Stine Marie Berge

Doctoral theses at NTNU, 2021:285
Doctoral thesis
Stine Marie Berge
Stine Marie Berge

Quantitative Unique
Continuation and Eigenvalue
Bounds for the Laplacian
ISBN 978-82-326-6943-1 (printed ver.)

ISBN 978-82-326-5414-7 (electronic ver.)
ISSN 1503-8181 (printed ver.)
ISSN 2703-8084 (online ver.)
NTNU
Norwegian University of Science and Technology
Thesis for the Degree of
Philosophiae Doctor
Faculty of Information Technology and Electrical
Engineering
Department of Mathematical Sciences
Stine Marie Berge
Quantitative Unique
Thesis for the Degree of Philosophiae Doctor
Trondheim, September 2021

Faculty of Information Technology and Electrical Engineering
NTNU
Thesis for the Degree of Philosophiae Doctor
Faculty of Information Technology and Electrical Engineering

© Stine Marie Berge
ISBN 978-82-326-6943-1 (printed ver.)

Printed by NTNU Grafisk senter

Abstract
In this thesis, we are going to study several aspects of the Laplace operator, es-
pecially related to eigenvalues and eigenfunctions. A large part of the thesis is
dedicated to quantitative unique continuation for harmonic functions and eigen-
functions of the Laplace operator. In particular, we determine the optimal growth
rate with respect to the wave number for the three-ball inequality of the Laplace
operator on Riemannian model spaces. Additionally, we show eigenvalue inequal-
ities for several eigenvalue problems (e.g. Dirichlet, Neumann, Steklov) where the
Laplacian is heavily present. Throughout the thesis, the curvature of the underlying
Riemannian manifold will play a key role.
Sammendrag
I denne avhandlingen skal vi studere flere aspekter ved laplaceoperatoren, spe-
sielt med hensyn på egenverdier og egenfunksjoner. En stor del av avhandlingen
er dedikert til kvantitativ unik utvidelse ulikheter for harmoniske funksjoner og
egenfunksjoner til laplaceoperatoren. For eksempel, vi bestemmer den best mulige
vekstraten med hensyn på bølgetallet for tre-ball ulikheter for laplaceoperatoren
på riemannske modell-mangfoldigheter. I tillegg skal vi vise egenverdi ulikheter
for flere egenverdiproblemer (f.eks. dirichlet-, neumann-, steklovproblemet) der
laplaceoperatoren er i stor grad tilstedeværende. Krumningen til den underliggende
riemannske mangfoldigheten vil spille en sentral rolle fra start til slutt.
i
Preface
This thesis is submitted in partial fulfillment of the requirements for the degree of
Philosophiae Doctor (PhD) in Mathematical Sciences at the Norwegian University
of Science and Technology (NTNU). The research presented here was conducted
at the Department of Mathematical Sciences at NTNU, under the supervision of
Professor Eugenia Malinnikova and Professor Peter Lindqvist.
Structure of the Thesis

Part I of the thesis consists of two chapters: In Chapter 1 we give an overview of
topics related to the papers included in the thesis. Rather than being comprehensive,
we aim for a succinct approach. In Chapter 2 we give summaries for the three papers
included in the thesis, as well as an outline for the further results given in Chapter D.
Part II of the thesis consists of three papers, namely Paper A, Paper B, and Paper C.
Moreover, in Chapter D we present some further work not yet contained in a paper.
Acknowledgements
First and foremost, I would like to express my deepest gratitude to my main
supervisor, Eugenia Malinnikova, for her insightfulness and immense help during
the last four years. I would also like to thank Peter Lindqvist for being my co-
supervisor.
I greatly appreciated the helpfulness and hospitality that Dan Mangoubi showed
me when I visited the Einstein Institute of Mathematics in Jerusalem. I would like
to thank Franz Luef for inviting me into the time-frequency analysis community at
NTNU.
During the last four years, Are Austad and Eirik Skrettingland have made the
PhD experience more enjoyable and I am very grateful for the time I have shared an
office with them. I am also thankful for the many illuminating discussions I have
had with Stefano Decio during the last years. Last but not least, I am very grateful
iii
for the endless support and patience from my husband, as well as his optimism
regarding my work.
Stine Marie Berge

Trondheim, June 2021
iv
Contents
Abstract i
Preface iii
Contents v
I Introduction 1
1 Preliminaries 3
1.1 Riemannian Manifolds . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.1 Riemannian Connection and Curvature . . . . . . . . . . 3
1.1.2 Calculus on Manifolds . . . . . . . . . . . . . . . . . . . 5
1.1.3 Flow of Vector Fields . . . . . . . . . . . . . . . . . . . . 8
1.1.4 Hypersurfaces . . . . . . . . . . . . . . . . . . . . . . . 9
1.1.5 Spherical Coordinates and Model Spaces . . . . . . . . . 11
1.1.6 Comparison Theorems . . . . . . . . . . . . . . . . . . . 14
1.1.7 Spherically Symmetric Manifolds . . . . . . . . . . . . . 15
1.2 The Laplace Operator . . . . . . . . . . . . . . . . . . . . . . . . 15
1.2.1 The Helmholtz Equation . . . . . . . . . . . . . . . . . . 15
1.2.2 Associated Harmonic Function . . . . . . . . . . . . . . . 16
1.2.3 Self-Adjoint Operators . . . . . . . . . . . . . . . . . . . 17
1.2.4 Eigenvalue Problems . . . . . . . . . . . . . . . . . . . . 19
1.2.5 Steklov Problem . . . . . . . . . . . . . . . . . . . . . . 23
1.2.6 Rellich Identity . . . . . . . . . . . . . . . . . . . . . . . 24
1.3 Almgren’s Frequency Function . . . . . . . . . . . . . . . . . . . 26
1.3.1 Geometric Convexity . . . . . . . . . . . . . . . . . . . . 27
1.3.2 Three Spheres Theorem and the Frequency Function . . . 27
1.3.3 Doubling Index . . . . . . . . . . . . . . . . . . . . . . . 31
1.3.4 Quantitative Unique Continuation Results . . . . . . . . . 33
v
1.3.5 Inverse Three Sphere Theorem . . . . . . . . . . . . . . . 35
2 Summary of Papers 37
2.1 Paper A: Convexity Properties of Harmonic Functions . . . . . . . 37
2.2 Paper B: Three Ball Theorem . . . . . . . . . . . . . . . . . . . . 37
2.3 Paper C: Kuttler-Sigillito’s Inequalities . . . . . . . . . . . . . . . 38
2.4 Further Results D: Spherically Symmetric Manifolds . . . . . . . 38
II Research Papers 39
A Convexity Properties of Harmonic Functions 43

A.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
A.2 The Convexity Result . . . . . . . . . . . . . . . . . . . . . . . . 45
A.2.1 Prerequisites . . . . . . . . . . . . . . . . . . . . . . . . 45
A.2.2 The Main Theorem . . . . . . . . . . . . . . . . . . . . . 48
A.2.3 Corollaries . . . . . . . . . . . . . . . . . . . . . . . . . 53
A.2.4 An Inequality of Hörmander . . . . . . . . . . . . . . . . 54
A.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
A.3.1 Geodesic Spheres . . . . . . . . . . . . . . . . . . . . . . 56
A.3.2 1-Homogeneous Functions . . . . . . . . . . . . . . . . . 61
A.3.3 Example of the distance function of 𝑆 𝑘 ⊂ R𝑛 . . . . . . . 66
A.3.4 Non-Positive Eigenvalues of −Δ . . . . . . . . . . . . . . 68
B Three Ball Theorem 73

B.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
B.2 The Three Ball Inequality . . . . . . . . . . . . . . . . . . . . . . 75
B.2.1 Bessel Functions and the Helmholtz Equation in R𝑛 . . . . 75
B.2.2 Solutions of the Helmholtz Equation on Model Spaces . . 78
B.3 The Reverse Three Ball Inequality . . . . . . . . . . . . . . . . . 84
B.4 Appendix: The First Positive Zero of the Bessel Function . . . . . 88
B.5 Appendix: Comparison Theorems for Sturm-Liouville Equations . 88
C Kuttler-Sigillito’s Inequalities and Rellich-Christianson Identity 93

C.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
C.2 Kuttler-Sigillito’s Type Inequalities . . . . . . . . . . . . . . . . . 95
C.2.1 Steklov-Dirichlet and Neumann-Dirichlet Comparison . . 96
C.2.2 Robin-Dirichlet Comparison . . . . . . . . . . . . . . . . 102
C.3 Hadamard Formula for the Mixed Neumann-Dirichlet Problem . . 104
C.4 Rellich Identity . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
C.5 Rellich-Christianson Type Identity . . . . . . . . . . . . . . . . . 108
vi
D Eigenvalues on Spherically Symmetric Manifolds 115
D.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
D.2 Dirichlet Eigenvalues on Model Spaces . . . . . . . . . . . . . . 116
D.3 Eigenvalues on Spherically Symmetric Manifolds . . . . . . . . . 120
D.4 Eigenvalues on Balls . . . . . . . . . . . . . . . . . . . . . . . . 122
D.5 Spherical Symmetric Manifolds . . . . . . . . . . . . . . . . . . 125
Bibliography 129
vii
Part I
Introduction
Chapter 1
Preliminaries
The Laplace operator is arguably the most important and studied differential oper-
ator of all times, only challenged by the derivative. It is renowned for its intimate
relation to several physical problems, its involvement in stochastic processes, and,
my personal favorite, its relationship with Riemannian geometry. In this thesis we
are going to different aspects with the Laplacian. The two main themes we are
going to study are eigenvalue inequalities involving the Laplacian and quantitative
unique continuation result solutions to the Helmholtz equation Δ + 𝑘 2 . We will
start by giving an introduction to these themes.
1.1 Riemannian Manifolds

In this section, we will give an admittedly rather bare bone introduction to Rieman-
nian geometry. The goal of this section is to introduce the reader to the notations
and conventions used later in this text. We will use [Lee18] as our main reference,
although other standard books such as [Pet16; Jos11] would also suffice. To avoid
repetition, the following chapters from [Lee18] are used: Section 1.1.1 is taken
from [Lee18, Chap. 2 and 5–7]; Section 1.1.2 is taken from [Lee18, Chap. 2 and 4];
Section 1.1.4 is mostly taken [Lee18, Chap. 8]; Section 1.1.5 is taken from [Lee18]
and Section 1.1.6 is taken from [Lee18, Chap. 11].
1.1.1 Riemannian Connection and Curvature

Throughout the thesis we will let (𝑀, g) denote a 𝐶 ∞ -smooth complete Riemannian
manifold. In coordinates we can write the Riemannian metric g as 𝑔𝑖 𝑗 𝑑𝑥 𝑖 ⊗ 𝑑𝑥 𝑗 ,
where the matrix (𝑔𝑖 𝑗 ) is a positive definite matrix for each point in 𝑀. For
p given two vectors 𝑢, 𝑣 ∈ 𝑇 𝑝 𝑀 we
simplicity,
∗
will write g(𝑢, 𝑣) = h𝑢, 𝑣i and
|𝑢| = g(𝑢, 𝑢). We will use the notation g for the co-metric and denote the
3
Chapter 1. Preliminaries
components of g∗ in coordinates by 𝑔 𝑖 𝑗 . In coordinates the components of the

co-metric can be computed by the formula (𝑔 𝑖 𝑗 ) = (𝑔𝑖 𝑗 ) −1 .
Using the Riemannian metric g one can define the Riemannian distance function
dist, making (𝑀, dist) into a metric space. Associated with every Riemannian
manifold (𝑀, g) is a unique connection called the Riemannian connection ∇ (also
called Levi-Civita connection). For vector fields 𝑋, 𝑌 , and 𝑍 the Riemannian
connection ∇ is characterized by the following two conditions:
Torsion free
∇𝑋 𝑌 − ∇𝑌 𝑋 = [𝑋, 𝑌 ],
Compatibility of the metric
𝑋 (g(𝑌 , 𝑍)) = g(∇𝑋 𝑌 , 𝑍) + g(𝑌 , ∇𝑋 𝑍).
Arguably, the most important concept in Riemannian geometry is curvature.

The Riemannian curvature tensor is defined by
𝑅(𝑋, 𝑌 )𝑍 = ∇𝑋 ∇𝑌 𝑍 − ∇𝑌 ∇𝑋 𝑍 − ∇ [𝑋 ,𝑌 ] 𝑍,
for vector fields 𝑋, 𝑌 , and 𝑍. The curvature (𝑅(𝑋, 𝑌 )𝑍) 𝑝 at a point 𝑝 ∈ 𝑀 only
depends on 𝑋 𝑝 , 𝑌 𝑝 , and 𝑍 𝑝 . As such, it is possible to unambiguously define
𝑅(𝑢, 𝑣)𝑤 for 𝑢, 𝑣, 𝑤 ∈ 𝑇 𝑝 𝑀. Using the curvature tensor one can define:
Sectional curvature
g(𝑅(𝑢, 𝑣)𝑣, 𝑢)
Sec(𝑢, 𝑣) = , 𝑢 ≠ 𝑐𝑣,
g(𝑢, 𝑢)g(𝑣, 𝑣) − g(𝑢, 𝑣) 2
Ricci curvature
Ric(𝑢, 𝑣) = trg g(𝑅(×, 𝑢)𝑣, ×),
Scalar curvature
S( 𝑝) = trg Ric(×, ×),
for 𝑢, 𝑣 ∈ 𝑇 𝑝 𝑀 and 𝑐 ∈ R. We have used the notation trg to denote the trace of
a tensor with respect to g. The entries that are being traced are marked with the
symbol ×. Notice that the sectional curvature Sec(𝑢, 𝑣) for 𝑢 ≠ 𝑐𝑣 only depends
on the two-dimensional plane Π spanned by the vectors 𝑢, 𝑣 ∈ 𝑇 𝑝 𝑀. If we write
inequalities such as e.g. Sec ≤ 𝐾 for 𝐾 ∈ R, this is a shorthand to way of stating
that Sec(Π) ≤ 𝐾 for all two-dimensional planes Π ⊂ 𝑇 𝑝 𝑀 and all 𝑝 ∈ 𝑀.
4
1.1. Riemannian Manifolds
A geodesic can be defined as a locally distance minimizing curve. For every

𝑣 ∈ 𝑇 𝑝 𝑀 we denote by 𝛾 𝑣 the unique maximal geodesic satisfying 𝛾 𝑣 (0) = 𝑝 and
𝛾¤ 𝑣 (0) = 𝑣. One defines the exponential map exp 𝑝 centered at 𝑝 ∈ 𝑀 by
exp 𝑝 (𝑣) = 𝛾 𝑣 (1), 𝑣 ∈ 𝑇 𝑝 𝑀.
The exponential map is a diffeomorphism from a neighborhood 𝑈 ⊂ 𝑇 𝑝 𝑀 of

0 ∈ 𝑇 𝑝 𝑀 to a neighborhood of 𝑝 ∈ 𝑀. A related quantity is the injectivity radius
of 𝑝 denoted by Inj( 𝑝). The injectivity radius Inj( 𝑝) is the largest number such
that the exponential map is injective on the ball
𝐵Inj( 𝑝) (0) = {𝑣 ∈ 𝑇 𝑝 𝑀 : |𝑣| < Inj( 𝑝)}.
1.1.2 Calculus on Manifolds

We will now extend certain parts of the Euclidean calculus to the Riemannian
setting. Denote by 𝐶 ∞ (𝑀) the space of real-valued smooth functions on 𝑀. For
𝑓 ∈ 𝐶 ∞ (𝑀) we can take the exterior differential of 𝑓 , denoted by 𝑑𝑓 . The gradient
of 𝑓 , denoted by grad 𝑓 , is defined by
g(grad 𝑓 , 𝑣) = 𝑑𝑓 (𝑣), 𝑣 ∈ 𝑇 𝑝 𝑀.
Using the product rule for the exterior differential we get that for 𝑓 , 𝑔 ∈ 𝐶 ∞ (𝑀)
that
grad( 𝑓 𝑔) = 𝑓 grad 𝑔 + 𝑔 grad 𝑓 .
Written in coordinates the gradient is given by
𝜕𝑓 𝜕
grad 𝑓 = 𝑔 𝑖 𝑗 ,
𝜕𝑥𝑖 𝜕𝑥 𝑗
where we use the Einstein summing convention. A point 𝑡 ∈ R such that grad 𝑓 ≠ 0
for all 𝑝 ∈ 𝑓 −1 (𝑡) is called a regular value of 𝑓 . Whenever 𝑡 ∈ R is a regular value
we have that 𝑓 −1 (𝑡) is a smooth submanifold of 𝑀. If every point in Im( 𝑓 ) ⊂ R is
a regular value of 𝑓 , then we say that 𝑓 is a regular function.
The Hessian of a function 𝑓 is defined by
∇2 𝑓 (𝑋, 𝑌 ) = ∇2 𝑓 (𝑌 , 𝑋) = ∇𝑋 ∇𝑌 𝑓 − ∇ ∇𝑋 𝑌 𝑓 = g(∇𝑋 grad 𝑓 , 𝑌 )
for vector fields 𝑋 and 𝑌 . We can show that ∇2 𝑓 at a point 𝑝 ∈ 𝑀 only depends on
𝑋 𝑝 and 𝑌 𝑝 . Hence ∇2 𝑓 (𝑢, 𝑣) for 𝑢, 𝑣 ∈ 𝑇 𝑝 𝑀 is well defined. The product rule for
the Hessian gives that
∇2 ( 𝑓 𝑔) (𝑢, 𝑣) = 𝑔∇2 𝑓 (𝑢, 𝑣) + 𝑓 ∇2 𝑔(𝑢, 𝑣) + 𝑑𝑓 (𝑢)𝑑𝑔(𝑣) + 𝑑𝑔(𝑢)𝑑𝑓 (𝑣), (1.1.1)
5
where 𝑓 , 𝑔 ∈ 𝐶 ∞ (𝑀) and 𝑢, 𝑣 ∈ 𝑇 𝑝 𝑀.

A manifold is defined to be orientable if the bundle top forms is trivial. The
orientable give us the notion a non-vanishing top form. Locally, every manifold is
locally orientable. Locally one can define a Riemannian volume form that depends
on thepchoice of orientation which is given in coordinates by |𝑔|𝑑𝑥1 ∧. . . 𝑑𝑥 𝑛 , where
|𝑔| = det(𝑔𝑖 𝑗 ). It can be shown that integrating functions with respect to a local
volume form is independent of the orientation. Hence one introduces the concept
of a Riemannian volume density, denoted by vol , to be able to integrate functions
on all Riemannian manifolds. For more information see e.g. [Lee13, p. 405]. We
will denote by 𝐿 2 (𝑀) the 𝐿 2 -space with respect to the volume density.
We can locally define the divergence of a vector field 𝑋 by
div(𝑋) vol = 𝐿 𝑋 vol ,
where 𝐿 𝑋 denotes the Lie derivative with respect to the vector field 𝑋. Notice that
the definition of divergence is independent of choice of orientation, and hence can
be defined globally even when 𝑀 is not orientable. Using the product rule for the
Lie derivative one obtains for 𝑓 ∈ 𝐶 ∞ (𝑀) that
div( 𝑓 𝑋) = 𝑓 div(𝑋) + 𝑋 ( 𝑓 ).
𝜕
Let us denote by 𝑋 = 𝑋 𝑖 𝜕𝑥 𝑖
, then the divergence written in coordinates becomes
1 𝜕 √ 𝑖
div(𝑋) = √ g𝑋 .
g 𝜕𝑥𝑖
Let us now introduce the main operator of interest in this thesis. The Laplace
operator Δ on (𝑀, g) can now be defined as
Δ 𝑓 = trg ∇2 𝑓 (×, ×) = div(grad 𝑓 ),
for 𝑓 ∈ 𝐶 ∞ (𝑀). In local coordinates we have that the Laplacian can be written as

1 𝜕 √ 𝑖𝑗 𝜕 𝑓
Δ𝑓 = √ g𝑔 . (1.1.2)
g 𝜕𝑥𝑖 𝜕𝑥 𝑗
Using (1.1.1) one can verify that
Δ( 𝑓 𝑔) = 𝑓 Δ𝑔 + 𝑔Δ 𝑓 + 2hgrad 𝑓 , grad 𝑔i, 𝑓 , 𝑔 ∈ 𝐶 ∞ (𝑀).
Let Ω denote a bounded domain in 𝑀 with 𝐶 1 -boundary denoted by 𝜕Ω.

Throughout this thesis, we will use the integration by parts formula
∫ ∫ ∫
𝑓 div(𝑋) dvol = − h𝑋, grad 𝑓 i dvol + 𝑓 h𝑋, ni dS, (1.1.3)
Ω Ω 𝜕Ω
6
where n is the outward unit normal and dS is the surface measure on 𝜕Ω. For
𝑓 , 𝑔 ∈ 𝐶 ∞ (𝑀) we will often use the following variant of the integration by parts
formula:
∫ ∫ ∫
𝑓 Δ𝑔 dvol = − hgrad 𝑓 , grad 𝑔i dvol + 𝑓 hgrad 𝑔, ni dS . (1.1.4)
Ω Ω 𝜕Ω
Let Ω be a pre-compact domain in 𝑀 with 𝐶 1 -boundary. Denote by 𝐶0∞ (Ω)

the space of smooth functions on Ω with compact support. Motivated by (1.1.4),
we say that a function 𝑓 ∈ 𝐿 2 (Ω) has a weak derivative grad 𝑓 if
∫ ∫
hgrad 𝑓 , grad 𝜙i dvol = − 𝑓 Δ𝜙 dvol for all 𝜙 ∈ 𝐶0∞ (Ω).
Ω Ω
Define the Sobolev norm of a function 𝑓 ∈ 𝐿 2 (Ω) with a weak derivative to be

∫ ∫
k 𝑓 k 2𝐻 1 (Ω) = 𝑓 2 dvol + | grad 𝑓 | 2 dvol .
Ω Ω
The Sobolev space 𝐻 1 (Ω) consists of all functions in 𝐿 2 (Ω) with finite Sobolev
norm. For 𝐹 ⊂ 𝜕Ω we use the notation
𝐶 ∞ (Ω, 𝐹) = 𝜙 ∈ 𝐶 ∞ (Ω) : 𝜙| 𝐹 = 0 .

Denote by 𝐻 1 (Ω, 𝐹) the completion of 𝐶 ∞ (Ω, 𝐹) in the Sobolev norm. We will

refer to the elements in 𝐻 1 (Ω, 𝐹) as the Sobolev functions which are zero on 𝐹.
When 𝐹 = 𝜕Ω we will use the notation
𝐻01 (Ω) B 𝐻 1 (Ω, 𝜕Ω).
Since the Laplacian is a second differentiable operator, we will also need the
second Sobolev space 𝐻 2 (Ω). Let 𝑢 be a 𝐶 2 (Ω) function, and define the norm
∫
2 2
k𝑢k 𝐻 2 (Ω) = k𝑢k 𝐻 1 (Ω) + |∇2 𝑢| 2 dvol,
Ω
where is locally defined by
|∇2 𝑢| 2 = |𝑔 𝑖 𝑗 ∇2 𝑢(𝜕𝑖 , 𝜕 𝑗 )| 2 .
𝐻 2 (Ω) is defined as the closure of 𝐶 2 (Ω) in the k · k 2𝐻 2 (Ω) norm. For more in-depth
theory of Sobolev spaces, see e.g. [Aub82, Chap. 2].
In later sections, we will use the following Sobolev embedding theorem to
prove that certain eigenvalue problems only have point-spectrum.
7
Theorem 1.1.1 (Rellich-Kondrachov Theorem [Aub82, Thm. 2.34 and 2.33]).

Let Ω ⊂ 𝑀 be a pre-compact domain with 𝐶 1 -boundary. Then the inclusion
𝐻 1 (Ω) ⊂ 𝐿 2 (Ω) is compact. Additionally, in the case that 𝑀 = R𝑛 , the inclusion
𝐻 1 (Ω) ⊂ 𝐿 2 (Ω) is compact even when 𝜕Ω is only Lipschitz.
Notice that since the inclusion 𝐻 1 (Ω, 𝐹) ⊂ 𝐻 1 (Ω) is continuous, the inclusion
𝐻 1 (Ω, 𝐹)⊂ 𝐿 2 (Ω) is also compact.
1.1.3 Flow of Vector Fields

To be able to deform submanifolds in a rigorous manner, we give some basic
definitions from the theory of flows. An integral curve 𝛾 : 𝐼 ⊂ R → 𝑀 of a vector
field 𝑋 through the point 𝑝 is a curve satisfying 𝛾(0) = 𝑝 and
𝑋 (𝛾(𝑡)) = 𝛾(𝑡),
¤ 𝑡 ∈ 𝐼.
We will often use a 1-parameter family of diffeomorphisms or a flow to collect the

information of integral curves. A 1-parameter family of diffeomorphisms defined
on a neighborhood of zero 𝐼 ⊂ R is a smooth map Φ : 𝐼 × 𝑀 → 𝑀 such that;
1. Φ0 = Id,
2. Φ𝑡 is a diffeomorphism for every 𝑡 ∈ 𝐼.
A flow on 𝑀 is a 1-parameter family of diffeomorphisms Φ : 𝐼 × 𝑀 → 𝑀

additionally satisfying
Φ𝑡+𝑠 = Φ𝑡 ◦ Φ𝑠
for 𝑡, 𝑠, 𝑡 + 𝑠 ∈ 𝐼. We say that Φ is the flow corresponding to the vector field 𝑋 if
𝛾(𝑡) = Φ𝑡 ( 𝑝) is the integral curve with respect to 𝑋 through the point 𝑝 ∈ 𝑀.
Given a 1-parameter family of diffeomorphism Φ𝑡 , one can define the corre-
sponding time-dependent vector field 𝑋𝑡 by
𝑑Φ𝑡 ( 𝑝)
𝑋𝑡 (Φ𝑡 ( 𝑝)) = , 𝑝 ∈ 𝑀.
𝑑𝑡
We are mainly interested in 1-parameter families of diffeomorphisms deforming
submanifolds of 𝑀. Let 𝑆 be a submanifold in 𝑀 with boundary 𝜕𝑆. Then we will
use the notation 𝑆𝑡 = Φ𝑡 (𝑆) and 𝜕𝑆𝑡 = Φ𝑡 (𝜕𝑆).
Example 1.1.2. Consider a regular function 𝑓 : Ω ⊂ 𝑀 → R such that 0 ∈ 𝑓 (Ω).

Then one can define the family of submanifolds 𝑆𝑡 = 𝑓 −1 (𝑡). In this case, the unit
normal n to 𝑆𝑡 is given by
grad 𝑓
n= .
| grad 𝑓 |
8
Taking the flow Φ𝑡 of the vector field 𝑋 given by
grad 𝑓
𝑋= ,
| grad 𝑓 | 2
we claim that 𝑆𝑡+𝑠 = Φ𝑡 (𝑆 𝑠 ). To see this, let 𝑝 ∈ 𝑆 𝑠 and let 𝛾 be an integral curve
of the vector field 𝑋 with 𝛾(0) = 𝑝. Then

𝑑 grad 𝑓
𝑓 (𝛾(𝑡)) = hgrad 𝑓 , 𝛾(𝑡)i
¤ = grad 𝑓 , = 1.
𝑑𝑡 | grad 𝑓 | 2
Hence the claim follows since 𝑓 (𝛾(𝑡)) = 𝑡 + 𝑠.
The time derivative of integrals over time depending surfaces with respect to
the flow satisfies the following identity.
Theorem 1.1.3 ([Fra12, Thm. 4.42]). Let 𝑆 be an 𝑚-dimensional submanifold of

𝑀. Assume that Φ𝑡 is a 1-parameter family of diffeomorphisms corresponding to
a vector field 𝑋 on 𝑀, and denote by 𝑆𝑡 = Φ𝑡 (𝑆). Let 𝛼𝑡 be a 1-parameter family
of 𝑚-forms on 𝑀. Then
∫ ∫ ∫
𝑑 𝜕
𝛼𝑡 = 𝛼𝑡 + 𝐿 𝑋 𝛼𝑡 .
𝑑𝑡 𝑆𝑡 𝑆𝑡 𝜕𝑡 𝑆𝑡
1.1.4 Hypersurfaces
We will throughout the thesis be particularly interested in submanifolds that are
hypersurfaces. A hypersurface 𝑁 ⊂ 𝑀 is a smooth Riemannian submanifold of
co-dimension 1. In this context, we will refer to 𝑀 as the ambient space. We will
use ˜ above the metric, connection and curvature when the objects are defined with
respect to 𝑁. Since 𝑁 is a hypersurface we can locally define the normal vector
field n. The volume distribution on 𝑁 is locally given by vol
g = 𝜄n vol , where 𝜄 is
the contraction of the form.
Using the relationship with 𝑀 we can define the second fundamental form.
Then the second fundamental form is defined by
II(𝑋, 𝑌 ) = h∇𝑋 𝑌 , nin,
for vector fields 𝑋 and 𝑌 defined on a neighborhood of 𝑁. The second fundamental

form satisfies the following properties:
1. II is independent of the choice of n, and hence can be defined globally.
2. II(𝑋, 𝑌 ) is bilinear in both 𝑋 and 𝑌 .
9
3. II is symmetric.
4. II(𝑋, 𝑌 ) at the point 𝑝 ∈ 𝑁 depends only on 𝑋 𝑝 and 𝑌 𝑝 .
The second fundamental form is exactly the difference between the connection
on 𝑀 and the connection on 𝑁. To be precise, for vector fields 𝑋 and 𝑌 on 𝑁
smoothly extended to a neighborhood of 𝑁 we have that
∇𝑋 𝑌 = ∇˜ 𝑋 𝑌 + II(𝑋, 𝑌 ).
We can use this result to figure out when a geodesic on 𝑁 is also a geodesic on 𝑀. If
¤
𝛾 is a geodesic on 𝑁, then it is also a geodesic on 𝑀 if and only if II( 𝛾(𝑡), ¤
𝛾(𝑡)) = 0.
The submanifold 𝑁 is said to be totally geodesic in 𝑀 if every geodesic in 𝑁 is also
a geodesic in 𝑀. By the above remarks, we can hence view the second fundamental
form Π as an obstruction to 𝑁 being totally geodesic in 𝑀.
Assume that 𝑁 is orientable with a chosen global unit normal vector n. The
second fundamental form Π gives rise to the shape operator 𝑠, defined by
h𝑠(𝑋), 𝑌 i = h∇𝑋 𝑌 , ni = hII(𝑋, 𝑌 ), ni.
Note that the shape operator restricted to a point 𝑠 𝑝 : 𝑇 𝑝 𝑁 → 𝑇 𝑝 𝑁 is a symmetric
linear map. The Gaussian curvature 𝐾 is defined by 𝐾 ( 𝑝) = det(𝑠 𝑝 ), while the
mean curvature is defined by
tr(𝑠 𝑝 )
𝐻 ( 𝑝) = .
dim(𝑁)
Note that in changing the orientation changes the sign of the mean curvature and
the Gaussian curvature is multiplied by (−1) 𝑛 . Mean curvature can tell us when
the hypersurface is area-minimizing. Recall, that we say that a hypersurface 𝑁 is
area minimizing if it has the minimal area compared to all other surfaces with the
same boundary. In this case we have that mean curvature of 𝑁 is zero.
Example 1.1.4 (Example 1.1.2 continued). We will use the same notation as in
Example 1.1.2. Our goal is to compute the second fundamental form of 𝑆𝑡 . We
will denote by n = grad 𝑓 /| grad 𝑓 | and let 𝑋 be a vector field on 𝑀. Denote by
𝑋 > the part of 𝑋 that is tangential to 𝑁. Then any vector field can be written as
𝑋 = 𝑋 > + h𝑋, nin
when restricted to 𝑁. Let 𝑋 and 𝑌 be vector fields that are tangent to 𝑁. Then
hII(𝑋, 𝑌 ), ni = h∇𝑋 𝑌 , grad 𝑓 /| grad 𝑓 |i
1 1
= ∇𝑋 h𝑌 , grad 𝑓 i − h𝑌 , ∇𝑋 grad 𝑓 i
| grad 𝑓 | | grad 𝑓 |
1
=− ∇2 𝑓 (𝑌 , 𝑋).
| grad 𝑓 |
10
In the case when 𝑌 = h𝑌 , nin we get that
h𝑌 , ni 2
hII(𝑋, 𝑌 ), ni = ∇ 𝑓 (n, 𝑋).
| grad 𝑓 |
If 𝑋 = 𝑋 > + h𝑋, nin and 𝑌 = 𝑌 > + h𝑌 , nin we hence obtain
1 h𝑌 , ni 2
hII(𝑋, 𝑌 ), ni = − ∇2 𝑓 (𝑌 > , 𝑋 > ) + ∇ 𝑓 (n, 𝑋 > )

h𝑋, ni 2 > 1
+ ∇ 𝑓 (n, 𝑌 ) − h𝑋, nih𝑌 , ni grad 𝑓 , grad .
Let us compute the mean curvature of 𝑆𝑡 . Let 𝐸 1 , . . . , 𝐸 𝑛 be a local orthonormal

frame with 𝑛 = dim(𝑀), where 𝐸 𝑛 restricted to 𝑁 is given by grad 𝑓 /| grad 𝑓 |.
This gives that
𝑛
1 Õ
𝐻 ( 𝑝) = hII(𝐸 𝑖 , 𝐸 𝑖 ), ni
𝑛 − 1 𝑖=1

1 div(grad 𝑓 ) 1
=− + grad 𝑓 , grad
𝑛−1 | grad 𝑓 | | grad 𝑓 |

1 grad 𝑓
=− div ,
𝑛−1 | grad 𝑓 |
where the last equality follows from the product rule for the divergence.
When integrating over domains we will use the co-area formula, see e.g. [Fed59,
Thm. 3.1]or [Cha06, Thm. VIII3.3]. Recall by Sard’s theorem that a smooth
function defined on a pre-compact domain Ω ⊂ 𝑀 and a function 𝑓 : Ω → R is
almost every value is a regular values.
Theorem 1.1.5. Let Ω be a pre-compact domain in (𝑀, g), and let 𝑓 : Ω → R be

a smooth function. Then we have that
∫ ∫ ∞∫
𝑢| grad 𝑓 | dvol = 𝑢(𝑥) dS dt. (1.1.5)
Ω −∞ 𝑓 −1 (𝑡)
1.1.5 Spherical Coordinates and Model Spaces

We now introduce spherical coordinates on Riemannian manifolds as a convenient
local tool. After defining the model spaces in Riemannian geometry, we will see
how spherical coordinates can be used to give them a uniform description.
11
Spherical Coordinates
For a Riemannian manifold (𝑀, g) we fix a point 𝑝 ∈ 𝑀. Inside the ball
𝐵Inj( 𝑝) ( 𝑝) ⊂ 𝑀 we can work with spherical coordinates. This is done by utilizing
the standard spherical coordinates in 𝐵Inj( 𝑝) (0) ⊂ 𝑇 𝑝 𝑀 and using the exponential
map. To elaborate, let us define the radial distance function 𝑟 : 𝑀 → R with
respect to the point 𝑝 ∈ 𝑀 by
𝑟 (𝑥) = dist(𝑥, 𝑝).
An important fact we are going to use often is that | grad 𝑟 | = 1. Additionally,

the function 𝑟 2 (𝑥) is a smooth function on 𝐵Inj( 𝑝) ( 𝑝). Denote the sphere with
radius 𝑡 and center 𝑝 in 𝑀 by 𝑆𝑡𝑛−1 = 𝑟 −1 (𝑡). Let (𝑟, 𝜃) be spherical coordinates
on 𝐵Inj( 𝑝) (0) ⊂ 𝑇 𝑝 𝑀 so that 𝜃 ∈ {𝑥 ∈ R𝑛 : |𝑥| = 1}. Then the association
(𝑟, 𝜃) ↦→ exp(𝑟𝜃) gives us spherical coordinates on 𝐵Inj( 𝑝) ( 𝑝) ⊂ 𝑀. In this case,
we have that
𝑆𝑡𝑛−1 = exp(𝑡{𝑥 ∈ R𝑛 : |𝑥| = 1}).
When integrating over 𝐵 𝑅 ( 𝑝) for 𝑅 ≤ Inj( 𝑝) we will use the fact that the co-area
formula (1.1.5) simplifies to
∫ ∫ 𝑅∫
𝑢 dvol = 𝑢(𝑥) dS 𝑑𝑡. (1.1.6)
𝐵 𝑅 ( 𝑝) 0 𝑆𝑡𝑛−1
Model spaces
Before going any further let us go through some important examples, namely the
so called model spaces. The model spaces are intuitively the maximally symmetric
Riemannian manifolds and are used as prototypical examples of Riemannian ge-
ometries. We will use the model spaces in Section 1.1.6 for comparison theorems.
The model spaces can be defined as the connected, simply connected, and complete
Riemannian manifolds with constant sectional curvature 𝐾. If we fix the dimension
𝑛, then there are only three classes of model spaces:
K = 0 (Euclidean Space): The Euclidean space R𝑛 is the Riemannian manifold
where the Riemannian metric is given by the standard Euclidean inner prod-
uct gR𝑛 . The sectional and Ricci curvatures of this space are both identically
zero.
K = R−2 (Round Sphere): The sphere S𝑛𝑅 = {𝑥 ∈ R𝑛+1 : |𝑥| = 𝑅} can naturally
be made into a Riemannian manifold with radius 𝑅 > 0. On the sphere S𝑛𝑅 ,
the sectional curvature is the positive number
1
Sec = .
𝑅2
12
Moreover, the Ricci curvature is precisely given by
𝑛−1
Ric = gS𝑛𝑅 .
𝑅2
K = −R−2 (Hyperbolic Space): The hyperbolic space H𝑛𝑅 with radius 𝑅 > 0 can
be realized as the upper half plane R𝑛−1 × (0, ∞) with the metric
gR𝑛 | R𝑛−1 ×(0,∞)

gH𝑛𝑅 = 𝑅 2 .
𝑥 𝑛2
On the hyperbolic space H𝑛𝑅 , the sectional curvature is the negative number
1
Sec = − .
𝑅2
Moreover, the Ricci curvature is precisely given by
𝑛−1
Ric = − gH𝑛𝑅 .
𝑅2
The following result shows that the model spaces can be locally given a uniform
description in terms of spherical coordinates.
Theorem 1.1.6 (Metric in Spherical Coordinates). Let (𝑀𝐾 , g𝐾 ) be the model

space with constant sectional curvature 𝐾. Define the associated function
√


 sin 𝑟 𝐾


 √ 𝐾 > 0,


 𝐾
sin𝐾 (𝑟) = 𝑟 𝐾 = 0,
 √
sinh 𝑟 −𝐾


𝐾 < 0.

 √

 −𝐾
In spherical coordinates we can write
g𝐾 = 𝑑𝑟 ⊗ 𝑑𝑟 + sin𝐾 (𝑟) 2 gS𝑛−1 .

