Topics in Spectral Geometry

to appear in the AMS Graduate Studies in Mathematics series

michael levitin
University of Reading
michaellevitin.net

dan mangoubi
The Hebrew University
www.math.huji.ac.il/~mangoubi/

spectral adj. \ "spEktr@l \ iosif polterovich

Université de Montréal
dms.umontreal.ca/~iossif/

2. a. Having the character of a spectre or phantom;

ghostly, unsubstantial, unreal.

5. a. Of or pertaining to, appearing or observed in, the

spectrum. Also applied to a property or parame-
ter which is being considered as a function of fre-
quency or wave-length, or which pertains to a given
frequency range or value within the spectrum.

From Oxford English Dictionary typeset May 29, 2023

Preliminary version (May 29, 2023)

 Please refer to this version as
M. Levitin, D. Mangoubi, and I. Polterovich, Topics in Spectral Geometry, preliminary version
dated May 29, 2023. https://michaellevitin.net/Book/.
 Contents

Preface 1

Introduction 5

Chapter 1. Strings, drums, and the Laplacian 11

§1.1. Basic examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
§1.1.1. The Laplace operator . . . . . . . . . . . . . . . . . . . . . . . . . 11
§1.1.2. Vibrating strings . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
§1.1.3. Vibrating drums . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
§1.2. The Laplacian on a Riemannian manifold . . . . . . . . . . . . . . . . 22
§1.2.1. The Laplace–Beltrami operator . . . . . . . . . . . . . . . . . . 22
§1.2.2. The Laplacian on a flat torus . . . . . . . . . . . . . . . . . . . . 26
§1.2.3. The Laplace operator on spheres . . . . . . . . . . . . . . . . . 30
Chapter 2. The spectral theorems 35
§2.1. Weak spectral theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
§2.1.1. Spectral theorems: an overview and the roadmap . . . . . . 35
§2.1.2. Weak derivatives and Sobolev spaces . . . . . . . . . . . . . . . 36
§2.1.3. Weak solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
§2.1.4. The weak spectral theorem for the Dirichlet Laplacian . . 42
§2.1.5. Self-adjoint unbounded linear operators . . . . . . . . . . . . 44
§2.1.6. The Dirichlet Laplacian via the Friedrichs extension . . . . 46
§2.1.7. The weak spectral theorem: Neumann case . . . . . . . . . . 47
§2.1.8. The weak spectral theorem: Riemannian manifolds . . . . 49
§2.2. Elliptic regularity and strong spectral theorems . . . . . . . . . . . . . 51
§2.2.1. Elliptic regularity for the Dirichlet Laplacian . . . . . . . . . 51
§2.2.2. Proof of the local regularity . . . . . . . . . . . . . . . . . . . . . 52
§2.2.3. A priori estimates and the method of difference quo-
tients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 53
§2.2.4. Global regularity of Dirichlet eigenfunctions . . . . . . . .. 57

iii
iv Contents

§2.2.5. Continuity up to the boundary of Dirichlet eigenfunc-

tions on Lipschitz domains . . . . . . . . . . . . . . . . . . . . . 61
§2.2.6. Regularity of Neumann eigenfunctions . . . . . . . . . . . . 62
§2.2.7. Strong spectral theorems. . . . . . . . . . . . . . . . . . . . . . . 63
Chapter 3. Variational principles and applications 65
§3.1. Variational principles for Laplace eigenvalues . . . . . . . . . . . . . . 65
§3.1.1. The Rayleigh quotient . . . . . . . . . . . . . . . . . . . . . . . . 65
§3.1.2. Variational principles . . . . . . . . . . . . . . . . . . . . . . . . . 69
§3.1.3. The Robin and Zaremba problems. . . . . . . . . . . . . . . . 71
§3.2. Consequences of variational principles . . . . . . . . . . . . . . . . . . . 75
§3.2.1. Domain monotonicity and Dirichlet-Neumann brack-
eting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
§3.2.2. Symmetry tricks . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
§3.2.3. Counting functions . . . . . . . . . . . . . . . . . . . . . . . . . . 89
§3.2.4. Inequalities between the Dirichlet and Neumann eigen-
values for Euclidean domains. . . . . . . . . . . . . . . . . . . . 91
§3.3. Weyl’s law and Pólya’s conjecture. . . . . . . . . . . . . . . . . . . . . . . 94
§3.3.1. Weyl’s law. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
§3.3.2. The two-term asymptotic formula and Weyl’s conjec-
ture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
§3.3.3. Pólya’s conjecture . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
§3.3.4. The Berezin–Li–Yau inequalities . . . . . . . . . . . . . . . . . 106
Chapter 4. Nodal geometry of eigenfunctions 111
§4.1. Courant’s nodal domain theorem . . . . . . . . . . . . . . . . . . . . . . 111
§4.1.1. Nodal domains and nodal sets . . . . . . . . . . . . . . . . . . . 111
§4.1.2. Courant’s theorem. . . . . . . . . . . . . . . . . . . . . . . . . . . 113
§4.1.3. Restriction of an eigenfunction to a nodal domain . . . . . 115
§4.1.4. Proof of Courant’s theorem . . . . . . . . . . . . . . . . . . . . 117
§4.1.5. Properties of subharmonic and harmonic functions . . . . 120
§4.1.6. Corollaries of Courant’s theorem. . . . . . . . . . . . . . . . . 124
§4.2. Density of nodal sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
§4.2.1. Geometric features of nodal sets. . . . . . . . . . . . . . . . . . 126
§4.2.2. A lower bound on the size of the nodal set in dimension
two . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 129
§4.3. Yau’s conjecture on the volume of nodal sets . . . . . . . . . . . . . .. 131
§4.3.1. Nodal volume and doubling index . . . . . . . . . . . . . . .. 131
§4.3.2. The Donnelly–Fefferman growth bound: a sketch of
the proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 134
§4.3.3. Distribution of doubling indices: a combinatorial ap-
proach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 141
§4.3.4. A polynomial upper bound: an overview. . . . . . . . . . .. 144
Contents v

§4.4. Nodal sets on surfaces and eigenvalue multiplicity bounds . . . . . 146

§4.4.1. Local structure of the nodal set . . . . . . . . . . . . . . . . . . 146
§4.4.2. Multiplicity bounds . . . . . . . . . . . . . . . . . . . . . . . . . . 148
Chapter 5. Eigenvalue inequalities 151
§5.1. The Faber–Krahn inequality. . . . . . . . . . . . . . . . . . . . . . . . . . 151
§5.1.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
§5.1.2. The co-area formula . . . . . . . . . . . . . . . . . . . . . . . . . . 153
§5.1.3. Symmetric decreasing rearrangement . . . . . . . . . . . . . . 155
§5.1.4. Proof of the Faber–Krahn inequality . . . . . . . . . . . . . . 156
§5.1.5. Applications of the Faber–Krahn inequality . . . . . . . . . 161
§5.2. Cheeger’s inequality and its applications. . . . . . . . . . . . . . . . . . 164
§5.2.1. Cheeger’s inequality. . . . . . . . . . . . . . . . . . . . . . . . . . 164
§5.2.2. Examples and further results . . . . . . . . . . . . . . . . . . . . 168
§5.2.3. The first eigenvalue and the inradius of planar domains . . 170
§5.3. Upper bounds for Laplace eigenvalues . . . . . . . . . . . . . . . . . . . 174
§5.3.1. The Szegő–Weinberger inequality . . . . . . . . . . . . . . . . 174
§5.3.2. Hersch’s theorem for the first eigenvalue on the sphere . . 180
§5.3.3. Topological upper bounds for eigenvalues on surfaces . . . 183
§5.4. Universal inequalities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
§5.4.1. The Payne–Pólya–Weinberger inequality . . . . . . . . . . . 188
§5.4.2. Abstract commutator identities . . . . . . . . . . . . . . . . . . 189
§5.4.3. Applications to Dirichlet eigenvalues . . . . . . . . . . . . . . 192
§5.4.4. Spectral prescription . . . . . . . . . . . . . . . . . . . . . . . . . 195
Chapter 6. Heat equation, spectral invariants, and isospectrality 197
§6.1. Heat equation and spectral invariants. . . . . . . . . . . . . . . . . . . . 197
§6.1.1. Heat kernel on a Riemannian manifold. . . . . . . . . . . . . 197
§6.1.2. Heat kernel asymptotics . . . . . . . . . . . . . . . . . . . . . . . 203
§6.1.3. Weyl’s law on a Riemannian manifold. . . . . . . . . . . . . . 205
§6.2. Isospectral manifolds and domains . . . . . . . . . . . . . . . . . . . . . 207
§6.2.1. Isospectrality. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
§6.2.2. Milnor’s example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
§6.2.3. Sunada’s construction . . . . . . . . . . . . . . . . . . . . . . . . 211
§6.2.4. Transplantation of eigenfunctions and mixed Dirichlet–
Neumann isospectrality . . . . . . . . . . . . . . . . . . . . . . . 215
§6.2.5. Isospectral drums . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
§6.2.6. Spectral rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map 227
§7.1. The Steklov eigenvalue problem . . . . . . . . . . . . . . . . . . . . . . . 227
§7.1.1. Definition and variational principle . . . . . . . . . . . . . . . 227
§7.1.2. The sloshing problem. Steklov eigenvalues of a square. . . 233
vi Contents

§7.1.3. Isoperimetric inequalities for the Steklov eigenvalues. . . . 237

§7.2. The Dirichlet-to-Neumann map and the boundary Laplacian . . . 244
§7.2.1. Weyl’s law for Steklov eigenvalues. . . . . . . . . . . . . . . . . 244
§7.2.2. The Hörmander–Pohozhaev identities . . . . . . . . . . . . . 245
§7.2.3. The Steklov spectrum and the spectrum of the bound-
ary Laplacian. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 247
§7.3. Steklov spectra on domains with corners. . . . . . . . . . . . . . . . . . 252
§7.3.1. Asymptotics of Steklov eigenvalues for curvilinear poly-
gons. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
§7.3.2. Sloping beach problems . . . . . . . . . . . . . . . . . . . . . . . 258
§7.3.3. Peters solutions of the Robin problem in an infinite
sector. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
§7.3.4. Quasimode construction for the Steklov problem in a
curvilinear polygon . . . . . . . . . . . . . . . . . . . . . . . . . . 261
§7.3.5. Asymptotics of the sloshing eigenvalues . . . . . . . . . . . . 265
§7.3.6. Inverse spectral problem for curvilinear polygons . . . . . . 267
§7.4. The Dirichlet-to-Neumann map for the Helmholtz equation . . . 268
§7.4.1. Definition and basic properties . . . . . . . . . . . . . . . . . . 268
§7.4.2. Dependence of the eigenvalues of the Dirichlet-to-Neumann
map on the parameter. . . . . . . . . . . . . . . . . . . . . . . . . 272
§7.4.3. The Dirichlet-to-Neumann map and the eigenvalue count-
ing functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
Appendix A. A short tutorial on numerical spectral geometry 281
§A.1. Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
§A.1.1. The Finite Element Method . . . . . . . . . . . . . . . . . . . . 281
§A.1.2. Solving spectral problems with Mathematica . . . . . . . . 284
§A.2. Learning FreeFEM by example . . . . . . . . . . . . . . . . . . . . . . . . 288
§A.2.1. The basics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 288
§A.2.2. The structure of a FreeFEM script: the Neumann prob-
lem in a rectangle . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
§A.2.3. Curvilinear boundaries and holes . . . . . . . . . . . . . . . . . 294
§A.2.4. The Dirichlet, Zaremba, and Robin problems . . . . . . . . 295
§A.2.5. The Laplace–Beltrami operator on manifolds . . . . . . . . 296
§A.2.6. The Steklov problem, the sloshing problem, and the
spectrum of the Dirichlet-to-Neumann map . . . . . . . . . 299
§A.3. List of downloadable scripts . . . . . . . . . . . . . . . . . . . . . . . . . . 300
Appendix B. Background definitions and notation 303
§B.1. Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
§B.2. Function spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
§B.3. Regularity of the boundary . . . . . . . . . . . . . . . . . . . . . . . . . . 306
Contents vii

List of margin portraits 307

Bibliography 311
Index 335
 Preface

Various distinct physical phenomena, such as wave propagation, heat diffusion, electron move-
ment in quantum physics, oscillations of fluid in a container, can be modelled mathematically
using the same differential operator — the Laplacian. Its spectral properties depend in a subtle
way on the geometry of the underlying object, e.g. a Euclidean domain or a Riemannian mani-
fold, on which the operator is defined. This dependence — or, rather, the interplay between the
geometry and the spectrum — is the main subject of spectral geometry.
The roots of spectral geometry go back to the famous experiments of the physicist Ernst
Chladni with vibrating plates in the late eighteenth – early nineteenth century, as well as to the in-
vestigations of Lord Rayleigh on the theory of sound some decades later. The celebrated question
of Mark Kac “Can one hear the shape of a drum?” motivated a lot of research in the second half of
the twentieth century and helped spectral geometry to emerge as a separate branch of geometric
analysis.
Modern spectral geometry is a rapidly developing area of mathematics, with close connec-
tions to other fields, such as differential geometry, mathematical physics, number theory, dynam-
ical systems and numerical analysis. It is a vast subject, and by no means this book pretends to
be comprehensive. Our goal was to write a textbook that can be used for a graduate or an ad-
vanced undergraduate course, starting from the basics but at the same time covering some of the
exciting recent developments in the area which can be explained without too many prerequisites.
The authors have taught such courses over the past few years at different locations, in particular
at the Université de Montréal and the Hebrew University of Jerusalem, and shorter courses at
the Universities of Cardiff and Reading, as well as at several summer schools and instructional
conferences, see e.g. [BouLev07]. The present book is based in part on our lecture notes.

1
 2 Preface

Acknowledgements
We gratefully acknowledge the influence of many earlier books on spectral geometry and related
subjects, by Courant and Hilbert [CouHil89], Berger, Gauduchon, and Mazet [BerGauMaz71],
Reed and Simon [ReeSim75], Bandle [Ban80], Chavel [Cha84], Bérard [Bér86], Davies [Dav89],
[Dav95], Schoen and Yau [SchYau94], Rosenberg [Ros97], Henrot [Hen06], Helffer [Hel13], and
Shubin [Shu20], to name just a few, as well as some of the more recent lecture notes by Lauge-
sen [Lau12], Canzani [Can13], Buhovsky [Buh16], Logunov and Malinnikova [LogMal20]. Of
course, the standard disclaimer is that the choice of the topics in this book reflects the personal
tastes and preferences of the authors.
Many people contributed to this book in different ways. It is a pleasure to thank our mentors,
Robert Brooks, Yakar Kannai, Victor Lidskii, Leonid Polterovich (to whom we are particularly
Robert Wolfe
Brooks thankful for encouraging this book project since its very early stages), and Dmitri Vassiliev, for
(1952—2002) introducing us to geometric spectral theory. Through the years, we have been also greatly influ-
enced by collaborations and innumerable helpful discussions with Michiel van den Berg, E. Brian
Davies, Lennie Friedlander, Nikolai Nadirashvili, Leonid Parnovski, and Mikhail Sodin, among
others.
While preparing the manuscript we used the notes taken by Simon St-Amant at a course given
by I.P. at the Université de Montréal. We are especially grateful to Matteo Capoferri, Philippe
Charron, Stefano Decio, Emily Dryden, Alexandre Girouard, Asma Hassannezhad, Mikhail Kar-
pukhin, Jean Lagacé, Antoine Métras, Nilima Nigam, and David Sher, who made a lot of insight-
ful comments and suggestions on the preliminary draft of the book.
We also thank Lyonell Boulton, Dorin Bucur, Simon Chandler-Wilde, Bruno Colbois, Dmit-
ry Faifman, Bernard Helffer, Antoine Henrot, Dmitry Jakobson, Alexander Logunov, Eugenia
Malinnikova, Marco Marletta, Egor Shelukhin, and Vukašin Stojisavljević for many useful dis-
cussions.
All the remaining errors, omissions, and inaccuracies are of course entirely ours.
At various stages of work on this book, M.L. has been visiting the Centre de recherches math-
Victor Borisovich
Lidskii ématiques in Montréal. Its hospitality is gratefully acknowledged.
(1924—2008) Last but not least, we would like to thank our families for their patience and support. Without
them this project would have never been accomplished.

Notes on typesetting1
The main text is typeset using EBGaramond2 font package. The mathematics is typeset using
Fourier–GUTenberg3 font package. Theorems, lemmata, definitions, etc., as well as proofs,
figures, and listings use tcolorbox4 package. Headers, footers, chapter and section titles, and the
table of contents are styled using titlesec and titletoc5 packages. The index was produced
1 The reader is of course welcome to either like or dislike it.
2 by Georg Duffner and Octavio Pardo, see CTAN
3 by Michel Bovani, see CTAN
4 by Thomas F. Sturm, see CTAN
5 by Javier Bezos, see CTAN
Preface 3

using imakeidx6 and idxlayout7 . Most images are created in Mathematica 138 . For the list
of margin portraits and their sources see page 307.

6 by Enrico Gregorio, see CTAN

7 by Thomas Titz, see CTAN
8 see Wolfram Mathematica
 Introduction

Overview

The central theme of the book is spectral geometry of the Laplace operator on bounded Euclidean
domains and compact Riemannian manifolds. Most of the time, we consider the classical Dirich-
let or Neumann boundary conditions, except for the last chapter, where instead of the spectral
parameter in the equation we look at the less explored Steklov problem with the spectral parame-
ter in the boundary conditions.
The main topics discussed in the book can be summarised as follows:

• spectral theorems;

• eigenvalue inequalities;

• spectral asymptotics;

• nodal geometry;

• isospectrality and spectral invariants.

To cover these subjects we use a variety of techniques, such as variational principles, elliptic
regularity, symmetrisation, conformal maps, harmonic analysis, heat equation methods. Through-
out the presentation we tried to keep a balance between the following principles:

• Focus on phenomena. For that reason, in many cases the proofs are given in the Euclidean
setting, with indications on how the argument can be extended to the Riemannian case.

• Avoid black boxes as much as possible. While it is often unfeasible to present all the details, we
at least tried to explain the main ideas behind the proofs.

• Keep generality reasonably wide to include most interesting examples. In particular, in the
Euclidean setting we mostly consider Lipschitz boundaries, whilst on manifolds we deal with
smooth Riemannian metrics.

The highlights of the book include:

5
6 Introduction

• Spectral theorems and elliptic regularity. In particular, we discuss in detail both interior and
boundary regularity of eigenfunctions.

• Weyl’s law for the eigenvalue counting function.

• Friedlander–Filonov inequalities between Dirichlet and Neumann eigenvalues.

• Polya’s conjecture for tiling domains and Berezin–Li–Yau–inequalities.

• Courant and Pleijel nodal domain theorems.

• Yau’s conjecture on the size of the nodal sets.

• Isoperimetric inequalities for eigenvalues: Faber–Krahn, Cheeger, Szegő–Weinberger, Hersch.

• Universal inequalities for eigenvalues.

• Heat trace asymptotics.

• Isospectrality and transplantation of eigenfunctions.

• Spectral geometry of the Steklov problem.

While many of these topics can be found in other books, having all these subjects under one
cover makes this book quite different from the others. At times, our exposition of classical re-
sults contains some features which have not been emphasised previously. For example, we prove
Courant’s nodal domain theorem for Dirichlet eigenfunctions without any regularity assump-
tions on the boundary. Moreover, some of the material is based on recent research and therefore
cannot be found in textbooks, such as the section on Yau’s conjecture and essentially the entire
chapter on the Steklov problem.

Plan of the book

The book is organised is follows.
In Chapter 1 we introduce our main hero, the Laplacian, and discuss several examples for
which its eigenvalues and eigenfunctions can be calculated explicitly.
In Chapter 2 we lay the foundations for the further material and explain the proofs of the
weak and the strong spectral theorems for the Laplacian. This chapter includes mini-crash courses
on the theory of self-adjoint unbounded linear operators, as well as on the Sobolev spaces and
elliptic regularity. Our emphasis is on presenting the main tools and ideas, such as the Friedrichs
extension, the a priori estimates and Nirenberg’s method of difference quotients, while referring
the reader interested in full details to the existing literature.
Chapter 3 is concerned with the variational principles for eigenvalues and their applications.
Apart from basic results such as domain monotonicity, Dirichlet-Neumann bracketing and Weyl’s
law, we prove the Friedlander–Filonov inequalities between Dirichlet and Neumann eigenvalues,
the Berezin–Li–Yau inequalities and Pólya’s conjecture for tiling domains.
Introduction 7

Chapter 4 focuses on the nodal geometry of eigenfunctions. We give a complete proof of

Courant’s nodal domain theorem, explaining some delicate issues arising for domains with non-
smooth boundary that have often been omitted in other sources. We also discuss Yau’s conjecture
on the volume of nodal sets, including recent breakthrough developments due to Logunov and
Malinnikova. In particular, we give a sketch of the proof of a polynomial upper bound on the size
of the nodal set. Some related topics, such as the density of the nodal set, and the lower bound
on the size of the nodal set in dimension two, are also presented. As an application of results on
the local structure of the nodal set we prove multiplicity bounds for eigenvalues on surfaces.
In Chapter 5 we collect various geometric eigenvalue inequalities, such as the Faber–Krahn
inequality, Cheeger’s inequality, the Szegő–Weinberger inequality, as well as Hersch’s inequality
and other isoperimetric inequalities of surfaces. The latter is an actively developing subject and
several recent advances are discussed in detail. This chapter also includes the universal inequalities,
as well as related commutator identities.
The heat equation and results on heat kernel asymptotics are presented in Chapter 6. As an
application, we prove Weyl’s law on Riemannian manifolds. The spectral invariants arising from
the heat asymptotics naturally lead us to the study of isospectrality. Some partial answers are given
to the question “Can one hear the shape of a drum?" mentioned above. We present Milnor’s
example of flat isospectral tori which has fascinating connections to the theory of modular forms,
and the celebrated Sunada construction of isospectral manifolds based on algebraic ideas. We also
describe a rather elementary but ingenious transplantation technique that yields isospectral but
not isometric planar domains. Some recent results on spectral rigidity are also discussed.
In the past decade, the study of the Steklov problem and of the Dirichlet-to-Neumann map
became one of the most active directions in spectral geometry. This is the subject of Chapter 7. We
define the Steklov spectrum and prove isoperimetric inequalities for Steklov eigenvalues. Using
the connection between the Dirichlet-to-Neumann map and the boundary Laplacian, we obtain
results on the asymptotics of the Steklov spectrum by means of the Hörmander–Pohozhaev iden-
tities and the Weyl’s law for the Laplacian on manifolds. We also provide a detailed exposition of
recent results on the asymptotics of sloshing eigenvalues as well as Steklov eigenvalues on curvilin-
ear polygons. Finally, we discuss the Dirichlet-to-Neumann map for the Helmholtz operator, and
use its properties to give another proof of the Friedlander–Filonov inequalities between Dirichlet
and Neumann eigenvalues originally presented in Chapter 3.
Appendix A contains a short introduction to numerical spectral geometry, which provides
the students with all the necessary tools for quick numerical calculation of eigenvalues and eigen-
functions of planar domains.
In Appendix B we collect some standard background definitions and notation which we use
throughout the book.
8 Introduction

Possible courses based on this book

The book is to a large extent self-contained and is accessible to students and researchers with basic
knowledge of PDEs, functional analysis, and differential geometry. We do not really require the
prior knowledge of the theory of distributions and Sobolev spaces and explain the main notions
we need. Throughout the book we often stay in the Euclidean setting, and, where necessary,
provide references for a reader unfamiliar with the fundamentals of Riemannian geometry. While
graduate students in mathematics are the main target audience for the book, it could also be used,
in parts, for teaching an advanced undergraduate course, as well as for both introductory and
advanced mini-courses.
In our experience, essentially the whole book with the exception of the most advanced sec-
tions (§§2.2, 4.3 and 7.2–7.4) can be covered in a one-semester course. There are various ways to
create shorter courses using the following diagram of dependencies.

Chapter 1

Chapter 2
§2.2

Chapter 3

Chapter 4 Chapter 5 Chapter 6 Chapter 7

§4.3 §§7.2–7.4
Appendix A (for numerical exercises)

Appendix B (background notation)

The diagram of chapter dependencies. The subsections in shaded boxes may be

omitted for all but advanced courses. The dashed arrows indicate that while there
is some dependency of material there, the corresponding chapters may be taught
separately from each other.

For example, one could teach the first three chapters only, or the first three chapters followed
by one of the chapters 4–7, with some minor additions and adjustments. Finally, the material
of each of the chapters 1–3 can be taught as an introductory level mini-course, and each of the
remaining chapters as a more advanced one.
Introduction 9

Last, but not least, the book contains many exercises! The more difficult ones are provided
with references and hints. A user-friendly tutorial on numerical spectral geometry presented in
Appendix A could also help teachers who would like to introduce a computational component
into their classes.

What is not in this book: some further reading

Spectral geometry is a vast subject, and by no means this book pretends to fully cover it. Below
we discuss some interesting and important topics for further reading.
In order to keep the prerequisites to a minimum, we focused on results that can be presented
without using pseudodifferential operators and microlocal analysis. As a consequence, apart from
nodal geometry, we did not explore much the properties of eigenfunctions. We refer to [Sog17]
and [Zwo12] for an exposition of results on asymptotic eigenfunction bounds, as well as questions
arising in the fascinating area of mathematical quantum chaos, such as Shnirelman’s quantum
ergodicity theorem.
Throughout the book, we have almost exclusively dealt with the case of bounded domains and
compact manifolds, for which the Laplace spectrum is discrete. A lot of interesting phenomena
occur in other geometric set-ups. We refer to [Bor16] and [DyaZwo19] for recent developments of
the spectral theory on infinite area hyperbolic spaces and the mathematical theory of resonances.
In this book, we focus on the Laplacian and the Dirichlet-to-Neumann map and do not touch
other important operators. A modern exposition of the spectral theory of Schrödinger opera-
tors, with a particular focus on the celebrated Lieb–Thirring inequalities (closely linked to the
Berezin–Li–Yau inequalities featured in Chapter 3), can be found in [FraLapWei22]. Many inter-
esting geometric questions arise in the study of the spectrum of the Dirac operator, and we refer to
[BerGetVer04, Fri00, Gin09] for further reading on this subject. Recent results on spectral geom-
etry of potential operators, which are related to the Dirichlet-to-Neumann map, can be found in
[RuzSadSur20]. A detailed introduction to the rich and actively developing theory of quantum
graphs, which makes a cameo appearance in §7.3 of this book, can be found in [BerKuc13].
 CHAPTER 1
Strings, drums, and the Laplacian
In this chapter, we introduce the Laplacian, both in the Euclidean
space and on a Riemannian manifold, and consider the eigenvalue
problems with the Dirichlet and Neumann boundary conditions.
We discuss the related models of vibrating strings and drums, and
consider a few examples in which spectral problems can be explicitly
solved.

§1.1. Basic examples

§1.1.1. The Laplace operator

In the Euclidean space Rd of dimension d with Cartesian coordinates x = (x 1 , . . . , x d ), let
d ∂2 f
∆f =
X
, (1.1.1)
j =1 ∂x 2j

where f = f (x 1 , . . . , x d ) is a twice differentiable function. Pierre-Simon de Laplace

(1749—1827)
Definition 1.1.1: The Laplacian
The operator −∆ is called the Laplace operator (or the Laplacian) in Rd .

Remark 1.1.2
There is no unique sign convention for ∆. In this book, we define ∆ by (1.1.1), that is in
the analyst’s sense; geometers often incorporate the minus sign into the definition of ∆.
(The authors have argued long and hard about which notation to adopt.) Additionally,
the term Laplacian may also be applied to the negative of our Laplacian.

11
 12 Chapter 1. Strings, drums, and the Laplacian

One can rewrite (1.1.1) as

∆ f = div ∇ f ,
where div denotes the divergence of a vector field, and ∇ is the gradient of a scalar function, see
§B.1. We will use this representation later on in order to define the Laplacian on a Riemannian
manifold.
The Laplace operator appears in major partial differential equations arising in mathematical
physics. Here are some examples; in all of them we set ∆ := ∆x , i. e. the operator acts only in the
x variable.

• Wave equation:
∂2U (t , x)
= ∆U (t , x).
∂t 2
Here U (t , x) denotes the displacement from the equilibrium of the vibrating object at the point
x ∈ Rd at time t .

• Heat (or diffusion) equation:

∂U (t , x)
= ∆U (t , x).
∂t
Here U (t , x) denotes the temperature of the object (or the density of the matter) at the point
x at time t .

• Laplace equation:
∆U (x) = 0.
The solutions of the Laplace equation are called harmonic functions. In hydrodynamics, the
velocity potential U (x) of an incompressible fluid flow is a solution of the Laplace equation.

• Poisson equation:
−∆U (x) = f (x).
In electrostatics, U (x) is interpreted as an electric potential corresponding to a given charge
distribution f .
Erwin Rudolf
Josef Alexander • Schrödinger equation:
Schrödinger ∂U (t , x)
i = −∆U (t , x),
(1887—1961) ∂t
where i2 = −1. In quantum mechanics, the solution U (t , x) of this equation is called the wave
function. Note that U (t , x) is complex-valued; the quantity |U (t , x)|2 describes the probability
density for a particle to be at the position x at time t .

Let us start with two simple real life examples, which are also among the most relevant ones
from the viewpoint of spectral geometry: the vibrating strings and drums.
§1.1. Basic examples 13

§1.1.2. Vibrating strings

Even if you never played a guitar yourself, you probably know that thicker guitar strings produce
lower sounds, and that pressing down on a string rises the pitch. These phenomena could be easily
explained using a mathematical model of a vibrating string, given by the one–dimensional wave
equation.
Consider a string of length l and uniform density ρ , fixed at both ends. Let U : R+ ×[0, l ] → R
be a function, whose value U (t , x) is equal to the deviation from the equilibrium of a transversally
vibrating string at the point x ∈ [0, l ] at the time t ∈ R+ (transversal vibrations mean that each
point of the string moves along the vertical line orthogonal to the equilibrium position). The
function U (t , x) satisfies the one-dimensional wave equation

U t t = a 2 ∆U = a 2U xx , (1.1.2)

where the constant a can be expressed in terms of the tension τ of the string and the density ρ :

τ/ρ.
p
a=

Since the string is attached at both ends, we impose the Dirichlet boundary conditions:

U (t , 0) = U (t , l ) = 0, t ∈ R+ . (1.1.3) Johann Peter Gustav

Lejeune Dirichlet
In order to find a solution of this equation we use the Fourier method. The first step is to separate (1805—1859)
the variables and to look for a solution in the form

U (t , x) = T (t )X (x).

This is a so-called standing wave. From the equation (1.1.2) we get,

T ′′ (t )X (x) = a 2 T (t )X ′′ (x),

and, since X (x) and T (t ) are not identically zero, we obtain

X ′′ (x) T ′′ (t )
= 2 = −λ,
X (x) a T (t )

where λ is some constant (the choice of the minus sign will become clear later). Indeed, the left- Jean-Baptiste
hand side of the equality does not depend on t , and the middle part is independent of x , so both Joseph Fourier
are equal to a constant. (1768—1830)
We now consider the equations for the functions X (x) and T (t ) separately.
Taking into account (1.1.3), we obtain a Sturm–Liouville eigenvalue problem for the function
X (x) with Dirichlet boundary conditions:
(
−X ′′ (x) = λX (x),
(1.1.4)
X (0) = X (l ) = 0.
 14 Chapter 1. Strings, drums, and the Laplacian

Definition 1.1.3
A non-trivial solution X (x) of the Sturm–Liouville problem (1.1.4) is called an eigenfunc-
tion corresponding to an eigenvalue λ.

Exercise 1.1.4
Show that the eigenvalues and eigenfunctions of the Sturm–Liouville problem (1.1.4) are
given by
³ πm ´2 ³ πm ´
λm = , X m (x) = sin x , m = 1, 2, . . . .
l l

Exercise 1.1.5
Jacques Charles
François Sturm Show that for all natural numbers k ̸= m ,
(1803—1855)
ˆl
X k (x)X m (x)dx = 0.
0

Resolving a similar Sturm–Liouville problem for T (t ) = Tm (t ) we obtain

³ aπm ´ ³ aπm ´
Tm (t ) = A m cos t + B m sin t ,
l l
Joseph Liouville
(1809—1882) where A m and B m are arbitrary constants. Taking a superposition of the standing waves Um (t , x) =
Tm (x)X m (x), we get a formal solution of the wave equation (1.1.2):
∞ ³
X ³ aπm ´ ³ aπm ´´ ³ πm ´
U (t , x) = A m cos t + B m sin t sin x . (1.1.5)
m=1 l l l

Exercise 1.1.6
Show that the constants A m and B m , m ∈ N, are uniquely determined by the initial con-
ditions u(0, x) = ϕ(x) (initial position), u t (0, x) = ψ(x) (initial velocity). Calculate A m
and B m using the Fourier decompositions of the functions ϕ and ψ.

We are now in a position to address the questions about sounds emitted by a guitar string
raised at the beginning of this section. As can be easily seen from (1.1.5), the natural frequencies
of the string are given by
p aπm
ωm = a λm = , m ∈ N. (1.1.6)
l
The frequency ω1 is called the principal frequency, or the fundamental tone of the string, and
the higher frequencies are called overtones. It follows immediately from (1.1.6) that the frequencies
§1.1. Basic examples 15

decrease as the length l increases: in other words, shorter strings produce higher notes. This is
precisely what we observe when pressing down on a guitar string (pressing down is essentially
a way to change the length of the vibrating part of the string). Recall now that the constant a
decreases as the density of a string increases. Therefore, the thicker the string is, the lower are the
sounds that it emits. Similarly, the higher is the tension of the string, the higher is the pitch.
The eigenfunctions X m (x) describe the shape of the pure vibration modes. In particular, one
may observe that for each m = 1, 2, . . . , the eigenfunction X m (x) has precisely m − 1 zeros on the
open interval (0, l ), see Figure 1.1. This fact has interesting higher–dimensional generalisations
that we will discuss later.

Figure 1.1: First four eigenfunctions of a fixed vibrating string

Exercise 1.1.7
The vibrations of a free string of length l are modelled by the equation (1.1.2) with Neu-
mann boundary conditions

U x (t , 0) = U x (t , l ) = 0, t ∈ R+ . (1.1.7)

Find the eigenfrequencies of a free vibrating string and compare them with the (Dirichlet)
eigenfrequencies given by (1.1.6).
Carl Gottfried Neumann
Let us explain the physical meaning of the Neumann condition (1.1.7). As follows from the (1832—1925)
model leading to the wave equation (1.1.2), the tension force acting at the point x is equal to τU x .
Free vibration means that the endpoints of the string experience no tension, and therefore at these
points U x must vanish.
16 Chapter 1. Strings, drums, and the Laplacian

Example 1.1.8
Consider the vibrations of a string whose ends are neither fixed nor free but joined to-
gether in a circular loop. If the length of the string is 2π, we arrive, after the separation of
variables, at the spectral problem
(
−X ′′ (x) = λX (x),
(1.1.8)
X (x) is 2π-periodic.

Looking for the values of λ for which (1.1.8) has a non-trivial solution, we obtain

λ0 = 0, X 0 (x) = 1,

and also eigenvalues m 2 , m ∈ N, for each of which there are two linearly independent
eigenfunctions X m,1 (x) = sin mx and X m,2 (x) = cos mx .

§1.1.3. Vibrating drums

Consider now a two-dimensional analogue of the problem discussed in the previous section.
Imagine a drum with a membrane (drumhead) shaped as a bounded domain Ω ⊂ R2 . The func-
tion
U (t , x, y) : R+ × Ω → R

describing the vibration of the drumhead satisfies the wave equation

(
U t t − a 2 ∆U = 0,
U |∂Ω = 0,

where the constant a depends on the physical characteristics of the membrane. Again, searching
for solutions in the form U (t , x, y) = T (t ) u(x, y), we get a familiar (ordinary) Sturm–Liouville
equation for T (t ) and a Dirichlet eigenvalue problem for the function u(x, y), that is the eigen-
value problem for the Laplacian in Ω,

−∆u = λu, (1.1.9)

subject to the Dirichlet condition

u|∂Ω = 0. (1.1.10)
We say, as in Definition 1.1.3, that λ is an eigenvalue of the Dirichlet problem (1.1.9)–(1.1.10) if this
problem has a non-trivial solution u(x, y).
Unlike (1.1.4), the problem (1.1.9)–(1.1.10) usually cannot be explicitly solved. However, for
certain geometries — for example, for a rectangle or for a disk — that could be done by using
once again the separation of variables (in this case, the spatial variables x and y ).
§1.1. Basic examples 17

Exercise 1.1.9
Let R a,b = (0, a) × (0, b) be a rectangle with sides a and b . Show that

k 2 m2
µ ¶
λD
k,m = π
2
+ , k, m = 1, 2, . . . , (1.1.11)
a2 b2

are the eigenvalues of the Dirichlet problem (1.1.9)–(1.1.10) on R a,b , and the corresponding
eigenfunctions are given by

D kπ mπ
u k,m (x, y) = sin x sin y. (1.1.12)
a b

Prove that these functions form an orthogonal basis in L 2 (R a,b ).

Remark 1.1.10
The separation of variables does not immediately imply that (1.1.11) and (1.1.12) provide all
eigenvalues and eigenfunctions of the Dirichlet problem (1.1.9) on a rectangle. This has to
be shown separately and, indeed, it follows from the fact that the set (1.1.12) forms a basis
in L 2 (R a,b ).
More generally, the fact that eigenfunctions of (1.1.9)–(1.1.10) in a bounded domain Ω
can be chosen to form a basis in L 2 (Ω) follows from the spectral theorems, see Chapter 2.

Definition 1.1.11
The multiplicity of an eigenvalue λ is the dimension of the corresponding eigenspace. If
the dimension is equal to one, the eigenvalue is called simple.

Exercise 1.1.12
2
Show that if ba 2 is irrational, then all the Dirichlet eigenvalues of a rectangle R a,b are sim-
ple.
2
Note that if ba 2 is rational, then the multiplicities of the Dirichlet eigenvalues of R a,b
can be arbitrary large. This follows from number-theoretic results on representation of
integers as binary quadratic forms. In the case of a square, the precise answer could be
found using the so-called sum of squares function, see [HarWri79, p. 241], and also Re-
mark 1.2.14 below. For example, if a = b = π, one can check that the eigenvalue λ = 52k−1 ,
k ∈ N, has multiplicity 2k .
18 Chapter 1. Strings, drums, and the Laplacian

Example 1.1.13
Since for an eigenvalue of multiplicity m we have an m -dimensional linear space of corre-
sponding eigenfunctions, particular eigenfunctions may look quite unlike each other, see
Figure 1.2.

Figure 1.2: Two eigenfunctions corresponding to the same

Dirichlet eigenvalue 85π2 of the unit square [0, 1]2 : on the
left, the eigenfunction sin(2πx) sin(9πy), and on the right, the
eigenfunction p15 (sin(2πx) sin(9πy) − sin(9πx) sin(2πy) −
sin(6πx) sin(7πy) + 2 sin(7πx) sin(6πy))

Along with the Dirichlet boundary condition u|∂Ω = 0 corresponding to a membrane with
a fixed boundary, one may consider the vibration of a free membrane. This problem gives rise
to the Neumann boundary condition, which can be viewed as an appropriate generalisation of
(1.1.7):
∂n u = 0, (1.1.13)
where from now on we set
∂n u := 〈(∇u)|∂Ω , n〉
to denote the normal derivative of u . Here n is the exterior unit normal to the boundary ∂Ω, and
〈·, ·〉 stands for the standard vector inner product in Rd (or Cd ), see §B.1. It is clear that in order
for the Neumann condition (1.1.13) to be well defined, certain regularity of the boundary has to
be assumed. For instance, if one assumes the boundary to be Lipschitz (i.e., locally representable
as a graph of a Lipschitz function, see §B.3 for the definition), the normal derivative is well de-
fined at almost every point of the boundary. More general conditions under which the Neumann
problem is well defined will be discussed later.

Rudolf Otto Exercise 1.1.14

Sigismund
Lipschitz Show that
k 2 m2
µ ¶
(1832 –1903) λN
k,m = π
2
+ , k, m ∈ N0 . . . , (1.1.14)
a2 b2
are the eigenvalues of the Neumann problem (1.1.9), (1.1.13) in the rectangle R a,b , with the
§1.1. Basic examples 19

corresponding eigenfunctions

N kπ mπ
u k,m (x, y) = cos x cos y.
a b

Note that the indices k, m of the Neumann eigenvalues may take the value zero, while in the
Dirichlet case they start with one. In particular, the lowest Neumann eigenvalue is zero and the
corresponding eigenfunction is a constant. In fact, this is true for any bounded domain Ω on
which the Neumann problem is well defined.

Exercise 1.1.15
Using the formula (1.1.11) for the eigenvalues of the Laplacian in an arbitrary rectangle
with Dirichlet boundary conditions, find which rectangle minimises the first Dirichlet
eigenvalue among all rectangles of fixed area. Similarly, using (1.1.14), find which rectangle
of a fixed area maximises the first nonzero Neumann eigenvalue. What happens if we
interchange minimisation and maximisation in these questions?

Exercise 1.1.16
Compute the Dirichlet and Neumann eigenvalues and eigenfunctions of a rectangular
box in Rd .

Example 1.1.17
Let us describe the eigenvalues and eigenfunctions of the Dirichlet and Neumann prob-
lems in the unit disk D. Switching to polar coordinates (r, ϕ), using the standard expres-
sion
∂2 1 ∂ 1 ∂2
∆= + +
∂r 2 r ∂r r 2 ∂ϕ2
for the Laplacian in planar polar coordinates, and looking for solutions of (1.1.9) in the
form
+∞
u(r, ϕ) = u m (r )eimϕ ,
X
m=−∞
we arrive at the equations
m2
µ ¶
′′ 1 ′
um (r ) + u m (r ) + λ − 2 u m (r ) = 0 (1.1.15)
r r
for unknown functions u m .
The equations (1.1.15) are closely related to the Bessel equation
m2
µ ¶
′′ 1 ′
y (r ) + y (r ) + 1 − 2 y(r ) = 0. (1.1.16)
r r
 20 Chapter 1. Strings, drums, and the Laplacian

For m ∈ N0 , equation (1.1.16) possesses, up to a multiplicative constant, only one solution

regular at r = 0. A specific choice of that constant corresponds to the solution defined via
a power series
∞
³ r ´m X (−1)k ³ r ´2k
J m (r ) = , (1.1.17)
2 k=0 k!Γ(m + k + 1) 2

which is called the Bessel function of the first kind of order m . In fact, Bessel functions
Friedrich Wilhelm J ν (r ) can be defined in a similar manner for ν ∈ R by taking m = ν in (1.1.17), see [Wat95,
Bessel Chapter 3] for details, and it follows that J −m (r ) = (−1)m J m (r ) for m ∈ N. We refer to
(1784—1846) [Wat95] for a complete treatment of the theory of Bessel functions, and recall only some
facts which we will use in the sequel.
One can show that Bessel functions have infinitely many real zeros. Denote by j m,k the
′
k th positive zero of the m th Bessel function J m (r ), and by j m,k the k th positive zero of
′ ′
the derivative J m (r ) (with the exception j 0,1 = 0 for the first zero of J 0′ (r ), see [DLMF22,
§10.21(i)]), cf. Figure 1.3.
Returning now to the equations (1.1.15) and comparing to (1.1.16), one can easily
see that the ³regular´ solutions of (1.1.15) are given, modulo a multiplicative constant, by
p
u m (r ) = J m λr .
³p ´
Imposing the Dirichlet condition (1.1.10) now implies u m (1) = J m λ = 0, and
therefore the Dirichlet eigenvalues of the unit disk D are given by
2
j m,k , m ∈ N0 , k ∈ N.

For m > 0, the eigenvalues should be repeated with multiplicity two and the correspond-
ing linearly independent eigenfunctions can be chosen either as
¡ ¢ ¡ ¢
J m j m,k r sin mϕ, J m j m,k r cos mϕ. (1.1.18)
¡ ¢
For m = 0 each eigenvalue is simple, with the corresponding eigenfunction J 0 j 0,k r be-
ing radially symmetric. To ensure that we have found all the eigenfunctions we also need
to prove that they form a basis in L 2 (D) as discussed in Remark 1.1.10; this is not entirely
trivial and follows from the Sturm–Liouville theory, see [CouHil89, §V.5.5]. ³ ´
p ′ p
Similarly, imposing the Neumann condition (1.1.13) implies u m′
(1) = λJ m λ = 0,
′2
and therefore the Neumann eigenvalues of the unit disk D are given by j m,k , m ∈ N0 ,
k ∈ N, where for m > 0 the eigenvalues should be repeated with multiplicity two. The
′2
eigenfunctions corresponding to j m,k are given by either
³ ´ ³ ´
′ ′
J m j m,k r sin mϕ, J m j m,k r cos mϕ (1.1.19)

(as before we have only one eigenfunction for m = 0).

§1.1. Basic examples 21

Finally, let us also note that the zeros of Bessel functions of different orders (respec-
tively, of their derivatives) never coincide, and therefore there are no “accidental” mul-
tiplicities in the Dirichlet (respectively, Neumann) spectrum. In the Dirichlet case this
follows from the proof of the celebrated Bourget hypothesis (1866) found by C. L. Siegel
back in 1929, see [Sie29] and also [Wat95, pp. 484–485]. Essentially, Siegel proved a rather
deep number-theoretic result: if x ̸= 0 is an algebraic number, J m (x) is transcendental. At
the same time, using relations between Bessel functions of different orders, one can show
that if J m and J k share a common zero, it has to be an algebraic number. Therefore, the
only possible common zero may be x = 0. The Neumann analogue of this result is also
known, see [HelSun16].

Figure 1.3: The graphs of some Bessel functions, with zeros of J 0 (x) and J 1 (x)
′
marked. Note that j 1,k = j 0,k+1 for k ∈ N.

Exercise 1.1.18
Using integrals [DLMF22, formulae 10.22.37–38], check the orthogonality in L 2 (D) of the
eigenfunctions (1.1.18) or (1.1.19) in either the Dirichlet or Neumann case. This is just an
illustration of a much more general phenomenon which we will encounter later in Theo-
rems 2.1.20, 2.1.36, and 2.2.21: the eigenfunctions of the Dirichlet or Neumann Laplacian
can always be chosen to form an orthonormal basis in L 2 .
 22 Chapter 1. Strings, drums, and the Laplacian

Remark 1.1.19
In the same manner, the eigenvalues of the Dirichlet and Neumann Laplacians on circular
sectors and annuli can be expressed in terms of zeros of some Bessel functions or their
combinations, or of their derivatives. Similarly, the variables separate for ellipses, and the
eigenvalues can be expressed in terms of zeros of some special functions, see [GreNgu13]
and [KutSig84].

Remark 1.1.20
Apart from the Dirichlet and Neumann boundary conditions, there exist other types of
self-adjoint boundary conditions, for example the Robin ones or Zaremba (mixed) ones,
which we discuss later in §3.1.3. The Robin conditions arise, for example, when the bound-
ary is neither free nor fixed, but attached by a spring or some elastic material. Dirichlet,
Neumann and Robin conditions have also other physical interpretations, notably in terms
of the heat equation, see [Str07] for further details.

§1.2. The Laplacian on a Riemannian manifold

§1.2.1. The Laplace–Beltrami operator

In this section we use various basic notions from Riemannian geometry which can be found in
standard textbooks. In particular, lecture notes [Bur98] contain a concise and clear exposition of
essentially everything that is needed.
Consider a smooth closed (that is, compact without boundary) manifold M of dimension
dim M = d endowed with the Riemannian metric g = {g i j }, i , j = 1, . . . , d .
Georg Friedrich For any differentiable function f on M one can define the gradient ∇ f : it is a vector field,
Bernhard Riemann
such that for any p ∈ M and for any vector ξ ∈ T p M the following identity holds,
(1826–1866)
∇ f , ξ g = d f p (ξ) =: ξ f ,
­ ®
(1.2.1)

where 〈·, ·〉g is a scalar product on T p M defined by the Riemannian metric; we will usually omit
the subscript g . We say that ξ f is the directional derivative of the function f in the direction of
the vector ξ at the point p . It is easy to check that for the Euclidean space (1.2.1) yields the usual
definition of the gradient.
Let us now introduce the divergence div X of a vector field X on a Riemannian manifold. Let
dVg be the volume density on (M , g ). In local coordinates x 1 , . . . , x d it takes the form
q
dVg = det g dx 1 dx 2 . . . dx d .

We will sometimes write this as

q
dV = dVg = det g dx
§1.2. The Laplacian on a Riemannian manifold 23

for brevity. Given a smooth vector field X , one can define div X as a smooth function on M
satisfying the identity ˆ ˆ
­ ®
f div X dVg = − ∇ f , X dVg (1.2.2)
M M

for all f ∈ C 1 (M ). To verify that the divergence exists, we note that using a partition of unity it
suffices to check (1.2.2) for functions f supported in a coordinate chart, which is done below. We
refer to [Ros97, §1.2.3] for a discussion concerning this approach.
Let us calculate the gradient and the divergence in³ local coordinates
´ (x 1 , . . . , x d ). The corre-
∂ ∂
sponding basis in the tangent bundle T M is given by ∂x1 , . . . , ∂xd satisfying

∂ ∂
¿ À
, = gi j . (1.2.3)
∂x i ∂x j

The gradient ∇ f in this basis is given by

d ∂
c j (x)
X
∇f = (1.2.4)
j =1 ∂x j

for some coefficients c j (x). Applying formula (1.2.1) we get

* +
d d ∂ ∂ ∂ ∂f
µ ¶
j
cj
X X
c (x)g j i = , = df = .
j =1 j =1 ∂x j ∂x i ∂x i ∂x i

Applying the inverse matrix {g i j } and substituting the values of c j into (1.2.4) we obtain
d ∂f ∂
gi j
X
∇f = . (1.2.5)
i , j =1 ∂x i ∂x j

Let us now calculate the divergence. Let f be a differentiable function compactly supported
in a coordinate chart. Applying formula (1.2.2) to a vector field X = a (x), . . . , a d (x) and sub-
¡ 1 ¢

stituting (1.2.5) in the right-hand side we obtain

ˆ q
f div X det g dx 1 . . . dx d
M
ˆ * +
d
ij ∂f ∂ X d ∂ q
ai
X
=− g , det g dx 1 . . . dx d
i , j =1 ∂x i ∂x j i =1 ∂x i
M
ˆ d ∂f q (1.2.6)
a i det g dx 1 . . . dx d
X
=−
i =1 ∂x i
M
ˆ
X d ∂ ³ iq ´
= f a det g dx 1 . . . dx d .
i =1 ∂x i
M
 24 Chapter 1. Strings, drums, and the Laplacian

The second equality follows from (1.2.3), and the last equality is a result of the integration by parts.
Since formula (1.2.6) holds for any such function f , comparing its left- and right-hand sides we
get
1 X d ∂ ³ iq ´
div X = p a det g . (1.2.7)
det g i =1 ∂x i
Recall that for a vector field X in the Euclidean space Rd ,
X d ∂a i
div X = . (1.2.8)
i =1 ∂x i

It is easy to check that (1.2.7) agrees with (1.2.8) in this case.

Remark 1.2.1: Definitions of the divergence

There are several equivalent ways to define the divergence. Note that the right-hand side
of (1.2.8) can be represented as the trace of the operator ξ 7→ ξX := (ξa 1 , . . . , ξa d ) acting
on vector fields. On a Riemannian manifold, the analogue of the directional derivative
ξX is the covariant derivative ∇ ξ X , where ∇ denotes the Levi–Civita connection. Thus,
a standard way to define the divergence in Riemannian geometry is

div X = trace ξ 7→ ∇ ξ X ,
£ ¤
(1.2.9)

see, for example, [Bur98, §2.2] or [Cha84, §I.1].

On an orientable manifold one can also define the divergencepin a coordinate-free way
using differential forms, see [BerGauMaz71, §II.G.I]. Let ωg = det g dx 1 dx 2 . . . dx d =
dVg be the volume form corresponding to the Riemannian metric g on M . One can show
(see, for instance, [Pet06, Corollary 46]) that the Lie derivative of ωg in the direction of a
vector field X is given by
L X (ωg ) = (div X )ωg . (1.2.10)
This formula explains the meaning of the term divergence: it measures the rate of expan-
sion of the volume element as it flows along the vector field X .

Exercise 1.2.2
Show that formulas (1.2.9) and (1.2.10) yield the same expression (1.2.7) for the divergence
in local coordinates. See [Cha84, §I.1] and [Ros97, §1.2.3] for a solution.

Let us now state the main definition of this subsection.

Definition 1.2.3
Eugenio Beltrami The operator −∆ := − div ∇ defined on smooth functions is called the Laplacian (or the
(1835–1900) Laplace–Beltrami operator) on the manifold (M , g ). We will sometimes write it as −∆g =
§1.2. The Laplacian on a Riemannian manifold 25

−∆M to distinguish a particular manifold or metric.

Combining the formulas (1.2.5) and (1.2.7) we obtain the following expression for the Lapla-
cian:
d ∂ ∂f
µ q ¶
1 X ij
−∆ f = − p g det g . (1.2.11)
det g i , j =1 ∂x i ∂x j

Example 1.2.4
Let g i j = δi j , where δi j is the Kronecker symbol. Then the metric is flat and the Lapla-
cian takes the form

∂f
µ
∂f
¶
X d ∂2 f
−∆ f = − div ,..., =− ,
∂x 1 ∂x n i =1 ∂x i
2

and we recover the usual definition (1.1.1) of the Laplace operator in the Euclidean space.

Exercise 1.2.5
Recall that given two Riemannian manifolds (M , g ) and (N , h), a diffeomorphism F :
(M , g ) → (N , h) is called an isometry if it preserves the Riemannian metric, i.e. F ∗ h =
g , where F ∗ h denotes the pullback metric, see, for example [BerGauMaz71, Definition
A.2]. Using the invariance properties of the divergence and the gradient, show that the
Laplace operator commutes with isometries: −∆g (u ◦ F ) = (−∆h u) ◦ F for any function
u ∈ C ∞ (N ).

Exercise 1.2.6
Given u, v ∈ C ∞ (M ), show that

∆(uv) = v∆u + 2〈∇u, ∇v〉g + u∆v.

Example 1.2.7
Suppose that the Riemannian metric in local coordinates (x, y) on a surface is given by
ds 2 = h(x, y)(dx 2 + dy 2 ), where h(x, y) > 0. Such coordinates are called isothermal and
they locally exist on any surface, see [Spi88, Addendum 1, Chapter 9]. Show that the Lapla-
cian in isothermal coordinates has the form
µ 2
∂ ∂2
¶
1
−∆ = − + .
h(x, y) ∂x 2 ∂y 2
26 Chapter 1. Strings, drums, and the Laplacian

Remark 1.2.8: Manifolds with boundary

In what follows, we will also consider compact Riemannian manifolds M with bound-
ary ∂M ̸= ;. Note that in contrast to domains, which are open sets, by definition
∂M ⊂ M . Somewhat abusing notation, when talking about differential expressions or
function spaces on a Riemannian manifold M with boundary, we always have in mind
the interior of M , that is M \ ∂M , without indicating this explicitly. Let us also mention
that the definition of the divergence given above has to be adjusted accordingly in case of
a manifold with boundary: the equality (1.2.2) should hold for all f ∈ C 01 (M ), where by
our convention C 01 (M ) := C 01 (M \ ∂M ).

§1.2.2. The Laplacian on a flat torus

Consider a two-dimensional flat square torus T2a = R2 /(a Z)2 . Separating variables, and using Ex-
2π i 〈x,m〉
ample 1.1.8, we can find its eigenfunctions using complex notation: they are of the form e a ,
where x = (x 1 , x 2 ) ∈ T2a , and m = (m1 , m2 ) ∈ Z2 is a vector with integer coordinates. The eigen-
2
values are given by λm1 ,m2 = 4π a2
(m 12 + m 22 ). In particular, we have a constant eigenfunction
coming from the vector m = (0, 0) and corresponding to the eigenvalue zero. The first nonzero
2
eigenvalue λ1 = 4π a2
is of multiplicity four, and comes from the eigenfunctions with m = (±1, 0)
and m = (0, ±1). The corresponding eigenfunctions may be chosen to be real as

2πx 1 2πx 1 2πx 2 2πx 2

cos , sin , cos , sin .
a a a a

Numerical Exercise 1.2.9

Show that the multiplicity of an eigenvalue λ ∈ N of the torus T22π is equal to the sum of
squares function
r 2 (λ) := # (m 1 , m 2 ) ∈ Z2 : λ = m 12 + m 22 ,
© ª
(1.2.12)
cf. Exercise 1.1.12. Use this to compile a table of all the distinct eigenvalues of T22π less than
2, 500 together with their multiplicities.

Exercise 1.2.10
Calculate the eigenvalues of the Laplacian on a flat rectangular d -dimensional torus

Td(a1 ,...ad ) = R/(a 1 Z) × · · · × R/(a d Z),

using separation of variables and the spectrum of the Laplacian on a circle from Example
1.1.8.
§1.2. The Laplacian on a Riemannian manifold 27

Exercise 1.2.11
Find the eigenvalues and eigenfunctions of an arbitrary flat d -dimensional torus TdΓ =
Rd /Γ, where Γ is an arbitrary lattice in Rd . (You can find the answer in [Cha84, §II.2],
[BerGauMaz71, §III.B.1], [Can13, §5.2].)

A flat torus is a rare example of a manifold for which the eigenvalues and eigenfunctions can
be calculated explicitly. However, even in this case, some basic questions regarding the properties
of eigenvalues turn out to be very difficult.
Let us introduce, for a closed manifold M , the counting function of the eigenvalues of the
Laplace–Beltrami operator on M ,

N M (λ) = N (λ) := #{ j : λ j (M ) ≤ λ}.

Each eigenvalue is counted with its multiplicity. The behaviour of the function N M (λ) for large
values of λ describes the asymptotic distribution of eigenvalues as λ → +∞. Understanding the
properties of the counting function is one of the fundamental questions in spectral geometry.
Let us estimate N (λ) := NT2a (λ) for a flat square torus. Each eigenvalue

4π2 ¡ 2
λm1 ,m2 = m 1 + m 22
¢
a 2

corresponds to a point with integer coordinates (m1 , m2 ) on the plane, and we are counting the
number
G (ρ) := # (m 1 , m 2 ) ∈ Z2 : m 12 + m 22 ≤ ρ 2
© ª

p
of such points inside a circle of radius ρ := a2πλ : we have
à p !
a λ
NT2a (λ) = G . (1.2.13)
2π

Clearly, an approximate number of integer points inside the circle is given by the area of the circle.
Therefore, in this case
à p !
a λ a2λ Area(T2a )λ
N (λ) = G = + R(λ) = + R(λ), (1.2.14)
2π 4π 4π

where R(λ) = o(λ) as λ → ∞. Note the appearance of area in this asymptotic formula — as we
will see later, this is not a coincidence. The asymptotic formula (1.2.14) for the counting function
of the torus is known as Weyl’s law, see §3.3.1.
What more can be said about the size of the remainder R(λ)?
 28 Chapter 1. Strings, drums, and the Laplacian

Lemma 1.2.12
The remainder in Weyl’s law (1.2.14) on a square torus satisfies the estimate
p
R(λ) = O( λ) as λ → +∞.

Proof
For simplicity, set a = 2π; the result would follow for an arbitrary a by rescaling, see Exer-
cise 2.1.42. Let us identify each unit square with integer coordinates in the plane with its
left bottom corner (m, n). Then if m 2 + n 2 < pλ thepwhole square (corresponding to that
corner) is contained insidep p of radius λ + 2, see Figure 1.4.
the disk
Therefore, N (λ) < π( λp+ 2)p2 . Similarly, if the square has a nontrivial intersection
2 2
with the open disk of radius λ − 2,p then mp + n < λ. Note that the union p ofpsuch
squares fully covers the disk of radius λ − 2, and therefore N (λ) > π( λ − 2)2 .
Johann Carl Friedrich
Combining the two bounds on N (λ) we get
Gauss p
(1777–1855) |N (λ) − πλ| ≤ 2π 2 λ + 2π,

which implies the statement of the Lemma.

This result was known to C. F. Gauss, and the problem of counting the number G (ρ) of
integer points inside a disk of radius ρ is called Gauss’s circle problem. However, the estimate
given by Lemma 1.2.12 is quite far from the optimal one.

Conjecture 1.2.13
For any ε > 0, we have R(λ) = O(λ1/4+ε ) as λ → +∞.

Godfrey Harold This conjecture is due to G. H. Hardy (1916) and has remained wide open for more than a
Hardy
century. It is one of the most famous open problems in analytic number theory. It is known that
(1877–1947)
without ε in the exponent the conjecture is false — this follows from a quite nontrivial lower
bound due to Hardy and E. Landau. It was shown by G. Voronoi (1903), W. Sierpiński (1906)
and J. G. van der Corput (1923) that the upper bound holds with the exponent 13 . At present,
the best upper bound for R(λ) is due to J. Bourgain and N. Watt [BouWat17] with the exponent
approximately equal to 0.3137.

Remark 1.2.14
There is a surprising link between the eigenvalue counting function for a flat square torus
and the Bessel functions which appear in the spectral problems in the disk, see Example
1.1.17. Consider, once more, the torus T22π . Its eigenvalue counting function NT22π (λ)
Jean Bourgain coincides with the disk lattice point counting function G (λ) by (1.2.13). Consider, for an
(1954–2018)
§1.2. The Laplacian on a Riemannian manifold 29

Figure 1.4: Estimating the number of integer

p
points in a disk.
p
p
The radii of the three concentric circles are λ − 2, λ, and
p p
λ + 2.

integer m ≥ 0, the sum of squares function defined by (1.2.12). Then

2
⌊ρ
X⌋
G (ρ) = r 2 (m),
m=0

where ⌊·⌋ denotes the integer part. The function G (ρ) experiences a jump whenever ρ 2 is
an integer with r 2 (ρ 2 ) > 0. The identity due to Hardy [Har15] (in some form suggested
by S. Ramanujan) is then

r 2 (ρ 2 ) ∞ r (n) ¡
2 p ¢
G (ρ) − = πρ 2 + ρ
X
p J 1 2πρ n ,
2 n=1 n
Srinivasa
thus bringing the Bessel function J 1 into play, see also [BerDKZ18] for some historical Ramanujan
remarks and generalisations involving other Bessel functions. (1887–1920)
30 Chapter 1. Strings, drums, and the Laplacian

§1.2.3. The Laplace operator on spheres

This section is based on the material that can be found in [BerGauMaz71, §III.C.1], [Shu01, §III.22],
[Cha84, §II.4], [AxlBouWad01, Chapter 5].
Let (ξ1 , . . . , ξd ) be local coordinates on the unit sphere Sd ⊂ Rd +1 centred at the origin.
Consider the corresponding spherical coordinates (r, ξ1 , . . . , ξd ) defined in some open cone in
Rd +1 , where r > 0 is the radial variable. The standard Euclidean coordinates can be expressed as
x i = r ϕi (ξ1 , . . . , ξd ), i = 1, . . . , d +1, where ϕi , i = 1, . . . , d +1, are smooth functions parametris-
ing the unit sphere. Given a function f ∈ C ∞ (Rd +1 ), we obtain using the chain rule:

∂ f dX +1 ∂ f ∂x
i
dX +1 ∂f
= = ϕi ,
∂r i =1 ∂x i ∂r i =1 ∂x i
(1.2.15)
∂f +1 ∂ f ∂x
dX
i
dX +1 ∂ϕ ∂ f
i
= =r , j = 1, . . . , d .
∂ξ j i =1 ∂x i ∂ξ j i =1 ∂ξ j ∂x i
³ ´
∂ ∂ ∂
Consider the basis ∂r , ∂ξ1 , . . . , , ∂ξd in the tangent space T x Rd . Then formulas (1.2.15) imply

∂ ∂ dX
+1
¿ À
, = ϕ2k = 1,
∂r ∂r g Rd +1 k=1

∂ ∂ dX
+1 ∂ϕk
¿ À
, =r ϕk = 0,
∂r ∂ξ j g Rd +1 k=1 ∂ξ j

∂ ∂ +1 ∂ϕ
dX ∂ϕk ∂ ∂
¿ À ¿ À
2 k 2
, =r =r , ,
∂ξi ∂ξ j g Rd +1 k=1 ∂ξi ∂ξ j ∂ξi ∂ξ j g
Sd

where g Sd denotes the standard round metric on the sphere Sd , that is, the metric induced by
the Euclidean metric g Rd +1 . Note that the last equality on the first line is simply the equation of
the unit sphere; differentiating it with respect to ξ j we obtain the last equality on the second line.
In view of the formulas above, the Euclidean metric in spherical coordinates (r, ξ1 , . . . , ξd ) is
given by
µ ¶
1 0
g Rd +1 = .
0 r 2 g Sd
Therefore, applying formula (1.2.11) for the Laplace operator we obtain

∂2 d ∂ 1
∆g Rd +1 = + + 2 ∆g Sd . (1.2.16)
∂r 2 r ∂r r

Let P m be the space of homogeneous polynomials in Rd +1 of degree m . By definition, P ∈

P m if and only if P = r m · P |Sd . In particular,

∂P ∂2 P
= mr m−1 · P |Sd , = m(m − 1)r m−2 · P |Sd . (1.2.17)
∂r ∂r 2
§1.2. The Laplacian on a Riemannian manifold 31

We will denote by
P
fm := P m | d = P | d : P ∈ P m
© ª
S S

the restriction of P m to the sphere Sd .

Let n o
H m := P ∈ P m : ∆g Rd +1 P = 0

be the space of all harmonic homogeneous polynomials of degree m , and let

H
fm := H m | d = P | d : P ∈ H m
© ª
S S

be the space of their restrictions to the sphere Sd . It is easy to check that the spaces H m and H
fm
are isomorphic: indeed, the restriction map H m → H fm has an inverse given by

Pe 7→ r m Pe. (1.2.18)

Moreover, applying the left- and the right-hand sides of (1.2.16) to r m Pe and taking into account
(1.2.17) we obtain:
³ ´
0 = r m−2 −∆g Sd Pe − m(d + m − 1)Pe ,

which immediately implies that Pe is an eigenfunction of the Laplacian on the sphere with the
eigenvalue m(d + m − 1). In other words, we have proved the following

Proposition 1.2.15
Any element of the space Hfm is an eigenfunction of the Laplacian on the sphere corre-
sponding to the eigenvalue λ = m(d + m − 1).

The space H fm of such eigenfunctions is called the space of spherical harmonics of degree m .
Let us now calculate the multiplicities of the eigenvalues m(d + m − 1), m ∈ N0 , and show that
there are no other eigenvalues of the Laplacian on the sphere.

Theorem 1.2.16
The eigenvalues of the Laplace operator on the standard sphere Sd are given by m(d +
m − 1), m ∈ N0 , and the corresponding eigenspaces coincide with H
fm . The multiplicity
of the eigenvalue λ = m(d + m − 1) is equal to
Ã! Ã !
d + m d + m − 2
κd ,m := dim H
fm = − . (1.2.19)
d d

In order to prove this theorem we use the following proposition.

32 Chapter 1. Strings, drums, and the Laplacian

Proposition 1.2.17
For any m ≥ 0, the following decomposition of P m into a direct sum holds:

P m = H m ⊕ r 2 P m−2 .

Here and further on we assume that P m = {0} if m < 0.

Proof
We prove the statement by induction in m . For m = 0, 1 the result is trivially true. Assume
that it is true for all l < m and let us show that it holds for l = m . First, let us show that

H m ∩ r 2 P m−2 = {0}. (1.2.20)

Indeed, suppose there exists P ∈ H m ∩ r 2 P m−2 . Consider its restriction on the sphere
Pe ∈ Hfm ∩ P fm−2 . Note that P
g m is isomorphic to P m , with the inverse to the restriction
map given by the same formula as (1.2.18).
As we have already shown, the space H fm is contained in the eigenspace of the Lapla-
cian corresponding to the eigenvalue λ = m(m + d − 1). At the same time, by induction
hypothesis, the space P
fm−2 could be represented as a direct sum of certain spaces H fj , and
for all of them j < m . Using integration by parts it is easy to show that Laplace eigen-
functions corresponding to distinct eigenvalues are orthogonal in L 2 (Sd ). Therefore, we
conclude that Pe ≡ 0. Since P = r m Pe by (1.2.18), we obtain P ≡ 0, which implies (1.2.20).
We have thus shown that P m ⊃ H m ⊕ r 2 P m−2 , and therefore

dim H m ≤ dim P m − dim P m−2 . (1.2.21)

At the same time, consider the Laplace operator as a map ∆ : P m → P m−2 . Its kernel is
precisely H m , and therefore

dim H m ≥ dim P m − dim P m−2 . (1.2.22)

Combining (1.2.21) and (1.2.22) we conclude that

dim H m = dim P m − dim P m−2 . (1.2.23)

It then follows that the map ∆ : P m → P m−2 is surjective, and by the dimension count
P m = H m ⊕ r 2 P m−2 , which completes the proof of the proposition.
§1.2. The Laplacian on a Riemannian manifold 33

Proof of Theorem 1.2.16

Let us show first that
∞
L 2 (Sd ) = H
M
fm . (1.2.24)
m=1

Indeed, applying inductively (1.2.20) and taking restriction to the sphere we get
∞ ∞
H P
M M
fm = fm .
m=1 m=1

Note that the direct sum on the right is isomorphic to the space of all polynomials in Rd +1
restricted to Sd . Formula (1.2.24) then holds since polynomials are dense in L 2 (Rd +1 ).
Hence, the first assertion of Theorem 1.2.16 follows from Proposition 1.2.15.
It remains to note that (1.2.19) follows from (1.2.23), with account of the isomorphism
Hm ∼ =H fm and of Lemma 1.2.18 below.

Lemma 1.2.18
The dimension of the space P m of homogeneous polynomials of order m in Rd +1 is given
by à !
d +m (m + d )(m + d − 1) · · · (m + 1)
dim P m = = . (1.2.25)
d d!

Proof
The basis in P m is given by monomials x 1m1 . . . x dm+1
d +1
, such that m1 + · · · + md +1 = m .
Therefore, the dimension of P m is the number of ordered partitions of m into a sum of
d + 1 non-negative integers. Finding it is equivalent to finding the number of sequences
of zeros and ones of length d + m with exactly d zeros (summing up the ones between the
neighbouring zeros we get precisely the required partition of m ), which is clearly given by
(1.2.25).

Exercise 1.2.19
Show that the coordinate functions x 1 , . . . , x d +1 restricted to the sphere Sd form a basis
of the first eigenspace on Sd .

Exercise 1.2.20
Show that the eigenvalue counting function of the Laplacian on the sphere Sd satisfies
the asymptotics
2 d ³ d −1 ´
NSd (λ) = λ 2 +O λ 2 , (1.2.26)
d!
34 Chapter 1. Strings, drums, and the Laplacian

and the power in the remainder estimate cannot be improved. Hint: find the asymptotic
behaviour of multiplicities. A complete solution to this exercise can be found in [Shu01,
§III.22].

Exercise 1.2.21
Using formula (1.2.16) and separation of variables, find eigenvalues and eigenfunctions of
the Dirichlet and Neumann Laplacian for Euclidean balls in Rd . In particular, show that
for the d -dimensional unit ball Bd , the Dirichlet eigenvalues are
³ ´2
d
λD
m,k ( B ) = j m+ −1,k ,
d m ∈ N0 , k ∈ N,
2

with multiplicity κd −1,m given by (1.2.19), where j m+ d −1,k is the k th positive zero of the
2
Bessel function J m+ d −1 (x), see Example 1.1.17. Show also that the Neumann eigenvalues
2
are ³ ´ 2
d
λN ′
m,k (B ) = p d ,m,k , m ∈ N0 , k ∈ N,

with the same multiplicity κd −1,m , where p d′ ,m,k is the k th positive zero of the derivative
Ud′ ,m (x) of the ultraspherical Bessel function

d
Ud ,m (x) := x 1− 2 J m+ d −1 (x),
2

with the exception p d′ ,0,1 := 0. For d = 2, compare your results with those given in Exam-
ple 1.1.17.
 CHAPTER 2
The spectral theorems
In this chapter, we present the weak and strong spectral theorems for
the Dirichlet and Neumann Laplacians, as well as for the
Laplace–Beltrami operator on a Riemannian manifold. We present
the fundamentals of the theory of Sobolev spaces and define the
notion of weak solutions. We also recall some basic facts about
self-adjoint unbounded linear operators and introduce the
Friedrichs extension. In order to prove local and global regularity of
eigenfunctions we give a brief overview of the theory of elliptic
regularity, based on a priori estimates and Nirenberg’s method of
difference quotients.

§2.1. Weak spectral theorems

§2.1.1. Spectral theorems: an overview and the roadmap

Generally speaking, a spectral theorem is a result stating that subject to certain conditions an oper-
ator can be in some sense diagonalised. More specifically, in application to the eigenvalue problem
(1.1.9) for the Laplacian in a bounded domain Ω ⊂ Rd with Dirichlet (1.1.10) boundary conditions,
it says that the eigenvalues form a discrete sequence with the only limit point at +∞, and that the
corresponding eigenfunctions can be chosen to form an orthonormal basis in L 2 (Ω). A similar
result holds also for Neumann (1.1.13) boundary conditions, under some mild regularity assump-
tions on the boundary ∂Ω. We emphasise that we have not yet formally put the eigenproblem
(1.1.9) subject to either (1.1.10) or (1.1.13) in the framework of operator theory, and for now con-
sider eigenvalues and eigenfunctions as those of a boundary value problem — we will call the cor-
responding spectral theorems the strong spectral theorems, and postpone their formulation until
§2.2.7.
The analysis behind the strong spectral theorems is somewhat delicate, and we perform it in
the following steps. First of all, we switch to the so-called weak spectral problems (i.e. understood

35
36 Chapter 2. The spectral theorems

in the distributional sense), introduced first for the Dirichlet boundary value problem in §2.1.2,
together with required preliminaries from the theory of Sobolev spaces. The Dirichlet case is easier
as no conditions on the boundary are required; this allows us to formulate and prove the weak
Dirichlet spectral Theorem 2.1.20 in §2.1.4. Along the way we give a brief reminder of basic spec-
tral theory of unbounded self-adjoint operators in §2.1.5, and use it to put the Dirichlet spectral
problem in the operator-theoretic framework via the construction of the Friedrichs extension in
§2.1.6.
The formulation of the weak spectral theorem for the Neumann problem is a bit more subtle
and is dealt with in §2.1.7, see Theorem 2.1.36.
In §2.1.8 we establish the weak spectral theorem for the Laplacian acting on a Riemannian
manifold. This would allow us to treat the strong spectral theorem in this case later on within the
general framework.
The weak spectral theorems do not imply the strong ones on their own. The essential missing
ingredient is the so called elliptic regularity, which we review in §2.2. In essence, this fundamental
property of elliptic PDEs allows us to establish that the weak eigenfunctions of either Dirichlet
Laplacian, Neumann Laplacian, or a Laplace–Beltrami operator on a compact Riemannian man-
ifold, for which we have already deduced some minimal regularity in weak spectral theorems, are
in fact infinitely smooth in the interior. Together with the results on regularity near the bound-
ary (which may require some additional conditions on ∂Ω) this allows us to show that the weak
eigenfunctions are in fact the strong ones, finally leading to the strong spectral Theorem 2.2.21.

§2.1.2. Weak derivatives and Sobolev spaces

Definition 2.1.1
Let Ω ⊂ Rd be a domain. Let u, v ∈ L 1loc (Ω). Suppose that for any ϕ ∈ C 01 (Ω),
ˆ ˆ
u ∂ j ϕ dx = − v ϕ dx,
Ω Ω

where ∂ j := ∂x∂ j . Then we say that ∂ j u exists in Ω in the weak sense and is equal to v .

Remark 2.1.2
If u ∈ C 1 (Ω) then the weak and the classical derivatives coincide. In the theory of dis-
tributions, weak derivatives are also referred to as distributional derivatives. To keep the
presentation more accessible, in what follows we do not use the language of distributions.
§2.1. Weak spectral theorems 37

Definition 2.1.3
Set H 0 (Ω) := L 2 (Ω). Let m ∈ N. The Sobolev spaces H m (Ω) are defined recursively as

H m (Ω) := {u ∈ L 2 (Ω) : ∂ j u exists in the weak sense,

and ∂ j u ∈ H m−1 (Ω) for all j = 1, . . . , d }.

Equipped with the inner products

ˆ ˆ Sergei Lvovich Sobolev
(u, v)H 1 (Ω) := uv dx + 〈∇u, ∇v〉 dx, (1908—1989)
Ω Ω
d ¡
∂ j u, ∂ j v H m−1 (Ω) ,
X ¢
(u, v)H m (Ω) := (u, v)L 2 (Ω) + m ≥ 2,
j =1

and the induced norms

∥u∥2H 1 (Ω) := ∥u∥2L 2 (Ω) + ∥∇u∥2L 2 (Ω) ,
d (2.1.1)
∥u∥2H m (Ω) := ∥u∥2L 2 (Ω) + ∥∂ j u∥2H m−1 (Ω) ,
X
m ≥ 2,
j =1

the Sobolev spaces H m (Ω) become Hilbert spaces.

Remark 2.1.4
As we mostly deal with real-valued functions, we omit the complex conjugation in the
definition of the Sobolev inner product and elsewhere.

Remark 2.1.5
The Sobolev norm may be alternatively defined as

∥u∥2H m (Ω) = ∥∂α u∥2L 2 (Ω) ,

X
(2.1.2)
|α|≤m

where we use the multi-index notation (B.2.1). It can be easily checked that the norms
(2.1.1) and (2.1.2) are equivalent, and in fact coincide for m = 1.

Remark 2.1.6
It turns out that one may also define the Sobolev space H m (Ω) as the completion of

{u ∈ C ∞ (Ω) : ∥u∥H m (Ω) < ∞}.

This result is due to Meyers and Serrin [MeySer64], see also [AdaFou03, Theorem 3.17].
38 Chapter 2. The spectral theorems

We denote by H0m (Ω) the closure of C 0m (Ω) with respect to the norm (2.1.1). The following
important compactness result holds.

Theorem 2.1.7
Let Ω ⊂ Rd be a bounded domain.

(i) The space H01 (Ω) is compactly embedded in L 2 (Ω).

(ii) If, in addition, ∂Ω is Lipschitz then H 1 (Ω) is compactly embedded in L 2 (Ω).

Theorem 2.1.7 is the Rellich–Kondrachov compactness theorem, see [AdaFou03, Theorem

6.3]. To get some intuition, one can compare it with a version of the Arzela–Ascoli theorem
which states that for a bounded domain Ω, the Banach space C 1 (Ω) is compactly embedded in
the Banach space C (Ω). In fact, one can prove the Rellich–Kondrachov theorem by mollifying
and reducing it to the Arzela–Ascoli Theorem, see [Bre11, Theorem 9.16] or [Eva10, §II.5.7].

Remark 2.1.8
Statement (ii) of Theorem 2.1.7 is still valid under some weaker conditions on the regu-
larity of the boundary ∂Ω, namely that Ω satisfies the so-called extension property. For a
comprehensive discussion of the extension property see, e.g., [EdmEva18, §V.4].

In many cases, the notion of a weak derivative is much more convenient to work with than
the notion of a classical derivative. The remarkable Sobolev embedding theorem below connects
these two notions. In particular, it shows that classical derivatives of all orders exist in a domain Ω
if and only if weak derivatives of all orders belong to L 2loc (Ω).

Theorem 2.1.9: The Sobolev embedding theorem [AdaFou03, Theorem 4.12]

Let Ω ⊂ Rd be a bounded domain. Then for m > d2 +k we have a continuous embedding

H0m (Ω) ⊂ C k (Ω). If, in addition, ∂Ω is Lipschitz then H m (Ω) is continuously embedded
in C k (Ω).

The following characterisation of Sobolev spaces in terms of the Fourier transform can be
used to give a proof of Theorem 2.1.9. Here the Fourier transform of a function u ∈ L 1 (Rd ) is
defined as
ˆ
− d2
(F u)(ξ) := (2π) e−i〈x,ξ〉 u(x) dx, ξ ∈ Rd . (2.1.3)
Rd

It can be shown (see e.g. [Eva10, §4.3.1]) that the formula (2.1.3) defines an isometry F : L 1 (Rd ) ∩
L 2 (Rd ) → L 2 (Rd ) and that F extends to an isometric isomorphism F : L 2 (Rd ) → L 2 (Rd ). Its
§2.1. Weak spectral theorems 39

inverse on L 1 (Rd ) ∩ L 2 (Rd ) is given by

ˆ
d
ei〈x,ξ〉 v(ξ) dξ , x ∈ Rd .
¡ −1 ¢
F v (x) = (2π)− 2
Rd

Proposition 2.1.10: [Shu20, Proposition 8.3]

Let u ∈ L 2 (Rd ). Then u ∈ H m (Rd ) if and only if its Fourier transform F u satisfies
¢m
1 + |ξ|2 2 (F u)(ξ) ∈ L 2 (Rd ) .
¡
(2.1.4)

m
Now, if m > d /2 and (1+|ξ|2 ) 2 (F u) ∈ L 2 (Ω), then F u ∈ L 1 (Ω), which easily follows from
the Cauchy–Schwarz inequality and the fact that (1 + |ξ|2 )−m ∈ L 1 (Ω). Then u is the inverse
Fourier transform of an L 1 (Ω)-function F u , and in particular it is continuous. This gives an
idea of the proof of Theorem 2.1.9.
It is sometimes desirable to define Sobolev spaces H m for fractional (non-negative) values of
the parameter m . The characterisation (2.1.4) leads to a natural definition of H m (Rd ) for frac-
tional m , see [Fol95, Chapter 6] or [McL00, Chapter 3].
For a domain Ω ⊂ Rd and m ∈ N we define H −m (Ω) as the dual Hilbert space of H0m (Ω).
m
In what follows, we will also say that u ∈ Hloc (Ω) if u|U ∈ H m (U ) for any open set U ⋐ Ω.
We also need to define Sobolev spaces H m (∂Ω) on the boundary ∂Ω of a Lipschitz domain
Ω ⊂ Rd . This is a delicate and technically involved construction, and we refer to [Gri11] and in
particular to [ChWGLS12, §A.3] for full details. Let us briefly explain the main ideas.
First, if Ω = Rd −1 × R+ is a half-space, then ∂Ω = Rd −1 and no additional work is required.
Second, let n o
Ω = (x ′ , x d ) ∈ Rd −1 × R : x d > f (x ′ )

be a “curved” half-space whose boundary ∂Ω = (x ′ , f (x ′ )) : x ′ ∈ Rd −1 is represented as the graph

© ª

of a Lipschitz function f : Rd −1 → R. Given u ∈ L 2 (∂Ω) we define u f ∈ L 2 (Rd −1 ) by u f (x ′ ) =

u(x ′ , f (x ′ )), x ′ ∈ Rd −1 . Then we set
n o
H m (∂Ω) = u ∈ L 2 (∂Ω) : u f ∈ H m (Rd −1 ) . (2.1.5)

One can check that for for a Lipschitz hypersurface ∂Ω this definition makes sense only if 0 ≤
m ≤ 1, whereas for smooth hypersurfaces one can take an arbitrary m ≥ 0.
Finally, in order to define Sobolev spaces on ∂Ω for a bounded Lipschitz domain Ω ⊂ Rd ,
we represent the boundary locally using graphs of Lipschitz functions as in §B.3, and use (2.1.5)
together with a partition of unity argument, cf. §2.1.8 below.
The following trace theorem gives a natural example where the boundary Sobolev spaces ap-
pear.
40 Chapter 2. The spectral theorems

Theorem 2.1.11: The trace theorem [Eva10, Section 5.5], [Gri11, Section 1.5]
Let Ω ⊂ Rd be a bounded domain with a Lipschitz boundary. There exists a bounded
1
linear operator T : H 1 (Ω) → H 2 (∂Ω) (called the trace operator) such that Tu = u|∂Ω if
u ∈ H 1 (Ω) ∩C (Ω).

Note that in view of Theorem 2.1.9, functions from H m (Ω) have pointwise boundary values
for m > d /2.

§2.1.3. Weak solutions

We will use the following standard integration by parts formula.

Lemma 2.1.12: Integration by parts

Let Ω ⊂ Rd be a smooth bounded domain. Let u, v ∈ C 1 (Ω). Then

ˆ ˆ ˆ
u(∂k v) dx = − (∂k u)v dx + uvn k dσ, (2.1.6)
Ω Ω ∂Ω

where n k is the k th coordinate of the outward unit normal vector on ∂Ω.

Lemma 2.1.12 remains valid also for Lipschitz domains [Nec12, §3.1, Theorem 1.1]). It implies

Lemma 2.1.13: Green’s formula [EvaGar15, §4.3], [ChWGLS12, formula (A.26)]

For a bounded domain Ω ⊂ Rd with a Lipschitz boundary, and for any real valued u ∈
H 2 (Ω), v ∈ H 1 (Ω),
ˆ ˆ ˆ
− ∆u v dx = 〈∇u, ∇v〉 dx − (∂n u) v ds. (2.1.7)
Ω Ω ∂Ω

Exercise 2.1.14
Prove (2.1.6) and (2.1.7) for a smooth domain Ω.

Of course, formula (2.1.7) re-written as

(−∆u, v)L 2 (Ω) = (∇u, ∇v)L 2 (Ω) − (∂n u, v)L 2 (∂Ω) ,

remains valid for complex valued u ∈ H 2 (Ω), v ∈ H 1 (Ω) as well.

§2.1. Weak spectral theorems 41

Remark 2.1.15
If v ∈ H01 (Ω), a simple argument shows that for any bounded domain Ω, with no regular-
ity assumptions on its boundary, and for any v ∈ H01 (Ω), Green’s formula is still valid in
the form ˆ ˆ
− ∆u v dx = 〈∇u, ∇v〉 dx. (2.1.8)
Ω Ω

We leave the proof of (2.1.8) as an exercise for the reader.

The concept of a weak solution of a boundary value problem is standard and can be found in
numerous textbooks, see for example [Shu20]. Let Ω ⊂ Rd be a bounded domain, let f ∈ C (Ω),
and suppose that u ∈ C 2 (Ω) ∩C (Ω) is a solution of the boundary value problem
(
−∆u + u = f ,
(2.1.9)
u|∂Ω = 0.

Then for any test function v ∈ C 01 (Ω) we get

ˆ ˆ ˆ
− ∆u v dx + u v dx = f v dx. (2.1.10)
Ω Ω Ω

Applying now Green’s formula (2.1.8) with v ∈ C 01 (Ω) to (2.1.10), we obtain

ˆ ˆ ˆ
uv dx + 〈∇u, ∇v〉 dx = f v dx. (2.1.11)
Ω Ω Ω

Note that both sides of (2.1.11) are well defined if u ∈ H 1 (Ω), v ∈ H01 (Ω) and f ∈ L 2 (Ω). A func-
tion u ∈ H 1 (Ω) which satisfies (2.1.11) for any test function v ∈ H01 (Ω) is called a weak solution of
the equation −∆u + u = f . To make it a weak solution of the Dirichlet boundary value problem
we also require u ∈ H01 (Ω).

Definition 2.1.16: Weak Dirichlet solution and weak Dirichlet spectral problem
We say that u ∈ H01 (Ω) is a weak solution of the boundary value problem (2.1.9) if (2.1.11)
holds for all v ∈ H01 (Ω) (or, equivalently for all v ∈ C 01 (Ω)). The weak Dirichlet spectral
problem is to find λ ∈ R and u ∈ H01 (Ω) \ {0} such that
ˆ ˆ
〈∇u, ∇v〉 dx = λ uv dx for all v ∈ H01 (Ω). (2.1.12)
Ω Ω
42 Chapter 2. The spectral theorems

Exercise 2.1.17

´ unique. Hint: apply1 the Riesz

Prove that a weak solution of (2.1.9) always exists and is
representation theorem to the linear functional F (v) = f v dx defined on H0 (Ω).
Ω

§2.1.4. The weak spectral theorem for the Dirichlet Laplacian

Existence and uniqueness of a weak solution of (2.1.9) allow us to define the solution opera-
tor Ke : L 2 (Ω) → H01 (Ω), Ke f := u . Now we let K be the composition of Ke with the inclusion
H01 (Ω) ⊂ L 2 (Ω). Note that (informally) K = (−∆+1)−1 , and hence it is a resolvent of the Dirich-
let Laplacian (see §2.1.6 for a formal definition). It is easy to check that Ke is bounded and, due to
Theorem 2.1.7, the operator K is compact. Moreover, the following proposition holds, see §2.1.5
for a brief overview of notions in functional analysis.

Proposition 2.1.18
The operator K : L 2 (Ω) → L 2 (Ω) is compact, symmetric and positive.

Exercise 2.1.19
Prove that K is positive and symmetric, and that ∥K ∥ ≤ 1.

Compactness of the resolvent operator K is a crucial ingredient of the proof of the spectral
theorem. By the Hilbert–Schmidt theorem, L 2 (Ω) admits an orthonormal basis consisting of
eigenfunctions of the compact symmetric operator K . The corresponding eigenvalues form a
sequence of positive real numbers converging to zero. Note that if w is an eigenfunction of K
with an eigenvalue µ, we get from (2.1.9) that w ∈ H01 (Ω) is a weak solution of the equation
−µ∆w + µw = w. (2.1.13)
Dividing now (2.1.13) by µ and re-arranging, we deduce that w is a weak solution of
1−µ
−∆w = w.
µ
We therefore arrive at

Theorem 2.1.20: The weak spectral theorem for the Dirichlet Laplacian
Let Ω ⊂ Rd be a bounded domain. There exists an orthonormal basis of L 2 (Ω) composed
of weak eigenfunctions of the Dirichlet spectral problem. The corresponding eigenvalues
are non-negative and form a non-decreasing sequence which tends to +∞.

In fact, we additionally have

§2.1. Weak spectral theorems 43

Proposition 2.1.21
The first eigenvalue of the weak Dirichlet spectral problem (2.1.12) is strictly positive.

To prove Proposition 2.1.21 we rely on the following important bound.

Proposition 2.1.22: Poincaré’s inequality, see e.g. [Shu20, Proposition 8.8]

If Ω ⊂ Rd is a bounded domain, then there exists a constant C Ω > 0 such that
ˆ ˆ
|u|2 dx ≤ C Ω |∇u|2 dx (2.1.14)
Ω Ω

for all u ∈ H01 (Ω).

The integral in the right-hand side of (2.1.14) is called the Dirichlet energy of u .

Exercise 2.1.23
Prove Poincaré’s inequality, first for functions in C 01 (Ω). Show that in fact a stronger ver-
sion of (2.1.14) holds: for any j = 1, . . . , d ,
ˆ ˆ
|u|2 dx ≤ C Ω |∂ j u|2 dx Jules Henri
Ω Ω Poincaré
(1854–1912)
for all u ∈ H01 (Ω).

Substituting λ := λ1 and u = v := u 1 into the weak Dirichlet spectral problem (2.1.12) (where
λ1 and u 1 are its first eigenvalue and eigenfunction), we immediately deduce from Poincaré’s
inequality that
∥∇u 1 ∥2L 2 (Ω) 1
λ1 = ≥ > 0,
∥u 1 ∥2L 2 (Ω) CΩ

thus proving Proposition 2.1.21.

44 Chapter 2. The spectral theorems

§2.1.5. Self-adjoint unbounded linear operators

We very briefly review a few basic notions from functional analysis. For more details see e.g. [Lax02].

Let H be a complex Hilbert space with an inner product (·, ·)H . By a (possibly, unbounded)
linear operator A in H we understand a linear map A : Dom(A) → H defined on a dense (but
not necessarily closed) subspace Dom(A) ⊂ H called the domain of A . If two linear operators
A, B in H satisfy Dom(A) ⊂ Dom(B ) and Bu = Au whenever u ∈ Dom(A), we say that B is an
extension of A and write A ⊂ B .
The adjoint operator A ∗ of A is defined to have the domain
David Hilbert
Dom(A ∗ ) := u ∈ H : there exists f ∈ H such that
©
(1862–1943) ¡ ¢ ª (2.1.15)
(u, Av)H = f , v H for all v ∈ Dom(A) ,

and then by setting A ∗ u := f , where u , f are as above ( f is uniquely defined since Dom(A) is
dense in H ). Therefore, we have

A ∗ u, v for all u ∈ Dom(A ∗ ), v ∈ Dom(A).

¡ ¢
H = (u, Av)H

An operator A is called symmetric if

(Au, v)H = (u, Av)H for all u, v ∈ Dom(A).

