Mathematical

Surveys
and
Monographs
Volume 186

Introduction to Quantum
Graphs

'REGORY "ERKOLAIKO
Peter Kuchment

American Mathematical Society

Introduction to Quantum
Graphs
Mathematical
Surveys
and
Monographs
Volume 186

Introduction to Quantum
Graphs

Gregory Berkolaiko
Peter Kuchment

American Mathematical Society

Providence, Rhode Island
 EDITORIAL COMMITTEE
Ralph L. Cohen, Chair Benjamin Sudakov
Michael A. Singer Michael I. Weinstein

2010 Mathematics Subject Classification. Primary 34B45, 35R02, 81Q35, 35Pxx, 58J50,
05C50, 82D77, 81Q50, 47N60, 82D80, 82D55.

For additional information and updates on this book, visit

www.ams.org/bookpages/surv-186

Library of Congress Cataloging-in-Publication Data

Berkolaiko, Gregory, 1976–
Introduction to quantum graphs / Gregory Berkolaiko, Peter Kuchment.
pages cm. – (Mathematical surveys and monographs ; volume 186)
Includes bibliographical references and index.
ISBN 978-0-8218-9211-4 (alk. paper)
1. Quantum graphs. 2. Boundary value problems. I. Kuchment, Peter, 1949– II. Title.

QA166.165.B47 2012
530.12015115–dc23 2012033434

Copying and reprinting. Individual readers of this publication, and nonprofit libraries
acting for them, are permitted to make fair use of the material, such as to copy a chapter for use
in teaching or research. Permission is granted to quote brief passages from this publication in
reviews, provided the customary acknowledgment of the source is given.
Republication, systematic copying, or multiple reproduction of any material in this publication
is permitted only under license from the American Mathematical Society. Requests for such
permission should be addressed to the Acquisitions Department, American Mathematical Society,
201 Charles Street, Providence, Rhode Island 02904-2294 USA. Requests can also be made by
e-mail to [email protected].

c 2013 by the American Mathematical Society. All rights reserved.
The American Mathematical Society retains all rights
except those granted to the United States Government.
Printed in the United States of America.

∞ The paper used in this book is acid-free and falls within the guidelines
established to ensure permanence and durability.
Visit the AMS home page at http://www.ams.org/
10 9 8 7 6 5 4 3 2 1 18 17 16 15 14 13
To our families and our teachers
 Contents

Preface xi

Introduction xiii
Chapter 1. Operators on Graphs. Quantum graphs 1
1.1. Main graph notions and notation 2
1.2. Difference operators. Discrete Laplace operators 5
1.3. Metric graphs 7
1.4. Differential operators on metric graphs. Quantum graphs 12
1.4.1. Vertex conditions. Finite graphs. 15
1.4.2. Scale invariance 22
1.4.3. Quadratic form 22
1.4.4. Examples of vertex conditions 24
1.4.5. Infinite graphs 27
1.4.6. Non-local vertex conditions 32
1.5. Further remarks and references 33

Chapter 2. Quantum Graph Operators. Special Topics 37

2.1. Quantum graphs and scattering matrices 37
2.1.1. Scattering on vertices 37
2.1.2. Bond scattering matrix and the secular equation 41
2.2. First order operators and scattering matrices 44
2.3. Factorization of quantum graph Hamiltonians 51
2.4. Index of quantum graph operators 52
2.5. Dependence on vertex conditions 54
2.5.1. Variations in the edge lengths 58
2.6. Magnetic Schrödinger operator 59
2.7. Further remarks and references 62

Chapter 3. Spectra of Quantum Graphs 65

3.1. Basic spectral properties of compact quantum graphs 66
3.1.1. Discreteness of the spectrum 66
3.1.2. Dependence on the vertex conditions 67
3.1.3. Eigenfunction dependence 68
3.1.4. An Hadamard-type formula 68
vii
viii CONTENTS

3.1.5. Generic simplicity of the spectrum 71

3.1.6. Eigenvalue bracketing 72
3.1.7. Dependence on the coupling constant at a vertex 76
3.2. The Shnol’ theorem 79
3.3. Generalized eigenfunctions 82
3.4. Failure of the unique continuation property. Scars 84
3.5. The ubiquitous Dirichlet-to-Neumann map 85
3.5.1. DtN map for a single edge 85
3.5.2. DtN map for a compact graph with a “boundary” 87
3.5.3. DtN map for a single vertex boundary 89
3.5.4. DtN map and the secular equation 89
3.5.5. DtN map and number of negative eigenvalues 90
3.6. Relations between quantum and discrete graph spectra 90
3.7. Trace formulas 92
3.7.1. Secular equation 93
3.7.2. Weyl’s law 95
3.7.3. Derivation of the trace formula 96
3.7.4. Expansion in terms of periodic orbits 99
3.7.5. Other formulations of the trace formula 100
3.8. Further remarks and references 101

Chapter 4. Spectra of Periodic Graphs 105

4.1. Periodic graphs 105
4.2. Floquet-Bloch theory 107
4.2.1. Floquet transform on combinatorial periodic graphs 109
4.2.2. Floquet transform of periodic difference operators 112
4.2.3. Floquet transform on quantum periodic graphs 113
4.2.4. Floquet transform of periodic operators 114
4.3. Band-gap structure of spectrum 115
4.3.1. Discrete case 115
4.3.2. Quantum graph case 116
4.3.3. Floquet transform in Sobolev classes 117
4.4. Absence of the singular continuous spectrum 117
4.5. The point spectrum 118
4.6. Where do the spectral edges occur? 121
4.7. Existence and location of spectral gaps 123
4.8. Impurity spectra 124
4.9. Further remarks and references 124

Chapter 5. Spectra of Quantum Graphs. Special Topics 129

5.1. Resonant gap opening 129
5.1.1. “Spider” decorations 133
 CONTENTS ix

5.2. Zeros of eigenfunctions and nodal domains 134

5.2.1. Some basic results 137
5.2.2. Bounds on the nodal count 138
5.2.3. Nodal count for special types of graphs 142
5.2.4. Nodal deficiency and Morse indices 143
5.3. Spectral determinants of quantum graphs 148
5.4. Scattering on quantum graphs 150
5.5. Further remarks and references 153

Chapter 6. Quantum Chaos on Graphs 157

6.1. Classical “motion” on graphs 158
6.2. Spectral statistics and random matrix theory 159
6.2.1. Form factor of a unitary matrix 161
6.2.2. Random matrices 162
6.3. Spectral statistics of graphs 164
6.4. Periodic orbit expansions 167
6.4.1. On time-reversal invariance 170
6.4.2. Diagonal approximation 170
6.4.3. The simplest example of an off-diagonal term 173
6.5. Further remarks and references 179

Chapter 7. Some Applications and Generalizations 181

7.1. Inverse problems 181
7.1.1. Can one hear the shape of a quantum graph? 183
7.1.2. Quantum graph isospectrality 184
7.1.3. Can one count the shape of a graph? 185
7.1.4. Inverse scattering 185
7.1.5. Discrete “electrical impedance” problem 185
7.2. Other types of equations on metric graphs 186
7.2.1. Heat equation 186
7.2.2. Wave equation 186
7.2.3. Control theory 186
7.2.4. Reaction-diffusion equations 187
7.2.5. Dirac and Rashba operators 187
7.2.6. Pseudo-differential Hamiltonians 187
7.2.7. Non-linear Schrödinger equation (NLS) 188
7.3. Analysis on fractals 188
7.4. Equations on multistructures 188
7.5. Graph models of thin structures 188
7.5.1. Neumann tubes 190
7.5.2. Dirichlet tubes 192
7.5.3. “Leaky” structures 194
x CONTENTS

7.6. Quantum graph modeling of various physical phenomena 200

7.6.1. Simulation of quantum graphs by microwave networks 200
7.6.2. Realizability questions 200
7.6.3. Spectra of graphene and carbon nanotubes 201
7.6.4. Vacuum energy and Casimir effect 205
7.6.5. Anderson localization 207
7.6.6. Bose-Einstein condensates 207
7.6.7. Quantum Hall effect 207
7.6.8. Flat band phenomena and slowing down light 208
Appendix A. Some Notions of Graph Theory 209
A.1. Graph, edge, vertex, degree 209
A.2. Some special graphs 210
A.3. Graphs and digraphs 210
A.4. Paths, closed paths, Betti number 211
A.5. Periodic graph 211
A.6. Cayley graphs and Schreier graphs 212
Appendix B. Linear Operators and Operator-Functions 213
B.1. Some notation concerning linear operators 213
B.2. Fredholm and semi-Fredholm operators. Fredholm index 213
B.3. Analytic Fredholm operator functions 215
B.3.1. Some notions from the several complex variables theory 215
B.3.2. Analytic Fredholm operator functions 216
Appendix C. Structure of Spectra 219
C.1. Classification of the points of the spectrum 220
C.2. Spectral theorem and spectrum classification 220
Appendix D. Symplectic Geometry and Extension Theory 223
Bibliography 227
Index 267
 Preface

