Commun. Math. Phys.

317, 827–869 (2013)

Communications in
Digital Object Identifier (DOI) 10.1007/s00220-012-1557-1
Mathematical
Physics

An Algorithm for the Computation of Eigenvalues,

Spectral Zeta Functions and Zeta-Determinants
on Hyperbolic Surfaces
Alexander Strohmaier, Ville Uski
Department of Mathematical Sciences, Loughborough University, Loughborough,
Leicestershire LE11 3TU, UK. E-mail: [email protected]; [email protected]

Received: 24 November 2011 / Accepted: 7 March 2012

Published online: 30 August 2012 – © Springer-Verlag 2012

Abstract: We present a rigorous scheme that makes it possible to compute eigenvalues

of the Laplace operator on hyperbolic surfaces within a given precision. The method
is based on an adaptation of the method of particular solutions to the case of locally
symmetric spaces and on explicit estimates for the approximation of eigenfunctions on
hyperbolic surfaces by certain basis functions. It can be applied to check whether or not
there is an eigenvalue in an -neighborhood of a given number λ > 0. This makes it
possible to find all the eigenvalues in a specified interval, up to a given precision with
rigorous error estimates. The method converges exponentially fast with the number of
basis functions used. Combining the knowledge of the eigenvalues with the Selberg
trace formula we are able to compute values and derivatives of the spectral zeta function
again with error bounds. As an example we calculate the spectral determinant and the
Casimir energy of the Bolza surface and other surfaces.

1. Introduction
Let (M, g) be a compact Riemannian manifold of dimension n and let  be the (positive)
Laplace operator on functions, which in local coordinates is given by

n
1 ∂  ∂
=− √ |g|g ik .
i,k=1
|g| ∂ xi ∂ xk

Here |g| denotes the determinant of the metric tensor and g ik are the components of the
dual metric on the cotangent bundle. This operator is formally self-adjoint on the space
of smooth functions with inner product
 
 f1 , f2  = f 1 (x) f 2 (x) |g|d x1 , · · · d xn
M
 This work was supported by the Leverhulm grant F/00 261/Z.
828 A. Strohmaier, V. Uski

√
defined by the Riemannian measure |g|d x1 · · · d xn . Then  extends to a self-adjoint
operator L 2 (M) ⊃ H 2 (M) → L 2 (M) with compact resolvent. This means there exists
an orthonormal basis {φ j | j ∈ N0 } in L 2 (M) consisting of eigenfunctions
φ j = λ j φ j ,
which we assume to be ordered such that 0 = λ0 ≤ λ1 ≤ λ2 ≤ · · ·.
It is a classical problem to compute the eigenvalues and the eigenfunctions of the
Laplace operator on a given manifold. Its solution allows complete control over the
functional calculus and thus over solutions to the heat equation, the wave equation, or
the Schrödinger equation.
In this paper we show how an adaptation of the method of particular solutions can
be used to compute eigenvalues on manifolds with high accuracy, beyond what can
currently be achieved by finite element methods or boundary element methods. We will
focus primarily on the case of two dimensional oriented surfaces of constant curvature.
These are the simplest examples of manifolds with non-trivial spectral geometry. They
are topologically classified by their genus. By the Gauss-Bonnet theorem the curvature
κ of a surface of genus g can always be normalized to be κ = sign(g − 1) by multiplying
the metric by a positive constant. This divides oriented surfaces of constant curvature
into three categories:
• κ = +1, g = 0: up to isometries there is only one such manifold: the round sphere S 2 .
The spectrum of the Laplace operator on the round sphere is of course well known
and the spherical harmonics provide a basis consisting of eigenfunctions.
• κ = 0, g = 1: such manifolds are isometric to flat tori which are obtained as quo-
tients of R2 by co-compact lattices. The moduli space of equivalence classes of
flat metrics on a given topological torus is the modular surface SL(2, Z)\H, where
H = {x + iy ∈ C | y > 0} is the upper half space. For a surface R2 /L a basis
consisting of eigenfunctions is obtained by taking the Fourier modes associated with
points in the dual lattice L ∗ and the corresponding eigenvalues are the squares of the
norms of these points.
• κ = −1, g > 1: these are the so called hyperbolic surfaces. They can be obtained
either as quotients \H by co-compact hyperbolic lattices in SL(2, R) or by glue-
ing hyperbolic pairs of pants. The moduli space of hyperbolic metrics on a given
topological surface is a quotient of the Teichmüller space Tg by a discrete group, the
mapping class group. The Teichmüller space for a genus g surface has dimension
6g − 6 so that a constant curvature metric on a topological surface may be given by
specifying 6g − 6 parameters. Unlike in the previous two cases the spectrum of the
Laplace operator of a hyperbolic surface can not currently be explicitly computed.
The main purpose of this article is to establish an algorithm to compute the eigen-
values and eigenfunctions of the Laplace operator on a hyperbolic surface up to a certain
precision and with mathematically rigorous error bounds. Once this is achieved for a
number of eigenvalues it is then possible to compute values of the spectral zeta function
and the spectral determinant of the Laplace operator on such surfaces. We demonstrate
how this can be done using the Selberg trace formula in such a way that the error can
again be bounded rigorously.
The method we use is adapted to a pants decomposition of the surface. For a particular
basis of functions we establish various bounds and estimates that prove that the method
of particular solutions can be applied to an interval and yields all eigenvalues in that
interval. In particular we prove a novel estimate that allows us to show non-existence of
eigenvalues in a certain interval.
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 829

1.1. Organization of the article and results. In Sect. 2 we describe how an adaptation of
the method of particular solutions can be applied to manifolds. The method uses a finite
dimensional space of test functions that solve the eigenvalue equation with eigenvalue
λ on open submanifolds that are glued together along co-dimension one hypersurfaces.
The glueing conditions for the eigenfunctions, continuity of the eigenfunctions and their
normal derivatives, is measured in terms of the differences of the function values and
normal derivatives along the hypersurfaces. For a test function φ we introduce in Eq. (1)
the number F−1/2,−3/2 (φ) that measures these differences in suitable Sobolev norms.
We prove that if φ L 2 (M) = 1 and F−1/2,−3/2 (φ) is small then λ must be close to an
eigenvalue. Our Theorem 2.1 together with the subsequent estimates on the constants
gives a quantitative version of this statement for hyperbolic surfaces that are glued along
geodesic segments.
In Sect. 3 we review the construction of hyperbolic surfaces from pairs of pants.
This gives an explicit parametrization of Teichmüller space in terms of Fenchel-Nielsen
coordinates and shows that hyperbolic surfaces can be cut open along geodesic segments
into pairs of pants and subsequently into subsets of hyperbolic cylinders.
In Sect. 4 we describe a set of basis functions for a surface of genus g that is decom-
posed into pairs of pants. We prove that true eigenfunctions can be approximated by
linear combinations of this set of basis functions with explicit bounds on the error. Sec-
tion 4.3 deals with the construction of m × k matrices B0λ , Bλ and Cλ with the following
properties:
(1) The distance of λ to the spectrum can be bounded from above in terms of the first
singular value σ1 (B0λ ) of B0λ and its singular vector (Theorem 4.5).
(2) The smallest relative singular value σ1 (Bλ , Cλ ) is bounded from above by

c1 (k, λ) + c2 (λ)dist(spec(), λ),

where c1 (k, λ) and c2 (λ) are explicitly computable constants and c1 is exponen-
tially decaying in k (Theorem 4.4).
Whereas the first property allows to prove that an eigenvalue is in a certain interval,
the second property allows to determine intervals in which there are no eigenvalues.
Both estimates together can be used to find all eigenvalues in a specified interval. We
demonstrate that these matrices can be computed within a given precision and show that
the singular values can be bounded from above and below using interval arithmetics.
As a proof of concept we implemented our method in Fortran and in Mathematica. This
resulted in programs that allow to compute eigenvalues rather accurately for a surface
of genus g with given Fenchel-Nielsen coordinates. A Mathematica program was used
to compute the first eigenvalues of the Bolza surface with extremely high accuracy.
The spectral zeta function ζ (s) is defined as the meromorphic continuation of the
function
∞

ζ (s) = λi−s ,
i=1

where (λi )i∈N are the non-zero eigenvalues of the Laplace operator repeated according
to their multiplicities. The (zeta-regularized) spectral determinant detζ () is defined by
log det ζ () = −ζ (0) (zero is not a pole of the meromorphic continuation). Since the
meromorphic continuation is contructed from the full spectrum it is a priori not enough
to know a finite part of the spectrum to compute the spectral determinant up to a certain
830 A. Strohmaier, V. Uski

accuracy. However, we show in Sect. 6 that the Selberg trace formula may be used in
conjunction with the list of eigenvalues up to a certain threshold to calculate values of
the spectral zeta function, in particular the Casimir energy and the spectral determinant,
up to a certain precision depending on that threshold. The spectral determinant for the
Laplace operator on hyperbolic surfaces is of particular importance. In the seminal paper
[OPS88] Osgood, Phillips and Sarnak showed that the spectral determinant is maximized
at the hyperbolic metric in each conformal class of a given volume. Thus, understand-
ing the extremal properties of the determinant as a function on the space of metrics of
fixed volume in dimension two is equivalent to the understanding of the spectral deter-
minant as a function on the Teichmüller space of hyperbolic metrics. Of course values
of the determinant for non-hyperbolic surfaces can be computed from the value for the
corresponding uniformized hyperbolic surface using the Polyakov formula.
Finally, Sect. 7 contains a collection of examples of interesting surfaces of genus two
and three for which we computed the spectral determinant and the value of the spectral
zeta function at the point −1/2. The set of examples contains the isolated surfaces of
genus two with large symmetry group which are well known to be critical points for any
value of the spectral zeta function or the spectral determinant (see e.g. [KKK09]). We
conjecture that the value of the spectral determinant is maximized at the Bolza surface
and we compute the value of the spectral determinant rather accurately. Such a conjec-
ture would imply that the global maximum of the spectral determinant as a function on
the space of all metrics of fixed volume is attained at this point. The appendix contains
an explicit estimate for various resolvents as well as explicit estimates of the L ∞ norm
of eigenfunctions and derivatives of eigenfunctions. These estimates are needed to make
the constants in our estimates explicit but they may also be interesting in their own right.

1.2. Discussion. Eigenvalues of hyperbolic surfaces of genus two have been calculated
in the physics literature by Aurich and Steiner in [AS89] using the finite element method,
and in [AS93] using the boundary element method. In both cases the authors relied on the
realization of the surface by geodesic octagons. It is quite interesting that the Hadamard-
Gutzwiller model discussed in [AS89] is actually the same as the Bolza surface. Using
the group action and the decomposition into geodesics triangles Ninnemann [Nin95]
computed the eigenvalues of the regular geodesic octagon with every second side iden-
tified. The resulting surface, which he also refers to as the Gutzwiller octagon, does
however not coincide with the Bolza surface. Its Fenchel Nielsen coordinates are given
in Sect. 7.4. There, as well as in other parts of the physics literature, the Bolza surface
is referred to as the regular octagon.
The method of particular solutions is based on an article by Fox, Henrici and Moler
[FHM67]. It was subsequently further developed and revived by Betcke and Trefethen
[BT05] to achieve high accuracy in eigenvalue computations on domains with Dirichlet
boundary conditions. Further versions, including domain decompositions and match-
ing of boundary data including the normal derivative were described in [BB10] and
[Bet07]. Explicit error bounds for the eigenvalues in the method of particular solutions
for domains in Rn were given in [MP68] and further improved on by Alex Barnett
[Bar09]. We would also like to refer to the latter article for further background and
literature on the method of particular solutions. Our estimate for the error bound does
not use [MP68] but is instead based on the more refined information contained in the
resolvent of the Laplace operator on the manifold. Thus, for generic eigenfunctions it
is expected to give an improvement of the order of the square root of the eigenvalues.
Estimates that provide such an improvement without this genericity assumption have
recently been obtained for domains in Rn in [Bar09] and [BH11].
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 831

Finally, for hyperbolic surfaces with cusps Hejhal ([Hej92] and [Hej99]) introduced
a method to compute Maass cusp forms which is based on the group action and the
expansion of the embedded eigenvalues on the cusp. After most of the work in this
paper was completed we learned that Booker, Strömbergsson and Venkatesh [BSV06]
have recently used the pre-trace formula together with the Taylor expansion of boundary
data of quasi-modes to rigorously verify embedded eigenvalues for the modular surface
S L(2, R)\H. Their way of certifying eigenvalues is in spirit similar to the method we
use to prove that our computed eigenvalues are accurate within the error bounds. Apart
from the fact that their method applies to a different geometric situation (surfaces with at
least one cusp) the method of bounding the error is different when it comes to technical
detail: for example in [BSV06] it is estimated in terms of the L ∞ -norm of the boundary
data instead of the L 2 -norm. It would be very interesting in the future to combine these
ideas. In particular a modification of our estimates applies to hyperbolic surfaces with
cusps. The problem of adjusting the step size in the search for Maass cusp forms on
S L(2, Z)\H in such a way that no eigenvalues are missed could be tackled in this way.
Formulae for the spectral determinant based on the length spectrum were given in the
mathematics literature by Fried [Fri86] and by Pollicott and Rocha [PR97], and in the
physics literature by Aurich and Steiner [St87,AS92]. In [PR97], exponential conver-
gence was proved and a numerical algorithm was established for the case of surfaces of
genus 2 in mw-Fenchel Nielsen coordinates without twisting. We are not able to confirm
the numerical values obtained in [PR97] in the three examples there but obtain quite dif-
ferent values. We believe that the part of the length spectrum computed in [PR97] was
not sufficient to obtain the correct values. In contrast to the other methods employed,
our approach to compute the spectral determinant and the values for the zeta function
allows for explicit error estimates even if the length spectrum is unknown.