Observe that if A is symmetric then A ⊂ A ∗ . A symmetric operator A is called self-adjoint if

Dom(A) = Dom(A ∗ ) (so A = A ∗ ). Note that not all symmetric unbounded operators are self-
adjoint, as seen in the following

Example 2.1.24
2
d 2
Consider the operator A 0 := − dx 2 with the domain C 0 (R) acting in the Hilbert space
2
L (R). It is easily checked that A 0 is symmetric; however it is not self-adjoint as the func-
2
tion e−x belongs to the domain of A ∗0 but not to the domain of A 0 .

The resolvent set of a linear operator A in H is the set of complex numbers λ ∈ C such that
the operator A −λI maps its domain bijectively to H and such that R(λ) = (A −λ)−1 : H → H
is bounded. The operator R(λ) is called the resolvent operator. The spectrum of an operator A ,
denoted by Spec(A), is defined as the complement of the resolvent set. A number λ ∈ Spec(A)
is called an eigenvalue of A if dim Ker(A − λI ) > 0 — in other words, if there exists a non-trivial
solution u ∈ H \ {0} of the equation
Au = λu.
This dimension is then called the multiplicity of the eigenvalue λ, and non-trivial elements of
Ker(A −λI ) are called the eigenvectors (or the eigenfunctions if H consists of functions) of A cor-
responding to the eigenvalue λ. The discrete spectrum of A is the set of all isolated eigenvalues of A
of finite multiplicity. The complement of the discrete spectrum inside the full spectrum is called
§2.1. Weak spectral theorems 45

the essential spectrum of A . We say that the operator A has discrete spectrum if its essential spec-
trum is empty. Importantly, the spectrum of a self-adjoint operator is always real. Additionally,
the spectrum is discrete if there is at least one point of the resolvent set λ0 at which the resolvent
(A − λ0 I )−1 is compact.
Suppose that an operator A 0 is symmetric and semi-bounded from below, that is, there exists
a constant c (not necessarily positive) such that

(A 0 u, u)H ≥ c(u, u)H for all u ∈ Dom(A 0 ). (2.1.16)

If c > 0, we say that the operator A 0 is positive. We want to specify a particular self-adjoint exten-
sion A of A 0 . Without loss of generality we will assume that c = 1 in (2.1.16); if this is not the case
we may consider instead the operator A 0 +(1−c)I and subtract (1−c)I at the end. We introduce
a new inner product on Dom(A 0 ) by using the bilinear form of A 0 ,

(u, v) A 0 := (A 0 u, v)H = (u, A 0 v)H for all u, v ∈ Dom(A 0 ).

Let H 0 be the completion of Dom(A 0 ) with respect to (·, ·) A 0 . Then there is a natural embedding
H 0 ⊂ H with the norm of the embedding operator not greater than one.
We now define the Friedrichs extension A of A 0 by setting

Dom(A) := u ∈ H 0 : there exists f ∈ H such that

©
(2.1.17) Kurt Otto
(u, v) A 0 = f , v H for all v ∈ H 0 ,
¡ ¢ ª
Friedrichs
(1901–1982)
and Au := f for u ∈ Dom(A) and f as above.

Remark 2.1.25
Let us compare the definition of the Friedrichs extension (2.1.17) with the definition of
the adjoint operator (2.1.15). They look similar, but we note that in (2.1.17) we take u from
H 0 instead of a larger space H , and take v also from H 0 rather than from a smaller space
Dom(A 0 ). We therefore have
A 0 ⊂ A ⊂ A ∗0 .

The following result holds.

Theorem 2.1.26: [Lax02, §33.3]

The Friedrichs extension of a symmetric semi-bounded from below operator is self-
adjoint.

Remark 2.1.27
The construction of the Friedrichs extension shows that every symmetric semi-bounded
below operator has at least one self-adjoint extension. There exist, however, symmetric
46 Chapter 2. The spectral theorems

operators which are not semi-bounded and which have no self-adjoint extensions at all,
see [Lax02, §33.2].

§2.1.6. The Dirichlet Laplacian via the Friedrichs extension

We start by describing explicitly the construction of the Dirichlet Laplacian via the Friedrichs
extension following the general scheme given in §2.1.5. Let Ω be an open bounded set in Rd ,
and let A 0 be the operator −∆ defined on Dom(A 0 ) := C 02 Ω) ⊂ L 2 (Ω). Green’s formula (2.1.8)
immediately implies that A 0 is symmetric; it is not however self-adjoint, cf. Example 2.1.24.
Proposition 2.1.22 together with Green’s formula (2.1.8) also implies that A 0 is semi-bounded
from below, since then
∥u∥2A 0 = (A 0 u, u)L 2 (Ω) = (−∆u, u)L 2 (Ω) = ∥∇u∥2L 2 (Ω)
1
≥ ∥u∥2L 2 (Ω) for all u ∈ C 02 (Ω).
CΩ
Therefore,
(1 +C Ω )∥u∥2A 0 ≥ ∥u∥2H 1 (Ω) = ∥∇u∥2L 2 (Ω) + ∥u∥2L 2 (Ω) ≥ ∥u∥2A 0 ,
and so the norms ∥·∥ A 0 and ∥·∥H 1 (Ω) are equivalent on Dom(A 0 ). Hence, the completion H 0 of
C 02 (Ω) with respect to the norm ∥·∥ A 0 , appearing in the construction of the Friedrichs extension,
is the Sobolev space H01 (Ω).
Using now (2.1.17), we deduce that the Friedrichs extension of A 0 is the operator A , which
we from now on will denote as −∆D := −∆D Ω and will call the Dirichlet Laplacian on Ω, with the
domain (see also Definition 2.1.16)
n
Dom(−∆D ) = u ∈ H01 (Ω) : there exists f ∈ L 2 (Ω) such that
o
(∇u, ∇v)L 2 (Ω) = f , v L 2 (Ω) for all v ∈ H01 (Ω)
¡ ¢

= u ∈ H01 (Ω) : there exists f ∈ L 2 (Ω) such that (2.1.18)

©
ª
−∆u = f in the weak sense
= u ∈ H01 (Ω) : ∆u ∈ L 2 (Ω) .
© ª

Repeating now word by word the construction of the compact operator K from §2.1.4, we
conclude that we indeed have K = (−∆D − 1)−1 . Therefore we arrive at the following equivalent
formulation of Theorem 2.1.20.

Theorem 2.1.28
Let Ω ⊂ Rd be a bounded domain. The Dirichlet Laplacian −∆D defined as the Friedrichs
extension with domain (2.1.18), is a self-adjoint operator in L 2 (Ω) with a discrete spectrum
of eigenvalues accumulating to +∞, and the first eigenvalue being positive. The eigen-
functions can be chosen to form an orthonormal basis in L 2 (Ω).
§2.1. Weak spectral theorems 47

Remark 2.1.29
Theorem 2.1.28 remains valid if Ω is just an open subset of Rd of a finite volume, not nec-
essarily bounded. Moreover, the spectrum of −∆D is still discrete if an even less restrictive
condition
lim sup¯B x,1 ∩ Ω¯d = 0
¯ ¯
|x|→∞
x∈Ω

is satisfied, see [EdmEva18, Remark V.5.18(4)].

The following simple results will be often needed later on.

Lemma 2.1.30
Given a bounded open set Ω ⊂ Rd , denote by Ωρ a homothety
³ ´of Ω with the coefficient
¡ D¢ D
ρ > 0. Then λ ∈ Spec −∆Ω if and only if ρ λ ∈ Spec −∆Ωρ .
−2

Lemma 2.1.31
Let Ω³∈ Rd be of two bounded domains Ω1 and Ω2 . Then Spec −∆D
¡ ¢
´
a disjoint
³
union
´ Ω =
Spec −∆D D
Ω1 ∪ Spec −∆Ω2 with account of multiplicities.

Exercise 2.1.32
Prove Lemmas 2.1.30 and 2.1.31.

§2.1.7. The weak spectral theorem: Neumann case

In this section we discuss the analog of the weak Dirichlet Spectral Theorem 2.1.20 in the case of
the Neumann boundary condition. Unlike the Dirichlet case, in the Neumann case some regular-
ity conditions need to be imposed on the boundary from the start, and we will assume throughout
that the boundary is Lipschitz, see the discussion below.

Definition 2.1.33: Weak Neumann solution and weak Neumann spectral problem
Let Ω ⊂ Rd be a bounded domain with a Lipschitz boundary. Let f ∈ L 2 (Ω). We say that
u : Ω → R is a weak solution of the boundary value problem
(
−∆u + u = f ,
(2.1.19)
∂n u = 0.
48 Chapter 2. The spectral theorems

if u ∈ H 1 (Ω) and
ˆ ˆ ˆ
〈∇u, ∇v〉 dx + uv dx = f v dx for all v ∈ H 1 (Ω).
Ω Ω Ω

The weak Neumann spectral problem is to find λ ∈ R and u ∈ H 1 (Ω) \ {0} such that
ˆ ˆ
〈∇u, ∇v〉 dx = λ uv dx for all v ∈ H 1 (Ω). (2.1.20)
Ω Ω

Remark 2.1.34
We note that the boundary condition ∂n u = 0 “disappears” in the weak statement. How-
ever, note that (2.1.19) is required to hold for all v ∈ H 1 (Ω), not only for v ∈ H01 (Ω) (cf.
the Dirichlet case in Definition 2.1.16). We also note that the Neumann spectrum always
starts with λ1 = 0, with the corresponding eigenfunction u 1 being a constant.

Exercise 2.1.35
Prove that if u ∈ C 2 (Ω) ∩C 1 (Ω) is a weak solution of the problem (2.1.19) then it is also a
classical solution of it.

The argument given in the Dirichlet case for the existence of weak solutions works in the
Neumann case as well. The Riesz representation theorem guarantees the existence of a unique
solution u ∈ H 1 (Ω) for any given f ∈ L 2 (Ω). The composition of the solution operator Ke :
L 2 (Ω) → H 1 (Ω) with the inclusion H 1 (Ω) ⊂ L 2 (Ω) is compact, see Remark 2.1.37 below. As a
result, we prove

Theorem 2.1.36: The weak spectral theorem for the Neumann Laplacian
Let Ω ⊂ Rd be a bounded domain with a Lipschitz boundary. There exists an orthonormal
basis of L 2 (Ω) composed of weak eigenfunctions of the Neumann spectral problem. The
corresponding eigenvalues are non-negative and form a non-decreasing sequence which
tends to +∞.

As in the Dirichlet case, we can equivalently reformulate the spectral theorem in the operator
theory sense by constructing the Neumann Laplacian using the Friedrichs extension, see [Hel13,
§4.4.4] or [AreCSVV18, §7.4]. Given u ∈ H 1 (Ω) such that −∆u ∈ L 2 (Ω), we say that ∂n u ∼ 0
on ∂Ω if ˆ ˆ
− ∆u v dx = 〈∇u, ∇v〉 dx for all v ∈ H 1 (Ω).
Ω Ω
§2.1. Weak spectral theorems 49

Note that while ∂n u may not be defined in L 2 (∂Ω) even weakly, both the right- and the left-hand
sides of this formula are well-defined, and hence the relation ∂n u ∼ 0 still makes sense. This allows
us to define a self-adjoint operator −∆N := −∆N Ω , which we call the Neumann Laplacian on Ω, as
the weak Laplacian with the domain (cf. (2.1.18))
n
Dom(−∆N ) = u ∈ H 1 (Ω) : there exists f ∈ L 2 (Ω) such that
o
(∇u, ∇v)L 2 (Ω) = f , v L 2 (Ω) for all v ∈ H 1 (Ω)
¡ ¢

= u ∈ H 1 (Ω) : −∆u ∈ L 2 (Ω) and ∂n u ∼ 0 on ∂Ω .

© ª

For a slightly different approach to Neumann boundary value problems see [Fol95] or [Dav95].

Remark 2.1.37
We emphasise that our assumption that the boundary ∂Ω is Lipschitz is crucial for the va-
lidity of Theorem 2.1.36. It guarantees that Theorem 2.1.7(ii) holds, and therefore −∆N has
a compact resolvent, thus ensuring the discreteness of the spectrum. Although this con-
dition can be slightly relaxed, see [Dav95, Chapter 7] for details, it cannot be dismissed
altogether: without any regularity assumptions on ∂Ω one can construct examples of
bounded domains for which the spectrum of the Neumann Laplacian is no longer dis-
crete, see e.g. [HemSecSim91].

Exercise 2.1.38
Prove the analogues of Lemmas 2.1.30 and 2.1.31 for the Neumann Laplacian.

§2.1.8. The weak spectral theorem: Riemannian manifolds

Let (M , g ) be a smooth compact Riemannian manifold, possibly with boundary. If the boundary
is non-empty we assume that either Dirichlet or Neumann boundary conditions are imposed on
∂M , and recall Remark 1.2.8.
By Green’s identity, the Laplacian acting on functions from C 2 (M ) is a symmetric differential
operator in the space L 2 (M ):
ˆ ˆ ˆ
(−∆u)v dV = 〈∇u, ∇v〉g dV = u(−∆v) dV.
M M M

Note that the boundary term vanishes due to the boundary conditions. Setting u = v we also
observe that the Laplacian is a non-negative symmetric operator.
Let us introduce the Sobolev space

H 1 (M ) = {u ∈ L 2 (M ), ∇u ∈ L 2 (M )},
50 Chapter 2. The spectral theorems

where the gradient is understood in the weak sense. The norm in H 1 (M ) is defined by

ˆ ˆ
∥u∥2H 1 (M ) := u 2 dV + |∇u|2 dV.
M M

Moreover, for any m ∈ N, one can extend the definition of the Sobolev space H m (M ) to mani-
folds using coordinate charts and a partition of unity. We refer to [Tay11, Section 4.3] and [Shu01,
Section I.7] for details.
Let us define also the space H01 (M ) as the closure of the space C 01 (M ) in the norm of H 1 (M ).
Clearly, H01 (M ) ⊂ H 1 (M ) ⊂ L 2 (M ).
We can define the weak spectral problem for the Laplace operator on a closed manifold, or the
weak Neumann spectral problem on a manifold with boundary by analogy with (2.1.20), and the
weak Dirichlet spectral problem on a manifold with boundary by analogy with (2.1.12). Acting
similarly to Theorems 2.1.36 and 2.1.20, we obtain

Theorem 2.1.39: The weak spectral theorem for a Riemannian manifold

Let (M , g ) be a smooth compact Riemannian manifold, possibly with boundary. If the
boundary is non-empty we assume that either Dirichlet or Neumann boundary condi-
tions are imposed on ∂M . In each of these cases, there exists and orthonormal basis
of L 2 (M ) composed of weak eigenfunctions of the corresponding Laplace spectral prob-
lem. The corresponding eigenvalues are non-negative and form a non-decreasing sequence
tending to +∞.

Exercise 2.1.40
Show that on a compact connected Riemannian manifold the only harmonic function
is a constant. In particular, this implies that the Laplace eigenvalue zero on a compact
connected manifold always has multiplicity one.

Notation 2.1.41
Let M be a closed Riemannian manifold. We will be enumerating the eigenvalues of the
Laplace-Beltrami operator on M starting with λ0 = 0 as

0 = λ0 ≤ λ1 ≤ . . . .

In particular, for a connected manifold λ1 > 0 by Exercise 2.1.40. This enumeration is

traditional, and is motivated by the importance of the first non-zero eigenvalue λ1 .
§2.2. Elliptic regularity and strong spectral theorems 51

Exercise 2.1.42
Let (M , g ) be a compact Riemannian manifold of dimension d . Show that for any ρ > 0,
j ∈ N,
λ j (M , g )
λ j (M , ρg ) = , j ∈ N, (2.1.21)
ρ
and, consequently, the quantity λ j (M , g ) Vol(M , g )2/d is invariant under rescaling. This
is a Riemannian analogue of Lemma 2.1.30, see also Exercise 2.1.38.

§2.2. Elliptic regularity and strong spectral theorems

§2.2.1. Elliptic regularity for the Dirichlet Laplacian

We want to show that the weak eigenfunctions of the Dirichlet problem (2.1.12) found in Theo-
rem 2.1.20 are in fact smooth. This is due to an important phenomenon known as elliptic regu-
larity. We present an overview of this theory below.
We have

Theorem 2.2.1: Smoothness of Dirichlet eigenfunctions

Let Ω ⊂ Rd be a bounded open set, and let u 1 , u 2 , . . . , be the eigenfunctions of the weak
Dirichlet spectral problem (2.1.12). Then

(i) Each eigenfunction u j , j ∈ N, belongs to C ∞ (Ω).

(ii) Moreover, each eigenfunction is real analytic in Ω.

(iii) The eigenfunctions are smooth up to the boundary near the smooth parts of the
boundary: if ∂∞ Ω is the C ∞ part of ∂Ω, then u j and all its derivatives can be con-
tinuously extended to Ω ∪ ∂∞ Ω.

(iv) If ∂Ω is Lipschitz, u j ∈ C (Ω), and u j |∂Ω = 0 pointwise.

Parts (i) and (ii) of Theorem 2.2.1 follow from what is usually referred to as local or interior
elliptic regularity. Clearly, (ii) implies (i), however we present an independent proof of the latter
statement, as it can be generalised to the setting of smooth Riemannian manifolds. Parts (iii) and
(iv) follow from global elliptic regularity, or regularity up to the boundary.
52 Chapter 2. The spectral theorems

§2.2.2. Proof of the local regularity

The goal of this subsection is to prove parts (i) and (ii) of Theorem 2.2.1. We start with the proof
of the latter as it can be easily deduced from the real analyticity of harmonic functions.

Proof of Theorem 2.2.1, part (ii)

We use the so-called
p
lifting trick (cf. Exercise 4.3.17) and consider the harmonic function
λt
h(x, t ) := u(x)e in Ω × R ⊂ Rd +1 . Since harmonic functions are real analytic [Axl-
BouWad01, Theorem 1.28], it follows that u(x) = h(x, 0) is real analytic.

We note that this argument can be adjusted to work for solutions of −∆u − λu = 0 with
negative λ and in any case does not require any boundary conditions. Let us also remark that the
analogue of this statement holds for uniformly elliptic operators with real analytic coefficients, see
[Fri69, Theorem III.1.2], [Joh81, Ch. 7], [MorNir57], and hence applies to the Laplace–Beltrami
operators on Riemannian manifolds with real analytic metrics.
In order to prove part (i) we use a fundamental regularity result from the theory of elliptic
partial differential equations. First, we need to define weak solutions for a wider class of problems.
Consider an open set Ω ⊂ Rd , and a uniformly elliptic equation in divergence form,
− div A∇u = f in Ω, (2.2.1)
¢d
where f ∈ L 2 (Ω) and A = a i j is a positive definite symmetric matrix with entries a i j ∈
¡
i , j =1
L ∞ (Ω) which satisfies
〈Aξ, ξ〉 ≥ α0 |ξ|2 (2.2.2)
for all x ∈ Ω and ξ ∈ Rd , and some fixed α0 > 0.

Definition 2.2.2: Weak solution of a uniformly elliptic equation in divergence

form
We say that u ∈ H 1 (Ω) is is a weak solution of the equation (2.2.1) (or alternatively that u
satisfies the equation (2.2.1) in the weak sense) if
ˆ ˆ
〈A∇u, ∇v〉 dx = f v dx
Ω Ω

for all v ∈ H01 (Ω).

Remark 2.2.3
If we take A to be the identity matrix, then equation (2.2.1) becomes the standard Laplace
equation −∆u = f .

The fundamental result mentioned above is

§2.2. Elliptic regularity and strong spectral theorems 53

Theorem 2.2.4: Local elliptic regularity for the Laplacian [GilTru01, Theorem
8.10], [Fol95, Lemma 6.32], [Eva10, §6.3.1, Theorem 2]
Let Ω ⊂ Rd be an open set. Suppose that for some m ≥ 0 and f ∈ Hloc m
(Ω), a function
1 m+2
u ∈ H (Ω) satisfies the equation −∆u = f in Ω in the weak sense. Then u ∈ Hloc (Ω).

Assuming this result for the moment, let us show how it implies what we need.

Proof of Theorem 2.2.1, part (i)

Let u ∈ H 1 (Ω) be a weak solution of the equation −∆u = λu . Then applying iteratively
k
Theorem 2.2.4 to u we conclude that u ∈ Hloc (Ω) for all k ≥ 1 (this procedure is called
elliptic bootstrapping). Therefore, by Theorem 2.1.9 it follows that u ∈ C ∞ (U ) for any open
set U ⋐ Ω, and hence u ∈ C ∞ (Ω).

Remark 2.2.5: Local regularity of eigenfunctions for other eigenvalue problems

Note that the proof does not use boundary conditions, and hence local regularity holds
also for Neumann eigenfunctions. Moreover, arguments of elliptic regularity are robust in
a sense that Theorem 2.2.4 can be extended to second order elliptic operators with smooth
coefficients such as the Laplace-Beltrami operator, see Theorem 2.2.17 below.

§2.2.3. A priori estimates and the method of difference quotients

The proof of Theorem 2.2.4 uses two key ideas: an a priori estimate and Nirenberg’s method of
difference quotients. Let us start with the latter. Following [Nir59], let the difference quotient be
defined as
u(x + h) − u(x)
D h u(x) := ,
|h|
where h ∈ Rd \ {0}. Since D h commutes with differentiations, we get

−∆(D h u) = D h f .

Given f ∈ H 1 (Rd ), it is not difficult to verify that if t > 0 and e k is the k th unit coordinate vector
Louis Nirenberg
in Rd ,
(1925–2020)
∥D t e k ∥L 2 (Rd ) ≤ ∥∂k f ∥L 2 (Rd ) , (2.2.3)
and hence
∥D h f ∥L 2 (Rd ) ≤ ∥∇ f ∥L 2 (Rd ) . (2.2.4)

Exercise 2.2.6
Prove estimate (2.2.3). Hint: Prove it first for C 01 -functions using the fundamental the-
orem of calculus and Fubini’s theorem, and then use the fact that C 01 (Rd ) is dense in
54 Chapter 2. The spectral theorems

H 1 (Rd ) (see [GilTru01, Lemma 7.23] or [Bre11, Proposition 9.3]).

The following important theorem gives a condition for showing that an L 2 (Rd ) function
belongs to the Sobolev space H 1 (Rd ). It is proved using weak compactness of closed bounded
sets in L 2 (Rd ) (cf. proof of Lemma 2.2.14 for a similar argument).

Theorem 2.2.7: The method of difference quotients [GilTru01, Lemma 7.24],

[Bre11, Prop. 9.3]
Let u ∈ L 2 (Rd ). If there exists C > 0 such that for all h ∈ Rd with 0 < |h| ≤ 1 we have

∥D h u∥L 2 (Rd ) ≤ C ,

then u ∈ H 1 (Rd ). In particular, if ∥D t e k u∥L 2 (Rd ) ≤ C for all |t | < 1, then ∂k u ∈ L 2 (Rd ).

Exercise 2.2.8: Leibniz rule and integration by parts for difference quotients
Show that for u, v ∈ H 1 (Rd ) and h ∈ Rd ,

(i)
D h (uv) = (D h u)v + u(D h v) + |h|(D h u)(D h v);

(ii) ˆ ˆ
(D h u)v dx = − u(D −h v) dx.
Rd Rd

Let us move to the second part of the argument. In order to formulate an a priori estimate we
recall that we have defined the negative order Sobolev space H −1 (Ω) as the dual space of H01 (Ω),
with the usual norm of the dual Hilbert space.

Example 2.2.9
Let f ∈ L 2 (Ω). The formula
ˆ
F f (v) := f v dx, v ∈ H01 (Ω),
Ω

defines an element of H −1 (Ω). Moreover, ∥F f ∥H −1 (Ω) ≤ ∥ f ∥L 2 (Ω) . Slightly abusing nota-

tion, we write ∥ f ∥H −1 (Ω) ≤ ∥ f ∥L 2 (Ω) .

We can now state the following

§2.2. Elliptic regularity and strong spectral theorems 55

Lemma 2.2.10: An a priori estimate in H 1 (Rd ) [Fol95, Theorem 6.28]

Let f ∈ L 2 (Rd ), and let u ∈ H 1 (Rd ) be a weak solution of the equation −∆u = f in Rd .
Then ³ ´
∥u∥2H 1 (Rd ) ≤ 2 ∥u∥2L 2 (Rd ) + ∥ f ∥2H −1 (Rd ) . (2.2.5)

Proof
We have
ˆ ˆ
2
|∇u| dx = f u dx ≤ ∥ f ∥H −1 (Rd ) ∥u∥H 1 (Rd )
Rd Rd
1 1 1
≤ ∥ f ∥2H −1 (Rd ) + ∥u∥2L 2 (Rd ) + ∥∇u∥2L 2 (Rd ) .
2 2 2
Rearranging the terms yields the result.

Remark 2.2.11
One way to think about a priori estimate (2.2.5) is as follows: an L 2 bound on a function
and H −1 bound on its Laplacian imply L 2 bounds on all its first derivatives. Another
illustration of a similar phenomenon is a more elementary estimate
ˆ ¯ ˆ
¯ ∂2 u ¯2
¯
¯ ¯ dx ≤ |∆u|2 dx, (2.2.6)
¯ ∂x ∂x ¯
k l
Rd Rd

which holds for any u ∈ C 02 (Rd ) and k, l , = 1, . . . , d .

The reason (2.2.6) holds is the ellipticity of the Laplace operator. Consider also an a
priori estimate for the first order elliptic Cauchy–Riemann operator: for a complex valued
u ∈ C 01 (R2 ),
ˆ ¯ ˆ ¯
¯ ∂u ¯2
¯2
¯ dx ≤ ¯ ∂u + i ∂u ¯ dx,
¯
¯ ¯
¯ j = 1, 2. (2.2.7)
¯ ∂x j
¯ ¯ ∂x 1 ∂x 2
¯
R2 R2
∂2 u
On the other hand, no a priory estimate is possible for the operator A := u 7→ ∂x 12
−
∂ u
2

∂x 22
, which is not elliptic:

ˆ ¯¯ 2 ¯2
¯ ˆ
¯∂ u ¯
for any C > 0 there exists u ∈ C 02 (R2 ) such that ¯ 2 ¯ dx ≥ C |A u|2 dx. (2.2.8)
¯ ∂x ¯
1
Rd Rd

We leave the proofs of (2.2.6)–(2.2.8) as an exercise for the reader.

56 Chapter 2. The spectral theorems

In the proof of Theorem 2.2.4 we will require the following

Lemma 2.2.12
Let f ∈ L 2 (Rd ) and h ∈ Rd . Then

∥D h f ∥H −1 (Rd ) ≤ ∥ f ∥L 2 (Rd ) .

Proof
For any v ∈ H 1 (Rd ) we have:
¯ˆ ¯ ¯ˆ
¯ ¯ ¯ ¯
¯
¯ ¯ ¯ ¯
¯ D h f v dx ¯ = ¯ f D −h v dx ¯ ≤ ∥ f ∥L 2 (Rd ) ∥D −h v∥L 2 (Rd ) ≤ ∥ f ∥L 2 (Rd ) ∥v∥H 1 (Rd ) ,
¯ ¯ ¯ ¯
¯ ¯ ¯ ¯
R R
¯ d ¯ ¯ d ¯

which implies the desired estimate. Here the first equality follows from Exercise 2.2.8 and
the last inequality follows from (2.2.4).

We now have all the required tools to prove Theorem 2.2.4.

Proof of Theorem 2.2.4

Assume first that u ∈ H 1 (Rd ) is a weak solution of the equation −∆u = f in Rd with
f ∈ L 2 (Rd ). Taking difference quotients we obtain the equation

−∆D h u = D h f ,

and after applying Lemma 2.2.10 on this new equation we obtain

³ ´
∥D h ∂k u∥2L 2 (Rd ) ≤ ∥D h u∥2H 1 (Rd ) ≤ 2 ∥D h u∥2L 2 (Rd ) + ∥D h f ∥2H −1 (Rd )
≤ 2∥u∥2H 1 (Rd ) + 2∥ f ∥2L 2 (Rd ) ,

for any k = 1, . . . , d . Here we have used (2.2.4) and Lemma 2.2.12 to estimate the right-
hand side. Applying Theorem 2.2.7, we deduce that ∂k u ∈ H 1 (Rd ) for any k = 1, . . . , d ,
and hence u ∈ H 2 (Rd ). This proves the assertion of the theorem for Ω = Rd and m = 0.
Since −∆∂k u = ∂k f we deduce the result by induction for any m ≥ 1.
Now, in order to prove the theorem for an arbitrary Ω, we use the standard localisa-
1
tion argument. Suppose that u ∈ Hloc (Ω) satisfies −∆u = f in Ω in the weak sense with
2
f ∈ L loc (Ω). Take a cut-off function ϕ ∈ C 0∞ (Ω). It is immediate that the function ϕu , ex-
tended by zero onto Rd , belongs to H 1 (Rd ). Then, −∆(ϕu) = g in the weak sense, where
g = ϕ f − 2〈∇ϕ, ∇u〉 − (∆ϕ)u . Note that g ∈ L 2 (Rd ), and we deduce that ϕu ∈ H 2 (Rd ),
2
and hence u ∈ Hloc (Ω). Iterating the argument as above completes the proof of the theo-
rem.
§2.2. Elliptic regularity and strong spectral theorems 57

§2.2.4. Global regularity of Dirichlet eigenfunctions

So far, we have shown that if u is a weak solution of the equation −∆u = λu in Ω, then u ∈
k
Hloc (Ω) for all k ∈ N, and hence u ∈ C ∞ (Ω). Note that the boundary conditions as well as
boundary regularity are irrelevant for this property. Our goal is to prove part (iii) of Theorem
2.2.1 which states u is smooth up to the boundary near smooth parts of the boundary, provided
the Dirichlet condition is imposed.
After a partition of unity argument, we can assume that u is localised in a small neighbour-
hood of the boundary. Using an appropriate change of variables we can “straighten” the smooth
part of the boundary, i.e. transform it into a part of a hyperplane. At the same time, the Euclidean
Laplacian is transformed into a certain Laplace–Beltrami operator. Indeed, if −∆u(x) = f (x),
x = ϕ(y) denotes a change of variables, and u = u ◦ ϕ, f = f ◦ ϕ, then u satisfies the equation
−Lu = f, where

1 d X
d ³ q ´
∂i g i j det g ∂ j u ,
X
Lu := p (2.2.9)
det g i =1 j =1

®ªd
the matrix g := g i j = ∂i ϕ, ∂ j ϕ i , j =1 , g i j = g −1 , and det g = ¯det(Jac ϕ)¯, where
© ª ©­ © ª p ¯ ¯

Jac ϕ is the Jacobian matrix of ϕ.

k
³ ´ we would like to show that u ∈ H (Ω) for all k , and hence, by Theorem 2.1.9,
As before,
u ∈ C ∞ Ω . Similarly to the local regularity, the main tools are an a priori estimate in a half-
space Rd+ := {(x 1 , . . . , x d ) ∈ Rd : x d > 0}, and an appropriately adjusted Nirenberg’s method of
difference quotients.
The equation
³q ´ p
det g L u = g f

is of divergence type as in (2.2.1).

Proposition 2.2.13: An a priori estimate in half space

Let m ≥ 0. Consider the equation (2.2.1), where we additionally assume that the entries
of the matrix A satisfy a i j ∈ C m R +d and have compact supports.
¡ ¢

u ∈ H m+1 Rd, + ∩H01 Rd+

¡ ¢ ¡ ¢
Let u be a¡ weak solution of this equation, and suppose that
and f ∈ H m Rd+ . Then
¢

³ ´
∥u∥H m+1 ¡Rd ¢ ≤ C ∥u∥H m ¡Rd ¢ + ∥ f ∥H m−1 ¡Rd ¢ ,
+ + +

with some constant C > 0 which depends only on the constant α0 from (2.2.2) and
bounds on ∥a i j ∥C m ¡Rd+ ¢ .
58 Chapter 2. The spectral theorems

Proof
Consider first the case m = 0. Then
ˆ ˆ ˆ
2
α0 |∇u| dx ≤ 〈A∇u, ∇u〉 dx = f u dx ≤ ∥ f ∥H −1 ¡Rd ¢ ∥u∥H 1 ¡Rd ¢
+ +

Rd+ Rd+ Rd+

1 α0 ³ ´
≤ ∥ f ∥2H −1 ¡Rd ¢ + ∥u∥2L 2 ¡Rd ¢ + ∥∇u∥2L 2 ¡Rd ¢ .
2α0 + 2 + +

After rearranging and collecting terms, one gets

1
∥∇u∥2L 2 ¡Rd ¢ ≤ ∥ f ∥2H −1 ¡Rd ¢ + ∥u∥2L 2 ¡Rd ¢ ,
+ α20 + +

or, equivalently,
1
∥u∥2H 1 ¡Rd ¢ ≤ ∥ f ∥2H −1 ¡Rd ¢ + 2∥u∥2L 2 ¡Rd ¢ .
+ α20 + +

For m > 0, we use an inductive argument. By differentiating equation (2.2.1), it is easy

to check that the equation

− div A∇∂k u = ∂k f + div((∂k A)∇u)

is satisfied in Rd+ in the weak sense. Let 1 ≤ k ≤ d − 1, i.e. consider tangential directions
with respect to the hyperplane Rd −1 × {0} bounding R+ d
. Using Lemma 2.2.14 below we
m
¡ d¢ 1 d
get that ∂k u ∈ H R+ ∩ H0 R+ , and by induction
¡ ¢

³ ´
∥∂k u∥H m ¡Rd ¢ ≤ C ∥∂k u∥H m−1 ¡Rd ¢ + ∥∂k f ∥H m−2 ¡Rd ¢ + ∥ div((∂k A)∇u)∥H m−2 ¡Rd ¢
+ + + +
³ ´
≤ C ∥u∥H m ¡Rd ¢ + ∥ f ∥H m−1 ¡Rd ¢ .
+ +

(2.2.10)

Equivalently, we have an estimate on ∥∂i ∂ j u∥H m−1 ¡Rd+ ¢ for all 1 ≤ i ≤ d and 1 ≤ j ≤ d − 1.
It remains to estimate ∥∂2d u∥H m−1 ¡Rd+ ¢ , which can be done by using the partial differen-
tial equation (2.2.1) once more. We isolate this derivative,
d dX
−1
−a d d ∂2d u = f + ∂i (a i j ∂ j u) + (∂d a d d )(∂d u).
X
(2.2.11)
i =1 j =1

Hence, the desired estimate follows from the fact that a d d ≥ α0 and the existence of the
bounds on ∥∂i ∂ j u∥H m−1 ¡Rd+ ¢ for 1 ≤ i ≤ d and 1 ≤ j ≤ d − 1 given by (2.2.10).

It remains to state and prove the auxiliary lemma used in the proof of Proposition 2.2.13.
§2.2. Elliptic regularity and strong spectral theorems 59

Lemma 2.2.14: [Bre11, Lemma 9.7]

Let u ∈ H 2 Rd+ ∩ H01 Rd+ . Then ∂k u ∈ H01 Rd+ for 1 ≤ k ≤ d − 1.
¡ ¢ ¡ ¢ ¡ ¢

Proof
Given 1 ≤ ¡k ≤¢d − 1, we set h = t e k¡ , where e k is the k th unit coordinate vector. Then,
for v ∈ H01 Rd+ , we have D h v ∈ H01 Rd+ , since e k is parallel to the hyperplane
¢
bounding
d 1 d
R+ . Due to (2.2.3) and the weak compactness of the unit ball in H0 R+ , we can find
¡ ¢

w ∈ H01 Rd+ and a sequence (h n )∞ 1 d

n=1 = (t n e k )¡n=1¢such that D h n v → w weakly in H0 R+
∞
¡ ¢ ¡ ¢

as n → ∞. On the other hand, for all ϕ ∈ C 0∞ Rd+ we have

ˆ ˆ ˆ
t n →0
(D hn v)ϕ dx = − vD −hn ϕ dx −−−→ − v∂k ϕ dx.
Rd+ Rd+ Rd+

It follows that ˆ ˆ
wϕ dx = − v∂k ϕ dx.
Rd+ Rd+

Hence, ∂k v = w , and in particular ∂k v ∈ H01 Rd+ .

¡ ¢

Combining the a priori estimate in Proposition 2.2.13 with Nirenberg’s argument we obtain
the global regularity statement.

Theorem 2.2.15: Global regularity in half space

Let u be a weak solution of equation (2.2.1), where we assume the¡ conditions on the
ij m+1 d 1 d
R+ ∩ H0 R+ and
¢ ¡ ¢
entries a¡ imposed in Proposition 2.2.13. Suppose that u ∈ H
f ∈ H m Rd+ for some m ≥ 0. Then
¢

³ ´
u ∈ H m+2 Rd+ .

Proof
Let h = t e k as above. For 1 ≤ k ≤ d − 1, we have D h u ∈ H m+1 Rd+ ∩ H01 Rd+ , and
¡ ¢ ¡ ¢

therefore we can apply the a priori estimate of Proposition 2.2.13 to the equation

− div A∇D h u = D h f + div((D h A)∇u) + |h| div((D h A)∇D h u).

Here we have used the analogue of Leibniz rule, see Exercise 2.2.8. Hence, for small enough
60 Chapter 2. The spectral theorems

|h| we have
³ ´
∥D h u∥H m+1 ¡Rd ¢ ≤ C ∥D h u∥H m ¡Rd ¢ + ∥D h f ∥H m−1 ¡Rd ¢ + ∥ div((D h A)∇u)∥H m−1 ¡Rd ¢
+ + + +

1
+ ∥D h u∥H m+1 ¡Rd ¢ .
2 +

Rearranging terms and recalling that the norms ∥D h f ∥H m−1 are bounded by ∥ f ∥H m in
view of (2.2.4), we obtain that ∥D h u∥H m+1 ¡Rd+ ¢ is bounded independently of h . It follows
2.2.7 that ∂k u ∈ H m+1 Rd+ for all 1 ≤ k ≤ d − 1, or equivalently, ∂i ∂ j u ∈
¡ ¢
from¡ Theorem
H m Rd+ for all 1 ≤ i ≤ d and 1 ≤ j ≤ d − 1. Finally, we can express ∂2d u as in (2.2.11) and
¢

deduce that ∂2d u ∈ H m Rd+ . Summarising, we have shown that u ∈ H m+2 Rd+ .
¡ ¢ ¡ ¢

Corollary 2.2.16: Global regularity for the Dirichlet problem

Let m ≥ 0, and let Ω ⊂ Rd be a bounded domain with C m+2 boundary. Let u ∈ H01 (Ω)
satisfy
−∆u = f in Ω
in the weak sense, where f ∈ H m (Ω). Then, u ∈ H m+2 (Ω).

Proof
A partition of unity argument and a change of coordinates leading to (2.2.9) reduces the
problem to Theorem 2.2.15 . We obtain an equation

− div A∇v = w g in Rd+ ,

³ ´ ³ ´
with a positive definite C m+1 Rd+ matrix A (see (2.2.1)), a positive C m+1 Rd+ weight
function w , and g ∈ H m¡ Rd+¢ . Hence w g ∈ H m Rd+ , and we can apply Theorem 2.2.15.
¡ ¢ ¡ ¢

It follows that v ∈ H m+2 Rd+ and finally that u ∈ H m+2 (Ω).

We can now finish the proof of Theorem 2.2.1(iii).

Proof of Theorem 2.2.1, part (iii)

Applying Corollary 2.2.16 with f = λu and elliptic bootstrapping shows that if Ω has
C m+2 boundary, then u ∈ H m+2 (Ω). If the boundary of Ω is C ∞ , the Sobolev embedding
Theorem 2.1.9 shows that u ∈ C ∞ (Ω).

Recall that the Laplace–Beltrami operator on a Riemannian manifold is a multiple of a second

order elliptic operator in divergence form. Hence, the proof of Corollary 2.2.16 can be extended
to this case verbatim, and we obtain
§2.2. Elliptic regularity and strong spectral theorems 61

Theorem 2.2.17: Smoothness of eigenfunctions of the Laplace–Beltrami operator

Let (M , g ) be a closed Riemannian manifold. Then the eigenfunctions of the weak spec-
tral problem for the Laplace–Beltrami operator are C ∞ on M .

Clearly, global regularity also holds for Dirichlet eigenfunctions on compact Riemannian
manifolds with boundary with the same proof as in the Euclidean case.

§2.2.5. Continuity up to the boundary of Dirichlet eigenfunctions on Lipschitz do-

mains
Boundary regularity may fail near corners of piecewise-smooth domains. A standard example is a
domain with a re-entrant corner.

Exercise 2.2.18
Let Ω = (r, ϕ) : 0 < r < 1, 0 < ϕ < 3π
© ª
2 be a three-quarter disk. Find its Dirichlet eigen-
functions by separation of variables, and show that they do not lie in H 2 (Ω).

Still, Dirichlet eigenfunctions on Lipschitz domains are continuous up to the boundary. The
proof of this result uses a different set of ideas from the usual boundary regularity. We present
them below.9

Sketch of the proof of Theorem 2.2.1, part (iv)

First, one can show that a Dirichlet eigenfunction u ∈ L ∞ (Ω). One way to do it is due to
Moser [Mos60] (the so-called Moser iteration method), see [GilTru01, Th. 8.15]. Another
approach uses the fact that the heat kernel (to be defined in Chapter 6) in Ω at any fixed
positive time is bounded, see [Dav89, Example 2.1.8].
Let B be a ball containing Ω. Let us extend u to B by zero, and denote this extension
ue. Observe that due to the boundedness of u the extension ue ∈ L p (B ) for any p . Let us
solve the Dirichlet problem −∆θ = λue in B with θ|∂B = 0. By an L p analogue of the local
elliptic regularity Theorem 2.2.4, it follows that θ ∈ W 2,p (B ) (see [GilTru01, Theorem
9.15]; here W 2,p is an L p analogue of the Sobolev space H 2 = W 2,2 ). In particular, by the
Sobolev embedding theorem ([Eva10, §5.6.3]), θ ∈ C 1 (B ).
Consider now the unique harmonic function h in Ω such that h − θ ∈ H01 (Ω). In
other words h is a weak solution of the Dirichlet problem −∆h = 0 in Ω and h = θ³ on´
∂Ω. Since ∂Ω is Lipschitz, all its boundary points are regular in the sense that h ∈ C Ω
and the boundary values of h are given by θ (see [HeiKilMar93, Theorems 6.31 and 6.27],
[ArmGar01, Theorem 6.6.15]).
Finally, set v = θ − h . Note that v ∈ H01 (Ω), while −∆v = λu in Ω. Since −∆u = λu
in Ω, we conclude that −∆(v −u) = 0 in Ω for v −u ∈ H01 (Ω), and hence u = v = θ −h in

9 We are grateful to Dorin Bucur for outlining this argument.

62 Chapter 2. The spectral theorems

Ω. Since θ, h ∈ C (Ω) and θ = h on ∂Ω, we obtain that u ∈ C (Ω), and it vanishes on ∂Ω.

§2.2.6. Regularity of Neumann eigenfunctions

Consider now the Neumann Laplacian. We have

Theorem 2.2.19: Elliptic regularity in the Neumann case

Let Ω ⊂ Rd be a bounded domain with a Lipschitz boundary. Then

(i) The eigenfunctions of the weak Neumann spectral problem (2.1.20) are C ∞ in Ω.
Moreover, they are real analytic in Ω.

(ii) The eigenfunctions are smooth up to the boundary near the smooth parts of the
boundary.

The local regularity has been already established in §2.2.2. The proof of the global regularity
for the Neumann problem is essentially the same as that for the Dirichlet problem. One observes
that an a priori estimate in H 1 Rd+ still holds due to ellipticity assumption and that H k Rd+ is
¡ ¢ ¡ ¢

invariant under translations or differentiation along the boundary (see (2.2.10)), and proceeds in
the same way. Moreover, if Ω is smooth, any eigenfunction u ∈ H k (Ω) for all k . It follows that
the eigenfunctions are smooth near the smooth parts of the boundary.

Remark 2.2.20
We can additionally deduce that for a bounded domain Ω with a Lipschitz boundary,
every weak Neumann eigenfunction u ∈ H 1 (Ω) corresponding to an eigenvalue λ in fact
belongs to the Sobolev space H 3/2 (Ω). To do so, consider an auxiliary problem
(
−∆v = λu in Ω,
v =0 on ∂Ω.

This is an inhomogeneous Dirichlet problem for v , and by [JerKen95, Theorem B, part 2]

we have v ∈ H 3/2 (Ω). Then by [ChWGLS12, Lemma A.10], ∂n v ∈ L 2 (∂Ω). Set w = v −u ,
then w solves (
−∆w = 0 in Ω,
2
∂n w = ∂n v ∈ L (∂Ω) on ∂Ω.

Hence by [JerKen81], w ∈ H 3/2 (Ω), which implies u ∈ H 3/2 (Ω). We also note that the
weak Dirichlet eigenfunctions belong to H 3/2 (Ω) as follows directly from [JerKen95].
§2.2. Elliptic regularity and strong spectral theorems 63

§2.2.7. Strong spectral theorems

Elliptic regularity Theorems 2.2.1, 2.2.19, and 2.2.17, together with the weak spectral Theorems
2.1.20, 2.1.36, and 2.1.39, immediately imply that subject to some assumptions on the regularity of
the boundary, where applicable, the eigenvalues and eigenfunctions of the corresponding weak
spectral problems in fact satisfy the relevant equations and boundary conditions in the strong
sense. More precisely, we have

Theorem 2.2.21: Strong spectral theorem

Consider one of the following eigenvalue problems:

• The Dirichlet eigenvalue problem

(
−∆u = λu in Ω,
(2.2.12)
u=0 on ∂Ω,

for a bounded Lipschitz domain Ω ⊂ Rd .

• The Neumann eigenvalue problem

(
−∆u = λu in Ω,
∂n u = 0 on ∂Ω,

for a smooth bounded domain Ω ⊂ Rd .

• The Laplace–Beltrami eigenvalue problem

−∆g u = λu,

for a closed Riemannian manifold (M , g ).

Then the eigenvalues and the eigenfunctions of each eigenvalue problem understood
in the strong sense (i.e., the eigenvalue equations and boundary conditions are satisfied
pointwise) coincide with those of the corresponding weak eigenvalue problem.

One can also show that the same result holds for the Dirichlet and Neumann eigenvalue prob-
lems on a compact Riemannian manifold with boundary, see [Tay11, sections 5.1 and 5.7].

Remark 2.2.22
A Neumann eigenfunction u of a Lipschitz domain Ω can be thought to satisfy the Neu-
mann condition pointwise almost everywhere in the following sense. Given a boundary
point x ∈ ∂Ω at which the normal derivative exists, consider a sequence of points y i ∈ Ω
64 Chapter 2. The spectral theorems

which approach x nontangentially. Then ∂n u(x) = lim n, ∇u(y i ) = 0. We refer to

­ ®
y i →x
[JerKen81, ChWGLS12] for details including the formal definition of nontangential con-
vergence.

The immediate corollary of Theorem 2.2.21 is that in each case the “strong” spectrum is dis-
crete, consists of eigenvalues of finite multiplicity accumulating only to +∞, and the eigenfunc-
tions can be chosen to form a basis in L 2 (Ω) or L 2 (M ), as appropriate.
It is important to emphasise that a restriction on the smoothness of the boundary in the Eu-
clidean case is essential. We assume the boundary to be Lipschitz which is not optimal and can be
slightly relaxed at a cost of extra technicalities — but this condition cannot be omitted altogether.
We have already remarked that dropping it in the Neumann problem may lead to undesirable con-
sequences even in the weak form: the spectrum may no longer be discrete. Although the weak
Dirichlet spectral theorem works without any restrictions on the smoothness of the boundary,
this may not be the case for the strong one, as the following example indicates.

Example 2.2.23
Consider the eigenvalue problem (2.2.12) in a punctured disk Ω = D \{0}. The weak eigen-
values and eigenfunctions of this problem are the same as for the whole disk, see Example
1.1.17. However, the eigenfunctions J 0 ( j 0,k r ) do not satisfy the boundary condition at the
origin in the strong (pointwise) sense.
 CHAPTER 3
Variational principles and
applications
In this chapter, we introduce the variational principles for
eigenvalues and discuss their applications. These include domain
monotonicity, Dirichlet–Neumann bracketing, and Weyl’s law for
general domains. Along the way, we also introduce the Robin and
Zaremba eigenvalue problems and consider some applications of
variational principles on symmetric domains. We also prove the
Friedlander–Filonov inequalities between Dirichlet and Neumann
eigenvalues, the Berezin–Li–Yau inequalities, and discuss Pólya’s
conjecture.

§3.1. Variational principles for Laplace eigenvalues

§3.1.1. The Rayleigh quotient

Let H be a real (or complex) Hilbert space with an inner product (·, ·)H . Consider a symmetric
bilinear (respectively, sesquilinear) form Q[u, v], Q : U ×U → R, defined on a dense linear sub-
space U =: Dom(Q) of H , which we from now on refer to as the domain of the form Q . Of
particular importance to us is the corresponding quadratic form Q[u, u], u ∈ Dom(Q).

Definition 3.1.1: Semi-bounded quadratic form

We say that the quadratic form Q[u, u] is semi-bounded from below if there exists a con-
stant c ∈ R, such that

Q[u, u] ≥ c(u, u)H for all u ∈ Dom(Q).

65
66 Chapter 3. Variational principles and applications

In what follows, we assume that U = Dom(Q) is complete in the norm induced by the inner
product
(u, v)U := Q[u, v] + (1 − c)(u, v)H . (3.1.1)
Consider an abstract eigenvalue problem for a symmetric semi-bounded from below bilinear form
Q : we are looking for λ ∈ R and u ∈ Dom(Q) \ {0} such that

Q[u, v] = λ(u, v)H for all v ∈ Dom(Q). (3.1.2)

Assume in addition that the embedding U ,→ H is compact (here U is endowed with the norm
induced by (3.1.1)). Then all the eigenvalues of (3.1.2) are of finite multiplicity, their sequence may
have an accumulation point only at +∞, and the corresponding eigenfunctions may be chosen
to form an orthogonal basis in H (see [Ban80, §III.1.2]). We enumerate the eigenvalues of (3.1.2)
in non-decreasing order with account of multiplicities as

λ1 (Q) := λ1 ≤ λ2 ≤ . . . .

The basic examples are the forms Q D and Q N for the weak Dirichlet spectral problem (2.1.12)
and the weak Neumann spectral problem (2.1.20), respectively, in a bounded Euclidean domain
domain Ω (which we assume to be Lipschitz in the Neumann case). These forms are defined by
the same differential expression
ˆ
D N
Q Ω [u, v] = Q Ω [u, v] := (∇u, ∇v)L (Ω) = 〈∇u, ∇v〉 dx,
2 (3.1.3)
Ω

and act in the same Hilbert space H = L 2 (Ω), but have different domains: Dom(Q D ) = H01 (Ω),
and Dom(Q N ) = H 1 (Ω) .
The following simple result is often useful, see Remark 3.1.21 below for particular applications.

Proposition 3.1.2
Let {u j } be a basis of eigenfunctions of the eigenvalue problem (3.1.2), chosen to be or-
thogonal in H . Then distinct eigenfunctions are also orthogonal in U equipped with the
inner product (3.1.1).

Proof
Take u = u j and v = u k in (3.1.1) and (3.1.2), with k ̸= j . Then

u j , u k U = Q[u j , u k ] + (1 − c) u j , u k H = (λ j + 1 − c) u j , u k H = 0.
¡ ¢ ¡ ¢ ¡ ¢

For each u ∈ Dom(Q) \ {0}, we define its Rayleigh quotient

Q[u, u]
R[u] := . (3.1.4)
∥u∥2H
§3.1. Variational principles for Laplace eigenvalues 67

Then the following variational (or min-max) principle (variously associated in the literature with
the names of Lord Rayleigh, W. Ritz, R. Courant, and H. Poincaré, among others) for the eigen-
values of the weak spectral problem (3.1.2) holds.

Proposition 3.1.3: The variational principle for a quadratic form [Dav95, §4.5],
[Ban80, §III.1.2]
We have
λk (Q) = min max R[u], k ∈ N. (3.1.5)
L ⊂Dom(Q) u∈L \{0}
dim L =k
John William Strutt,
3rd Baron Rayleigh
Remark 3.1.4 (1842–1919)
We will use the following additional properties of the variational principle.

(i) For the principal eigenvalue λ1 , (3.1.5) becomes

λ1 = min R[u]. (3.1.6)

u∈Dom(Q)\{0}

Any given u 0 ∈ Dom(Q) \ {0} becomes a test function for λ1 in the sense that

λ1 ≤ R[u 0 ]. Walther Heinrich

Wilhelm Ritz
(ii) If u 1 , . . . , u k−1 are eigenvectors of the weak spectral problem (3.1.2) corresponding (1878–1909)
to the eigenvalues λ1 , . . . , λk−1 , and Lk−1 := Span{u 1 , . . . , u k−1 }, then (3.1.5) with
k ≥ 2 is equivalent to
λk = min R[u]. (3.1.7)
u∈Dom(Q)\{0}
u⊥L k−1

The minimum in (3.1.6) and (3.1.7) is attained by u if and only if u is an eigenvector

of (3.1.2) corresponding to the eigenvalues λ1 and λk , respectively. The minimum
in (3.1.5) is attained by L = Span{u 1 , . . . , u k }, however it may not be the only min-
imiser, see [Ste70].

Remark 3.1.5
If Q[u, v] = (Au, v)H is a bilinear form associated with a non-negative self-adjoint op-
¡p ¢
erator A , such as the Dirichlet or Neumann Laplacian, one has Dom(Q) = Dom A
(see [Dav95, §4.4]), and one can replace Dom(Q) in the variational principles above
by Dom(A), replacing at the same time min and max by inf and sup, respectively, see
[Dav95, Theorem 4.5.3].

Let us illustrate the idea of the proof of the abstract Proposition 3.1.3. Let u 1 , u 2 , . . ©. , beª the
∞
eigenfunctions of Q chosen to form an orthonormal basis in H . By Proposition 3.1.2, u j j =1
68 Chapter 3. Variational principles and applications

is also an orthogonal basis in U = Dom(Q) with respect to the inner product (3.1.1). Let us take
u ∈ Dom(Q) and expand it in this basis,
∞
αj u j , α j = u, u j H .
X ¡ ¢
u=
j =1

Then it is easy to see that

∞
λ j |α j |2
P
Q[u, u] j =1
R[u] = = ∞ ,
∥u∥2H P
|α j |2
j =1

and Proposition 3.1.3 follows immediately.

Exercise 3.1.6
Use the method outlined above to prove Proposition 3.1.3 for Q[u, v] := 〈Au, v〉, where
A is a Hermitian d × d matrix acting in Rd .

Exercise 3.1.7
Show that the eigenvectors of the weak spectral problem (3.1.2) are the critical points of
the functional u 7→ Q[u, u] subject to the constraint ∥u∥H = 1, with the corresponding
critical values being the eigenvalues of (3.1.2). We refer to [Lau12, Chapter 9] for a solution.

The following comparison principle immediately follows from Proposition 3.1.3.

Proposition 3.1.8: Abstract eigenvalue comparison principle

Let Q1 and Q2 be two bilinear forms as above such that Dom(Q2 ) ⊆ Dom(Q1 ) and

Q1 [u, u] ≤ Q2 [u, u] for all u ∈ Dom(Q2 ).

Then the eigenvalues of the corresponding weak spectral problems satisfy

λk (Q1 ) ≤ λk (Q2 ) for all k ∈ N.

Simply speaking, Proposition 3.1.8 states that either narrowing the domain of a quadratic
form or increasing the value of the form may only push all the eigenvalues up but not down.
§3.1. Variational principles for Laplace eigenvalues 69

§3.1.2. Variational principles

Let Ω ⊂ Rd be a bounded open set, and consider the weak spectral problem for the Dirichlet
Laplacian −∆D Ω on Ω. As was mentioned above, the corresponding quadratic form has the domain
H01 (Ω) and is given by ˆ
Q D [u, u] = ∥∇u∥2L 2 (Ω) = |∇u|2 dx.
Ω
Hence, its Rayleigh quotient is
´
|∇u|2 dx
∥∇u∥2L 2 (Ω) Ω
R[u] = = ´ , u ∈ H01 (Ω) \ {0}.
∥u∥2L 2 (Ω) u 2 dx
Ω

Proposition 3.1.3 then leads to the variational characterisation of the eigenvalues of −∆D
Ω.

Theorem 3.1.9: The variational principle for the eigenvalues of the Dirichlet
Laplacian
Let λk = λD
k
(Ω), k ∈ N, be the eigenvalues of the Dirichlet Laplacian −∆D
Ω on a bounded
open set Ω ⊂ Rd . Then

∥∇u∥2L 2 (Ω)
λk = min max , k ∈ N. (3.1.8)
L ⊂H01 (Ω) u∈L \{0} ∥u∥2L 2 (Ω)
dim L =k

If additionally Lk−1 := Span{u 1 , . . . , u k−1 } is the linear subspace of H01 (Ω) spanned by
the first k − 1 eigenfunctions of −∆DΩ , then we also have

∥∇u∥2L 2 (Ω)
λk = min , k ∈ N. (3.1.9)
u∈H01 (Ω)\{0} ∥u∥2L 2 (Ω)
u⊥L k−1

The minimum in (3.1.9) is attained by u if and only if u is an eigenfunction of −∆D

Ω cor-
responding to λk .

Exercise 3.1.10
Finish the proof of Theorem 3.1.9 using the weak Dirichlet spectral Theorem 2.1.20.

For the Neumann Laplacian, the Rayleigh quotient is the same as in the Dirichlet case, and
we have a direct analogue of Theorem 3.1.9, the only difference being that the space H01 (Ω) should
be replaced by H 1 (Ω). Note that the Neumann spectrum always starts with the eigenvalue µ1 =
λN
1 (Ω) = 0, for which the corresponding eigenfunction is a constant.
70 Chapter 3. Variational principles and applications

Theorem 3.1.11: The variational principle for the eigenvalues of the Neumann
Laplacian
Let µk = λN
k
(Ω), k ∈ N, be the eigenvalues of the Neumann Laplacian −∆N
Ω on a bounded
d
open set Ω ⊂ R with Lipschitz boundary. Then

∥∇u∥2L 2 (Ω)
µk = min max , k ∈ N. (3.1.10)
L ⊂H 1 (Ω) u∈L \{0} ∥u∥2L 2 (Ω)
dim L =k

If additionally Lk−1 := Span{u 1 , . . . , u k−1 } is the linear subspace of H 1 (Ω) spanned by

the first k − 1 eigenfunctions of −∆NΩ , then we also have

∥∇u∥2L 2 (Ω)
µk = min , k ∈ N,
u∈H 1 (Ω)\{0} ∥u∥2L 2 (Ω)
u⊥L k−1

and in particular
∥∇u∥2L 2 (Ω)
µ2 = min . (3.1.11)
u∈H 1
´ (Ω)\{0} ∥u∥2L 2 (Ω)
Ω
u dx=0

The minima in (3.1.10) and (3.1.11) are attained by u if and only if u is an eigenfunction of
−∆N
Ω corresponding to µk and µ2 , respectively.

Remark 3.1.12
In practice, one can replace Dom(Q) in (3.1.5) by its dense subspace, simultaneously re-
placing min by inf and max by sup. In particular, H01 (Ω) appearing in Theorem 3.1.9 can
be replaced by C 0∞ (Ω), and H 1 (Ω) appearing in Theorem 3.1.11 can be replaced by C ∞ (Ω),
see also Appendix A.

Finally, the case of the Laplace–Beltrami operator on a smooth closed Riemannian manifold
(M , g ) is essentially identical to that of the Neumann Laplacian. We however have to remember
our notational convention 2.1.41 on the enumeration of the eigenvalues of the Laplace–Beltrami
operator on a closed Riemannian manifold.

Theorem 3.1.13: The variational principle for the eigenvalues of the the Laplace–
Beltrami operator on a closed Riemannian manifold
Let 0 = λ0 (M ) < λ1 (M ) ≤ . . . , be the eigenvalues of the Laplace–Beltrami operator −∆M
§3.1. Variational principles for Laplace eigenvalues 71

on a smooth closed Riemannian manifold M := (M , g ). Then

∥∇u∥2L 2 (M )
λk = min max , k ∈ N0 .
L ⊂H 1 (M ) u∈L \{0} ∥u∥2L 2 (M )
dim L =k+1

If additionally Lk := Span{u 0 = 1, . . . , u k−1 } is the linear subspace of H 1 (M ) spanned by

the first k eigenfunctions of −∆M , then we also have

∥∇u∥2L 2 (M )
λk = min , k ∈ N, (3.1.12)
u∈H 1 (M )\{0} ∥u∥2L 2 (M )
u⊥L k

and in particular
∥∇u∥2L 2 (M )
λ1 = min . (3.1.13)
u∈H 1
´ (M )\{0} ∥u∥2L 2 (M )
M
u dV =0

The minima in (3.1.12) and (3.1.13) are attained by u if and only if u is an eigenfunction of
−∆M corresponding to λk and λ1 , respectively.

Exercise 3.1.14
Let Σ be a smooth surface, and let g 1 and g 2 be two Riemannian metrics on Σ which are
conformally equivalent, i.e., g 1 = α(x)g 2 for some smooth positive function α. Show
that the Dirichlet energy is conformally invariant, i.e., that
ˆ ˆ
¯ ¯2
¯∇ f ¯ dVg = ¯∇ f ¯2 dVg .
¯ ¯
g1 1 g2 2
(3.1.14)
Σ Σ

§3.1.3. The Robin and Zaremba problems

As we have briefly mentioned in Remark 1.1.20, one can consider other types of boundary condi-
tions for the Laplacian apart from the Dirichlet and Neumann ones. We now discuss some of the
many possible generalisations.
Let Ω ⊂ Rd be a bounded domain with a Lipschitz boundary. Fix a parameter γ ∈ R, and
consider the spectral problem
(
−∆u = λu in Ω,
(3.1.15)
∂n u + γu = 0 on ∂Ω.

The boundary condition in (3.1.15) is known as the Robin condition, and the problem (3.1.15) as the
Robin spectral problem (see [GusAbe98] for a fascinating historical investigation into the origins
of this terminology). We note that for γ = 0 the Robin condition becomes the Neumann one.
72 Chapter 3. Variational principles and applications

Acting as in §2.1.7 for the Neumann Laplacian, we can construct the Robin Laplacian −∆R,γ
as the Friedrichs extension with the domain

Dom(−∆R,γ ) = u ∈ H 1 Ω) : −∆u ∈ L 2 (Ω) and ∂n u + γu ∼ 0 on ∂Ω ,

© ª

where the condition ∂n u + γu ∼ 0 is understood in the sense

ˆ ˆ ˆ
∆u v dx + 〈∇u, ∇v〉 dx + γuv ds = 0
Ω Ω ∂Ω

for all v ∈ H 1 (Ω) (see [AreCSVV18, §7.5]). The corresponding bilinear form is given by

Q R,γ [u, v] = −∆R,γ u, v L 2 (Ω) = (∇u, ∇v)L 2 (Ω) + γ(u, v)L 2 (∂Ω) ,
¡ ¢
(3.1.16)

and has the same form domain H 1 (Ω) as the Neumann Laplacian; the corresponding quadratic
form is obviously semi-bounded from below by zero for γ ≥ 0. For each fixed γ ≥ 0, the spectrum
of the Robin Laplacian is discrete and consists of eigenvalues
R,γ R,γ
0 ≤ λ1 ≤ λ2 ≤ . . . ,

accumulating to +∞ which can be found from the variational principle analogous to (3.1.10),

R,γ
∥∇u∥2L 2 (Ω) + γ∥u∥2L 2 (∂Ω)
λk (Ω) = min max , k ∈ N. (3.1.17)
L ⊂H 1 (Ω) u∈L \{0} ∥u∥2L 2 (Ω)
dim L =k

Taking in (3.1.17) k = 1 and L = Span{1} we immediately obtain the bound

R,γ Vold −1 (∂Ω)

λ1 (Ω) ≤ γ . (3.1.18)
Vold (Ω)

Remark 3.1.15
It can be shown using a Sobolev trace inequality [Gri11, Theorem 1.5.1.10], that the Robin
Laplacian is semi-bounded from below also for γ < 0, see [AreCSVV18, Theorem 7.15] for
details. It is then not hard to check that the variational formula (3.1.17) holds for γ < 0 as
R,γ
well, see [BucFreKen17, formula (4.5)]. The principal eigenvalue λ1 (Ω) is negative for
γ < 0; moreover, inequality (3.1.18) still holds.

Exercise 3.1.16
Write down transcendental equations whose roots are the eigenvalues of −∆R,γ for the
interval (0, L) ⊂ R1 .
§3.1. Variational principles for Laplace eigenvalues 73

Numerical Exercise 3.1.17

By separating the variables in polar coordinates, write down transcendental equations
whose roots are the eigenvalues of −∆R,γ for the unit disk D, and hence reproduce Fig-
ure 3.1.

Figure 3.1: Some eigenvalues of the Robin Laplacian −∆R,γ

for the unit disk as functions of γ. The dashed black curves
correspond to single eigenvalues, and the solid curves to dou-
ble eigenvalues. The horizontal dotted lines are placed at the
ordinates coinciding with the Dirichlet eigenvalues of the unit
disk.

Exercise 3.1.18
Note that the scaling for the Robin eigenvalues is not the same as in the Dirichlet and
Neumann cases, cf. Lemma 2.1.30 and Exercise 2.1.38. Namely, prove that for a scaled
copy Ωρ , ρ > 0, of a Lipschitz domain Ω ⊂ Rd and for j ∈ N we have

R,γ 1 R,ργ
λ j (Ωρ ) = λ (Ω).
ρ2 j

We will shortly obtain further bounds on Robin eigenvalues. For the moment, we observe
only that for any fixed k ∈ N,
R,γ
lim λk (Ω) = λDk (Ω). (3.1.19)
γ→+∞

We omit a formal proof of this fact (see [BucFreKen17, Proposition 4.5]), but it can be easily
 74 Chapter 3. Variational principles and applications

deduced from the variational principle (3.1.17): for very large γ the minimisation procedure elim-
inates the dominant term γ∥u∥2L 2 (∂Ω) in the numerator of the Rayleigh quotient, thus forcing
u|∂Ω = 0.

Remark 3.1.19
An alternative approach to the Robin problem (3.1.15) is to consider λ as a given parame-
ter, and to treat γ (or, more precisely, σ = −γ) as a spectral parameter. This is the spectral
problem for the so-called Dirichlet-to-Neumann map which we study extensively in Chap-
ter 7.

We will also need to consider spectral problems with mixed Dirichlet–Neumann boundary
conditions, often called Zaremba problems, which first appeared in [Zar10]. Let Ω be a bounded
domain in Rd with a Lipschitz boundary ∂Ω which we decompose into the Dirichlet boundary
Γ := ∂D Ω and the Neumann boundary ∂N Ω := ∂Ω \ Γ. To avoid unnecessary complications we
assume that each of ∂D,N Ω consists of a finite number of connected components and that the
interface between the two parts, ∂D Ω ∩ ∂N Ω, is sufficiently regular for d ≥ 3, see [OttBro13] for
more precise conditions. We consider a mixed Dirichlet–Neumann spectral problem

−∆u = λu in Ω,
Stanisław Zaremba 

(1863—1942)
u=0 on ∂D Ω, (3.1.20)

∂ u = 0

on ∂N Ω.
n

Obviously, if Γ = ∂Ω we get the standard Dirichlet problem, and if Γ = ; — the standard

Neumann problem.
To give an operator-theoretic form of (3.1.20) and to obtain its variational formulation, we
first define the space
∞
C 0,Γ (Ω) := {u ∈ C ∞ (Ω) : supp u ∩ Γ = ;},
1
and then the Sobolev space H0,Γ ∞
(Ω) as the completion of C 0,Γ (Ω) in the H 1 (Ω) norm. Then
the Zaremba (or mixed Dirichlet–Neumann) Laplacian −∆ZΩ;Γ = −∆ZΓ can be defined via the
Friedrichs extension with the domain

Dom −∆ZΓ = u ∈ H0,Γ

1
(Ω) : ∆u ∈ L 2 (Ω) and ∂n u ∼ 0 on ∂N Ω ,
¡ ¢ © ª

1
where the last condition is understood in the sense that (2.1.8) holds for any v ∈ H0,Γ (Ω). It
Z
is easy to check that the bilinear form Q [u, v] corresponding to the weak Zaremba problem
coincides with the bilinear form for the Dirichlet and Neumann Laplacians, with the difference
that its domain is given by Dom(Q Z ) = H0,Γ
1
(Ω). Hence, the eigenvalues λZk (Ω, Γ) of −∆ZΓ can be
determined from the variational principle

∥∇u∥2L 2 (Ω)
λZk (Ω, Γ) = min max , k ∈ N, (3.1.21)
L ⊂H0,Γ
1
(Ω) u∈L \{0} ∥u∥2L 2 (Ω)
dim L =k
§3.2. Consequences of variational principles 75

which is identical to the Dirichlet variational principle (3.1.8) with H01 (Ω) replaced by H0,Γ
1
(Ω).

Exercise 3.1.20

(i) Find the eigenvalues of the one-dimensional mixed Laplacian on the interval (0, L)
with the Dirichlet condition imposed at one end and the Neumann one at the other.

(ii) Use (i) to find the eigenvalues of the Zaremba Laplacian −∆ZΓ in the unit square in
the following cases:

(a) Γ is a single side of the square;

(b) Γ is the union of two adjacent sides;
(c) Γ is the union of two opposite sides;
(d) Γ is the union of three sides of the square.

Remark 3.1.21
Let {u j } is a basis of eigenfunctions of either Dirichlet, Neumann, or Zaremba Laplacian
in a bounded domain Ω ⊂ R¡d , chosen to ¢
be orthogonal in L 2 (Ω). It immediately follows
from Proposition 3.1.2 that ∇u j , ∇u k L 2 (Ω) = 0 for j ̸= k , and therefore distinct eigen-
functions are also orthogonal in H 1 (Ω). This is however not true for the eigenfunctions
of the Robin Laplacian −∆R,γ with γ ̸= 0.

§3.2. Consequences of variational principles

§3.2.1. Domain monotonicity and Dirichlet-Neumann bracketing

We start with the following simple but immensely important application of the variational prin-
ciple for the Dirichlet Laplacian.

Theorem 3.2.1: Domain monotonicity for the Dirichlet Laplacian

Let Ω1 ⊂ Ω2 be two bounded domains. Then their Dirichlet eigenvalues satisfy λD
k
(Ω2 ) ≤
D
λk (Ω1 ) for all k ∈ N.

Proof
We have a natural embedding H01 (Ω1 ) ⊂ H01 (Ω2 ): if u ∈ H01 (Ω1 ), extending u by zero onto
Ω2 we obtain a function ue ∈ H01 (Ω2 ). Moreover R Ω1 [u] = R Ω2 [u]
e . The result then follows
immediately from Proposition 3.1.8.
76 Chapter 3. Variational principles and applications

Proposition 3.2.2: Strict domain monotonicity for the Dirichlet Laplacian

Let Ω Ú Ω e ⊂ Rd be two bounded domains such that Ω e \ Ω contains an open set. Then
D e D
their Dirichlet eigenvalues satisfy λk (Ω) < λk (Ω) for all k ∈ N.

Proof
This was first observed in [CouHil89, footnote on p. 409]. We mostly follow the argument
in [Wel72]. Firstly, by non-strict domain monotonicity Theorem 3.2.1 we have λD k
(Ω)
e ≤
D
λk (Ω) for all k ∈ N. Suppose, for contradiction with the statement of proposition, that
for some number k ,
λ := λD D
k (Ω) = λk (Ω).
e (3.2.1)
Since the spectrum of the Dirichlet Laplacian −∆De is unbounded above, there exists m ∈ N
Ω
such that
λDm (Ω) > λ.
e (3.2.2)
Choose a nested sequence of m domains

Ω =: Ω
e1 Ú Ω
e2 Ú ··· Ú Ω
e m := Ω,
e

such that Ω
e i +1 \ Ω
e i contains an open set, i = 1, . . . , m −1, see Figure 3.2. By domain mono-
tonicity and (3.2.1),
λ = λD D e D
k (Ω) ≤ λk (Ωi ) ≤ λk (Ω) = λ,
e

and therefore λDk

(Ω
e i ) = λ for all i = 1, . . . , m .
Let u i ∈ H0 (Ωi ) be an eigenfunction of −∆De corresponding to the eigenvalue λ, and
1 e
Ωi
let uei ∈ H01 (Ω) e . We claim that the set {uei }m is linearly
e be its extension by zero onto Ω
i =1
independent. Indeed, suppose that
m
αi uei
X
f := (3.2.3)
i =1

is identically zero in Ω
e for some coefficients α1 , . . . , αm ∈ R. The restriction of f to Ω
em \
Ω
e m−1 is equal to αm uem = αm u m , and since the eigenfunction u m cannot vanish on an
open subset by real domain analyticity, we have αm = 0. We therefore have a shorter linear
combination f ; repeating the argument we at the end conclude that αm = αm−1 = · · · =
α1 = 0. Thus for L = Span{uei }m i =1 we have dim L = m .
§3.2. Consequences of variational principles 77

Let now f ∈ L be given by (3.2.3), and let us evaluate its Rayleigh quotient. We have
à !
m iX
−1
∥2L 2 (Ω) α2i ∥∇u i ∥2L 2 (Ω αi α j ∇u i , ∇u j L 2 (Ω
X ¡ ¢
∥∇ f = ei) + 2 j)
e e
i =1 j =1
à !
m iX
−1
α2i (−∆u i , u i )L 2 (Ω αi α j −∆u i , u j L 2 (Ω
X ¡ ¢
= ei) + 2 e j)
i =1 j =1
à !
m iX
−1
=λ α2i ∥u i ∥2L 2 (Ω αi α j u i , u j L 2 (Ω = λ∥ f ∥2L 2 (Ω)
X ¡ ¢
ei) + 2 e j) e ,
i =1 j =1

and therefore R Ωe [ f ] = λ for all f ∈ L . Thus, by the variational principle λD

m (Ω) ≤ λ,
e
which contradicts (3.2.2), and our assumption (3.2.1) is incorrect.

Figure 3.2: A nested sequence of domains appearing in the

proof of Proposition 3.2.2.

Remark 3.2.3

(i) For Dirichlet eigenfunctions for domains in Riemannian manifolds, the analogue of
Proposition 3.2.2 holds as well, where the non-vanishing on open sets follows from
the Aronszajn unique continuation property [Aro57], see also Remark 4.1.14.

(ii) Strict domain monotonicity does not hold for disconnected sets. For example, if
Ω = Ω1 ⊔ Ω2 , then
λD D D
1 (Ω) = min λ1 (Ω1 ), λ1 (Ω2 ) .
© ª
78 Chapter 3. Variational principles and applications

Exercise 3.2.4
Use domain monotonicity and Exercise 1.2.21 to find explicit two-sided estimates, in terms
of d = 2, 3, . . . , for the first positive zero j d −1,1 of the Bessel function J d −1 (x).
2 2

Exercise 3.2.5
Show that domain monotonicity does not generally hold for Neumann eigenvalues. Hint:
compare Neumann eigenvalues of a square and of a thin rectangle inscribed along a diag-
onal of the square, see [Lau12].

Example 3.2.6
Despite the result of Exercise 3.2.5, there are particular situations when Neumann domain
monotonicity holds and can be once more deduced from Proposition 3.1.8. Consider a
family of planar domains

Ω f := {(x, y) : 0 < x < 1, − f (x) < y < f (x)},

where f is a positive Lipschitz continuous function on (0, 1) such that ∂Ω f is Lipschitz

as well. Fix any such function f , and a number ρ > 1; obviously Ω f ⊂ Ωρ f , see Figure 3.3.
We claim that
λN N
k (Ωρ f ) ≤ λk (Ω f ), for all k ∈ N. (3.2.4)
Indeed, we first of all can establish a bijection between the spaces H 1 (Ω f ) and H 1 (Ωρ f )
by identifying u ∈ H 1 (Ω f ) with ue := u(x, ρy) ∈ H 1 (Ωρ f ). Moreover, a simple change of
variables shows the monotonicity of the Rayleigh quotients:
¶2
∂u 1 ∂u 2
Ï µµ µ ¶ ¶
+ dx dy
∂x ρ 2 ∂y
Ωf
R Ωρ f [u]
e = Ï ≤ R Ω f [u].
2
u dx dy
Ωf

Then (3.2.4) follows from a simple re-wording of Proposition 3.1.8.

Exercise 3.2.7
As a particular oapplication of Example 3.2.6, consider the ellipse E ρ :=
y2
n
(x, y) : x 2 + ρ 2 < 1 , ρ > 1. By using the above construction and recalling Lemma 2.1.30
§3.2. Consequences of variational principles 79

Figure 3.3: An example of the domains Ω f (shaded) and Ω2 f ,

with λN
k
(Ω2 f ) ≤ λN
k
(Ω f ), k ∈ N.

as well as Exercise 2.1.38, prove that

1 N
λ (D) ≤ λN N
k (E ρ ) ≤ λk (D) for all k ∈ N.
ρ2 k

Note that similar inequalities hold for the eigenvalues of the Dirichlet Laplacian in E ρ
2
directly by domain monotonicity D ⊂ E ρ ⊂ B 0,ρ and Lemma 2.1.30.

Numerical Exercise 3.2.8

Verify the inequalities in Exercise 3.2.7 for the first few k and ρ = 2 numerically.

Another important corollary of Proposition 3.1.8 is the following result establishing the in-
equalities between the eigenvalues of the Dirichlet and Robin Laplacians in the same region.

Theorem 3.2.9
Let Ω ⊂ Rd be a bounded open set with a Lipschitz boundary, and let γ2 ≥ γ1 . Then
R,γ1 R,γ2
λk ≤ λk ≤ λD
k for k ∈ N.
80 Chapter 3. Variational principles and applications

Proof
The inequality between the eigenvalues of the Robin Laplacians follows directly from
Proposition 3.1.8: they have the same domains, and the quadratic form (3.1.16) is monotone
increasing in γ. To establish the inequality between the Robin and the Dirichlet eigenval-
ues, we re-write the Dirichlet quadratic form as
¡ D
−∆ u, u L 2 (Ω) = −∆R,γ u, u L 2 (Ω)
¢ ¡ ¢

for any u ∈ H01 (Ω) and any γ ∈ R since in this case u|∂Ω = 0, and use the fact that H01 (Ω) ⊂
H 1 (Ω).

Taking γ2 = 0 in Theorem 3.2.9 immediately implies the following

Corollary 3.2.10
Let Ω ⊂ Rd be a bounded open set with a Lipschitz boundary. Then λN
k
(Ω) ≤ λD
k
(Ω).

In fact, as we will show in §3.2.4, a much stronger inequality holds between the Dirichlet and
Neumann eigenvalues.
Let us now discuss the Dirichlet–Neumann bracketing. Informally, its idea is as follows: given
a Laplacian on a domain, adding some extra Dirichlet conditions yields higher eigenvalues, and
adding some extra Neumann conditions yields lower eigenvalues. Let us illustrate this by two
specific examples.
The first result illustrates the effect of changing the boundary conditions from Dirichlet to
Neumann (or vice versa) on a part of the boundary.

Proposition 3.2.11: Dirichlet–Neumann bracketing, version 1

Let Ω ⊂ Rd be a bounded domain with a Lipschitz boundary, and let Γ1 ⊂ Γ2 ⊂ ∂Ω. Then

λZk (Ω, Γ1 ) ≤ λZk (Ω, Γ2 ) for all k ∈ N.

This result follows immediately from the variational principle for a mixed eigenvalue problem
1 1
(3.1.21) and Proposition 3.1.8 with account of the inclusion H0,Γ 2
(Ω) ⊂ H0,Γ1
(Ω).
The second version illustrates the effect of adding Dirichlet or Neumann conditions on a
hypersurface inside the domain. Namely, let Ω ⊂ Rd be a bounded domain, and consider the
Dirichlet Laplacian −∆D Ω in Ω. Let Γ ⊂ Ω be a Lipschitz hypersurface. Let Ω = Ω \ Γ, so that
e
∂Ω = ∂Ω ∪ Γ, see Figure 3.4 for some possible configurations of Γ within Ω. (In particular, Γ
e
may separate Ω into two subdomains. This case will be particularly important, for example in
§3.2.2.) We consider the Dirichlet Laplacian −∆De , obtained from −∆D Ω by imposing the addi-
Ω
tional Dirichlet conditions on Γ, and the mixed Laplacian −∆ e on Ω, obtained from −∆D
Z e
Ω by
Ω,∂Ω
imposing the additional Neumann conditions on Γ and preserving the Dirichlet conditions on
§3.2. Consequences of variational principles 81

∂Ω ⊂ ∂Ω
e.

Proposition 3.2.12: Dirichlet–Neumann bracketing, version 2

In the geometry described above, we have

λZk (Ω,
e ∂Ω) ≤ λD (Ω) ≤ λD (Ω)
k k
e for all k ∈ N.

Figure 3.4: Three possible configurations of a hypersurface Γ inside Ω. On the left, Γ is

a closed hypersurface; in the middle, ∂Γ ⊂ ∂Ω; and on the right, ∂Γ ⊂ Ω

Remark 3.2.13
We note that for the middle and the right domains in Figure 3.4, the boundary part Γ
of Ω
e is not Lipschitz with respect to Ω e at the points of ∂Γ. Nevertheless, the extension
property, see Remark 2.1.8, still holds, and therefore all the operators are well-defined and
have discrete spectra.

Exercise 3.2.14

(i) Prove Proposition 3.2.12.

(ii) Prove a version of Proposition 3.2.12 in which some arbitrary combination of Dirich-
let, Neumann and Robin conditions is originally imposed on parts of ∂Ω.
(iii) Suppose that a Lipschitz domain Ω is partitioned into N disjoined Lipschitz do-
mains Ωn , n = 1, . . . , N , in the sense that Ω is the interior of the closure of the union
of Ωn , see Figure 3.5. Prove that
à !
N
λD D
k (Ω) ≤ λk Ωn , k ∈ N,
[
n=1
82 Chapter 3. Variational principles and applications

and à !
N
λN N
k (Ω) ≥ λk Ωn , k ∈ N.
[
n=1

Figure 3.5: An example of partitioning a domain into sub-

N
Ωn , the
S
domains. Note that in the spectral problems on n=1
boundary conditions are imposed both on the exterior and the
interior boundaries.

Remark 3.2.15
Imposing boundary conditions on sets of co-dimension two or higher does not affect the
eigenvalues. Indeed, such sets have zero capacity (see Definition 4.1.8), and hence do not
influence the spectrum (see [RauTay75]).

Exercise 3.2.16
Use domain monotonicity and Dirichlet–Neumann bracketing to derive two-sided esti-
mates on the first few Dirichlet and Neumann eigenvalues of the L -shaped domain and
the Π-shaped domain shown in Figure 3.6.
§3.2. Consequences of variational principles 83

Figure 3.6: The L -shaped domain constructed from three

unit squares and the Π-shaped domain constructed from five
unit squares

Numerical Exercise 3.2.17

Compute the first ten Dirichlet and Neumann eigenvalues for the L -shaped and the Π-
shaped domains and compare them with bounds you have derived in Exercise 3.2.16.

§3.2.2. Symmetry tricks

Let Ω be a Euclidean domain which is symmetric with respect to a hyperplane S . Consider a

Laplacian in Ω subject to some combination of Dirichlet, Neumann, and Robin boundary con-
ditions which are also imposed symmetrically with respect to S . It turns out that one can choose
a basis of eigenfunctions of the Laplacian on Ω in such a way that each eigenfunction is either
symmetric with respect to S (and therefore satisfies the Neumann condition on S ∩ Ω) or anti-
symmetric with respect to S (and therefore satisfies the Dirichlet condition on S ∩Ω). In this way,
the spectral problem for the Laplacian on Ω decomposes into two mixed problems on a half Ω′
of Ω lying to one side of S , with the Neumann and Dirichlet conditions, respectively, imposed
on S ∩ Ω, see Figure 3.7.
The spectral decomposition illustrated above follows from the following abstract result.

Theorem 3.2.18
Let A be a self-adjoint operator with a discrete spectrum acting in a Hilbert space H ,
and let J be a self-adjoint involution in H which commutes with A on Dom A , that is,
J 2 = Id, and J A − A J = 0. Then one can choose an orthogonal basis of eigenfunctions of
A in such a way that every eigenfunction u of A is either symmetric with respect to J , i.e.
Ju = u , or antisymmetric with respect to J , i.e. Ju = −u .
84 Chapter 3. Variational principles and applications

Figure 3.7: An example of a spectrum decomposition for a

symmetric domain. The solid lines denote the Dirichlet con-
ditions, and the dashed ones the Neumann conditions. The
union of spectra is understood in the sense of multisets, with
account of multiplicities.

Proof
Fix an eigenvalue λ of A , and denote by U the corresponding eigenspace. We start with the
case of a simple eigenvalue λ, so that dim U = 1. If u is a corresponding eigenfunction,
Au = λu , and since J commutes with A , we also have A Ju = J Au = λJu . Therefore, u
and Ju should be linearly dependent, Ju = cu , c = const. As J 2 u = u , we have c = ±1,
and either Ju−u or Ju+u vanishes identically. Therefore, an eigenfunction corresponding
to a simple eigenvalue is automatically either symmetric or antisymmetric.
If dim U > 1, we first remark that any u ∈ U can be decomposed into a sum of sym-
metric and antisymmetric elements with respect to J :

u + Ju u − Ju
u= + .
2 2
Let U ± := {v ∈ U : J v = ±v}, and note that the subspaces U ± are orthogonal: for any
u ± ∈ U ± we have

(u + , u − )H = (Ju + , u − )H = (u + , Ju − )H = −(u + , u − )H ,

which implies (u + , u − )H = 0. Since A commutes with J , the finite-dimensional operator

A|U decomposes into the direct sum

A|U = A|U + ⊕ A|U −

of two self-adjoint operators, and the result follows immediately.

§3.2. Consequences of variational principles 85

Exercise 3.2.19
Let Ω ⊂ Rd be an open set which is symmetric with respect to either a hyperplane or a
point in Rd . If τ : Ω → Ω is a corresponding symmetry reflection, prove that the operator
J : L 2 (Ω) → L 2 (Ω) defined by Ju = u◦τ is a self-adjoint involution which commutes with
the Laplacian on H 1 (Ω).

Let us now return to the example considered in the beginning of this subsection and illus-
trated in Figure 3.7, assuming for definiteness that the Dirichlet conditions are imposed on ∂Ω.
Let τS : Ω → Ω be the mirror symmetry with respect to S . We choose the involution J on H01 (Ω)
to be Ju = u ◦ τS . Applying now Theorem 3.2.18, we immediately obtain
³ ´
Spec −∆D
¡ ¢ ¡ D¢ Z
Ω ⊆ Spec −∆ Ω ′ ∪ Spec −∆ Ω ;∂1 Ω ,
′ ′ (3.2.5)

where we set ∂1 Ω′ = ∂Ω′ \S to be the part of the boundary of Ω′ excluding the extra “cut” along S .
We recall that −∆ZΩ′ ;∂1 Ω′ denotes the mixed, or Zaremba, Laplacian, with the Dirichlet condition
imposed on ∂1 Ω′ , and the Neumann one on the rest of the boundary, see §3.1.3.
To show the opposite inclusion, we need to demonstrate that every eigenfunction of the
Laplacian on Ω′ subject to the Dirichlet or Neumann condition on S ∩ Ω can be reflected anti-
symmetrically or symmetrically, respectively, across S to produce an eigenfunction on the whole
domain Ω.

Proposition 3.2.20: Reflection principle

Let Ω ⊂ Rd be a domain symmetric with respect to a hyperplane S which divides it into
two disjoint parts Ω′ and τS Ω′ . Decompose the boundary of Ω′ into ∂1 Ω′ = ∂Ω′ \ S and
∂2 Ω′ = ∂Ω′ ∩ S . Then

(i) If u ∈ H01 (Ω′ ) is an eigenfunction of the Dirichlet Laplacian −∆D

Ω′
corresponding to
an eigenvalue λ, then

u(x),

 if x ∈ Ω′ ,
v(x) = −u(τS x)), if x ∈ τS (Ω′ ),

if x ∈ ∂2 Ω′ ,

0,

is an eigenfunction of the Dirichlet Laplacian on Ω corresponding to the same eigen-

value.
1 ′ D
(ii) If u ∈ H0,∂ 1Ω
′ (Ω ) is an eigenfunction of the mixed Laplacian −∆Ω′ ,∂ Ω′ with the
1
Dirichlet condition imposed on ∂1 Ω′ and the Neumann condition on ∂2 Ω′ , then
(
u(x), if x ∈ Ω′ ,
v(x) =
u(τS x)), if x ∈ τS (Ω′ ),
86 Chapter 3. Variational principles and applications

extended by continuity to ∂2 Ω′ , is an eigenfunction of the Dirichlet Laplacian on Ω

corresponding to the same eigenvalue.

Exercise 3.2.21
Prove this proposition by showing first that in both cases v(x) is a weak eigenfunction of
the Dirichlet problem in Ω, and then apply elliptic regularity.

Remark 3.2.22
Note that the elliptic regularity of eigenfunctions is essential in the above argument, and
a reflection of an arbitrary smooth function does not necessarily yield a smooth function.
For example, consider in (0, +∞) the function u(x) = x 2 +x , which satisfies the Dirichlet
condition at the origin. Reflecting this function in an odd fashion with respect to the
origin yields
(
x 2 + x, if x ≥ 0,
f (x) = 2
−x + x, if x < 0,
which is a C 1 (R) function, but does not belong to C 2 near the origin.

Proposition 3.2.20 immediately implies

³ ´
Spec −∆D
¡ ¢ ¡ D¢ Z
Ω ⊇ Spec −∆Ω′ ∪ Spec −∆Ω′ ;∂1 Ω′ . (3.2.6)

Combining (3.2.5) and (3.2.6) gives the symmetry decomposition (or symmetry reduction) formula
for symmetric domains:
³ ´
Spec −∆D
¡ ¢ ¡ D¢ Z
Ω = Spec −∆ Ω ′ ∪ Spec −∆ Ω ,∂1 Ω .
′ ′ (3.2.7)

Remark 3.2.23
The same symmetry reduction method is applicable on a Riemannian manifold: for ex-
ample, the spectrum of the Laplace–Beltrami operator on the sphere Sd decomposes into
the union of the spectra of the Dirichlet and Neumann problems on the hemisphere. It
also works for other boundary conditions on ∂Ω (for example, in a Robin or in a Zaremba
problem) as long as they are imposed symmetrically with respect to S .

Remark 3.2.24
As an immediate application of the reflection principle, consider the Dirichlet problem for
the right isosceles triangle with legs of length π. By Proposition 3.2.20(i), any eigenfunc-
tion on the triangle, reflected antisymmetrically with respect to the hypothenuse, extends
§3.2. Consequences of variational principles 87

to an eigenfunction of the Dirichlet Laplacian on the square of side π. Therefore, the

Dirichlet eigenvalues of this triangle coincide with those of the square corresponding to
an eigenfunction antisymmetric with respect to the diagonal. It is easy to verify that these
eigenvalues are given by

λk,m = k 2 + m 2 , k, m ∈ N, k > m.

A similar approach works in the Neumann case, as well as for the equilateral triangles, see
[Lam33], [Mak70], [Pin80], and [Pin85]. We refer to [McC11] for a historical overview of
the reflection method in application to polygons.

There are two main applications of the symmetry decomposition. One is pretty straightfor-
ward and is often used in numerical analysis for reducing the underlying mesh sizes (since one can
consider a smaller domain).
Gabriel Lamé
(1795–1870)
Numerical Exercise 3.2.25
Compute the eigenvalues of the Dirichlet Laplacian on an ellipse by two methods: first,
directly, and second, by decomposing the problem into four problems on a quarter-ellipse,
with Dirichlet and Neumann conditions imposed on the semi-axes.

The second application of the symmetry decomposition is often used in conjunction with
the Dirichlet–Neumann bracketing.

Proposition 3.2.26
Let Ω ⊂ Rd be a domain symmetric with respect to a hyperplane S , and consider a Lapla-
cian in Ω with some boundary conditions imposed symmetrically with respect to S . Then
its first eigenfunction is symmetric with respect to S .