1
Additionally, the volume form can be written by
vol g𝐾 = sin𝐾 (𝑟) 𝑛−1 𝑑𝑟 ∧ vol gS𝑛−1 .

1
In particular, the volume of a sphere is given by
vol 𝑛−1 (𝑆𝑟𝑛−1 ) = sin𝐾 (𝑟) 𝑛−1 vol 𝑛−1 (S1𝑛−1 ).
13
1.1.6 Comparison Theorems

In this section we are going to discuss two comparison theorems. Conceptually,
comparison theorems are results that compare a geometric quantity of a Riemannian
manifold (such as volume growth) with the corresponding quantity on a model
space. Define for 𝐾 ∈ R the utility function
cot𝐾 (𝑟) = 𝜕𝑟 log(sin𝐾 (𝑟)).
Then we have from Theorem 1.1.6 that the growth of spheres for the model spaces
are given by
𝜕𝑟 vol 𝑛−1 (𝑆𝑟𝑛−1 ) = (𝑛 − 1) cot𝐾 (𝑟) vol 𝑛−1 (𝑆𝑟𝑛−1 ).
Although we do not have such simple formulas for more general Riemannian
manifolds, one can use comparison theorems to retain some control.
Both in Paper A and in Paper B the surface area of spheres will be of importance.
In the case that our Riemannian manifold is not a model space we can still get that
∫
𝑛−1
𝜕𝑟 vol 𝑛−1 (𝑆𝑟 ) = Δ(𝑟) dS .
𝑆𝑟𝑛−1
Notice that since ∫
lim+ Δ(𝑟) dS = 0,
𝑟 →0 𝑆𝑟𝑛−1
one has that ∫ ∫
𝑅
vol 𝑛−1 (𝑆 𝑛−1
𝑅 ) = Δ(𝑟) dS 𝑑𝑟.
0 𝑆𝑟𝑛−1
We can estimate Δ𝑟 by comparing with the volume growth on the model spaces.
Theorem 1.1.7 (Rauch Comparison Theorem). Let (𝑀, g) be an 𝑛-dimensional
Riemannian manifold where 𝜅 ≤ Sec ≤ 𝐾 for 𝜅, 𝐾 ∈ R. Then
cot𝐾 (𝑟)g𝑆𝑟𝑛−1 ≤ ∇2𝑟 ≤ cot 𝜅 (𝑟)g𝑆𝑟𝑛−1 . (1.1.7)
Taking the trace of (1.1.7) gives
(𝑛 − 1) cot𝐾 (𝑟) ≤ Δ𝑟 ≤ (𝑛 − 1) cot 𝜅 (𝑟).
The curvature condition in Theorem 1.1.7 is rather restrictive. Even with
relaxed curvature conditions, one still can get an estimate for Δ𝑟 as shown in the
following result.
Theorem 1.1.8 (Volume Comparison theorem). Assume that (𝑀, g) is an 𝑛-
dimensional Riemannian manifold such that for any vector field 𝑋 one has
Ric(𝑋, 𝑋) ≥ (𝑛 − 1)𝜅|𝑋 | 2 , 𝜅 ∈ R.
Then
Δ𝑟 ≤ (𝑛 − 1) cot 𝜅 (𝑟).
14
1.2. The Laplace Operator
1.1.7 Spherically Symmetric Manifolds

Let 𝑓 : [0, 𝑅) → R be a function satisfying
𝑓 (0) = 0, 𝑓 0 (0) = 1, 𝑓 00 (0) = 0.
We say that a Riemannian manifold (𝑀, g) is spherically symmetric with respect

to the point 𝑝 ∈ 𝑀 if g can be written in 𝐵 𝑅 ( 𝑝) for 𝑅 ≤ Inj( 𝑝) as
g = 𝑑𝑟 ⊗ 𝑑𝑟 + 𝑓 (𝑟) 2 gS𝑛−1 .
1
We will see in Example 1.2.9 that this added symmetry is particularly helpful
when studying eigenvalue problems on balls centered at the point 𝑝. The relevant
formulas for spherically symmetric manifold can be given through the function 𝑓
by
Volume Form vol = 𝑓 (𝑟) 𝑛−1 𝑑𝑟 ∧ 𝑑𝑆1 ,
0 2 00 (𝑟 )
Scalar Curvature S( 𝑝) = (𝑛 − 1) (𝑛−2)−(𝑛−2) ( 𝑓𝑓 )(𝑟 (𝑟
)2
)−2 𝑓 𝑓 (𝑟 )
,
0
Mean Curvature of 𝑆𝑟 1
𝑛−1 Δ𝑟 = (𝑛 − 1) 𝑓𝑓 (𝑟(𝑟)) ,
0
𝜕2
Laplace Operator Δ = 𝜕𝑟 2
+ (𝑛 − 1) 𝑓𝑓 (𝑟(𝑟)) 𝜕𝑟
𝜕
+ 1
Δ 𝑛−1 .
𝑓 2 (𝑟 ) S1
Examples of spherically symmetric manifolds are the model spaces. In this case
we have 𝑓 (𝑟) = sin𝐾 (𝑟). For more information and for a proof of the formulas
presented in this section, see [Pet16, Sec. 4.2.3] where the spherically symmetric
spaces go under the name rotationally symmetric spaces.
1.2 The Laplace Operator

In Section 1.1.2 we defined the Laplace operator Δ on a Riemannian manifold
(𝑀, g) by
Δ( 𝑓 ) = div(grad( 𝑓 )), 𝑓 ∈ 𝐶 ∞ (𝑀).
In this section we are going to explore properties of the Laplace operator, as well
as certain associated eigenvalue problems.
1.2.1 The Helmholtz Equation

We say that a twice differentiable function 𝑓 defined on a bounded domain 𝑈 ⊂ 𝑀
is a solution to the Helmholtz equation with wave number 𝑘 2 if
Δ 𝑓 + 𝑘 2 𝑓 = 0.
15
In the special case that 𝑘 2 = 0 we say that the function 𝑓 is harmonic. The operator
Δ + 𝑘 2 is elliptic when written out in coordinates. Hence Δ + 𝑘 2 inherits local
properties from elliptic theory such as the following result.
Theorem 1.2.1 (Harnack’s Inequality). Let (𝑀, g) be a Riemannian manifold with

bounded sectional curvature 𝜅 ≤ Sec ≤ 𝐾 for 𝜅, 𝐾 ∈ R. Assume that there exists
a constant 𝛿 > 0 such that Inj( 𝑝) > 𝛿 for all 𝑝 ∈ 𝑀. Then there exists 𝑅 > 0
independent of 𝑝 such that for all 𝑟 satisfying 4𝑟 < 𝑅 and any strictly positive
harmonic function ℎ : 𝐵4𝑟 ( 𝑝) → R we have
sup ℎ ≤ 𝐶 inf ℎ
𝐵𝑟 ( 𝑝) 𝐵𝑟 ( 𝑝)
where the constant 𝐶 > 0 only depends on 𝑀.
Proof. By [Eld13, Thm. 2.4] we have that when written out in exponential coor-
dinates there exists some 𝑅 > 0 such that 𝑔𝑖 𝑗 is bounded in 𝐵 𝑅 ( 𝑝) for all 𝑝 ∈ 𝑀.
This implies that for any vector 𝑣 = 𝑣 𝑖 𝜕𝑖 ∈ 𝑇𝑥 𝑀 and any 𝑥 ∈ 𝐵 𝑅 ( 𝑝) we have that
𝑛 𝑛 𝑛
Õ √ Õ 𝑖𝑗 𝑖 𝑗 1 Õ 𝑖 2
𝛾 |𝑣 𝑖 | 2 ≤ g 𝑔 𝑣𝑣 ≤ |𝑣 |
𝑖=1 𝑖, 𝑗=1
𝛾 𝑖=1
for some constant 𝛾 ∈ (0, 1). The result now follows from [GL86, Thm. 8.20].
By definition, we know that the eigenfunctions are twice differentiable. How-

ever it turns out that they are indeed smooth.
Theorem 1.2.2 (Regularity [GL86, Thm. 8.10]). Let 𝑓 ∈ 𝐶 2 (𝑈) be a solution to

the Helmholtz equation defined on a bounded domain 𝑈 with wave number 𝑘 2 .
Then for all open sets 𝑉 ⊂⊂ 𝑈 we have that 𝑓 is 𝐶 ∞ (𝑉).
1.2.2 Associated Harmonic Function

Solutions to the Helmholtz equation on the Riemannian manifold (𝑀, g) can always
be extended to harmonic functions on the Riemannian manifold (𝑀 × R, g × gR ).
To be precise, assume that 𝑓 is a solution to the Helmholtz equation on the bounded
domain 𝑈 ⊂ 𝑀 with wave number 𝑘 2 . Then the function ℎ(𝑥, 𝑡) = cosh(𝑘𝑡) 𝑓 (𝑥)
is harmonic on 𝑀 × R since
𝜕 2 (cosh(𝑘𝑡) 𝑓 (𝑥))
Δg×gR (ℎ(𝑥, 𝑡)) = Δg (cosh(𝑘𝑡) 𝑓 (𝑥)) +
𝜕𝑡 2
2
= cosh(𝑘𝑡)Δg 𝑓 (𝑥) + 𝑘 cosh(𝑘𝑡) 𝑓 (𝑥) = 0.
16
Define a set 𝑁 ⊂ 𝑈 to be 𝐶-dense for a positive constant 𝐶 if we have that

every point 𝑥 ∈ 𝑈 satisfies
dist(𝑥, 𝑁 ∪ 𝜕𝑈) ≤ 𝐶.
The nodal set of a solution to the Helmholtz equation 𝑓 is defined as the zero set
N ( 𝑓 ) = 𝑓 −1 (0) = {𝑥 ∈ 𝑈 : 𝑓 (𝑥) = 0}.
Using the extension trick together with Harnack’s inequality we can show that the
nodal set is 𝐶𝑘 -dense.
Theorem 1.2.3 (Density of the Nodal Set). Let 𝑓 : 𝑈 → R be a solution to the

Helmholtz equation with wave number 𝑘 2 . Assume that the conditions in Theorem
1.2.1 is satisfied for (𝑀, g). Then there exists a constant 𝐷 > 0 only depending on
𝑈 and g such that the nodal set N ( 𝑓 ) is 𝐷𝑘 -dense.
Proof. Define ℎ(𝑥, 𝑡) = cosh(𝑘𝑡) 𝑓 (𝑥) to be the associated harmonic function of

𝑓 . Assume without loss of generality that ℎ( 𝑝) > 0 for some 𝑝 = (𝑥, 0) ∈ 𝑈 × R.
Denote by
𝐵4√2𝑅 ( 𝑝) ⊂ 𝑈 × R
√
the ball centered at the point 𝑝 with radius 4 2𝑅 so that ℎ > 0 on 𝐵4√2𝑅 ( 𝑝). In
other words, we assume that 𝐵4√2𝑅 ( 𝑝) does not intersect the nodal set N (ℎ). By
Theorem 1.2.1 and the definition of ℎ we have that
exp (𝑘 𝑅)
inf 𝑓 ≤ cosh (𝑘 𝑅) sup 𝑓
2 𝐵 𝑅 ( 𝑥) 𝐵 𝑅 ( 𝑥)
= sup ℎ
𝐵√2𝑅 ( 𝑝)
≤𝐶 inf ℎ
𝐵√2𝑅 ( 𝑝)
= 𝐶 inf 𝑓 .
𝐵 𝑅 ( 𝑥)
log(𝐶) 𝐷
This implies that exp (𝑘 𝑅) ≤ 𝐶 or equivalently 𝑅 ≤ 𝑘 = 𝑘.
1.2.3 Self-Adjoint Operators

Before jumping into eigenvalue problems associated with the Laplace operator it
will be beneficial to recap some theory on self-adjoint operators. The material in
this section is fairly standard, and can for example be found in [Bor20, Ch. 3 & 4].
This general framework allows us to study the eigenvalue problems abstractly, as
17
well as making the next section less repetitive. In this section we will let H
denote a Hilbert space with inner product h·, ·i. An densely defined linear operator
𝐴 : 𝐷 ( 𝐴) ⊂ H → H with domain 𝐷 ( 𝐴) is said to be symmetric if
h𝐴𝑢, 𝑣i = h𝑢, 𝐴𝑣i for all 𝑢, 𝑣 ∈ 𝐷 ( 𝐴).
The domain of the adjoint of a densely defined linear operator 𝐴 : 𝐷 ( 𝐴) ⊂ H → H

with domain 𝐷 ( 𝐴) is defined to be the set
𝐷 ( 𝐴∗ ) = {𝑣 ∈ H : 𝑢 ↦→ h𝑣, 𝑇𝑢i is bounded for all 𝑢 ∈ 𝐷 ( 𝐴)}.
Notice that one always has that 𝐷 ( 𝐴) ⊂ 𝐷 ( 𝐴∗ ). A symmetric operator 𝐴 is called

self-adjoint if 𝐷 ( 𝐴) = 𝐷 ( 𝐴∗ ). Showing that a given operator is indeed self-adjoint
is often cumbersome. Hence the following results will be useful:
Theorem 1.2.4 ([Bor20, Lem. 3.27 and Thm. 3.33]). Let 𝐴 : 𝐷 ( 𝐴) ⊂ H → H be
a densely defined symmetric operator.
• If 𝐴 is self-adjoint and 𝐵 is a bounded symmetric operator, then 𝐴 + 𝐵 is a
self-adjoint operator on 𝐷 ( 𝐴).
• If 𝐴 is such that
h𝐴𝑣, 𝑣i ≥ 𝐶 k𝑣k 2 , 𝑣 ∈ 𝐷 ( 𝐴),
for some constant 𝐶 > 0, then there exists a self-adjoint extension of 𝐴. This
extension is known as the Friedrichs extension.
Define the spectrum of a densely defined linear operator 𝐴 : 𝐷 ( 𝐴) ⊂ H → H
to be the collection
𝜎( 𝐴) = {𝜆 ∈ C : 𝐴 − 𝜆𝐼 does not have a bounded inverse}.
We can further divide the spectrum into the essential spectrum and the discrete
spectrum. The essential spectrum consists of all the points 𝜆 ∈ 𝜎( 𝐴) where one
can find a sequence 𝑢 𝑖 ∈ 𝐷 ( 𝐴) satisfying
• k𝑢 𝑖 k = 1,
• 𝑢 𝑖 converges weakly to zero,
• lim𝑖→∞ k ( 𝐴 − 𝜆)𝑢 𝑖 k = 0.
The complement of the essential spectrum is by definition the discrete spectrum.
If 𝜆 is in the discrete spectrum, then there exists an element 𝑢 ∈ 𝐷 ( 𝐴) such that
𝐴𝑢 = 𝜆𝑢. In this case, we refer to 𝜆 as an eigenvalue of 𝐴. When 𝐴 is a self-adjoint
operator it is well-known that 𝜎( 𝐴) ⊂ R.
18
Theorem 1.2.5 (Min-Max Theorem [Bor20, Thm. 5.15]). Let 𝐴 : 𝐷 ( 𝐴) ⊂ H →

H be a densely defined non-negative self-adjoint linear operator. Then
h𝐴𝑢, 𝑢i
𝛼𝑘 = min max
𝑉𝑘 ⊂𝐷 ( 𝐴) 𝑢 ∈𝑉𝑘 \{0} k𝑢k 2
dim(𝑉𝑘 )=𝑘
is either the 𝑘’th smallest eigenvalue counting multiplicity, or the infimum over the
essential spectrum.
In the spectral theory of operators, an important special case is the case of

compact operators. Recall that an operator 𝑆 : H → H is said to be compact
if for every bounded sequence 𝑢 𝑖 ∈ H we have that the sequence 𝑆𝑢 𝑖 contains a
convergent subsequence.
Theorem 1.2.6 (Spectral Theorem for compact operators). Let 𝑆 : H → H be

a compact self-adjoint operator. Then there exists an orthonormal basis for H
consisting of eigenvectors for 𝑆. Additionally, the eigenvalues when ordered in
decreasing order the eigenvalues goes towards 0.
1.2.4 Eigenvalue Problems

In this section we are going to explore global properties of eigenvalue problems for
the Laplacian. Among the problems we are going to study are eigenvalue problems
for compact manifolds, the Neumann problem, and the Dirichlet problem. We are
also going to study the Steklov problem, which is a boundary value problem for
harmonic functions.
Compact Manifolds
Let (𝑀, g) be a Riemannian manifold where 𝑀 is compact and without boundary.
Then we define an eigenfunction 𝑢 ∈ 𝐶 ∞ (𝑀) with corresponding eigenvalue 𝑘 2 to
be the global solution to the problem
Δ𝑢 + 𝑘 2 𝑢 = 0. (1.2.1)
The collection of all eigenvalues is called the spectrum of (𝑀, g). Notice that the
constant functions are eigenfunctions with corresponding eigenvalue 𝑘 2 = 0. We
can order the eigenvalues counting multiplicity
0 = 𝑘 12 < 𝑘 22 ≤ 𝑘 32 ≤ · · · 𝑘 𝑚
2
→ ∞.
One can find an orthonormal basis for 𝐿 2 (𝑀) consisting of eigenfunctions. We

refer the reader to [Cha84, Chap. 1.3] for the claims made above.
19
Example 1.2.7 (Sphere, see [Cha84, Chap. 2 Sec. 4] or [Shu01, Sec. 22.3]). Of
the model spaces, the spheres S𝑛𝑅 for 𝑅 > 0 are the only compact spaces. Let
us for simplicity only consider 𝑅 = 1 with Laplace operator ΔS1𝑛 . When the
sphere is viewed as a Riemannian embedding into R𝑛+1 , the eigenfunctions are
given as restrictions of homogeneous harmonic functions. Recall that a function
𝑓 ∈ 𝐶 ∞ (R𝑛+1 ) is homogeneous of degree 𝑚 (or simply 𝑚-homogeneous) if it
satisfies
𝑓 (𝑟𝑥) = 𝑟 𝑚 𝑓 (𝑥) for all 𝑟 ∈ [0, ∞) and 𝑥 ∈ R𝑛+1 .
One has that the restriction of 𝑚-homogeneous harmonic polynomials to the sphere
have eigenvalue 𝑘 2 = 𝑚(𝑚 + 𝑛 − 1). Moreover, each of the eigenspaces 𝐸 𝑘 2 have
dimension1
𝑛+𝑚 𝑛+𝑚−2
dim (𝐸 𝑘 2 ) = − .
𝑚 𝑛
We will refer to the eigenfunctions of the sphere as spherical harmonics.
In the case that 𝑛 = 1, the non-constant eigenfunctions corresponding to the
eigenvalue 𝑚 2 for 𝑚 ∈ N are spanned by cos(𝑚𝜃) and sin(𝑚𝜃). As such, the
eigenfunctions are precisely the classical Fourier basis.
We can also view the Laplacian as a densely defined self-adjoint operator on

the space Δ : 𝐷 (Δ) ⊂ 𝐿 2 (𝑀) → 𝐿 2 (𝑀). It can be shown that the Laplacian
is essentially self adjoint when acting on 𝐶 2 (𝑀), meaning that there exists a
unique self-adjoint extension. By using regularity theory of elliptic operators,
one can show that the eigenfunctions has a smooth representative and corresponds
to solutions to (1.2.1), see [Bor20, Thm. 9.27]. Using the theory of self-adjoint
operators one can Rayleigh quotient.
Theorem 1.2.8. The 𝑚’th eigenvalue of the problem (1.2.1) can be written as
∫
2 𝑀
| grad 𝑢| 2 dvol
𝑘𝑚 = inf sup ∫ .
𝑉 ⊂𝐻 1 ( 𝑀 ) 𝑢 ∈𝑉 \{0}
𝑀
𝑢 2 dvol
dim(𝑉 )=𝑚
Outline of proof. Let us show that the operator (Δ − 1) −1 extends to a compact

operator (Δ − 1) −1 : 𝐿 2 (𝑀) → 𝐿 2 (𝑀). For 𝑢 ∈ 𝐶 ∞ (𝑀) we have
hΔ𝑢 − 𝑢, 𝑢i = −k𝑢k 2𝐻 1 ( 𝑀 ) .
Hence
k𝑢k 𝐻 1 ( 𝑀 ) ≤ kΔ𝑢 − 𝑢k 𝐿 2 ( 𝑀 ) .
1The reader should be made aware of the typo in [Cha84] for the dimension of the 𝐸 𝑘 2 .
20
Using that 𝐶 ∞ (𝑀) is dense in 𝐿 2 (𝑀) we get that
k (Δ − 1) −1 𝑢k 𝐻 1 ( 𝑀 ) ≤ k𝑢k 𝐿 2 ( 𝑀 ) .
By Theorem 1.1.1 this shows that (Δ − 1) −1 : 𝐿 2 (𝑀) → 𝐿 2 (𝑀) is compact. This

means that Δ can only have eigenvalues, and no essential spectrum. For more
details, see [Cha84, Sec. I.5].
Boundary Value Problems

In this section we are going to talk about the Dirichlet and Neumann boundary
value problems. For simplicity, we will assume that Ω is a pre-compact domain
where Ω ⊂ 𝑀 has 𝐶 ∞ -boundary.
Eigenfunctions of the Dirichlet problem is a function 𝑢 ∈ 𝐶 2 (Ω) ∩ 𝐶 0 (Ω)
satisfies (
Δ𝑢 + 𝜆𝑢 = 0 in Ω
. (1.2.2)
𝑢=0 on 𝜕Ω
It is well known that the eigenvalues when ordered are positive, discrete, and go
towards infinity. We also have that the eigenfunctions become smooth by using
elliptic regularity and we have that the first eigenfunctions minimizes the Dirichlet
energy. In general we have that the 𝑚’th eigenvalue is equal to the Rayleigh quotient
∫
Ω
| grad 𝑢| 2 dvol
𝜆 𝑚 = inf sup ∫ .
𝑉 ⊂𝐻01 (Ω) 𝑢 ∈𝑉 \{0}
Ω
𝑢 2 dvol
dim(𝑉 )=𝑚
As a reference, for the statement in this section see [Cha84, Chap. 1] and [Bor20,
Chap. 9].
Another standard problem is the Neumann problem, where a solution 𝑢 ∈
𝐶 2 (Ω) ∩ 𝐶 1 (Ω) is given by
(
Δ𝑢 + 𝜇𝑢 = 0 in Ω
. (1.2.3)
𝑢𝑛 = 0 on 𝜕Ω
In this case we have that the first eigenvalue is zero, while the rest is positive.
Writing out the Rayleigh quotient, see [Cha84, Chap. 1], for the 𝑚’th eigenvalue
gives ∫
| grad 𝑢| 2 dvol
𝜇 𝑚 = inf sup Ω ∫ .
𝑉 ⊂𝐻 1 (Ω) 𝑢 ∈𝑉 \{0}
Ω
𝑢 2 dvol
dim(𝑉 )=𝑚
By using [GL86, Thm. 9.19] together with the Sobolev embedding theorem one get
that the eigenfunctions are smooth up to the boundary. Comparing the Rayleigh
21
quotient of the Dirichlet and Neumann problems we immediately see that 𝜇 𝑚 ≤ 𝜆 𝑚

for all 𝑚 ∈ N.
Example 1.2.9 (Model Spaces). In this section we will let (𝑀𝐾 , g𝐾 ) be the model
space with curvature 𝐾. We will start by computing the eigenfunctions for the
Dirichlet and Neumann problem in the case when Ω = 𝐵(𝑟 0 , 𝑝). In the case that
𝐾 > 0 we have the additional requirement that 𝑟 0 < √𝜋 . Since the manifolds are
𝐾
spherically symmetric we will use separation of variables to find the eigenfunctions.
Let 𝑢(𝑟, 𝜃) = 𝑅(𝑟)Θ(𝜃) be an eigenfunction written in spherical coordinates with
eigenvalue 𝜆. Then we have that
1
Δ𝑢(𝑟, 𝜃) = 𝑅 00 (𝑟)Θ(𝜃) + (𝑛 − 1) cot𝐾 (𝑟)𝑅 0 (𝑟)Θ(𝜃) + 𝑅(𝑟) ΔS𝑛−1 Θ(𝜃)
sin2𝐾 (𝑟) 1
= −𝜆𝑅(𝑟)Θ(𝜃),
which simplifies to
sin2𝐾 (𝑟)𝑅 00 (𝑟) + (𝑛 − 1) cos𝐾 (𝑟) sin𝐾 (𝑟)𝑅 0 (𝑟) + 𝜆 sin2𝐾 (𝑟)𝑅(𝑟) ΔS𝑛−1 Θ(𝜃)
=− 1 .
𝑅(𝑟) Θ(𝜃)
Since the left hand-side is independent of 𝜃 we have that Θ is a spherical harmonic

function. Using that the eigenvalues on the (𝑛−1)-sphere have the form 𝑚(𝑚+𝑛−2)
we get that
sin2𝐾 (𝑟)𝑅 00 (𝑟) + (𝑛−1) cos𝐾 (𝑟) sin𝐾 (𝑟)𝑅 0 (𝑟) + (𝜆 sin2𝐾 (𝑟) −𝑚(𝑚+𝑛−2))𝑅(𝑟) = 0.
(1.2.4)
Hence we get that if 𝑅(𝑟) is a solution to (1.2.4) satisfying 𝑅(𝑟 0 ) = 0 one has that
𝑢 is a solution to the Dirichlet problem. If on the other hand 𝑅 0 (𝑟 0 ) = 0 we get
that 𝑢 is a solution to the Neumann problem. In Paper D, we will show that all the
solutions are a product of a radial and a spherical function.
Denote by
𝑛−2 √
𝐽 (𝜌) = sin𝐾2/𝜆 (𝜌)𝑅(𝜌/ 𝜆),
0
and cos𝐾 /𝜆 (𝜌) = sin𝐾 (𝜌). Notice that
/𝜆
00 𝐾
sin𝐾 /𝜆 (𝜌) = − 𝜆
sin𝐾 /𝜆 (𝜌)
and
𝐾
cos2𝐾 /𝜆 (𝜌) + sin2𝐾 /𝜆 (𝜌) = 1.
𝜆
22
Then we have that
0 = (sin𝐾 /𝜆 (𝜌)𝐽 0 (𝜌)) 0+ (1.2.5)

2
𝑛−2
2 𝐾/𝜆 + 1 sin2𝐾 /𝜆 (𝜌) − (𝑚 + 𝑛/2 − 1) 2 + 𝑛−2
2 𝐾/𝜆
+ 𝐽 (𝜌).
sin𝐾 /𝜆 (𝜌)
In the case that 𝐾 = 0 we get that (1.2.5) becomes
(𝑚 + 𝑛/2 − 1) 2

0 0
0 = (𝑟 𝐽𝑚+𝑛/2−1 (𝜌)) + 𝑟 − 𝐽𝑚+𝑛/2−1 (𝜌). (1.2.6)
𝑟
The equation above is known as the Bessel equation and the solutions is known as
Bessel functions. Denote by 𝑗 𝑚+𝑛/2−1
𝑘 the 𝑘’th zero of the Bessel function 𝐽𝑚+𝑛/2−1 .
Then we have the solution

𝑛−2 𝑟
𝑢(𝑟, 𝜃) = 𝑟 𝐽𝑚+𝑛/2−1 𝑗 𝑚+𝑛/2−1
2 𝑌𝑚 (𝜃),
𝑟0
where 𝑌𝑚 is a spherical harmonic function with eigenvalue 𝑚(𝑚 + 𝑛 − 2). In this
case we have that the Dirichlet eigenvalues are
𝑘 !2
𝑗 𝑚+𝑛/2−1
𝜆= .
𝑟0
𝑘,1
If we denote 𝑗 𝑚+𝑛/2−1 0
to be the 𝑘’th zero of 𝐽𝑚+𝑛/2−1 then the solution to the
Neumann problem is given by

𝑛−2 𝑘,1 𝑟
𝑢(𝑟, 𝜃) = 𝑟 2 𝐽𝑚+𝑛/2−1 𝑗 𝑚+𝑛/2−1 𝑌𝑚 (𝜃).
𝑟0
In this case the Neumann eigenvalues the first eigenvalue is 0 and the rest are given
by
𝑘,1 2
© 𝑗 𝑚+𝑛/2−1 ª
𝜇= ® .
𝑟0
« ¬
1.2.5 Steklov Problem
Another problem concerning the Laplace operator is the Steklov problem. The
Steklov problem on a 𝐶 1 -domain Ω ⊂ 𝑀 in a Riemannian manifold (𝑀, g) is the
problem (
Δℎ = 0 in Ω
, (1.2.7)
ℎ 𝑛 = 𝜎ℎ on 𝜕Ω
23
for 𝜎 ∈ R. Notice that for 𝜎 = 0 the constant function ℎ = 1 is a trivial solution to

(1.2.7). Even though this formulation of the Steklov problem is not an eigenvalue
problem, we will for the sake of simplicity refer to 𝜎 as eigenvalues. We can also
interpret the eigenvalues 𝜎 to be eigenvalues for the Dirichlet to Neumann map,
see [GP17, Sec. 5.1.2]. For the Steklov problem, we have that by [Agr06] that the
eigenvalues can be ordered in the fashion
0 = 𝜎1 < 𝜎2 ≤ · · · → ∞.
The Rayleigh quotient to this problem is given by
∫
| grad 𝑢| 2 dvol
𝜎𝑚 = inf sup Ω ∫ .
𝑉 ⊂𝐻 1 (Ω) 𝑢 ∈𝑉 \{0} 𝑢 2 dS
dim(𝑉 )=𝑚 𝜕Ω
Example 1.2.10 (Steklov Problem on the Cylinder, [CEG11, Lem. 6.1]). For a
compact Riemannian manifold (𝑀, g) we consider the cylinder S𝐿 (𝑀) = 𝑀 ×
[−𝐿, 𝐿] with the usual product metric g 𝐿 = g + 𝑑𝑡 ⊗ 𝑑𝑡. The Steklov problem on
the cylinder (S𝐿 (𝑀), g 𝐿 ) is given by
(
Δ S𝐿 ( 𝑀 ) ℎ(𝑥, 𝑡) = 0 in S𝐿 (𝑀)
. (1.2.8)
ℎ 𝑛 (𝑥, 𝑡) = 𝜎ℎ(𝑥, 𝑡) on 𝜕S𝐿 (𝑀)
As always, we have for 𝜎 = 0 the constant solution ℎ(𝑥, 𝑡) = 1. For ℎ(𝑥, 𝑡) = 𝑡 we
have the corresponding eigenvalue 𝜎 = 1/𝐿. Define
ℎ 𝑠 (𝑥, 𝑡) = sinh(𝑘𝑡)𝑢(𝑥), ℎ 𝑐 (𝑥, 𝑡) = cosh(𝑘𝑡)𝑢(𝑥),
where 𝑢 ∈ 𝐶 ∞ (𝑀) is a solution to Δ𝑢 + 𝑘 2 𝑢 = 0 on 𝑀 for some 𝑘 2 > 0. Then
ℎ 𝑠 and ℎ 𝑐 solve (1.2.8) with eigenvalues 𝜎𝑠 = 𝑘 coth(𝑘 𝐿) and 𝜎𝑐 = 𝑘 tanh(𝑘 𝐿),
respectively. All the solutions to (1.2.8) are on the form given above.
1.2.6 Rellich Identity

Let 𝑢 be a solution to (1.2.2) with Ω ⊂ R𝑛 . In [Rel40] Rellich showed that the
associated eigenvalue 𝜆 to 𝑢 can be given by the integral
∫
2
1 𝜕Ω 𝑢 𝑛 h𝑥, ni dS
𝜆= ∫ . (1.2.9)
2 𝑢 2 dvol
Ω
One refers to (1.2.9) as the Rellich identity for the Dirichlet problem. In this section
we will derive Rellich identities on general Riemannian manifolds (𝑀, g). In this
section let 𝑟 denote the radial distance function with respect to a point 𝑝 ∈ 𝑀
defined in Section 1.1.5. We will work inside the domain Ω ⊂ 𝐵Inj( 𝑝) ( 𝑝) so that
𝑟 2 is a smooth function. Furthermore, we will assume that the boundary of Ω is
𝐶 2 . We will start by showing a general kind of inequality first shown in [HS20].
24
Theorem 1.2.11 ([HS20][Thm. 3.1]). Assume that 𝑢 ∈ 𝐶 2 (Ω) ∩ 𝐶 1 (Ω) satisfies

Δ𝑢 + 𝜆𝑢 = 0. Then
∫
𝜆 2 1
𝑢 Δ(𝑟 2 ) + ∇2𝑟 2 (grad 𝑢, grad 𝑢) − | grad 𝑢| 2 Δ𝑟 2 dvol
Ω 2 2
∫
= 𝑢 𝑛 hgrad 𝑟 2 , grad 𝑢i + (𝜆𝑢 2 − | grad 𝑢| 2 )h𝑟 grad 𝑟, ni dS .
𝜕Ω
Proof. Notice that we have by assumption that

∫
0 = (Δ𝑢 + 𝜆𝑢)hgrad 𝑟 2 , grad 𝑢i dvol . (1.2.10)
Ω
The goal of the proof is the expand (1.2.10) appropriately and obtain the claim. A
straightforward computation shows that
∫
𝐼𝜆 = 𝜆 𝑢hgrad 𝑟 2 , grad 𝑢i dvol
∫Ω
𝜆
= hgrad 𝑟 2 , grad 𝑢 2 i dvol
2 Ω
∫ ∫
2 𝜆 2
= 𝜆𝑢 h𝑟 grad 𝑟, ni dS − 𝑢 Δ(𝑟 2 ) dvol .
𝜕Ω Ω 2
We now turn to the quantity

∫
𝐼Δ = Δ𝑢hgrad 𝑟 2 , grad 𝑢i dvol .
Ω
By the divergence theorem we obtain

∫ ∫
𝐼Δ = 𝑢 𝑛 hgrad 𝑟 2 , grad 𝑢i dS − hgradhgrad 𝑟 2 , grad 𝑢i, grad 𝑢i dvol
𝜕Ω Ω
= 𝐼 𝜕Ω − 𝐼Ω .
We can simplify the term 𝐼Ω to become
∫
𝐼Ω = h∇grad 𝑢 grad 𝑟 2 , grad 𝑢i + hgrad 𝑟 2 , ∇grad 𝑢 grad 𝑢i dvol
∫Ω
= ∇2𝑟 2 (grad 𝑢, grad 𝑢) + ∇2 𝑢(grad 𝑟 2 , grad 𝑢) dvol .
Ω
Using that
div(| grad 𝑢| 2 grad 𝑟 2 ) = | grad 𝑢| 2 div(grad 𝑟 2 ) + hgrad 𝑟 2 , grad | grad 𝑢| 2 i
= | grad 𝑢| 2 Δ(𝑟 2 ) + 2h∇grad 𝑟 2 grad 𝑢, grad 𝑢i
= | grad 𝑢| 2 Δ(𝑟 2 ) + 2∇2 𝑢(grad 𝑟 2 , grad 𝑢),
25
we get from another application of the divergence theorem that

∫
1
𝐼Δ = − ∇2𝑟 2 (grad 𝑢, grad 𝑢) + div(| grad 𝑢| 2 grad 𝑟 2 ) dvol
Ω∫ 2
∫
1 2 2
+ | grad 𝑢| Δ(𝑟 ) dvol + 𝑢 𝑛 hgrad 𝑟 2 , grad 𝑢i dS
2 Ω 𝜕Ω
∫
1
=− ∇ 𝑟 (grad 𝑢, grad 𝑢) − | grad 𝑢| 2 Δ(𝑟 2 ) dvol
2 2
2
∫Ω
+ 𝑢 𝑛 hgrad 𝑟 2 , grad 𝑢i − | grad 𝑢| 2 h𝑟 grad 𝑟, ni dS .
𝜕Ω
Adding together 𝐼Δ and 𝐼𝜆 proves the result.
In the case of Dirichlet and Neumann problems in R𝑛 we get the following

corollary.
Corollary 1.2.12. Assume that Ω ⊂ R𝑛 has 𝐶 ∞ -boundary. If 𝑢 ∈ 𝐶 2 (Ω) ∩ 𝐶 0 (Ω)

solves (1.2.2) with eigenvalue 𝜆, then
∫
1 𝜕Ω
𝑢 2𝑛 h𝑥, ni dS
𝜆= ∫ .
2 𝑢 2 dvol
Ω
In the case that 𝑢 solves (1.2.3) we have that

∫
𝜕Ω
| grad𝜕Ω 𝑢| 2 dS
𝜇=∫ ∫ ,
𝜕Ω
𝑤2 dS −2 Ω 𝑤 2 dvol
where grad𝜕Ω 𝑢 = grad 𝑢 − 𝑢 𝑛 n.
Notice that using the comparison Theorem 1.1.7 one can get an inequality
version of Corollary 1.2.12, see Lemma [HS20][Lem. 3.7].
1.3 Almgren’s Frequency Function

In the 70’s Almgren [Alm79] noticed that harmonic functions on R𝑛 satisfy a
convexity property. The convexity property implies some unique continuation
results which are referred to as quantitative unique continuation. In both Paper A
and Paper B we will study these kinds of convexity result or related inequalities.
26
1.3. Almgren’s Frequency Function
1.3.1 Geometric Convexity

Definition 1.3.1. A continuous function 𝑓 : (0, ∞) → (0, ∞) is said to be geomet-
rically convex if for all 𝑠 ∈ [0, 1] and 𝑎, 𝑏 ∈ (0, ∞) we have that

𝑓 𝑎 𝑠 𝑏 1−𝑠 ≤ 𝑓 (𝑎) 𝑠 𝑓 (𝑏) 1−𝑠 .
A special case of the geometric convexity is the inequality;

√ p
𝑓 𝑟 𝑠 ≤ 𝑓 (𝑟) 𝑓 (𝑠). (1.3.1)
In fact, if (1.3.1) holds for all positive 𝑟 and 𝑠 then 𝑓 is geometrically convex. This
can be showed by using the result of Sierpinski, which says that for real valued
measurable functions midpoint convexity implies convexity (see [Don69, p.12]).
Geometrically convex can be reformulated in terms of a differential inequality. The
proof of this is straightforward and hence will be omitted.
Proposition 1.3.2. Assume that 𝑓 : (0, ∞) → (0, ∞) is twice differentiable. Then

the following are equivalent;
1. The function 𝑓 is geometrically convex.
2. The composite function log ( 𝑓 (𝑒 𝑥 )) is convex.

𝑡 𝑓 0 (𝑡)
3. The function 𝑓 (𝑡) is increasing.
4. The inequality 𝑓 00 (𝑡) 𝑓 (𝑡) + 𝑓 0 (𝑡) 𝑓 (𝑡) /𝑡 − 𝑓 0 (𝑡) 2 ≥ 0 holds.
It is obvious that the product of two geometrically convex functions is again

geometrically convex. Moreover, with a bit more work, it can also be shown that
the sum of two geometrically convex functions is again geometrically convex.
1.3.2 Three Spheres Theorem and the Frequency Function

We will start by proving the classical result for harmonic functions which can be
found in [Han07]. The reader should consult [Man13] for a generalization of the
result to Riemannian manifolds. Let ℎ : 𝐵 𝑅 (0) ⊂ R𝑛 → R be a harmonic function
on the ball 𝐵 𝑅 (0) of radius 𝑅 centered at 0. Define the function 𝐻 ℎ : (0, 𝑅) →
[0, ∞) by ⨏ ∫
2 1
𝐻 ℎ (𝑟) = ℎ dS = ℎ2 dS .
S𝑟𝑛−1 vol 𝑛−1 (S𝑟𝑛−1 ) S𝑟𝑛−1
We will later do the convexity results without averaging the integral, however, the
formulas become slightly simpler when averaging the integral.
27
Proposition 1.3.3. The function 𝐻 ℎ (𝑟) is geometrically convex for all 0 < 𝑟 < 𝑅.
Proof. We will show that 𝐻 ℎ satisfies the last inequality in Proposition 1.3.2.
Define ⨏ ⨏
𝐷 ℎ (𝑟) = 𝑟 | grad ℎ| 2 dvol = ℎℎ 𝑛 dS, (1.3.2)
𝐵𝑟 (0) S𝑟𝑛−1
where ℎ 𝑛 is the normal derivative of ℎ with respect to the outward unit normal
vector on the sphere. Using the divergence theorem and the fact that ℎ is harmonic
gives us
• 𝐻 ℎ0 (𝑟) = 2𝐷 ℎ (𝑟) ,
−2(𝑛−1)
⨏
• 𝐻 ℎ00 (𝑟) = 𝑟 𝐷ℎ (𝑟) + 2 S𝑟𝑛−1
|grad ℎ| 2 dS .
Furthermore, we have that
⨏ ⨏
2
2 |grad ℎ| 2 dS = h𝑥 |grad ℎ| 2 , grad 𝑟i dS
𝑛−1 𝑟 𝑛−1
⨏ 𝑟
S𝑟 S
2
= div |grad ℎ| 2 𝑥 dvol
𝑟 𝐵𝑟 (0)
⨏
2𝑛𝐷 ℎ (𝑟)
= +4 𝑥 𝑖 ℎ 𝑗 ℎ𝑖 𝑗 dvol .
𝑟 𝐵𝑟 (0)
Using integration by parts we get that

⨏ ⨏ ⨏
4
4 𝑥 𝑖 ℎ 𝑗 ℎ𝑖 𝑗 dvol = −4 𝜕 𝑗 𝑥 𝑖 ℎ 𝑗 ℎ𝑖 dvol + 2 𝑥 𝑗 𝑥 𝑖 ℎ 𝑗 ℎ𝑖 dS
𝐵𝑟 (0) 𝐵𝑟 (0) 𝑟 S𝑟𝑛−1
⨏
−4
= 𝐷 ℎ (𝑟) + 4 ℎ2𝑛 dS .
𝑟 𝑛−1
S𝑟
This implies that ⨏

−2
𝐻 00 (𝑟) = 𝐷 ℎ (𝑟) + 4 ℎ2𝑛 dS .
𝑟 S𝑟𝑛−1
Plugging everything into Proposition 1.3.2 we obtain
⨏
00 0 0 2
𝐻 (𝑟)𝐻 (𝑟) + 𝐻 (𝑟)𝐻 (𝑟)/𝑟 − 𝐻 (𝑟) = 4 ℎ2𝑛 dS 𝐻 ℎ (𝑟) − 4𝐷 ℎ (𝑟) 2 .
S𝑟𝑛−1
The claim now follows from the Cauchy-Schwartz inequality.
Remark 1.3.4. Note that the proof hinges on the fact that
⨏ 𝑛−2
|grad ℎ| 2 − 2ℎ2𝑛 dS = 𝐷 ℎ (𝑟) .
S𝑟𝑛−1 𝑟
We will do a similar trick in Paper A, however in that case the equality will be
replaced by an inequality.
28
By setting 𝑠 = 1/2, 𝑎 = 𝑟/4, and 𝑏 = 𝑟 into the definition of geometrically

convex, we get the important special case
𝐻 ℎ (𝑟/2) ≤ 𝐻 ℎ (𝑟/4) 1/2 𝐻 ℎ (𝑟) 1/2 . (1.3.3)
This and similar result will throughout this thesis be known as three spheres
theorems, referring to the three spheres involved.
Definition 1.3.5. The frequency function of a harmonic function ℎ is defined to be
𝑟 𝐷 ℎ (𝑟)
𝑁 ℎ (𝑟) = ,
𝐻 ℎ (𝑟)
where 𝐷 ℎ (𝑟) is defined in Equation (1.3.2).
The frequency function was introduced in [Alm79] by Almgren. Note that
log(𝐻 ℎ (𝑒 𝑥 )) 0
𝑁 ℎ (𝑒 𝑥 ) = .
2
Hence by the second equivalence in Proposition 1.3.2 together with Proposition
1.3.3, we have that 𝑁 ℎ (𝑟) is increasing.
Example 1.3.6. We can show that for homogeneous harmonic polynomials the
frequency function is constant. Denote by ℎ : R𝑛 → R a homogeneous harmonic
polynomial of degree 𝑚 on the form
Õ
ℎ(x) = 𝑎 𝑖1 ,...,𝑖𝑛 𝑥 1𝑖1 · · · 𝑥 𝑛𝑖𝑛 .
𝑖1 +...+𝑖𝑛 =𝑚
Then we have that ⨏

𝑚𝐻 ℎ (𝑟)
𝐷 ℎ (𝑟) = ℎℎ 𝑛 dS =
S𝑟𝑛−1 𝑟
and ⨏
𝑚 2 𝐻 ℎ (𝑟)
ℎ2𝑛 dS = .
S𝑟𝑛−1 𝑟2
Hence using the computations done in the proof of Proposition 1.3.2 we get that
𝐻 ℎ00 (𝑟)𝐻 ℎ (𝑟) − 𝐻 ℎ0 (𝑟) 2 + 𝐻 ℎ0 (𝑟) 𝐻 ℎ (𝑟) /𝑟 = 0.
By Proposition 1.3.2 this means that (1.3.3) becomes an equality. Additionally,

we have that the frequency function in this case coincides with the degree of the
polynomial ℎ. This is the motivation for calling 𝑁 ℎ the frequency function.
29
In the cases when ℎ is not a polynomial, we can still think of the frequency 𝑁 ℎ
as being the “degree” of ℎ. This can be made precise by noticing that
𝐻 ℎ0 (𝑟) 2𝑁 ℎ (𝑟)
(log 𝐻 ℎ (𝑟)) 0 = = .
𝐻 ℎ (𝑟) 𝑟
Integrating the expression from 𝑠 to 𝑟, for thereafter taking the exponential on both
sides, we get ∫ 𝑟
𝐻 ℎ (𝑟) 𝑁 ℎ (𝑥)
= exp 2 𝑑𝑥 .
𝐻 ℎ (𝑠) 𝑠 𝑥
Using that 𝑁 ℎ is increasing we get the following corollary:
Corollary 1.3.7. For any 0 < 𝑠 < 𝑟 and any harmonic function ℎ we have that
𝑟 2𝑁ℎ (𝑠) 𝐻 ℎ (𝑟) 𝑟 2𝑁ℎ (𝑟 )
≤ ≤ .
𝑠 𝐻 ℎ (𝑠) 𝑠
In particular, for 𝑟 = 2𝑠 we have the following doubling condition
𝐻 ℎ (2𝑟)
22𝑁ℎ (𝑟 ) ≤ ≤ 22𝑁ℎ (2𝑟 ) .
𝐻 ℎ (𝑟)
Thinking of the example 1.3.6 where ℎ is a homogeneous harmonic polynomial
it should come as no surprise that the frequency controls the vanishing order. Recall
that a function 𝑢 has vanishing order 𝑚 at 𝑥 0 if 𝑚 is the largest value where there
exist 𝑟 𝑚 and 𝑐 𝑚 such that
|𝑢(𝑥)| ≤ 𝑐 𝑚 |𝑥0 − 𝑥| 𝑚
for all |𝑥 0 − 𝑥| < 𝑟 𝑚 .
Corollary 1.3.8. The value of lim𝑟 →0+ 𝑁 ℎ (𝑟) is equal to the vanishing order of ℎ
at zero.
Proof. By fixing 𝑟 in Corollary 1.3.7 we get that the vanishing order for ℎ at zero is
at least lim𝑟 →0+ 𝑁 ℎ (𝑟) . Note that the limit lim𝑟 →0+ 𝑁 ℎ (𝑟) always exists, since 𝑁 ℎ is
bounded from below by zero and increasing. Note that if ℎ (𝑥) = ∞
Í
𝑗=0 𝑝 𝑗 (𝑥) is the
Taylor series of ℎ at zero, then 𝑝 𝑗 must be homogeneous harmonic polynomials.
If ℎ vanishes at 𝑥 0 = 0 of order 𝑚 we need 𝑝 𝑗 = 0 for all 𝑗 < 𝑚. Hence
ℎ (𝑥) = 𝑝 𝑚 (𝑥) + ∞
Í
𝑗=𝑚+1 𝑝 𝑗 (𝑥) and
⨏
𝑟 S𝑟𝑛−1
𝑚 𝑝 𝑚 (𝑥) 2 /𝑟 + "higher order terms"/𝑟 dS
lim 𝑁 ℎ (𝑟) = lim+
𝑟 →0+ 𝑟 →0
⨏
2
= 𝑚.
S𝑛−1
𝑝 𝑚 (𝑥) + "higher order terms" dS
𝑟
30
1.3.3 Doubling Index

In this section we extend the previous results to the 𝐿 ∞ -setting and also define the
doubling index of a harmonic function. For the ball 𝐵 𝑅 (0) ⊂ R𝑛 we consider a
harmonic function ℎ : 𝐵 𝑅 (0) → R. Define the function 𝑀ℎ : (0, 𝑅) → [0, ∞) by
𝑀ℎ (𝑟) = max |ℎ|.