2. The Method of Particular Solutions on Manifolds

The method of particular solutions is a method to approximate eigenvalues and eigen-

functions of a differential operator. We will consider here the case of the Laplace operator
 : C ∞ (M) → C ∞ (M) acting on functions on a compact oriented Riemannian mani-
fold M.
To fix notations suppose that ⊂ M is a closed subset which is the finite union
of oriented compact codimension one submanifolds α , possibly with boundary. We
will assume here for simplicity that the α intersect at most at boundary points, i.e.
α ∩ β ⊂ ∂ α ∩ ∂ β . Removing from M results in an open manifold M\ that is
the interior of a manifold with piecewise smooth boundary. This manifold might how-
ever have corners or other singularities. We would like to define the class of functions
that are "smooth up to the boundary" on M\ . In order to do this let us be more precise
about how the manifold with piecewise smooth boundary is constructed from M and .
The space M\ equipped with the geodesic distance is not complete and we denote by
M\ its abstract metric completion. Since M is complete there is a natural continuous
surjection π : M\ → M. We define a function f ∈ C(M\ ) to be smooth if and only
if for every point x ∈ M\ there exists an open neighborhood U such that there exists
a smooth function g on M with f |U = π ∗ (g)|U . The corresponding sheaf of smooth
functions endows M\ with the structure of a smooth manifold with piecewise smooth
boundary. We write C ∞ (M\ ) for the space of smooth functions on M\ .
Of course an element φ ∈ C ∞ (M\ ) can be understood as a function in L ∞ (M) as
the boundary of M\ has zero measure and the interior of M\ can be identified with
832 A. Strohmaier, V. Uski

M\ . We will now make this identification without further mention. Since each α is
oriented there is a natural unit normal vector field n α to α and a closed tubular neigh-
borhood of α diffeomorphic to α × [−, ]. A function φ ∈ C ∞ (M\ ) therefore has
two boundary values on α , a right boundary value φα+ and a left boundary value φα− .
Both are smooth functions on α . Of course the function φ ∈ C ∞ (M\ ) is continuous
on M if and only if for all indices α,
∀x ∈ α : φα+ (x) = φα− (x).
Similarly, φ ∈ C 1 (M) if and only if for all α
∀x ∈ α : φα+ (x) = φα− (x), and
∀x ∈ α : (n α φ)+ (x) = (n α φ)− (x).
Here, (n α φ)± denotes the right and left limits of the normal derivative of φ.
Let us define Dα φ to be the function on each manifold α given by φα+ − φα− and let
Dn α φ be the function given by (n α φ)+ − (n α φ)− . Let Dφ and Dn φ be the corresponding
functions in L ∞ ( ). As the boundary of has zero measure in these functions are
well defined and they are smooth on each α .
Let s, t ≤ 0. The functional Fs,t (φ) defined by

Fs,t (φ) := ( Dφ 2
Hs( ) + Dn φ H t ( ))
2 1/2
, (1)
measures the continuity of the function and its normal derivatives in different Sobolev
norms. Here the H s norm on is defined as

f 2H s ( ) = f α 2H s ( α ) .
α

Let now as above φ ∈ C ∞ (M\ ) and let χ ∈ C ∞ (M\ ) be defined by χ (x) =

( − λ)φ(x) for x ∈ M\ . Suppose that f ∈ C ∞ (M) is any test function. Then, by
Green’s formula (see e.g. Prop. 4.1 in [Tay96]),

φ(( − λ) f )(x)dμ(x)
M
 
= (( − λ)φ)(x) f (x)dμ(x) + (nφ(x) f (x) − φ(x)n f (x)) dν(x)
M ∂ M\
  
= χ (x) f (x)dμ(x) + (Dn φ)(x) f (x)dν (x) − (Dφ)(x)(n f )(x)dν (x),
M
√
where n denotes the unit normal vector field along , dμ(x) = gd x is the measure
induced by the Riemannian metric on M, and dν(x), dν (x) are the measures on ∂ M\
and induced by the Riemannian metrics there.
If we think of φ as a distribution on M, then the above simply means that we have in
the distributional sense
( − λ)φ = χ + Dφ δ + Dn φ δ ,
where δ is the Dirac delta distribution on and δ = nδ is its normal derivative so
that for a test function f ∈ C ∞ (M) the above distributions are given by

(Dn φ δ ) ( f ) = (Dn φ)(x) f (x)dν (x)
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 833

and

 
Dφ(δ ) ( f ) = − (Dφ)(x)(n f )(x)dν (x).

Note that the distribution

h = Dφ δ + Dn φ δ

is in H −2 (M). Denote by F̃ (φ) its H −2 -norm, where the H s -norm is defined by

ψ H s (M) = (1 + )s/2 ψ 2 .
By the trace theorem for Sobolev spaces (see e.g. Prop. 4.5 in [Tay96]) the maps
f → f | and f → (n f )| extend to bounded maps from H 2 (M) → H 3/2 ( ) and
H 2 (M) → H 1/2 ( ) respectively. Hence, by duality,

Dφ δ H −2 (M) ≤ C̃ Dφ 1
H− 2 ( )

and
Dn φ δ H −2 (M) ≤ C̃ Dn φ 3 ,
H− 2 ( )

for some constants C̃, C̃ > 0. Thus, there exists a constant C > 0 depending only on
the geometry of M and such that

F̃ (φ) ≤ C · F− 1 ,− 3 (φ). (2)

2 2

A constant C in this estimate can be explicitly found for a given manifold, for example
by using gluing functions and the Fourier transform. We will illustrate in Sect. 4 how to
get a bound on C for a given hyperbolic surface when consists of geodesic segments
(cf. (4) and (5)).
The method of particular solutions relies on the following simple observation.
Theorem 2.1. Suppose that φ ∈ C ∞ (M\ ) ⊂ L 2 (M) with φ 2 = 1 such that
∀x ∈ M\ : ( − λ)φ(x) = χ (x),
χ ∈ L 2 (M), 0 ≤ χ ≤ η.

If F̃ (φ) <  < 1, then there is an eigenvalue of  in the interval

(1 + λ) + η (1 + λ) + η
[λ − ,λ + ].
1− 1−
Proof. As before let
h = Dφ δ + Dn φ δ .

Then g = ( + 1)−1 h ∈ L 2 (M) and by our assumptions

g 2 ≤ .
Now note that
( − λ)(φ − g) = χ + (1 + λ)g.
834 A. Strohmaier, V. Uski

Of course
χ + (1 + λ)g 2 ≤ η + (1 + λ) g 2

and
φ−g 2 ≥1− g 2

as φ 2 = 1. From this it immediately follows that in case ( − λ)−1 exists its operator
norm is bounded from below by
1− g 2
.
η + (1 + λ) g 2

This implies the theorem since the operator norm of the resolvent at the point λ is
bounded from above by the inverse of the distance of λ to the spectrum. 

Remark 2.2. The statement of the previous theorem also applies to situations where
eigenvalues might be close to each other or have high multiplicities. Not to overload
notation we state this here only as a remark. If there is an orthonormal set (φi )i=1...,k in
L 2 (M) such that for each of the φ the conditions of Theorem 2.1 hold, then the same
proof shows that there are at least k eigenvalues (counting multiplicities) in a small
interval around λ.
To apply the method of particular solutions one constructs normalized functions φ
which are eigenfunctions of the Laplace operator on each of the Mi with eigenvalue λ
and for which F 1 3 (φ) is small. The above theorems then show that λ must be close
− 2 ,− 2
to an eigenvalue and φ is close to an eigenfunction. We will show in the following how
this can be done for oriented hyperbolic surfaces of genus g.

3. Hyperbolic Surfaces
A hyperbolic surface is a 2-dimensional orientable Riemannian manifold endowed with
a metric of constant negative curvature equal to −1. Such surfaces can be obtained by
factorizing the hyperbolic plane H = {x + iy ∈ C | y > 0} with metric y −2 (d x 2 + dy 2 )
by a discrete co-compact hyperbolic subgroup of S L(2, R). The isometric action of
S L(2, R) is given by fractional linear transformations
 
a b az + b
z= .
c d cz + d
Another way to obtain a surface of genus g with hyperbolic metric is to glue 2g − 2
hyperbolic pair of pants. A hyperbolic pair of pants is a genus 0 surface with bound-
˙ 1 ∪S
ary S 1 ∪S ˙ 1 equipped with a metric of constant negative curvature −1 such that the
boundary curves are geodesics (see Fig. 1). We will also refer to such a surface as a
Y -piece.
Such hyperbolic pairs of pants are, up to isometry, uniquely determined by the length
of their boundary geodesics (1 , 2 , 3 ). Any hyperbolic surface can be decomposed into
pairs of pants by cutting along 3g − 3 non-intersecting simple geodesics on the surface
(see Fig. 2). This results in 2g − 2 pairs of pants.
A pair of pants can be glued from a right angled geodesic octagon in the upper half
space (see Fig. 3) by identifying the sides b and h, as well as e and g respectively.
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 835

Fig. 1. Y -piece with boundary geodesics

Fig. 2. Pants decomposition of a surface of genus three into four Y -pieces

Fig. 3. Y -piece glued from a right angled geodesic octagon

As indicated in the figure the geodesic octagon is constructed from two right angled
geodesic hexagons.
Note that identification of the sides b and h only yields a subset of a hyperbolic
cylinder so that every pair of pants can be constructed from a hyperbolic cylinder by
cutting and gluing.

4. The Method of Particular Solutions on a Surface of Genus g

4.1. Explicit error estimates for hyperbolic surfaces. The first ingredient in our algo-
rithm is an estimate on the constant in the estimate (2) for F̃ in the particular case
when M is a hyperbolic surface and is a finite union of geodesic segments. To start
let us assume that consists of exactly one geodesic segment. The general case of a
finite union of geodesic segments will later be reduced to this case. We can assume
that M is realized as a quotient of the upper half plane H by some co-compact hyper-
bolic subgroup in SL(2, R) and that D ⊂ H is an (open) fundamental domain whose
boundary is a geodesic polygon. This means we can identify functions on M with func-
tions on H that are invariant under the action of the group. We can also assume that the
836 A. Strohmaier, V. Uski

geodesic segment is a segment on the imaginary axis. Then there is a tesselation of

H by translates Di of D. Let L > 0 and suppose that D1 , . . . , D N is a finite number
of translates of D such that dist(x, D) ≤ L implies that x is contained in the closure of
∪i=1
N D . If L is chosen as the smallest distance between non-adjacent sides of D then
i
N − 1 would be the minimal number of fundamental domains of D needed to cover
an open neighborhood of D (and thus the number of neighboring fundamental domains
sharing either an edge or a corner with D). Let χ : H → [0, 1] be a smooth compactly
supported function which is equal to one in an open neighborhood O of D with support
contained in D̃ = ∪i=1
N D . Now suppose that h is a distribution on M. Then,
i

|h( f )|
h H −2 (M) = sup
0= f ∈C ∞ (M) ( + 1) f L 2 (M)
|h(χ f )|
= sup .
0= f ∈C ∞ (M) ( + 1)χ f L 2 (D)

Using

( + 1)(χ f ) = f (χ ) + χ ( + 1) f + 2(∇χ , ∇ f ),

where the inner product is with respect to the metric, we get

( + 1)χ f L 2 ( D̃) ≤ χ ( + 1) f L 2 ( D̃) + f χ L 2 ( D̃) + 2 ∇χ , ∇ f  L 2 ( D̃) .

Of course

2 ∇χ , ∇ f  L 2 ( D̃) ≤ 2 ∇χ ∞ ∇f L 2 ( D̃) = 2 ∇χ ∞  f,  f  L 2 ( D̃)

≤ ∇χ ∞ ( + 1) f L 2 ( D̃) .

Since f is invariant under the group action

√
( + 1) f L 2 ( D̃) = N ( + 1) f L 2 (D) ,

and therefore we obtain

( + 1)χ f L 2 ( D̃) ≤ C̃ ( + 1)χ f L 2 (D)

with
√
C̃ = N (1 + ∇χ ∞ + χ ∞) .

Consequently,

|h(χ f )|
h H −2 (M) ≤ C̃ sup . (3)
f ∈C ∞ (M) ( + 1)(χ f ) L 2 ( D̃)

To continue we use a particular coordinate system on the upper half space. Namely, we
identify the upper half space with (−π/2, π/2) × R using coordinates (ϕ, t) and the
metric (cos ϕ)−2 (dϕ 2 + dt 2 ) such that ϕ = 0 coincides with the imaginary axis and the
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 837

coordinate system is centered at the point i. Note that (ϕ, t) is related to Fermi coordi-
nates (ρ, t) by (cosh ρ, t) = ((cos ϕ)−1 , t). We assume that in these coordinates is
identified with the segment {0} × [0, ] so that h has the form
     
d
h( f ) = h 1 ( f ) + h 2 ( f ) = F1 (t) f (0, t)dt − F2 (t) f (0, t)dt.
0 0 dϕ
Moreover in these coordinates  has the form (cos2 ϕ)e , where
 2 
∂ ∂2
e = − +
∂ϕ 2 ∂t 2
is the Euclidean Laplace operator in coordinates (ϕ, t). Then we have

( + 1)(χ f ) 2
L 2 ( D̃)
= | cos ϕ|e (χ f ) 2
L 2 ((−π/2,π/2)×R)
+ 2χ f, e (χ f ) L 2 ((−π/2,π/2)×R)
+ | cos ϕ|−1 χ f 2
L 2 ((−π/2,π/2)×R)
≥ (Ĉ (e + 1)(χ f ) L 2 ((−π/2,π/2)×R) )
2
,

where Ĉ = inf{| cos ϕ| | χ (ϕ, t) = 0}. Collecting all terms we obtain

|h( f )|
h H −2 (M) ≤ Ĉ −1 C̃ sup .
f ∈C0∞ (R2 ) (e + 1) f L 2 (R2 )

However,

1 | F̂1 (η)|
|h 1 ( f )| = √ (1 + ξ 2 + η2 ) fˆ(ξ, η)dξ dη
2π R2 1 + ξ 2 + η2
1
≤ F1 H −3/2 (R) (e + 1) f ,
2

1 | F̂2 (η)ξ |
|h 2 ( f )| = √ (1 + ξ 2 + η2 ) fˆ(ξ, η)dξ dη
2π R 2 1 + ξ 2 + η2

1
≤ F2 H −1/2 (R) (e + 1) f ,
2
where F1 and F2 are understood as functions on R extended by zero from [0, ]. There-
fore, we finally obtain

h H −2 (M) ≤C F1 2
H −3/2 (R)
+ F2 2
H −1/2 (R)
, (4)

where the constant C is given by

1 √
C = √ ( sup cosh ρ) N (1 + ∇χ ∞ + χ ∞) . (5)
2 (ρ,t)∈suppχ
If consists of several geodesic segments then of course the same inequality holds with
the constant being the maximum of the constants for the individual segments.
For low lying eigenvalues it is in practice easier to compute the L 2 -norms of F1 and
F2 . Of course since the negative Sobolev norms are dominated by the L 2 -norms the
above gives estimates of h H −2 (M) in terms of the L 2 -norms of F1 and F2 . However, in
838 A. Strohmaier, V. Uski

this case a slightly better constant can be obtained using Theorems B.4 and B.5. Indeed,
Eq. 3 means by duality that
h H −2 (M) ≤ C̃ (H + 1)−1 h L 2 ( H) ,

where H is the Laplace operator on the hyperbolic plane. By Theorem B.4 and B.5 we
have

(H + 1)−1 h L 2 ( H) ≤ C12 + C22 F1 2

L 2 (R )
+ F2 2
L 2 (R )
, (6)

where C1 and C2 are the constants in these theorems.

