 1 = λK + λ2 K 2 + · · ·

where the convergence is for the operator norm. This defines a holomorphic function of λ in the open
disk B(0, kKk). We cannot expect this identity to extend to larger values of λ, simply because 1 − λK
is usually not invertible for certain values of λ.
We will first discuss smoothness, and then present two approaches to study holomorphic and meromor-
phic dependence on λ, without restricting to the disk B(0, kKk). The first, perhaps the most direct
approach, is abstract in nature and relies on notions of differentiability in Banach spaces. It applies to
all kernels that define a bounded operator. The second approach is due to Fredholm, who constructed
an explicit meromorphic continuation of the inverse (1 − λK)−1 when K is a compact operator.

C.1 Regularity
If λ ∈ C such that 1 − λK is invertible, the Fredholm equation (C.1) has a unique L2 solution

g = (1 − λK)−1 f

for all f ∈ L2 . The hope is that when f has nice properties, then so does g. In general, applying
a (reasonable) smooth kernel to a smooth function yields a smooth function. What’s special about
1 + λK, a ‘perturbation’ of the identity, is that the converse holds:
Theorem C.3 (Smoothness of the solution). Let k and f be as above but λ arbitrary. Let g be
any L2 solution to (1 − λK)g = f . Suppose that k is of class C n (n > 0) and that it is compactly
supported in the sense of (3.12)(4): for every compact V ⊆ M , the restriction k : V × M → C has
compact support. Suppose also that vol(M ) < ∞. Then f is of class C n iff g is of class C n .
Proof. We have Z
f (x) = g(x) − λ k(x, y)g(y)dy
M
The key is that the second term is always of class C n . Indeed, we fix x0 ∈ M and we restrict our
attention to a compact neighborhood V of x0 contained in a coordinate chart. Then the support of
the integrand is contained in some compact subset of M × M , independently of x ∈ V . Because M has

87
finite volume, g ∈ L2 implies g ∈ L1 by Cauchy-Schwarz. Because the integrand has compact support
and k is of class C n , its derivatives w.r.t. x up to order n are uniformly bounded by the integrable
function R · |g(y)| for some R > 0 independent of x ∈ V . By the dominated convergence theorem, we
conclude that the second term is of class C n .
Thus f and g differ by a C n function, and the conclusion follows.
Note how the parameter λ has no relevance in the above result; we could absorb it in the kernel k. It
is just there for the presentation.

C.2 The Fredholm equation for bounded operators

We now let λ and f depend holomorphically on some variable s, and prove that g depends also
holomorphically on s. While it may seem more natural to write our family of operators as µ(s) − K,
the problem is that when µ(s0 ) = 0 this can not be seen as a ‘perturbation of the identity’, and (C.3)
does no longer guarantee that the solutions g(·, s0 ) are continuous when f and k are.
Lemma C.4 (Holomorphic operators and holomorphic functions). Let X, Y, C be complex Banach
spaces and U ⊆ C open. Let A : U → L(X, Y ) and f : U → X be holomorphic at s0 ∈ U . Then
Af : U → Y is holomorphic at s0 with derivative

(Af )0 (s0 ) = A0 (s0 )f + A(f 0 (s0 ))

Proof. Analogous to the proof of the product rule. If we don’t want to repeat the proof, we can in
fact deduce it from the product rule by making L(X, Y ) ⊕ X into a Banach algebra by defining

(A, x)(B, w) = (AB, Aw + Bx)

with submultiplicative norm k(A, x)k := kAk + kxk, so that L(X, Y ) and X embed isometrically into
this space.

Theorem C.5 (Holomorphy of the solution). Let M and k be as in (C.3): M has finite volume and k
has compact support in the sense that for V compact, k : V × M → C has compact support. Suppose
in addition that k is continuous. Let U ⊆ C be open and λ : U → C a holomorphic function such that
all operators 1 − λ(s)K are invertible. Let f : M × U → C be continuous and supported in T × U for
some compact T ⊆ M . So f (·, s) ∈ L2 (M ) for all s ∈ U , and we can define

g(·, s) = (1 − λ(s)K)−1 f (·, s) ∈ L2 (M )

By (C.3), the assumptions on k, vol(M ) and the continuity of f imply that g(·, s) is continuous for all
s, and in particular we can evaluate it in w ∈ M . We then have:
1. The following are equivalent:

(a) f is pointwise holomorphic

(b) f is L2 -holomorphic
(c) g is L2 -holomorphic
(d) g is pointwise holomorphic and jointly continuous

2. If the above equivalent statements hold and k is smooth, then f is jointly smooth iff g is jointly
smooth.16
Proof. 1. We show that (a) =⇒ (b) =⇒ (c) =⇒ (d) =⇒ (a):
(a) =⇒ (b): Follows from the continuity of f and the condition on its support (B.27).
16 Note that, under the assumption of pointwise holomorphy, joint smoothness can be formulated in terms of the partial

derivatives w.r.t. w only (B.3)(3) and that joint smoothness implies C ∞ -holomorphy (B.23).

88
 (b) =⇒ (c): This is elementary. Because λ(s) is holomorphic, so is the the function B(s) =
1 − λ(s)K, which takes values in the Banach algebra of bounded operators on L2 (M ). By
assumption, the operators 1 − λ(s)K are invertible, so A(s) := B(s)−1 is holomorphic in s.
By (C.4), the function

g = Af : U → L2 (M )
s 7→ A(s)f (·, s)

is holomorphic.
(c) =⇒ (d): For holomorphy for fixed w, we use (B.27): it suffices that g(w, s) is jointly con-
tinuous. Fix (w1 , s1 ) and let (w2 , s2 ) vary in a compact neighborhood of (w1 , s1 ). We
write
Z Z 
 

 k(w 1 , y)g(y, s1 ) − k(w 2 , y)g(y, s )
2 

M M
Z Z
6 |k(w1 , y) − k(w2 , y)||g(y, s1 )|dy + |k(w2 , y)||g(y, s1 ) − g(y, s2 )|dy
M M

When (w2 , s2 ) → (w1 , s1 ), the first term approaches 0 by the dominated convergence theo-
rem, where we use that k has compact support and that g(·, s1 ) ∈ L2 ⊂ L1 . For the second
term we have
Z  Z
 

 k(w, y) (g(y, s1 ) − g(y, s2 )) dy  6 |k(w, y)| · |g(y, s1 ) − g(y, s2 )|dy
M M
6 R vol(M ) kg(·, s1 ) − g(·, s2 )kL1

where R = maxy∈M |k(w, y)| is finite because k(w, ·) has compact support. By Cauchy-
Schwarz and because vol(M ) < ∞,

kg(·, s1 ) − g(·, s2 )kL1  kg(·, s1 ) − g(·, s2 )kL2

Because g : U → L2 (M ) is holomorphic, it is in particular continuous, and the we conclude

that the second term goes to 0.
(d) =⇒ (a): Fix w. It suffices to show that the integral
Z
(C.6) k(w, y)g(y, s)dy
M

defines a holomorphic function. Therefore it suffices to show that the integrand is bounded
by an integrable function locally in s independently of s. Because k has compact support
and g is jointly continuous, we can take a constant function.
2. Once again, it suffices to show that the integral (C.6) is jointly smooth. We want apply dominated
convergence to switch the order of differentiation and integration. Because k is smooth, it remains
to argue that g (n) (y, s) is jointly continuous for all n, the derivative being with respect to s. This
follows from the joint continuity of g, and by repeatedly applying (B.3).
Suppose that (1 + λ(s)K)−1 has isolated singularities in U , which is the case when the spectrum of
K is discrete, possibly with the exception of 0 as an accumulation point. Then the above result says
that when f is (pointwise, say) holomorphic in U , then g is holomorphic with the exception of isolated
singularities at those s for which λ(s)−1 is in the spectrum of K. We are interested in meromorphy of
g, and thus we need to know when the isolated singularities of (1 + λK)−1 are poles.
Theorem C.7. Let K be a compact operator on a complex Banach space. Then every nonzero point
of its spectrum is a pole of its resolvent

R(s, K) = (K − s)−1

89
Sketch of proof. The proof uses holomorphic functional calculus to show that the restriction of K to
the range of the residue E(λ) at an eigenvalue λ 6= 0 is compact, invertible and has spectrum {λ}.
Thus this range is finite-dimensional, and by linear algebra in finite dimension over C, there exists n
with (K − λ)n E(λ) = 0. By investigating the Laurent-coefficients, this says precisely that λ is a pole
of order at most n.

Corollary C.8 (Meromorphy of the solution). Let all variables be as in (C.5), except for K, which
we assume to be a compact operator and λ which we assume nowhere constant. Then:
1. If f is pointwise holomorphic, g is uniformly meromorphic, and even C 0 -meromorphic and L2 -
meromorphic.

2. If f is in addition (jointly) smooth, then g is C ∞ -meromorphic.

Proof. 1. If f is pointwise holomorphic, then g = (1 + λ(s)K)−1 f is L2 -meromorphic because f is
L -holomorphic and (1 + λ(s)K)−1 is meromorphic. Because g has isolated singularities and is
2

continuous outside of them, we know from (B.31) that it is C 0 -meromorphic. In particular, it is

uniformly meromorphic.

2. Because g is C 0 -meromorphic and its Laurent-coefficients are automatically smooth.

C.3 Fredholm theory for compact operators

If K is Hilbert–Schmidt and λ < kKk2 < ∞, then (C.2) is also convergent for the Hilbert–Schmidt
norm. Recall that Hilbert–Schmidt operators form a Hilbert space (A.54) so that the limit will be
Hilbert–Schmidt.
We thus make the following elementary observation:
Proposition C.9. If K is Hilbert–Schmidt and λ < kKk2 , the inverse does not only exist but is also
given by a Hilbert–Schmidt integral operator:
Z ∞
X
(C.10) (1 − λK)−1 f (x) = f (x) + λ λj−1 kj (x, y)f (y)dy
M j=1

where the iterated kernels kj are given by k1 = k and

Z
kj (x, y) = k(x, z)kj−1 (z, y)dz (j > 2)
M
P∞
The hope is to meromorphically continue the kernel Rλ = j=1 λj−1 kj (x, y). Using the same notation
for the associated integral operator, we have

(C.11) (1 − λK)Rλ = K

When M is, say, the interval [0, 1], one can discretize the linear equation (1+λK)g = f , by replacing K
by the (n+1)×(n+1) matrix (Kij ) = (K( ni , nj ))06i,j6n , replacing f by the column vector (f ( ni ))06i6n
and similarly for g. By applying Cramer’s rule to this linear system and taking the limit n → ∞, sums
turn into integrals, and Fredholm obtained, for a general finite-volume Riemannian manifold M :
Theorem C.12. Suppose vol(M ) < ∞ and that k is continuous and bounded. There exists an entire
function, the Fredholm determinant D(λ), and an entire function Dλ (·, ·) with values in the Banach
space of continuous bounded kernels, such that

(C.13) (1 − λK) ◦ Dλ = D(λ) · K

Moreover, when k is smooth, D( · ) (·, ·) is jointly smooth.

90
Proof. See e.g. [Iwaniec, 2002, Appendix A.4]. The statement about smoothness is not mentioned
there, but can be seen from the explicit formula given for the Taylor coefficients of Dλ (x, y): applying
a differential √
operator L to those coefficients can make them larger, but they still satisfy a bound of
the form  ( mCL (x, y))m /m!, where m is the index, for some function CL (depending on L) which
can be taken constant locally in (x, y).

Comparing this with (C.11) we obtain, still under the assumption that vol(M ) < ∞:
Corollary C.14. D(λ)−1 Dλ defines a meromorphic continuation of Rλ , which is jointly smooth away
from poles of D. In particular, 1 − λK is invertible when λ is not a pole of D.
Corollary C.15. Let vol(M ) < ∞, λ : C → C be entire and f : M × C → C continuous and pointwise
holomorphic, with f (·, s) ∈ L2 for fixed s. Let k be a bounded kernel on M , compactly supported in
the sense that for compact V ⊆ M , the restriction k : V × M → C has compact support. Then:
1. For s ∈ C not a pole of D ◦ λ, there exists a unique solution g(·, s) to the Fredholm equation

(1 − λ(s)K)g(·, s) = f (·, s)

(which is automatically continuous, by (C.3)).