𝐵𝑟 (0)
From local equivalence of 𝐿 𝑝 norms, see [GT01, p. 194 Thm. 8.17], we get that
s⨏
p
2
3𝑟
𝑀ℎ (𝑟) ≥ 𝐻 ℎ (𝑟) ≥ 𝐶 ℎ dvol ≥ 𝑐𝑀ℎ , (1.3.4)
𝐵𝑟 (0) 4
for some constant 𝑐 < 1. Using the geometric convexity we can show the following
quantitative unique continuation result:
Proposition 1.3.9. For any 𝛼 ∈ (0, 1) we have that
1 𝛼 1−𝛼

𝛼 1−𝛼 4 4
𝑀ℎ 𝑟 𝑠 ≤ 𝑀ℎ 𝑟 𝑀ℎ 𝑠 , (1.3.5)
𝑐 3 3
whenever 𝑠 < 𝑟 < 3/4𝑅. If we have 𝜔 ⊂⊂ 𝐵 ⊂⊂ Ω for domains 𝜔, 𝐵, Ω ⊂ 𝐵 𝑅 (0)
we get that for some 𝛼 ∈ (0, 1) we have that
𝛼 1−𝛼
max |ℎ| ≤ 𝐶 max |ℎ| max |ℎ| ,
𝐵 𝜔 Ω
where 𝐶 = 𝐶 (𝜔, 𝐵, Ω).

Proof. The inequality (1.3.5) follows from (1.3.4) and Proposition 1.3.2.
Without loss of generality we can assume that maxΩ |ℎ| = 1. Cover 𝐵 by balls
such that
𝑁
Ø
𝐵⊂ 𝐵𝑟𝑖 (𝑦 𝑖 ) ⊂⊂ Ω.
𝑖=1
Our goal is to show that for each 𝑖 there exists a ball 𝐵𝜌𝑖 (𝑦 𝑖 ) ⊂ 𝜔 such that
𝛼𝑖
max |ℎ| ≤ 𝐶 max |ℎ| .
𝐵𝑟𝑖 ( 𝑥𝑖 ) 𝐵𝜌𝑖 ( 𝑦𝑖 )
We will do this by moving along the piece-wise linear curve 𝑙𝑖 (𝑡) inside Ω such
that 𝑙𝑖 (0) = 𝑥 𝑖 and 𝑙𝑖 (1) = 𝑦 𝑖 . Denote by

3
𝑅𝑖 = min dist ({𝑙𝑖 (𝑡) : 𝑡}, 𝜕Ω) , 𝜌𝑖 .
4
31
𝑅𝑖 /2 𝑙𝑖 (𝑡)
𝑥𝑖 𝑥3
𝑥1 𝑥2
Figure 1.1: The figure illustrates three iterations of the chain argument.
Without loss of generality assume that 𝑟 𝑖 ≤ 𝑅𝑖 /2, if not we can use (1.3.5) to show
that ! 𝛼0

𝐵𝑟𝑖 ( 𝑥1 ) 𝐵 𝑅𝑖 /2 ( 𝑥𝑖 )
Go along the curve 𝑙𝑖 to the first point 𝑥 1 such that dist 𝑥 𝑖 , 𝑥 1 = 𝑅𝑖 /4. Then we

have that
𝐵 𝑅𝑖 /4 (𝑥 𝑖 ) ⊂ 𝐵 𝑅𝑖 𝑥 1 .
Let 𝛼1 be such that

𝛼1 1−𝛼1
3 1
𝑅𝑖 /2 = 2 𝑅𝑖 ,
4 4
then we have by (1.3.5) that
! 𝛼1 ! 𝛼1
max |ℎ| ≤ 𝐶 max |ℎ| ≤𝐶 max |ℎ| .
𝐵 𝑅𝑖 /2 ( 𝑥 𝑖 ) 𝐵 𝑅𝑖 /4 ( 𝑥𝑖 ) 𝐵 𝑅𝑖 ( 𝑥 1 )
Doing the same procedure until we reach 𝑦 𝑗 we obtain
𝛼𝑛 𝑗
1
𝐵 𝑅𝑖 /2 ( 𝑥 𝑛 𝑗 ) 𝐵 𝑅𝑖 ( 𝑦𝑖 )
32
Definition 1.3.10. The doubling index Nℎ : (0, 𝑅/2) → [0, ∞) of a harmonic

function ℎ : 𝐵 𝑅 (0) → R is defined by

𝑀ℎ (2𝑟)
Nℎ (𝑟) = log .
𝑀ℎ (𝑟)
At first glance the log-part of the doubling index might seem unnatural. How-
ever, this makes the frequency function 𝑁 ℎ and the doubling index Nℎ comparable:
Note that by (1.3.5) and Corollary 1.3.7 there exist constants 𝐶, 𝑐 > 0 such that
v
u
©t 𝐻 ℎ (2𝑟) ª
𝑁 ℎ (𝑟) − 𝑐 ≤ log
®®
𝐶𝐻 ℎ 4𝑟3
« ¬
s !
𝐶𝐻 ℎ (4𝑟)
≤ Nℎ (𝑟) ≤ log ≤ log (4) 𝑁 ℎ (4𝑟) + 𝑐.
𝐻 ℎ (𝑟)
1.3.4 Quantitative Unique Continuation Results

One can also show quantitative unique continuation results for more general solu-
tions to elliptic equations. In this section we will discuss some of the results in this
direction.
Helmholtz Equation
We can use the extension approach to get a result for the Helmholtz equation. Let
𝑢 be a solution to a
Δ𝑢 + 𝑘 2 𝑢 = 0
on a ball 𝐵 𝑅 (0) ⊂ R𝑛 . Using Proposition 1.3.9 together with the associated
harmonic function given in Section 1.2.2 one get for some 𝛼 ∈ (0, 1) and all 𝑟 < 𝑅
that 𝛼 1−𝛼
max |𝑢| ≤ 𝐴 exp(𝐵𝑘𝑟) max |𝑢| max |𝑢| , (1.3.6)
𝐵𝑟/2 (0) 𝐵𝑟/4 (0) 𝐵𝑟 (0)
for some constants 𝐴 and 𝐵 independent of 𝑢.

Proposition 1.3.2 have been generalized to Riemannian manifolds with bounded
curvature in [Man13]. This result will be discussed in details in Section A.3.1.
One can use the associated harmonic function in the case of Riemannian manifolds
to show a similar result to (1.3.6), see [Man13, Thm. 3.2]. We will show in
Paper B that one cannot replace the exponential function exp(𝐵𝑘𝑟) in (1.3.6) with
a function with slower growth. We will also show that the same is true for the
analogous problems on the sphere and the hyperbolic space.
33
Elliptic Equations
Consider 𝐵 𝑅 (0) ⊂ R𝑛 for 𝑛 ≥ 2 and let 𝐴 : 𝐵 𝑅 (0) × R𝑛 → R𝑛 be such that for
each 𝑥 ∈ 𝐵 𝑅 (0) the map 𝑣 ↦→ 𝐴(𝑥, 𝑣) is a symmetric matrix. For convenience, we
will simply denote the map 𝑣 ↦→ 𝐴(𝑥, 𝑣) by 𝐴(𝑥). Assume for some 𝐿 > 0 that the
matrix entries 𝑎 𝑖 𝑗 (𝑥) satisfy the Lipschitz condition
|𝑎 𝑖 𝑗 (𝑥) − 𝑎 𝑖 𝑗 (𝑦)| ≤ 𝐿|𝑥 − 𝑦| for all 0 ≤ 𝑖, 𝑗 ≤ 𝑛. (1.3.7)
Moreover, assume that

𝐴(0) = Id. (1.3.8)
We will assume that the elliptic operator
div( 𝐴(𝑥) grad ·)
is uniformly elliptic, meaning that for some 𝛾 > 0 we have that
𝛾 −1 |𝑣| 2 ≤ h𝐴(𝑥)𝑣, 𝑣i ≤ 𝛾|𝑣| 2 , 𝑣 ∈ R𝑛 . (1.3.9)
For a non-zero 𝑥 ∈ 𝐵 𝑅 (0) we define
h𝐴(𝑥)𝑥, 𝑥i
𝜇(𝑥) = .
|𝑥| 2
Notice that for the Laplacian Δ we have 𝜇 ≡ 1. For a solution 𝑢 : 𝐵 𝑅 (0) → R to

the problem div( 𝐴(𝑥) grad 𝑢) = 0 we define the function
∫
𝐻𝑢 (𝑟) = 𝑢 2 𝜇 dS, 𝑟 < 𝑅. (1.3.10)
S𝑟𝑛−1
This allows us to define the corresponding frequency function
𝑟 𝐻𝑢0 (𝑟)
𝑁𝑢 (𝑟) = , 𝑟 < 𝑅. (1.3.11)
𝐻𝑢 (𝑟)
In this case we have the classical result by Garofalo and Lin:
Theorem 1.3.11 ([GL86]). Let 𝑢 be a solution to the equation div( 𝐴(𝑥) grad 𝑢) = 0
on the ball 𝐵 𝑅 (0) as above. Then there exists a constant 𝐶 = 𝐶 (𝑛, 𝛾, 𝐿, 𝑅) such
that the function
exp(𝐶𝑟)𝑁𝑢 (𝑟)
is non-decreasing for 𝑟 ∈ (0, 𝑅0 ), where 𝑅0 is a fixed constant 𝑅0 < 𝑅 .
34
Using this result Garofalo and Lin showed a doubling inequality for solutions
to elliptic equations. This result was generalized to more general elliptic solution
in [Kuk98] as follow: Let 𝑤 ∈ 𝐶 2 (𝐵 𝑅 (0)) be a solution to
div( 𝐴(𝑥) grad 𝑤) (𝑥) + 𝑘 2 𝑤(𝑥) = 0, (1.3.12)
where the matrix 𝐴 satisfies (1.3.7), (1.3.8), and (1.3.9).
Theorem 1.3.12. Let 𝑤 be a solution to (1.3.12). Define 𝑁 𝑤 as in (1.3.11). Then

there exists a constant 𝐶 = 𝐶 (𝑛, 𝛾, 𝐿, 𝑅) independent of 𝑤 such that the function

exp(𝐶𝑟) 𝑁 𝑤 (𝑟) + 𝑘 2 + 1
is non-decreasing for 𝑟 ∈ (0, 𝑅0 ), where 𝑅0 is a fixed constant 𝑅0 < 𝑅 .
1.3.5 Inverse Three Sphere Theorem

We will show in Paper B that one cannot do better than in (1.3.6). Interestingly,
one can do better if one works with annuli instead of balls. This will also be the
focus of Paper B. Let us start with some initial results. Consider the problem
(
Δ𝑢 + 𝑉 (𝑥)𝑢 = 0 𝑥 ∈ 𝐵1 (0) ⊂ R𝑛
,
𝑢=0 𝑥 ∈ S1𝑛−1
where 𝑉 ∈ 𝐶 ∞ (𝐵1 (0)).
Theorem 1.3.13 ([ALM16, Thm. 1.2 and Thm. 1.3]). Let Ω be a non-empty open
subset of 𝐵1 (0) such that Ω∩S1𝑛−1 ≠ ∅. Then there exists a constant 𝐶1 = 𝐶1 (𝑉, Ω)
such that
k𝑢k 2𝐿 2 (𝐵 (0)) ≤ 𝐶1 k𝑢k 2𝐿 2 (Ω) .
1
Assume that Γ is a non-empty open subset of S1𝑛−1 . Then there exists a constant
𝐶2 = 𝐶2 (Γ, 𝑉) such that
k𝑢k 2𝐻 1 (𝐵 ≤ 𝐶2 k𝑢 𝑛 k 2𝐿 2 (Γ) .
1 (0))
In the case that we not have Dirichlet conditions on the boundary, [HI04]
showed the increased stability for the Helmholtz equation: Let us consider the
cylinder Ω = 𝐵𝑟 (0) × (0, 2𝑟) and let Γ be an open set in (𝜕𝐵𝑟 (0)) × (0, 2𝑟). Denote
by
Ω(𝑑) = 𝐵𝑟 (0) × (𝑑, 2𝑟) ⊂ Ω, 𝑑 < 2𝑟.
35
Theorem 1.3.14 ([HI04, Thm. 1.2]). Let 𝑢 ∈ 𝐶 2 (Ω) be a solution to the Helmholtz
equation
Δ𝑢 + 𝑘 2 𝑢 = 0 on Ω,
with wave number 𝑘 ≥ 1. Assume that

1 1
√ 𝑘+ k𝑢k 𝐿 2 (Γ) + k grad 𝑢k 𝐿 2 (Γ) ≤ k𝑢k 𝐻 1 (Ω) .
𝑑 𝑑
Then for some 𝐶 > 0 and 𝛼 ∈ (0, 1) we have
k𝑢k 𝐻 1 (Ω) 1−𝛼

𝛼
1 1
k𝑢k 𝐻 1 (Ω(𝑑)) ≤𝐶 √ 𝑘+ k𝑢k 𝐿 2 (Γ) + k grad 𝑢k 𝐿 2 (Γ) .
𝑑 𝑑 𝑑2
36
Chapter 2
Summary of Papers
2.1 Paper A: Convexity Properties of Harmonic Functions

on Parameterized Families of Hypersurfaces [Ber21]
In Paper A we are going to generalize Proposition 1.3.3 to more general families
of compact hypersurfaces. The main result given in Theorem A.2.5 produces a
versatile differential inequality for the 𝐿 2 -norms of harmonic functions over these
hypersurfaces. As for solutions to elliptic equations in Section 1.3.4, we will get an
additional weight on the integrals. To get a parameterized family of hypersurfaces
we will use level surfaces of functions, see Example 1.1.2 and Example 1.1.4.
As a consequence, we will in Theorem A.3.1 slightly improve the result [Man13,
Thm. 2.2] in the positive curvature case for spheres in Riemannian manifolds. The
paper will utilize the comparison result given in Theorem 1.1.7. We apply the main
result to the level surfaces of 1-homogeneous functions, e.g. ellipses with constant
eccentricity. Moreover, we will also consider tori as level surfaces of the distance
function from a fixed lower-dimensional sphere.
2.2 Paper B: On the Three Ball Theorem for Solutions of

the Helmholtz Equation [BM21]
As we saw in Section 1.3.4 we can get a quantitative unique continuation result
for the solutions to the Helmholtz equation. In Paper B we will use the solutions
given in Example 1.2.9 to show Theorem B.2.2 stating that we cannot replace the
exponential function in (1.3.6) with a function of slower growth. We will show the
optimal growth for all the model spaces. To do this we will use Sturm-Liouville
theory, which will be summarized in Appendix B.5. When using annuli instead of
balls, we will show in (B.1.5) that the growth function in (1.3.6) can be chosen to
37
Chapter 2. Summary of Papers
be constant. The result obtained is similar to those presented in Section 1.3.5.
2.3 Paper C: Kuttler-Sigillito’s Inequalities and Rellich-

Christianson Identities
In Paper C we study eigenvalues of boundary value problems where the bound-
ary conditions are mixed. The first goal is to continue the work of Kuttler and
Sigillito regarding eigenvalue inequalities. We will show in Theorem C.2.3 that
all eigenvalues for the mixed Steklov-Dirichlet problem can be bounded below by
an expression involving the eigenvalues for the Neumann-Dirichlet problem. In
Example C.2.7 we will consider a variant of Example 1.2.10 in the setting of the
mixed Steklov-Dirichlet problem. Throughout Paper C we will heavily use the
Rayleigh quotient, see Section 1.2.4 for examples. The second goal is to derive the
Rellich identity for polytopes with mixed Neumann-Dirichlet conditions. This can
be shown by simplifying Theorem 1.2.11 in the case of polytopes.
2.4 Further Results D: Eigenvalues on Spherically Sym-

metric Manifolds
In Chapter D we are going to explore Dirichlet eigenvalues on balls in spherically
symmetric manifolds. Our main result in Theorem D.2.2is an eigenvalue inequality
relating the eigenvalues on the model spaces to the eigenvalues on balls in R𝑛
described in Example 1.2.9.
38
Part II
Research Papers
Paper A
Convexity Properties of Harmonic Functions on

Parameterized Families of Hypersurfaces
Stine Marie Berge
Published in
The Journal of Geometric Analysis, 2021, Volume 31
Paper A
Convexity Properties of
Harmonic Functions on
Parameterized Families of
Hypersurfaces
Abstract
It is known that the 𝐿 2 -norms of a harmonic function over spheres satis-
fies some convexity inequality strongly linked to the Almgren’s frequency
function. We examine the 𝐿 2 -norms of harmonic functions over a wide
class of evolving hypersurfaces. More precisely, we consider compact level
sets of smooth regular functions and obtain a differential inequality for the
𝐿 2 -norms of harmonic functions over these hypersurfaces. To illustrate our
result, we consider ellipses with constant eccentricity and growing tori in
R3 . Moreover, we give a new proof of the convexity result for harmonic
functions on a Riemannian manifold when integrating over spheres. The
inequality we obtain for the case of positively curved Riemannian manifolds
with non-constant curvature is slightly better than the one previously known.
A.1 Introduction
Since the paper by Almgren [Alm79], the frequency function have been intensively
used to study harmonic functions in R𝑛 and, more generally, solutions to second
order elliptic equations. For a harmonic function ℎ on R𝑛 we let 𝐻 (𝑡) denote the
𝐿 2 -norm of ℎ over the sphere of radius 𝑡. In [Agm66], and later in [Alm79], it was
shown that the function 𝐻 is geometrically convex, i.e.

𝐻 𝑟 𝛼 𝑠1−𝛼 ≤ 𝐻 (𝑟) 𝛼 𝐻 (𝑠) 1−𝛼 , 0 ≤ 𝛼 ≤ 1, 𝑟, 𝑠 > 0. (A.1.1)
43
Paper A. Convexity Properties of Harmonic Functions
Inequality (A.1.1) is equivalent to the statement that the frequency function

𝑡𝐻 0 (𝑡)
𝑁 (𝑡) = , 𝑡>0
𝐻 (𝑡)
is increasing. The notion of frequency function was generalized to solutions of
elliptic operators of divergence form by Garafalo and Lin in [GL86] and was shown
to be almost increasing for 𝑡 < 𝑡 0 . They further used the result to show that the
squares of solutions of the elliptic equations are Muckenhoupt weights on the ball
𝐵 𝑅 with radius 𝑅 > 0.
In the paper of Mangoubi [Man13], a more explicit convexity result on Rie-
mannian manifolds was obtained by using comparison geometry. Using this result
and extending eigenfunctions to harmonic functions, Mangoubi gave a new proof
that a solution 𝑢 to div (grad 𝑢) = −𝑘 2 𝑢 satisfies
𝛼 1−𝛼
𝐶2 𝑟 𝑘
max |𝑢| ≤ 𝐶1 𝑒 max |𝑢| max |𝑢| . (A.1.2)
𝐵𝑟 ( 𝑝) 𝐵2𝑟 ( 𝑝) 𝐵𝑟/2 ( 𝑝)
In (A.1.2) the positive constants 𝐶1 , 𝐶2 and 0 < 𝛼 < 1 only depend on the
dimension and curvature of the Riemannian manifold. Inequality (A.1.2) was first
shown by Donnelly and Fefferman in [DF88].
The main aim of this work is to study the 𝐿 2 -norm of harmonic functions over
families of surfaces, generalizing the geometric convexity inequality (A.1.1). Let
ℎ be a harmonic function on a domain Ω in a Riemannian manifold (M, g) and
fix a point 𝑝 ∈ M. Consider for 𝑅 > 0 a smooth function 𝑓 : Ω → [0, 𝑅) that is
regular and have compact level surfaces. Let
∫
𝐻 (𝑡) = ℎ2 |grad 𝑓 | 𝜎𝑡
𝑓 −1 (𝑡)
be the squared 𝐿 2 -norm of ℎ over the level surface 𝑓 −1 (𝑡) with the weight |grad 𝑓 |.
Our main theorem states that 𝐻 satisfies an inequality of the type
(log 𝐻 (𝑡)) 00 + 𝜏 (𝑡) (log 𝐻 (𝑡)) 0 ≥ 𝜌 (𝑡) , (A.1.3)
where 𝜏 and 𝜌 are independent of ℎ. In fact, the functions 𝜏 and 𝜌 only depend on
explicit estimates on the derivatives of 𝑓 and are given in Theorem A.2.5. These
kinds of inequalities when integrated imply that 𝐻 satisfies a variant of Inequality
(A.1.1). In particular, when 𝑓 is the Riemannian distance function from a fixed
point, we give a new proof of [Man13, Thm. 2.2]. For the case when the curvature
is positive we obtain a slight improvement of his inequality, see Section A.3.1.
Next we illustrate our result by considering 1-homogeneous functions, that is,
functions that satisfy 𝑓 (𝑡x) = 𝑡 𝑓 (x) for 𝑡 > 0. A way to construct 1-homogeneous
44
A.2. The Convexity Result
functions is to choose a compact and star convex (with respect to the origin) set 𝑅
with smooth boundary 𝑆 ⊂ R𝑛 \ {0}. Define a function 𝑓 by 𝑓 (x) = 1 for all x ∈ 𝑆
and extend this to a 1-homogeneous function on the whole R𝑛 \ {0}. In this case,
we will show that there exist constants 𝐴 and 𝐵 such that the function 𝐻 satisfies
𝐴 𝐵
(log 𝐻 (𝑡)) 00 + (log 𝐻 (𝑡)) 0 ≥ − 2 .
𝑡 𝑡
For the special case where 𝑆 is an ellipsoid in R𝑛 we find in Section A.3.2 the
explicit values of 𝐴 and 𝐵.
To give an example of level surfaces not diffeomorphic to the sphere we take
the distance function of the submanifold
q
𝑘 𝑛 𝑘
𝑆 × {0} × · · · × {0} ⊂ R , where 𝑆 = x ∈ R 𝑘+1 2 2
: 𝑥 1 + · · · + 𝑥 𝑘+1 = 1 .
In particular, whenever 𝑘 = 1 and 𝑛 = 3 the level surfaces form a family of tori.

Let 𝑓 (x) = dist x, 𝑆 𝑘 , and 𝐻 be as above. Then for a fixed 0 < 𝜀 < 1 we have
that for all 𝑡 < 1 − 𝜀 the function 𝐻 satisfies (A.1.3) for some 𝐴 and 𝐵. Lastly, in
Section A.3.4 we show that if div (grad 𝑢) = 𝑘 2 𝑢 then the spherical 𝐿 2 -norm of 𝑢
satisfies (A.1.1).
Acknowledgment
The author would like to thank Eugenia Malinnikova for her guidance and Dan
Mangoubi for his insightful suggestions. The author was partially supported by the
BFS/TFS project Pure Mathematics in Norway.
A.2 The Convexity Result

A.2.1 Prerequisites
In this article (M, g) will always denote a smooth Riemannian manifold. The
volume density and its respective divergence will be denoted by vol and div. We
will use the notation ∇ to denote the Levi-Civita connection, and define the Hessian
of a function ℎ ∈ 𝐶 ∞ (M) by
∇2 ℎ (𝑋, 𝑌 ) = ∇𝑋 ∇𝑌 ℎ − ∇ ∇𝑋 𝑌 ℎ = h∇𝑋 grad ℎ, 𝑌 i,
where 𝑋 and 𝑌 are vector fields and grad ℎ denotes the gradient of the function ℎ.
The Laplace operator Δ is given by
Δℎ = div (grad ℎ) = trg ∇2 ℎ (×, ×) ,
45
where trg denotes the trace with respect to the metric g.

The idea of the proof of Theorem A.2.5 is to emulate the proof of the well known
special case (which is presented in details in Section A.3.1): Let ℎ : 𝐵 𝑅 ( 𝑝) ⊂
M → R be a harmonic function on the ball with radius 𝑅 centered at 𝑝 and define
∫
𝐻 (𝑡) = ℎ2 𝜎𝑡 ,
𝑆𝑡
where 𝑆𝑡 is the geodesic sphere centered at 𝑝 with radius 𝑡 and 𝜎𝑡 is its surface
measure. In [Man13] it was shown that 𝐻 satisfies a convexity property, which in
the case of constant curvature spaces is on the form
(log 𝐻 (𝑡)) 00 + log (sin K (𝑡)) 0 (log 𝐻 (𝑡)) 0 ≥ − (𝑛 − 1) K + (𝑛 − 2) min (0, K) ,
where K is the sectional curvature and sin K (𝑟) is the function defined by Equation
(A.3.1) in Section A.3.1.
Our goal is to obtain a similar inequality for other families of parameterized
surfaces than geodesic spheres. Since an important step in [Man13] depends
indirectly on the co-area formula, we will assume that this family of surfaces is
given as the level surfaces of a Lipschitz function 𝑓 : Ω → [0, 𝑅) , where Ω ⊂ M
is an open set. To ensure that the preimages 𝑆𝑡 = 𝑓 −1 (𝑡) are hypersurfaces for
𝑡 ∈ (0, 𝑅), we will assume that every value in (0, 𝑅) is regular (see [Lee13,
Thm. 5.12]), meaning that |grad 𝑓 | > 0 for all 𝑥 ∈ 𝑓 −1 (0, 𝑅). We will also need
that the integral over the hypersurfaces are finite. Thus we assume that the surfaces
𝑆𝑡 are closed manifolds, that is, compact manifolds without boundary. Finally,
to be able to use the divergence theorem on the surfaces 𝑆𝑡 we will assume that
𝑅𝑡 = 𝑓 −1 ([0, 𝑡)) is open and compactly embedded in Ω for all 𝑡 ∈ (0, 𝑅) and 𝑆𝑡 is
given as the boundary of 𝑅𝑡 .
Definition A.2.1. We say that the function 𝑓 : Ω → [0, 𝑅) is a parameterizing

function if it satisfies the following properties;
1. 𝑓 is Lipschitz continuous in Ω and smooth on 𝑓 −1 ((0, 𝑅)),
2. all values in (0, 𝑅) are regular values of 𝑓 , and 𝑆𝑡 = 𝑓 −1 (𝑡) are closed
hypersurfaces in M,
3. 𝑅𝑡 = 𝑓 −1 ([0, 𝑡)) are compactly embedded submanifolds of Ω with boundary

𝑆𝑡 . Furthermore, we need that |grad1 𝑓 | is an integrable function on 𝑅𝑡 for all
𝑡.
Under the assumptions on the function 𝑓 we can formulate the co-area formula
on manifolds.
46
Lemma A.2.2 ([Fed59, Thm. 3.1]). Let 𝑓 : Ω ⊂ M → R be a parameterizing

function. Define 𝑆𝑡 = 𝑓 −1 (𝑡) and let 𝜎𝑡 be the area measure on 𝑆𝑡 . Then for all
integrable functions 𝜑 : Ω → R𝑛 we have that
∫ ∫ 𝑡∫
𝜑
𝜑 vol = 𝜎𝑠 𝑑𝑠.
𝑅𝑡 0 𝑆𝑠 | grad 𝑓 |
It will be beneficial for us to view 𝑆𝑡 as variations of hypersurfaces following the

grad 𝑓
flow of the vector field . To make this precise, we formulate the following
|grad 𝑓 | 2
lemma.
Lemma A.2.3. Let 𝑓 : Ω → [0, 𝑅) be a parameterizing function. Let 𝜑𝑡 denote
grad 𝑓
the flow of the vector field and fix a value 𝑡0 ∈ (0, 𝑅). Then for all 𝑡0 + 𝑡 < 𝑅
|grad 𝑓 |2
we have that 𝜑𝑡 𝑆𝑡0 = 𝑆𝑡0 +𝑡 .
grad 𝑓
Proof. Let 𝛾 (𝑡) be an integral curve of such that 𝛾 (0) = 𝑥 ∈ 𝑆𝑡0 . We need
|grad 𝑓 | 2
to show that 𝑓 (𝛾 (𝑡)) = 𝑡0 + 𝑡. Taking the derivative we obtain

𝑑 grad 𝑓
𝑓 (𝛾 (𝑡)) = 𝑑𝑓 ( 𝛾¤ (𝑡)) = hgrad 𝑓 , 𝛾¤ (𝑡)i = grad 𝑓 , = 1. (A.2.1)
𝑑𝑡 |grad 𝑓 | 2

Integrating (A.2.1) shows that 𝜑𝑡 𝑆𝑡0 ⊂ 𝑆𝑡+𝑡0 . To see that 𝜑𝑡 𝑆𝑡0 = 𝑆𝑡+𝑡0 we
pick 𝑝 ∈ 𝑆𝑡+𝑡0 . Since 𝜑𝑡 is a diffeomorphism with 𝜑−1 𝑡 = 𝜑−𝑡 the element 𝜑−𝑡 ( 𝑝)
is in 𝑆𝑡0 by repeating the argument above. Hence 𝜑𝑡 (𝜑−𝑡 ( 𝑝)) = 𝑝 and the result
follows.
We remind the reader that the Lie derivative of a 𝑘-form 𝜔 in the direction of
the vector field 𝑋 evaluated at the point 𝑝 ∈ 𝑀 is given by
𝜑∗𝑡 𝜔 𝜑𝑡 ( 𝑝) − 𝜔 𝑝

(L 𝑋 𝜔) 𝑝 = lim ,
𝑡→0 𝑡
where 𝜑∗𝑡 denotes the pull back with respect to the flow 𝜑𝑡 of 𝑋. We will use some
standard properties of the Lie derivative acting on forms which can be found in
[Lee13, p. 372]. Let 𝑋 be a 𝐶 1 vector field and 𝜔 and 𝜈 be differentiable 𝑘- and
𝑙-forms, respectively. Then
L 𝑋 (𝜔 ∧ 𝜈) = (L 𝑋 𝜔) ∧ 𝜈 + 𝜔 ∧ (L 𝑋 𝜈) , (A.2.2)
Cartan’s Magic Formula: L 𝑋 𝜔 = 𝜄𝑋 𝑑𝜔 + 𝑑𝜄𝑋 𝜔, (A.2.3)

where 𝜄 denotes the interior product and 𝑑 is the exterior differential. The Cartan’s
magic formula implies that for any function 𝑓 we have the formula
L 𝑓 𝑋 𝜔 = 𝑓 L 𝑋 𝜔 + 𝑑𝑓 ∧ (𝜄𝑋 𝜔) ,
47
where we have used that 𝜄 𝑓 𝑋 = 𝑓 𝜄𝑋 and 𝑑 ( 𝑓 𝜔) = 𝑑𝑓 ∧ 𝜔 + 𝑓 𝑑𝜔. The reason for

going from the level surfaces of a function to a variation of surfaces by using the
flow point of view is to utilize the following differentiation theorem.
Lemma A.2.4. Let 𝛼 be an (𝑛 − 1)-form and let 𝑆 be an oriented closed smooth

hypersurface in M. Denote by 𝑋 a vector field and denote by 𝜑𝑡 the flow generated
by 𝑋. Then ∫ ∫
𝑑
𝛼= L 𝑋 𝛼,
𝑑𝑡 𝜑𝑡 (𝑆) 𝜑𝑡 (𝑆)
where L 𝑋 denotes the Lie derivative with respect to 𝑋.

𝑑
Proof. By using the definition of 𝑑𝑡 we get
∫ ∫
𝑑
∫
𝜑𝑡+ℎ (𝑆)
𝛼− 𝜑𝑡 (𝑆)
𝛼
𝛼 = lim
𝑑𝑡 𝜑𝑡 (𝑆) ℎ→0 ℎ
∫ ∫
𝜑∗ 𝛼
𝜑𝑡 (𝑆) ℎ
− 𝜑 (𝑆)
𝛼
𝑡
= lim
ℎ→0 ℎ
𝜑∗ℎ 𝛼 − 𝛼
∫
= lim
𝜑𝑡 (𝑆) ℎ→0 ℎ
∫
= L 𝑋 𝛼.
𝜑𝑡 (𝑆)
We refer the reader to [Fra12, p.139] where the result is proved for more general
variations of submanifolds.
A.2.2 The Main Theorem

Let ℎ : Ω → R be a harmonic function where Ω ⊂ M is an open set and let
𝑓 : Ω → [0, 𝑅) be a parametrization function from Section A.2.1. Define the
function ∫
𝐻 (𝑡) = ℎ2 |grad 𝑓 | 𝜎𝑡 ,
𝑆𝑡
where 𝑆𝑡 = 𝑓 −1 (𝑡) and 𝜎𝑡 is its surface measure. The goal of this section is to
show that 𝐻 satisfies a convexity property.
We will need the following version of a result of Hörmander, [Hör18, p. 38
Thm. 1]: Let 𝑓 be a parameterizing function and 𝑆𝑡 and 𝜎𝑡 be as above. Then there
exists a function 𝐾 only depending on 𝑓 such that for any harmonic function ℎ,
2
grad𝑆𝑡 ℎ − ℎ2𝑛
∫ ∫
𝜎𝑡 ≥ −𝐾 (𝑡) |grad ℎ| 2 vol , (A.2.4)
𝑆𝑡 |grad 𝑓 | 𝑅𝑡
48
where grad𝑆𝑡 ℎ and ℎ 𝑛 denote the gradient with respect to 𝑆𝑡 and the unit normal
derivative, respectively. Inequality (A.2.4) is proved in the end of this section, see
Lemma A.2.11. The following theorem is the main result of the paper; it shows
that for any harmonic function ℎ the 𝐿 2 -norms 𝐻 satisfy some convexity inequality
only depending on the function 𝑓 .
Theorem A.2.5. Let (M, g) be a Riemannian manifold, and let the functions ℎ,
𝑓 and 𝐻 be as described earlier in this section. Define the functions 𝑚, 𝑀 and 𝑔
such that:
Δ𝑓
4. 𝑚 (𝑡) ≤ ≤ 𝑀 (𝑡) on 𝑆𝑡 , and
|grad 𝑓 | 2
D E
Δ𝑓 grad 𝑓
5. grad , ≥ 𝑔 (𝑡) on 𝑆𝑡 .
|grad 𝑓 | 2 |grad 𝑓 | 2
Then 𝐻 satisfies the growth estimate

∫ ∫
0 Δ𝑓
𝐻 (𝑡) = 2 ℎℎ 𝑛 𝜎𝑡 + ℎ2 |grad 𝑓 | 𝜎𝑡 (A.2.5)
𝑆𝑡 𝑆𝑡 |grad 𝑓 | 2
∫
≥2 |grad ℎ| 2 vol +𝑚 (𝑡) 𝐻 (𝑡) .
𝑅𝑡
Moreover, if 𝐾 is the function given in Inequality (A.2.4) and 𝐾 (𝑡) + 𝑀 (𝑡) ≥ 0

then
(log 𝐻 (𝑡)) 00 + (𝐾 (𝑡) + 𝑀 (𝑡)) (log 𝐻 (𝑡)) 0 ≥ 𝑔 (𝑡) + 𝑚 (𝑡) 𝑀 (𝑡) + 𝑚 (𝑡) 𝐾 (𝑡) .
(A.2.6)
Proof. Using Lemma A.2.4 to take the derivative of 𝐻 we obtain

∫
0
𝐻 (𝑡) = L (grad 𝑓 )/|grad 𝑓 |2 ℎ2 |grad 𝑓 | 𝜎𝑡
𝑆
∫𝑡
grad 𝑓 2
= 2
, grad ℎ |grad 𝑓 | 𝜎𝑡
𝑆𝑡 |grad 𝑓 |
∫
+ ℎ2 L (grad 𝑓 )/|grad 𝑓 |2 (|grad 𝑓 | 𝜎𝑡 )
𝑆𝑡
∫ ∫
=2 ℎℎ 𝑛 𝜎𝑡 + ℎ2 L (grad 𝑓 )/|grad 𝑓 |2 (|grad 𝑓 | 𝜎𝑡 )
𝑆𝑡 𝑆𝑡
The following lemma takes care of the last term in the above computation and
finishes the proof of Equation (A.2.5). In the literature a version of the next lemma
is known as the first variation of area for hypersurfaces (see [CLN06, p. 51]).
49
Lemma A.2.6. Using the notation above, we have that

Δ𝑓
Lgrad 𝑓 /| grad 𝑓 |2 (|grad 𝑓 | 𝜎𝑡 ) = 𝜎𝑡
|grad 𝑓 |

= hgrad 𝑓 , grad |grad 𝑓 |i /|grad 𝑓 | 2 − 𝑛𝐻𝑡 𝜎𝑡 ,
where 𝐻𝑡 is the mean curvature of 𝑆𝑡 and 𝑛 is the dimension of M.
Proof. Using the properties of the Lie derivative given by Equation (A.2.2) and
(A.2.3) together with the definition of the divergence we obtain

grad 𝑓
L grad 𝑓 (|grad 𝑓 | 𝜎𝑡 ) = , grad |grad 𝑓 | 𝜎𝑡 + L grad 𝑓 (𝜎𝑡 )
| grad 𝑓 | 2 |grad 𝑓 | 2 | grad 𝑓 |
+ |grad 𝑓 | 𝑑 (1/|grad 𝑓 |) ∧ 𝜄 grad 𝑓 (𝜎𝑡 )

|grad 𝑓 |

grad 𝑓
= , grad |grad 𝑓 | 𝜎𝑡 + L grad 𝑓 𝜄 grad 𝑓 vol
|grad 𝑓 | 2 | grad 𝑓 | |grad 𝑓 |

grad 𝑓
= , grad |grad 𝑓 | 𝜎𝑡 + 𝜄 grad 𝑓 L grad 𝑓 vol
|grad 𝑓 | 2 |grad 𝑓 | |grad 𝑓 |

grad 𝑓 grad 𝑓
= , grad |grad 𝑓 | 𝜎𝑡 + div 𝜎𝑡
|grad 𝑓 | 2 |grad 𝑓 |
Δ𝑓
= 𝜎𝑡 .
|grad 𝑓 |
This concludes the proof of the Identity (A.2.5), note that the expression for 𝐻 0
holds for an arbitrary function ℎ not necessarily harmonic.
To prove the differential inequality (A.2.6) we differentiate (A.2.5). We rewrite
the first term by using the divergence formula and applying the co-area formula
given in Lemma A.2.2 when 𝜑 (𝑥) = ℎ2 (𝑥) and obtain
∫ ∫ ∫ 𝑡∫
2 1
𝐷 (𝑡) := ℎℎ 𝑛 𝜎𝑡 = |grad ℎ| vol = |grad ℎ| 2 𝜎𝑠 𝑑𝑠.
𝑆𝑡 𝑅𝑡 0 𝑆𝑠 |grad 𝑓 |
Computing the second derivative of 𝐻 by applying Lemma A.2.4 and A.2.6 once
more gives
|grad ℎ| 2
∫ ∫
Δ𝑓
𝐻 00 (𝑡) = 2 𝜎𝑡 + 2 ℎℎ 𝑛 |grad 𝑓 | 𝜎𝑡
𝑆𝑡 |grad 𝑓 | 𝑆𝑡 |grad 𝑓 | 3
∫ 2
2 Δ𝑓
+ ℎ |grad 𝑓 | 𝜎𝑡
𝑆𝑡 |grad 𝑓 | 2
∫
2 Δ𝑓 grad 𝑓
+ ℎ grad , |grad 𝑓 | 𝜎𝑡 .
𝑆𝑡 | grad 𝑓 | 2 |grad 𝑓 | 2
50
2
Using that |grad ℎ| 2 = 2ℎ2𝑛 + grad𝑆𝑡 ℎ − ℎ2𝑛 and denoting

Δ𝑓 grad 𝑓
𝐺 (𝑥) = grad , ,
| grad 𝑓 | 2 |grad 𝑓 | 2
we have
∫ ∫
00 2
1
𝐻 (𝑡) = 2 ℎ2𝑛 /|grad 𝑓 | 𝜎𝑡 + 2 grad𝑆𝑡 ℎ − ℎ2𝑛 𝜎𝑡
𝑆𝑡 𝑆𝑡 |grad 𝑓 |
∫ 2 ∫
1 Δ𝑓
+ ℎ2 2
|grad 𝑓 | 𝜎𝑡 + ℎ2 𝐺 |grad 𝑓 | 𝜎𝑡
2 𝑆𝑡 |grad 𝑓 | 𝑆𝑡
∫ 2
ℎ𝑛 ℎΔ 𝑓
+2 + |grad 𝑓 | 𝜎𝑡 .
𝑆𝑡 |grad 𝑓 | 2 |grad 𝑓 | 2
Applying Inequality (A.2.4) we get

2
∫ grad𝑆𝑡 ℎ − ℎ2𝑛 ∫
𝜎𝑡 ≥ −𝐾 (𝑡) ℎℎ 𝑛 𝜎𝑡 = −𝐾 (𝑡) 𝐷 (𝑡) , (A.2.7)
𝑆𝑡 |grad 𝑓 | 𝑆𝑡
and by Cauchy-Schwarz we have the inequalities

∫ 2
ℎ𝑛 ℎΔ 𝑓 1 0 2
2 + |grad 𝑓 | 𝜎𝑡 𝐻 (𝑡) ≥ 𝐻 (𝑡) (A.2.8)
𝑆𝑡 |grad 𝑓 | 2 |grad 𝑓 | 2 2
and
2
ℎ2 Δ 𝑓
∫ ∫
1 0 2 1 2 Δ𝑓
𝐻 (𝑡) − 2𝐷 (𝑡) 2
|grad 𝑓 | 𝜎𝑡 − ℎ |grad 𝑓 | 𝜎𝑡
2 𝑆𝑡 |grad 𝑓 | 2 𝑆𝑡 |grad 𝑓 | 2
ℎ2𝑛
∫
≤2 𝜎𝑡 𝐻 (𝑡) . (A.2.9)
𝑆𝑡 | grad 𝑓 |
A straightforward computation combining (A.2.7), (A.2.8) and (A.2.9) shows that

∫
Δ𝑓
𝐻 00 (𝑡) 𝐻 (𝑡) − 𝐻 0 (𝑡) 2 ≥ −2𝐾 (𝑡) 𝐷 (𝑡) 𝐻 (𝑡) − 2𝐷 (𝑡) ℎ2 |grad 𝑓 | 𝜎𝑡
(A.2.10)
∫
Δ𝑓 grad 𝑓
+ 𝐻 (𝑡) ℎ2 grad 2
, |grad 𝑓 | 𝜎𝑡 .
𝑆𝑡 |grad 𝑓 | |grad 𝑓 | 2
51
Applying the estimates (4) and (5) and noting the fact that 𝐾 + 𝑀 is non-negative,
implies
𝐻 00 (𝑡) 𝐻 (𝑡) − 𝐻 0 (𝑡) 2 + (𝐾 (𝑡) + 𝑀 (𝑡)) 𝐻 0 (𝑡) 𝐻 (𝑡)

∫
Δ𝑓 grad 𝑓
≥ 𝐻 (𝑡) ℎ2 grad , |grad 𝑓 | 𝜎𝑡
∫
Δ𝑓
+ (𝑀 (𝑡) + 𝐾 (𝑡)) 𝐻 (𝑡) ℎ2 |grad 𝑓 | 𝜎𝑡
≥ 𝑔 (𝑡) 𝐻 (𝑡) 2 + 𝑚 (𝑡) (𝑀 (𝑡) + 𝐾 (𝑡)) 𝐻 (𝑡) 2
≥ (𝑔 (𝑡) + 𝑚 (𝑡) 𝑀 (𝑡) + 𝑚 (𝑡) 𝐾 (𝑡)) 𝐻 (𝑡) 2 .
Dividing both sides of the equation by 𝐻 (𝑡) 2 , we obtain (A.2.6) and thus finish the
proof of Theorem A.2.5.
Remark A.2.7. Sometimes it is beneficial to replace Inequality (A.2.4) by
2
grad𝑆𝑡 ℎ − ℎ2𝑛
∫
𝜎𝑡 ≥ −𝐾 (𝑡) 𝐷 (𝑡) + 𝑘 (𝑡) 𝐻 (𝑡) ,
𝑆𝑡 |grad 𝑓 |
to obtain a better result. Using this inequality in the proof above replaces Inequality
(A.2.10) with
𝐻 00 (𝑡) 𝐻 (𝑡) − 𝐻 0 (𝑡) 2 ≥ −2𝐾 (𝑡) 𝐷 (𝑡) 𝐻 (𝑡) + 2𝑘 (𝑡) 𝐻 (𝑡) 2

∫
2 Δ𝑓 grad 𝑓
+ 𝐻 (𝑡) ℎ grad , |grad 𝑓 | 𝜎𝑡
∫
Δ𝑓
− 2𝐷 (𝑡) ℎ2 |grad 𝑓 | 𝜎𝑡 .
Completing the proof in the same manner as before gives
(log 𝐻 (𝑡)) 00 + (𝐾 (𝑡) + 𝑀 (𝑡)) (log 𝐻 (𝑡)) 0 ≥ 𝑔 (𝑡) + 𝑚 (𝑡) (𝑀 (𝑡) + 𝐾 (𝑡)) + 2𝑘 (𝑡)
as a generalization of Inequality (A.2.6) in Theorem A.2.5. We will use this

modified version of Theorem A.2.5 in Section A.3.1 when the upper bound of the
sectional curvature K is negative.
Remark A.2.8. If M is an oriented manifold, then 𝑆𝑡 is always orientable. In
general, when M is orientable any hypersurface that can be described as the level
set of a regular value of a smooth function is orientable (see [Lee13, Prop. 15.23]).
52
A.2.3 Corollaries
Before proving Inequality (A.2.4), we provide some corollaries and remarks.
Corollary A.2.9. Let 𝑓 : Ω → [0, ∞) be a convex and parameterizing function.
Then 𝑚 is non-negative, and hence 𝐻 is increasing. In this case, the sets 𝑅𝑡 =
𝑓 −1 ( [0, 𝑡)) are (totally) convex.
Proof. That 𝑓 is convex means that the Hessian of 𝑓 satisfies ∇2 𝑓 (𝑣, 𝑣) ≥ 0 for

all 𝑣 ∈ 𝑇 𝑝 M and 𝑝 ∈ M. Taking the trace of the Hessian of 𝑓 shows that Δ 𝑓 ≥ 0,

and hence 𝑚 (𝑡) ≥ 0. Thus Inequality (A.2.5) implies that 𝐻 is increasing. We
say that 𝑅𝑡 is (totally) convex if any geodesic in Ω starting and ending in 𝑅𝑡 is
contained in 𝑅𝑡 . For a geodesic 𝛾 a straightforward computation gives that
𝑑2
𝑓 (𝛾 (𝑠)) = ∇2 𝑓 ( 𝛾¤ (𝑠) , 𝛾¤ (𝑠)) ≥ 0.
𝑑𝑠2
Hence if 𝛾 is geodesic such that 𝛾 (0) = 𝑥 ∈ 𝑅𝑡 and 𝛾 (1) = 𝑦 ∈ 𝑅𝑡 , then
𝑓 (𝛾 (𝜆)) ≤ 𝜆 𝑓 (𝛾 (1)) + (1 − 𝜆) 𝑓 (𝛾 (0)) ≤ 𝑡.
In conclusion, we have that 𝛾 ⊂ 𝑅𝑡 and hence 𝑅𝑡 is (totally) convex.