Example 4.1. For a surface of genus 2 we use a right angled 12-gon and thus have
N = 25. Let L be the minimum of the distance between non-adjacent sides. For each
side si we use a cutoff function χi (ρ, t) = P(ρ/L), where P is the function
⎧
⎨ 1, x <0
P(x) = 0, x >1
⎩ 1 − 3x 2 + 2x 3 , x ∈ [0, 1],

and ρ are Fermi-coordinates with respect to the infinite extension of the side si oriented
in such a way that the region ρ > 0 does not intersect the polygon. Note that ∇χi ≤ 2L 3

and χi ≤ L62 + 2L 3

. With χ = χ1 · · · χ12 we then get
3
∇χ ∞ ≤ , (7)
L
33 3
χ ∞ ≤+ , (8)
L2 L
where we used the well known formula (χ1 χ2 ) = (χ1 )χ2 + 2∇χ1 , ∇χ2  + (χ2 )χ1
and the fact that near a fixed point all but at most two of the functions χi are constant.
Since χ is not smooth we can not directly use it in our estimate. However, we can use
a suitable regularization χ and the fact that the second distributional derivatives of χ
are in L ∞ to show that the estimate above still holds with the function χ . Collecting all
terms we get
 
33 6
h H −2 (M) ≤ 5 (C1 + C2 )
2 2 + +1 F1 2L 2 (R) + F2 2L 2 (R) .
L2 L

Since 5 (C12 + C22 ) ≤ 12 this further simplifies into the easy step to use estimate
 
33 6
h H −2 (M) ≤ 12 + +1 F1 2L 2 (R) + F2 2L 2 (R) . (9)
L2 L
This estimate improves by a factor of 4.69 if one is willing to believe the numerical
values computed at the end of Appendix Appendix B.

4.2. Basis functions on hyperbolic cylinders. Let  > 0. Then the hyperbolic cylinder
Z  is the non-compact manifold obtained by factorizing
 /2 the upper
 half space by the
e 0
subgroup of S L(2, R) generated by the element . We will use Fermi
0 e−/2
coordinates (ρ, t) which are related to the usual coordinates by
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 839

1
(x, y) = et (tanh ρ, ).
cosh ρ

For L > 0 denote by Z L the finite hyperbolic cylinder

Z L = {x ∈ Z  | −L ≤ ρ(x) ≤ L}.

For λ ∈ R let V (λ) (Z L ) be the space of functions f in C ∞ (Z L ) that satisfy

( − λ) f = 0.

Note that f ∈ V (λ) (Z L ) has a Fourier expansion of the form


f (ρ, t) = (ak φk (ρ, t) + bk ψk (ρ, t)) ,
k∈Z

where (φk )k∈Z satisfy

( − λ)φk = 0

and have initial data

φk (t, 0) = e2π ikt/ ,

∂ρ φk (t, 0) = 0,

and the functions ψk satisfy

( − λ)ψk = 0

and have initial data

ψk (t, 0) = 0,
∂ρ ψk (t, 0) = e2π ikt/ .

These functions are of the form k (ρ)e2π ikt/ , where k (ρ) solves the ordinary differ-
ential equation
 
1 d d 4π 2 k 2
− cosh ρ + − λ k (ρ) = 0 (10)
cosh ρ dρ dρ 2 cosh2 ρ
with the corresponding initial conditions. A fundamental system of (non-normalized)
solutions of this equation, consisting of an even and an odd function, can be given
explicitly in terms of hypergeometric functions
 
2π ik s π ik 1 − s π ik 1
k (ρ) = (cosh ρ)
even  2 F1 + , + ; ; − sinh ρ ,
2
2  2  2
 
2π ik 1 + s π ik 2 − s π ik 3
k (ρ) = sinh ρ(cosh ρ)
odd  2 F1 + , + ; ; − sinh ρ ,
2
2  2  2
where λ = s(1 − s) (see [Bor10], where these functions are analysed). Normalization
gives the corresponding solutions to the initial value problems.
840 A. Strohmaier, V. Uski

(λ)
Denote by VN the 4N + 2 dimensional subspace of V (λ) spanned by φk and ψk with
|k| ≤ N . Given f ∈ V (λ) we can truncate the Fourier expansion to obtain an element
(λ)
f (N ) ∈ VN given by

f (N ) (ρ, t) = (ak φk (ρ, t) + bk ψk (ρ, t)) .
|k|≤N

The C 1 -norm on the cylinder Z L is defined as

1
f 2
C1
= f 2
L∞ + ∂t f 2
L∞ + ∂ρ f L∞ .
2
cosh ρ
The following result is crucial for our approximation of eigenfunctions.
Theorem 4.2. Suppose φ ∈ V (λ) (Z  ) is bounded on Z  . Let L > 0 and suppose that N
2 2
is an integer such that 24πcoshN2 L > λ ≥ 0. Then,

φ − φ (N ) L ∞ (Z L ) ≤ β̃λ (N ) φ L ∞ (Z  ) ,

φ − φ (N ) C 1 (Z L ) ≤ βλ (N ) φ L ∞ (Z  ) ,

where β̃λ (N ) = A1 (N ), βλ (N ) = A1 (N )2 + A2 (N )2 + A3 (N )2 ,
∞

A1 (N ) = 4 (cosh ck (ϕ0 ))−1 ,
k=N +1
∞
2π k
A2 (N ) = 4 (cosh ck (ϕ0 ))−1 ,

k=N +1
∞
 4π 2 k 2 λ
A3 (N ) = 4 − (sinh ck (ϕ0 ))−1 ,
2 cos (ϕ0 )
2
k=N +1

and
⎛ ⎞ ⎛ √ ⎞
2π k 2π k
sin(ϕ) √ λ sin(ϕ)
ck (ϕ) = arccos ⎝  ⎠− λ arccot ⎝ ⎠.
 4π 2 k 2
−λ 4π 2 k 2
cos2 (ϕ) − λ
2 2

Remark 4.3. Both βλ (N ) and β̃λ (N ) are decreasing exponentially fast in N for fixed λ
and give very good bounds for reasonably large N (see Fig. 4).
Proof of Theorem 4.2. By rescaling we can assume in the proof that φ L ∞ (Z  ) ≤ 1. Of
course ψ = φ − φ (N ) has a convergent expansion into a Fourier series

ψ(ρ, t) = (ak φk (ρ, t) + bk ψk (ρ, t)) (11)
|k|>N

which converges uniformly on compact sets. Both φk and ψk are of the form
k (ρ)e2π ikt/ , where k solves the ordinary differential equation (10) with initial con-
ditions either k (0) = 1, k (0) = 0 or k (0) = 0, k (0) = 1 respectively. Thus, k
is either even or odd with respect to the reflection ρ → −ρ.
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 841

4
10

6
10

8
10

10
10

12
10

14
10
N
50 60 70 80 90
√
Fig. 4. βλ (N ) on a logarithmic scale for λ = 30,  = 2 arccosh(3 + 2 2) and L = 3/2. The corresponding
cylinder covers a fundamental domain for the Bolza surface

Substituting ϕ = 2 arctan(tanh ρ/2) we write k (ρ) = k (ϕ). Let

4π 2 k 2 λ
Vk (ϕ) = −
2 cos2 ϕ
√
 λ
and define sk = arccos 2π |k| so that Vk is non-negative on the interval [0, sk ]. Then, by
Lemma A.2, we have for all k > N and φ < sk :
1
k (sk ) ≥ k (ϕ) cosh ck (ϕ) + k (ϕ) √ sinh ck (ϕ),
Vk (ϕ)
where
 sk 
ck (ϕ) = Vk (r )dr.
ϕ
This integral can be computed explicitly and gives the expression in the theorem. There-
fore
|k (ϕ)| ≤ (cosh ck (ϕ))−1 k (sk ),

|k (ϕ)| ≤ Vk (ϕ) (sinh ck (ϕ))−1 k (sk ).
Since the ψ(ρ, t) is bounded by 1 we have |ak ||φk (ρ, t)| ≤ 1 and |bk ||ψk (ρ, t)| ≤ 1,
for any ρ > 0. With ϕ0 = 2 arctan tanh L2 the above bound then implies that for any
ρ ∈ [−L , L] both |ak ||φk (ρ, t)| and |bk ||ψk (ρ, t)| are smaller than or equal to
(cosh ck (ϕ0 ))−1
∂ ∂
and |ak || ∂ρ φk (ρ, t)| and |bk || ∂ρ ψk (ρ, t)| are smaller than or equal to

Vk (ϕ0 ) (sinh ck (ϕ0 ))−1 .
Summing the Fourier series one obtains on Z L the bounds |ψ| ≤ A1 , | ∂t∂ ψ| ≤ A2 and
∂
| ∂ρ ψ| ≤ A3 . 

842 A. Strohmaier, V. Uski

0
–1 0 1 2 3 4 5 6

Fig. 5. A Y -piece that was cut open along two geodesic segments covered by a hyperbolic cylinder. The hyper-
bolic cylinder is obtained from the shaded region by identification of the two dashed circles. The inscribed
finite cylinder is double shaded

4.3. Main estimates for the algorithm for a surface of genus g. Suppose that M is a
surface of genus g. Then M can be decomposed into X and Y pieces (pairs of pants).
Each of these pieces can be cut open along a shortest geodesic that connects two different
boundary components to obtain a surface that can be identified with a closed subset of a
hyperbolic cylinder (see Fig. 5). In this way we obtain a collection (Mi )i=1,...,m of closed
subsets of hyperbolic cylinders Z i with geodesic boundary components in such a way
that the defining unique simple closed geodesic γi on Z i is contained in Mi . Note that
we use disjoint cylinders here, i.e. we use a different copy of Z i even if the geodesics
have the same length and correspond to the same model in the upper half space. As γi
is a simple closed geodesic in M the subgroup it generates in the fundamental group
will define a covering map Z i → M. Therefore, every eigenfunction φ on M lifts to
a bounded smooth function on Z i which is an eigenfunction of the Laplace operator
on Z i .
The surface M can of course be constructed from the Mi by glueing, i.e. by identi-
fying the different boundary geodesics. Suppose that we have a collection ( α )α=1,...,M
of geodesic segments and a collection ( α̃ )α=1,...,M so that the geodesic segment α
is identified with the segment α̃ . Thus, = ∪i i can be thought of as afinite union
of geodesic segments on M and M\ is the interior of the disjoint union i Mi of the
Mi . This is exactly the setting described in Sect. 2, where M\ gets identified with

i Mi andthe segments α and α̃ form the boundary of M\ . Thus, for a function
f ∈ C ∞ ( i Mi ) which is smooth up to the boundary, the boundary values at α and
α̃ represent the two different boundary values it has along α if we think of it as a
function on M.
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 843

On each segment α we choose a set {xα,1 , xα,2 , . . . , xα,n α } consisting of an even

number of equidistant points such that xα,1 is the origin of the segment and xα,n α is the
endpoint. Let {xα̃,1 , xα̃,2 , . . . , xα̃,n α̃ } be the corresponding points on α̃ . We denote by
δα the distance between neighbouring points on the segment α and by δ the maximum
of (δα )α .
Let N > 0 be a fixed integer and suppose λ > 0. Of course on each cylinder Z i and
for each integer m such that |m| ≤ N there exists two linearly independent functions on
this cylinder which are both eigenfunctions of the Laplace operator with eigenvalue λ
and which are of the form
2iπ mti
(ρi , ti ) = (ρi )e i
,
where (ρi , ti ) are Fermi coordinates on the corresponding cylinder. We are choosing
the basis functions orthonormal on the largest subset of Mi which is invariant under
the rotation symmetry of the cylinder (for a Y -piece this will be one half of a cut-off
hyperbolic cylinder and corresponds to the double shaded region in Fig. 5). Since such a
subset will always have the form [L 1 , L 2 ] × S 1 , this can be achieved by choosing (ρi )
orthonormal on this interval.
We extend these functions by zero to the disjoint union of all the cylinders and think
of the collection of all these functions as a basis (k )k=1,...,(4N +2)m in the space
(λ)
⊕i=1
m
VN (Z i ).