2. We have Z
λ(s)
g(w, s) = f (w, s) + Dλ(s) (w, y)f (y, s)dy
D(λ(s)) M

3. g(w, s) is meromorphic for fixed w. More precisely, it is a holomorphic function divided by

D(λ(s)).
4. If k is smooth and either:
(a) f has support contained in T × C for some compact T ⊆ M
(b) M has a global chart (xi ) in which the derivatives of k are still bounded kernels, and f (·, s)
is locally bounded by an L2 function independent of s
then f is jointly smooth iff g is.
Proof. 1. Because 1 − λK is invertible when λ is not a pole of D.
2. We have (1 − λK)−1 = 1 + λRλ for small λ, so that indeed 1 + λDλ D(λ)−1 is the meromorphic
continuation of (1 − λK)−1 .
3. Because s 7→ Dλ(s) is holomorphic, it is in particular bounded by a constant function locally in
s, so that the integral defines a holomorphic function for fixed w.
4. Once again, it suffices to show that the integral defines a jointly differentiable function. If the
integrand has uniformly compact support, which is the case if (a) holds, then we can freely choose
the order if integration and differentiation, and we are done. If (b) holds, then investigating the
Taylor coefficients of Dλ shows that its derivatives are still bounded kernels. Together with the
uniform integrability condition on f this allows us to differentiate inside the integral.
Note how the meromorphic continuation of the kernel Rλ gives us meromorphy for free, while with
the previous method we had to invoke holomorphic functional calculus to prove that the isolated
singularities are indeed poles.

91
D Riemannian geometry
The below can be found in [Lee, 1991, Chapters 1–6].
Notation D.1. Let M be a smooth manifold. Its tangent bundle will be denoted T M , vector fields on
M are global sections of the tangent bundle and form a module Γ(T M ) over C ∞ (M ). Smooth 1-forms
are global sections of the dual T M ∗ , and form a C ∞ (M )-module Ω1 (M ) = Γ(M, Ω1 ). It contains the
differentials df of smooth real-valued functions f ∈ C ∞ (M ).

D.1 Riemannian manifolds

Definition D.2. A Riemannian metric on M is a smooth global section g of the vector bundle
(T M ⊗ T M )∗ = T M ∗ ⊗ T M ∗ such that for all p ∈ M , g(p) is a positive definite symmetric bilinear
form on the tangent space Tp M . A pseudo-Riemannian metric is one which is not necessarily positive
definite. An isometry between Riemannian manifolds (M, g), (N, h) is a diffeomorphism σ for which
the pullback σ ∗ h = g. That is, such that if p ∈ M with σ(p) = q then gp (X, Y ) = hq (dσ|p X, dσp Y )
for tangent vectors X, Y ∈ Tp M . The isometry group is denoted Isom(M ).

Example D.3 (Submanifolds). A submanifold of a Riemannian manifold inherits a metric by restrict-

ing the metric to the tangent space of the submanifold. The restriction of a positive definite symmetric
bilinear form to a subspace has indeed the same properties.
Example D.4 (Covering maps). Let π : E → B be a covering map between manifolds, or more
generally a local diffeomorphism. If B has a Riemannian metric h, we can locally pull it back by π and
obtain a Riemannian metric on E. In particular, the universal cover of B has a canonical structure of
a Riemannian manifold. Conversely, suppose E has a Riemannian metric g and π is a Galois cover,
i.e. its automorphism group Aut(π) acts transitively on fibers. Suppose also that Aut(π) ⊆ Isom(E).
Then we can locally pushforward g to B in a well-defined way.
Example D.5 (Euclidean space). The Euclidean space Rn is a Riemannian manifold with metric
(dxi )2 defined in the standard (global) chart.
P

Example D.6 (Spheres). Let R > 0. Inside Rn+1 the n-sphere SR

n
inherits a Riemannian metric from
n
R , by (D.3).
Example D.7 (Hyperbolic space). Let R > 0. Hyperbolic n-space HnR can be defined in the following
equivalent ways:
(a) (Hyperboloid model) As the upper sheet {τ > 0} of the hyperboloid τ 2 − |ξ|2 = R2 in Rn × R
which inherits the pseudo-Riemannian Minkowski metric
X
(dξ i )2 − (dτ )2
i

from Rn+1 , as in (D.3).

(b) (Poincaré ball model) As the open ball B(R) in Rn with the metric in the standard chart given
by
i 2
P
4 i (du )
4R
(R2 − |u|2 )2

(c) (Poincaré half-space model) As the upper half-space HnR = Rn−1 ×R>0 with coordinates ((xi ), y))
and metric
i 2 2
P
2 i (dx ) + (dy)
R
y2

92
Proof of equivalence. In [Lee, 1991, Proposition 3.5] it is shown by direct computation that the hy-
perbolic stereographic projection gives an isometry between the hyperboloid and the Poincaré ball:
it sends a point P on the upper sheet to the unique intersection point of the segment [P, Q] with
B(R) ⊆ Rn = Rn × {0}, where S = (0, −R) is the hyperbolic south pole.
To compare the ball model and the half-space model, one shows that for n = 2 with the standard
complex structure on these varieties, the map
w + iR
w 7→ −iR
w − iR
defines a biholomorphism that respects the metric. In real coordinates (u, v) on B(R) it takes the form

2R2 u R2 − |u|2 − v 2
 
(u, v) 7→ , R
|u|2 + (v − R)2 |u|2 + (v − R)2

with inverse
2R2 x |x|2 + y 2 − R2
 
(x, y) 7→ , R
|x|2 + (y + R)2 |x|2 + (y + R)2
One then checks that these smooth maps are still inverses of each other when u and x take values in
Rn−1 , and that they still preserve the metric.
Remark D.8. With the Poincaré ball model, homotheties do induce diffeomorphisms between balls
B(R1 ), B(R2 ) of different radius, but they are not isometries. Likewise for spheres.

D.2 Connections
Definition D.9. Let M be a smooth manifold and E be a vector bundle over M . A connection in E
is a map
∇ : Γ(T M ) × Γ(E) → Γ(E)
denoted ∇(X, Y ) = ∇X Y , the covariant derivative of Y in the direction of X such that:
1. ∇X Y is linear over C ∞ (M ) in X.
2. ∇X Y is linear over R in Y .
3. We have the product rule:
∇X (f y) = f ∇X Y + (Xf ) · Y
for f ∈ C ∞ (M ).
A linear connection is a connection in T M .
Proposition D.10. For a linear connection ∇ in a vector bundle E at a point p:
1. ∇X Y |p depends only on the value of X at p and the values of Y in a neighborhood of p.
2. For a linear connection and a smooth curve with tangent vector X in a neighborhood of p, it
depends only on the values of Y along the curve.
Proof. See [Lee, 1991, Lemma 4.1, Lemma 4.2, Exercise 4.7].
Definition D.11. Let γ be a smooth curve in M defined on some open interval I. A vector field
along γ is a smooth map I → T M assigning to each t ∈ I a tangent vector at γ(t). Let Γ(γ) be the
set of vector fields along γ.
A vector field V along a curve can always be locally extended to a vector field on an open set of M
whose restriction to γ is (locally) V .
Example D.12. The derivative dγ/dt is a vector field along γ. More generally, f (t)dγ/dt is one. If
f is an immersion (i.e. df is injective at every point) then every vector field along γ is of this form.

93
Definition D.13 (Covariant derivative along a curve). Let ∇ be a linear connection on M and γ be
a smooth curve in M defined on some open interval, whose image . The covariant derivative along γ
is the map
Dt : Γ(γ) → Γ(γ)
defined by
(Dt V )(s) = ∇γ̇(s) Ve

where Ve is a smooth extension of V to a neighborhood of γ(s).

Proposition D.14. 1. Dt is linear over R.
2. It satisfies the product rule Dt (f V ) = f˙V + f Dt V .

Proof. See [Lee, 1991, Lemma 4.9].

Example D.15. In Rn , the Euclidean connection ∇X Y is defined by letting X act on the components
Y i of Y in the standard basis. Then differentiation of V along a curve γ means differentiating the
V i ◦ γ.

The linear combination of connections by a partition of unity is again a connection, so every manifold
admits a linear connection. ([Lee, 1991, Proposition 4.5])

D.3 Geodesics and parallel transport

Definition D.16 (Acceleration along a curve). Let M be a manifold with a linear connection ∇. Let
γ be a curve in M . The acceleration of γ is the vector field Dt γ̇ along γ.
Definition D.17 (Geodesic). Let M be a manifold with a linear connection. A geodesic with respect
to ∇ is a curve with acceleration 0.

By uniqueness, we can consider maximal geodesics: those who are defined on a largest possible interval.
Example D.18. In Rn with the Euclidean connection from (D.15), the geodesics are the straight
lines: those whose second derivative vanishes.
Proposition D.19 (Existence and uniqueness of geodesics). Let M be a manifold with a linear
connection. For any tangent vector V at any point p, there exists a neighborhood in which there is a
unique geodesic with derivative V at p.
Proof. Using the theory of differential equations. See [Lee, 1991, Theorem 4.19].
Definition D.20. A vector field V along a curve γ is parallel along γ iff Dt V = 0.
Proposition D.21 (Existence and uniqueness of parallel transport). For any curve γ on a manifold
M , point p = γ(t0 ) and tangent vector V0 at p, there is a unique parallel vector field along the whole
of γ extending V0 at t0 , called the parallel transport of V0 , denoted t 7→ Ptt0 V0 .
Proof. See [Lee, 1991, Theorem 4.11].
Thus a curve is geodesic iff the parallel transport of the derivative at some point of the curve, equals
the derivative at all other points.

94
D.4 Geodesics on Riemannian manifolds
From a linear connection on M , there is a natural way to obtain a connection on each T M ⊗a (T M ∗ )⊗b ,
see [Lee, 1991, Lemma 4.6]. In particular we can consider the connection on T M ⊗ T M ∗ and apply
the connection to a Riemannian metric considered as a section of that bundle.

Definition D.22. Let M be a Riemannian manifold with metric g. A linear connection ∇ on M is

compatible with g if the following equivalent conditions hold:
1. ∇X hY, Zi = h∇X Y, Zi + hY, ∇X Zi for all vector fields X, Y, Z.
2. ∇g = 0, with the action of ∇ defined as above.

3. For vector fields V, W along a curve:

d
hV, W i = hDt V, W i + hV, Dt W i
dt

4. For parallel vector fields along a curve, hV, W i is constant.

5. For any curve γ, parallel transport Pts is an isometry between the tangent spaces at γ(t) and
γ(s).
Proof of equivalence. See [Lee, 1991, Lemma 5.2].

Definition D.23. Let M be a manifold. A linear connection ∇ is symmetric if

∇X Y − ∇Y X = [X, Y ]

Proposition D.24. Let M be a Riemannian manifold. Then it has a unique compatible symmetric
linear connection, the Riemannian connection or Levi-Civita connection.
Proof. See [Lee, 1991, Theorem 5.4].
On Riemannian manifolds we will always work with the Riemannian connection and the derived notions
of (Riemannian) geodesics and parallel transport.

Example D.25. On Rn with the standard metric (dxi )2 , the Euclidean connection ∇ is compatible
P
(this is Leibniz’ rule) and symmetric (because the coordinate-wise definition of ∇ gives precisely the
Lie bracket).
Proposition D.26. An isometry between two Riemannian manifolds takes geodesics to geodesics.
Proof. Because everything we defined so far is functorial. See also [Lee, 1991, Proposition 5.6].