Proof
By (3.2.7) and Remark 3.2.23, the first eigenfunction will satisfy either the Dirichlet or the
Neumann condition on S . However, imposing the Dirichlet condition on S increases the
eigenvalues compared to imposing the Neumann condition, therefore the eigenfunction
corresponding to the minimal eigenvalue is symmetric. We note that in the case of the Neu-
mann problem in Ω, the result is trivially true since the first eigenfunction is a constant.
88 Chapter 3. Variational principles and applications

Exercise 3.2.27
Let Ω be a planar domain symmetric with respect to a line S passing through the origin O
and such that the set Ω∩S is centrally symmetric with respect to O . Impose some bound-
ary conditions on ∂Ω symmetrically with respect to S , and denote the first eigenvalue of
the corresponding problem by λ1 (Ω). Now take a half Ω′ of Ω lying to one side of S , and
let Ω
e be the union of Ω′ and its centrally symmetric reflection τ(Ω′ ) around O ; reflect
the boundary conditions in the same way, see Figure 3.8. Show that λ1 (Ω)e ≥ λ1 (Ω). A
solution can be found in [JakLNP06].

Figure 3.8: An example of a symmetric domain Ω and a cen-

trally symmetric domain Ωe , obtained by adding to the right half
of Ω its copy reflected with respect to the point O . The solid
lines denote the Dirichlet conditions, and the dashed ones the
Neumann conditions.

Exercise 3.2.28
Modify the argument in Example 3.2.27 to show that λ1 (Ω)
e ≥ λ1 (Ω), where λ1 (Ω) and
λ1 (Ω) refer to two boundary value problems on the quarter-sphere shown in Figure 3.9.
e
§3.2. Consequences of variational principles 89

Figure 3.9: Two boundary value problems on a quarter-

sphere, Ω on the left and Ω
e on the right. The solid lines denote
the Dirichlet conditions, and the dashed ones the Neumann
conditions.

§3.2.3. Counting functions

We have already encountered the counting function of eigenvalues of a flat torus in §1.2.2. Study-
ing counting functions as opposed to individual eigenvalues provides an alternative, and often
more convenient, approach to certain problems in spectral geometry.

Definition 3.2.29: Eigenvalue counting function

Let A be a self-adjoint semi-bounded from below operator with a discrete spectrum con-
sisting of eigenvalues λ1 ≤ λ2 ≤ . . . . The eigenvalue counting function of A is the function
N : R → N0 defined as

N (λ) = N A (λ) := #{ j : λ j (A) ≤ λ}.

It is clear that N (λ) is right-continuous and monotone non-decreasing. Importantly, know-

ing N A (λ) for all λ ∈ R we can recover the eigenvalues of A : if N A (λ+0)−N A (λ−0) = 0, then
λ ̸∈ Spec(A), and if N A (λ+0)−N A (λ−0) = m > 0, then λ is an eigenvalue of A of multiplicity
m.
Sometimes, we will deal instead with the left-continuous eigenvalue counting function

Nf(λ) = NfA (λ) := #{ j : λ j (A) < λ}, (3.2.8)

whose values differ from those of N (λ) only at eigenvalues of A : if λ is an eigenvalue of A of

multiplicity mλ then N A (λ) = NfA (λ) + mλ .
 90 Chapter 3. Variational principles and applications

Remark 3.2.30
Given any two self-adjoint semi-bounded from below operators A and B with discrete
spectra, the inequalities λk (A) ≤ λk (B ), k ∈ N, could be equivalently rewritten as
N A (λ) ≥ N B (λ) for all λ ∈ R: indeed, the smaller are the eigenvalues, the larger is the
counting function. This simple observation will be very useful in the sequel.

Similarly, one can define an eigenvalue counting function N Q (λ) of the weak spectral prob-
lem (3.1.2) associated with a bilinear form Q . The following important result shows that the vari-
ational principle from Proposition 3.1.3 can be reformulated in terms of the eigenvalue counting
function.

Lemma 3.2.31: Glazman’s Lemma

Consider the weak spectral problem (3.1.2) associated with a symmetric bilinear semi-
bounded from below form as defined in §3.1.1. Then the counting function of the cor-
responding weak eigenvalues satisfies
Q
N (λ) = max dim L ,
L ⊂Dom(Q)
R[u]≤λ for all u∈L \{0}

where L is a finite-dimensional linear subspace of Dom(Q), and R[u] is the Rayleigh

quotient (3.1.4).

Exercise 3.2.32
Prove Glazman’s Lemma, see [Shu20, Proposition 9.5].
Israel Markovich
Glazman
Since we will be mostly dealing with the counting functions of Dirichlet and Neumann Lapla-
(1916–1968)
cians, we introduce a shorthand notation for them.

Notation 3.2.33: Eigenvalue counting functions for the Laplacians

If Ω is a bounded domain (with sufficiently regular boundary in the Neumann case) in a
Euclidean space or in a Riemannian manifold, we will write for brevity
−∆D −∆N
NΩD (λ) := N Ω (λ), NΩN (λ) := N Ω (λ),

and so on. Similarly, for a closed Riemannian manifold (M , g ) we will write

−∆(M ,g )
N(M ,g ) (λ) = N M (λ) = N g (λ) := N (λ),

depending on the context.

§3.2. Consequences of variational principles 91

Numerical Exercise 3.2.34

Plot N D (λ) and N N (λ) for the planar unit disk, unit square, or any other domain of
your choice, with eigenvalues computed either analytically or numerically.

§3.2.4. Inequalities between the Dirichlet and Neumann eigenvalues for Euclidean do-
mains

The goal of this subsection is to prove

Theorem 3.2.35: The Friedlander–Filonov inequality

Let Ω ⊂ Rd , d ≥ 2, be a bounded open set with Lipschitz boundary, and let λk := λDk
(Ω),
µk := λN
k
(Ω) . Then
µk+1 (Ω) < λk (Ω), k ∈ N. (3.2.9)

This inequality was first proposed by L. Payne in 1955 [Pay55]. Its non-strict version was
proved by L. Friedlander in 1991 [Fri91] for C 1 domains. Friedlander’s original proof is very in-
structive, and we will re-visit it in §7.4.3. In 2004, N. Filonov [Fil04] found a strikingly simple
and elegant argument that proved Theorem 3.2.35 as stated above.
Before proceeding to Filonov’s proof, we start with the following simple lemma.

Lemma 3.2.36
Let u be an eigenfunction of the Neumann Laplacian on Ω ⊂ Rd . Then u ∉ H01 (Ω).

Proof
Suppose, for contradiction, that u is an eigenfunction of the Neumann Laplacian in Ω
corresponding to an eigenvalue µ and u ∈ H01 (Ω). Let w be an extension of u by zero to
the whole Rd . Then w ∈ H 1 (Rd ), and, given v ∈ C 0∞ (Rd ), we have

(∇w, ∇v)L 2 (Rd ) = (∇u, ∇v)L 2 (Ω) = (−∆u, v)L 2 (Ω)

(3.2.10)
= µ(u, v)L 2 (Ω) = µ(w, v)L 2 (Rd ) .

Note that the boundary term vanishes because u is a Neumann eigenfunction. Compar-
ing the left- and the right-hand sides of (3.2.10) we deduce that w is a weak solution of the
equation −∆w = µw in Rd . By elliptic regularity it is therefore real analytic, and since
w|Rd \Ω = 0, w is identically zero. Hence u is identically zero, and therefore not an eigen-
function.
92 Chapter 3. Variational principles and applications

Exercise 3.2.37
Modify the proof of Lemma 3.2.36 to show that a Neumann eigenfunction on Ω cannot
1
belong to the space H0,Γ (Ω), where Γ is an open subset of ∂Ω.

Let us also state the following exercise which we will use later.

Exercise 3.2.38
Let Ξ be any finite non-empty subset of Rd . Prove that the set of exponential functions
© i〈ω,x〉
: ω ∈ Ξ is linearly independent over C.
ª
e

Proof of Theorem 3.2.35

In this proof we, exceptionally, work with complex-valued functions, and therefore all
scalar products are understood over C.
By Glazman’s Lemma 3.2.31 applied to the Dirichlet Laplacian on Ω, we have
N D (λ) = max dim L . (3.2.11)
L ⊂H01 (Ω)
R[u]≤λ for all u∈L \{0}

Fix λ ≥ λ1 , and let Vλ ⊂ H01 (Ω) be a maximising L in (3.2.11), that is a linear subspace
of H01 (Ω) such that dimVλ = N D (λ), and R[u] ≤ λ for all u ∈ Vλ \ {0}. Let also Fλ =
Ker(−∆N −λ) ⊂ H 1 (Ω): that is, F λ = {0} if λ ̸∈ Spec(−∆N ), otherwise F λ is the eigenspace
of dimension mλ corresponding to the Neumann eigenvalue λ of multiplicity mλ ≥ 1.
According to Lemma 3.2.36, Fλ ∩ Vλ = {0}; also Vλ + Fλ = Vλ ⊕ Fλ is finite-dimensional:
dim(Vλ + F λ ) = N D (λ) + m λ .
Consider now the set of functions ei〈ω,x〉 : ω ∈ Rd , |ω|2 = λ . By the result of Exercise
© ª

3.2.38, this set is infinite-dimensional if d ≥ 2, and we therefore can choose a particular

vector ω with |ω|2 = λ in such a way that g := ei〈ω,x〉 does not belong to Vλ ⊕ Fλ . Set
Wλ := Vλ + F λ + c g : c ∈ C ,
© ª

and consider an arbitrary w ∈ Wλ \ {0}, w = v + f + c g , where v ∈ Vλ and f ∈ Fλ .

Let us estimate the Rayleigh quotient R[w], taking into account, firstly, that by the
definition of Vλ we have ∥∇v∥2 ≤ λ∥v∥2 for any v ∈ Vλ , secondly that ∥∇ f ∥2 = λ∥ f ∥2
for any f ∈ Fλ , and lastly that ∇g = ig ω and −∆g = |ω|2 g = λg (all norms and inner
products here and for the rest of the proof are in L 2 (Ω)).
In the numerator of R[w] we have
∥∇(v + f + c g )∥2 = ∥∇v∥2 + ∥∇ f ∥2 + ∥c∇g ∥2
| {z }
=:I1
¡¡ ¢ ¡ ¢¢
+ 2 Re ∇ f , ∇(v + c g ) + ∇(c g ), ∇v .
| {z }
=:I2
§3.2. Consequences of variational principles 93

We further simplify
I1 = ∥∇v∥2 + λ∥ f ∥2 + |c|2 |ω|2 Vold (Ω)
(3.2.12)
= ∥∇v∥2 + λ∥ f ∥2 + |c|2 λ Vold (Ω),

and, using Green’s formula,

I2 = 2 Re −∆ f , v + c g + c −∆g , v
¡¡ ¢ ¡ ¢¢
¡¡ ¢ ¡ ¢¢ (3.2.13)
= 2λ Re f , v + c g + c g , v

(the boundary terms vanish since f is a Neumann eigenfunction or zero, and v ∈ H01 (Ω)).
In the denominator of R[w] we have
∥v + f + c g ∥2 = ∥v∥2 + ∥ f ∥2 + ∥c g ∥2 + 2 Re f , v + c g + c g , v ,
¡¡ ¢ ¡ ¢¢
(3.2.14)
| {z } | {z }
=:I1′ =:I2′

where after a simplification

I1′ = ∥v∥2 + ∥ f ∥2 + |c|2 Vold (Ω). (3.2.15)
Note that with account of ∥∇v∥2 ≤ λ∥v∥2 , the comparison of (3.2.12) and (3.2.15) yields
I1 ≤ λI1′ ,

and the comparison of (3.2.13) and (3.2.14) yields

I2 = λI2′ .

Thus, we deduce the bound on the Rayleigh quotient,

I1 + I2
R[w] = ≤ λ, (3.2.16)
I1′ + I2′

valid for all w ∈ Wλ \ {0}.

We now re-state Glazman’s Lemma for the Neumann Laplacian in Ω:
N N (λ) = max dim L . (3.2.17)
L ⊂H 1 (Ω)
R[u]≤λ for all u∈L \{0}

By (3.2.16), we can take L = Wλ in (3.2.17), giving

N N (λ) ≥ dimW = N D (λ) + m λ + 1.

Substituting into this inequality λ = λk , for which we have N D (λk ) ≥ k , we obtain

N N (λk ) = NfN (λk ) + m λk ≥ k + 1 + m λk ,

or NfN (λk ) ≥ k + 1. In other words, on the semi-open interval [0, λk ) there are at least
k + 1 Neumann eigenvalues, which means that µk+1 < λk .
 94 Chapter 3. Variational principles and applications

Remark 3.2.39

pthe proof hinges upon the existence of a function g such that −∆g = λg and
Note that
∥∇g ∥ ≤ λ∥g ∥. For Euclidean domains, one can take an exponential function as we do.
As was shown in [Maz91], such a function does not always exist on Riemannian manifolds,
and the Friedlander–Filonov inequality may fail there. For instance, it fails for spherical
caps that are larger than a hemisphere.

Remark 3.2.40
In dimension d = 1 the inequality (3.2.9) turns into an equality for each k ≥ 1.

Exercise 3.2.41
Inspect the proof of Theorem 3.2.35 and explain why the strict inequality (3.2.9) fails in
dimension one.

Let us conclude this section with the following open problem, which gives a stronger version
of (3.2.9).

Conjecture 3.2.42
For any bounded domain Ω ⊂ Rd , we have µk+d ≤ λk , k ≥ 1.

This result was proved by H. A. Levine and H. F. Weinberger [LevWei86] for convex domains,
but for arbitrary domains it remains a challenging open question.

§3.3. Weyl’s law and Pólya’s conjecture

§3.3.1. Weyl’s law

Weyl’s law for the asymptotic distribution of eigenvalues is one of the most important results
Hermann Klaus Hugo in spectral geometry. In its original form it was proved by Hermann Weyl in 1911, confirming
Weyl a conjecture proposed in 1905 by Lord Rayleigh (with a constant corrected by J. H. Jeans, see
(1885–1955) [SafVas97] for a discussion).
Weyl’s law is quite universal, in a sense that its versions apply to a wide variety of situations:
Riemannian manifolds, Euclidean domains, various self-adjoint boundary conditions, and dif-
ferent elliptic operators. Below we prove Weyl’s law for the Dirichlet Laplacian on Euclidean
domains and leave its generalisations for later.
§3.3. Weyl’s law and Pólya’s conjecture 95

Theorem 3.3.1
d
Let −∆DΩ be the Dirichlet Laplacian on a bounded domain Ω ⊂ R . Then its eigenvalue
D
counting function NΩ (λ) satisfies the asymptotic formula
d
NΩD (λ) = C d Vold (Ω)λ 2 + R(λ), (3.3.1)
³ d
´
where R(λ) = o λ 2 as λ → +∞. Here

ωd 1
C d := = (3.3.2)
(2π)d d
³ ´
(4π) Γ d2 + 1
2

is the Weyl constant, and ωd denotes the volume of the unit ball in Rd , see (B.1.1).

Proof
Let us split the proof into three steps. First, arguing in a similar way as in the proof of
the asymptotic formula (1.2.14) for the flat torus, we prove (3.3.1) for cubes with either the
Dirichlet or the Neumann boundary condition. The only difference compared to the torus
case is that now one needs to take into account points with positive integer coordinates in
the Dirichlet case, and non-negative ones in the Neumann case. We leave the details as an
exercise.
The next step is to consider domains that could be represented as an almost disjoint
union of cubes (this means that if K is a finite collection of disjoint open cubes, then Ω
is the interior of the closure of K , and therefore ∂Ω ⊂ ∂K ). Let Ω be such a domain, see
Figure 3.10. Consider its partition into cubes (that is, the region Ω
e := Ω\∂K ) and impose
the Dirichlet (respectively, the Neumann) boundary conditions on ∂Ω e.
By Dirichlet–Neumann bracketing and Remark 3.2.30 we then have, for all λ,
D D N
NΩ
e (λ) ≤ NΩ (λ) ≤ NΩ
e (λ).

The result then follows by noticing that the counting functions N eD (λ) and N eN (λ) are
Ω Ω
sums of the corresponding counting functions for the cubes and applying the first step of
the argument.
Finally, let Ω be an arbitrary bounded domain. Let ΩE ,a and ΩI ,a be two domains that
can be represented as almost disjoint unions of cubes of side a > 0, such that ΩI ,a ⊂ Ω ⊂
ΩE ,a , see Figure 3.11.
By the domain monotonicity for the Dirichlet eigenvalues,

NΩDI ,a (λ) ≤ NΩD (λ) ≤ NΩDE ,a (λ).

96 Chapter 3. Variational principles and applications

Therefore, applying the result obtained on step two, we get

NΩD (λ)
≤ C d Vold ΩE ,a ,
¡ ¢
lim sup d
λ→∞ λ 2

and
NΩD (λ)
≥ C d Vold ΩI ,a .
¡ ¢
lim inf d
λ→∞ λ 2

The result then follows by taking the limit a → 0 and observing that one can choose ΩE ,a
and ΩI ,a in such a way that

lim Vold ΩE ,a = lim Vold ΩI ,a = Vold (Ω).

¡ ¢ ¡ ¢
a→0 a→0

This completes the proof of the theorem in the Dirichlet case.

Figure 3.10: An almost disjoint union of open squares.

Remark 3.3.2
This proof could be found, for instance, in [ReeSim75, Chapter XIII], [CouHil89, Chap-
ter VI.4], [Bér86, Chapter 3]. As shown in [Roz72] (see also [Fri21]), Theorem 3.3.1 holds
in fact for arbitrary Euclidean domains of finite volume.

Remark 3.3.3: Weyl’s law for the Neumann Laplacian

An analogue of Theorem 3.3.1 holds for the Neumann eigenvalue problem in bounded Eu-
clidean domains with Lipschitz boundary, see [NetSaf05] for a detailed discussion. In fact,
§3.3. Weyl’s law and Pólya’s conjecture 97

Figure 3.11: An example of a domain Ω, with its bound-

ary shown as a solid curve, and corresponding domains ΩI ,a
(darker shading, boundary is shown as a dotted line) and ΩE ,a
(lighter shading, boundary is shown as a dashed line.)

for piecewise C 2 planar domains one can prove Weyl’s law for the Neumann Laplacian us-
ing a modification of the argument presented above, see [CouHil89, §VI.4.4]. Note that
a direct generalisation of the proof to the Neumann case does not work, as the last step
involves domain monotonicity for the Dirichlet eigenvalues. Instead, one can approxi-
mate Ω by a union of cubes (in the interior) and right triangles (near the boundary), and
show that small perturbations of triangles do not change the asymptotics of the eigenvalue
counting function assuming that the boundary is sufficiently regular.

Theorem 3.3.1 admits various extensions and improvements. In particular, for Euclidean do-
mains with piecewise smooth boundaries, the remainder estimate can be improved to
³ d −1 ´
R(λ) = O λ 2 ,

for both Dirichlet and Neumann boundary conditions, see [Vas86]. Further improvements of the
remainder estimates will be discussed in the next subsection. There exist also remainder estimates
98 Chapter 3. Variational principles and applications

for domains with very rough boundaries, including some fractal ones, see e.g. [Mét77], [Lap91],
[LevVas96].
Weyl’s law holds also in the Riemannian setting.

Theorem 3.3.4
Let M be a d -dimensional smooth compact Riemannian manifold. If ∂M ̸= ;, assume
that either the Dirichlet or the Neumann boundary conditions are imposed on the bound-
ary. Then the eigenvalue counting function for M has the asymptotics
d
³ d −1 ´
N M (λ) = C d Vol(M )λ 2 + O λ 2 , (3.3.3)

where C d is again defined by (3.3.2).

Remark 3.3.5
The error estimate in (3.3.3) is sharp, as follows from the eigenvalue asymptotics on the
round sphere, see (1.2.26). The proof of the sharp Weyl’s law uses the theory of pseudod-
ifferential operators and is beyond the scope of this book. We refer to [Shu01], [Tré82],
[SafVas97] for further details. We will revisit Weyl’s law on manifolds in Chapter 6, and
will explain how to deduce (3.3.3) with a weaker remainder estimate from the heat trace
asymptotics.

Exercise 3.3.6
Prove that Theorem 3.3.1 is equivalent to the asymptotic law
³ 2´
− d2 d2
λD
k (Ω) = (C d Vol d (Ω)) k + o kd as k → ∞. (3.3.4)

The same asymptotics also holds for the Neumann eigenvalue λN

k
(Ω), and the remainder
estimates can be improved.

§3.3.2. The two-term asymptotic formula and Weyl’s conjecture

Let us recall Weyl’s law on a square: can one get a better remainder estimate in this case? Now we
will be more careful and take boundary conditions into account.
Consider a square K π of side p π. In the Dirichlet case, the eigenvalues correspond to integer
points inside the cirle of radius λ lying in the first quadrant excluding the coordinate axes; in the
Neumann case, the points on the axes (i.e. points having zero as one of the coordinates) should
be counted as well. p
Howpmany integer points lie on the coordinate axes inside the circle of radius λ? Approx-
imately, λ on each semi-axis. There are four semi-axes, and therefore the quarter of integer
points inside a circle is equal to the number of integer points in the interior of a quadrant plus the
§3.3. Weyl’s law and Pólya’s conjecture 99

number of integer points on a single semi-axis. Therefore, in the Dirichlet case we need to take a
quarter of integer points inside a circle and subtract the contribution of one semi-axis, while in
the Neumann case we need to add the contribution of one semi-axis. Therefore, for a square K π
we get
π p π p
N D (λ) = λ − λ + R D (λ), N N (λ) = λ + λ + R N (λ).
4 4
Note that these
³ptwo-term
´ asymptotic formulas would be meaningful only if the remainders R D,N (λ)
are of order o λ . This is indeed true and could be deduced from the number-theoretic results
on Gauss’s circle problem, see discussion after Conjecture 1.2.13.
In 1911, H. Weyl conjectured that a similar two-term asymptotic formula holds for an arbitrary
Euclidean domain, and that the second term arises from the boundary.

Conjecture 3.3.7: Weyl’s conjecture

Let Ω ⊂ Rd be a piecewise smooth Euclidean domain. Then
d d −1
³ d −1 ´
N (λ) = C d Vold (Ω)λ 2 ±C b,d Vold −1 (∂Ω)λ 2 +o λ 2 , (3.3.5)

where the minus sign corresponds to the Dirichlet boundary conditions, and the plus sign
to the Neumann boundary conditions. Here
1
C b,d = d −1
³ ´. (3.3.6)
2d +1 π 2 Γ d +1
2

The expression (3.3.6) can be deduced from the heat trace asymptotics, see Remark 6.1.11 and
Exercise 6.1.12.
In dimension two, (3.3.5) takes a particularly simple form,

Area(Ω) Length(∂Ω) p ³p ´
N (λ) = λ± λ+o λ . (3.3.7)
4π 4π

Example 3.3.8
In practice, at least for relatively simple planar domains, both the Dirichlet and Neumann
asymptotic formulae (3.3.7) work remarkably well even for low values of λ. To illustrate
this, we plot in Figure 3.12 the actual Dirichlet and Neumann counting functions to-
gether with one-term Weyl asymptotics (3.3.1) and the corresponding two-term asymp-
totics (3.3.7) for the rectangle R π,2π and for the unit disk D.

In full generality Weyl’s conjecture remains open. There has been a significant progress on it
in the past decades, in particular, due to V. Ivrii [Ivr80], R. Melrose [Mel84], Yu. Safarov and D.
Vassiliev [SafVas97]. The key observation here is that the growth of the error term is closely linked
to the dynamical properties of the billiard flow (or, in a more general setting of a Riemannian
100 Chapter 3. Variational principles and applications

Figure 3.12: The actual counting functions and the one- and two-term Weyl’s asymp-
totics for the rectangle R π,2π (left) and for the unit disk D (right). In both figures, blue
curves correspond to the Dirichlet Laplacian and the magenta curves to the Neumann
one. The graphs of the actual N (λ) are shown as solid, and the graphs of the two-term
asymptotics as dotted lines. The dashed black line corresponds to the one-term Weyl’s
asymptotics.

manifold, of the geodesic flow). From the physical standpoint, this can be explained via Bohr’s
correspondence principle in quantum mechanics. Mathematically, the connection could be made
via the wave trace. A rigorous treatment of this subject is way beyond the scope of this book, and
we refer the reader to [SafVas97] for details. We shall simply state the main result of this theory,
which is essentially due to V. Ivrii [Ivr80] with some improvements and generalisations due to D.
Vassiliev [Vas86].
A billiard trajectory satisfying the usual law of reflection in a bounded Euclidean domain Ω ⊂
Rd is uniquely determined by the initial point x ∈ Ω and the initial direction ξ ∈ Sd −1 . Consider
the Liouville measure on the unit (co)tangent bundle of Ω, which in this case can be simply viewed
as the measure dxdξ on Ω × Sd −1 . We say that Ω satisfies the non-periodicity condition if the set
of pairs (x, ξ) corresponding to periodic billiard trajectories has Liouville measure zero.

Theorem 3.3.9
Let Ω ⊂ Rd be a bounded domain with piecewise smooth boundary satisfying the non-
periodicity condition. Then the two-term asymptotics (3.3.5) holds.
§3.3. Weyl’s law and Pólya’s conjecture 101

Remark 3.3.10
It was conjectured by V. Ivrii (see also [SafVas97, Conjecture 1.3.35]) that the non-
periodicity condition holds for any Euclidean domain. This is an outstanding open prob-
lem in billiard dynamics. The affirmative answer is known just for a few specific classes of
domains, such as convex analytic domains, piecewise-concave domains and polygons.

Exercise 3.3.11
Show that a rectangle satisfies the non-periodicity condition.

Remark 3.3.12
Under conditions of Theorem 3.3.9, the Neumann two-term asymptotic formula (3.3.5)
remains valid for the eigenvalue counting function N R,γ (λ) of the Robin Laplacian for
any fixed γ. This is due to the fact that the second Weyl asymptotic term (for an elliptic
boundary value problem in general) depends only upon the leading order differentiations
in the boundary conditions and ignores the lower order differentiations, see [SafVas97].

Theorem 3.3.9 admits a generalisation to Riemannian manifolds with boundary. However,

in this case the non-periodicity condition is not always satisfied, and is essential for the two-term
asymptotics (3.3.5) to hold.

Exercise 3.3.13

(i) Show that all the trajectories of the geodesic flow on a hemisphere are periodic.

(ii) Using Theorem 1.2.16 and formula (1.2.26) show that the two-term asymptotics does
not hold for a hemisphere with either the Dirichlet or the Neumann boundary con-
ditions. Hint: Show that the eigenfunctions on a hemisphere with the Dirichlet
(respectively, the Neumann) conditions are precisely the eigenfunctions of the full
sphere which are antisymmetric (respectively, symmetric) with respect to the equa-
torial plane bounding the hemisphere. Full details can be found in [BérBes80].

Finally, there is a version of Theorem 3.3.9 for closed Riemannian manifolds, see [DuiGui75].
In this case the second term is equal to zero, and we simply obtain a refinement of the error term
(big O is replaced by little o ) in (3.3.3) under the non-periodicity assumption. Once again, this
assumption is essential, as we have already seen in the example of the round sphere.
102 Chapter 3. Variational principles and applications

§3.3.3. Pólya’s conjecture

Assuming that (3.3.5) holds (say, under the conditions of Theorem 3.3.9), it follows immediately
that for Ω ⊂ Rd with a sufficiently regular boundary and for a sufficiently large λ, we have
d
N D (λ) ≤ C d Vold (Ω)λ 2 ≤ N N (λ). (3.3.8)

In 1954, George Pólya [Pól54] conjectured that the inequalities (3.3.8) hold for all λ ≥ 0. In fact
Pólya’s original conjecture was stated only for planar domains, and in a slightly different form.
There exist other versions of these inequalities, re-written, for example, using strict inequal-
ities in (3.3.8). We will state Pólya’s conjecture as the inequalities for the k th Dirichlet eigenvalue
λk = λD k
(Ω) and the k th nonzero Neumann eigenvalue µk+1 = λN k+1
(Ω):

µ ¶2
1 d 2
µk+1 ≤ k d ≤ λk , (3.3.9)
George Pólya C d Vold (Ω)
(1887–1985)
cf. (3.3.4).

Conjecture 3.3.14: Pólya’s Conjecture

The inequalities (3.3.9) hold for any k ≥ 1.

In fact, it is expected that (3.3.9) hold with strict inequalities, see [FreLagPay21].
We will start by showing that the two forms of Pólya’s Conjecture, the bounds on the eigen-
value counting functions, and the bounds on eigenvalues, are in fact equivalent.

Lemma 3.3.15
The inequalities (3.3.8) hold for all λ ≥ 0 if and only if the inequalities (3.3.9) hold for all
k ≥ 1.

Proof
Obviously, the Dirichlet and Neumann cases can be treated independently. We will give
the proof in the Dirichlet case only, and will leave the Neumann one as an exercise. First,
assume that (3.3.8) holds. Substitute, for any k ≥ 1, λ = λk into (3.3.8), and note that
N D (λk ) ≥ k . Then we have
d
k ≤ N D (λk ) ≤ C d Vold (Ω)λk2 ,

and the second inequality (3.3.9) follows by raising both sides to the power d2 . Thus, (3.3.8)
implies (3.3.9).
Assume now that (3.3.9) holds for all k ≥ 1. We will prove (3.3.8) by induction in the
intervals of the non-negative λ-axis between consecutive distinct Dirichlet eigenvalues. To
§3.3. Weyl’s law and Pólya’s conjecture 103

start with, as λ1 > 0, we automatically get

d
0 = N D (λ) ≤ C d Vold (Ω)λ 2 for λ ∈ [0, λ1 ).

Assume now additionally that (3.3.8) holds for λ ∈ [0, λk ) with some k ≥ 1. Let λk = · · · =
λk+m < λk+m+1 be a Dirichlet eigenvalue of multiplicity m + 1, where m ≥ 0. Then by
(3.3.9),
µ ¶2
1 d 2
λk = · · · = λk+m ≥ (k + m) d . (3.3.10)
C d Vold (Ω)
Moreover,
N D (λ) = k + m for λ ∈ [λk , λk+m+1 ),
giving, with account of (3.3.10),
d d
N D (λ) ≤ C d Vold (Ω)λk2 ≤ C d Vold (Ω)λ 2 for λ ∈ [λk , λk+m+1 ).

This completes the induction step, therefore (3.3.9) implies (3.3.8).

Exercise 3.3.16

(i) Prove that the original inequalities (3.3.8) are equivalent to their analogues for the
left-continuous counting functions NfD (λ) and NfN (λ), see (3.2.8).

(ii) Prove Lemma 3.3.15 in the Neumann case. You may find it easier to work with
NfN (λ) instead of N N (λ) and use the result of part (i) at the end.

In a paper [Pól61] written a few years after stating his conjecture, G. Pólya proved Conjecture
3.3.14 for any tiling domain Ω ⊂ Rd — that is, a domain such that the whole space Rd is an almost
disjoint union of an infinite number of non-intersecting copies (shifted and possibly rotated) of
Ω, with some additional restrictions in the Neumann case (these restrictions were later removed
in [Kel66]). We emphasise that Pólya’s Conjecture 3.3.14 still remains open in full generality.

Theorem 3.3.17: Pólya conjecture holds for tiling domains

Let Ω ⊂ Rd be a tiling domain. Then the inequalities (3.3.9) hold for any k ≥ 1.

Proof
We present Pólya’s proof in the Dirichlet case, and refer to [Kel66] for the Neumann one.
Suppose that Ω ⊂ Rd is a tiling domain; somewhat abusing notation, we will denote
its shifted (and possibly rotated) non-intersecting copies by the same symbol. Let also Ωh
104 Chapter 3. Variational principles and applications

denote a copy of Ω scaled with a factor h > 0. Obviously, if Ω tiles the space, so does Ωh
(for a fixed h ); also,
Vold (Ωh ) = h d Vold (Ω).
Fix for the moment h > 0 and some tiling of Rd by Ωh . Let K be a unit cube, let

Ωh := Ωh
G
Ωh ⊂K

be a disjoint union of copies of Ωh fully inside K , and let

M h := #{Ωh ⊂ K }

be the number of such copies.

By the Dirichlet domain monotonicity and Dirichlet–Neumann bracketing, we have

λℓ (K ) ≤ λℓ (Ωh )

for any ℓ ∈ N. Fix now k ∈ N, and choose ℓ = kMh . As Ωh is a disjoint union of Mh copies
of Ωh , we have
1
λℓ (Ωh ) = λkMh (Ωh ) = λk (Ωh ) = λk (Ω),
h2
and so
h 2 λkMh (K ) ≤ λk (Ω). (3.3.11)
We now take the limit as h → 0+ , noting two limiting identities. Firstly, we have

lim M h h d Vold (Ω) = lim+ M h Vold (Ωh ) = Vold (K ) = 1.

h→0+ h→0

Secondly, by one-term Weyl’s Law for the eigenvalues of the cube,

¶2
λkMh (K )
µ
1 d 1
lim+ 2
= = 2
.
h→0 (kM h ) d C d Vold (K ) d
Cd

Passing now to the limit h → 0+ in the left-hand side of (3.3.11) and using the two
limiting identities above, we obtain
1 2
λk (Ω) ≥ lim+ h 2 λkMh (K ) = 2
lim+ h 2 (kM h ) d
h→0 d h→0
Cd
2 2
kd 2 kd
= 2
lim+ h 2 (h −d ) d = 2
,
(C d Vold (Ω)) d h→0 (C d Vold (Ω)) d

proving the second inequality in (3.3.9).

§3.3. Weyl’s law and Pólya’s conjecture 105

Numerical Exercise 3.3.18

Use any software capable of finding zeros of Bessel functions and their derivatives to verify
that Pólya conjecture holds for the first thousand eigenvalues of the unit disk.

Remark 3.3.19: Pólya’s conjecture for disks and balls

We note that Pólya’s conjecture for the planar disk and, in the Dirichlet case, for balls
in Rd , d ≥ 3, has been recently proved in [FilLPS23], thus making the disk the first non-
tiling planar domain for which it is known. The proofs in [FilLPS23] are based on relations
between the Dirichlet and Neumann eigenvalue counting functions for the balls and some
lattice counting problems, and, in the Neumann case for the disk, is partially computer-
assisted.

We cite the following result which in a sense complements Theorem 3.3.17.

Theorem 3.3.20: [FilLPS23, Theorem 1.8]

Let Ω ⊂ Rd be a domain for which either the Dirichlet or the Neumann Pólya’s conjecture
holds, and let Ω′ be a domain which tiles Ω. Then the same Pólya’s conjecture also holds
for Ω′ .

Proof
Assume that Ω can be tiled by M ≥ 2 congruent copies of Ω′ , so that Vold (Ω) =
M Vold (Ω′ ). We have, by Dirichlet–Neumann bracketing and since the eigenvalues of all
the congruent copies coincide with those of Ω′ ,

M NΩD′ (λ) ≤ NΩD (λ) ≤ NΩN (λ) ≤ M NΩN′ (λ).

Assuming now (3.3.8) for all λ ≥ 0, we get

M NΩD′ (λ) ≤ C d Vold (Ω)λd = C d M Vold (Ω′ )λd ≤ M NΩN′ (λ),

and the result follows by cancelling M .

Theorem 3.3.20 and the validity of Pólya’s conjecture for the disk imply that Pólya’s conjec-
ture is also valid for planar sectors of an aperture 2π/n , n ∈ N [FilLPS23]. A more complicated
argument also shows it to be true for sectors of any aperture.
 106 Chapter 3. Variational principles and applications

§3.3.4. The Berezin–Li–Yau inequalities

Using the inequality for the sum of the first k Dirichlet eigenvalues, which originated in the stud-
ies of the Schrödinger operator, one can deduce slightly weakened (in comparison to Pólya’s con-
jecture) bounds for the Dirichlet eigenvalues which are always true. Our exposition here follows
[LieLos97, §12.11], see also [Nam21, §5.1].

Theorem 3.3.21: The Berezin–Li–Yau inequality [Ber72], [LiYau83]

Let Ω ⊂ Rd be a bounded domain. Then its Dirichlet eigenvalues λm = λD
m (Ω) satisfy

2
Felix Alexandrovich k d k 1+ d
λm ≥
X
(3.3.12)
Berezin
m=1 d + 2 (C d Vold (Ω)) d2
(1931–1980)
for all k ∈ N.

k
An immediate consequence of Theorem 3.3.21, obtained from (3.3.12) by using λk ≥ k1 λm ,
P
m=1
is

Corollary 3.3.22
For any k ∈ N,
¶2
d k
µ
d
λk ≥ .
d + 2 C d Vold (Ω)
In other words, Polya’s conjecture for Dirichlet eigenvalues, that is, the right inequality in
(3.3.9), holds in a weakened form with an additional factor dd+2 .

Before proceeding to the proof of Theorem 3.3.21, we introduce the following notation, which
we will also need further on.

Notation 3.3.23
Let F : O → R by a real valued function defined on an open set O ⊂ Rd . We set, for t ∈ R,
L F (t ) := {y ∈ O : F (y) = t },
to denote its level sets,
U F (t ) := {y ∈ O : F (y) < t },
to denote its sublevel sets, and
VF (t ) := {y ∈ O : F (y) > t },
to denote its superlevel sets. We additionally denote the volume of a sublevel set by
UF (t ) := Vold (U F (t )),
§3.3. Weyl’s law and Pólya’s conjecture 107

and the volume of a superlevel set by

VF (t ) := Vold (VF (t )).

Exercise 3.3.24
Let O = (−2, 2) × (−1, 1) ⊂ R2 , and let F : O → R be defined by F (x, y) =
p
x 2 + y 2 . Plot
the graphs of UF (t ) and VF (t ).

We will also need the result of the following

Proposition 3.3.25: A variant of the Bathtub principle [LieLos97, Theorem 1.14]

Let f : Rd → R be a measurable function such that for all t ∈ R, Vold (L f (t )) = 0 and
U f (t ) is finite, and let g : Rd → [0, 1] ∈ L 1 (Rd ). Set
ˆ
© ª
A := g (ξ) dξ, s := sup t : U f (t ) ≤ A . (3.3.13)
Rd

Then ˆ ˆ
f (ξ)g (ξ) dξ ≥ f (ξ) dξ. (3.3.14)
Rd U f (s)

Proof of Proposition 3.3.25

Let h(ξ) := χU f (s) (ξ) be the characteristic function of the set U f (s). Proving (3.3.14) is
equivalent to showing that for any g satisfying the conditions of the Proposition we have
ˆ
¡ ¢
f (ξ) h(ξ) − g (ξ) dξ ≤ 0.
Rd
Since Vold (L f (s)) = 0, we can re-write the integral above as
ˆ ˆ ˆ
 
¡ ¢ ¡ ¢
f (ξ) h(ξ) − g (ξ) dξ =  +  f (ξ) h(ξ) − g (ξ) dξ.
 
(3.3.15)
Rd V f (s) U f (s)

Note that when ξ ∈ V f (s) we have f (ξ) ≥ s and h(ξ) − g (ξ) = −g (ξ) ≤ 0. Similarly, for
ξ ∈ U f (s) we have f (ξ) ≤ s and h(ξ) − g (ξ) = 1 − g (ξ) ≥ 0. Therefore, replacing f (ξ) by
s in both integrals in the right-hand side of (3.3.15) leads to an upper bound, yielding
ˆ ˆ
¡ ¢ ¡ ¢ ¡ ¢
f (ξ) h(ξ) − g (ξ) dξ ≤ s h(ξ) − g (ξ) dξ = s U f (s) − A = 0
Rd Rd
108 Chapter 3. Variational principles and applications

by (3.3.13), which completes the proof.

Proof of Theorem 3.3.21

D
Let u m = u m be an orthonormal sequence of Dirichlet eigenfunctions corresponding to
the eigenvalues λm , m ∈ N. Then we get
k k k
λm = ∥∇u m ∥2L 2 (Ω) = ∥ξ(F u m )∥2L 2 (Rd ) ,
X X X
(3.3.16)
m=1 m=1 m=1

where (F u m )(ξ) is the Fourier transform (see (2.1.3)) of u m extended by zero onto the
whole Rd . The first equality in (3.3.16) follows from the variational principle, and the sec-
ond one from Plancherel’s theorem.
Denote
Vold (Ω) (2π)d X k
f (ξ) := |ξ|2 , g (ξ) := |(F u m )(ξ)|2 ≥ 0. (3.3.17)
(2π)d Vold (Ω) m=1

Then (3.3.16) may be re-written as

k
ˆ
λm =
X
f (ξ)g (ξ) dξ. (3.3.18)
m=1
Rd

We want to estimate the integral in the right-hand side of (3.3.18) using (3.3.14), but need
to show first that the function g (ξ) defined by (3.3.17) satisfies the conditions of Proposi-
tion 3.3.25. By the definition of the Fourier transform, and using the fact that {u m } is an
orthonormal basis in L 2 (Ω), we have

(2π)d X k ¯³ d
´ ¯2
g (ξ) := ¯ (2π)− 2 e−i〈x,ξ〉 , u m 2 ¯
¯ ¯
Vold (Ω) m=1 L (Ω)

1 ° °2
° −i〈x,ξ〉 °
≤ °e ° 2 =1
Vold (Ω) L (Ω)

by Bessel’s inequality, so that (3.3.14) is indeed applicable. By Plancherel’s theorem

∥F u m ∥2L 2 (Rd ) = ∥u m ∥2L 2 (Ω) = 1, and we therefore have
ˆ ˆ
(2π)d X k (2π)d k
A= g (ξ) dξ = |(F u m )(ξ)|2 dξ = .
Vold (Ω) m=1 Vold (Ω)
Rd Rd

³ ´1
(2π)d t 2
Further on, since U f (t ) for t > 0 is the ball of radius r t = Vold (Ω) , the constant s ap-
pearing in (3.3.13) satisfies ³ ´
Vold B rds = (r s )d ωd = A,
§3.3. Weyl’s law and Pólya’s conjecture 109

with ωd given by (B.1.1), from where

¶1
k
µ
d
r s = 2π . (3.3.19)
ωd Vold (Ω)

We now apply (3.3.14) to the right-hand side of (3.3.18):

ˆ ˆ ˆ
Vold (Ω) 2 Vold (Ω)r sd +2
f (ξ)g (ξ) dξ ≥ |ξ| dξ = |ξ′ |2 dξ′
(2π)d (2π)d
Rd B rds Bd

ˆ ˆ1
Vold (Ω)r sd +2 (3.3.20)
= ρ 1+d dρ dκ
(2π)d
Sd −1 0
Vold (Ω)r sd +2 σd −1
= ,
(2π)d d +2

where we have used the changes of variables ξ = r s ξ′ and ξ′ = ρκ, ρ ∈ [0, 1), κ ∈ Sd −1 ,
and have used σd −1 to denote the volume of Sd −1 , see (B.1.2). Substituting (3.3.19) into
(3.3.20) and simplifying with account of σωd d−1 = d , we obtain
ˆ 2
4π2 k 1+ d d
f (ξ)g (ξ) dξ ≥ 2
· .
(Vold (Ω)ωd ) d d +2
Rd

Finally, recalling the definition (3.3.2) of the Weyl constant C d and using (3.3.18), we re-
write the last inequality as (3.3.12).

Remark 3.3.26
The approach of Theorem 3.3.21 can be adapted to prove similar inequalities for the eigen-
values of the Neumann Laplacian, see [Krö92]. In this case
2
k d k 1+ d
λN
X
m (Ω) ≤ ,
m=1 d + 2 (C d Vold (Ω)) d2

and ¶2 µ ¶2
d +2 k
µ
d d
λN
k+1 (Ω) ≤ , k ∈ N.
2 C d Vold (Ω)
For further details, and other applications of Berezin–Li–Yau inequalities, including their
relation to the Lieb–Thirring inequalities and to the asymptotics of the Riesz means, see
[Lap97], [Lie73], [LapSaf96], [LapWei00], and [FraLapWei22].
 CHAPTER 4
Nodal geometry of eigenfunctions
In this chapter, we present nodal geometry of eigenfunctions. We
prove Courant’s nodal domain theorem and show that the nodal set
of an eigenfunction is dense on the wave-length scale. We also obtain
a lower bound for the size of the nodal set in dimension two, and give
an overview of results concerning Yau’s conjecture on the volume of
nodal sets of Laplace–Beltrami eigenfunctions. In particular, we
discuss Donnelly–Fefferman’s estimate on the doubling index of
eigenfunctions and its relation to the nodal volume. We also outline
the proof, following the work of A. Logunov and E. Malinnikova, of
a polynomial upper bound on the nodal volume.

§4.1. Courant’s nodal domain theorem

§4.1.1. Nodal domains and nodal sets

Let Ω ⊂ Rd be a bounded domain, and let u be an eigenfunction of either the Dirichlet or Neu-
mann Laplacian. Consider the set

Z u := {x ∈ Ω : u(x) = 0},

called the nodal set of u . A connected component of Ω \ Z u is called a nodal domain of u . Sim-
ilarly, one defines the nodal domains and nodal sets for Laplace–Beltrami eigenfunctions on a
Riemannian manifold. For an illustration, the nodal set and the nodal domains of some partic-
ular Dirichlet and Neumann eigenfunctions of a unit square are shown in Figure 4.1. See also
Figure 4.2 for the nodal set and the nodal domains of a Laplace–Beltrami eigenfunction on the
sphere.

111
 112 Chapter 4. Nodal geometry of eigenfunctions

Figure 4.1: The nodal sets and the nodal domains of the eigenfunction u D =
p1 (sin(2πx) sin(9πy)−sin(9πx) sin(2πy)−sin(6πx) sin(7πy)+2 sin(7πx) sin(6πy))
5
D 2 2
corresponding to the Dirichlet eigenvalue λ = 85π of the unit square [0, 1]
(left, cf. Figure 1.2) and of the eigenfunction u N = p15 (cos(6πx) cos(43πy) −
cos(11πx) cos(42πy)+cos(38πx) sin(21πy)+2 cos(27πx) sin(34πy)) corresponding
to the Neumann eigenvalue λN = 1885π2 (right).
Ernst Florens Friedrich
Chladni
(1756–1827)
Numerical Exercise 4.1.1
Plot your own analogue of Figure 4.1 for some eigenfunctions of a Laplacian, computed
either using separation of variables or numerically, on a domain of your choice.

The nodal sets and the nodal domains are important geometric characteristics which can be
used to measure “complexity” of eigenfunctions. Their investigation goes back to E. Chladni’s
experiments with vibrating plates at the end of the 18th – beginning of the 19th century (while
Chladni’s figures do not exactly correspond to nodal sets of Laplace eigenfunctions, they illustrate
the same phenomenon).
Marie-Sophie Germain
We refer to [Stö07] for a fascinating story about Chladni’s work, his meeting with Napoleon,
(1776–1831)
and a prize won by Sophie Germain, see also Figure 4.3.
In what follows we assume for simplicity that Ω is a Euclidean domain, though essentially all
the results hold for Riemannian manifolds, either closed or with boundary. Where necessary we
will indicate adjustments that are needed in the Riemannian case.
§4.1. Courant’s nodal domain theorem 113

Figure 4.2: The nodal set and the nodal

domains of an eigenfunction of the Laplace–
Beltrami operator −∆S2 on the round sphere
corresponding to the eigenvalue λ = 17 × 18.

§4.1.2. Courant’s theorem

Let us start with the following simple one-dimensional example.

Example 4.1.2
Consider the Dirichlet problem on Ω = (0, ℓ). Its eigenfunctions are given by

πkx
u k (x) = sin ,
ℓ

with eigenvalues λk = πℓk2 , for k ∈ N. Therefore, Z uk consists of k − 1 zeros equidis-

2 2

tributed on (0, ℓ), and u k has k nodal domains.

Exercise 4.1.3
Consider the Sturm–Liouville eigenvalue problem on the interval (a, b) ⊂ R,

−(pu ′ )′ + qu = λwu in (a, b),

u(a) = u(b) = 0,
114 Chapter 4. Nodal geometry of eigenfunctions

Figure 4.3: Drawing from W. H. Stone, Elementary Lessons on Sound, Macmillan

and Co., London (1879), p. 26, showing how vibrations are excited in a Chladni plate
with a violin bow to create the sand figures of nodal lines.

where p, q, w ∈ C 2 ([a, b]), and p, w are positive functions. The eigenvalues form a se-
quence λ1 ≤ λ2 ≤ . . . ↗ +∞. Using the Sturm oscillation theorem, prove that the num-
ber of nodal domains of an eigenfunction u k corresponding to the eigenvalue λk is equal
to k . For a solution, see [Shu20, Chapter 3].

Example 4.1.2 shows that in one dimension, the k th eigenfunction has precisely k nodal do-
mains. One can easily check using Exercise 1.1.9 that this is no longer true for the square. However,
the following fundamental theorem due to R. Courant [Cou23] holds in all dimensions.
§4.1. Courant’s nodal domain theorem 115

Theorem 4.1.4: Courant’s nodal domain theorem

Let Ω ⊂ Rd be a bounded domain. Suppose that u is a Dirichlet eigenfunction on Ω
corresponding to an eigenvalue λk . Then u has at most k nodal domains.

Remark 4.1.5
Richard Courant
We state Courant’s theorem for the Dirichlet boundary conditions for the sake of simplic-
(1888—1972)
ity. Under additional assumptions on the regularity of ∂Ω, the argument presented below
can be generalised to other self-adjoint boundary conditions, such as Neumann, Robin
or Zaremba.

§4.1.3. Restriction of an eigenfunction to a nodal domain

A non-trivial technical step in the proof of Courant’s theorem is

Theorem 4.1.6
Let u ∈ H01 (Ω) ∩C (Ω), and let Ω1 ⊂ Ω be a nodal domain of u . Then, u|Ω1 ∈ H01 (Ω1 ).

Theorem 4.1.6 immediately follows from Lemma 4.1.7 below under the additional assump-
tions that u ∈ C (Ω) ∩ C 1 (Ω) and u = 0 on ∂Ω. Note that these assumptions are satisfied on Eu-
clidean domains with Lipschitz boundaries by Theorem 2.2.1, part (iv), and on closed manifolds
by Theorem 2.2.17.

Lemma 4.1.7
Let Ω ⊂ Rd be a bounded domain. Suppose that u ∈ C (Ω) ∩ C 1 (Ω) and u = 0 on ∂Ω.
Then u ∈ H01 (Ω).

Proof
We follow the argument in [Buh16]. Let h : R → R be a smooth monotone function such
that h(t ) = 0 on (−1, 1), and h(t ) = t if |t | > 2. Set hε (t ) := εh(t /ε). The function
v ε := h ε ◦u is an element of C 01 (Ω), due to the assumptions on u . We leave it as an exercise
for the reader to show that v ε → u in H 1 (Ω) as ε tends to zero, which implies u ∈ H01 (Ω).

The proof of Theorem 4.1.6 in the general case uses some fine properties of Sobolev spaces
which are discussed below. 10 First, let us recall the following notions.

10 We thank Dorin Bucur for outlining this argument.

116 Chapter 4. Nodal geometry of eigenfunctions

Definition 4.1.8: Capacity

Let E ⊂ Rd . The capacity of E is the number

cap(E ) := inf ∥u∥2H 1 (Rd ) ,

u∈A(E )

where
A(E ) = {u ∈ C 01 (Rd ) ¯ u ≥ 1 in a neighbourhood of E }.
¯

The capacity is an outer measure and it may be used to refine the notion of zero measure,
since cap(E ) = 0 implies that the Lebesgue measure of E vanishes. Note that in R2 , a point has
both zero measure and zero capacity, whereas a segment has zero measure and a positive capacity,
cf. Remark 3.2.15.

Definition 4.1.9: Quasi-everywhere

A property P holds quasi-everywhere in X ⊂ Rd if there exists E ⊂ X such that cap(E ) = 0
and P holds in X \ E .

Definition 4.1.10: Quasi-continuity

A function u : Rd → R is quasi-continuous if for all ε > 0 there exists E ⊂ Rd such that
cap(E ) < ε, and the restriction u|Rd \E is a continuous function.

One can show that any function from the Sobolev space H 1 (Rd ) has a quasi-continuous rep-
resentative. Further, the above notions are useful for characterising the space H01 (Ω) for an open
subset Ω ⊂ Rd or for defining the restriction of an H 1 (Rd ) function on an arbitrary subset of Rd .

Theorem 4.1.11: [HeiKilMar93, Theorem 4.5], [Kin21, Corollary 4.31], see also
[Hed81] and references therein
Let Ω be an open subset of Rd . Then the function u belongs to the Sobolev space H01 (Ω)
if and only if there exists a quasi-continuous function v ∈ H 1 (Rd ) such that v(x) = 0
quasi-everywhere outside Ω and v(x) = u(x) almost everywhere in Ω.

The quasi-continuous representatives are unique in the following sense.

Theorem 4.1.12: [HeiKilMar93, Theorem 4.12], [Kin21, Theorem 4.23]

Let U ⊂ Rd be open. Let v 1 , v 2 be quasi-continuous functions defined in U . If v 1 = v 2
almost everywhere, then v 1 = v 2 quasi-everywhere.

We can now apply these notions in order to prove Theorem 4.1.6.

§4.1. Courant’s nodal domain theorem 117

Proof of Theorem 4.1.6

Since u ∈ H01 (Ω), we can find v as in Theorem 4.1.11. Let

F = {x ∈ Ωc : v(x) ̸= 0} ∪ {x ∈ Ω : v(x) ̸= u(x)},

where Ωc := Rd \ Ω. Since u ∈ C (Ω) we can deduce from Theorem 4.1.12 that cap(F ) = 0.
Let (
v(x), if x ∈ Ω1 ,
w(x) :=
0, if x ̸∈ Ω1 .
We will show that w is quasi-continuous.
Let ε > 0 be given. There exists a set E such that cap(E ) < ε, and v|E c is a continuous
function. Consider the function w restricted to (E ∪F )c . We pick an arbitrary converging
sequence x k → x 0 , where x k and x 0 are points in (E ∪ F )c . Consider the possible cases:

• If x k ∈ Ω1 ∩ (E ∪ F )c and x 0 ∈ Ω1 ∩ (E ∪ F )c , then w(x k ) = v(x k ), w(x 0 ) = v(x 0 ), and

the convergence w(x k ) → w(x 0 ) follows from the continuity of v|E c .

• If x k ∈ Ω1 ∩ (E ∪ F )c and x 0 ∈ (∂Ω1 ) ∩ Ω ∩ (E ∪ F )c , then w(x k ) = v(x k ), w(x 0 ) = 0

and v(x 0 ) = u(x 0 ) = 0. Thus, the continuity of v|E c implies the convergence w(x k ) →
w(x 0 ).

• If x k ∈ Ω1 ∩(E ∪F )c and x 0 ∈ (∂Ω1 )∩(∂Ω)∩(E ∪F )c , then w(x k ) = v(x k ), w(x 0 ) = 0

and v(x 0 ) = 0. Again, we have as above w(x k ) → w(x 0 ).

• If x k ∈ Ωc1 ∩ (E ∪ F )c and x 0 ∈ Ωc1 ∩ (E ∪ F )c , then w(x k ) = 0 and w(x 0 ) = 0, and

trivially w(x k ) → w(x 0 ).

It follows that w|(E ∪F )c is continuous. We have found a quasi-continuous function w such

that w = 0 everywhere in Ωc1 , and w = u almost everywhere in Ω1 . Hence, by Theorem
4.1.11, u ∈ H01 (Ω1 ).

§4.1.4. Proof of Courant’s theorem

Below we give two slightly different proofs of Courant’s theorem: one uses the strict domain
monotonicity, see Proposition 3.2.2, and the other one directly relies on the unique continuation
property of eigenfunctions, see also Remark 4.1.14. Since the latter is needed for the proof of the
strict domain monotonicity, in the end the two arguments use the same set of ideas.

First proof of Theorem 4.1.4

Let u be an eigenfunction corresponding to an eigenvalue λ = λ(Ω) and suppose it has at
least k + 1 nodal domains Ω1 , . . . , Ωk , Ωk+1 , . . . . To prove the theorem, if suffices to show
118 Chapter 4. Nodal geometry of eigenfunctions

that λ > λk . Set (

u(x) if x ∈ Ωi ,
ψi (x) =
0 otherwise.
By Theorem 4.1.6, ψi is an element of H01 (Ωi ). Let L = Span{ψ1 , . . . , ψk }. Since ψ ∈
H01 (Ωi ) and −∆ψi = λψi in Ωi , we deduce that ψi is a Dirichlet eigenfunction in Ωi
with the eigenvalue λ. Therefore,

∥∇ψi ∥2L 2 (Ω)

R[ψi ] = = λ. (4.1.1)
∥ψi ∥2L 2 (Ω)

Set
k
Ω Ωi ,
[
e=
i =1
Pk
then for any linear combination ψ = i =1 c i ψi ∈ L , we have

k k
∥∇ψ∥2L 2 (Ω) |c i |2 ∥∇ψi ∥2L 2 (Ω ) = λ |c i |2 ∥ψi ∥2L 2 (Ω ) = λ∥ψ∥2L 2 (Ω)
X X
e = i i e .
i =1 i =1

Hence, R Ωe [ψ] = λ, and by the variational principle λ ≥ λk (Ω)

e . However, since there are
at least k +1 nodal domains, Ω\ Ω contains a non-empty open set Ωk+1 , and thus by strict
e
domain monotonicity (Proposition 3.2.2),

λ ≥ λk (Ω)
e > λk (Ω),

which completes the first proof of Theorem 4.1.4.

Remark 4.1.13
We recall that since the variational principle for the Dirichlet Laplacian can be applied
without any assumptions on the regularity of the boundary, and since ψi ∈ H01 (Ωi ), we
do not need to impose any smoothness conditions on ∂Ωi for the validity of (4.1.1).

Second proof of Theorem 4.1.4

This proof is due to Å. Pleijel [Ple56]. We argue essentially in the same way as above until
Åke Vilhelm
Carl Pleijel the last step. As before, with L = Span{ψ1 , . . . , ψk } ⊂ H01 (Ω), we have that for all f ∈
(1913–1989) L , R[ f ] = λ. Assume that λ = λk , and that u has at least k + 1 nodal domains. Since
dim L = k , we can choose f ∈ L such that f is orthogonal in L 2 (Ω) to the first k − 1
§4.1. Courant’s nodal domain theorem 119

Dirichlet eigenfunctions u 1 , . . . , u k−1 . Then (see Remark 3.1.21)

∞
λi f i2
P
i =k
λk = R[ f ] = ∞ ,
f i2
P
i =k
¡ ¢
where f i = f , u i L 2 (Ω) are the coefficients in the expansion of f in the basis {u i }. By The-
orem 3.1.9, this equality implies that f is an eigenfunction corresponding to the eigenvalue
λk . Hence, f is real analytic in Ω. But f |Ωk+1 ≡ 0 by construction. It follows that f ≡ 0 on
all Ω, and we get a contradiction. Therefore, an eigenfunction corresponding to λk can
have at most k nodal domains, which completes the second proof of Courant’s theorem.

Remark 4.1.14
Some changes are needed in the above argument in order to prove Courant’s theorem
on Riemannian manifolds. Note that Laplace eigenfunctions on smooth Riemannian
manifolds are smooth but not necessarily real analytic. In this case, in the last step of the
proof above one should use N. Aronszajn’s unique continuation principle, see [Aro57].
It implies that eigenfunctions of elliptic operators with smooth coefficients may vanish at
a given point only to a finite order and, as a consequence, cannot vanish on an open set.
Later on we will also discuss a quantitative version of the unique continuation principle,
see Theorem 4.3.7 and Remark 4.3.19.

Exercise 4.1.15
Deduce from the second proof of Theorem 4.1.4 that without using the unique contin-
uation property one can prove a weaker version of Courant’s bound with k replaced by
k + m(λk ) − 1, where m(λk ) is the multiplicity of the eigenvalue λk .

Let us also make a few historical remarks. The proof of Courant’s theorem in the Rieman-
nian setting appeared first in an influential paper by S.-Y. Cheng [Che75]. The argument relied on
a claim regarding the regularity of the nodal set that was used to justify the application of Green’s
formula, cf. Remark 4.1.13. However, as was pointed out by Y. Colin de Verdière, the proof of this
claim was incomplete in dimensions three and higher. A corrected proof of Courant’s theorem
was presented several years later by P. Bérard and D. Meyer in [BérMey82]. Cheng’s claim regard-
ing the regularity of nodal sets has been finally proved in [HarSim89] by R. Hardt and L. Simon.
For Laplace–Beltrami eigenfunctions, their result can be stated as follows.
120 Chapter 4. Nodal geometry of eigenfunctions

Theorem 4.1.16
Let u be an eigenfunction of the Laplacian on a smooth Riemannian manifold of dimen-
sion d . Then its nodal set decomposes into a regular part Z u ∩ {|∇u| > 0}, which is a
smooth (d −1)-dimensional submanifold having a finite (d −1)-dimensional volume, and
a singular part Z u ∩ {|∇u| = 0}, which is a closed countably (d − 2)-rectifiable subset (see
[Fed14, §3.2.14] for the definition) of the manifold.

Remark 4.1.17
In general, there is no nontrivial lower bound for the number of nodal domains. Antonie
Stern proved in 1925 that for a square and for a round sphere, there exist eigenfunctions
with two nodal domains, corresponding to eigenvalues lying arbitrarily high in the spec-
trum. We refer to [BérHel14] for a recent exposition of these results.

§4.1.5. Properties of subharmonic and harmonic functions

In order to deduce several important corollaries from Courant’s theorem we need to review some
properties of subharmonic and harmonic functions.

Definition 4.1.18: Subharmonic and harmonic functions

Let Ω be an open set. A function u ∈ C 2 (Ω) is called subharmonic in Ω if ∆u ≥ 0 in Ω. If
∆u = 0 in Ω we say that u is harmonic in Ω.

In fact, the notion of subharmonicity can be extended to continuous functions using the
inequality (4.1.4) below, see [AxlBouWad01, p. 224].

Example 4.1.19
Let u be a Laplace eigenfunction on some domain, corresponding to an eigenvalue λ ≥ 0,
and let Ω be a nodal domain of u such that u|Ω < 0. Then u is subharmonic in Ω, since
−∆u = λu ≤ 0 in Ω.

Exercise 4.1.20
Prove that if h is a harmonic function, then |h|2 is subharmonic.

Subharmonic and harmonic functions satisfy a mean value property and a maximum princi-
ple that we discuss below. Given x ∈ Rd , let, as before, S x,r = S dx,r−1 and B x,r = B x,r
d
be the sphere
and the open ball of radius r centred at x .
§4.1. Courant’s nodal domain theorem 121

Definition 4.1.21: Means over spheres and balls

The spherical mean of a locally integrable function u at the point x ∈ Rd is the function
ˆ
1
M u,x (r ) = u := u(x) dS r .
Vold −1 (S x,r )
S x,r S x,r

We will consider also the mean over a ball

ˆ
1
A u,x (r ) = u := u(x) dx.
Vold (B x,r )
B x,r B x,r

For a function u defined on a domain Ω ⊂ Rd , we assume in Definition 4.1.21 that x ∈ Ω, and

that r is chosen small enough for B x,r ⊂ Ω.

Lemma 4.1.22
The derivative of a spherical mean is given by
ˆ
′ 1
M u,x (r ) = ∆u(y) dy. (4.1.2)
Vold −1 (S x,r )
B x,r

Proof
This is a standard result, and we follow the proof of [Shu20, Theorem 6.1]. Let us rewrite
the spherical mean as an average over a unit sphere. Let σd −1 be the volume of a unit sphere
y−x
given by (B.1.2), and set z = r . Then switching to the variable z yields
ˆ ˆ
1 1
M u,x (r ) = u(y) dS r (y) = u(x + r z) dS 1 (z).
σd −1 r d −1 σd −1
S x,r S 0,1

Therefore,
ˆ ˆ
′ 1 d 1
M u,x (r ) = u(x + r z) dS 1 (z) = ∂n u(y) dS r (y),
σd −1 dr σd −1 r d −1
S 0,1 S x,r

where ∂n is the outward normal derivative. Here we used that z is the unit normal at
y ∈ S x,r and made a reverse change of variables. Taking now v ≡ 1 and Ω = B x,r in Green’s
formula (2.1.7), we get
ˆ
′ 1
M u,x (r ) = ∆u(y) dy,
Vold −1 (S x,r )
B x,r

which completes the proof of the lemma.

122 Chapter 4. Nodal geometry of eigenfunctions

Corollary 4.1.23
Let u be a subharmonic function. Then Mu,x (r ) and A u,x (r ) are monotone non-
decreasing in r .

Proof
Indeed, by (4.1.2) and Definition 4.1.18 the derivative of Mu,x (r ) is non-negative for a sub-
harmonic u , and vanishes if u is harmonic. Moreover, it is easy to check that A u,x is a
weighted average of Mu,x : namely, in Rd we have

ˆ1
σd −1
A u,x (r ) = t d −1 M u,x (t r ) dt , (4.1.3)
ωd
0

where ωd , σd −1 are the volumes of the unit ball Bd and the unit sphere Sd −1 , respectively.
Hence it follows that A u,x is monotone non-decreasing as well.

From Corollary 4.1.23, the fact that Mu,x (r ) tends to u(x) as r tends to zero, and the iden-
tity (4.1.3), we readily deduce

Corollary 4.1.24: The mean value inequality for subharmonic functions

Let u be a subharmonic function in B R . Then

u(x) ≤ A u,x (r ) ≤ M u,x (r ) (4.1.4)

for all 0 < r < R . Additionally, if u is harmonic, then the inequalities are replaced by
equalities.

We are now in a position to prove

Theorem 4.1.25: The maximum principle for subharmonic functions

Let Ω ⊂ Rd be a domain and let u ∈ C 2 (Ω) be a subharmonic function. Then u cannot
attain a maximum in Ω unless it is constant.

Proof
Let x 0 ∈ Ω be such that u(x 0 ) ≥ u(x) for all x ∈ Ω. Set m = u(x 0 ) and consider the level
set Z := Lu (m). We want to show that Z = Ω. This follows from the fact that Z is both
open and closed in Ω. Firstly, since u is continuous and Z = u −1 ({m}), it is immediate that
Z is closed. Let us show that Z is also open. Indeed, let y ∈ Z and choose ρ > 0 such that
B y,ρ ⊂ Ω. Then, for all 0 < r < ρ , we have that M u,y (r ) ≥ m by the mean value property.
§4.1. Courant’s nodal domain theorem 123

Therefore, u|S y,r ≡ m for all 0 < r < ρ since m is the maximum. Thus, we get u|B y,ρ ≡ m ,
and so B y,ρ ⊂ Z . It follows that Z is open, and since Ω is connected we have Z = Ω. This
completes the proof of the theorem.

Corollary 4.1.26
Let u satisfy −∆u = λu in a domain Ω, and let x 0 ∈ Ω be such that u(x 0 ) = 0. Then
either u vanishes in a neighbourhood of x 0 or u attains both positive and negative values
in every neighbourhood of x 0 .

Proof
Suppose u does not change sign in a ball B x0 ,r . We can assume that u is non-positive there.
The function u is subharmonic in B (see Example 4.1.19). Then, by Theorem 4.1.25 u is
identically zero in B x0 ,r .

Remark 4.1.27
The maximum principle holds for second order elliptic operators in divergence form, in
particular, for the Laplace–Beltrami operator on a Riemannian manifold. The proof of
this fact uses Hopf’s lemma, see [Eva10, §6.4.2].

The next theorem shows that for harmonic functions the L 2 and L ∞ norms are in a sense
comparable. Such a comparison is also possible for solutions of other elliptic equations, and can
be viewed as part of elliptic regularity.

Theorem 4.1.28: Comparison of L 2 and L ∞ norms

Let h be harmonic in a ball B R(1+δ) ⊂ Rd , with R, δ > 0. Then,

1 d
µ ¶
2 2
|h| ≤ sup |h(x)| ≤ 1 + |h|2 .
x∈B R δ
BR B R(1+δ)

Proof
The left inequality is trivially true for any function. Let x ∗ ∈ B R be such that |h(x ∗ )| =
124 Chapter 4. Nodal geometry of eigenfunctions

supx∈B R |h(x)|. Then by the mean value property and the Cauchy–Schwartz inequality,
¯ ¯2
¯ ¯
¯ ¯ Vold (B R(1+δ) )
|h(x ∗ )|2 = ¯ h¯ ≤ |h|2 ≤ |h|2
¯ ¯
¯
¯B
¯
¯ B Vol d (B x ∗ ,δR )
x ∗ ,δR x ∗ ,δR B R(1+δ)
µ ¶d
1
= 1+ |h|2 .
δ
B R(1+δ)

We record the following important property of positive harmonic functions (see also [GilTru01,
Theorem 2.5]).

Theorem 4.1.29: Harnack’s inequality in concentric balls

Let h be a positive harmonic function in a ball B x0 ,R ⊂ Rd . Then for all x ∈ B x0 ,R/2 we
have h(x) ≤ 2d h(x 0 ).

Proof
By the mean value property of harmonic functions,

h(x) = A h,x (R/2) ≤ 2d A h,x0 (R) = 2d h(x 0 ),

where we have used the fact that B x,R/2 ⊂ B x0 ,R and the positivity to compare the integrals
over these balls.

§4.1.6. Corollaries of Courant’s theorem

Using Courant’s theorem one can show that the first eigenvalue and the first eigenfunctions have
some special features.

Theorem 4.1.30
An eigenfunction corresponding to the eigenvalue λD
1 (Ω) does not vanish in Ω.

Proof
By Courant’s theorem, the first eigenfunction has exactly one nodal domain, i.e. it does
not change sign. The assertion of the theorem then follows from Corollary 4.1.26.
§4.1. Courant’s nodal domain theorem 125

Exercise 4.1.31
Show that an eigenfunction of the Dirichlet Laplacian cannot have nonpositive values at
local maxima or non-negative values at local minima.

Corollary 4.1.32
The first eigenvalue λD
1 is simple.

Proof
By contradiction, assume that u 1 , u 2 are two linearly independent first eigenfunctions.
We can choose u 1 ⊥ u 2 in L 2 (Ω). But this is impossible since they do not vanish.

Corollary 4.1.33
The only Dirichlet eigenfunction that does not change sign is the first eigenfunction. In
particular, if Ω′ ⊂ Ω is a nodal domain of an eigenfunction in Ω with eigenvalue λ, then
λ1 (Ω′ ) = λ.

We leave the proof of Corollary 4.1.33 as an exercise for the reader.

Corollary 4.1.34
The second eigenfunction of the Dirichlet Laplacian has precisely two nodal domains.

Proof
Indeed, it cannot have one nodal domain by Corollary 4.1.33, and it cannot have more than
two nodal domains by Courant’s theorem.

Remark 4.1.35
An eigenvalue λk is called Courant-sharp if it has an eigenfunction with exactly k nodal
domains. In one-dimension, all eigenvalues are Courant-sharp. Furthermore, λ1 and λ2
are always Courant-sharp. How many Courant-sharp eigenvalues can there be? We will
return to this question in Remark 5.1.23.
126 Chapter 4. Nodal geometry of eigenfunctions

§4.2. Density of nodal sets

§4.2.1. Geometric features of nodal sets

In the previous section we focused on the properties of nodal domains of Laplace eigenfunctions.
Let us now explore the geometric features of the nodal sets. Looking at Figure 4.1 we observe that
the nodal lines become more dense as the eigenvalue grows. This is also seen from looking at the
eigenfunction u m,1 (x, y) = sin mx sin y , corresponding to the eigenvalue λ = λm,1 = m 2 + 1
of the square (0, π)2 : it has the nodal set composed of m equally spaced vertical lines. Let us
investigate this phenomenon in more detail.

Definition 4.2.1
Given a set X in a metric space, we say that X is ε-dense (or dense at the scale ε) for some
ε > 0, if any open ball of radius bigger than ε intersects X .

Returning to the eigenfunctions of the square (0, π)2 , we see that Z um,1 is m1 -dense, and there-
fore the scale at which the nodal set is dense is approximately p 1 .
λm,1

Theorem 4.2.2
Let f be a solution of the equation −∆ f = λ f with λ > 0 in a domain Ω ⊂ Rd . Then the
nodal set of f is pcd -dense where
λ
q
c d = j d −1,1 = λ1 (Bd ). (4.2.1)
2

The first proof

We follow the argument in [BérMey82, Appendix D]. Let Ω′ ⋐ Ω be a smooth bounded
subdomain such that f does not vanish in Ω′ . Without loss of generality, suppose that
f > 0 in Ω′ . Let u 1 > 0 be the first Dirichlet eigenfunction in Ω′ , whose corresponding
first eigenvalue is λD ′
1 (Ω ). By Green’s formula (2.1.7) we have
¡ D ′
λ1 (Ω ) − λ u 1 , f L 2 (Ω′ ) = −∆u 1 , f L 2 (Ω′ ) − u 1 , −∆ f L 2 (Ω′ )
¢¡ ¢ ¡ ¢ ¡ ¢
ˆ
¡ ¢
=− (∂n u 1 ) f − u 1 (∂n f ) ds
∂Ω′
ˆ
=− (∂n u 1 ) f ds ≥ 0.
∂Ω′

Indeed, since u 1 |∂Ω′ = 0 and u 1 > 0 in Ω′ , the exterior normal derivative satisfies ∂n u 1 ≤ 0,
and f |∂Ω′ ≥ 0 by continuity. Since (u 1 , f )L 2 (Ω′ ) > 0, we find that
λD ′
1 (Ω ) ≥ λ. (4.2.2)
§4.2. Density of nodal sets 127

Taking Ω′ to be a ball B r and recalling that λD

1 (B r ) = c d r
−2
, we conclude from (4.2.2)
that r ≤ c d λ−1/2
.

The second proof

This proof is essentially taken from [BerNirVar94]. Let Ω′ ⋐ Ω be a smooth bounded
domain, where f is positive on the closure of Ω′ . Let u 1 be the first Dirichlet eigenfunction
of Ω′ , so that u 1 > 0 in Ω′ . Consider the quotient g = uf1 . A direct computation shows
that ­ ®
∇g , ∇ f
−∆g = λD ′
1 (Ω ) − λ g + 2
¡ ¢
.
f

The maximum of g on Ω′ is attained at an interior point x 0 ∈ Ω′ , since g vanishes on ∂Ω′ .

Since ∆ is the trace of the Hessian, one has −∆g (x 0 ) ≥ 0, while g (x 0 ) > 0 and ∇g (x 0 ) = 0.
Hence we deduce (4.2.2) and conclude the argument as in the first proof.

The third proof

Consider the spherical mean (see Definition 4.1.21)

A(r ) := A f ,x (r ) = f.
∂B r

By Lemma 4.1.22, or simply by superposition, the radial function A satisfies the equation

−∆A = λA,

with
d −1 ′
∆A(x) = A ′′ (r ) + A (r ),
r
p
and r = |x|. Let Je(ρ) := A(r ), with ρ = r λ as a dimensionless quantity. Then, Je satisfies
the equation
d −1 ′
Je′′ (ρ) + Je (ρ) + Je(ρ) = 0.
ρ
d
Finally, we set J (ρ) = ρ 2 −1 Je(ρ), and we find that J satisfies the Bessel equation

(d /2 − 1)2
µ ¶
′′ 1 ′
J (ρ) + J (ρ) + 1 − J (ρ) = 0.
ρ ρ2

We conclude that ³ p ´1−d /2 ³ p ´

A(r ) = Cu(x 0 ) r λ J d /2−1 r λ ,
128 Chapter 4. Nodal geometry of eigenfunctions

with C = 2d /2−1 Γ(d /2). Hence, for

j d −1,1
2
r0 = p
λ
n o
we have A(r 0 ) = 0, and it follows that f must vanish at a point on the circle x : |x| = pcd .
λ

The fourth proof

The following elegant proof based on Harnack’s inequality and lifting to harmonic func-
tions (cf. Exercise 4.3.17) is due to T. Colding and W. Minicozzi [ColMin11]. This argument
gives a density result without the sharp constant. Assume p that f is positive in a ball B x0 ,r .
Consider a harmonic function h(x, t ) := f (x) cosh (t λ) in the (d + 1)-dimensional ball
d +1
B (x . Since h is positive there, by Harnack’s inequality (Theorem 4.1.29)
0 ,0),r

h(x 0 , r /2) ≤ 2d +1 h(x 0 , 0) = 2d +1 f (x 0 ).

It follows that p
f (x 0 ) cosh (r λ/2) ≤ 2d +1 f (x 0 ).
p p
Equivalently, cosh(r λ/2) ≤ 2d +1 , or r ≤ (2 arccosh 2d +1 )/ λ.

Given a bounded domain Ω, let ρ Ω denote its inradius. The following result is an immediate
corollary of the density of nodal sets. In view of Corollary 4.1.33, it also easily follows from the
domain monotonicity for the first Dirichlet eigenvalue.

Proposition 4.2.3: [PólSze51, p. 98]

Let Ω ⊂ Rd be a bounded domain, and let u λ be a Dirichlet eigenfunction corresponding
to an eigenvalue λ. Let Ωλ ⊂ Ω be a nodal domain of u λ . Then
cd
ρ Ωλ ≤ p .
λ

We refer to §5.2.3 for further results relating the inradius and the first Dirichlet eigenvalue.

Exercise 4.2.4
Prove the analogue of Proposition 4.2.3 for compact Riemannian manifolds (if the bound-
ary is nonempty, assume Dirichlet boundary conditions). Hint: Use the fact that any Rie-
mannian metric is locally close to Euclidean. A complete proof can be found in [Man08].
§4.2. Density of nodal sets 129

§4.2.2. A lower bound on the size of the nodal set in dimension two

Let us prove the following lower bound on the size of the nodal sets for Dirichlet eigenfunctions
of planar domains.

Theorem 4.2.5: [BrüGro72]

Let Ω ⊂ R2 be a bounded domain, and let u λ be an eigenfunction of −∆D Ω corresponding
p
to an eigenvalue λ > λ1 . Then, the total length of the nodal set satisfies L(Z uλ ) ≥ C λ,
where C is a positive constant independent of λ.

Proof
Let c 2 be defined by (4.2.1), and let us partition the domain Ω using a square grid of size
2c 2
h := p . (4.2.3)
λ

Choose a grid square Q ⊂ Ω and consider the bigger square 3Q of side length 3h formed
by Q and all its neighbours, see Figure 4.4; we assume that Q is such that 3Q ⊂ Ω. By
Theorem 4.2.2, there exists p ∈ Q ∩ Z uλ . If u is identically zero in a neighbourhood of p ,
the theorem is trivially true (in fact, this situation is impossible since the eigenfunctions
are real analytic). Otherwise, consider a nodal line passing through the point p . There are
two possibilities.
If this nodal line leaves 3Q , then its length is at least h .
If the nodal line stays in 3Q , by Corollary 4.1.26 there exists a nodal domain Ω′ such
that p ∈ ∂Ω′ and Ω′ ⊂ 3Q . Let D r be a disk of minimal radius r which contains Ω′ . By
the domain monotonicity, Corollary 4.1.33, and (4.2.3),

c 22 4c 22
= λD D ′
1 (D r ) ≤ λ1 (Ω ) = λ = ,
r2 h2

thus r ≥ h2 . Therefore,

Length(∂Ω′ ∩ 3Q) > 2diam(Ω′ ) ≥ 2r ≥ h.

In either case, we get that the size of the nodal set contained in each square 3Q is at least
2c
h= p 2 . Since for large λ there are O(λ) such squares inside Ω, there exists C > 0 such that
λ p
L(Z uλ ) ≥ C λ.
130 Chapter 4. Nodal geometry of eigenfunctions

Figure 4.4: Grid squares Q (darker shading) and 3Q (lighter

shading) inside a planar domain, with p ∈ Q , and a nodal line
passing through p and existing 3Q on the left, or staying closed
in 3Q on the right.

Remark 4.2.6
For Euclidean domains with Neumann boundary conditions the proof of Theorem 4.2.5
can be repeated essentially verbatim. In order to generalise it for surfaces with a Rieman-
nian metric, some further observations are required. Note that all the measurements
³ ´ in
1
the proof of Theorem 4.2.5 are made in small neighbourhoods of size O p . Due to
λ
the existence of local isothermal coordinates on a surface, we may assume that in¢ each
neighbourhood the Riemannian metric has the form ds 2 = h(x, y) dx 2 + dy 2 with
¡
1
K ≤ h(x, y) ≤ K for some K > 0. Then the Riemannian lengths and their Euclidean
counterparts are comparable, i.e. they differ by at most a factor which is controlled by K .
Moreover, as follows from the variational principle and the conformal equivalence of the
Dirichlet energy in two dimensions (3.1.14), the eigenvalues of the Laplacian in the Rie-
mannian metric ds 2 are comparable to the corresponding eigenvalues of the Euclidean
Laplacian. Hence, the proof of Theorem 4.2.5 could be adapted to the Riemannian case.
§4.3. Yau’s conjecture on the volume of nodal sets 131

This result was obtained in [Brü78] and independently by S.-T. Yau.

Interestingly enough, the analogue of Theorem 4.2.5 for surfaces can be proved with an ex-
plicit universal constant. The following result is due to A. Savo [Sav01].

Theorem 4.2.7
Let M be a compact Riemannian surface without boundary. Then
¢ 1 p
L Z uλ > Area(M ) λ
¡
(4.2.4)
11
for sufficiently large λ.

It is a challenging open question to find the optimal constant in inequality (4.2.4). It is

suggested in [Sav01] that the possible answer is π1 with equality attained by the eigenfunctions
u m (x, y) = sin mx , m → ∞, on a flat square torus.

§4.3. Yau’s conjecture on the volume of nodal sets

§4.3.1. Nodal volume and doubling index

In higher dimensions, the method of the proof of Theorem 4.2.5 fails for the following reason. It is
easy to see that the above argument does not rule out “needle-like” nodal sets, for which the diam-
eter is large, but the volume could be made arbitrary small. Still, in 1982, S.-T. Yau [Yau82] made
a conjecture that the following two-sided inequality holds for an arbitrary closed d -dimensional
Riemannian manifold M :
p p
C 1 λ ≤ H d −1 (Z uλ ) ≤ C 2 λ, (4.3.1)

with some constants C 1 ,C 2 > 0 depending only on the metric. Here H d −1 (·) denotes the (d −1)-
dimensional Hausdorff measure, which is a generalisation of the notion of the (d −1)–dimensional
volume (see [Fed14, Introduction and §3.2.46] for the definition). Yau’s conjecture has attracted
a lot of attention in the past decades. In 1988, the following fundamental result was proved by H.
Donnelly and C. Fefferman.

Theorem 4.3.1: [DonFef88]

Assume that the Riemannian metric on M is real analytic. Then Yau’s conjecture (4.3.1)
holds.

In particular, this proves the upper bound in Yau’s conjecture for the standard two dimen-
sional sphere and both upper and lower bounds for higher dimensional spheres, all previously
unknown cases.
132 Chapter 4. Nodal geometry of eigenfunctions

The approach of Donnelly–Fefferman has been recently significantly developed by A. Lo-

gunov and E. Malinnikova (see [LogMal18b] and references therein), who obtained several break-
through results for smooth manifolds.

Theorem 4.3.2: [Log18a, Log18b]

Let M be a closed d -dimensional Riemannian manifold endowed with a smooth Rieman-
nian metric. Then p
C 1 λ ≤ H d −1 (Z uλ ) ≤ C 2 λS , (4.3.2)
where S = S(M ) is a positive constant.

In particular, the lower bound in Yau’s conjecture holds. The polynomial upper
³ bound
p ´
in
c λ
(4.3.2) is a breakthrough compared with the Hardt–Simon exponential estimate O λ that
has been known earlier [HarSim89, Theorem 5.3]. Note that the upper bound O λ1/2 in (4.3.1)
¡ ¢

is still not proved even in two dimensions. In the planar case, the best known exponent is 34 −ε for
a certain small ε > 0 [LogMal18a]. H. Donnelly and C. Fefferman [DonFef90], and R.-T. Dong
[Don92], have previously proved a two-dimensional upper bound with the exponent 34 .
The goal of this section is to explain some ideas behind the proofs of Theorems 4.3.1 and
4.3.2, with a particular focus on the upper bound in (4.3.2) which we discuss in detail. One of
the key observations is that in order to estimate the nodal volume one needs to understand well
the growth properties of the eigenfunctions, see Remark 4.3.9 below. Recall that a geodesic ball
B := B x,r ⊂ M is the image of the Euclidean ball B 0,r ⊂ T x M under the exponential map (see
[Bur98, §3.3]), where r > 0 is small enough so that this map is a diffeomorphism. Similarly to the
Euclidean balls, we write cB := B x,cr .

Definition 4.3.3: The doubling index

Let B ⊂ M be a geodesic ball such that 2B ⊂ M is also a geodesic ball, and assume that
f ∈ C (2B ) is not the zero function. The L ∞ -doubling index of f (or simply its doubling
index) is the number
 ¯ ¯
sup ¯ f (x)¯
∥ f ∥L ∞ (2B )
µ ¶
 x∈2B
β( f , B ) := log2  ¯  = log2 .

¯
sup¯ f (x)¯ ∥ f ∥L ∞ (B )
x∈B

Example 4.3.4
If P n is a homogeneous polynomial of degree n in d variables, then

β P n , B 0,r = n.
¡ ¢

The doubling index is closely related to the vanishing order of a smooth function. The van-
§4.3. Yau’s conjecture on the volume of nodal sets 133

¡ ¢
ishing order ordx f of a function f at the point x is defined as the maximal integer k such that
all the derivatives of f of order smaller than
¡ ¢
k vanish at x . If no such k exists we say that f vanishes
¡ ¢
to infinite order at x . For instance, ordx f = 0 if f (x) ̸= 0, ordx f = 1 if x is a simple zero of
2
f , and f (x) = e −1/x vanishes to infinite order at x = 0.

Exercise 4.3.5: Doubling index and vanishing order

Let f be a smooth function.

(i) Show that if f has a finite vanishing order at x , then

ordx f = lim β f , B x,r .

¡ ¢ ¡ ¢
r →0

(ii) Show that if there exists¡a ¢constant C > 0 such that β( f , B x,r ) ≤ C for all small
enough r > 0, then ordx f ≤ C .

The following important fact of independent interest established in [DonFef88] is heavily

used in the proofs of both Theorems
p 4.3.1 and 4.3.2. Roughly speaking, it says that eigenfunctions
grow like polynomials of degree λ, similarly to the spherical harmonics.

Theorem 4.3.6: The Donnelly–Fefferman growth bound

Let u λ be a Laplace eigenfunction on a Riemannian manifold M . Then for any geodesic
ball B ⊂ M such that 2B is also a geodesic ball in M ,
p
β(u λ , B ) ≤ C M λ,

where C M is a constant depending only on the geometry of M .

We review the main ideas involved in its proof in §4.3.2. In view of the second part of Exercise
4.3.5, Theorem 4.3.6 immediately implies

Theorem 4.3.7
Let u λ be a Laplace eigenfunction on a smooth Riemannian manifold M . Then
p
ordx (u λ ) ≤ C M λ

at any point x ∈ M .

Theorem 4.3.7 could be viewed as a quantitative version of the Aronszajn’s unique continu-
ation result for Laplace eigenfunctions, see Remark 4.1.14.
134 Chapter 4. Nodal geometry of eigenfunctions

Exercise 4.3.8
Use spherical harmonics to show that the bounds in Theorems 4.3.6 and 4.3.7 are sharp.

Remark 4.3.9: The doubling index and nodal volume

There is a natural link between the degree of a polynomial and the size of its zero set. In
one real dimension, a polynomial of degree d has at most d zeros; in higher dimensions,
Milnor’s bound on the number of connected components of the zero set in terms of the
degree yields an estimate on the nodal volume [HarSim89, Theorem 2.1]. In one com-
plex dimension the number of zeros, counted with multiplicities, equals the degree. For
a holomorphic function in C, the number of zeros is bounded by its growth (Jensen’s
formula), see e.g. [LogMal18b, §4.2]. This result and the Crofton formula play an im-
portant role in the proof of the upper bound in Yau’s conjecture in the real analytic case.
In the smooth case, one needs to develop other methods which connect the growth of a
harmonic function to the size of its nodal set. For solutions of second order elliptic equa-
tions with smooth coefficients one has an important result of R. Hardt and L. Simon
[HarSim89, Theorem 1.7], which together with Theorem 4.3.6 implies the existence of an
upper bound on the size of the nodal set. In particular, if h is a solution of such an equation
(e.g. a harmonic function or an eigenfunction of a Laplacian) in a ball 2B , then

H d −1 (Z h ∩ B ) ≤ C β(h, B )C β(h,B ) , (4.3.3)

with some constant C > 0 independent of h .

§4.3.2. The Donnelly–Fefferman growth bound: a sketch of the proof

In this section we prove Theorem 4.3.6 under the simplifying assumption that M is endowed with
a locally Euclidean metric, which allows us to consider only harmonic functions. The proof we
give illustrates the main ideas and is adaptable to general smooth Riemannian manifolds using
the standard techniques of elliptic theory, since our arguments do not rely on the real analyticity
of harmonic functions.
The proof of Theorem 4.3.6 is based on a monotonicity property of the doubling index of a
harmonic function, which goes back to T. Carleman [Car33], S. Agmon [Agm65] and F. J. Alm-
gren [Alm00]. The monotonicity in the context of general elliptic equations of second order and
related applications are due to N. Garofalo and F.-H. Lin [GarLin86].

Definition 4.3.10: The height and frequency functions

Consider a continuous function f defined in a ball B x0 ,R ⊂ Rd . The height function of f
§4.3. Yau’s conjecture on the volume of nodal sets 135

is given by
H f (r ) = H f (x 0 , r ) := f 2, r ∈ (0, R)
∂B x0 ,r

(see Definition 4.1.21).

The frequency function of f is defined by

r H ′f (x 0 , r )
N f (r ) = N f (x 0 , r ) := , r ∈ (0, R). (4.3.4)
2H f (x 0 , r )

The height function of a harmonic function h is monotonically non-decreasing, since h 2

is subharmonic (see Lemma 4.1.22). Recall that a function is called logarithmically convex if its
logarithm is a convex function. The following result holds.

Theorem 4.3.11: Monotonicity of the frequency function

Let h be a harmonic function defined in a Euclidean ball B R . The function t 7→ Hh (et )
defined in R is logarithmically convex. Equivalently, the frequency function Nh (r ) is
monotonically non-decreasing.

Proof
In dimension¡ two, working in polar coordinates (r, θ) one easily verifies the convexity of
t |m| imθ
¡ t¢
t 7→ log Hhm e for h m (r, θ) := r e . Indeed, in this case log Hhm e = 2|m|t is just
¢

linear. Then, the orthogonal decomposition

h(r, θ) = a m h m (r, θ)
X
m∈Z

shows that
H h et = |a m |2 e2|m|t ,
¡ ¢ X
m∈Z
the logarithm of which is convex (see Exercise 4.3.13 below).
Similarly, for a ball B R ⊂ Rd +1 , d ≥ 2, one uses the expansion into spherical harmonics
to write
∞ µ ¶
x
c k, j |x|k Pek, j
X X
h(x) = ,
k=0 Pek, j ∈H
fk |x|

where H fk is a space of homogeneous harmonic polynomials (or spherical harmonics)

of degree k whose elements {Pek, j } are chosen to be orthonormal in L 2 Sd ; the dimen-
¡ ¢

sion of Hfk is given in Theorem 1.2.16. The height function Hh (et ) for each term
³ ´ k, j
x
h k, j = c k, j |x|k Pek, j
in this expansion is equal to |c k, j |2 e2kt and hence is logarithmi-
|x|
cally convex. Therefore, as above, Hh (et ) is also logarithmically convex. The equivalence
136 Chapter 4. Nodal geometry of eigenfunctions

of this property to the monotonicity of the frequency function follows immediately by

noting that
d
log Hh (et ) = 2Nh (et ).
dt

Remark 4.3.12
Theorem 4.3.11 may be proved using integration by parts, which is adaptable to general
manifolds, where no orthogonal decomposition is available, see, e.g., [Agm65], [Log-
Mal20].

Exercise 4.3.13
Show that if f 1 and f 2 are positive functions in some open interval I ⊂ R, and log f 1 , log f 2
are convex, then log( f 1 + f 2 ) is also convex. Hint: use the geometric–arithmetic mean
inequality.

Exercise 4.3.14
Show that the frequency function can be expressed as
´ 2
rB (x 0 ,r ) |∇h| dx
Nh (x 0 , r ) = ´ 2
.
∂B (x 0 ,r ) |h| dS r

In what follows we often use a shortcut notation B r := B x0 ,r for concentric balls provided
the centre x 0 can be an arbitrary fixed point.

Theorem 4.3.15
Let R > 0, and let h be a harmonic function in Ω ⊃ B cR for some fixed c > 1. Then, the
quantity ffl 2
1 ∂B cr |h| 1 Hh (cr )
N (h, B r , c) := log2 ffl = log2 (4.3.5)
2 2 2 Hh (r )
∂B |h|
r

is monotonically non-decreasing in r for r ∈ (0, R).