In this book, the name “quantum graph” refers to a graph con-

sidered as a one-dimensional simplicial complex and equipped with a
differential operator (“Hamiltonian”). Works that currently would be
classified as discussing quantum graphs have been appearing since at
least the 1930s in various areas of chemistry, physics, and mathemat-
ics. However, as a coherent and actively pursued topic, the area of
quantum graphs has experienced an explosive growth only in the last
couple of decades. There are manifold reasons for this surge. Quantum
graphs arise naturally as simplified models in mathematics, physics,
chemistry, and engineering when one considers propagation of waves
of various nature through a quasi-one-dimensional (e.g., “meso-” or
“nano-scale”) system that looks like a thin neighborhood of a graph.
One can mention in particular the free-electron theory of conjugated
molecules, quantum wires, photonic crystals, carbon nano-structures,
thin waveguides, some problems of dynamical systems, system theory
and number theory, and many other applications that have led inde-
pendently to quantum graph models. Quantum graphs also play a role
of simplified, although still non-trivial, models for studying difficult
issues, for instance, Anderson localization and quantum chaos.
There are fruitful relations of quantum graphs with the older spec-
tral theory of “standard” (combinatorial) graphs [191, 195, 213–215,
415] and with what is sometimes called discrete geometric analysis
[682]. Quantum graphs present new non-trivial mathematical chal-
lenges, which makes them dear to a mathematician’s heart. As the
reader will see, work on quantum graph theory and applications has
brought together tools and intuition coming from graph theory, com-
binatorics, mathematical physics, PDEs, and spectral theory.
In the new millennium, these relations between the various topics
leading to quantum graphs were noticed, which has triggered a series
of interdisciplinary meetings and intensive communication and coop-
eration among researchers coming from different areas of science and
engineering. Surveys and collections of papers on quantum graphs and
related issues have started to appear (e.g., [98,121,122,126,161,353,

xi
xii PREFACE

421, 438, 477–480, 536, 614, 615, 622]). These surveys, however, usu-
ally focus on special features of quantum graph theory and there is still
no comprehensive introduction to the topic, which is why the authors
decided to write this text. The book is intended to serve a dual pur-
pose: to provide an introduction to and survey of the current state of
quantum graph theory, as well as to serve as a reference text, where
the main notions and techniques are collected.
The authors are indebted to many colleagues, from whom they have
learned over the years a great deal about quantum graph theory and
related issues. This includes, in particular, M. Aizenman, R. Band,
J. Bolte, R. Carlson, Y. Colin de Verdiere, P. Exner, A. Figotin,
M. Freidlin, L. Friedlander, S. Fulling, S. Gnutzmann, R. Grigorchuk,
J.M. Harrison, J. P. Keating, E. Korotyaev, V. Kostrykin, M. Kotani,
T. Kottos, L. Kunyansky, S. Molchanov, V. Nekrashevich, S. Novikov,
K. Pankrashkin, B. Pavlov, O. Post, H. Schantz, R. Schrader, M. Shu-
bin, U. Smilansky, A. Sobolev, M. Solomyak, T. Sunada, A. Teplyaev,
B. Vainberg, I. Veselić, and B. Winn. The second author has been sig-
nificantly influenced by the late M. Birman, R. Brooks, and V. Geyler.
We cordially thank S. Fulling, W. Justice and our graduate students
N. Do, W. Liu and T. Weyand for their critical reading of the manu-
script and numerous corrections.
We are grateful to the reviewers, especially those who had to endure
the cruel and unusual punishment of refereeing several versions, for
many important suggestions. Our thanks go to the people from the
AMS office for invaluable help. We are especially grateful to S. Gelfand,
without whose insistence this book would have never been finished, and
C. Thivierge for constant support during the production.
Some parts of this work were supported by grants from the National
Science Foundation. The authors express their gratitude to the NSF
for the support.
Finally, we are grateful to our families for their limitless patience.

Gregory Berkolaiko and Peter Kuchment

College Station, TX
July 2012
 Introduction

In this book, our goal is to introduce the main notions, structures,

and techniques used in quantum graph studies, as well as to provide
a brief survey of more special topics and applications. This task has
shaped the book as follows: we present in detail the basic constructions
and frequently used technical results in Chapters 1 and 2, devoted to
quantum graph operators, and Chapters 3 – 5, which address various
issues of the spectral theory of quantum graphs. The remaining two
chapters are of review nature and thus less detailed; in most cases the
reader will be directed to the cited literature for precise formulations
and proofs. Using graphs as models for quantum chaos is considered in
Chapter 6. Chapter 7 provides a brief survey of various generalizations
and applications. The reader will notice that the area is developing
very fast; had we tried to be more specific in this chapter, it would be
outdated by the time of publications anyway.
Our intent was to make the book accessible to graduate and ad-
vanced undergraduate students in mathematics, physics, and engineer-
ing.
Since a variety of techniques are used, for the benefit of the reader
we introduce the main notions and relevant results in graph theory,
functional analysis, and operator theory in a series of Appendices.
In order to make reading smoother, we normally do not include
references in the main text of the chapters, collecting them, as well
as additional comments, in the specially devoted last section of each
chapter. We also have not tried to make the considerations too gen-
eral. For instance, we mostly treat the second derivative operators on
quantum graphs, while considerations could be easily extended to the
more general Schrödinger operators. When we do mention more gen-
eral operators, we do not look for the most general conditions on the
coefficients (potentials), settling for some reasonable conditions that
make the techniques work.

xiii
 CHAPTER 1

Operators on Graphs. Quantum graphs

In this chapter, we introduce the main players of the quantum graph

theory: metric graphs and differential operators on them. A graph
consists of a set of points (vertices) and a set of segments (edges) con-
necting some of the vertices (Fig. 1).
More notions and results concerning graph theory can be found in
Section 1.1) and Appendix A. Most mathematicians are already famil-
iar with combinatorial graphs (which we survey briefly in Section 1.1),
where the vertices are the main players and the edges merely indicate
some relations between them. In a metric graph, in contrast, atten-
tion is focused on the edges. Metric graphs are introduced in Section
1.3. Quantum graphs are essentially metric graphs equipped with dif-
ferential operators. Such operators (Hamiltonians) are considered in
Section 1.4. The main operator under consideration acts as the second
derivative along the edges with “appropriate” conditions at junctions
(vertices). These conditions generalize the boundary conditions for
ODEs. Here, a lot of attention is devoted to describing what are the
“appropriate” conditions. Considering the quantum graph from the
point of view of waves propagating along edges and scattering at ver-
tices and other more advanced (but fundamental) topics are deferred
to Chapter 2.

Figure 1. A graph
1
2 1. OPERATORS ON GRAPHS

1.1. Main graph notions and notation

We start by introducing some common graph notions and notation
used throughout the text. The reader is also referred to Appendix A
for these and some other notions from the graph theory.
A graph Γ consists of a finite or countably infinite set of vertices
V = {vi } and a set E = {ej } of edges connecting the vertices. The edges
are undirected. As we will see in Section 1.4, in the quantum graph
situation one can usually assume absence of loops and multiple edges,
since if these are present, one can break them into pieces by introducing
new intermediate vertices. We thus will be assuming mostly that loops
and multiple edges between vertices are not present. We will use the
notation E := |E| and V := |V| for the number of edges and vertices
correspondingly. The notation v ∈ e will be taken to mean that v is
a vertex of the edge e. Two vertices u and v will be called adjacent
(denoted u ∼ v) if there exists an edge connecting them. A graph Γ is
fully specified by its |V| × |V| adjacency matrix AΓ . In the simplest
case when there are no loops or multiple edges, the elements of the
adjacency matrix are given by

1 if u ∼ v,
(1.1.1) Au,v =
0 otherwise.
More generally,
(1.1.2) Au,v = |{e ∈ E : u ∈ e, v ∈ e}| .
The degree  du of a vertex u is the number of edges emanating
from it, du = v∈V Au,v . All degrees are assumed to be finite (local
finiteness of the graph).
We will denote by DΓ the degree matrix, i.e. the diagonal |V|×|V|
matrix with the diagonal entries dv :
(1.1.3) Du,v = dv δu,v ,
where δu,v is the Kronecker delta

1 if u = v,
δu,v =
0 otherwise.
Sometimes it becomes necessary to consider directed edges. A graph
is called directed graph or digraph, if each of its edges is assigned a
direction. In other words, each edge has one origin and one terminal
vertex. Directed edges will be called bonds. The set of all bonds is
denoted by B. We will use the shorthand notation B := |B| for the
total number of bonds in a directed graph Γ. The origin and terminal
 1.1. MAIN GRAPH NOTIONS AND NOTATION 3

points of a bond are specified via functions o : B → V and t : B → V,

i.e. a bond b begins at vertex o(b) and ends at t(b).
We define the set of incoming bonds at a vertex v as the set of
bonds satisfying t(b) = v. If o(b) = v, the bond b is called outgoing at
vertex v. The number of incoming (respectively, outgoing) bonds at a
vertex v is called incoming (resp.outgoing) degree of v and denoted
div (resp. dov ). Clearly, dov + div = dv .
Most frequently, in our considerations a non-directed graph will be
considered as a digraph by assigning two bonds b and b with opposite
directions to each edge e, as shown in Figure 2. We denote the resulting
directed graph by Γ. 

Figure 2. An undirected graph Γ (left) and the corre-

 obtained by duplicating bonds
sponding directed one Γ
(right).

The graph Γ  satisfies the condition that at each vertex v the numbers of
incoming and outgoing bonds are equal: dov = div = dv /2 (the notation
dv := dv /2 will be used in this situation). The set of bonds of Γ  is
symmetric in the sense that b ∈ B if and only if there is another bond
b ∈ B such that o(b) = t(b) and t(b) = o(b). The bond b is called the
reversal of b. The operation of reversal is reflexive: b = b.
We will be considering graphs from two different prospectives. The
first one, standard in graph theory and combinatorics, is where edges
are considered as relations between vertices, rather than any physical
or geometrical links. To emphasize this, we will use the term discrete
or combinatorial graph in such situations. On the other hand, in
most of the text graphs will be considered as 1D complexes, and thus
edges will be treated as 1D segments (or could be thought of as physical
“wires”). Such graphs will be equipped with additional structures that
will make them metric or quantum graphs (see Sections 1.3 and
1.4).
If one considers graphs as 1D complexes, which we will often do,
there exists a natural projection π : Γ  → Γ that maps the points x1
4 1. OPERATORS ON GRAPHS

and x2 of the two bonds arising from the same point x of an edge to
the point x (see Fig. 3).