D.4.1 The exponential map

Definition D.27. Let M be a Riemannian manifold and p ∈ M . The exponential map at p sends
a tangent vector V to the point γV (1) of the corresponding geodesic γV at time 1, if the geodesic is
defined for that time.
Example D.28. For M = R, the exponential at t sends s to s + t. The exponential on T0 R is thus
very different from the Lie exponential.

Uniqueness implies γλV (1) = γV (λ) so that exp is defined on a star-shaped subset of Tp M centered
around the origin.
Proposition D.29. The exponential map at p ∈ M is smooth and restricts to a diffeomorphism
between a neighborhood of 0 ∈ Tp M and one of p ∈ M , called a normal neighborhood of p.

95
Proof. See [Lee, 1991, Proposition 5.7, Lemma 5.10]. The first statement comes from the theory of
differential equations, the second statement follows from the observation that, the differential of exp
at 0 is the identity map of Tp M , as seen by letting d exp act on tangent vectors in terms of germs of
curves, by composition with exp.

Given an normal neighborhood U of a point, normal coordinates at the point are the components of
exp−1 : U → Tp M in an orthonormal basis. If we call them (xi ), the radial distance function is
sX
r(x) = (xi )2
i

which is defined on U and the unit radial vector field

∂ X xi ∂ 1
= i
= · exp−1
∂r i
r ∂x r

which is defined on U − {p}. They do not depend on the choice of normal coordinates (i.e. of an
orthonormal basis of Tp M ). The norm of ∂/∂r is 1, because by construction (gij ) = (δij ) at the point
p, in normal coordinates.
Proposition D.30. With U as above, and q ∈ U − {p}, the vector ∂/∂r|q is the velocity vector of
the unit speed geodesic from p to q. That is, t 7→ exp(t∂/∂r|q )

Proof. Let r be the radius of q. Then indeed, exp(r∂/∂r|q ) = q, so that γ∂/∂r|q (r) = q. See also [Lee,
1991, Proposition 5.11].

D.4.2 Geodesics and distance

A curve (segment) defined on a closed bounded interval of R is one that extends smoothly to an open
interval containing it. A curve is regular if its differential is injective, i.e. it is an immersion.
Definition D.31 (Length of a curve). Let M be a Riemannian manifold and γ : [a, b] → M a
smooth curve. Its length is the integral of the norm of its derivative from a to b. It is invariant under
reparametrization.
Definition D.32. The Riemannian distance d(p, q) between two points p, q ∈ M is the infimum of
lengths of piecewise regular (equivalently, regular) curve segments joining the two.

Proposition D.33. This defines a metric which induces the original topology of M .
Proof. That it defines a pseudometric follows from the triangle inequality. For the topology and
positivity of the metric, see [Lee, 1991, Lemma 6.2].
This gives rise to the notions of geodesic ball and geodesic sphere.
One can consider smooth families of piecewise regular closed curves joining two points (with a common
finite set of possibly non-smooth points) indexed by an open interval. We call a curve critical if the
derivative of the length is zero at the curve, for every smooth family. A curve is minimizing if its
length equals the geodesic distance between its endpoints. A curve is locally minimizing if every point
of its interval of definition has a neighborhood to which the restriction is minimizing.
Using calculus of variations, one shows:

Proposition D.34. Every critical piecewise regular curve is in fact regular, and geodesic when we
give it constant speed parametrization.
Proof. See [Lee, 1991, Theorem 6.6, Corollary 6.7]. Note that constant speed reparametrization is
unique for regular curves.
Note also that:

96
Proposition D.35. Geodesics have constant speed.
Proof. Follows by (D.22) of the compatibility of the Riemannian connection. See [Lee, 1991, Lemma
5.5].
Proposition D.36. Let p ∈ M and q contained in a ball {r 6 } of a normal neighborhood of p,
where r is the radial distance to p. Then the radial geodesic from (D.30) is the unique minimizing
piecewise regular curve from p to q. Consequently, inside such a geodesic ball the radius r equals the
geodesic distance to p, and every geodesic is locally minimizing.
Proof. See [Lee, 1991, Proposition 6.10, Corollary 6.11, Theorem 6.12].
Corollary D.37. Let p ∈ M and U a normal neighborhood of p. Let r denote the geodesic distance
to p on U . Then r2 is smooth on U and r is smooth on U − {p}.
Proof. If q ∈ U has normal coordinates (xi ) then
X
r2 = (xi )2

which is smooth in q. Next, at q 6= p we have r 6= 0 so r too is smooth there.

D.4.3 Completeness
Because geodesics have constant speed (D.35), they can only be reparametrized by homotheties and
translations of the interval of definition, and the property of being defined on R is an intrinsic notion,
which only depends on the image of the geodesic.
Definition D.38. A Riemannian manifold is geodesically complete if all geodesics can be defined on
the whole of R.
Theorem D.39 (Hopf–Rinow). For a connected Riemannian manifold M , TFAE:
1. M is geodesically complete.
2. M has the Heine-Borel property: every closed bounded subset is compact.
3. M is complete as a metric space for the Riemannian distance.
4. There exists a point p for which the exponential is defined on the whole of Tp M .
5. Every two points can be joined by a (not necessarily unique) minimizing geodesic segment.
6. Every two points can be joined by a geodesic segment.
Proof. See [Lee, 1991, Theorem 6.13, Corollary 6.14, Corollary 6.15] and [Petersen, 2016, Theorem
5.7.1] for the second statement.
Proposition D.40. An isometry φ between complete Riemannian manifolds X, Y is determined by
the image of one point and the differential at that point.
Proof. Let p, q ∈ X with p fixed. Let φ, ψ be isometries with equal differential at p. Let γ be a geodesic
from p to q with initial velocity V . By (D.26) both φ ◦ γ and ψ ◦ γ are geodesics with initial vector
dφ|p V joining φ(p) with φ(q) resp. ψ(q). By uniqueness, they are equal. In particular their endpoints
are equal, and φ(q) = ψ(q).
Proposition D.41. Let M be a complete Riemannian manifold. TFAE:
1. All geodesics are minimizing.
2. All points are joined by a unique geodesic.
3. The exponential map expp : Tp M → M is a diffeomorphism for all p ∈ M .
Proof. See the Math Stackexchange post [Manifolds with geodesics which minimize length globally
2018].

97
D.5 Integration
Definition D.42. Let (M, g) be a Riemannian manifold with an orientation. Its associated √ volume
form is the unique volume form that equals 1 on positive orthonormal p bases. That is, ω = det gdx1 ∧
n 1 n 1 n
· · · dx in any chart (x , . . . , x ), or more generally ω = | det g|φ ∧ · · · φ for any local frame
(E1 , . . . , En ) with dual (φ1 , . . . , φn ). Integrating functions against this volume forms gives rise to the
Riemannian measure.
This makes M into a measure space, which gives us a notion of integration of measurable functions,
and not just of differential forms.
Example D.43. 1. For Rn , we have ω = dx1 ∧ · · · ∧ dxn , and we get the Lebesgue measure.
2. For SnR we have ω = iN (dx1 ∧ · · · ∧ dxn ) where N is the outward unit normal.
3. For M = Hn+1R with the Poincaré half-space model, we have ω = (R/y)n+1 dx1 ∧ · · · ∧ dxn ∧ dy,
with associated hyperbolic measure for R = 1.

D.6 The Laplace–Beltrami operator

The choice of a nondegenerate bilinear form g on a finite-dimensional real vector space V determines
a isomorphism with its dual, which sends v 7→ g(v, ·). When the bilinear form g is clear from the
context, we will denote the isomorphism by [ : v 7→ v [ , called flat.
P If (ei ) is
Pa basis of V with dual
basis (e∗i ), and g has matrix A = (aij ) in this basis, then [ sends i λi ei to i,j λi aij e∗j . The inverse
map is denoted ].
If g is a Riemannian metric on a manifold M , then to each vector field V on X we can associate its
flat V [ which is a 1-form. Indeed, the explicit formula above ensures that it is smooth. Likewise, a
1-form gives rise to a vector field by applying ].
Definition D.44. Let M be a Riemannian manifold with volume form ω, let f : M → R smooth and
X ∈ Γ(T M ) a vector field.
• The gradient of f is the vector field grad f = (df )] .
• The divergence of X is the unique smooth function div X for which the interior derivative
d(iX ω) = div X · ω.
• The Laplace-Beltrami operator (or simply Laplacian) of f is −∆ defined as:17

−∆f = − div(grad f )

∂


Proposition D.45. In a chart (x1 , . . . , xn ) such that g has matrix (gij ) in the basis ∂xi with inverse
(g ij ), we have
X 1 ∂ p 
ij ∂f
−∆f = − |g|g
|g| ∂xi ∂xj
p
i,j

where |g| is the determinant of the matrix (gij ).

P ∂f j P ij ∂f ∂
Proof. We have df = j ∂x j dx , hence grad f = i,j g ∂xi ∂xj . We have
!
^ X ^
i
iY dx = (−1)j+1 Yj dxk
i j k6=j

and thus:
  
p ∂ ∂ X ∂f ∂
d(igrad f (ω)) = d  |g| 1 ∧ · · · ∧ n−1 ∧  g ij i j 
∂x ∂x i,j
∂x ∂x

17 Also defined without the minus sign. This way we obtain a positive operator; see (G.17).

98
  
p X ∂f ^
= d  |g| (−1)j+1 g ij i dxk 
i,j
∂x
k6=j
 
X ∂ p X ij ∂f ^ k
=  |g| g  dx
∂xj ∂xi
i,j i,j k

Example D.46 (The Laplacian on hyperbolic spaces). Let n ∈ N>0 . Consider the hyperbolic space
Hn+1 as in (D.7)(c). We have from (D.45):
 2
∂2

∂

2

 −y + :n=1

 ∂x2 ∂y 2
−∆ = X ∂2
!
 2 ∂2 ∂
−y + 2 + (1 − n)y :n>1


i
∂(x )2 ∂y ∂y
i

D.7 Isometry groups

Theorem D.47 (Myers, Steenrod). The isometry group G of a Riemannian manifold M is a (real)
Lie group for the compact-open topology, and its action on M is smooth. If M is compact, then so is
G.
Proof. See their original paper [Myers and Steenrod, 1939, Theorem 10].
Note that a continuous group homomorphism between Lie groups is smooth, so that a Lie group
structure, if it exists, is unique.
One can show that the connected component of G still acts transitively. This is a result from topology,
which relies on the Baire category theorem:, if a group acts continuously and transitively on a locally
compact Hausdorff space X that is σ-compact (countable union of compact sets) then the action of G
on a fixed x ∈ X is open. So if X is connected, open subgroups act transitively.

D.8 Stabilizers
Proposition D.48. Let M be a Riemannian manifold with isometry group G, p ∈ M and K the
stabilizer of p.
1. The group homomorphism G → O(Tp M ) : σ 7→ (dσ)p is continuous.
2. K is closed in G.
3. If M is connected and complete, the homomorphism G → O(Tp M ) is injective.
Proof. 1. Let U be an open geodesic ball with center p which is a normal neighborhood. Then
the action of G on M restricts to U , and by functoriality of the compact-open topology, the
restriction map G → Isom(U ) is continuous. For σ ∈ K we have
expp ◦dσ = σ ◦ expp
and again by functoriality of the compact-open topology, Isom(U ) → Diff(exp−1 (U )) : σ 7→ dσ
is a homeomorphism on its image. Such a dσ extends uniquely to a linear map in O(Tp M ). It
remains to show that taking the unique extension defines a continuous map. Note first that the
norm topology on GL(Tp M ) is the same as the compact-open topology: both are the topology of
compact convergence. The restriction map O(Tp M ) → Diff(exp−1 (U )) is continuous by functo-
riality of the compact-open topology, and it is a homeomorphism on its image, because we can
recover the norm of a linear operator by its action on a neighborhood of 0. That is, the image
is a metric space and the restriction map is an isometry on its image. We conclude that the
composition
Isom(M ) → Isom(U ) → Diff(exp−1 (U )) → O(Tp M )
is continuous.