Δ𝑓
In the case when |grad 𝑓 | is constant, the term |grad 𝑓 | coincides with the
mean curvature, giving a geometric interpretation to the functions 𝑚, 𝑀 and 𝑔.
When 𝑓 is given as the distance function from a compact submanifold (e.g. radial
distance function) we have that |grad 𝑓 | = 1 (see [Lee18, Thm. 6.38]). Letting 𝑓
be the distance function from a point, then 𝑚 (𝑡) = 𝑀 (𝑡) is equivalent with ⨏ the
Riemannian manifold (M, g) being locally harmonic at 𝑝, meaning that 𝑆 ℎ𝜎𝑡
𝑡
is constant for all ℎ and 𝑡 less than some fixed 𝜀. When |grad 𝑓 | is not constant the
geometric interpretation of 𝑚, 𝑀 and 𝑔 becomes somewhat more diffuse. However,
the following proposition tells us that the difference 𝑀 (𝑡) − 𝑚 (𝑡) measures how
far the level sets of 𝑓 are from satisfying the mean value theorem.
Proposition A.2.10. Assume that 𝑓 : Ω → R is a parameterizing function such
that Δ 𝑓 2 = 𝑀 (𝑡) on 𝑆𝑡 = 𝑓 −1 (𝑡). Then
|grad 𝑓 |
∫
𝑆
ℎ |grad 𝑓 | 𝜎𝑡
𝐹 (𝑡) = ∫𝑡
𝑆𝑡
|grad 𝑓 | 𝜎𝑡
satisfies the mean value property, i.e. 𝐹 0 (𝑡) = 0.

When 𝑓 (𝑥) = 𝑟 (𝑥) is the radial distance function centered at the point 𝑝,
then Δ𝑟 = 𝑀 (𝑡) for all 𝑡 less that some fixed 𝜀 is equivalent with the Riemannian
manifold being locally harmonic at 𝑝.
53
Proof. The derivative of 𝐹 is equal to

Δ𝑓 Δ𝑓
∫ ∫
ℎ
𝑆𝑡 |grad 𝑓 | 2 |grad 𝑓 | 𝜎𝑡 𝑆𝑡 |grad 𝑓 | 2
0
𝐹 (𝑡) = ∫ − ∫ 𝐹 (𝑡) ,
𝑆
|grad 𝑓 | 𝜎𝑡 𝑆
𝑡 𝑡
by using Lemma A.2.6. Hence if Δ 𝑓 2 = 𝑀 (𝑡) we get that 𝐹 0 (𝑡) = 0, and 𝐹 (𝑡)
|grad 𝑓 |
is constant.
For the last claim we utilize that the manifold is locally harmonic if and only
if the geodesic spheres centered at 𝑝 have constant mean curvature (see [Kre10,
Prop. 3.1.2]). The mean curvature 𝐻𝑡 of a hypersurface given as a level surface of
a function 𝑓 at the value 𝑡 satisfies

grad 𝑓 grad 𝑓
Δ 𝑓 − ∇2 𝑓 | grad 𝑓 | , | grad 𝑓 |
𝐻𝑡 = −
𝑛 |grad 𝑓 |

1 grad 𝑓
= − div ,
𝑛 |grad 𝑓 |
(see [Lee18, Exercise 8-2 b)]). Since the gradient of 𝑟 has norm one we get
𝐻𝑡 = − 𝑛1 Δ 𝑓 , which proves the claim.
A.2.4 An Inequality of Hörmander

The only thing left is to prove Inequality (A.2.4). The statement and its proof can
be found in Hörmander’s works, see [Hör18, p. 38 Thm. 1]. We need a weighted
version of the inequality and provide a proof for the convenience of the reader.
Lemma A.2.11. Let Δℎ = 0, and assume that R is an open compactly embedded
manifold. Denote by 𝑆 := 𝜕R and by 𝜎 its area measure. Let 𝜕𝑛 denote any smooth
extension of the outward unit normal vector of 𝑆 to R. Then
∫
2
grad𝑆 ℎ − ℎ2𝑛 𝑤𝜎 (A.2.11)
𝑆
∫
2 grad ℎ
= | grad ℎ| div (𝑤𝜕𝑛 ) − 2 ∇ grad ℎ (𝑤𝜕𝑛 ) , vol ,
R |grad ℎ| |grad ℎ|
where 𝑤 (𝑥) 𝜕𝑛 is a smooth vector field defined on R. Since R is compact there
exists a minimum (and maximum) of
div (𝑤 (𝑥) 𝜕𝑛 ) − 2h∇ 𝑣 (𝑤 (𝑥) 𝜕𝑛 ) , 𝑣i
where 𝑣 ∈ 𝑇𝑥 M and |𝑣| = 1. Hence there exists a constant 𝐾 such that
∫ ∫
2 2
grad𝑆 ℎ − ℎ 𝑛 𝑤𝜎 ≥ −𝐾 ℎℎ 𝑛 𝜎.
𝑆 𝑆
54
A.3. Examples
Proof. Denote by 𝑋 = 𝑤 (𝑥) 𝜕𝑛 and 𝑉 = 2𝑋 (ℎ) grad ℎ − |grad ℎ| 2 𝑋. Then com-

puting the divergence of 𝑉 we get
div (𝑉) = 2𝑋 (ℎ) Δℎ + 2hgrad 𝑋 (ℎ) , grad ℎi − |grad ℎ| 2 div (𝑋)
− grad |grad ℎ| 2 , 𝑋
= 2h∇grad ℎ 𝑋, grad ℎi + 2 𝑋, ∇grad ℎ grad ℎ − |grad ℎ| 2 div (𝑋)
− 2 h∇𝑋 grad ℎ, grad ℎi
= 2h∇grad ℎ 𝑋, grad ℎi − |grad ℎ| 2 div (𝑋)

+ 2 ∇2 ℎ (𝑋, grad ℎ) − ∇2 ℎ (𝑋, grad ℎ)

2 grad ℎ
= |grad ℎ| 2 ∇grad ℎ/|grad ℎ | 𝑋, − div (𝑋) .
|grad ℎ|
Applying the divergence theorem we get
∫ ∫
2 grad ℎ
h𝑉, 𝜕𝑛 i 𝜎 = |grad ℎ| 2 ∇grad ℎ/|grad ℎ | 𝑋, − div (𝑋) vol .
𝑆 R |grad ℎ|
Using the definition of 𝑉 gives

2
h𝑉, 𝜕𝑛 i = −𝑤 (𝑥) grad𝑆 ℎ − ℎ2𝑛 ,
when 𝑥 ∈ 𝑆. Hence
∫
2
grad𝑆 ℎ − ℎ2𝑛 𝑤𝜎
𝑆
∫
2 grad ℎ
= | grad ℎ| div (𝑤𝜕𝑛 ) − 2 ∇ grad ℎ (𝑤𝜕𝑛 ) , vol .
R |grad ℎ| |grad ℎ|

Remark A.2.12. For Lemma A.2.11 to hold it is not enough for the function 𝑓
to be Lipschitz. Consider for example the function 𝑓 : R2 → R is defined by
𝑓 (𝑥, 𝑦) = |𝑥| + |𝑦|. In this case we have that the level surfaces are squares.
Considering the family of harmonic functions ℎ (𝑥, 𝑦) = 𝑒 𝑛𝑥 cos (𝑛𝑦 + 𝜋/2) we get
that ∫
2 √
grad𝑆1 ℎ − ℎ2𝑛 𝜎1 = −𝑛 2 (2 sinh(2𝑛) − 2 sin (2𝑛))
𝑆1
and ∫
ℎℎ 𝑛 𝜎1 = cosh (2𝑛) − 1.
𝑆1
Thus there is no 𝐾 such that
∫ ∫
2 2
grad𝑆1 ℎ − ℎ 𝑛 𝜎1 ≥ −𝐾 ℎℎ 𝑛 𝜎1
𝑆1 𝑆1
holds for all functions in this family.
55
A.3 Examples
Although Theorem A.2.5 is rather technical, it has several novel applications which
are explored in this section. As stated in the introduction, we start with an ap-
plication to geodesic spheres on Riemannian manifolds. In this case, we will use
results from comparison geometry to find the functions 𝑀, 𝑚, 𝑔 and 𝐾 in Theorem
A.2.5. Thereafter we consider level surfaces of 1-homogeneous functions which
cover ellipsoids with constant eccentricity. The distance function for closed lower
dimensional spheres will be an example of level surfaces that are not homeomor-
phic to spheres. Finally, we will show if we have upper and lower estimates on the
sectional curvature we have that eigenfunctions of the Laplacian corresponding to
positive eigenvalues satisfy the same type of convexity as harmonic functions.
A.3.1 Geodesic Spheres

Using exponential coordinates centered at a point 𝑝 ∈ M we can introduce polar
coordinates in a neighborhood of 𝑝. Define the radial distance function on a normal
neighborhood of 𝑝 by
q
𝑟 (𝑥) = dist (𝑥, 𝑝) = 𝑥 12 (𝑥) + · · · + 𝑥 𝑛2 (𝑥),
where 𝑥 𝑖 are the coordinate functions in the normal neighborhood. In this example
we let the function 𝑓 given in Theorem A.2.5 be 𝑓 (𝑥) = 𝑟 (𝑥). The level surfaces
of 𝑟 are precisely the geodesic spheres 𝑆𝑡 = 𝑟 −1 (𝑡) of radius 𝑡. Moreover, the
Riemannian metric in polar coordinates can be written as g = 𝑑𝑟 2 + 𝑔𝑡 where 𝑔𝑡
is the induced metric on 𝑆𝑡 . Let Inj ( 𝑝) denote the injectivity radius at the point
𝑝, i.e. the supremum over the radius of all balls centered at 0 ∈ 𝑇 𝑝 M where the
exponential map is injective. Then 𝑟 is smooth in 𝐵Inj( 𝑝) ( 𝑝) \ {𝑝}. We will use
the notation √

 sin K𝑡

 √ , when K > 0
K




sin K (𝑡) = 𝑡, when K = 0 (A.3.1)
 √
sinh −K𝑡


when K < 0,

 √
 ,
 −K
cos K (𝑡)
cos K (𝑡) = (sin K (𝑡)) 0 , and cot K (𝑡) = (log (sin K (𝑡))) 0 = .
sin K (𝑡)
Theorem A.3.1. Assume that (M, g) is an 𝑛-dimensional Riemannian manifold
with 𝑝 ∈ M and with sectional curvature Sec satisfying
𝜅 ≤ Sec ≤ K, (A.3.2)
56
A.3. Examples

where 𝜅, K ∈ R. Set 𝑅 := min Inj ( 𝑝) , √𝜋 if K > 0 and 𝑅 := Inj ( 𝑝) whenever
2 K
K ≤ 0. Let ℎ be a harmonic function defined
∫ on 𝐵 𝑅 ( 𝑝). If 𝑟 (𝑥) = dist (𝑥, 𝑝) is
the radial distance function and 𝐻 (𝑡) = 𝑟 −1 (𝑡) ℎ2 𝜎𝑡 , then
𝐻 0 (𝑡) ≥ (𝑛 − 1) cot K (𝑡) 𝐻 (𝑡) . (A.3.3)
Moreover, we have
(log 𝐻 (𝑡)) 00 + (cot K (𝑡) + (𝑛 + 1) (cot 𝜅 (𝑡) − cot K (𝑡))) (log 𝐻 (𝑡)) 0 (A.3.4)
≥ − (𝑛 − 1) K + (𝑛 − 2) min (K, 0) − (𝑛 − 1) (K − 𝜅) ,
for every 𝑡 ∈ (0, 𝑅).
Remark A.3.2. 1. Note that Equation (A.3.3) implies that 𝐻 is increasing.
When K > 0 Inequality (A.3.3) is also valid when 𝑡 < 𝑅˜ := min Inj ( 𝑝) , √𝜋 .
K
However, when √𝜋 < 𝑡 < √𝜋 the function cot K (𝑡) is negative. To
2 K K
see that 𝐻 is not necessarily increasing for values 𝑡 > 𝑅 we consider the
unit sphere M = 𝑆 2 and ℎ (𝑥) = 1. In this case, we have precisely that
𝐻 0 (𝑡) = cot (𝑡) 𝐻 (𝑡). This shows the necessity of the constraint 𝑅 since
cot (𝑡) 𝐻 (𝑡) is negative whenever 𝑡 > 𝑅.
2. Equation (A.3.4) is slightly better than the one presented in [Man13, Thm. 2.2
(ii)] whenever K > 0 and K ≠ 𝜅. The Inequality (A.3.3) in [Man13] is proved
with the right hand side equal to
𝑛
− (𝑛 − 1) K − (𝑛 − 1) 1 + (𝑛 − 2) (K − 𝜅)
2
instead of our improvement − (𝑛 − 1) K − (𝑛 − 1) (K − 𝜅).
Proof. To prove this theorem, we will apply Theorem A.2.5 and use comparison
geometry to find 𝑀, 𝑚, 𝑔 and 𝐾. When K < 0, we will need to adapt the Theorem
A.2.5 slightly, see Remark A.2.7.
Rauch Comparison Theorem states that the following estimate hold under the
sectional curvature bounds given in (A.3.2)
cot K (𝑡) 𝑔𝑡 ≤ ∇2𝑟 ≤ cot 𝜅 (𝑡) 𝑔𝑡 for 𝑡 < 𝑅.

˜ (A.3.5)
The proof of Rauch Comparison Theorem can be found in [Pet16, Thm. 6.4.3] or
[Lee18]. Inequality (A.3.5) implies that
𝑚 (𝑡) = (𝑛 − 1) cot K (𝑡) ≤ Δ𝑟 ≤ (𝑛 − 1) cot 𝜅 (𝑡) = 𝑀 (𝑡) .
57
To find 𝑔 we use the following identity, see [Pet16, Equation (2) p. 276],
2
hgrad Δ𝑟, grad 𝑟i = − ∇2𝑟 − Ric (grad 𝑟, grad 𝑟) ,
for all functions with | grad 𝑟 | = 1. Using the Rauch Comparison Theorem we
obtain
2
(𝑛 − 1) cot2K (𝑡) ≤ ∇2𝑟 ≤ (𝑛 − 1) cot2𝜅 (𝑡) .
Hence we conclude that
hgrad Δ𝑟, grad 𝑟i ≥ − (𝑛 − 1) cot2𝜅 (𝑡) − (𝑛 − 1) K = 𝑔 (𝑡) .
Next we need to find 𝐾 which exists by Inequality (A.2.4). To do this, we will

use the following version of Lemma A.2.11.
grad 𝑟
Lemma A.3.3. Let 𝜑 (𝑟 ( 𝑥)) be a smooth vector field, then
∫
| grad𝑆𝑡 ℎ| 2 − ℎ2𝑛 𝜎𝑡
𝑆𝑡
0
∫
2 𝜑 (𝑟 (𝑥)) Δ𝑟 − 𝜑 (𝑟 (𝑥))
= 𝜑 (𝑡) | grad ℎ| vol
𝐵𝑡 𝜑2 (𝑟 (𝑥))
∇2𝑟 (grad ℎ, grad ℎ) 𝜑 (𝑟 (𝑥)) − 𝜑 0 (𝑟 (𝑥)) hgrad 𝑟, grad ℎi 2
∫
− 2𝜑 (𝑡) vol .
𝐵𝑡 𝜑 (𝑟 (𝑥)) 2
Proof. Fix 𝑡0 . Using Lemma A.2.11 with the extension of 𝜕𝑛 to 𝐵𝑡0 be equal to
𝜑 (𝑡0 ) grad 𝑟
𝜑 (𝑟 ( 𝑥)) gives
∫
2
grad𝑆 ℎ − ℎ2𝑛 𝜎𝑡
𝑆𝑡
∫
2 grad 𝑟 grad 𝑟 grad ℎ
= 𝜑 (𝑡0 ) | grad ℎ| div − 2 ∇ grad ℎ , vol .
𝐵𝑡 𝜑 (𝑟 (𝑥)) |grad ℎ| 𝜑 (𝑟 (𝑥)) |grad ℎ|
Using the product rule for the divergence and covariant derivative finishes the
proof.
1
Using Lemma A.3.3 with 𝜑 (𝑡) = sinK (𝑡) implies that
∫ ∫
2
1
| grad𝑆𝑡 ℎ| − ℎ2𝑛 𝜎𝑡 = | grad ℎ| 2 (sin K (𝑠) Δ𝑟 + sin K (𝑠) cot K (𝑠))
𝑆𝑡 sin K (𝑡) 𝐵𝑡

− 2 sin K (𝑠) ∇2𝑟 (grad ℎ, grad ℎ) − cot K (𝑠) sin K (𝑠) ℎ2𝑛 vol .
58
A.3. Examples
Applying Rauch Comparison Theorem gives

∫ ∫
1
| grad𝑆𝑡 ℎ| 2 − ℎ2𝑛 𝜎𝑡 ≥ | grad ℎ| 2 (𝑛 cos K (𝑠) − 2 sin K (𝑠) cot 𝜅 (𝑠)) vol
𝑆𝑡 sin K (𝑡) 𝐵𝑡
≥ 2 (cot K (𝑡) − cot 𝜅 (𝑡)) 𝐷 (𝑡)
∫
𝑛−2
+ cos K (𝑟 (𝑥)) |grad ℎ| 2 vol
sin K (𝑡) 𝐵𝑡
where we have used that
sin K (𝑡) (cot K (𝑡) − cot 𝜅 (𝑡))
is decreasing for 𝑡 < 𝑅. Using integration by part on the last term together with the
observation that ∫
𝐻 0 (𝑡) − 𝑆 ℎ2 Δ𝑟𝜎𝑡
𝐷 (𝑡) = 𝑡
2
we get that
∫
| grad𝑆𝑡 ℎ| 2 − ℎ2𝑛 𝜎𝑡 ≥ 2 (cot K (𝑡) − cot 𝜅 (𝑡)) 𝐷 (𝑡) + (𝑛 − 2) cot K (𝑡) 𝐷 (𝑡)
𝑆𝑡
∫ 𝑡
K
+ (𝑛 − 2) sin K (𝑠) 𝐷 (𝑠) 𝑑𝑠
sin K (𝑡) 0
≥ (𝑛 cot K (𝑡) − 2 cot 𝜅 (𝑡)) 𝐷 (𝑡)
∫
𝐻 0 (𝑠) − 𝑆 ℎ2 Δ𝑟𝜎𝑠
∫ 𝑡 !
(𝑛 − 2) min(K, 0)
+ sin K (𝑠) 𝑠
𝑑𝑠
sin K (𝑡) 0 2
min(K, 0)
≥ (𝑛 cot K (𝑡) − 2 cot 𝜅 (𝑡)) 𝐷 (𝑡) + (𝑛 − 2) 𝐻 (𝑡) ,
2
where we have used that
(cos K (𝑡)) 0 = −K sin K (𝑡) .
Setting
𝐾 (𝑡) = 2 cot 𝜅 (𝑡) − 𝑛 cot K (𝑡)
we finish the case when K > 0. When K < 0, we use Remark A.2.7 with
min (K, 0)
𝑘 (𝑡) = (𝑛 − 2) .
2
Hence we have that
𝑀 (𝑡) + 𝐾 (𝑡) = cot K (𝑡) + (𝑛 + 1) (cot 𝜅 (𝑡) − cot K (𝑡))
59
and
𝑔 (𝑡) + 𝑚 (𝑡) (𝑀 (𝑡) + 𝐾 (𝑡)) + 2𝑘 (𝑡) = − (𝑛 − 1) K + (𝑛 − 2) min (K, 0)

− (𝑛 − 1) cot2𝜅 (𝑡) + (𝑛 − 1) cot K (𝑡) (cot K (𝑡) + (𝑛 + 1) (cot 𝜅 (𝑡) − cot K (𝑡)))

≥ − (𝑛 − 1) K + (𝑛 − 2) min (K, 0) + (𝑛 − 1) cot2K (𝑡) − cot2𝜅 (𝑡) .
Using that
cot2K (𝑡) − cot2𝜅 (𝑡) ≥ 𝜅 − K,
see [Man13, p.652], we get that
𝑔 (𝑡) + 𝑚 (𝑡) (𝑀 (𝑡) + 𝐾 (𝑡)) + 2𝑘 (𝑡) ≥ − (𝑛 − 1) K + (𝑛 − 2) min (K, 0)

+ (𝑛 − 1) (𝜅 − K) .
Let us briefly discuss the sharpness of our results in Theorem A.3.1. Remem-
ber that in R2 the homogeneous harmonic polynomials can be written in polar
coordinates as
ℎ (𝑡, 𝜃) = 𝑡 𝑘 (𝑎 cos (𝑘𝜃) + 𝑏 sin (𝑘𝜃)) ,
where 𝑎, 𝑏 ∈ R. In this case we have that K = 𝜅 = 0 and Theorem A.3.1 becomes
1
(log 𝐻 (𝑡)) 00 + (log 𝐻 (𝑡)) 0 ≥ 0.
𝑡
For the homogeneous polynomials we have that the inequality is sharp. Let K = 𝜅
and define tan K (𝑡) = cotK1 (𝑡) . The equivalent of homogeneous harmonic polyno-
mials for the two dimensional constant curvature spaces are
𝑡 𝑘
ℎ (𝑡, 𝜃) = tan K (𝑎 cos (𝑘𝜃) + 𝑏 sin (𝑘𝜃)).
2
In this case we have that Theorem A.3.1 becomes
(log 𝐻 (𝑡)) 00 + cot K (𝑡) (log 𝐻 (𝑡)) 0 ≥ −K,
and again we have that for the functions ℎ we have that the inequality is sharp.
When K ≥ 0 we have that for the constant harmonic function Theorem A.3.1 is
sharp for all 𝑛. In the case when K < 0 doing the example of constant harmonic
functions would suggest that the inequality could be improved to the right hand
side being − (𝑛 − 1) K. When K < 0 and 𝑛 ≥ 2 the radial solutions using spherical
harmonics can be found in [Min75, Prop. 4.2]. However, the solutions are expressed
using hypergeometric functions and it is thus no trivial task to see if the result is
sharp for these solutions.
60
A.3. Examples
A.3.2 1-Homogeneous Functions

The natural next step from looking at spheres in R𝑛 is to look at families of surfaces
in R𝑛 where the domains bounded by the surfaces are star convex with respect to
the origin. Fix a smooth and compact surface 𝑆 ⊂ R𝑛 such that the origin is not
contained in 𝑆. Moreover, assume that for each point x ∈ 𝑆 the ray {𝑡x : 𝑡 ≥ 0}
intersects the surface 𝑆 precisely once, namely at 𝑡 = 1. The we can unambiguously
define the inside of 𝑆 to be the collection of points
Ø Ø
Inn(𝑆) := 𝑡x.
x∈𝑆 0≤𝑡 <1
It is clear from its definition that Inn(𝑆) is star convex with respect to the origin.
A function 𝑔 : R𝑛 → R is called 𝑘-homogeneous for 𝑘 ∈ Z if 𝑔(𝑡x) = 𝑡 𝑘 𝑔(x)
for all 𝑡 ≥ 0. Define 𝑓 to be 𝑓 : R𝑛 → [0, ∞) by requiring that 𝑓 ≡ 1 on 𝑆 and
𝑓 (𝑡x) := 𝑡 · 𝑓 (x) = 𝑡, x ∈ 𝑆, 𝑡 ≥ 0.
Then 𝑓 is a 1-homogeneous function since 𝑓 (𝑡x) = 𝑡 𝑓 (x) for every 𝑡 ≥ 0 and

x ∈ R𝑛 . Given a 1-homogeneous with smooth compact level surface 𝑓 . Denote by
𝑆𝑡 = 𝑓 −1 (𝑡), then Inn(𝑆𝑡 ) are star convex with respect to 0.
Proposition A.3.4. Let 𝑓 : R𝑛 → [0, ∞) be a 1-homogeneous function which is

smooth in R𝑛 \ {0} with compact smooth level surfaces 𝑆𝑡 . Consider a harmonic
function ℎ : R𝑛 → R and set
∫
𝐻 (𝑡) = ℎ2 (x) |grad 𝑓 | 𝜎𝑡 .
𝑆𝑡
Then the function 𝐻 satisfies
𝐻 00 (𝑡) 𝐻 (𝑡) − 𝐻 0 (𝑡) 2 + 𝐴 0

𝑡 𝐻 (𝑡) 𝐻 (𝑡)
2
≥ −𝐵/𝑡 2 , (A.3.6)
𝐻 (𝑡)
where the positive constants 𝐴 and 𝐵 only depend on 𝑓 .
Proof. To apply Theorem A.2.5 we will use the fact that the derivative of a 𝑘-
homogeneous function is a (𝑘 − 1)-homogeneous function. Thus the derivatives of
𝑓 satisfy 𝑓 𝑥𝑖 (𝑡x) = 𝑓 𝑥𝑖 (x), 𝑓 𝑥𝑖 𝑥 𝑗 (𝑡x) = 1𝑡 𝑓 𝑥𝑖 𝑥 𝑗 (x) and 𝑓 𝑥𝑖 𝑥 𝑗 𝑥𝑘 (𝑡x) = 𝑡12 𝑓 𝑥𝑖 𝑥 𝑗 𝑥𝑘 (x).
Note also that all 𝑘-homogeneous functions are uniquely determined by their re-
strictions to the unit sphere 𝑆 𝑛−1 ⊂ R𝑛 . This implies that the estimates we need
to satisfy in Theorem A.2.5 are given by taking the minimum or maximum of
the derivatives over 𝑆 𝑛−1 . Hence we can take 𝑚(𝑡) = 𝐶1 /𝑡, 𝑀 (𝑡) = 𝐶2 /𝑡 and
𝑔(𝑡) = 𝐶3 /𝑡 2 for some constants 𝐶1 , 𝐶2 and 𝐶3 .
61
Fix 𝑡0 > 0. To find 𝐾 (𝑡) we extend the normal vector field on 𝑆𝑡0 to the inside
𝑓 grad 𝑓
Inn(𝑆𝑡0 ) by 𝜕𝑛 = 𝑡0 |grad 𝑓 | . Then by Equation (A.2.11) we obtain

2
∫ grad𝑆𝑡 ℎ − ℎ2𝑛
0
𝜎𝑡0
𝑆𝑡0 |grad 𝑓 |
∫
2 𝑓 grad 𝑓 𝑓 grad 𝑓
= | grad ℎ| div − 2 ∇grad ℎ , grad ℎ vol
𝑅𝑡0 𝑡0 |grad 𝑓 | 2 𝑡0 |grad 𝑓 | 2
∫
−𝐶4
≥ |grad ℎ| 2 vol ,
𝑡0 𝑅𝑡
for some constant 𝐶4 where the last inequality follows from the components of
𝑓 grad 𝑓
|grad 𝑓 | being 1-homogeneous in each component. Using Theorem A.2.5 we obtain
Inequality (A.3.6) with 𝐴 = 𝐶2 + 𝐶4 and 𝐵 = 𝐶3 + 𝐶1𝐶2 + 𝐶1𝐶4 .
Note that if ℎ is a homogeneous harmonic functions of degree 𝑘 then 𝐻 becomes

an (𝑛 − 1 + 2𝑘) homogeneous function. Thus 𝐻 (𝑡) = 𝑡 𝑛−1+2𝑘 𝐻 (1). Hence
𝐻 00 (𝑡) 𝐻 (𝑡) − 𝐻 0 (𝑡) 2 + 𝐴 0

𝑡 𝐻 (𝑡) 𝐻 (𝑡)
= (− (𝑛 − 1 + 2𝑘) + 𝐴 (𝑛 − 1 + 2𝑘)) /𝑡 2
𝐻 (𝑡) 2
= ( 𝐴 − 1) (𝑛 − 1 + 2𝑘) ≥ −𝐵/𝑡 2 .
Since this holds for all 𝑡 we have that 𝐴 ≥ 1.

We can integrate the inequality in Proposition A.3.4 and get a convexity property
for 𝐻. Doing this we get the following corollary.
Corollary A.3.5. 1. When we have that 𝐴 > 1, 𝐻 satisfies

𝐵
𝛼 𝐴−1 𝐵
(1−𝛼) 𝐴−1
𝑡0 𝑡2
𝐻 (𝑡1 ) ≤ 𝐻 (𝑡0 ) 𝛼 𝐻 (𝑡2 ) 1−𝛼 ,
𝑡1 𝑡1
where 𝐴−1 𝐴−1

𝑡1 𝑡1
(1 − 𝛼) +𝛼 = 1, 𝑡0 ≤ 𝑡1 ≤ 𝑡2 .
𝑡2 𝑡0
In this case, the function
𝑡𝐻 0 (𝑡)

𝐴−1 𝐵
𝑁 𝐻 (𝑡) := 𝑡 +
𝐻 (𝑡) 𝐴−1
is increasing.
62
A.3. Examples
2. When 𝐴 = 1 we have that

𝐵 𝑡0 𝑡2
𝐻 (𝑡1 ) ≤ exp − log log 𝐻 (𝑡0 ) 𝛼 𝐻 (𝑡2 ) 1−𝛼 ,
2 𝑡1 𝑡1
where
𝑡1 𝑡1
(1 − 𝛼) log + 𝛼 log = 0.
𝑡2 𝑡0
In this case, the function
𝑡𝐻 0 (𝑡)
𝑁 𝐻 (𝑡) := + 𝐵 log (𝑡)
𝐻 (𝑡)
is increasing.
Proof. Assume first that 𝐴 > 1. By using the integrating factor 𝑡 𝐴 inequality
(A.3.6) becomes
0
𝐴 0 𝐵 𝐴−1
𝑡 (log 𝐻 (𝑡)) + 𝑡 ≥ 0.
𝐴−1
Hence the function

𝑡𝐻 0 (𝑡)

𝐴−1 𝐵
𝑁 𝐻 (𝑡) = 𝑡 +
𝐻 (𝑡) 𝐴−1
is increasing. Define
𝐵
𝐺 (𝑡) = 𝑡 𝐴−1 𝐻 (𝑡) .
Then 𝑡 𝐴 (log (𝐺 (𝑡))) 0 = 𝛾 (𝑡) is an increasing function and
𝑡01−𝐴 − 𝑡11−𝐴
∫ 𝑡1
log 𝐺 (𝑡1 ) − log 𝐺 (𝑡0 ) = 𝛾 (𝑡) 𝑡 −𝐴 𝑑𝑡 ≤ 𝛾 (𝑡1 ) .
𝑡0 𝐴−1
Similarly,
𝑡11−𝐴 − 𝑡21−𝐴
log 𝐺 (𝑡2 ) − log 𝐺 (𝑡1 ) ≥ 𝛾 (𝑡1 ) .
𝐴−1
We also know that

𝛼 𝑡01−𝐴 − 𝑡11−𝐴 = (1 − 𝛼) 𝑡11−𝐴 − 𝑡21−𝐴 .
This implies the required inequality 𝐺 (𝑡1 ) ≤ 𝐺 (𝑡0 ) 𝛼 𝐺 (𝑡2 ) 1−𝛼 .

Whenever 𝐴 = 1 we obtain through similar computations that
𝑡 2 00

log 𝐻 𝑒 𝑡 + 𝐵 ≥ 0.
2
63
𝑡2
Since 𝐻 (𝑒 𝑡 ) + 2𝐵 is convex we get
2 2
/2+(1−𝛼) 𝐵 (log 𝑡0 ) 2 /2
𝑒 𝐵 (log 𝑡1 ) /2
𝐻 (𝑡1 ) ≤ 𝑒 𝛼𝐵 (log 𝑡0 ) 𝐻 (𝑡0 ) 𝛼 𝐻 (𝑡2 ) 1−𝛼 .
Using that
(1 − 𝛼) log (𝑡2 ) = log (𝑡1 ) − 𝛼 log (𝑡0 )
and
𝛼 log (𝑡2 ) = log (𝑡1 ) − (1 − 𝛼) log (𝑡2 )
we get that

𝐵 𝑡0 𝑡2
𝐻 (𝑡1 ) ≤ exp − log log 𝐻 (𝑡0 ) 𝛼 𝐻 (𝑡2 ) 1−𝛼 .
2 𝑡1 𝑡1
𝑡 𝐻 0 (𝑡)
In this case, the function 𝑁 𝐻 (𝑡) = 𝐻 (𝑡) + 𝐵 log (𝑡) is increasing.
Ellipsoids with Constant Eccentricity

We will now specialize to the case of ellipsoids with constant eccentricity. Define
the dilation matrix
𝑎 1 0 . . . 0 
 
 0 𝑎2 . . . 0 
𝐷=.
 
. . . ,
.
 . 0 
0 ... 0 𝑎 𝑛 

where 0 < 𝑎 1 ≤ 𝑎 2 ≤ · · · ≤ 𝑎 𝑛 . The function 𝑓 (x) = 𝐷 −1 x is 1-homogeneous
and its level surfaces are ellipsoids in R𝑛 centered at the origin. We wish to illustrate
Proposition A.3.4 and will hence need to find the values 𝐴 and 𝐵 in Proposition
A.3.4 explicitly. To find 𝐴 and 𝐵 we will use Theorem A.2.5.
A straightforward calculation gives the gradient, Hessian and Laplacian of 𝑓
as
2
tr 𝐷 −2 𝐷 −2 x
𝑇
𝐷 −2 x 𝐷 −2 x

𝐷 −2 x 2 𝐷 −2
grad 𝑓 = , ∇ 𝑓 = − , Δ 𝑓 = − .
𝑓 (x) 𝑓 (x) 𝑓 3 (x) 𝑓 (x) 𝑓 3 (x)
Hence we obtain the estimates
𝑎 21 tr 𝐷 −2 − 1 𝑎 2𝑛 tr 𝐷 −2 − 1

Δ𝑓
≤ ≤ .
𝑓 (x) |grad 𝑓 | 2 𝑓 (x)
𝑎2 tr ( 𝐷 −2 ) −1 𝑎12 tr ( 𝐷 −2 ) −1
We may now set 𝑀 (𝑡) = 𝑛 𝑡 and 𝑚 (𝑡) = 𝑡 , such that 𝑀 and 𝑚
are the functions in Equation (4).
64
A.3. Examples
To find a candidate for 𝑔 in Equation (5) we compute that
2
tr 𝐷 −2 2 tr 𝐷 −2 𝑓 2 (x) 𝐷 −3 x

Δ𝑓 grad 𝑓 1
grad , = 2 + −
|grad 𝑓 | 2 |grad 𝑓 | 2 𝑓 (x) 𝐷 −2 x
2
𝐷 −2 x
6
1 + 𝑎 21 tr 𝐷 −2 − 2 𝑎 4𝑛 /𝑎 21 tr 𝐷 −2

≥ .
𝑓 2 (x)
This allows us to set
1 + 𝑎 21 tr 𝐷 −2 − 2 𝑎 4𝑛 /𝑎 21 tr 𝐷 −2

𝑔 (𝑡) = .
𝑡2
Next we want to use Lemma A.2.11 to find the function 𝐾. Fix 𝑡0 and extend
𝑓 (x) grad 𝑓
the unit normal of the ellipsoid 𝑆𝑡0 to the inside of 𝑆𝑡0 by 𝜕𝑛 = 𝑡0 |grad 𝑓 | . Then
2 !
2 𝑓 2 (x) 𝐷 −3 x tr 𝐷 −2 𝑓 2 (x)

𝑓 (x) grad 𝑓 1
div = 2− +
𝑡0 |grad 𝑓 | 2 𝑡0 𝐷 −2 x
4
𝐷 −2 x
2
!
1 𝑎2
≥ 2 − 2 𝑛2 + 𝑎 21 tr 𝐷 −2 .
𝑡0 𝑎1
Furthermore, we have that

𝑓 (x) grad 𝑓
∇grad ℎ , grad ℎ
𝑡0 |grad 𝑓 | 2
D E
2𝑓 grad ℎ grad 𝑓 grad ℎ grad 𝑓
2 hgrad 𝑓 , grad ℎ i 2 2 𝑓 (x) ∇ , ,
|grad ℎ| | grad ℎ | |grad ℎ | |grad 𝑓 | |grad ℎ | |grad 𝑓 |
= 2
− 2
𝑡0 |grad 𝑓 | |grad 𝑓 |

𝑓 (x) grad ℎ grad ℎ
+ 2
∇2 𝑓 ,
|grad 𝑓 | |grad ℎ| |grad ℎ|
|grad ℎ| 2 𝐷 −2

grad ℎ grad 𝑓 grad 𝑓 grad ℎ
= 2 , , 𝐼−
𝑡0 |grad ℎ| |grad 𝑓 | |grad 𝑓 | |grad 𝑓 | 2 |grad ℎ|
D E
grad ℎ −2 grad ℎ
|grad ℎ | , 𝐷 |grad ℎ |
+
|grad 𝑓 | 2
!
|grad ℎ| 2 𝑎 2𝑛 2𝑎 21
≤ 2+ 2 − 2 .
𝑡0 𝑎1 𝑎𝑛
65
Hence using Lemma A.2.11 we get that

2 ! !
grad𝑆𝑡 ℎ − ℎ2𝑛 𝑎 2𝑛 𝑎 21
∫ ∫
1 2

−2
𝜎𝑡 ≥ − 2 + 4 2 − 2 − 𝑎 1 tr 𝐷 ℎ 𝑛 ℎ𝜎𝑡
𝑆𝑡 |grad 𝑓 | 𝑓 (x) 𝑎1 𝑎 𝑛 𝑆𝑡
∫
= −𝐾 (𝑡) ℎℎ 𝑛 𝜎𝑡 .
𝑆𝑡
Using Theorem A.2.5 we find the explicit values

!
𝑎 2𝑛 𝑎 21
𝐴 = 1 + 4 2 − 2 + 𝑎 2𝑛 − 𝑎 21 tr 𝐷 −2
𝑎1 𝑎 𝑛
and
! !
𝑎 21
𝑎 2𝑛 𝑎 4 𝑎 4
1
2
𝐵 = 4 2 − 2 + 3𝑎 21 + 3𝑎 2𝑛 − 2 2 − 4 2 tr 𝐷 −2 + 𝑎 21 𝑎 2𝑛 − 𝑎 21 tr 𝐷 −2 .
𝑛
𝑎 𝑛 𝑎1 𝑎1 𝑎𝑛
Note that 𝐴 = 1 if and only if 𝑓 (x) = 𝑐 |x|. In this case, we are integrating over
spheres. In all other cases we have 𝐴 > 1.
Let ℎ be a harmonic function. Define ℎ (𝐷y) = 𝑣 (y). Then div 𝐷 −2 grad 𝑣 =

0. By using change
𝐷 −2 x
∫
𝐻 (𝑡) = ℎ2 (x) 𝜎𝑡
𝑆𝑡 𝐷 −1 x
h𝐷 −2 y, yi
∫
= 𝑣 2 (y) 𝜎˜𝑡 ,
𝐷 −1 (𝑆𝑡 ) |y| 2
where 𝐷 −1 (𝑆𝑡 ) is the sphere with radius 𝑡 and 𝜎˜𝑡 is the spherical measure. This is
the same measure as was considered in [GL86], however they only considered the
case the when div ( 𝐴 (x) grad 𝑢) = 0 when 𝐴 (0) = 𝐼.
A.3.3 Example of the distance function of 𝑆 𝑘 ⊂ R𝑛

Let 𝑘 < 𝑛 and
𝑆 𝑘 = (𝑥 1 , . . . , 𝑥 𝑛 ) : 𝑥 12 + · · · + 𝑥 2𝑘+1 = 1, 𝑥 𝑘+2 = · · · = 𝑥 𝑛 = 0 .