→
Figure 3. Projection π : Γ  Γ illustrated on a single edge.
Another notion that we will frequently encounter is of the cyclo-
matic number:
Definition 1.1.1. Let Γ be a connected graph. The cyclomatic
number βΓ (or the 1st Betti number, or simply Betti number)
of Γ is the number of edges that have to be removed from E (without
reducing the number of vertices) to turn Γ into a tree1 [90].
Remark 1.1.2. Several simple observations are in order.
• It is immediate from the definition that βΓ = 0 if and only if
Γ is a tree.
• Since the number of vertices is preserved and a tree on |V|
vertices has |V| − 1 edges, we see that to turn the graph Γ into
a tree we need to remove
(1.1.4) βΓ = |E| − |V| + 1
edges. This explicit expression implies that the Betti number
is well-defined, in that it does not depend on which set of edges
is removed.
• The Betti number is equal to the rank of the fundamental
group2 of Γ considered as a 1D complex.
• In the case of a graph with k connected components, one can
generalize the concept of the Betti number to mean the number
of edges to be removed in order to turn Γ into a forest (a dis-
joint union of trees) with the same number of components. It
is easy to see that the Betti number of the disconnected graph
is equal to the sum of the Betti numbers of its components.
Formula (1.1.4) is thus replaced with
(1.1.5) βΓ = |E| − |V| + k.
1A tree is a connected graph without cycles, see also Appendix A
2See the definition of the fundamental group, e.g., in [241] or any book on
elements of algebraic topology.
 1.2. DIFFERENCE OPERATORS 5

1.2. Difference operators. Discrete Laplace operators

Let Γ be a discrete (as opposed to “metric”, see Section 1.3) graph.
A real or complex valued function f on Γ takes its values at vertices
and thus can be identified with an element of R|V| or C|V| correspond-
ingly.
We recall some basic difference operators acting on functions defined
on Γ. The main one is the discrete Laplace operator, which has several
versions, reviewed below. Different versions are useful under different
circumstances.
We assume initially that the graph has no loops or multiple edges.
Definition 1.2.1.
• The combinatorial Laplacian LΓ (or just L, if no confusion
can arise) acting on functions f (u) defined on Γ is given by
the self-adjoint V × V -matrix3 L with entries
⎧
⎪
⎨du if u = v,
(1.2.1) Lu,v = −1 if u ∼ v,
⎪
⎩0 otherwise,
where du is the degree of the vertex u. In other words,
(1.2.2) (LΓ f )(u) = Lu,v f (v) = du f (u) − f (v).
v∈V v∼u

• The quadratic form associated with the combinatorial Lapla-

cian is
(1.2.3) l[x, x] := (Lx, x) = |xu − xv |2 = Au,v |xu − xv |2 ≥ 0,
u∼v u,v∈V

where Au,v are the entries of the adjacency matrix AΓ (see

(1.1.2)).
• The harmonic Laplacian ΔΓ is defined as
(1.2.4) ΔΓ := DΓ−1 L,
where DΓ is the diagonal degree matrix (1.1.3).
Remark 1.2.2.
• Notice that
(1.2.5) LΓ = DΓ − AΓ ,
where AΓ is the adjacency matrix of Γ. For graphs with loops
or multiple edges, this formula can be taken as the definition of
3As before, we use the shorthand notation V for the number of vertices |V|.
6 1. OPERATORS ON GRAPHS

the Laplacian. The last expression in (1.2.3) gives the correct

quadratic form also for such graphs.
• As the result, the harmonic Laplacian can be represented as
(1.2.6) ΔΓ = I − DΓ−1 AΓ .
The operator P = DΓ−1 AΓ appearing in the right hand side of
(1.2.6) is sometimes called the Markov operator, because of
its role in random walks and Markov chains theory.
• Notice that the positivity of the quadratic form (see (1.2.3))
implies that the operator LΓ is non-negative.  Comparing with
the standard Laplace operator in Rn , Δ = j ∂x2j , one notices
that the latter is in fact non-positive. Operator LΓ is analogous
to what is sometimes called the “geometer’s Laplacian”, which
in Rn is −Δ.
Also widely used is the following version of the Laplacian:
Definition 1.2.3. The normalized Laplacian L has the matrix
with entries
⎧
⎪
⎨1 if u = v,
(1.2.7) Lu,v = − √du dv if u ∼ v,
1
⎪
⎩
0 otherwise.
In other words,
(1.2.8) L = D−1/2 LD−1/2 .
If we orient the edges of the graph Γ in an arbitrary fashion, we
can define the operator that is an analog of exterior derivative. It is
represented by the B × V -matrix d
⎧
⎪
⎨1 if v = t(b),
(1.2.9) db,v = −1 if v = o(b),
⎪
⎩0 otherwise,
and acts from functions on vertices to functions on bonds. Then the
Laplacian L can be written as
(1.2.10) L = d∗ d,
where d∗ : C|B| → C|V| denotes the dual operator to d : C|V| → C|B|
with respect to the standard inner products in C|V| and C|B| . Relation
(1.2.10) is a discrete analog of the representation of the Laplace opera-
tor Δ = div grad = ∇ · ∇. The matrix of the operator d∗ is sometimes
called the incidence matrix of the directed graph Γ.
 1.3. METRIC GRAPHS 7

If the graph Γ is considered as a 1D cell complex, the definition

of the operator d allows a topological interpretation. Namely, let us
denote by C 0 (Γ) the space of 0-co-chains on Γ, i.e. functions f (v) on
the set of vertices V. The set of 1-co-chains C 1 (Γ) consists of skew-
symmetric functions h(b) on the set B of bonds, i.e. such that h(b) =
−h(b). Then
d : C 0 (Γ) → C 1 (Γ), df (b) := f (t(b)) − f (o(b))
is what is called in algebraic topology the co-boundary operator.
The representation (1.2.10) suggests another generalization, which
often happens to be useful in applications. Namely, assume that some
positive weight functions m(v) on the set V of vertices and μ(b) on the
set B of bonds are assigned. One can define an operator by (1.2.10),
where, however, the dual operator d∗ is defined using weighted inner
|V| |B|
products on the vector spaces C and C : (f, g) = v f (v)g(v)m(v)

and (p, q) = b p(b)q(b)μ(b) correspondingly. This leads to the follow-
ing definition:
Definition 1.2.4. The weighted Laplacian Δm,μ,Γ acts as fol-
lows:
1
(1.2.11) Δm,μ,Γ f (u) = μ(e)(f (u) − f (v)).
m(u)
e=(u,v)∈Eu

Sometimes we will need to deal with more general difference oper-

ators.
Definition 1.2.5. A linear finite order difference operator M
on Γ is given by a matrix {Mu,v } with u, v ∈ V such that the following
conditions hold:
(1) Each row and column of the matrix M has a finite number of
non-zero entries.
(2) There is a positive integer ρ (the order of the operator) such
that Mu,v = 0 for any pair of vertices u, v such that the length
of the shortest path connecting u and v exceeds ρ.

1.3. Metric graphs

We have dealt so far with graphs considered as discrete (combina-
torial) objects. Here we add a structure that equips Γ with a topology
and metric. Roughly speaking, now one will need to imagine the edges
of Γ not as abstract relations between vertices, but rather as physical
“wires” connecting them.
Definition 1.3.1. A graph Γ is said to be a metric graph, if
8 1. OPERATORS ON GRAPHS

(1) each bond (i.e. a directed edge) b is assigned a positive length

Lb ∈ (0, ∞). We will often need to consider edges of infinite
length. Such an edge has only one vertex (the other end goes
to “infinity”) and will be called a lead;
(2) the lengths of the bonds that are reversals of each other are
assumed to be equal: Lb = Lb , and thus the length Le of an
edge e is also defined;
(3) a coordinate xb ∈ [0, Lb ] increasing in the direction of the bond
is assigned on each bond;
(4) the relation xb = Lb − xb holds between the coordinates on
mutually reversed bonds (in other words, xb and Lb − xb are
mapped to the same point on e by the projection π, see Fig. 3).
In most cases, when this does not lead to any confusion, we will
drop the subscript in the coordinate xb and denote it by x.
As the reader will see, it is often important how the lengths of
different edges are related. In some topics (e.g., quantum chaos or
inverse problems) it helps when the lengths are as unrelated as possible,
more precisely, are rationally independent. In some other instances, for
example when trying to relate spectra of quantum and discrete graphs,
the opposite happens, and one benefits from assuming that all lengths
are the same. This justifies the following notion:
Definition 1.3.2. A metric graph Γ is said to be equilateral, if
the lengths of all its bonds (equivalently, edges) are equal: Lb ≡ L.
The introduced metric graph structure enables one to interpret the
graph Γ as a topological space (simplicial complex) that is the union
of all edges, with the ends of the edges that correspond to the same
vertex v being identified.
The reader should note that we do not assume the graph to be em-
bedded in any way into a Euclidean space or a Riemannian manifold.
In some applications such a natural embedding does exist (e.g., in mod-
eling quantum wire circuits, carbon nanotubes, or photonic crystals),
and in such cases the coordinate along an edge is usually the induced
arc length. In some other applications (e.g., in quantum chaos) the
graph does not need to be embedded anywhere and can be considered
as an abstract complex.
As the name suggests, any metric graph Γ can be equipped with a
 If a sequence of edges {ej }j=1 forms a path, its length
M
natural metric.
is defined as Lj . For two vertices v and w, the distance ρ(v, w) is
defined as the minimal length of the path connecting them. Since along
each edge the distance is determined by the coordinate x, it is easy to
 1.3. METRIC GRAPHS 9

define the distance ρ(x, y) between two points x and y of the graph
that are not necessarily vertices. We leave this to the reader.
In the metric or quantum graph setting, we will be always assuming
that the degree of each vertex is finite and positive. In particular,
vertices that are not incident to any edge are prohibited.
Remark 1.3.3. If a metric graph has multiple edges connecting two
vertices, or a loop at a vertex, one can introduce additional vertices in-
side parallel edges and inside the loop. These vertices will break down
the relevant edges into smaller pieces and eliminate loops and multi-
plicity (see Fig 4). Thus, in metric (and quantum) graph situations
one can usually assume without loss of generality absence of multiple
edges and loops.

Figure 4. Breaking multiple edges and loops: a graph

with multiple edges and a loop (left) and the same graph
after breaking them (right).