§4.3. Yau’s conjecture on the volume of nodal sets 137

Proof
Using the definition (4.3.4), we have

ˆc ˆc ˆc
Nh (t r ) r Hh′ (t r ) 1 d¡ ¢
dt = dt = log Hh (t r ) dt
t 2Hh (t r ) 2 dt
1 1 1
1¡ ¢
= log Hh (cr ) − log Hh (cr ) = (log 2)N (h, B r , c),
2
and therefore
ˆc
1 Nh (t r )
N (h, B r , c) = dt . (4.3.6)
log 2 t
1

Since by Theorem 4.3.11 the frequency function Nh is monotone, (4.3.6) shows that
N (h, B r , c) is monotone in r .

In view of (4.3.6), we call N (h, B r , c) the integrated frequency. It can be also viewed as an L 2
analogue of the doubling index.

Exercise 4.3.16
One may define versions of the height and frequency functions Hh , Nh for balls as
¡ ¢′
r Hhb (r )
Hhb (r ) := 2
|h| , Nhb (r ) = .
2Hhb (r )
B

Show using essentially the same arguments as above that

ffl 2
1 B |h|
N b (h, B r , c) := log2 ffl cr 2
2 B |h|
r

is monotonically non-decreasing in r . Show also that

Hhb (r ) ≤ Hh (r ), Nhb (r ) ≤ N (r ), N b (h, B r , c) ≤ N (h, B r , c).

Hint: Observe that

ˆ1
Hhb (r ) 1 Hh (t r )
= Vol(∂B t ) dt ,
Hh (r ) Vol(B 1 ) Hh (r )
0

where the integrand is monotonically non-increasing in r by Theorem 4.3.15.

To prove Theorem 4.3.6 we will apply the following lifting trick to reduce it to the case of
138 Chapter 4. Nodal geometry of eigenfunctions

harmonic functions (cf. proof of Theorem 2.2.1, part (ii)).

Exercise 4.3.17: Lifting trick

Consider an open product Riemannian manifold M × I , where I ⊂ R, and let u λ be an
eigenfunction of the Laplace–Beltrami operator on M corresponding to an eigenvalue λ.
Show that the function
³p ´
h(x, t ) := u λ (x) cosh λt , (x, t ) ∈ M × (−1, 1),

is harmonic in M × I .

Theorem 4.3.18: A local version of the Donnelly–Fefferman growth bound

Let u satisfy the equation −∆u = λu in a ball B R . Then the following statements hold.

(i) For all 0 < r ≤ 32 s < R3 one has

p
β(u, B r ) ≤ C (β(u, B s ) + s λ + 1).

(ii) The following three–ball inequality holds for all 0 < r < R4 ,
à !α à !1−α
p
Cr λ
sup |u(x)| ≤ C e sup |u(x)| sup |u(x)| ,
x∈B 2r x∈B r x∈B 4r

with some α ∈ (0, 1) independent of λ and u .

Here, C > 0 denotes some constants (possibly, different) which are independent of λ and
u.

Remark 4.3.19: Aronszajn’s unique continuation principle

In particular, by fixing s and letting r tend to zero, it follows from the corresponding
version of Theorem 4.3.18 for smooth manifolds and Exercise 4.3.5 that if −∆g u = λu
and u has a zero of infinite order then u is identically zero (see also [Aro57]).

Proof of Theorem 4.3.18

We follow [Man13]. Let B = B x0 ,r , and let B rd +1 be the (d + 1)-dimensional ball with cen-
tre (x 0 , 0) and radius r . Lift the eigenfunction u(x) to a harmonic function h(x, t ) on
B rd +1 as in Exercise 4.3.17.
§4.3. Yau’s conjecture on the volume of nodal sets 139

We observe that for any r, δ with r (1 + δ) ∈ (0, R),

sup |u(x)|2 ≤ sup |h(x, t )|2

x∈B r (x,t )∈B rd +1

1 d +1
µ ¶
Theorems 4.1.28, 4.1.24 (4.3.7)
≤ 1+ |h|2 ,
δ
∂B rd(1+δ)
+1

while
p p
|h|2 ≤ sup |h|2 ≤ sup |u|2 · (cosh(r λ))2 ≤ sup |u|2 · e2r λ . (4.3.8)
∂B rd +1 Br Br
∂B rd +1

Applying Theorem 4.3.15 to h ,

ffl 2
supx∈B 2r |u(x)|2 (4.3.7),(4.3.8) d +1 2r pλ ∂B 3rd +1 |h(x, t )|
≤ 3 e ffl
supx∈B r |u(x)|2 ∂B d +1 |h(x, t )|
2
ffl
r
ffl 2
d +1 |h(x, t )|
2 d +1 |h(x, t )|
p ∂B p
d +1 2r λ ffl∂B 3r
· ffl
3r
=3 e 2 2
d +1 |h(x, t )|
∂B p ∂B rd +1 |h(x, t )|
3r
p
p
(4.3.5) d +1 2r λ Hh (3r ) Hh ( 3r )
= 3 e p ·
Hh ( 3r ) Hh (r )
Theorem 4.3.15 d +1 2r pλ Hh (3r ) 2
µ ¶
≤ 3 e p
Hh ( 3r )
Theorem 4.3.15 d +1 4s pλ/3 Hh (2s) 2
µ ¶
≤ 3 e p
Hh (2s/ 3)
!2
supx∈B 2s |u(x)|2
Ã
(4.3.8),(4.3.7) p
28s λ/3
≤ Ce ,
supx∈B s |u(x)|2

where C denotes a positive constant independent of λ. Taking logarithms on both sides

we obtain the first part of the theorem. By substituting s = 2r in the preceding inequality
we arrive at
à !1/3 à !2/3
p
19r λ
sup |u(x)| ≤ C e sup |u(x)| sup |u λ (x)| ,
x∈B 2r x∈B r x∈B 4r

thus establishing statement in part (ii) with α = 13 .

Finally, we apply Theorem 4.3.18 together with the fact that the manifold M is compact in
order to prove Theorem 4.3.6.
140 Chapter 4. Nodal geometry of eigenfunctions

Proof of Theorem 4.3.6

Fix r 0 > 0 such that every ball of radius 3r 0 is geodesic, i.e., 3r 0 is smaller than the injec-
tivity radius of M , see [Cha84, p. 118]. We first show that Theorem 4.3.6 holds for any
ball B p,r of radius r ≥ r 0 . Normalise u λ so that supM |u λ | = 1. Let x ∗ ∈ M be a point
where |u λ (x ∗ )| = 1. Let x 0 , x 1 , . . . , x N be a sequence of points such that x 0 = p, x N =
x ∗ , d (x j , x j +1 ) < r 0 and such that N depends on the geometry of M and r 0 only. Ob-
serve that B x j ,2r 0 ⊃ B x j +1 ,r 0 . The three-ball inequality of Theorem 4.3.18 with α = 31 gives,
taking into account that |u λ | ≤ 1,
à !3
p
−1 −3C r 0 λ
sup |u λ (x)| ≥ C e sup |u λ (x)|
x∈B x j ,r 0 x∈B x j ,2r 0
à !3
p
−1 −3C r 0 λ
≥C e sup |u λ (x)| .
x∈B x j +1 ,r 0

Using this inequality recursively for j = N − 1, N − 2, . . . , 0, we arrive at

à !3 N
p p
′ −C ′′ r 0 λ ′′
r0 λ
sup |u λ (x)| ≥ C e sup |u λ (x)| = C ′ e−C ,
x∈B x0 ,r 0 x∈B x∗ ,r 0

where
N 3N − 1
C ′ = C −(3 −1)/2
, C ′′ = 3 .
2
The preceding inequality shows that for all r ≥ r 0 ,
p
′′
r0 λ
sup |u λ (x)| ≥ sup |u λ (x)| ≥ C ′ e−C
x∈B x0 ,r x∈B x0 ,r 0
p
′′
r0 λ
≥ C ′ e−C sup |u λ (x)|.
x∈B x0 ,2r

Recalling Definition 4.3.3, we have proved, in other words, that for all r ≥ r 0 ,
p
β(u λ , B p,r ) ≤ C 1 λ +C 2 , (4.3.9)

where C 1 ,C 2 depend only on r 0 and the geometry of M . For 0 < r < r 0 we apply part (i) of
Theorem 4.3.18 with s = 32 r 0 and inequality (4.3.9) to get that for any ball B = B (x, r ) ⊂ M
p
β(u λ , B ) ≤ C 3 λ +C 4 ,

where C 3 ,C 4 depend only on M . Finally, since we may assume that λ ≥ λ1 (M ) > 0 (the
case λ = 0 being trivial), we can absorb the additive constant C 4 in the multiplicative con-
stant C 3 .
§4.3. Yau’s conjecture on the volume of nodal sets 141

§4.3.3. Distribution of doubling indices: a combinatorial approach

Spectacular recent progress on Yau’s conjecture due to Logunov and Malinnikova [LogMal18a,
Log18a, Log18b] is based on a better understanding of the distribution of doubling indices. The-
orem 4.3.6 gives a worst-case scenario, but in reality, in most of the balls the doubling index is
much smaller. In view of Remark 4.3.9 this should lead to better nodal estimates. We note that
this observation in various forms is also key for the proof of Theorem 4.3.1, as well as the lower
bound in Theorem 4.3.2. For instance, in dimension two, the upper bound in Yau’s conjecture
p is equivalent to showing that the doubling indices of an eigenfunction u λ on balls of radii
(4.3.1)
C / λ are bounded on average, see [NazPolSod05, RoF15].
Below we survey some of the important insights on the distribution of doubling indices that
has led to the proof of the polynomial upper bound in Theorem 4.3.2. Remarkably, a key idea
discussed in this subsection is purely combinatorial.
In what follows, we work with cubes rather than with balls: it does not make an essential
difference and is more convenient for combinatorial purposes. However, minor technical issues
appear. We denote by Q a cube in Rd , and by αQ a concentric cube with parallel sides of length
αs(Q) , where s(Q) is the side length of Q .
Slightly abusing notation, given a continuous function f : ℓQ → R, we define the doubling
index of a cube Q by
∥ f ∥L ∞ (ℓQ)
β( f ,Q) = log2 ,
∥ f ∥L ∞ (Q)
p
where ℓ is a fixed large odd integer depending on dimension (one can take ℓ > 2 d ). The integer
ℓ appears in order to allow the comparison of β( f ,Q) with relevant quantities of the inscribed
and circumscribed balls.

Lemma 4.3.20: Combinatorial lemma for an arbitrary function

Let f be a continuous function in ℓQ ⊂ Rd . Subdivide ℓQ into (ℓK )d equal subcubes of
side length K1 s(Q). Assume that β(h, q) > β0 for each subcube q with ℓq ⊂ ℓQ . Then
β(h,Q) > K β0 .

Proof
Find a subcube q0 of Q and a point x 0 ∈ q0 such that | f (x 0 )| = maxx∈Q | f (x)|. Since
β(h, q 0 ) > β0 we can find a point x 1 ∈ ℓq 0 such that | f (x 1 )| > 2β0 | f (x 0 )|, and a subcube
q 1 such that x 1 ∈ q 1 . Observe that if K > 1 then ℓq 1 ⊂ ℓQ and β(h, q 1 ) > β0 . At the
( j +1)th step, as long as j < K we find a point x j +1 ∈ ℓq j such that | f (x j +1 )| > 2β0 | f (x j )|
and a subcube q j +1 such that x j +1 ∈ q j +1 . For j = K − 1 we get | f (x K )| > 2K β0 | f (x 0 )|,
see Figure 4.5.

In order to iterate Lemma 4.3.20 one has to know that an upper bound on the doubling index
does not grow after a subdivision. In general this is obviously false, but for harmonic functions
it is essentially the content of the monotonicity Theorem 4.3.15 after replacing the L 2 estimates
142 Chapter 4. Nodal geometry of eigenfunctions

Figure 4.5: Cubes Q and ℓQ , the latter subdivided into

(ℓK )d equal subcubes, shown here for d = 2, ℓ = 3, and K = 4.
The subcubes q satisfying ℓq ⊂ ℓQ are shaded grey. Also, an
example of the sequence q j of subcubes appearing in the proof
of Lemma 4.3.20; the corresponding cubes ℓq j are shown by
dashed lines.

by L ∞ ones (with the same arguments in the proofs of Theorems 4.3.18 and 4.3.6). One obtains
Lemma 4.3.21 stated below which provides such a monotonicity result when the cubes are not
p of the exterior one (cf. the case λ = 0
concentric and when the inner cube is far from the boundary
of Theorem 4.3.18). We have fixed ℓ to be larger than 2 d above exactly in order for this lemma
to hold.

Lemma 4.3.21: [LogMal18a, Hal22]

There exist a positive constant C 0 and a positive odd integer T such that for any harmonic
function h in a cube ℓQ ⊂ Rd and for any subcube q ⊂ T1 Q ,

β(h, q) ≤ C 0 β(h,Q) +C 0 .
§4.3. Yau’s conjecture on the volume of nodal sets 143

Given a harmonic function h : ℓQ → R, we introduce the notation

βsup (h,Q) := sup β(h, q).
q⊂Q

The quantity βsup is convenient since it is monotonic with respect to the inclusion of cubes.
Lemma 4.3.21 implies that
βsup h, q ≤ 2C 0 max{β(h,Q), 1}
¡ ¢
(4.3.10)
for a function h harmonic in ℓQ and any q ⊂ T1 Q .
Set β0 = 2C 0 . Iterating Lemma 4.3.20 we get

Lemma 4.3.22: Combinatorial lemma for harmonic functions

Let a cube Q 0 ⊂ Rd be subdivided into A md equal subcubes Q m , where m ∈ N, and A ∈ N
is greater than some constant A 0 . For any harmonic function h in ℓQ 0 , one can regroup
the subcubes Q m into m + 1 disjoint subsets G 0m , . . . ,G m
m
such that

βsup (h,Q 0 )
½ ¾
β sup
(h,Q m
j ) ≤ max , β0 for all Q m m
j ∈Gj , (4.3.11)
2j

and à !
m ³ d ´m− j
#G m
j = · A −1 .
j

Proof
Set β := βsup (h,Q 0 ) and s 0 = s(Q 0 ). We argue by induction. For m = 0 there is nothing to
prove. Suppose G 0m , . . . ,G m
m
are defined and satisfy the required properties. Partition each
subcube Q j ∈ G j of side length A −m s 0 into (ℓK )d equal sized subcubes q m
m m
j ,k
with side
³ ´
s0
length A m ℓK as in Lemma 4.3.20. Since β h, ℓ1 Q m
j
≤ βsup (h,Q m
j
) ≤ max{2− j β, β0 } by
(4.3.11), we can apply the contrapositive of Lemma 4.3.20 to the cube ℓ1 Q m
j
of side length
s0 m m
A m ℓ , and therefore we can find a subcube q j ,k 0 of Q j such that
³ ´ 1
β h, q m
j ,k 0 ≤ max{2− j β, β0 },
K
s0
and ℓq m
j ,k 0
⊂ Qm
j
. Consider the cube T1 q m
j ,k 0
of side length A m T ℓK . Applying (4.3.10) with
1
Q = qm
j ,k 0
and q = T qm
j ,k 0
, we have, given that β0 = 2C 0 , and choosing K > 4C 0 ,
µ ¶
1 m n ³ ´ o
β sup
h, q j ,k0 ≤ 2C 0 max β h, q m
j ,k 0 , 1
T
½ ¾ (4.3.12)
2C 0 − j 2C 0 n o
≤ max 2 β, β0 , 2C 0 ≤ max 2− j −1 β, β0 .
K K
144 Chapter 4. Nodal geometry of eigenfunctions

We have proved that if we partition Q m j

into (T ℓK 0 )d equal subcubes with K 0 := ⌈4C 0 ⌉,
there exists at least one such subcube T1 q m
j ,k 0
for which (4.3.12) holds (here we used the fact
that T was chosen to be odd in Lemma 4.3.21).
Let us now re-partition Q m j
into A d equal subcubes q with

A ≥ A 0 := 3T ℓK 0 = 3T ℓ⌈4C 0 ⌉

(which corresponds to partitioning the original cube Q 0 into A (m+1)d subcubes). Then
1 m
T q j ,k 0 contains at least one such subcube q , and from (4.3.12) and the monotonicity of
βsup we have βsup (h, q) ≤ max{2− j −1 β, β0 }.
We add q to G m+1
j +1
. We add the other A d − 1 remaining subcubes of Q m
j
to G m+1
j
.
Counting the contributions of G m
j
and G m
j +1
to G m+1
j +1
, we arrive at the following recursion,

#G m+1 m d m
j +1 = #G j + (A − 1) · #G j +1 .

This is a classical recursion of a weighted Pascal triangle with initial condition #G 00 = 1. Its
¡m ¢ ¡ d ¢m− j
solution is #G m
j
= j · A − 1 , see Exercise 4.3.23 below. This completes the proof
of the theorem.

Exercise 4.3.23: Weighted Pascal triangle

Let g (m, j ) be a function defined for all pairs of non-negative integers (m, j ) such that
0 ≤ j ≤ m . Assume that it satisfies the recursion g (m, j ) = ag (m−1, j )+bg (m−1, j −1),
where g (m, j ) is interpreted as 0 when j < 0 or j > m , and a, b > 0 are fixed constants.
¡m ¢ m− j j
In addition assume that g (0, 0) = 1. Prove that g (m, j ) = j a b .

§4.3.4. A polynomial upper bound: an overview

The proof of the polynomial upper bound on the size of the nodal set is based on further im-
provements of Lemma 4.3.22. Fix a number A . When m is large enough (independently of
the harmonic function h ) one observes that #G 0 , the number of subcubes for which the dou-
bling index is is greater than β0 and is not guaranteed to decrease by Lemma 4.3.22, is arbitrarily
small compared to the total number of subcubes in the subdivision of Q = Q 0 . While the total
number of subcubes is A md , at most #G 0 = (A d − 1)m < 0.01A md subcubes q := Q 0m satisfy
βsup (h, q) > max βsup (h,Q)/2, β0 . It turns out that the number of these “bad” cubes is even
© ª

smaller by an order of magnitude, i.e., it can be compared to the number of subcubes on ∂Q .

Theorem 4.3.24: [Log18a]

Subdivide a cube Q ⊂ Rd into A d equal subcubes q , where A is greater than some con-
stant A 1 . Let h be a harmonic function in ℓQ . Then, the number of subcubes such that
§4.3. Yau’s conjecture on the volume of nodal sets 145

βsup (h, q) > max βsup (h,Q)/2, β0 is at most 0.9A d −1 .

© ª

The proof of Theorem 4.3.24 is based on the following two ideas. For a harmonic function
one can improve Lemma 4.3.20 as follows: if there exist only a few (say, d + 1) bad subcubes (i.e.
where the doubling index is large) that are well distributed in ℓQ then one can still deduce that the
doubling index of Q is even bigger (the Simplex lemma [Log18a, §2]). Accordingly, if the doubling
index of Q is small, the bad subcubes should be spread along a hyperplane. This idea can be applied
to deduce that the number of bad subcubes is at most A d −1 . In turn, the cubes along a hyperplane
cannot be all bad, since otherwise, a quantitative version of the Cauchy data uniqueness theorem
can be applied to show the doubling index of Q would be too big (the Hyperplane lemma [Log18a,
§4]). As a consequence one shows that at most 0.9A d −1 of the subcubes are bad.
We have now all the required ingredients to complete the overview of the polynomial upper
bound in Theorem 4.3.2 provided the metric on M is flat. While the proof in the general case is
more technical, the argument in the flat case highlights essentially all the conceptual ideas.

Proof of the upper bound in Theorem 4.3.2 for flat metrics

We start by applying the lifting trick to an eigenfunction. Let Q 1 ⊂ Rd be a unit cube. In
view of Theorem 4.3.6, it is sufficient to prove that for any harmonic function h : ℓQ 1 → R
for which βsup (h,Q 1 ) ≤ β, we have

H d −1 (Z h ∩Q 1 ) ≤ C β2S

for some S > 0 independently of h (although Theorem 4.3.6 refers to the doubling indices
on balls, the doubling indices on cubes are essentially equivalent, as mentioned earlier). Let

F (β) := sup H d −1 (Z h ∩Q 1 ),
h∈Hβ

where
Hβ := h : ℓQ 1 → R, h is harmonic with βsup (h,Q 1 ) < β .
© ª

It follows from (4.3.3) that F (β) is finite. Note that for any cube q and a harmonic function
h : ℓq → R one has

H d −1 (Z h ∩Q 1 ) ≤ F βsup (h, q) s(q)d −1 .

¡ ¢

Let ho : ℓQ 1 → R be harmonic, where βsup (ho ,Q 1 ) ≤ β optimises F (β) up to a small

positive ε, i.e., it satisfies

H d −1 (Z ho ∩Q 1 ) ≤ F (β) < H d −1 (Z ho ∩Q 1 ) + ε. (4.3.13)

To give a bound on the size of its nodal set, subdivide Q 1 into A d equal subcubes q where
A is large as in Theorem 4.3.24. Collecting the contributions to the nodal set from all
146 Chapter 4. Nodal geometry of eigenfunctions

subcubes q , it is clear that

1 1
H d −1 (Z ho ∩Q 1 ) ≤ A d F (β/2) + 0.9A d −1 F (β) , (4.3.14)
A d −1 A d −1
and we derive from (4.3.13) and (4.3.14) that

F (β) < 10AF (β/2) + 10ε.

Since ε > 0 is arbitrary, we get

F (β) ≤ 10AF (β/2).

This inequality implies (see Exercise 4.3.25) that F is bounded by a polynomial in β of

degree log2 (10A) =: 2S , which completes the proof of the theorem.

Exercise 4.3.25
Let f : [1, ∞) → R be a non-negative monotonically non-decreasing function. Suppose
that f (2x) < A f (x) for all x ≥ 1 and some A > 1. Prove that f (x) < C x S for all x > 1,
where C , S are positive constants depending on A only.

§4.4. Nodal sets on surfaces and eigenvalue multiplicity bounds

§4.4.1. Local structure of the nodal set

Let u be a smooth function in a neighbourhood of the origin in Rd , and suppose that it has
vanishing order N ∈ N at x = 0. Then, by Taylor’s Theorem,

u(x) = P N (x) + o |x|N

¡ ¢
as x → 0, (4.4.1)

where P N is a non-zero homogeneous polynomial of degree N . If u is a solution of a linear partial

differential equation, we have the following simple result.

Theorem 4.4.1
Let A be a linear differential operator with C ∞ smooth coefficients in a neighbourhood
0 ∋ W ⊂ Rd , and let A0 be its principal part with the coefficients fixed at x = 0. Suppose
that u ∈ C ∞ (W ) is a solution of the equation A u = 0 that has vanishing order N ∈ N at
x = 0. Then
A0 P N = 0, (4.4.2)
where P N is defined in (4.4.1).
§4.4. Nodal sets on surfaces and eigenvalue multiplicity bounds 147

Proof
We follow the argument in [Alb71, Theorem 2.12]. Let m be the degree of A , and let us
represent u(x) in the form u(x) = P N (x)+R(x). Write A = A0 +A1 +A2 where A0 +A1
is the principal part of A and A2 is of a smaller degree. Then

0 = A u = A0 P N + (A1 + A2 )P N + A R.

Since A0 P N is the Taylor polynomial of¡A u of degree N −m , we may conclude by Taylor’s

N −m
theorem that (A1 + A2 )P N + A R = o |x| . It follows that 0 = A0 P N + o |x|N −m .
¢ ¡ ¢

This is possible only if A0 P N = 0.

Remark 4.4.2: Bers’s theorem

Lipman Bers
Theorem 4.4.1 can be viewed as an elementary version of the celebrated Bers’s theorem (1914–1993)
[Ber55], which guarantees that any solution u of an elliptic equation with Hölder coeffi-
cients has a polynomial asymptotics (4.4.1) near its zero set, as if u were a smooth function.
In addition (4.4.2) is also satisfied.

We can now prove the following result which in a way is a two-dimensional version of Theo-
rem 4.1.16.

Theorem 4.4.3: [Che75]

Let M be a compact Riemannian surface. The nodal set of a Laplace eigenfunction on
M consists of C 1 immersed circles. The nodal critical points of an eigenfunction (i.e. the
zeros of its gradient lying on the nodal set) are isolated, and at each such point the nodal
lines divide the angle 2π equally.

Proof
Let us apply Theorem 4.4.1 to a Laplace eigenfunction u(x) with eigenvalue λ on a Rie-
mannian manifold (with A = ∆g +λ). Note that by elliptic regularity (see Theorem 2.2.17)
u(x) is smooth and by Theorem 4.3.7 it has a finite vanishing order N . Choose coordinates
at a neighbourhood of a point x 0 in which the Riemannian metric g i j (x 0 ) = δi j . The
principal part of ∆g at the point x 0 is the Euclidean Laplacian. Then, P N is a harmonic
homogeneous polynomial of degree N . Note that in two dimensions, homogeneous har-
monic polynomials of degree N have a particularly simple form Re Az N , where z = x 1 +ix 2
and A ∈ C. The nodal set of such a harmonic polynomial is a union of straight lines going
through the origin and dividing the unit disk into 2N congruent sectors.
It was shown in [Che75] (see also [BérMey82, Appendix E] and the discussion before
Theorem 4.1.16) that in two dimensions there exists a C 1 -diffeomorphism f near x such
that u(x) = P N ( f (x)). Hence, the nodal set Z u is locally diffeomorphic to the nodal set of
148 Chapter 4. Nodal geometry of eigenfunctions

a harmonic polynomial, and the first statement follows. Note also that f maps nodal criti-
cal points of u to nodal critical points of P N , which are isolated, and the second statement
follows.
It remains to prove the equiangular property of nodal lines. Take a path
(r (t ) cos ϕ(t ), r (t ) sin ϕ(t )) lying in the nodal set Z u and starting at a critical point r (0) =
0. We can write

0 = u(r (t ) cos ϕ(t ), r (t ) sin ϕ(t ))

= A 1 r (t )N cos N ϕ(t ) + A 2 r (t )N sin N ϕ(t ) + R r (t ) cos ϕ(t ), r (t ) sin ϕ(t )
¡ ¢

= Ar (t )N sin N ϕ(t ) + α + R r (t ) cos ϕ(t ), r (t ) sin ϕ(t ) ,

¡ ¢ ¡ ¢

where A 1 , A 2 , A , and α are some constants, and R(r (t ) cos ϕ, r (t ) sin ϕ) = o r (t )N as

¡ ¢

t → 0. It follows that
lim sin(N ϕ(t ) + α) = 0,
t →0

from which one concludes that ϕ(0) can take only values of the form (kπ − α)/N for
some k ∈ Z. Recall that the nodal set of the harmonic polynomial P N consists of 2N rays
emanating from the critical point. Since f is a C 1 –diffeomorphism, the images of different
rays under f can not yield the same value of ϕ(0) mod 2π, and the equiangular property
follows.

§4.4.2. Multiplicity bounds

We have the following

Lemma 4.4.4: [Nad87], [KarKokPol14]

Let M be a Riemannian surface, and let u 1 , . . . , u 2n be a collection of linearly independent
eigenfunctions corresponding to some eigenvalue λ. Then, for a given point x ∈ M , there
exists a non-trivial linear combination 2n
i =1 αi u i with vanishing order at x of at least n .
P

Proof
Let V = Span{u 1 , . . . , u 2n }, and let Vi be the subspace of elements u ∈ V such that
ordx (u) ≥ i . Clearly, Vi +1 ⊂ Vi . We need to show that Vn ̸= {0}. Suppose the contrary.
Let us calculate dimV . We have
n−1
X
dimV = dim(V j /V j +1 ).
j =0

As follows from the proof of Theorem 4.4.3, V j /V j +1 can be identified with a subspace of
the space of harmonic homogeneous polynomials of degree j . In turn, the latter space is of
dimension one for j = 0 and of dimension two for j ≥ 1. Therefore, dimV ≤ 1+2(n−1) <
2n , which is a contradiction.
§4.4. Nodal sets on surfaces and eigenvalue multiplicity bounds 149

It is useful to think about the nodal set of an eigenfunction on a Riemannian surface as a graph
with edges being the arcs of the nodal lines and the vertices being the critical points. If there is a
closed nodal line without critical points on it, we may introduce an artificial vertex with the edge
being a cycle. The graph constructed this way is called the nodal graph of an eigenfunction.
Let us recall some general facts about graphs on surfaces. Given a graph Γ, let the degree of a
vertex x , denoted degΓ x , be the number of edges incident to x ; if there is an edge that starts and
ends at x , it is counted twice. Let e be the number of edges in the graph. Then
X
2e = degΓ x.
x

Let f be the number of faces of Γ, i.e., the number of connected components of M \ Γ. Euler’s
inequality states that
v − e + f ≥ χ(M ),

where χ(M ) is the Euler characteristic of M . It becomes an equality (the well-known Euler’s
formula) if all the faces are topological disks. The following theorem is due to N. Nadirashvili
[Nad87]); weaker versions were earlier obtained by G. Besson [Bes80] and S.-Y. Cheng [Che75].

Theorem 4.4.5
The multiplicity m(λk ) of the eigenvalue λk on a Riemannian surface M satisfies the
inequality
m(λk ) ≤ 2k − 2χ(M ) + 5. (4.4.3)

Proof
Suppose the contrary. Then there exist 2k − 2χ(M ) + 6 linearly independent λk -
eigenfunctions. By Lemma 4.4.4, there exists an eigenfunction with the vanishing order
k − χ(M ) + 3 at some point x 0 . Consider the nodal graph of this eigenfunction. The
number of faces of this graph is the number of nodal domains. Therefore, by Courant’s
theorem, since we number our eigenvalues as 0 = λ0 < λ1 ≤ λ2 ≤ . . . , we have

k + 1 ≥ f ≥ χ(M ) + e − v.

At the same time, e = 21 x degΓ (x), and the degree of each vertex is at least two. Hence,
P

in order to obtain a lower bound on the right-hand side we can assume that x 0 is the only
vertex. Since degΓ (x 0 ) = 2(k − χ(M ) + 3), we get

k + 1 ≥ χ(M ) + k − χ(M ) + 3 − 1 = k + 2,

which is a contradiction.

In some cases, further refinements of the bound (4.4.3) can be obtained using a careful anal-
ysis of the structure of the nodal graph. Multiplicity estimates could be also proved in a similar
150 Chapter 4. Nodal geometry of eigenfunctions

way for the Dirichlet and Neumann eigenvalues on surfaces with boundary, see [KarKokPol14,
§6] for details.
The estimate (4.4.3) in general is not sharp.

Exercise 4.4.6
Deduce from Weyl’s law (see Theorem 3.3.4) that the multiplicity m(λ) on a d -
dimensional manifold satisfies

m(λk ) = o(k) as k → ∞. (4.4.4)

It follows from (4.4.4) that the estimate (4.4.3) is not of the correct order in k asymptotically.
Yet, in a few cases it yields sharp multiplicity bounds.

Corollary 4.4.7
On the sphere, m λ1 (S¡2 ) ≤ 3, which
¡ ¢
is sharp and is attained by the round metric. On
the projective plane, m λ1 (RP2 ) ≤ 5, which is again sharp and is attained by the round
¢

metric.

We leave the proof of Corollary 4.4.7 as an exercise for the reader.

 CHAPTER 5
Eigenvalue inequalities
In this chapter, we prove various geometric eigenvalue inequalities,
in particular, due to Faber–Krahn, Cheeger and Szegő–Weinberger.
We also present the results of Hersch and Yang–Yau, as well as other
isoperimetric inequalities for Laplace–Beltrami eigenvalues on
surfaces. Furthermore, we discuss universal inequalities for Dirichlet
eigenvalues on Euclidean domains, and related commutator
identities.

§5.1. The Faber–Krahn inequality

Throughout the chapter, we will use

Definition 5.1.1
Let Ω ⊂ Rd be a measurable set of finite volume. Its symmetric rearrangement is an open
ball Ω∗ = B Rd ∗ , where the radius R ∗ = R Ω
∗
is determined by the condition Vold (Ω∗ ) =
Ω
Vold (Ω). Therefore
∗
¢ d1
= Vold (Ω)ω−1
¡
RΩ d ,
where ωd is the volume of the unit ball Bd , see (B.1.1).

We will also use Notation 3.3.23 for level and superlevel sets of a function and the volume of
a superlevel set.

151
152 Chapter 5. Eigenvalue inequalities

§5.1.1. Motivation
The Faber–Krahn inequality states that among all Euclidean domains of given volume, the first
Dirichlet eigenvalue is minimal for the ball.

Theorem 5.1.2: Faber–Krahn inequality

Let Ω ⊂ Rd be a bounded domain. Then

λD D ∗
1 (Ω) ≥ λ1 (Ω ). (5.1.1)

Inequality (5.1.1) was conjectured in 1877 by Lord Rayleigh in his famous book on the theory
of sound [Ray77]. Moreover, he proved, using perturbation theory, that a ball is a local minimiser
for λ1 = λD1 among all domains of a given volume. A complete proof of (5.1.1) was obtained
independently by G. Faber and E. Krahn [Fab23], [Kra25].

Remark 5.1.3
In view of Definition 5.1.1 and Exercise 1.2.21, Theorem 5.1.2 can be reformulated as follows:
for any bounded domain Ω ⊂ Rd ,

λ1 (Ω) Vol(Ω)2/d ≥ ω2/d

d jd
2
,
2
−1,1
Georg Faber
d
(1877–1966) where j d −1,1 is the first zero of the Bessel function of the first kind of order 2 − 1. In
2
particular, for d = 2,
2
λ1 (Ω) Area(Ω) ≥ π j 0,1 ≈ 5.76π.
Note that this estimate confirms Pólya’s Conjecture 3.3.14 for the first Dirichlet eigenvalue
of a planar domain, which in this case reads λ1 (Ω) Area(Ω) ≥ 4π.

In order to get some physical intuition, it is instructive to look at the Faber–Krahn inequality
from the viewpoint of the heat equation on a bounded domain Ω ∈ Rd :

∂u(t ,x)

 ∂t = ∆u(t , x) for (t , x) ∈ (0, ∞) × Ω,


u=0 on ∂Ω,


u(0, x) = u (x).
0
Edgar Krahn
(1894–1961) Here u(t , x) is the temperature at the point x ∈ Ω at the time t > 0, and u 0 (x) is the initial
temperature distribution. Using the Fourier method, we obtain
∞
c k e−λk t u k (x),
X
u(t , x) =
k=1

where λk and u k are the Dirichlet eigenvalues and eigenfunctions, respectively, and the coeffi-
§5.1. The Faber–Krahn inequality 153

cients c k are determined by the initial condition u 0 . Consider the heat content of Ω,
ˆ
Q Ω (t ) := u(t , x) dx, (5.1.2)
Ω

and the rate of the relative heat loss,

′
QΩ (t )
αΩ (t ) := − .
Q Ω (t )

Clearly,
lim αΩ (t ) = λ1 (Ω).
t →∞

In other words, the smaller is λ1 , the smaller is the long-term heat loss. At the same time, it is
natural to assume that in order to minimise the heat loss due to the fact that the boundary is
kept at the zero temperature, one needs to minimise the boundary surface of Ω. This leads to the
isoperimetric problem: given the fixed interior volume, minimise the (d −1)-dimensional volume
of the boundary. It is well known that the solution of this problem is a ball, which is in agreement
with the Faber–Krahn inequality.
Interestingly enough, while the argument above is in no way rigorous, the isoperimetric in-
equality indeed plays the key role in the proof of the Faber–Krahn inequality. We present the
details below.

§5.1.2. The co-area formula

One of the technical tools used in the proof of the Faber–Krahn inequality is an important result
from geometric measure theory called the co-area formula (see, for instance, [Maz85, §1.2.4]).

Theorem 5.1.4: The co-area formula

Let Ω ⊂ Rd be a domain, let h : Ω → R be an integrable function, and let F : Ω → [a, b] ⊂
R be a smooth function. Then

ˆ ˆb ˆ
h(x)|∇F (x)| dx = h(x) dΣt dt , (5.1.3)
Ω a L F (t )

where LF (t ) are the level sets of F , see Notation 3.3.23, and dΣt is the surface measure on
L F (t ).

Note that since F is smooth, the set of critical values have measure zero by Sard’s theorem.
Therefore, by implicit function theorem, the level sets LF (t ) are smooth hypersurfaces for almost
all t . One can also check that the interior integral on the right is an integrable function of t , and
hence the iterated integral is well defined.
154 Chapter 5. Eigenvalue inequalities

Remark 5.1.5
The smoothness assumption on F in Theorem 5.1.4 can be relaxed. In particular, the co-
area formula holds if F is Lipschitz or if it is a function of bounded variation, see [Eva-
Gar15].

The co-area formula can be viewed as a kind of a “curvilinear Fubini theorem”, as the follow-
ing examples shows.

Example 5.1.6
x
Let Ω = B R ⊂ Rd and F (x) = |x|. Then ∇F (x) = |x| and |∇F (x)| = 1 for all x . In view of
Remark 5.1.5 one can apply the co-area formula. It follows that LF (r ) = S r := {x ∈ Rd :
|x| = r }, and thus
ˆ ˆR ˆ
h(x) dx = h(x) dS r dr,
BR 0 Sr

which is the usual integration formula in spherical coordinates.

Suppose that the set of critical points

C F := {x ∈ Ω : ∇F = 0}

of a function F has measure zero. Substituting formally h(x) = |∇F1(x)| into (5.1.3) we obtain

ˆb ˆ
1
Vol(Ω) = dΣt dt . (5.1.4)
|∇F (x)|
a L F (t )

To justify this result (see, for instance, [Dan11]), take ε > 0 and set

1
h ε (x) = .
|∇F (x)| + ε

Applying (5.1.3) to hε we get

ˆ ˆ ˆ
h ε (x)|∇F (x)| dx = h ε (x) dΣt dt . (5.1.5)
Ω\C F [a,b]\F (C F ) L F (t )

Using the monotone convergence theorem as ε → 0 and taking into account that C F has measure
zero, we obtain (5.1.4).
§5.1. The Faber–Krahn inequality 155

Remark 5.1.7
As was pointed out in [CadFar18], the assumption that the set of critical points of F has
measure zero has been often neglected in the literature, though it is necessary for the va-
lidity of (5.1.4).

§5.1.3. Symmetric decreasing rearrangement

The proof of the Faber–Krahn inequality also uses the notion of symmetric decreasing rearrange-
ment of a function. There are several equivalent ways to define it; we essentially follow the ap-
proach of [LieLos97]. First, given a set A ⊂ Rd of finite volume, we define the symmetric rear-
rangement of its characteristic function by χ∗A := χ A ∗ .

Definition 5.1.8: Symmetric decreasing rearrangement

Let u : Ω → R be a measurable non-negative function on an open bounded set Ω ⊂ Rd .
The symmetric decreasing rearrangement of u is a function u ∗ : Ω∗ → R defined by the
relation
ˆ+∞
∗
u (x) = χ∗Vu (t ) (x) dt , (5.1.6)
0

where Vu (t ) are the superlevel sets of u , see Notation 3.3.23.

Exercise 5.1.9
Show that u ∗ (x) is a lower semi-continuous radially symmetric function which is non-
increasing in |x|.

Recall the “layer cake representation” formula (see [LieLos97, Theorem 1.13]):
ˆ+∞
u(x) = χVu (t ) (x) dt . (5.1.7)
0

Comparing the two formulas above, we observe that u ∗ (x) is obtained from u(x) by symmetri-
sation of its superlevel sets. It then easily follows that the functions u and u ∗ are equimeasurable,
i.e. Vu (t ) = Vu ∗ (t ) for any t ∈ R, see Notation 3.3.23.

Exercise 5.1.10
Symmetric decreasing rearrangement of a function u is sometimes alternatively defined as

u ∗ (x) := sup t : x ∈ (Vu (t ))∗ .

© ª

Show that the two definitions are equivalent.

156 Chapter 5. Eigenvalue inequalities

Integrating both sides of the layer cake representation (5.1.7) over Ω, applying Fubini theorem
and making a change of variables t = s p yields
ˆ ˆ+∞
p
u(x) dx = p s p−1Vu (s) ds. (5.1.8)
Ω 0

for any p ≥ 1. Since u and u ∗ are equimeasurable, (5.1.8) implies

∥u∥L p (Ω) = ∥u ∗ ∥L p (Ω∗ ) . (5.1.9)

§5.1.4. Proof of the Faber–Krahn inequality

We follow the argument that essentially goes back to E. Krahn [Kra25], see also [Dan11].
We will first prove

Proposition 5.1.11: The Pólya–Szegő principle

Let Ω ⊂ Rd be a bounded domain, and let u be the first Dirichlet eigenfunction on Ω.
Then
∥∇u∥L 2 (Ω) ≥ ∥∇u ∗ ∥L 2 (Ω∗ ) . (5.1.10)

Proof of Proposition 5.1.11

Without loss of generality, we can assume u > 0 in Ω. Let Ω∗ = B R ∗ be the symmetric
rearrangement of Ω, and let v(x) := u ∗ (x) be the symmetric decreasing rearrangement of
u.
Since u is real analytic by Theorem 2.2.1(ii), the set of its critical points C u has measure
zero, and therefore by (5.1.4) we have

ˆ maxˆ
x∈Ω u(x) ˆ
1
Vu (t ) = dx = dΣs ds. (5.1.11)
|∇u|
Vu (t ) t L u (s)

Since Vu (t ) is a non-increasing function of t , it is differentiable almost everywhere. In view

of (5.1.11), its derivative is given by
ˆ
1
Vu′ (t ) = − dΣt .
|∇u|
L u (t )

Note that the integral on the right is well-defined for almost all t since Vu (t ) ≤ Vol(Ω) < ∞;
this also follows from Sard’s theorem, implying that |∇u| > 0 on the level set Lu (t ) for
almost all t .
We would like to obtain an analogue of (5.1.11) for v . However, we cannot apply (5.1.4)
to the function v directly, since a priori the set C v of the critical points of v may have a
§5.1. The Faber–Krahn inequality 157

positive measure. Since v is radially decreasing and hence of bounded variation, one can
apply the co-area formula. Arguing as in (5.1.5), we obtain

ˆ ˆ ∗ v(x)
maxx∈Ω ˆ
1
Vv (t ) = ρ v (t ) + dx = ρ v (t ) + dΣs ds, (5.1.12)
|∇v|
Vv (t )\C v t L v (s)

where
ρ v (t ) := Vold (Vv (t ) ∩ C v ).
By [CiaFus02, Lemma 2.4]) it follows that ρ ′v (t ) = 0 for almost all t . Differentiating both
sides of (5.1.12) with respect to t we get
ˆ
1
Vv′ (t ) = − dΣt (5.1.13)
|∇v|
L v (t )

for almost all t , as in (5.1.11).

Since Vu (t ) = Vv (t ) for all t , their derivatives must coincide provided they are well
defined. Hence, Vu′ (t ) = Vv′ (t ) for almost all t , which implies
ˆ ˆ
1 1
dΣt = dΣt . (5.1.14)
|∇u| |∇v|
L u (t ) L v (t )

Let us show that ˆ ˆ

|∇u| dΣt ≥ |∇v| dΣt (5.1.15)
L u (t ) L v (t )

for almost all t . Indeed, by the Cauchy–Schwarz inequality,

2
ˆ ˆ ˆ
   
1
dΣt  |∇u| dΣt  ≥  dΣt  = (Vold −1 (L u (t )))2 .
    
 (5.1.16)
|∇u|
L u (t ) L u (t ) L u (t )

However, by the isoperimetric inequality, Vold −1 (Lu (t )) ≥ Vold −1 (L v (t )), since, by the
definition of the symmetric decreasing rearrangement, the sets Lu (t ) and L v (t ) bound
the same volume, and L v (t ) is a sphere because v is a radial function. Furthermore, for
the same reason, |∇v| is constant on the spheres L v (t ), which leads to the case of equality
in the Cauchy–Schwarz inequality analogous to (5.1.16),

ˆ ˆ
  
1
(Vold −1 (L v (t )))2 =  dΣt  |∇v| dΣt .
  
(5.1.17)
|∇v|
L v (t ) L v (t )
158 Chapter 5. Eigenvalue inequalities

Hence, (5.1.15) follows from (5.1.14) combined with (5.1.16) and (5.1.17).
Applying the co-area formula once again and taking into account that maxx∈Ω u(x) =
maxx∈Ω∗ v(x) we get

ˆ maxˆ
x∈Ω u(x) ˆ
2
|∇u| dx = |∇u| dΣt dt
Ω 0 L u (t )
maxˆ
x∈Ω u(x) ˆ ˆ
≥ |∇v| dΣt dt = |∇v|2 dx,
0 L v (t ) Ω∗

where the inequality follows fom (5.1.15). This completes the proof of the Pólya–Szegő
principle (5.1.10).

Remark 5.1.12
The justification of (5.1.13) in the proof of the Pólya–Szegő principle follows the approach
of [Fus08, formula (3.14)]. It is omitted in most available proofs of the Faber–Krahn equal-
ity, cf. Remark 5.1.7.

Remark 5.1.13
Given a non-negative measurable function u : Rd → R of compact support one can define
its symmetric rearrangement u ∗ : Rd → R by formula (5.1.6). A more general version of
the Pólya–Szegő principle holds: for any p ≥ 1 and any non-negative u ∈ W 1,p (Rd ) of
compact support, one has
ˆ ˆ
¯ ∗ ¯p
|∇u|p dx ≥ ¯∇u ¯ dx.

Rd Rd

We refer to [Fus08, Theorem 3.1] and [Kaw85, Remark 2.16] for details.

We are now in the position to prove Theorem 5.1.2. The inequality (5.1.10) together with the
equality (5.1.9) for L 2 norms yields the inequality

R Ω [u] ≥ R Ω∗ [v]

for the Rayleigh quotients of u and v . Note that for x ∈ ∂Ω∗ = S R ∗ , we have

ˆ+∞
v(x) = χ∗Vu (t ) (x) dt = 0.
0
§5.1. The Faber–Krahn inequality 159

since Vu (t )∗ (x) ⊂ B R ∗ for any t > 0. It remains to show that v ∈ H01 (Ω∗ ). Let us extend u ∈
H01 (Ω) by zero to the whole Rd and apply the Pólya–Szegő principle to this extension (cf. Remark
5.1.13). The resulting function is the extension of v by zero and it lies in H 1 (Rd ). Given that v is
radially decreasing, it follows that it is continuous up to the boundary ∂Ω∗ where it vanishes, and
hence it belongs to H01 (Ω∗ ). Therefore, one can use v as a test function for the first eigenvalue of
the Dirichlet problem on the ball Ω∗ . Hence,

λ1 (Ω) = R Ω [u] ≥ R Ω∗ [v] ≥ λ1 (Ω∗ ),

which proves the Faber–Krahn inequality.

Remark 5.1.14: Equality in the Faber–Krahn inequality

Let us inspect the proof of Theorem 5.1.2 in order to characterise the case of equality in the
Faber–Krahn inequality. Note that the case of equality in the geometric isoperimetric in-
equality implies that the domain is a ball up to a set of measure zero (see [Fus04, Theorem
4.11]). Therefore, it follows from (5.1.16) that if the equality in the Faber–Krahn inequality
is attained, Vu (t ) are balls up to sets of measure zero for almost all t . At the same time,
since Ω = Vu (t ), it follows that Ω is a ball up to a set of measure zero. In particular, if
S
t >0
Ω is sufficiently regular (for example, Lipschitz), it has to be a ball.
In fact, with some extra work one can prove an even more precise characterisation: the
equality in the Faber–Krahn inequality implies that Ω is a ball up to a set of zero capacity,
cf. Remark 3.2.15. Indeed, suppose Ω is an open set achieving the equality. As was shown
above, it is equal to a ball B up to a set of zero measure. Therefore, Ω ⊂ B , and the result
follows from the following characterisation of domain monotonicity for Dirichlet eigen-
values proved in [AreMon95, Theorem 3.1]: λ1 (Ω) = λ1 (B ) if and only if the capacity of
B \ Ω is equal to zero. We refer to [Dan11, Remark 5.1] for more details.

We want to address the stability of the Faber–Krahn inequality. Namely, suppose that λ1 (Ω)
is close to λ1 (Ω∗ ) for an open set Ω. Does it imply that Ω is in some sense close (up to rigid
motions) to the ball Ω∗ ? The answer to this question is positive. In order to state it properly, we
need the following

Definition 5.1.15: The Fraenkel asymmetry

The Fraenkel asymmetry of a set Ω is defined by
³ ´
Vold Ω∆B y,R ∗
A (Ω) = inf , (5.1.18)
y∈Rd Vold (Ω)
Ludwig Edward
where R ∗ is the radius of the symmetric rearrangement Ω∗ of Ω, and Ω∆B := (Ω \ B ) ∪ Fraenkel
(B \ Ω) denotes the symmetric difference of sets Ω and B . (1927–2019)
 160 Chapter 5. Eigenvalue inequalities

The following result had been conjectured independently by N. Nadirashvili [Nad97, p. 200]
and by T. Bhattacharya and A. Weitsman [BhaWei99, §8] in the late 1990s, and was recently proved
in [BraDePVel15].

Theorem 5.1.16
There exists ad > 0 such that for any bounded domain Ω ⊂ Rd ,

a d A (Ω)2 ≤ Vol(Ω)2/d λ1 (Ω) − λ1 (Ω∗ ) .

¡ ¢
(5.1.19)

Moreover, one can check that the power two in the left-hand side of (5.1.19) is the smallest
possible.

Remark 5.1.17: Torsional rigidity

Given a bounded domain Ω, the quantity
¡´ ¢2
Ω u dx
T (Ω) := sup ´ 2
u∈H01 (Ω)\{0} Ω |∇u| dx

is called the torsional rigidity of Ω (see [PólSze51]). Its physical meaning is as follows: T (Ω)
measures the amount of resistance of a beam with a cross-section Ω against torsional de-
formation. The celebrated inequality of A. de Saint-Venant, proved by G. Pólya, states
that the ball has the maximal torsional rigidity among all domains of given volume, that is

Adhémar Jean T (Ω) ≤ T (Ω∗ ).

Claude Barré
de Saint-Venant
(1797–1886)

Exercise 5.1.18
Prove the Saint-Venant inequality using an adaptation of the proof of the Faber–Krahn
inequality.

Apart from the symmetric rearrangement, there exist other symmetrisation techniques which
are used to prove eigenvalue inequalities. Probably the most important one is the Steiner sym-
metrisation of a set, which is a symmetrisation with respect to a hyperplane, see [PólSze51, Chap-
ter 1]. The corresponding Steiner rearrangement of a function shares the essential features with
the symmetric decreasing rearrangement: in particular, it preserves the L 2 norm of a function and
does not increase the Dirichlet energy. Therefore, the Steiner rearrangement does not increase the
fundamental tone. Motivated by this approach, in 1951 G. Pólya and G. Szegő made the following
well-known conjecture (see [PólSze51, page 159]), which they proved for n = 3 and n = 4.
§5.1. The Faber–Krahn inequality 161

Conjecture 5.1.19: Pólya–Szegő conjecture

Among all polygons with n sides and a given area, λ1 is minimised by a regular n -gon.

For n = 3, one can show that given any triangle, there exists a sequence of Steiner symmetri-
sations under which it converges to an equilateral triangle. For n = 4 the argument is even easier:
any quadrilateral can be transformed into a rectangle using a sequence of not more than three
symmetrisations. We refer to [Hen06, §3.3.2] for details of the proof in these two cases. However,
this method no longer works for a higher number of vertices n of a polygon: indeed, it is easy to
check that in this case a Steiner symmetrisation may increase the number of sides of an n -gon.
Therefore new ideas will be required to prove Conjecture 5.1.19 for n ≥ 5. See also §A.1.2 for a
discussion about the asymptotics of the first eigenvalue of the regular n -gon as n → ∞.

§5.1.5. Applications of the Faber–Krahn inequality

Faber–Krahn inequality combined with the Courant nodal domain theorem implies a sharp isoperi-
metric inequality for the second Dirichlet eigenvalue. It was proved by E. Krahn in [Kra26], and
later rediscovered independently by P. Szego (the son of G. Szegő) and I. Hong, see [Hen06, §4.1]

Theorem 5.1.20: Krahn–Szego inequality

Among all (possibly, disconnected) Euclidean domains of a given volume, the second
Dirichlet eigenvalue is minimised by the disjoint union of two identical balls.

Proof
Let Ω be a connected bounded domain of given volume, which by rescaling we may assume
to be equal to one. By Corollary 4.1.34 of the Courant nodal domain theorem, the second
Dirichlet eigenfunction has precisely two nodal domains Ω1 and Ω2 . Let Ω∗1 and Ω∗2 be
the symmetric decreasing rearrangements of Ω1 and Ω2 , respectively. Applying the Faber–
Krahn inequality one obtains
λ2 (Ω) = λ1 (Ω1 ) = λ1 (Ω2 ) ≥ max λ1 (Ω∗1 ), λ1 (Ω∗2 ) .
© ª

At the same time, since Vol(Ω∗1 ) + Vol(Ω∗2 ) = Vol(Ω), one has

max λ1 (Ω∗1 ), λ1 (Ω∗2 ) ≥ λ1 (B R ) = λ2 B R ⊔ B R′ ,
¡ ¢ ¡ ¢

where B R , B R′ are identical balls such that Vol(B R ) = Vol(B R′ ) = 21 Vol(Ω). Here the first
step follows from rescaling and the last step uses the fact that the spectrum of a disjoint
union of domains is a union of their spectra. This completes the proof of the Krahn–Szego
inequality for connected domains.
If Ω = Ω′1 ⊔Ω′2 is not connected, we can modify the above argument as follows. In this
case
λ2 (Ω) = max{λ1 (Ω1 ), λ1 (Ω2 )}
162 Chapter 5. Eigenvalue inequalities

for some disjoint sets Ω1 ⊔Ω2 ⊂ Ω, which are either connected components of Ω, or nodal
domains of the second eigenfunction of a connected component. Applying the Faber–
Krahn inequality and rescaling if necessary we again arrive at the conclusion that the min-
imum of λ2 is attained by a domain which is a disjoint union of two identical balls of
volume 21 Vol(Ω). This completes the proof of the theorem.

Remark 5.1.21
In view of Remark 5.1.14, it follows from the proof of Theorem 5.1.20 that the minimum
of the second Dirichlet eigenvalue is attained if and only if the domain is equal to a disjoint
union of two identical balls up to a set of zero capacity. In particular, the minimum of λ2
is not attained in the class of connected domains.

Let us now discuss an application of the Faber–Krahn inequality to the nodal geometry. The
following result is due to Å. Pleijel [Ple56] and could be viewed as an asymptotic refinement of
Courant’s nodal domain theorem.

Theorem 5.1.22: Pleijel’s nodal domain theorem

Let Ω ⊂ Rd be a bounded domain and let u k be an orthogonal basis of Dirichlet eigen-
functions corresponding to eigenvalues λD
k
. Let η k be the number of nodal domains of
u k . Then
ηk
lim sup < 1. (5.1.20)
k→∞ k

Proof
For simplicity, assume d = 2; the proof in higher dimensions is analogous (see [BérMey82,
Lemme 9] for the last step). Let Ω ⊂ R2 , and let Ωl ⊂ Ω, l = 1, . . . , η k , be the nodal domains
of an eigenfunction u k . Then, λDk
(Ω) = λD1 (Ωl ) for all l . At the same time, by the Faber–
Krahn inequality,
Area(Ωl ) 1 1
≥ = . (5.1.21)
π j 0,1
2
λD
1 (Ωl ) λD
k
(Ω)
Summing up the inequalities (5.1.21) over l = 1, . . . , η k , we get

Area(Ω) ηk
≥ .
π j 0,1
2
λk (Ω)
D

By Weyl’s law,
λD
k
(Ω) 4π
lim = ,
k→∞ k Area(Ω)
§5.1. The Faber–Krahn inequality 163

and hence
ηk 4
lim sup ≤ 2 ≃ 0.691 < 1. (5.1.22)
k→∞ k j 0,1

Remark 5.1.23: Courant-sharp eigenvalues and nodal deficiency

Pleijel’s theorem implies that in dimension d ≥ 2, only finitely many eigenvalues λD k
admit
eigenfunctions satisfying η k = k . We recall that such eigenvalues are called Courant-sharp,
see Remark 4.1.35. The non-negative quantity k − η k is called the nodal deficiency of an
eigenfunction and it admits interesting interpretations in terms of the Morse indices of
certain functionals and operators, see [BerKucSmi12, CoxJonMar17, BerCHS22] and ref-
erences therein.

Remark 5.1.24: Pleijel’s nodal domain theorem in other settings

Inequality 5.1.20 also holds in the case of Neumann boundary conditions. The main diffi-
culty in the Neumann case is to handle the nodal domains which touch the boundary. On
those nodal domains the corresponding Neumann eigenvalue is equal to the first eigen-
value of a mixed Dirichlet–Neumann problem, and therefore the Faber–Krahn inequality
cannot be applied. Pleijel’s theorem for Neumann boundary conditions was first estab-
lished in [Pol09] for piecewise analytic planar domains. The result was later extended in
[Lén19] to arbitrary dimensions and more general Robin boundary conditions for do-
mains with C 1,1 boundary. Analogues of Pleijel’s theorem exist in other settings as well, in
particular, for compact Riemannian manifolds [BérMey82], and for certain Schrödinger
operators in Rd [Cha18, ChaHelHoO18].

Remark 5.1.25: Optimal Pleijel’s constant

One may wonder whether Pleijel’s constant 0.691 in the right-hand side of (5.1.22) is close
to being optimal for planar domains (a similar question could be also asked in arbitrary
dimension). By taking separable eigenfunctions on rectangles, it is easy to check that Plei-
jel’s constant is not smaller than π2 ≃ 0.636. It was conjectured in [Pol09] that this value
is optimal for planar domains with either Dirichlet or Neumann boundary conditions.
Slight improvements of the constant in (5.1.22) were obtained in [Bou15, Ste14] by using
quantitative stability results for the Faber–Krahn inequality and estimates on the packing
density by disks.
164 Chapter 5. Eigenvalue inequalities

§5.2. Cheeger’s inequality and its applications

§5.2.1. Cheeger’s inequality

By the variational principle, in order to estimate the first eigenvalue from above it is sufficient to
find an appropriate test function. Estimating eigenvalues from below is, a priori, a more difficult
task. The importance of Cheeger’s inequality [Che71] is that it provides a rather simple geometric
lower bound for the first eigenvalue. In order to state the result we need to introduce the following
definition.

Definition 5.2.1: The Dirichlet Cheeger constant

Let Ω be a compact Riemannian manifold with boundary or a bounded Euclidean do-
main, of dimension d . The Dirichlet Cheeger constant is defined by

Vold −1 (∂A)
h D := h D (Ω) = inf , (5.2.1)
A Vold (A)

where the infimum is taken over all compactly embedded open subsets A ⋐ Ω with
smooth boundary.

The subsets A appearing in (5.2.1) are not assumed to be connected. Let us also remark that
the smoothness assumption on ∂A is not restrictive, since any set of bounded perimeter can be
approximated by sets with smooth boundary. We note as well that the Dirichlet Cheeger constant
is somewhat reminiscent of the isoperimetric constant

d
Vold −1 (∂A) d −1
,
Vold (A)

however, unlike the latter it is not scaling invariant.

Theorem 5.2.2: Dirichlet Cheeger’s inequality [Che71]

Let Ω be a compact Riemannian manifold with boundary or a bounded Euclidean do-
main. Then the first Dirichlet eigenvalue of Ω satisfies Cheeger’s inequality
1 2
λD
1 (Ω) ≥ h D (Ω). (5.2.2)
4

In order to prove Theorem 5.2.2 we will need

§5.2. Cheeger’s inequality and its applications 165

Lemma 5.2.3
Let ϕ ≥ 0 be a smooth function such that ϕ|∂Ω = 0. Then
ˆ ˆ
ϕ dx.
¯ ¯
¯∇ϕ¯ dx ≥ h D (Ω)
Ω Ω

Proof of Lemma 5.2.3

Applying a version of the co-area formula (5.1.3) for Riemannian manifolds (see, for in-
stance, [CadFar18, §2.3]), we obtain
ˆ ˆ∞ ˆ ˆ∞
¯ ¯
¯∇ϕ¯ dx = dΣt dt = Vold −1 (L ϕ (t )) dt
Ω 0 L ϕ (t ) 0
ˆ∞ ˆ
≥ hD Vϕ (t ) dt = h D ϕ dx,
0 Ω

where the last equality follows from the layer-cake representation (5.1.7).

Proof of Theorem 5.2.2

We follow the argument presented in [SchYau94, §III.1], see aslo [Bus80]. Let u be the first
Dirichlet eigenfunction of Ω. Then
ˆ ˆ
¯ ¡ 2 ¢¯
¯∇ u ¯ dx = 2 |u||∇u| dx ≤ 2∥u∥L 2 (Ω) ∥∇u∥L 2 (Ω)
Ω Ω (5.2.3)
q
2
= 2 λD
1 (Ω)∥u∥L 2 (Ω) .

Here the first inequality is simply the Cauchy–Schwarz inequality, and the last equality
holds since λD
1 (Ω) = R[u], where R[u] is the Rayleigh quotient of u .
We now use Lemma 5.2.3 with ϕ := u 2 , which implies
ˆ
¯ ¡ 2 ¢¯
¯∇ u ¯ dx ≥ h D ∥u∥2 2 .
L (Ω)
Ω
q
Combining this with the inequality (5.2.3), we get 2 λD
1 (Ω) ≥ h D , and hence (5.2.2).

Consider now the case of the Neumann boundary conditions.

166 Chapter 5. Eigenvalue inequalities

Definition 5.2.4: The Neumann Cheeger constant

Let Ω be a compact Riemannian manifold (with or without boundary) or a bounded
Euclidean domain, of dimension d . The Neumann Cheeger constant is defined by

Vold −1 (Γ)
h N := h N (Ω) = inf , (5.2.4)
Γ min{Vold (Ω1 ), Vold (Ω2 )}

where the infimum is taken over all smooth hypersurfaces Γ (not necessarily connected)
separating Ω into two open sets Ω1 and Ω2 , see Figure 5.1.

Figure 5.1: A hypersurface Γ splitting a domain Ω into two

open sets Ω1 and Ω2 .

Theorem 5.2.5: Neumann Cheeger’s inequality

Let Ω be a compact Riemannian manifold with boundary or a bounded Lipschitz do-
main, of dimension d , and let λN
2 (Ω) be its first nonzero eigenvalue of the Neumann
Laplacian −∆NΩ . Then
1 2
λN
2 (Ω) ≥ h N (Ω). (5.2.5)
4
§5.2. Cheeger’s inequality and its applications 167

Proof
By Corollary 4.1.34 of Courant’s nodal domain theorem, an eigenfunction u correspond-
ing to the first nonzero eigenvalue has exactly two nodal domains Ω+ and Ω− separated
by the nodal set Z u . Without loss of generality, assume that Vold (Ω+ ) ≤ Vold (Ω− ). The
function u satisfies mixed boundary conditions on ∂Ω+ : the Dirichlet one on Z u , and
the Neumann one on ∂Ω+ \ Z u = ∂Ω+ ∩ ∂Ω (if this part of ∂Ω+ is non-empty). The first
eigenvalue of this mixed (Zaremba) problem satisfies λZ1 (Ω+ , Z u ) = λN
2 (Ω). Let us define
the mixed Cheeger constant (cf. [Bus82])

Vold −1 (∂A ∩ Ω+ )
h DN (Ω+ ) = inf ,
A Vold (A)

where the infimum is taken over all open sets A ⊂ Ω+ with smooth boundary such that
∂A ∩ Z u = ;. Arguing as in the proof of Theorem 5.2.2 we obtain

1 2
λZ1 (Ω+ , Z u ) ≥ h DN (Ω+ ).
4
At the same time,
h DN (Ω+ ) ≥ h N (Ω).
Indeed, the volume of any subdomain A ⊂ Ω+ is smaller than Vold (Ω+ ) ≤ Vold (Ω− ), and
Γ := ∂A can be taken as a separating hypersurface for Ω in (5.2.4). Therefore,

1 2 1 2
λN Z
2 (Ω) = λ1 (Ω+ , Z u ) ≥ h DN (Ω+ ) ≥ h N (Ω),
4 4
which completes the proof of (5.2.5).

An exact analogue of Theorem 5.2.5 holds for closed Riemannian manifolds.

Theorem 5.2.6: Cheeger’s inequality for closed manifolds

Let M be a closed Riemannian manifold of dimension d , and let λ1 (Ω) be the first
nonzero eigenvalue of the Laplace-Beltrami operator −∆M (we recall that for a closed con-
nected manifold, in our Notation 2.1.41, 0 = λ0 < λ1 ). Then
1 2
λ1 (M ) ≥ h N (M ),
4
where hN (M ) is the Neumann Cheeger constant (5.2.4).

The proof of Theorem 5.2.6 is almost identical to that of Theorem 5.2.5, the only difference
being that instead of the mixed problem on Ω+ we have a pure Dirichlet problem and should use
the Cheeger constant hD (Ω+ ) instead of hDN (Ω+ ) in the intermediate bounds.
168 Chapter 5. Eigenvalue inequalities

§5.2.2. Examples and further results

The following example shows that in the Riemannian setting Cheeger’s inequality is sharp in any
dimension.

Example 5.2.7
Let Hd be the hyperbolic space of constant sectional curvature −1 (see [Cha84, §2.5] and
[Bur98, §3.4] for the definitions). Let B r ⊂ Hd be a geodesic ball of radius r . An explicit
computation shows that
Vold −1 (∂B r )
> d −1
Vold (B r )
for any r > 0. Moreover, the isoperimetric inequality for the hyperbolic space (see [Oss78,
formula (4.23)]) states that a geodesic ball has the minimal volume of the boundary among
all smooth domains of given volume. Therefore, hD (B r ) > d −1 for any r > 0. At the same
time, another computation yields

(d − 1)2
µ ¶
1
λD
1 (B r ) = +O 2 as r → ∞.
4 r

Therefore, the inequality (5.2.2) is sharp, with the equality attained in the limit as the ra-
dius of the geodesic ball in the hyperbolic space tends to infinity. We refer to [Bus80] for
further details on this example, as well as its generalisation to the case of closed manifolds.

Exercise 5.2.8
Using the isoperimetric inequality for the sphere, show that
³ ´ 2
h N Sd = ³ ´,
d 1
B 2,2

where B is the Euler beta function [DLMF22, §5.12]. In particular, show that hN S2 = 1.
¡ ¢

Example 5.2.7 admits the following important extension ([Yau75], see also [Cha84]). Let M
be a complete simply connected d -dimensional manifold with all sectional curvatures bounded
above by some −κ < 0. Using comparison theorems, one can generalise the isoperimetric in-
equality mentioned above to manifolds of variable negative curvature. For any bounded domain
Ω ⊂ M we have
Vold −1 (∂Ω) p
> (d − 1) κ.
Vold (Ω)
Cheeger’s inequality then implies McKean’s inequality [McK70],

(d − 1)2 κ
λD
1 (Ω) > ,
4
§5.2. Cheeger’s inequality and its applications 169

for any bounded domain Ω ⊂ M . Note that this inequality has no analogue in the Euclidean
space: there exists no nontrivial uniform bound for the first Dirichlet eigenvalue of a bounded
domain in Rd .
It follows from Cheeger’s inequality that if the first eigenvalue is small, the Cheeger constant is
small as well. In fact, for closed manifolds the converse is also true. Recall that the Ricci curvature
Ric of a Riemannian manifold (M , g ) is a 2-tensor which is the trace of the Riemann curvature
tensor (see [Bur98, §4.1.1]). We write Ric ≥ −κ if Ric(ξ, ξ) ≥ −κ|ξ|2g for any ξ ∈ T M .

Theorem 5.2.9: Buser’s inequality [Bus82]

Let M be a closed Riemannian manifold of dimension d , with Ricci curvature Ric ≥ −κ,
κ ≥ 0. Then p
2
λ1 (M ) ≤ 2 (d − 1)κh N (M ) + 10h N (M ). (5.2.6)

As indicated in [Bus82], there is no direct analogue of Theorem 5.2.9 for manifolds with
boundary.

Example 5.2.10: Cheeger’s dumbbell

Buser’s inequality can be illustrated by the following example. Let Mε be a surface, ob-
tained by taking two identical round spheres and smoothly attaching them to each other
by a thin cylinder of length one and radius ε > 0, see Figure 5.2. Take a test function which
is equal to 1 on one sphere, −1 on the other and is changing linearly along the cylinder in
such a way that it is orthogonal to constants on Mε . It then follows from the variational
principle that λ1 (Mε ) = O(ε) as ε → 0. Moreover, with some extra work one can show
that λ1 (Mε ) ̸= o(ε), see [JimMor92]. At the same time, it is clear that hN (Mε ) = O(ε).
This explains the presence of the first term (which is linear in hN ) in Buser’s inequality.

Figure 5.2: Cheeger’s dumbbell.

170 Chapter 5. Eigenvalue inequalities

Exercise 5.2.11
2
Use Exercise 5.2.8 to show that the term containing hN is essential in Buser’s inequal-
ity (5.2.6). Another way to verify this is to consider a sequence of flat square tori Tn =
R2 /(n Z)2 as n → ∞. See [Ben15, Example 3.6] for further details on Cheeger’s constants
of the flat tori and the Klein bottles.

Remark 5.2.12
The Ricci curvature assumption in Buser’s inequality is also necessary: there exists a se-
quence of metrics on a torus with Ricci curvature unbounded from below, such that
their Cheeger constants hN tend to zero, and the first nonzero eigenvalues are uniformly
bounded from below. We refer to [Col17, Example 23] for details. Let us also note that a
lower bound on the Ricci curvature often arises as an assumption in spectral inequalities,
see [HasKokPol16]. At the same time, the dependence on the dimension in the first term
of (5.2.6) can be removed: as was shown in [Led04, Theorem 5.2], see also [DePMon21,
formula (7)], p 2
λ1 (M ) ≤ max{6 κh N , 36h N }.

§5.2.3. The first eigenvalue and the inradius of planar domains

Let us present another application of Cheeger’s inequality. Our exposition closely follows [Gri06].
Let Ω be a simply connected planar domain, and let ρ Ω be its inradius, i.e. the radius of the
largest disk contained inside Ω. Define the reduced inradius
ρΩ
ρeΩ = ,
πρ 2Ω
1+ |Ω|

πρ 2Ω ρΩ
where |Ω| = Area(Ω). Clearly, 0 < |Ω| ≤ 1 and hence 2 < ρeΩ ≤ ρ Ω .

Theorem 5.2.13: [Gri06]

The first Dirichlet eigenvalue of simply connected Ω ⊂ R2 satisfies
1
λD
1 (Ω) ≥ . (5.2.7)
4ρe2Ω

Remark 5.2.14
j2
Note that by domain monotonicity, λD 0,1
1 (Ω) ≤ ρ 2 , where the right-hand side is the first
Ω
Dirichlet eigenvalue of a disk of radius ρ Ω , cf. Proposition 4.2.3. Together with (5.2.7), it
means that λ1 (Ω)ρ 2Ω is uniformly bounded away both from zero and infinity for all sim-
ply connected planar domains. Earlier versions of (5.2.7) were obtained in [Mak65] and
§5.2. Cheeger’s inequality and its applications 171

[Hay78]; see also an improvement in [BañCar94]. In [Oss77], the bound was extended
to non-simply connected planar domains, for which the constant on the right-hand side
depends on the connectivity. In higher dimensions, a straightforward generalisation of
(5.2.7) is false. Indeed, take a unit cube, split it into small cubes with the side length n1 and
remove all the vertices of those cubes. The remaining open set is simply connected, its
inradius tends to zero as n → ∞, while the first Dirichlet eigenvalue remains unchanged,
since a point has capacity zero in R3 (see Remark 3.2.15). We refer to [MazShu05] for a
more delicate higher-dimensional generalisation of (5.2.7).

Theorem 5.2.13 immediately follows from Cheeger’s inequality combined with

Proposition 5.2.15
Let Ω ⊂ R2 be a simply connected domain. Then
1
h D (Ω) ≥ . (5.2.8)
ρeΩ

The proof of the proposition is based on a Bonnesen-type isoperimetric inequality originally

due to A. Besicovitch, which could be viewed as a strengthening of the usual isoperimetric in-
equality for planar domains.

Theorem 5.2.16: [Gri06], [Oss78]

Let Ω ⊂ R2 be simply connected. Then
¢2
|∂Ω|2 − 4π|Ω| ≥ |∂Ω| − 2πρ Ω ,
¡
(5.2.9)

where |Ω| := Vol2 (Ω) denotes the area of Ω and |∂A| = Vol1 (∂A) is its perimeter.

Proof of Proposition 5.2.15

Recall the definition of the Dirichlet Cheeger constant in the planar setting,
½ ¾
|∂A|
h D (Ω) = inf : A ⋐ Ω smooth .
|A|
Since Ω is simply connected, it suffices to consider only simply connected A . Indeed, if A
is not simply connected, filling in the holes increases the area and decreases the perimeter,
while keeping the set inside Ω.
Let us now show that A ⊂ Ω implies that ρeA ≤ ρeΩ . Indeed, consider a function
ρ
f a (ρ) := .
πρ 2
1+ a
172 Chapter 5. Eigenvalue inequalities

It is easy to check that f a′ (ρ) ≥ 0 if πρ 2 ≤ a . Hence, for these values of ρ , f a (ρ) in increas-
ing. Since |A| ≤ |Ω| := Vol2 (Ω),

ρeA = f |A| (ρ A ) ≤ f |Ω| (ρ A ) ≤ f |Ω| (ρ Ω ) = ρeΩ .

Now, applying (5.2.9) to A , we get

¢2
|∂A|2 − 4π|A| ≥ |∂A| − 2πρ A .
¡

Therefore,
ρ A |∂A| ≥ |A| + πρ 2A ,
and hence
A πρ 2
|∂A| 1 + |A| 1 1
≥ = ≥ .
|A| ρA ρeA ρeΩ
Since A ⋐ Ω is arbitrary, this completes the proof of the proposition.

Let us show that Proposition 5.2.15 gives a sharp estimate. We claim that

1
h D (D) = 2 =
ρeD

for the unit disk. One can compute the Cheeger constant hD (D) for the unit disk using the isoperi-
metric inequality. The following useful lemma provides a more elementary way to do this.

Lemma 5.2.17: [Gri06, Proposition 1]

Let Ω ⊂ Rd , let V be a smooth vector field on Ω, and let h ≥ 0. Assume that |V (x)| ≤ 1
and divV (x) ≥ h for all x ∈ Ω. Then hD (Ω) ≥ h .

Proof
Let A ⋐ Ω be an open set with smooth boundary. Then
ˆ ˆ
Vold (∂A) ≥ 〈V (x), n〉 ds = divV (x) dx ≥ h Vold (A),
∂A A

where the equality in the middle holds by the divergence theorem, and the inequalities
follow from the assumptions. The result then immediately follows from (5.2.2).
§5.2. Cheeger’s inequality and its applications 173

Example 5.2.18: [Gri06]

Let Bd ⊂ Rd be the unit ball. Choosing A ⋐ Bd arbitrary close to Bd in (5.2.2) and taking
d
into account that Vol d −1 (S )
Vol (Bd )
= d , we find that h D (Bd ) ≤ d . At the same time, applying
d
the lemma above to the vector field V (x) = x we get hD (Bd ) ≥ div x = d . Therefore,
h D (Bd ) = d .

The following two remarks give some more information on the optimality of the constant in
Cheeger’s inequality.11

Remark 5.2.19
Recall that by Exercise 1.2.21, λ1 (Bd ) = j 2d , i.e. the square of the first zero of the
2
−1,1
Bessel function J d −1 . Using the asymptotics of the first Bessel zero as the order of the
2
Bessel function tends to infinity [Wat95, §15.81], [DLMF22, Eq. 10.21.40], we observe that
2
λ1 (Bd ) = d4 (1 + o(1)) as d → ∞. Since h D (Bd ) = d , this shows that the constant 1/4 in
Cheeger’s inequality (5.2.2) is asymptotically sharp for Euclidean domains as the dimen-
sion d → ∞. We refer to [Fto21, BriButPri22] for further discussion and related results.

Remark 5.2.20
It would be interesting to understand whether the constant 14 in Cheeger’s inequality
(5.2.2) admits an improvement for Euclidean domains of a given dimension. For convex
planar domans, such a result was obtained in [Par17]. In the same paper, a nice way to
unify the inequalities (5.2.2), (5.2.7) and (5.2.8) is presented. Indeed, all these inequalities
can be viewed as relations between the first Dirichlet eigenvalues of the p -Laplacians −∆p ,
which are nonlinear operators defined by

∆p u = div |∇u|p−2 ∇u .
¡ ¢

Note that for p = 2 it is the usual Laplace operator. Moreover, one can show that
λ1 (−∆p , Ω) → h D (Ω) as p → 1 and λ1 (−∆p , Ω) → ρ1Ω as p → ∞.

In conclusion, let us note that we have covered just a few aspects of Cheeger’s inequality. In
particular, aside from its significance in analysis and geometry, it has important applications to
probability and graph theory. For further reading on this topic see, for instance, [Chu97].
11 We thank Dorin Bucur and Dmitry Faifman for useful discussions on Remarks 5.2.19 and 5.2.20.
174 Chapter 5. Eigenvalue inequalities

§5.3. Upper bounds for Laplace eigenvalues

§5.3.1. The Szegő–Weinberger inequality

The Faber–Krahn inequality has stimulated further research on isoperimetric inequalities for
Laplace eigenvalues in various settings. Let us start with the Neumann problem for bounded
Euclidean domains Ω. The Neumann spectrum 0 = µ1 < µ2 ≤ . . . , µ j = µ j (Ω) = λNj (Ω), always
starts with the zero eigenvalue, and therefore the Neumann analogue of the fundamental tone is
the second (i.e., first nonzero) Neumann eigenvalue µ2 . Recall that by formula (3.1.11), the first
nonzero Neumann eigenvalue is given by
∥∇u∥2L 2 (Ω)
µ2 (Ω) = min . (5.3.1)
u∈H 1 (Ω)\{0} ∥u∥2L 2 (Ω)
u⊥1

Remark 5.3.1: Physical interpretation of µ2

Recall that the Neumann boundary conditions for the heat equation correspond to a per-
fectly insulated boundary. Therefore, as the time t → ∞, the temperature distribution
becomes constant at each point of the domain; mathematically, this follows from the fact
that the first Neumann eigenvalue µ1 is equal to zero. The first nonzero Neumann eigen-
value µ2 defines the exponential rate of convergence to this constant distribution.

As follows from Exercise 1.1.15, Neumann (respectively, Dirichlet) eigenvalues do not admit
nontrivial lower (respectively, upper) bounds under the volume constraint. Therefore, while in
the Dirichlet case we were looking for a minimum of the first eigenvalue, in the Neumann case we
should be looking for a maximum. The following theorem was first proved by G. Szegő [Sze54] for
simply connected planar domains, and later generalised by H. F. Weinberger [Wei56] to arbitrary
domains in any dimension.

Gábor Szegő Theorem 5.3.2: Szegő–Weinberger inequality

(1895–1985)
Let Ω ⊂ Rd be a Lipschitz domain, and let Ω∗ ⊂ Rd be a ball of the same volume. Then
µ2 (Ω) ≤ µ2 (Ω∗ ) with equality attained if and only if Ω is a ball.

The following exercise forms one of the crucial steps in the proof of Theorem 5.3.2.

Exercise 5.3.3
Consider the ball B Rd ⊂ Rd of radius R . Using your solution of Exercise 1.2.21 or directly
by separation of variables in spherical coordinates, show that
Hans F. !2
p d′ ,1,1
Ã
Weinberger ³ ´
(1928–2017) µ2 B Rd = , (5.3.2)
R
§5.3. Upper bounds for Laplace eigenvalues 175

where p d′ ,1,1 is the first zero of the derivative of an ultraspherical Bessel function
d
Ud ,1 (r ) := r 1− 2 J d (r ). Moreover, show that the multiplicity of the eigenvalue µ2 B Rd
¡ ¢
2
g (r )x i
is equal to d , and that the corresponding eigenfunctions are given by u i (x) = r ,
i = 1, . . . , d , where x i are the coordinate functions, and

p d′ ,1,1 r
à !
1− d2
g (r ) = r Jd . (5.3.3)
2 R

Proof of Theorem 5.3.2

Let R be the radius of the ball Ω∗ , and let µ∗2 := µ2 (Ω∗ ) = µ2 (B Rd ) be its first non-zero
eigenvalue, given by (5.3.2).
It is easy to show using Bessel equation (1.1.16) that the function g (r ) defined by (5.3.3)
satisfies the equation

d −1 ′ d −1
µ ¶
g ′′ (r ) + g (r ) + µ∗2 − 2 g (r ) = 0. (5.3.4)
r r

In particular, r = R is the first zero of g ′ (r ), and it follows from (1.1.17) that g (r ) is mono-
tone increasing and positive for 0 < r < R . Let us define an extension of g (r ):
(
g (r ) if r ≤ R,
G(R) :=
g (R) if r > R.

It is clear that G(r ) ∈ C ([0, +∞)) and it follows from (1.1.17) that G(r )
r has bounded deriva-
tives as r → 0+ . Therefore, the functions f i (x) := G(rr )xi , i = 1, . . . , d , are in H 1 (Ω).
We will use the following

Lemma 5.3.4: The “centre of mass” lemma

There exists a choice of the origin O of the coordinate system such that
ˆ ˆ
G(r )x i
f i (x) dx = dx = 0 for all i = 1, . . . , d .
r
Ω Ω

This lemma is proved using a topological argument. In fact, an argument of this kind
appears also in the proof of Szegő, as well as in the proof of Hersch’s inequality, see §5.3.2.
Let us postpone the proof of Lemma 5.3.4 for later, and note that for the choice of the
origin O given by this lemma, the functions f i (x), i = 1, . . . , d , are orthogonal to constants,
and hence admissible for the variational characterisation (5.3.1).
176 Chapter 5. Eigenvalue inequalities

∂r xj
Let us calculate the Rayleigh quotients R[ f i ]. Taking into account that ∂x j = r , we
have:
∂ f i (x) G ′ (r )x i x j G(r )x i x j G(r )
= − + δi j , i , j = 1, . . . , d .
∂x j r 2 r 3 r
Therefore, a direct computation yields
x2
³ ´
d µ ∂ f ¶2 G ′
(r ) 2 2
x G(r )2 1 − r i2
¯ ¯2 X
¯∇ f i ¯ = i i
= + , i = 1, . . . , d .
j =1 ∂x j r 2 r2

Thus by the variational principle we get

ˆ ˆ x2
   ³ ´
2 ′
(r )2 x i2 G(r )2 1 − r i2
G(r ) 2  G
x dx µ2 (Ω) ≤ +  dx.
 
r2 i r2 r 2

Ω Ω

Summing up for i = 1, . . . , d , we obtain

´ ³ )2
´
Ω G ′ (r )2 + (d −1)G(r
r2
dx
µ2 (Ω) ≤ ´ . (5.3.5)
2 dx
Ω G(r )
d
Let Ω1 := Ω ∩ Ω∗ and Ω2 := Ω \ Ω∗ , where we assume that Ω∗ = BO,R is now centred
at O , see Figure 5.3.
Then ˆ ˆ ˆ
G(r )2 dx = G(r )2 dx +G(R)2 dx,
Ω Ω1 Ω2
and, since G(r ) is non-decreasing,
ˆ ˆ ˆ
2 2
G(r ) dx = G(r ) dx + G(r )2 dx
Ω∗ Ω1 Ω∗ \Ω1
ˆ ˆ (5.3.6)
≤ G(r )2 dx +G(R)2 dx.
Ω1 Ω∗ \Ω1

Note that Vol(Ω) = Vol(Ω ), and hence Vol(Ω2 ) = Vol(Ω∗ \ Ω1 ). Therefore,

∗
ˆ ˆ ˆ
G(r )2 dx ≥ G(r )2 dx = g (r )2 dx.
Ω Ω∗ Ω∗

Let us investigate the numerator in the Rayleigh quotient (5.3.5). Differentiating the inte-
grand, we get
G(r )2
µ ¶
d ′ 2
G (r ) + (d − 1) 2
dr r
(5.3.7)
r G(r )G ′ (r ) −G(r )2
= 2G ′ (r )G ′′ (r ) + 2(d − 1) .
r3
§5.3. Upper bounds for Laplace eigenvalues 177

For r > R this expression is negative since G(r ) is constant. For r ≤ R , we use the Bessel-
type equation (5.3.4), which yields
d −1 ′ d −1
µ ¶
g ′′ (r ) = − g (r ) + − µ∗
2 g (r ).
r r2
Substituting it into (5.3.7) gives
G(r )2
µ ¶
d ′ 2
G (r ) + (d − 1) 2
dr r
¡ ′ ¢2
∗ ′ r g (r ) − g (r )
= −2µ2 g (r )g (r ) − 2(d − 1)
r3
¡ ′ ¢2
∗ 2 ′ r g (r ) − g (r )
= −µ2 (g (r ) ) − 2(d − 1) < 0,
r3
since g (r )2 is monotone increasing. Therefore, the integrand in the numerator in the (5.3.5)
is monotone decreasing for r > 0, and arguing as in (5.3.6), we get
ˆ µ ˆ µ
(d − 1)G(r )2 (d − 1)g (r )2
¶ ¶
′ 2 ′ 2
G (r ) + 2
dx ≤ g (r ) + 2
dx. (5.3.8)
r r
Ω Ω∗

It follows that ´ ³ (d −1)g (r )2

´
Ω∗ g ′ (r )2 + r2
dx
µ2 (Ω) ≤ ´ . (5.3.9)
2
Ω∗ g (r ) dx
g (r )x
At the same time, since r i is an eigenfunction on Ω∗ corresponding to the first non-
zero eigenvalue µ∗2 , it realises the equality in the variational characterisation (5.3.1):
ˆ ˆ x2
³ ´
g (r )2 x i2 ′
g (r ) 22 g (r )2 1 − r i2
µ∗2 dx = xi + dx.
r2 r2 r2
Ω∗ Ω∗

Summing up for i = 1, . . . , d , we get

ˆ ˆ µ
(d − 1)g (r )2
¶
2 2
µ∗2 g (r ) dx = ′
g (r ) + dx.
r2
Ω∗ Ω∗

In view of (5.3.9), this implies

µ2 (Ω) ≤ µ∗2 = µ2 (Ω∗ ). (5.3.10)
Moreover, it is easy to see that the equality in both (5.3.6) and (5.3.8) is attained if and only
if Ω = Ω∗ up to a set of measure zero. Since by assumption Ω has Lipschitz boundary
(which is a common assumption for the Neumann boundary value problem, but in fact is
not necessary for the validity of (5.3.10), see Remark 5.3.6), µ2 (Ω) = µ2 (Ω∗ ) if and only if
Ω = Ω∗ .
178 Chapter 5. Eigenvalue inequalities

d
Figure 5.3: An example of a domain Ω and a ball Ω∗ = BO,R .
Ω is decomposed into Ω1 := Ω∩Ω∗ (lighter shading) and Ω2 :=
Ω \ Ω∗ (darker shading).

It remains to prove the “centre of mass” lemma.

Proof of Lemma 5.3.4

Let O 1 ∈ Rd be the origin of some initial coordinate system. Consider a ball B d ⊃ Ω, and
let F = (F1 , . . . , Fd ) : B d → Rd be a map defined by
ˆ
G(|x − y|)(x i − y i )
F i (y 1 , . . . , y d ) = dx.
|x − y|
Ω

We want to show that there exists y = (y 1 , . . . , y d ) ∈ B d such that F (y) = 0. Indeed, if this
is the case, choosing O = y as the new origin of the coordinate system proves the result.
Clearly, F is continuous. Take y ∈ ∂B d . The outward unit normal at y is given by
y
n = y . Then,
| |
ˆ ­
x, y − |y|2
®
­ ® X d yi
F (y), n = F i (y) = G(|x − y|) dx.
i =1 |y| |x − y||y|
Ω
§5.3. Upper bounds for Laplace eigenvalues 179

Since Ω ®⊂ B d and y ∈ ∂B d , we have |y| > |x| and hence x, y − |y|2 < 0. Therefore,
­ ®

F (y), n < 0 for all y ∈ ∂B d . Therefore, F is a continuous vector field on B d which

­

points inward on the boundary ∂B d . Then there exists ε > 0 such that the continuous
transformation y 7→ y +εF (y) maps B d into itself. Recall that by Brouwer’s theorem (see,
for example, [Mil98, §2]) such a transformation has a fixed point. Moreover, since F points
inward on the boundary, there are no fixed points on ∂B d . Hence, there exists y ∈ B d such
that F (y) = 0. This completes the proof of the lemma.

Remark 5.3.5
For simply connected planar domains, G. Szegő [Sze54] proved Theorem 5.3.2 using the
Riemann mapping theorem. While his method cannot be extended to higher dimen-
sions, it has been generalised to other contexts, in particular, by R. Weinstock for the first
nonzero Steklov eigenvalue, see §7.1.3, as well as by J. Hersch for the first nonzero Laplace
eigenvalue on a sphere, see §5.3.2. Note also that Szegő’s approach yields a stronger result
(cf. Proposition 5.3.11):
1 1 2
+ ≥ ,
µ2 (Ω) µ3 (Ω) µ2 (Ω∗ )
for any simply connected planar domain Ω.

Remark 5.3.6: Stability of the Szegő–Weinberger inequality

Similarly to the Faber–Krahn inequality, the Szegő–Weinberger inequality is stable: it was
shown in [BraPra12, Theorem 4.1] (see also [BraDeP17]) that for any bounded open set
Ω ⊂ Rd ,
2 2
Vol(Ω∗ ) d µ2 (Ω∗ ) − Vol(Ω) d µ2 (Ω) ≥ c d A (Ω)2 ,
where A (Ω) is the Fraenkel asymmetry defined by (5.1.18), cf. Theorem 5.1.16. Moreover,
the exponent 2 on the right-hand side is sharp. Note that the stability result, as well as
the Szegő–Weinberger inequality itself, could be stated for arbitrary open bounded do-
mains, with µ2 defined by (5.3.1); assuming that the Neumann spectrum is discrete is not
necessary.

Remark 5.3.7: Higher eigenvalues

Among all Euclidean domains of fixed volume, the second nonzero Neumann eigenvalue
µ3 is maximised by a disjoint union of two identical balls. This result was proved in [Gir-
NadPol09] for simply connected planar domains using an argument inspired by Szegő’s
proof, and extended in [BucHen19] to arbitrary Euclidean domains using an argument
inspired by Weinberger’s proof presented above. This result, together with the Szegő–
Weinberger inequality, as well as with the Faber–Krahn and the Krahn–Szego inequal-
180 Chapter 5. Eigenvalue inequalities

ities, implies that Pólya’s conjecture (3.3.8) is true for k = 1, 2. For higher eigenvalues,
both for the Dirichlet and the Neumann boundary conditions, little is known apart from
some numerics, showing that some peculiar shapes may arise as extremal geometries (see
[AntFre12, Figures 1 and 2]).

§5.3.2. Hersch’s theorem for the first eigenvalue on the sphere

The Faber–Krahn and Szegő–Weinberger inequalities gave rise to a new direction in spectral ge-
ometry called isoperimetric inequalities for Laplace eigenvalues. In the following two subsections
we are going to review some of the main results in this subject.
Let (M , g ) be a closed d -dimensional Riemannian manifold, and let λ1 (g ) := λ1 (M , g ) be
the first nonzero eigenvalue of the Laplacian.12 As was shown in (2.1.21), the quantity

λ1 (M , g ) := λ1 (g ) Vol(M , g )2/d

is invariant under rescaling. Adapting the Cheeger’s dumbbell example (see Example 5.2.10) it is
easy to see that inf λ1 (M , g ) = 0 for any M , where the infimum is taken over all Riemannian met-
g
rics on M . Note that the eigenvalues of the closed eigenvalue problem satisfy the same variational
principle as the Neumann eigenvalues, and therefore it is natural to consider the maximisation
problem in this setting.
As it turns out, sup λ1 (M , g ) = +∞ on any compact Riemannian manifold M of dimension
g
d ≥ 3 [ColDod94]. We will therefore restrict ourselves to the case of surfaces. Note that if d = 2,
λ1 (M , g ) = λ1 (M , g ) Area(M , g ). Let us start with the simplest surface, namely, the 2-sphere.