Then the distance from a point x ∈ R𝑛 to the surfaces 𝑆 𝑘 is given by

q
𝑓 (x) = (𝑟 𝑘+1 (x) − 1) 2 + 𝑥 2𝑘+2 + · · · + 𝑥 𝑛2 ,
66
A.3. Examples
where q
𝑟 𝑘+1 (x) = 𝑥 12 + · · · + 𝑥 2𝑘+1 .
This is a special case of Fermi coordinates, see [Cha06], where the submanifold is
𝑆 𝑘 . In the case when 𝑘 = 0 the set 𝑆 𝑘 consists only of two points. In this case, the
function 𝑓 (x) is the usual distance function from x to the nearest of the two points
in 𝑆 0 . When 𝑘 = 1 and 𝑛 = 3 the level surfaces 𝑆𝑡 = 𝑓 −1 (𝑡) for small 𝑡 are tori.
Note that 𝑓 is not smooth along the set of points
{(0, . . . , 0, 𝑥 𝑘+2 , . . . , 𝑥 𝑛 ) : 𝑥 𝑘+2 , . . . , 𝑥 𝑛 ∈ R𝑛 } .
Hence we will only consider values in the range of 𝑓 in [0, 1 − 𝜀) for some 0 <
𝜖 < 1. ∫ Let ℎ : 𝑓 −1 ( [0, 1 − 𝜀)) → R be a harmonic function and consider
𝐻 (𝑡) = 𝑆 ℎ2 (x) 𝜎𝑡 , where 𝑆𝑡 = 𝑓 −1 (𝑡). Again, we wish to apply Theorem A.2.5.
𝑡
The gradient of 𝑓 is given by
x grad 𝑟 𝑘+1
grad 𝑓 = − .
𝑓 (x) 𝑓 (x)
It follows from a computation that |grad 𝑓 | = 1 and the Laplacian of 𝑓 is given by
𝑛 − 1 𝑘/𝑟 𝑘+1 (x)

Δ 𝑓 (x) = − .
𝑓 (x) 𝑓 (x)
We can similarly compute the gradient of the Laplacian and we find that
− (𝑛 − 1) + 2𝑘/𝑟 𝑘+1 (x) − 𝑘/𝑟 𝑘+1 (x) 2

hgrad Δ 𝑓 , grad 𝑓 i = .
𝑓 2 (x)
𝑓 (x) grad 𝑓
Assume that 0 < 𝑡0 < 1 − 𝜖 is fixed. Let 𝜕𝑛 = 𝑡0 be the extension of 𝜕𝑛 to
Inn(𝑆𝑡0 ). If e is a unit vector then
𝑛 − 𝑘/𝑟 𝑘+1 (x) + 2∇2𝑟 𝑘+1 (e, e) − 2

div (𝜕𝑛 ) − 2h∇e 𝜕𝑛 , ei =
𝑡0
𝑛 − 2 − 𝑘/𝜀
≥ .
𝑡0
In short, if we assume that 𝜀 < 𝑟 𝑘+1 < 2 − 𝜀 we obtain the expressions
𝑛 − 1 − 𝑘/𝜀 𝑛 − 1 − 𝑘/(2 − 𝜀)
𝑚 (𝑡) = , 𝑀 (𝑡) = ,
𝑡 𝑡
− (𝑛 − 1) + 2𝑘/(2 − 𝜀) − 𝑘/𝜀 2 𝑘/𝜀 − (𝑛 − 2)

𝑔 (𝑡) = , 𝐾 (𝑡) = .
𝑡2 𝑡
67
Setting
1 1
𝐶 =1+𝑘 −
𝜀 2−𝜀
and
(2𝑛 − 3) (1 − 𝜀) 2
𝐵=𝑘 − 2
𝜀 (2 − 𝜀) 𝜀
we get that 𝐻 satisfy the convexity property
𝐻 00 (𝑡) 𝐻 (𝑡) − 𝐻 0 (𝑡) 2 + 𝐶

𝑡 𝐻 (𝑡) 𝐻 0 (𝑡) 𝐵
2
≥ .
𝐻 (𝑡) 𝑡2
Note that this is again an equation on the same form as (A.3.6). Hence we get the
convexity inequality for 𝐻 given by Corollary A.3.5.
A.3.4 Non-Positive Eigenvalues of −Δ

Let (M, g) be a non-compact 𝑛-dimensional Riemannian manifold with sectional
curvature bounded by
𝜅 ≤ Sec ≤ K,
where 𝜅 and K are constants. Assume that 𝑢 : M → R is a solution to Δ𝑢−𝑘 2 𝑢 = 0.
We will show that the spherical 𝐿 2 -norm of 𝑢 satisfies a convexity property similarly
to harmonic functions.
Denote by 𝑆1/𝑘 ⊂ R2 the circle with radius 1/𝑘. Let 𝑌 (𝜃) be the normalized
first eigenfunction for 𝑆1/𝑘 with eigenvalue −𝑘 2 . Extend the function 𝑢 (𝑥) to a
harmonic function on M × 𝑆1/𝑘 by ℎ (𝑥, 𝜃) = 𝑢 (𝑥) 𝑌 (𝜃), and denote by H𝑘 all
harmonic functions created this way. Define the function 𝑓 (𝑥, 𝜃) = 𝑟 M (𝑥), where
𝑟 M is the radial distance corresponding to a fixed point 𝑝 ∈ M. Then
𝑆𝑡 = 𝑓 −1 (𝑡) = 𝑟 −1
M (𝑡) × 𝑆 1/𝑘 .
In this case, we obtain

∫ ∫
2 2
𝐻 (𝑡) = 𝑌 (𝜃) 𝑢 (𝑥) 𝜎𝑡 = 𝑢 2 (𝑥) 𝜎˜𝑡 ,
𝑆𝑡 −1 (𝑡)
𝑟M
Ñ
M
where 𝜎˜𝑡 is the measure on the geodesic sphere on M.

Note that we can use Theorem A.2.5 on H𝑘 with the function 𝑓 as described
above. Since 𝑓 does not depend on 𝜃, to find 𝑚, 𝑀, and 𝑔 we can use Rauch
Comparison Theorem on 𝑟 M . Using the argumentation found in Section A.3.1 we
get that
𝑚 (𝑡) = (𝑛 − 1) cot K (𝑡) ≤ Δ 𝑓 = Δ𝑟 M (𝑥) ≤ (𝑛 − 1) cot 𝜅 (𝑡) = 𝑀 (𝑡)
68
A.3. Examples
and
hgrad Δ 𝑓 , grad 𝑓 i = hgrad Δ𝑟 M , grad 𝑟 M i ≥ 𝑔 (𝑡) = − (𝑛 − 1) cot2𝜅 (𝑡) − (𝑛 − 1) K.

sin (𝑟 ( 𝑥)) grad 𝑟
Fix 𝑡0 > 0 and let 𝜕𝑛 = K sinK (𝑡0 ) M . Denote by 𝑅𝑡 = 𝑓 −1 ([0, 𝑡)), then we have
that
∫ ∫
2 2
1
| grad𝑆𝑡 ℎ| − ℎ 𝑛 𝜎𝑡 = | grad ℎ| 2 (sin K ( 𝑓 (𝑥)) Δ𝑟 M + cos K ( 𝑓 (𝑥)))
𝑆𝑡 sin K (𝑡) 𝑅𝑡

− 2 sin K ( 𝑓 (𝑥)) ∇2𝑟 M (grad ℎ, grad ℎ) + cos K ( 𝑓 (𝑥)) ℎ2𝑛 vol
𝑛−2
≥ (𝑛 cot K (𝑡) − 2 cot 𝜅 (𝑡)) 𝐷 (𝑡) + min (0, K) 𝐻 (𝑡) ,
2
by using the same argument as in Section A.3.1 and that ℎ ∈ H𝑘 . Using Theorem
A.2.5 on the family H𝑘 we get that 𝐻 satisfies
(log 𝐻 (𝑡)) 00 + (cot K (𝑡) + (𝑛 + 1) (cot 𝜅 (𝑡) − cot K (𝑡))) (log 𝐻 (𝑡)) 0
≥ − (𝑛 − 1) K + (𝑛 − 2) min (K, 0) − (𝑛 − 1) (K − 𝜅) .
In short, solutions to Δ𝑢 = 𝑘 2 𝑢 satisfy the same convexity estimate as harmonic

functions.
69
Paper B
On the Three Ball Theorem for Solutions of the

Helmholtz Equation
Stine Marie Berge and Eugenia Malinnikova
Published in
Complex Analysis and its Synergies, 2021
Paper B
On the Three Ball Theorem for

Solutions of the Helmholtz
Equation
Abstract
Let 𝑢 𝑘 be a solution of the Helmholtz equation with the wave number 𝑘,
Δ𝑢 𝑘 + 𝑘 2 𝑢 𝑘 = 0, on (a small ball in) either R𝑛 , S𝑛 , or H𝑛 . For a fixed
point 𝑝, we define 𝑀𝑢𝑘 (𝑟) = max𝑑 ( 𝑥, 𝑝) ≤𝑟 |𝑢 𝑘 (𝑥)|. The following three ball
inequality
𝑀𝑢𝑘 (2𝑟) ≤ 𝐶 (𝑘, 𝑟, 𝛼) 𝑀𝑢𝑘 (𝑟) 𝛼 𝑀𝑢𝑘 (4𝑟) 1−𝛼
is well known, it holds for some 𝛼 ∈ (0, 1) and 𝐶 (𝑘, 𝑟, 𝛼) > 0 independent of
𝑢 𝑘 . We show that the constant 𝐶 (𝑘, 𝑟, 𝛼) grows exponentially in 𝑘 (when 𝑟 is
fixed and small). We also compare our result with the increased stability for
solutions of the Cauchy problem for the Helmholtz equation on Riemannian
manifolds.
B.1 Introduction
In the present work we study constants in the three ball inequality for solutions of
the Helmholtz equation. We begin by recalling Hadamard’s celebrated three circle
theorem. Let 𝑓 be a holomorphic function in the disk D𝑅 = {𝑧 ∈ C : |𝑧| < 𝑅}.
Then its maximum function
𝑀 𝑓 (𝑟) = max | 𝑓 (𝑧)|

|𝑧 | ≤𝑟
satisfies the convexity condition
𝑀 𝑓 (𝑟 0𝛼 𝑟 11−𝛼 ) ≤ 𝑀 𝑓 (𝑟 0 ) 𝛼 𝑀 𝑓 (𝑟 1 ) 1−𝛼 , (B.1.1)
73
Paper B. Three Ball Theorem
for any 𝑟 0 , 𝑟 1 < 𝑅 and 𝛼 ∈ (0, 1). The proof of (B.1.1) is based on the fact that
log | 𝑓 | is a subharmonic function. Note that by the maximum principle (B.1.1)
also holds when the maximum is taken over circles.
Surprisingly, Hadamard’s theorem generalizes to other classes of functions,
such as solutions of second order elliptic equations and their gradients. We refer
the reader to the article [Lan63] of Landis and to the survey [Ale+09]. Three
spheres theorems for the gradients of harmonic functions and, more generally,
harmonic differential forms can be found in [Mal00]. The three ball theorem
for solutions of the Helmholtz equation on Riemannian manifolds was studied in
[Man13]. This has various applications, for example it was one of the tools used
to estimate the Hausdorff measure of the nodal sets of Laplace eigenfunctions, see
[Log18a; Log18b].
We consider the Helmholtz equation
Δ𝑀 𝑢 𝑘 + 𝑘 2𝑢 𝑘 = 0 (B.1.2)
on a domain 𝐷 in a Riemannian manifold (𝑀, g). For 𝐷 = 𝑀 and 𝑀 being a

closed manifold without boundary, solutions of (B.1.2) are 𝐿 2 -eigenfunctions of
the Laplacian. One of the important facts for analysis on closed manifolds is the
existence of an orthonormal basis for 𝐿 2 (𝑀) consisting of eigenfunctions of the
Laplacian. The classical example is the Fourier basis on the circle S1 . Such an
orthonormal basis can be used to solve the heat, wave, and Schrödinger equations
on closed manifolds, under certain conditions.
We study properties of functions that satisfy the Helmholtz equation on some
geodesic ball in the manifold. Fix a point 𝑝 ∈ 𝑀 and denote by 𝐵𝑟 ( 𝑝) the geodesic
ball of radius 𝑟 centered at 𝑝. Then for a function 𝑢 we define
𝑀𝑢 (𝑟) = max |𝑢(𝑥)|.

𝑥 ∈𝐵𝑟 ( 𝑝)
The following doubling inequality holds for Laplace eigenfunctions on a closed

manifold
𝑀𝑢𝑘 (2𝑟) ≤ 𝐶1 𝑒𝐶2 𝑘 𝑀𝑢𝑘 (𝑟), (B.1.3)
where 𝐶1 and 𝐶2 are constants only depending on the Riemannian manifold (𝑀, g).
Inequality (B.1.3) was first shown by Donnelly and Fefferman in [DF88]. Later
Mangoubi [Man13, Theorem 3.2] gave a new proof by showing the stronger local
inequality
𝑀𝑢𝑘 (3𝑟) ≤ 𝐶3 𝑒𝐶4 𝑘𝑟 𝑀𝑢𝑘 (2𝑟) 𝛼 𝑀𝑢𝑘 (8𝑟) 1−𝛼 , (B.1.4)
for small 𝑟, some fixed 𝛼 ∈ (0, 1), and constants 𝐶3 and 𝐶4 only depending on the
curvature. Further results on the propagation of smallness for eigenfunctions were
obtained in [LM18]. In this article we show that (B.1.4) is sharp in the following
74
B.2. The Three Ball Inequality
sense: The coefficient 𝐶3 𝑒𝐶4 𝑘𝑟 in (B.1.4) cannot be replaced by a function growing

subexponentially in 𝑘𝑟 as 𝑘 grows. This is done by constructing special families of
solutions of the Helmholtz equation on Euclidean spaces, hyperbolic spaces, and
the standard spheres.
We also compare (B.1.4) with the increased stability for solutions of the Cauchy
problem for the Helmholtz equation studied in [HI04; IK11; BNO19]. Roughly
speaking, the idea is that one can estimate the solution in the interior of some
convex domain from an a priori bound and an estimate of the Cauchy data on some
part of the boundary. Moreover, the estimate does not depend on 𝑘. For solutions
of the Helmholtz equation in a geodesic ball 𝐵 𝑅 ( 𝑝) we prove for 𝑟 1 < 𝑅1 < 𝑅 that
∫ ∫
2
𝑢 𝑘 dvol ≤ 𝐶 (𝑟, 𝑅) 𝑢 2𝑘 dvol, (B.1.5)
𝐵𝑟1 ( 𝑝) 𝐵 𝑅1 ( 𝑝)\𝐵𝑟1 ( 𝑝)
and call (B.1.5) the reverse three ball inequality. A more general result can be
found in [ALM16, Section 1.3], where delicate questions regarding localization of
solutions of the Schrödinger equation are considered. We deduce (B.1.5) from a
similar estimate for the 𝐻 1 norms where the constant does not depend on 𝑘. The 𝐻 1
estimate is proved by a Carleman estimate, that can be found in [Isa17; BNO19].
The structure of the paper is as follows. We prove the sharpness of the three
ball inequality (B.1.4) in Section B.2. In Section B.2.1 we present the argument for
the Euclidean space, while the arguments for the hyperbolic space and the sphere
are given in Section B.2.2. We prove inequality (B.1.5) in Section B.3. Finally, we
list some properties of the Bessel functions, and collect some comparison theorems
for solutions of the Sturm-Liouville equations in Appendix.
Acknowledgements
The authors are very grateful to the anonymous referee for useful comments. Their
suggestions, in particular, substantially improved the presentation in Section B.3.
B.2 The Three Ball Inequality

B.2.1 Bessel Functions and the Helmholtz Equation in R𝑛
Let 𝐽𝑙 denote the Bessel function of the first kind. We have collected some facts
about the Bessel functions in Appendix B.4. If 𝑌𝑚 is an eigenfunction of the
Laplace operator on the sphere S𝑛−1 with eigenvalue 𝑚(𝑚 + 𝑛 − 2) then
𝑢 𝑘 (𝑟, 𝜃) = 𝑟 1−𝑛/2 𝐽𝑚+𝑛/2−1 (𝑘𝑟)𝑌𝑚 (𝜃)
solves the Helmholtz equation (B.1.2). Moreover, any solution of (B.1.2) in R𝑛 (or
in the unit ball) can be decomposed into a series of such solutions.
75
In order to study the constant in the three ball inequality (B.1.4) that involves
the maximum function, we analyze the behavior of the Bessel functions. From
now on we assume that 𝑛 = 2 for simplicity. Our results can be easily extended to
all dimensions 𝑛 ≥ 2.
√
Lemma B.2.1. Let 0 < 𝛾 < 𝛿 < 1 and set 𝛽 = 1 − 𝛿2 . Then there exists a
constant 𝐶, only depending on 𝛾 and 𝛿, such that for any positive number 𝑚 we
have 𝛾 𝛽𝑚
𝐽𝑚 (𝛾𝑚) < 𝐶 𝐽𝑚 (𝛿𝑚).
𝛿
Proof. The strategy is to apply the Sturm comparison theorem, see Theorem B.5.2.
We apply the theorem to the Bessel function 𝐽𝑚 solving the Bessel equation
0 0 𝑥2 − 𝑚2
𝑥𝐽𝑚 (𝑥) + 𝐽𝑚 (𝑥) = 0,
𝑥
and a solution of the Euler equation
𝛿2 − 1 𝑚 2

0 0
(𝑥𝑦 (𝑥)) + 𝑦 (𝑥) = 0. (B.2.1)
𝑥
Let 𝑦 be the solution of (B.2.1) satisfying the initial conditions
𝑦(𝛾𝑚) = 𝐽𝑚 (𝛾𝑚) and 𝑦 0 (𝛾𝑚) = 𝐽𝑚

0
(𝛾𝑚).
We know that 𝐽𝑚 is positive and increasing on [0, 𝑚]. The latter can be verified by
using the second derivative test and inserting the argument of the first maximum
of 𝐽𝑚 into the equation
00 0
𝑥 2 𝐽𝑚 (𝑥) + 𝑥𝐽𝑚 (𝑥) + (𝑥 2 − 𝑚 2 )𝐽𝑚 (𝑥) = 0.
Moreover, notice that for 𝑥 ∈ [𝛾𝑚, 𝛿𝑚] we have
𝑥 2 − 𝑚 2 ≤ (𝛿2 − 1)𝑚 2 .
Hence all the conditions in the comparison theorem are satisfied and we conclude
that 𝑦 (𝑥) ≤ 𝐽𝑚 (𝑥) on [𝛾𝑚, 𝛿𝑚].
Any solution of the Euler equation (B.2.1) is on the form
𝑦(𝑥) = 𝑐 1 𝑥 𝑚𝛽 + 𝑐 2 𝑥 −𝑚𝛽 .
Using that
0
𝐽𝑚 (𝛾𝑚) = 𝑦 (𝛾𝑚) > 0 and 𝐽𝑚 (𝛾𝑚) = 𝑦 0 (𝛾𝑚) > 0,
76
we conclude that 𝑐 1 > 0 and |𝑐 2 | < 𝑐 1 𝛾 2𝑚𝛽 𝑚 2𝑚𝛽 . Thus
𝐽𝑚 (𝛾𝑚) = 𝑐 1 (𝛾𝑚) 𝑚𝛽 + 𝑐 2 (𝛾𝑚) −𝑚𝛽 < 2𝑐 1 (𝛾𝑚) 𝑚 𝛽
and
𝑦(𝛿𝑚) > 𝑞𝑐 1 (𝛿𝑚) 𝑚𝛽 ,
where 𝑞 = 𝑞(𝛾, 𝛿) > 0. It follows that
2 𝛾 𝑚𝛽 2 𝛾 𝑚𝛽
𝐽𝑚 (𝛾𝑚) < 2𝑐 1 (𝛾𝑚) 𝑚𝛽 < 𝑦 (𝛿𝑚) < 𝐽𝑚 (𝛿𝑚) .
𝑞 𝛿 𝑞 𝛿

We can now prove the main result of this section.
Theorem B.2.2. Assume that there is an 𝛼 ∈ (0, 1) and a constant 𝐶 (𝑘, 𝑟, 𝛼) such
that for any solution 𝑢 𝑘 of the Helmholtz equation (B.1.2) the following three ball
inequality holds
𝑀𝑢𝑘 (2𝑟) ≤ 𝐶 (𝑘, 𝑟, 𝛼) 𝑀𝑢𝑘 (𝑟) 𝛼 𝑀𝑢𝑘 (4𝑟) 1−𝛼 . (B.2.2)
Then 𝐶 (𝑘, 𝑟, 𝛼) grows at least exponentially in 𝑘𝑟. More precisely, 𝐶 (𝑘, 𝑟, 𝛼) ≥

𝑐𝑒 𝑑 𝛼𝑘𝑟 , where 𝑐 and 𝑑 are absolute constants.
Proof. Consider solutions of the Helmholtz equation on the form
𝑢 𝑘 (𝑟, 𝜃) = 𝐽𝑚 (𝑘𝑟) sin(𝑚𝜃).
The maximal function then simplifies to
𝑀𝑢𝑘 (𝑟) = max |𝐽𝑚 (𝑥)|.

0≤𝑥 ≤𝑘𝑟
We now use the fact that for 𝑚 > 0 the maximum of 𝐽𝑚 (𝑥) is attained in the interval
(𝑚, 𝑚 (1 + 𝜀 (𝑚))), where 𝜀 (𝑚) → 0 as 𝑚 → ∞. This is a well known result
on the asymptotic of the first zero of the Bessel functions, for the convenience of
the reader we include a simple proof in Appendix B.4. We choose 𝑚 0 such that
𝜀 (𝑚) ≤ 1/3 when 𝑚 ≥ 𝑚 0 . Assume first that
𝑘𝑟 > 𝑚 1 = max{4, 2𝑚 0 /3}.
Then given 𝑟 we can find 𝑚 ≥ 𝑚 0 such that
6𝑘𝑟/5 < 𝑚 < 3𝑘𝑟/2.
77
This implies 𝑘𝑟 < 5𝑚/6 and 2𝑘𝑟 > 4/3𝑚. Then 𝑀𝑢𝑘 (4𝑟) = 𝑀𝑢𝑘 (2𝑟) and we can
reduce (B.2.2) to
𝑀𝑢𝑘 (2𝑟) 𝛼

≤ 𝐶 (𝑘, 𝑟, 𝛼).
𝑀𝑢𝑘 (𝑟)
1+𝛾 11
Set 𝛾 = 5/6 and 𝛿 = 2 = 12 . Applying Lemma B.2.1 together with
𝑀𝑢𝑘 (2𝑟) > 𝐽𝑚 (𝑚) > 𝐽𝑚 (𝛿𝑚)
and 𝑀𝑢𝑘 (𝑟) < 𝐽𝑚 (𝛾𝑚) we conclude that

𝛼
𝑀𝑢𝑘 (2𝑟)
𝐶 (𝑘, 𝑟, 𝛼) ≥ ≥ 𝑐𝑒 𝛼𝑑𝑚 ,
𝑀𝑢𝑘 (𝑟)
for some positive constant 𝑑 that can be computed. Finally, since 𝑚 > 6𝑘𝑟/5 we
get the required estimate when 𝑟 > 𝑘 −1 𝑚 1 .
Now for 𝑟 ≤ 𝑘 −1 𝑚 1 we consider the solution 𝑢 𝑘 (𝑟, 𝜃) = 𝐽0 (𝑘𝑟). Then 𝑀𝑢𝑘 (𝑟) =
𝐽0 (0) since
𝐽0 (0) = max |𝐽0 (𝑥)|,
𝑥 ≥0
and we conclude that 𝐶 (𝑘, 𝑟, 𝛼) ≥ 1 for any 𝑟 > 0. Choosing 𝑐 < 𝑒 −𝛼𝑑𝑚1 we have
for all 𝑟 > 0 that
𝐶 (𝑘, 𝑟, 𝛼) ≥ 𝑐𝑒 𝛼𝑑𝑘𝑟 .
B.2.2 Solutions of the Helmholtz Equation on the Sphere and Hyper-

bolic Space
In this section we repeat the argument of the sharpness of the three ball inequality
on the hyperbolic space and sphere. We show in particular that assumptions on
the sign of the curvature do not lead to better behavior of the constant in the three
ball inequality. Again, we use the spherical symmetry of the spaces and separation
of variables to construct a solution of (B.1.2) that is the product of a radial and a
spherical factor. On the sphere the radial part is given by Legendre polynomials.
For the hyperbolic space the radial part is also explicitly known, see [CH94, p. 4222
eq. (2.26)]. Once again, in our argument we only use the differential equation for
the radial part.
We define √

 sin 𝐾 𝑟


 √ , when 𝐾 > 0


 𝐾
sin𝐾 (𝑟) = 𝑟, when 𝐾 = 0 .
 √
sinh −𝐾

 𝑟
when 𝐾 < 0

 √
 ,
 −𝐾
78
Furthermore, we use the associated functions cos𝐾 (𝑟) = (sin𝐾 (𝑟)) 0 , cot𝐾 (𝑟) =
cos𝐾 (𝑟 ) 1
sin𝐾 (𝑟 ) , and tan𝐾 (𝑟) = cot𝐾 (𝑟 ) . Then the Laplacian of a simply connected 𝑛-
dimensional Riemannian manifold (𝑀, g) with constant sectional curvature 𝐾 is
given in polar coordinates by
𝑑2 𝑑 1
Δ𝑀 = + (𝑛 − 1) cot𝐾 (𝑟) + 2 Δ 𝑛−1 .
𝑑𝑟 2 𝑑𝑟 sin𝐾 (𝑟) S
In this section we work in dimension two. Assume that 𝑢 𝑘 (𝑟, 𝜃) = 𝑅 (𝑟) Θ(𝜃) is a
solution of the Helmholtz equation. Then 𝑅 (𝑟) satisfies the equation
00
𝑅 (𝑟) + cot𝐾 (𝑟) 𝑅 0 (𝑟)

2 2 Δ 1 Θ (𝜃)
sin𝐾 (𝑟) +𝑘 =− S = 𝑚2 . (B.2.3)
𝑅(𝑟) Θ (𝜃)
Let 𝜅 = 𝐾/𝑘 2 and let 𝐿 𝜅 ,𝑚 (𝜌) be the solution of the differential equation
sin2𝜅 (𝜌) 𝐿 00 0
𝜅 ,𝑚 (𝜌) + sin 𝜅 (𝜌) cos 𝜅 (𝜌) 𝐿 𝜅 ,𝑚 (𝜌)

+ sin2𝜅 (𝜌) − 𝑚 2 𝐿 𝜅 ,𝑚 (𝜌) = 0, (B.2.4)
where 𝐿 𝜅,𝑚 is well defined at 𝜌 = 0 and positive on some interval (0, 𝜀). Then
for 𝑚 > 0 we have 𝐿 𝜅 ,𝑚 (0) = 0. Note that when 𝐾 = 0 this equation becomes
the Bessel equation. Setting 𝑅 (𝑟) = 𝐿 𝜅,𝑚 (𝑘𝑟) we get a solution to (B.2.3). Then
(B.2.4) can be rewritten in the Sturm-Liouville form as
0 sin2𝜅 (𝜌) − 𝑚 2
sin 𝜅 (𝜌) 𝐿 0𝜅 ,𝑚 (𝜌) + 𝐿 𝜅 ,𝑚 (𝜌) = 0. (B.2.5)
sin 𝜅 (𝜌)
We begin by estimating the maximum point of 𝐿 𝜅 ,𝑚 from below. Let
(
∞, 𝜅≤0
𝑅𝜅 = 𝜋 .
√ , 𝜅 > 0
2 𝜅
Note that for 𝜌 ≤ 𝑅 𝜅 we have that sin 𝜅 (𝜌) is increasing, or equivalently that
cos 𝜅 (𝜌) ≥ 0.
Proposition B.2.3. Let 0 < 𝜌1∗ < 𝜌2∗ < · · · < 𝑅 𝜅 be the points where 𝐿 𝜅 ,𝑚 attains
local maximums and minimums before 𝑅 𝜅 . Then 𝐿 𝜅 ,𝑚 𝜌𝑖∗ is a decreasing

sequence in 𝑖. Moreover, the first local maximum 𝜌1∗ satisfies 𝜌1∗ ≥ sin−1
𝜅 (𝑚).
Proof. At 𝜌1∗ we have 𝐿 0𝜅 ,𝑚 (𝜌1∗ ) = 0 and (B.2.5) implies that

sin2𝜅 𝜌1∗ 𝐿 00 ∗ 2 ∗ 2
𝐿 𝜅 ,𝑚 𝜌1∗ = 0.

𝜅 ,𝑚 𝜌 1 + sin 𝜅 𝜌 1 − 𝑚
79
By the second derivative test it is not possible to have a maximum before sin−1
𝜅 (𝑚),
implying the lower bound for the first local extremum.
The remaining part of the proposition follows from Sonin-Pólya oscillation
theorem, see Theorem B.5.3. The conditions in the oscillation theorem are satisfied
on the interval sin−1
𝜅 (𝑚) , 𝑅 𝜅 since sin 𝜅 (𝜌) > 0,
(sin2𝜅 (𝜌) − 𝑚 2 )/sin 𝜅 (𝜌) ≠ 0,
and 0
sin2𝜅 (𝜌) − 𝑚 2

sin 𝜅 (𝜌) = 2 cos 𝜅 (𝜌) sin 𝜅 (𝜌) > 0.
sin 𝜅 (𝜌)
Thus the sequence |𝐿 𝜅,𝑚 (𝜌𝑖∗ )| is decreasing.
Remark B.2.4. For 𝑚 = 0, by analyzing the differential equation (B.2.4), we see

that 𝐿 0𝜅 ,0 (0) = 0. Then the proof of Proposition B.2.3 implies that 𝐿 𝜅 ,0 (𝜌) satisfies
𝐿 𝜅 ,0 (0) ≥ |𝐿 𝜅 ,0 (𝜌)| for 𝜌 > 0.
Now our aim is to prove an analog of Lemma B.2.1. The next four results show
how we can control the ratio of two values of 𝐿 𝜅 ,𝑚 .
Lemma B.2.5. Let 𝜌2 ∈ (0, 𝑅 𝜅√) and 𝛿 ∈ (0, 1) satisfy the inequality sin 𝜅 (𝜌2 ) ≤
𝛿𝑚. Then for 𝜌1 < 𝜌2 and 𝛽 = 1 − 𝛿2 we have the bound
" 𝛽𝑚 −𝛽𝑚 #
𝐿 𝜅 ,𝑚 (𝜌2 ) 1 tan 𝜅 (𝜌2 /2) tan 𝜅 (𝜌2 /2)
≥ − .
𝐿 𝜅,𝑚 (𝜌1 ) 2 tan 𝜅 (𝜌1 /2) tan 𝜅 (𝜌1 /2)
Proof. We compare the function 𝐿 𝜅 ,𝑚 to a solution of the equation
𝑚 2 (𝛿2 − 1)
(sin 𝜅 (𝜌)𝑦 0 (𝜌)) 0 + 𝑦(𝜌) = 0. (B.2.6)
sin 𝜅 (𝜌)
By the assumption we have
sin2𝜅 (𝜌) − 𝑚 2 ≤ (𝛿2 − 1)𝑚 2 = −𝛽2 𝑚 2
on the interval [𝜌1 , 𝜌2 ]. Let 𝑦 be the solution to (B.2.6) that satisfies the initial
conditions
𝑦(𝜌1 ) = 𝐿 𝜅 ,𝑚 (𝜌1 ) and 𝑦 0 (𝜌1 ) = 𝐿 0𝜅 ,𝑚 (𝜌1 ).
Then the comparison theorem implies that 𝐿 𝜅 ,𝑚 (𝜌2 ) > 𝑦(𝜌2 ).
The explicit solution to (B.2.6) is given by
𝛽𝑚 −𝛽𝑚
𝑦(𝜌) = 𝑐 1 tan 𝜅 (𝜌/2) + 𝑐 2 tan 𝜅 (𝜌/2).
80
The first maximum 𝜌1∗ of 𝐿 𝜅 ,𝑚 satisfies sin 𝜅 (𝜌1∗ ) ≥ 𝑚 implying that 𝜌2 < 𝜌1∗ .
Therefore 𝐿 𝜅 ,𝑚 (𝜌1 ) > 0 and 𝐿 0𝜅 ,𝑚 (𝜌1 ) > 0. Thus we have the inequality
2𝛽𝑚 2𝛽𝑚
−𝑐 1 tan 𝜅 (𝜌1 /2) < 𝑐 2 < 𝑐 1 tan 𝜅 (𝜌1 /2),
since (tan 𝜅 (𝜌/2)) 0 > 0. We conclude that 𝑐 1 > 0 and similarly to Lemma B.2.1
we get
𝛽𝑚 2𝛽𝑚 −𝛽𝑚
𝐿 𝜅 ,𝑚 (𝜌2 ) > 𝑦(𝜌2 ) > 𝑐 1 (tan 𝜅 (𝜌2 /2) − tan 𝜅 (𝜌1 /2) tan 𝜅 (𝜌2 /2)).
The estimate of 𝑐 2 from below implies that
𝛽𝑚
𝐿 𝜅 ,𝑚 (𝜌1 ) = 𝑦(𝜌1 ) < 2𝑐 1 tan 𝜅 (𝜌1 /2).
Combining the last two inequalities gives the result.
Corollary B.2.6. Suppose
√ that 𝐾 > 0 and that 𝜌1 < 𝜌2 < min{𝑅 𝜅 , 𝑚𝛿} for some
2
𝛿 ∈ (0, 1). For 𝛽 = 1 − 𝛿 we have the estimate
" −𝛽𝑚 #
𝛽𝑚
𝐿 𝜅 ,𝑚 (𝜌2 ) 1 𝜌2 𝜌2
≥ − . (B.2.7)
𝐿 𝜅 ,𝑚 (𝜌1 ) 2 𝜌1 𝜌1
Proof. We note that sin 𝜅 (𝜌2 ) < 𝜌2 < 𝑚𝛿. Applying Lemma B.2.5 and using the
elementary inequality 𝑏 tan 𝑥 ≥ tan 𝑏𝑥 for 𝑏 ∈ (0, 1), the result follows since
√
tan 𝜅 (𝜌2 /2) tan( 𝜅𝜌2 /2) 𝜌2
= √ ≥ .
tan 𝜅 (𝜌1 /2) tan( 𝜅𝜌1 /2) 𝜌1
Corollary B.2.7. Let 𝐾 < 0 and suppose that
𝜌1 < 𝜌2 < min{𝑅 |𝜅 | , 2𝑚𝛿/3}
√
for some 𝛿 ∈ (0, 1). Then for 𝛽 = 1 − 𝛿2 and 𝐴 = sin 𝜅 (𝜌2 ) /𝜌2 we have
𝐿 𝜅 ,𝑚 (𝜌2 ) 1 𝜌2 𝜌2
≥ − . (B.2.8)
𝐿 𝜅 ,𝑚 (𝜌1 ) 2 𝐴𝜌1 𝐴𝜌1
p
Proof. Since |𝜅|𝜌 < 𝜋/2 and sinh is convex we have
sin 𝜅 (𝜌2 ) ≤ 2𝜌2 sinh (𝜋/2) /𝜋 < 3𝜌2 /2 < 𝑚𝛿.
Applying Lemma B.2.5 together with
√
0 𝜌2 −𝜅 √
(log(tanh 𝑥)) ≥ √ for 𝑥 < 𝜌2 −𝜅/2,
sinh 𝜌2 −𝜅 𝑥
gives (B.2.8), since
p
tan 𝜅 (𝜌2 /2) tanh( |𝜅|𝜌2 /2) 𝜌2
= ≥ .
tan 𝜅 (𝜌1 /2) tanh( |𝜅|𝜌1 /2)
p
𝐴𝜌1
81
We want to estimate the ratio of the values of 𝐿 𝜅 ,𝑚 at two points 𝜌2 > 𝜌1 >
sin−1
𝜅 (𝑚). In contrast with the Bessel functions, we do not locate the maximum
precisely.
Lemma B.2.8. Suppose that 0 < 𝜌1 < 𝑅 |𝜅 | and sin 𝜅 (𝜌1 ) > 𝜉𝑚, where 𝜉 > 1.
There is an absolute constant 𝐶 > 0 such that
max𝜌 |𝐿 𝜅 ,𝑚 (𝜌)| 𝐶
≤ 1+ . (B.2.9)
max𝜌 ≤𝜌1 |𝐿 𝜅,𝑚 (𝜌)| (𝜉 − 1)𝑚
Proof. Let 𝜌1∗ be the first local maximum of 𝐿 𝜅 ,𝑚 . By Proposition B.2.3 if 𝜌1 > 𝜌1∗
then the left-hand side of (B.2.9) is one and the statement becomes trivial.
The rest of the proof relies on the comparison of 𝐿 𝜅 ,𝑚 and a solution of the
equation
𝑚 2 (𝜉 2 − 1)
(sin 𝜅 (𝜌)𝑦 0) 0 + 𝑦=0 (B.2.10)
sin 𝜅 (𝜌)
on the interval (𝜌1 , ∞). Solutions to (B.2.10) are of the form
𝑦(𝜌) = 𝑐 1 cos(𝛾 log(tan 𝜅 (𝜌/2))) + 𝑐 2 sin(𝛾 log(tan 𝜅 (𝜌/2))),
where 𝛾 2 = 𝜉 2 − 1. Let 𝑑 = 𝛾 log(tan 𝜅 (𝜌1 /2)) and choose a solution 𝑦 of (B.2.10)

on the form
𝑦(𝜌) = 𝐶1 cos(𝛾 log(tan 𝜅 (𝜌/2)) − 𝑑) + 𝐶2 sin(𝛾 log(tan 𝜅 (𝜌/2)) − 𝑑),
with initial data 𝑦(𝜌1 ) = 𝐿 𝜅 ,𝑚 (𝜌1 ) and 𝑦 0 (𝜌1 ) = 𝐿 0𝜅 ,𝑚 (𝜌1 ). This gives the values
𝐿 0𝜅 ,𝑚 (𝜌1 ) sin 𝜅 (𝜌1 )