99
 2. Because G has the compact-open topology.
3. From (D.40) an isometry is determined by its differential at p.
Theorem D.49 (van Dantzig, van der Waerden). Let M be a locally compact connected metric space
with isometry group G, given the compact-open topology. Then G acts properly on M .
Proof. See [Dantzig and Waerden, 1928].
Proposition D.50 (Compactness of isotropy groups). Let M be a Riemannian manifold and p ∈ M .
1. The isotropy subgroup K of p is compact.

2. Suppose M is complete. Then the identification from (D.48) has closed image and gives an
isomorphism of Lie groups between K and a closed subgroup of O(Tp M ).
Proof. 1. Because the action G y M is proper by (D.49).
2. An bijective continuous group homomorphisms between Lie group is automatically an isomor-
phism. It suffices to show that the identification is continuous. Its image will be closed (hence a
Lie group) because K is compact. The continuity is precisely (D.48)(1).

100
E Symmetries of manifolds
E.1 Isotropic manifolds
Definition E.1 (Isotropic manifold). A Riemannian manifold is isotropic (at a point p) if the stabilizer
(isotropy subgroup) of every point (resp. the point p) acts transitively on unit tangent vectors, with
the action from (D.48).
Proposition E.2. Let M be a Riemannian manifold and p ∈ M with stabilizer K. TFAE:

1. M is isotropic at p.
2. There exist arbitrarily small δ > 0 such that K acts transitively on the geodesic sphere S(p, δ).
3. For all δ > 0 for which the geodesic ball B(p, δ) is a normal neighborhood, K acts transitively
on S(p, δ).

Proof. The exponential map at p commutes with the action of K: for g ∈ K:

exp(dgV ) = g exp(V )

and if one side is defined, so is the other.

E.2 Homogeneous spaces

Definition E.3 (Homogeneous Riemannian manifold). A Riemannian manifold is homogeneous if its
isometry group acts transitively on points.
Proposition E.4. A homogeneous Riemannian manifold is complete.

Proof. By (D.29), for a point p there exists δ > 0 such that all geodesics through p are defined at time
[−δ, δ]. By homogeneity and (D.26), we can take the same δ for all points.
More generally we define:
Definition E.5. Let G be a Lie group. A manifold M together with a smooth and transitive action of
G is a homogeneous (G-)space. A morphism of homogeneous G-spaces is a smooth map that respects
the G-action.
A morphism of G-spaces is necessarily surjective.
Theorem E.6 (Construction of homogeneous spaces). Let G be a Lie group and H a closed subgroup.
The left coset space G/H has a unique differentiable structure for which the projection is a smooth
submersion. Its dimension is dim G − dim H and it is a homogeneous space for the action g1 · (g2 H) =
(g1 g2 )H.
Proof. See [Lee, 2012, Theorem 21.17].
Theorem E.7. Every homogeneous G-space is of the above form, up to isomorphism. More precisly,
the stabilizer K of a point p is a closed subgroup of G and the bijection from the orbit-stabilizer
theorem provides the isomorphism.
Proof. See [Lee, 2012, Theorem 21.18].
Proposition E.8 (Local parametrization by a Lie group). Let M be a homogeneous G-space and
x0 , y0 ∈ M . Then there exists an open neighborhood U of y0 and a smooth embedding φ : U → G
such that φ(y)x0 = y for all y ∈ U .

101
Proof. We may suppose x0 = y0 : for arbitrary x0 it suffices to replace φ(y) by φ(y) · σ where σ ∈ G is
∼
such that σx0 = y0 . Let K be the stabilizer of y0 . By (E.7), we have an isomorphism ψ : G/K − → S as
G-spaces and the projection π : G  G/K is a submersion. By the local normal form for submersions,
there exists an open neighborhood V of π(e) = [e] ∈ G/K and a smooth local section τ : V → G of π
which is an embedding.
τ G
π

y
V G/K φ
S

Let U = φ(V ) and ψ = τ ◦ φ−1 . This does what we want: for y ∈ U we have by G-equivariance of φ
and φ([e]) = y0 that:
τ (φ−1 (y)) · y0 = φ(τ (φ−1 (y)) · [e])
= φ(π(τ (φ−1 (y))))
= φ(φ−1 (y))
=y

E.3 Symmetric spaces

For a point p on a Riemannian manifold, we can take a geodesic ball B inside a normal neighborhood
U (§D.4.1) which is stable by the geodesic inversion which sends exp(V ) to exp(−V ), equivalently,
γV (1) to γV (−1) for V ∈ exp(B), equivalently, it sends small geodesics through p to the geodesics at
the same speed in the other direction. It is smooth with differential − id at p. Nothing guarantees
that it is an isometry: normal coordinates tell little about the metric at points other than the center.
Definition E.9. A (Riemannian) locally symmetric space is a connected Riemannian manifold for
which the following equivalent conditions hold:
1. For every point, the geodesic inversion of any geodesic ball contained in a normal neighborhood
is an isometry.
2. For every point, the geodesic inversion of some small geodesic ball contained in a normal neigh-
borhood is an isometry.
3. For every point, there exists a local isometry defined on an open neighborhood that fixes the
point and has differential − id.
A symmetric space is one for which it extends to a global isometry.
Note that by (D.40), an isometry with differential − id at a fixpoint is uniquely determined by that
property, on a complete manifold. By (D.29) and Hopf–Rinow (D.39), every Riemannian manifold is
locally complete.
Proposition E.10. 1. A symmetric space is complete.
2. A symmetric or complete isotropic manifold is homogeneous.
Proof. 1. Take a point p and a geodesic γ defined on an interval [−s, s]. The geodesic reflection
around γ(s/2) defines, by uniqueness and by (D.26), an extension to the interval [−s, 2s].
2. Take two points p, q and a geodesic γ joining the two. It has finite length and we may consider
the geodesic inversion σ around its midpoint γ(t0 ), or in the isotropic case, any isometry fixing
γ(t0 ) whose differential sends γ̇(t0 ) to −γ̇(t0 ). Applying this to γ and following the curve in the
opposite direction, we obtain a geodesic with the same initial velocity as γ. Thus it coincides with
γ on the intersection of their domains. Comparing lengths (which are preserved by isometries)
we see that p and q are interchanged by σ.
Every symmetric space is homogeneous and thus of the form G/K where K is the isotropy subgroup
of a point, by (E.7).

102
F Differential operators
i
Notation F.1. Let M be a manifold.
Q Let (x ) be a chart defined on an open set U . For a multi-index
a ∈ Nn we have the operator Da = (∂/∂xi )ai which acts on C ∞ (U ).
Definition F.2 (Differential operator). A (smooth) differential operator D on a smooth manifold M
is a linear operator on C ∞ (M ) for which the following equivalent conditions hold:
1. If (xi ) is a local chart on an open set U :
• D followed by the restriction to U is a formal power series in the ∂/∂xi with coefficients in
C ∞ (U ) (applied to the restriction to U ).
• Locally around every point, only finitely many of the coefficients are nonzero functions (i.e.
attain nonzero values). Equivalently, on every relatively compact open subset W of U with
W ⊂ U , only finitely many coefficients are nonzero. That is, it is a finite linear combination
of the Da with coefficients in C ∞ .
2. D is a local operator: If φ and ψ coincide on an open set V , then so do Dφ and Dψ.

3. For all smooth φ we have supp(Dφ) ⊆ supp(φ).

4. If φ has compact support, then so does Dφ and supp(Dφ) ⊆ supp(φ).
Proof of equivalence. We clearly have 1 =⇒ 2 =⇒ 3 =⇒ 4. For 4 =⇒ 1, see [Helgason, 1984,
Theorem II.1.4].

Way may state some results only for differential operators on real-smooth functions, but it should be
noted that the analogous statements for complex differential operators hold as well.
They form an associative algebra under composition and pointwise addition with center R, denoted
E(M ).
Definition F.3. Let p ∈ M . Denote D|p : C ∞ (M ) → R : f 7→ D(f )(p) for D composed with
evaluation at p. The vector space of all these R-linear forms on germs of functions at p is denoted
E(M )p , the differential operators at p. See (F.12) for a motivation of the notation.
Proposition F.4 (Unique representation of differential operators). Let M be a smooth n-dimensional
manifold and U a coordinate neighborhood for a chart (xi ). let D ∈ E(M ) be a differential operator.
1. There exists a unique formal power series in n variables with coefficients in C ∞ (U ) with locally
only finitely many nonzero coefficients, that equals D when evaluated in the ∂/∂xi .
2. Let p ∈ U . Then there is a unique real polynomial P (ti ) in n variables such that D|p =
P (∂/∂xi )|p . In particular, by sending ti 7→ ∂/∂xi |p we get a linear isomorphism
∼
(F.5) Sym(Tp M ) −
→ E(M )p

Proof. 1. The existence is by definition of a differential operator. For uniqueness, we obtain the
coefficients of the power series by letting D act on monomials in the xi . 2. For the exact same
reason.
∼
Remark F.6. 1. The R-linear isomorphism Sym(Tp M ) −
→ E(M )p depends on the chart.
∼
2. E(M ) and Sym(Tp M ) are algebras, but the composition E(M )  E(M )p − → Sym(Tp M ) (defined
by a chart) is only a morphism of vector spaces. Had it been a morphism of algebras, then E(M )
would be commutative, which it clearly need not be.
Definition F.7 (Pushforward of a differential operator). Let σ : M → N be a diffeomorphism. For
D ∈ E(M ) define the pushforward by the following two equivalent definitions:

103
 1. For f ∈ C ∞ (N ), define σ∗ D ∈ E(M ) by σ∗ D(f ) = D(f ◦ σ) ◦ σ −1 .
2. For f ∈ C ∞ (N ) and p ∈ M , define (σ∗ D)|σ(p) (f ) = D|p (f ◦ σ).
Proof of equivalence. The two are equal because the second definition is obtained by evaluating the
first at σ(p).

Note that σ∗ D is smooth from the first definition, and it decreases supports, so it is indeed a differential
operator.
Proposition F.8 (Properties of the pushforward). Let σ : M → N be a diffeomorphism.

1. σ∗ sends C ∞ -functions to C ∞ functions and vector fields to vector fields: it is the usual push-
forward.
2. (σ ◦ τ )∗ = σ∗ ◦ τ∗ when τ : P → M is another diffeomorphism.
3. σ∗ is an isomorphism of algebras with inverse (σ −1 )∗ .

Thus the pushforward gives a (left) action of the diffeomorphism group Diff(M ) y E(M ). One also
denotes the pullback σ ∗ D := (σ −1 )∗ D by Dσ ; one then has the exponentiation rule (Dσ )τ = Dστ .
If φ = (φi ) is a chart, then φ∗ (∂/∂φi ) = ∂/∂xi , where (xi ) is the standard chart on open sets of Rn .

F.1 Grading
Lemma F.9. Let U ⊆ Rn open, f ∈ C ∞ (U ) smooth and Dα , Dβ monomials in the ∂/∂xi of degrees
a and b.
1. Dα f Dβ is a polynomial in the ∂/∂xi of degree 6 a + b with coefficients in C ∞ (U )
2. [Dα , f ] is a polynomial in the ∂/∂xi of degree < a with coefficients in C ∞ (U ).

3. More precisely, if Dα = j ∂/∂xij then

Q

X Y
[Dα , f ] = ∂f /∂xij ∂/∂xik + P (∂/∂xj )
j k6=j

where P has degree 6 n − 2.

Proof. The statements are trivial for a = 0. Let a > 0 and write Dα = ∂/∂xi Dγ for some i.
1. By induction on a. For a = 1 this is true by the product rule. In the induction step, we apply
the a = 1 case to all monomials appearing in Dγ f Dβ .

2, 3 This is the product rule applied to Dα (f g).

F.1.1 Global grading

Definition F.10 (Grading of differential operators). A differential operator D ∈ E(M ) is of degree
at most n ∈ N if the following equivalent conditions hold:
1. In every chart (U, (xi )), D is a polynomial of degree 6 n in the ∂/∂xi with coefficients in C ∞ (U ).