Theorem 5.3.8: Hersch’s theorem [Her70]

Let (S2 , g ) be a sphere endowed with a Riemannian metric g . Then

λ1 (S2 , g ) ≤ 8π, (5.3.11)

Joseph Hersch
(1925–2012) with the equality attained if and only if g is a round metric.

Proof
We follow the argument given in [SchYau94]. Let g 0 be the standard round metric on
S2 . Then Area(S2 , g 0 ) = 4π and, as was shown in §1.2.3, λ1 (g 0 ) = 2 with multiplicity
three. The corresponding eigenspace is generated by the restriction to the sphere of the
coordinate functions x 1 , x 2 , x 3 . see Exercise 1.2.3. Let g be an arbitrary metric on S2
normalised in such a way that Area(S2 , g ) = 4π. We need to show that λ1 (g ) ≤ 2. We
claim that there exists a conformal map ϕ : (S2 , g ) → (S2 , g 0 ) such that the pull-back met-

12 We recall once more that this is a standard way to enumerate eigenvalues of closed Riemannian manifolds, which
is different from the one we used for the Neumann problem on Euclidean domains.
§5.3. Upper bounds for Laplace eigenvalues 181

ric ϕ∗ g 0 = α(x)g with α(x) > 0. Indeed, by the uniformisation theorem, S2 admits a
unique complex structure up to a diffeomorphism, see [Tay11b, Proposition 9.8]. At the
same time, there is a one-to-one correspondence between complex structures and confor-
mal classes on a Riemannian surface (see, for instance, [Bob11, Theorem 4]). Therefore,
up to a diffeomorphism, there is a unique conformal class on S2 , and the claim follows.
Set y i = x i ◦ϕ for i = 1, 2, 3. Recall that by (3.1.14), the Dirichlet energy is conformally
invariant in two dimensions, and hence
ˆ ˆ ˆ
¯∇y i ¯2 dVg = |∇x i |2 dV = 2 x 2 dV = 8π ,
¯ ¯
g i (5.3.12)
3
S2 S2 S2

where dVg and dV are the area forms corresponding to the metrics g and g 0 , respectively.
Note that the last equality follows from the symmetry considerations.
For each p ∈ S2 , we have y 12 (p) + y 22 (p) + y 32 (p) = 1. Thus,
´ 2
3 3
y i dVg ˆ 3
X 1 X S2 3 X 2 3 3
= ´¯ ¯2 = y i dVg = · 4π = . (5.3.13)
i =1 R[y i ] i =1 ¯∇y i ¯ dVg
g
8π i =1 8π 2
S2
S2

Therefore, for at least one of i = 1, 2, 3, we have R[y1 i ] ≥ 21 , and hence R[y i ] ≤ 2. If we were
able to take this particular y i´as a test function for λ1 , that would have been the end of
the proof. However, a priori y i dVg ̸= 0. At the same time, we still have the freedom to
S2
choose the conformal map ϕ. Our goal is to do it in such a way that
ˆ
y i dVg = 0, i = 1, 2, 3. (5.3.14)
S2

In other words, the map ϕ must keep the center of mass at the origin, cf. Lemma 5.3.4.
In order to construct such a map ϕ, we use the group of conformal automorphisms
of the sphere. Let B3 ⊂ R3 be the open unit ball. Given ξ ∈ B3 , define a transformation
K ξ : B3 → B3 by the formula

(1 − |ξ|2 )x + (1 + 2(ξ, x) + |x|2 )ξ

K ξ (x) = . (5.3.15)
1 + 2(ξ, x) + |ξ|2 |x|2

Exercise 5.3.9
Show that K ξ is a conformal map from S2 to itself.

Let us now define a map H = (H1 , H2 , H3 ) : B3 → B3 as

ˆ
1
Hi (ξ) = x i ◦ K ξ ◦ ϕ dVg , i = 1, 2, 3.
4π
S2
182 Chapter 5. Eigenvalue inequalities

Lemma 5.3.10: Hersch’s lemma

There exists ξ ∈ B3 such that H (ξ) = 0.

Proof of Hersch’s lemma

Let ξ ∈ ∂B3 = S2 . Then it follows from (5.3.15) that K ξ (S2 \ {−ξ}) = ξ. Therefore,
ˆ ˆ
1 1
Hi (ξ) = x i ◦ K ξ ◦ ϕ dVg = x i (ξ) dVg = ξi .
4π 4π
S2 S2

In other words, the map H : B3 → B3 can be continuously extended to ∂B3 = S2 ,

and it is the identity on the boundary. Suppose there is no ξ ∈ B3 such that H (ξ) =
H (ξ)
0. Then the map |H (ξ)| : B3 → S2 is a retraction, i.e. a continuous map which is
identity on S2 . We get a contradiction with no-retraction theorem, or, equivalently,
with Brouwer’s fixed point theorem (cf. the proof of Lemma 5.3.4). Indeed, if such
a map exists, we can compose it with a central symmetry with respect to the origin
and get a continuous map of B3 into itself without fixed points, which is impossible.

Replacing now ϕ in y i = x i ◦ ϕ by K ξ ◦ ϕ, where ξ is given by Hersch’s lemma, yields

(5.3.14). It remains to show that the equality in (5.3.11) is attained if and only if the metric
g is round. Without loss of generality, assume that Area(S2 , g ) = 4π. Suppose also that ϕ
keeps the center of mass at the origin (if not, we replace it by K ξ ◦ ϕ as above). Then the
functions y i = x i ◦ ϕi are orthogonal to constants with respect to the measure dVg . The
equality λ1 (g ) = 2 together with the variational principle implies that R[y i ] = 2, i = 1, 2, 3,
and that y i are the first nontrivial eigenfunctions of the Laplacian −∆g with the eigenvalue
2. At the same time, consider the pull-back of the standard round metric ϕ∗ g 0 = α(x)g .
Since x i are the first nontrivial eigenfunctions of −∆g 0 with the eigenvalue 2, the functions
y i are the first nontrivial eigenfunctions of −∆α(x)g with the same eigenvalue. Therefore,

1 2
2y i = −∆α(x)g y i = − ∆g y i = yi , i = 1, 2, 3,
α(x) α(x)

which implies α(x) ≡ 1, and hence g = ϕ∗ g 0 is a round metric.

The proof of Hersch’s theorem implies, in fact, a stronger statement.

§5.3. Upper bounds for Laplace eigenvalues 183

Proposition 5.3.11
For any metric g on S 2 ,
3
X 1 3Area(S 2 , g )
≥ .
i =1 λi (S , g )
2 8π

The proof of this result uses the following generalisation of the variational principle.

Exercise 5.3.12: Variational principle for the sum of eigenvalue reciprocals

Show that
k
X 1 X k 1
= sup ,
i =1 λi (M , g ) i =1 R[ϕi ]
´
where the supremum ¡
is taken
¢
over all 0 ̸= ϕi ∈ H 1 (M , g ), such that M ϕi dV = 0 for
i = 1, . . . , k and ∇ϕi , ∇ϕ j L 2 (M ,g ) = 0 for i ̸= j . A proof of this statement can be found
in [Ban80, formula (3.7)], see also [YanYau80].

Proof of Proposition 5.3.11

In view of (5.3.13), it remains to check that y i , i = 1, 2, 3, can be taken as test functions
in the variational characterisation given in Exercise 5.3.12. Indeed, y i are orthogonal to
constants by Hersch’s lemma. Moreover,
ˆ ˆ
¡ ¢ ­ ® ­ ®
∇y i , ∇y j L 2 (S2 ,g ) = ∇x i , ∇x j dV = 2 x i , x j dV = 0
S2 S2

for i ̸= j , where the first equality follows from the conformal equivalence of the Dirichlet
­ ® ¯ ¯2 ¯ ¯2 ¯ ¯2
energy via the relation 2 ∇y i , ∇y j = ¯∇(y i + y j )¯ − ¯∇y i ¯ − ¯∇y j ¯ .

§5.3.3. Topological upper bounds for eigenvalues on surfaces

Hersch’s theorem has been the starting point for the study of the isoperimetric inequalities for
eigenvalues on surfaces. This is an active area of research, with a number of important recent
advances. The goal of this subsection is to review some of the results in this subject.
Recall that each orientable surfaces is homeomorphic to a sphere with γ ≥ 0 handles. The
number of handles γ is called the genus of a surface. In particular, the sphere itself has genus zero.
Let us start with an extension of Hersch’s estimate to surfaces of higher genus.
184 Chapter 5. Eigenvalue inequalities

Theorem 5.3.13: The Yang–Yau bound [YanYau80]

Let M be an orientable surface of genus γ. Then, for any Riemannian metric g on M ,
3
X 1 3 Area(M , g )
≥ i , (5.3.16)
λ γ+3
h
i =1 i (M , g ) 8π 2

where [·] denotes the integer part. As a consequence,

γ+3
· ¸
λ1 (M , g ) ≤ 8π . (5.3.17)
2

Proof
We follow the argument in [YanYau80]. Assume that there exists a conformal branched
covering (or, equivalently, a non-constant holomorphic map) ψ : (M , g ) → (S2 , g 0 ) of
degree m (see [Bob11, §1.2] for definitions and background). Away from a finite number
of branch points, ψ is a covering map with m sheets. Consider the push-forward metric
m ³ ´∗
ψ−1
X
g∗ = j g (5.3.18)
j =1

on S2 . Here ψ j is a mapping from the j th sheet of the covering to S2 which is well de-
fined by ψ away from the branch points. The metric g ∗ is a smooth metric away from
the branch points, and at those points it has conical singularities, see [KarNPP19, §6] for
details. In fact, one can show that g ∗ = ρg 0 , where 0 ≤ ρ ∈ L p (S2 , g 0 ) for some p > 1,
and the Laplace eigenvalues for such metrics can be defined using the variational principle
in the same manner as for the smooth metrics. However, for the purpose of the present
argument, it suffices to verify that the area defined by g ∗ is finite, which can be done by a
direct computation [YanYau80, p. 58].
It is also not hard to check that for any u ∈ C 1 (S2 ),
ˆ ˆ
u dV∗ = (u ◦ ψ) dVg , (5.3.19)
S2 M

and ˆ ˆ ˆ
2 1 ¯∇(u ◦ ψ)¯2 dVg ,
|∇u|2g ∗ dV∗
¯ ¯
|∇u| dV = = g (5.3.20)
m
S2 S2 M

where dVg and dV∗ are the area forms corresponding to the metrics g on M and g ∗ on
S2 , respectively. Indeed, (5.3.19) follows from the definition of the pull-back measure dV∗ ,
and (5.3.20) follows from (5.3.18) and conformal equivalence of the Dirichlet energy on each
sheet of the covering.
§5.3. Upper bounds for Laplace eigenvalues 185

Let us now proceed as in the proof of Hersch’s theorem. As before, let x i , i = 1, 2, 3, be

the coordinate functions on the round sphere S2 . By Hersch’s lemma, choose a conformal
map ϕ : S2 → S2 such that the center of mass of the measure dV∗ is at the origin. Then
the functions x i ◦ ϕ , i = 1, 2, 3, are orthogonal to constants on (S2 , g ∗ ), and hence by
(5.3.19), the functions v i = x i ◦ ϕ ◦ ψ are orthogonal to constants on (M , g ). Therefore,
setting λi := λi (M , g ), and arguing as in Proposition 5.3.11, we obtain:
´ 2
v i dVg
3 1
M
≥´
X
. (5.3.21)
i =1 λi |∇v i |2g dVg
M

Note that by (5.3.20) and (5.3.12), the denominator in each term on the right-hand side is
3 3
8πm
v i2 = x i2 = 1 pointwise, it follows from (5.3.21) that
P P
equal to 3 . Moreover, since
i =1 i =1

X 3 1 3 Area(M , g )
≥ .
i =1 λi 8πm

To complete the proof of (5.3.16) it remains

h to i note that, as known from the theory of
γ+3
Riemann surfaces, one can choose m = 2 [Gun72, p. 186], see also Remark 5.3.14.
Inequality (5.3.17) immediately follows from (5.3.16) since λ1 is the smallest of the three
eigenvalues.

Remark 5.3.14
γ+3
h i
In the context of the Yang–Yau inequality, a possibility of choosing m = 2 was first
observed in [ElSIli84]. Originally, the inequality was stated in [YanYau80] with m = γ+1.

Substantial new ideas are needed to extend the Yang–Yau theorem to non-orientable surfaces.
This has been done in [Kar16], improving upon the approach of [LiYau82].

Theorem 5.3.15: [Kar16]

Let M be a non-orientable surface with an orientable double cover of genus γ. Then

γ+3
¸ ·
λ1 (M , g ) ≤ 16π . (5.3.22)
2

Estimates (5.3.17) and (5.3.22) imply that the quantity

Λ1 (M ) = sup λ1 (M , g ) (5.3.23)
g

is finite for any surface M . If there exists a metric attaining the supremum in (5.3.23) on a given a
 186 Chapter 5. Eigenvalue inequalities

surface M , we say that this metric is λ1 -maximal. The study of λ1 -maximal metrics on surfaces
is a rapidly developing subject, see [KarNPP21, §2] and references therein. It turns out that such
metrics give rise to minimal isometric immersions of surfaces into spheres Sr by the first eigen-
functions, where r +1 is the multiplicity of the corresponding first eigenvalue. For the time being,
λ1 -maximal metrics are explicitly known only for a few surfaces of low genus:

• Λ1 (S2 ) = 8π, attained on the standard round metric (Hersch’s theorem).

• Λ1 (RP2 ) = 12π, attained on the standard round metric [LiYau82].

2
8π
• Λ1 (T2 ) = p ,
attained on the flat equilateral torus. This was conjectured by M. Berger in
3
Marcel Berger [Ber73] and proved by N. Nadirashvili in [Nad96].
(1927–2016) ³ p ´
• Λ1 (K) = λ1 (K, g max ) = 12πE 2 3 2 , where K is the Klein bottle, g max is a certain metric of
revolution, and E is a complete elliptic integral of the second kind. The Klein bottle (K, g max )
is a bipolar surface for the Lawson τ3,1 -torus and admits a minimal immersion by the first eigen-
functions into S4 . Unlike the examples above, this metric does not have constant curvature.
It was proved to be extremal for the first eigenvalue in [JakNadPol06] and conjectured to be
maximal. It was proved in [ElSGiaJaz06] that there are no other extremal metrics on K, and it
was shown to be maximal in [CiaKarMed19].
Interestingly enough, all the λ1 -maximal metrics above also maximise the multiplicity of the
first eigenvalue on their respective surfaces. On the sphere and on the projective plane it was
Christian Felix
proved in Corollary 4.4.7, and on the torus as well as on the Klein bottle it follows from a
Klein
refinement of (4.4.3) obtained in [Nad87].
(1849–1925)
• Λ1 (Σ2 ) = 16π, where Σ2 is the surface of genus two. The maximum is attained on a metric
with conical singularities on the Bolza surface, induced from the round metric on the sphere
using the standard branched double covering. This result was first stated in [JakLNNP05],
however, the last step of the proof there hinged upon a numerical calculation. A complete
analytic proof was obtained in [NaySho19] using new ideas from algebraic geometry. Note
that the Bolza surface is characterised among all surfaces of genus two as the one having the
largest automorphism group.
Finding the explicit values of Λ1 (M ) and the corresponding maximising metrics is an open
question for all other surfaces.

Remark 5.3.16
All the λ1 -maximising metrics above are unique up to isometries and dilations, except for
the surface of genus two, on which there exists a continuous family of maximisers. More-
Oskar Bolza over, it was shown in [KarNPS21] that all these maximisers, once again with the exception
(1857–1942) of the genus two case, satisfy certain stability properties. We also note that all λ1 -maximal
metrics are highly symmetric, and the multiplicity of the first eigenvalue in all the examples
except for the surface of genus two is maximal possible (cf. Corollary 4.4.7 and [Nad87]).
§5.3. Upper bounds for Laplace eigenvalues 187

One can observe as well that the equality in Yang–Yau inequality (5.3.17) is attained for
γ = 0 and γ = 2; as was shown in [Kar19], this is not the case for all other genera. Further
improvements have been recently obtained in [Ros22a, Ros22b] and [KarVin22].

Let us now present a brief overview of related results for higher eigenvalues. It was conjectured
in [Yau82] and proved by N. Korevaar in [Kor93] (see also [GriNetYau04]), that there exists a
constant C > 0, such that for any k ≥ 1,
λk (M , g ) ≤ C k(γ + 1), (5.3.24)
on any Riemannian surface (M , g ). A substantial improvement of Korveaar’s bound was ob-
tained in [Has11]:
λk (M ) ≤ C (k + γ).
As in the case of the first eigenvalue, these results lead to the question regarding the existence of
λk -maximising metrics and the values of

Λk (M ) := sup λk (M , g )
g

for various k and M . The latter question has been recently completely answered for the sphere
and for the real projective plane. It was conjectured in [Nad02] and shown in [KarNPP21] that
Λk (S2 ) = 8πk, k ≥1 (5.3.25)
(see also [Nad02, Pet14] for k = 2 and [NadSir17] for k = 3), with the supremum attained in the
limit by a sequence of metrics degenerating to a disjoint union of k identical round spheres, see
Figure 5.4.

Figure 5.4: A disjoint union of five identical round

spheres maximising λ5 (S2 , g ).

This is a manifestation of the “bubbling phenomenon” which arises for the maximisers of
higher eigenvalues, see [NadSir15, Pet18, KarNPP19, KarSte20]. Similarly, it was conjectured in
[KarNPP21] and proved in [Kar21] (see also [NadPen18] for k = 2) that
Λk (RP2 ) = 4π(2k + 1), k ≥ 1. (5.3.26)
 188 Chapter 5. Eigenvalue inequalities

For k ≥ 2 the supremum is attained in the limit by a sequence of metrics degenerating to a union
of k − 1 identical round spheres and a standard projective plane touching each other, such that
the ratio of the areas of the projective plane and the spheres is equal to Λ1 (RP2 ) : Λ1 (S2 ) = 3 : 2.

Remark 5.3.17: Korevaar’s bound with an explicit constant

As was noted in [KarNPP19], using (5.3.25) and (5.3.26) one can make the constant C
in the Korevaar’s bound (5.3.24) explicit. Indeed, a slightiadaptation of the proof of
γ+3
h
Theorem 5.3.13 yields that (5.3.24) holds with C = 8π 2 for orientable surfaces and
γ+3
h i
C = 16π 2for non-orientable ones. In the latter case, γ is understood as the genus of
the orientable double cover.

As was mentioned earlier, it was shown in [ColDod94] that Λ1 (M ) = +∞ for any Rieman-
nian manifold M of dimension d ≥ 3. Therefore, in higher dimensions, one needs to restrict the
class of metrics over which the supremum is taken. For example, maximisation of the Laplace
eigenvalues among metrics within a fixed conformal class is an interesting question in any dimen-
sion, see [ColElS03, Kim22, KarSte22, Pet22].

§5.4. Universal inequalities

§5.4.1. The Payne–Pólya–Weinberger inequality

In 1956, L. E. Payne, G. Pólya, and H. F. Weinberger [PayPólWei56] proved the following

Theorem 5.4.1: The Payne–Pólya–Weinberger inequality

For any domain Ω ⊂ Rd , the eigenvalues of the Dirichlet Laplacian λm = λD
m (Ω) satisfy
the gap estimates
4 X m
λm+1 − λm ≤ λj (5.4.1)
Lawrence Edward d m j =1
Payne
(1923–2011) for each m ∈ N.

The inequality (5.4.1) was improved to

m
X λj md
≥ (5.4.2)
j =1 λm+1 − λ j 4

by G. N. Hile and M. H. Protter [HilPro80]. This is indeed stronger than (5.4.1), which can be
obtained from (5.4.2) by replacing all the λ j in the denominators in the left-hand side by λm .
Later, Hongcang Yang [Yan91] proved an even stronger inequality
m ¡ µ µ
4
¶ ¶
λm+1 − λ j λm+1 − 1 + ,λj ≤ 0
X ¢
(5.4.3)
j =1 d
§5.4. Universal inequalities 189

which after some modifications implies an explicit estimate

µ ¶
4 1 Xm
λm+1 ≤ 1 + λj . (5.4.4)
d m j =1

These two inequalities are known as Yang’s first and second inequalities, respectively. We note
that (5.4.3) still holds if we replace λm+1 by an arbitrary z ∈ (λm , λm+1 ] (see [HarStu97]), and
that the sharpest so far known explicit upper bound on λm+1 is also derived from (5.4.3), see
[Ash99, formula (3.33)].
The Payne–Pólya–Weinberger, Hile–Protter and Yang’s inequalities are commonly referred
to as universal estimates for the eigenvalues of the Dirichlet Laplacian. These estimates are valid
uniformly over all bounded domains in Rd and depend only upon the dimension d . The deriva-
tion of all four results is similar and uses the variational principle with ingenious choices of test
functions, as well as the Cauchy-Schwarz inequality. We refer the reader to the survey [Ash99]
which provides the detailed proofs as well as the proof of the implication

(5.4.3) =⇒ (5.4.4) =⇒ (5.4.2) =⇒ (5.4.1) .

In 1997, E. M. Harrell and J. Stubbe [HarStu97] showed that all of these results are conse-
quences of a certain abstract operator identity and that this identity has several other applications.
This approach was further simplified in [LevPar02], and we outline it in the next subsection.
For an alternative proof of Theorem 5.4.1 and other related results, see also [SchYau94, §3.7] and
[Ura17, Chapter 5].

§5.4.2. Abstract commutator identities

We start with

Theorem 5.4.2: [LevPar02, Theorem 2.2]

Let H and G be self-adjoint operators acting in a Hilbert space H with an inner product
(·, ·) := (·, ·)H and a norm ∥·∥ := ∥·∥H . Assume that G(Dom(H )) ⊆ Dom(H ) ⊆ Dom(G)
and that H is semi-bounded from below. Let λ j , j ∈ N, be the eigenvalues of H (ordered
non-decreasingly), and let u j be the corresponding orthonormal eigenvectors. Then for
each fixed j ∈ N

∞ ¯ [H ,G]u , u ¯2
¯¡ ¢¯
∞
j k ¢¯2
(λk − λ j )¯ Gu j , u k ¯
X X ¯¡
= (5.4.5)
k=1 λk − λ j k=1
1¡ ¢
=− [[H ,G],G]u j , u j . (5.4.6)
2
190 Chapter 5. Eigenvalue inequalities

Remark 5.4.3
Note that all the terms in the left-hand side of (5.4.5) with λk = λ j have vanishing denom-
inators. However, as will be shown in the proof, these terms also have vanishing numera-
tors and should be simply dropped from this and similar sums in the sequel.

Proof of Theorem 5.4.2

Obviously, we have

[H ,G]u j = HGu j −G Hu j = (H − λ j )Gu j . (5.4.7)

Therefore,
G[H ,G]u j , u j = G(H − λ j )Gu j , u j .
¡ ¢ ¡ ¢
(5.4.8)
Since G is self-adjoint, we have

G(H − λ j )Gu j , u j = (H − λ j )Gu j ,Gu j

¡ ¢ ¡ ¢

∞ ¡ ∞ ¢¯2 (5.4.9)
(H − λ j )Gu j , u k u k ,Gu j = (λk − λ j )¯ Gu j , u k ¯ .
X ¢¡ ¢ X ¯¡
=
k=1 k=1

We note that [H ,G] is skew-adjoint, since

[H ,G]∗ = (HG −G H )∗ = G H − HG = −[H ,G],

and therefore the left-hand side of (5.4.8) can be rewritten as

¡ ¢ ¡ ¢ ¡ ¢
G[H ,G]u j , u j = − [[H ,G],G]u j , u j + [H ,G]Gu j , u j
¡ ¢ ¡ ¢
= − [[H ,G],G]u j , u j − u j ,G[H ,G]u j ,

so that
¡ ¢ 1¡ ¢
G[H ,G]u j , u j = − [[H ,G],G]u j , u j
2
¡ ¢
(notice that G[H ,G]u j , u j is real, see (5.4.8) and (5.4.9)). This proves (5.4.6).
Since (5.4.7) implies

[H ,G]u j , u k = (λk − λ j ) Gu j , u k ,
¡ ¢ ¡ ¢

this also proves (5.4.5). Obviously, [H ,G]u j , u k = 0 whenever λk = λ j , and the nota-
¡ ¢

tional convention of Remark 5.4.3 therefore applicable.

We can now establish an abstract version of the Payne–Pólya–Weinberger inequality.

§5.4. Universal inequalities 191

Theorem 5.4.4
Under the conditions of Theorem 5.4.2,
m ¡ m ° °2
−(λm+1 − λm )
X ¢ X
[[H ,G],G]u j , u j ≤ 2 °[H ,G]u j ° (5.4.10)
j =1 j =1

for each m ∈ N.

Proof
Let us sum up the equations (5.4.6) over j = 1, . . . , m . Then we have

∞ ¯ [H ,G]u , u ¯2
¯¡ ¢¯
m X 1X m ¡
X j k ¢
=− [[H ,G],G]u j , u j . (5.4.11)
j =1 k=1 λk − λ j 2 j =1

To estimate the left-hand side of (5.4.11) from above, we first note that since [H ,G] is skew-
adjoint,
¯ [H ,G]u j , u k ¯2 = ¯ [H ,G]u k , u j ¯2 ,
¯¡ ¢¯ ¯¡ ¢¯
k, j ≥ 1,
all the terms with k ≤ m cancel out. Then we replace all the positive denominators by the
smallest one λm+1 − λm and use Parseval’s equality, giving

∞ ¯ [H ,G]u , u ¯2 ¯ [H ,G]u j , u k ¯2
¯¡ ¢¯ ¯¡ ¢¯
m X m ∞
X j k X X
=
j =1 k=1 λk − λ j j =1 k=m+1 λk − λ j
1 m ∞ ¯¡
¯ [H ,G]u j , u k ¯2
X X ¢¯
≤
λm+1 − λm j =1 k=m+1
1 m X ∞ ¯¡
¯ [H ,G]u j , u k ¯2
X ¢¯
≤
λm+1 − λm j =1 k=1
1 m °
°[H ,G]u j °2 .
X °
=
λm+1 − λm j =1

Combining this with (5.4.11) proves (5.4.10).

An abstract version of Yang’s inequality (5.4.3) is somewhat more complicated, for the proof
of a slightly more general version see [LevPar02, Corollary 2.8].

Theorem 5.4.5
Under the condition of Theorem 5.4.2,
m °2 1X m ¡ ¢2 ¡
(λm+1 − λm )°[H ,G]u j ° ≥ − λm+1 − λ j [[H ,G],G]u j , u j
X ° ¢
j =1 2 j =1
192 Chapter 5. Eigenvalue inequalities

for all m ∈ N.

Although abstract inequalities in Theorems 5.4.4 and 5.4.5 are valid for any self-adjoint op-
erators H and G such that the commutators involved make sense, in order to obtain meaning-
ful bounds, a choice of G should be adjusted to a particular H , as illustrated below for the case
H = −∆D Ω.

§5.4.3. Applications to Dirichlet eigenvalues

Fix a bounded domain Ω ⊂ Rd , let H = −∆D Ω be the Dirichlel Laplacian on Ω with eigenvalues
λm and orthonormalised eigenfunctions u m . Let G be an operator of multiplication by the co-
ordinate x l , where l is between 1 and d . Obviously, the action of G preserves the domain H01 (Ω)
of −∆D Ω.
An easy computation shows that in this case
∂u
[H ,G]u = −∆(x l u) + x l ∆u = −2〈∇x l , ∇u〉 = −2 ,
∂x l
and
∂(x l u) ∂u
[[H ,G],G]u = −2 + 2x l = −2u,
∂x l ∂x l
therefore (5.4.5)–(5.4.6) simplify to
´
µ ¶2
∂u j
u k dx ˆ
 2
∞ ∂x l ∞
Ω
(λk − λ j )
X X
4 = x l u j u k dx  = 1 (5.4.12)
k=1 λk − λ j k=1
Ω

for any fixed j ∈ N. These relations have a long history — the second equation in (5.4.12), in the
context of a Schrödinger operator acting in Rd is known as the Thomas–Reiche–Kuhn sum rule
in the physics literature. It was derived by W. Heisenberg in 1925 [Hei30]. The name attached
to the sum rule comes from the fact that W. Thomas, F. Reiche, and W. Kuhn had derived some
semiclassical analogues of this formula in their study of the width of the lines of the atomic spectra
[Kuh25, ReiTho25].
We are now in position to prove the original Payne–Pólya–Weinberger inequality (5.4.1).

Proof of Theorem 5.4.1

We use Theorem 5.4.4 with H = −∆D and G = x l , which gives
m °° ∂u j °2
°
4 X
λm+1 − λm ≤ ° ° .
m j =1° ∂x l °L 2 (Ω)

Summing up these inequalities over l = 1, . . . , d and using ∥∇u j ∥2L 2 (Ω) = λ j gives (5.4.1).
§5.4. Universal inequalities 193

Using in the similar manner Theorem 5.4.5 produces (5.4.3).

We further demonstrate the use of commutator trace identity by deducing a bound on a sum
of d consecutive eigenvalues, where d is the dimension. Fix j ∈ N, l ∈ {1, . . . , d }, and consider
again the first equality (5.4.12) re-written as

∞
X w l2k
1
= , (5.4.13)
k=1 λk − λ j 4

where
ˆ
∂u j
w l k := u k dx.
∂x l
Ω

Consider the d ×d matrix W = (w l k ), l = 1, . . . , d , k = j +1, . . . , j +d . We can re-write it for

brevity as
ˆ
W= (∇u j )U dx,
Ω
¡ ¢
where ∇u j is the gradient written as a column vector, U is the row vector u j +1 , . . . , u j +d , and
the integration is performed entry-by-entry. Let Q be matrix of an orthogonal coordinate change
x 7→ Qx . Under this coordinate change, the gradient vector is transformed as ∇u j 7→ Q t ∇u j ,
and therefore the matrix W is transformed as W 7→ Q t W . On the other hand, we can always
choose an orthogonal matrix Q 0 such that W = Q 0 R , where R is an upper-triangular matrix (QR-
decomposition), and choosing the change of coordinates with Q = Q 0 thus makes W upper-
triangular. We now fix this coordinate system, so that

wl k = 0 for l = 1, . . . , d , k = j + 1, . . . , j + l − 1.

We proceed to estimate the left-hand side of (5.4.13) by dropping all the negative terms (with
k < j ), replacing all the denominators in non-zero terms by the lowest possible one, extending
summation to all k starting from one, and using Parseval’s equality, arriving at

° ∂u j °2 ∞ w l2k
° °
1 X 1
° ° ≥ = ,
λ j +l − λ j ° ∂x l °L 2 (Ω) k=1 λk − λ j 4

or
° ∂u j °2
° °
λ j +l − λ j ≤ 4°
° ° .
∂x l °L 2 (Ω)

Summing up these inequalities over l = 1, . . . , d , we obtain

194 Chapter 5. Eigenvalue inequalities

Theorem 5.4.6: [LevPar02]

The eigenvalues λ j of the Dirichlet Laplacian on a bounded domain Ω ⊂ Rd satisfy

d
λ j +l ≤ (4 + d )λ j
X
(5.4.14)
l =1

for all j ∈ N. In particular, in the planar case d = 2,

λ j +1 + λ j +2 ≤ 6λ j .

One of the main drawbacks of the type of universal estimates we have considered is that by
their very nature they are not supposed to be sharp. For example, the Payne–Pólya–Weinberger
bound (5.4.1), the Hile–Protter bound (5.4.2), and Yang’s bound (5.4.4), taken with m = 1, all
yield, for the Dirichlet eigenvalues of bounded domains in Rd ,
λ2 d + 4
≤ . (5.4.15)
λ1 d
At the same time, M. S. Ashbaugh and R. D. Benguria proved, using more accurate approach
involving symmetrisation, the optimal bound for the ratio of the first two Dirichlet eigenvalues,
originally conjectured by Payne, Pólya, and Weinberger.

Theorem 5.4.7: [AshBen91]

Let Ω ⊂ Rd be a bounded domain. Then its first two Dirichlet eigenvalues λm = λD
m (Ω),
m = 1, 2, satisfy
j 2d
λ2 λ2 (Bd ) ,1
≤ = 22 , (5.4.16)
λ1 λ1 (B ) j d
d
2
−1,1

where j p,1 is the first positive zero of the Bessel function J p (x). The equality in (5.4.16) is
attained if and only if Ω is a ball.

The bound (5.4.16) is stronger than (5.4.15): for example, in dimension d = 2 the constants in
the right-hand sides of these bounds are 2.539 (approximately) and 3, respectively. Non-optimality
of universal estimates is even more noticeable for higher eigenvalues, see for example [LevYag03].

Remark 5.4.8: Fundamental gap

Inequality (5.4.16) means in a way that the first and the second Dirichlet eigenvalues of
a Euclidean domain cannot be too far apart. Can they be arbitrary close to each other?
Without further restrictions, the answer is positive: indeed, take a domain which is a
union of two identical balls joined by a thin short passage. However, under the additional
convexity assumption this question can be made interesting if instead of the ratio we con-
§5.4. Universal inequalities 195

sider the difference λD D

2 − λ1 . This quantity is called the fundamental gap. It was shown
in [AndClu11] that
3π2
λD D
2 (Ω) − λ1 (Ω) ≥ , (5.4.17)
diam(Ω)2
where diam(Ω) denotes the diameter of Ω. This inequality has been previously known
as the fundamental gap conjecture, which originated in [vdB83, Yau86, AshBen89]. The
equality in (5.4.17) is attained in the limit as a thin rectangular box degenerates into an
interval. There is also a Neumann analogue of (5.4.17) called the Payne–Weinberger in-
equality:
π2
λN
2 (Ω) ≥ , (5.4.18)
diam(Ω)2
with equality once again achieved in the limit as a thin rectangular box degenerates into
an interval. Since λN
1 (Ω) = 0, it can be viewed as a bound on the Neumann fundamental
gap. Inequality (5.4.18) was proved in [PayWei60], see also [Beb03] for a slight correction
in dimensions d ≥ 3.

§5.4.4. Spectral prescription

d
What about universal bounds for the eigenvalues of the Neumann Laplacian −∆N Ω , Ω ⊂ R ? One
technical difficulty in applying commutator trace identities in this situation is making sure that the
commutators are well defined: necessarily, a choice of G such that G Dom −∆N
¡ ¡ ¢¢ ¡ N¢
Ω ⊆ Dom −∆ Ω
is more complicated than in the Dirichlet case. The resulting bounds are not, strictly speaking,
universal, but depend on some geometric properties of either Ω of M , see, e.g., [HarMic95] and
some further improvements in [LevPar02] by analogy with [ChuGriYau96].
There is however a fundamental obstacle for the existence of universal eigenvalue bounds in
the Neumann case. Consider the following general question of spectral prescription: given a finite
monotone sequence of positive (or non-negative) real numbers, can it coincide with the beginning
of the sequence of eigenvalues of either the Dirichlet or Neumann Laplacian in a domain Ω ⊂
Rd ? Obviously, in the Dirichlet case, the universal bounds (5.4.4) and (5.4.14) for the eigenvalues
should hold: if a finite positive sequence {λ1 , . . . , λK } does not satisfy either of these conditions,
it cannot form the lower part of the spectrum of a Dirichlet Laplacian for a domain in Rd .
Rather surprisingly, in the Neumann case for higher dimensions there are no significant ob-
structions to spectral prescription as demonstrated by the following result of Y. Colin de Verdière.

Theorem 5.4.9: [CdV87, Theorem 1.4]

Let 0 = η 1 < η 2 ≤ · · · ≤ η K be a finite monotone increasing sequence of real numbers.
Then for any d ≥ 3 there exists a domain Ω ⊂ Rd with piecewise C 1 boundary such that
η j = λNj
(Ω) for j = 1, . . . , K . The same is true for d = 2 if and only if K ≤ 4. If, moreover,
the sequence {η j }Kj=1 is strictly increasing, then such a domain exists for any d ≥ 2 and
any K .
196 Chapter 5. Eigenvalue inequalities

A similar result holds in the Riemannian case.

Theorem 5.4.10: [CdV87, Theorems 1.2 and 1.3]

Let M be a closed manifold of dimension d ≥ 3, and let 0 = µ0 < µ1 ≤ µ2 ≤ · · · ≤ µK be
a finite monotone increasing sequence of real numbers. Then there exists a Riemannian
metric g on M such that µ j = λ j (M , g ) for j = 1, . . . , K . If, moreover, the sequence
{µ j }Kj=1 is strictly increasing, this is also true in dimension two.

Note that in dimension two the condition that the sequence is strictly increasing cannot be
completely removed in either the Riemannian or the Neumann case due to the multiplicity bound
(4.4.3) and its Neumann analogue.
 CHAPTER 6
Heat equation, spectral invariants,
and isospectrality
In this chapter, we construct the heat kernel on a Riemannian
manifold and study its asymptotics at small times. As an
application, we prove Weyl’s law for eigenvalues of the
Laplace–Beltrami operator on a closed manifold. We also discuss
spectral invariants arising from the heat asymptotics and the related
question “Can one hear the shape of a drum?”, leading to the notion
of isospectrality. We present Milnor’s example of isospectral
sixteen-dimensional tori as well as a more general Sunada’s
construction of isospectral manifolds. The transplantation of
eigenfunctions and related examples of isospectral planar domains
with Dirichlet, Neumann and mixed boundary conditions are also
presented. We conclude the chapter by a brief overview of results and
open problems concerning spectral rigidity.

§6.1. Heat equation and spectral invariants

§6.1.1. Heat kernel on a Riemannian manifold

Let (M , g ) be a closed Riemannian manifold. Consider the initial-value problem for the heat
equation,
∂u
(
∂t (t , y) = ∆ y u(t , y), t ∈ R+ = (0, +∞), y ∈ M ,
(6.1.1)
u(0, y) = ϕ(y), y ∈ M.
Recall that the physical meaning of the heat equation is as follows: given initial temperature distri-
bution ϕ(y), find the temperature u(t , y) at the point y at the time t . Equation (6.1.1) is also often
referred to as diffusion equation: in this case u(t , y) is understood as the density of the diffusing
substance.

197
198 Chapter 6. Heat equation, spectral invariants, and isospectrality

To simplify notation, throughout this section when integrating over the Riemannian mea-
sure dVg with respect to some variable z we denote the measure simply by dz .

Definition 6.1.1
A fundamental solution of the heat equation (or the heat kernel) is a function e(t , x, y)
for t ∈ R+ , (x, y) ∈ M × M , which is continuous in all three variables, C 1 in t , C 2 in y ,
and satisfies (6.1.1) for all (t , x, y) ∈ R+ × M × M with the initial temperature distribution
ϕ(y) = δx (y). The initial condition is understood in a weak sense:
ˆ
lim+ e(t , x, y) f (y) dy = f (x) (6.1.2)
t →0
M

for any f ∈ C (M ). Here δx denotes the Dirac δ-function supported at the point x ∈ M .

The following important result holds.

Theorem 6.1.2: Existence and uniqueness of a heat kernel

Let (M , g ) be a closed Riemannian manifold. There exists a unique heat kernel e(t , x, y)
on R+ × M × M which is a C ∞ function. Moreover,
∞
e−λ j t u j (x)u j (y),
X
e(t , x, y) = (6.1.3)
j =0

© ª∞
where u j j =0 is an orthonormal basis of eigenfunctions of the Laplace–Beltrami op-
erator −∆M corresponding to the eigenvalues λ j , and the series in the right-hand side
converges pointwise in R+ × M × M .

We follow the exposition in [BerGauMaz71] and [Ros97]. Let us first assume that a heat kernel
exists, and use the method of [Gaf58] to prove that it is unique and is given by (6.1.3).

Proposition 6.1.3
Let M be a closed Riemannian manifold. Suppose that a heat kernel e(t , x, y) exists. Then
it is unique, and the series (6.1.3) converges pointwise in R+ × M × M .

Proof
For any fixed t > 0 and x ∈ M , we can write, by expanding in an orthonormal basis of
eigenfunctions u j in L 2 (M ),
∞
X
e(t , x, y) = e j (t , x)u j (y)
j =0
§6.1. Heat equation and spectral invariants 199

as a function of y . The coefficients of this expansion are given by

ˆ
e j (t , x) = e(t , x, y)u j (y) dy. (6.1.4)
M

Therefore, differentiating with respect to t , we get

ˆ
d
∆ y e(t , x, y) u j (y) dy
¡ ¢
e j (t , x) =
dt
M
ˆ
e(t , x, y) ∆ y u j (y) dy = −λ j e j (t , x),
¡ ¢
=
M

where we first used the fact that e(t , x, y) solves the heat equation, and then integrated by
parts. Hence, we get an ordinary differential equation for e j (t , x) which yields

e j (t , x) = c j (x)e−λ j t ,

with the coefficients c j (x) still to be determined. From the expression (6.1.4) and prop-
erty (6.1.2) we get that c j (x) = u j (x). Hence,
∞
e−λ j t u j (x)u j (y)
X
e(t , x, y) = (6.1.5)
j =0

in L 2 (M ) (in the variable y for given t , x ). The convergence of the series in L 2 (M ) implies
that for any fixed t , x there exists a subsequence j m → ∞ such that
jm
e−λ j t u j (x)u j (y) → e(t , x, y)
X
(6.1.6)
j =0

for almost every y . At the same time, by Parseval’s theorem,

µ µ
t
¶ µ
t
¶¶ ∞ λj t λj t
e− 2 u j (x)e− 2 u j (y)
X
e , x, z , e , y, z =
2 2 L 2 (M ) j =0
∞ (6.1.7)
−λ j t
X
= e u j (x)u j (y)
j =0

for any x, y ∈ M . In particular, the right-hand side of (6.1.7) converges pointwise. Since, by
definition, the heat kernel is continuous in all three variables, the left-hand side of (6.1.7) is
a continuous function in t , x, y . Therefore, the right-hand side defines a continuous func-
tion in R+ ×M ×M . Combining this with the almost everywhere convergence of the series
(6.1.6), we obtain that the right-hand side of (6.1.5) converges pointwise everywhere (since
two continuous functions which are equal almost everywhere are equal). In particular, this
implies that the heat kernel is unique provided it exists.
200 Chapter 6. Heat equation, spectral invariants, and isospectrality

Definition 6.1.4
The heat trace of a closed Riemannian manifold (M , g ) is defined by
∞
e−λ j t = Tr et ∆M .
X
e M (t ) :=
j =0

Corollary 6.1.5
´
The heat trace e M (t ) is a convergent series for t > 0, and its sum equals e(t , x, x) dx .
M

Proof
Setting x = y in the heat kernel expression, we get
∞
e−λ j t u j (x)2 .
X
e(t , x, x) =
j =0

Since all terms are non-negative, we can integrate the series in the right-hand side term by
term, and obtain
ˆ ∞
ˆ ∞
e−λ j t u j (x)2 dx = e−λ j t
X X
e(t , x, x) dx =
j =0 j =0
M M

given that all the eigenfunctions have been chosen to have the unit L 2 norm.

Let us now describe the main ideas of the proof of the existence of the heat kernel.

Existence of the heat kernel: sketch of the proof

First, recall that on Rd ,
d r2
e(t , x, y) = (4πt )− 2 e− 4t ,
¯ ¯
where r = ¯x − y ¯. Note that the Euclidean heat kernel is small unless both r and t are
small. We expect a similar property to hold on an arbitrary Riemannian manifold. More-
over, any Riemannian metric is locally close to a Euclidean one. Hence, we may attempt to
construct an approximate heat kernel for x close to y and t small, by using an appropriate
perturbation of the Euclidean heat kernel, and then modify it slightly to obtain a global
solution.
the Riemannian metric g in Riemannian normal coordinates centred at
Let us expressp
x and set θ(y) = det g (y). We look for approximations of the heat kernel as t → 0+ of
§6.1. Heat equation and spectral invariants 201

the form
d r2
³ ´
S k (t , x, y) = (4πt )− 2 e− 4t v 0 (x, y) + v 1 (x, y)t + · · · + v k (x, y)t k , (6.1.8)

where k ∈ N0 , r = dist(x, y) < ε is now the Riemannian distance, ε > 0 is small enough,
and the functions v j (x, y) depend on the local geometry of the manifold. We choose ε <
ρ inj (M ), where ρ inj (M ) denotes the injectivity radius, so that B x,ε is a geodesic ball for
any x ∈ M . Let us define v j (x, y) recursively as follows, see [BerGauMaz71, §III.E.III]. Set
1
v 0 (x, y) = θ − 2 (y) and
ˆr
− 12 −j 1
v j (x, y) = θ (y)r θ 2 (γ(s))∆ y v j −1 ((γ(s), y)s j −1 ds, j ∈ N,
0

where γ(s) is a unit speed minimal geodesic emanating from x to y . Then for k large
enough, S k is “almost” a solution of the heat equation as t → 0+ in the following sense:
d d −r 2 d
³ ´
L y S k (x, y, t ) = (4π)− 2 t k− 2 e 4t ∆ y v k (x, y) = O t k− 2 , (6.1.9)

where L y = ∂t∂ − ∆ y is the heat operator.

Let Hk = ηS k , where η is a smooth cut-off function with η ≡ 1 near the diagonal x = y ,
and η ≡ 0 when dist(x, y) ≥ ε. One can show that
(i) the functions Hk are smooth for x, y ∈ M and t > 0,
(ii) lim Hk (t , x, y) = δx (y) for all y ∈ M (as in (6.1.2) with e replaced by Hk ).
t →0+

The properties (i) and (ii) hold for any k ≥ 0. Moreover,

(iii) for any k > d2 , L y Hk can be extended to a continuous function in R≥0 × M × M .
Note that t = 0 is included: this is the most nontrivial point of the statement (iii) which
can be deduced using (6.1.9).

Remark 6.1.6
A function satisfying the conditions (i)–(iii) is called a parametrix for the heat
equation. In fact, one can show that L y Hk ∈ C l (R≥0 × M × M ) for k > d2 + l ,
l ≥ 0.

Let us now modify a parametrix to a fundamental solution. Recall the notion of a

convolution of two continuous functions F, H ∈ C (R≥0 × M × M ):
ˆt ˆ
(F ∗ H )(t , x, y) := F (s, x, z)H (t − s, z, y) dz ds.
0 M
202 Chapter 6. Heat equation, spectral invariants, and isospectrality

We will also denote the iterated convolutions by F ∗ j = F ∗ · · · ∗ F , where F is repeated

j ≥ 1 times.

Exercise 6.1.7
Let F ∈ C (R≥0 × M × M ). Show that for any k > d2 +2, F ∗ Hk ∈ C 2 (R+ × M × M )
and L y (F ∗ Hk ) = F + F ∗ (L y Hk ). For a solution, see [BerGauMaz71, Lemme
E.III.7]. Compare this exercise with Duhamel’s principle [Eva10, §2.3.1.c], [Cha84,
§VI.1].

∞
d
(−1) j +1 (L y Hk )∗ j . One can show that the
P
Fix some k > 2 + 2, and set F = F k =
j =1
series defining Fk converges, and Fk ∈ C 2 (R≥0 × M × M ). We claim that the function
P k (t , x, y) := Hk − F k ∗ Hk is the fundamental solution of the heat equation. Indeed, by
Exercise 6.1.7, P k (t , x, y) ∈ C 2 (R+ × M × M ) and

L y P k = L y (Hk − F k ∗ Hk ) = L y (Hk ) − L y (F k ∗ Hk )
= L y Hk − F k − F k ∗ (L y Hk )
∞ ∞
= L y Hk − (−1) j +1 (L y Hk )∗ j − (−1) j +1 (L y Hk )∗( j +1) = 0.
X X
j =1 j =1

It remains to check that

P k (t , x, y) → δx (y) (6.1.10)
as t → 0+ . Indeed, lim+ Hk (t , x, y) = δx (y). At the same time, one can show that there
t →0
exists C > 0 such that
d
F k (t , x, y) ≤ C t k− 2 (6.1.11)
for all x, y ∈ M and 0 ≤ t < 1, see [BerGauMaz71, Lemme E.III.6]. A direct computation
then implies that
F k ∗ Hk → 0 as t → 0+
for any k > d2 (where convergence is understood in the sense of measures), which proves
(6.1.10). Since the uniqueness of the heat kernel has already been established, we have
P k (t , x, y) = e(t , x, y) (note that this implies that the definition of P k (t , x, y) does not
depend on the choice of k > d2 + 2). This completes the proof of Theorem 6.1.2.
§6.1. Heat equation and spectral invariants 203

§6.1.2. Heat kernel asymptotics

From the viewpoint of spectral geometry, of particular interest is the behaviour of the heat kernel
on the diagonal x = y as t → 0+ .

Theorem 6.1.8: Minakshisundaram–Pleijel asymptotic expansion [MinPle49]

Let (M , g ) be a closed Riemannian manifold, dim M = d . The following asymptotic
expansion of the heat kernel holds for t → 0+ :
à !
k Subbaramiah
− d2
³ ´
j k+1
X
e(t , x, x) = (4πt ) a j (x)t + O t , Minakshisundaram
j =0 (1913–1968)

for all k > 0. The heat kernel coefficients a j (x) are called the local heat invariants and are
calculated in terms of the local geometry of M near x .

Proof
We have e(t , x, y) = Hk − Fk ∗ Hk for all k > d2 + 2. Since on the diagonal y = x one has
Hk (t , x, x) = S k (t , x, x), with S k (t , x, y) given by (6.1.8), we obtain

d
k+1
v j (x, x)t j .
X
(4πt ) 2 Hk+1 (t , x, x) =
j =0

Set
a j (x) := v j (x, x),
then

d
k d
a j (x)t j + a k+1 (x)t k+1 − (4πt ) 2 (F k+1 ∗ Hk+1 )(t , x, x).
X
(4πt ) 2 e(t , x, x) =
j =0

In view of (6.1.11), we get, for 0 < t < 1,

¯ˆ ˆ
¯ t ¯
¯
¯ ¯
|(F k+1 ∗ Hk+1 )(t , x, x)| = ¯
¯ F k+1 (s, x, z)Hk+1 (t − s, z, x) dz ds ¯¯
¯ ¯
0 M
ˆt ˆ
k+1− d2
≤ C1 t |Hk+1 (t − s, z, x)| dz ds
0 M
ˆt ˆ
d
= C 1 t k+1− 2 |Hk+1 (s, z, x)| dz ds,
0 M
204 Chapter 6. Heat equation, spectral invariants, and isospectrality

where throughout this proof C j denote some positive constants which may depend on k .
Note that Hk+1 (s, z, x) is non-zero only near the diagonal z = x , so we can assume that
dist(z, x) < ρ with ρ ∈ (ε, ρ inj (M )), where ε is defined after (6.1.8). Then |Hk+1 (s, z, x)|
2
is bounded by C s −d /2 e− dist(z,x) /(4s) , with C independent of z and x . We therefore get

ˆt ˆ
dist(z,x)2
k+1− d2 d
|(F k+1 ∗ Hk+1 )(t , x, x)| ≤ C 2 t s − 2 e− 4s dz ds
0 B x,ρ

ˆ1 ˆ
|y|2
k+1− d2 d
≤ C3 t s − 2 e− 4s dy ds
0 B (0,ρ)⊂Rd

ˆ1 ˆ
|w|2
k+1− d2
≤ C3 t e− 4 dw ds
0 Rd
k+1− d2
= C4 t ,
p
where we changed the variables as w = y/ s . This completes the proof of the theorem.

∞
e−λ j t u j (x)2 . Therefore, as t → 0+ ,
P
Recall now that e(t , x, x) =
j =0

∞ d
∞
e−λ j t = (4πt )− 2 aj t j,
X X
(6.1.12)
j =0 j =0

´
where a j := a j (M ) = a j (x) dx . The coefficients a j are called the heat invariants of the Rie-
M
mannian manifold M .
The heat trace asymptotics is an important tool in the study of the inverse spectral problem,
which is concerned with the recovery of the geometric properties of the manifold M from the
spectrum of the corresponding Laplace–Beltrami operator. Following Mark Kac, this problem is
Mark Kac often described by the celebrated question: “Can one hear the shape of a drum?” [Kac66]. We say
(1914–1984) that a property of M is a spectral invariant (or that it can be “heard”) if it is completely determined
by the Laplace spectrum. For example, the left-hand side in (6.1.12) is determined by the Laplace
eigenvalues of M . This immediately implies that the dimension d and the heat invariants a j are
spectral invariants. Using explicit calculations in Riemannian normal coordinates one obtains
(see [Ros97, §3.3])
1
a 0 (x) = 1, a 1 (x) = τ(x),
6
where τ(x) is the scalar curvature. Hence, a0 = Vol(M ), and therefore ´the volume of a Rieman-
nian manifold is a spectral invariant. Similarly, the total scalar curvature M τ(x) dx is determined
§6.1. Heat equation and spectral invariants 205

by the spectrum. Moreover, if M is two-dimensional, its Euler characteristic is given by

ˆ
1 3
χ(M ) = τ(x) dx = a 1 (M ).
4π 2π
M

Therefore, the Euler characteristic of a surface is a spectral invariant; in particular, one can hear
the number of handles of an orientable surface!
There is a vast literature on the computation of heat invariants (see, for instance, [Gil04],
[Pol00] and references therein), and there exist various ways to express them. Geometrically, the
most natural way is to present the local heat invariants in terms of curvatures and their derivatives.
However, the complexity of this task rapidly increases for higher heat invariants, and the geometric
information becomes difficult to extract. Still, heat invariants are quite useful in the study of
spectral rigidity, see §6.2.6 for further details.

§6.1.3. Weyl’s law on a Riemannian manifold

Let us now use the heat trace expansion (6.1.12) to prove Weyl’s law for the eigenvalue counting
function on closed manifolds. We have already stated this result with a sharp remainder estimate,
see Theorem 3.3.4). As was mentioned in Remark 3.3.5, its proof uses techniques that are beyond
the scope of this book. Below we present a proof of Weyl’s law based on the heat trace expansion,
albeit with a weaker remainder estimate.

Theorem 6.1.9: Weyl’s law for manifolds

Let M be a closed Riemannian manifold, dim M = d . The counting function N M (λ) of
Laplace–Beltrami eigenvalues on M satisfies the asymptotics
d
³ d´
N M (λ) = C d Vol(M )λ 2 + o λ 2 . (6.1.13)

1 ωd
As before, the numerical coefficient is C d = d = (2π)d
, where ωd is the volume of
Γ d2 +1
¡ ¢
(4π) 2
the unit ball in Rd .
The proof of Theorem 6.1.9 will use the following well-known result, see, for example, [Fel71,
§XIII.5].

Theorem 6.1.10: Hardy–Littlewood–Karamata Tauberian theorem

Let N (λ) be a monotone increasing function such that
ˆ∞
e−λt dN (λ) = c t −α + o t −α as t → 0+ .
¡ ¢

Then
c
λα + o λα as λ → ∞.
¡ ¢
N (λ) =
Γ(α + 1)
206 Chapter 6. Heat equation, spectral invariants, and isospectrality

Proof of Theorem 6.1.9

Since a0 = Vol(M ), it follows from the heat trace expansion that
ˆ∞ ∞ 1
e−λt dN (λ) = e−λ j t =
X
d
(Vol(M ) + O(t )). (6.1.14)
j =0 (4πt ) 2
0

Taking α = d2 and applying the Hardy–Littlewood–Karamata Theorem to the right-hand

side of (6.1.14) completes the proof of Theorem 6.1.9.

Remark 6.1.11
The heat trace expansion (6.1.12) can be extended to manifolds with Dirichlet or Neumann
boundary conditions, see [Gil04]. For manifolds with boundary, the expansion has twice
as many terms:
∞ d
∞ k
e−λ j t ∼ + (4πt )− 2
X X
ak t 2.
j =1 t →0 k=0
2

As before, a0 = Vol(M ), but the terms inside the sum corresponding to k = m + 21 with
integer m , arise from the boundary contributions. In particular,
p
π
a1 =± Vold −1 (∂M ), (6.1.15)
2 2
where the plus sign is taken for the Neumann boundary condition and the minus sign for
the Dirichlet boundary condition. It follows that the volume of the boundary is a spectral
invariant.

Exercise 6.1.12
Assume that the conjectured two-term asymptotic formula (3.3.5) in Weyl’s law holds. Use
Theorem 6.1.10 to show that formula (6.1.15) agrees with the second term in (3.3.5) .

Remark 6.1.13
The main term of the heat trace asymptotics (and, hence, of Weyl’s asymptotics (6.1.13)
for the eigenvalue counting function) is not affected by the boundary condition. This
can be explained using Kac’s principle of “not feeling the boundary”. It is best illustrated
using the model of diffusion: for small times, the particles in the interior do not feel the
boundary, and the diffusion process is not influenced by the boundary conditions. We
refer to [Kac51] for further details.
§6.2. Isospectral manifolds and domains 207

Example 6.1.14: Heat trace asymptotics for planar domains

Let Ω be a smooth planar domain with r boundary components. Then the Dirichlet heat
trace of Ω satisfies
∞ Area(Ω) L(∂Ω) (2 − r )
X −λDj (Ω)t
e = − p + + o(1).
j =1 4πt 8 πt 6

For the Neumann boundary condition, the second term should be taken with a plus sign:
∞ Area(Ω) L(∂Ω) (2 − r )
X −λNj (Ω)t
e = + p + + o(1).
j =1 4πt 8 πt 6

In the presence of corners, the third term becomes more complicated and depends on the
angles at the corner points, see [vdBSri88], [NurRowShe19] and references therein.

§6.2. Isospectral manifolds and domains

§6.2.1. Isospectrality

We start with

Definition 6.2.1: Isospectral manifolds

We say
¡
that two
¢
closed
¡
Riemannian
¢
manifolds (M , g ) and (N , h) are isospectral if
Spec −∆(M ,g ) = Spec −∆(N ,h) , understood as the equality of multisets with account
of multiplicities.

Similarly, one can define isospectrality for manifolds with boundary and for Euclidean do-
mains: in this case, boundary conditions have to be specified. One of the central questions in
spectral geometry is to understand the possible mechanisms of isospectrality: how to construct
manifolds or domains that are isospectral and not isometric? A counterpoint to this question is
spectral rigidity: which manifolds or domains are uniquely defined by their spectrum, or at least
do not admit isospectral deformations? We focus on these problems in the present section.
It turns out that the heat trace is an important tool in the study of isospectrality. The follow-
ing simple observation is useful.

Exercise 6.2.2
Let (M , g ) and (N , h) be two compact Riemannian manifolds; if their boundaries are
non-empty, we assume that the same self-adjoint boundary condition is specified on each
boundary. Suppose that the corresponding heat traces coincide for all times: e M (t ) =
e N (t ), t > 0. Then (M , g ) and (N , h) are isospectral.
208 Chapter 6. Heat equation, spectral invariants, and isospectrality

Below we present two elegant constructions of isospectral and not isometric Riemannian
manifolds, relying on the heat trace. The first one is due to J. Milnor [Mil64] and the second one
was discovered by T. Sunada [Sun85]. In fact, Sunada’s construction has lead to a whole variety of
examples of isospectral manifolds and domains. Somewhat surprisingly, Milnor’s and Sunada’s
examples are based on methods coming from different areas of mathematics which are seemingly
distant from spectral geometry: the theory of modular forms and group theory.

§6.2.2. Milnor’s example

In this subsection we follow the exposition of [BerGauMaz71, §III.B.III]. The argument is based
on the Poisson summation formula for lattices. First, let us recall the usual Poisson summation
formula: given a Schwartz (see §B.2) function f ∈ S (Rd ), we have

fˆ(m).
X X
f (k) = (6.2.1)
k∈Zd m∈Zd

Here ˆ
fˆ(y) := f (x)e−2πi〈x,y 〉 dx = (2π)d /2 (F f ) 2πy
¡ ¢

Rd
Siméon Denis
Poisson is the rescaled Fourier transform of f , cf. (2.1.3).
(1781—1840)
Exercise 6.2.3
Prove the Poisson summation formula (6.2.1) for d = 1. Hint: compute the Fourier coef-
P
ficients of the 1-periodic function F (x) := f (x +k) and evaluate the resulting Fourier
k∈Z
series at x = 0.

The Poisson summation formula can be generalised to an arbitrary lattice Γ in Rd (that is, a
discrete additive subgroup of Rd such that Rd /Γ is compact). If Γ is a lattice, let Γ∗ be the dual
n
lattice, i.e. Γ consists of all elements x ∈ R such that the scalar product x, y ∈ Z for all y ∈ Γ.
∗
­ ®

The Poisson summation formula for lattices states that

1
fˆ(m),
X X
f (k) =
k∈Γ Vol(Γ) m∈Γ∗

2
where the volume of a lattice is understood as the volume of Rd /Γ. Take f (x) = e−a|x| ∈ S (Rd ),
where a > 0. Then,
³ π ´ d π2 | y | 2
2 −
fˆ(y) = e a .
a
1
Plugging f (x) with a = 4t into the Poisson summation formula and switching the variables x
and y , we obtain
Vol(Γ) X | y |2
2
t |x|2
e−4π e−
X
= d
4t . (6.2.2)
x∈Γ∗ (4πt ) 2 y∈Γ
§6.2. Isospectral manifolds and domains 209

Note that the left-hand side of (6.2.2) is precisely the heat trace of the flat torus Rd /Γ, because its
2 2
¯ given by 4π |x| , x ∈ Γ , cf. Exercise 1.2.10.d The right-hand side can be interpreted
∗
eigenvalues¯are
as follows: y are the lengths of the closed geodesics in R /Γ, and in the sum we take one closed
¯ ¯
geodesic in each free homotopy class.

Remark 6.2.4
The Poisson formula is a manifestation of a link between the spectral (quantum) and dy-
namical (classical) quantities, which can be explained via Bohr’s correspondence principle
in quantum mechanics. This important connection has already been mentioned in §3.3.2,
and we will revisit it in §6.2.6. There exist various generalisations of the Poisson formula,
such as the Selberg trace formula, the Balian–Bloch trace formula, the wave-trace formula,
etc. For a generalisation based on the heat trace we refer to [CdV73].

Consider now the following special class of lattices in Rd with d = 8k , k ∈ N. Let Γ2 be the
lattice in Rd consisting of (x 1 , . . . , x d ) ∈ Zd such that dj=1 x j is even. It is a sublattice (i.e., a
P

subgroup)
¡1
of Zd of index two. Let Γ(d ) be the lattice in Rd generated by Γ2 and the vector w d =
1
2 , . . . , 2 . Since 2w d ∈ Γ2 (recall that d = 8k is even), it is easy to check that Γ2 is a sublattice of
¢

index two in Γ(d ). Hence Vol(Γ(d )) = 12 Vol(Γ2 ) = Vol(Zd ) = 1.

Exercise 6.2.5
Let Γ = Γ(d ) for d = 8k , k ∈ N. Show that

(i) for all x ∈ Γ, |x|2 is even;

(ii) Γ∗ = Γ.

Consider two 16-dimensional lattices Γ(16) and Γ(8, 8) := Γ(8) ⊕ Γ(8).

Exercise 6.2.6
p
Show that the lattice Γ(8) is generated by the elements of norm 2, while Γ(16) is not.

Theorem 6.2.7: [Mil64]

The two flat 16-dimensional tori, M1 = R16 /Γ(16) and M2 = R16 /Γ(8, 8), are isospectral
but non-isometric.

Proof
It immediately follows from Exercise 6.2.6 that the tori M1 and M2 are not isometric. Let
us show that M1 and M2 are isospectral by comparing their heat traces. Given an arbitrary
210 Chapter 6. Heat equation, spectral invariants, and isospectrality

lattice Γ ⊂ Rd , consider its theta-function

t
µ ¶
X −π|x|2 t
θΓ (t ) := e Rd /Γ = e .
4π x∈Γ∗

Let Γ ⊂ Z16 be a lattice satisfying the properties (i) and (ii) in Exercise 6.2.5; clearly, this is
true for both Γ(16) and Γ(8, 8). Property (ii) implies, in particular, that Γ is unimodular,
i.e. Vol(Γ) = 1. Therefore, the Poisson summation formula yields
2 | y |2
θΓ (t ) = e−π|x| t = t −8 e−π
X X
t .
x∈Γ y∈Γ

Hence, θΓ (t ) = t −8 θΓ t −1 . One can show, using the Weierstrass theorem, that θΓ (t )

¡ ¢

extends to a holomorphic function on the complex half-plane Re z > 0, and that

θΓ (z) − z −8 θΓ (z −1 ) = 0.

Indeed, this equality holds for any real positive z , and since holomorphic functions have
isolated zeros, it must hold for all Re z > 0. Set
θeΓ (z) := θΓ (−iz).

The function θeΓ is holomorphic in the upper half-plane Im z > 0, and satisfies
θeΓ (z) = z −8 θeΓ −z −1 .
¡ ¢
(6.2.3)
Moreover,
2 2
θeΓ (z + 1) = eiπ|x| z eiπ|x| = θeΓ (z),
X
(6.2.4)
x∈Γ

since |x|2 is even by the first assertion of Exercise 6.2.5.

Exercise 6.2.8
Using (6.2.3) and (6.2.4), show that

az + b
µ ¶
θΓ
e = (c z + d )8 θeΓ (z)
cz + d

for any a, b, c, d ∈ Z such that ad − bc = 1, i.e., the matrix

µ ¶
a b
∈ SL2 (Z).
c d

Note also that if z = u + iv and v → ∞, then all the terms in the sum (6.2.4) vanish in
the limit except for |x| = 0, and hence
θeΓ (u + iv) → 1 as v → ∞. (6.2.5)
§6.2. Isospectral manifolds and domains 211

This condition, together with the result of Exercise 6.2.8, implies that θeΓ (z) is a modular
form of weight 8. However, it is known that such a form is unique up to multiplication,
see [Ser73, §VII.3.2, Theorem 4]. Therefore, condition (6.2.5) determines θeΓ (z) uniquely,
and hence θΓ (z) does not depend on the choice of Γ. In particular, the heat traces for
Γ = Γ(16) and Γ = Γ(8) ⊕ Γ(8) coincide. Therefore, it follows from Exercise 6.2.2 that the
corresponding 16-dimensional tori M1 and M2 are isospectral. This completes the proof
of the theorem.

Exercise 6.2.9
Show that any two isospectral two-dimensional flat tori are isometric.

In fact, the minimal dimension in which there exist isospectral but not isometric flat tori is
equal to four, see [Sch90, ConSlo92, Sch97].

§6.2.3. Sunada’s construction

In this subsection we follow the exposition of R. Brooks [Bro88], [Bro98]. Let M and N be
two closed smooth manifolds. Recall that p : M → N is a covering map if it is a surjective and
continuous map such that every point in N has an open neighbourhood whose pre-image is a
disjoint union of open sets, and the restriction of p to each of them is a homeomorphism. A
covering (or deck) transformation corresponding to a smooth covering map p is a diffeomorphism
ψ such that p ◦ ψ = p :
ψ
M M

p p
N

In other words, a deck transformation permutes the elements of the fiber p −1 (x), x ∈ N . The set
of all covering transformations is called a covering group. If, in addition, the manifolds M and N
are Riemannian, and p is a local isometry, we say that p is a Riemannian covering map. If ω is a
Riemannian metric on N , then ω e = p ∗ ω is a Riemannian metric on M which is invariant under
the deck transformations, and p : (M , ω)
e → (N , ω) is a Riemannian covering.

Example 6.2.10
If p : Sd → RP d is the standard double cover, its deck transformation group is Z2 .

Let π1 (N , b) be the fundamental group of N with the base point b ¡∈ N¡ , and¢¢let be ∈ M be

e = b . A covering map p : M → N is called normal if p ∗ π1 M , b
such that p(b) e is a normal
subgroup of π1 (N , b). It is easy to verify that this definition does not depend on the choice of
the base points. One can show that a covering map p is normal if and only if its group of deck
212 Chapter 6. Heat equation, spectral invariants, and isospectrality

transformations G acts transitively on the fibers, i.e. for any x ∈ N and any xe1 , xe2 such that p(xei ) =
x , i = 1, 2, there exists g ∈ G such that g xe1 = xe2 , see [Hat01, Proposition 1.39].

Theorem 6.2.11
Let p : M → N be a normal Riemannian covering with a finite covering group G . Then
the heat kernels on M and N are related by
X
e N (t , x, y) = e M (t , xe, g ye), (6.2.6)
g ∈G

where p(xe) = x and p( ye) = y .

Note that since p is a normal covering, the right-hand side of (6.2.6) does not depend on the
particular choice of the pre-images xe and ye.

Exercise 6.2.12
Prove Theorem 6.2.11. Hint: Use a direct computation to show that the right-hand side
of (6.2.6) satisfies the heat equation and the initial condition.

Therefore, the heat trace on the Riemannian manifold N can be represented as

ˆ ˆ
1 X
e N (t ) = e N (t , x, x) dx = e M (t , xe, g xe) dxe,
cardG g ∈G
N M

where cardG is the cardinality of the group G . The last equality follows by replacing the integra-
tion over M by the integration over (cardG) copies of N .
Let h be an isometry of M . Then e M (t , h xe, h ye) = e M (t , xe, ye) and
ˆ ˆ ˆ
e M (t , xe, hg h −1 xe) dxe = e M (t , h −1 xe, g h −1 xe) dxe = e M (t , xe, g xe) dxe.
M M M

Therefore, one can rewrite the formula for the heat trace as
ˆ
X card[g ]
e N (t ) = e M (t , xe, g xe) dxe, (6.2.7)
[g ]⊂G cardG
M

where [g ] denotes the conjugacy class of the element g ∈ G .

Definition 6.2.13: Sunada triple

Let G be a finite group and let H1 , H2 be two subgroups of G . We say that (G, H1 , H2 ) is
§6.2. Isospectral manifolds and domains 213

a Sunada triple if for any g ∈ G ,

card{[g ] ∩ H1 } = card{[g ] ∩ H2 }.

Definition 6.2.13 implies that if (G, H1 , H2 ) is a Sunada triple, then card H1 = card H2 .
In group theory, the subgroups satisfying Definition 6.2.13 have been first considered by F. Gass-
mann, and thus Sunada triples are sometimes referred to as Gassmann triples.

Exercise 6.2.14: Gassmann’s example [Gas26]

Let G = Sym(6), a symmetric group acting on six elements {a, b, c, d , e, f }, and let H1 =
{1, (ab)(cd ), (ac)(bd ), (ad )(bc)}, H2 = {1, (ab)(cd ), (ab)(e f ), (cd )(e f )} be two sub-
groups of G . Show that (G, H1 , H2 ) is a Sunada triple and the subgroups H1 and H2 are
not conjugate in G (i.e. there is no g ∈ G such that g H1 g −1 = H2 ).

We can now describe the Sunada construction of isospectral manifolds. Consider the follow-
ing diagram of coverings where p is normal (and hence p 1 and p 2 are normal as well):

M
H1 H2
p1 p2
N1 G p N2 (6.2.8)

For example, we may assume that N is a four-dimensional manifold with the fundamental group
G (it is known that any finite group can be realised as the fundamental group of a four-manifold),
and M is its universal cover.

Theorem 6.2.15: [Sun85]

Suppose that (G, H1 , H2 ) is a Sunada triple, and let manifolds M , N , N1 , N2 be as on the
diagram (6.2.8). Take any Riemannian metric on N and lift it to the coverings N1 and N2 .
Then the Riemannian manifolds N1 and N2 are isospectral.
214 Chapter 6. Heat equation, spectral invariants, and isospectrality

Proof
In view of formula (6.2.7) for the heat trace, we have for i = 1, 2,
ˆ
X card([g ])
e Ni = e M (t , xe, g xe) dxe
[g ]⊂Hi card Hi
M
ˆ
X card([g ] ∩ Hi )
= e M (t , xe, g xe) dxe,
[g ]⊂G card Hi
M

where the metric on M is the lift of the metric on N . Since (G, H1 , H2 ) is a Sunada triple,
the right-hand side is independent of i . Therefore, e N1 (t ) = e N2 (t ) for all t > 0, and by
Exercise 6.2.2 it follows that N1 and N2 are isospectral.

It remains to show that there exist Sunada triples leading to non-isometric manifolds N1 and
N2 . Suppose that H1 and H2 are not conjugate in G (cf. Exercise 6.2.14) and M is the universal
cover of N . If the metric on N (which we are free to choose) is bumpy enough so that M has no
isometries that are not in G , then N1 and N2 are not isometric. Indeed, in that case any isometry
between N1 and N2 lifts to an isometry of M which conjugates H1 and H2 and hence does not
belong to the deck transformation group G . Moreover, there exist examples of Sunada triples
such that H1 and H2 are not isomorphic (see [Sun85], [Ros97] for details). In this case, N1 and
N2 have non-isomorphic fundamental groups, and are thus non-homeomorphic and hence non-
isometric.
While isospectral and non-isometric manifolds have been known prior to Sunada’s work (like
Milnor’s example described in the previous subsection), Sunada’s construction provided the first
“machine” to produce an abundance of such examples. Moreover, an adaptation of Sunada’s
method to planar domains has lead to a breakthrough paper [GorWebWol92] by C. Gordon,
D. Webb, and S. Wolpert, who have produced the first examples of isospectral non-isometric pla-
nar domains with either Dirichlet or Neumann boundary conditions, see Figure 6.1. We discuss
some related examples in the next subsection, and show that the algebraic techniques of Sunada
can be in fact replaced by a rather elementary idea called the transplantation of eigenfunctions,
originating in [Bér92].

Numerical Exercise 6.2.16

Compute the eigenvalues of the domains in Figure 6.1 to check their isospectrality, with
either Dirichlet or Neumann boundary conditions. Check if isospectrality still holds for
Robin conditions.
§6.2. Isospectral manifolds and domains 215

Figure 6.1: Isospectral domains of C. Gordon, D. Webb, and

S. Wolpert, each constructed of seven isosceles right-angled tri-
angles.

§6.2.4. Transplantation of eigenfunctions and mixed Dirichlet–Neumann isospectral-

ity
Let us first apply the transplantation technique to a simplified problem: find isospectral non-
isometric domains with mixed Dirichlet and Neumann boundary conditions. The possibility to
impose mixed conditions, as shown in [LevParPol06], provides more freedom and leads to simpler
examples, while capturing the main idea of the method.

Theorem 6.2.17: [LevParPol06]

The following two boundary value problems, see Figure 6.2, are isospectral:

(i) A unit square Ω, with the Dirichlet condition imposed on three sides and the Neu-
mann condition on the remaining side.

(ii) A right isosceles triangle Ω

e with the Dirichlet condition imposed on the hypotenuse
of length 2 and on one of the sides, and the Neumann condition on the other side.

Figure 6.2: A pair of isospectral domains with mixed bound-

ary conditions. The solid lines denote the Dirichlet boundaries,
and the dashed lines denote the Neumann ones.
216 Chapter 6. Heat equation, spectral invariants, and isospectrality

Figure 6.3: The construction for the proof of Theorem

6.2.17.

Proof
It is convenient to position Ω and Ω
e as shown in Figure 6.3. Let K = Ω ∩ Ω e be the triangle
shown, with the vertical side denoted a , the horizontal side b , and the hypotenuse c , so
that
Ω = K ∪ c ∪ τc K , Ω
e = K ∪ a ∪ τa K ,

where τa and τc denote the mirror symmetries with respect to a and c .

Let u be some eigenfunction of the corresponding mixed problem on Ω. We represent
u as a pair of functions (u 1 , u 2 ) : K × K → R as follows:
(
u 1 (x), x ∈K,
u(x) =
u 2 (τc x), x ∈ τc K .

Note that since u is smooth inside Ω we have the matching conditions

u 1 |c = u 2 |c , ∂n u 1 |c = − ∂n u 2 |c . (6.2.9)

The minus sign appears in the second condition since reflections change the direction of
the normal (and therefore the sign of the normal derivative) to the opposite one. We also
have the boundary conditions

u 1 |a = ∂n u 2 |a = 0, u 1 |b = u 2 |b = 0. (6.2.10)

Let us now transplant u to Ω

e . Introduce a new pair of functions (v 1 , v 2 ) : K × K → R by
setting
v 1 (x) = u 2 (x) − u 1 (x), v 2 (x) = u 1 (x) + u 2 (x).
§6.2. Isospectral manifolds and domains 217

Consider the function v on Ω

e defined by
(
v 1 (x), x ∈K,
v(x) = (6.2.11)
v 2 (τa x), x ∈ τa K ,

and let us show that v is an eigenfunction of the corresponding mixed problem on Ω e . It is

easy to check that v satisfies the correct boundary conditions on ∂Ω
e . Indeed, we have

v|b = v 1 |b = (u 2 − u 1 )|b = 0, v|τa b = v 2 |b = (u 1 + u 2 )|b = 0,

v|c = v 1 |c = (u 2 − u 1 )|c = 0, ∂n v|τa c = ∂n v 2 |c = ∂n (u 1 + u 2 )|c = 0,

by (6.2.10) and (6.2.9).

Obviously, v satisfies the eigenvalue equation on Ω e \ a but we need to verify that it is
true on the whole domain Ω e . A nontrivial point here is that u extends smoothly across the
line of reflection a , cf. Remark 3.2.22. Recall that the function u 1 satisfies the Dirichlet
condition on a , and u 2 the Neumann condition. Therefore, by the reflection principle of
Proposition 3.2.20, u 1 reflects antisymmetrically about a , and u 2 reflects symmetrically.
As a result, v 1 = u 2 − u 1 becomes u 2 + u 1 after the reflection, thus matching v 2 , and the
definition (6.2.11) therefore indeed produces a smooth eigenfunction on Ω e.
In order to complete the proof it remains to note that the operations used to construct
v out of u are invertible and linear, and hence there is a one-to-one correspondence be-
tween linearly independent eigenfunctions of the two problems, which therefore have the
same eigenvalue. Thus, the domains Ω1 and Ω2 with the boundary conditions specified
above are isospectral.

Exercise 6.2.18
Prove Theorem 6.2.17 by an explicit computation of the spectra for both problems using
separation of variables.

Remark 6.2.19
Alternative approaches to proving Theorem 6.2.17 and its generalisations can be found in
[LevParPol06] and [BanParBSh09].

Exercise 6.2.20
Show, in each case, that the following Zaremba problems are isospectral.
(a) Two domains shown in the top row of Figure 6.4, one simply connected and another
not simply connected.
(b) Two Zaremba problems on half-disk, shown in the second row of Figure 6.4, obtained
218 Chapter 6. Heat equation, spectral invariants, and isospectrality

from each other by swapping Dirichlet and Neumann boundary conditions. The
central arc where the boundary conditions change is a quarter-circle. This result, first
stated in [JakLNP06], plays a role in studying the first eigenvalue of the Laplace–
Beltrami operator on the Bolza surface mentioned in §5.3.3.

(c) Four Zaremba problems shown in the last two rows of Figure 6.4.

Figure 6.4: Two pairs and a quadruple of mixed isospectral

problems from Exercise 6.2.20. In each case, the solid lines de-
note the Dirichlet boundaries, and the dashed lines — the Neu-
mann ones.

Remark 6.2.21
One can show using the heat trace asymptotics that isospectral planar domains with mixed
boundary conditions must have the same area (corresponding to the coefficient a0 in the
heat trace expansion) and the same difference between the lengths of the Dirichlet and
Neumann parts of the boundary (this quantity corresponds to the heat trace coefficient
§6.2. Isospectral manifolds and domains 219

a 1 of the mixed problem, see [NurRowShe19]). One can observe that this is indeed the
2
case in all the examples above. At the same time, it was shown in [vdBDryKap14, Exam-
ple 6] that isospectral problems on Figure 6.2 can be distinguished by their heat contents
(5.1.2) corresponding to the unit initial temperature distributions.

§6.2.5. Isospectral drums

Let us now apply the transplantation method to the case of pure Dirichlet (or Neumann) bound-
ary conditions. We start with the following simple example, where the isospectral regions are dis-
connected.

Example 6.2.22: [Cha95]

With either Neumann or Dirichlet boundary conditions, the disjoint union of a square
of side 1 and an isosceles right triangle of side 2 is isospectral
p to the disjoint union of a
1 × 2 rectangle and an isosceles right triangle of side 2, see Figure 6.5. This is essentially
a variation of the construction presented in Theorem 6.2.17.

Figure 6.5: Two disjoint isospectral regions from Example 6.2.22

As we have mentioned previously, the first examples of planar isospectral connected domains
were constructed in [GorWebWol92], see Figure 6.1. A bit later, a whole zoo of isospectral pairs
was produced using a similar approach in [BusCDS94]. In fact, one can find an underlying
Sunada triple behind each of those pairs. At the same time, in this case isospectrality can be also
verified directly using the elementary transplantation method.
The simplest example of isospectral domains constructed in [BusCDS94] is presented in Fig-
ure 6.6. These domains are called “warped propellers”, and we will denote them by Ω and Ω e.
13
Each of the warped propellers is a union of seven identical copies of the same given scalene
triangle14 , arranged in a particular manner; we will denote these copies by A j and Ae j , with j =
13 Strictly speaking, the interior of the closure of the union.
14 Some restrictions on the angles of this given triangle are required in order to avoid self-intersecting propellers.
220 Chapter 6. Heat equation, spectral invariants, and isospectrality

Figure 6.6: Two isospectral warped propellers Ω and Ω

e.

0, . . . , 6. To construct Ω, we start with the given triangle A 0 , enumerating its sides from one to
three. We then construct15 A j , j = 1, 2, 3, as

A j = τ0, j A 0 ,

where τm,n denotes the reflection with respect to the straight line containing the n th side of A m .
We do the same for Ae j , j = 1, 2, 3, starting from Ae0 = A 0 , so at this stage the propellers are iden-
tical. We preserve the enumeration of sides under reflections.16
We now construct the remaining triangles, numbered four to six, in two different ways. For
Ω, we set
A 4 = τ1,2 A 1 , A 5 = τ2,3 A 2 , A 6 = τ3,1 A 3 ,

whereas for Ω
e we reflect as

Ae4 = τ1,3 Ae1 , Ae5 = τ2,1 Ae2 , Ae6 = τ3,2 Ae3 .

Theorem 6.2.23: [BusCDS94]

The domains Ω and Ω
e are non-isometric and isospectral for both Dirichlet and Neumann
boundary conditions.

15 Our enumeration of triangles and other notation differ sometimes from those in [BusCDS94].
16 To help distinguishing the sides, the different sides of the original triangles and their reflections are marked in
different line styles in Figures 6.6 and 6.7.
§6.2. Isospectral manifolds and domains 221

Proof
Since the triangles are chosen to be scalene, it is easy to check that Ω and Ω e are not isomet-
ric.
We first give the proof of the isospectrality of the Dirichlet Laplacians on Ω and Ω e,
and will mention the modifications required in the Neumann case at the end. Let u be an
eigenfunction of the Dirichlet Laplacian on Ω. Similarly to what we have done in the proof
of Theorem 6.2.17, we identify u with a collection of seven functions u j : A 0 → R, where
u| A j = u j ◦κ j , and κ j : A j → A 0 is a unique (since triangles are scalene) isometry between
triangles, j = 0, . . . , 6, κ0 = Id. The functions u j satisfy some boundary and matching
conditions. Firstly, if a side of the triangle A j is part of the external boundary ∂Ω , then on
that side we have u j = 0. Secondly, if two triangles A j and A k are reflections of each other
across a common side, then on that side
u j = uk and ∂n u j = −∂n u k .
We now describe the transplantation of the eigenfunction u from Ω to an eigenfunc-
tion v on Ω
e . We once more identify v with a collection of seven functions v j : Ae0 → R,
where
v| Ae j = v j ◦ κ
ej , (6.2.12)
and κe j : Ae j → Ae0 is a unique isometry between triangles, j = 0, . . . , 6, κ
e 0 = Id. We assume
for simplicity that the propellers are positioned in such a way that Ae0 = A 0 .
We start by assigning
v 0 = u1 + u2 + u3 . (6.2.13)
We now have to “propagate” this eigenfunction across the boundary of the triangles in the
following way. We start by reflecting (6.2.13) across the joint side 1 of Ae0 and Ae1 . We note
that on Ω, u 1 smoothly matches u 0 across side 1 and u 3 smoothly matches u 6 across the
common side 1 of triangles A 3 and A 6 . Finally, side 1 of the triangle A 2 is a part of the
exterior boundary of ∂Ω, thus by the reflection principle of Proposition 3.2.20, u 2 reflects
antisymmetrically across side 1 and becomes −u 2 . We therefore assign
v 1 = u0 − u2 + u6 , (6.2.14)
see Figure 6.7.
We now reflect (6.2.13) across the joint side 2 of Ae0 and Ae2 . In the same manner, u 1
smoothly reflects to u 4 across the joint side 2 of triangles A 1 and A 4 , u 2 smoothly reflects
to u 0 across the joint side 2 of triangles A 2 and A 0 , and since side 2 is an exterior side of
triangle A 3 , u 3 smoothly reflects to −u 3 across this side. We therefore assign
v 2 = u4 + u0 − u3 . (6.2.15)
Continuing in the same manner, we further obtain
v 3 = −u 1 + u 5 + u 0 , v 4 = u3 − u5 − u6 ,
(6.2.16)
v 5 = −u 4 + u 1 − u 6 , v 6 = −u 4 − u 5 + u 2 ,
222 Chapter 6. Heat equation, spectral invariants, and isospectrality

see Figure 6.7.

By construction, the resulting function v defined by (6.2.12)–(6.2.16) satisfies the eigen-
value equation and is smooth in Ω e . It remains to verify that it satisfies the Dirichlet bound-
ary condition on ∂Ω e . This is done triangle-by-triangle. Let us look, for example, at the
triangle A 4 which has two exterior sides: side 1 and side 2. Recalling the definition of v 4
e
in (6.2.16), we observe that on side 1 we have u 3 = u 6 since they match across this side;
we also have on this side u 5 = 0, since side 1 is an exterior side of the triangle A 5 . Thus
v 4 = u 3 − u 5 − u 6 = 0 on side 1. Similarly, on side 2 we have u 3 = u 5 = u 6 = 0 since it is
an exterior side for all three triangles A 3 , A 5 , and A 6 , and therefore v 4 = 0 on side 2. The
remaining triangles and their exterior sides are checked similarly.
Thus, our transplantation u 7→ v indeed generates an eigenfunction v of the Dirichlet
Laplacian on Ω e corresponding to the same eigenvalue as u . Moreover, as in the proof of
Theorem 6.2.17, the operator u 7→ v is linear and invertible, and hence we obtain that Ω
and Ω e are indeed Dirichlet isospectral.
For the Neumann boundary conditions the argument is the same, but instead of re-
flecting the functions u j antisymmetrically across the sides of the triangle A j which lie on
the boundary of Ω, we apply the symmetric reflection. As a result, all minuses in formulae
(6.2.14)–(6.2.16) and in Figure 6.7 should be replaced by pluses. Verifying that the resulting
function v satisfies the Neumann conditions on ∂Ω e is straightforward.

The starting transplantation (6.2.13) used in the proof of Theorem 6.2.23 above is not unique.

Exercise 6.2.24
Give another proof of this theorem by choosing a different starting transplantation de-
fined by
v 0 = u0 + u4 + u5 + u6 . (6.2.17)
Show that any non-trivial linear combination of the transplantations defined by (6.2.13)
and (6.2.17) is also a transplantation.

Note that the transplantation method uses in an essential way the fact that the boundary
conditions on each part of the boundary are either Dirichlet or Neumann. In particular, it does
not work for the Robin eigenvalue problem, cf. Exercise 6.2.16.

Open Problem 6.2.25

Do there exist Robin isospectral planar domains with a non-zero Robin parameter?

A similar question is also open for the Steklov problem that will be considered in Chapter 7.
Some higher-dimensional examples of Robin and Steklov isospectral manifolds can be found in
[GorHerWeb21].
§6.2. Isospectral manifolds and domains 223

Figure 6.7: A transplantation of a Dirichlet eigenfunction u on a warped propeller Ω

to a Dirichlet eigenfunction v on Ω e . The number j inside the triangle A j is a shorthand
for writing u| A j = u j ◦ κ j . The expression of the form ±l ± m ± n inside the triangle Ae j
is a shorthand for writing v| Ae j = ±u l ◦ κ
el ± um ◦ κ
em ± un ◦ κ
e n . For a transplantation of a
Neumann eigenfunction, replace all minuses by pluses.

§6.2.6. Spectral rigidity

In this subsection we discuss some results in the opposite direction to isospectrality. Namely, we
would like to understand which manifolds and domains are uniquely determined (in an appropri-
ate sense) by their spectra. This is an active area of research, and there is still very little known on
this subject. For example, while we have seen in the previous subsection that there exist isospec-
tral non-isometric planar domains, all known examples of isospectral pairs are non-smooth and
non-convex.

Open Problem 6.2.26

(i) Do there exist smooth Dirichlet (or Neumann) isospectral non-isometric planar do-
mains? (In the Dirichlet case, this is precisely the question posed in [Kac66].)

(ii) Do there exist Dirichlet (or Neumann) isospectral non-isometric convex planar do-
mains?

Let us discuss some results in the negative direction.

224 Chapter 6. Heat equation, spectral invariants, and isospectrality

Theorem 6.2.27
Let Ω ⊂ Rd be a bounded Lipschitz domain, and suppose that the Dirichlet (respectively,
Neumann) spectrum of Ω coincides with the Dirichlet (respectively, Neumann) spectrum
of a ball B ⊂ Rd . Then Ω coincides with B up to a rigid motion.

Proof
Consider the Dirichlet case first. It follows from Weyl’s law that Vol(Ω) = Vol(B ). Putting
this together with the equality λ1 (Ω) = λ1 (B ), and recalling that the equality in Faber–
Krahn’s theorem is attained among Lipschitz domains only for a ball (see Remark 5.1.14),
it follows that Ω is a ball of the same volume as B = Ω∗ .
The argument in the Neumann case is identical, with the equality in the Faber–Krahn
inequality replaced by the equality in the Szegő–Weinberger inequality, see Theorem 5.3.2.

Remark 6.2.28
As follows from the Ashbaugh–Benguria Theorem 5.4.7, the ratio between the first two
Dirichlet eigenvalues attains its maximum if and only if the domain is a ball. Therefore, in
the Dirichlet case, the ball is uniquely determined by only two lowest eigenvalues. An ana-
logue of this result in the Neumann case is false in dimensions n ≥ 3; in two dimensions,
it is not known whether a disk is uniquely determined by any finite part of its Neumann
spectrum.

Remark 6.2.29
One can alternatively prove Theorem 6.2.27 using the heat trace asymptotics (see [Bro93]
for the two-term heat trace expansion on Lipschitz domains) and the classical isoperimet-
ric inequality. Indeed, the heat trace coefficients a0 and a 1 determine the volumes of Ω
2
and of ∂Ω, and the equality in the isoperimetric inequality is attained among Lipschitz
domains if and only if the domain is a ball.

Beyond Theorem 6.2.27, rather little is known about domains which are spectrally deter-
mined in full generality. Some important advances have been achieved in the class of real ana-
lytic domains satisfying certain symmetry assumptions, see, for example, [Zel09]. To illustrate
how difficult the questions on spectral rigidity are, let us note that it is unknown whether any
ellipse is spectrally determined among all smooth planar domains. Recently, this has been shown
in [HezZel22] for ellipses of small eccentricity (i.e., that are close to a disk) using a highly sophisti-
cated machinery coming from billiard dynamics developed in [KalSor18], [AviDSiKal16]. As we
have mentioned earlier in Remark 6.2.4, there is a deep connection between the Laplace spectrum
and the dynamics of the geodesic (or billiard) flow. In particular, the Laplace spectrum contains
a lot information about the length spectrum, i.e. the set of lengths of closed trajectories, which in
some cases allows control of the geometry.
§6.2. Isospectral manifolds and domains 225

Consider also a “local” version of Open Problem 6.2.26 which was formulated by P. Sarnak.

Conjecture 6.2.30: [Sar90]

There exist no non-isometric isospectral continuous deformations of smooth planar do-
mains.

In other words, the claim is that all isospectral pairs are “isolated”. For domains close to a
disk, some progress on this conjecture and its dynamical counterpart, for which isospectrality is
understood in the sense of the length spectrum, has been obtained in [DeSKalWei17]. The best
known general result in this direction is the compactness in C ∞ topology of the set of Dirich-
let isospectral planar domains [OsgPhiSar88]. In the same paper, a similar compactness result
was also obtained for Riemannian metrics on closed surfaces. Interestingly enough, the proof in
[OsgPhiSar88] uses a certain property of the heat trace coefficients. Another related result in the
Riemannian setting states that closed negatively curved manifolds do not admit non-isometric
isospectral deformations. It was proved in [GuiKaz80] in dimension two, and in [CroSha98] in
arbitrary dimensions.
Let us conclude this chapter by the following interesting result obtained by S. Tanno [Tan73].
It uses the explicit expressions for the heat trace coefficients a1 , a2 and a3 of a closed Riemannian
manifold.