𝐶1 = 𝐿 𝜅 ,𝑚 (𝜌1 ), 𝐶2 = .
𝛾
Applying the comparison theorem, we get
𝐿 𝜅 ,𝑚 (𝜌2 ) ≤ 𝐶1 cos(𝛾 log(tan 𝜅 (𝜌2 /2)) − 𝑑) + 𝐶2 sin(𝛾 log(tan 𝜅 (𝜌2 /2)) − 𝑑).
Since sin and cos are bounded by 1, we estimate 𝐶2 /𝐿 𝜅 ,𝑚 (𝜌1 ) from above to
prove (B.2.9). In order to estimate 𝐶2 , we see that the assumption 𝜌1 < 𝜌1∗ implies
𝐿 𝜅 ,𝑚 (𝜌) > 0 and 𝐿 0𝜅 ,𝑚 (𝜌) > 0 on the interval (0, 𝜌1 ). Equation (B.2.5) shows that
𝐿 00 −1
𝜅 ,𝑚 (𝜌) < 0 when 𝜌 ∈ (𝜌0 , 𝜌1 ), where 𝜌0 = sin 𝜅 (𝑚). Then the Taylor formula
gives
𝐿 𝜅 ,𝑚 (𝜌0 ) − 𝐿 𝜅 ,𝑚 (𝜌1 ) + (𝜌1 − 𝜌0 )𝐿 0𝜅 ,𝑚 (𝜌1 ) < 0.
Consequently,
𝐿 𝜅,𝑚 (𝜌1 ) − 𝐿 𝜅 ,𝑚 (𝜌0 ) 𝐿 𝜅 ,𝑚 (𝜌1 )
𝐿 0𝜅 ,𝑚 (𝜌1 ) < < .
𝜌1 − 𝜌0 𝜌1 − 𝜌0
82
For 𝐾 > 0 the inequality
𝑥1 − 𝑥 0 > sin 𝑥 1 − sin 𝑥 0 when 𝑥 1 > 𝑥0
implies that 𝜌1 − 𝜌0 > (𝜉 − 1)𝑚. For 𝐾 < 0, we note that
𝑥 1 − 𝑐 sinh 𝑥 1 > 𝑥0 − 𝑐 sinh 𝑥 0 when 𝑥 1 > 𝑥0 ,
if cosh 𝑥 1 < 𝑐−1 . Using the assumption 𝜌1 < 𝑅 |𝜅 | , we conclude that 𝜌1 − 𝜌0 >
𝑐(𝜉 − 1)𝑚, where 𝑐 = (cosh 𝜋/2) −1 . Finally, we obtain
q
2 2 𝐶
max |𝐿 𝜅 ,𝑚 (𝜌)| ≤ 𝐶1 + 𝐶2 ≤ 𝐿 𝜅 ,𝑚 (𝜌1 ) 1 + .
𝜌 ≥𝜌1 (𝜉 − 1)𝑚
Now we are ready to prove that the coefficient in the three ball theorem grows
exponentially in 𝑟 𝑘 if we restrict ourselves to balls with sufficiently small radius 𝑟.
Theorem B.2.9. Let (𝑀, g) be either a hyperbolic plane or a sphere and denote
its curvature by 𝐾. Suppose that for some 𝛼 ∈ (0, 1) there exists a constant
𝐶 𝛼 (𝑘, 𝑟, 𝐾) such that for any solution 𝑢 𝑘 to the Helmholtz equation (B.1.2) the
following inequality holds
𝜋
𝑀𝑢𝑘 (2𝑟) ≤ 𝐶 𝛼 (𝑘, 𝑟, 𝐾) 𝑀𝑢𝑘 (𝑟) 𝛼 𝑀𝑢𝑘 (4𝑟) 1−𝛼 , 0<𝑟 < p .
8 |𝐾 |
Then
𝐶 𝛼 (𝑘, 𝑟, 𝐾) ≥ 𝑐 1𝛼 𝑒 𝑐2 𝛼𝑘𝑟 , (B.2.11)
where 𝑐 1 and 𝑐 2 only depend on 𝐾.
Proof. Consider the family of functions
𝑢 𝑘,𝑚 (𝑟, 𝜃) = 𝐿 𝜅,𝑚 (𝑘𝑟) sin (𝑚𝜃) ,
where 𝑚 is a non-negative integer. By construction, 𝑢 𝑘,𝑚 solves the Helmholtz

equation. Thus for any 𝑚 we have the inequality
𝛼 1−𝛼
𝑀𝑚 (2𝑘𝑟) 𝑀𝑚 (2𝑘𝑟)
𝐶 𝛼 (𝑘, 𝑟, 𝐾) ≥ ,
𝑀𝑚 (𝑘𝑟) 𝑀𝑚 (4𝑘𝑟)
where
𝑀𝑚 (𝜌) = max |𝐿 𝜅 ,𝑚 (𝑥)|.
𝑥 ≤𝜌
Note that choosing 𝑚 = 0 gives 𝐶 𝛼 (𝑘, 𝑟, 𝐾) ≥ 1 by Remark B.2.4. Thus if we

assume that 𝑘𝑟 < 𝐶1 for some constant 𝐶1 , we may choose 𝑐 2 and 𝑐 1 small enough
such that the inequality holds.
83
Assume first that 𝐾 < 0 so that (𝑀, g) is the hyperbolic plane. If 𝑘𝑟 > 𝐶1 we
choose a positive integer 𝑚 such that 10𝑚 < 18𝑘𝑟 < 11𝑚. We apply (B.2.8) with
𝜌1 = 𝑘𝑟 and 𝜌2 = 𝜇𝑘𝑟 < 2𝑘𝑟, where 𝜇 = 19/17. Then 𝜌2 < 2/3𝑚𝛿 with 𝛿 < 1.
We obtain
𝑀𝑚 (2𝑘𝑟) 𝐿 𝜅,𝑚 (𝜌2 ) 1 𝜌2 𝜌2
≥ ≥ − ,
𝑀𝑚 (𝑘𝑟) 𝐿 𝜅 ,𝑚 (𝜌1 ) 2 𝐴𝜌1 𝐴𝜌1
√
where 𝛽 = 1 − 𝛿2 and
𝐴 = sin 𝜅 (𝜌2 ) /𝜌2 < sin 𝜅 (2𝑘𝑟) /(2𝑘𝑟) < 4 sinh(𝜋/4)/𝜋 < 10/9.
Therefore 𝑞 = 𝜇/𝐴 > 1 is an absolute constant and we have
𝑀𝑚 (2𝑘𝑟) 1
≥ (𝑞 𝛽𝑚 − 𝑞 −𝛽𝑚 ). (B.2.12)
𝑀𝑚 (𝑘𝑟) 2
Thus there are 𝑐 1 > 0 and 𝑐 2 > 0 such that
𝑀𝑚 (2𝑘𝑟) ≥ 𝑐 1 exp(𝑐 2 𝑚)𝑀𝑚 (𝑘𝑟).
On the other hand, we have 2𝑘𝑟 > 𝜉𝑚 for 𝜉 = 10/9. Applying (B.2.9) we get
𝑀𝑚 (4𝑟) max𝜌 ≤4𝑘𝑟 |𝐿 𝜅 ,𝑚 (𝜌)|

= ≤ 1 + 𝐶/𝑚 ≤ 𝐶0 ,
𝑀𝑚 (2𝑟) max𝜌 ≤2𝑘𝑟 |𝐿 𝜅 ,𝑚 (𝜌)|
where 𝐶0 is an absolute constant. Note also that 𝑚 & 𝑘𝑟. Then (B.2.11) follows
for negative curvature.
Assume now that 𝐾 > 0 so that (𝑀, g) is a sphere. If 𝑘𝑟 > 𝐶1 we choose
𝑚 to be a positive integer such that 10𝑚 < 12𝑘𝑟 < 11𝑚. We first let 𝜌1 = 𝑘𝑟
and 𝜌2 = 13𝑘𝑟/12 and apply (B.2.7) with 𝛿 = 143/144. Thus (B.2.12) follows
whenever 𝑘𝑟 > 𝐶1 . Using (B.2.9) with 𝜌1 = 2𝑘𝑟 we need to check that 2𝜌1 > 𝜉𝜋𝑚
for some 𝜉 > 1. Note that 2𝜌1 = 4𝑘𝑟 > 10/3𝑚 and choose 𝜉 < 310𝜋 . Then (B.2.11)
follows for positive curvature.
B.3 The Reverse Three Ball Inequality

The question of stability of the solution to the Cauchy problem for the Helmholtz
equation and the dependence of the estimates on the wave number 𝑘 was studied
by many authors, see e.g. [HI04; SI07; IK11; BNO19]. We include a special
case of the results adapted to the case of Riemannian manifolds to demonstrate the
difference between the usual three ball theorem and the reverse one.
84
B.3. The Reverse Three Ball Inequality
Let (𝑀, g) be a Riemannian manifold with sectional curvature satisfying
𝜅 ≤ sec ≤ 𝐾.
We denote by grad 𝑀 and Δ 𝑀 the gradient and Laplace operators on functions on

𝑀. Let 𝐵 be a geodesic ball with diameter less than the injectivity radius of (𝑀, g).
Additionally, in the case that 𝐾 > 0 we assume that the diameter of 𝐵 is strictly
less than √𝜋 .
2 𝐾
We say that a function 𝜙 : 𝐵 → R is convex if its Hessian is positive definite.
We choose a point 𝑝 such that 𝑝 ∉ 𝐵 but dist( 𝑝, 𝑥) is less that the injectivity radius
and √𝜋 for the case 𝐾 > 0, and consider 𝜙(𝑥) = dist(𝑥, 𝑝) this function is smooth
2 𝐾
on 𝐵 (since the metric on the Riemannian manifold is assumed to be smooth and
𝑝 ∉ 𝐵 while 𝐵 is contained in the ball of the injectivity radius around 𝑝). Moreover,
𝐻𝑒𝑠𝑠(𝜙) is strictly positive definite on 𝐵 and 𝜙 has no critical points, see [Pet16,
Theorem 6.4.8] and the preceding discussions. By repeating the computations of
[BNO19, Lemma 1], where it is also pointed out that the result holds on Riemannian
manifolds, we obtain the following point-wise inequality, where we also use the
notation 𝑣 = 𝑒 𝑡 𝜙 𝑤,
𝑒 2𝑡 𝜙 (Δ 𝑀 𝑤 + 𝑘 2 𝑤) 2 ≥ 2 div(𝑏 grad 𝑀 𝑣 + 𝑎) + 4𝑡h∇2𝑀 (𝜙) grad 𝑀 𝑣, grad 𝑀 𝑣i 𝑀

+ 4𝑡 3 h∇2𝑀 𝜙 grad 𝑀 𝜙, grad 𝑀 𝜙i 𝑀 𝑣 2 + 𝑡hgrad 𝑀 Δ 𝑀 𝜙, grad 𝑀 𝑣i 𝑀 𝑣,
where 𝑏 = −𝑡𝑣Δ 𝑀 𝜙 − 2𝑡hgrad 𝑀 𝑣, grad 𝑀 𝜙i 𝑀 and
𝑎 = 𝑡 (| grad 𝑀 𝑣| 2𝑀 − (𝑘 2 + 𝑡 2 | grad 𝑀 𝜙| 2𝑀 )𝑣 2 ) grad 𝑀 𝜙.
Lemma B.3.1. Let (𝑀, g) and 𝐵 be as above. Then there exists a constant 𝑐 0 > 0
such that for any function 𝑤 ∈ 𝐶02 (𝐵) and 𝑘 ≥ 0 the following inequality holds
∫ ∫
2 2
|Δ 𝑀 𝑤 + 𝑘 𝑤| dvol ≥ 𝑐 0 |𝑤| 2 + | grad 𝑀 𝑤| 2𝑀 dvol . (B.3.1)
𝐵 𝐵
Proof. We repeat the argument given in [BNO19, Corollary 1]. Integrating the
last inequality over a ball 𝐵 and taking into account that functions 𝑏 and 𝑎 have
compact support in 𝐵, we conclude that the divergence term disappears. For the
next two terms, which contain the Hessian of 𝜙, we use the convexity inequality
h∇2𝑀 𝜙 grad 𝑀 𝑓 , grad 𝑀 𝑓 i 𝑀 ≥ 𝑐 𝜙 | grad 𝑀 𝑓 | 2𝑀
and the computation
grad 𝑀 𝑣 = 𝑒 𝑡 𝜙 (grad 𝑀 𝑤 + 𝑡𝑤 grad 𝑀 𝜙).
85
Finally, the last term is estimates as
𝑡hgrad 𝑀 Δ 𝑀 𝜙, grad 𝑀 𝑣i 𝑀 𝑣 ≤ 𝜀𝑡| grad 𝑀 𝑣| 2𝑀 + 𝜀 −1 𝑡| grad 𝑀 Δ 𝑀 𝜙| 2𝑀 𝑣 2 .
Combining these estimates, we get

∫ ∫
2 2 2𝑡 𝜙
|Δ 𝑀 𝑤 + 𝑘 𝑤| 𝑒 dvol ≥ 𝑐 1 𝑡 3 |𝑤| 2 𝑒 2𝑡 𝜙 + 𝑡| grad 𝑤| 2 𝑒 2𝑡 𝜙 dvol,
𝐵 𝐵
when 𝑡 > 𝑡0 . The powerful feature of the last inequality is that 𝑐 1 and 𝑡0 do
not depend on 𝑘 (but depend on 𝜙 which we fix). Finally, we fix some 𝑡 > 𝑡0
and let 𝑀 = max 𝐵 𝑒 2𝑡 𝜙 and 𝑚 = min 𝐵 𝑒 2𝑡 𝜙 . Then (B.3.1) holds with 𝑐 0 =
𝑐 1 𝑚 min{𝑡 3 , 𝑡}𝑀 −1 .
Suppose now that 𝑢 𝑘 is a solution of the Helmholtz equation (B.1.2) in a ball

𝐵 that satisfies the conditions in Lemma B.3.1. We apply inequality (B.3.1) to
𝑤 = 𝑢 𝑘 𝜒, where 𝜒 ∈ 𝐶02 (𝐵) is compactly supported on 𝐵 and equals to one on a
smaller ball 𝐵1 ⊂⊂ 𝐵. This gives the inequality
∫ ∫
2 2 1
|𝑢 𝑘 | + | grad 𝑢 𝑘 | dvol ≤ |𝑢 𝑘 Δ 𝑀 𝜒 + 2 grad 𝑢 𝑘 · grad 𝜒| 2 dvol .
𝐵1 𝐶0 𝐵\𝐵1
The last inequality implies that for any 𝑟 < 𝑅 such that 𝐵 𝑅 = 𝐵 𝑅 (𝑥) and 𝐵𝑟 = 𝐵𝑟 (𝑥)
are geodesic balls satisfying the conditions in Lemma B.3.1, there is a constant
𝐶2 (𝑟, 𝑅) such that
∫
|𝑢 𝑘 | 2 + | grad 𝑢 𝑘 | 2 dvol ≤
𝐵𝑟
∫
𝐶2 (𝑟, 𝑅) |𝑢 𝑘 | 2 + | grad 𝑢 𝑘 | 2 dvol . (B.3.2)
𝐵 𝑅 \𝐵𝑟
Inequality (B.3.2) shows that if 𝑢 2𝑘 + | grad 𝑢 𝑘 | 2 is small on the annulus 𝐵 𝑅 \ 𝐵𝑟 ,

then it is small on the whole ball 𝐵 𝑅 . For the Euclidean space an alternative proof
can be obtained by decomposing a solution 𝑢 𝑘 into series of products of Bessel
functions and spherical harmonics. From this, one can deduce (B.3.2) from the
Debye asymptotic of the Bessel functions.
To compare with the previous section, we can also use Caccioppoli’s inequality
to control the Sobolev norm of 𝑢 𝑘 by its 𝐿 2 norm.
Lemma B.3.2 (Caccioppoli’s inequality). Let 𝜀 > 0 and let 𝑅 = 𝑅(𝑀) be small
enough. Furthermore, let 𝜀 < 𝑟 < 𝑅 − 𝜀. We denote Ω = 𝐵 𝑅 (𝑥) \ 𝐵𝑟 (𝑥) and
Ω+ = 𝐵 𝑅+𝜀 (𝑥) \ 𝐵𝑟 −𝜀 (𝑥), Ω− = 𝐵 𝑅−𝜀 (𝑥) \ 𝐵𝑟 +𝜀 (𝑥).
86
B.4. Appendix: The First Positive Zero of the Bessel Function
Assume that 𝑢 𝑘 ∈ 𝐶 2 (Ω) and Δ 𝑀 𝑢 𝑘 + 𝑘 2 𝑢 𝑘 = 0 in Ω+ . Then there exists a constant

𝐶 = 𝐶 (𝑀) such that
∫ ∫ ∫
𝐶
𝑘2 𝑢 2𝑘 dvol − 2 𝑢 2𝑘 dvol ≤ |grad 𝑢 𝑘 | 2 dvol
Ω− 𝜀 Ω Ω
∫
2 𝐶
≤ 𝑘 + 2 𝑢 2𝑘 dvol .
𝜀 Ω+
Proof. There exists a smooth function 𝜑+ with compact support in Ω+ that satisfy
𝜑+ = 1 on Ω and |grad 𝜑+ | ≤ 𝐶𝜀 and |Δ 𝑀 𝜑+ | ≤ 𝜀𝐶2 . Then, using the divergence
theorem, we have
∫ ∫
2 2
𝑘 𝜑+ 𝑢 𝑘 dvol = − 𝜑+ 𝑢 𝑘 Δ 𝑀 𝑢 𝑘 dvol
Ω+ Ω+
∫
= hgrad 𝑢 𝑘 , grad (𝜑+ 𝑢 𝑘 )i dvol
Ω+
∫ ∫
1
= 𝜑+ |grad 𝑢 𝑘 | 2 dvol − 𝑢 2𝑘 Δ 𝑀 𝜑+ dvol .
Ω+ 2 Ω+
Hence ∫ ∫
|grad 𝑢 𝑘 | 2 dvol ≤ 𝑘 2 + 𝐶𝜀 −2 𝑢 2𝑘 dvol .
Ω Ω+
On the other hand, choosing a similar function 𝜑− ∈ 𝐶0∞ (Ω) such that 𝜑− = 1 on
Ω− , we conclude that
∫ ∫ ∫
2 2 2 −2
| grad 𝑢 𝑘 | dvol ≥ 𝑘 𝑢 𝑘 − 𝐶𝜀 𝑢 2𝑘 dvol .
Ω Ω− Ω
We now go back to the inequality (B.3.2), and apply the Caccioppoli inequality.
Rename 𝑅1 = 𝑅 + 𝜀 and 𝑟 1 = 𝑟 − 𝜀. This gives the following estimate of the 𝐿 2
norm of a solution to the Helmholtz equation by its 𝐿 2 norm on an annulus
∫ ∫ ∫
𝑘2 𝑢 2𝑘 dvol ≤ | grad 𝑢 𝑘 | 2 dvol +𝐶𝜀 −2 𝑢 2𝑘 dvol
𝐵𝑟1 𝐵𝑟 𝐵𝑟
∫ ∫
2 −2 −2
≤ (𝑘 + 𝐶𝜀 ) 𝑢 2𝑘 dvol +𝐶𝜀 𝑢 2𝑘 dvol,
𝐵 𝑅1 \𝐵𝑟1 𝐵𝑟1
for any 𝑟 1 < 𝑅1 . Then for 𝑘 > 𝐶𝜀 −1 the inequality (B.1.5) follows. For 𝑘 < 𝐶𝜀 −1 ,
we use (B.3.2) and the Caccioppoli inequality again, to see that
∫ ∫
2 2 2
|𝑢 𝑘 | dvol ≤ 𝐶 (𝑟, 𝑅)(𝑘 + 𝐶𝜀 ) |𝑢 𝑘 | 2 dvol .
𝐵𝑟 𝐵 𝑅1 \𝐵𝑟1
Thus, since 𝑘 < 𝐶𝜀 −1 , the inequality (B.1.5) follows.
87
B.4 Appendix: The First Positive Zero of the Bessel Func-

tion
Let 𝑙 be a non-negative half-integer, and let Γ denotes the gamma function. The
Bessel function 𝐽𝑙 is a solution to the second order ODE

𝜌 2 𝐽𝑙00 (𝜌) + 𝜌𝐽𝑙0 (𝜌) + 𝜌 2 − 𝑙 2 𝐽𝑙 (𝜌) = 0,
which is bounded at the origin and normalized by the condition

lim 𝜌 −𝑙 𝐽𝑙 (𝜌) = 2−𝑙 Γ(𝑙 + 1) −1 .
𝜌→0
For alternative definitions and many useful asymptotic formulas for the Bessel
functions we refer the reader to [Olv97].
It is well known that the first positive zero of 𝐽𝑙 , usually denoted by 𝑗𝑙 , satisfies
𝑗𝑙 𝑙 + 𝑐𝑙 1/3 as 𝑙 → ∞. We already explained in the proof of Lemma B.2.1 that
𝑗𝑙 > 𝑙 and we give a simple proof of the inequality 𝑗𝑙 ≤ 𝑙 + 𝑐𝑙 1/3 . This will be done
by comparing the Bessel equation to the following equation
𝜌 2 𝑦 𝑙00 (𝜌) + 𝜌𝑦 𝑙0 (𝜌) + (𝑎 𝑙2 𝜌 2 − 1/4)𝑦 𝑙 (𝜌) = 0 (B.4.1)
on the interval 𝜌 ∈ [𝑙 + 𝑙 1/3 , +∞).
Suppose that 𝑎 𝑙 < 1 satisfies
(𝑙 + 𝑙 1/3 ) 2 (1 − 𝑎 𝑙2 ) ≥ 𝑙 2 − 1/4. (B.4.2)
Then the Sturm-Picone comparison theorem, see Theorem B.5.1, implies that
between any two zeros of 𝑦 𝑙 there is a zero of 𝐽𝑙 . It is easy to check that 𝑦 𝑙 (𝜌) =
𝜌 −1/2 cos(𝑎 𝑙 𝜌) solves (B.4.1) and has roots at (𝜋/2 + 𝑘 𝜋)/𝑎 𝑙 for 𝑘 ∈ Z. We choose
𝑎 𝑙 = 𝑙 −1/3 . Then (B.4.2) holds for 𝑙 large enough. Hence 𝐽𝑙 has a root on the
interval [𝑙 + 𝑙 1/3 , 𝑙 + 𝑙 1/3 + 𝜋𝑙 1/3 ] and 𝑗𝑙 ≤ 𝑙 + (𝜋 + 1)𝑙 1/3 .
B.5 Appendix: Comparison Theorems for Sturm-Liouville

Equations
Classical Sturm-Liouville theory is concerned with second order differential equa-
tions on the form
( 𝑝(𝑥)𝑦 0 (𝑥)) 0 + 𝑞(𝑥)𝑦(𝑥) = 0 on [𝑎, 𝑏].
Special cases of Sturm-Liouville equations are the radial solutions to the Helmholtz
equation, see (B.2.3). To estimate solutions to these radial equations in Section
B.2.2, we compare them to some more simple Sturm-Liouville equations. To do
this we use the following classical theorems:
88
B.5. Appendix: Comparison Theorems for Sturm-Liouville Equations
Theorem B.5.1 (Sturm-Picone Comparison Theorem, [Hin05, Theorem B]). Let

𝑦 1 and 𝑦 2 be non-zero solutions to
( 𝑝 1 (𝑥)𝑦 10 (𝑥)) 0 + 𝑞 1 (𝑥)𝑦 1 (𝑥) = 0,

( 𝑝 2 (𝑥)𝑦 20 (𝑥)) 0 + 𝑞 2 (𝑥)𝑦 2 (𝑥) = 0,
on the interval [𝑎, 𝑏]. Assume that 0 < 𝑝 2 ≤ 𝑝 1 and 𝑞 1 ≤ 𝑞 2 , and let 𝑧 1 and 𝑧 2 be
two consecutive zeros of 𝑦 1 . Then either 𝑦 2 has a zero in the interval (𝑧 1 , 𝑧2 ), or
𝑦1 = 𝑦2.
Theorem B.5.2 (Sturm Comparison Theorem, [Dur12, Chapter 13.7]). Let 𝑦 1 > 0
on (𝑎, 𝑐) and 𝑦 2 be non-zero solutions to
( 𝑝(𝑥)𝑦 10 (𝑥)) 0 + 𝑞 1 (𝑥)𝑦 1 (𝑥) = 0,

( 𝑝(𝑥)𝑦 20 (𝑥)) 0 + 𝑞 2 (𝑥)𝑦 2 (𝑥) = 0,
on the interval (𝑎, 𝑐) ⊂ [𝑎, 𝑏]. Assume that 𝑝 > 0 and 𝑞 2 ≤ 𝑞 1 on [𝑎, 𝑏].
Furthermore, assume that
𝑦 1 (𝑎) = 𝑦 2 (𝑎) ≥ 0 and 𝑦 10 (𝑎) = 𝑦 20 (𝑎) ≥ 0.
Then 𝑦 1 (𝑥) < 𝑦 2 (𝑥) for all 𝑥 ∈ (𝑎, 𝑐).
It is also important for us to estimate the maximum of some solutions to the

Helmholtz equation. To limit the search, we use the following theorem:
Theorem B.5.3 (Sonin-Pólya Oscillation Theorem, [Dur12, Chapter 13.7]). Let 𝑦

be a solution of the differential equation
( 𝑝(𝑥)𝑦 0 (𝑥)) 0 + 𝑞(𝑥)𝑦(𝑥) = 0,
where 𝑝 and 𝑞 are continuously differentiable functions on [𝑎, 𝑏]. Suppose that
𝑝 > 0, 𝑞 ≠ 0 and ( 𝑝𝑞) 0 > 0 on (𝑎, 𝑏). Then the successive local maximums of
|𝑦(𝑥)| form a decreasing sequence.
89
Paper C
Kuttler-Sigillito’s Inequalities and

Rellich-Christianson Identity
Stine Marie Berge
Paper C
Kuttler-Sigillito’s Inequalities
and Rellich-Christianson Identity
Abstract
We show Kuttler-Sigillito’s inequalities for the mixed Neumann-Dirichlet,
mixed Steklov-Dirichlet, and mixed Robin-Dirichlet eigenvalue problems.
Additionally, we show a Rellich identity for the mixed Neumann-Dirichlet
problem. Finally, we show a Rellich-Christianson identity for the Neumann-
Dirichlet problem.
C.1 Introduction
In this paper we will study various eigenvalue problems with mixed conditions on
the boundary. The first problem of interest is the eigenvalue problem for the Laplace
operator on a bounded domain Ω with mixed Neumann-Dirichlet conditions on the
boundary 𝜕Ω. This problem has for example been studied in [LR17; Jak+06].
In the first part of the article we are going to compare these eigenvalues with
eigenvalues of the mixed Steklov-Dirichlet problem. More precisely, we consider
the following problem Δ𝑢 = 0 on Ω with boundary conditions 𝑢 𝑛 = 𝜎𝑢 on part of
the boundary 𝐹 ⊂ 𝜕Ω and Dirichlet conditions 𝑢 = 0 on 𝜕Ω \ 𝐹. This problem
has been studied in e.g. [Seo21; Bañ+10]. When the Dirichlet conditions are not
present the problem is known as the Steklov problem, for historical context see
[Kuz+14].
Kuttler and Sigillito showed in [KS68] several bounds for eigenvalue problems
in R2 . One of their results is as follows: Let Ω ⊂ R2 be a bounded star convex
domain with 𝐶 2 -boundary. Denote by 𝜇2 the first non-zero eigenvalue of the
Laplace eigenvalue problem with Neumann boundary conditions and let 𝜎2 be the
first non-zero eigenvalue of the Steklov problem Δ𝑢 = 0 on Ω and 𝜎2 𝑢 = 𝑢 𝑛 on
93
Paper C. Kuttler-Sigillito’s Inequalities and Rellich-Christianson Identity
𝜕Ω. Then
ℎmin 𝜇2
𝜎2 ≥ √ , (C.1.1)
2𝑟 max 𝜇2 + 2
where ℎmin and 𝑟 max are geometric constants depending explicitly on Ω. In [HS20]
Hassannezhad and Siffert generalized (C.1.1) to Riemannian manifolds of arbitrary
dimension.
One of the tools used in [KS68] and later in [HS20] is the Rellich identity
which expresses the eigenvalue in terms of the eigenfunction and its gradient on
the boundary. The simplest example is the Laplace Dirichlet eigenvalue problem
Δ𝑢 + 𝜆𝑢 = 0 on the 𝐶 2 -domain Ω with 𝑢 = 0 on the boundary 𝜕Ω. In this case 𝜆
can be expressed as ∫
1
𝜆= 𝑥 · n 𝑢 2𝑛 dS, (C.1.2)
2 𝜕Ω
where dS is the surface measure on the boundary with normal n. Rellich proved
(C.1.2) in [Rel40]. In 2007 similar Rellich identities were shown to hold for
other eigenvalue problems, e.g. the Neumann, clamped plate and buckling problem
[Liu07]. The Rellich identities in [Rel40; Liu07] were also recently extended to
eigenvalue problems on Riemannian manifolds [HS20].
For Dirichlet eigenfunctions on the triangle one can get an interesting refor-
mulation of the Rellich identity as shown in [Chr19]. It was further extended to
polytopes in [Mét18], which referred to the identity as Rellich-Christianson type
identity. For a regular polytope 𝑃 this result gives
∫
𝐶
𝜆= 𝑢 2 dS,
2 𝜕𝑃 𝑛
where 𝐶 is the distance from the center to the faces.

The first goal of this article is to generalize (C.1.1) to all eigenvalues. In
particular for a star-shaped domain Ω ⊂ R2 this means that if 𝜎𝑘 and 𝜇 𝑘 are the
𝑘’th eigenvalues to the Steklov problem and the Neumann problem, respectively
we have
ℎmin 𝜇 𝑘
𝜎𝑘 ≥ √ .
2𝑟 max 𝜇 𝑘 + 2
More precisely, we prove an inequality similar to (C.1.1) for mixed Neumann-
Dirichlet problem and Steklov-Dirichlet problem on sub-domains in Riemannian
manifolds of arbitrary dimension, generalizing the result of [HS20]. Moreover,
we will show an inequality for the first mixed Steklov-Dirichlet eigenvalue, the
mixed Neumann-Dirichlet, and the mixed Robin-Dirichlet eigenvalues. Another
goal of the article is to compute the Rellich-Christianson identity for the mixed
Neumann-Dirichlet problem.
94
C.2. Kuttler-Sigillito’s Type Inequalities
An outline of the article is as follows: In Section C.2 we will give the Kuttler-
Sigillito’s type inequalities. The inequality will be obtained on Riemannian man-
ifolds satisfying a curvature bound. In Section C.3 we will discuss the Hadamard
formula and prove it in the case of the Laplace eigenvalue problem with mixed
Neumann-Dirichlet boundary conditions. Section C.4 will be devoted to proving
the Rellich identities using Hadamard formulas for the mixed Neumann-Dirichlet
problem. In the last two sections we will prove the Rellich-Christianson identity
on polytopes for the mixed Neumann-Dirichlet problem.
C.2 Kuttler-Sigillito’s Type Inequalities

Let (𝑀, g) be a smooth 𝑛 dimensional Riemannian manifold. Denote by Ric the
Ricci curvature of 𝑀, and assume that (𝑀, g) satisfies the curvature bound
(𝑛 − 1)𝜅|𝑋 | 2 ≤ Ric(𝑋, 𝑋), (C.2.1)
where |𝑋 | 2 = g(𝑋, 𝑋) for a vector field 𝑋 and 𝜅 ∈ R. The Laplacian with respect
to g of a twice differentiable function 𝜙 is defined by
Δ𝜙 = div(grad 𝜙),
where div and grad denotes the divergence and gradient with respect to g.
In this text we will let 𝑟 (𝑥) denote the radial distance function from a chosen
point 𝑝 ∈ 𝑀. We will always work with domains which are contained inside a
geodesic ball centered at 𝑝 with radius less than the injectivity radius Inj( 𝑝). Some
properties of the radial distance function are that | grad 𝑟 | = 1 and 𝑟 2 is a smooth
function on 𝐵Inj( 𝑝) ( 𝑝). For manifolds satisfying (C.2.1) we have the following
volume comparison result:
Lemma C.2.1 (Volume Comparison [Pet16, Lemma 7.1.2]). Assume that (𝑀, g)
is an 𝑛-dimensional Riemannian manifold satisfying (C.2.1), and denote by
√ √


 coth(𝑟 −𝜅)/ −𝜅, 𝜅<0


cot 𝜅 (𝑟) = 1/𝑟, 𝜅=0 .
 cot(𝑟 √𝜅)/√𝜅,

𝜅>0


Then for all 𝑥 ∈ 𝑀 such that 𝑟 (𝑥) < Inj( 𝑝) we have
Δ𝑟 ≤ (𝑛 − 1) cot 𝜅 (𝑟).
95
Denote by vol the volume density and let Ω ⊂ 𝐵Inj( 𝑝) ( 𝑝) be a domain with
𝐶 2 -boundary. If 𝑢, 𝑣 ∈ 𝐻 1 (Ω) then by the trace theorem 𝑣 ∈ 𝐻 1/2 (𝜕Ω) and
𝑢𝑛 ∈ 𝐻 −1/2 (𝜕Ω). If additionally, we have that Δ𝑢 ∈ 𝐿 2 (Ω) then
∫ ∫
grad 𝑢 · grad 𝑣 dvol = − 𝑣Δ𝑢 dvol +h𝑢 𝑛 , 𝑣i𝜕Ω , (C.2.2)
Ω Ω
where h𝑢 𝑛 , 𝑣i𝜕Ω is the paring between 𝐻 −1/2 (𝜕Ω) and 𝐻 1/2 (𝜕Ω), see [McL00,
Chap. 4].
Remark C.2.2. In the case when Ω ⊂ R𝑛 one can relax the smoothness condition
of the boundary to be just Lipschitz, see [McL00, Chap. 4].
C.2.1 Steklov-Dirichlet and Neumann-Dirichlet Comparison

Let Ω ⊂⊂ 𝐵Inj( 𝑝) ( 𝑝) with 𝐶 2 -boundary and assume 𝐹 is a finite union of connected
open subset of 𝜕Ω where 𝜕𝐹 is Lipschitz continuous. Denote by
𝐻01 (Ω, 𝜕Ω \ 𝐹) = {𝑢 ∈ 𝐻 1 (Ω) : 𝑢 = 0 on 𝜕Ω \ 𝐹},
where the boundary values are well defined by using the trace theorem. Con-
sider weak solutions in 𝐻 1 (Ω), see [McL00, Chap. 4] to the following eigenvalue
problems: Steklov-Dirichlet problem


 Δ𝑢(𝑥) = 0 for 𝑥 ∈ Ω


𝑢 𝑛 (𝑥) = 𝜎 𝐹 𝑢(𝑥) for 𝑥 ∈ 𝐹 (C.2.3)

 𝑢(𝑥) = 0 for 𝑥 ∈ 𝜕Ω \ 𝐹,


and Neumann-Dirichlet problem


 Δ𝑣(𝑥) + 𝜇 𝐹 𝑣(𝑥) = 0 for 𝑥 ∈ Ω


𝑣 𝑛 (𝑥) = 0 for 𝑥 ∈ 𝐹 (C.2.4)

 𝑣(𝑥) = 0 for 𝑥 ∈ 𝜕Ω \ 𝐹.


Notice that in the case 𝐹 = 𝜕Ω we get the Steklov problem and the Neumann
problem, respectively. Additionally, in the case that 𝐹 = ∅ the Steklov-Dirichlet
problem have only the trivial solution 𝑢 = 0, and the mixed Neumann-Dirichlet
problem is the Dirichlet problem.
From [Agr06] it follows that
0 ≤ 𝜎1𝐹 ≤ · · · ≤ 𝜎𝑛𝐹 → ∞.
For Neumann-Dirichlet eigenvalues we also have that
0 ≤ 𝜇1𝐹 ≤ · · · ≤ 𝜇 𝑛𝐹 → ∞.
96
This can be shown by using the min-max theorem [Bor20, Theorem 5.15] together
with a standard argument showing that (−Δ + 1) −1 is a compact operator on 𝐿 2 (Ω).
The Rayleigh quotients for the 𝑘’th eigenvalues are given by
∫
𝐹 Ω
| grad 𝑢| 2 dvol
𝜎𝑘 = inf sup ∫
𝑉𝑘 ⊂𝐻01 (Ω,𝜕Ω\𝐹 ) 𝑢 ∈𝑉𝑘 \{0}
𝐹
𝑢 2 dS
dim(𝑉𝑘 )=𝑘
and ∫
Ω
| grad 𝑣| 2 dvol
𝜇 𝐹𝑘 = inf sup ∫ .
𝑉𝑘 ⊂𝐻01 (Ω,𝜕Ω\𝐹 ) 𝑣 ∈𝑉𝑘 \{0}
Ω
𝑣 2 dvol
dim(𝑉𝑘 )=𝑘
The Rayleigh quotient for the mixed Steklov-Dirichlet can be found in [Agr06],
while for the Neumann-Dirichlet problem the Rayleigh quotient follows from the
min-max theorem [Bor20, Theorem 5.15]. Notice that the 𝑘’th Neumann-Dirichlet
eigenvalue is bounded below by the 𝑘’th Neumann eigenvalue, and above by the
𝑘’th Dirichlet eigenvalue by using the Rayleigh quotient together with
𝐻 1 (Ω, 𝜕Ω) ⊂ 𝐻 1 (Ω, 𝜕Ω \ 𝐹) ⊂ 𝐻 1 (Ω).
Our first main result is the following generalization of (C.1.1).
Theorem C.2.3. Assume that (𝑀, g) satisfies (C.2.1) and let Ω ⊂ 𝑀 be a bounded
𝐶 2 domain contained inside 𝐵Inj( 𝑝) ( 𝑝) for a fixed point 𝑝 ∈ 𝑀. Denote by 𝑟 the
radial distance function with respect to 𝑝 and let 𝜇 𝑘 and 𝜎𝑘 be as in (C.2.4) and
(C.2.3), respectively. Then
ℎmin 𝜇 𝐹𝑘
𝜎𝑘𝐹 ≥ q , (C.2.5)
2𝑟 max 𝜇 𝐹𝑘 + 𝐶0
where
𝑟 max = sup 𝑟 (𝑥), ℎmin = inf n · 𝑟 grad 𝑟,
𝑥 ∈Ω 𝑥 ∈𝐹
and
𝐶0 = 1 + (𝑛 − 1) sup 𝑟 (𝑥) cot 𝜅 (𝑟 (𝑥)).
𝑥 ∈Ω
Remark C.2.4.
• If 𝜅 ≥ 0 we have that 𝐶0 = 𝑛 since 𝑟 cot 𝜅 (𝑟) is non-increasing in 𝑟. Addi-

tionally, for 𝜅 < 0 the function 𝑟 cot 𝜅 (𝑟) is increasing, hence the maximum
is obtained on the boundary.
97
• Inequality (C.2.5) is non-trivial if and only if ℎmin > 0. In the case that
𝑀 = R𝑛 and 0 ∈ Ω the condition ℎmin > 0 means that 𝑥 cannot be tangent to
the part of the boundary 𝐹. In the case that 𝐹 = 𝜕Ω this condition implies
that Ω is strictly star convex with respect to the point 0. We define a domain
to be strictly star convex with respect to a point 𝑝 if any straight line from 𝑝
to the boundary intersect the boundary only at the end point.
It should also be noted that all convex sets containing zero satisfies ℎmin > 0.
• The Weyl laws on Ω ⊂ R𝑛 in the case of the Steklov problem [GP17,
Sec. 5.2.1] and Neumann problem [Cha84, p. 9] are
1
𝑛−1
𝜎𝑘 1
lim 1
= 2𝜋
𝑘→∞ 𝑘 𝑛−1 𝜔 𝑛−1 vol 𝑛−1 (𝜕Ω)
and
𝑛1
𝜇𝑘 2 2
lim = (2𝜋) , (C.2.6)
𝑘→∞ 𝑘 2/𝑛 𝜔 𝑛 vol 𝑛 (Ω)
where 𝜔 𝑚 denotes the volume of the 𝑚-dimensional unit ball. In this case
1
we have that the right hand-side of (C.2.5) behaves like 𝐶0 (Ω)𝑘 𝑛−1 and the
1
left hand-side behaves like 𝐶1 (Ω)𝑘 𝑛 . This shows that the result is trivially
true when 𝑘 goes to infinity.
Proof. When 𝑘 = 1 and 𝐹 = 𝜕Ω we have that 𝜎1𝐹 = 𝜇1𝐹 = 0 and the inequality
trivially holds. Hence we will assume that either 𝑘 ≥ 2 or 𝐹 ≠ 𝜕Ω.
Let 𝜎𝑚𝐹 be the 𝑚’th eigenvalue of the Steklov-Dirichlet problem with corre-
sponding eigenfunction 𝑢 𝑚 , and denote by
𝑉𝑘 = span{𝑢 1 , . . . , 𝑢 𝑘 } ⊂ 𝐻01 (Ω, 𝜕Ω \ 𝐹).
Then for any 𝑣 ∈ 𝑉𝑘 \ {0} we have that

∫
Ω
| grad 𝑣| 2 dvol
∫ ≤ 𝜎𝑘𝐹 .
𝑣 2 dS
𝐹
Choose 𝑣 ∈ 𝑉𝑘 \ {0} to satisfy

∫ ∫
| grad 𝑣| 2 dvol | grad 𝑢| 2 dvol
Ω Ω
∫ = sup ∫ .
𝑣 2 dvol 𝑢 ∈𝑉 \{0} 𝑢 2 dvol
Ω 𝑘 Ω
By the definition of the eigenvalues of the mixed Neumann-Dirichlet problem we

get that ∫ ∫
𝜇 𝐹𝑘 𝑣 2 dvol ≤ | grad 𝑣| 2 dvol . (C.2.7)
Ω Ω
98
Since we are inside the ball 𝐵Inj( 𝑝) ( 𝑝) we have that 𝑟 2 is a smooth function. Using
the integration by parts (C.2.2) and Lemma C.2.1 we obtain
∫ ∫
2
2 𝑣 n · 𝑟 grad 𝑟 dS = 𝑣 2 n · grad(𝑟 2 ) dS (C.2.8)
𝐹 ∫𝐹 ∫
= grad 𝑣 2 · grad 𝑟 2 dvol + 𝑣 2 Δ(𝑟 2 ) dvol
Ω∫ Ω∫
2
= 2 grad 𝑣 · 𝑟 grad 𝑟 dvol + 𝑣 2 (2 + 2𝑟Δ𝑟) dvol
∫Ω Ω ∫
2
≤ 2 grad 𝑣 · 𝑟 grad 𝑟 dvol +2𝐶0 𝑣 2 dvol .
Ω Ω
The Cauchy-Schwartz inequality together with (C.2.7) gives

∫ ∫
𝑟 max
𝑣 grad 𝑣 · 𝑟 grad 𝑟 dvol ≤ q | grad 𝑣| 2 dvol . (C.2.9)
Ω 𝜇 𝐹𝑘 Ω
Using (C.2.7), (C.2.8), and (C.2.9) gives

∫ ∫
2 𝐹 −1/2
𝑣 n · 𝑟 grad 𝑟 dS ≤ 2𝑟 max (𝜇 𝑘 ) 𝐹
+ 𝐶0 /𝜇 𝑘 | grad 𝑣| 2 dvol .
𝐹 Ω
Simplifying the expression

∫ ∫
ℎmin 𝑣 2 dS ≤ 2𝑟 max (𝜇 𝐹𝑘 ) −1/2 + 𝐶0 /𝜇 𝐹𝑘 𝜎𝑘𝐹 𝑣 2 dS
𝐹 𝐹
gives the result.
In the special case when Ω = 𝐵 𝑅 ( 𝑝) we get the following corollary.

Corollary C.2.5. Let (𝑀, g) be an 𝑛-dimensional Riemannian manifold and let
Ω = 𝐵 𝑅 ( 𝑝) ⊂ 𝑀 where 𝑅 < Inj( 𝑝) and 𝐹 ⊂ 𝜕𝐵 𝑅 ( 𝑝). Then
𝑅𝜇 𝐹𝑘
𝜎𝑘𝐹 ≥ q ,
2𝑅 𝜇 𝐹𝑘 + 𝐶0
where (
𝑛 𝜅≥0
𝐶0 = .
1 + (𝑛 − 1)𝑅 cot 𝜅 (𝑅) 𝜅<0
By using Remark C.2.2 we have that for bounded domains in R𝑛 with Lipschitz
boundary the proof of Theorem C.2.3 still holds. Hence we can explore the
following examples.
99
Example C.2.6. Let 𝐽𝑙 (𝑟) denote the 𝑙’th Bessel function for 𝑙 ∈ N \ {0}, which
satisfies the differential equation
𝑟 2 𝐽𝑙00 (𝑟) + 𝑟 𝐽𝑙0 (𝑟) + (𝑟 2 − 𝑙 2 )𝐽𝑙 (𝑟) = 0.
Consider the half-ball
𝐷 + = {(𝑥, 𝑦) ∈ R2 : 𝑥 2 + 𝑦 2 < 1, 𝑦 > 0}.
On this domain we will consider the problems


 Δ𝑢(𝑥, 𝑦) = 0 on 𝐷 +


𝑢 𝑛 (𝑥, 𝑦) = 𝜎𝑢 on 𝐹 = {(𝑥, 𝑦) ∈ R2 : 𝑥 2 + 𝑦 2 = 1, 𝑦 > 0}

 𝑢(𝑥, 𝑦) = 0 on {(𝑥, 0) ∈ R2 : |𝑥| ≤ 1},


and


 Δ𝑣(𝑥, 𝑦) + 𝜇𝑣(𝑥, 𝑦) = 0 on 𝐷 +


𝑣 𝑛 (𝑥, 𝑦) = 0 on 𝐹 = {(𝑥, 𝑦) ∈ R2 : 𝑥 2 + 𝑦 2 = 1, 𝑦 > 0}

 𝑣(𝑥, 𝑦) = 0 on {(𝑥, 0) ∈ R2 : |𝑥| ≤ 1}.