2. There exists a cover by charts (U, (xi )) in which D is a polynomial of degree 6 n in the ∂/∂xi
with coefficients in C ∞ (U ).
3. Define degree 0 operators to be C ∞ (M ) functions and define inductively D to have degree 6 n
if [D, f ] has degree < n for all f ∈ C ∞ (M ).
Note that degree 6 n − 1 according to the third definition implies degree 6 n by (F.9)(1).

104
Proof of equivalence. 1 =⇒ 2: Ok. 2 =⇒ 3: Degree 0 elements according to definition 2 are indeed
C ∞ -functions. The condition in 3. can be verified inductively and locally, and we conclude using
(F.9)(2). 3 =⇒ 1: A C ∞ function indeed has degree 0 according to definition 1. Suppose definition 3
has degree < n for all f ∈ C ∞ (M ), it is a polynomial
implies definition 1 up to degree n − 1. If [D, f ] P
of degree
Qm< n in any chart (U, (x )). Let D = α gα Dα in this chart. Let m = dim M and denote
i
α i αi
D = i=1 (∂/∂x ) . For the sake of contradiction, suppose gα 6= 0 for some α of degree N > n
maximal. Among those α, select those with α1 maximal, among those consider the ones with α2
maximal, etc. I.e. take α maximal for the lexicographic order. By (F.9).3, for every f ,
m m  αk −δkj
X ∂f Y ∂
[Dα , f ] = αj + P (∂/∂xi )
j=1
∂xj ∂xk
k=1

where P is a polynomial (depending on f ) of degree 6 N − 2. Take j maximal with αj > 0. Then

j  αk −δkj
∂f Y ∂
[D, f ] = gα αj + Q(∂/∂xi )
∂xj ∂xk
k=1

where Q is a linear combination of monomials Dβ with β < α for the lexicographic order. By assump-
tion, [D, f ] has degree 6 n − 1 < N − 1 so the first term should be zero. But if we take a point p ∈ U
for which gα (p) 6= 0 and choose f with ∂f /∂xj (p) 6= 0, we see that it is nonzero. Contradiction.
Notation F.11. We denote the C ∞ (M )-module of differential operators on U of degree at most n by
E6n (M ); their union as E<∞ (M ).
Remark F.12 (Differential operators and jet bundles). For any vector bundle E → M we can define
the kth jet bundle J k E which carries the information of partial derivatives op to order k of sections
of E. It comes with a natural map j k : Γ(E) → Γ(J k E). We are interested in the case E = M × R,
where Γ(E) = C ∞ (M ). For a differential operator D of degree 6 k on M there exists a unique
homomorphism of vector bundles D e : J k E → E such that D(f ) = D
e ◦ j k (f ). That is, E(M ) is linearly
isomorphic with the space of sections of the Hom-bundle Hom(J k E, E). Taking D|p corresponds to
taking the fiber of D
e (as an section of the bundle).

Proposition F.13 (The filtered algebra of differential operators). E<∞ is filtered by the E6n .
Proof. From (F.9)(1).
Proposition F.14 (Pushforward and filtration). If σ : M → N is a diffeomorphism between manifolds,
σ∗ : D<∞ (M ) → D<∞ (N ) is an isomorphism of filtered algebras.
The key is that the grading is defined in charts, and composing a chart of N with σ yields a chart of
M . That is, the hard work was to prove the equivalence of the definitions at (F.10).
Proof. Let (U, φ) be a chart of N , then (σ −1 (U ), φ ◦ σ) is a chart of M . Let D ∈ E6n (M ), then
D ∈ E6n (σ −1 (U )) and hence φ∗ σ∗ D = (φ ◦ σ)∗ D ∈ E6n (φ(U )) because φ ◦ σ is a chart. That is,
σ∗ D ∈ E6n (U ). Because U is arbitrary, σ∗ D ∈ E6n (N ). It remains to show that σ∗ does not decrease
degrees. This follows by symmetry, by applying what we just proved to σ∗−1 , which is the inverse of
σ∗ .

F.1.2 Grading at a point

Definition F.15 (Degree of a differential operator at a point). Let M be a manifold and p ∈ M .
Then P ∈ E(M )p has degree at most 6 n if the following equivalent conditions hold:
∼
1. For every chart around p, the image of P under the isomorphism E(M )p −
→ Sym(Tp M ) from
(F.4) has degree 6 n.
2. The above holds for one chart (U, (xi )) around p.

105
 3. It is the image of a differential operator of degree 6 n under the map E(M ) → E(M )p : D 7→ D|p .
Proof of equivalence. 1 =⇒ 2: Ok. 2 =⇒ 3: Because P (∂/∂xi ) extends to U , and using a bump
function smoothly to M . 3 =⇒ 1: Ok.

Proposition F.16 (Pushforward and filtration at a point). If σ : M → N is a diffeomorphism between

manifolds, p ∈ M with σ(p) = q and D ∈ E(M ), then D|p and (σ∗ D)|q have the same degree.
Proof. This is similar to the proof of (F.14). Let (U, φ) be a chart around q. Then (σ −1 (U ), φ ◦ σ) is a
chart of M around p. By assumption, D has degree deg D at p in this chart, i.e. (φ ◦ σ)∗ D has degree
deg D at φ(q). Thus σ∗ D has degree deg D at q.

Again, the hard work has been done in the equivalence of definition of the global degree of a differential
operator, where a crucial argument is the characterization of degrees with commutators. An alternative
proof, which does not rely on that characterization, uses the following calculation:
Lemma F.17. Let σ : U → V be a diffeomorphism between open sets of Rn , and p ∈ U with
∼ ∼
σ(p) = q. (F.4) gives R-linear isomorphisms E(M )p − → Sym(Tq N ). Let Dα
→ Sym(Tp M ) and E(N )q −
i
be a monomial of degree a in the ∂/∂x . Then Sym(dσ|p )D|p and (σ∗ D)|q differ by a polynomial of
degree 6 a − 1 in the ∂/∂xi |q .
Proof. By induction on a. For a 6 1 we have Sym(dσ|p )D|p = (σ∗ D)|q . Let a > 1 and write
Dα = Dγ ∂/∂xi . We have for f ∈ C ∞ (N ):
 
∂
(σ∗ D)|q (f ) = Dγ i (f ◦ σ)(p)
∂x
 
X ∂f ∂σ j
= D γ |p  j
◦σ· i
j
∂x ∂x
  X  ∂   ∂f 
X ∂f ∂σj
= D γ |p ◦ σ · (p) + P j |p ◦ σ
j
∂xj ∂xi j
∂xk ∂xj

with Pj of degree 6 a − 2, by (F.9)(2).

 
X
γ ∂f ∂σj
= Sym(dσ|p )D |p · (p)
j
∂xj ∂xi
X  ∂   ∂f  ∂σj
+ Q |q · (p)
j
∂xk ∂xj ∂xi
 
∂
+P |q
∂xk

with deg Q 6 a − 2 by the induction hypothesis, and deg P 6 a − 1.

Second proof of (F.16). Choose charts (xi ) and (y j ) about p and q. (F.4) gives R-linear isomorphisms
∼ ∼
E(M )p −→ Sym(Tp M ) and E(N )q − → Sym(Tq N ). The idea is that taking σ∗ is approximately the same
as applying Sym(dσ|p ), which is a graded isomorphism. WLOG assume a := deg(D|p ) 6 deg(σ∗ D|q ) =:
b. We have σ∗ D|q = Sym(dσ|p )(D|p ) + P (∂/∂y j ) with P of degree 6 a − 1 6 b − 1 by (F.17). Thus

a = deg(D|p ) = deg(Sym(dσ|p )(D|p )) = deg σ∗ D|q − P (∂/∂y j ) = b

Proposition F.18 (Kernel of a differential operator). Let D be a differential operator with real-
valued coefficients on an open interval I ⊆ R, of degree n > 0, and whose highest degree coefficient is
nonzero on I. Then any local solution to Df = 0 extends uniquely to a global solution on I, and the
kernel ker D has dimension exactly n. If D has complex-valued coefficients, the same holds, where the
complex dimension equals n.

106
Proof. We can divide by the highest degree coefficient and write the equation Df = 0 in the form
d n

d k


dt f = F dt f for some smooth F : Rn → R. By the uniqueness and existence theorem for
k<n
d k


linear ODE’s, for any t0 ∈ I and any initial values dt f (t0 ) for k ∈ {0, . . . , n − 1} there is a unique
local solution and it has a unique global extension. The claims follow.
For D complex, the same theorem says that the kernel has real dimension 2n. It has indeed complex
structure, and so it is of complex dimension n.

F.1.3 Symbols
Proposition F.19. Let M be a manifold with a global chart (xi ) and p ∈ M . The C ∞ (M )-linear
map
C ∞ (M ) ⊗R Sym(Tp M ) → E<∞ (M )
obtained by sending ∂/∂xi |p 7→ ∂/∂xi is an isomorphism of graded C ∞ (M )-modules, where the LHS
inherits the grading from Sym(Tp M ).
Proof. In general, if V and W are real vector spaces and W has basis (ei ), then every element of
V ⊗R W can be written uniquely as a finite sum of v ⊗ ei with v ∈ V − {0}. We conclude using
(F.4).
Definition F.20 (Symbols). Fix a point p ∈ M , a global chart φ = (xi ) and use it to identify
Sym(Tp M ) with R[ξ1 , . . . , ξn ] by sending ∂/∂xi |p 7→ ξi . The isomorphism of graded C ∞ (M )-modules
∼ ∼
→ C ∞ (M ) ⊗R Sym(Tp M ) −
E<∞ − → C ∞ (M ) ⊗R R[ξ1 , . . . , ξn ]

is called the (total) symbol map. If D ∈ E<∞ has degree d, the degree d component of the total symbol
is called the principal symbol σ(D).
The total symbol depends on the choice of the chart, but not on the point. The way in which it changes
when taking a different chart ψ = (y i ) is complicated and hard to understand. For the principal symbol
however, we know by (F.17) that it transforms simply by the action of Sym(d(ψ ◦ φ−1 )):

(F.21) σψ (D) = Sym(d(ψ ◦ φ−1 ))σφ (D)

F.2 Elliptic regularity

P Let M be a manifold with a global chart φ and D ∈ E<∞ of degree d, with principal
Definition F.22.
symbol σ(D) = α∈Nd aα (x)ξ α . We call D:
1. elliptic if σ(D)(x, ξ) is nonzero for all x ∈ φ(M ) and ξ 6= 0.
d
2. uniformly elliptic if σ(D)(x, ξ) > C kξk for some C > 0 independent of x.
Note that by (F.21), these conditions do not depend on the choice of a chart. Many theorems about
existence and regularity of solutions to differential equations, are known under the name “elliptic
regularity”. We mention a few results that are relevant to us.
Theorem F.23 (Elliptic regularity for degree 2 operators). Let Ω ⊆ Rn be open18 and D ∈ E<∞ (Ω)
of degree 2 uniformly elliptic and f ∈ C 2 (Ω) with Df = 0.
1. If D has C ∞ coefficients (which we have always assumed) then f ∈ C ∞ (Ω).
2. If D has real analytic coefficients, then f is real analytic.
Proof. See e.g. [Evans, 2010, §6.3.1, Theorem 3] resp. [Petrowsky, 1939].
Proposition F.24. Let M be a Riemannian manifold and −∆ its Laplacian. Then ∆ is locally
uniformly elliptic of degree 2.
18 Not necessarily bounded, contrary to what many texts assume.