Definition 1.3.4. We will call a metric graph infinite if it has in-

finitely many vertices (equivalently, infinitely many edges). Otherwise
the graph will be called finite. A finite graph, whose edges all have
finite lengths will be called compact.
A compact graph is also compact as a topological space. One should
notice that not every finite graph is compact. Indeed, a finite graph,
although having finitely many edges, might have some edges of infinite
length and thus be non-compact.
For infinite graphs (i.e., those with infinitely many vertices), some-
times the following additional condition is adopted:
Assumption 1.3.5 (Finite ball condition). For any positive number
r and any vertex v, there is only a finite number of vertices w at a
distance smaller than r from v.
This implies, in particular, that the distance between any two dis-
tinct vertices is strictly positive, and no finite length paths with infin-
itely many edges exist. This restriction matters only for graphs with
infinitely many edges and is normally satisfied in applications.4
4Sometimes (e.g., when studying fractals or infinite quantum trees), this as-
sumption is too restrictive and is abandoned.
10 1. OPERATORS ON GRAPHS

As we have already mentioned, a small peculiarity arises when edges

of infinite length are present. Then it is usually assumed that each of
these edges has only one vertex and extends to infinity in the other
direction. Thus, such a graph can be thought of as a graph with edges
of finite lengths, with one or more additional infinite “leads” attached
to some vertices. Such graphs arise naturally, for instance in scattering
theory.
One should notice that the points of a metric graph Γ are not
only its vertices, but all intermediate points x on the edges as well.
Thus, when we speak about functions f (x) on Γ, we consider them as
defined along the edges (rather than at the vertices only, as in discrete
models). One can in particular talk about continuous functions and
thus define the standard space of continuous functions C(Γ). Presence
of the coordinate x along the edges enables one to define in the natural
way the Lebesgue measure dx on the graph. Having this measure, one
can define some other standard function spaces on Γ. We recall first
the notation H k (e) for the Sobolev space5 of functions on the segment
e that have all their distributional derivatives up to the order k belong
to L2 (e).
Definition 1.3.6.
(1) The space L2 (Γ) on Γ consists of functions that are measur-
able and square integrable on each edge e and such that
(1.3.1) f 2
L2 (Γ) := f 2
L2 (e) < ∞.
e∈E

In other words, L2 (Γ) is the orthogonal direct sum of spaces

L2 (e).
(2) The Sobolev space H 1 (Γ) consists of all continuous functions
on Γ that belong to H 1 (e) for each edge e and such that
(1.3.2) f 2
H 1 (Γ) := f 2
H 1 (e) < ∞.
e∈E

The reader notices that the conditions of finiteness of the sums in

this definition are relevant for infinite graphs only.
The continuity condition imposed on functions from the Sobolev
space H 1 (Γ) means that any function f from this space assumes the
same value at a vertex v on all edges adjacent to v, and thus f (v)
is uniquely defined. This is a natural condition for one-dimensional
H 1 -functions, which are known to be continuous in the standard 1D
setting.
5See the standard definitions of the Sobolev spaces, e.g., in [7].
 1.3. METRIC GRAPHS 11

There seems to be no natural definition of Sobolev spaces H k (Γ) of

order k higher than 1. The reason is that, unlike the 1D case, there
are no natural conditions such functions should satisfy at vertices. And
indeed, as we shall see further in this book, such junction conditions
will depend on a particular Hamiltonian studied.
We will often employ the spaces where smoothness is enforced along
edges only, without any junction conditions at the vertices at all. We
thus introduce the shorthand notation that places a tilde on the top of
the space to signify this:
 k (Γ) the space
Definition 1.3.7. We denote by H

(1.3.3)  k (Γ) :=
H H k (e),
e∈E

which consists of the functions f on Γ that on each edge e belong to

the Sobolev space H k (e) and such that

(1.3.4) f 2
k
H
:= f 2
H k (e) < ∞.
e∈E

(The latter finiteness condition is needed for infinite graphs only.)

We will need sometimes the following well known trace estimate,
which we prove here for completeness.
Lemma 1.3.8. Let f ∈ H 1 [0, a], then
2
(1.3.5) |f (0)|2 ≤ f 2
L2 [0,a] + l f 2
L2 [0,a]
l
for any l ≤ a.
Proof. Due to H 1 -continuity of both sides of the inequality, it
is sufficient to prove it for smooth functions. Let us start with the
representation
x

(1.3.6) f (0) = f (x) − f  (t)dt, x ∈ [0, l]

0

and estimate the integral there using the Cauchy-Schwarz inequality:

x 2

f  (t)dt ≤ f 2
L2 [0,a] χ[0,x] 2
L2 [0,a] = x f 2
L2 [0,a] .
0
12 1. OPERATORS ON GRAPHS

This implies
x 2 l
 l2 
f (t)dt ≤ f  2L2 [0,a] x dx = f 2
L2 [0,a] .
2
0 L2 [0,l] 0

Now, taking L2 [0, l]-norms in both sides of (1.3.6) and using the in-
equality (a + b)2 ≤ 2a2 + 2b2 , we get the estimate
l|f (0)|2 ≤ 2 f 2
L2 [0,a] + l2 f  2
L2 [0,a] ,

which implies the statement of the lemma. 

1.4. Differential operators on metric graphs. Quantum

graphs
A metric graph becomes a quantum one after being equipped with
an additional structure: assignment of a differential (or sometimes more
general) operator on Γ. This operator will be often called the Hamil-
tonian. In most cases, but not always, the Hamiltonian is required to
be self-adjoint. In the studies of quantum graphs, the most frequently
used operator is the negative second derivative acting on each edge:
d2 f
(1.4.1) f (x) → − 2 .
dx
Here, as before, x is the coordinate x along an edge.
A more general Schrödinger operator
d2 f
(1.4.2) f (x) → −
+ V (x)f (x),
dx2
where V (x) is an electric potential, is also very common.
Notice that for both of these operators the direction of the edge is
irrelevant. This is not so anymore if one wants to include first order
derivative terms, e.g. magnetic potentials, or to consider first order
operators like d/dx (see Section 2.2). In such cases, we need to assume
the graph to be directed and consider bonds b (directed edges) and the
corresponding coordinates xb . The magnetic Schrödinger operator can
then be defined as
 2
1 d
(1.4.3) f (xb ) → − Ab (xb ) f (xb ) + V (xb )f (xb ),
i dxb
with the electric potential V and magnetic potential A. Note that
A is a vector field (in one dimension); in other words, it changes sign
when the direction of the bond is reversed: Ab (xb ) = −Ab (xb ) (the
reader should recall that xb = Lb − xb ).
 1.4. DIFFERENTIAL OPERATORS. QUANTUM GRAPHS 13

The classical case of differential operators on a single segment (i.e.,

graph with one edge) makes clear that the definition of the quan-
tum graph Hamiltonian is not complete until its domain is described.
The same experience shows that the domain description should involve
smoothness conditions along the edges and some junction conditions
at the vertices (analogs of boundary conditions for a single interval).
Moreover, the self-adjointness requirement will impose additional re-
strictions on the vertex conditions.
For “decent” potentials V and A, the natural smoothness require-
ment coming from the ODE theory is that f belongs to the Sobolev
space H 2 (e) on each edge e. Appropriate boundary value conditions at
the vertices (vertex conditions) still need to be added. We postpone
discussion of these till the next section.
Definition 1.4.1. Quantum graph is a metric graph equipped
with a differential operator H (Hamiltonian), accompanied by “appro-
priate” vertex conditions. That is, a quantum graph Γ is a triple
{metric graph Γ, Hamiltonian H, vertex conditions}.
In many (probably most) cases we will assume that H acts as the
(negative) second order derivative (1.4.1) along the edges. More gen-
eral Hamiltonians (1.4.2)-(1.4.3) also arise, although we will consider
magnetic Schrödinger operator in more detail in Section 2.6 and will
find that it can be reduced to the non-magnetic operator on a slightly
modified graph. Possibilities of even more general scalar or matrix (e.g.,
Dirac) differential or pseudo-differential operators will be mentioned at
a later stage (see Chapter 7). However, in the most of the text we
will concentrate on the most common scalar second order differential
operators, and for simplicity of exposition, on (1.4.1).
As we have mentioned already, in order for the definition of an op-
erator to be complete, one needs to describe its domain. Assuming that
we are dealing with the second derivative operator (or with Schrödinger
with sufficiently “nice” potentials), it is natural to require that f be-
longs to the Sobolev space H 2 (e) on each edge e. Indeed,
 consider the
second derivative operator defined on the direct sum e∈E C0∞ (e) of the
spaces of smooth functions on edges vanishing with all their derivatives
6
at the vertices. It is easy
 to check that the closure of the operator in
L2 (Γ) has the domain e∈E H0 (e), where H02 (e) is the Sobolev space of
2

order 2 on the edge e of functions vanishing with their first derivatives

at the ends of the edge. The operator thus defined is clearly symmetric.
6In order to show this, one can squeeze the functions to a smaller segment than
e and then mollify them to make them infinitely differentiable.
14 1. OPERATORS ON GRAPHS

Now our task boils down to finding boundary value conditions at the
vertices that would produce self-adjoint extensions of this operator.
Before discussing general vertex conditions, we describe the most
commonly used example. Those are called the “Neumann” or “Kirch-
hoff” or even “standard” conditions:
⎧
⎪
⎨f (x) is continuous on Γ
(1.4.4) and
⎪
⎩at each vertex v one has  df
e∈Ev dxe
(v) = 0.

Here the sum is taken over the set Ev of all edges e incident to the
vertex v and the derivatives are assumed to be taken in the
directions away from the vertex (i.e., into the edge), which we
will call the outgoing directions. We will assume this agreement
about outgoing differentiations whenever such conditions are involved.
The first condition of (1.4.4) is called the continuity condition while
the second is sometimes termed the current conservation condition.

We find that the more suitable name for (1.4.4) is Neumann con-
dition (sometimes the conciliatory name Neumann–Kirchhoff con-
ditions will be used). Why is it Neumann? Indeed, at the “loose
ends” (vertices of degree 1) this turns into the actual Neumann condi-
tion. Moreover, as the Neumann boundary condition for the Laplace
operator, it is natural. This means that, as we will see in Section 1.4.4,
the domain of the quadratic form of the corresponding operator does
not require any conditions on a function besides its being in H 1 (Γ).

Remark 1.4.2. When the boundary conditions (1.4.4) hold at a

vertex of degree 2, the vertex can be eliminated, thus combining two
adjacent edges into one smooth edge. Then (1.4.4) exactly guarantees
that the adjacent H 2 -pieces of the function match into a single H 2 -
function on the resulting long edge: the continuity condition ensures
continuity (naturally) while the current conservation condition ensures
continuity of the first derivative.

Example 1.4.3 (Spectrum of compact Neumann star graph). Con-

sider the simplest non-trivial graph, the star graph of Fig. 6. We im-
pose Neumann conditions at every vertex of the graph and would like
to characterize the spectrum of the Laplacian (1.4.1). It will be shown
in the subsequent sections that with these conditions the operator is
self-adjoint and non-negative and its spectrum is discrete.
 1.4. DIFFERENTIAL OPERATORS. QUANTUM GRAPHS 15

Figure 5. A compact star graph.