The solutions to the Neumann-Dirichlet problem are given by

0
𝑣(𝑟, 𝜃) = 𝐽𝑙 ( 𝑗𝑙,𝑚 𝑟) sin(𝑙𝜃) 𝑙 ∈ N,
0
where 0 < 𝑗𝑙,𝑚 is the 𝑚’th zero of 𝐽𝑙0 and 𝜇 = ( 𝑗𝑙,𝑚
0 ) 2 . The Steklov-Dirichlet
problem have the solutions
𝑢(𝑟, 𝜃) = 𝑟 𝑘 sin(𝑘𝜃) 𝑘∈N
0
with corresponding eigenvalues 𝜎𝑘 = 𝑘 > 0. If we order the zeroes 𝑗𝑙,𝑚 we have
that
𝑗𝑙02𝑘 ,𝑚𝑘 𝜇𝑘
𝜎𝑘 = 𝑘 ≥ = √ .
2( 𝑗𝑙0𝑘 ,𝑚𝑘 + 1) 2 𝜇 𝑘 + 2
We can extend the Neumann-Dirichlet eigenfunctions above to Neumann eigen-

function on the ball by using a reflection. Recall that for the Neumann problem all
the eigenspaces corresponding to non-constant eigenfunctions have dimension 2.
The solutions given by the extension from the Neumann-Dirichlet problem gives
us one of the two linearly independent solutions of the eigenspace of the Neumann
problem. The Weyl law presented in (C.2.6) shows that 𝜇 𝑘 𝑘.
100
Example C.2.7. Let Ω = [0, 1] × [0, 1] and let 𝐹 = {0, 1} × (0, 1). We pose the
problem


 Δ𝑣(𝑥, 𝑦) + 𝜇𝑣(𝑥, 𝑦) = 0 on Ω


𝑣 𝑛 (𝑥, 𝑦) = 0 on 𝐹 (C.2.10)

 𝑣(𝑥, 𝑦) = 0 on [0, 1] × {0, 1}.


Then we have the solutions
𝑣(𝑥, 𝑦) = cos (𝜋𝑚𝑥) sin (𝜋𝑛𝑦)
with corresponding eigenvalues 𝜇 = (𝜋𝑚) 2 + (𝜋𝑛) 2 , where 𝑚 ∈ N ∪ {0} and

𝑛 ∈ N \ {0}.
The Steklov-Dirichlet problem on the same domain is given by


 Δ𝑢(𝑥, 𝑦) = 0 on Ω


𝑢 𝑛 (𝑥, 𝑦) = 𝜎𝑢(𝑥, 𝑦) on 𝐹

 𝑢(𝑥, 𝑦) = 0 on [0, 1] × {0, 1}.


Computing the eigenvalues gives
𝜋𝑛
! ±1
cosh 2
𝜎 = 𝜋𝑛 𝜋𝑛

sinh 2
with corresponding eigenfunctions

!
𝜎
𝑢(𝑥, 𝑦) = cosh (𝜋𝑛𝑥) − sinh (𝜋𝑛𝑥) sin (𝜋𝑛𝑦) ,
𝑛𝜋
for all 𝑛 ∈ N \ {0}. Using that

𝜋(𝑛 − 1) 𝜋𝑛
𝜋(𝑛 − 1) coth < 𝜋𝑛 tanh
2 2
for all 𝑛 ∈ N \ {0, 1}, we can find the order of the eigenvalues. Hence we get

𝜋𝑘
𝜎2𝑘 = 𝜋𝑘 coth
2
!
𝜋𝑘
𝑢 2𝑘 (𝑥, 𝑦) = cosh (𝜋𝑘𝑥) − coth sinh (𝜋𝑘𝑥) sin (𝜋𝑘 𝑦)
2
101
and
𝜋𝑘
𝜎2𝑘−1 = 𝜋𝑘 tanh
2
!
𝜋𝑘
𝑢 2𝑘−1 (𝑥, 𝑦) = cosh (𝜋𝑘𝑥) − tanh sinh (𝜋𝑘𝑥) sin (𝜋𝑘 𝑦) .
2
In this case we have that 𝑟 max = √1 , ℎmin = 1/2, and 𝐶0 = 2. Using Theorem C.2.3
2
we obtain
𝜋 2 (𝑚 2 + 𝑛22𝑘 )

𝜋𝑘
𝜎2𝑘 = 𝜋𝑘 coth ≥ √ q 2𝑘 .
2 2 + 𝑛2 + 4
2 2𝜋 𝑚 2𝑘 2𝑘
and
𝜋 2 (𝑚 2 + 𝑛22𝑘−1 )

𝜋𝑘
𝜎2𝑘−1 = 𝜋𝑘 tanh ≥ √ q 2𝑘−1 ,
2 2
2 2𝜋 𝑚 2𝑘−1 + 𝑛22𝑘−1 + 4
where 𝜇 𝑗 = 𝜋 2 (𝑚 2𝑗 + 𝑛2𝑗 ).
C.2.2 Robin-Dirichlet Comparison

In this section we will assume that 𝜕Ω \ 𝐹 has non-empty interior in 𝜕Ω. The
Robin-Dirichlet problem for any 𝛼 > 0 is defined by


 Δ𝑤(𝑥) + 𝜆 𝛼 𝑤(𝑥) = 0 for 𝑥 ∈ Ω


𝑤 𝑛 (𝑥) + 𝛼𝑤 = 0 on 𝐹

 𝑤(𝑥) = 0 on 𝜕Ω \ 𝐹.


Note that we have suppressed the domain 𝐹 in the Robin-Dirichlet eigenvalues 𝜆 𝛼
for readability. By the min-max theorem [Bor20, Theorem 5.15] we have that the
Rayleigh quotient is given by
∫ ∫
𝛼 Ω
| grad 𝑤| 2 dvol +𝛼 𝐹 𝑤 2 dS
𝜆𝑘 = inf sup ∫ .
𝑉𝑘 ⊂𝐻01 (Ω,𝜕Ω\𝐹 ) 𝑤 ∈𝑉𝑘 \{0} 𝑤 2 dvol
Ω
dim(𝑉𝑘 )=𝑘
The eigenvalue of the Robin-Dirichlet problem is always between the Dirichlet

eigenvalues and the Neumann-Dirichlet eigenvalues with the Dirichlet eigenvalue
being the limit as 𝛼 → ∞. Again we have only point spectrum since (−Δ + 1) −1 is
a compact operator on 𝐿 2 (Ω).
102
Theorem C.2.8. Assume that 𝐹¯ ≠ 𝜕Ω, then eigenvalues satisfy
𝜆 𝑘𝛼 − 𝜇 𝐹𝑘 𝜇 𝐹𝑘
≤ .
𝛼 𝜎1𝐹
Proof. For all 𝑢 ∈ 𝐻01 (Ω, 𝜕Ω \ 𝐹) we have that

∫ ∫
𝜎1𝐹 𝑢 2 dS ≤ | grad 𝑢| 2 dvol .
𝐹 Ω
Denote by 𝑣 𝑚 the eigenfunctions counting multiplicity of the Neumann-Dirichlet

problem with eigenvalue 𝜇 𝑚 . Let 𝑉𝑘 consists of the span of 𝑣 1 , . . . , 𝑣 𝑘 . By the
Rayleigh quotient the eigenvalue 𝜆 𝑘𝛼 satisfies
∫ ∫
| grad 2 dvol +𝛼 𝑣 2 dS
Ω
𝑣| 𝜇𝐹
𝜆 𝑘𝛼 ≤ sup ∫ 𝐹
≤ 𝜇 𝐹𝑘 + 𝛼 𝑘𝐹 .
𝑣 ∈𝑉𝑘 \{0} Ω
𝑣 2 dvol 𝜎1
Remark C.2.9. Taking the limit when 𝛼 → 0 and assuming that 𝜆 𝑘𝛼 is differentiable
at 0, we have that
𝑑𝜆 𝑘𝛼 𝜇 𝐹𝑘
≤ .
𝑑𝛼 𝛼=0 𝜎1𝐹
Example C.2.10. In this example we will continue using some of the notation from
Example C.2.6. Consider the equation
𝛼𝐽𝑙 (𝑟) + 𝑟 𝐽𝑙0 (𝑟) = 0.
Denote by 𝑗𝑙,𝑚
𝛼 the 𝑚’th zero of the equation. Posing the problem


 Δ𝑤(𝑥, 𝑦) + 𝜆 𝛼 𝑤(𝑥, 𝑦) = 0 for (𝑥, 𝑦) ∈ 𝐷 +


𝑤 𝑛 (𝑥, 𝑦) + 𝛼𝑤(𝑥, 𝑦) = 0 on 𝐹 = {(𝑥, 𝑦) ∈ R2 : 𝑥 2 + 𝑦 2 = 1, 𝑦 > 0}

 𝑤(𝑥, 𝑦) = 0 on {(𝑥, 0) ∈ R2 : |𝑥| ≤ 1},


𝛼 𝑟) sin(𝑙𝜃), with eigenvalue 𝜆 𝛼 = ( 𝑗 𝛼 ) 2 .
we get the solution 𝑤(𝑟, 𝜃) = 𝐽𝑙 ( 𝑗𝑙,𝑚 𝑙,𝑚
Using Theorem C.2.8 we obtain
( 𝑗𝑙𝛼𝑘 ,𝑚𝑘 ) 2 − ( 𝑗𝑙0𝑘 ,𝑚𝑘 ) 2

≤ ( 𝑗𝑙0𝑘 ,𝑚𝑘 ) 2 = 𝜇 𝑘 .
𝛼
Taking the derivative of the equation
𝛼
𝛼𝐽𝑙 ( 𝑗𝑙,𝑚 𝛼
) + 𝑗𝑙,𝑚 𝐽𝑙0 ( 𝑗𝑙,𝑚
𝛼
) = 0,
103
with respect to 𝛼 gives

𝛼
𝑑 𝑗𝑙,𝑚 𝛼
𝑑 𝑗𝑙,𝑚
𝛼
𝐽𝑙 ( 𝑗𝑙,𝑚 ) +𝛼 𝐽𝑙0 ( 𝑗𝑙,𝑚
𝛼
) + 𝐽𝑙0 ( 𝑗𝑙,𝑚
𝛼
) + 𝛼
𝑗𝑙,𝑚 𝐽𝑙00 ( 𝑗𝑙,𝑚
𝛼
) = 0.
𝑑𝛼 𝑑𝛼 𝑑𝛼
Taking the limit when 𝛼 → 0 shows that
𝛼
𝑑 𝑗𝑙,𝑚 0 )2
𝑙 2 − ( 𝑗𝑙,𝑚
0 0
𝐽𝑙 ( 𝑗𝑙,𝑚 )+ 𝑗𝑙,𝑚 𝐽𝑙00 ( 𝑗𝑙,𝑚
0 0
) = 𝐽𝑙 ( 𝑗𝑙,𝑚 0
)+ 𝑗𝑙,𝑚 0 )2
0
𝐽𝑙 ( 𝑗𝑙,𝑚 ) = 0,
𝑑𝛼 𝑑𝛼 ( 𝑗𝑙,𝑚
𝛼=0 𝛼=0
where we have used that 𝐽𝑙0 ( 𝑗𝑙,𝑚

0 ) = 0. Simplifying the derivative we get
𝑑𝜆 𝛼 𝑑 ( 𝑗𝑙𝛼𝑘 ,𝑚𝑘 ) 2
=
𝑑𝛼 𝛼=0 𝑑𝛼
𝛼=0
𝑑 𝑗𝑙𝛼𝑘 ,𝑚𝑘
= 2 𝑗𝑙0𝑘 ,𝑚𝑘
𝑑𝛼
𝛼=0
( 𝑗𝑙0 ,𝑚 ) 2
= 2 0 𝑘 2𝑘 2 .
( 𝑗𝑙𝑘 ,𝑚𝑘 ) − 𝑙 𝑘
Using Remark C.2.9 and Example C.2.6 we get
𝑑𝜆 𝛼 ( 𝑗𝑙0𝑘 ,𝑚𝑘 ) 2 𝜇𝑘
=2 ≤ = ( 𝑗𝑙0𝑘 ,𝑚𝑘 ) 2 . (C.2.11)
𝑑𝛼 𝛼=0 ( 𝑗𝑙0𝑘 ,𝑚𝑘 ) 2 − 𝑙 2𝑘 𝜎1
√
Simplifying this expression we get 𝑗𝑙,𝑚0 ≥ 𝑙 2 + 2. This inequality follows from
p
0
the stronger inequality 𝑗𝑙,𝑚 0 >
> 𝑗𝑙,1 𝑙 (𝑙 + 2), which can be found in [Wat95,
0 ≤ 𝑙 + 𝐶𝑙 1/3 , see e.g. [EL97], we
p. 486]. In particular, since we also have that 𝑗𝑙,1
get
0
𝑗𝑙,1
lim √ = 1.
𝑙→∞ 𝑙 2 + 2
This makes the inequality sharp in the limit.
C.3 Hadamard Formula for the Mixed Neumann-Dirichlet

Problem
Hadamard formulas explain how the eigenvalues change when the domain is de-
formed as an integral of the eigenfunctions. In this and the following sections we
will only be working in R𝑛 , however, there exists a Hadamard formula on Rieman-
nian manifolds for the Dirichlet problem see e.g. [EI07]. Let 𝑉 be a (possibly time
104
C.3. Hadamard Formula for the Mixed Neumann-Dirichlet Problem
dependent) vector field and let Ω𝑡 be the deformation of the domain Ω0 with respect
to the corresponding one-parameter family of diffeomorphisms 𝜙 with speed 𝑉.
Assume that Ω𝑡 is pre-compact with smooth boundary 𝜕Ω𝑡 , and 𝐹𝑡 = 𝜙𝑡 (𝐹). We
will use the superscript 𝑡 to denote the time-dependence.
Let 𝑣 𝑡 ∈ 𝐻 2 (Ω)∩𝐶 1 (Ω) be a solution to the mixed Neumann-Dirichlet problem


 Δ𝑣 𝑡 (𝑥) + 𝜇𝑡 𝑣 𝑡 (𝑥) = 0 in Ω𝑡


𝑣 𝑡𝑛 (𝑥) = 0 on 𝐹𝑡

 𝑣 𝑡 (𝑥) = 0

on 𝜕Ω𝑡 \ 𝐹𝑡 .

We will prove the Hadamard formula

∫ ∫
𝑡 2 𝑡 𝑡 2
𝑡
𝜕𝑡 𝜇 = | grad𝜕Ω𝑡 𝑣 | − 𝜇 (𝑣 ) dS − n · 𝑉 (𝑣 𝑡𝑛 ) 2 dS . (C.3.1)
𝐹𝑡 𝜕Ω𝑡 \𝐹𝑡
Formula (C.3.1) is written down in [Gri10, Eq. (11)]. We will assume the eigen-
𝑡
values are simple and that 𝜇𝑡 and 𝑣 𝑡 is differentiable in 𝑡, where 𝑑𝑣 1
𝑑𝑡 ∈ 𝐶 (Ω𝑡 ).
For the case of non-simple eigenvalues see [Gri10]. When taking the derivative of
integrals it will be useful to have the following lemma in mind:
Lemma C.3.1 ([Fra12, p. 138]). Let 𝑉 be a possibly time dependent vector field,
and let Ω𝑡 be the variation of Ω0 with respect to 𝑉. Assume furthermore that
Ω𝑡 is pre-compact with smooth boundary 𝜕Ω𝑡 . Then given a smooth function
ℎ𝑡 ∈ 𝐶 1 (Ω𝑡 ) which is differentiable in 𝑡 we have that
∫ ∫ ∫
𝑑 𝑡 𝑡
ℎ dvol = 𝜕𝑡 ℎ dvol + n · 𝑉 ℎ𝑡 dS .
𝑑𝑡 Ω𝑡 Ω𝑡 𝜕Ω𝑡
Proof of (C.3.1). Note that since

∫
(𝑣 𝑡 ) 2 vol = 1
Ω𝑡
with the boundary conditions we get that

∫ ∫
2 𝑡 𝑡
𝑣 𝜕𝑡 𝑣 dvol + 𝑉 · n (𝑣 𝑡 ) 2 dS = 0.
Ω𝑡 𝐹𝑡
The (normalized) Rayleigh quotient gives us that

∫
𝜇𝑡 = | grad 𝑣 𝑡 | 2 dvol .
Ω𝑡
105
Taking the derivative of 𝜇𝑡 with respect to 𝑡 and using the divergence theorem gives
∫ ∫
𝑡
𝜕𝑡 𝜇 = 2 𝑡 𝑡
grad 𝜕𝑡 𝑣 · grad 𝑣 dvol + n · 𝑉 | grad 𝑣 𝑡 | 2 dvol
∫Ω𝑡 ∫ 𝜕Ω𝑡
∫
=2 𝑡
𝑣 𝑛 𝜕𝑡 𝑣 𝑡 dS −2 𝑡 𝑡
(Δ𝑣 )𝜕𝑡 𝑣 dvol + n · 𝑉 | grad 𝑣 𝑡 | 2 dS
𝜕Ω𝑡 \𝐹𝑡 Ω𝑡 𝜕Ω𝑡
∫ ∫ ∫
=2 𝑣 𝑡𝑛 𝜕𝑡 𝑣 𝑡 dS −𝜇𝑡 𝑉 · n(𝑣 𝑡 ) 2 dS + n · 𝑉 | grad 𝑣 𝑡 | 2 dS .
𝜕Ω𝑡 \𝐹𝑡 𝐹𝑡 𝜕Ω𝑡
For 𝑥 ∈ 𝜕Ω𝑡 \ 𝐹𝑡 using the chain rule gives
𝑑 𝑡 𝑑𝑣 𝑡 𝑑𝑣 𝑡
0= (𝑣 (𝜙𝑡 (𝑥))) = (𝜙𝑡 (𝑥)) + 𝑉 · grad 𝑣 𝑡 = (𝜙𝑡 (𝑥)) + 𝑉 · n𝑣 𝑡𝑛 ,
𝑑𝑡 𝑑𝑡 𝑑𝑡
where we have used that gradient in the tangent direction to 𝜕Ω𝑡 \ 𝐹𝑡 of 𝑣 is 0. This
gives that
∫ ∫ ∫
𝜕𝑡 𝜇𝑡 = −2 𝑉 · n (𝑣 𝑡𝑛 ) 2 dS −𝜇𝑡 𝑉 · n(𝑣 𝑡 ) 2 dS + n · 𝑉 | grad 𝑣 𝑡 | 2 dS
𝜕Ω \𝐹 𝐹𝑡 𝜕Ω𝑡
∫ 𝑡 𝑡 ∫
𝑡 2 𝑡 𝑡 2 𝑡 2
= | grad𝜕Ω𝑡 𝑣 | − 𝜇 (𝑣 ) dS − n · 𝑉 (𝑣 𝑛 ) dS .
𝐹𝑡 𝜕Ω𝑡 \𝐹𝑡
C.4 Rellich Identity

Applying the Hadamard formulas in the case that 𝑉 is the vector field 𝑥 ↦→ 𝑥 we can
obtain Rellich type identities. The time independent version of the problems in the
previous section are given as follows: Denote by Ω ⊂ R𝑛 a domain with piece-wise
smooth boundary 𝜕Ω. We will assume that the solutions of the mixed Neumann-
Dirichlet are normalized in 𝐿 2 (Ω). We are going to show that the eigenvalue 𝜇 𝐹
with solution 𝑣 ∈ 𝐻 2 (Ω) ∩𝐶 1 (Ω) of the mixed Neumann-Dirichlet problem (C.2.4)
satisfies
∫ ∫ ∫
2 2
𝜇 𝐹
n · 𝑥 𝑣 dS −2 = n · 𝑥 | grad𝜕Ω 𝑣| dS − n · 𝑥 𝑣 2𝑛 dS .
𝐹 𝐹 𝜕Ω\𝐹
Remark C.4.1. In [HS20, Theorem 1.5] the authors show the Rellich identity by
using the divergence theorem. This means that when we have solutions on a
Lipschitz domain and a solution in 𝑣 ∈ 𝐻 2 (Ω), where | grad 𝑣| 2 on the boundary is
interpreted by using the trace theorem.
Denote by Ω ⊂ R𝑛 a smooth pre-compact domain. The idea of going from
the Hadamard formula to the corresponding Rellich type identity is to evolve the
domain by using the vector field 𝑥 ↦→ 𝑥, which by abuse of notation we are going
106
C.4. Rellich Identity
denote to 𝑥. We will denote the perturbation of the domain by Ω𝑡 , which is explicitly

given by 𝑒 𝑡 Ω. Since the domain does not change shape, only size, we can find
out how the eigenvalues explicitly change. Hence setting 𝑡 = 0 in the Hadamard
formula will give us an expression for the eigenvalues using the boundary data.
Let 𝑣 be a solution to


 Δ𝑣(𝑥) + 𝜇 𝐹 𝑣(𝑥) in Ω


𝑣 𝑛 (𝑥) = 0 on 𝐹 .

 𝑣(𝑥) = 0 on 𝜕Ω \ 𝐹


Then we have that 𝑣 𝑡 (𝑥) = 𝑒 −𝑡 𝑛/2 𝑣(𝑒 −𝑡 𝑥) and 𝜇𝑡 = 𝑒 −2𝑡 𝜇 𝐹 solves


 Δ𝑣 𝑡 (𝑥) + 𝜇𝑡 𝑣 𝑡 (𝑥) = 0 in Ω𝑡


𝑣 𝑡𝑛 (𝑥) = 0 on 𝐹𝑡 ,

 𝑣 𝑡 (𝑥) = 0

on 𝜕Ω \ 𝐹𝑡

∫
and additionally satisfy Ω (𝑣 𝑡 ) 2 dvol = 1. Using the Hadamard formula we have
𝑡
that
∫ ∫
2
𝜕𝑡 𝜇𝑡 = −2𝑒 −2𝑡 𝜇 𝐹 = n·𝑥 grad𝜕Ω𝑡 𝑣 𝑡 − 𝜇𝑡 (𝑣 𝑡 ) 2 dS − n·𝑥 (𝑣 𝑡𝑛 ) 2 dS .
𝑒𝑡 𝐹 𝑒𝑡 𝜕Ω\𝐹
(C.4.1)
Setting 𝑡 = 0 and collecting the 𝜇 𝐹 -terms on the left-hand side we have
∫ ∫ ∫
2 2
𝜇 𝐹
n · 𝑥 𝑣 dS −2 = n · 𝑥 | grad𝜕Ω 𝑣| dS − n · 𝑥 𝑣 2𝑛 dS .
𝐹 𝜕Ω 𝜕Ω\𝐹
Remark C.4.2. • For the general

∫ mixed Neumann-Dirichlet problem we know
that in certain cases the 𝐹 𝑥 · n𝑣 2 dS = 2. For an example of when this
happens see Example C.5.3.
• Note that in the case that 𝐹 = 𝜕Ω and Ω is star-convex with respect to the
origin, we have that n · 𝑥 ≥ 0 with some open part of the boundary having
n · 𝑥 > 0. Hence assuming that 𝜇 ≠ 0 and using (C.4.1) we get
∫ ∫
2 2
n · 𝑥𝑣 dS −2 𝜇 = n · 𝑥 grad𝜕Ω 𝑣 dS > 0.
𝜕Ω 𝜕Ω
∫
In particular we have that 𝜕Ω
n · 𝑥𝑣 2 dS −2 > 0, meaning that the denomi-
nator is not identically 0.
107
• In the case that 𝐹 = ∅, we get that

∫
−2𝜇 = −𝐹
n · 𝑥 𝑣 2𝑛 dS,
𝜕Ω\𝐹
which simplifies to ∫
1
𝐹
𝜇 = n · 𝑥 𝑣 2𝑛 dS .
2 𝜕Ω\𝐹
C.5 Rellich-Christianson Type Identity

In this section we will look on the Rellich identities applied to polytopes. It should
be noted that if the solution is regular enough, the Rellich identities also work on
Lipschitz, due to an alternative proof using the divergence theorem, see [HS20;
Liu07], and Section C.4. For a good reference on the mixed Neumann-Dirichlet
problem on Lipschitz domains see [LR17].
A polytope is by definition a domain restricted by a finite number of hyper-
surfaces in R𝑛 with finite volume. We will let 𝑃 denote a polytope with faces 𝐹𝑖 ,
𝑖 = 1, . . . , 𝑑. The hypersurfaces corresponding to 𝐹𝑖 will be denoted 𝐻𝑖 . Let 𝑝 be
an arbitrary point in R𝑛 . Then we define the signed distance dist( 𝑝)𝑖 from 𝑝 to 𝐻𝑖
to be minus the distance from 𝑝 to 𝐻𝑖 in the case that 𝑝 and 𝑃 are on the same side
of 𝐻𝑖 , and the distance otherwise.
In the rest of this section we will assume that 𝑝 is at the origin, which we will
denote by 0. When this is the case we have that if 𝑥 ∈ 𝐹𝑖 then n · 𝑥 = dist(0)𝑖 . In
the next couple of subsections we will use this fact to simplify the Rellich identity.
The Rellich-Christianson identity for the mixed Neumann-Dirichlet generalizes the
Rellich-Christianson identity for the Dirichlet problem found in [Mét18]. We are
going to present a similar proof to the one found in [Mét18].
𝑃 𝐹𝑖
𝑥
𝐻𝑖
n · 𝑥 = dist(0)𝑖
108
C.5. Rellich-Christianson Type Identity
We are going to denote by 𝐹𝑖𝑛 = 𝐹𝑖 a face where we have posed the Neumann
boundary conditions and 𝐹𝑖𝑑 = 𝐹𝑖+ 𝑗 the faces where we have posed Dirichlet
conditions. In this case the mixed Neumann-Dirichlet problem becomes


 Δ𝑣(𝑥) + 𝜇𝑣(𝑥) = 0 in 𝑃


𝑣 𝑛 (𝑥) = 0 on 𝐹1𝑛 ∪ 𝐹2𝑛 ∪ · · · ∪ 𝐹 𝑗𝑛 . (C.5.1)

 𝑣(𝑥) = 0 on 𝐹1𝑑 ∪ 𝐹2𝑑 ∪···∪ 𝐹𝑘𝑑


To get the Rellich identity we will need that 𝑣 ∈ 𝐻 2 (𝑃).
Proposition C.5.1. Let 𝑣 ∈ 𝐻 2 (𝑃) be a solution to (C.5.1) which is normalized in

𝐿 2 (𝑃). In this case we have that 𝜇 and 𝑣 satisfy
𝑗 ∫ ! 𝑗 ∫
Õ Õ 2
2
𝜇 dist( 𝑝)𝑖 𝑣 dS −2 = dist( 𝑝)𝑖 grad𝜕𝑃 𝑣 dS
𝑖=1 𝐹𝑖𝑛 𝑖=1 𝐹𝑖𝑛
𝑘
Õ ∫
− dist( 𝑝)𝑖+ 𝑗 𝑣 2𝑛 dS . (C.5.2)
𝑖=1 𝐹𝑖𝑑
Remark C.5.2.
• In general, one can not assume that solutions are in 𝐻 2 (Ω) for 𝜕Ω is Lipschitz
domain. An example of this is the problem on the half disk 𝐷 + given by


 Δ𝑣 + 𝜇𝑣 = 0 on 𝐷 +


𝑣𝑛 = 0 on {(𝑟, 𝜋) : 𝑟 < 1}

𝑣 = 0 on {(𝑟, 0) : 𝑟 < 1} ∪ {(1, 𝜃) : 0 < 𝜃 < 𝜋}.


A solution to this problem is given by
𝑣(𝑟, 𝜃) = 𝐽1/2 ( 𝑗1,1/2𝑟) sin(𝜃/2),
where 𝑗1,1/2 is the first zero of the Bessel function 𝐽1/2 . Using the Taylor
series
∞
Õ (−1) 𝑚 𝑥 𝑚+1/2
𝐽1/2 (𝑥) =
𝑚=0
𝑚!Γ(𝑚 + 3/2) 2
is an eigenfunction on the half-ball 𝐷 + which is not in 𝐻 2 (𝐷 + ).
• The assumption that 𝑣 ∈ 𝐻 2 (𝑃) is for example true in the case of convex
polygons in R2 where we assume that the angle between two consecutive
sides 𝐹𝑖𝑛 and 𝐹 𝑗𝑑 have angle less than 𝜋/2, see [Dym79].
109
Proof. Using the Rellich identity for the mixed Neumann-Dirichlet problem we
have that
𝑗 ∫
! 𝑗 ∫ 𝑘 ∫
Õ Õ 2
Õ
2
𝜇 n · 𝑥 𝑣 dS −2 = n · 𝑥 grad𝜕𝑃 𝑣 dS − n · 𝑥 𝑣 2𝑛 dS .
𝑖=1 𝐹𝑖𝑛 𝑖=1 𝐹𝑖𝑛 𝑖=1 𝐹𝑖𝑑
Using the above computation that n · 𝑥 = dist(0)𝑖 on each face of the polytope we
get
𝑗 ∫ ! 𝑗 ∫
Õ Õ 2
2
𝜇 dist(0)𝑖 𝑣 dS −2 = dist(0)𝑖 grad𝜕𝑃 𝑣 dS
𝑘
Õ ∫
− dist(0) 𝑗+𝑖 𝑣 2𝑛 dS .
𝑖=1 𝐹𝑖𝑑
where | grad𝜕𝑃 𝑣| 2 = | grad 𝑣| 2 − 𝑣 2𝑛 . By translating the formula to an arbitrary

point 𝑝 we get
𝑗 ∫ ! 𝑗 ∫
Õ Õ 2
2
𝜇 dist( 𝑝)𝑖 𝑣 dS −2 = dist( 𝑝)𝑖 grad𝜕𝑃 𝑣 dS
𝑘
Õ ∫
− dist( 𝑝) 𝑗+𝑖 𝑣 2𝑛 dS .
𝑖=1 𝐹𝑖𝑑
Example C.5.3 (Example C.2.7 continued). We saw in Example C.2.7 that the
solutions to the problem (C.2.10) are given by
𝑣(𝑥, 𝑦) = 2 cos (𝜋𝑚𝑥) sin (𝜋𝑛𝑦)
with corresponding eigenvalues 𝜇 = (𝜋𝑚) 2 + (𝜋𝑛) 2 , where 𝑚 ∈ N ∪ {0} and

𝑛 ∈ N \ {0}. In this case we have that 𝐹1𝑛 = {0} × [0, 1], 𝐹2𝑛 = {1} × [0, 1],
𝐹1𝑑 = [0, 1] × {0} and 𝐹2𝑑 = [0, 1] × {1}. In this case we will let 𝑝 be arbitrary. On
∫ ∫
| grad𝜕𝑃 𝑣| 2 dS = | grad𝜕𝑃 𝑣| 2 dS = 2(𝜋𝑛) 2
𝐹1𝑛 𝐹2𝑛
and ∫ ∫
𝑣 2𝑛 dS = 𝑣 2𝑛 dS = 2(𝜋𝑛) 2 .
𝐹1𝑑 𝐹2𝑑
110
C.5. Rellich-Christianson Type Identity
In this case we have that

∫ ∫
2
dist( 𝑝)1 | grad𝜕𝑃 𝑣| dS + dist( 𝑝)2 | grad𝜕𝑃 𝑣| 2 dS = 2(𝜋𝑛) 2 ,
𝐹1𝑛 𝐹2𝑛
and ∫ ∫
dist( 𝑝)3 𝑣 2𝑛 dS + dist( 𝑝)4 𝑣 2𝑛 dS = 2(𝜋𝑛) 2 .
𝐹1𝑑 𝐹2𝑑
This means that

2
Õ ∫ 2
Õ ∫
2
dist( 𝑝)𝑖 grad𝜕𝑃 𝑣 dS − dist( 𝑝)𝑖+2 𝑣 2𝑛 dS = 0.
𝑖=1 𝐹𝑖𝑛 𝑖=1 𝐹𝑖𝑑
We have also have that

∫ ∫
𝑣 2 dS = 𝑣 2 dS = 2,
𝐹1𝑛 𝐹2𝑛
and ∫ ∫
dist( 𝑝)1 𝑣 2 dS + dist( 𝑝)2 𝑣 2 dS = 2.
𝐹1𝑛 𝐹2𝑛
Hence we end up with zero on both sides in (C.5.2).
111
Further Results D
Eigenvalues on Spherically Symmetric Manifolds

Stine Marie Berge
Further Results D
Eigenvalues on Spherically
Symmetric Manifolds
Abstract
In this note we will explore eigenvalues on balls for spherically symmetric
manifolds. We will compare the eigenvalues to the eigenvalues on balls in
Euclidean space.
D.1 Introduction
In this chapter we are going to compare the Dirichlet eigenvalues for the Laplacian
on balls in spherically symmetric (also called rotationally symmetric) manifolds to
the Euclidean space. Recall that the Dirichlet eigenvalues on the ball 𝐵𝑟0 (0) ⊂ R𝑛
are given by
𝑙 !2
𝑗 𝑚+𝑛/2−1
,
𝑟0
where 𝑗 𝑚+𝑛/2−1
𝑙 is the 𝑙’th zero of the Bessel function 𝐽𝑚+𝑛/2−1 . In [Bag90; Bag91]
the author compared all the Dirichlet eigenvalues for 𝐵𝑟0 ( 𝑝) ⊂ S2 to the Dirichlet
eigenvalues in the ball in the plane with the same radius using comparison theo-
rems for Sturm-Liouville equations. In particular, the author showed that the first
eigenvalue 𝜆1 (𝑟 0 ) on 𝐵𝑟0 ( 𝑝) ⊂ S2 satisfies
!2 ! !2
𝑗01 1 1 1 𝑗01 1
− 1+ 2 − 2 ≤ 𝜆1 (𝑟 0 ) ≤ − . (D.1.1)
𝑟0 4 sin (𝑟 0 ) 𝑟 0 𝑟0 3
More recently, the first eigenvalue of a ball on a spherically symmetric manifold
was compared to the first eigenvalue of a ball in Euclidean space in [BF17]. The
result in [BF17] generalizes (D.1.1) arbitrary symmetric manifolds.
115
Further Results D. Eigenvalues on Spherically Symmetric Manifolds
In this article we will present two different proofs of similar inequalities to

(D.1.1) for all eigenvalues for both the spheres and hyperbolic space. We will give
the first proof in Section D.2 using Sturm-Liouville theory as outlined in [Bag90].
In Section D.3 we will represent eigenfunctions of balls on spherically symmetric
manifolds using spherical harmonics. The goal of Section D.4 will be to express the
eigenvalues as zeros of radial functions satisfying a Sturm-Liouville equation. In
the last section we will generalize the inequality found in [BF17] to all eigenvalues,
and use it to give a new proof for the inequality on spheres and hyperbolic space.
D.2 Dirichlet Eigenvalues on Model Spaces

In this section we will work on the model space (𝑀𝐾 , g𝐾 ) with curvature 𝐾 and
dimension 𝑛. Denote by sin𝐾 (𝑟) the function defined by
√ √


 sin(𝑟 𝐾)/ 𝐾 for 𝐾 > 0


sin𝐾 (𝑟) = 𝑟 for 𝐾 = 0
 √ √
 sinh(𝑟 −𝐾)/ −𝐾

for 𝐾 < 0,

and cos𝐾 (𝑟) = sin𝐾 (𝑟) 0. We will see in Section D.3 that the solutions to the
Dirichlet eigenvalue problem
(
Δ𝑢𝑟0 + 𝜆(𝑟 0 )𝑢𝑟0 = 0 in 𝐵𝑟0 ( 𝑝) for 𝑝 ∈ 𝑀𝐾
(D.2.1)
𝑢𝑟0 = 0 at 𝑆𝑟𝑛−1
0
( 𝑝)
given on the form 𝑢𝑟0 (𝑟, 𝜃) = 𝑅𝑟𝑚0 (𝑟)Θ(𝜃) form a basis for 𝐿 2 (𝐵𝑟0 ( 𝑝)). The function
Θ is a spherical harmonic solving
ΔS𝑛−1 Θ + 𝑚(𝑛 − 2 + 𝑚)Θ = 0,

1
where 𝑚 ∈ N and the function 𝑅𝑟𝑚0 solves the equation
sin2𝐾 (𝑟)𝑅𝑟𝑚0 (𝑟) 00 + (𝑛 − 1) sin𝐾 (𝑟) cos𝐾 (𝑟)𝑅𝑟𝑚0 (𝑟) 0

+ (𝜆(𝑟 0 ) sin2𝐾 (𝑟) − 𝑚(𝑚 + 𝑛 − 2))𝑅𝑟𝑚0 (𝑟) = 0 (D.2.2)
with the condition 𝑅𝑟𝑚0 (𝑟 0 ) = 0 and that 𝑅𝑚 is bounded at 𝑟 = 0.