We will use this basis to identify vectors in C(4N +2)m with functions on the disjoint union
of the cylinders, but also with L ∞ functions on M, by restricting the functions to the
pieces Mi . For v ∈ C(4N +2)m let (v) be the corresponding function.
For each geodesic segment α we form the matrices

δα bi
(A1,α )ik = k (xi ),
3

δα bi
(A2,α )ik = nα k (xi ),
3

δα bi
(A3,α )ik = tα k (xi ),
3
where n is the outward normal unit vector field along the geodesic segment, t the nor-
malized tangential vector field, and bi are the coefficients in the composite Simpson
rule,1 i.e. b1 = 1, b2 = 4, b3 = 2, . . . , bn α −1 = 4, bn α −1 = 1. Next arrange A1,α
and A2,α into one matrix Aα = A1,α ⊕ A2,α , and add also the tangential derivatives
Atα = A1,α ⊕ A2,α ⊕ A3,α . All the matrices involved map into the same space and we
understand the direct sum here as a direct sum of linear maps mapping into C(4N +2)m .
Thus, the matrix A1 ⊕ A2 is obtained by attaching all the rows of A2 as rows to the
matrix A1 (similar to the Matlab notation [A1 ; A2 ]).
Similarly, but with one change of sign

δα̃ bi
(A1,α̃ )ik = k (xi ),
3
1 Other coefficients such as those of the Gauss rule may also be used here and lead to analogous error
estimates.
844 A. Strohmaier, V. Uski


δα̃ bi
(A2,α̃ )ik = − nα̃ k (xi ),
3

δα̃ bi
(A3,α̃ )ik = tα̃ k (xi ),
3
and again Aα̃ = A1,α̃ ⊕ A2,α̃ , Atα̃ = A1,α̃ ⊕ A2,α̃ ⊕ A3,α̃ . Now define

A = ⊕α Aα , At = ⊕α Atα
à = ⊕α Aα̃ , Ãt = ⊕α Atα̃

as well as Bλ,N = At − Ãt , B0λ,N = A − Ã and Cλ,N = A ⊕ Ã. Thus, Bλ,N as compared
to B0λ,N contains also the tangential derivatives.
For each v ∈ C(4N +2)m the vector B0λ,N v contains the discretization of the jump
of boundary data of (v) across whereas Cλ,N v contains the discretization of the
boundary data itself. We will see below that the L 2 norm of an eigenfunction on M can
be bounded from above and below by the L 2 -norm of the boundary data and thus the
norm of Cλ,N v will serve as a normalization in our method. To be more precise, the
matrices are chosen such that the norms B0λ,N v and Cλ,N v are the approximations
of F0,0 ((v)) and of the L 2 norm of the boundary data of (v) using the Simpson
rule. The error of composite Simpson integration of a function g on an interval [a, b]
with step size δ is bounded from above by δ 180
4 |b−a|
supξ ∈[a,b] |g (4) (ξ )| (see e.g. [PT96],
p. 133). Thus, we have
| B0λ,N v − (F0,0 ((v)))2 |
2

δ4   4 
≤ Dt ((v)2 )| L ∞ ( ) + Dt4 n ((v)2 )| L∞( ) , (12)
180  
| Cλ,N v 2 − ((v)| 2L 2 ( ) + n(v))| 2L 2 ( ) |
δ4   4 
≤ Dt ((v)2 ) L∞( ) + Dt4 n ((v)2 )| L∞( ) , (13)
180
where Dt denotes the tangential derivative.
Moreover, as we choose the basis functions orthonormal with respect to the L 2 -inner
product of a subset of M, we have
(v) L 2 (M) ≥ v . (14)
Assume now that λ is an eigenvalue of the Laplace operator and suppose that φ is
an L 2 -normalized eigenfunction. We may truncate the Fourier expansion of this eigen-
function on each of the cylinders to obtain a function φ (N ) which then corresponds to a
fixed vector v such that φ (N ) = (v). We choose N large enough such that Theorem
4.2 applies and from the fact that φ is bounded by an explicitly computable constant (see
Appendix, Cor. E.3) the following estimates are explicit
(φ (N ) − φ)| L∞( ) ≤ βλ (N ) φ ∞, (15)
(N )
(nφ − nφ)| L∞( ) ≤ βλ (N ) φ ∞. (16)
The error of the composite Simpson integration of a function g on an interval [a, b]
with step size δ is also bounded by δ|b − a| supξ ∈[a,b] |g (ξ )| as one can easily see by
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 845

replacing g in each sub-interval [xk , xk+1 , xk+2 ] of length 2δ by the constant function
g(xk+1 ). Therefore,2 we obtain
 
| Cλ,N v 2 − φ| 2L 2 ( ) + nφ| 2L 2 ( ) |
 
≤ 2βλ (N )2 φ 2∞ + (2δ ) Dt (φ 2 )| L ∞ ( ) + Dt n (φ 2 )| L ∞ ( ) .

From the bound (23) of the Appendix and the fact that φ is L 2 -normalized we conclude
that
 1
2
φ| 2L 2 ( ) + nφ| 2L 2 ( ) > c1 (λ),
where c1 is explicitly computable and depends only on λ and the geometry of M and
. Using the bound L ∞ and the C 1 bound in the Appendix of eigenfunctions we can
compute constants Nc and δc such that for all N > Nc and δ < δc we have
c1 (λ)
Cλ,N v > .
2
Note that this is a very crude estimate and we have used only the linear error estimate in
Simpson’s rule. This was done so that we can use the explicit estimates that we proved in
the Appendix. The critical constants δc and Nc guarantee that the numerical integration
using the Simpson rule together with the approximation of φ by the truncated function
together yield a relative error of not more than 50%. For the method it will be enough
to show that Cλ,N v is bounded from above so that this estimate is sufficient.
Let σ1 (B0λ,N ) be the smallest singular value of B0λ,N and σ1 (Bλ,N , Cλ,N ) the smallest
generalized singular value (see e.g. [GL96] or also [Bet07] in the context of MPS) of
the pair (Bλ,N , Cλ,N ). Here the generalized singular values of Bλ,N ∈ Mat(m 2 × m 1 )
with respect to Cλ,N ∈ Mat(m 3 × m 1 ) are defined as the singular values of Bλ,N as an
operator from Cm 1 to Cm 2 where the Euclidean norm · on Cm 1 is replaced by the
norm · Cλ,N defined by v Cλ,N := Cλ,N v . Note that
Bλ,N v
σ1 (Bλ,N , Cλ,N ) = inf .
v=0 Cλ,N v
If N is large enough, so that the assumptions of Theorem 4.2 are satisfied, we have
√
2 2
σ1 (Bλ,N , Cλ,N ) ≤ βλ (N ) φ ∞
c1 (λ)
in case λ is an eigenvalue. This follows immediately from the fact that an exact eigen-
function satisfies the boundary conditions.
Let as usual λ = s(1 − s) with Im(s) ≥ 0 and define σ = Re(s) and r = Im(s).
Assume that λ is not an eigenvalue and let s , σ and r be defined as above. For each i
we choose a hyperbolic cylinder Z L with L large enough such that Mi ⊂ Z L and such
that Mi is relatively compact in the interior of Z L , so that the distance between ∂ Z L and
Mi greater than a positive number d > 0. For the truncated eigenfunction φ (N ) consider
(N ) (N )
the boundary data φ Z ⊕ nφ Z on ∂ Z L . Using the resolvent kernel of the Laplace
3
operator on the hyperbolic cylinder (see Appendix Appendix D) we can reconstruct
φ (N ) from its boundary data
2 We use this estimate instead of (13) because a bound on the first derivative of eigenfunctions is easier to
obtain explicitly and we do not need good accuracy for this bound.
3 This is also often referred to as the Green’s function of the Helmholz operator.
846 A. Strohmaier, V. Uski


φ (N ) (x) = ksZ (x, x )nx φ (N ) (N )
Z (x ) − φ Z (x )nx ks (x, x )d x .
Z
∂ZL
Now define

(N ) (N ) (N )
φs (x) = ksZ (x, x )nx φ Z (x ) − φ Z (x )nx ksZ (x, x )d x .
∂ZL
Then
(N )
φ (N ) (x)−φs (x)
 s
∂ Z (N ) (N ) ∂ Z
= ks̃ (x, x )nx φ Z (x ) − φ Z (x )nx k (x, x )d x d s̃.
s ∂Z L ∂ s̃ ∂ s̃ s̃
Using Lemma D.2 we obtain
 1
(N ) (N ) 2 (N ) 2
φ (N ) − φs
2
C 1 (M) ≤ |s − s |C̃ M,s,s φZ L 2 (∂ Z )
+ nφ Z L 2 (∂ Z )
.

Of course the L 2 -norm of the boundary data is bounded from above by the L 2 -norm of
the boundary data of the exact eigenfunction φ so that we obtain the bound
(N )

φ (N ) − φs C 1 (M) ≤ |s − s |C̃ M,s,s Vol( ) φ C 1 (M) .
Theorem 4.4. With the notation above we have for all N > Nc and δ < δc ,
√
2 2   
σ1 (Bλ ,N , Cλ ,N ) ≤ βλ (N ) φ ∞ + |s − s |C̃ M,s,s Vol( ) φ C1
c1 (λ)
if λ is an eigenvalue with normalized eigenfunction φ. Here φ ∞ and φ C1 are
bounded by explicit constants (Corollary E.3 and E.5).
Moreover, by Theorem 2.1 and the inequalities (12) and (14) we have:
Theorem 4.5. Let λ ≥ 0. Suppose that v ∈ C(4N +2)m is a unit vector and  = B0λ,N v
and let
δ4  4 
τ = C( + Dt ((v)2 ) L ∞ ( ) + Dt4 n ((v)2 )| L ∞ ( ) ) < 1,
180
where C = C̃ C12 + C22 are the constants in Sect. 4.1. Then, there exists an eigenvalue
in the interval [λ − (1+λ)τ (1+λ)τ
1−τ , λ + 1−τ ]. In particular the estimate holds if  = σ1 (Bλ,N )
0
and v is a corresponding normalized singular vector.
4  
Remark 4.6. The term  + δ180 Dt4 ((v)2 ) L ∞ ( ) + Dt4 n ((v)2 )| L ∞ ( ) can be
minimized as another singular value problem, using the bounds on the derivatives of the
basis functions that follow from the differential equation together with Lemmata A.4
and A.1. This results in a matrix, whose singular value is precisely the error estimate.
Actual computations show that in practice the major contribution of the error comes
from the first term.
Remark 4.7. An analog of Theorem 4.4 holds for the n th relative singular values Bλ ,N
with respect to Cλ ,N . In this case the term |s − s | gets replaced by the maximum of
{|sk − s | | k = 1 . . . , n}, where sk correspond to n pairwise orthogonal eigenvalues.
We do not state this here as we want to keep the exposition simple. In the same way
Theorem 4.5 applies to situations with multiplicities and eigenvalues that are close to
each other (see Remark 2.2).
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 847

4.4. Description of the algorithm for a surface of genus g. Since the fundamental solu-
tions of the differential equation (10) satisfy the bounds of Lemmata A.4 and A.1,
their higher derivatives can also be bounded explicitly simply by using the differential
equation. We use coordinates r = sinh ρ and t on the hyperbolic cylinders, so that the
differential equation is of the form
 
d d 4π 2 k 2
−(1 + r 2 ) (1 + r 2 ) + − (1 + r 2
)λ k (r ) = 0. (17)
dr dr 2
Using subdivision into smaller intervals and the Taylor expansion around the boundary
points of these intervals yields an approximation (M)k (r ) to the exact solution by piece-
wise defined polynomial splines. Using the bounds on the higher derivatives the error
in the C 1 -norm can be explicitly bounded from above. Since the measure on the hyper-
bolic cylinder is equal to dr dt in these coordinates, normalization of the basis functions
can be done within this error, simply by integrating these piecewise polynomials. In this
way, the matrices B0λ,N , Bλ,N and Cλ,N can be computed with explicit error bounds. The
generalized singular value decomposition can then be done using any method from linear
algebra. Once a numerical generalized singular value decomposition has been obtained
the error of the singular values can be estimated from above as follows. Suppose that UDV
is a numerically obtained generalized singular value decomposition of B with respect to
C. Thus, D is diagonal and U is unitary within a certain precision. Moreover, V is unitary
with respect to the inner product induced by C, again within a certain precision. Thus,
U∗ U − 1 ≤ δ1 and V∗ C∗ CV − C∗ C ≤ δ2 , and B − UDV ≤ δ3 , where these
three numbers reflect the errors. All three numbers can be estimated from above using
interval arithmetics for example by bounding the operator norm by the Hilbert-Schmidt
norm. Since the matrices U0 = U(U∗ U)−1/2 and V0 = V(V∗ C∗ CV)−1/2 (C∗ C)1/2 are
unitary in the respective inner products the generalized singular values of U0 DV0 are
exactly the diagonal entries of D. By the characterization of generalized singular values
of B as the norm distance of B to the rank k operators, it follows that the error of the
generalized singular values of B bounded by

(B − U0 DV0 )(C∗ C)−1/2 .