107
Proof. By (D.45), we have, in a chart (xi ):
 
X 1 ∂ p ∂
∆= |g|g ij j
|g| ∂xi
p
i,j
∂x

whose principal symbol in that chart is X

g ij ξi ξj
i,j

That is, it is a bilinear form whose matrix is the inverse of that of the Riemannian metric. In particular,
it is positive definite.
Example F.25. For M = Hn+1 with the half-space model, we have from (D.46):
 2
∂2

∂

2

 −y + :n=1

 ∂x2 ∂y 2
−∆ = X ∂2
!
 2 ∂2 ∂
−y + 2 + (1 − n)y :n>1


i
∂(x )2 ∂y ∂y
i

Its coefficients are real analytic, hence elliptic regularity implies that C 2 Laplacian eigenfunctions are
real analytic.
Theorem F.26 (Elliptic regularity for weak solutions). [Agmon, 1965, Theorem 6.6] Let Ω ∈ Rn be
open and D ∈ E<∞ (Ω) be elliptic., f ∈ C ∞ (Ω) and u ∈ L2loc a locally square integrable weak solution
to Du = f , in the sense that Z Z
uDφ = fφ
Ω Ω
for all compactly supported test functions φ ∈ C0∞ (Ω). Then u ∈ C ∞ , that is has a smooth represen-
tative. More generally, this holds if u is a simultaneous solution of an overdetermined elliptic system
of differential operators Di , meaning that their principal symbols do nowhere simultaneously vanish
[Agmon, 1965, Definition 6.3].

F.3 Invariant differential operators

Definition F.27 (Invariant differential operator). Let G be a Lie group and M a homogeneous G-
space. An invariant differential operator on M is one for which, for all f ∈ C ∞ (M ) and g ∈ G:

D(f ◦ g) = (Df ) ◦ g

That is, for all x ∈ M :

D|x (f ◦ g) = D|gx f
or in terms of pullback: Dg = D.
This applies in particular to a homogeneous Riemannian manifold. A Lie group G acts on itself by
left-translations Lg : h 7→ gh but also by right translations Rg : h 7→ hg; this gives rise to the notions
of left- and right-invariance.
Notation F.28. The algebra of invariant differential operators will be denoted D(M ), the group G
being implicit.
Remark F.29. Take a Lie group G with identity e, fix a metric on Te G and transport it to obtain
a Riemannian metric that is invariant by G. Then G is contained in the isometry group, but usually
not equal, for example for G = Rn . We see that differential operators that are invariant by G, need
not be invariant by its isometry group.
Proposition F.30. Take a homogeneous G-space M and p ∈ M . A left-invariant extension D of a
differential operator D|p at a point, if it exists, is unique.

108
Proof. Because D|gp (f ) = D|p (f ◦ g) and G acts transitively.
Proposition F.31. For D ∈ D(M ) and n ∈ N, TFAE:
1. D ∈ E6n .
2. There exists a point p ∈ M with D|p ∈ Sym(Tp M )6n .
Proof. From (F.16).
Example F.32. Let M be a Riemannian manifold with isometry group G. The Laplacian is G-
invariant of degree 2.
Proof. It is a differential operator of degree at most 2 by the explicit formula from (D.45). It has
degree exactly 2 by (F.9)(2). The invariance follows from (F.33) below.
Proposition F.33. Let M, N be Riemannian manifolds and φ : M → N an isometry. Then for
f ∈ C ∞ (N ) and X ∈ Γ(M ):
1. dφ grad(f ◦ φ) = grad f .
2. div(dφX) ◦ φ = div X.
3. −∆(f ◦ φ) = −∆(f ) ◦ φ.
Proof. 1. For a vector field X ∈ Γ(M ) we have

hgrad(f ◦ φ), Xi = d(f ◦ φ)X

= df dφX
= hgrad f, dφXi
= (dφ)−1 grad f, X



2. Let α resp. β be the Riemannian volume form of M resp. N . We have

φ∗ (idφX β) = iX (φ∗ β)

and φ∗ β = α. The exterior derivative commutes with φ∗ , so

φ∗ (div(dφX)β) = div(X)α

and we conclude using φ∗ β = α once more.

3. From the first two formulas applied to φ and φ−1 :

div grad(f ◦ φ) = div (dφ)−1 grad f = (div grad f ) ◦ φ

F.4 Differential operators on Lie groups

Proposition F.34 (Invariant vector fields). Let G be a Lie group with identity e. For every tangent
vector X ∈ Te G = g, there is a unique left-invariant smooth extension to a vector field on G by defining
Xeg = dLg (Xe ).

Proof. Uniqueness follows from (F.30). For existence, we show that the construction in the statement
d eg (f ) := d f (g exp(tX)) |t=0
is indeed smooth. For X ∈ g we have X(f ) = dt f (exp(tX))|t=0 . Hence X dt
is smooth in g. It is a left-invariant extension by construction.
e of a tangent vector X ∈ g gives an R-linear map λ : g →
Taking the unique left-invariant extension X
D(G).
Lemma F.35. For:

109
 1. a real function f that is smooth in a neighborhood of t0 ∈ R:
d d dn
··· f (t1 + · · · + tn )|ti =t0 = n f (t)|t=nt0
dt1 dtn dt

2. a function f : Rn → R smooth in a neighborhood of t0 and x1 , . . . , xn ∈ R:

k
dk
X
∂
f (x1 t, . . . , xn t)|t=t0 = xi f (x1 t0 , . . . , xn t0 )
dtk ∂ti

Proof. 1. By induction: write the innermost derivative as

d
f (tn )|tn =t0 +t1 +···+tn−1
dtn
d
and apply the induction hypothesis to dt f (t0 + t).
2. Again by induction. The case n = 1 is the chain rule, and we apply the induction hypothesis to
d
dt f (x1 t, . . . , xn t).

Proposition F.36 (The symmetrization map). Let n = dim(G). There exists a unique extension of
λ to a linear map λ : Sym(g) → D(G) satisfying λ(X m ) = X e m for X ∈ g, called symmetrization.
Moreover, it is a filtered linear isomorphism that respectsP degrees, and we have the following explicit
formula: Let (Xi ) be a basis of g. Then φ : (ti ) 7→ exp ( i ti Xi ) is a local parametrization of G around
e. Denote ∂i = ∂/∂ti so that ∂i |e = Xi . Then for a real polynomial P in n variables and f ∈ C ∞ (G),

λ(P (Xi ))|g (f ) = P (∂i )f (g exp (t1 X1 + · · · + tn Xn )) |ti =0

Proof. [Helgason, 1984, Theorem II.4.3] It is a general fact that the symmetric algebra Sym(V ) of a
vector space over a field of characteristic 0 is linearly generated by the powers of elements of V . (More
precisely, the set of elements of degree n is linearly generated by nth powers.) Thus there is at most
one such linear map.
The definition of λ(P (Xi )) in the statement sends a smooth f to a smooth function. It is a differential
operator because it is linear and decreases supports. It is left-invariant, essentially by construction.
The map λ is linear. For each i we have λ(Xi )g f = ∂i |e f (g exp(ti Xi ) = X ei |g (f ) by definition of exp
and Xei . By linearity, λ(X) = X e for all X ∈ g, so we have an extension of the map λ we defined earlier.
The map λ is surjective because it gives a left-invariant extension of any Q ∈ Sym(g), which is unique.
For injectivity, take a monomial in the Xi of maximal degree whose coefficient at e is nonzero. Choose
f such that f (exp (t1 X1 + · · · + tn Xn )) equals that monomial in the ti in a neighborhood of e. Then
λ(P )|e (f ) 6= 0.
Consider now λ(X k ) for X ∈ g with X =
P i
x Xi . Using (F.35) we have:
X k  X 
λ(X k )|g (f ) = xi ∂i f g exp ti Xi |ti =0
dk  X 
= k f g exp txi Xi |t=0
dt
dk
= k f (g exp(tX)) |t=0
dt
d d
= ··· f (g exp (t1 X + · · · + tk X))
dt1 dtk
d d
= ··· f (g exp(t1 X) · · · exp(tk X))
dt1 dtk
=Xe k |g (f )

Regarding degrees, we have that λ(P )|e = P so P and λ(P ) have the same degree by (F.31).

110
Remark F.37. The symmetrization map need not be an algebra isomorphism.
Corollary F.38. The R-vector space D(G)6k has dimension k+n


n .

Proof. Because the symmetrization map is a filtered linear isomorphism that respects degrees.

It is natural to consider the induced algebra homomorphism T (g) → D(G) from the tensor algebra.
The kernel contains the elements of the form XY − Y X − [X, Y ]. Call I the two-sided ideal generated
by these elements. Then U (g) = T (g)/I is the universal enveloping algebra of g.
Proposition F.39. The filtered algebra homomorphism U (g) → D(G) is an isomorphism.

Proof. It is definitely surjective. More precisely, the image of T (g)6k in the quotient U (g) surjects
onto D(G)6k . Hence it suffices to find a generating set of T (g)6k /I with k+n


n elements, and it will
be automatically a basis. This is the easy part of Poincaré–Birkhoff–Witt! See below.
Theorem F.40 (Poincaré–Birkhoff–Witt). Let G be a real Lie group of dimension n with Lie algebra
g. Let (ei ) be a basis of g.
Q
1. The products eij of length m 6 k with i1 6 . . . 6 im generate T (g)6k /I linearly over R.

2. They form a basis of T (g)6k /I.

Proof. 1. We represent every element Y ∈ T (g)6k /I as a linear combination, by induction on the
degree of the highest occuring monomials in Y and the amount of them.
2. There are k+n of them. They are linearly independent because T (g)6k /I surjects onto D(G)6k ,


n
which has dimension n+k


n by (F.38).

111
G Spectral theory of the Laplacian
When studying differential operators and their spectral properties, the theory of bounded operators
is rarely applicable or satisfactory: Given a Riemannian manifold M and a differential operator D ∈
E(M ), one would like to study the action of D on C ∞ (M ), or at least on square-integrable functions:
C ∞ (M ) ∩ L2 (M ). This is no longer a Banach space, and there is no reason to assume that D applied
to a smooth square-integrable function yields again a square-integrable function. This leads to the
notion of unbounded operators. We will give special attention to the Laplacian.

G.1 Unbounded operators

Definition G.1 (Unbounded operator). An unbounded operator between Banach spaces X and Y is
a linear map A from a subspace (not necessarily closed) D ⊆ X to Y . We do not require A to be
continuous on D. We call D the domain of A.
In this subsection, “operator” always means “unbounded operator”. If D is dense in X, we say A is
densely defined.
Example G.2. The space C([0, 1], R) of continuous R-valued functions on [0, 1] is a Banach space for
the supremum norm k·k∞ . It contains the continuously differentiable functions as a subspace. (But
not a closed subspace: by stone Weierstrass already the polynomials are dense.) The differentiation
operator defined on this subspace is unbounded, because there exist uniformly bounded functions with
arbitrary steep slope at some point.
Definition G.3 (Extensions and closure). Let X be a Hilbert space. When A and B are operators
on X, B is defined on the domain of A and coincides with A on its domain, we call B an extension of
A, denoted A ⊆ B. We say A is closed when its graph

gra A ⊆ X × X

is closed. Otherwise, consider the closure gra A. When this is the graph of an operator B, we call A
closable and B its closure.
By the closed graph theorem (A.14), bounded operators are always closed. Contrary to what the
terminology suggests, the closure (when it exists) is in general not defined on the closure of the
original domain. Otherwise, together with the closed graph theorem that would imply that every
closable densely-defined operator is bounded.
Definition G.4 (Symmetric operator). Let X be a Hilbert space and A a densely defined operator
with domain D. We say A is symmetric if

hAx, yi = hx, Ayi ∀x, y ∈ D

Proposition G.5. A symmetric operator on a Hilbert space is closable, and its closure is symmetric.
Proof. See [Davies, 1995, Lemma 1.1.4].
Definition G.6. A densely defined operator A with domain D on a complex Hilbert space is positive
iff its associated quadratic form is nonnegative:

hAx, xi > 0 ∀x ∈ D

Proposition G.7. Let A be a densely defined symmetric positive operator on a complex Hilbert space
H. Then its closure is positive.
Proof. Let D be the domain of A, then by definition, the domain of its closure A is the set of f ∈ H
for which there exists g ∈ H and a sequence (fn ) ∈ D such that fn → f and Afn → g, in which case
Af = g. Then, by continuity of the inner product, we have

hAf, f i = lim hAfn , fn i ∈ R>0

n→∞

112
It is tempting to refer to symmetric operators as self-adjoint ones. Indeed, in the case of bounded
operators, these notions coincide:
Definition G.8 (Adjoint of an unbounded operator). Let X be a Hilbert space and A a densely
defined operator with domain D. Its adjoint is the operator A∗ with domain

D0 = {y ∈ X : ∃z ∈ X : ∀x ∈ X : hAx, yi = hx, zi}

and one defines A∗ y = z for such y. We say A is self-adjoint if D = D∗ and A = A∗ .