Taking these properties as given, the eigenvalue equation can be

written as
d2 f
(1.4.5) − = k 2 f (xb ), k ∈ R.
dx2b
Consider the function f on the bond b going from a peripheral to
the central vertex. At the peripheral vertex we have the Neumann
condition f  (0) = 0; together with equation (1.4.5) this implies that on
the bond b the eigenfunction f must have the form f (xb ) = Ab cos(kxb ).
The outgoing derivative of f (xb ) at the central vertex xb = Lb is equal
to Ab k sin(kLb ). Altogether, conditions (1.4.4) at the central vertex
become

A1 cos(kL1 ) = A2 cos(kL2 ) = . . . = C,
E
b=1 Ab k sin(kLb ) = 0.

Dividing the second equation by C, we get that k 2 is an eigenvalue if

E
(1.4.6) F (k) := tan(kLb ) = 0.
b=1

The case C = 0 requires a separate consideration which leads to the

conclusion that if a value of k is a pole for n out of E summands in
F (k), then k 2 is an eigenvalue of multiplicity n − 1. In particular, if Lb
are incommensurate (i.e. Lb /Lb ∈ Q), then all eigenvalues of the star
graph are simple and k 2 is an eigenvalue if and only if F (k) = 0.
1.4.1. Vertex conditions. Finite graphs. We will now study
what boundary conditions one can add to the differential expression
(1.4.1) in order to create a self-adjoint operator. Everything we will
do here applies without change to the case of the operator (1.4.2) with
appropriate conditions on the electric potential V (e.g., continuous or
locally L2 ).
We have agreed on assuming that functions from the domain of the
operator belong to the Sobolev space H 2 (e) on each edge e. Then the
standard Sobolev trace theorem (e.g., [7, 249]) implies that f ∈ H 2 (e)
16 1. OPERATORS ON GRAPHS

and its first derivative have correctly defined values at the endpoints7
of the edge e. Hence, the vertex conditions may involve only the values
of f and df /dx at a vertex v. One could (and sometimes does) allow
the vertex conditions to involve the values at different vertices. For
instance, the periodicity condition for a function f on the segment [0, 1]
does just that: f (0) = f (1), which involves two vertices. We, however,
will deal at the moment with the local vertex conditions only, i.e.
those that involve the values of functions and their derivatives at a
single vertex at a time. We will see in Section 1.4.6 that the general,
non-local, case can be reduced to the local one by modifying the graph.
For simplicity of consideration, we start with the case of finite
graphs and will deal with infinite graphs in Section 1.4.5. That is,
we assume that the number of edges E = |E| and the number of ver-
tices V = |V| are finite. The edges are still allowed to have infinite
length.
If we are interested in local vertex conditions, when different vertices
do not interact, it is sufficient to concentrate on a single junction of
d edges at a vertex v. An arbitrary graph locally (i.e., near a vertex)
looks like a star graph. This is why quantum star graphs are frequent
subjects of study and thus will feature prominently in several parts of
this book.

Figure 6. Locally around a vertex every graph looks

like a star graph

Let us introduce some notations first. For an H 2 -function f defined

along the edges of the star graph, we denote by F (v) the column vector
⎛ ⎞
f1 (v)
⎜ ... ⎟
(1.4.7) F (v) := (f1 (v), . . . , fd (v))t = ⎜
⎝ ... ⎠
⎟
fd (v)

of the values at the vertex v that the function f attains along the edges
incident to v (e.g., if f is continuous, all these values will be the same).

7Second order derivatives of H 2 functions, though, do not have values at single

points.
 1.4. DIFFERENTIAL OPERATORS. QUANTUM GRAPHS 17

Analogously,
⎛ ⎞
f1 (v)
⎜ ... ⎟
(1.4.8) F  (v) := (f1 (v), ..., fd (v))t = ⎜
⎝ ... ⎠
⎟
fd (v)

is the column vector of the values at v of the derivatives of f taken

along these edges in the outgoing directions.
Since H acts as a second order operator on each edge, one needs
to establish two conditions per edge. Therefore, at each vertex v the
number of conditions coincides with the degree dv of the vertex. Since
for functions in H 2 on each edge the conditions may involve only the
boundary values of the function and its derivative, the most general
form of such (homogeneous) condition is

(1.4.9) Av F (v) + Bv F  (v) = 0.

Here Av and Bv are (dv × dv )-matrices. In order to guarantee that the

correct number of independent conditions (equal to the degree dv of
the vertex) is imposed, the rank of the dv × 2dv matrix (Av , Bv ) must
be equal to dv , i.e., maximal.
What other restrictions should be imposed on the matrices enter-
ing (1.4.9) to guarantee self-adjointness of the resulting operator? This
question can be completely answered. This can be done, for instance,
by using the von Neumann theory of extensions of symmetric opera-
tors, or the equivalent symplectic geometry approach (Appendix D).
In particular, according to the theory of extensions of symmetric op-
erators, one expects all self-adjoint conditions at a vertex of degree d
to be parameterized by elements of the unitary group U (d), which has
real dimension d2 . Indeed, assuming the edges of the star-graph to be
infinite, and thus leaving only one vertex, one can find out that the de-
ficiency index8 of the operator is (d, d). We, however, will address the
issue of vertex conditions without explicitly referring to the extension
theory or symplectic geometry.
In the next theorem, we provide different equivalent descriptions
of all possible self-adjoint vertex conditions, all of which are useful in
various circumstances. Another useful description of these conditions
was obtained in [184].

8See, e.g., Appendix C and [16, 436]

18 1. OPERATORS ON GRAPHS

Theorem 1.4.4. Let Γ be a metric graph with finitely many edges.

d2
Consider the operator H acting as − 2 on each edge e, with the do-
dxe
main consisting of functions that belong to H 2 (Γ) and satisfying some
local vertex conditions involving vertex values of functions and their
derivatives. The operator is self-adjoint if and only if the vertex condi-
tions can be written in one (and thus any) of the following three forms:
A: For every vertex v of degree dv there exist dv × dv matrices
Av and Bv such that
(1.4.10) the dv × 2dv matrix (Av Bv ) has the maximal rank,

(1.4.11) the matrix Av Bv∗ is self-adjoint,

and the boundary values of f satisfy
(1.4.12) Av F (v) + Bv F  (v) = 0.
B: For every vertex v of degree dv , there exists a unitary dv × dv
matrix Uv such that the boundary values of f satisfy
(1.4.13) i(Uv − I)F (v) + (Uv + I)F  (v) = 0,
where I is the dv × dv identity matrix.
C: For every vertex v of degree dv , there are three orthogonal
(and mutually orthogonal) projectors PD,v , PN,v and PR,v :=
I − PD,v − PN,v (one or two projectors can be zero) acting in
Cdv and an invertible self-adjoint operator Λv acting in the
subspace PR,v Cdv , such that the boundary values of f satisfy
⎧
⎪
⎨PD,v F (v) = 0 “Dirichlet part”,
(1.4.14) 
PN,v F (v) = 0 “Neumann part”,
⎪
⎩P F  (v) = Λ P F (v) “Robin part”.
R,v v R,v

Remark 1.4.5. Conditions B and C are obviously special cases

of the condition of the type A. The restrictions in condition A are
the weakest (and thus the easiest to check). Its drawback is the non-
uniqueness of the choice of A and B. The other conditions are more
tight in the sense that the constituent matrices are defined uniquely.
Remark 1.4.6. Vertex conditions can also be described in a differ-
ent manner, by prescribing how waves of various frequencies scatter at
each vertex. This approach is popular in studies of quantum chaos on
graphs. We will review this approach in Sections 2.1 and 2.2.
 1.4. DIFFERENTIAL OPERATORS. QUANTUM GRAPHS 19

The following lemma is used in the proof of Theorem 1.4.4. It also

introduces the matrix σ(k) that takes the center stage in Section 2.1
and features prominently thereafter.
Lemma 1.4.7. Let matrices A and B satisfy conditions (1.4.10) and
(1.4.11). Then for any real k = 0 the matrix A + ikB is invertible and
the matrix σ(k) = −(A + ikB)−1 (A − ikB) is unitary.
Proof. We observe that (A + ikB)∗ = (A∗ − ikB ∗ ) and, since
AB ∗ = BA∗ (condition (1.4.11)),
(A + ikB)(A∗ − ikB ∗ ) = AA∗ + k 2 BB ∗
(1.4.15)
= (A − ikB)(A∗ + ikB ∗ ) = (A kB)(A∗ kB ∗ )T ,
where (A kB) denotes the dv × 2dv matrix composed of A and kB
written side-by-side. We now use the fact that rank(T T ∗ ) = rank(T )
for any matrix T to show that A + ikB has the maximal rank:
 ∗ ∗

rank(A + ikB) = rank (A + ikB)(A
 − ikB
 )
= rank (A kB)(A kB)∗ = rank (A kB) = dv .
Having thus shown that A + ikB is invertible, the matrix σ(k) is well-
defined and we can use (1.4.15) to transform it into
σ(k) = −(A + ikB)−1 (A − ikB)
= −(A + ikB)−1 (A − ikB)(A∗ + ikB ∗ )(A∗ + ikB ∗ )−1
= −(A + ikB)−1 (A + ikB)(A∗ − ikB ∗ )(A∗ + ikB ∗ )−1
= −(A∗ − ikB ∗ )(A∗ + ikB ∗ )−1 .
Unitarity of σ(k) can now be verified directly. 