Theorem 6.2.31: [Tan73]

¡ ¢
Let M
³ be a closed´ Riemannian manifold of dimension d ≤ 6 with Spec −∆(M ,g ) =
Spec −∆(Sd ,g 0 ) , where g 0 is a round metric. Then (M , g ) is isometric to (Sd , g 0 ).

In dimension d ≥ 7, the geometric information contained in the first three heat invariants
becomes insufficient to prove the result of Theorem 6.2.31.

Open Problem 6.2.32

Is a round sphere uniquely determined by its Laplace–Beltrami spectrum among all com-
pact closed Riemannian manifolds in any dimension?
 CHAPTER 7
The Steklov problem and the
Dirichlet-to-Neumann map
In this chapter, we focus on the spectral geometry of the Steklov
eigenvalue problem and the Dirichlet-to-Neumann map. We state
the variational principle for the Steklov spectrum and prove the
Hersch–Payne–Schiffer inequalities for Steklov eigenvalues on
simply connected planar domains. We also use the
Hörmander–Pohozhaev identity to investigate the link between the
Dirichlet-to-Neumann map and the boundary Laplacian. As an
application, we derive the spectral asymptotics for the Steklov
problem on smooth Riemannian manifolds with boundary. We also
discuss the asymptotics of Steklov eigenvalues on planar domains
with corners, as well as the spectrum of the Dirichlet-to-Neumann
map for the Helmholtz equation.

§7.1. The Steklov eigenvalue problem

§7.1.1. Definition and variational principle

Let Ω be a bounded domain in a complete Riemannian manifold of dimension d ≥ 2. This in-
cludes bounded Euclidean domains and compact Riemannian manifolds with boundary. We de-
note the boundary of Ω by M = ∂Ω and assume that M is at least Lipschitz. The Steklov eigenvalue
problem on Ω is stated as follows,
(
∆U = 0 in Ω,
(7.1.1)
∂n U = σU on M .

Note that, unlike the Dirichlet and Neumann problems, the spectral parameter σ for the Steklov
problem is in the boundary condition. Sometimes, a more general Steklov-type boundary condi-

227
 228 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

tion is considered,
∂n U = σρU , on M , (7.1.2)
where L ∞ (M ) ∋ ρ ≥ 0 is a non-zero weight function.
The Steklov problem was introduced by Vladimir Steklov at the turn of the twentieth century,
see [KuzKKNPPS14] for a historical overview. It arises in various contexts, in particular, in inverse
problems, hydrodynamics and differential geometry. Some of these applications will be discussed
later on. We can alternatively interpret the Steklov eigenvalue problem as a spectral problem for
the Dirichlet-to-Neumann map D0 defined in the following way. Let u ∈ H 1/2 (M ), and let us
consider the non-homogeneous Dirichlet problem
(
Vladimir Andreevich ∆U = 0 in Ω,
(7.1.3)
Steklov (or Stekloff) U =u on M .
(1864 – 1926)
This problem has a unique (weak) solution U ∈ H 1 (Ω), see, e.g., [McL00, Theorem 4.10]. We
will call this solution the harmonic extension of u into Ω, and denote it by

U = E 0 u.

The Dirichlet-to-Neumann map for the Laplacian,

D0 : H 1/2 (M ) → H −1/2 (M ),

is defined as a linear operator D0 : u 7→ (∂n U )|M = (∂n (E 0 u))|M , which maps the boundary
Dirichlet datum of a harmonic function U into its Neumann datum. Here, we define the normal
derivative ∂n U ∈ H −1/2 (M ) by the relation
ˆ ˆ
(∂n U )v ds = 〈∇U , ∇V 〉 dx
M Ω

for every V ∈ H 1 (Ω) such that ∆V ∈ L 2 (Ω), where v := V |M ∈ H 1/2 (M ), see [ChWGLS12, p.
280].
Note that the operator D0 is non-local, and thus is not differential. If the boundary M is
smooth, then D0 is an elliptic self-adjoint pseudodifferential operator of order one. Its principal
symbol is given by |ξ|, which is the square root of the principal symbol of the boundary Laplacian
p
−∆M . The close link between D0 and −∆M will be particularly important for spectral asymp-
totics; see also Remark 7.1.5.

Remark 7.1.1
It is customary to call the function U ̸= 0 in (7.1.1) an eigenfunction of the Steklov prob-
lem corresponding to an eigenvalue σ. At the same time, an eigenfunction of the cor-
responding Dirichlet-to-Neumann map D0 (which acts on the functions defined on the
boundary) is U |M .
§7.1. The Steklov eigenvalue problem 229

Let
H 0 (Ω) := {U ∈ H 1 (Ω) : ∆U = 0} = E 0 u : u ∈ H 1/2 (M )
© ª
(7.1.4)
be the subspace of harmonic functions in H 1 (Ω). If U ∈ H 0 (Ω) satisfies (7.1.1), i.e. it is a Steklov
eigenfunction, then by Green’s formula we get

(∇U , ∇V )L 2 (Ω) = (−∆U ,V )L 2 (Ω) + (∂n U ,V )L 2 (M ) = σ(U ,V )L 2 (M )

for any V ∈ H 1 (Ω), since ∆U = 0. The weak spectral problem

(∇U , ∇V )L 2 (Ω) = σ(U ,V )L 2 (M ) for all V ∈ H 1 (Ω) (7.1.5)

is a weak version of the Steklov problem (7.1.1). Any weak eigenfunction U ∈ H 1 (Ω) of (7.1.5)
automatically belongs to H 0 (Ω) and is therefore harmonic, see e. g. [AreMaz12].
Using a similar approach to that in §2.1, one can show that the spectrum of the Steklov prob-
lem (or of the Dirichlet-to-Neumann map D0 ) is discrete provided that the composition of the
trace map and the embedding H 1 (Ω) → H 1/2 (M ) ,→ L 2 (M ) is compact. This condition will be
assumed throughout this chapter. It is true, for instance, if Ω has Lipschitz boundary M , in which
case the trace map is continuous and the embedding is compact (see, for example, [AreMaz12]).
Moreover, taking in (7.1.5) V = U , we immediately deduce that the eigenvalues of the Steklov
problem are non-negative. We denote the Steklov eigenvalues by

0 = σ1 = σ1 (Ω) < σ2 = σ2 (Ω) ≤ · · · ↗ +∞,

where the eigenfunction corresponding to σ1 = 0 is constant, as for the Neumann boundary

conditions. The eigenfunctions u j = U j |M of the Dirichlet-to-Neumann map (which coincide
with the boundary traces of the Steklov eigenfunctions U j ) form an orthogonal basis in L 2 (M ).

Exercise 7.1.2
Let Ω ⊂ Rd be a bounded domain, and let Ωa be its homothety with a coefficient a > 0.
Show that σk (Ωa ) = a1 σk (Ω), cf. Lemma 2.1.30.

Example 7.1.3: The Steklov eigenvalues of the unit disk

The Steklov eigenvalues of the unit disk D are given by

σ1 (D) = 0, σ2k (D) = σ2k+1 (D) = k, k ∈ N.

The eigenfunction corresponding to σ1 = 0 is a constant function, and the eigenspace

corresponding to σ2k = σ2k+1 = k is spanned by the functions r k sin kθ and r k cos kθ ,
written in polar coordinates (r, θ). Indeed, recall that

∂2 1 ∂ 1 ∂2
∆= + + ,
∂r 2 r ∂r r 2 ∂θ 2
230 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

and it is easy to see that all these functions are harmonic. This can alternatively be seen
from the fact that these functions are just the real and imaginary parts of holomorphic
functions z k , where z = r eiθ = x + iy . We note that the eigenspace corresponding to
σ2 = σ3 = 1 is spanned by the Cartesian coordinate functions x and y , cf. Exercise 1.2.3
for a basis of the first eigenspace of the Laplace–Beltrami operator on the round sphere.
Moreover, since the normal derivative on the boundary coincides with the partial
derivative with respect to r ,

∂³ k
µ ´¶¯¯ ³ ´¯
r sin kθ ¯¯ = k r k sin kθ ¯ ,
¯
∂r r =1
¶¯r =1
∂ k
µ ³ ´ ³ ´¯
= k r k cos kθ ¯ .
¯
r cos kθ ¯¯
¯
∂r r =1 r =1

There are no other eigenvalues as the boundary traces of the Steklov eigenfunctions

{1, sin θ, cos θ, . . . , sin kθ, cos kθ, . . . }

form a basis in L 2 (M ) = L 2 (S1 ).

Remark 7.1.4: Steklov–Robin duality

It is easily seen that σ is a Steklov eigenvalue if and only if 0 is an eigenvalue of the Robin
Laplacian −∆R,−σ ; moreover, the dimensions of the corresponding eigenspaces coincide.
See also Remark 3.1.19.

Throughout this chapter, let

0 = ν1 (M ) ≤ ν2 (M ) ≤ . . . ,

denote the eigenvalues of the Laplace–Beltrami operator −∆M on the boundary M = ∂Ω, assum-
ing that this boundary is sufficiently smooth.17

Remark 7.1.5
Note that σ2k (D) = νk (S1 ), k ∈ N. Moreover, if Uk are the Steklov eigenfunctions on D,
then u k = Uk |S1 are the Laplace–Beltrami eigenfunctions on S1 .
Let us mention as well that the Steklov eigenfunctions Uk behave as r k for r < 1, i.e.,
they decay rapidly in the interior. This decay is a general feature of Steklov eigenfunctions,
see [HisLut01, Theorem 1.1].

17 This enumeration of eigenvalues differs from the standard one used in the rest of the book, cf. footnote on page
180. In terms of our usual notation, νk (M ) = λk−1 (M ), k ∈ N.
§7.1. The Steklov eigenvalue problem 231

Exercise 7.1.6
Calculate the Steklov eigenvalues and eigenfunctions of the unit ball Bd in Rd , and com-
pare the results with the Laplace–Beltrami eigenvalues and eigenfunctions of the round
sphere Sd −1 .

Exercise 7.1.7: Steklov problem on a generalised cylinder [ColElSGir11]

Let Σ be a closed Riemannian manifold. Let 0 = ν1 < ν2 ≤ . . . be its Laplace–Beltrami
spectrum, and let {u k } be the corresponding orthonormal basis of eigenfunctions satisfy-
ing −∆Σ u k = νk u k . Given any l > 0, consider a cylinder Ω = (−l , l ) × Σ ⊂ R × Σ. Show
that the Steklov spectrum of Ω is given by
1 p p p p
0, , νk tanh( νk l ), νk coth( νk l ), k ≥ 2,
l
and the corresponding eigenfunctions are
p p
1, t , cosh( νk t )u k (x), sinh( νk t )u k (x), t ∈ (−l , l ), x ∈ Σ.

Compare also with Exercise 4.3.17.

For the remainder of this subsection we assume for simplicity that Ω ⊂ Rd is a Euclidean
domain. The extension of the variational principles to the Riemannian case is essentially verbatim.
Let u ∈ H 1/2 (M ) = Dom(D0 ). The quadratic form of the Dirichlet-to-Neumann map is
given by
(D0 u, u)L 2 (M ) = (∂n U , u)L 2 (M ) = ∥∇U ∥2L 2 (Ω) , (7.1.6)

where U = E 0 u ∈ H 0 (Ω). Therefore, the Rayleigh quotient for the Dirichlet-to-Neumann map
is given by
∥∇E 0 u∥2L 2 (Ω)
S
R [u] := , u ∈ H 1/2 (M ) \ {0}. (7.1.7)
∥u∥L 2 (M )

Using (7.1.7) and arguing in the same way as in §3.1, we obtain the following variational char-
acterisation of the Steklov eigenvalues:

∥∇E 0 u∥2L 2 (Ω) ∥∇U ∥2L 2 (Ω)

σk = min max = min max , k ∈ N. (7.1.8)
L
f⊂H 1/2 (M ) u∈L
f\{0} ∥u∥2L 2 (M ) L ⊂H 0 (Ω) U ∈L ∥U |M ∥2L 2 (M )
dim L
f=k dim L =k U = ̸ 0

Note that in the first min-max of (7.1.8) the minimum is taken over subspaces L f of H 1/2 (M ),
and in the second one over subspaces L of the space H 0 (Ω) of harmonic functions. We can in
fact replace H 0 (Ω) there by the usual Sobolev space H 1 (Ω) but to show this we need
232 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

Proposition 7.1.8
Let Ω ⊂ Rd be a bounded open set. Then

H 1 (Ω) = H 0 (Ω) ⊕ H01 (Ω). (7.1.9)

The direct sum in (7.1.9) is not orthogonal, however

(∇U , ∇V )L 2 (Ω) = 0 for any U ∈ H 0 (Ω),V ∈ H01 (Ω). (7.1.10)

Proof
Let W ∈ H 1 (Ω). Set u = W |M , and let U = E 0 u ∈ H 0 (Ω) be the unique solution of
(7.1.3). Then V = W − U belongs to H01 (Ω) since V |M = 0. As H 0 (Ω) ∩ H01 (Ω) = {0},
(7.1.9) follows.
To prove (7.1.10), we integrate by parts:

(∇U , ∇V )L 2 (Ω) = (−∆U ,V )L 2 (Ω) + (∂n U ,V )L 2 (M ) = 0,

since ∆U = 0 in Ω and V |M = 0.

We can now prove

Theorem 7.1.9: The variational principle for the Steklov problem

Let Ω be a bounded open set in Rd with a Lipschitz boundary M = ∂Ω, and let σk (Ω) be
the eigenvalues of the Steklov problem on Ω. Thena

∥∇W ∥2L 2 (Ω)

σk (Ω) = min max , k ∈ N. (7.1.11)
L ⊂H 1 (Ω) W ∈L ∥W |M ∥2L 2 (M )
dim L =k W = ̸ 0

In particular,
∥∇W ∥2L 2 (Ω)
σ2 (Ω) = min .
1
´0̸=W ∈H (Ω) ∥W |M ∥2L 2 (M )
M
W |M ds=0

a In what follows, we use a convention Q = +∞ for Q > 0.

0

Proof
Using Proposition 7.1.8, we represent any W ∈ H 1 (Ω) as W = U + V , where U ∈ H 0 (Ω),
V ∈ H01 (Ω). We note that ∥(U +V )|M ∥2L 2 (M ) = ∥U |M ∥2L 2 (M ) . Moreover, by (7.1.10) and the
§7.1. The Steklov eigenvalue problem 233

variational principle for the principal Dirichlet eigenvalue on Ω we have

∥∇(U + V )∥2L 2 (Ω) = ∥∇U ∥2L 2 (Ω) + ∥∇V ∥2L 2 (Ω)

≥ ∥∇U ∥2L 2 (Ω) + λD 2
1 (Ω)∥V ∥L 2 (Ω) .

The minimisation procedure now requires taking V = 0, and thus (7.1.11) is equivalent to
(7.1.8).

§7.1.2. The sloshing problem. Steklov eigenvalues of a square

Similarly to Zaremba problems considered in §3.1.3, we will be also looking at the mixed Steklov–
Neumann–Dirichlet spectral problems, with the spectral parameter in the boundary conditions,
stated as follows. Let Ω ⊂ Rd be a bounded simply connected domain with a Lipschitz boundary
M = ∂Ω. We decompose M into a disjoint union M = S ⊔ WD ⊔ WN , where either of WD,N may
be empty, and consider the spectral problem

∆U = 0 in Ω,




∂ U = σU

on S ,
n
(7.1.12)


∂nU = 0 on WN ,

U =0 on WD .


Sir Horace Lamb

The problem (7.1.12) goes back to the important special case considered by H. Lamb and
(1849–1934)
A. G. Greenhill already in the 19th century [Lam93, Chapter 9], [Gre86], see also [FoxKut83,
LevPPS22a]. Let Ω ⊂ R2 be in the lower half-plane, let S be an interval of the real line, called the
sloshing surface, and let WD = ;, see Figure 7.1. Then (7.1.12) models small gravitational oscillations
of an ideal fluid in an infinite canal with the cross-section Ω and the walls W = WN , and is called the
sloshing problem. The square roots of sloshing eigenvalues σ are proportional to the frequencies
of the fluid oscillations, and the sloshing eigenfunctions U represent the fluid velocity potential.
We may equivalently consider the mixed problem (7.1.12) as an example of a Steklov problem
with a variable weight boundary condition (7.1.2), where we formally take

1

 on S ,
ρ= 0 on WN ,

on WD .

+∞

1
The weak statement of the mixed problem (7.1.12) is to find σ ∈ R and U ∈ H0,W D
(Ω) \ {0} Sir Alfred George
such that Greenhill
1
(∇U , ∇V )L 2 (Ω) = σ(U ,V )L 2 (S ) for all V ∈ H0,W D
(Ω). (1847–1927)

Similarly to the Steklov problem, the spectrum of (7.1.12) is discrete and non-negative, and the
eigenfunctions can be chosen so that their traces on S form an orthogonal basis in L 2 (S ).
234 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

Figure 7.1: Geometry of a sloshing problem.

The angles between the sloshing surface S and
the walls W (which will be convenient to denote
β
by α2 and 2 ) play a significant role in the asymp-
totics of the sloshing eigenvalues, see §7.3.5.

Exercise 7.1.10
Let Ω be a rectangle (0, 1) × (−h, 0), h > 0, and let S = (0, 1). Find the eigenvalues and
eigenfunctions of (7.1.12) assuming either the Neumann or the Dirichlet boundary con-
ditions on the rest of the boundary.

We will now use the properties of some mixed Steklov–Neumann–Dirichlet problems to find
the Steklov spectrum of a square, following [GirPol17]. Let Ω = (−1, 1)2 ⊂ R2 be a square of side
2. Looking for the eigenfunctions of the Steklov problem on Ω using separation of variables,
we easily obtain the following eigenfunctions, and the equations for the separation parameter κ,
which is assumed to be positive; the eigenvalues are then easily expressed in terms of the positive
solutions of the corresponding equations, see Table 7.1 and Figure 7.2.
It remains to prove that there are no other eigenvalues. To do so, it is sufficient to demonstrate
j
that the traces of the eigenfunctions U 0 , U 1 , Uκ , j = 2, . . . , 9, form a basis in L 2 (∂Ω). We observe
that the Steklov problem on the square is symmetric with respect to the two diagonals {(x, y) :
x = ±y}. Reasoning as in the proof of the symmetry decomposition (3.2.7) for the Dirichlet
Laplacian, we obtain that the Steklov problem on the square decomposes into the union of four
mixed Neumann–Steklov, Dirichlet–Steklov,p or Neumann–Dirichlet–Steklov problems on an
isosceles right-angled triangle of side 2, with the Steklov condition on the hypothenuse, see
Figure 7.3.
We can now identify the eigenfunctions of the Steklov problem given in the first column of
Table 7.1, after transformations of the basis, with each of the mixed problems from Figure 7.3, see
Table 7.2.
Consider now the mixed problem I, which is in fact a sloshing problem. To ensure that we
have encountered all of its eigenvalues it is enough to demonstrate that the traces on S = (−1, 1)×
{1} of the corresponding Steklov eigenfunctions

U 0 ,U 1 ,Uκ2 +Uκ3 ,Uκ8 +Uκ9 , (7.1.13)

§7.1. The Steklov eigenvalue problem 235

Multi-
Eigenfunction Equation for κ Eigenvalue σ
plicity
U 0 := 1 0 1
U 1 := x y 1 1
Uκ2 := cos(κx) cosh(κy)
tan κ + tanh κ = 0 κ tanh κ 2
Uκ3 := cosh(κx) cos(κy)
Uκ4 := sin(κx) cosh(κy)
tan κ − coth κ = 0 κ tanh κ 2
Uκ5 := cosh(κx) sin(κy)
Uκ6 := cos(κx) sinh(κy)
tan κ + coth κ = 0 κ coth κ 2
Uκ7 := sinh(κx) cos(κy)
Uκ8 := sin(κx) sinh(κy)
tan κ − tanh κ = 0 κ coth κ 2
Uκ9 := sinh(κx) sin(κy)

Table 7.1: The Steklov eigenfunctions and eigenvalues of the square (−1, 1)2
obtained by the separation of variables.

Figure 7.2: Equations for the Steklov eigenvalues of the square: the κ coordinate
of each intersection of a dashed curve with one of the solid curves for κ > 0 corre-
sponds to a double Steklov eigenvalue of the square.

selected from Table 7.2, form an orthogonal basis in L 2 (S ). To do so, we use the following result
which was already known to Lamb, see also [FoxKut83], [LevPPS22a]: the traces on S of the
eigenfunctions of the sloshing problem I coincide with the eigenfunctions of the one-dimensional
236 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

Figure 7.3: Decomposition of the Steklov problem on the

square into four mixed problems on isosceles right-angled tri-
angles.

The Steklov eigenfunction The corresponding mixed problem

0
U Mixed problem I
U1 Mixed problem I
Uκ2 +Uκ3 Mixed problem I
Uκ2 −Uκ3 Mixed problem IV
Uκ4 +Uκ5 Mixed problem II
Uκ4 −Uκ5 Mixed problem III
Uκ6 +Uκ7 Mixed problem II
Uκ6 −Uκ7 Mixed problem III
Uκ8 +Uκ9 Mixed problem I
Uκ8 −Uκ9 Mixed problem IV

Table 7.2: The correspondence between the Steklov eigenfunctions from

Table 7.1 and the mixed problems from Figure 7.3.

vibrating free beam problem

(
f (iv) (x) = κ4 f (x), x ∈ (−1, 1),
′′ ′′′
(7.1.14)
f (±1) = f (±1) = 0,

where κ4 plays the role of the spectral parameter. It is now easy to verify that the traces of (7.1.13),
§7.1. The Steklov eigenvalue problem 237

that is, the functions

1, x,
cosh(1) cos(κx) + cos(1) cosh(κx) with tan κ + tanh κ = 0,
sinh(1) sin(κx) + sin(1) sinh(κx) with tan κ − tanh κ = 0,

are indeed the only eigenfunctions of (7.1.14). As (7.1.14) is a self-adjoint fourth order Sturm–
Liouville problem, its eigenfunctions form a basis in L 2 (S ) as required.
The mixed problems II–IV can be treated in the similar manner: they are again linked to
the boundary value spectral problems for the fourth derivative on the sloshing surface, the only
difference being the boundary conditions: at the ends adjoining the Dirichlet walls we need to
impose the Dirichlet conditions f = f ′ = 0 rather than the free ones as in (7.1.14). Combining all
the results together we confirm that Table 7.1 gives the full list of eigenvalues and eigenfunctions
of the Steklov problem on (−1, 1)2 .

Exercise 7.1.11
Using Table 7.1 and Figure 7.2, show that asymptotically the Steklov eigenvalues of the
square (−1, 1)2 satisfy

1 π
µ ¶
σ4m−k = m − + O m −∞ ,
¡ ¢
k = 0, 1, 2, 3, as m → +∞.
2 2

Remark 7.1.12
A calculation of Steklov eigenvalues of rectangles and higher-dimensional boxes, and an
alternative proof of completeness of the set of eigenfunctions which uses the Steklov–
Robin duality mentioned in Remark 7.1.4, can be found in [GirLPS19].

§7.1.3. Isoperimetric inequalities for the Steklov eigenvalues

As was indicated in §7.1.1, the Steklov eigenvalue problem shares some common features with the
Neumann problem. Recall that, in two dimensions, the Neumann problem models vibrations of
a homogeneous free membrane. Similarly, the Steklov problem (7.1.1) can be viewed as a model
for a vibrating free membrane with all the weight uniformly distributed along the boundary (or
with a density ρ , if a more general boundary condition (7.1.2) is considered). It is therefore natural
to look for an analogue of the Szegő–Weinberger inequality (Theorem 5.3.2) in the Steklov case.
The following result was obtained by R. Weinstock in [Wei54], using a modification of Szegő’s
approach.

Theorem 7.1.13: Weinstock’s inequality [Wei54]

Let Ω ⊂ R2 be a bounded simply connected domain with a Lipschitz boundary of length
238 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

L(∂Ω) := Vol1 (∂Ω). Then

σ2 (Ω)L(∂Ω) ≤ 2π,
with the equality attained if and only if Ω is a disk.

Exercise 7.1.14
Prove Theorem 7.1.13 by adapting the proof of Hersch’s theorem (Theorem 5.3.8). Note
that the first nontrivial Steklov eigenfunctions of the disk are the coordinate functions
(see Example 7.1.3), similarly to the first nontrivial Laplace eigenfunctions on the round
sphere. For a solution, see [GirPol10a], as well as [FreLau20, §7] for details on the equality
case for general Lipschitz boundaries.

While Weinstock’s Theorem is a direct analogue of Szegő’s result, there are significant dif-
ferences between the isoperimetric inequalities for the Steklov and the Neumann eigenvalues. In
particular, one can observe that Weinstock’s inequality does not admit a generalisation to non-
simply connected planar domains.

Example 7.1.15
Using separation of variables, one can investigate the first Steklov eigenvalues and eigen-
functions of an annulus A ε := D \B ε2 , see [GirPol17, Example 4.2.5]. In particular, if ε > 0
is small enough, then σ2 (A ε )L(∂A ε ) > 2π.

Remark 7.1.16: Non-simply connected planar domains

The previous example indicates that an appropriate perforation of a domain increases the
first nontrivial Steklov eigenvalue. This is indeed the case: as was shown in [GirKarLag21],
there exists a sequence of planar domains Ωk , with the number of boundary components
of Ωk tending to infinity as k → ∞, and such that σ2 (Ωk )L(Ωk ) → 8π. Moreover, this is
the maximal possible value of the limit, since, as was shown in [Kok14],

σ2 (Ω)L(Ω) ≤ 8π (7.1.15)

on any surface with boundary. The proof of (7.1.15) uses Hersch’s estimate, which explains
why the constant on the right-hand side of (7.1.15) is precisely the same as in (5.3.11).

Remark 7.1.17: Higher dimensions

Weinstock’s theorem does not admit a straightforward generalisation to higher dimen-
sions. However, among convex domains of given boundary volume, the ball maximises
the first nonzero Steklov eigenvalue [BucFNT21]. One can also use a different normalisa-
§7.1. The Steklov eigenvalue problem 239

tion: fix the volume of the domain itself, rather than of its boundary. F. Brock has shown
in [Bro01] that the ball maximises σ2 among all Euclidean domains of given volume. Note
that for simply connected planar domains this result is an easy consequence of Theorem
7.1.13 and the classical isoperimetric inequality. It is also interesting to note that Brock’s
inequality is stable similarly to the Szegő–Weinberger inequality (see [BraDeP17, §7.5]),
whereas Weinstock’s inequality is extremely unstable [BucNah21].

For the remainder of this subsection let us focus on simply connected planar domains. Sur-
prisingly enough, in this case one can obtain sharp isoperimetric inequalities for all Steklov eigen-
values.

Theorem 7.1.18: The Hersch–Payne–Schiffer inequality [HerPaySch75]

Let Ω ⊂ R2 be a bounded simply connected domain with a Lipschitz boundary. Then for
any p, q ≥ 1,
(
2 π2 (p + q − 1)2 if p + q is odd,
σp+1 σq+1 L(∂Ω) ≤ 2 2
(7.1.16)
π (p + q) if p + q is even.

In particular, with p = q = k , Menahem Max

Schiffer
σk+1 (Ω)L(∂Ω) ≤ 2πk (7.1.17) (1911–1997)

for all k ∈ N.

Remark 7.1.19
Note that (7.1.17) is precisely Weinstock’s inequality for k = 1. Moreover, it was shown
in [GirPol10b] that this inequality is sharp for any k . The equality is attained in the limit
by a union of k identical disks touching each other (cf. (5.3.25) and the corresponding
construction of maximisers for the Laplace eigenvalues on the sphere).

Before proceeding to the proof of Theorem 7.1.18, we give a brief reminder of some facts from
complex analysis.

Definition 7.1.20: Harmonic conjugate

Given a harmonic function U : Ω → R defined in a simply connected planar domain Ω,
its harmonic conjugate V : Ω → R is a harmonic function such that U +iV is holomorphic
in Ω.
240 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

Exercise 7.1.21
Show that for any harmonic function on a bounded simply connected planar domain, its
harmonic conjugate exists and is uniquely defined up to an additive constant.

Exercise 7.1.22
Let Ω be a bounded simply connected planar domain with Lipschitz boundary. Let u ∈
H 1 (Ω) be a harmonic function and v be its harmonic conjugate. It easily follows from
the Cauchy–Riemann equations that v ∈ H 1 (Ω). Show that

∂n u = −∂τ v (7.1.18)

on ∂Ω, where the normal derivative ∂n u and the tangential derivative of ∂τ v are under-
stood as elements of the Sobolev space H −1/2 (∂Ω). Hint: Use the Cauchy–Riemann
equations and Lemma 2.1.12. For a complete solution, see [BarBouLeb16, §6.2.1].

Proof of Theorem 7.1.18

Since Ω is simply connected, by the Riemann mapping theorem there exists a confor-
mal diffeomorphism ψ : Ω → D. Moreover, by Carathéodory’s theorem, ψ extends con-
tinuously to the boundary, see [Pom92, Chapter 2]. Let ds be the measure on ∂Ω and
dµ = ψ∗ ds be the push-forward measure on S1 = ∂D.
Let us introduce the “mass parameter”
ˆθ
m(θ) = dµ,
0
1
where θ is the coordinate on S . Then dµ = m ′ (θ)dθ , and
ˆ
m(2π) = dµ = L(∂Ω).
S1
Let h : R → R be a smooth periodic function (to be chosen later) with period L := L(∂Ω).
Let u : S1 → R be defined by
u(θ) = h(m(θ)).
The function u admits a unique harmonic extension to the disk, which we denote by U =
E0u.
Choosing´ an appropriate additive constant we can choose a harmonic conjugate V of
U such that S1 V dµ = 0. Let A, B : Ω → R be defined as A = U ◦ ψ and B = V ◦ ψ. By
conformal equivalence of the Dirichlet energy,
ˆ ˆ
|∇A| dx = |∇U |2 dx
2
(7.1.19)
Ω D
§7.1. The Steklov eigenvalue problem 241

and ˆ ˆ
|∇B |2 dx = |∇V |2 dx. (7.1.20)
Ω D

Let x = (x 1 , x 2 ). By the Cauchy–Riemann equations, we have

∂U ∂V ∂U ∂V
= , =− ,
∂x 1 ∂x 2 ∂x 2 ∂x 1

and hence |∇U |2 = |∇V |2 . Therefore, denoting

v(θ) := V |S1 ,

we get ˆ ˆ ˆ
2 2 ∂V
|∇U | dx = |∇V | dx = v dθ, (7.1.21)
∂r
D D S1

where the last equality follows from Green’s formula since V is harmonic. Putting together
(7.1.19), (7.1.20) and (7.1.21) yields
2
ˆ ˆ ˆ
   
∂V
|∇A|2 dx  |∇B |2 dx  =  v dθ  .
 
 (7.1.22)
∂r
Ω Ω S1

Recall that U = u on S1 . It follows from (7.1.18) that

∂V ¯¯
¯
= −u ′ (θ)
∂r ¯S1

as elements of H −1/2 (S1 ). Plugging the last equation into (7.1.22) and taking into account
that u ′ (θ) = h ′ (m(θ))m ′ (θ), we get
2
ˆ ˆ ˆ
   

|∇A|2 dx  |∇B |2 dx  =  v(θ)h ′ (m(θ))m ′ (θ) dθ 

  

Ω Ω S1
ˆ ˆ
  

≤ v(θ)2 dµ h ′ (m(θ))2 dµ,

  

S1 S1

where we have used the Cauchy–Schwarz inequality and the fact that dµ = m ′ (θ) dθ . At
the same time, by the definition of the push-forward measure dµ,
ˆ ˆ ˆ ˆ
a 2 ds = u 2 dµ, b 2 ds = v 2 dµ,
∂Ω S1 ∂Ω S1
242 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

where we set
a := A|∂Ω , b := B |∂Ω .
Therefore, it follows that the product of the Steklov Rayleigh quotients on Ω of A and B
can be estimated as
¡´ 2
¢¡´ ′ 2
¢
v dµ 1 h (m(θ)) dµ
R S [A]R S [B ] ≤ ¡´S ¢¡´S
1
¢ = R L [h],
2 2
S1 v dµ S1 h(m(θ)) dµ

where ´L ′ 2
0 h (η) dη
R L [h] = ´L
2
0 h(η) dη
is the usual Rayleigh quotient of h with respect to the Laplacian on the circle of length L
(to simplify notation below, we have introduced a new variable η := m(θ)
´ ).2 In other words,
we have reduced the problem to the boundary. Note that the term S1 v dµ cancels out,
which is the key feature of the method.
Let Φ j denote the eigenfunctions of the Steklov problem on Ω corresponding to eigen-
values σ j , j ∈ N, and chosen in such a way that their boundary traces ϕ j := Φ j |∂Ω form an
orthogonal basis in L 2 (∂Ω). We will now specify the choice of the function h . The main
idea is to use the resulting functions A and B as the test functions for σp+1 and σq+1 re-
spectively. Therefore, a should be orthogonal, in L 2 (∂Ω), to ϕ j with j = 1, . . . , p , and b
should be orthogonal to ϕ j with j = 1, . . . , q .
Let hk : R → R, k ∈ N, be the Laplace–Beltrami eigenfunctions on the circle of length
L , extended by periodicity,
 ³ ´
cos 2πnη , if k = 2n + 1,
h k (η) = ³ L ´
sin 2πnη , if k = 2n,
L

¡ 2πn ¢2
where n ∈ N0 (we ignore the function h0 = 0). Clearly, R L [h2n ] = R L [h2n+1 ] = L .
Set N = p + q , and consider
N
X
u= ck uk ,
k=2

where c k ∈ R, and the functions u k : S1 → R are defined by u k (θ) = hk (m(θ)), k ∈ N.

The functions u k are dµ-orthogonal and hence linearly independent. Therefore, their
harmonic extensions Uk = E 0 u k onto the unit disk are also linearly independent, and so
are the harmonic conjugates Vk of Uk . Moreover,
´ since u 1 = 1, the functions u k are dµ-
orthogonal to constants for k ≥ 2, and hence ∂Ω ak ds = 0 for all k ≥´2, where ak = u k ◦ψ.
At the¡ same¢ time, by our normalisation of harmonic conjugates, ∂Ω bk ds = 0, where
b k = V k | S1 ◦ ψ .
§7.1. The Steklov eigenvalue problem 243

Set now
N
X N
X N
X
U= c k Uk , V= c k Vk , h= ck hk .
k=2 k=2 k=2

We have p − 1 + q − 1 = N − 2 orthogonality conditions on a = (U ◦ ψ)|∂Ω and b = (V ◦

ψ)|∂Ω , left to be satisfied, and there are N −1 coefficients to choose. Therefore, there exists
a nontrivial choice of the coefficients c k for k = 1, . . . , N such that all the orthogonality
conditions are fulfilled, and hence

σp+1 σq+1 ≤ R LS [A]R LS [B ] ≤ R L [h]

(
³ π ´2 (p + q − 1)2 , if p + q is odd,
≤ R L [h N ] =
L (p + q)2 , if p + q is even.

This completes the proof of the theorem.

Remark 7.1.23
It has been already mentioned in Remark 7.1.19 that the inequalities (7.1.16) are sharp for
p = q = k . It immediately follows that they are also sharp for p = k , q = k + 1. In
particular, σ2 σ3 L 2 ≤ 4π2 .

Remark 7.1.24
There exist various generalisations of the Hersch–Payne–Schiffer inequalities. In partic-
ular, for the Steklov problem with a weight ρ ≥ 0 in the boundary condition (7.1.2), the
inequalities (7.1.16) hold provided the perimeter is replaced by the “mass”
ˆ
L ρ (∂Ω) = ρ(s) ds.
∂Ω

One can also extend the inequalities (7.1.16) to arbitrary surfaces with boundary, see [Gir-
Pol12]. In particular, it was shown in [Kar18] that

σk L(∂Σ) ≤ 2π(k + γ + ℓ − 1),

where γ is the genus of the surface Σ and ℓ is the number of its boundary components.
244 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

§7.2. The Dirichlet-to-Neumann map and the boundary Laplacian

§7.2.1. Weyl’s law for Steklov eigenvalues

The goal of this section is to prove Weyl’s law for the Steklov eigenvalues of a bounded domain Ω,
or, equivalently, for the eigenvalues of the Dirichlet-to-Neumann map D0 acting on its boundary
M = ∂Ω. Let
D0
N S (σ) = NΩS (σ) = N (σ) := #{ j : σ j (Ω) ≤ σ}
be the counting function of the Steklov eigenvalues.

Theorem 7.2.1: Weyl’s law for Steklov eigenvalues

Let Ω be a bounded domain in a complete Riemannian manifold, and assume that ∂Ω =
M is smooth. Then the following asymptotic relation holds,
³ ´
N S (σ) = C d −1 Vol(M )σd −1 + o σd −1 as σ → +∞. (7.2.1)

ωd −1
Here, as before, C d −1 = (2π)d −1
denotes the Weyl constant, and ωd −1 is the volume of a
unit ball in Rd −1 .

The standard approach to establishing Theorem 7.2.1 uses the theory of pseudodifferential
operators, which is beyond the scope of this book. The key observation is that the principal sym-
bol of the operator D0 is precisely the square root of the principal symbol of the boundary Lapla-
p
cian −∆M on M . This implies that D0 and −∆M have similar eigenvalue asymptotics. Here
we take a different route which is based on rather elementary tools, and at the same time pro-
vides a more geometric way to understand the link between the Dirichlet-to-Neumann operator
and the boundary Laplacian. Our exposition mostly follows [GirKLP22], and is based on the so-
called Pohozhaev’s identity [Poh65] and its generalisations, which in turn is an application of the
method of multipliers going back to F. Rellich (see [ChWGLS12, p. 205] for a discussion), and to
an old unpublished work of L. Hörmander [Hör18] that was originally written in the 1950s (see
also [Hör54] where an identity similar to Pohozhaev’s has been obtained).
For simplicity, we will prove Theorem 7.2.1 in the Euclidean setting, and will outline the nec-
essary modifications for the Riemannian case, and some relaxations of the conditions of the the-
orem at the end, see Remark 7.2.11. We also note that for Euclidean domains, Weyl’s law was first
obtained by L. Sandgren in in [San55] using a different approach under the assumption that the
boundary is C 2 regular. Using heavier machinery, the result can be also proved for Euclidean
domains with piecewise C 1 boundaries [Agr06].

Franz Rellich Remark 7.2.2

(1906–1955)
The validity of Weyl’s law for the Steklov problem in a Lipschitz domain Ω ⊂ Rd has been
a long-standing open problem attributed to M. S. Agranovich. In the two-dimensional
case, it was proved in [KarLagPol22] using the theory of conformal mappings. While this
§7.2. The Dirichlet-to-Neumann map and the boundary Laplacian 245

book was in the final preparation stage, G. Rozenblum [Roz23] established Weyl’s law for
domains with Lipschitz boundary in any dimension.

§7.2.2. The Hörmander–Pohozhaev identities

We start with a reminder of some notions and identities from vector calculus. Let Ω ⊂ Rd be
an open set, and let a : Ω → R and A, B : Ω → Rd be a scalar function and vector fields on Ω,
respectively, which we assume to be sufficiently smooth.18 We denote by

∂A i
µ ¶
JacA :=
∂x j i , j =1,...,d

the Jacobian of A, and by

∂2 a
µ ¶
Hesa := Jac∇a =
∂x i ∂x j i , j =1,...,d

the Hessian of a . Additionally, for any linear operator (that is, a matrix) C acting in Rd , we will Stanislav Ivanovich
denote by C∗ its adjoint (that is, a transposed matrix), and by Pohozhaev
(1935–2014)
C[A, B] := 〈CA, B〉

its quadratic form.

Exercise 7.2.3
Prove the following identities:

div(aA) = 〈∇a, A〉 + a div A, (7.2.2)

∇(〈A, B〉) = Jac∗A B + Jac∗B A, (7.2.3)
∇ |∇a|2 = 2Hesa ∇a.
¡ ¢
(7.2.4)

Lars Valter
Let now state a useful Pohozhaev-type identity which has various applications, see [ColGirHas18, Hörmander
Lemma 20]. (1931–2012)

Theorem 7.2.4: Generalised Pohozhaev’s identity for the Laplacian

Let Ω ⊂ Rd be a bounded domain with smooth boundary M = ∂Ω. Let F be a smooth
vector field on Ω, let u ∈ H 1 (M ), and let U = E 0 u be the unique harmonic extension of

18 Throughout this chapter, we distinguish the vector fields by bold font, in particular the exterior normal vector on
the boundary of Ω will be denoted n.
246 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

u onto Ω. Then
ˆ ˆ
1
〈F, ∇U 〉∂n U ds − |∇U |2 〈F, n〉 ds
2
ˆM ˆ M
(7.2.5)
1
+ |∇U |2 div F dx − JacF [∇U , ∇U ] dx = 0.
2
Ω Ω

Proof
Since ∆U = div ∇U = 0 in Ω, using (7.2.2) and (7.2.3), we obtain

div(〈F, ∇U 〉∇U ) = 〈∇〈F, ∇U 〉, ∇U 〉 = JacF [∇U , ∇U ] + HesU [F, ∇U ].

(note that the Hessian of U is well-defined since U is harmonic). At the same time, using
(7.2.2) once more together with (7.2.4),
1 1
div |∇U |2 F = HesU [F, ∇U ] + |∇U |2 div F.
¡ ¢
2 2
Subtracting the second equality from the first one, we get
µ ¶
1 1
div 〈F, ∇U 〉∇U − |∇U | F = JacF [∇U , ∇U ] − |∇U |2 div F.
2
2 2

Finally, we integrate this identity over Ω and use the divergence theorem, noting that
(∇U )|M ∈ L 2 (M ) since we have assumed u = U |M ∈ H 1 (M ) (see [ChWGLS12, Theorem
A.5]).

We now make a choice of a vector field F in Theorem 7.2.4, leading to the following result,
which was originally obtained by L. Hörmander in the 1950s.

Theorem 7.2.5: Hörmander’s identity [Hör18]

Let Ω ⊂ Rd be a bounded domain with a smooth boundary M = ∂Ω. Let F be a smooth
vector field on Ω which on the boundary of Ω coincides with the exterior unit normal,
F|M = n. Let u ∈ H 1 (M ), and let U = E 0 u be the unique harmonic extension of u onto
Ω. Then

(D0 u, D0 u)L 2 (M ) − (−∆M u, u)L 2 (M )

ˆ
2JacF [∇U , ∇U ] − |∇U |2 div F dx. (7.2.6)
¡ ¢
=
Ω
§7.2. The Dirichlet-to-Neumann map and the boundary Laplacian 247

Proof
Using F|M = n and the definition of the Dirichlet-to-Neumann map D0 , we substitute
into (7.2.5) the following relations,

〈F|M , n〉 = 1, 〈F, ∇U 〉|M = ∂n U ,

|∇U |2 ¯ = |∇M u|2 + (∂n U )2 , D0 u = ∂n U ,
¯
M

and (7.2.6) then follows immediately taking into account the expression for the quadratic
form of the Laplace–Beltrami operator on M , (−∆M u, u)L 2 (M ) = (∇M u, ∇M u)L 2 (M ) .

§7.2.3. The Steklov spectrum and the spectrum of the boundary Laplacian
Theorem 7.2.5 almost immediately implies

Corollary 7.2.6
Let Ω ⊂ Rd be a bounded domain with a smooth boundary M = ∂Ω. Then there exists a
constant C > 0 such that for any u ∈ H 1 (M ),
¯(D0 u, D0 u) 2
¯ ¯
L (M ) − (−∆M u, u)L 2 (M ) ≤ C (D0 u, u)L 2 (M ) . (7.2.7)
¯

Proof
We note that the integrand in the right-hand side of (7.2.6) is a quadratic form in ∇U with
bounded coefficients, since the vector field F is smooth. Hence, there exists a constant
C > 0 such that
¯ˆ
¯ ¯
¯
¯ 2JacF [∇U , ∇U ] − |∇U |2 div F dx¯ ≤ C ∥∇U ∥2 2
¯ ¡ ¢ ¯
¯ ¯ L (Ω)
¯ ¯
Ω
= C (D0 u, u)L 2 (M ) .

Remark 7.2.7
In fact, the constant C appearing in the right-hand side of (7.2.7) may be chosen to depend
only on the geometry¢ of Ω in a small neighbourhood of M . To see this, we may choose
F(x) = ∇ d M (x)χ(x) , where d M (x) is a distance from x to the boundary, and χ(x) is a
¡

smooth cut-off function equal to one near M and zero outside a small neighbourhood of
M . Then F(x) satisfies the assumptions of Theorem 7.2.5, see [ProStu19, §5.3]. For explicit
expressions on C in terms of geometric characteristics of Ω and M see [ProStu19], [Xio18],
[ColGirHas18].

Corollary 7.2.6 already links, in a way, the Dirichlet-to-Neumann map and the Laplace-Beltrami
248 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

operator on M . We will now use it to compare the eigenvalues of this operator, using the following
abstract result, essentially due to L. Hörmander.

Proposition 7.2.8: [GirKLP22, Proposition 3.3], generalising [Hör18]

Let H be a Hilbert space with an inner product (·, ·)H . Let A , B be two non-negative
self-adjoint operators in H with discrete spectra Spec(A ) = {α1 ≤ α2 ≤ . . . } and
Spec(B) = {β1 ≤ β2 ≤ . . . } and the corresponding orthonormal bases of eigenfunctions
{a k }, {b k }. Assume additionally that a k ∈ Dom(B) and b k ∈ Dom(A 2 ), k ∈ N, where
the domains are understood in the sense of quadratic forms. Suppose that for some C > 0,

|(A u, A u)H − (Bu, u)H | ≤ C (A u, u)H

(7.2.8)
for all u ∈ D := Dom(B) ∩ Dom(A 2 ).

Then ¯ 2
¯α − βk ¯ ≤ C αk ,
¯
k (7.2.9)
and consequently q ¯
¯
¯αk − βk ¯ ≤ C
¯ ¯
(7.2.10)
for all k ∈ N, with the same constant C as in (7.2.8).

Proof
We note that (7.2.8) is equivalent to
(
(Bu, u)H ≤ (A u, A u)H +C (A u, u)H ,
(7.2.11)
(A u, A u)H −C (A u, u)H ≤ (Bu, u)H ,
and (7.2.9) is equivalent to
(
βk ≤ α2k +C αk ,
(7.2.12)
βk ≥ α2k −C αk .
From the variational principle for the eigenvalues of B and the first inequality in (7.2.11)
we have
(Bu, u)H
βk ≤ sup
0̸=u∈Vk ⊂Dom(B) (u, u)H
(7.2.13)
(A u, A u)H +C (A u, u)H
≤ sup
0̸=u∈Vk ⊂Dom(B) (u, u)H
for any subspace Vk with dimVk = k . Take Vk = Span{a1 , . . . , ak }. As for any u = c 1 a1 +
· · · + c k a k ∈ Vk with |c 1 |2 + · · · + |c k |2 = 1 we have due to orthogonality

(A u, A u)H +C (A u, u)H k
|c j |2 (α2j +C α j ) ≤ α2k +C αk ,
X
=
(u, u)H j =1
§7.2. The Dirichlet-to-Neumann map and the boundary Laplacian 249

the first inequality (7.2.12) follows immediately from (7.2.13).

We now prove the second inequality (7.2.12). Let K 0 := max{k ∈ N : αk ≤ C }. We
note that for k ≤ K 0 the second inequality (7.2.12) is automatically satisfied since in this
case βk ≥ 0 ≥ αk (αk −C ), so we need to consider only k > K 0 . We re-write the second
inequality (7.2.11) as
³ ´ C2
Afu, Afu ≤ (Bu, u)H + (u, u)H ,
H 4
2
where Af:= A − C2 . Let α
e2k denote the eigenvalues of Af enumerated in non-decreasing
¢2
order. We note that α e2k = αk − C2 for k > K 0 (this may not be the case for k ≤ K 0 but as
¡

mentioned above we can ignore these values of k ). Writing down the variational principle
e2k similarly to (7.2.13) and choosing a test subspace Vk = Span{b 1 , . . . , b k } leads in a
for α
similar manner to
2
C C2
µ ¶
e2k = αk −
α ≤ βk + ,
2 4
which gives the second inequality (7.2.12) after a simplification.
Finally, we note that (7.2.9) implies, for αk βk ̸= 0,
¯ q ¯ αk
¯αk − βk ¯ ≤ C p ≤ C,
¯ ¯
αk + βk

yielding (7.2.10). Note that αk = 0 implies βk = 0 by (7.2.12).

Using Proposition 7.2.8, we are now able to obtain a uniform bound comparing the Steklov
eigenvalues with the ones of the Laplace–Beltrami operator on the boundary.

Theorem 7.2.9
Let Ω ⊂ Rd be a bounded domain with a smooth boundary M = ∂Ω, and let σk , νk , k ∈
N, be the Steklov eigenvalues of Ω and the eigenvalues of the Laplace–Beltrami operator
on M , respectively. Then
¯σk − pνk ¯ ≤ C
¯ ¯
(7.2.14)
holds for all k ∈ N with the same constant C as in (7.2.7).

Proof
We apply Proposition 7.2.8 with A = D0 , B = −∆M , and therefore αk = σk , and βk = νk ,
taking into account Corollary 7.2.6 and choosing D = H 1 (M ).

We can now finish the proof of Theorem 7.2.1.

250 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

Proof of Theorem 7.2.1

It follows from Theorem 7.2.9 that

N M (σ −C )2 ≤ N S (σ) ≤ N M (σ2 +C σ),

¡ ¢
(7.2.15)

where N M (·) is the eigenvalue counting function of the Laplace–Beltrami operator on

M . Indeed, to prove the left inequality (7.2.15) we observe that if νk ≤ (σ − C )2 , then
p
σ ≥ νk +C ≥ σk by (7.2.14). To prove the right inequality (7.2.15), we note that if σk ≤ σ,
then νk ≤ σ2k + C σk ≤ σ2 + C σk , once more using (7.2.14). An application of Theorem
6.1.9 to both sides of (7.2.15) then yields the result.

Remark 7.2.10
We have stated Theorem 6.1.9 in the Riemannian setting but have proved it in the Eu-
clidean case only. The Riemannian argument goes through identically, with the only
modification required is in (7.2.5) where JacF [∇U , ∇U ] in the last integral should be re-
placed by (∇∇U , F)∇U , where ∇∇U denotes a covariant derivative in the direction ∇U , see
[GirKLP22] for details.

Remark 7.2.11
There exist various improvements and extensions of the results presented in this subsec-
tion. In particular, Theorem 7.2.4 can be proved verbatim under the assumption that Ω
has Lipschitz boundary and F is a Lipschitz vector field. Consequently, Theorem 7.2.5
holds if F is a Lipschitz vector field and Ω has C 1,1 boundary, so that the normal field
on the boundary is Lipschitz. As a result, the regularity assumptions in Theorem 7.2.1
can be significantly relaxed; moreover, the error term estimate in (7.2.1) can be improved
¡ d −2 ¢
to O σ using the sharp Weyl’s law for the boundary Laplacian. This improvement
holds for domains with C 2,α boundary for some α > 0 in arbitrary dimension, and for
domains with C 1,1 boundary in dimension two. We refer to [GirKLP22] for a detailed
exposition of these results.

Remark 7.2.12: The two-dimensional case

Let Ω be a smooth simply-connected planar domain. Then M = ∂Ω pis one-dimensional,
and hence locally isometric to a circle. Hence, by Remark 7.1.5, νk (M ) = σk (Ω⋆ ),
where Ω⋆ is a disk of the same perimeter as Ω. Therefore, in this case (7.2.14) yields
¯σk (Ω) − σk (Ω⋆ )¯ < C , k ∈ N.
¯ ¯

This bound admits a significant asymptotic improvement:

¯σk (Ω) − σk (Ω⋆ )¯ = o k −N ,
¯ ¯ ¡ ¢
(7.2.16)
§7.2. The Dirichlet-to-Neumann map and the boundary Laplacian 251

for any N > 0, see [Roz86], [Edw93a], which is proved using pseudodifferential tech-
niques. In particular, this implies that in this case the remainder estimate in (7.2.1) can
be replaced by o σ−N for any N > 0.
¡ ¢

We refer also to [GirPPS14] for a generalisation of (7.2.16) to arbitrary Riemannian

surfaces with boundary.

Remark 7.2.13: Steklov isospectrality

Similarly to Definition 6.2.1, we can say that two Riemannian manifolds with boundary,
or two Euclidean domains, are Steklov isospectral if their Steklov spectra coincide. For
some examples of Steklov isospectral manifolds see e.g. [GirPPS14]. It is immediately clear
from (7.2.1) that the volume Vol(M ) of the boundary M = ∂Ω of a complete Riemannian
manifold is a Steklov spectral invariant.
Interestingly enough, no examples of non-isometric Steklov isospectral planar do-
mains are presently known [GirPol17, Open problem 6]; we refer also to [Edw93b, Mal-
Sha15, JolSha14, JolSha18] for some related results and conjectures. At the same time,
Steklov spectral invariants of planar domains are also quite scarce — we know, in addition
to the perimeter, that if the boundary of Ω ⊂ R2 is smooth, its Steklov spectrum deter-
mines the number of boundary components of M = ∂Ω and their lengths [GirPPS14].
However, for smooth simply connected planar domains, extracting further geometric in-
formation from the Steklov problem is quite difficult. In part, the reason for that lies
in formula (7.2.16): any two smooth simply connected planar domains ΩI and ΩII of
the same perimeter will have Steklov eigenvalues which differ as |σk (ΩI ) − σk (ΩII )| =
O(k −∞ ) as k → ∞. As a result, no other spectral invariants except the perimeter can be
obtained from the eigenvalue asymptotics on the polynomial scale.
We will say that two (not necessarily smooth) planar domains ΩI and ΩII are asymp-
totically Steklov isospectral if

|σk (ΩI ) − σk (ΩII )| = o(1) as k → ∞.

We will consider some further examples of asymptotically Steklov isospectral planar do-
mains and corresponding Steklov spectral invariants in §7.3.6.
252 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

§7.3. Steklov spectra on domains with corners

§7.3.1. Asymptotics of Steklov eigenvalues for curvilinear polygons

In this section, we mostly follow [LevPPS22b]. Let P = P α,ℓ be a (simply connected) curvilinear
polygon in R2 with n vertices V1 , . . . ,Vn numbered clockwise, corresponding internal angles 0 <
α j < π at V j , and smooth sides I j of length ℓ j joining V j −1 and V j . Here, α = (α1 , . . . , αn ) ∈
(0, π)n , ℓ = (ℓ1 , . . . , ℓn ) ∈ Rn+ , and we will use cyclic subscript identification n + 1 ≡ 1. Our
choice of orientation ensures that an internal angle α j is measured from I j to I j +1 in the counter-
clockwise direction, as in Figure 7.4. The perimeter of P is L(∂P ) = L = ℓ1 + · · · + ℓn .

Figure 7.4: A curvilinear polygon.

We will give an improved asymptotics of the Steklov eigenvalues σm (P ) as m → +∞, which

takes into account not just the perimeter of a curvilinear polygon but also the lengths of individ-
ual sides and the angles between them. The philosophy behind this result is somewhat similar
to the principle of Theorem 7.2.9: we will compare (asymptotically only, and using a completely
different set of techniques) the Steklov eigenvalues of P to the eigenvalues of a particular “bound-
ary Laplacian” on M = ∂P . More precisely, the role of this boundary Laplacian is played here by
a certain quantum graph Laplacian (see [BerKuc13] and references therein for a comprehensive
spectral theory of quantum graphs).
Let us associate with the boundary of a curvilinear polygon P α,ℓ a cyclic metric graph Mα,ℓ
with n vertices V1 , . . . ,Vn and n edges I j (joining V j −1 and V j , with V0 identified with Vn ) of
length ℓ j , j = 1, . . . , n . Let s be the arc-length parameter on Mα,ℓ starting at V1 and going in the
clockwise direction, see Figure 7.5.
Consider the spectral problem for a quantum graph Laplacian on Mα,ℓ ,

d2 f
− = νf , (7.3.1)
ds 2
§7.3. Steklov spectra on domains with corners 253

Figure 7.5: A quantum graph.

with matching conditions

µ 2 ¶ µ 2 ¶
π π
sin f |V j +0 = cos f |V j −0 ,
4α j 4α j
µ 2 ¶ µ 2 ¶ (7.3.2)
π ′ π
cos f |V j +0 = sin f ′ |V j −0 .
4α j 4α j

Hereinafter at each vertex V j , j = 1, . . . , n , g |V j −0 and g |V j +0 denote the limiting values of a quan-

tity g (s) as s approaches the vertex V j from the left and from the right, respectively, in the direc-
tion of s .
d2 f
We will denote the operator f 7→ − ds 2 subject to matching conditions (7.3.2) by −∆M . It
is easy to check that −∆M is self-adjoint and non-negative. Therefore, its spectrum is given by a
sequence of non-negative real eigenvalues

0 ≤ ν1 ≤ ν2 ≤ . . . νm ≤ · · · ↗ +∞,

listed with multiplicity.

254 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

Remark 7.3.1
The eigenvalues νm also satisfy a standard variational principle: if
( µ 2 ¶ µ 2 ¶ )
n
M 1 π π
Dom(QM ) := f ∈ H (I j ) : sin f |V j +0 = cos f |V j −0
j =1 4α j 4α j

denotes the domain of the quadratic form

n
ˆ
QM [ f ] := ( f ′ (s))2 ds
X
j =1
Ij

of −∆M , then
QM [ f ]
νm = min max n ´
.
S⊂Dom(QM ) 0̸= f ∈S P 2 ds
dim S=m Ij ( f (s))
j =1

We now have

Theorem 7.3.2: Eigenvalue asymptotics for curvilinear polygons [LevPPS22b,

Theorem 1.4]
Let P = P α,ℓ be a curvilinear polygon defined above, let σm , m ∈ N, be its Steklov eigen-
values, and let νm , m ∈ N, be the eigenvalues of the associated quantum graph problem
(7.3.1), (7.3.2). Then there exists ε > 0 such that we have
p
σm = νm + O m −ε
¡ ¢
as m → +∞.

From now on, we will call the numbers19

p
τm := νm

the quasi-eigenvalues of the Steklov problem on P .

Theorem 7.3.2 immediately implies

Corollary 7.3.3
I II
Let P α,ℓ and P α,ℓ be two curvilinear polygons with the same angles α and the same side
lengths ℓ. Then there exists ε > 0 such that

σm P I − σm P II = O m −ε
¡ ¢ ¡ ¢ ¡ ¢
as m → +∞.

19 We emphasise that in [LevPPS22b] and a related paper [LevPPS22a], the Steklov eigenvalues are denoted by λ
(rather than σ) and the quasi-eigenvalues by σ (rather than τ).
§7.3. Steklov spectra on domains with corners 255

As it turns out, the Steklov quasi-eigenvalues τm can be determined as the roots of a particular
trigonometric function which depends only on the side lengths ℓ and angles α of the curvilinear
polygon P . To define this trigonometric function, we need to introduce some combinatorial
notation.
Let
Zn = {±1}n ,
and for a vector ζ = (ζ1 , . . . , ζn ) ∈ Zn with cyclic identification ζn+1 ≡ ζ1 , let

Ch(ζ) := { j ∈ {1, . . . , n} | ζ j ̸= ζ j +1 }

denote the set of indices of sign change in ζ, e.g.

Ch((1, 1, 1)) = ∅; Ch((−1, −1, 1, 1)) = {2, 4}.

Given a curvilinear polygon P α,ℓ , we now define the following trigonometric function in
real variable σ:
n π2
µ ¶
pζ cos(〈ℓ, ζ〉τ) −
X Y
F α,ℓ (τ) := sin , (7.3.3)
ζ∈Zn j =1 2α j
ζ1 =1

where µ 2 ¶
πY
pζ = pζ (α) := cos ,
j ∈Ch(ζ) 2α j
Q
and we assume the convention = 1.
;
We can now state

Theorem 7.3.4: [LevPPS22b, Theorem 2.16]

Let P α,ℓ be a curvilinear polygon. Then τ ≥ 0 is its quasi-eigenvalue if and only if it
is a root of the trigonometric function Fα,ℓ (τ). The multiplicity of a quasi-eigenvalue
τ > 0 coincides with its multiplicity as a root of (7.3.3), and the multiplicity of the quasi-
eigenvalue τ = 0 (if present) is half its multiplicity as a root of (7.3.3).

Theorem 7.3.4 is proved by a rather complicated but straightforward computation of the sec-
ular equation of the quantum graph problem (7.3.1), (7.3.2) using the methods of [KotSmi99,
p
KurNow10, BerKuc13, Ber17], which shows that Fα,ℓ ( νk ) = 0 with the same multiplicities as
in Theorem 7.3.4.

Example 7.3.5
¡ p
Let P be the isosceles right-angled triangle with α = π4 , π4 , π2 and ℓ = 1, 2, 1 . For
¡ ¢ ¢

each ζ ∈ Z3 with ζ1 = 1 we list the corresponding set Ch(ζ), and the quantities 〈ℓ, ζ〉 and
pζ in the table below:
256 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

ζ 〈ℓ, ζ〉 Ch(ζ) pζ
p
(1, 1, 1) 2+ 2 ; 1
p
(1, 1, −1) 2 {2, 3} −1
p
(1, −1, 1) 2− 2 {1, 2} 1
p
(1, −1, −1) − 2 {1, 3} −1

3 ³ 2´
Q π
Since in this case we also have sin 2α j
= 0, the definition (7.3.3) yields
j =1

³³ p ´ ´ ³p ´ ³³ p ´ ´
F α,ℓ (τ) = cos 2 + 2 τ − 2 cos 2τ + cos 2 − 2 τ
³p ´
= −4 cos2 τ − 1 cos 2τ ,
¡ ¢

where the second equality follows from some elementary trigonometry. Therefore, by
solving Fα,ℓ (τ) = 0 and using Theorem 7.3.4, we deduce that we have a single quasi-
eigenvalue τ = 0, a subsequence of quasi-eigenvalues τ =¡πm , m¢ ∈ N, of multiplicity
two, and another subsequence of quasi-eigenvalues τ = pπ2 m − 12 , m ∈ N, of multiplic-
ity one. See also Remark 7.3.7.

Exercise 7.3.6
For each of the following polygons, write down the trigonometric function Fα,ℓ (τ) and
hence find the quasi-eigenvalues, with multiplicities.
¡π π π
(i) The equilateral triangle with α = and ℓ = (1, 1, 1).
¢
3, 3, 3
¡π π π
p ¢
(ii) The right-angled triangle with α = and ℓ = 1, 2, 3 .
¢ ¡
3, 6, 2

(iii) The square with α = π2 , π2 , π2 , π2 and ℓ = (2, 2, 2, 2). In this case, additionally com-
¡ ¢

pare the quasi-eigenvalues with those implied by Exercise 7.1.11.

Remark 7.3.7
Although it is not immediately transparent from the statements of Theorems 7.3.2 and
7.3.4, the asymptotics of the Steklov eigenvalues and eigenfunctions of a curvilinear poly-
gon is strongly affected by the arithmetic properties of its angles, in particular by the pres-
π
ence or absence of the so-called exceptional angles of the form 2k , k ∈ N, and special angles
π
of the form 2k−1 , k ∈ N. Firstly, in the absence of exceptional angles a multiplicity of
every quasi-eigenvalue is either one or two, whereas in the presence of K exceptional an-
gles a multiplicity of a quasi-eigenvalue may be as high as K (compare with the results
of Exercise 7.3.6(iv): the square has four exceptional angles, and the multiplicity of every
§7.3. Steklov spectra on domains with corners 257

quasi-eigenvalue is in fact four). Secondly, the asymptotic behaviour of the eigenfunc-

tions u m of the Dirichlet-to-Neumann map (that is, of the boundary traces Um |∂P of
the Steklov eigenfunctions) as m → ∞ may also be different: if all angles are special, then
the Dirichlet-to-Neumann eigenfunctions u m are asymptotically equidistributed on the
boundary in the sense that for any arc I ⊂ ∂P (not necessarily a side),

∥u m ∥L 2 (I ) Length(I )
lim = ,
m→∞ ∥u m ∥L 2 (∂P ) Length(∂P )

whilst in the presence of exceptional angles the eigenfunctions tend to concentrate on the
exceptional components of ∂P : the parts of the boundary between two consecutive excep-
tional angles. For an illustration of this phenomenon see Figures 7.6 and 7.7, which show
some numerically computed eigenfunctions u m for the equilateral triangle from Exercise
7.3.6(i), and for the isosceles right-angled triangle from Example 7.3.5. In the former case
all angles are special, and one observes that the eigenfunctions are more or less equally dis-
tributed on all sides, whereas in the latter case there are three exceptional angles, and the
eigenfunction u 18 is mostly concentrated on the union of two sides, and the eigenfunc-
tion u 19 is mostly concentrated on the hypothenuse.

Figure 7.6: Numerically computed eigenfunctions of the Dirichlet-to-Neumann map

on the equilateral triangle corresponding to the eigenvalues σ18 ≈ 17.8023 and σ19 ≈
16.6608, which in turn correspond to the quasi-eigenvalues τ18 = 5π and τ19 = 19π
3 (both
of which are in fact double, τ17 = τ18 and τ19 = τ20 ).

The complete proofs of Theorems 7.3.2 and 7.3.4 are rather difficult and lie well beyond the
scope of this book. In the next subsections, we explain some main ideas underlying these proofs
and their links to some classical problems in hydrodynamics, including the sloshing problem we
have mentioned already.
258 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

Figure 7.7: Numerically computed eigenfunctions of the Dirichlet-to-Neumann map

on the isosceles right-angled triangle from Example 7.3.5, corresponding to the eigenval-
ues σ18 ≈ 15.708 and σ19 ≈ 19.8968, which in turn correspond to the quasi-eigenvalues
τ18 = 17π 15π
3 (which is in fact double, τ17 = τ18 ), and τ19 = 2 2 (which is single).
p

§7.3.2. Sloping beach problems

Let (x, y) be Cartesian coordinates in R2 , and let (ρ, θ) denote the polar coordinates. Let

Sα = {(r, θ) : r > 0, −α < θ < 0}

denote an infinite sector of angle α with the vertex at the origin, where 0 < α ≤ π. For future use,
we denote its sides as

I in := {(r, −α) : r > 0}, I out := {(r, 0) : r > 0},

and call them the incoming and outgoing side, respectively, so that the angle α is measured counter-
clockwise from I in to I out . We also denote the bisector by I b := {(r, −α/2) : r > 0}, and introduce
the boundary coordinate s on ∂Sα = I in ∪ {(0, 0)} ∪ I out as shown in Figure 7.8, with s = 0 at the
vertex, s negative on I in , and positive on I out .
Restricting for the moment our attention to the half-sector S α2 , we consider two problems
there: a mixed Robin–Neumann problem
∂Φ
µ ¶¯
∆Φ = 0 − Φ ¯¯ = 0, ∂n Φ|I b = 0,
¯
in S ,
α (7.3.4)
2 ∂y I out

and a similar mixed Robin–Dirichlet problem

∂Φ
µ ¶¯
∆Φ = 0 − Φ ¯¯ = 0, Φ|I b = 0.
¯
in S α2 , (7.3.5)
∂y I out

We are particularly interested, in each case, in the existence of solutions which are bounded in the
closed sector S α2 and behave far from the origin as cos(x − ξ)e y , with some constant ξ. More
§7.3. Steklov spectra on domains with corners 259

Figure 7.8: Infinite sectors Sα and S α2 .

precisely, we additionally impose the conditions

Φ(x, y) = cos(x − ξ)e y + R(x, y), (7.3.6)

where
R(x, y) + |ρ∇R(x, y)| = O ρ −r as ρ → ∞,
¡ ¢
(7.3.7)
with some constant r > 0 (which may depend on the angle of the sector) to be determined.
The Robin–Neumann problem (7.3.4), (7.3.6), (7.3.7) is known as the sloping beach or the
floating mat problem, and has a long and storied history in hydrodynamics, see [Lew46] and ref-
erences therein20 . We will also refer to the Robin–Dirichlet problem (7.3.5)–(7.3.7) as a sloping
beach problem, somewhat abusing terminology. In particular, each of these problems has a solu-
tion of the required form if the parameter ξ takes a specific value which depends on the angle α2 :
in the Robin–Neumann case, one needs to take
π π2
ξ = ξ α2 ,N = − , (7.3.8)
4 4α
and in the Robin–Dirichlet case,
π π2
ξ = ξ α2 ,D = + . (7.3.9)
4 4α
The first result is due to A. S. Peters [Pet50], which was extended to the second problem in [LevPPS22a,
Theorem 2.1]; in both cases one can take r = απ in (7.3.7). We denote the corresponding solutions
of (7.3.4), (7.3.6), (7.3.7) and (7.3.5)–(7.3.7) by Φ α2 ,N (x, y) and Φ α2 ,D (x, y), respectively.
20 Remarkably, Lewy’s paper also recovers an elementary proof of the number-theoretical quadratic reciprocity law
as a corollary of his hydrodynamics results.
260 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

In §7.3.5, we will outline how to use the solutions Φ α2 ,N (x, y) and Φ α2 ,D (x, y) of the slop-
ing beach problems to obtain the asymptotics of the eigenvalues of the sloshing problem (7.1.12),
WD = ;.

§7.3.3. Peters solutions of the Robin problem in an infinite sector

We will now use the sloping beach solutions Φ α2 ,N (x, y) and Φ α2 ,D (x, y) in the “half" sector S α2
to construct some specific solutions of the Robin problem

∆Φ
e =0 in Sα , ∂n Φ
e = τΦ
e on I in ∪ I out , (7.3.10)

in the “full” sector Sα for large values of the Robin parameter τ. To do so, we start by extend-
ing the rescaled Robin–Neumann solution Φ α2 ,N (τx, τy) symmetrically across I b to a symmetric
Peters solution Φ e s (x, y) of (7.3.10). Similarly, we extend the rescaled Robin–Dirichlet solution
Φ α2 ,D (τx, τy) antisymmetrically across I b to an antisymmetric Peters solution Φ e a (x, y) of (7.3.10).
Let us now consider an arbitrary non-trivial linear combination Φ(x,e y) of Φ e s (x, y) and Φ
e a (x, y)
with constant complex coefficients. It is a solution of the Robin problem (7.3.10) which we call
its Peters solution. It is also clear from (7.3.6), (7.3.7), by converting the cosines into the complex
exponentials, that, as τ → +∞, the leading terms of the traces of Φ(x, e y) on the boundary rays I in
and I out are oscillatory in the variable s ,
¿ µ −iτs ¶À
iτs −iτs e
Φ I in (s) = h in,1 e + h in,2 e
¯
e ¯ + o(1) = hin , iτs + o(1),
e C2
¿ µ −iτs ¶À (7.3.11)
iτs −iτs e
Φ I out (s) = h out,1 e + h out,2 e
¯
e ¯ + o(1) = hout , iτs + o(1),
e C2

with some vectors µ ¶ µ ¶

h in,1 h out,1
hin := , hout := ∈ C2 .
h in,2 h out,2
We will denote such a Peters solution by

Φ
e τ (x, y; hin , hout ).

We now ask what should be the relations (if any) between vectors h+ and h− for the existence
of a Peters solution Φe τ (x, y; hin , hout ) of (7.3.10) with asymptotics (7.3.11). The equations (7.3.6),
(7.3.8), and (7.3.8) imply, after some linear algebra, that the relations we seek in fact depend upon
the arithmetic properties of the angle α: more precisely, they depend upon whether or not the
angle is exceptional, see Remark 7.3.7.

Theorem 7.3.8: [LevPPS22b, Theorem 3.1]

(i) Let α be a non-exceptional angle. Then for every hin ∈ C2 there exists a Peters solu-
§7.3. Steklov spectra on domains with corners 261

tion Φ
e τ (x, y; hin , hout ) of (7.3.10) satisfying (7.3.11) with

hout = A(α)hin , (7.3.12)

where
π 2
π 2
 
cosec 2α −i cot 2α
A(α) :=  . (7.3.13)
π2 π2
i cot 2α cosec 2α

π
(ii) Let α = 2k , k ∈ N, be an exceptional angle. Then a Peters solution Φ e τ (x, y; hin , hout )
of (7.3.10) satisfying (7.3.11) exists if the vectors hin , hout satisfy
­ ®
〈hin , X〉C2 = hout , X C2 = 0, (7.3.14)

where à k+1
!
e(−1) iπ/4
X := k .
e(−1) iπ/4

Remark 7.3.9
In both cases in Theorem 7.3.8, we obtain the existence of a Peters solution
Φ
e τ (x, y; hin , hout ) by fixing two out of the four components of the vectors hin , hout . The
difference is that in the non-exceptional case we fix the two components of the same vec-
tor and find the other vector from (7.3.12) (it does not in fact matter whether we fix either
of the two vectors as the matrix A(α) is invertible), whereas in the exceptional case we fix
exactly one component of each of hin and hout , and recover the other ones from (7.3.14).

Remark 7.3.10
It may be shown that the conditions on hin , hout in Theorem 7.3.8 are not only sufficient
but also necessary for the existence of Peters solutions.

§7.3.4. Quasimode construction for the Steklov problem in a curvilinear polygon

We are now outline the main ideas behind the proofs of Theorems 7.3.2 and 7.3.4 following the
exposition in [LevPPS22b]. As in the sloshing problem, we start by describing the construction
of the corresponding quasimodes.
Assume for simplicity that the polygon P = P (α, ℓ) has straight sides, and that all angles are
non-exceptional. We introduce on ∂P near each vertex V j the local coordinate s j such that s j is
zero at V j , negative on the side I j , and positive on the side I j +1 , see Figure 7.9. Note that on each
side I j joining V j −1 and V j we have effectively two coordinates: the coordinate s j running from
262 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

−ℓ j to 0, and the coordinate s j −1 running from 0 to ℓ j , related as

s j = s j −1 − ℓ j . (7.3.15)

This emphasises the fact that I j is the outgoing side of the sector with the vertex at V j −1 and the
incoming side of the sector with the vertex at V j .

• •

Figure 7.9: A straight polygon with local boundary coordi-

nates.

Let V j be the orientation-preserving isometry of the plane which maps the sector V j −1V j V j +1
into the sector Sα j with the vertex at the origin, and let (x ′j , y ′j ) := V j (x, y) be the local Cartesian
coordinates with the origin at V j . We will seek the quasimodes Ueτ (z) of the Steklov problem on
P which coincide, in the vicinity of each vertex V j , with a Peters solution

e τ (x ′ , y ′ ; h j ,in , h j ,out ),
Φ j j

where suitable values of the quasi-eigenvalues τ and the coefficient vectors h j ,in , h j ,out ∈ C2 are
to be determined. By Theorem 7.3.8(i), these vectors should be related by

h j ,out := A(α j )h j ,in (7.3.16)

to ensure the existence of the Peters solutions.

As a consequence of (7.3.11),

as τ → ∞,
¯
eτ ¯ = ue + o(1)
U ∂P

where we can write u| e I j as a trigonometric function of the variable s j involving the vectors h j ,in ,
h j ,out (using Φ
e τ (x ′ , y ′ ; h j ,in , h j ,out )) or as a trigonometric function of the variable s j −1 involving
j j
the vectors h j −1,in , h j −1,out (using Φ e τ (x ′ , y ′ ; h j −1,in , h j −1,out )).
j −1 j −1
These expressions should match, so an easy computation shows that we must have, with ac-
count of (7.3.15),
h j ,in = B(ℓ j , τ)h j −1,out , (7.3.17)
§7.3. Steklov spectra on domains with corners 263

where the side transfer matrices B(ℓ j , τ) are defined by

µ ¶
exp(iℓτ) 0
B(ℓ, τ) := (7.3.18)
0 exp(−iℓτ)

(the relations (7.3.17) and (7.3.18) essentially manifest just a change of variables on I j ). We will call
the vector h j ,in , the boundary quasi-wave incoming into V j (from V j −1 ), and the vector h j −1,out ,
the boundary quasi-wave outgoing from V j −1 (towards V j ). In order for our Peters solutions on
I j to match, these must be related by (7.3.17).
This formulation allows us to think of our problem as a transfer problem. Consider a bound-
ary quasi-wave b := hn,out outgoing from the vertex Vn towards V1 . It arrives at the vertex V1 as
an incoming quasi-wave h1,in = B(ℓ1 , τ)b, and, according to (7.3.16), leaves V1 towards V2 as an
outgoing boundary quasi-wave

h1,out = A(α1 )h1,in = A(α1 )B(ℓ1 , τ)b.

It then arrives at V2 as an incoming boundary quasi-wave

h2,in = B(ℓ2 , τ)A(α1 )B(ℓ1 , τ)b,

and leaves V2 towards V3 as an outgoing boundary quasi-wave

h2,out = A(α2 )B(ℓ2 , τ)A(α1 )B(ℓ1 , τ)b.

Continuing the process, we conclude that it arrives at Vn from Vn−1 as an incoming boundary
quasi-wave
hn,in = B(ℓn , τ)A(αn−1 )B(ℓn−1 , τ) · · · A(α1 )B(ℓ1 , τ)b
and leaves Vn towards V1 as an outgoing boundary quasi-wave

hn,out = A(αn )B(ℓn , τ)A(αn−1 )B(ℓn−1 , τ) · · · A(α1 )B(ℓ1 , τ)b = T(α, ℓ)b,

where we have denoted

T(α, ℓ) := A(αn )B(ℓn , τ)A(αn−1 )B(ℓn−1 , τ) · · · A(α1 )B(ℓ1 , τ).

The boundary quasi-wave hn,out must match the original outgoing boundary quasi-wave b, which
imposes the following quantisation condition on τ:

the matrix T(α, ℓ) has an eigenvalue 1. (7.3.19)

Using the explicit definitions (7.3.13) of the matrices A(α j ) and (7.3.18) of the matrices B(ℓ j , τ), it
is easily seen that (7.3.19) is equivalent to

Tr(T(α, ℓ)) = 2. (7.3.20)

Some rather elaborate calculations then demonstrate that every non-negative solution τ of (7.3.20)
is a root of the trigonometric equation Fα,ℓ (τ) = 0 and vice versa, with Fα,ℓ defined by (7.3.3),
264 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

and with multiplicities as stated in Theorem 7.3.4, thus giving the first hint of the validity of that
Theorem.
The full proof of Theorem 7.3.4 is highly non-trivial, and we only mention the remaining
steps briefly. First, after a rigorous construction of quasimodes Uem using appropriate cut-offs,
and with τm being the roots of (7.3.20), it is relatively easy to see that Um approximately sat-
isfy the Laplace equation and the Steklov boundary condition with suitably diminishing errors
as m → ∞. That allows us to conclude, in a standard manner, that τm are indeed the approxi-
mate eigenvalues of the Steklov problem on the curvilinear polygon in a sense that there exists a
subsequence of exact Steklov eigenvalues σi m such that |τm − σi m | = o(1) as m → ∞.
The most difficult part of the proof consists in establishing the correct enumeration of quasi-
eigenvalues by showing that i m = m . This is done with the help of Dirichlet–Neumann bracket-
ing: a suitably chosen sequence of cuts perpendicular to the boundary is added to ∂P , on which
either the Dirichlet or Neumann conditions are imposed, see Figure 7.10. These cuts are intro-
duced not simultaneously but in a particular order, allowing at each step a quantitative compari-
son with the known asymptotics of sloshing problems (mixed Steklov–Neumann problems) and
other mixed Steklov–Dirichlet and Steklov–Neumann–Dirichlet problems obtained in Theorem
7.3.11 and Remark 7.3.12 below, either directly, or using transplantation tricks similar to those used
in the proof of Theorem 6.2.17.

Figure 7.10: An example of a polygon with cuts.

Then all the results are extended from straight polygons to curvilinear polygons; here the
curvature of the boundary at the vertices requires special treatment using potential theory. Finally,
each step should be adjusted for the case of polygons with exceptional angles which need to be
analysed separately.
§7.3. Steklov spectra on domains with corners 265

§7.3.5. Asymptotics of the sloshing eigenvalues

We are now able to outline, following [LevPPS22a], how to use the solutions Φ α2 ,N (x, y) and
Φ α2 ,D (x, y) of the sloping beach problem to obtain the asymptotics of eigenvalues of the sloshing
problem (7.1.12), WD = ; — this does not require the full machinery of §7.3.4 and is, in fact, a
preliminary step for that. For simplicity, we assume that Ω is a triangle, the sloshing surface S
coincides with the interval (A, B ) = (0, L) of the horizontal axis, and that the walls W form the
β
angles α2 and 2 with the sloshing surface at the points A and B , respectively, see Figure 7.11.

Figure 7.11: A sloshing problem in a triangular

domain.

We are looking for quasimodes (approximate solutions) of (7.1.12), WD = ;, which are con-
structed, in the first approximation, by gluing together a sloping beach solution ±Φ α2 ,N (σx, σy)
near the corner point A and a sloping beach solution ±Φ β ,N (σ(L − x), σy) near the corner point
2 ³ ´
B . As the traces of these two solutions on S behave asymptotically as ± cos σx − ξ α2 ,N and
³ ´
± cos σx − σL + ξ β ,N for σ → +∞, cf. (7.3.6) and (7.3.8), the phases of the cosines should
2
match. This matching condition yields an asymptotic quantisation condition for the eigenvalues
σm subject to which the quasimodes can be rigorously constructed. The quasimode analysis can
be extended to the more general (no longer triangular, and with possibly curved walls) sloshing
domains, such as the one shown in Figure 7.1, which eventually leads to

Theorem 7.3.11: [LevPPS22a, Theorem 1.1]

Let Ω ⊂ Rd be a bounded simply connected domain with a Lipschitz boundary M = ∂Ω,
the sloshing surface S ⊂ M which is a straight line interval (A, B ) of length L , and the
β
walls W = M \ S which form the interior angles 0 < α2 , 2 < π2 with S at the points A
and B . Then the eigenvalues σm , m ∈ N, of the sloshing problem (7.1.12), WD = ;, have
the asymptotics

π2 2 2
µ ¶ µ ¶
1
Lσm = π m − − + + o(1) as m → ∞. (7.3.21)
2 8 α β
266 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

Remark 7.3.12
A similar method of constructing the quasimodes can also be applied in a full mixed
Steklov–Neumann–Dirichlet problem (7.1.12), with the following modifications: if the
Dirichlet condition is imposed on W near the corner A , we use a sloping beach solution
±Φ α2 ,D (σx, σy) there, and similarly near corner B . The result is the asymptotic formula
for the eigenvalues similar to (7.3.21), see [LevPPS22a, Theorem 1.8],

π2
µ ¶ µ ¶
1 2 2
Lσm = π m − + ± ± + o(1) as m → ∞, (7.3.22)
2 8 α β

β
where the contributions from the angles α2 and 2 appear with the plus sign if a Dirichlet
condition is imposed on W adjacently to the corner points A , B , respectively, and with a
minus sign in case of a Neumann condition.

Remark 7.3.13
As was additionally shown in [LevPPS22b], the remainder estimates in (7.3.21) and (7.3.22)
can be improved if the walls are straight near the corner. The formula (7.3.21) is also ap-
plicable if the walls form right angles with the sloshing surface subject to some additional
geometric constrains.

Remark 7.3.14
Numerical evidence suggests that asymptotics (7.3.21) and (7.3.22) remain valid for angles
α β
£π ¤
2 , 2 ∈ 2 , π . In the same vein, numerics suggest that Theorem 7.3.2 also remains valid
if the restriction α j < π on the angles of a curvilinear polygon is replaced by α j < 2π.
However, there is no proof of that in either case as the exponent r in the error estimate
(7.3.7) is not good enough to implement the quasimode argument.

Exercise 7.3.15
Verify the asymptotics (7.3.21) and (7.3.22) for the sloshing problems allowing separation
β
of variables: the rectangle (0, 1)×(−h, 0) from Exercise 7.1.10 (in which L = 1 and α2 = 2 =
π
2 ), and the mixed problems I–IV on the triangular domains from Figure 7.2 (in which
β
L = 2 and α2 = 2 = π4 , see also the discussion at the end of §7.1.2).

Remark 7.3.16
Very little is known about the spectral asymptotics for sloshing eigenvalues in higher di-
mensions beyond the leading term. We refer to [MaySenStA22] for some partial results
in that direction, as well as to [GirLPS19], [Ivr19] for related developments in the case of
§7.3. Steklov spectra on domains with corners 267

Steklov eigenvalues when the boundary has edges.

§7.3.6. Inverse spectral problem for curvilinear polygons

Here we follow [KryLPPS21]. Recalling, first of all, the definition of asymptotically Steklov isospec-
tral domains from Remark 7.2.13, we note that two curvilinear polygons with the same side lengths
ℓ and angles α are asymptotically Steklov isospectral by Corollary 7.3.3 (of course, at the same time
they need not be isospectral).
We further have

Theorem 7.3.17
Two curvilinear polygons are asymptotically Steklov isospectral if and only if their
trigonometric characteristic functions (7.3.3) coincide. Moreover, the trigonometric char-
acteristic functions of two curvilinear polygons coincide if and only if their non-negative
real roots (that is, the quasi-eigenvaluies τm of the polygons) coincide with account of
multiplicities. Additionally, the trigonometric characteristic function Fα,ℓ (τ) of a curvi-
linear polygon P (α, ℓ) can be uniquely reconstructed from the Steklov spectrum of
P (α, ℓ).

The proof of Theorem 7.3.17 is based on the application of the Hadamard–Weierstrass fac-
torisation theorem for entire functions and the property of almost periodic real functions with all
real zeros: if two such functions have asymptotically close zeros, they have exactly the same zeros
[KurSuh20, Theorem 6].
We will now describe what information on the geometry of a curvilinear polygon P (α, ℓ)
may be deduced from its Steklov spectrum (or equivalently, in accordance with Theorem 7.3.17,
from a characteristic trigonometric function F (τ). To do so, we need to work within a generic
class of curvilinear polygons, which we call admissible polygons, and which satisfy the following
two conditions:

the lengths ℓ1 , . . . , ℓn are incommensurable over {−1, 0, +1} (7.3.23)

(that is, only the trivial linear combination of ℓ1 , . . . , ℓn with these coefficients vanishes), and

all angles α1 , . . . , αn are not special (7.3.24)

(see Remark 7.3.7 for the definition).

Theorem 7.3.18
Given a characteristic trigonometric function F (τ) = Fα,ℓ (τ) of an admissible curvilinear
polygon, we can constructively recover, in a finite number of steps,

(i) the number of vertices n , and the number of exceptional angles K ;

268 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

(ii) if K = 0, then the vector of side lengths ℓ, in the correct order, subject to a cyclic
shift and a change of orientation, and further, once the enumeration of ℓ is fixed,
the vector
π2 π2
µ ¶
c(α) = cos , . . . , cos ,
2α1 2αn
modulo a global change of sign;

(iii) if K > 0, then we can recover the same information as in (ii) for each exceptional
component of ∂P (a part of the boundary between two exceptional angles) but not
the order in which the exceptional components are joined together.

If either (or both) of the admissibility conditions (7.3.23) and (7.3.24) is not satisfied, then
Theorem 7.3.18 is no longer applicable.

Example 7.3.19

(i) Consider a family of straight parallelograms P a depending on a parameter a ∈ (0, 1),

with angles π5 (which is special) and 4π
5 , and side lengths a and 1− a . In this case the
characteristic function
1
F (τ) = cos(2τ) − p
2
is independent of a , and we therefore cannot reconstruct side lengths from it —
all these parallelograms are asymptotically Steklov isospectral. In this example, both
conditions (7.3.23) and (7.3.24) are not satisfied.

of the same perimeter and with angles α = π7 , 63

π 53π
¡ ¢
(ii) Two ¡straight triangles , 63 and
π π 53π
α
¢
e = 9 , 21 , 63 , respectively (in each case there are two special angles), have the
same characteristic function F (τ) and are therefore asymptotically Steklov isospec-
tral.

§7.4. The Dirichlet-to-Neumann map for the Helmholtz equation

§7.4.1. Definition and basic properties

Let, as in §7.1, Ω be a bounded domain in a complete Riemannian manifold of dimension d ≥

2, with a Lipschitz boundary M := ∂Ω. Let us choose a real parameter Λ ̸∈ Spec(−∆D Ω ), and
consider, for a given u ∈ H 1/2 (Ω), a non-homogeneous Dirichlet problem

(
−∆U = ΛU in Ω,
(7.4.1)
U =u on M .
§7.4. The Dirichlet-to-Neumann map for the Helmholtz equation 269

This problem has a unique solution U ∈ H 1 (Ω) which we will call the Λ-Helmholtz extension of
u , and which we denote as
U := E Λ u ∈ H Λ (Ω),

where by analogy with (7.1.4) we define

H Λ (Ω) := {U ∈ H 1 (Ω) : −∆U = ΛU } = E Λ u : u ∈ H 1/2 (M )

© ª
(7.4.2)

to be the subspace of Λ-harmonic functions in H 1 (Ω).

Definition 7.4.1
Let Λ ̸∈ Spec(−∆D
Ω ). The linear operator

DΛ : H 1/2 (Ω) → H −1/2 (Ω), DΛ : u 7→ ∂n (E Λ u)|M ,

which maps u into the trace of the normal derivative of its Λ-Helmholtz extension, is
called the Dirichlet-to-Neumann map for the Helmholtz equation.

We want to extend Definition 7.4.1 to the case when Λ ∈ Spec(−∆D Ω ). We only do it briefly,
outlining the major steps; for the full rigorous definition in terms of the so called linear relations,
see [BehtEl15], and also [AreMaz12]. Let

K Λ := ∂n U : U ∈ H Λ (Ω) ∩ H01 (Ω)

© ª

be the finite-dimensional linear space of the Neumann boundary traces of eigenfunctions of −∆D
corresponding to a Dirichlet eigenvalue Λ. The non-homogenous problem (7.4.1) is solvable if
and only if u is orthogonal in L 2 (M ) to K Λ , see [McL00, Theorem 4.10]. The necessity of this
condition is immediate by Green’s formula: if U D is an eigenfunction of −∆D corresponding to
Λ, then from (7.4.1)

Λ U ,U D L 2 (Ω) = −∆U ,U D L 2 (Ω) = U , −∆U D L 2 (Ω) + u, ∂n U D L 2 (M )

¡ ¢ ¡ ¢ ¡ ¢ ¡ ¢

= Λ U ,U D L 2 (Ω) + u, ∂n U D L 2 (M ) ,
¡ ¢ ¡ ¢

implying u, ∂n U D L 2 (M ) = 0.
¡ ¢

Let K Λ⊥ denote an orthogonal complement to K Λ in L 2 (M ), and let ΠK ⊥ denote the or-

Λ
thogonal projection onto it. For any u ∈ H 1 (M ) ∩ K Λ⊥ , a solution to (7.4.1) exists but is not
unique as it is defined modulo an addition of an eigenfunction of −∆D corresponding to the
eigenvalue Λ. If we however treat E Λ u as a multi-valued map, then the map u 7→ ΠK ⊥ ∂n E Λ u is
Λ
still uniquely defined for u ∈ H 1 (M ) ∩ K Λ⊥ , and we will call it the Dirichlet-to-Neumann map
for the Helmholtz equation for Λ ∈ Spec(−∆D Ω ). We note that this construction relies on the fact
that the eigenfunctions of the Dirichlet Laplacian on Ω belong to the space H 3/2 (Ω), see Remark
2.2.20.
270 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

For any fixed Λ ∈ R, the Dirichlet-to-Neumann map DΛ is a self-adjoint operator in L 2 (M )

with a discrete spectrum of real eigenvalues

σΛ Λ
1 ≤ σ2 ≤ . . . ,

see [BehtEl15], [AreMaz12], and also [GréNédPla76]. The eigenvalues and the corresponding
eigenfunctions u Λj , j = 1, . . . , ∞, satisfy
(
−∆U = ΛU in Ω,
(7.4.3)
∂n U = σΛj u j on M ,

with U := E Λ u j , and the basis of eigenfunctions may be chosen to be orthogonal in L 2 (M ). The

analogue of the weak Steklov spectral problem (7.1.5) for (7.4.3) is

(∇U , ∇V )L 2 (Ω) − Λ(U ,V )L 2 (Ω) = σ(U ,V )L 2 (M ) for all V ∈ H 1 (Ω). (7.4.4)

Let now (
H 1/2 (M ), if Λ ̸∈ Spec(−∆D
Ω ),
u ∈ Dom(DΛ ) =
H 1 (M ) ∩ K Λ⊥ , if Λ ∈ Spec(−∆DΩ ).
The quadratic form of the Dirichlet-to-Neumann map DΛ is given by

(DΛ u, u)L 2 (Ω) = (∂n U , u)L 2 (M ) = ∥∇U ∥2L 2 (Ω) − Λ∥U ∥2L 2 (Ω) , (7.4.5)

cf. (7.1.6). We have the following analogue of (7.1.11) and Theorem 7.1.9.