Note that A∗ is well-defined because A is assumed densely defined: if

hx, z1 − z2 i = 0 ∀x ∈ D

then z1 = z2 . For a symmetric operator, we always have A ⊆ A∗ . A self-adjoint operator is always

symmetric, but the converse is not true.
Proposition G.9. If A is closed and densely defined, then A∗ is also closed and densely defined.
Proof. See [Davies, 1995, Lemma 1.2.1].
Definition G.10. Let X be a complex Hilbert space and A a densely defined operator on X. We call
A essentially self-adjoint if it satisfies the following equivalent conditions:
1. A is symmetric and its closure is self-adjoint.
2. A has a unique self-adjoint extension.
Proof of equivalence. See [Davies, 1995, Theorem 1.2.7].
Proposition G.11. Let A : X → Y be an everywhere defined operator between Banach spaces. If it
has a bounded inverse, then kAxk  kxk uniformly in x.
Proof. We have, for x ∈ X,
kxk = A−1 Ax 6 A−1 kAxk

G.2 The Laplacian as a symmetric operator

Let M be a Riemannian manifold. We want to understand the spectrum of the Laplacian −∆. For
the moment, −∆ is defined on C ∞ (M ), and the subspace C0∞ (M ) of compactly supported smooth
functions is stable. We have an inclusion

C0∞ (M ) ⊆ L2 (M )

and one can wonder what the theory of extensions of symmetric unbounded operators tells us about
the action of −∆ on L2 (M ), or at least, L2 (M ) ∩ C ∞ (M ).
Note that the Laplacian of a smooth L2 function need not be L2 again, take for example x−1/3 on
]0, 1] with the standard Euclidean metric.
Definition G.12. Let M be an oriented Riemannian manifold with smooth boundary ∂M , given the
induced orientation. The outward unit normal is the unique section N : ∂M → T M for which if
p ∈ ∂M and (ei ) is a positive orthonormal basis for Tp ∂M then (N, (ei )) is a positive orthonormal
basis of Tp M .
The boundary ∂M is a Riemannian manifold for the induced metric; we denote its volume form by
dVe .
is iN (dV )|∂M , where dV is the volume form of M . (Since both agree on an orthonormal basis of
Tp ∂M .)
Proposition G.13. Let M be a Riemannian manifold with volume form dV and smooth boundary
j
∂M ,→ M . Let X be a vector field on M and u smooth on M . Then

113
 1. dVe = j ∗ (iN (dV )).

2. j ∗ (iX dV ) = hX, N i dVe .

3. div(uX) = u · div X + hgrad u, Xi.
Proof. 1. Since both agree on an orthonormal basis of Tp ∂M , for each p ∈ ∂M .
The same argument: take a local orthonormal frame (Ei ) of T ∂M and write X = hX, N i N +
2. P
Xi Ei . Evaluating both sides on the frame (Ei ) gives the same result.
3. We have div(uX)dV = d(iuX dV ) = d(u · iX dV ) = ud(iX dV ) + du ∧ iX dV . The first term
is u div XdV ; the second is iX (du) ∧ dV = X(u)dV where we use the fact that iX (α ∧ β) =
iX α ∧ β + (−1)|α| α ∧ iX β.

Theorem G.14 (Stokes). Let ω be a compactly supported (n − 1)-form on an oriented manifold M

of dimension n with (or without) smooth boundary. Then
Z Z
dω = ω
M ∂M

for the induced orientation on ∂M . In particular if M has no boundary, then the LHS is 0.
Proposition G.15 (Green’s identities). Let M be an oriented Riemannian manifold with measure
dV , smooth (possibly empty) boundary ∂M with measure dVe and u, v smooth functions, at least one
of which has compact support. Then
1. Z Z Z
u(−∆v)dV − hgrad u, grad vi dV = − uN vdVe
M M ∂M

2. Z Z
(u(−∆v) − (−∆u)v)dV = − (uN v − vN u)dVe
M ∂M

where N denotes the outward unit normal vector field.

Proof. 1. From Stokes and (G.13). 2. Immediate from 1.
Definition G.16. A harmonic function is one whose Laplacian is 0.
There are a few things that one can deduce from Green’s identities. We mention the ones that are
relevant to us:
Corollary G.17. Let M be an oriented Riemannian manifold without boundary.
1. 0 is an eigenvalue of −∆.
2. −∆ is symmetric for the L2 -inner product: if u, v ∈ C0∞ (M, C) have compact support, then
Z Z
u · −∆vdV = −∆u · vdV
M M

3. −∆ is a positive operator on compactly supported functions: if u ∈ C0∞ (M, C) has compact

support, then Z
u · −∆udV ∈ R>0
M
In particular, its eigenvalues are nonnegative.
Proof. 1. A nonzero constant function is harmonic.

114
 2. For real-valued u and v this follows directly from Green’s identities. For u = u1 + iu2 and
v = v1 + iv2 complex-valued, this follows by taking linear combinations, as in the lemma below.
3. For real-valued functions this follows by the first of Green’s identities. For complex-valued
functions we conclude again by taking linear combinations, or using the general lemma below.

Lemma G.18. Let X be a Hilbert space over R and A a densely defined symmetric unbounded
operator on X with domain D, that is,

hAu, vi = hu, Avi ∀u, v ∈ D

Let XC be its complexification and AC the induced operator on XC with domain DC , so that AC (u1 +
iu2 ) = Au1 + iAu2 . Then:
1. AC is symmetric.
2. If A is positive in the sense that hAu, ui ∈ R>0 for all u ∈ D, then so is AC .

Proof. 1. From the definition of AC and the properties of the complexified inner product.
2. Because A is assumed symmetric:

hA(u + iv), u + ivi = hAu, ui + hAv, vi + hAu, ivi + hiAv, ui

= hAu, ui + hAv, vi > 0

G.3 Extensions and essential self-adjointness

We have established that Laplacian is a positive symmetric unbounded operator on L2 (M ), whose
domain contains the compactly supported smooth functions. It is densely defined:
Proposition G.19. Let M be an orientable Riemannian manifold, and write it as the countable
increasing union of compact subsets Kn . Let f ∈ L2 (M ). Let δn : M → [0, 1] be compactly supported
smooth functions which are 1 on Kn . Then
Z
lim |f − f δn |2 = 0
n→∞ M

Proof. Follows immediately from the dominated convergence theorem.

Note that if M is complete, it is σ-compact: Fix x0 ∈ M , then the closed geodesic balls B(x0 , n) are
compact by Hopf–Rinow (D.39), and they cover M .
Theorem G.20. Let M be a complete Riemannian manifold. Then the Laplacian, as a densely
defined unbounded operator on L2 (M ) with domain C0∞ (M ), is essentially self-adjoint.

Proof. This was first shown by Gaffney in [Gaffney, 1951] and Roelcke in [Roelcke, 1960]. A different
proof is given by Strichartz in [Strichartz, 1983]. A discussion of essential self-adjointness of the
Laplacian can also be found in the blog post [Tao, 2011].
Assume that M is complete from now on, which is in particular the case if M is homogeneous (E.4),
for example when M = H is the hyperbolic plane, and we observed that this implies that finite-volume
quotients Γ\H are also complete. The Laplacian is then essentially self-adjoint, so it has a unique self-
adjoint extension, which is closed: it is the closure of the Laplacian on compactly supported smooth
functions. It is called the Dirichlet–Laplacian. We will keep calling it simply the Laplacian and denote
it by −∆.
Proposition G.21. The domain of −∆ contains

{f ∈ C ∞ (M ) : f ∈ L2 (M ), −∆f ∈ L2 (M )}

115
Proof. Follows from the discussion in [Strichartz, 1983].
One can show that the domain of −∆ is exactly equal to the Sobolev space H 2 , and that on smooth
L2 -functions, the closure of the Laplacian coincides with the usual Laplacian. Finally, we conclude,
using (G.7) on the closure of a positive operator:
Proposition G.22. The Laplacian is a positive densely-defined operator on L2 (M ). In particular,
its eigenvalues are nonnegative.

G.4 Operators with compact resolvent

We call an unbounded operator T : X → Y between normed spaces bounded below if there exists C > 0
for which
kT xk > C kxk ∀x ∈ X
By Hahn-Banach, this is equivalent to T having a bounded left inverse. In particular, a bounded below
operator is injective. A closed bounded below unbounded operator between Banach spaces has closed
range.
Definition G.23. Let H be a complex Hilbert space and T a densely defined operator on H with
domain D.
1. The resolvent set ρ(T ) is the set of all λ ∈ C for which λ − T
(a) is injective
(b) has dense range ran(λ − T )
(c) its inverse (λ − T )−1 : ran(λ − T ) → D is bounded.
2. The spectrum σ(T ) = C − ρ(T ) is its complement.
We say λ ∈ σ(T ) belongs to
• the point spectrum σp (T ) if λ − T is not injective, that is, if λ is an eigenvalue.
• the continuous spectrum σc (T ) if λ − T is injective with dense range but not surjective.
• the residual spectrum σr (T ) if λ − T is injective and its range is not dense.
For λ ∈ ρ(T ), we call Rλ = (λ − T )−1 the resolvent at λ. It commutes with T for all such λ.
Definition G.24 (Compact resolvent). A densely defined operator T on H with domain D has compact
resolvent if Rλ is a compact operator H → D for some λ ∈ ρ(T ).
Note that such T cannot be bounded on D, unless H is finite-dimensional: when T is bounded, the
compact bounded operator (λ − T )−1 : H → D ⊂ H has a bounded left inverse (an extension of λ − T
from D to H, given by Hahn-Banach). But then 1D is compact, which implies dim D = dim H < ∞.
When T has compact resolvent, its spectrum inherits many properties from compact operators:
Proposition G.25. Let T be densely defined and closed on H with compact resolvent. Then the
spectrum σ(T )
1. is discrete (hence countable)
2. has no accumulation point
3. consists of eigenvalues only, that is, equals the point spectrum σp .
4. The eigenspaces of T are finite-dimensional.
If T is in addition self-adjoint, then
5. H has an orthonormal basis of eigenvectors of T . In particular, H is separable.

116
Proof. [Taylor, 1997, Proposition 8.8]
1,2 Let ζ 6= 0 be such that ζ − T has a compact inverse Rζ . We know from (A.37) that σ(Rζ )
is countable and can only have 0 as an accumulation point. A computation shows that for
λ ∈ ρ(Rζ ) − {0} we have ζ − λ−1 ∈ ρ(T ) with

(G.26) (ζ − λ−1 − T )−1 = λ(λ − Rζ )−1 Rζ

The first two statements follow.

3 In fact, the inclusion
 
1
(G.27) {ζ} ∪ ζ − ⊆ ρ(T )
ρ(Rζ )

is an equality: when λ ∈ σ(Rζ ) is nonzero, then it is an eigenvalue of (ζ − T )−1 , hence ζ − λ−1

is an eigenvalue of T . In particular, ζ − λ−1 ∈ σp (T ). Note how from (G.26) and (G.27) it now
follows that all resolvents of T are compact.