Using functional calculus, a longer calculation shows that this is bounded by

 
δ3 δ1 δ2 + δ1 δ2
+ D + ,
σ1 (C) 2 σ1 (C)3

where σ1 (C) is the smallest singular value of C. Here we assumed that δ2 ≤ σ1 (C). Note
that the smallest singular value of C can also be estimated from below using exactly
the same procedure for the ordinary singular value decomposition of C. Thus, interval
arithmetics can be applied to obtain a lower bound for the generalized singular values
of B with respect to C.
An interval I is then tested for eigenvalues by computing a lower bound for the value
of σ1 (Bx,N , Cx,N ) for a discrete set of points x in I . By Theorem 4.4 for large enough N
the distance between the points can be chosen such that an eigenvalues can occur only
near points where the lower bound is small enough. For a fixed  > 0 this algorithm
then will normally find a discrete set of points yi such that eigenvalue can occur only
in -neighborhoods of these points. The presence of eigenvalues and their multiplicities
can then be tested using the singular values of B0λ,N . An upper bound for small singular
848 A. Strohmaier, V. Uski

values can easily be obtained by first finding the singular value decomposition numeri-
cally and then using the numerically obtained singular vectors v to compute B0λ,N v .
Since the derivatives of the basis functions can be bounded using the differential equa-
tion (10) and the bounds of Lemmata A.4 and A.1 the error terms in Theorem 4.5 can
directly be estimated from above. Using interval arithmetics one can therefore obtain
rigorous interval inclusions for the eigenvalues.
Remark 4.8. Once the presence of a single eigenvalue is established in a small interval
Theorem 4.4 can also be used to establish an error bound by narrowing the interval in
which the eigenvalue can be located.
The analysis of rigorous error bounds can be simplified considerably by replacing
the exact basis functions with the piecewise defined polynomial basis functions them-
selves. These are of course orthogonal on any cylinder, thus normalization involves only
integration of piecewise defined polynomials which can be done explicitly. Thus, the
quasi-mode becomes a finite linear combination of products of piecewise defined poly-
nomials in r and Fourier modes in t. The so constructed function does not satisfy the
equation ( − λ) = 0 but instead ( − λ) = χ , where due to the nature of the differ-
ential equation (1 + r 2 )χ is again constructed out of piecewise defined polynomials and
Fourier modes. Its L 2 -norm can very easily be estimated from above by integrating over
the smallest cylinders that contain Mi . The fourth derivatives in the error of Simpson’s
rule in Theorem 4.5 can also be estimated more directly in this case as a bound can be
easily obtained from the coefficients of the polynomial.

5. Implementation and Examples

We have implemented our method in three different programs.

5.1. Non-rigorous implementations. The first two programs are Fortran programs. One
works solely for genus 2 surfaces and takes as coordinates mw-Fenchel-Nielsen coor-
dinates. The second program takes as an input a labeled graph (encoding the way the
surface is glued from Y -pieces) and the Fenchel-Nielsen coordinates. Both programs
then follow the algorithm described in this article. The basis functions are computed
for each value of λ by numerically solving the differential equation (10) using the
DLSODA routine from ODEPACK (see [Hin83,Pet83]). Normalization of the basis
functions is achieved by numerical integration of the solution of (10). In order to com-
pute the singular values σ1 (B0λ,N ) we use LAPACK routines. The generalized singular
values σ1 (Bλ,N , Cλ,N ) are computed using QR-decomposition. A search algorithm then
looks for the small minima of σ1 (Bλ,N , Cλ,N ) as a function of λ. A small enough step
size guarantees that no eigenvalues are missed in the process. The other singular values
are also necessary in the search: a small second singular value of B0λ,N implies that
either the eigenvalue has multiplicity greater than one, or another eigenvalue is very
close. Our method can not distinguish between multiplicities and close eigenvalues if
they can not be separated within the given precision. Both programs achieve accuracies
close to machine precision for the low lying eigenvalues when the surface is far enough
from the boundary of Teichmüller space (so that the DLSODA algorithm can compute
the basis functions accurately enough). The code for the genus two program is available
to the scientific community under GPLv3 and can be obtained from http://www-staff.
lboro.ac.uk/~maas3/hyperbolic-surfaces/hypermodes.html.
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 849

Another program√ was written in Mathematica. It uses Taylor’s method on intervals

of size of order 1/ λ in order to solve the differential equation (10) and compute the
matrices Bλ,N and Cλ,N with arbitrarily high precision. The linear algebra such as QR-
decomposition and search for the minimum is done within Mathematica. The achieved
accuracy depends on the computing time and the machine precision used. We were able
to compute the first 70 digits of the first eigenvalue for the Bolza surface (see Sect. 5.3)
using this program.

5.2. Rigorous numerical implementations. To demonstrate the power of our method we

modified the genus two Fortran and Mathematica codes to give rigorous error estimates
for the eigenvalues. As we described in the previous section we use the Taylor series
method to solve the differential equation (10) within a given accuracy and then follow
the analysis described in Sect. 4.4. The corresponding eigenvalue inclusions can be con-
sidered rigorous if one carefully includes rounding errors in the analysis and is willing to
trust the implementation of basic mathematical functions in Fortran and Mathematica.

5.3. The Bolza surface. In order to demonstrate our method we consider a special sur-
face of genus 2. The so-called Bolza surface has Fenchel-Nielsen mw-coordinates

(1 , t1 ; 2 , t2 ; 3 , t3 )
√ 1 √ √
= (2 arccosh(3 + 2 2), ; 2 arccosh(1 + 2), 0; 2 arccosh(1 + 2), 0).
2
This surface is known to maximize the length of the systole in genus 2 and it also
maximizes the order of the symmetry group (see e.g. [Jen84,Sch93,Sch94]). The first
eigenvalue on the Bolza surface was estimated by Jenni [Jen84] remarkably accurately
to be in the interval [3.83, 3.85]. Aurich and Steiner [AS89] (see also [ASS88]) give the
value 3.8388 which was obtained using the finite element method. Figure 6 shows a plot
of the singular value after QR-decomposition (N = 60, δ = 0.001) of the matrix A ⊕ B
and projection onto the first direct summand. Note that at the minima the function is
very small and order 10−12 .
The search algorithm in the Fortran program finds the first eigenvalue within 8 digits
of precision. Using multiple precision in Mathematica we obtained the value:

λ1 ≈ 3.8388872588421995185866224504354645970819150157,

where we believe that all digits are correct. The rigorous implementation that uses very
simple bounds in its current form gives a rigorous error bound of 10−6 (which can be
improved by simply investing more computing time).
We believe that λ1 is important as we conjecture that this is the maximal value of the
first eigenvalue of the Laplace operator for constant negative curvature −1 in genus 2.
This conjecture is supported by our calculations in Teichmüller space (the correspond-
ing analysis will be discussed elsewhere) and is in line with the findings in ([JLN+ 05]),
where it is shown4 that the maximal value of the first eigenvalue of the Laplace operator
with respect to any metric fixing the volume is attained for a singular metric on the Bolza
surface ([JLN+ 05]).
4 This proof relies on a lower bound for the eigenvalue of a mixed Dirichlet-Neumann problem that was
obtained numerically from a finite element method.
850 A. Strohmaier, V. Uski

Fig. 6. Smallest singular value after QR-decomposition for the Bolza surface as a function of λ

6. Calculation of the Spectral Zeta Function and the Spectral Determinant

The spectral zeta function ζ (s) is defined as the meromorphic continuation of the
function
∞

ζ (s) = λi−s ,
i=1

where (λi )i∈N0 are the eigenvalues repeated according to their multiplicities. The above
sum converges for (s) > 1. As usual a meromorphic continuation to the whole complex
plane is constructed using the Mellin transform
 ∞
1  
ζ (s) = t s−1 tr(e−t ) − 1 dt,
(s) 0
and the fact that the heat kernel has an asymptotic expansion, by splitting the above
integral into an integral over [0, 1] and one over [1, ∞). From the heat expansion it fol-
lows that zero is not a pole, which allows one to define the zeta-regularized determinant
det ζ () of the Laplace operator by

log det ζ () = −ζ (0).

In the examples we will also consider the value of the zeta function at the point −1/2.
In physics this value is often referred to as the Casimir energy or vacuum energy.
For our purposes we use the following splitting:
   ∞ 
1  −t   −t 
ζ (s) = t s−1
tr(e ) − 1 dt + t s−1
tr(e ) − 1 dt ,
(s) 0 

where  > 0.
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 851

The Selberg trace formula applied to the heat trace (see e.g. Eq. (193) in [M04]) is
t  ∞
  e−t/4 (γ )e− 4t
n 2 (γ )2
Vol(M)e− 4 ∞ π e−r t
2
−t
tr(e )= dr + √ )
, (18)
4π t 0 cosh2 (πr ) n=1 γ
4π t 2 sinh n(γ
2

where the second sum in the second term is over the set of primitive closed geodesics γ .
Splitting the integral and using the standard analytic continuation gives for (s) > −N ,
1
ζ (s) = (T  (s) + T2,N (s) + T3,N (s) + T4,N (s)),
(s) 1
where
∞

T1 (s) = λi−s (s, λi ),
i=1
N
ak  s+k−1
T2,N (s) = ,
s+k−1
k=0

,N Vol(M) ∞
T3 (s) = I N (r )dr,
4π 0

∞   n 2 (γ )2
−t/4 (γ )e− 4t
,N s−1 e
T4 (s) = t √ )
dt,
n=1 γ 0 4π t 2 sinh n(γ2

where
  
 
N
(−1)k 1 k k
−(r 2 + 14 )t
I N (r ) = t s−2
e − (r + ) t dt,
2
0 k! 4
k=0

the ak are given by

 ∞
Vol(M) (−1)k π(r 2 + 1/4)k
ak = dr − δ1,k ,
4π 0 k! cosh2 (πr )
and (x, y) denotes the incomplete Gamma function
 ∞
(x, y) = t x−1 e−t dt.
y

Differentiation results in the following formula for the spectral determinant:

− log det ζ  = ζ (0) = L 1 + L 2 + L 3 ,

where
∞

L 1 = (0, λi ),
i=1
852 A. Strohmaier, V. Uski

 
Vol(M) Vol(M) Vol(M)
L 2 = − − + 1 (γ + log()) +
4π  12π
  2 1  4 
 ∞  
1−E2 (r + ) 1
× sech (πr )
2 4
+ (r + ) γ −1+log((r + 1/4)) dr,
2 2
0  4
∞ 

n 2 (γ )2
 i e− 4t
L 3 = e −t/4
√   dt,
n=1 γ 0 4 π t 3/2 sinh 21 n(γ )

and E2 (x) is the generalized exponential integral which equals x (−1, x). All the inte-
grals have analytic integrands and can be truncated with exponentially small error. They
can therefore be evaluated to high accuracy using numerical integration.
For fixed s and  > 0 not too small the sums over the eigenvalues converge rather
quickly and therefore T1 (s) and L 1 can be computed quite accurately from the first
eigenvalues only. The error made by summing only over finitely many eigenvalues can
be estimated directly by Lemma E.2. For not too large  > 0 the sums over the length
spectrum are also very fast convergent. As they are exponentially decaying as  tends
to zero one can either neglect this contribution by choosing  small enough or one can
compute them using a finite part of the length spectrum. The error introduced in this
way is also exponentially decreasing as  → 0 and it can be explicitly estimated using
the Selberg trace formula as follows. If the length of the shortest closed geodesic
√ is L
the second term in the Selberg trace formula Eq. (18) is for all t < T < L 2 + 1 − 1
bounded by

T T L2 L2
FT (t) = tr(e−T )e 4 + 4T e− 4t (19)
t
and the heat trace for a fixed T can be estimated by Lemma E.2. Therefore, we have for
all T > ,
 
,N
|T4 (s)| ≤ t |s−1| FT (t)dt,
0
 
1
|L 3 | ≤ FT (t)dt,
0 2t
and the right hand side converges exponentially fast to zero as  → 0. An analogous
formula holds for the error if elements in the length spectrum up to length L are taken
into account.
Of course the short geodesics dominate the contribution from the length spectrum.
Note however that as a consequence of the collar theorem there are at most 3g − 3 closed
geodesics of length smaller that 2 arcsinh(1) ≈ 1.7627 on a surface of genus g > 1 (see
e.g. [Bus92]).

7. Spectral Determinants and Casimir Energy for Special Surfaces

In this section we compute the values of the Casimir energy and the spectral determinant
for the three isolated surfaces of genus two with large symmetry group as well as for
other examples of genus two and genus three surfaces that have been treated in the lit-
erature before. All Fenchel-Nielsen coordinates are given in mw-form, i.e. the 3-valent
graph describing this decomposition consists of two vertices that are connected by three
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 853

edges. For all the surfaces we first used the Fortran double precision implementation
of our program to generate a list of eigenvalues. Even though our algorithm allows us
to choose the stepsize in such a way that we are guaranteed not to miss an eigenvalue,
the current implementation does not achieve that automatically. In our computations we
have chosen to check completeness of the list of eigenvalues by using the Selberg trace
formula applied to the heat trace. This is possible as the absolute value of the contribu-
tion of the length spectrum in Selberg’s trace formula is bounded from above by FT (t),
defined in Eq. (19), so that missed eigenvalues can easily be detected comparing the
heat trace computed from the Selberg trace formula. Note that unlike heuristic checks
using Weyl’s law this gives a rigorous way to check completeness up to a certain value.
The data files for all the examples below can be found at http://www-staff.lboro.ac.uk/
~maas3/publications/eigdata/datafile.html.

7.1. The Bolza surface. The Bolza surface has symmetry group S4 × Z2 . Its order,
48, thus maximizes the order of the symmetry group amongst all genus 2 hyperbolic
surfaces. The Bolza surface is referred to as the regular octagon (and sometimes as
the Hadamard-Gutzwiller model) in the physics literature as it can be obtained from a
regular octagon in hyperbolic space by identifying opposite sides. Its Fenchel-Nielsen
coordinates are
(1 , t1 ; 2 , t2 ; 3 , t3 )
√ 1 √ √
= (2 arccosh(3 + 2 2), ; 2 arccosh(1 + 2), 0; 2 arccosh(1 + 2), 0).
2
We computed the spectral determinant as well as the Casimir energy ζ (−1/2) and
obtained
det ζ () ≈ 4.72273280444557,
ζ (−1/2) ≈ −0.65000636917383,
where we believe that all decimal places are correct (see also Fig. 7 for other ζ values).
This accuracy was obtained by computing the first 10 eigenvalues with a 30 digit preci-
sion, the first 500 eigenvalues in quad precision and subsequently by using the simple
length spectrum up to  = 9. We relied on the multiplicities of the length spectrum as
computed in [AS88]. Note that the spectral determinant and the Casimir energy can both
be computed accurately within 5 decimal places without taking the length spectrum into
account at all, using the eigenvalues only. Completeness of the list of the first 500 eigen-
values was checked using a list of 700 computed eigenvalues. To compute the first 500
eigenvalues of the Bolza surface to a precision of 12 digits about 10000 λ-evaluations
of generalized singular value decomposition were needed. This took about 10 minutes
on a 2.5 GHz Intel Core i5 quad core processor (where parallelization was used).
We conjecture that the spectral determinant is maximized in genus 2 for the Bolza
surface and we therefore believe that the above number is important. Our computations
show that Bolza surface as a point in Teichmüller space is indeed a local maximum.
Note that the Bolza surface is known to be a critical point for the spectral determinant.