Proof of Theorem 1.4.4. In what follows, we will work with a

single vertex at a time and thus will drop all the v-subscripts in Av ,
Bv , etc. We will establish the implications SA ⇒ A ⇒ B ⇒ C ⇒ SA,
where “SA” stands for self-adjointness of the operator H.
Self-adjointness ⇒ A: Let 2 (Γ) satis-
us choose a function f ∈ H
∞
fying (1.4.9) and a function g ∈ e∈E C (e) that is non-zero in a small
vicinity of the vertex v only (and so vanishes near all other vertices).
We will use our standard notations F, F  and G, G for the vectors of
their values and derivatives at v. Integrating by parts twice in the
expression
d2 fe (x)
ge (x)dx,
e∈E
dx2
e
20 1. OPERATORS ON GRAPHS

one sees that in order for the non-integral term to cancel, the following
equality must be satisfied:
(1.4.16) fe (v)ge (v) − fe (v)ge (v) = F  , G − F, G  = 0,
e∈E e∈E

where ,  denotes the standard hermitian inner product in Cdv . Thus,

in order for g to be in the domain of the adjoint operator H∗ , the
expression given by (1.4.16) (the boundary form) should vanish for
any f as above. Thus, self-adjointness of H means the following: if f
satisfies (1.4.12) then condition (1.4.16) is equivalent to g also satisfying
the conditions (1.4.12):
(1.4.17) AG + BG = 0.
Notice, first of all, that this implies the maximality of the rank of
(A B). Indeed, the consideration in the previous paragraph shows that
the number of conditions on the vector (G G ) for any function g in the
domain of H∗ is equal to 2d minus the number of conditions imposed on
(F F  ) for functions f in the domain of H. Since self-adjointness means
in particular that these numbers are equal, we need to have exactly d
conditions imposed on f , which means that the rank of (A B) is d, and
thus maximal.
Since the values of functions ge and their derivatives at v are inde-
pendent, one can produce any pair of vectors G, G varying the choice
of the function g. In particular, choosing any vector h ∈ Cdv , one can
produce a function gh such that G = −B ∗ h and G = A∗ h. Then
(1.4.16) is automatically satisfied, due to (1.4.9). Self-adjointness im-
plies then that such vectors (G G ) should themselves satisfy (1.4.9).
This gives (−AB ∗ + BA∗ )h = 0 for any h, and thus AB ∗ = BA∗ , which
is condition (1.4.11).
A ⇒ B: Let us multiply equation (1.4.12) on the left by the matrix
−2i(A + ikB)−1 , which exists according to Lemma 1.4.7. We obtain
1 
−2i(A + ikB)−1 A = −2i(A + ikB)−1 (A − ikB) + (A + ikB)
  2
= i σ(k) − I ,
where the matrix σ(k) = −(A + ikB)−1 (A − ikB) is unitary for all real
k = 0, according to Lemma 1.4.7. Similarly,
1  
−2i(A + ikB)−1 B = −2i(A + ikB)−1 (A + ikB) − (A − ikB)
2ik
1 
= − I + σ(k) .
k
 1.4. DIFFERENTIAL OPERATORS. QUANTUM GRAPHS 21

Thus, setting k = −1 we obtain a condition of the type (1.4.13) with

U = σ(−1) = σ(1)−1 .
B ⇒ C: Let PD be the orthogonal projector onto the eigenspace of
the matrix U corresponding to the eigenvalue −1, PN be the orthogonal
projector onto the eigenspace of the eigenvalue +1. Then PR = I−PD −
PN is the projector onto the eigenspace of all eigenvalues other than
±1. Note that all three projectors commute with U and, therefore,
with U − I and U + I.
Multiplying equation (1.4.13) by PD and commuting, we get
(1.4.18) i(U − I)PD F (v) + (U + I)PD F  (v) = 0.
From the definition of PD we have (U + I)PD = 0 and (U − I)PD =
−2PD , reducing (1.4.18) to
−2iPD F (v) = 0.
Similarly we obtain PN F (v) = 0.
Let (U + I)R be the restriction of (U + I) to the eigenspace PR Cd ;
it is obviously invertible. Multiplying equation (1.4.13) by PR ,
i(U − I)PR F (v) + (U + I)PR F  (v) = 0,
and setting Λ = −i(U +I)−1 
R (U −I) we deduce PR,v F (v) = Λv PR,v F (v).
The matrix Λ is invertible on PR Cd and is self-adjoint as the Cayley
transform of a unitary matrix.
C ⇒ self-adjointness: As discussed in the first part of the proof,
to prove self-adjointness we need to establish the following two impli-
cations: (a) if f and g satisfy (1.4.14) then the boundary form (1.4.16)
vanishes and (b) if f satisfies (1.4.14) and the boundary form (1.4.16)
vanishes then g also satisfies (1.4.14).
For the first implication, we use I = PD + PN + PR and self-
adjointness of the projectors to evaluate F  , G,
F  , G = F  , (PD + PN + PR )G
= F  , PD G + PN F  , G + PR F  , G = ΛPR F, PR G.
A similar calculation yields F, G  = PR F, ΛPR G. Since Λ is self-
adjoint, the boundary form vanishes.
For the second implication we re-use an argument from the first
part of the proof. We let A = PD − ΛPR and B = PN + PR , writing
condition (1.4.14) in the form (1.4.12). Then, for any h, F = −B ∗ h
and F  = A∗ h satisfy condition (1.4.14). Substituting this into the
boundary form we get
0 = A∗ h, G − −B ∗ h, G  = h, AG + BG .
Since h is arbitrary, we infer AG + BG = 0. 
22 1. OPERATORS ON GRAPHS

Remark 1.4.8. To go directly from A to C, we note that PD is the

projector on the kernel of I + σ(−1) = −2i(A − iB)−1 B. This obviously
coincides with the kernel of B. A similar relationship connects PN and
A. By looking at the components of the formula Λ = −i(U +I)−1 R (U −I),
we can also conclude that B −1 is well-defined on the range of APR and
maps it back to PR Cd . To summarize,
(1.4.19) Ran PD = Ker B, Ran PN = Ker A, Λ = B −1 APR ,
and, of course, PR = I − PD − PN .
1.4.2. Scale invariance. It is sometimes useful to know whether
the vertex conditions are scale invariant.
Definition 1.4.9. Vertex conditions are said to be scale invari-
ant, if for any function f (x) defined in a neighborhood of a vertex v
and satisfying the conditions at v and for any r > 0 the function f (rx)
also satisfies the vertex condition at v (the re-scaling x → rx is done
along all edges adjacent to v simultaneously).
The above rescaling leaves the F (v) unchanged and multiplies F  (v)
by r. This observation results in the following criterion.
Theorem 1.4.10. Vertex conditions written in the form (1.4.14)
are scale invariant if and only if they do not contain any Robin part,
i.e. PR,v = 0.
As it will be seen in the next several sections, absence of Robin con-
ditions is crucial not only for scale invariance, but also for factorization
of the Hamiltonian and several other interesting properties, such as for
instance energy independence of vertex scattering matrices.

1.4.3. Quadratic form. We will describe now the quadratic form

of the operator H corresponding to any self-adjoint realization of the
(negative) second derivative operator along each edge. For this task,
the self-adjoint vertex conditions written in the form (C) of Theorem
1.4.4 are the most convenient.
Theorem 1.4.11. The quadratic form h of H is given as
2
df
(1.4.20) h[f, f ] = dx + Λv PR,v F, PR,v F  ,
e∈E e
dx v∈V

where ,  denotes the standard hermitian inner product in Cdim PR,v .

The domain of this form consists of all functions f that belong to H 1 (e)
on each edge e and satisfy at each vertex v the condition PD,v F = 0.
 1.4. DIFFERENTIAL OPERATORS. QUANTUM GRAPHS 23

Correspondingly, the sesqui-linear form of H is

df dg
(1.4.21) h[f, g] = dx + Λv PR,v F, PR,v G .
e∈E e
dx dx v∈V

Proof. Lemma 1.3.8 implies that (1.4.20) with the domain de-
scribed in the theorem defines a closed quadratic form (see Appendix
C). Indeed, the norm

M f 2L2 (Γ) + h[f, f ]
with a sufficiently large M on the domain of h is equivalent to the
norm of the space H = H 1 (Γ). Therefore, the form h corresponds to a
self-adjoint operator M in L2 (Γ). We will now show that M = H.
According to the definition of operator M, for any f ∈ D(M) ⊂ H
there exists p ∈ L2 (Γ) (which is then denoted by Mf ) such that for
any g ∈ D(h) one has

(1.4.22) h[f, g] − p(x)g(x)dx = 0.

e e

Let now g be any function smooth on each edge and equal to zero in
a neighborhood of each vertex. Then clearly g ∈ D(h). Choosing only
such functions in the previous equality, substituting the definition of h
for the left hand side, and integrating by parts, one concludes that
d2 f
(1.4.23) p(x) = Mf (x) = −
dx2
on each edge, where the derivatives are meant in the distributional
2
2 (Γ).
sense. Thus it is required that ddxf2 ∈ L2 (Γ), and therefore f ∈ H
Since f ∈ D(h), the conditions PD,v F (v) = 0 is automatically en-
forced. We need to show that f satisfies also the remaining vertex
conditions in (1.4.14). One does this using a test function g ∈ D(h)
that is non-zero in small neighborhood of a single vertex v only.
Integration by parts of h in equation (1.4.22) transforms it into
(1.4.24) − F  , G + Λv PR,v F, G = 0
v∈V v∈V

Taking into account that then G(v) can be an arbitrary vector such
that PD,v G(v) = 0, this means that for each v the equality
(1.4.25) PN,v F  (v) + PR,v F  (v) − Λv PR,v F (v) = 0
needs to be satisfied. This implies conditions (1.4.14) for the function
f.
24 1. OPERATORS ON GRAPHS

It is also easy to check in a similar fashion that as soon as a function

f belongs to H 2 on each edge and satisfies (1.4.14), it belongs to the
domain of M. This proves that M in fact coincides with the previously
described operator H. 
1.4.4. Examples of vertex conditions. In this section we list
some examples of commonly used vertex conditions. We start with the
δ-type condition at vertex v is defined as follows:

f (x) is continuous at v,
(1.4.26)  df
e∈Ev dxe (v) = αv f (v),

where αv is a fixed number. One can recognize this condition as being

an analog of the conditions one obtains for the Schrödinger operator on
the line with a δ potential (e.g., [20]), which explains the name. When
αv = 0, one arrives to the previously considered Neumann-Kirchhoff
condition.
The δ-type condition can be written in the form (1.4.12) with
⎛ ⎞
1 −1 0 . . . 0 0
⎜ 0 1 −1 . . . 0 0⎟
⎜ ⎟
Av = ⎜
⎜ . . . . . . . . . . . . . . . . . . ⎟
⎟
⎝ 0 0 0 . . . 1 −1⎠
−αv 0 0 ... 0 0
and ⎛ ⎞
0 0 ... 0
⎜. . . . . . . . . . . .⎟
Bv = ⎜
⎝0 0
⎟.
... 0 ⎠
1 1 ... 1
Since ⎛ ⎞
0 ... 0 0
⎜... ... ... ...⎟
Av Bv∗ = ⎜
⎝ 0
⎟,
... 0 0 ⎠
0 ... 0 −α
the self-adjointness condition (1.4.11) is satisfied if and only if αv is
real.
In order to write the vertex conditions in the form (1.4.14), one
introduces the orthogonal projection PD,v onto the kernel of Bv . If
α = 0 then PN,v = I − PD,v and there is no Robin part. For α = 0,
on the contrary, there is no Neumann part and I − PD,v plays the role
of the projector onto the Robin part, PR,v . A straightforward calcu-
lation shows that I − PD,v is the one-dimensional orthogonal projec-
tor onto the space of vectors with equal coordinates. More explicitly,
 1.4. DIFFERENTIAL OPERATORS. QUANTUM GRAPHS 25