4 The finite-dimensionality follows from the same argument: for µ 6= ζ, the µ-eigenspace of T is
the (ζ − µ)−1 -eigenspace of Rζ , which is finite-dimensional. For µ = ζ, it suffices to repeat the
argument with another ζ ∈ ρ(T ).
5 We know from the spectral theorem (A.49) that H has an orthonormal basis of eigenvectors of
Rζ , hence of T . Because the eigenspaces are finite-dimensional, H must be separable.
Conversely, if H has an orthonormal basis of eigenvectors for T , and the eigenvalues of T are nonzero
and tend to ∞ with multiplicities, then T has compact resolvent. Indeed, for λ not an eigenvalue, the
resolvent Rλ exists, is diagonalizable and its eigenvalues tend to 0 with multiplicities. By (A.42), it
follows that Rλ is compact.

G.5 The spectrum of the Laplacian

One can show that:
Theorem G.28. If M is a compact Riemannian manifold, then the resolvent of the (closure of the)
Laplacian is compact. In particular:
1. The spectrum of −∆ equals its point spectrum, which is a discrete closed infinite subset of R>0 .
2. Each eigenspace is finite-dimensional and the eigenspaces corresponding to distinct eigenvalues
are L2 -orthogonal.

Sketch of proof. There are two possible arguments. One is by explicitly writing the resolvent of the
Laplacian as an integral operator in terms of what is called a Green function. See e.g. [Chavel, 1984,
VI§1]. A different proof uses a general result in functional analysis to reduce the compactness of
the resolvent to the Rellich–Kondrachov embedding theorem, which says that the embedding of the
Sobolev space H 2 in L2 is compact. See e.g. [Taylor, 2018, Proposition 2.8], where it is also shown
that [1, ∞] ⊂ ρ(−∆).
Once we know that −∆ has compact resolvent, the remaining statements follow from (G.25) and the
positivity of the Laplacian.
While (G.28) is no longer true for noncompact manifolds, the same techniques generalize to certain
noncompact manifolds:
Theorem G.29. [Bump, 1996, Theorem 2.3.5] Let Γ be a lattice in PSL2 (R). Even when the quo-
tient Γ\H is not compact, −∆ has countably many eigenvalues, and they tend to infinity, counting
multiplicities. In particular, its eigenspaces are finite-dimensional.

117
H Whittaker functions
We study the differential equation
G00 + (λy −2 − 1)G = 0
when λ ∈ C, which occurred naturally in the Fourier expansion of periodic Laplacian eigenfunctions
on H (2.27). Substituting G(y) = W (2y) gives
 
00 λ 1
W (2y) + − W (2y) = 0
(2y)2 4

This equation in W was proposed by Whittaker in the form

1 2
4 −m

00 1
(H.1) W + − W =0
y2 4

Note that when W (y) is a solution on R − {0}, then so is W (−y).

19

is given by the Whittaker function W0,m (y), which is analytic
Proposition H.2. A solution of (H.1)
1
for y ∈ C − R60 and for <e m + 2 > 0 it equals

∞ m−1/2
e−y/2
Z 
t
W0,m (y) = tm−1/2 1 + e−t dt
Γ 12 + m


0 y
 
t
Where for log 1 + y we take the principal branch, with branch cut R60 .

Proof. See [Whittaker and Watson, 1915, Chapter XVI §16·12].

Thus the Witthaker function is analytic with the exception of a branch. We can turn its branch a
little bit counterclockwise, say to z0 R60 for some z0 of modulus 1 with 0 < arg z0 < π2 , by shifting
the contour to the half-line z0−1 · R>0 . (The exponential decay of the integrand ensures
 that we
 don’t
change the Witthaker function.) To do this, a branch must then be chosen for log 1 + tzy0 . The
principal branch of the logarithm makes the Witthaker function analytic with branch z0 R60 , and we
have only changed its values for −π 6 arg y 6 −π + arg z0 . In particular, and most importantly,
we haven’t changed the branch in the angular region − arg z0 6 arg y 6 0, where the shifting of the
contour happens.
Similarly, we can turn the branch clockwise.

H.1 Asymptotic expansion

1


Proposition H.3. The Whittaker function W0,m , <e m + 2 > 0 satisfies

W0,m (z) = e−z/2 1 + Oα (z −1 )

(H.4)

as |z| → ∞ in the angular region | arg z| 6 π − α < π.

Proof. See [Whittaker and Watson, 1915, Chapter XVI §16·3], where one can also find the full series
expansion for the O(z −1 ) term.
Similarly, one proves that this asymptotic relation still holds when we turn the branch of W0,m a little
bit clockwise or counterclockwise.
19 One of the parameters in Whittaker’s original equation is 0 in our case.

118
H.2 Second solution
When W0,m (y) is a solution to (H.1) and it is holomorphic on a neighborhood of the half-line R<0 ,
then so is W0,m (−y). Moreover, the asymptotic expansion (H.4) and the discussion about shifting the
branch cut shows that, for <e m + 12 > 0,

W0,m (−z) = ez/2 1 + Oα (z −1 )

(H.5)

for z in an angular region containing the positive real line.

From the asymptotic expansions we conclude in particular that W0,m (z) and W0,m (−z) are linearly
independent. We conclude:

Proposition H.6. The equation (H.1) has two linearly independent solutions on R>0 , W0,m (y) and
W0,m (−y), the first decays exponentially, the other grows exponentially.

119
References
Agmon, S. (1965). Elliptic boundary value problems.
Bade, W. G. (1954). “Weak and strong limits of spectral operators.” In: Pacific J. Math. 4.3, pp. 393–
413.
Beardon, A. F. (1983). The geometry of discrete groups. Springer.
Bernstein, J. (1984). Meromorphic Continuation of Eisenstein Series. Lecture notes.
Borel, A. (1997). Automorphic Forms on SL2 (R). Cambridge University Press.
Brumley, F. (2015). “Eisenstein series and functoriality”. Lecture notes.
Bump, D. (1996). Automorphic Forms and Representations. Cambridge Studies in Advanced Mathe-
matics 55. Cambridge University Press.
Charollois, P. (2017). “Introduction aux formes modulaires”. Lecture notes.
Chavel, I. (1984). Eigenvalues in Riemannian Geometry. Vol. 115. Pure and Applied Mathematics.
Clark, P. (2018). Lectures on Shimare curves 2: General theory of Fuchsian groups. url: http://
math.uga.edu/~pete/SC2-Fuchsian.pdf (visited on 09/06/2018).
Cohen, P. and Sarnak, P. (1980). Selberg trace formula, Chapter 6 and 7: Eisenstein Series for Hyper-
bolic Manifolds. url: http://web.math.princeton.edu/sarnak/ (visited on 07/15/2018).
Conway, J. B. (1990). A Course in Functional Analysis. 2nd ed. Graduate Texts in Mathematics 96.
Springer.
Dantzig, D. van and Waerden, B. L. van der (Dec. 1928). “Über metrisch homogene Räume”. In:
Abhandlungen aus dem Mathematischen Seminar der Universität Hamburg 6, pp. 374–376.
Davies, E. B. (1995). Spectral Theory and Differential Operators. Cambridge University Press.
Evans, L. C. (2010). Partial Differential Equations. 2nd ed. Vol. 19. Graduate Studies in Mathematics.
Ford, L. R. (1929). Automorphic functions. 1st ed. McGrawHill Book Company.
Gaffney, M. P. (1951). “The harmonic operator for exterior differential forms”. In: Proceedings of the
National Academy of Sciences of the United States of America 37, pp. 48–50.
Garrett, P. (2005). Holomorphic vector-valued functions. url: http://www- users.math.umn.edu/
~garrett/m/fun/Notes/09_vv_holo.pdf (visited on 08/09/2018).
– (2012a). Meromorphic continuation of Eisenstein series. url: https://mathoverflow.net/questions/
58386 (visited on 09/05/2018).
– (2001). Meromorphic continuation of Eisenstein series for SL(2). url: http://www-users.math.
umn.edu/~garrett/m/v/mero_contn_eis.pdf (visited on 08/23/2018).
– (2014a). Meromorphic continuation of higher-rank Eisenstein series. url: http : / / www - users .
math.umn.edu/~garrett/m/v/tel_aviv_talk.pdf (visited on 09/04/2018).
– (2012b). Most-continuous automorphic spectrum for GLn . url: http://www-users.math.umn.edu/
~garrett/m/v/gln_cont_spec.pdf (visited on 09/06/2018).
– (2011). The simplest Eisenstein series. url: http://www- users.math.umn.edu/ ~garrett/m/
mfms/notes_c/simplest_eis.pdf (visited on 09/06/2018).
– (2014b). Vector-valued integrals. url: http://www-users.math.umn.edu/~garrett/m/fun/notes_
2012-13/07e_vv_integrals.pdf (visited on 08/09/2018).
Helgason, S. (1984). Groups and Geometric Analysis: Integral Geometry, Invariant Differential Oper-
ators, and Spherical Functions. Pure and Applied Mathematics. Academic Press.
Iwaniec, H. (2002). Spectral Methods of Automorphic Forms. 2nd ed. Vol. 53. Graduate Studies in
Mathematics. American Mathematical Society.
Johnson, B. E. (Feb. 1969). “Separate Continuity and Measurability”. In: Proceedings of the American
Mathematical Society 20.2, pp. 420–422.
Kowalski, E. (2009). Spectral theory in Hilbert spaces. url: https : / / people . math . ethz . ch /
~kowalski/spectral-theory.pdf (visited on 05/14/2018).
Lee, J. M. (2012). Introduction to Smooth Manifolds. 2nd ed. Graduate Texts in Mathematics 218.
Springer.
– (1991). Riemannian Manifolds: An Introduction to Curvature. Graduate Texts in Mathematics 176.
Springer.
Manifolds with geodesics which minimize length globally (2018). url: https://math.stackexchange.
com/questions/1529968 (visited on 07/16/2018).

120
 √
Mathoverflow (2011). Smoothness of f ( x). url: https://mathoverflow.net/questions/77120
(visited on 06/30/2018).
Myers, S. B. and Steenrod, N. E. (Apr. 1939). “The Group of Isometries of a Riemannian Manifold”.
In: Annals of Mathematics 40.2, pp. 400–416.
Nagy, G. (2014). Vector-Valued Calculus Notions. url: https://www.math.ksu.edu/ ~nagy/func-
an-2007-2008/bs-6.pdf (visited on 08/09/2018).
Petersen, P. (2016). Riemannian Geometry. 3rd ed. Graduate Texts in Mathematics 171. Springer.
Petrowsky, I. G. (1939). “Sur l’analyticité des solutions des systèmes d’équations différentielles”. In:
Rec. Math. [Mat. Sbornik] N.S. 5 (1), pp. 3–70.
Roelcke, W. (1960). “Über den Laplace-Operator auf Riemannschen Mannigfaltigkeiten mit diskon-
tinuierlichen Gruppen”. In: Mathematische Nachrichten 21 (3–5), pp. 131–149.
Selberg, A. (1956). “Harmonic analysis and discontinuous groups in weakly symmetric Riemannian
spaces with applications to Dirichlet series”. In: Journal of the Indian Mathematical Society 20,
pp. 47–87.
Strichartz, R. S. (1983). “Analysis of the Laplacian on the complete Riemannian manifold”. In: Journal
of Functional Analysis 52, pp. 48–79.
Takeuchi, K. (1975). “A characterization of arithmetic Fuchsian groups”. In: J. Math. Soc. Japan 27.4.
Tao, T. (Dec. 2011). The spectral theorem and its converses for unbounded symmetric operators. url:
https://terrytao.wordpress.com/2011/12/20/the-spectral-theorem-and-its-converses-
for-unbounded-symmetric-operators/ (visited on 07/15/2018).
Taylor, M. E. (1997). Partial Differential Equations I: Basic Theory. Vol. 115. Applied Mathematical
Sciences.
– (2018). Spectral Theory. url: http://www.unc.edu/math/Faculty/met/chap8.pdf (visited on
08/18/2018).
Vernaeve, H. (2015). “Banachruimten en Banachalgebra’s”. Lecture notes.
Whittaker, E. T. and Watson, G. N. (1915). A Course of Modern Analysis. 2nd ed. Cambridge Uni-
versity Press.
Yosida, K. (1980). Functional Analysis. 6th ed. Springer-Verlag.

121
122