7.2. The surface with symmetry group D6 × Z2 . This surface has Fenchel-Nielsen coor-
dinates (see [BS05], Sect. 3.5)
(1 , t1 ; 2 , t2 ; 3 , t3 ) = (2 arccosh(2), 0; 2 arccosh(2), 0; 2 arccosh(2), 0).
854 A. Strohmaier, V. Uski

Fig. 7. ζ (s) as a function of s for the Bolza surface

It is also known to be a critical point for any modular invariant such as the Casimir
energy or the spectral determinant. Computing the first 500 eigenvalues with double
precision and neglecting the contribution from the length spectrum other than from the
12 primitive closed geodesics of length 2 arccosh(2) we obtain
det ζ () ≈ 4.428000668,
ζ (−1/2) ≈ −0.67250924.

7.3. The surface with symmetry group Z5 × Z2 . This surface has Fenchel-Nielsen coor-
dinates (see [BS05], Sect. 3.5)
(1 , t1 ; 2 , t2 ; 3 , t3 )
√ √ √
3+ 5 1+ 5 1 9+3 5 1
= (2 arccosh , 0; 2 arccosh , ; 2 arccosh , .
2 2 2 4 2
It is also known to be a critical point for any modular invariant such as the Casimir energy
or the spectral determinant. Computing the first 500 eigenvalues with double precision
and neglecting the contribution from the length spectrum we obtain
det ζ () ≈ 4.358289,
ζ (−1/2) ≈ −0.67914508.

7.4. The Gutzwiller octagon. This is a surface obtained from a regular hyperbolic octa-
gon in the upper half space using a side identification different from that for the Bolza
surface (see [Nin95] where also the low lying spectrum of this surface was computed).
The Fenchel-Nielsen coordinates of this surface are given by
(1 , t1 ; 2 , t2 ; 3 , t3 )
√ √ √
2+1 1 2+1 1 2+1 1
= (4 arccosh √ , ; 2 arccosh √ , ; 2 arccosh √ , ).
2 4 2 2 2 2
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 855

Computing the first 500 eigenvalues with double precision and neglecting the con-
tribution from the length spectrum we obtain
det ζ () ≈ 3.76048,
ζ (−1/2) ≈ −0.72747.

7.5. The example of Aurich and Steiner. Eigenvalue statistics of a certain generic octa-
gon was investigated in [AS93]. The Fenchel-Nielsen parameters of this surface are
roughly (see [ABCKS]),5
(1 , t1 ) ≈ (3.717414183638, 0.0304758025243),
(2 , t2 ) ≈ (3.303402988815, 0.2711895405183),
(3 , t3 ) ≈ (3.408324727953, 0.2187479424048).
Computing the first 500 eigenvalues with double precision and neglecting the contribu-
tion from the length spectrum we obtain
det ζ () ≈ 3.959168,
ζ (−1/2) ≈ −0.715195.

7.6. Examples of Rocha and Pollicott. These are the examples treated in [PR97], where
approximate values for their spectral determinants are given.
The first example has Fenchel-Nielsen coordinates
(1 , t1 ; 2 , t2 ; 3 , t3 ) = (5.05, 0; 1, 0; 0.9, 0).
Computing the first 500 eigenvalues with double precision and taking into acount only
the short primitive geodesics in the length spectrum we obtain
det ζ () ≈ 0.395833,
ζ (−1/2) ≈ −1.817507.
The second example has Fenchel-Nielsen coordinates
(1 , t1 ; 2 , t2 ; 3 , t3 ) = (0.98, 0; 3.5, 0; 0.98, 0).
Computing the first 500 eigenvalues with double precision and taking into acount only
the short primitive geodesics in the length spectrum we obtain
det ζ () ≈ 0.6114618,
ζ (−1/2) ≈ −1.6541313.
The third example has Fenchel-Nielsen coordinates
(1 , t1 ; 2 , t2 ; 3 , t3 ) = (5, 0; 1, 0; 1, 0).
Computing the first 500 eigenvalues with double precision and taking into acount only
the short primitive geodesics in the length spectrum we obtain
det ζ () ≈ 0.5124672,
ζ (−1/2) ≈ −1.6591527.
The values do not agree with the values obtained in [PR97]. We believe that not
enough lengths of primitive geodesics were taken into account there.
5 See Acknowledgements.
856 A. Strohmaier, V. Uski

7.7. An example of genus three. The following example has genus three and is a double
cover of the Bolza surface. It is obtained by cutting the√two copies of the Bolza surface
open along the two geodesics of length 2 arccosh(1 + 2) used in its pants decomposi-
tion (see Sect. 7.2) The two resulting surfaces of type (0, 4) are then glued by identifying
each boundary with the corresponding boundary on the other copy of the surface. We
computed the first 600 eigenvalues on this surface and obtained

det ζ () ≈ 9.6507,

ζ (−1/2) ≈ −0.7802.

For this accuracy the contribution of the length spectrum can be neglected.

Acknowledgements. We would like to thank Peter Buser for very interesting discussions and also for helping
us to compute the Fenchel Nielsen coordinates of surfaces given in various parametrizations. We are very
grateful to Ralf Aurich for providing the eigenvalues of surfaces for testing purposes. Our thanks also go to
Alex Barnett, Frank Steiner and Andreas Strömbergsson for useful comments and stimulating discussions.

Appendix A. Estimates on the Basis Functions

This section deals with growth estimates and bounds on the function (ρ) that satisfies
the differential equation
 
1 d d A
− cosh ρ + − λ (ρ) = 0 (20)
cosh ρ dρ dρ cosh2 ρ
with initial conditions
d
(0) = a, (0) = b.
dρ
Substituting ϕ = 2 arctan(tanh ρ/2) this can also be written as

d2 λ
(− + A− )(ϕ) = 0,
dϕ 2 cos2 ϕ

where (ϕ) = (ρ(ϕ)) satisfies (0) = a and dρ (0)

d
= b. The derivatives in
different coordinates are related by cosh ρ dρ
d
= d
dϕ . Now define a potential

λ
V (ϕ) = A − .
cos2 ϕ
Lemma A.1. Suppose that λ > 0 and V (ϕ) does not vanish on an interval I ⊂ [0, π/2).
Then the function E(ϕ) = |(ϕ)|2 − V (ϕ)
1
| (ϕ)|2 is non-increasing on I .

Proof. Simple differentiation gives

1 2λ sin ϕ
E (ϕ) = − | (ϕ)|2 ≤ 0,
(V (ϕ)) cos3 ϕ
2

which implies the statement. 


Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 857

Lemma A.2. Suppose that λ > 0 and V (ϕ) ≥ 0 on some interval [ϕ0 , ϕ1 ] ⊂ [0, π/2)
and assume (ϕ0 ) ≥ 0 and  (ϕ0 ) ≥ 0. Define
 ϕ
cϕ0 (ϕ) = V (t)dt.
ϕ0

Then for all ϕ ∈ [ϕ0 , ϕ1 ] we have

  1  
(ϕ) ≥ (ϕ0 ) cosh cϕ0 (ϕ) + √  (ϕ0 ) sinh cϕ0 (ϕ) ,
V (ϕ0 )
and
   V (ϕ)  
 (ϕ) ≥ V (ϕ)(ϕ0 ) sinh cϕ0 (ϕ) +  (ϕ0 ) cosh cϕ0 (ϕ) .
V (ϕ0 )

Proof. Since every solution can be obtained as a sum of one with (ϕ0 ) = 0 and one
with  (ϕ0 ) = 0, we can assume without loss of generality that either (ϕ0 ) = 0 or
 (ϕ0 ) = 0. We also assume that (ϕ) is non-zero. Define
  1  
h(ϕ) := (ϕ0 ) cosh cϕ0 (ϕ) + √  (ϕ0 ) sinh cϕ0 (ϕ) .
V (ϕ0 )
Then h(ϕ) and (ϕ) have the same initial data at ϕ0 . It follows from the differential
equation and the assumption (ϕ0 ) ≥ 0 and  (ϕ0 ) ≥ 0 that (ϕ),  (ϕ) and  (ϕ)
are positive on some interval (ϕ0 , δ). Moreover, where (ϕ) > 0 and  (ϕ) > 0 the
function (ϕ) is convex and therefore, (ϕ),  (ϕ) and  (ϕ) are non-decreasing
functions on [ϕ0 , ϕ1 ]. In particular (ϕ) > 0 on (ϕ0 , ϕ1 ]. The differential equation
implies
h (ϕ)
( (ϕ)h(ϕ) − h (ϕ)(ϕ)) = (V (ϕ) − )(ϕ)h(ϕ).
h(ϕ)
The right-hand side of this is non-negative as one can easily see by direct calculation
that
h (ϕ) tan(ϕ)
(V (ϕ) − ) = λ√ T ((cϕ0 (ϕ)) ≥ 0,
h(ϕ) V (ϕ) cos2 (ϕ)
where T (ϕ) is either tanh(ϕ) or coth(ϕ) depending on the inital conditions. Thus,
(ϕ)
(ln ) = ( (ϕ)h(ϕ) − h (ϕ)(ϕ)) ≥ 0
h(ϕ)
on (ϕ0 , ϕ1 ] and consequently (ϕ) ≥ h(ϕ). The equation also implies that (ϕ) − h(ϕ)
is non-decreasing which gives the inequality for the derivatives. 


The following lemma is a direct consequence of Gronwall’s inequality applied to the

first order system
    
d (ϕ) 0 c (ϕ)
= .
dϕ c  (ϕ)
1 V (ϕ)/c 0 c  (ϕ)
1
858 A. Strohmaier, V. Uski

Lemma A.3. Suppose that λ > 0, c > 0 and |V (ϕ)| ≤ c2 on some interval [ϕ0 , ϕ1 ] ⊂
[0, π/2). Then for all ϕ ∈ [ϕ0 , ϕ1 ] we have
1 1 1 1
(|(ϕ)|2 + | (ϕ)|2 ) 2 ≤ ec|ϕ−ϕ0 | (|(ϕ0 )|2 + 2 | (ϕ0 )|2 ) 2 .
c2 c
For intervals (ϕ0 , ϕ1 ) where V (ϕ) is positive this can be slightly refined.
Lemma A.4. Suppose that λ > 0 and V (ϕ) > 0 on some interval (ϕ0 , ϕ1 ) ⊂ [0, π/2).
Then for all ϕ ∈ (ϕ0 , ϕ1 ) we have
1 1 V (ϕ0 ) cϕ (ϕ) 1 1
(|(ϕ)|2 + | (ϕ)|2 ) 2 ≤ e 0 (|(ϕ0 )|2 + | (ϕ0 )|2 ) 2 .
V (ϕ) V (ϕ) V (ϕ0 )
Proof. The equation is equivalent to the system
   √  
d (ϕ) 0 V (ϕ) (ϕ)
= √ V (ϕ) √ 1  (ϕ) .
dϕ √V (ϕ)  (ϕ)
1
V (ϕ) − 2V (ϕ) V (ϕ)

The operator norm of the family of matrices in this system at the point ϕ is given by

V (ϕ) V (ϕ) 2
− + ( ) + V (ϕ)
2V (ϕ) 2V (ϕ)
which is bounded from below by
V (ϕ) 
| | + V (ϕ)
V (ϕ)
given that V (ϕ) ≤ 0. The lemma now follows immediately from Gronwall’s inequality.



Appendix B. Estimates for the Resolvent on the Upper Half Space

Let H be the Laplace operator on the upper half space equipped with the Poincare
metric. Then the integral kernel ks (z, w) of the resolvent
Rs(1−s) (H ) = (H − s(1 − s))−1
is given by
ks (z, w) = Q −s (cosh dist(z, w)).
Here Q s is the Legendre Q-function. Using the point pair invariant u(z, w) =
cosh dist(z,w)−1 |z−w|2
2 = 4(z)(w) , this can be expressed as
1
ks (z, w) = Q −s (1 + 2u(z, w)) = Fs (u(z, w)),
2π
where Fs (u) is given by an integral (see e.g. [Iwa02])
 1
1 1
Fs (u) := Q −s (cosh ρ) = (ξ(1 − ξ ))s−1 (ξ + u)−s dξ.
2π 4π 0
This formula can be used to construct a meromorphic continuation of the resolvent as a
function of s to the whole complex plane. We will need the following estimates on this
function all of which follow right from the above integral formula.
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 859

Lemma B.1. Suppose that σ = Re(s) > 0. Then

1 ( (σ ))2 −σ
|Fs (u)| ≤ u ,
4π (2σ )
d |s| ( (σ ))2 −σ −1
| Fs (u)| ≤ u .
du 4π (2σ )
Lemma B.2. Suppose that σ = Re(s) > 1. Then
1 1+u
|Fs (u)| ≤ log ,
4π u
d |s| 1
| Fs (u)| ≤ .
du 4π u(1 + u)
A similar estimate as in the above lemma also holds for Re(s) > 0 (c.f. [Iwa02],
Lemma 1.7, where however (1.48) seems to be inaccurate).
Lemma B.3. Suppose that σ = Re(s) > 0. Then, there are constants C1 (σ ), C3 (σ ),
depending only on σ , and a constant C2 (s), depending only on s, such that
1 1+u
|Fs (u)| ≤ log + C1 (σ ),
4π u
d C2 (s) 1
| Fs (u)| ≤ + C3 (σ ).
du 4π u
Moreover, C2 (s) can be bounded by C4 (σ )|s|, where C4 (σ ) depends only on σ .
In the important case σ = 1/2 one can choose C1 (1/2) = 1, C2 (1/2) = 2 and
C3 (1/2) = 1, although these choices are not optimal. The statement that C2 (s) can be
bounded by C4 (σ )|s| is not optimal either for large values of (s), where cancellations
determine the asymptotic behaviour as u → 0.
Theorem B.4. Let be a geodesic segment in the upper half space and let (ρ, t) be
Fermi coordinates with respect to the infinite extension of . Suppose a distribution h 1
is defined as

h1( f ) = F1 (t) f (0, t)dt,

where F1 ∈ L 2 (R). Then

(H + 3/2)−1 h 1 L 2 ( H) ≤ C1 F1 L 2 (R ) ,

where the constant C1 is given by (21) and satisfies C1 ≤ 1.75.