I − PD,v = d−1
v E, where E is the dv × dv matrix of ones. The self-adjoint
operator Λv = Bv−1 Av on the range of PR,v is simply the multiplication
αv
by the number . In particular, the quadratic form of the operator
dv
H (assuming δ-type conditions on all vertices of the graph) is
2
df
h[f, f ] = dx + Λv F, F 
e∈E e dx v∈V
2
df
(1.4.27) = dx + αv |f (v)|2 ,
e∈E e dx v∈V

defined on f ∈ H 1 (Γ), which are automatically continuous, and so

F (v) = (f (v), . . . , f (v))t . Note that equation (1.4.27) is valid for the
αv = 0 case, when h[f, f ] reduces to its “natural” form.
The vertex Dirichlet condition requires that the function van-
ishes at the vertex: f (v) = 0. At the first glance, it might look like
it is significantly different from the δ-type conditions, but a closer in-
spection shows that this is not the case. Indeed, since the function
must vanish when approaching the vertex from any edge, the vertex
Dirichlet condition can be recast in the following form:

f (x) is continuous at v,
(1.4.28)
f (v) = 0,
Now one finds resemblance with (1.4.26), and indeed, if one divides the
equality in (1.4.26) by αv and then takes the limit when αv → ∞, one
arrives to (1.4.28).
Hence, the vertex Dirichlet condition seems to be the limit case
of (1.4.26) when αv → ∞. We thus introduce the extended δ-type
conditions by allowing αv = ∞. In order to avoid considering infinite
values of αv , the two types of conditions can be also written in the form
df
(1.4.29) cos(γv ) (v) = sin(γv )f (v).
e∈E
dxe
v

Here γv = 0 corresponds to the Neumann condition and γv = π/2

corresponds to the Dirichlet one, with more general δ-type conditions
in between.
The importance of considering the Dirichlet condition as a part of
the family of δ-type conditions becomes clear in spectral theory, as will
be illustrated, for instance, in Theorem 3.1.8.
Interpreting the Dirichlet condition in terms of the corresponding
projectors, as in part C of Theorem 1.4.4, one notices that here PD,v = I
and, correspondingly, PR,v = 0. Hence there is no additive contribution
26 1. OPERATORS ON GRAPHS

to the quadratic form h[f, f ] coming from the vertex v. Instead, the
condition f (v) = 0 is introduced directly into the domain D(h).
To summarize, the quadratic form for a graph with the extended
δ-type conditions (i.e., allowing αv = ∞) at all vertices, can be written
as
2
df
h[f, f ] = dx + αv |f (v)|2 ,
e∈E e
dx
{v∈V | αv <∞}

where
f ∈ H 1 (Γ) (in particular, it is continuous on Γ),
and
f (v) = 0 whenever αv = ∞.
The vertex Dirichlet condition presents also an example of a de-
coupling condition (see discussion below), since it essentially removes
any connection between the edges attached to the vertex.
The δ  -type condition is similar to the δ-type one, with the roles
of the vertex values of functions and their derivatives switched:
⎧ df
⎨ the values dx (v) are independent of e at the vertex v,

e
(1.4.30) df
⎩ fe (v) = αv dx e
(v).
e∈Ev

When dv = 2, these conditions correspond to the symmetrized version

of what is usually called δ  conditions at a point. The true counterpart
of δ  conditions is provided in [265].
In comparison with the δ-type case, the matrices Av and Bv get
switched and the self-adjointness still holds for real αv only.
In order to rewrite the condition in the form (1.4.14), we look first
at the case when αv = 0 for some vertex v. Then the kernel of Bv
is one-dimensional and consists of all vectors with equal coordinates.
The projector PD,v projects orthogonally onto this subspace, projector
PN,v = I − PD,v and projector PR,v = 0. Thus the vertex v gives no
contribution to the quadratic form. Instead, the domain is restricted
to the functions from H 1 (Γ) that have the sum of the vertex values at
v along all entering edges equal to zero. In particular, if αv = 0 for all
vertices, we get the quadratic form
2
df
h[f, f ] = dx.
e e∈E
dx

Let us now look at the case of a non-zero value of αv . In this case

the operator Bv is invertible and so PD,v = 0 and PR,v = I. It is not
 1.4. DIFFERENTIAL OPERATORS. QUANTUM GRAPHS 27

hard to compute that (Λv )ij = α(−1) for all indices i, j. This leads to
the non-integral term
2
1
fe (v) .
αv
{e∈Ev }

One can think of the case αv = 0 as a “particular case” of this one,

if one assumes that the denominator being equal to zero forces the
condition that the sum in the numerator also vanishes.
To summarize, the quadratic form for a graph with the δ  -type
conditions on all vertices and a general choice of real coefficients αv
can be written as follows:
2
2
df 1
h[f, f ] = dx + fe (v) ,
e dx αv
e∈E {v∈V | αv =0} {e∈Ev }

where
 1 (Γ) and
f ∈H fe (v) = 0 whenever αv = 0.
e∈Ev

As in the case of δ-type conditions, it is often convenient to re-write

the δ  -type ones as follows:
⎧ df
⎨ the values dx (v) are independent of e at the vertex v,

e
(1.4.31) df
⎩ cos γv fe (v) = sin(γv ) dx e
(v).
e∈Ev

Decoupling conditions. Some vertex conditions essentially dis-

connect the edges entering the vertex. One example is the vertex
Dirichlet condition, considered above. If the Dirichlet condition is im-
posed at all vertices of a graph, the operator is the direct sum of the
operators on each edge e with Dirichlet conditions at the ends. In this
cases, the topology of the graph becomes irrelevant and the resulting
spectrum is the union of the spectra of individual edges.
The situation is similar if one enforces the vertex Neumann con-
ditions (not to be mixed up with the Neumann-Kirchhoff conditions),
requesting that derivatives along each incoming edge vanish at the ver-
tex v: df /dxe (v) = 0. This corresponds to γv = π/2 in (1.4.31). One
then has PN,v = I. Other examples of decoupling conditions can be
easily constructed.
1.4.5. Infinite graphs. We will now allow the number of vertices
and edges of a metric graph Γ to be (countably) infinite (the reader
should remember that each vertex degree is assumed to be finite). Our
goal is to define a self-adjoint operator H on Γ in a manner similar
28 1. OPERATORS ON GRAPHS

to the one used for finite graphs. In other words, H should act as
the (negative) second derivative along each edge, and the functions
from its domain should satisfy (now infinitely many) vertex conditions
(1.4.9) or equivalently (1.4.14). This would turn a metric graph Γ
into a quantum graph. However, unless additional restrictions on the
graph and vertex conditions are imposed, the situation can become
more complex than in the finite graph case. This is true even for such
“simple” graphs as trees (see Appendix A), where additional boundary
conditions at infinity may or may not be needed depending on geometry
(see [170, 563, 564, 673, 674]). In other words, the operator defined
on D(Γ) is not essentially self-adjoint. On the other hand, if one looks
at the naturally arising infinite graphs, one can notice that in many
cases there is an automorphism group acting on the graph such that the
orbit space (which is a graph by itself) is compact. This is the case for
instance with periodic graphs (see Chapter 4) and Cayley graphs of
finitely generated groups and Schreier graphs of their homogeneous
spaces (see Appendix A). We do not need exactly the homogeneity, but
rather that the geometry does not change drastically throughout the
graph. The assumptions that we introduce below capture this idea and
cover all cases mentioned above. It also enables one to establish nice
properties of the corresponding Hamiltonians. This class of graphs is
in some sense an analog of the so called manifolds of bounded geometry
[654, 658]. Studying differential operators on such manifolds is easier
than on more general ones.

Assumption 1.4.12. The lengths of all edges are uniformly bounded

from below:

(1.4.32) 0 < l0 ≤ le ≤ ∞.

Remark 1.4.13.
(1) This assumption is always satisfied for finite graphs.
(2) There is no upper bound assumed on the lengths of edges. In
fact, some edges may be of infinite length, as they often are,
e.g. in scattering problems.

Let us now equip a metric graph Γ with the negative second deriv-
ative operator and boundary conditions (1.4.14), which will make it a
quantum graph. The precise definition of the operator will be provided
a little bit later in this section. In order to do so, another “bounded
geometry” type assumption will be needed.
 1.4. DIFFERENTIAL OPERATORS. QUANTUM GRAPHS 29

Assumption 1.4.14. The following estimate holds uniformly for all

vertices v:
(1.4.33) Λv ≤ S < ∞,
where Λv is the operator entering vertex conditions (1.4.14). The norm
in (1.4.33) is the operator norm with respect to the standard l2 norm
induced from the space Cdv .
Remark 1.4.15. If the vertex conditions are given in the form
(1.4.9), then the assumption above should be replaced by the following:
(1.4.34) Bv(−1) Av Qv ≤ S < ∞,
where Qv denotes the orthogonal projection onto the range of Bv∗ and
is the inverse to the operator Bv acting from the range of Bv∗ to
(−1)
Bv
the range of Bv .
Let us now define the operator H more precisely.
Definition 1.4.16. The Hamiltonian H in L2 (Γ) acts as the neg-
ative second derivative along the edges, defined on the domain D(H)
consisting of functions f such that:
(1) f ∈ H2 (Γ) (i.e., f ∈ H 2 (e) for each edge e),
(2)
f 2H2 (e) < ∞,
e
(3) for each vertex v, conditions (1.4.14) are satisfied:
⎧
⎪
⎨PD,v F (v) = 0 “Dirichlet part”,
(1.4.35) PN,v F  (v) = 0 “Neumann part”,
⎪
⎩P F  (v) = Λ P F (v) “Robin part”.
R,v v R,v

It will be shown below that this operator is self-adjoint and in

fact is the only “reasonable” self-adjoint realization of our Hamiltonian
(i.e., its restriction to the appropriate subspace of compactly supported
functions is essentially self-adjoint). We, however, want to describe the
corresponding quadratic form first. The definition will be similar to
the one we gave in the case of finite graphs:
Definition 1.4.17. The quadratic form h is defined as
2
df
(1.4.36) h[f, f ] = dx + Λv PR,v F, PR,v F ,
e∈E e
dx v∈V

where ,  denotes the standard hermitian inner product in Cdv .