Theorem 7.4.2: The variational principle for the eigenvalues of the Dirichlet-to-
Neumann map
Let Ω be a bounded open set in Rd , with a Lipschitz boundary M = ∂Ω, let Λ ∈ R, and
let σΛ
k
be the eigenvalues of the Dirichlet-to-Neumann map for the Helmholtz equation
in Ω. Then
∥∇E Λ u∥2L 2 (Ω) − Λ∥E Λ u∥2L 2 (Ω)
σΛ
k = min max
L
f⊂Dom(DΛ ) u∈L
f\{0} ∥u∥2L 2 (M )
dim L
f=k
(7.4.6)
∥∇U ∥2L 2 (Ω) − Λ∥U ∥2L 2 (Ω)
= min max , k ∈ N.
L ⊂H Λ (Ω) U ∈L ∥U |M ∥2L 2 (M )
dim L =k U = ̸ 0

Moreover, if Λ < λD 1 (Ω), then H Λ (Ω) in the right-hand side of (7.4.6) may be replaced
by H 1 (Ω), and we have
∥∇W ∥2L 2 (Ω) − Λ∥W ∥2L 2 (Ω)
σΛ
k (Ω) = min max , k ∈ N. (7.4.7)
L ⊂H 1 (Ω) W ∈L ∥W |M ∥2L 2 (M )
dim L =k W = ̸ 0
§7.4. The Dirichlet-to-Neumann map for the Helmholtz equation 271

Proof
The formula (7.4.6) is just the standard variational principle taking into account (7.4.5),
(7.4.2), and the definition of E Λ . In order to prove the validity of (7.4.7) we first need,
assuming Λ < λD 1 (Ω), the following analogue of Proposition 7.1.8: we have H (Ω) =
1
1
H Λ (Ω) ⊕ H0 (Ω) and

(∇U , ∇V )L 2 (Ω) = Λ(U ,V )L 2 (Ω) for any U ∈ H Λ (Ω),V ∈ H01 (Ω).

Taking now in (7.4.7) H 1 (Ω) ∋ W = U + V , with U ∈ H Λ (Ω), V ∈ H01 (Ω), we obtain

∥∇W ∥2L 2 (Ω) − Λ∥W ∥2L 2 (Ω) ≥ ∥∇U ∥2L 2 (Ω) − Λ∥U ∥2L 2 (Ω) + (λD 2
1 (Ω) − Λ)∥V ∥L 2 (Ω) ,

and the minimisation procedure now requires taking V = 0.

Exercise 7.4.3
By separating variables in polar coordinates (r, θ), show that the spectrum of the Dirichlet-
to-Neumann map DΛ in the unit disk consists of the single eigenvalues
 ′ ¡p ¢
I 0 −Λ

 ¡p ¢ , if Λ < 0,
 I 0 −Λ


0, if Λ = 0,
J′ Λ

 ¡p ¢
 0 ¡p ¢ , if Λ > 0,


J0 Λ

with the corresponding eigenfunction u(θ) = 1, and the double eigenvalues

 ′ ¡p ¢
I m −Λ

 ¡p ¢ , if Λ < 0,
 I m −Λ


m, if Λ = 0, m ∈ N,
Λ

J′
 ¡p ¢
 m ¡p ¢ , if Λ > 0,


Jm Λ

with the corresponding eigenfunctions u(θ) = cos mθ and u(θ) = sin mθ , where J m
and I m are the Bessel functions and the modified Bessel functions, respectively. Use these
expressions to reproduce Figure 7.12, and compare it to Figure 3.1, cf. also Exercise 3.1.17.
272 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

Figure 7.12: Some eigenvalues of the Dirichlet-to-Neumann

map DΛ for the unit disk as functions of Λ. The dashed black
curves correspond to single eigenvalues, and the solid curves to
double eigenvalues. The vertical dotted lines are placed at the
abscissae coinciding with the Dirichlet eigenvalues of the unit
disk.

§7.4.2. Dependence of the eigenvalues of the Dirichlet-to-Neumann map on the param-

eter

The behaviour of eigenvalues of the Dirichlet-to-Neumann map DΛ as functions of Λ shown in

Figure 7.12 for the unit disk is in fact typical (except for the multiplicities of the eigenvalues) for a
generic Lipschitz domain Ω ⊂ Rd . We start by revisiting Remark 3.1.19 and re-stating it rigorously.

Proposition 7.4.4: Robin–Dirichlet-to-Neumann duality [AreMaz12, Theorem

3.1], [HasShe22]
Let Ω ⊂ Rd be a Lipschitz domain, and let Λ, σ ∈ R. Then σ is an eigenvalue of the
Dirichlet-to-Neumann map DΛ if and only if Λ is an eigenvalue of the Robin Laplacian
−∆R,−σ . Moreover, the multiplicities of σ as an eigenvalue of DΛ and of Λ as an eigenvalue
of −∆R,−σ coincide.

Proposition 7.4.4 is almost immediately obvious (at least when Λ ̸∈ Spec(−∆D )) from the
fact that the mapping T : H Λ (Ω) → H 1/2 (M ) which acts as T : U → U |M , is an isomorphism
between the corresponding eigenspaces (as well as its inverse E Λ : H 1/2 (M ) → H Λ (Ω)).
We now go back to the Robin problem and state the following extension of (3.1.19).
§7.4. The Dirichlet-to-Neumann map for the Helmholtz equation 273

Proposition 7.4.5: [AreMaz12, Proposition 3]

R,γ
Let Ω ⊂ Rd be a Lipschitz domain. For a fixed k ∈ N, the eigenvalues λk (Ω) of the
Robin Laplacian on Ω are continuous strictly monotone increasing functions of γ ∈ R,
and satisfy
n o
R,γ R,γ
lim λk (Ω) = sup λk (Ω) : γ ∈ R = λD
k (Ω), (7.4.8)
γ→+∞
R,γ
lim λ (Ω) = −∞. (7.4.9)
γ→−∞ k

Proof
For an illustration, see once more Figure 3.1. We have already established the (non-strict)
monotonicity of the Robin eigenvalues as functions of γ in Theorem 3.2.9. To prove the
R,γ R,γ
strict monotonicity, assume for contradiction that for some k ∈ N we have λk 2 = λk 1 =:
Λ with some γ1 < γ2 . Then by Proposition 7.4.4,
£ ¤
−γ2 , −γ1 ⊂ Spec(DΛ ),

which contradicts the fact that the spectrum of the Dirichlet-to-Neumann map DΛ is dis-
crete.
The limiting behaviour (7.4.8) of the Robin eigenvalues as γ → +∞ has been already
discussed in §3.1.3. To prove the limiting identity (7.4.9), assume for contradiction that for
γ,R γ,R
some k ∈ N, the eigenvalue λk is bounded below by Λ := infγ∈R λk > −∞. Then by
Proposition 7.4.4,
γ,R
n o
Spec(DΛ ) ⊆ −γ : λ j = Λ, j = 1, . . . , k ,

which is a finite set, and therefore impossible, thus proving (7.4.9).

Remark 7.4.6
R,γ
As can be seen from Figure 3.1, the k th Robin eigenvalue λk is only continuous in γ, and
not necessarily smooth. If however we follow the eigenvalue branches correctly through
their crossings, forsaking the ordering of eigenvalues, then the union over γ of spectra of
the Robin Laplacians −∆R,γ may be decomposed into the union of analytic eigencurves,
R,γ
see [BucFreKen17, §4.4.2] for details. Moreover, if λk is a simple eigenvalue of the Robin
Laplacian −∆R,γ , and Uk is the corresponding eigenfunction, then
° °2
d R,γ ° Uk |M °L 2 (M )
λ = .
dγ k ∥Uk ∥2L 2 (Ω)

R,γ
Let us now consider the functions γk : (−∞, λD
k
) → R which are the inverses of λk viewed
274 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

as functions of γ. These inverses are well-defined due to the strict monotonicity of the Robin
eigenvalues established in Proposition 7.4.5. The functions −γk (Λ) are continuous and strictly
monotone decreasing for Λ ∈ −∞, λD
¡ ¢
k
, and satisfy
¡ ¢ ¡ ¢
lim −γk (Λ) = +∞, lim
¡ ¢− −γk (Λ) = −∞.
Λ→−∞ Λ→ λD
k

Using Proposition 7.4.4, we can now explicitly find the spectrum of the Dirichlet-to-Neumann
map DΛ in terms of the functions −γk (Λ).

Proposition 7.4.7: [AreMaz12, Proposition 5]

Let Ω ⊂ Rd be a Lipschitz domain, and let Λ ∈ R. Choose m ∈ N such that λD
m−1 ≤ Λ <
λD
m , where we assume the convention λD
0 := −∞ . Then
© ª
Spec(DΛ ) = −γk (Λ) : k ≥ m .

Using Proposition 7.4.7 we immediately deduce

Theorem 7.4.8
Let Ω ⊂ Rd be a Lipschitz domain. The eigenvalues σΛ k
(Ω) of the Dirichlet-to-Neumann
map DΛ are continuous and strictly monotone decreasing
¡ D¢
functions of Λ on each interval
of the real line not containing the points of Spec −∆Ω . As Λ approaches from below a
Dirichlet eigenvalue λD of multiplicity m , the first m eigenvalues σΛ Λ
1 , . . . , σm of DΛ tend
to −∞.

Remark 7.4.9
In the smooth case, Theorem 7.4.8 was first stated in [Fri91, Lemma 2.3]. Further re-
¡ D ¢−
Λ Λ
sults on the asymptotics of eigenvalues σ1 , . . . , σm as Λ → λ can be deduced from
[BelBBT18], see also [GirKLP22, §4.4].

Remark 7.4.10
As we already know that the eigenvalues of the Steklov problem (or the operator D0 ) are
non-negative, Theorem 7.4.8 immediately implies that

σΛ
k >0 for all Λ < 0 and all k ∈ N.

The behaviour of the Dirichlet-to-Neumann eigenvalues σΛ k

as functions of Λ will be dis-
cussed also below in §7.4.3. We now concentrate on the analogue of Theorem 7.2.9 in order to
compare the eigenvalues of the Dirichlet-to-Neumann map DΛ with Λ ≤ 0 with those of the
boundary Laplace–Beltrami operator −∆M . Namely, we state
§7.4. The Dirichlet-to-Neumann map for the Helmholtz equation 275

Theorem 7.4.11: [GirKLP22, Theorem 4.2]

Let Ω ⊂ Rd be a bounded domain with a smooth boundary M = ∂Ω, and let σΛ k
and
νk , k ∈ N, be the eigenvalues of the Dirichlet-to-Neumann map D Λ on Ω, and of the
Laplace–Beltrami operator on M , respectively. Then there exists a constant C > 0 such
that ¯ ¯
¯ Λ p
¯σk − νk − Λ¯ ≤ C
¯
(7.4.10)
uniformly over all Λ ∈ (−∞, 0] and all k ∈ N.

We note that in two dimensions, much more precise results are available as k → ∞ [LagStA21],
cf. Remark 7.2.12 in the case Λ = 0.
The proof of Theorem 7.4.11 relies on the following generalisation of Hörmander’s identity
of Theorem 7.2.5.

Theorem 7.4.12: The generalised Hörmander’s identity [GirKLP22, Theorem

4.3]
Let Ω ⊂ Rd be a bounded domain with a smooth boundary M = ∂Ω. Let F be a smooth
vector field on Ω which on the boundary of Ω coincides with the exterior unit normal,
F|M = n. Let u ∈ H 1 (M ), let Λ ≤ 0, and let U = E Λ u be the unique Λ-Helmholtz exten-
sion of u onto Ω. Then
(DΛ u, DΛ u)L 2 (M ) − (−∆M u, u)L 2 (M ) + Λ(u, u)L 2 (M )
ˆ
2JacF [∇U , ∇U ] − |∇U |2 div F + ΛU 2 div F dx. (7.4.11)
¡ ¢
=
Ω

Exercise 7.4.13
Prove Theorem 7.4.12 by first showing that after replacing the harmonic extension U =
E 0 u by the Λ-Helmholtz extension U = E Λ u in Theorem 7.2.4 the formula (7.2.5) be-
comes
ˆ ˆ
1
〈F, ∇U 〉∂n U ds − |∇U |2 〈F, n〉 ds
2
M M
ˆ ˆ
Λ 1
+ u 2 〈F, n〉 ds + |∇U |2 div F dx (7.4.12)
2 2
M Ω
ˆ ˆ
Λ
− JacF [∇U , ∇U ] dx − U 2 div F dx = 0
2
Ω Ω

(see [HasSif20, Theorem 3.1]), and then using (7.4.12) and repeating the arguments in the
proof of Theorem 7.2.5, keeping track of Λ-dependent terms. See also Exercise 7.4.15 for
276 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

further applications of (7.4.12).

Proof of Theorem 7.4.11

We first note that under the conditions of Theorem 7.4.12 there exists a constant C > 0
such that
¯(DΛ u, DΛ u) − Λ)u, u)L 2 (M ) ¯ ≤ C (DΛ u, u)L 2 (M ) .
¯ ¯
L 2 (M ) − ((−∆M (7.4.13)

Indeed, taking the absolute value of the left-hand side of (7.4.11) gives the left-hand side of
(7.4.13). Taking the absolute value of the right-hand side of (7.4.11) and estimating the first
two terms as in Corollary 7.2.6 produces an upper bound C (∇U , ∇U )L 2 (Ω) for them; the
last term can be estimated as C |Λ|(U ,U )L 2 (Ω) (possibly with a different constant C but also
depending on F and the geometry of Ω only). Combining the two bounds with account
of |Λ| = −Λ, the total bound on the right-hand side becomes

C (∇U , ∇U )L 2 (Ω) − Λ(U ,U )L 2 (Ω) = C (DΛ u, u)L 2 (M ) ,

¡ ¢

thus establishing (7.4.13). The bound (7.4.10) now follows from (7.4.13) by a direct appli-
cation of Proposition 7.2.8 with A = DΛ and B = −∆M −Λ, which are both non-negative
for Λ ≤ 0.

We illustrate Theorem 7.4.11 in Figure 7.13.

Figure 7.13: Some eigenvalues of the Dirichlet-to-Neumann map DΛ for the unit disk as
functions of Λ (solid curves), and, for comparison, the plots of νk − Λ (dashed curves).
p

In the left figure, Λ ∈ [−20, 0], and k is chosen in the set {1, 3, 5, 7, 9}. In the right figure,
Λ ∈ [−2 × 106 , −2 × 106 + 103 ], and k is chosen in the set {100, 102, 104, 106, 108}.
§7.4. The Dirichlet-to-Neumann map for the Helmholtz equation 277

Remark 7.4.14
The boundary regularity assumed in the conditions of Theorem 7.2.11 may be relaxed
slightly to allow for C 1,1 boundary, cf. Remark 7.2.11. On the other hand, [GirKLP22,
Proposition 4.6] shows that for curvilinear polygons the bound (7.4.10) cannot hold uni-
formly over all k ∈ N and Λ ≤ 0 for any fixed choice of the sequence {νk }. This observation
is based on comparison of the asymptotics of the eigenvalues σΛ k
as Λ → −∞ imposed by
(7.4.10) with the actual asymptotics for polygons which can be obtained from the results
of [LevPar08, Kha18, KhaPan18, KhaOuBPan20, Pan20, Pop20] on the asymptotics of
Robin eigenvalues.

Exercise 7.4.15
The generalised Pohozhaev’s identity (7.4.12) for the Helmholtz equation has some fur-
ther applications. Use it first to prove the classical Rellich’s identity [Rel40]: if Ω ⊂ Rd is a
domain with a smooth boundary M = ∂Ω, and Λ, U are an eigenvalue and a correspond-
ing normalised eigenfunction of −∆D Ω , then
ˆ
2Λ = 〈x, n〉(∂n U )2 dVM . (7.4.14)
M

Then use (7.4.12) and (7.4.14) to prove the following result of A. Hassell and T. Tao [Has-
Tao02]: there exist constants C 1 ,C 2 > 0 such that for any eigenvalue Λ and a correspond-
ing normalised eigenfunction U of the Dirichlet Laplacian −∆D Ω one has

C 1 Λ ≤ ∥∂n U ∥2L 2 (M ) ≤ C 2 Λ.

In a similar manner, one can estimate the boundary norm ∥U ∥2L 2 (M ) for a Neumann (or,
more generally, Robin) eigenfunction in a domain Ω, see [RudWigYes21].

§7.4.3. The Dirichlet-to-Neumann map and the eigenvalue counting functions

To make a full circle, we note that the Dirichlet-to-Neumann map is also useful for the study
of Laplace eigenvalues. We will sketch Friedlander’s original proof of the non-strict version of
the inequality (3.2.9) between the Neumann and Dirichlet eigenvalues of a Euclidean domain Ω
which we stated in Theorem 3.2.35. Let N D (Λ) and N N (Λ) denote the usual counting functions
of the Dirichlet and Neumann Laplacians on Ω, respectively, and let
DΛ
n(Λ) := N (0) = # k ∈ N : σΛ
© ª
k ≤0

be the number of non-positive eigenvalues of the Dirichlet-to-Neumann map DΛ . We have al-

ready established (Remark 7.4.10) that n(Λ) = 0 for Λ < 0. Note also that by the Robin–Dirichlet-
to-Neumann duality, zero is an eigenvalue of DΛ if and only if Λ is an eigenvalue of the Neumann
Laplacian on Ω, and that the multiplicities of these eigenvalues coincide. Thus, as a varying Λ
278 Chapter 7. The Steklov problem and the Dirichlet-to-Neumann map

passes through a Neumann eigenvalue of multiplicity m , exactly m eigenvalue curves σΛ

k
cross
from the upper half-plane to the lower one.
The key result of [Fri91] is the formula relating the three counting functions.

Lemma 7.4.16: [Fri91, Lemma 1.2], [AreMaz12, Proposition 4]

Let Ω ⊂ Rd be a Lipschitz domain, and let Λ ∈ R. Then

N N (Λ) − N D (Λ) = n(Λ).

R,γ
Proof. Since a Robin eigenvalue λk is strictly monotone increasing in the interval λN , λD
£ ¢
k k
as
γ increases from zero to +∞, we have

I Λ := k ∈ N : λN
k ≤ Λ < λk
D
© ª
n o
R,γ
= k ∈ N : there exists γ ≥ 0 such that λk = Λ .

By the definition of the eigenvalue counting functions and the first expression for the set I Λ , we
have #I Λ = N N (Λ) − N D (Λ). At the same time, by the Robin–Dirichlet-to-Neumann duality
and the second expression for I Λ , we have #I Λ = n(λ), and the result follows.

We further have

Lemma 7.4.17: [Fri91, Lemma 1.3], [AreMaz12, Lemma 3.2]

Let Ω ⊂ Rd be a bounded domain with a Lipschitz boundary M = ∂Ω, and let Λ > 0.
Then n(Λ) ≥ 1.

Proof
Consider, as in the original proof of Theorem 3.2.35, ¡a function g = ei〈ω,x〉 , where ω ∈ Rd ,
and |ω|2 = Λ. We have −∆g − Λg = 0 in Ω, and DΛ g ¯M = i 〈ω, n〉g ¯M . Thus,
¯ ¢ ¯

ˆ
DΛ g ¯M , g ¯M L 2 (M ) = i 〈ω, n〉 dVM = 0
¡ ¡ ¯ ¢ ¯ ¢
(7.4.15)
M

by the divergence theorem. On the other hand, assuming n(Λ) = 0 immediately implies
(DΛ u, u)L 2 (M ) > 0 for every u ∈ H 1/2 (M ), thus contradicting (7.4.15).

The proof of the non-strict version λN k+1

≤ λDk
of (3.2.9) now follows¡ immediately from
Lemmas 7.4.16 and 7.4.17: assuming that it is false and choosing any Λ0 ∈ λD λN
¢
k
, k+1
, we have
N D
N (Λ0 ) = N (Λ0 ), and so n(Λ0 ) = 0 by Lemma 7.4.16, thus contradicting Lemma 7.4.17. The
proof of the strict version can be achieved with minor modifications of this argument, see [Are-
Maz12, Theorem 3.3].
§7.4. The Dirichlet-to-Neumann map for the Helmholtz equation 279

Spectral geometry of the Steklov problem and the Dirichlet-to-Neumann map is an actively
developing subject, and many interesting questions remain beyond the scope of this chapter. For
further reading we refer to survey papers [GirPol17], [ColGGS22].
 APPENDIX A
A short tutorial on numerical spectral
geometry
After a brief overview of the Finite Element Method, we give a
hands-on tutorial on solving numerically some of the spectral
problems presented in this book using Mathematica and FreeFEM.

§A.1. Overview

§A.1.1. The Finite Element Method

The aim of this short tutorial is to provide the readers (who may be unfamiliar with numerical
analysis or any aspects of computer programming) a direct route to practical calculation of eigen-
values of some of the problems considered in this book. To this end, we neither pretend to give
a comprehensive survey of numerical spectral theory nor keep the presentation rigorous, concen-
trating instead on the practicalities of the Finite Element Method (FEM) in its most basic form
and in dimension two only, and ignoring numerous other available techniques (the finite differ-
ences, the method of fundamental solutions, spectral methods, the boundary element method,
to name just a few). For a comprehensive survey of both theoretical and practical foundations of
FEM applied to spectral problems see [SunZho17].
The Finite Element Method is based on the Galerkin (also called the Ritz–Galerkin) method
of solving a weak eigenvalue problem (3.1.2); we suppose that all the assumptions made in §3.1.1
about the bilinear form Q are fulfilled. Boris Grigoryevich
Let V ⊂ U be a finite-dimensional subspace of U = Dom Q . We consider the restriction of Galerkin
(3.1.2) to V : namely, we want to find λ ∈ R and u ∈ V \ {0} such that (1871–1945)

Q[u, v] = λB[u, v] for all v ∈ V, (A.1.1)

281
282 Appendix A. A short tutorial on numerical spectral geometry

where we set
B[u, v] := (u, v)H . (A.1.2)
If the subspace V approximates well the span of some eigenfunctions of Q , we expect that the
eigenvalues of (A.1.1) will approximate well the corresponding eigenvalues of (3.1.2). One usually
studies a family of approximating subspaces Vh depending on a real parameter h > 0 in such a
way that the projector U → Vh converges to the identity map as h → 0. Then various estimates
of convergence of eigenvalues and eigenfunctions are available. In particular, if λk is a simple
eigenvalue of (3.1.2) with the corresponding eigenfunction u , and λk,h is the k th eigenvalue of
(A.1.1) with V = Vh , then with some constant C independent of h we have
2
λk ≤ λk,h ≤ λk +C inf ∥u − v∥U ,
v∈Vh

where ∥ · ∥U is the norm induced by (3.1.1), see [SunZho17, §1.4.3] and [BabOsb91, §8].
Suppose now that {v 1 , . . . , v m } is a basis in V , not necessarily an orthogonal one. Looking for
an eigenvector of (A.1.1) in the form u = m
P
j =1 c j v j with unknown constants c j , j = 1, . . . , m ,
and taking v = v k , k = 1, . . . , m , we rewrite (A.1.1) as a generalised matrix eigenvalue problem

c1
 

c =  ..  ∈ Rm ,
 . 
Sc = λMc, (A.1.3)
cm

where
S := Q[v j , v k ]
¡ ¢
k, j =1,...,m
(A.1.4)
is the so-called stiffness matrix, and

M := B[v j , v k ]
¡ ¢
k, j =1,...,m
(A.1.5)

is called the mass matrix. We now solve the eigenvalue problem (A.1.3) using some numerical
linear algebra method.
The Finite Element Method (specifically, in application to spectral problems for the Lapla-
cian in a bounded domain Ω ⊂ R2 , and in its simplest form) is usually understood as a particular
realisation of the Galerkin method subject to the following conditions:

(a) Ω is represented (or approximated) by a union Th of closed triangles, called a mesh, where a
real parameter h provides an upper bound on the diameter (or some other linear size) of each
T ∈ Th . The different triangles may only have a common side or a common vertex, see Figure
A.1. If Ω is not a polygon, approximating it by a union of triangles obviously introduces
some additional errors. There are many alternative choices to triangles, such as quadrilaterals
or curvilinear elements, which we do not discuss.

(b) Let T ∈ Th be a triangle in the chosen mesh, and let P k = P k (T ) be the subspace of all
polynomials in two variables of degree at most k . Then dim P k = 12 (k + 1)(k + 2) =: s k . We
choose s k points z 1 , . . . , z sk ∈ T , called nodes, which lie on k + 1 straight lines. In particular
§A.1. Overview 283

when k = 1 we have s 1 = 3 and choose the nodes at the vertices of the triangle, and when
k = 2 we have s 2 = 6 and choose additionally the nodes at the midpoints of the sides. For a
polynomial p ∈ P k (T ), the set of functionals N := {N j : p 7→ p(z j ), j = 1, . . . , s k } is the set
of degrees of freedom which is unisolvent: knowing N (p) := {N j (p), j = 1, . . . , s k } uniquely
determines p . In principle, this choice of degrees of freedom is just a specific realisation of
the general principle of using any unisolvent set of functionals N .

(c) We can now choose a local basis {p 1 , . . . , p sk } in P k (T ) by requiring p i (z j ) = N j (p i ) = δi j ,

i , j = 1, . . . , s k . Finally, we set V to be the space of continuous functions on Ω whose restric-
tions to each T ∈ Th coincides with P k (T ): thus, the elements of V are piecewise polyno-
mials. Note that continuity is required to ensure V ⊂ H 1 (Ω) (thus providing the so-called
conforming finite elements). Such basis functions are called Lagrangian finite elements.

Joseph-Louis
Lagrange
(1736–1813)

Figure A.1: Examples of automatically constructed meshes for a disk and a domain with
a hole.

Remark A.1.1
The term finite elements is variably applied to the whole method, a choice of mesh subdo-
mains (e.g. triangular or quadrilateral finite elements), or a choice of local basis functions
(e.g. linear or quadratic conforming finite elements or some other non-conforming finite
elements).
284 Appendix A. A short tutorial on numerical spectral geometry

§A.1.2. Solving spectral problems with Mathematica

There is a large number of software packages, either commercial or free to use, which implement
the FEM for solving partial differential equations including spectral problems. For an up-to-date
review see the corresponding Wikipedia page. In particular, widely available commercial pack-
ages Matlab21 (with PDE Toolbox22 ) and Mathematica23 (starting from version 10.2) allow
one to compute eigenvalues and eigenfunctions of various boundary value problems with rela-
tive ease. Mathematica is particularly easy to use as it provides two commands, DEigenvalues
24 and NDEigenvalues25 for calculating the eigenvalues of a boundary value problem analytically

(if possible) and numerically, respectively. The numerical version effectively “hides” all the FEM
machinery from the user. The version NDEigensystem26 allows additionally to compute the eigen-
functions.
We do not intend to give any further details of Mathematica commands, restricting our-
selves to several examples below.

Remark A.1.2
All the scripts listed or discussed in this Appendix are available for download, see §A.3.

Listing A.1 gives some examples of using Mathematica for finding eigenvalues and eigen-
functions analytically.27

Listing A.1: Finding eigenvalues analytically with Mathematica

1 (* Neumann eigenvalues for the unit square *)

2 DEigenvalues [ - Laplacian [ u [x , y ] , {x , y }] , u [x , y ] , {x , y } ∈
Rectangle [{0 , 0} , {1 , 1}] , 10]
3 (* Dirichlet eigenvalues for the unit disk *)
4 DEigenvalues [{ - Laplacian [ u [x , y ] , {x , y }] , DirichletCondition [ u [
x , y ] == 0 , True ]} , u [x , y ] , {x , y } ∈ Disk [] , 10]
5 (* Dirichlet eigenvalues for the isosceles right triangle with
sides \[ Pi ] *)
6 DEigenvalues [{ - Laplacian [ u [x , y ] , {x , y }] , DirichletCondition [ u
[x , y ] == 0 , True ]} , u [x , y ] , {x , y } ∈ Triangle [{{0 , 0} , { Pi
, 0} , {0 , Pi }}] , 10]

21 https://www.mathworks.com/products/matlab.html
22 https://www.mathworks.com/products/pde.html
23 https://www.wolfram.com/mathematica
24 https://reference.wolfram.com/language/ref/DEigenvalues.html
25 https://reference.wolfram.com/language/ref/NDEigenvalues.html
26 https://reference.wolfram.com/language/ref/NDEigensystem.html
27 Listings A.1– A.5 may also be copy-pasted into Mathematica.
 §A.1. Overview 285

Our main geometric example throughout this tutorial will be the domain
Ω = Ω′ \ B,
n ³ x ´o
Ω′ = (x, y) : 0 < x < π, 0 < y < π + x 1 − , (A.1.6)
π
B = B ¡ π , π ¢, π ,
3 2 4

see Figure A.2.

Figure A.2: Domain Ω given by (A.1.6).

Listing A.2 shows how to compute the first ten Dirichlet and Neumann eigenvalues of Ω
with Mathematica.

Listing A.2: Computing Dirichlet and Neumann eigenvalues of Ω with

Mathematica

1 Ω = RegionDifference [ ImplicitRegion [0 < x < Pi && 0 < y < Pi + x

( Pi - x ) / Pi , {x , y }] , Disk [{ Pi /3 , Pi /2} , Pi /4]];
2 (* numerical Neumann eigenvalues for Ω *)
3 NDEigenvalues [ - Laplacian [ u [x , y ] , {x , y }] , u [x , y ] , {x , y } ∈ Ω ,
10]
4 (* numerical Dirichlet eigenvalues for Ω *)
5 NDEigenvalues [{ - Laplacian [ u [x , y ] , {x , y }] , DirichletCondition [ u
[x , y ] == 0 , True ]} , u [x , y ] , {x , y } ∈ Ω , 10]
286 Appendix A. A short tutorial on numerical spectral geometry

Further on, Listings A.3 and A.4 demonstrate how to compute the Robin and Zaremba eigen-
values and eigenfunctions, respectively. The graphical outputs of these scripts are shown in Fig-
ures A.3 and A.4. (The actual graphical outputs from these scripts have been slightly edited for
presentation purposes.)

Listing A.3: Computing Robin eigenvalues and eigenfunctions of Ω with

Mathematica

1 Ω = RegionDifference [ ImplicitRegion [0 < x < Pi && 0 < y < Pi + x

( Pi - x ) / Pi , {x , y }] , Disk [{ Pi /3 , Pi /2} , Pi /4]];
2 (* numerical Robin ( ∂n u = γu ) eigenvalues and contour plots of
eigenfunctions for Ω *)
3 γ = 2;
4 { eval , efun } = NDEigensystem [ - Laplacian [ u [x , y ] , {x , y }] +
NeumannValue [ γ u [x , y ] , True ] , u [x , y ] , {x , y } ∈ Ω , 6];
5 GraphicsGrid [ Table [ ContourPlot [ efun [[3 ( i - 1) + j ]] , {x , 0 , Pi
} , {y , 0 , 5 Pi /4} , Contours -> { -0.5 , -0.25 , 0 , 0.25 , 0.5} ,
RegionFunction -> Function [{ x , y , z } , {x , y } ∈ Ω ] , Frame ->
False , AspectRatio -> Automatic , BoundaryStyle -> Thick ,
PlotLabel -> " λ = " <> ToString [ eval [[3 ( i - 1) + j ]]]] , {i ,
1 , 2} , {j , 1 , 3}]]

Listing A.4: Computing Zaremba eigenvalues and eigenfunctions of Ω with

Mathematica

1 Ω = RegionDifference [ ImplicitRegion [0 < x < Pi && 0 < y < Pi + x

( Pi - x ) / Pi , {x , y }] , Disk [{ Pi /3 , Pi /2} , Pi /4]];
2 (* numerical Zaremba eigenvalues and eigenfunctions for Ω *)
3 (* Dirichlet condition on all sides except Neumann on the curved
upper side *)
4 { eval , efun } = NDEigensystem [{ - Laplacian [ u [x , y ] , {x , y }] ,
Di rich letCo nditi on [ u [x , y ] == 0 , y <= Pi ]} , u [x , y ] , {x , y } ∈
Ω , 6];
5 GraphicsGrid [ Table [ Plot3D [ efun [[3 ( i - 1) + j ]] , {x , 0 , Pi } , {y ,
0 , 5 Pi /4} , RegionFunction -> Function [{ x , y , z } , {x , y } ∈ Ω
, BoundaryStyle -> Thick , Boxed -> False , Axes -> True ,
AxesOrigin -> {0 , 0 , 0} , AspectRatio -> Automatic , Ticks ->
None , PlotLabel -> " λ = " <> ToString [ eval [[3 ( i - 1) + j ]]]] ,
{i , 1 , 2} , {j , 1 , 3}]]

To conclude the Mathematica part of our tutorial, we verify in Listing A.5 the Faber–Krahn
inequality for regular n -gons P n for n = 5, . . . , 20. We additionally compare our numerical result
with the asymptotics [BerGMR21]
³ ´
2
λD 12 − 2 j 0,1 ζ(5)
1 (P n ) 4ζ(3)
= 1+ + + O(n −6 ) as n → ∞, (A.1.7)
λ1 (P n )
D ∗ n3 n5
 §A.1. Overview 287

Figure A.3: Contour plots of Robin eigenfunctions (γ = 2)

of Ω given by (A.1.6).

where P n∗ is the symmetric rearrangement of P n and ζ(·) is the Riemann zeta function. The graph-
ical output from this script (once more, slightly edited for presentation purposes) is shown in
Figure A.5.

Listing A.5: Verifying the Faber–Krahn inequality for regular polygons with
Mathematica

1 (* verifying the Faber - - Krahn inequality for regular n - gons , n

=5 ,... ,20 *)
2 evs = Table [ NDEigenvalues [{ - Laplacian [ u [x , y ] , {x , y }] ,
Di rich letCo nditi on [ u [x , y ] == 0 , True ]} , u [x , y ] , {x , y } \[
Element ] RegularPolygon [ n ] , 1][[1]] , {n , 5 , 20}];
3 asympt = 1 + 4 Zeta [3]/ n ^3 + (12 - 2 BesselJZero [0 , 1]^2) Zeta
[5]/ n ^5;
4 Show [ ListPlot [ Table [{ n , evs [[ n - 4]] /( Pi / Area [ RegularPolygon [ n
]] BesselJZero [0 , 1]^2) } , {n , 5 , 20}] , PlotStyle -> { Black ,
PointSize [ Large ]} , AxesOrigin -> {4 , 1}] , Plot [ asympt , {n , 5 ,
20} , PlotRange -> All ]]
288 Appendix A. A short tutorial on numerical spectral geometry

Figure A.4: Height plots of Zaremba eigenfunctions of Ω given by (A.1.6), with the
Neumann condition imposed on the curved part of the outer boundary, and the Dirichlet
condition elsewhere. Note some spurious oscillations introduced by the numerics.

Figure A.5: The ratio λD1 (P n )/λD1 (P n∗ ) for regular polygons

P n plotted as a function of n (black dots) and compared to
asymptotics (A.1.7).

§A.2. Learning FreeFEM by example

§A.2.1. The basics

For the rest of this tutorial, we will concentrate on describing the FEM package FreeFEM, see
[Hec12] and the product website https://freefem.org/. As the name suggests, the package
 §A.2. Learning FreeFEM by example 289

is freely available for download. It is powerful enough to cover most of the problems considered
in this book, within the usual limitations of the finite element method — for example, one should
not expect to perform a reliable computation of sufficiently large eigenvalues of any problem. At
the same time, it is easy enough to learn very quickly without any prior knowledge of program-
ming or numerical analysis.
Giving a full description of FreeFEM is well outside the scope of this tutorial. One should
also consult the package documentation for installation instructions and additional details. We
will instead show, starting in the next subsection, various examples which should allow the reader
to produce their own scripts by mimicking ours.
The general flow of working with FreeFEM is somewhat similar of that of LATEX: one creates
a FreeFEM script (a text file with extension .edp) in an appropriate editing programme; one then
executes FreeFEM (many editors allow to do so directly from the editing window); corrects any
script errors reported, and then repeats the process until everything works as intended.

§A.2.2. The structure of a FreeFEM script: the Neumann problem in a rectangle

The standard structure of a FreeFEM script used in spectral problems is more or less the same,
and roughly complies with the following pattern.
A. Declarations: all user’s variables (identifiers) should be declared (and possibly assigned values
to) before or during their first appearance.
B. Boundary description.
C. Mesh creation.
D. Choice of FEM basis functions.
E. Description of quadratic forms Q and B for (A.1.1).
F. Creation of matrices S and M for (A.1.3).
G. Solving (A.1.3).
H. Results output and/or visualisation.
Listing A.6 shows the script which computes the Neumann eigenvalues for a rectangle (0, Lπ)×
(0, π) (with L = 1 as shown), and we will go through it in detail to illustrate the realisation of each
of the steps A–H in the general scheme.

Listing A.6: The Neumann Laplacian in a rectangle

1 / This FreeFEM script computes eigenvalues of the Neumann

Laplacian in the rectangle [0. , L pi ] x [0 , pi ].
2 //
3 // ---- A . DECLARATIONS ----
4 // IMPORTANT - every command ends with the semicolon !
290 Appendix A. A short tutorial on numerical spectral geometry

5 // IMPORTANT - all variables must be declared before ( or during )

first use
6 //
7 real L =1.; // declare variable L to be real and assign
value 1.0 to it ; decimal point indicates it is not integer
8 int npoints =30; // declare variable npoints to be integer and
assign value 30 to it
9 int N =50; // declare variable N to be integer and assign value
50 to it
10 real [ int ] Evalues ( N ) ; // an array of N real numbers parametrised
by integers ; in FreeFem , indices of an array of length N run
from 0 to N -1
11 real t ; // real uninitialised variable used later
12 // ---- END OF DECLARATIONS so far ----
13 // some more to come later
14 //
15 // ---- B . BOUNDARY DEFINITIONS ----
16 // just use parametric curve definitions going in
COUNTERCLOCKWISE direction ; ( t = t0 , t1 ) in lines below means
that t changes from t0 to t1 for this piece
17 //
18 border d1Omega ( t =0 , L ) { x = pi * t ; y = 0; label =1; };
19 border d2Omega ( t =0 , 1) { x = pi * L ; y = pi * t ; label =1; };
20 border d3Omega ( t =L , 0) { x = pi * t ; y = pi ; label =1; };
21 border d4Omega ( t = pi , 0) { x = 0; y = t ; label =1; };
22 // " label =1;" part can be omitted in lines above but will be
useful in other examples
23 // ---- END OF BOUNDARY DEFINITIONS ----
24 //
25 // ---- C . MESH CREATION USING buildmesh COMMAND ----
26 // format of the argument : pi e ce _d ef i ne d_ by _ bo rd e r (
nu m b er _o f _ me sh _p oi nt s ) +...
27 //
28 mesh Th = buildmesh ( d1Omega ( npoints * L * pi ) + d2Omega ( npoints * pi ) +
d3Omega ( npoints * L * pi ) + d4Omega ( npoints * pi ) ) ; // the number of
mesh points per side need not be proportional to side length
and may not be integer ( it ’s rounded down )
29 // ---- END OF MESH DEFINITIONS ----
30 //
31 // ---- VISUALISE THE MESH , may be commented out
32 //
33 plot ( Th , wait =1) ;
34 //
35 // ---- D . DECLARE THE FEM SPACE AND FEM VARIABLES ----
36 //
37 fespace Vh ( Th , P2 ) ;
38 Vh u , v ;
39 // ---- END OF FEM SPACE DEFINITIONS ----
40 //
41 // ---- E . DEFINE THE QUADRATIC FORMS ----
 §A.2. Learning FreeFEM by example 291

42 varf q (u , v ) = int2d ( Th ) ( dx ( u ) * dx ( v ) + dy ( u ) * dy ( v ) ) ;
43 varf b (u , v ) = int2d ( Th ) ( u * v ) ;
44 // ---- END OF QUADRATIC FORMS DEFINITIONS ----
45 //
46 // ---- F . CREATE THE MATRICES ----
47 matrix S = q ( Vh , Vh ) ;
48 matrix M = b ( Vh , Vh ) ;
49 // ---- END OF MATRIX CREATION ----
50 //
51 // ---- DECLARE THE ARRAY TO HOLD EIGENFUNCTIONS ----
52 Vh [ int ] Efunctions ( N ) ;
53 //
54 // ---- G . SOLVE THE PROBLEM ----
55 int k = EigenValue (S ,M , sym = true , value = Evalues , vector = Efunctions ) ;
56 //
57 // ---- END OF SOLVER ----
58 // ---- H . PRINT THE EIGENVALUES ----
59 cout << " We asked for " << N << " eigenvalues and computed " <<
k << " eigenvalues :\ n " << Evalues ;
60 //
61 // ---- PLOT THE 6 th EIGENFUNCTION ----
62 plot ( Efunctions [5]) ;
63 // press ’? ’ on the image to see options for graphics
64 //
65 // ---- END ----

The first eight lines of the script are just the comments, in fact every line starting with the
double slash (or any text at the end of a line after a double slash) is ignored by FreeFEM, and is
there just for the ease of reading the script. By the way, empty lines and spaces are also ignored.
Group A of commands, in lines 7–11, contains some declarations. Let us look at them line by
line, ignoring the comments.
The line

7 real L =1.;

declares a variable L to be real, and assigns value 1.0 to it. Variable names can be of arbitrary length
and consist of upper- and lower-case letters, numbers, and underscore, and start with a letter. One
can declare several variables at once, not necessarily assigning any values to them, for example one
can have

real L1 , L2 =0.5 , L3 ;

to define three real variables L1 (unassigned), L2 (with the value 0.5), and L3 (unassigned).
292 Appendix A. A short tutorial on numerical spectral geometry

Remark A.2.1
It is very important to remember that every individual command should end with the
semicolon!

The line

8 int npoints =30;

declares a variable npoints to be integer, and assigns value 30 to it; this variable will be used later.
The line

9 int N =50;

declares a variable N to be integer, and assigns value 50 to it. This variable will denote the number
of eigenvalues we want to compute.
The line

10 real [ int ] Evalues ( N ) ;

declares Evalues to be an array of real numbers of length N indexed by integers from 0 to N−1,
which will eventually hold the eigenvalues. Note that interchanging lines 10 and 9 would give an
error — we cannot declare an array until we know its size.
The last declaration in line

11 real t ;

declares t as another real variable, left unassigned.

Group B of commands, describing the boundary, is in lines 18–21. The boundary should be
defined as a collection of smooth parametrised curves swept in such a way that the domain lies to
the left of the direction of parametrisation: since in this case we have a simply connected domain,
we parametrise in the counterclockwise direction. A definition of a boundary piece usually takes
the form

border border_name ( t = t0 , t1 ) { x = a_function ( t ) ; y =

another_function ( t ) ; label = natural_number ;}

to define a parametric curve. Note that we can have t1<t0 as in lines 20 and 21. Note also that
parametrisation parameters are of course a matter of choice, compare lines 19 and 21. The part
“label=...;” in lines 18–21 is optional — but labels are important if we want to integrate over the
boundary, or impose different boundary conditions on different boundary pieces, allowing us to
group them together, as we will do later.
 §A.2. Learning FreeFEM by example 293

The meshing is done in Group C consisting of one line 28. Once all the boundary pieces
are defined, we create the mesh by executing buildmesh command in and assigning the output to
variable Th declared to be a mesh. The general format of buildmesh command is

buildmesh ( boundary1 ( points1 ) +...+ boundaryX ( pointsX ) ) ;

where each boundary1, . . . , boundaryX has been previously defined as a border, and points1, . . . ,
pointsX indicate how many mesh points to place on each border, thus determining mesh coarse-
ness. We have used a previously defined variable npoints to indicate the number of boundary
points per unit length of the boundary, but such a choice is not compulsory, albeit convenient.
We now proceed to describing the finite element space in Group D of commands. The line

37 fespace Vh ( Th , P2 ) ;

defines the FEM space Vh on the mesh Th consisting of Lagrangian quadratic finite elements, as
indicated by parameter P2. We may have chosen instead Lagrangian linear finite elements (replace
P2 with P1) or many other types of finite elements described in FreeFEM manual. Essentially this
command introduces the new type Vh, and the following command in line 38 declares u and v to
be variables of that type.
Group E, consisting of two commands

42 varf q (u , v ) = int2d ( Th ) ( dx ( u ) * dx ( v ) + dy ( u ) * dy ( v ) ) ;
43 varf b (u , v ) = int2d ( Th ) ( u * v ) ;

defines the quadratic forms (varf)

ˆ
Q[u, v] =
¡ ¢
(∂x u)(∂x v) + (∂ y u)(∂ y v) dxdy,
Ω
ˆ
B[u, v] = uv dxdy
Ω

in accordance with (3.1.3) and (A.1.2), where we use build-in two-dimensional integration com-
mand int2d and differentiation commands dx and dy.
We now create, in group F of commands in lines 47–48, the matrices S and M associated with
the quadratic forms, see (A.1.4) and (A.1.5).
We now proceed to group G of actual computations, first declaring the array Efunctions (of
type Vh and length N, parametrised by integers) in line 52, and then solving the problem in line
55. The parameter sym=true to EigenValue indicates that the problem is symmetric (in principle,
FreeFEM is capable of solving non-self-adjoint problems as well). The output parameter k of
EigenValue gives the number of eigenvalues actually computed; in most cases it will coincide with
the requested number of eigenvalues N, that is, the length of the output array Evalues declared
294 Appendix A. A short tutorial on numerical spectral geometry

earlier in line 10. If one is not interested in eigenfunctions but only in eigenvalues, line 52 and the
input vector=Efunctions may be omitted.
Finally, group H provides the output: first, the number of eigenvalues computed and the
eigenvalues themselves are printed to the standard output (that is, the screen) cout in line 59 (seee
FreeFEM manual for details on output to a file), and then the contour plot of the sixth eigenfunc-
tion is plotted in line 62; press "?" on the plot for help on changing its appearance.
Everything going to plan, one should see, after executing the script, output similar to
-- mesh : Nb of Triangles = 20768 , Nb of Vertices 10573
Real symmetric eigenvalue problem : A * x - B * x * lambda
We asked for 50 eigenvalues and computed 50 eigenvalues :
50
-7.779660247 e -15 1.000000001 1.000000001 2.000000011 4.00000009
4.000000092 5.000000167 5.00000017 8.000000683 9.000001004
9.000001013 10.00000134 10.00000137 13.00000296 13.00000304
16.00000569 16.00000585 17.00000663 17.00000679 18.00000783
20.00001082 20.00001104 25.00002128 25.00002163 25.00002211
25.00002219 26.00002377 26.00002461 29.00003341 29.00003427
32.00004473 34.0000529 34.00005502 36.00006399 36.00006524
37.00006881 37.00006912 40.00008553 40.00008746 41.00009245
41.00009489 45.00012209 45.00012669 49.00016273 49.00016362
50.00016672 50.00016833 50.00017239 52.00019103 52.00019161
times : compile 0.011743 s , execution 4.45535 s , mpirank :0
######## ...
Ok : Normal End

One can see that the accuracy of FreeFEM is, at least in this case, very reasonable!

Numerical Exercise A.2.2

Experiment with modifying the script from Listing A.6: vary the parameter L, the number
of mesh points per unit length of the boundary npoints, the requested number of eigen-
values N, and change the type of finite element from P2 to P1, in various combinations, to
see how these modifications affect computational accuracy and time.

§A.2.3. Curvilinear boundaries and holes

We now discuss, first, how to modify the FreeFEM script from Listing A.6 in order to compute
the Neumann eigenvalues of the domain Ω′ given by (A.1.6). To do so, we need to change the
definition of the boundary piece d3Omega in line 20 to

20 border d3Omega ( t = pi *L , 0) { x = t ; y = pi + t *( pi - t ) / pi ; label =1;

};

As the boundary piece d3Omega is now slightly longer, we may additionally increase the number
of mesh points on this piece by using . . . +d3Omega(1.2*npoints*L*pi)+. . . in line 28.
§A.2. Learning FreeFEM by example 295

Secondly, to incorporate additionally the circular hole and thus consider the domain Ω de-
fined by (A.1.6), we add to group B the command

border d5Omega ( t =0 , 2* pi ) { x = pi /3 + ( pi /4) * cos ( t ) ; y = pi /2 -

( pi /4) * sin ( t ) ; label =1;}

Note that in order to keep the domain to the left of this part of the boundary as t changes from
0 to 2π we parametrise the circle clockwise. We now change the mesh creation command to

mesh Th = buildmesh ( d1Omega ( npoints * L * pi ) + d2Omega ( npoints * pi ) +

d3Omega (1.2* npoints * L * pi ) + d4Omega ( npoints * pi ) + d5Omega (2* pi * pi
/4* npoints ) ) ;

Sample scripts may be downloaded following the links in §A.3.

§A.2.4. The Dirichlet, Zaremba, and Robin problems

The Dirichlet conditions are imposed at the stage of defining the quadratic form varf q: if all the
boundary pieces have the same label label=1;, then changing the definition of q to

varf q (u , v ) = int2d ( Th ) ( dx ( u ) * dx ( v ) + dy ( u ) * dy ( v ) ) + on (1 , u =0) ;

will impose the Dirichlet conditions on the whole boundary. The general format of the on com-
mand is +on(some_label, u=0) or +on(some_label, another_label, ..., last_label, u=0), the
latter version imposing Dirichlet conditions on boundaries with all the labels listed. It is impor-
tant to note (and to remember) that the Dirichlet boundary conditions are imposed only in the
definition of varf q and not of varf b, and only on the first variable of the quadratic form, in our
case u.
For Zaremba problem, we use the same approach but we have to change the labels of the
boundary pieces where we do not want to impose the Dirichlet condition to something else, say
label=2;, and keep on(1, u=0) in the form definition.
The Robin boundary conditions are also imposed by modifying the quadratic form varf q
according to (3.1.16): to impose this condition with γ = 2, say, we change the definition of q to

real gamma = 2.;

varf q (u , v ) = int2d ( Th ) ( dx ( u ) * dx ( v ) + dy ( u ) * dy ( v ) ) + int1d ( Th ,1) (
gamma * u * v ) ;

It is important to know that the factor γ should appear inside the integral according to FreeFEM
syntax.
For the sample scripts, see §A.3.
296 Appendix A. A short tutorial on numerical spectral geometry

§A.2.5. The Laplace–Beltrami operator on manifolds

FreeFEM can additionally handle periodic boundary conditions, which allows us to solve some
problems on Riemannian manifolds. All the examples in this subsection can be also done analyti-
cally (thus allowing an easy control on the accuracy of the numerics) but we encourage the reader
to modify them further in order to create more interesting examples, see also [LevStr21] for an
illustration of the use of FreeFEM in computations of eigenvalues and resonances on hyperbolic
manifolds.
We start from our basic script for a Neumann problem in a rectangle (Listing A.6), and relabel
the sides individually by replacing the lines 18–21 by

18 border d1Omega ( t =0 , L ) { x = pi * t ; y = 0; label =1; };

19 border d2Omega ( t =0 , 1) { x = pi * L ; y = pi * t ; label =2; };
20 border d3Omega ( t =L , 0) { x = pi * t ; y = pi ; label =3; };
21 border d4Omega ( t = pi , 0) { x = 0; y = t ; label =4; };

We now want to identify the sides labelled 2 and 4, thus turning the problem into the one on a
flat cylinder. This is achieved at the stage of declaring the FEM space, using FreeFEM command
periodic, by replacing the original line 37 with

37 fespace Vh ( Th , P2 , periodic =[[4 , y ] ,[2 , y ]]) ;

which basically tells FreeFEM to identify the value of y on sides 4 and 2.

To solve instead the spectral problem on the flat torus, we additionally have to identify sides
1 and 3 by replacing line 37 with

37 fespace Vh ( Th , P2 , periodic =[[4 , y ] ,[2 , y ] ,[1 , x ] ,[3 , x ]]) ;

If we instead identify sides 4 and 2 by the mapping y 7→ π − y as in

37 fespace Vh ( Th , P2 , periodic =[[4 , y ] ,[2 , pi - y ]]) ;

we solve the Neumann problem on the Möbius strip.

We note that the boundary pieces identified by periodic command need not be either parallel,
or straight, or even have the same length (but must have the same number of boundary mesh
points).
The sample scripts are listed in §A.3. We remark that Mathematica is also able to handle
periodic boundary conditions via PeriodicBoundaryCondition28 command, see a sample script.
We finish this subsection by showing how to compute, in FreeFEM, the eigenvalues of the
Laplace–Beltrami operator on the sphere S2 . This may be done in several ways; in order to sim-
plify the calculations we will use, first of all, a symmetry trick from §3.2.2 and decompose the
28 https://reference.wolfram.com/language/ref/PeriodicBoundaryCondition.html
 §A.2. Learning FreeFEM by example 297

spectrum into the union of spectra of the two problems on the hemisphere, one with the Neu-
mann condition imposed on the boundary, and another one with the Dirichlet one. We now use
the stereographic projection of the hemisphere onto the unit disk arriving at the Dirichlet and
Neumann problems for
−∆u = c(x, y)λu, (A.2.1)
where the conformal factor c is given by
4
c(x, y) = ,
(1 + x 2 + y 2 )2
and the Laplacian in the left hand-side of (A.2.1) is the usual Cartesian one. Thus, when formu-
lating the corresponding weak problems we need to replace the quadratic form (A.1.2) with
ˆ
B[u, v] := c(x, y)u(x, y)v(x, y) dxdy.
D

The final trick, in order to solve two problems simultaneously, is to solve (A.2.1) in the disjoint
union of two unit disks centred at (0, ±2), with the Dirichlet condition imposed on one of the
circles, and to adjust the conformal factor to
¶ ¶−2
y 2
µ µ
2
c(x, y) = 4 1 + x + y − 2 .
|y|
The resulting script (this time, uncommented) is shown in Listing A.7.

Listing A.7: The spectrum of the Laplace–Beltrami operator on S2

1 // This FreeFEM script computes eigenvalues of the Laplace -

Beltrami operator on the unit sphere
2 //
3 // ---- A . DECLARATIONS ----
4 int npoints =30;
5 int N =50;
6 real [ int ] Evalues ( N ) ;
7 real t ;
8 // ---- END OF DECLARATIONS so far ----
9 //
10 // ---- B . BOUNDARY DEFINITIONS ----
11 border dOmega1 ( t =0 , 2* pi ) { x = cos ( t ) ; y = 2+ sin ( t ) ; label =1;
};
12 border dOmega2 ( t =0 , 2* pi ) { x = cos ( t ) ; y = -2+ sin ( t ) ; label =2;
};
13 // ---- END OF BOUNDARY DEFINITIONS ----
14 //
15 // ---- C . MESH CREATION USING buildmesh COMMAND ----
16 mesh Th = buildmesh ( dOmega1 ( npoints *2* pi ) + dOmega2 ( npoints *2* pi ) ) ;
17 // ---- END OF MESH DEFINITIONS ----
298 Appendix A. A short tutorial on numerical spectral geometry

18 //
19 // ---- VISUALISE THE MESH , may be commented out
20 plot ( Th , wait =1) ;
21 //
22 // ---- D . DECLARE THE FEM SPACE AND FEM VARIABLES ----
23 fespace Vh ( Th , P2 ) ;
24 Vh u , v ;
25 //
26 // ---- E . DECLARE THE QUADRATIC FORMS ----
27 varf q (u , v ) = int2d ( Th ) ( dx ( u ) * dx ( v ) + dy ( u ) * dy ( v ) ) + on (1 , u =0) ;
28 varf b (u , v ) = int2d ( Th ) (4/(1+ x ^2+( y -2* y / abs ( y ) ) ^2) ^2* u * v ) ;
29 //
30 // ---- F . CREATE THE MATRICES ----
31 matrix S = q ( Vh , Vh ) ;
32 matrix M = b ( Vh , Vh ) ;
33 //
34 // ---- DECLARE THE ARRAY TO HOLD EIGENFUNCTIONS ----
35 Vh [ int ] Efunctions ( N ) ;
36 //
37 // ---- G . SOLVE THE PROBLEM ----
38 int k = EigenValue (S ,M , sym = true , value = Evalues , vector = Efunctions ) ;
39 //
40 // ---- H . PRINT THE EIGENVALUES ----
41 cout << Evalues ;
42 //

Executing this script produces an output similar to

-- mesh : Nb of Triangles = 12342 , Nb of Vertices 6361
Real symmetric eigenvalue problem : A * x - B * x * lambda
50
-9.550276753 e -15 2.000139729 2.000139735 2.000282213 6.000350531
6.000350568 6.000701052 6.000707766 6.000707795 12.00061962
12.00061991 12.00125069 12.0012509 12.0013633 12.00136486
12.00150391 20.00095108 20.00095132 20.00191427 20.00191754
20.00217762 20.00217802 20.00245631 20.00247364 20.0024823
30.00136367 30.00136728 30.00273512 30.00273578 30.00318084
30.00319296 30.00368812 30.00369269 30.00378001 30.00380983
30.00397306 42.00190156 42.00190932 42.00377552 42.00378548
42.00447397 42.00449093 42.00524994 42.00527772 42.00551986
42.00553103 42.00587206 42.00590262 42.00600163 56.00263505
times : compile 0.010895 s , execution 3.14007 s , mpirank :0
######## ...
Ok : Normal End

This is in a good agreement with Theorem 1.2.16 which gives in this case the eigenvalues k(k + 1),
k ∈ {0} ∪ N of multiplicity 2k + 1.
 §A.2. Learning FreeFEM by example 299

§A.2.6. The Steklov problem, the sloshing problem, and the spectrum of the Dirichlet-
to-Neumann map
Our example domain in this subsection is the half-disk
D − = {(x, y) ∈ R2 : x 2 + y 2 < 1, y < 0}.

To find the eigenvalues of the Steklov problem in D − we need to recall its weak formulation
(7.1.5). Therefore, we need to re-define the form B in this case as
ˆ
B[u, v] := uv ds. (A.2.2)
∂D −

Otherwise, the treatment is standard, see Listing A.8.

Listing A.8: The spectrum of the Steklov problem in the half-disk

1 // This FreeFEM script computes eigenvalues of the Steklov

problem in the half - disk .
2 //
3 // ---- A . DECLARATIONS ----
4 int npoints =30;
5 int N =50;
6 real [ int ] Evalues ( N ) ;
7 real t ;
8 // ---- END OF DECLARATIONS ----
9 //
10 // ---- B . BOUNDARY DEFINITIONS ----
11 border d1Omega ( t = pi , 2* pi ) { x = cos ( t ) ; y = sin ( t ) ; label =1; };
12 border d2Omega ( t =1 , -1) { x = t ; y = 0; label =2; };
13 // ---- END OF BOUNDARY DEFINITIONS ----
14 //
15 // ---- C . MESH CREATION USING buildmesh COMMAND ----
16 mesh Th = buildmesh ( d1Omega ( npoints * pi ) + d2Omega ( npoints *2) ) ;
17 // ---- END OF MESH DEFINITIONS ----
18 //
19 // ---- VISUALISE THE MESH , may be commented out
20 plot ( Th , wait =1) ;
21 //
22 // ---- D . DECLARE THE FEM SPACE AND FEM VARIABLES ----
23 fespace Vh ( Th , P2 ) ;
24 Vh u , v ;
25 //
26 // ---- E . DECLARE THE QUADRATIC FORMS ----
27 varf q (u , v ) = int2d ( Th ) ( dx ( u ) * dx ( v ) + dy ( u ) * dy ( v ) ) ;
28 varf b (u , v ) = int1d ( Th ,1 ,2) ( u * v ) ; // note that b changes for
Steklov and that we integrate over both parts of the boundary
29 //
30 // ---- F . CREATE THE MATRICES ----
31 matrix S = q ( Vh , Vh ) ;
300 Appendix A. A short tutorial on numerical spectral geometry

32 matrix M = b ( Vh , Vh ) ;
33 //
34 // ---- G . SOLVE THE PROBLEM ----
35 int k = EigenValue (S ,M , sym = true , value = Evalues ) ;
36 //
37 // ---- H . PRINT THE EIGENVALUES ----
38 cout << Evalues ;
39 //
40 // ---- END --------

To consider the sloshing problem in D − , with the Steklov condition on the straight part of
the boundary and the Neumann condition on the arc, we just need to adjust the definition of
the form B in (A.2.2) in order to integrate over the straight part of the boundary only, therefore
replacing line 28 in Listing A.8 by

28 varf b (u , v ) = int1d ( Th ,2) ( u * v ) ;

Finally, to compute the spectrum of the Dirichlet-to-Neumann map DΛ for a given value of
Λ (say, 1.5), we recall the weak statement (7.4.4) and replace line 27 in Listing A.8 by

27 real Lambda =1.5; varf q (u , v ) = int2d ( Th ) ( dx ( u ) * dx ( v ) + dy ( u ) * dy ( v

) - Lambda * u * v ) ;

leaving line 28 unchanged. Note that if Λ is chosen very close to but lower than a Dirichlet eigen-
value of D − , some low negative eigenvalues of DΛ may be lost.

§A.3. List of downloadable scripts

All scripts mentioned in this appendix are available for download from

https://michaellevitin.net/Book/Scripts

or by clicking directly on the script name. The domains Ω′ and Ω are defined by (A.1.6).

Filename Description Reference

script1.nb Mathematica: computing eigenvalues analytically Listing A.1
script2.nb Mathematica: Neumann and Dirichlet eigenvalues of Listing A.2
Ω
script3.nb Mathematica: Robin eigenvalues of Ω Listing A.3
script4.nb Mathematica: Zaremba eigenvalues of Ω Listing A.4
script5.nb Mathematica: verifying the Faber–Krahn inequality Listing A.5
for regular n -gons
script6.edp FreeFEM: Neumann eigenvalues of (0, π)2 Listing A.6
script7.edp FreeFEM: Neumann eigenvalues of Ω′ §A.2.3
§A.3. List of downloadable scripts 301

Filename Description Reference

script8.edp FreeFEM: Neumann eigenvalues of Ω §A.2.3
script9.edp FreeFEM: Dirichlet eigenvalues of Ω §A.2.4
script10.edp FreeFEM: Zaremba eigenvalues of Ω §A.2.4
script11.edp FreeFEM: Robin eigenvalues of Ω §A.2.4
script12.edp FreeFEM: Laplace–Beltrami eigenvalues of a flat cylinder §A.2.5
script13.edp FreeFEM: Laplace–Beltrami eigenvalues of a flat torus §A.2.5
script14.edp FreeFEM: Laplace–Beltrami eigenvalues of a Möbius §A.2.5
strip
script15.nb Mathematica: eigenvalues of periodic problems §A.2.5
script16.edp FreeFEM: Laplace–Beltrami eigenvalues of S2 Listing A.7
script17.edp FreeFEM: Steklov eigenvalues in the half-disk Listing A.8
script18.edp FreeFEM: Sloshing eigenvalues in the half-disk §A.2.6
script19.edp FreeFEM: Eigenvalues of DΛ in the half-disk §A.2.6
 APPENDIX B
Background definitions and notation
We list here some of the standard definitions and notation used
throughout the book.

§B.1. Sets
We use the standard symbols N, Z, R, C, for the sets of natural, integer, real, and complex numbers,
respectively. Our natural numbers do not include zero. We sometimes write

N0 := N ∪ {0}

and
R+ := (0, +∞).
The coordinates of a point x ∈ Rd are usually denoted by (x 1 , . . . , x d ). For x, y ∈ Rd , we write

­ ® X d
x, y := xj yj
j =1

for the usual dot product; we use the same notation in Cd with
q
the additional complex conjuga-
tion over y j . The length of a vector y ∈ Rd is written as |y| =
­ ®
y, y .
The complement of a set X ⊂ Rd is denoted by X c := Rd \ X . The closure of an open set
U ⊂ Rd is denoted by U and its boundary by

∂U = U \U .

Throughout this book, we say that Ω ⊂ Rd is a domain if it is a non-empty connected open

set.
We let n o
d
B a,r = B a,r = x ∈ Rd : |x − a| < r

303
304 Appendix B. Background definitions and notation

denote the ball in Rd with centre a and radius r . We will also write

B rd := B 0,r
d

for the ball centred at the origin (or whenever the position of the centre is irrelevant), and

Bd := B 1d = B 0,1
d

for the unit ball in Rd . In the planar case, we will also use D := B2 for the unit disk.
We denote the volume of the unit ball by
d
d π2
ωd = Vold (B ) = ³ ´, (B.1.1)
Γ d2 + 1

where Γ is the Gamma function.

Similarly, n o
S da,r−1 = S a,r = x ∈ Rd : |x − a| = r

denotes the sphere in Rd with centre a and radius r , and will also write S rd −1 = S r := S 0,r when
it is centred at the origin (or when the position of the centre is irrelevant). We denote the unit
sphere in Rd by
Sd −1 := S 1d −1 ,
and its (d − 1)-dimensional volume by
d
d −1 2π 2
σd −1 := Vold −1 (S ) = d ωd = ³ ´ . (B.1.2)
Γ d2

§B.2. Function spaces

If U is an open subset of Rd , we denote by L p (U´), 1 ≤ p < ∞, the set of all Lebesgue measurable
functions u : U → R (or u : U → C) such that U |u(x)|p dx < ∞. The space L p (U ), equipped
with the norm
ˆ
 1/p

∥u∥L p (U ) :=  |u(x)|p dx 
U

is a Banach space, in which we identify elements which coincide almost everywhere. Similarly,
L ∞ (U ) is the Banach space of all essentially bounded functions u on U with

∥u∥L ∞ (U ) := ess sup u(x) < ∞.

x∈U

We also define
p
L loc (U ) := {u : u|K ∈ L p (K ) for all compact K ⊂ U }.
§B.2. Function spaces 305

In the special case p = 2, L 2 (U ) equipped with the inner product

ˆ
(u, v)L 2 (U ) := u(x)v(x) dx
U

is a Hilbert space. Since we are mostly dealing with real-valued functions, we will usually omit the
complex conjugation.
For an open set U ⊂ Rd , let C (U ) denote the space of continuous functions on u . We denote
the partial derivatives of of a function u (if they exist) by

∂|α| u
∂α u := α
, (B.2.1)
∂x 1 1 . . . ∂x d α
d

for a multi-index α = (α1 , . . . , αd ) ∈ Nd0 , where |α| := α1 + · · · + αd is the order of the derivative.
We also write
∂u
∂ j u := .
∂x j
We denote the space of k -times continuously differentiable functions on U by

C k (U ) :=
{u : U → R : ∂α u exists and is continuous in U for all α with |α| ≤ k},

for k ∈ N0 ; obviously, C 0 (U ) = C (U ).
For u : U → R, u ∈ C 1 (U ), we denote its gradient by

∂u ∂u
µ ¶
∇u := ,..., ,
∂x 1 ∂x d

and for a vector-valued function f : U → Rd , f = ( f 1 , . . . , f d ) ⊂ C 1 (U ), we denote its divergence

by
d ∂f
X j
div f := .
j =1 ∂x j

We also set
C ∞ (U ) := {u : u ∈ C k (U ) for all k ∈ N0 }
and
C 0k (U ) := {u ∈ C k (U ) : supp u ⋐ U },
C 0∞ (U ) := {u : u ∈ C 0k (U ) for all k ∈ N0 },

where X ⋐ Y means that X is a compact subset of Y . Further on, we define

C k (U ) := {u ∈ C k (U ) :
∂α u can be continuously extended to ∂U for all α with |α| ≤ k}.
306 Appendix B. Background definitions and notation

We say that u ∈ C (U ) is Hölder continuous with exponent β ∈ (0, 1] if there exists a constant
C ≥ 0 such that
|u(x) − u(y)| ≤ C |x − y|β for all x, y ∈ U .
We sometimes use C k,β (U ) to denote the subspace of functions from C k (U ) whose derivatives
of order k are Hölder continuous with exponent β. We say that u is Lipschitz continuous (or just
Lipschitz) if it is Hölder continuous with exponent one; thus the space of all Lipschitz continuous
functions on U coincides with C 0,1 (U ).
The Schwartz space of rapidly decreasing functions on Rd is defined as

S (Rd ) :=
( )
¯ ¯
∞ d ¯ α β ¯ d
u ∈ C (R ) : sup ¯x ∂ u ¯ < ∞ for all multi-indices α, β ∈ N0 ,
x∈Rd

where x α := x 1α1 · · · x dαd .

We will use the shorthand notation L 2 (U ) = (L 2 (U ))d , C k (U ) = (C k (U ))d , etc., for the
spaces of vector-valued functions f = ( f 1 , . . . , f d ) : U → Rd .

§B.3. Regularity of the boundary

We follow [McL00] and [ChWGLS12, Appendix A]. Let U ⊂ Rd be a non-empty open set. We
say that its boundary ∂U is Lipschitz if ∂U is compact, and there exist finite families of sets {Wi }
and {Ui }, and of functions { f i }, of the same cardinality, such that

(i) the family {Wi }, Wi ⊂ Rd , is a finite open cover of ∂U ;

(ii) the family {Ui }, Ui ⊂ Rd , is such that Ui ∩ Wi = U ∩ Wi ;

(iii) each function f i : Rd −1 → Rd is Lipschitz continuous;

(iv) for each i there exists a rigid motion R i : Rd → Rd such that

R i (Ui ) = {(x ′ , x d ) ∈ Rd −1 × R : x d > f i (x ′ )}.

We will sometimes say that Ω is a Lipschitz domain if it is a domain with a Lipschitz boundary
∂Ω.
In the same manner, we define the C k or C ∞ boundaries by replacing in part (iii) of the above
definition the family of Lipschitz functions { f i } by a family of C k or C ∞ functions, respectively.
We mention that, for example, all polyhedra have Lipschitz boundary (but not C 1 ), whereas
domains with cusps or slits are not Lipschitz.
 List of margin portraits

Author, attribution, licence

(where known and as given by the source of download),
Name Page
and the source of download.
CC = Creative Commons
Beltrami unknown, Public Domain, image from Wikipedia 24
Berezin unknown, image from geni.com 106
Berger by Konrad Jacobs, MFO, CC BY-SA 2.0 de, image from Wikipedia 186
Bers photograph by Paul Halmos, image from MacTutor 147
Bessel by Christian Albrecht Jensen — Ny Carlsberg Glyptotek, Public 20
Domain, image from Wikipedia
Bolza unknown, Public Domain, image from Wikipedia 186
Bourgain photo by George M. Bergman, CC BY-SA 4.0, image from 28
Wikipedia
Brooks unknown, image from Technion memorial site 2
Chladni unknown, Public Domain, image from Wikipedia 112
Courant by Konrad Jacobs, MFO, CC BY-SA 2.0 de, image from Wikipedia 115
Dirichlet unknown, Public Domain, image from Wikipedia 13
Faber unknown, image from MacTutor 152
Fourier by Louis-Léopold Boilly — Getty Images, Public Domain, image 13
from Wikipedia
Fraenkel by Wolfups — own work, CC BY-SA 4.0, image from Wikipedia 159
Friedrichs by Martinfriedrichs at English Wikipedia, Public Domain, image 45
from Wikipedia
Galerkin unknown, image from Wikipedia 281
Gauss by Christian Albrecht Jensen, Public Domain, image from 28
Wikipedia
Germain unknown, image from MacTutor 112
Glazman unknown, image from website of Applied Mathematics Department 90
of the National Technical University “Kharkiv Polytechnic Institute"
Greenhill unknown, image from MacTutor 233
Hardy unknown, Public Domain, image from Wikipedia 28
Hersch unknown, image from the Swiss Mathematical Society 180
Hilbert unknown — possibly in C. Reid, Hilbert (1970), Public Domain, im- 44
age from Wikipedia

307
308 List of margin portraits

Author, attribution, licence

(where known and as given by the source of download),
Name Page
and the source of download.
CC = Creative Commons
Hörmander by Konrad Jacobs, MFO, CC BY-SA 2.0 de, image from Wikipedia 245
Kac by Konrad Jacobs, CC BY-SA 2.0 de, image from Wikipedia 204
Klein by Gebruder Noelle — Archivio storico dell’Accademia delle Scienze 186
di Torino, Public Domain, image from Wikipedia
Krahn by Heinrich Riedel — TÜR fotokogu, Public Domain, image from 152
Wikipedia
Lagrange unknown, Public Domain, image from Wikipedia 283
Lamb by Maul & Fox, see B. Launder, Horace Lamb and the circumstances of 233
his appointment at Owens College, Notes Rec. Royal Soc. 67 (2012),
139–158, Public Domain, image from Wikipedia
Lamé unknown, Public Domain, image from Wikipedia 87
Laplace by J. Posselwhite, Public Domain, image from Wikipedia 11
Lidskii unknown, photo curtesy of family for arXiv: 1008.2669 2
Liouville unknown, Public Domain, image from Wikipedia 14
Lipschitz unknown, Public Domain, image from Wikipedia 18
Minakshisundaram unknown, image from MacTutor 203
Neumann by Photo Deutsches Museum — Green’s Functions with Applica- 15
tions by Dean G. Duffy, Second Edition, CC BY-SA 4.0, image from
Wikipedia
Nirenberg by Konrad Jacobs, CC BY-SA 2.0 de, image from Wikipedia 53
Payne unknown, image from archived personal website at Cornell Univer- 188
sity
Pleijel unknown, image from MacTutor 118
Pohozhaev unknown, image from personal website at Steklov Institute 245
Poincaré unknown — Popular Science Monthly 82 (1913), Public Domain, im- 43
age from Wikipedia
Poisson by François-Séraphin Delpech, Public Domain, image from Smith- 208
sonian
Pólya unknown, image from MacTutor 102
Ramanujan unknown, Public Domain, image from Wikipedia 29
Rayleigh unknown, Public Domain, image from Wikipedia 67
Rellich by Konrad Jacobs — MFO, CC BY-SA 2.0 de, image from Wikipedia 244
Riemann unknown, Public Domain, image from Wikipedia 22
Ritz unknown, Public Domain, image from Wikipedia 67
Saint-Venant engraving in Les travaux publics et les ingénieurs des ponts et chaussées 160
depuis le XVIIe siècle by A. Debauve, Bibliothèque Nationale De
France, Public Domain, image from Wikipedia
Schiffer unknown, image from MacTutor 239
Schrödinger unknown, Public Domain, image from Wikipedia 12
Sobolev by Konrad Jacobs — MFO, CC SA 1.0, image from Wikipedia 37
Steklov unknown, Public Domain, image from Russian Academy of Sci- 228
ences
Sturm portrait by Jean-Daniel Colladon, image from Wikipedia 14
Szegő unknown, image from MacTutor 174
List of margin portraits 309

Author, attribution, licence

(where known and as given by the source of download),
Name Page
and the source of download.
CC = Creative Commons
Weinberger unknown, image from archived personal website at the University of 174
Minnesota
Weyl by ETH Zürich — ETH-Bibliothek Zürich, Bildarchiv, CC BY-SA 94
3.07, image from Wikipedia
Zaremba unknown, Public Domain, image from Wikipedia 74
 Bibliography

Q W E R T Y U I O P
A S D F G H J K L
Z X C V B N M
[AdaFou03] R. A. Adams and J. J. Fournier, Sobolev Spaces. Pure and Applied Mathematics (Am-
sterdam) 140. Elsevier/Academic Press, Amsterdam, 2003. ☞ pp. 37 and 38 A
[Agm65] S. Agmon, Unicité et convexité dans les problèmes différentiels, Séminaire de Math-
ématiques Supérieures, No. 13 (Été, 1965). Les Presses de l’Université de Montréal,
Montreal, Que., 1966. ☞ pp. 134 and 136
[Agr06] M. S. Agranovich, On a mixed Poincare-Steklov type spectral problem in a Lipschitz
domain, Russ. J. Math. Phys. 13:3 (2006), 239–244. doi: 10.1134/S1061920806030010.
☞ p. 244
[Alb71] J. H. Albert, Topology of the nodal and critical point sets for eigenfunctions of elliptic
operators, Ph. D. thesis, Massachusetts Institute of Technology, 1971. MIT Libraries.
☞ p. 147
[Alm00] F. J. Almgren Jr., Almgren’s big regularity paper. Q -valued functions minimizing
Dirichlet’s integral and the regularity of area-minimizing rectifiable currents up to
codimension 2. World Scientific Monograph Series in Mathematics 1. World Scien-
tific, River Edge, N. J., 2000. doi: 10.1142/4253. ☞ p. 134
[AndClu11] B. Andrews and J. Clutterbuck, Proof of the fundamental gap conjecture, J. Amer.
Math. Soc. 24:3 (2011), 899–916. doi: 10.1090/S0894-0347-2011-00699-1. ☞ p. 195
[AntFre12] P. Antunes and P. Freitas, Numerical optimization of low eigenvalues of the Dirich-
let and Neumann Laplacians, J. Optim. Theory Appl. 154:1 (2012), 235–257. doi:
10.1007/s10957-011-9983-3. ☞ p. 180
[AreCSVV18] W. Arendt, R. Chill, C. Seifert, H. Vogt, and J. Voigt, Form Methods for Evolution
Equations, and Applications, 18th Internet Seminar on Evolution Equations, Lecture
Notes, available online at seminar’s website. ☞ pp. 48 and 72

311
312 Bibliography

[AreMaz12] W. Arendt and R. Mazzeo, Friedlander’s eigenvalue inequalities and the Dirichlet-
to-Neumann semigroup, Commun. Pure Appl. Anal. 11:1 (2012), 2201–2212. doi:
10.3934/cpaa.2012.11.2201. ☞ pp. 229, 269, 270, 272, 273, 274, and 278
[AreMon95] W. Arendt and S. Monniaux, Domain perturbation for the first eigenvalue of the
Dirichlet Schrödinger operator, in Partial Differential Equations and Mathemati-
cal Physics (Holzhau, 1994), M. Demuth and B. Schulze, eds. Operator Theory: Ad-
vances and Applications 78, 9–19. Birkhäuser, Basel, 1995. doi: 10.1007/978-3-0348-
9092-2_2. ☞ p. 159
[ArmGar01] D. H. Armitage and S. J. Gardiner, Classical potential theory. Springer Monographs
in Mathematics. Springer-Verlag, London, 2001. doi: 10.1007/978-1-4471-0233-5.
☞ p. 61
[Aro57] N. Aronszajn, A unique continuation theorem for solutions of elliptic partial differ-
ential equations or inequalities of second order, J. Math. Pures Appl. (9) 36 (1957),
235–249. ☞ pp. 77, 119, and 138
[Ash99] M. S. Ashbaugh, Isoperimetric and universal inequalities for eigenvalues, in Spectral
theory and geometry (Edinburgh, 1998), E. B. Davies and Yu. Safarov, eds. London
Math. Soc. Lecture Note Ser. 273, 95–139. Cambridge Univ. Press, Cambridge, 1999.
doi: 10.1017/CBO9780511566165.007. ☞ p. 189
[AshBen89] M. S. Ashbaugh and R. Benguria, Optimal lower bound for the gap between the first
two eigenvalues of one-dimensional Schrödinger operators with symmetric single-well
potentials, Proc. Amer. Math. Soc. 105:2 (1989), 419–424. doi: 10.2307/2046959.
☞ p. 195
[AshBen91] M. S. Ashbaugh and R. D. Benguria, A sharp bound for the ratio of the first two eigen-
values of Dirichlet Laplacians and extensions, Ann. of Math. (2) 135:3 (1992), 601–628.
doi: 10.2307/2946578. ☞ p. 194
[AviDSiKal16] A. Avila, J. De Simoi, and V. Kaloshin, An integrable deformation of an ellipse of small
eccentricity is an ellipse, Ann. of Math. (2) 184:2 (2016), 527–558. doi: 10.4007/an-
nals.2016.184.2.5. ☞ p. 224
[AxlBouWad01] S. Axler, P. Bourdon, and R. Wade, Harmonic function theory. Graduate Texts
in Mathematics 137. Springer Science+Business Media, New York, 2001. doi:
10.1007/978-1-4757-8137-3, also at author’s website. ☞ pp. 30, 52, and 120
[BabOsb91] I. Babuška and J. Osborn, Eigenvalue Problems, Handbook Num. Anal. 2 (1991),
B 641–787. doi: 10.1016/S1570-8659(05)80042-0. ☞ p. 282
[BanParBSh09] R. Band, O. Parzanchevski, and G. Ben-Shach, The isospectral fruits of representation
theory: quantum graphs and drums, J. Phys. A 42:17 (2009), 175202. doi: 10.1088/1751-
8113/42/17/175202. ☞ p. 217
[Ban80] C. Bandle, Isoperimetric inequalities and applications. Monographs and Studies in
Mathematics 7. Pitman (Advanced Publishing Program), Boston–London, 1980.
☞ pp. 2, 66, 67, and 183
[BañCar94] R. Bañuelos and T. Carroll, Brownian motion and the fundamental frequency of
a drum, Duke Math. J. 75:3 (1994), 575–602. doi: 10.1215/S0012-7094-94-07517-0.
☞ p. 171
Bibliography 313

[BarBouLeb16] L. Baratchart, L. Bourgeois, and J. Leblond, Uniqueness results for inverse Robin
problems with bounded coefficient, J. Funct. Anal. 270:7 (2016), 2508–2542. doi:
10.1016/j.jfa.2016.01.011. ☞ p. 240
[Beb03] M. Bebendorf, A note on the Poincaré inequality for convex domains, Zeitschift Anal.
Anwend. 22:4 (2003), 751–756. doi: 10.4171/ZAA/1170. ☞ p. 195
[BehtEl15] J. Behrndt and A. F. M. ter Elst, Dirichlet-to-Neumann maps on bounded Lipschitz do-
mains, J. Differential Equations 259 (2015), 5903–5926. doi: 10.1016/j.jde.2015.07.012.
☞ pp. 269 and 270
[BelBBT18] F. Belgacem, H. BelHadjAli, A. BenAmor, and A. Thabet, Robin Laplacian in the
large coupling limit: convergence and spectral asymptotic, Ann. Sc. Norm. Super. Pisa
Cl. Sci. (5) 18:2 (2018), 565–591. doi: 10.2422/2036-2145.201601_008. ☞ p. 274
[BerDKZ18] B. C. Berndt, A. Dixit, S. Kim, and A. Zaharescu, Sums of squares and products
of Bessel functions, Adv. Math. 338 (2018), 305–338. doi: 10.1016/j.aim.2018.09.001.
☞ p. 29
[Ben15] B. Benson, The Cheeger constant, isoperimetric problems, and hyperbolic surfaces,
arXiv: 1509.08993. ☞ p. 170
[Bér86] P. Bérard, Spectral geometry: direct and inverse problems. Lecture Notes in Mathe-
matics 1207. Springer-Verlag, 1986. ☞ pp. 2 and 96
[Bér92] P. Bérard, Transplantation et isospectralité I, Math. Ann. 292:3 (1992), 547–559. doi:
10.1007/BF01444635. ☞ p. 214
[BérBes80] P. Bérard and G. Besson, Spectres et groupes cristallographiques. II : domaines
sphériques, Ann. Inst. Fourier (Grenoble) 30(3) (1980), 237–248. doi: 10.5802/aif.800.
☞ p. 101
[BérHel14] P. Bérard and B. Helffer, Nodal sets of eigenfunctions, Antonie Stern’s results revisited,
Sém. de Théorie Spect. et Géom., Grenoble 32 (2014–15), 1–37. doi: 10.5802/tsg.302.
☞ p. 120
[BérMey82] P. Bérard and D. Meyer, Inégalités isopérimétriques et applications, Annales Sci. École
Norm. Sup. (4) 15:3 (1982), 513–541. doi: 10.24033/asens.1435. ☞ pp. 119, 126, 147, 162,
and 163
[BerNirVar94] H. Berestycki, L. Nirenberg, and S. R. S. Varadhan, The principal eigenvalue and
maximum principle for second-order elliptic operators in general domains, Comm.
Pure Appl. Math. 47:1 (1994), 47–92. doi: 10.1002/cpa.3160470105. ☞ p. 127
[Ber72] F. A. Berezin, Convex operator functions, Math. USSR Sb. 17:2 (1972), 269–277. doi:
10.1070/SM1972v017n02ABEH001504. ☞ p. 106
[vdB83] M. van den Berg, On condensation in the free-boson gas and the spectrum of the Lapla-
cian, J. Statist. Phys. 31:3 (1983), 623–637. doi: 10.1007/BF01019501. ☞ p. 195
[vdBDryKap14] M. van den Berg, E. Dryden, and T. Kappeler, Isospectrality and heat content, Bull.
Lon. Math. Soc. 46:4 (2014), 793–808. doi: 10.1112/blms/bdu035. ☞ p. 219
[vdBSri88] M. van den Berg and S. Srisatkunarajah, Heat equation for a region in R2 with a polyg-
onal boundary, J. London Math. Soc. (2) 37:1 (1988), 119–127. doi: 10.1112/jlms/s2-
37.121.119. ☞ p. 207
314 Bibliography

[Ber73] M. Berger, Sur les premières valeurs propres des variétés riemanniennes, Compos.
Math. 26 (1973), 129–149. Numdam. ☞ p. 186
[BerGauMaz71] M. Berger, P. Gauduchon, and E. Mazet, Le spectre d’une variété riemannienne. Lec-
ture Notes in Mathematics 194. Springer-Verlag, Berlin-New York, 1971. ☞ pp. 2, 24,
25, 27, 30, 198, 201, 202, and 208
[BerGMR21] D. Berghaus, B. Georgiev, H. Monien, and D. Radchenko, On Dirichlet eigenvalues
of regular polygons, arXiv: 2103.01057. ☞ p. 286
[Ber17] G. Berkolaiko, An elementary introduction to quantum graphs, in Geometric and
Computational Spectral Theory, eds. A. Girouard, D. Jakobson, M. Levitin, N.
Nigam, I. Polterovich, and F. Rochon, Contemporary Mathematics 700, American
Mathematical Society, Providence, RI, and Centre de Recherches Mathématique,
Montréal, Québec, 2017, 41–72. doi: 10.1090/conm/700/14182. ☞ p. 255
[BerCHS22] G. Berkolaiko, G. Cox, B. Helffer, and M. P. Sundqvist, Computing nodal deficiency
with a refined Dirichlet-to-Neumann map, J. Geom. Anal. 32:10 (2022), 1–36. doi:
10.1007/s12220-022-00984-2. ☞ p. 163
[BerKuc13] G. Berkolaiko and P. Kuchment, Introduction to quantum graphs. Mathematical Sur-
veys and Monographs 186. American Mathematical Society, Providence, RI, 2013.
doi: 10.1090/surv/186. ☞ pp. 9, 252, and 255
[BerKucSmi12] G. Berkolaiko, P. Kuchment, and U. Smilansky, Critical partitions and nodal de-
ficiency of billiard eigenfunctions, Geom. Funct. Anal. 22:6 (2012), 1517–1540. doi:
10.1007/s00039-012-0199-y. ☞ p. 163
[BerGetVer04] N. Berline, E. Getzler, and M. Vergne, Heat kernels and Dirac operators. Grundlehren
Text Editions. Springer-Verlag, Berlin, Heidelberg 2004. ☞ p. 9
[Ber55] L. Bers, Local behavior of solutions of general linear elliptic equations, Comm. Pure
Appl. Math. 8 (1955) 473–496. doi: 10.1002/cpa.3160080404. ☞ p. 147
[Bes80] G. Besson, Sur la multiplicit́e de la première valeur propre des surfaces riemanniennes,
Ann. Inst. Fourier (Grenoble) 30 (1980), 109–128. doi: 10.5802/aif.777. ☞ p. 149
[Bob11] A. Bobenko, Introduction to compact Riemann surfaces, in Computational approach
to Riemann surfaces, A. Bobenko and C. Klein, eds. Lecture Notes in Mathemat-
ics 2013, 3–64. Springer, Berlin, Heidelberg, 2011. doi: 10.1007/978-3-642-17413-1_1
☞ pp. 181 and 184
[BhaWei99] T. Bhattacharya and A. Weitsman, Estimates for Green’s function in terms of asym-
metry, in Applied Analysis, J. R. Dorroh, G. R. Goldstein, J. A. Goldstein, and
M. M. Tom, eds. AMS Contemporary Math. Series 221, 31–58. Amer. Math. Soc.,
Providence, RI, 1999. doi: 10.1090/conm/221. ☞ p. 160
[Bor16] D. Borthwick, Spectral theory of infinite-area hyperbolic surfaces. Second edition.
Progress in Mathematics 318. Birkhäuser/Springer, Cham, 2016. doi: 10.1007/978-
3-319-33877-4. ☞ p. 9
[BouLev07] L. Boulton and M. Levitin, Trends and Tricks in Spectral Theory. Ediciones IVIC,
Caracas, 2007. ☞ p. 1
[Bou15] J. Bourgain, On Pleijel’s nodal domain theorem, Int. Math. Res. Not. 2015:6 (2015),
1601–1612. doi: 10.1093/imrn/rnt241. ☞ p. 163
Bibliography 315

[BouWat17] J. Bourgain and N. Watt, Mean square of zeta function, circle problem and divisor
problem revisited, arXiv: 1709.04340. ☞ p. 28
[BraDeP17] L. Brasco and G. De Philippis, Spectral inequalities in quantitative form, in Shape
optimization and spectral theory, A. Henrot, ed., 201–281. De Gruyter Open Poland,
2017. doi: 10.1515/9783110550887-007. ☞ pp. 179 and 239
[BraDePVel15] L. Brasco, G. De Philippis and B. Velichkov, Faber–Krahn inequalities in sharp quan-
titative form, Duke Math. J. 164:9 (2015), 1777–1831. doi: 10.1215/00127094-3120167.
☞ p. 160
[BraPra12] L. Brasco and A. Pratelli, Sharp stability of some spectral inequalities, Geom. Funct.
Anal. 22:1 (2012), 107–135. doi: 10.1007/s00039-012-0148-9. ☞ p. 179
[Bre11] H. Brezis, Functional analysis, Sobolev spaces and partial differential equations. Uni-
versitext, Springer, New York, 2011. ☞ pp. 38, 54, and 59
[BriButPri22] L. Briani, G. Buttazzo, and F. Prinari, On a class of Cheeger inequalities, Ann. Mat.
Pura Appl. (2022). doi: 10.1007/s10231-022-01255-1. ☞ p. 173
[Bro01] F. Brock, An isoperimetric inequality for eigenvalues of the Stekloff problem, ZAMM
Z. Angew. Math. Mech. 81:1 (2001), 69–71.
doi: 10.1002/1521-4001(200101)81:1<69::AID-ZAMM69>3.0.CO;2-# ☞ p. 239
[Bro88] R. Brooks, Constructing isospectral manifolds, Amer. Math. Monthly 95:8 (1988),
823–839. doi: 10.1080/00029890.1988.11972094. ☞ p. 211
[Bro98] R. Brooks, The Sunada method, in Tel Aviv Topology Conference: Rothenberg
Festschrift (1998). Contemp. Math. 231, 25–35. Amer. Math. Soc., Providence, RI,
1999. doi: 10.1090/conm/231/03350. ☞ p. 211
[Bro93] R. M. Brown, The trace of the heat kernel in Lipschitz domains, Trans. Amer. Math.
Soc. 339:2 (1993), 889–900. doi: 10.2307/2154304. ☞ p. 224
[Brü78] J. Brüning, Über Knoten von Eigenfunktionen des Laplace-Beltrami-Operators, Math.
Z. 158:1 (1978), 15–21. doi: 10.1007/BF01214561. ☞ p. 131
[BrüGro72] J. Brüning and D. Gromes, Über die Länge der Knotenlienien schwingender Mem-
branen, Math. Z. 124 (1972), 79–82. doi: 10.1007/BF01142586. ☞ p. 129
[BucFNT21] D. Bucur, V. Ferone, C. Nitsch, and C. Trombetti, Weinstock inequality in higher
dimensions, J. Differential Geom. 118:1 (2021), 1–21. doi: 10.4310/jdg/1620272940.
☞ p. 238
[BucFreKen17] D. Bucur, P. Freitas, and J. Kennedy, The Robin problem, in Shape optimization
and spectral theory, A. Henrot, ed., 78–119. De Gruyter Open Poland, 2017. doi:
10.1515/9783110550887-004. ☞ pp. 72, 73, and 273
[BucHen19] D. Bucur and A. Henrot, Maximization of the second non-trivial Neumann
eigenvalue, Acta Math. 222:2 (2019), 337–361. doi: 10.4310/ACTA.2019.v222.n2.a2.
☞ p. 179
[BucNah21] D. Bucur and M. Nahon, Stability and instability issues of the Weinstock inequality,
Trans. Amer. Math. Soc. 374:3 (2021), 2201–2223. doi: 10.1090/tran/8302. ☞ p. 239
[Buh16] L. Buhovsky, Lecture notes on spectral geometry (2016), private communication.
☞ pp. 2 and 115
316 Bibliography

[Bur98] F. E. Burstall, Basic Riemannian geometry, in Spectral theory and geometry

(Edinburgh, 1998), E. B. Davies and Yu. Safarov, eds. London Math. Soc.
Lecture Note Ser. 273, 1–29. Cambridge Univ. Press, Cambridge, 1999. doi:
10.1017/CBO9780511566165.004, also at author’s website. ☞ pp. 22, 24, 132, 168,
and 169
[Bus80] P. Buser, On Cheeger’s inequality λ1 ≥ h 2 /4, in Geometry of the Laplace Opera-
tor, Proc. Symp. Pure Math. 36, 29–77. Amer. Math. Soc., Providence, R.I., 1980.
☞ pp. 165 and 168
[Bus82] P. Buser, A note on the isoperimetric constant, Ann. Sci. École Norm. Sup. (4) 15:2
(1982), 231–230. doi: 10.24033/asens.1426. ☞ pp. 167 and 169
[BusCDS94] P. Buser, J. Conway, P. Doyle, and K.-D. Semmler, Some planar isospec-
tral domains, Internat. Math. Res. Notices 1994:9 (1994), 391–400. doi:
10.1155/S1073792894000437. ☞ pp. 219 and 220
[CadFar18] L. Cadeddu and M. Farina, A brief note on the coarea formula, Abh. Math. Semin.
C Univ. Hambg. 88:1 (2018), 193–199. doi: 10.1007/s12188-017-0183-4. ☞ pp. 155 and 165
[Can13] Y. Canzani, Analysis on manifolds via the Laplacian, notes of the course given at
Harvard University (2013), available online at author’s website. ☞ pp. 2 and 27
[Car33] T. Carleman, Sur une inégalité différentielle dans la théorie des fonctions analytiques,
C. R. Acad. Sci., Paris 196 (1933), 995–997. ☞ p. 134
[ChWGLS12] S. N. Chandler-Wilde, I. G. Graham, S. Langdon, and E. A. Spence, Numerical-
asymptotic boundary integral methods in high-frequency acoustic scattering, Acta Nu-
merica 21 (2012), 89–305. doi: 10.1017/S0962492912000037. ☞ pp. 39, 40, 62, 64,
228, 244, 246, and 306
[Cha95] S. J. Chapman, Drums that sound the same, Amer. Math. Monthly 102:2 (1995), 124–
138. doi: 10.2307/2975346. ☞ p. 219
[Cha18] P. Charron, A Pleijel-type theorem for the quantum harmonic oscillator, J. Spectr. The-
ory 8:2 (2018), 715–732. doi: 10.4171/JST/211. ☞ p. 163
[ChaHelHoO18] P. Charron, B. Helffer, and T. Hoffmann-Ostenhof, Pleijel’s theorem for Schrödinger
operators with radial potentials, Ann. Math. Qué. 42:1 (2018), 7–29. doi:
10.1007/s40316-017-0078-x. ☞ p. 163
[Cha84] I. Chavel, Eigenvalues in Riemannian geometry, Pure and Applied Mathematics 115.
Academic Press, Inc., Orlando, FL, 1984. ☞ pp. 2, 24, 27, 30, 140, 168, and 202
[Che71] J. Cheeger, A lower bound for the smallest eigenvalue of the Laplacian, in Problems in
analysis (Papers dedicated to Salomon Bochner, 1969), R. C. Gunning, ed. Princeton
Mathematical Series 1495. Princeton Univ. Press, 1970. doi: 10.1515/9781400869312.
☞ p. 164
[Che75] S.-Y. Cheng, Eigenfunctions and nodal sets, Comment. Math. Helv. 51:1 (1976), 43–55.
doi: 10.1007/BF02568142. ☞ pp. 119, 147, and 149
[Chu97] F. R. K. Chung, Spectral graph theory, CBMS Regional Conference Series in Math-
ematics 92. Amer. Math. Soc., Providence, RI, 1997. ☞ p. 173
[ChuGriYau96] F. R. K. Chung, A. Grigor’yan, and S.-T. Yau, Upper bounds for eigenvalues of
discrete and continuous Laplace operators, Adv. Math. 117:2 (1996), 165–178. doi:
10.1006/aima.1996.0006. ☞ p. 195
Bibliography 317

[CiaFus02] A. Cianchi and N. Fusco, Functions of bounded variation and rearrangements, Arch.
Ration. Mech. Anal. 165:1 (2002), 1–40. doi: 10.1007/s00205-002-0214-9. ☞ p. 157
[CiaKarMed19] D. Cianci, M. Karpukhin, and V. Medvedev, On branched minimal immersions of
surfaces by first eigenfunctions, Ann. Global Anal. Geom. 56:4 (2019), 667–690. doi:
10.1007/s10455-019-09683-8. ☞ p. 186
[Col17] B. Colbois, The spectrum of the Laplacian: a geometric approach, in Geometric and
computational spectral theory. Contemp. Math. 700, 1–40. Centre Rech. Math.
Proc., Amer. Math. Soc., Providence, RI, 2017. doi: 10.1090/conm/700/14181.
☞ p. 170
[ColDod94] B. Colbois and J. Dodziuk, Riemannian metrics with large λ1 , Proc. Amer. Math.
Soc. 122:3 (1994), 905–906. doi: 10.2307/2160770. ☞ pp. 180 and 188
[ColElS03] B. Colbois and A. El Soufi, Extremal eigenvalues of the Laplacian in a conformal class
of metrics: the “conformal spectrum” , Ann. Global Anal. Geom. 24:4 (2003), 337–349.
doi: 10.1023/A:1026257431539. ☞ p. 188
[ColElSGir11] B. Colbois, A. El Soufi, and A. Girouard, Isoperimetric control of the Steklov spectrum,
J. Funct. Anal. 261:5 (2011), 1384–1399. doi: 10.1016/j.jfa.2011.05.006. ☞ p. 231
[ColGGS22] B. Colbois, A. Girouard, C. Gordon, and D. Sher, Some recent developments on the
Steklov eigenvalue problem, arXiv: 2212.12528. ☞ p. 279
[ColGirHas18] B. Colbois, A. Girouard, and A. Hassannezhad, The Steklov and Laplacian spectra
of Riemannian manifolds with boundary, J. Funct. Anal. 278:6 (2020), 108409. doi:
10.1016/j.jfa.2019.108409. ☞ pp. 245 and 247
[ColMin11] T. H. Colding and W. P. Minicozzi II, Lower bounds for nodal sets of eigenfunctions,
Comm. Math. Phys. 306:3 (2011), 777–784. doi: 10.1007/s00220-011-1225-x. ☞ p. 128
[CdV73] Y. Colin de Verdière, Spectre du laplacien et longueurs des géodésiques périodiques I,
Compos. Math. 27:1 (1973), 80–106. EuDML. ☞ p. 209
[CdV87] Y. Colin de Verdière, Construction de laplaciens dont une partie finie du spectre est don-
née, Ann. Sci. École Norm. Sup. (4) 20:4 (1987), 599–615. doi: 10.24033/asens.1546.
☞ pp. 195 and 196
[ConSlo92] J. H. Conway and N. J. A. Sloane, Four-dimensional lattices with the same theta series,
Int. Math. Res. Not. 1992:4 (1992), 93–96. doi: 10.1155/S1073792892000102. ☞ p. 211
[Cou23] R. Courant, Ein allgemeiner Satz zur Theorie der Eigenfunktionen selbstadjungierter
Differentialausdrücke, Nachr. Ges. Göttingen (1923), 81–84. EuDML. ☞ p. 114
[CouHil89] R. Courant and D. Hilbert, Methods of mathematical physics, Volume 1, John Wiley
& Sons, 1989. doi: 10.1002/9783527617210. ☞ pp. 2, 20, 76, 96, and 97
[CoxJonMar17] G. Cox, C. K. R. T. Jones, and J. L. Marzuola, Manifold decompositions and indices of
Schrödinger operators, Indiana Univ. Math. J. 66:5 (2017), 1573–1602. Jstore. ☞ p. 163
[CroSha98] C. Croke and V. Sharafutdinov, Spectral rigidity of a compact negatively curved
manifold, Topology, 37:6 (1998), 1265–1273. doi: 10.1016/S0040-9383(97)00086-4.
☞ p. 225
[Dan11] D. Daners, Krahn’s proof of the Rayleigh conjecture revisited, Arch. Math. (Basel) 96:2
(2011), 187–199. doi: 10.1007/s00013-010-0218-x. ☞ pp. 154, 156, and 159 D
318 Bibliography

[Dav89] E. B. Davies, Heat kernels and spectral theory. Cambridge Tracts in Mathematics
92. Cambridge University Press, Cambridge, 1989. doi: 10.1017/CBO9780511566158.
☞ pp. 2 and 61
[Dav95] E. B. Davies, Spectral Theory and Differential Operators. Cambridge Studies in
Advanced Mathematics 42. Cambridge University Press, Cambridge, 1995. doi:
10.1017/CBO9780511623721. ☞ pp. 2, 49, and 67
[DePMon21] N. De Ponti and A. Mondino, Sharp Cheeger–Buser type inequalities in RCD(K , ∞)
spaces, J. Geom. Anal. 31 (2021), 2416–2438. doi: 10.1007/s12220-020-00358-6.
☞ p. 170
[DeSKalWei17] J. De Simoi, V. Kaloshin and Q. Wei, Dynamical spectral rigidity among Z2 -
symmetric strictly convex domains close to a circle, Ann. Math. (2) 186:1 (2017),
277–314. Appendix B co-authored with H. Hezari. doi: 10.4007/annals.2017.186.1.7.
☞ p. 225
[DLMF22] NIST Digital Library of Mathematical Functions. dlmf.nist.gov. Release 1.1.8 of
2022-12-15. F. W. J. Olver, A. B. Olde Daalhuis, D. W. Lozier, B. I. Schneider, R. F.
Boisvert, C. W. Clark, B. R. Miller, B. V. Saunders, H. S. Cohl, and M. A. McClain,
eds. ☞ pp. 20, 21, 168, and 173
[Don92] R.-T. Dong, Nodal sets of eigenfunctions on Riemann surfaces, J. Differential Geom.
36:2 (1992), 493–506. doi: 10.4310/jdg/1214448750. ☞ p. 132
[DonFef88] H. Donnelly and C. Fefferman, Nodal sets of eigenfunctions on Riemannian mani-
folds, Invent. Math. 93:1 (1988), 161–183. doi: 10.1007/BF01393691. ☞ pp. 131 and 133
[DonFef90] H. Donnelly and C. Fefferman, Nodal sets for eigenfunctions of the Laplacian on sur-
faces, J. Amer. Math. Soc. 3:2 (1990), 333–353. doi: 10.2307/1990956. ☞ p. 132
[DuiGui75] H. Duistermaat and V. Guillemin, Spectrum of positive elliptic operators and periodic
bicharacteristics, Invent. Math. 29:1 (1975), 39–79. doi: 10.1007/BF01405172. ☞ p. 101
[DyaZwo19] S. Dyatlov and M. Zworski, Mathematical theory of scattering resonances. Grad-
uate Studies in Mathematics 200. Amer. Math. Soc., Providence, RI, 2019. doi:
10.1090/gsm/200. ☞ p. 9
[EdmEva18] D. E. Edmunds and W. D. Evans, Spectral theory and differential operators. Ox-
E ford Mathematical Monographs. Oxford University Press, Oxford, 2018. doi:
10.1093/oso/9780198812050.001.0001. ☞ pp. 38 and 47
[Edw93a] J. Edward, An inverse spectral result for the Neumann operator on planar domains, J.
Func. Anal. 111:2 (1993), 312–322. doi: 10.1006/jfan.1993.1015. ☞ p. 251
[Edw93b] J. Edward, Pre-compactness of isospectral sets for the Neumann operator on
planar domains, Comm. Part. Diff. Eqs. 18:7–8 (1993), 1249–1270. doi:
10.1080/03605309308820973. ☞ p. 251
[ElSGiaJaz06] A. El Soufi, H. Giacomini, and M. Jazar, A unique extremal metric for the least eigen-
value of the Laplacian on the Klein bottle, Duke Math. J. 135:1 (2006), 181–202. doi:
10.1215/S0012-7094-06-13514-7. ☞ p. 186
[ElSIli84] A. El Soufi and S. Ilias, Le volume conforme et ses applications d’après Li et Yau, in
Séminaire de Théorie Spectrale et Géométrie, Année 1983–1984, VII.1–VII.15. Univ.
Grenoble I, Saint-Martin-d’Héres, 1984. Numdam. ☞ p. 185
Bibliography 319

[Eva10] L. C. Evans, Partial differential equations. Graduate Studies in Mathematics 19.