Proof. Since the Laplace operator commutes with the S L(2, R)-action we can assume
without loss of generality that the geodesic segment is part of the imaginary axis. The
function (H + 3/2)−1 h 1 can be computed by convolving h 1 with the integral kernel
ks (z, w) for s = 3/2. We use Fermi coordinates (ρ, t) so that the integral kernel is given
by ks (ρ, t; ρ , t ) = ks (ρ, t − t ; ρ , 0). Thus,

((H + 3/2)−1 h 1 )(ρ, t) = F1 (t )ks (ρ, t − t , 0, 0)dt .
860 A. Strohmaier, V. Uski

Since this is a convolution the generalized Young inequality applies so that

  2
|((H + 3/2)−1 h 1 )(ρ, t)|2 dt ≤ F1 2L 2 |ks (ρ, t, 0, 0)|dt .

The value of the point pair invariant u(z, w) at w = i in these coordinates is given by
1
u(ρ, t, 0, 0) = (cosh t cosh ρ − 1) ,
2
so that the inequality follows with
  1
1 2
C1 = ( |Fs ( (cosh t cosh ρ − 1))|dt) cosh ρdρ .
2
(21)
2
The integral over t can be estimated for cosh ρ > 3 as

1
|Fs ( (cosh t cosh ρ − 1))|dt
2
 ∞
1 ( (3/2))2 3/2 3
≤ (2) (cosh ρ)−3/2 (e|t| /2 − 1/3)− 2 dt
4π (3) 0
   
3 3 √ 2 1
= 4 2 − π − 2 arcsin + 2 arctan √ (cosh ρ)−3/2 .
8 2 3 2

For cosh ρ < 3 we split the integral over t into two parts. One with cosh t > 3 and one
with cosh t < 3. The former gives

1
|Fs ( (cosh t cosh ρ − 1))|dt
cosh t>3 2

1 ( (3/2))2 3/2 3
≤ 2 (cosh t − 1)− 2 dt
4π (3)
 cosh t>3 
1 √ 1 √
= 2 + log tanh log(3 + 2 2)
16 4
and the latter

1
|Fs ( (cosh t cosh ρ − 1))|dt
2
cosh t<3

1 cosh t + 1
≤ log dt
4π cosh t<3 cosh t − 1
 arccosh3
1 t2 + 4
≤ log 2 dt
2π 0 t
   
4
arccosh(3) log 1 + arccosh(3) 2 + 4 arctan 21 arccosh(3)
= .
2π
Squaring and integrating finally gives

1  √  √
C1 ≤ 3 − 2 2 α12 + 2(α2 )2 ≈ 1.7485475,
8
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 861

where
 √  √
α1 = 3π + 6 arccot 2 2 − 12 2,
   
√ 1 √ 
α2 = 2 + log tanh log 3 + 2 2
4
    
4
8 arccosh(3) log 1 + arccosh(3) 2 + 4 arctan 21 arccosh(3)
+ .
π



Theorem B.5. Let be a geodesic segment in the upper half space; suppose (ρ, t) are
Fermi coordinates with respect to the infinite extension of . Suppose a distribution h 2
is defined as
   
∂
h2( f ) = − F2 (t) f (0, t)dt.
0 ∂ρ
Then

(H + 1)−1 h 2 2
L 2 ( H)
≤ C2 F2 L 2 (0,) ,

where the constant C2 is given by (22) and satisfies C2 ≤ 1.61.

Proof. As in the proof of the previous theorem we can assume without loss of generality
that the geodesic segment is part of the imaginary axis. As before we use Fermi coordi-
nates (ρ, t) so that the integral kernel is given by ks (ρ, t; ρ , t ) = ks (ρ, t − t ; ρ , 0).
Thus,

 
(H + 3/2)−1 h 1 )(ρ, t) = F1 (t ) −∂ρ ks (ρ, t − t , ρ , 0) |ρ =0 dt .

Now by the chain rule

 
−∂ρ ks (ρ, t − t , ρ , 0) |ρ =0
 
∂u sinh ρ
= − Fs (u) (ρ, t − t , 0, 0) = Fs (u(ρ, t − t , 0, 0)).
∂ρ 2
Again, by the generalized Young inequality
    2
−1 sinh ρ 2
|((H + 3/2) h 2 )(ρ, t)| dt = F1 L 2
2 2
|Fs (u(ρ, t, 0, 0))|dt .
2
Thus,

(H + 1)−1 h 2 2
L 2 ( H)
≤ C2 F2 L 2 (0,) ,

where
  2  2 1
sinh ρ 2
C2 = |Fs (u(ρ, t, 0, 0))|dt cosh ρdρ . (22)
2
862 A. Strohmaier, V. Uski

Splitting the integral into a region cosh ρ > 3 and cosh ρ < 3 and using the estimates
B.1 and B.2 one obtains
√  1/2
3 √  √ √ 2  √
C2 ≤ 27−16 2 5 2−6π +12 arctan 2 +48 2 2+arccosh(3) .
16

The numerical value of the right hand side of this inequality is about 1.6086. 


Both constants C1 and C2 can be computed using numerical integration and their
values are about

C1 ≈ 0.313, C2 ≈ 0.343,

where we believe that at least two digits are correct. The corresponding estimates are
non-rigorous however.

Appendix C. Estimates on Derivatives of the Resolvent

We will need the following bounds on the derivative of ks with respect to s.

Lemma C.1. Let Fs (u) be the function defined in the previous section. Then, for any
σ > 0 we have the estimates
∂
| Fs (u)| ≤ C̃1 (σ ) log(1 + u)u −σ + C̃2 (σ )u −σ ,
∂s
∂2  
| Fs (u)| ≤ C̃2 (σ )u −σ −1 + |s| C̃1 (σ ) log(1 + u) + C̃2 (σ ) u −σ −1 ,
∂u∂s
∂3  
| 2 Fs (u)| ≤ |2s + 1|C̃1 (σ )u −σ −2 + |s(s + 1)| C̃1 (σ ) log(1 + u) + C̃2 (σ ) u −σ −2 ,
∂u ∂s
where
 1
1
C̃1 (σ ) = (ξ(1 − ξ ))σ −1 dξ,
4π 0
 1
1
C̃2 (σ ) = − log(ξ(1 − ξ ))(ξ(1 − ξ ))σ −1 dξ,
4π 0

log(4)
and s = σ + ir . In particular C̃1 (1/2) = 1
4 and C̃2 (1/2) = 2 .

Proof. Simple differentiation shows that with s = σ + ir :

 1
∂ 1 ξ(1 − ξ )
| Fs (u)| = | log (ξ(1 − ξ ))s−1 (ξ + u)−s dξ |
∂s 4π 0 ξ +u
 1
1
≤− log(ξ(1 − ξ ))(ξ(1 − ξ ))σ −1 dξ u −σ
4π 0
 1
1
+ (ξ(1 − ξ ))σ −1 dξ u −σ log(1 + u),
4π 0
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 863

which yields the first estimate. The second estimate and the third estimate are obtained
in a similar manner from
 1
∂2 1
| Fs (u)| ≤ (ξ(1 − ξ ))σ −1 (ξ + u)−σ −1 dξ
∂u∂s 4π 0
  
|s| 1 ξ(1 − ξ )
− log (ξ(1 − ξ ))σ −1 (ξ + u)−σ −1 dξ,
4π 0 ξ +u
and

∂3 |2s + 1| 1
| Fs (u)| ≤ (ξ(1 − ξ ))σ −2 (ξ + u)−σ −1 dξ
∂u 2 ∂s 4π 0
  
|s(s + 1)| 1 ξ(1 − ξ )
− log (ξ(1 − ξ ))σ −1 (ξ + u)−σ −2 dξ.
4π 0 ξ + u



Near zero one obtains similarly:

Lemma C.2. Let Fs (u) be the function defined in the previous section. Then, for any
σ > 0 we have the estimates
∂
| Fs (u)| ≤ C̃3 (σ ),
∂s
∂2
|u Fs (u)| ≤ C̃4 (σ ) + C̃3 (σ )(r 2 + σ 2 )1/2 ,
∂u∂s
∂3
|u 2 2 Fs (u)| ≤ |2s + 1|C̃4 (σ ) + C̃3 (σ )|s(s + 1)|,
∂u ∂s
where s = σ + ir .

In the important case σ = 1/2 one obtains

π
C̃3 (1/2) = ,
4
 1
1 u 1
C̃4 (1/2) = sup √ dξ.
u>0 4π 0 (ξ + u) 3/2 ξ(1 − ξ )
The latter constant can numerically be calculated and is approximately given by

C̃4 (1/2) ≈ 0.167878.

Appendix D. Resolvent Estimates on Hyperbolic Cylinders

Let Z be the hyperbolic cylinder obtained by factorizing
 /2 theupper half space by the
e 0
subgroup of S L(2, R) generated by the element . We use Fermi-coordi-
0 e−/2
nates (ρ, t) which identifies the cylinder with R × R/(Z). The integral kernel of the
resolvent of the Laplace operator  Z can be constructed from the resolvent of H by
averaging over the group: Let x = (ρ, t) and x = (ρ , t ) be distinct points on the
864 A. Strohmaier, V. Uski

hyperbolic cylinder and denote by ks (x, x ) the integral kernel of (H − s(1 − s))−1 .
Then the integral kernel ksZ (x, x ) of ( Z − s(1 − s))−1 is given by
 
ksZ ((ρ, t), (ρ , t )) = ks ((ρ, t), (ρ , t + n)) = ks ((ρ, t − t ), (ρ , n)).
n∈Z n∈Z

The sum converges absolutely whenever (s) > 0 because of the estimate in Lemma
B.1 and the fact that the distance between the points (ρ, t) and (ρ , t + n) grows like
n as |n| → ∞.
Combining the estimates in Lemma C.1 with Lemma B.3 we obtain
Lemma D.1. For any compact subset M ⊂ Z we have on M × M the following esti-
mates:
1
|ksZ (x, x )| ≤ | log ρ(x, x )| + C1M (σ ),
2π
C M (s) 1
|grad x ksZ (x, x )| ≤ 2 + C3M (s),
2π ρ(x, x )

where ρ(x, x ) = 2 arcsinh(u(x, x )1/2 ) is the distance between x and x and σ = (s).
Similarly Lemma C.1 and Lemma C.2 imply the following estimates
Lemma D.2. Let Z L ⊂ Z be a truncated hyperbolic cylinder and let d > 0. Then for
all points x ∈ Z L and x ∈ ∂ Z L with ρ(x, x ) ≥ d the following estimates hold:
∂ Z
| k (x, x )| ≤ C1M,d (σ ),
∂s s
∂
| grad x ksZ (x, x )| ≤ C2M,d (s),
∂s
∂
| grad x grad x ksZ (x, x )| ≤ C3M,d (s).
∂s
Suppose now that M ⊂ Z is a subset of the hyperbolic cylinder Z whose boundary
∂ M is a finite union of geodesic segments. If ψ is a function on M, smooth up to the
boundary, such that

( Z − s(1 − s))ψ = 0

in the interior of M. Let φ1 and φ2 be the restriction of ψ to the boundary and its outward
normal derivative respectively. Of course, as functions on Z we have

( Z − s(1 − s))(χ M ψ) = h

in the sense of distributions, where h is the distribution defined as


h( f ) := (∂n f )(x)φ1 (x) + f (x)φ2 (x)d x.
∂M

From this it follows that ψ in the interior of M can be expressed as

ψ = ( Z − s(1 − s))−1 h.
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 865

This is sometimes written in integral form


ψ(x) = ∂n,x ksZ (x, x )φ1 (x ) + ksZ (x, x )φ2 (x )d x .
∂M
Using, Lemma D.1 and applying the generalized Young inequality we obtain the fol-
lowing estimate:
 1/2
ψ L 2 (M) ≤ C M (s) φ1 2L 2 (∂ M) + φ2 2L 2 (∂ M) . (23)

Appendix E. Appendix-Estimates on the Counting Function

Let H be the Heaviside step function defined by
⎧
⎨ 1 y>0
H(y) = 1/2 y = 0 .
⎩ 0 y<0
Suppose that X is a compact connected oriented hyperbolic surface and let (φi )i∈N0 be
an orthonormal basis of eigenfunctions for the Laplace operator on X . Let
0 = λ0 < λ 1 ≤ λ 2 ≤ · · ·
be the corresponding sequence of eigenvalues. Let e(x, y, τ ) be the integral kernel of
the operator sign(τ )H(τ 2 − ) and denote by N x (τ ) = e(x, x, τ ) its restriction to the
diagonal. Of course, by the spectral theorem,

N x (τ ) = sign(τ ) H(τ 2 − λi )|φi (x)|2 .
i

Integrating this over X yields the eigenvalue counting function N (τ ):


N (τ ) = sign(τ ) H(τ 2 − λi ).
i

By construction N x and N are non-decreasing odd functions on R. Let d(x) be twice the
injectivity radius at the point x ∈ X . If L is the length of a systole we have the estimate
r (x) ≤ L. Using Selberg’s pretrace trace formula and finite propagation speed one can
easily see that the cosine transform of N x (r ) coincides on the interval (−d(x), d(x))
with the cosine transform of the function F (r ), where F(t) is defined by
F(0) = 0,
1
F (t) = 2 H(t 2 − 1/4)|t| tanh(π t 2 − 1/4).
4π
Note that since
 ∞ 1
2t (tanh(π t 2 − 1/4) − 1)dt = −
1/2 12

and tanh(π t 2 − 1/4) − 1 is non-positive on [1/2, ∞) we obtain the estimate
1 1
− ≤ sign(t)G(t) ≤ 0,
3 4π
866 A. Strohmaier, V. Uski

where G(t) = F(t) − sign(t) 4π t . Let ν be the first Dirichlet eigenvalue of the operator
1 2

d4
dx4
on the interval [−1/2, 1/2] and let φ(x) be the corresponding normalized eigenfunc-
tion. We think of φ as a function on R by extending it by zero. It can be easily worked out
that ν is the smallest non-zero solution to the equation cosh(λ) cos(λ) = 1 for λ > 0. Its
numerical value is ν ≈ 4.73004074. The estimate ν < 5 is an immediate consequence
of the intermediate value theorem. As a test function in the Fourier Tauberian theorem
we use the function ρ = (φ̂)2 . Let ρδ and ρδ,0 be defined as in [Saf01] as

ρδ (r ) = δρ(δr ),
ρδ,0 (r ) = δρ1,0 (δr ),

where
 ∞
ρ1,0 (r ) = tρ(t)dt.
r

Using ρδ,0 (r ) > 0 and ρδ L1 = 1 we obtain

1
(ρδ ∗ G)(r ) ≤ G ∞ = , (24)
12π
(ρδ,0 ∗ G )(r ) ≤ 0. (25)

The Fourier Tauberian Theorem 1.3 in [Saf01] together with the estimates Eqs. (2.9)
and (2.10) in [Saf01] and our estimates (24) and (25) imply the following bounds of N .