Looking at the eigenvalue problem
sin2𝐾 (𝑟)𝑅𝑟𝑚0 (𝑟) 00 + (𝑛 − 1) sin𝐾 (𝑟) cos𝐾 (𝑟)𝑅𝑟𝑚0 (𝑟) 0

− 𝑚(𝑚 + 𝑛 − 2)𝑅𝑟𝑚0 (𝑟) = −𝜆(𝑟 0 ) sin2𝐾 (𝑟)𝑅𝑟𝑚0 (𝑟)
116
D.2. Dirichlet Eigenvalues on Model Spaces
we will let 𝑅𝑟𝑚,𝑙

0
denote the eigenfunction corresponding to the 𝑙’th eigenvalue
which we denote by 𝜆 𝑚,𝑙 (𝑟 0 ). Using [Cha84, p. 318] we have that
lim 𝑟 2 𝜆 𝑚,𝑙 (𝑟 0 ) 𝑙
= ( 𝑗 𝑚+𝑛/2−1 )2.
𝑟0 →0 0
To compare the eigenvalues we are going to use the Sturm-Picone comparison

theorem.
Theorem D.2.1 (Sturm-Picone Comparison Theorem, [Hin05, Theorem B]). Let
𝑦 1 and 𝑦 2 be non-zero solutions to
( 𝑝 1 (𝑥)𝑦 10 (𝑥)) 0 + 𝑞 1 (𝑥)𝑦 1 (𝑥) = 0,

( 𝑝 2 (𝑥)𝑦 20 (𝑥)) 0 + 𝑞 2 (𝑥)𝑦 2 (𝑥) = 0,
on the interval [𝑎, 𝑏]. Assume that 0 < 𝑝 2 ≤ 𝑝 1 and 𝑞 1 ≤ 𝑞 2 , and let 𝑧 1 and 𝑧 2 be
two consecutive zeros of 𝑦 1 . Then either 𝑦 2 has a zero in the interval (𝑧 1 , 𝑧2 ), or
𝑦1 = 𝑦2.
We are now ready to state the main theorem:
Theorem D.2.2. Let 𝜆 𝑚,𝑙 (𝑟 0 ) be defined as above. In the case that either 𝑚 > 0
or 𝑛 > 2 we have that
𝑙 !2
𝑗 𝑚+𝑛/2−1 2𝑚 2 + 2𝑚(𝑛 − 2) − 𝑛(𝑛 − 1)
+ 𝐾 ≤ 𝜆 𝑚,𝑙 (𝑟 0 )
𝑟0 6
𝑙 !2 2 2 ! !
𝑗 𝑚+𝑛/2−1 𝑛−1 𝑛−2 1 1 1
≤ − 𝐾+ +𝑚 − − .
𝑟0 2 2 4 sin2𝐾 (𝑟 0 ) 𝑟 02
When 𝑚 = 0 and 𝑛 = 2 we get

!2 ! !2
𝑗0𝑙 1 1 1 𝑗0𝑙 1
− 𝐾+ 2 − 2 ≤ 𝜆0,2 (𝑟 0 ) ≤ − 𝐾.
𝑟0 4 sin𝐾 (𝑟 0 ) 𝑟 0 𝑟0 3
Remark D.2.3.
• In the case of the sphere S2 this inequality is known from [Bag90; Bag91].
Additionally, for the first eigenvalue the theorem is known from [BF17,
Thm. 3.3].
• When 𝐾 < 0 we have that

1 1 1 1
− ≤0 and lim − = 0.
sin2𝐾 (𝑟) 𝑟2 𝑟 →∞ sin2 (𝑟)
𝐾
𝑟2
117
When 𝐾 > 0 we have that

1 1
lim − = ∞.
𝑟 → √𝜋
𝐾
sin2𝐾 (𝑟) 𝑟2
For both positive and negative curvature one has that

1 1 𝐾
lim − = .
𝑟 →0 sin2𝐾 (𝑟) 𝑟2 3
• For 𝑛 = 3 and 𝑚 = 0 one has that Theorem D.2.2 simplifies to

𝑙 !2
𝑗1/2
𝜆0,𝑙 (𝑟 0 ) = − 𝐾.
𝑟0
Proof. We will only show this result for 𝑟 0 < √𝜋 in this section. For the case
2 |𝐾 |
for larger values of 𝑟 0 see Example D.5.3.
To get the radial part of 𝑢𝑟0 on Sturm-Liouville form, see Section B.5, we
denote by
𝑛−1
𝑣 𝐾 (𝑟) = sin𝐾2 (𝑟)𝑅𝑟𝑚,𝑙
0
(𝑟),
𝑛−2
and let 𝑚˜ = 2 + 𝑚. Then we have that
2 !
1 − 4𝑚˜ 2

00 𝑛−1
0= 𝑣𝐾 (𝑟) + 𝐾 + 𝜆 𝑚,𝑙 (𝑟 0 ) + 𝑣 𝐾 (𝑟). (D.2.3)
2 4 sin2𝐾 (𝑟)
Our goal is to use Theorem D.2.1 with the Sturm-Liouville equation for the Eu-
clidean space to get an estimate of 𝑟 0 . In this case we will use that 𝑢𝐶 (𝑟) =
√
𝑟 𝐽𝑚˜ (𝐶𝑟) solves the equation
1 − 4𝑚˜ 2

00
0 = 𝑢𝐶 (𝑟) + 𝐶 2 + 𝑢𝐶 . (D.2.4)
4𝑟 2
We will use Sturm-Picone comparison with 𝐶 being two different constants 𝐶1 and
𝐶2 .
Notice that sin21 (𝑟 ) − 𝑟12 is increasing and also that
𝐾
1 1 𝐾
lim+ − = .
𝑟 →0 sin2 (𝑟) 𝑟 2 3
𝐾
Denote by
!
4𝑚˜ 2 − 1 1 1 𝐾 (4𝑚˜ 2 − 1)
𝑎 𝑚˜ = − , 𝑏 𝑚˜ = .
4 sin2𝐾 (𝑟 0 ) 𝑟 02 12
118
D.2. Dirichlet Eigenvalues on Model Spaces
For the next part we will need a lower bound for the eigenvalue. Since we have
assumed that 𝑟 0 < √𝜋 for positive 𝐾 when 𝑚 > 0 we have that 𝜆 𝑚,𝑙 (𝑟 0 ) >
2 𝐾
𝑚(𝑚+𝑛−2)
sin2𝐾 (𝑟0 )
. We get this inequality from that the first zero of 𝑅𝑟𝑚,𝑙
0
is after the first
00
maximum, where means that 𝑅𝑟𝑚,𝑙
0
< 0 which means that
(𝜆 𝑚,𝑙 (𝑟 0 ) sin2𝐾 (𝑟) − 𝑚(𝑚 + 𝑛 − 2))𝑅𝑟𝑚,𝑙

0
(𝑟) > 0
which implies the inequality. For the negatively curved spaces we will use that
(𝑛 − 1) 2
− 𝐾 < 𝜆 𝑚,𝑙 (𝑟 0 )
4
which can be found in [Cha84, p. 46]. When either 𝑛 > 2 or 𝑚 > 0 we have that
1 − 4𝑚˜ 2 ≤ 0, and we will set
s
2
𝑛−1
𝐶1 = 𝜆 𝑚,𝑙 (𝑟 0 ) + 𝐾 − 𝑎 𝑚˜
2
and s
2
𝑛−1
𝐶2 = 𝜆 𝑚,𝑙 (𝑟 0 ) + 𝐾 − 𝑏 𝑚˜ .
2
In this case we have that
2
1 − 4𝑚˜ 2 1 − 4𝑚˜ 2 1 − 4𝑚˜ 2

2 𝑛−1 2
𝐶1 + ≤ 𝐾 + 𝜆 𝑚,𝑙 (𝑟 0 ) + ≤ 𝐶 2 + .
4𝑟 2 2 4 sin2𝐾 (𝑟) 4𝑟 2
Using Theorem D.2.1 for solutions to the equations (D.2.3) and (D.2.4) leads to
the estimate of 𝑟 0 by
𝑙
𝑗𝑚 𝑗𝑙
˜
≤ 𝑟 0 ≤ 𝑚˜ .
𝐶2 𝐶1
Solving for 𝜆 𝑚,𝑙 (𝑟 0 ) we get
𝑙 2
!
𝑗𝑚˜ 2𝑚 2 + 2𝑚(𝑛 − 2) − 𝑛(𝑛 − 1)
+ 𝐾 ≤ 𝜆 𝑚,𝑙 (𝑟 0 )
𝑟0 6
𝑙 2
! 2 !
4𝑚˜ 2 − 1

𝑗𝑚 ˜ 𝑛−1 1 1
≤ − 𝐾+ − .
𝑟0 2 4 sin2𝐾 (𝑟 0 ) 𝑟 02
When 𝑚 = 0 and 𝑛 = 2 we have that

2
1 − 4𝑚˜ 2 1 − 4𝑚˜ 2 1 − 4𝑚˜ 2

2 𝑛−1 2
𝐶2 + ≤ 𝐾 + 𝜆 0,𝑙 (𝑟 0 ) + ≤ 𝐶 1 + .
4𝑟 2 2 4 sin2𝐾 (𝑟) 4𝑟 2
119
Hence we obtain
!2 ! !2
𝑗0𝑙 1 1 1 𝑗0𝑙 1
− 𝐾+ 2 − 2 ≤ 𝜆0,𝑙 (𝑟 0 ) ≤ − 𝐾.
𝑟0 4 sin𝐾 (𝑟 0 ) 𝑟 0 𝑟0 3
D.3 Eigenvalues on Spherically Symmetric Manifolds

In this section we will assume that the (𝑀, g) is spherically symmetric manifold
with respect to a point 𝑝, meaning that the metric can be written as g = 𝑑𝑟 ⊗
𝑑𝑟 + 𝑓 2 (𝑟)gS𝑛−1 where 𝑟 (𝑥) = dist( 𝑝, 𝑥). By [Pet16, Prop. 1.4.7] the function
𝑓 : [0, 𝜌) → R satisfies
𝑓 (0) = 0, 𝑓 0 (0) = 1, 𝑓 00 (0) = 0.
In this case the Laplacian can be written on the form
𝜕2 𝑓 0 (𝑟) 𝜕 1
Δ= 2
+ (𝑛 − 1) + 2 ΔS𝑛−1 .
𝜕𝑟 𝑓 (𝑟) 𝜕𝑟 𝑓 (𝑟) 1
For more information about spherically symmetric manifolds see [Pet16, Sec. 4.2.3].
We will show that there exists an 𝐿 2 -basis on the form 𝑅 𝑡 (𝑟)Θ(𝜃) consisting of
eigenfunctions for the Dirichlet and Neumann problem on the ball 𝐵𝑡 ( 𝑝). The
radial part 𝑅 𝑡 solves the equation
( 𝑓 𝑛−1 (𝑟) (𝑅 𝑡 ) 0 (𝑟)) 0 − 𝑓 𝑛−3 (𝑟)𝑚(𝑚 + 𝑛 − 2)𝑅 𝑡 (𝑟) = −𝜆(𝑡) 𝑓 𝑛−1 (𝑟)𝑅 𝑡 (𝑟), (D.3.1)
with 𝑅 𝑡 (𝑡) = 0 and where 𝑅 𝑡 (0) is bounded. Notice that in the norm
∫ 𝑡
2
k𝜙k 𝑓 = 𝜙(𝑟) 2 𝑓 𝑛−1 (𝑟)dr,
0
the operator
𝐿 𝑚 𝜙 = ( 𝑓 𝑛−1 (𝑟)𝜙 0 (𝑟)) 0 − 𝑓 𝑛−3 (𝑟)𝑚(𝑚 + 𝑛 − 2)𝜙(𝑟)
is an unbounded self adjoint operator on the space
∫ 𝑡
1 𝑛−1 2 0
𝑋𝑡 = 𝜙 ∈ 𝐻loc (0, 𝑡) : 𝑓 (𝑟)𝜙(𝑟) dr < ∞, 𝜙(𝑡) = 𝜙(0)𝜙 (0) = 0 .
0
The eigenvalues of 𝐿 𝑚 are simple, since if 𝑢, 𝑣 are two eigenfunctions with same
eigenvalue 𝜆 we have that
0 = 𝑢𝐿 𝑚 𝑣 − 𝑣𝐿 𝑚 𝑢
= 𝑢( 𝑓 𝑛−1 (𝑟)𝑣 0) 0 − 𝑣( 𝑓 𝑛−1 (𝑟)𝑢 0) 0
= ( 𝑓 𝑛−1 (𝑟)(𝑢𝑣 0 − 𝑣𝑢 0)) 0 .
120
D.3. Eigenvalues on Spherically Symmetric Manifolds
This means that 𝑓 𝑛−1 (𝑟) (𝑢𝑣 0 − 𝑣𝑢 0) is constant. Using the boundary conditions
means that the Wronskian 𝑢𝑣 0 − 𝑣𝑢 0 = 0, hence 𝑢 and 𝑣 are linearly dependent.
We will let 𝜆 𝑚,𝑙 (𝑡) be the 𝑙’th eigenvalue of the problem (D.3.1) with the
corresponding eigenfunction 𝑅𝑚,𝑙 𝑡 .
Proposition D.3.1. Let (𝑀, g) be a spherically symmetric manifold with respect to

the point 𝑝 and consider the ball 𝐵𝑡 ( 𝑝) ⊂ 𝑀. Then there exists an 𝐿 2 -basis on 𝐵𝑡 ( 𝑝)
which consists of Dirichlet eigenfunctions on the form 𝑢 𝑡 (𝑟, 𝜃) = 𝑅𝑚,𝑙 𝑡 (𝑟)Θ (𝜃),
𝑚
where
ΔS𝑛−1 Θ𝑚 + 𝑚(𝑚 + 𝑛 − 2)Θ𝑚 = 0,
and 𝑅𝑚,𝑙
𝑡 is the eigenfunction of the 𝑙’th eigenvalue of the problem (D.3.1).
Proof. Let 𝑢(𝑟, 𝜃) = 𝑅(𝑟)Θ(𝜃) be an eigenfunction written in spherical coordinates

with eigenvalue 𝜆(𝑡). Then we have that
𝑓 0 (𝑟) 0 1
Δ𝑢(𝑟, 𝜃) = 𝑅 00 (𝑟)Θ(𝜃) + (𝑛 − 1) 𝑅 (𝑟)Θ(𝜃) + 𝑅(𝑟) 2 ΔS𝑛−1 Θ(𝜃)
𝑓 (𝑟) 𝑓 (𝑟) 1
= −𝜆(𝑡)𝑅(𝑟)Θ(𝜃),
which simplifies to
𝑓 2 (𝑟)𝑅 00 (𝑟) + (𝑛 − 1) 𝑓 0 (𝑟) 𝑓 (𝑟)𝑅 0 (𝑟) + 𝜆(𝑡) 𝑓 2 (𝑟)𝑅(𝑟) ΔS𝑛−1 Θ(𝜃)

=− 1 . (D.3.2)
𝑅(𝑟) Θ(𝜃)
Since the left hand-side of (D.3.2) is independent of 𝜃 we have that Θ is a spherical

harmonic function. Using that the eigenvalues on the (𝑛 − 1)-sphere have the form
𝑚(𝑚 + 𝑛 − 2) we get 𝑅 satisfies (D.3.1). Hence we get that if 𝑅(𝑟) is a solution
to (D.2.2) satisfying 𝑅(𝑟 0 ) = 0 and 𝑅(0) is bounded, then 𝑢 is a solution to the
Dirichlet problem. If on the other hand 𝑅 0 (𝑟 0 ) = 0 and 𝑅(0) is bounded we get
that 𝑢 is a solution to the Neumann problem.
The only thing left to show is that all eigenfunctions can be written as a
sum of eigenfunctions on the form 𝑅(𝑟)Θ(𝜃). Since the spherical harmonics
are the eigenfunctions of S𝑛−1 , we know that the spherical harmonics form an
orthonormal basis for 𝐿 2 (S1𝑛−1 ). This means that an arbitrary eigenfunction 𝑢(𝑟, 𝜃)
with eigenvalue 𝜆(𝑡) can be written as
∞
Õ
𝑢(𝑟, 𝜃) = 𝑎 𝑖 (𝑟)𝑌𝑖 (𝜃),
𝑖=1
where 𝑌𝑖 is the 𝑖’th spherical harmonic function with eigenvalue 𝑚 𝑖 (𝑚 𝑖 + 𝑛 − 2).
121
Using the Laplacian written out in spherical coordinates gives

∫
−𝑚 𝑖 (𝑚 𝑖 + 𝑛 − 2)𝑎 𝑖 (𝑟) = 𝑢(𝑟, 𝜃)ΔS𝑛−1 𝑌𝑖 (𝜃) dS
1
S1𝑛−1
∫
= ΔS𝑛−1 𝑢(𝑟, 𝜃)𝑌𝑖 (𝜃) dS
1
S1𝑛−1
= −𝜆(𝑡) 𝑓 2 (𝑟)𝑎 𝑖 (𝑟) − 𝑓 2 (𝑟)𝑎 𝑖00 (𝑟) − (𝑛 − 1) 𝑓 0 (𝑟) 𝑓 (𝑟)𝑎 𝑖0 (𝑟).

In particular, 𝑎 𝑖 (𝑟)𝑌𝑖 (𝜃) is an eigenfunction for all 𝑖. By orthogonality of eigen-
functions with different eigenvalues we get that 𝑢 can be written on the form
𝑚
Õ
𝑢(𝑟, 𝜃) = 𝑎 𝑖 (𝑟)Θ𝑖 (𝜃),
𝑖= 𝑗
where 𝑎 𝑖 (𝑟)Θ𝑖 (𝜃) has the same eigenvalue.

D.4 Eigenvalues on Balls

Let (𝑀, g) be a Riemannian manifold (not necessarily spherically symmetric) and
consider the ball 𝐵𝑡 ( 𝑝) ⊂ 𝑀 for 𝑝 ∈ 𝑀. We consider the Dirichlet problem
(
Δ𝑢 𝑡 + 𝜆 (𝑡) 𝑢 𝑡 = 0 on 𝐵𝑡 ( 𝑝)
. (D.4.1)
𝑢𝑡 = 0 on 𝑆𝑡𝑛−1 ( 𝑝)
∫
We will assume that 𝐵 ( 𝑝) (𝑢 𝑡 ) 2 dvol = 1, in which case
𝑡
∫
𝜆(𝑡) = | grad 𝑢 𝑡 | 2 dvol .
𝐵𝑡 ( 𝑝)
By using the Hadamard formula presented in [EI07, Cor. 2.1] one has that in the
case that 𝜆(𝑡) and 𝑢 𝑡 is differentiable with respect to 𝑡 one gets that
∫
0 2
𝜆 (𝑡) = − 𝑢 𝑡𝑛 dS .
𝑆𝑡𝑛−1 ( 𝑝)
For solutions to the Dirichlet problem we have the following result.

Proposition D.4.1. Let 𝑢 𝑡 be a solution to (D.4.1) which is normalized in 𝐿 2 .
Denote by 𝑟 the radial distance from the point 𝑝 and let 𝑟 1 < Inj( 𝑝) and assume
that 𝜆𝑡 and 𝑢 𝑡 are differentiable in 𝑡 for 𝑟 0 ≤ 𝑡 ≤ 𝑟 1 . Then we have that
𝑟 12 𝜆(𝑟 1 ) = lim 𝑡 2 𝜆(𝑡)+

𝑡→𝑟0
∫ 𝑟1 ∫
Δ (𝑟Δ𝑟) 2
(𝑢 𝑡 ) 2 +2 grad𝑆𝑟𝑛−1 ( 𝑝) 𝑢 𝑡 −2𝑟∇2𝑟 grad 𝑢 𝑡 , grad 𝑢 𝑡 dvol dt.

𝑡
𝑟0 𝐵𝑡 ( 𝑝) 2
122
D.4. Eigenvalues on Balls
The above proposition was proved for the first the first eigenvalue in [BF17,
Lem. 3.1]. Their proof was based on variational methods. We will give another
proof for general eigenvalues. Before proving Proposition D.4.1 we first develop
the following lemma of independent interest.
Lemma D.4.2. Let 𝑢 be a solution to Δ𝑢 + 𝜆𝑢 = 0 on the ball 𝐵𝑡 ( 𝑝) where
grad 𝑟 ( 𝑥)
𝑡 ≤ Inj( 𝑝). Let 𝜑 : (0, ∞) → (0, ∞) be such that 𝜑 (𝑟 ( 𝑥)) is a smooth vector field.
Then
∫
| grad 𝑢| 2 − 𝑢 2𝑛 dS
𝑆𝑡𝑛−1 ( 𝑝)
𝜑 (𝑟 (𝑥)) Δ𝑟 − 𝜑 0 (𝑟 (𝑥))
∫
= 𝜑 (𝑡) | grad 𝑢| 2 dvol
𝐵𝑡 ( 𝑝) 𝜑2 (𝑟 (𝑥))
∇2𝑟 (grad 𝑢, grad 𝑢) − 𝜆𝑢𝑟 𝑢 𝜑 (𝑟 (𝑥)) − 𝜑 0 (𝑟 (𝑥)) 𝑢𝑟2
∫
− 2𝜑 (𝑡) dvol .
𝐵𝑡 ( 𝑝) 𝜑2 (𝑟 (𝑥))
𝜕𝑟
Proof. We will use the notation 𝑋 = 𝜙 (𝑟 ( 𝑥)) and
𝑉 = 2𝑋 (𝑢) grad 𝑢 − | grad 𝑢| 2 𝑋.
Taking the divergence of 𝑉 one obtains
div(𝑉) = 2hgrad 𝑋 (𝑢), grad 𝑢i − 2𝜆𝑢𝑋 (𝑢) − hgrad | grad 𝑢| 2 , 𝑋i

− | grad 𝑢| 2 div(𝑋)
= 2h∇grad 𝑢 𝑋, grad 𝑢i + 2h𝑋, ∇grad 𝑢 grad 𝑢i − 2h∇𝑋 grad 𝑢, grad 𝑢i
− 2𝜆𝑢𝑋 (𝑢) − | grad 𝑢| 2 div(𝑋)
= 2h∇grad 𝑢 𝑋, grad 𝑢i + ∇2 𝑢(grad 𝑢, 𝑋) − 2∇2 𝑢(𝑋, grad 𝑢) − 2𝜆𝑢𝑋 (𝑢)
− | grad 𝑢| 2 div(𝑋)
= 2h∇grad 𝑢 𝑋, grad 𝑢i − 2𝜆𝑢𝑋 (𝑢) − | grad 𝑢| 2 div(𝑋).
Expanding the term

𝜙(𝑟 (𝑥))∇2𝑟 (grad 𝑢, grad 𝑢) − 𝜙 0 (𝑟 (𝑥))𝑢𝑟2

𝜕𝑟
∇grad 𝑢 , grad 𝑢 =
𝜙(𝑟 (𝑥)) 𝜙(𝑟 (𝑥)) 2
we get
2(𝜙(𝑟 (𝑥))∇2𝑟 (grad 𝑢, grad 𝑢) − 𝜙 0 (𝑟 (𝑥))𝑢𝑟2 ) − 2𝜆𝜙(𝑟 (𝑥))𝑢𝑢𝑟
div(𝑉) =
𝜙(𝑟 (𝑥)) 2
| grad 𝑢| (𝜙(𝑟 (𝑥))Δ𝑟 − 𝜙 0 (𝑟 (𝑥)))
2
+ .
𝜙(𝑟 (𝑥)) 2
123
Using the divergence theorem with div(𝑉) together with
| grad𝑆𝑡 𝑢| 2 − (𝜕𝑟 𝑢) 2
h𝑉, 𝜕𝑟 i = ,
𝜙(𝑡)
gives the result.
Proof of Prop. D.4.1. Lemma D.4.2 with 𝜑(𝑟 (𝑥)) = 𝑟 −1 (𝑥) implies that
∫
2
𝑡𝜆 0 (𝑡) = grad 𝑢 𝑡 (𝑟Δ𝑟 + 1)
𝐵𝑡 ( 𝑝)

− 2 𝑟∇2𝑟 grad 𝑢 𝑡 , grad 𝑢 𝑡 + (𝑢 𝑡 ) 𝑛2 − 𝜆 (𝑡) 𝑟 (𝑢 𝑡 ) 𝑛 𝑢 𝑡

dvol .
Using the equation
∫ ∫ ∫
1 𝑡 2 1
𝑡 𝑡
𝑟 (𝑢 ) 𝑛 𝑢 dvol = − (𝑢 ) dvol − 𝑟Δ𝑟 (𝑢 𝑡 ) 2 dvol
𝐵𝑡 ( 𝑝) 2 𝐵𝑡 ( 𝑝) 2 𝐵𝑡 ( 𝑝)
gives
∫
0 2
grad 𝑢 𝑡 (𝑟Δ𝑟 + 1) − 2 𝑟∇2𝑟 grad 𝑢 𝑡 , grad 𝑢 𝑡 + (𝑢 𝑡 ) 𝑛2 dvol

𝑡𝜆 (𝑡) =
𝐵𝑡 ( 𝑝)
∫
− 𝜆 (𝑡) − 𝜆 (𝑡) 𝑟Δ𝑟 (𝑢 𝑡 ) 2 dvol .
𝐵𝑡 ( 𝑝)
Using the formulas above on (𝑡 2 𝜆(𝑡)) 0 = 𝑡 (2𝜆(𝑡) + 𝑡𝜆 0 (𝑡)) we get
0 ∫
2
𝑡 2 𝜆 (𝑡) grad 𝑢 𝑡 (𝑟Δ𝑟 + 2) − 2𝑟∇2𝑟 grad 𝑢 𝑡 , grad 𝑢 𝑡 dvol

=𝑡
𝐵𝑡 ( 𝑝)
∫ ∫
𝑡 2
−2 (𝑢 ) 𝑛 dvol −𝑡𝜆 (𝑡) 𝑟Δ𝑟 (𝑢 𝑡 ) 2 dvol .
𝐵𝑡 ( 𝑝) 𝐵𝑡 ( 𝑝)
124
D.5. Spherical Symmetric Manifolds
Further simplifications gives that

∫
1
(𝑡 2 𝜆 (𝑡)) 0 = 𝑡 Δ((𝑢 𝑡 ) 2 ) + 𝜆(𝑡)(𝑢 𝑡 ) 2 𝑟Δ𝑟 dvol
𝐵𝑡 ( 𝑝) 2
∫
2
𝑡 2 𝑡 𝑡
+ 2𝑡 grad𝑆𝑟𝑛−1 ( 𝑝) 𝑢 − 𝑟∇ 𝑟 grad 𝑢 , grad 𝑢 dvol
𝐵𝑡 ( 𝑝)
∫
− 𝑡𝜆 (𝑡) 𝑟Δ𝑟 (𝑢 𝑡 ) 2 dvol
𝐵 ( 𝑝)
∫ 𝑡
1
= −𝑡𝜆 (𝑡) grad(𝑢 𝑡 ) 2 , grad (𝑟Δ𝑟) dvol
𝐵𝑡 ( 𝑝) 2
∫
2
𝑡 2 𝑡 𝑡
+ 2𝑡 grad𝑆𝑟𝑛−1 ( 𝑝) 𝑢 − 𝑟∇ 𝑟 grad 𝑢 , grad 𝑢 dvol
𝐵𝑡 ( 𝑝)
∫
Δ (𝑟Δ𝑟) 2
=𝑡 (𝑢 𝑡 ) 2 + 2 grad𝑆𝑟𝑛−1 ( 𝑝) 𝑢 𝑡 dvol
𝐵𝑡 ( 𝑝) 2
∫
2𝑟∇2𝑟 grad 𝑢 𝑡 , grad 𝑢 𝑡 dvol .

−𝑡
𝐵𝑡 ( 𝑝)
Finally, integrating the identity above gives the result.
Remark D.4.3. Let (𝑀, g) be an analytic Riemannian manifold and let 𝜙 𝑠 :
𝐵𝑡 ( 𝑝) → 𝐵𝑠+𝑡 ( 𝑝) be the flow of the vector field 𝜕𝑟 which is analytic for 𝑟 + 𝑡 <
Inj( 𝑝). Then by [EI07, Lem. 3.1] we can find a differentiable family 𝜆(𝑡) and 𝑢 𝑡
for 𝑡 ∈ (𝑟 − 𝜖, 𝑟 + 𝜖).
D.5 Spherical Symmetric Manifolds

We now assume that (𝑀, g) is spherically symmetric with respect to the point
𝑝 ∈ 𝑀. Using Proposition D.4.1 we get the following corollary. For the first
eigenvalue the representation was shown in [BF17, Lem. 3.1]. Using [Cha84,
p. 318] one has that lim𝑡→0 𝑡 2 𝜆(𝑡) = ( 𝑗 𝑚+𝑛/2−1
𝑙 ) 2 for a continuous family 𝜆(𝑡).
We will assume that 𝜆 𝑚,𝑙 (𝑡) and 𝑅𝑚,𝑙𝑡 are differentiable in 𝑡.
𝑡 (𝑟)Θ(𝜃) be a 𝐿 2 -normalized solution to
Corollary D.5.1. Let 𝑢 𝑡 (𝑟, 𝜃) = 𝑅𝑚,𝑙
(D.4.1) on the ball 𝐵𝑡 ( 𝑝) where
ΔS𝑛−1 Θ = −𝑚(𝑚 + 𝑛 − 2)Θ,
1
and 𝑡
𝑅𝑚,𝑙 𝑡 (𝑡) = 0 and 𝑅 𝑡 (𝑡) is bounded. Denote by
satisfies (D.3.1) with 𝑅𝑚,𝑙 𝑚,𝑙
𝑛−1
𝐹 (𝑟) = 3 (3 − 𝑛)𝑟 𝑓 0 (𝑟) 3 + (𝑟 𝑓 000 (𝑟) + 2 𝑓 00 (𝑟)) 𝑓 2 (𝑟)
𝑓 (𝑟)

+ ((𝑛 − 4)𝑟 𝑓 00 (𝑟) + (𝑛 − 3) 𝑓 0 (𝑟)) 𝑓 0 (𝑟) 𝑓 (𝑟) .
125
Then for 0 < 𝑟 0 < 𝑟 1 we have

𝑓 0 (𝑟)
∫ 𝑟1 ∫
1 4𝑚(𝑚 − 2 + 𝑛)
𝜆(𝑟 1 ) = 2 𝑡 (𝑢 𝑡 ) 2 𝐹 (𝑟) + 1 − 𝑟 dvol dt
2𝑟 1 𝑟0 𝐵𝑡 ( 𝑝) 𝑓 2 (𝑟) 𝑓 (𝑟)
1
+ 2 lim 𝑠2 𝜆(𝑠).
𝑟 1 𝑠→𝑟0
In particular taking the limit of 𝑟 0 → 0 one has that

𝑓 0 (𝑟)
∫ 𝑟1 ∫
1 𝑡 2 4𝑚(𝑚 − 2 + 𝑛)
𝜆(𝑟 1 ) = 2 𝑡 (𝑢 ) 𝐹 (𝑟) + 1−𝑟 dvol dt
2𝑟 1 0 𝐵𝑡 ( 𝑝) 𝑓 2 (𝑟) 𝑓 (𝑟)
( 𝑗 𝑚+𝑛/2−1
𝑙 )2
+ .
𝑟 12
Proof. Using Proposition D.4.1 we need to compute Δ(𝑟Δ𝑟). Notice first that
𝑓 0 (𝑟)
Δ𝑟 = (𝑛 − 1) .
𝑓 (𝑟)
0
Moreover, the radial part of the Laplacian is 𝜕𝑟2 + (𝑛 − 1) 𝑓𝑓 (𝑟(𝑟)) 𝜕𝑟 . Hence we get that
𝑓 0 (𝑟) (𝑛 − 1)𝑟 𝑓 0 (𝑟)

Δ(𝑟Δ𝑟) = 𝜕𝑟2 + (𝑛 − 1) 𝜕𝑟 = 𝐹 (𝑟).
𝑓 (𝑟) 𝑓 (𝑟)
Using the divergence theorem with 𝑠 < 𝑡 we get that
∫ ∫
𝑡 2 𝑚(𝑚 + 𝑛 − 2)
| grad𝑆𝑠𝑛−1 𝑢 | dS = (𝑢 𝑡 ) 2 dS,
𝑆𝑠𝑛−1 𝑓 2 (𝑠) 𝑆𝑠𝑛−1
1
where we have used that Δ𝑆𝑠𝑛−1 = Δ 𝑛−1 .
𝑓 2 (𝑠) S1
Thus the result follows.
Remark D.5.2. For the first eigenfunction we have that 𝑚 = 0 and get the same
result as in [BF17, Lem. 3.1].
Example D.5.3. In the case that (𝑀𝐾 , g𝐾 ) is a model space we have that 𝐹 in
Corollary D.5.1 simplifies to
sin𝐾 (𝑟) − 𝑟 cos𝐾 (𝑟)
𝐹 (𝑟) = (−𝐾) (𝑛 − 1) 2 + (𝑛 − 1) (𝑛 − 3) .
sin3𝐾 (𝑟)
When 𝐾 > 0 we will assume that 𝑟 1 < √𝜋 . The function

𝐾
sin𝐾 (𝑟) − 𝑟 cos𝐾 (𝑟)

𝐺 (𝑟) =
sin3𝐾 (𝑟)
126
D.5. Spherical Symmetric Manifolds
−𝐾
is increasing and satisfies lim𝑟 →0 𝐺 (𝑟) = 3 . Hence 𝐺 (𝑟) satisfies
−𝐾 sin𝐾 (𝑡) − 𝑡 cos𝐾 (𝑡)

≤ 𝐺 (𝑟) ≤ ,
3 sin3𝐾 (𝑡)
on the interval 𝑟 ∈ (0, 𝑡].

We will assume that 𝜆 𝑚,𝑙 (𝑡) is the 𝑙’th eigenvalue of 𝑅𝑚,𝑙
𝑡 and that 𝑢 𝑡 (𝑟, 𝜃) =
𝑡 (𝑟)Θ(𝜃) where 𝑢 𝑡 is normalized in 𝐿 2 (𝐵 ( 𝑝)) and 𝑅 𝑡
𝑅𝑚,𝑙 𝑡 𝑚,𝑙 satisfies (D.3.1) and
Θ is a spherical harmonic function with eigenvalue 𝑚(𝑚 + 𝑛 − 2).
When 𝑛 ≥ 3 or 𝑚 ≥ 1 we get that
( 𝑗 𝑚+𝑛/2−1
𝑙 )2 (𝑛 − 1) 2
∫ 𝑟1
𝜆 𝑚,𝑙 (𝑟 1 ) ≤ −𝐾 𝑡 dt
𝑟 12 2𝑟 12 0
" #0
𝑡2
∫ 𝑟1
(𝑛 − 1) (𝑛 − 3) + 4𝑚(𝑚 + 𝑛 − 2)
+ dt
2𝑟 12 0 sin2𝐾 (𝑡)
)2
!
( 𝑗 𝑚+𝑛/2−1
𝑙
(𝑛 − 1) 2 4(𝑚 + (𝑛 − 2)/2) 2 − 1 1 1
= −𝐾 + − .
𝑟 12 4 4 sin2𝐾 (𝑟 1 ) 𝑟 12
This is the same inequality as in Theorem D.2.2. Using the lower bound of 𝐺 (𝑡)
we obtain
( 𝑗 𝑚+𝑛/2−1
𝑙 )2 2𝑚 2 + 2𝑚(𝑛 − 2) − 𝑛(𝑛 − 1)
𝜆 𝑚,𝑙 (𝑟 1 ) ≥ + 𝐾.
𝑟 12 6
For the case 𝑛 = 2 and 𝑚 = 0 we also get the same result as in Theorem D.2.2.
127
Bibliography
[Agm66] S. Agmon. Unicité et convexité dans les problèmes différentiels. Les

Presses de l’Université de Montréal, 1966.
[Agr06] M. S. Agranovich. “On a mixed Poincaré-Steklov type spectral problem
in a Lipschitz domain”. In: Russian Journal of Mathematical Physics
13 (2006), pp. 239–244.
[Ale+09] G. Alessandrini et al. “The stability for the Cauchy problem for elliptic
equations”. In: Inverse Problems 25.12 (2009), p. 123004.
[Alm79] F. J. Almgren Jr. “Dirichlet’s problem for multiple valued functions
and the regularity of mass minimizing integral currents”. In: Minimal
submanifolds and geodesics. North-Holland, Amsterdam-New York,
1979, pp. 1–6.
[ALM16] N. Anantharaman, M. Léautaud, and F. Macià. “Wigner measures and
observability for the Schrödinger equation on the disk”. In: Inventiones
Mathematicae 206.2 (2016), pp. 485–599.
[Aub82] T. Aubin. Nonlinear analysis on manifolds. Monge-Ampère equations.
Springer-Verlag New York, 1982.
[Bag91] F. E. Baginski. “Errata: Upper and lower bounds for eigenvalues of the
Laplacian on a spherical cap”. In: Quarterly of Applied Mathematics
49.2 (1991), p. 399.
[Bag90] F. E. Baginski. “Upper and lower bounds for eigenvalues of the Lapla-
cian on a spherical cap”. In: Quarterly of Applied Mathematics 48.3
(1990), pp. 569–573.
[Bañ+10] R. Bañuelos et al. “Eigenvalue inequalities for mixed Steklov prob-
lems”. In: Operator theory and its applications. Vol. 231. American
Mathematical Society, 2010, pp. 19–34.
[Ber21] S. M. Berge. “Convexity properties of harmonic functions on parame-
terized families of hypersurfaces”. In: The Journal of Geometric Anal-
ysis 31 (2021), pp. 953–979.
129
Bibliography
[BM21] S. M. Berge and E. Malinnikova. “On the three ball theorem for solu-
tions of the Helmholtz equation”. In: Complex Analysis and its Syner-
gies 7 (2021).
[BF17] D. Borisov and P. Freitas. “The spectrum of geodesic balls on spheri-
cally symmetric manifolds”. In: Communications in Analysis and Ge-
ometry 25.3 (2017), pp. 507–544.
[Bor20] D. Borthwick. Spectral theory. Springer International Publishing, 2020.
[BNO19] E. Burman, M. Nechita, and L. Oksanen. “Unique continuation for
the Helmholtz equation using stabilized finite element methods”. In:
Journal de Mathématiques Pures et Appliquées. Neuvième Série 129
(2019), pp. 1–22.
[CH94] R. Camporesi and A. Higuchi. “Spectral functions and zeta functions in
hyperbolic spaces”. In: Journal of Mathematical Physics 35.8 (1994),
pp. 4217–4246.
[Cha84] I. Chavel. Eigenvalues in Riemannian geometry. Including a chapter
by B. Randol, With an appendix by J. Dodziuk. Academic Press, 1984.
[Cha06] I. Chavel. Riemannian geometry. Second. Cambridge University Press,
2006.
[CLN06] B. Chow, P. Lu, and L. Ni. Hamilton’s Ricci flow. American Mathe-
matical Society, 2006.
[Chr19] H. Christianson. “Equidistribution of Neumann data mass on simplices
and a simple inverse problem”. In: Mathematical research letters 26.2
(2019), pp. 421–445.
[CEG11] B. Colbois, A. El Soufi, and A. Girouard. “Isoperimetric control of the
Steklov spectrum”. In: Journal of Functional Analysis 261.5 (2011),
pp. 1384–1399.
[DF88] H. Donnelly and C. Fefferman. “Nodal sets of eigenfunctions on
Riemannian manifolds”. In: Inventiones Mathematicae 93.1 (1988),
pp. 161–183.
[Don69] W. F. Donoghue Jr. Distributions and Fourier transforms. Academic
Press, New York, 1969.
[Dur12] P. Duren. Invitation to classical analysis. American Mathematical So-
ciety, 2012.
[Dym79] V. I. Dymčenko. “Asymptotic expansions of the solutions of the Helmholtz
equation in the neighborhood of angular boundary points”. In: Akademiya
Nauk SSSR 19.5 (1979), pp. 1205–1216.
130
Bibliography
[EI07] A. El Soufi and S. Ilias. “Domain deformations and eigenvalues of the

Dirichlet Laplacian in a Riemannian manifold”. In: Illinois Journal of
Mathematics 51.2 (2007), pp. 645–666.
[EL97] Á. Elbert and A. Laforgia. “An upper bound for the zeros of the
derivative of Bessel functions”. In: Rendiconti del Circolo Matematico
di Palermo. Serie II 46.1 (1997), pp. 123–130.
[Eld13] J. Eldering. Normally hyperbolic invariant manifolds. Atlantis Press,
Paris, 2013.
[Fed59] H. Federer. “Curvature measures”. In: Transactions of the American
Mathematical Society 93 (1959), pp. 418–491.
[Fra12] T. Frankel. The geometry of physics. Third. Cambridge University
Press, 2012.
[GL86] N. Garofalo and F.-H. Lin. “Monotonicity properties of variational
integrals, 𝐴 𝑝 weights and unique continuation”. In: Indiana University
Mathematics Journal 35.2 (1986), pp. 245–268.
[GT01] D. Gilbarg and N. S. Trudinger. Elliptic partial differential equations
of second order. Reprint of the 1998 edition. Springer-Verlag, Berlin,
2001.
[GP17] A. Girouard and I. Polterovich. “Spectral geometry of the Steklov
problem”. In: Shape optimization and spectral theory. De Gruyter
Open, Warsaw, 2017, pp. 120–148.
[Gri10] P. Grinfeld. “Hadamard’s formula inside and out”. In: Journal of Op-
timization Theory and Applications 146.3 (2010), pp. 654–690.
[Han07] Q. Han. “Nodal sets of harmonic functions”. In: Pure and Applied
Mathematics Quarterly 3.3 (2007), pp. 647–688.
[HS20] A. Hassannezhad and A. Siffert. “A note on Kuttler-Sigillito’s inequal-
ities”. In: Annales Mathématiques du Québec 44.1 (2020), pp. 125–
147.
[Hin05] D. Hinton. “Sturm’s 1836 oscillation results evolution of the theory”.
In: Sturm-Liouville theory. Birkhäuser, Basel, 2005, pp. 1–27.
[Hör18] L. Hörmander. Unpublished manuscripts—from 1951 to 2007. Springer,
Cham, 2018.
[HI04] T. Hrycak and V. Isakov. “Increased stability in the continuation of
solutions to the Helmholtz equation”. In: Inverse Problems 20.3 (2004),
pp. 697–712.
131
Bibliography
[Isa17] V. Isakov. Inverse problems for partial differential equations. Third.

Springer, Cham, 2017.
[IK11] V. Isakov and S. Kindermann. “Subspaces of stability in the Cauchy
problem for the Helmholtz equation”. In: Methods and Applications of
Analysis 18.1 (2011), pp. 1–29.
[Jak+06] D. Jakobson et al. “Spectral problems with mixed Dirichlet-Neumann
boundary conditions: isospectrality and beyond”. In: Journal of Com-
putational and Applied Mathematics 194.1 (2006), pp. 141–155.
[Jos11] J. Jost. Riemannian geometry and geometric analysis. Sixth. Springer-
Verlag Berlin Heidelberg, 2011.
[Kre10] P. Kreyssig. “An introduction to harmonic manifolds and the Lich-
nerowicz conjecture”. In: arXiv e-prints, arXiv:1007.0477 (2010),
arXiv:1007.0477. arXiv: 1007.0477 [math.DG].
[Kuk98] I. Kukavica. “Quantitative uniqueness for second-order elliptic opera-
tors”. In: Duke Mathematical Journal 91.2 (1998), pp. 225–240.
[KS68] J. R. Kuttler and V. G. Sigillito. “Inequalities for membrane and Stekloff
eigenvalues”. In: Journal of Mathematical Analysis and Applications
23.1 (1968), pp. 148–160.
[Kuz+14] N. Kuznetsov et al. “The legacy of Vladimir Andreevich Steklov”. In:
Notices of the American Mathematical Society 61.1 (2014), pp. 9–22.
[Lan63] E. M. Landis. “A three-spheres theorem”. In: Doklady Akademii Nauk
SSSR 148 (1963), pp. 277–279.
[Lee18] J. M. Lee. Introduction to Riemannian manifolds. Second. Springer,
New York, 2018.
[Lee13] J. M. Lee. Introduction to smooth manifolds. Second. Springer, New
York, 2013.
[Liu07] G. Liu. “Rellich type identities for eigenvalue problems and application
to the Pompeiu problem”. In: Journal of Mathematical Analysis and
Applications 330.2 (2007), pp. 963–975.
[Log18a] A. Logunov. “Nodal sets of Laplace eigenfunctions: polynomial up-
per estimates of the Hausdorff measure”. In: Annals of Mathematics.
Second Series 187.1 (2018), pp. 221–239.
[Log18b] A. Logunov. “Nodal sets of Laplace eigenfunctions: proof of Nadi-
rashvili’s conjecture and of the lower bound in Yau’s conjecture”. In:
Annals of Mathematics. Second Series 187.1 (2018), pp. 241–262.
132
Bibliography
[LM18] A. Logunov and E. Malinnikova. “Quantitative propagation of small-

ness for solutions of elliptic equations”. In: Proceedings of the Inter-
national Congress of Mathematicians—Rio de Janeiro 2018. Vol. III.
Invited lectures. World Sci. Publ., Hackensack, NJ, 2018, pp. 2391–
2411.
[LR17] V. Lotoreichik and J. Rohleder. “Eigenvalue inequalities for the Lapla-
cian with mixed boundary conditions”. In: Journal of Differential
Equations 263.1 (2017), pp. 491–508.
[Mal00] E. Malinnikova. “The theorem on three spheres for harmonic differ-
ential forms”. In: Complex analysis, operators, and related topics.
Birkhäuser, Basel, 2000, pp. 213–220.
[Man13] D. Mangoubi. “The effect of curvature on convexity properties of
harmonic functions and eigenfunctions”. In: Journal of the London
Mathematical Society. Second Series 87.3 (2013), pp. 645–662.
[McL00] W. McLean. Strongly elliptic systems and boundary integral equations.
Cambridge University Press, 2000.
[Mét18] A. Métras. “Rellich–Christianson type identities for the Neumann data
mass of Dirichlet eigenfunctions on polytopes”. In: Annales mathéma-
tiques du Québec 42.1 (2018), pp. 95–99.
[Min75] K. Minemura. “Eigenfunctions of the Laplacian on a real hyperbolic
space”. In: Journal of the Mathematical Society of Japan 27.1 (1975),
pp. 82–105.
[Olv97] F. W. J. Olver. Asymptotics and special functions. Reprint of the 1974
original. A K Peters, Ltd., Wellesley, MA, 1997.
[Pet16] P. Petersen. Riemannian geometry. Third. Springer, Cham, 2016.
[Rel40] F. Rellich. “Darstellung der Eigenwerte von Δ𝑢 + 𝜆𝑢 = 0 durch ein
Randintegral”. In: Mathematische Zeitschrift 46 (1940), pp. 635–636.
[Seo21] D.-H. Seo. “A shape optimization problem for the first mixed Steklov-
Dirichlet eigenvalue”. In: Annals of Global Analysis and Geometry
59.3 (2021), pp. 345–365.
[Shu01] M. A. Shubin. Pseudodifferential operators and spectral theory. Sec-
ond. Translated from the 1978 Russian original by Stig I. Andersson.
Springer-Verlag, Berlin, 2001, pp. xii+288.
[SI07] D. A. Subbarayappa and V. Isakov. “On increased stability in the
continuation of the Helmholtz equation”. In: Inverse Problems 23.4
(2007), pp. 1689–1697.
133
Bibliography
[Wat95] G. N. Watson. A treatise on the theory of Bessel functions. Reprint of

the second (1944) edition. Cambridge University Press, 1995.
134
Doctoral thesis
Stine Marie Berge
Stine Marie Berge

Quantitative Unique
ISBN 978-82-326-6943-1 (printed ver.)

NTNU
Thesis for the Degree of
Philosophiae Doctor
Faculty of Information Technology and Electrical
Engineering

Stine Marie Berge - PHD

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Stine Marie Berge - PHD

Uploaded by

Copyright:

Available Formats

Doctoral theses at NTNU, 2021:285

ISBN 978-82-326-6943-1 (printed ver.)

Doctoral theses at NTNU, 2021:285

Thesis for the Degree of Philosophiae Doctor

Trondheim, September 2021

Norwegian University of Science and Technology

Thesis for the Degree of Philosophiae Doctor

Faculty of Information Technology and Electrical Engineering