Amer. Math. Soc., Providence, RI, 2010. doi: 10.1090/gsm/019. ☞ pp. 38, 40, 53,
61, 123, and 202
[EvaGar15] L. C. Evans and R. Gariepy, Measure theory and fine properties of functions. Text-
books in Mathematics. CRC Press, Boca Raton, FL, 2015. ☞ pp. 40 and 154
[Fab23] G. Faber, Beweis, dass unter allen homogenen Membranen von gleicher Fläche und
gleicher Spannung die kreisförmige den tiefsten Grundton gibt, Sitzungberichte der F
mathematisch-physikalischen Klasse der Bayerischen Akademie der Wissenschaften
zu München Jahrgang (1923), 169–172. BAdW. ☞ p. 152
[Fed14] H. Federer, Geometric measure theory. Classics in Mathematics. Springer, Berlin, Hei-
delberg, 2014 doi: 10.1007/978-3-642-62010-2. ☞ pp. 120 and 131
[Fel71] W. Feller, An Introduction to Probability Theory and its Applications, vol. II. John
Wiley & Sons, Inc., New York–London–Sydney, 1971. ☞ p. 205
[Fil04] N. Filonov, On an inequality between Dirichlet and Neumann eigenvalues for the
Laplace operator, Algebra i Analiz 16:2 (2004), 172–176 (Russian). English transla-
tion St. Petersburg Math. J. 16:2 (2005), 413–416. doi: 10.1090/S1061-0022-05-00857-
5. ☞ p. 91
[FilLPS23] N. Filonov, M. Levitin, I. Polterovich, and D. A. Sher, Pólya’s conjecture for Euclidean
balls, Invent. Math. (2023). doi: 10.1007/s00222-023-01198-1. ☞ p. 105
[Fol95] G. B. Folland, Introduction to partial differential equations. Princeton University
Press, Princeton, NJ, second edition, 1995. ☞ pp. 39, 49, 53, and 55
[FoxKut83] D. W. Fox and J. R. Kuttler, Sloshing frequencies, Z. angew. Math. Phys. (ZAMP) 34
(1983), 668–696. doi: 10.1007/BF00948809. ☞ pp. 233 and 235
[FraLapWei22] R. L. Frank, A. Laptev, and T. Weidl, Schrödinger operators: Eigenvalues and Lieb–
Thirring inequalities. Cambridge Studies in Advanced Mathematics 200. Cam-
bridge University Press, Cambridge, 2022. doi: 10.1017/9781009218436. ☞ pp. 9
and 109
[FreLagPay21] P. Freitas, J. Lagacé, and J. Payette, Optimal unions of scaled copies of domains and
Pólya’s conjecture, Ark. Mat. 59 (2021), 11–51. doi: 10.4310/ARKIV.2021.v59.n1.a2.
☞ p. 102
[FreLau20] P. Freitas and R. Laugesen, From Steklov to Neumann and beyond, via
Robin: The Szegő way, Canadian J. Math. 72 (4) (2020), 1024–1043. doi:
10.4153/S0008414X19000154. ☞ p. 238
[Fri91] L. Friedlander, Some inequalities between Dirichlet and Neumann Eigenvalues, Arch.
Rational Mech. Anal. 116:2 (1991), 153–160. doi: 10.1007/BF00375590. ☞ pp. 91, 274,
and 278
[Fri21] L. Friedlander, On the Weyl asymptotic formula for Euclidean domains of finite vol-
ume, arXiv: 2106.07690. ☞ p. 96
[Fri69] A. Friedman, Partial differential equations. Holt, Rinehart and Winston, Inc., New
York-Montreal-London, 1969. ☞ p. 52
[Fri00] T. Friedrich, Dirac operators in Riemannian geometry. Graduate Studies in Mathe-
matics 25. Amer. Math. Soc., Providence, RI, 2000. doi: 10.1090/gsm/025. ☞ p. 9
320 Bibliography

[Fto21] I. Ftouhi, On the Cheeger inequality for convex sets, J. Math. Anal. Appl. 504:2 (2021),
125443. doi: 10.1016/j.jmaa.2021.125443. ☞ p. 173
[Fus04] N. Fusco, The classical isoperimetric theorem, Rend. Acad. Sci. Fis. Mat. Napoli (4)
71 (2004), 63–107. docenti.unina.it. ☞ p. 159
[Fus08] N. Fusco, Geometrical aspects of symmetrization, in Calculus of variations and non-
linear partial differential equations. Lecture Notes in Math. 1927, 155–181. Springer,
Berlin, 2008. doi: 10.1007/978-3-540-75914-0_5. ☞ p. 158
[Gaf58] M. Gaffney, Asymptotic distributions associated with the Laplacian for forms, Comm.
G Pure Appl. Math. 11 (1958), 535–545. doi: 10.1002/cpa.3160110405. ☞ p. 198
[GarLin86] N. Garofalo and F.-H. Lin, Monotonicity properties of variational integrals, A p
weights and unique continuation, Indiana Univ. Math. J. 35:2 (1986), 245–268. Jstore.
☞ p. 134
[Gas26] F. Gassmann, Bemerkungen zur vorstehenden Arbeit von Hurwitz, Math. Zeit. 25
(1926), 665–675. doi: 10.1007/BF01283860. ☞ p. 213
[GilTru01] D. Gilbarg and N. S. Trudinger, Elliptic partial differential equations of second order.
Classics in Mathematics. Springer-Verlag, Berlin, 2001. ☞ pp. 53, 54, 61, and 124
[Gil04] P. Gilkey, Asymptotic formulae in spectral geometry, Studies in Advanced Mathemat-
ics. Chapman & Hall/CRC, Boca Raton, FL, 2004. ☞ pp. 205 and 206
[Gin09] N. Ginoux, The Dirac spectrum. Lecture Notes in Mathematics 1976. Springer-
Verlag, Berlin, 2009. doi: 10.1007/978-3-642-01570-0. ☞ p. 9
[GirKarLag21] A. Girouard, M. Karpukhin, and J. Lagacé, Continuity of eigenvalues and shape opti-
misation for Laplace and Steklov problems, Geom. Func. Anal. 31 (2021), 513–561. doi:
10.1007/s00039-021-00573-5. ☞ p. 238
[GirKLP22] A. Girouard, M. Karpukhin, M. Levitin, and I. Polterovich, The Dirichlet-to-
Neumann map, the boundary Laplacian, and Hörmander’s rediscovered manuscript,
J. Spectr. Theory 12:1 (2022), 195–225. doi: 10.4171/JST/399. ☞ pp. 244, 248, 250,
274, 275, and 277
[GirLPS19] A. Girouard, J. Lagacé, I. Polterovich, and A. Savo, The Steklov spectrum of cuboids,
Mathematica 65:2 (2019), 272–310. doi: 10.1112/S0025579318000414. ☞ pp. 237
and 266
[GirNadPol09] A. Girouard, N. Nadirashvili, and I. Polterovich, Maximization of the second positive
Neumann eigenvalue for planar domains, J. Diff. Geom. 83:3 (2009), 637–661. doi:
10.4310/jdg/1264601037. ☞ p. 179
[GirPPS14] A. Girouard, L. Parnovski, I. Polterovich, and D. A. Sher, The Steklov spectrum of sur-
faces: asymptotics and invariants, Math. Proc. Cambridge Philos. Soc. 157:3 (2014),
379–389. doi: 10.1017/S030500411400036X. ☞ p. 251
[GirPol10a] A. Girouard and I. Polterovich, Shape optimization for low Neumann and Steklov
eigenvalues, Math. Methods Appl. Sci. 33:4 (2010), 501–516. doi: 10.1002/mma.1222.
☞ p. 238
[GirPol10b] A. Girouard and I. Polterovich, On the Hersch–Payne–Schiffer inequalities for Steklov
eigenvalues, Funct. Anal. Its Appl. 44 (2010), 106–117. doi: 10.1007/s10688-010-0014-
1. ☞ p. 239
Bibliography 321

[GirPol12] A. Girouard and I. Polterovich, Upper bounds for Steklov eigenvalues on surfaces,
Electron. Res. Announc. Math. Sci. 19 (2012), 77–85. doi: 10.3934/era.2012.19.77.
☞ p. 243
[GirPol17] A. Girouard and I. Polterovich, Spectral geometry of the Steklov problem, J. Spectr.
Theory 7:2 (2017), 321–359. doi: 10.4171/JST/164. ☞ pp. 234, 238, 251, and 279
[GorHerWeb21] C. Gordon, P. Herbrich, and D. Webb, Steklov and Robin isospectral manifold, J.
Spectr. Theory 11:1 (2021), 39–61. doi: 10.4171/jst/335. ☞ p. 222
[GorWebWol92] C. Gordon, D. Webb, and S. Wolpert, Isospectral plane domains and surfaces via
Riemannian orbifolds, Invent. Math. 110:1 (1992), 1–22. doi: 10.1007/BF01231320.
☞ pp. 214 and 219
[GreNgu13] D. S. Grebenkov and B.-T. Nguyen, Geometrical structure of Laplacian eigenfunc-
tions, SIAM Review 55:4 (2013), 601–667. doi: 10.1137/120880173. ☞ p. 22
[Gre86] A. G. Greenhill, Wave Motion in Hydrodynamics, Amer. J. Math. 9:1 (1886), 62–96.
doi: 10.2307/2369499. ☞ p. 233
[GréNédPla76] J. P. Grégoire, J. C. Nédélec, and J. Planchard, A method of finding the eigenvalues
and eigenfunctions of self-adjoint elliptic operators, Comp. Methods in Appl. Mech.
and Eng. 8 (1976), 201–214. doi: 10.1016/0045-7825(76)90045-1. ☞ p. 270
[Gri06] D. Grieser, The first eigenvalue of the Laplacian, isoperimetric constants, and the max
flow min cut theorem, Arch. Math. (Basel) 87:1 (2006), 75–85. doi: 10.1007/s00013-
005-1623-4. ☞ pp. 170, 171, 172, and 173
[GriNetYau04] A. Grigor’yan, Y. Netrusov, and S.-T. Yau, Eigenvalues of elliptic operators and geomet-
ric applications, in Surveys in differential geometry 9, 147–217. Int. Press, Somerville,
MA, 2004. doi: 10.4310/SDG.2004.v9.n1.a5. ☞ p. 187
[Gri11] P. Grisvard, Elliptic problems in nonsmooth domains, Classics in Applied Mathemat-
ics 69. Society for Industrial and Applied Mathematics (SIAM), Philadelphia, PA,
2011. doi: 10.1137/1.9781611972030. ☞ pp. 39, 40, and 72
[GuiKaz80] V. Guillemin and D. Kazhdan, Some inverse spectral results for negatively curved
2-manifolds, Topology 19:3 (1980), 301–312. doi: 10.1016/0040-9383(80)90015-4.
☞ p. 225
[Gun72] R. Gunning, Lectures on Riemann surfaces, Jacobi varieties, Mathematical Notes No.
12. Princeton University Press, Princeton, N.J., University of Tokyo Press, Tokyo,
1972. ☞ p. 185
[GusAbe98] K. Gustafson and T. Abe, The third boundary condition — was it Robin’s?, Math.
Intelligencer 20:1 (1998), 63–71. doi: 10.1007/BF03024402. ☞ p. 71
[Hal22] N. Halperin, Estimates of the Hausdorff measure of nodal sets of Laplace eigenfunc-
tions, M.Sc. thesis, Hebrew University, 2022. ☞ p. 142 H
[HarSim89] R. Hardt and L. Simon, Nodal sets for solutions of elliptic equations, J. Diff. Geometry
30 (1989), 505–522. doi: 10.4310/jdg/1214443599. ☞ pp. 119, 132, and 134
[Har15] G. H. Hardy, On the expression of a number as the sum of two squares, Quart. J. Pure
Appl. Math. 46 (1915), 263–283. ☞ p. 29
[HarWri79] G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers, Oxford
University Press, Oxford, 2008. ☞ p. 17
322 Bibliography

[HarMic95] E. M. Harrell II and P. L. Michel, Commutator bounds for eigenvalues of some dif-
ferential operators, in Evolution Equations (Baton Rouge, LA, 1992). Lecture Notes in
Pure and Appl. Math 168, 235–244. Dekker, New York, 1995. ☞ p. 195
[HarStu97] E. M. Harrell II and J. Stubbe, On trace identities and universal eigenvalue estimates
for some partial differential operators, Trans. Amer. Math. Soc. 349:5 (1997), 2037–
2055. doi: 10.1090/S0002-9947-97-01846-1. ☞ p. 189
[Has11] A. Hassannezhad, Conformal upper bounds for the eigenvalues of the Lapla-
cian and Steklov problem, J. Funct. Anal. 261:12 (2011), 3419–3436. doi:
10.1016/j.jfa.2011.08.003. ☞ p. 187
[HasKokPol16] A. Hassannezhad, G. Kokarev, and I. Polterovich, Eigenvalue inequalities on Rie-
mannian manifolds with a lower Ricci curvature bound, J. Spectr. Theory 6:4 (2016),
807–835. doi: 10.4171/JST/143. ☞ p. 170
[HasShe22] A. Hassannezhad and D. A. Sher, Nodal count for Dirichlet-to-Neumann operators
with potential, Proc. Amer. Math. Soc. (2022). doi: 10.1090/proc/16207. ☞ p. 272
[HasSif20] A. Hassannezhad and A. Siffert, A note on Kuttler–Sigillito’s inequalities, Ann. Math.
Qué. 44:1 (2020), 125–147. doi: 10.1007/s40316-019-00113-6. ☞ p. 275
[HasTao02] A. Hassell and T. Tao, Upper and lower bounds for normal derivatives
of Dirichlet eigenfunctions, Math. Res. Lett. 9:2–3 (2002), 289–305. doi:
10.4310/MRL.2002.v9.n3.a6. ☞ p. 277
[Hat01] A. Hatcher, Algebraic topology. Cambridge University Press, Cambridge, 2001. Avail-
able on author’s website. ☞ p. 212
[Hay78] W. K. Hayman, Some bounds for principle frequency, Appl. Anal. 7 (1978), 247–254.
doi: 10.1080/00036817808839195. ☞ p. 171
[Hec12] F. Hecht, New development in FreeFem++, J. Num. Math. 20:3–4 (2012), 251–265.
doi: 10.1515/jnum-2012-0013. ☞ p. 288
[Hed81] L. Hedberg, Spectral synthesis in Sobolev spaces, and uniqueness of solutions of the
Dirichlet problem, Acta Math. 147 (1981), 237–264. doi: 10.1007/BF02392874.
☞ p. 116
[HeiKilMar93] J. Heinonen, T. Kilpeläinen, and O. Martio, Nonlinear potential theory of degenerate
elliptic equations. Oxford Mathematical Monographs. Oxford Science Publications.
The Clarendon Press, Oxford University Press, New York, 1993. ☞ pp. 61 and 116
[Hei30] W. Heiseinberg, Über quantentheoretische Umdeutung kinematischer und mechanis-
cher Beziehungen, Z. Physik 33 (1925), 879–893. doi: 10.1007/BF01328377. ☞ p. 192
[Hel13] B. Helffer, Spectral Theory and its Applications. Cambridge University Press, Cam-
bridge, 2013. doi: 10.1017/CBO9781139505727. ☞ pp. 2 and 48
[HelSun16] B. Helffer and M. P. Sundqvist, On nodal domains in Euclidean balls, Proc. Amer.
Math. Soc. 144:11 (2016), 4777–4791. doi: 10.1090/proc/13098. ☞ p. 21
[HemSecSim91] R. Hempel, L. A. Seco, and B. Simon, The essential spectrum of Neumann Lapla-
cians on some bounded singular domains, J. Func. Anal. 102:2 (1991), 448–483. doi:
10.1016/0022-1236(91)90130-W. ☞ p. 49
[Hen06] A. Henrot, Extremum problems for eigenvalues of elliptic operators. Frontiers in Math-
ematics. Birkhäuser Verlag, Basel, 2006. ☞ pp. 2 and 161
Bibliography 323

[Her70] J. Hersch, Quatre propriétés isopérimétriques de membranes sphériques homogènes, C.

R. Acad. Sci. Paris Sér A-B 270 (1970), A1645–A1648. ☞ p. 180
[HerPaySch75] J. Hersch, L. Payne, and M. Schiffer, Some inequalities for Stekloff eigenvalues, Arch.
Rational Mech. Anal. 57 (1975), 99–114. doi: 10.1007/BF00248412. ☞ p. 239
[HezZel22] H. Hezari and S. Zelditch, One can hear the shape of ellipses of small eccentricity. Ann.
of Math. 196:3 (2022). doi: 10.4007/annals.2022.196.3.4. ☞ p. 224
[HilPro80] G. N. Hile and M. H. Protter, Inequalities for eigenvalues of the Laplacian, Indiana
Univ. Math. J. 29 (1980), 523–538. Jstor. ☞ p. 188
[HisLut01] P. D. Hislop and C. V. Lutzer, Spectral asymptotics of the Dirichlet-to-Neumann map
on multiply connected domains in R d , Inverse Problems 17:6 (2001), 1717–1741. doi:
10.1088/0266-5611/17/6/313. ☞ p. 230
[Hör54] L. Hörmander, Uniqueness theorems and estimates for normally hyperbolic partial
differential equations of the second order, in Tolfte Skandinaviska Matematikerkon-
gressen, Lunds Universitets Matematiska Institution, 1954, 105–115. ☞ p. 244
[Hör18] L. Hörmander, Inequalities between normal and tangential derivatives of harmonic
functions, in: Unpublished manuscripts, 37–41. Springer International Publishing,
2018. doi: 10.1007/978-3-319-69850-2_6 ☞ pp. 244, 246, and 248
[Ivr80] V. Ya. Ivrii, Second term of the spectral asymptotic expansion of the Laplace - Beltrami
operator on manifolds with boundary, Funct. Anal. Its Appl. 14 (1980), 98–106. doi: I
10.1007/BF01086550. ☞ pp. 99 and 100
[Ivr19] V. Ivrii, Spectral asymptotics for Dirichlet to Neumann operator in the domains with
edges, in V. Ivrii, Microlocal Analysis, Sharp Spectral Asymptotics and Applications V,
Springer, Cham, 2019, 513–539. doi: 10.1007/978-3-030-30561-1_31. ☞ p. 266
[JakLNNP05] D. Jakobson, M. Levitin, N. Nadirashvili, N. Nigam, and I. Polterovich, How large
can the first eigenvalue be on a surface of genus two?, Int. Math. Res. Not. 63 (2005), J
3967–3985. doi: 10.1155/IMRN.2005.3967. ☞ p. 186
[JakLNP06] D. Jakobson, M. Levitin, N. Nadirashvili, and I. Polterovich, Spectral problems with
mixed Dirichlet-Neumann boundary conditions: isospectrality and beyond, J. Com-
put. Appl. Math. 194:1 (2006), 141–155. doi: 10.1016/j.cam.2005.06.019. ☞ pp. 88
and 218
[JakNadPol06] D. Jakobson, N. Nadirashvili, and I. Polterovich, Extremal metric for the first eigen-
value on a Klein bottle, Canad. J. Math. 58:2 (2006), 381–400. doi: 10.4153/CJM-
2006-016-0. ☞ p. 186
[JerKen81] D. S. Jerison and C. E. Kenig, The Neumann problem on Lipschitz domains, Bull.
Amer. Math. Soc. (N.S.) 4:2 (1981), 203–207. doi: 10.1090/S0273-0979-1981-14884-9.
☞ pp. 62 and 64
[JerKen95] D. S. Jerison and C. E. Kenig, The inhomogeneous Dirichlet problem in Lipschitz do-
mains, J. Funct. Anal. 130:1 (1995), 161–219. doi: 10.1006/jfan.1995.1067. ☞ p. 62
[JimMor92] S. Jimbo and Y. Morita, Remarks on the behavior of certain eigenvalues on a singularly
perturbed domain with several thin channels, Comm. Partial Differential Equations
17:3–4 (1992), 189–226. doi: 10.1080/03605309208820852. ☞ p. 169
[Joh81] F. John, Plane waves and spherical means applied to partial differential equations.
Springer-Verlag New York, 1981. doi: 10.1007/978-1-4613-9453-2. ☞ p. 52
324 Bibliography

[JolSha14] A. Jollivet and V. Sharafutdinov, On an inverse problem for the Steklov spectrum of a
Riemannian surface, in Inverse Problems and Applications, P. Stefanov, A. Vasy, and
M. Zworski (Eds.). Contemporary Mathematics 615, 165–191. American Mathemati-
cal Society, Providence, RI, 2014. doi: 10.1090/conm/615. ☞ p. 251
[JolSha18] A. Jollivet and V. Sharafutdinov, Steklov zeta-invariants and a compactness theorem
for isospectral families of planar domains, J. Funct. Anal. 215:7 (2018), 1712–1755. doi:
10.1016/j.jfa.2018.06.019. ☞ p. 251
[Kac51] M. Kac, On some connections between probability theory and differential and integral
K equations, in Proceedings of the Second Berkeley Symposium on Mathematical Statis-
tics and Probability, 1950, 189–215. University of California Press, Berkeley and Los
Angeles, 1951. ☞ p. 206
[Kac66] M. Kac, Can one hear the shape of a drum?, Amer. Math. Monthly 73:4, part 2 (1966),
1–23. doi: 10.2307/2313748. ☞ pp. 204 and 223
[KalSor18] V. Kaloshin and A. Sorrentino, On the local Birkhoff conjecture for convex billiards,
Ann. Math. (2) 188:1 (2018), 315–380. doi: 10.4007/annals.2018.188.1.6. ☞ p. 224
[Kar16] M. Karpukhin, Upper bounds for the first eigenvalue of the Laplacian on non-
orientable surfaces, Int. Math. Res. Not. 2016:20 (2016), 6200–6209. doi: 10.1093/im-
rn/rnv345. ☞ p. 185
[Kar18] M. Karpukhin, Bounds between Laplace and Steklov eigenvalues on nonnegatively
curved manifolds, Electron. Res. Announc. Math. Sci. 24 (2017), 100–109. doi:
10.3934/era.2017.24.011. ☞ p. 243
[Kar19] M. Karpukhin, On the Yang–Yau inequality for the first eigenvalue, Geom. Funct.
Anal. 29:6 (2019), 1864 – 1885. doi: 10.1007/s00039-019-00518-z. ☞ p. 187
[Kar21] M. Karpukhin, Index of minimal spheres and isoperimetric eigenvalue inequalities,
Invent. Math. 223:1 (2021), 335–377. doi: 10.1007/s00222-020-00992-5. ☞ p. 187
[KarKokPol14] M. Karpukhin, G. Kokarev, and I. Polterovich, Multiplicity bounds for Steklov eigen-
values on Riemannian surfaces, Ann. Inst. Fourier (Grenoble) 64:6 (2014), 2481–
2502. doi: 10.5802/aif.2918. ☞ pp. 148 and 150
[KarLagPol22] M. Karpukhin, J. Lagacé, and I. Polterovich, Weyl’s law for the Steklov problem on
surfaces with rough boundary, arXiv: 2204.05294. ☞ p. 244
[KarNPP19] M. Karpukhin, N. Nadirashvili, A. Penskoi, and I. Polterovich, Conformally max-
imal metrics for Laplace eigenvalues on surfaces, Surveys in Diff. Geom. 24 (2019),
205–256. doi: 10.4310/SDG.2019.v24.n1.a6. ☞ pp. 184, 187, and 188
[KarNPP21] M. Karpukhin, N. Nadirashvili, A. Penskoi, and I. Polterovich, An isoperimetric in-
equality for Laplace eigenvalues on the sphere, J. Differential Geom. 118:2 (2021), 313–
333. doi: 10.4310/jdg/1622743142. ☞ pp. 186 and 187
[KarNPS21] M. Karpukhin, M. Nahon, I. Polterovich and D. Stern, Stability of isoperimetric in-
equalities for Laplace eigenvalues on surfaces, arXiv: 2106.15043. ☞ p. 186
[KarSte20] M. Karpukhin and D. L. Stern, Min-max harmonic maps and a new characterization
of conformal eigenvalues, arXiv: 2004.04086. ☞ p. 187
[KarSte22] M. Karpukhin and D. L. Stern, Existence of harmonic maps and eigenvalue optimiza-
tion in higher dimensions, arXiv: 2207.13635. ☞ p. 188
Bibliography 325

[KarVin22] M. Karpukhin and D. Vinokurov, The first eigenvalue of the Laplacian on orientable
surfaces, Math. Z. 301 (2022), 2733–2746. doi: 10.1007/s00209-022-03009-4. ☞ p. 187
[Kaw85] B. Kawohl, Rearrangements and convexity of level sets in PDE. Lecture Notes in
Mathematics 1150. Springer-Verlag, Berlin-Heidelberg-New York-Tokyo, 1985. doi:
10.1007/BFb0075060. ☞ p. 158
[Kel66] R. Kellner, On a theorem of Polya, Amer. Math. Monthly 73:8 (1966), 856–858. doi:
10.2307/2314181. ☞ p. 103
[Kha18] M. Khalile, Spectral asymptotics for Robin Laplacians on polygonal domains, J. Math.
Anal. Appl. 461:2 (2018), 1498–1543. doi: 10.1016/j.jmaa.2018.01.062. ☞ p. 277
[KhaOuBPan20] M. Khalile, T. Ourmières-Bonafos, and K. Pankrashkin, Effective operators for Robin
eigenvalues in domains with corners, Ann. Inst. Fourier (Grenoble) 70:5 (2020), 2215–
2301. doi: 10.5802/aif.3400. ☞ p. 277
[KhaPan18] M. Khalile and K. Pankrashkin, Eigenvalues of Robin Laplacians in infinite sectors,
Math. Nachr. 291:5–6 (2018), 928–965. doi: 10.1002/mana.201600314. ☞ p. 277
[Kim22] H. N. Kim, Maximization of the second Laplacian eigenvalue on the sphere, Proc.
Amer. Math. Soc. 150 (2022), 3501–3512. doi: 10.1090/proc/15908. ☞ p. 188
[Kin21] J. Kinnunen, Sobolev spaces, Lecture notes, Aalto University (2021), available online
at author’s website. ☞ p. 116
[Kok14] G. Kokarev, Variational aspects of Laplace eigenvalues on Riemannian surfaces, Adv.
Math., 258 (2014), 191–239. doi: 10.1016/j.aim.2014.03.006. ☞ p. 238
[Kor93] N. Korevaar, Upper bounds for eigenvalues of conformal metrics, J. Differential Geom.
37:1 (1993), 73–93. doi: 10.4310/jdg/1214453423. ☞ p. 187
[KotSmi99] T. Kottos and U. Smilansky, Periodic orbit theory and spectral statistics for quantum
graphs, Ann. Physics 274 (1999), 76–124. doi: 10.1006/aphy.1999.5904. ☞ p. 255
[Kra25] E. Krahn, Über eine von Rayleigh formulierte Minimaleigenschaft des Kreises, Math.
Ann. 94:1 (1925), 97–100. doi: 10.1007/BF01208645 ☞ pp. 152 and 156
[Kra26] E. Krahn, Über Minimaleigenschaften der Kugel in drei un mehr Dimensionen, Acta
Commun. Univ. Dorpat. A 9 (1926), 1–44. ☞ p. 161
[Krö92] P. Kröger, Upper bounds for the Neumann eigenvalues on a bounded domains in Eu-
clidean space, J. Funct. Anal. 106:2 (1992), 353–357. doi: 10.1016/0022-1236(92)90052-
K. ☞ p. 109
[KryLPPS21] S. Krymski, M. Levitin, L. Parnovski, I. Polterovich, and D. A. Sher, Inverse Steklov
spectral problem for curvilinear polygons, Int. Math. Res. Not. 2021:1 (2021), 1–37. doi:
10.1093/imrn/rnaa200. ☞ p. 267
[Kuh25] W. Kuhn, Über die Gesamtstarke der von einem Zustande ausgehenden Absorption-
slinien, Z. Physik 33:1 (1925), 408–412. doi: 10.1007/BF01328322. ☞ p. 192
[KurNow10] P. Kurasov and M. Nowaczyk, Geometric properties of quantum graphs and ver-
tex scattering matrices, Opus. Math. 30:3 (2010), 295–309. doi: 10.7494/Op-
Math.2010.30.3.295. ☞ p. 255
[KurSuh20] P. Kurasov and R. Suhr, Asymptotically isospectral quantum graphs and gener-
alised trigonometric polynomials, J. Math. Anal. Appl. 488:1 (2020), 124049. doi:
10.1016/j.jmaa.2020.124049. ☞ p. 267
326 Bibliography

[KutSig84] J. R. Kuttler and V. G. Sigillito, Eigenvalues of the Laplacian in two dimensions, SIAM
Review 26:2 (1984), 163–193. doi: 10.1137/1026033. ☞ p. 22
[KuzKKNPPS14] N. Kuznetsov, T. Kulczycki, M. Kwaśnicki, A. Nazarov, S. Poborchi, I. Polterovich,
and B. Siudeja, The legacy of Vladimir Andreevich Steklov, Notices Amer. Math. Soc.
61:1 (2014), 9–22. doi: 10.1090/noti1073. ☞ p. 228
[LagStA21] J. Lagacé, S. St-Amant, Spectral invariants of Dirichlet-to-Neumann operators on sur-
L faces, J. Spectr. Theory 11:4 (2021), 1627–1667. doi: 10.4171/JST/382. ☞ p. 275
[Lam93] H. Lamb, Hydrodynamics. Cambridge Mathematical Library. Cambridge University
Press, Cambridge, sixth edition, 1993. ☞ p. 233
[Lam33] G. Lamé, Mémoire sur la propagation de la chaleur dans les polyèdres, J. Éc. Polytech.
22 (1833), 194–251. ☞ p. 87
[Lap91] M. L. Lapidus, Fractal drum, inverse spectral problems for elliptic operators and a par-
tial resolution of the Weyl–Berry conjecture, Trans. Amer. Math. Soc. 325:2 (1991),
465–529. doi: 10.1090/S0002-9947-1991-0994168-5. ☞ p. 98
[Lap97] A. Laptev, Dirichlet and Neumann eigenvalue problems on domains in Euclidean
spaces, J. Funct. Anal. 151:2 (1997), 531–545. doi: 10.1006/jfan.1997.3155. ☞ p. 109
[LapSaf96] A. Laptev and Yu. Safarov, A generalization of the Berezin–Lieb Inequality, in Con-
temporary mathematical physics, R. Dobrushin, R. A. Milnos, M. A. Shubin, and A.
M. Vershik, eds. Amer. Math. Soc. Transl. (2) 175, 69–79. Amer. Math. Soc., Provi-
dence, RI, 1996. doi: 10.1090/trans2/175. ☞ p. 109
[LapWei00] A. Laptev and T. Weidl, Sharp Lieb–Thirring inequalities in high dimensions, Acta
Math. 184:1 (2000), 87–111. doi: 10.1007/BF02392782. ☞ p. 109
[Lau12] R. S. Laugesen, Spectral Theory of Partial Differential Equations — Lecture Notes,
arXiv: 1203.2344. ☞ pp. 2, 68, and 78
[Led04] M. Ledoux, Spectral gap, logarithmic Sobolev constant, and geometric bounds, Surveys
in Diff. Geom. 9 (2004), 219–240. doi: 10.4310/SDG.2004.v9.n1.a6. ☞ p. 170
[Lax02] P. D. Lax, Functional Analysis. John Wiley & Sons Inc., New York, 2002. ☞ pp. 44,
45, and 46
[Lén19] C. Léna, Pleijel’s nodal domain theorem for Neumann and Robin eigenfunctions,
Ann. Inst. Fourier (Grenoble) 69:1 (2019), 283–301. doi: 10.5802/aif.3243. ☞ p. 163
[LevWei86] H. A. Levine and H. F. Weinberger, Inequalities between Dirichlet and Neu-
mann eigenvalues, Arch. Rational Mech. Anal. 94:3 (1986), 193–208. doi:
10.1007/BF00279862. ☞ p. 94
[LevPar02] M. Levitin and L. Parnovski, Commutators, spectral trace identities, and univer-
sal estimates for eigenvalues, J. Funct. Anal. 192:2 (2002), 425–445. doi: 10.1006/j-
fan.2001.3913. ☞ pp. 189, 191, 194, and 195
[LevPar08] M. Levitin and L. Parnovski, On the principal eigenvalue of a Robin problem with a
large parameter, Math. Nachr. 281:2 (2008), 272–281. doi: 10.1002/mana.200510600.
☞ p. 277
[LevParPol06] M. Levitin, L. Parnovski and I. Polterovich, Isospectral domains with mixed bound-
ary conditions, J. Phys. A 39:9 (2006) 2073–2082. doi: 10.1088/0305-4470/39/9/006.
☞ pp. 215 and 217
Bibliography 327

[LevPPS22a] M. Levitin, L. Parnovski, I. Polterovich, and D. A. Sher, Sloshing, Steklov and cor-
ners: Asymptotics of sloshing eigenvalues, J. d’Anal. Math. 146 (2022), 65–125. doi:
10.1007/s11854-021-0188-x. ☞ pp. 233, 235, 254, 259, 265, and 266
[LevPPS22b] M. Levitin, L. Parnovski, I. Polterovich, and D. A. Sher, Sloshing, Steklov and corners:
Asymptotics of Steklov eigenvalues for curvilinear polygons, Proc. LMS 125:3 (2022),
359–487. doi: 10.1112/plms.12461. ☞ pp. 252, 254, 255, 260, 261, and 266
[LevStr21] M. Levitin and A. Strohmaier, Computations of eigenvalues and resonances on per-
turbed hyperbolic surfaces with cusps, Int. Math. Res. Not. 2021:6 (2021), 4003–4050.
doi: 10.1093/imrn/rnz157. ☞ p. 296
[LevVas96] M. Levitin and D. Vassiliev, Spectral asymptotics, renewal theorem, and the Berry con-
jecture for a class of fractals, Proc. London Math. Soc. (3) 72:1 (1996), 178–214. doi:
10.1112/plms/s3-72.1.188. ☞ p. 98
[LevYag03] M. Levitin and R. Yagudin, Range of the first three eigenvalues of the planar Dirich-
let Laplacian, LMS J. Math. Comp. 6 (2003), 1–17. doi: 10.1112/S1461157000000346.
☞ p. 194
[Lew46] H. Lewy, Water waves on sloping beaches, Bull. Amer. Math. Soc. 52(9) (1946), 737–
775. doi: 10.1090/S0002-9904-1946-08643-7. ☞ p. 259
[LiYau82] P. Li and S.-T. Yau, A new conformal invariant and its applications to the Willmore
conjecture and the first eigenvalue of compact surfaces, Invent. Math. 69:2 (1982), 269–
291. doi: 10.1007/BF01399507. ☞ pp. 185 and 186
[LiYau83] P. Li and S.-T. Yau, On the Schrödinger equation and the eigenvalue problem, Comm.
Math. Phys. 88:3 (1983), 309–318. doi: 10.1007/BF01213210. ☞ p. 106
[Lie73] E. H. Lieb, The classical limit of quantum spin systems, Comm. Math. Phys. 31:4
(1973), 327–340. doi: 10.1007/BF01646493. ☞ p. 109
[LieLos97] E. H. Lieb and M. Loss, Analysis. Graduate Studies in Mathematics 14 Amer. Math.
Soc., Providence, RI, 1997. doi: 10.2307/3621022. ☞ pp. 106, 107, and 155
[Log18a] A. Logunov, Nodal sets of Laplace eigenfunctions: polynomial upper estimates of
the Hausdorff measure, Ann. of Math. (2) 187:1 (2018), 221–239. doi: 10.4007/an-
nals.2018.187.1.4. ☞ pp. 132, 141, 144, and 145
[Log18b] A. Logunov, Nodal sets of Laplace eigenfunctions: proof of Nadirashvili’s conjecture
and of the lower bound in Yau’s conjecture, Ann. of Math. (2) 187:1 (2018), 241–262.
doi: 10.4007/annals.2018.187.1.5. ☞ pp. 132 and 141
[LogMal18a] A. Logunov and E. Malinnikova, Nodal sets of Laplace eigenfunctions: Estimates of
the Hausdorff measure in dimensions two and three, in 50 Years with Hardy Spaces. A
Tribute to Victor Havin. Operator Theory: Advances and Applications 261, 333–344.
Birkhäuser/Springer, Cham, 2018. doi: 10.1007/978-3-319-59078-3_17. ☞ pp. 132, 141,
and 142
[LogMal18b] A. Logunov and E. Malinnikova, Review of Yau’s conjecture on zero sets of Laplace
eigenfunctions, Current Developments in Mathematics 2018 (2018), 179–212. doi:
/10.4310/CDM.2018.v2018.n1.a4. ☞ pp. 132 and 134
[LogMal20] A. Logunov and E. Malinnikova, Lecture notes on quantitative unique continuation
for solutions of second order elliptic equations. in Harmonic analysis and applications,
C. E. Kenig, F. Lin, S. Mayboroda, and T. Toro, eds. IAS/Park City Math. Ser. 27,
328 Bibliography

1–33. Amer. Math. Soc., Providence, RI, 2020. doi: 10.1090/pcms/027/01. ☞ pp. 2
and 136
[Mak65] E. Makai, A lower estimation of the principal frequencies of simply connected mem-
M branes, Acta Math. Acad. Sci. Hungar. 16 (1965), 319–323. doi: 10.1007/BF01904840.
☞ p. 170
[Mak70] E. Makai, Complete orthogonal systems of eigenfunctions of three triangular mem-
branes, Studia Sci. Math. Hungar. 5 (1970), 51–62. ☞ p. 87
[MalSha15] E. Malkovich and V. Sharafutdinov, Zeta-invariants of the Steklov spec-
trum for a planar domain, Siberian Math. J. 56:4 (2015), 678–698. doi:
10.1134/S0037446615040114. ☞ p. 251
[Man08] D. Mangoubi, On the inner radius of a nodal domain, Canad. Math. Bull. 51:2 (2008),
249–260. doi: 10.4153/CMB-2008-026-2. ☞ p. 128
[Man13] D. Mangoubi, The effect of curvature on convexity properties of harmonic functions and
eigenfunctions, J. Lond. Math. Soc. (2) 87:3 (2013), 645–662. doi: 10.1112/jlms/jds067.
☞ p. 138
[MaySenStA22] J. Mayrand, C. Senécal, and S. St-Amant, Asymptotics of sloshing eigenvalues for a
triangular prism, Math. Proc. Cambridge Philos. Soc. 173:3 (2022), 539–571. doi:
10.1017/S0305004121000712. ☞ p. 266
[Maz85] V. G. Maz’ya, Sobolev spaces, Springer Series in Soviet Mathematics. Springer-Verlag,
Berlin, 1985. doi: 10.1007/978-3-662-09922-3. ☞ p. 153
[MazShu05] V. G. Maz’ya and M. A. Shubin, Can one see the fundamental frequency of a drum?,
Lett. Math. Phys. 74:2 (2005), 135–151. doi: 10.1007/s11005-005-0010-1. ☞ p. 171
[Maz91] R. Mazzeo, Remarks on a paper of L. Friedlander concerning inequalities between
Neumann and Dirichlet eigenvalues, Internat. Math. Res. Notices 1991:4 (1991), 41–
48. doi: 10.1155/S1073792891000065. ☞ p. 94
[McC11] B. J. McCartin, Laplacian eigenstructure of the equilateral triangle. Hikari Ltd.,
Ruse, 2011. Available at publisher’s website. ☞ p. 87
[McK70] H. P. McKean, An upper bound to the spectrum of ∆ on a manifold of negative curva-
ture, J. Diff. Geom. 4 (1970), 359–366. doi: 10.4310/jdg/1214429509. ☞ p. 168
[McL00] W. McLean, Strongly elliptic systems and boundary integral equations. Cambridge
University Press, Cambridge, 2000. ☞ pp. 39, 228, 269, and 306
[Mel84] R. Melrose, The trace of the wave group, in Microlocal analysis (Boulder, Colo.,
1983). Contemp. Math. 27, 12–167. Amer. Math. Soc., Providence, RI, 1984. doi:
10.1090/conm/027/741046. ☞ p. 99
[Mét77] G. Métivier, Valeurs propres de problèmes aux limites elliptiques irréguliers, Bull. Soc.
Math. France Suppl. Mém. 51–52 (1977), 125–219. doi: 10.24033/msmf.235. ☞ p. 98
[MeySer64] N. G. Meyers and J. Serrin, H = W . Proc. Nat. Acad. Sci. U.S.A. 51 (1964), 1055–1056.
doi: 10.1073/pnas.51.6.1055. ☞ p. 37
[Mil64] J. Milnor, Eigenvalues of the Laplace operator on certain manifolds, Proc. Nat. Acad.
Sci. U.S.A. 51 (1964), 542. doi: 10.1073/pnas.51.4.542. ☞ pp. 208 and 209
[Mil98] J. Milnor, Topology from the differentiable viewpoint. Princeton Landmarks in Math-
ematics and Physics 21. Princeton University Press, Princeton, N. J., 1998. ☞ p. 179
Bibliography 329

[MinPle49] S. Minakshisundaram and Å. Pleijel, Some properties of the eigenfunctions of the

Laplace-operator on Riemannian manifolds, Canadian J. Math. 1 (1949), 242–256.
doi: 10.4153/cjm-1949-021-5. ☞ p. 203
[MorNir57] C. B. Morrey Jr. and L. Nirenberg, On the analyticity of the solutions of linear elliptic
systems of partial differential equations, Comm. Pure Appl. Math. 10 (1957), 271–
290. doi: 10.1002/cpa.3160100204. ☞ p. 52
[Mos60] J. Moser, A new proof of De Giorgi’s theorem concerning the regularity problem for
elliptic differential equations, Comm. Pure Appl. Math. 13 (1960), 457–468. doi:
10.1002/cpa.3160130308. ☞ p. 61
[Nad87] N. Nadirashvili, Multiple eigenvalues of the Laplace operator, Math. USSR-Sb. 61:1
(1988), 225–238. doi: 10.1070/SM1988v061n01ABEH003204. ☞ pp. 148, 149, and 186 N
[Nad96] N. Nadirashvili, Berger’s isoperimetric problem and minimal immersions of surfaces,
Geom. Funct. Anal. 6:5 (1996), 877–897. doi: 10.1007/BF02246788. ☞ p. 186
[Nad97] N. Nadirashvili, Conformal maps and isoperimetric inequalities for eigenvalues of the
Neumann problem, in Proceedings of the Ashkelon Workshop on Complex Function
Theory (1996), 197–201. Israel Math. Conf. Proc. 11, Bar-Ilan Univ., Ramat Gan, 1997.
☞ p. 160
[Nad02] N. Nadirashvili, Isoperimetric inequality for the second eigenvalue of a sphere, J. Dif-
ferential Geom. 61:2 (2002), 335–340. doi: 10.4310/jdg/1090351388. ☞ p. 187
[NadPen18] N. Nadirashvili and A. Penskoi, An isoperimetric inequality for the second non-zero
eigenvalue of the Laplacian on the projective plane, Geom. Funct. Anal. 28:5 (2018),
1368–1393. doi: 10.1007/s00039-018-0458-7. ☞ p. 187
[NadSir15] N. Nadirashvili and Y. Sire, Maximization of higher order eigenvalues and applica-
tions, Mosc. Math. J. 15:4 (2015), 767–775. doi: 10.17323/1609-4514-2015-15-4-767-775.
☞ p. 187
[NadSir17] N. Nadirashvili and Y. Sire, Isoperimetric inequality for the third eigenvalue of the
Laplace-Beltrami operator on S2 , J. Differential Geom. 107:3 (2017), 561–571. doi:
10.4310/jdg/1508551225. ☞ p. 187
[Nam21] P. T. Nam, Functional Analysis II online lecture notes (2020–21) at LMU Munich,
available at author’s website. ☞ p. 106
[NaySho19] S. Nayatani and T. Shoda, Metrics on a closed surface of genus two which maximize
the first eigenvalue of the Laplacian, C. R. Math. Acad. Sci. Paris 357:1 (2019), 84–98.
doi: 10.1016/j.crma.2018.11.008. ☞ p. 186
[NazPolSod05] F. Nazarov, L. Polterovich, and M. Sodin, Sign and area in nodal geometry of Laplace
eigenfunctions, Amer. J. Math. 127:4 (2005), 879–910. doi: 10.1353/ajm.2005.0030.
☞ p. 141
[Nec12] J. Nečas, Direct methods in the theory of elliptic equations. Springer Monographs in
Mathematics, Springer-Verlag, 2012. ☞ p. 40
[NetSaf05] Yu. Netrusov and Yu. Safarov, Weyl asymptotic formula for the Laplacian on do-
mains with rough boundaries, Comm. Math. Phys. 253:2 (2005), 481–509. doi:
10.1007/s00220-004-1158-8. ☞ p. 96
[Nir59] L. Nirenberg, On elliptic partial differential equations, Ann. Scuola Norm. Sup. Pisa
Cl. Sci. (3) 13 (1959), 115–162. Numdam. ☞ p. 53
330 Bibliography

[NurRowShe19] M. Nursultanov, J. Rowlett, and D. A. Sher, The heat kernel on curvilinear polygonal
domains in surfaces, arXiv: 1905.00259. ☞ pp. 207 and 219
[OsgPhiSar88] B. Osgood, R. Phillips, and P. Sarnak, Compact isospectral sets of surfaces, J. Funct.
O Anal. 80:1 (1988), 212–234. doi: 10.1016/0022-1236(88)90071-7. ☞ p. 225
[Oss77] R. Osserman, A note on Hayman’s theorem on the bass note of a drum, Comment.
Math. Helv. 52:4 (1977), 545–555. doi: 10.1007/BF02567388. ☞ p. 171
[Oss78] R. Osserman, The isoperimetric inequality, Bull. Amer. Math. Soc. 84:6 (1978), 1182–
1238. doi: 10.1090/S0002-9904-1978-14553-4. ☞ pp. 168 and 171
[OttBro13] K. Ott and R. Brown, The mixed problem for the Laplacian in Lipschitz domains,
Potential Analysis 38 (2013), 1333–1364; see also Correction to: The mixed problem for
the Laplacian in Lipschitz domains, ibid. 54 (2021), 213–217. doi: 10.1007/s11118-012-
9317-6, doi: 10.1007/s11118-019-09822-7. ☞ p. 74
[Pan20] K. Pankrashkin, An eigenvalue estimate for a Robin p -Laplacian in C 1 domains,
P Proc. Amer. Math. Soc. 148:10 (2020), 4471–4477. doi: 10.1090/proc/15116. ☞ p. 277
[Par17] E. Parini, Reverse Cheeger inequality for planar convex sets, J. Convex Anal. 24:1
(2017), 107–122. heldermann.de. ☞ p. 173
[Pay55] L. E. Payne, Inequalities for eigenvalues of plates and membranes, J. Rational Mech.
Anal. 4 (1955), 517–529. doi: 10.1512/iumj.1955.4.54016. ☞ p. 91
[PayPólWei56] L. E. Payne, G. Pólya, and H. F. Weinberger, On the ratio of consecutive eigenvalues,
J. Math. and Phys. 35 (1956), 289–298. doi: 10.1002/sapm1956351289. ☞ p. 188
[PayWei60] L. E. Payne and H. F. Weinberger, An optimal Poincaré inequality for convex do-
mains, Arch. Rat. Mech. Anal. 5 (1960), 286–292. doi: 10.1007/BF00252910. ☞ p. 195
[Pet50] A. S. Peters, The effect of a floating mat on water waves, Comm. Pure and Appl. Math.
3:4 (1950), 319–354. doi: 10.1002/cpa.3160030402. ☞ p. 259
[Pet06] P. Petersen, Riemannian geometry. Graduate Texts in Mathematics 171. Springer-
Verlag, New York, 2006. doi: 10.1007/978-0-387-29403-2. ☞ p. 24
[Pet14] R. Petrides, Maximization of the second conformal eigenvalue of spheres, Proc. Amer.
Math. Soc. 142:7 (2014), 2385–2394. doi: 10.1090/S0002-9939-2014-12095-8. ☞ p. 187
[Pet18] R. Petrides, On the existence of metrics which maximize Laplace eigenvalue on surfaces,
Int. Math. Res. Not. 2018:14 (2018), 4261–4355. doi: 10.1093/imrn/rnx004. ☞ p. 187
[Pet22] R. Petrides, Maximizing one Laplace eigenvalue on n -dimensional manifolds, arXiv:
2211.15636. ☞ p. 188
[Pin80] M. A. Pinsky, The eigenvalues of an equilateral triangle, SIAM J. Math. Anal. 11
(1980), 819–827. doi: 10.1137/0511073. ☞ p. 87
[Pin85] M. A. Pinsky, Completeness of the eigenfunctions of the equilateral triangle, SIAM J.
Math. Anal. 16 (1985), pp. 848–851. doi: 10.1137/051606. ☞ p. 87
[Ple56] Å. Pleijel, Remarks on Courant’s nodal line theorem, Comm. Pure Appl. Math. 9
(1956), 543–550. doi: 10.1002/cpa.3160090324. ☞ pp. 118 and 162
[Poh65] S. I. Pohožaev, On the eigenfunctions of the equation ∆u + λ f (u) = 0, Soviet Math.
Dokl 6 (1965), 1408–1411. ☞ p. 244
Bibliography 331

[Pol00] I. Polterovich, Heat invariants of Riemannian manifolds, Israel J. Math. 119 (2000),
239–252. doi: 10.1007/BF02810670. ☞ p. 205
[Pol09] I. Polterovich, Pleijel’s nodal domain theorem for free membranes, Proc. Amer. Math.
Soc. 137:3 (2009), 1021–1024. doi: 10.1090/S0002-9939-08-09596-8. ☞ p. 163
[Pól54] G. Pólya, Mathematics and plausible reasoning, in two volumes, Princeton Univer-
sity Press, Princeton, N. J., 1954. ☞ p. 102
[Pól61] G. Pólya, On the eigenvalues of vibrating membranes, Proc. London Math. Soc. (3) 11
(1961), 419–433. doi: 10.1112/plms/s3-11.1.419. ☞ p. 103
[PólSze51] G. Pólya and G. Szegő, Isoperimetric inequalities in mathematical physics. Ann. Math.
Studies 27. Princeton University Press, Princeton, N. J., 1951. ☞ pp. 128 and 160
[Pom92] C. Pommerenke, Boundary behaviour of conformal maps. Grundlehren der Mathe-
matischen Wissenschaften 299. Springer–Verlag, Berlin, 1992. ☞ p. 240
[Pop20] N. Popoff, The negative spectrum of the Robin Laplacian. In Spectral Theory and
Mathematical Physics, P. Miranda, N. Popoff, and G. Raikov (eds), 229–242.
Lat. Amer. Math. Ser., Springer, Cham, 2020. doi: 10.1007/978-3-030-55556-6_12.
☞ p. 277
[ProStu19] L. Provenzano and J. Stubbe, Weyl-type bounds for Steklov eigenvalues, J. Spectr. The-
ory 9:1 (2019), 349–377. doi: 10.4171/JST/250. ☞ p. 247
[RauTay75] J. Rauch and M. Taylor, Potential and scattering theory on wildly perturbed domains,
J. Func. Anal. 18 (1975), 27–59. doi: 10.1016/0022-1236(75)90028-2. ☞ p. 82 R
[Ray77] J. W. Strutt, Baron Rayleigh, The Theory of Sound, 1st edition, Macmillan, London,
1877–1878. ☞ p. 152
[ReeSim75] M. Reed and B. Simon, Methods of modern mathematical physics, in four volumes.
Academic Press, New York-London, 1975–1979. ☞ pp. 2 and 96
[ReiTho25] F. Reiche and W. Thomas, Über die Zahl der Dispersionselektronen, die
einem stationären Zustand zugeordnet sind, Z. Physik 34 (1925), 510–525. doi:
10.1007/bf01328494. ☞ p. 192
[Rel40] F. Rellich, Darstellung der Eigenwerte von ∆u + λu durch ein Randintegral, Math.
Z. 46 (1940), 635–636. doi: 10.1007/BF01181459. ☞ p. 277
[Ros22a] A. Ros, On the first eigenvalue of the Laplacian on compact surfaces of genus three, J.
Math. Soc. Japan 74:3 (2022), 813-828. doi: 10.2969/jmsj/85898589. ☞ p. 187
[Ros22b] A. Ros, First eigenvalue of the Laplacian on compact surfaces for large genera, arXiv:
2211.15172. ☞ p. 187
[Ros97] S. Rosenberg, The Laplacian on a Riemannian manifold. London Math.
Soc. Student Texts 31. Cambridge University Press, Cambridge, 1997. doi:
10.1017/CBO9780511623783. ☞ pp. 2, 23, 24, 198, 204, and 214
[RoF15] G. Roy-Fortin, Nodal sets and growth exponents of Laplace eigenfunctions on surfaces,
Analysis & PDE 8:1 (2015), 223–255. doi: 10.2140/apde.2015.8.223. ☞ p. 141
[Roz72] G. V. Rozenbljum [= G. Rozenblum], On the eigenvalues of the first boundary
value problem in unbounded domains, Math. USSR-Sb. 18:2 (1972), 235–248. doi:
10.1070/SM1972v018n02ABEH001766 ☞ p. 96
332 Bibliography

[Roz86] G. V. Rozenblyum [= G. Rozenblum], On the asymptotics of the eigenvalues of certain

two-dimensional spectral problems, Sel. Math. Sov. 5:3 (1986), 233–244. ☞ p. 251
[Roz23] G. Rozenblum, Weyl asymptotics for Poincaré–Steklov eigenvalues in a domain with
Lipschitz boundary, arXiv: 2304.04047. ☞ p. 245
[RudWigYes21] Z. Rudnick, I. Wigman, and N. Yesha, Differences between Robin and Neumann
eigenvalues, Commun. Math. Phys. 388 (2021), 1603–1635. doi: 10.1007/s00220-021-
04248-y. ☞ p. 277
[RuzSadSur20] M. Ruzhansky, M. Sadybekov, and D. Suragan, Spectral geometry of par-
tial differential operators, Chapman and Hall/CRC, New York, 2020. doi:
10.1201/9780429432965. ☞ p. 9
[SafVas97] Yu. Safarov and D. Vassiliev, The asymptotic distribution of eigenvalues of partial dif-
S ferential operators. Translations of Mathematical Monographs 155. Amer. Math. Soc.,
Providence, RI, 1997. doi: 10.1090/mmono/155. ☞ pp. 94, 98, 99, 100, and 101
[San55] L. Sandgren, A vibration problem, Comm. Sém. Math. Univ. Lund 19 (1955), 1–84.
☞ p. 244
[Sar90] P. Sarnak, Determinants of Laplacians; heights and finiteness, in Analysis, et
cetera, 601–622. Academic Press, Boston, MA, 1990. doi: 10.1016/B978-0-12-574249-
8.50033-X. ☞ p. 225
[Sav01] A. Savo, Lower bounds for the nodal length of eigenfunctions of the Laplacian, Ann.
Global Anal. Geom. 19:2 (2001), 133–151. doi: 10.1023/A:1010774905973. ☞ p. 131
[Sch90] A. Schiemann, Ein Beispiel positiv definiter quadratischer Formen der Dimension
4 mit gleichen Darstellungszahlen, Arch. Math. (Basel) 54:4 (1990), 372–375. doi:
10.1007/BF01189584. ☞ p. 211
[Sch97] A. Schiemann, Ternary positive definite quadratic forms are determined by their theta
series, Math. Ann. 308:3 (1997), 507–517. doi: 10.1007/s002080050086. ☞ p. 211
[SchYau94] R. Schoen and S.-T. Yau, Lectures on differential geometry. Conference Proceedings
and Lecture Notes in Geometry and Topology, I. International Press, Cambridge,
MA, 1994. ☞ pp. 2, 165, 180, and 189
[Ser73] J.-P. Serre, A course in Arithmetic. Graduate Texts in Mathematics 7. Springer-Verlag,
New York, 1973. doi: 10.1007/978-1-4684-9884-4, ☞ p. 211
[Sie29] C. L. Siegel, Über einige Anwendungen diophantischer Approximationen. Abh.
Preuss. Akad. Wiss., Phys.-math. Kl., 1929. Reprinted in On some applications of Dio-
phantine approximations. Quad./Monogr 2, 81–138. Edizioni della Normale, Pisa,
2014. ☞ p. 21
[Shu01] M. Shubin, Pseudodifferential operators and spectral theory, Springer-Verlag, Berlin,
2001. doi: 10.1007/978-3-642-56579-3. ☞ pp. 30, 34, 50, and 98
[Shu20] M. Shubin, Invitation to Partial Differential Equations, edited by M. Braverman, R.
McOwen, and P. Topalov. Graduate Studies in Mathematics 205. American Math.
Soc., Providence, RI, 2020. ☞ pp. 2, 39, 41, 43, 90, 114, and 121
[Sog17] C. D. Sogge, Fourier integrals in classical analysis. Cambridge Tracts in Mathemat-
ics 210. Cambridge University Press, Cambridge, 2017. doi: 10.1017/9781316341186.
☞ p. 9
Bibliography 333

[Spi88] M. Spivak, A Comprehensive Introduction to Differential Geometry, vol. 4. Boston,

Publish or Perish, 1975. ☞ p. 25
[Ste14] S. Steinerberger, A geometric uncertainty principle with an application to Pleijel’s esti-
mate, Ann. Henri Poincaré 15:12 (2014), 2299–2319. doi: 10.1007/s00023-013-0310-4.
☞ p. 163
[Ste70] W. Stenger, Nonclassical choices in variational principles for eigenvalues, J. Funct.
Analysis 6:1 (1970), 157–164. doi: 10.1016/0022-1236(70)90053-4. ☞ p. 67
[Stö07] H. J. Stöckmann, Chladni meets Napoleon, Eur. Phys. J. Spec. Top. 145 (2007), 15–23.
doi: 10.1140/epjst/e2007-00144-5. ☞ p. 112
[Str07] W. A. Strauss, Partial differential equations. An introduction. John Wiley & Sons,
Chichester, 2007. ☞ p. 22
[SunZho17] J. Sun and A. Zhou, Finite element methods for eigenvalue problems. Monographs
and Research Notes in Mathematics. CRC Press, Boca Raton, FL, 2017. ☞ pp. 281
and 282
[Sun85] T. Sunada, Riemannian coverings and isospectral manifolds, Ann. of Math. (2) 121:1
(1985), 169–186. doi: 10.2307/1971195. ☞ pp. 208, 213, and 214
[Sze54] G. Szegő, Inequalities for certain eigenvalues of a membrane of given area, J. Rational
Mech. Anal. 3 (1954), 343–356. doi: 10.1512/iumj.1954.3.53017. ☞ pp. 174 and 179
[Tan73] S. Tanno, Eigenvalues of the Laplacian of Riemannian manifolds, Tôhoku Math. J.
25:3 (1973), 391–403. doi: 10.2748/tmj/1178241341. ☞ p. 225 T
[Tay11] M. Taylor, Partial differential equations. I. Basic Theory. Applied Math. Sciences 115.
Springer, New York, 2011. doi: 10.1007/978-1-4419-7055-8. ☞ pp. 50 and 63
[Tay11b] M. Taylor, Partial differential equations. II. Qualitative studies of linear equations.
Applied Math. Sciences 116. Springer, New York, 2011. doi: 10.1007/978-1-4419-7052-
7. ☞ p. 181
[Tré82] F. Tréves, Introduction to pseudodifferential operators and Fourier integral operators.
Vol. 2, Plenum Press, New York-London, 1980. ☞ p. 98
[Ura17] H. Urakawa, Spectral geometry of the Laplacian. World Scientific, Hackensack, NJ,
2017. doi: 10.1142/10018. ☞ p. 189 U
[Vas86] D. G. Vasil’ev [= D. Vassiliev], Two-term asymptotics of the spectrum of a boundary
value problem in the case of a piecewise smooth boundary, Soviet Math. Dokl. 33:1 V
(1986), 227–230. Mathnet.ru. ☞ pp. 97 and 100
[Wat95] G. N. Watson, A treatise on the theory of Bessel functions. Cambridge Mathematical
Library. Cambridge University Press, Cambridge, 1995 ☞ pp. 20, 21, and 173 W
[Wei56] H. F. Weinberger, An isoperimetric inequality for the N -dimensional free membrane
problem, J. Rational Mech. Anal. 5 (1956), 633–636. doi: 10.1512/iumj.1956.5.55021.
☞ p. 174
[Wei54] R. Weinstock, Inequalities for a classical eigenvalue problem, J. Rational Mech. Anal.
3 (1954), 745–753. doi: 10.1512/iumj.1954.3.53036. ☞ p. 237
[Wel72] W. Welsh, Monotonicity for eigenvalues of the Schrödinger operator on unbounded do-
mains, Arch. Rational Mech. Anal. 49 (1972/73), 129–136. doi: 10.1007/BF00281414.
☞ p. 76
334 Bibliography

[Xio18] C. Xiong, Comparison of Steklov eigenvalues on a domain and Laplacian eigenvalues

X on its boundary in Riemannian manifolds, J. Funct. Anal. 275:12 (2018), 3245–3258.
doi: 10.1016/j.jfa.2018.09.012. ☞ p. 247
[Yan91] H. Yang, An estimate of the difference between consecutive eigenvalues, preprint
Y IC/91/60 of the Intl. Centre for Theoretical Physics, Trieste, 1991 (revised preprint,
Academia Sinica, 1995). ☞ p. 188
[YanYau80] P. Yang and S.-T. Yau, Eigenvalues of the Laplacian of compact Riemann surfaces and
minimal submanifolds, Ann. Scuola Norm. Sup. Pisa Cl. Sci. 7:1 (1980), 55–63. Num-
dam. ☞ pp. 183, 184, and 185
[Yau75] S.-T. Yau, Isoperimetric constants and the first eigenvalue of a compact Riemannian
manifold, Ann. Sci. École Norm. Sup. 8:4 (1975), 487–507. doi: 10.24033/asens.1299.
☞ p. 168
[Yau82] S.-T. Yau, Problem section, in Seminar on Differential Geometry. Ann. of
Math. Stud. 102, 669–706. Princeton Univ. Press, Princeton, N.J., 1982. doi:
10.1515/9781400881918-035. ☞ pp. 131 and 187
[Yau86] S.-T. Yau, Nonlinear analysis in geometry. Monographies de L’Enseignement Math-
ématique, vol. 33. L’Enseignement Mathématique, Geneva, 1986. ☞ p. 195
[Zar10] S. Zaremba, Sur un problème mixte relatif à l’équation de Laplace, Bull. intern. Acad.
Z des Sciences de Cracovie. Classe des Sciences Math. et Naturelles, Serie A: Sciences
math. (1910), 313–344. ☞ p. 74
[Zel09] S. Zelditch, Inverse spectral problem for analytic domains. II. Z2 -symmetric domains,
Ann. of Math. (2) 170:1 (2009), 205–269. doi: 10.4007/annals.2009.170.205. ☞ p. 224
[Zwo12] M. Zworski, Semiclassical analysis. Graduate Studies in Mathematics 138. American
Mathematical Soc., Providence, R. I., 2012. ☞ p. 9
 Index

In most cases, only the first appearance of a term or its definition are listed.

Symbols e(t , x, y) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198

ffl E 0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 E Λ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
〈·, ·〉 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303 e M (t ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
(·, ·)H m (Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 η k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
A (Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 fˆ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
α . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252 F u . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
A ∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 F α,ℓ (τ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
A u,x (·) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 H 0 (Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
d
B a,r , B a,r , B rd , Bd . . . . . . . . . . . . . . . . . . . . . . 303 H Λ (Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
β( f ,Q) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 1
H0,Γ (Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
β( f , B ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
H d −1 (·) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
C , C k , C ∞ , C 0k , C 0∞ . . . . . . . . . . . . . . . . . . . . 305
∞ h D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
C 0,Γ (Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
H m (Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
C F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
H0m (Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
cap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
H m (∂Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
card . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 m
Hloc (Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
C b,d . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
h N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
C d . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Ch(ζ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255 Hm, H fm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

χ(M ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 Hes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245

D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304 H f (·) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
D0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228 I m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
DΛ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269 Jac . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
∂n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 J m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
′
∂α . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305 j m,k , j m,k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
∆ κd ,m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
Euclidean space . . . . . . . . . . . . . . . . . . . . . 11 ℓ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
Riemannian manifold . . . . . . . . . . . . . . . 25 L(∂Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
spherical coordinates . . . . . . . . . . . . . . . . 30 L F (t ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
p
−∆D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 L p , L ∞ , L loc . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
−∆N . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 λ1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
−∆R,γ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 Λ1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
−∆ZΓ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74 λk (M ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

335
336 Index

λD
k
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 VF (t ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Λk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187 Z u . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
λN
k
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
R,γ
λk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
λZk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
A
Mα,ℓ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252 angle
m(λk ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 exceptional . . . . . . . . . . . . . . . . . . . . . . . 256
µk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70 special . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
M u,x (·) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121 Ashbaugh–Benguria inequality . . . . . . . . . . 194
N0 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
N (λ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27, 89
Nf(λ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 B
N D (λ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
bathtub principle . . . . . . . . . . . . . . . . . . . . . . 107
N N (λ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Berezin–Li–Yau inequality . . . . . . . . . . . . . . 106
N S (σ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
Bers’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . 147
N f (·) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
Bessel equation . . . . . . . . . . . . . . . . . . . . . . . . . 19
N (h, B r , c) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Bessel function of the first kind . . . . . . . . . . . 20
N b (h, B r , c) . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
Bolza surface . . . . . . . . . . . . . . . . . . . . . . . . . . 186
νk
Bourget hypothesis . . . . . . . . . . . . . . . . . . . . . . 21
eigenvalues of quantum graph . . . . . . . 253
bubbling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
Laplace–Beltrami eigenvalues . . . . . . . 230 Buser’s inequality . . . . . . . . . . . . . . . . . . . . . . 169
Ω∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
Ω⋆ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
ωd .¡. . ¢. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
C
ordx f . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Pm, P fm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 capacity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
P α,ℓ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252 Cauchy–Riemann operator . . . . . . . . . . . . . . . 55
Q . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 Cheeger constant
D N Dirichlet . . . . . . . . . . . . . . . . . . . . . . . . . 164
QΩ ,Q Ω . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
R[·] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66 Neumann . . . . . . . . . . . . . . . . . . . . . . . . 166
R S [·] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 Cheeger’s dumbbell . . . . . . . . . . . . . . . . . . . . 169
ρ Ω , ρeΩ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 Cheeger’s inequality
R+ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303 closed manifolds . . . . . . . . . . . . . . . . . . 167
R ∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Dirichlet . . . . . . . . . . . . . . . . . . . . . . . . . 164
Ric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169 Neumann . . . . . . . . . . . . . . . . . . . . . . . . 166
S da,r−1 , S rd −1 , S r , Sd −1 . . . . . . . . . . . . . . . . . . . 304 co-area formula . . . . . . . . . . . . . . . . . . . . . . . . 153
commutator identities . . . . . . . . . . . . . . . . . . 189
S (Rd ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
conformal
σd −1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
branched covering . . . . . . . . . . . . . . . . . 184
σk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
invariance . . . . . . . . . . . . . . . . . . . . . . . . . 71
σΛ k
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
counting function . . . . . . . . . . . . . . . . . . . 27, 89
τm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
Courant’s theorem . . . . . . . . . . . . . . . . . . . . . . 115
τS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Courant-sharp eigenvalue . . . . . . . . . . . . 125, 163
T (Ω) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
UF (t ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
U F (t ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
u ∗ . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
D
VF (t ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 deck transformation . . . . . . . . . . . . . . . . . . . . 211
Index 337

difference quotient . . . . . . . . . . . . . . . . . . . . . . 53 G
diffusion equation . . . . . . . . . . . . . . . . . . . . . 197
Galerkin method . . . . . . . . . . . . . . . . . . . . . . . 281
directional derivative . . . . . . . . . . . . . . . . . . . . 22
Gauss’s circle problem . . . . . . . . . . . . . . . . . . . 28
Dirichlet boundary conditions . . . . . . . . . . 13, 16
genus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
Dirichlet energy . . . . . . . . . . . . . . . . . . . . . . . . 43
geodesic ball . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Dirichlet problem . . . . . . . . . . . . . . . . . . . . . . . 16
Glazman’s lemma . . . . . . . . . . . . . . . . . . . . . . . 90
weak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41 gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . 22, 305
Dirichlet–Neumann bracketing . . . . . . . . . . 80 Green’s formula . . . . . . . . . . . . . . . . . . . . . . . . 40
Dirichlet-to-Neumann map . . . . . . . . . . . . . 228
for Helmholtz equation . . . . . . . . . . . . 269
divergence . . . . . . . . . . . . . . . . . . . . . . . . . 22, 305
H
domain monotonicity . . . . . . . . . . . . . . . . . . . 75
Donnelly–Fefferman bound . . . . . . . . . . . . . 133 Hardy–Littlewood–Karamata theorem . . . 205
local version . . . . . . . . . . . . . . . . . . . . . . 138 harmonic
doubling index extension . . . . . . . . . . . . . . . . . . . . . . . . 228
for balls . . . . . . . . . . . . . . . . . . . . . . . . . . 132 function . . . . . . . . . . . . . . . . . . . . . . . 12, 120
for cubes . . . . . . . . . . . . . . . . . . . . . . . . . 141 Hausdorff measure . . . . . . . . . . . . . . . . . . . . . 131
heat equation . . . . . . . . . . . . . . . . . . . . 12, 152, 197
parametrix . . . . . . . . . . . . . . . . . . . . . . . . 201
heat invariants . . . . . . . . . . . . . . . . . . . . . . . . 204
E heat kernel . . . . . . . . . . . . . . . . . . . . . . . . . . . . 198
eigenfunction . . . . . . . . . . . . . . . . . . . . . . . . . . 14 heat trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
eigenvalue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 height function . . . . . . . . . . . . . . . . . . . . . . . . 134
elliptic bootstrapping . . . . . . . . . . . . . . . . . . . . 53 Helmholtz extension . . . . . . . . . . . . . . . . . . . 269
elliptic regularity Hersch’s
global . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 lemma . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
local . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 180
equimeasurable functions . . . . . . . . . . . . . . . . 155 Hersch–Payne–Schiffer inequality . . . . . . . . 239
Hile–Protter inequality . . . . . . . . . . . . . . . . . 188
homogeneous polynomials . . . . . . . . . . . . . . . 30
harmonic . . . . . . . . . . . . . . . . . . . . . . . . . . 31
F Hörmander’s identity . . . . . . . . . . . . . . . . . . 246
Faber–Krahn inequality . . . . . . . . . . . . . . . . . 152
FEM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
degrees of freedom . . . . . . . . . . . . . . . . . 283 I
mesh . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
injectivity radius . . . . . . . . . . . . . . . . . . . . . . . 140
nodes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
isospectral
finite elements
domains . . . . . . . . . . . . . . . . . . . . . . . . . 214
conforming . . . . . . . . . . . . . . . . . . . . . . . 283
manifolds . . . . . . . . . . . . . . . . . . . . . . . . 207
Lagrangian . . . . . . . . . . . . . . . . . . . . . . . 283
mixed boundary conditions . . . . . . . . . 215
linear . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
Steklov . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
quadratic . . . . . . . . . . . . . . . . . . . . . . . . . 283
triangular . . . . . . . . . . . . . . . . . . . . . . . . 282
Fraenkel asymmetry . . . . . . . . . . . . . . . . . . . . 159
frequency function . . . . . . . . . . . . . . . . . . . . . 135
K
Friedlander–Filonov inequality . . . . . . . . . . . 91 Klein bottle . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Friedrichs extension . . . . . . . . . . . . . . . . . . . . . 45 Korevaar’s bound . . . . . . . . . . . . . . . . . . . . . . 187
fundamental gap . . . . . . . . . . . . . . . . . . . . . . . 195 Krahn–Szego inequality . . . . . . . . . . . . . . . . . 161
338 Index

L O
Laplace equation . . . . . . . . . . . . . . . . . . . . . . . . 12 operator
Laplace operator . . . . . . . . . . . . . . . see Laplacian adjoint . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Laplace–Beltrami operator . . . . . . . . . . . . . . . 24 domain of . . . . . . . . . . . . . . . . . . . . . . . . 44
isothermal coordinates . . . . . . . . . . . . . . 25 extension . . . . . . . . . . . . . . . . . . . . . . . . . 44
on the sphere . . . . . . . . . . . . . . . . . . . . . . 30 positive . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Laplacian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 self-adjoint . . . . . . . . . . . . . . . . . . . . . . . . 44
Dirichlet . . . . . . . . . . . . . . . . . . . . . . . . . . 46 semi-bounded . . . . . . . . . . . . . . . . . . . . . 45
Neumann . . . . . . . . . . . . . . . . . . . . . . . . . 49 symmetric . . . . . . . . . . . . . . . . . . . . . . . . 44
p -Laplacian . . . . . . . . . . . . . . . . . . . . . . . 173
polar coordinates . . . . . . . . . . . . . . . . . . . 19
Robin . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72 P
Zaremba . . . . . . . . . . . . . . . . . . . . . . . . . . 74 Payne–Pólya–Weinberger inequality . . . . . . 188
lattice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208 Payne–Weinberger inequality . . . . . . . . . . . . 195
layer cake representation . . . . . . . . . . . . . . . . . 155 Peters solution . . . . . . . . . . . . . . . . . . . . . . . . 260
length spectrum . . . . . . . . . . . . . . . . . . . . . . . 224 Pleijel
level set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 constant . . . . . . . . . . . . . . . . . . . . . . . . . . 163
lifting trick . . . . . . . . . . . . . . . . . . . . . . . . . 52, 138 nodal domain theorem . . . . . . . . . . . . . 162
Lipschitz boundary . . . . . . . . . . . . . . . . . . . . 306 Pohozhaev’s identity . . . . . . . . . . . . . . . . . . . 245
Lipschitz function . . . . . . . . . . . . . . . . . . . . . 306 Poincaré’s inequality . . . . . . . . . . . . . . . . . . . . 43
Poisson equation . . . . . . . . . . . . . . . . . . . . . . . . 12
Poisson summation formula . . . . . . . . . . . . . 208
Pólya’s conjecture . . . . . . . . . . . . . . . . . . . . . . 102
M Pólya–Szegő
conjecture . . . . . . . . . . . . . . . . . . . . . . . . 161
mass matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 282 principle . . . . . . . . . . . . . . . . . . . . . . . . . 156
McKean’s inequality . . . . . . . . . . . . . . . . . . . . 168 principle of not feeling the boundary . . . . . 206
mean over a ball . . . . . . . . . . . . . . . . . . . . . . . . 121
method of difference quotients . . . . . . . . . . . . 53
Milnor’s example . . . . . . . . . . . . . . . . . . . . . . 208
Minakshisundaram–Pleijel expansion . . . . . 203
Q
multiplicity . . . . . . . . . . . . . . . . . . . . . . . . . 17, 44 quadratic form . . . . . . . . . . . . . . . . . . . . . . . . . 65
bounds . . . . . . . . . . . . . . . . . . . . . . . . . . 148 quantum graph . . . . . . . . . . . . . . . . . . . . . . . . 252
quasi-continuous . . . . . . . . . . . . . . . . . . . . . . . 116
quasi-eigenvalues . . . . . . . . . . . . . . . . . . . . . . 254
quasi-everywhere . . . . . . . . . . . . . . . . . . . . . . . 116
N quasimode . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
Neumann boundary conditions . . . . . . . . 15, 18
nodal
deficiency . . . . . . . . . . . . . . . . . . . . . . . . 163 R
domain . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 Rayleigh quotient . . . . . . . . . . . . . . . . . . . . . . 66
graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 rearrangement
set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 symmetric decreasing, of a function . . . 155
non-periodicity condition . . . . . . . . . . . . . . . 100 symmetric, of a set . . . . . . . . . . . . . . . . . . 151
normal covering . . . . . . . . . . . . . . . . . . . . . . . . 211 reduced inradius . . . . . . . . . . . . . . . . . . . . . . . 170
normal derivative . . . . . . . . . . . . . . . . . . . . . . . 18 Rellich identity . . . . . . . . . . . . . . . . . . . . . . . . 277
Index 339

Rellich–Kondrachov theorem . . . . . . . . . . . . 38 T
resolvent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Thomas–Reiche–Kuhn sum rule . . . . . . . . . 192
Riemannian metric . . . . . . . . . . . . . . . . . . . . . 22
tiling domain . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Robin problem . . . . . . . . . . . . . . . . . . . . . . . . . 71
torsional rigidity . . . . . . . . . . . . . . . . . . . . . . . 160
Robin–Dirichlet-to-Neumann duality . . . . 272 transplantation of eigenfunctions . . . . . . . . . 215

S U
Schrödinger equation . . . . . . . . . . . . . . . . . . . . 12 uniformisation theorem . . . . . . . . . . . . . . . . . 181
Schwartz space . . . . . . . . . . . . . . . . . . . . . . . . 306 universal inequalities . . . . . . . . . . . . . . . . . . . 189
sloping beach problem . . . . . . . . . . . . . . . . . . 258
sloshing
problem . . . . . . . . . . . . . . . . . . . . . . . . . . 233 V
surface . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
vanishing order . . . . . . . . . . . . . . . . . . . . . . . . 132
Sobolev
variational principle
embedding theorem . . . . . . . . . . . . . . . . 38
Dirichlet Laplacian . . . . . . . . . . . . . . . . . 69
space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Dirichlet-to-Neumann map . . . . . . . . 270
on a Riemannian manifold . . . . . . . . 49
Laplace–Beltrami operator . . . . . . . . . . 70
trace theorem . . . . . . . . . . . . . . . . . . . . . . 40 Neumann Laplacian . . . . . . . . . . . . . . . . 70
spectral invariant . . . . . . . . . . . . . . . . . . . . . . 204 quadratic form . . . . . . . . . . . . . . . . . . . . 67
spectral prescription . . . . . . . . . . . . . . . . . . . . 195 Steklov problem . . . . . . . . . . . . . . . . . . . 232
spectral theorem
strong . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
weak
Dirichlet . . . . . . . . . . . . . . . . . . . . . . . 42
W
Neumann . . . . . . . . . . . . . . . . . . . . . . 48 wave equation . . . . . . . . . . . . . . . . . . . . . . . . . . 12
on a Riemannian manifold . . . . . . . . 50 one-dimensional . . . . . . . . . . . . . . . . . . . . 13
spectrum weak derivative . . . . . . . . . . . . . . . . . . . . . . . . . 36
discrete . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 weak solution . . . . . . . . . . . . . . . . . 41, 42, 47, 52
essential . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 Weinstock inequality . . . . . . . . . . . . . . . . . . . 237
spherical harmonics . . . . . . . . . . . . . . . . . . . . . . 31 Weyl’s conjecture . . . . . . . . . . . . . . . . . . . . . . . 99
spherical mean . . . . . . . . . . . . . . . . . . . . . . . . . 121 Weyl’s law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
standing wave . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 on a Riemannian manifold . . . . . . . . . 205
Steiner symmetrisation . . . . . . . . . . . . . . . . . 160 Steklov problem . . . . . . . . . . . . . . . . . . 244
Steklov problem . . . . . . . . . . . . . . . . . . . . . . . 227
stiffness matrix . . . . . . . . . . . . . . . . . . . . . . . . 282
Sturm–Liouville problem . . . . . . . . . . . . . 13, 113 Y
subharmonic function . . . . . . . . . . . . . . . . . . 120 Yang’s inequalities . . . . . . . . . . . . . . . . . . . . . . 189
sublevel set . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 Yang–Yau bound . . . . . . . . . . . . . . . . . . . . . . . 184
sum of squares function . . . . . . . . . . . . . . . . . 26 Yau’s conjecture . . . . . . . . . . . . . . . . . . . . . . . . 131
Sunada
construction . . . . . . . . . . . . . . . . . . . . . . 211
triple . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
superlevel set . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Z
Szegő–Weinberger inequality . . . . . . . . . . . . 174 Zaremba problem . . . . . . . . . . . . . . . . . . . . . . 74
The photo courtesy of Elina Alexandrov. Nox the cat courtesy of Elena Kviatkovskaia.