Theorem E.1. The counting function on a hyperbolic surface satisfies.

   
1 4ν 2 + 2νπ ν 1
N x (r ) ≤ 2
r + r+ + ,
4π π d(x) d(x) 3
   
1 4ν 2 ν 1
N x (r ) ≥ r2 − r+ − .
4π π d(x) d(x) 3
√
Define the function Ñ x (λ) = H(λ)N x ( |λ|). The local heat kernel trace kt (x) for
t > 0 may be defined as follows

kt (x) = e−λ j t |φi (x)|2 .
λ j ≥0

If this sum is cut off at a point c > 0 the remainder is given by


Rtc (x) = e−λ j t |φi (x)|2 .
λ j ≥c

Then of course integration by parts gives the following formula

 ∞
−ct
Rt (x) = − lim+ Ñ x (c + )e
c
+t Ñ x (λ)e−λt dλ.
→0 c

Combining this with the estimate from Theorem E.1 we obtain

Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 867

Lemma E.2. The remainder Rtc (x) satisfies the estimate

 
√ 1
4π Rtc (x) ≤ e−ct Bu (x) + Bl (x) + (Au (x) + Al (x)) c +
t

π √
+Au (x) (1 − erf( ct)),
4t
where
4ν 2 + 2νπ 4ν 3 + 2ν 2 π 1
Au (x) = , Bu (x) = + ,
π d(x) π d(x)2 3
4ν 2 4ν 3 1
Al (x) = , Bl (x) = + .
π d(x) π d(x)2 3
Setting c = 0 one obtains a bound for the local heat kernel. The heat trace

tr(e−t ) = e−λ j t
λ j ≥0

is obtained by integrating the local heat trace. Lemma E.2 therefore yields bounds on
the heat trace in case c = 0 and on
 
−λ j t
e = Rtc (x)d x
λ j ≥c X

in general.
Theorem E.1 also immediately gives bounds on the eigenfunctions
Corollary E.3. Let ψ(x) be an eigenfunction of the Laplace operator with eigenvalue
λ such that ψ L 2 = 1. Then,
 
1 8ν 2 + 2νπ 1/2 ν 2
|ψ(x)|2 ≤ (λ + )+ .
4π π d(x) d(x) 3
For numerical purposes we will use the bound

11 1/2 5 2
|ψ(x)| ≤ (λ + ) + ,
2L L 3
where L is twice the radius of injectivity (which equals to the length of the shortest
closed geodesic).
Similarly, one gets a bound on the derivative of the eigenfunctions in the following
way. Fix a point x ∈ X and a unit vector n. As above we can define

N x1 (τ ) = sign(τ ) H(τ 2 − λi )|nφi (x)|2 .
i

Using the pre-trace formula and finite propagation speed it can easily worked out that
the cosine transform of (N x1 ) (r ) coincides on the interval (−d(x), d(x)) with the cosine
transform of the function F̃ (r ), where F̃(t) is defined by
F̃(0) = 0,
1
F̃ (t) = H(t 2 − 1/4)|t|3 tanh(π t 2 − 1/4).
4π
868 A. Strohmaier, V. Uski

Using
 ∞ 17
t 3 (tanh(π t 2 − 1/4) − 1)dt = − ,
1/2 960
we obtain in the same way as before
1 1
− ≤ sign(t)G̃(t) ≤ 0,
30 4π
where G̃(t) = F̃(t) − sign(t) 16π
1 4
t .
Again the Fourier Tauberian theorem in [Saf01] implies
Theorem E.4. The function N x1 satisfies
 3
1 4 2ν̃ 2 + π ν̃ ν̃ 1 1
N x1 (r ) ≤ r + r+ + ,
16π 4π 2 d(x) d(x) 8π 15
 3
1 4 ν̃ 2 ν̃ 1 1
N x1 (r ) ≥ r − 2
r+ − ,
16π 2π d(x) d(x) 8π 15
√
where ν̃ is the same as ν3 in [Saf01] and satisfies the esimtate ν3 ≤ 6 6 3 ≤ 8. This
results in the bound for the derivative of the eigenvalues.
Corollary E.5. Let ψ(x) be an eigenfunction of the Laplace operator with eigenvalue
λ such that ψ L 2 = 1 and let n x be a unit tangent vector at x ∈ X . Then,
 3
4ν̃ 2 + π ν̃ ν̃ 1 1
|nψ(x)| ≤
2
2
r+ + .
4π d(x) d(x) 4π 15
Again, for numerical purposes we will simply use the bound
 
6 8 3 1
|nψ(x)| ≤ r+ + .
L L 190

References
[ABCKS] Aigon-Dupuy, A., Buser, P., Cibils, M., Kuenzle, A.F., Steiner, F.: Hyperbolic octagons and
Teichmueller space in genus 2. J. Math. Phys. 46, 4901–4925 (2005)
[ASS88] Aurich, R., Sieber, M., Steiner, F.: Quantum chaos of the Hadamard-Gutzwiller model. Phys.
Rev. Lett. 61, 483–487 (1988)
[AS88] Aurich, R., Steiner, F.: On the periodic orbits of a strongly chaotic system. Phys. D 32(3), 451–
460 (1988)
[AS89] Aurich, R., Steiner, F.: Periodic-orbit sum rules for the Hadamard-Gutzwiller model. Physica
D 39, 169–193 (1989)
[AS92] Aurich, R., Steiner, F.: From classical periodic orbits to the quantization of chaos. Proc. R. Soc.
London A 437, 693–714 (1992)
[AS93] Aurich, R., Steiner, F.: Statistical properties of highly excited quantum eigenstates of a strongly
chaotic system. Phys. D 64(1–3), 185–214 (1993)
[Bar09] Barnett, A.H.: Perturbative analysis of the method of particular solutions for improved inclusion
of high-lying Dirichlet eigenvalues. SIAM J. Num. Anal. 47(3), 1952–1970 (2009)
[BB10] Barnett, A.H., Betcke, T.: An exponentially convergent nonpolynomial finite element method
for time-harmonic scattering from polygons. SIAM J. Sci. Comput. 32(3), 1417–1441 (2010)
[BH11] Barnett, A.H., Hassell, A.: Boundary quasi-orthogonality and sharp inclusion bounds for large
Dirichlet eigenvalues. SIAM J. Num. Anal. 49(3), 1046–1063 (2011)
Computation of Eigenvalues of Laplace Operation on Hyperbolic Surfaces 869

[Bet07] Betcke, T.: A gsvd formulation of a domain decomposition method for planar eigenvalue prob-
lems. IMA J. Num. Anal. 27(3), 451–478 (2007)
[Bor10] Borthwick, D.: Sharp geometric upper bounds on resonances for surfaces with hyperbolic ends.
Avilable at http://www.mathcs.emory.edu/~davidb/sharpsurf.pdf, 2010
[BS05] Buser, P., Silhol, R.: Some remarks on the uniformizing function in genus 2. Geom. Dedicata
115, 121–133 (2005)
[BSV06] Booker, A.R., Strömbergsson, A., Venkatesh, A.: Effective computation of Maass cusp forms.
Int. Math. Res. Not. 2006, Art. ID 71281, 1–34 (2006)
[BT05] Betcke, T., Trefethen, L.N.: Reviving the method of particular solutions. SIAM Review 47(3),
469–491 (2005)
[Bus92] Buser, P.: Geometry and spectra of compact Riemann surfaces. Boston, MA: Birkhäuser, Boston,
1992
[FHM67] Fox, L., Henrici, P., Moler, C.B.: Approximations and bounds for eigenvalues of elliptic
operators. SIAM J. Numer. Anal. 4, 89–102 (1967)
[Fri86] Fried, D.: Analytic torsion and closed geodesics on hyperbolic manifolds. Invent. Math. 84(3),
523–540 (1986)
[GL96] Golub, G.H., Van Loan, C.F.: Matrix computations. Johns Hopkins Studies in the Mathematical
Sciences, Third Edition, Baltimore, MD: Johns Hopkins University Press, 1996
[Hej92] Hejhal, D.A.: Eigenvalues of the Laplacian for Hecke triangle groups. Mem. Amer. Math.
Soc. 97(469), vi+165 (1992)
[Hej99] Hejhal, D.A.: On eigenfunctions of the Laplacian for Hecke triangle groups. In: Emerging
Applications of Number Theory, D. Hejhal, J. Friedman, M. Gutzwiller, and A. Odlyzko, eds.,
Berlin-Heidleberg-NewYork: Springer-Verlag, 1999, pp. 291–315
[Hin83] Hindmarsh, A.C.: ODEPACK, a systematized collection of ODE solvers, in scientific computing.
R. S. Stepleman et al. (eds.) Amsterdam: North-Holland, 1983, pp. 55–64
[Iwa02] Iwaniec, H.: Spectral methods of automorphic forms. Second ed., Graduate Studies in Mathe-
matics, Vol. 53, Providence, RI: American Mathematical Society, 2002
[Jen84] Jenni, F.: Über den ersten Eigenwert des Laplace-operators auf ausgewählten Beispielen ko-
mpakter Riemannscher Flächen. Comment. Math. Helv. 59(2), 193–203 (1984)
[JLN+ 05] Jakobson, D., Levitin, M., Nadirashvili, N., Nigam, N., Polterovich, I.: How large can the first
eigenvalue be on a surface of genus two? Int. Math. Res. Not. no. 63, 3967–3985 (2005)
[KKK09] Klein, C., Kokotov, A., Korotkin, D.: Extremal properties of the determinant of the Laplacian in
the Bergman metric on the moduli space of genus two Riemann surfaces. Math. Z. 261(1), 73–108
(2009)
[M04] Marklof, J.: Selberg’s trace formula: An Introduction. In: Hyperbolic Geometry and Appli-
cations in Quantum Chaos and Cosmology, eds. J. Bolte, F. Steiner, Cambridge: Cambridge
University Press, 2011, pp. 83-119
[MP68] Moler, C.B., Payne, L.E.: Bounds for eigenvalues and eigenvectors of symmetric opera-
tors. SIAM J. Num. Anal. 5(1), 64–70 (1968)
[Nin95] Ninnemann, H.: Gutzwiller’s octagon and the triangular billiard t ∗ (2, 3, 8) as models for the
quantization of chaotic systems by Selberg’s trace formula. Int. J. Mod. Phys. B 9, 1647–1753
(1995)
[Pet83] Petzold, L.R.: Automatic selection of methods for solving stiff and nonstiff systems of ordinary
differential equations. Siam J. Sci. Stat. Comput. 4 4, 136–148 (1983)
[PR97] Pollicott, M., Rocha, A.C.: A remarkable formula for the determinant of the Laplacian. Invent.
Math. 130(2), 399–414 (1997)
[OPS88] Osgood, B., Phillips, R., Sarnak, P.: Extremals of determinants of Laplacians. J. Funct. Anal.
80(1), 148–211 (1988)
[PT96] Phillips, G.M., Taylor, P.J.: Theory and applications of numerical analysis, Second. ed., London:
Academic Press Ltd., 1996
[Saf01] Safarov, Yu.: Fourier tauberian theorems and applications. J. Funct. Anal. 185(1), 111–128
(2001)
[Sch93] Schmutz, P.: Riemann surfaces with shortest geodesic of maximal length. Geom. Funct. Anal.
3(6), 564–631 (1993)
[Sch94] Schmutz, P.: Systoles on Riemann surfaces. Manus. Math. 85(3-4), 429–447 (1994)
[St87] Steiner, F.: On Selberg’s zeta function for compact Riemann surfaces. Phys. Lett. B 188(4), 447–
454 (1987)
[Tay96] Taylor, M.E.: Partial differential equations. I. Applied Mathematical Sciences, 115, Berlin,
Heidleberg-NewYork: Springer-Verlag, 1996

Communicated by S. Zelditch
Copyright of Communications in Mathematical Physics is the property of Springer Science & Business Media
B.V. and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright
holder's express written permission. However, users may print, download, or email articles for individual use.