30 1. OPERATORS ON GRAPHS

The domain of this form consists of all functions f that belong to

1
H (e) for each edge e, satisfy at each vertex v the condition PD,v F = 0,
and such that
(1.4.37) f 2
H 1 (e) < ∞.
e

One can write the sesqui-linear form for h:

df dg
(1.4.38) h[f, g] = dx + Λv PR,v F, PR,v G.
e∈E e
dx dx v∈V

Some remarks are due concerning this definition.

Remark 1.4.18.
(1) Analogous formulas can be easily written in terms of the ma-
trices Av , Bv of conditions (1.4.9).
(2) Due to Lemma 1.3.8 and the Assumption 1.4.12 on the lower
bound on the lengths of the edges, the norms of the trace
operators that associate to each function f ∈ H 1 (e) for e ∈
Ev its value at the vertex v ∈ e are bounded uniformly with
respect to v and e:
(1.4.39) |f (v)| ≤ C f 1
H (e).

(3) In order for the definition to be consistent, one needs to make

sure that both infinite sums in the formula for h converge. The
first one (the sum of the integrals) converges due to Cauchy-
Schwarz inequality and (1.4.37). Moreover, equations (1.4.37)
and (1.4.39) imply that for any f from the domain of the form
one has for its traces F (v) (recall that F (v) is a vector in Cdv )
the inequality

(1.4.40) F (v) 2
≤C f 2
H 1 (e) < ∞.
v e

Since the same applies to the function g, this, (1.4.33), and the
Cauchy-Schwarz inequality secure convergence of the second
sum.
We are now prepared for the discussion of the Hamiltonian.
Theorem 1.4.19. Let Γ be a quantum graph satisfying Assumptions
1.4.12 and 1.4.14. Under the definitions given above for the quadratic
form h and operator H, the following statements hold:
(1) The operator H is self-adjoint, bounded from below, and its
quadratic form is h.
 1.4. DIFFERENTIAL OPERATORS. QUANTUM GRAPHS 31

(2) Let H0 be the restriction of H onto the sub-domain consisting

of all functions from D(H) with compact support. Then H0 is
symmetric, essentially self-adjoint, and its closure is H.
Before we proceed to proving the theorem, let us mention that its
second statement implies that under the prescribed vertex conditions
there is essentially only one way to define our self-adjoint Hamiltonian
H, and it is the one we choose. One can also notice that according
to [170, 674] this is not true anymore for graphs that do not have
“bounded geometry” in terms of the Assumptions 1.4.12 and 1.4.14,
even for trees, where some boundary conditions at infinity might be
needed.
Proof. The proof of the first statement of the theorem is almost
identical the proof of Theorem 1.4.11. Below we give an outline and
point out the differences.
The closedness of the form h is implied by estimate (1.4.40). Hence,
the form h corresponds to a self-adjoint operator M. Integration by
parts against test functions g that are compactly supported and equal to
zero near vertices shows that M acts as the negative second derivative.
2
Thus it is required that ddxf2 ∈ L2 (Γ). Since f ∈ D(h), we conclude that
f ∈H 2 (Γ) and  f H 2 (e) < ∞. Proving that f satisfies the vertex
e
conditions carries over without any changes. Hence, M = H and the
first statement of the theorem is proven.
For the second statement, it is immediate to check that the operator
H0 is symmetric. We will show that the graph of H0 is dense in the
graph of H. Since H has already been shown to be self-adjoint (and
therefore closed), this will prove that the closure of H0 is H.
Let now f ∈ D(H). Our goal is to create a sequence of cut-off
functions φn (x) such that fn = φn f ∈ D(H0 ) and
(1.4.41) f − fn L2 (Γ) → 0, Hf − H0 fn L2 (Γ) → 0.
The idea of how functions φn should behave is clear: they must be equal
to 1 on an expanding and exhausting sequence Γn ⊂ Γ of compacta,
must have compact supports, must be constant in a neighborhood of
each vertex (in order not to destroy the vertex conditions), and must
fall off to zero “not too fast,” so that their first and second order deriva-
tives are uniformly bounded. Now, if a sequence of such functions is
constructed, then (1.4.41) is straightforward. The graph nature of our
variety causes some superficial complications, so we will describe this
construction (which could certainly be done in many different ways).
Let us fix a vertex o ∈ Γ and consider for any natural n the set Γn ⊂ Γ
that contains all (finite) edges e whose both endpoints are at a distance
32 1. OPERATORS ON GRAPHS

at most n from o and all points x of infinite edges such that ρ(x, o) ≤ n
(here ρ is the previously defined metric on Γ). This is clearly an ex-
panding sequence of compact sets that exhausts Γ. Let φ(x) be any
smooth function on [0, l0 /4] such that it is identically equal to 1 in a
neighborhood of 0 and identically equal to zero close to l0 /4. Here l0 is
the lower bound for the lengths of all edges of Γ, which is positive due
to the Assumption 1.4.12. We are ready to define the cut-off function
φn on Γ. It is equal to 1 on Γn and to 0 on all edges which do not
have vertices in Γn . We only need to define it along the edges that
have only one vertex in Γn . Let e be a finite edge whose one vertex v
is contained in Γn . The function φn is defined to be equal to 1 along e
starting from v till the middle of the edge, then it is continued by an
appropriately shifted copy of φ(x) (which by construction will become
zero at least at the distance le /4 from the end of the edge), and stays
zero after that. If e is an infinite edge starting at v ∈ Γn , then φn
is defined to be equal to 1 along e starting from v till the distance n
from o, then it is continued by an appropriately shifted copy of φ(x),
and stays zero after that. It is clear that all our requirements for the
sequence of functions are satisfied.
Let now f ∈ D(H) and fn = φn f . Then fn is in H 2 (e) for any
edge e and satisfies the boundary conditions. The reason for the latter
is that φn is constant around each vertex, and so multiplication by it
does not influence the vertex conditions. In other words, fn ∈ D(H0 ).
One gets the following simple conclusion:
lim f − fn L2 (Γ) = lim (1 − φn )f L2 (Γ)
n→∞ n→∞
≤ C lim f L2 (Γ−Γn ) = 0.
n→∞

Here C is the maximal value of |φ(x) − 1|. We also have

Hf − H0 fn = (1 − φn )f  − φn f  − φn f.
Since f , f  , and f  all belong to L2 (Γ) and the functions 1−φn , φn , and
φn are uniformly bounded and supported outside Γn , we also obtain the
second required limit
lim Hf − H0 fn L2 (Γ) = 0.
n→∞

This finishes the proof of the theorem. 

1.4.6. Non-local vertex conditions. The discussion in Section
1.4.4 of decoupling conditions leads to the understanding that the whole
topology of the quantum graph is in fact encoded in the vertex con-
ditions only. In particular, one can identify all vertices of a quantum
graph into one “super-vertex” v0 , so the graph becomes just a “rose”(or
 1.5. FURTHER REMARKS AND REFERENCES 33

“bouquet”) of several petals (each edge bends into a loop, see Fig. 7),
and all vertex conditions are written at this single vertex. If one starts

Figure 7. Graph Γ turned into a single “rose.”

with some special type of vertex conditions, for instance Neumann-

Kirchhoff ones, this procedure might not preserve the type. On the
other hand, if one needs to work with most general vertex conditions,
then one can as well assume that the graph is just a bouquet of sev-
eral loops at a single vertex. This also shows that non-local conditions
can be turned into local ones at this single “super-vertex”, and thus de-
scription of self-adjoint local conditions given in Theorem 1.4.4 extends
to non-local conditions as well.

1.5. Further remarks and references

Graphs
The literature on graph theory and its applications is vast and there
is no chance (or need) for us to survey it here. What is needed for this
text, can be found, for instance, in [90, 115, 358], see also Appendix
A.
Spectral graph theory
Spectral graph theory has been developed extensively, with appli-
cations both inside and outside graph theory itself. It is interesting to
observe how the discrete and analytic (e.g., PDE, geometric analysis)
techniques merge here. The books [191, 196, 215], devoted solely to
the spectra of graphs and their application, provide the reader with a
broad overview of this fascinating area.
Discrete geometric analysis
Discrete geometric analysis (DGA) is also an active area of research,
with applications ranging from number theory to such applied topics
as cryo-electron tomography [390, 391, 664] or visualization. One can
find a nice survey of DGA in Sunada’s paper [682].
Graph approximations have also been used to study various geo-
metric analysis problems on multi-dimensional domains and manifolds
(e.g., [194, 197, 199]), in particular for studying domains with cusp
singularities [257, 258]).
34 1. OPERATORS ON GRAPHS

Metric and quantum graphs

Metric graphs (also known in algebra as R-graphs) have been in use
in chemistry and physics for a long time (e.g., [21, 226, 522, 558, 594,
637]). They usually carried a differential operator over their edges,
with some junction conditions at the vertices. When much later such
graphs started being used as simplified models for studying quantum
chaos and quantum wires (see, e.g. [283,459,467]), the term “quantum
graph” started to appear, and by now seems to be widely accepted.
Vertex conditions
The general self-adjoint vertex conditions were described in [459]
and then went through several different versions [328,393], all of which
are collected in Theorem 1.4.4. Recently, it was discovered in [184] that
in some issues still another interpretation of vertex conditions is useful.
The reader is referred to [184] for details.
Von Neumann theory of extensions of symmetric operators, used
in [393] to describe vertex conditions, can be found in [16], or in its
more recent incarnation in terms of the complex symplectic geometry
(e.g. [259–263, 393, 572, 599, 618]), an approach that goes back at
least as far as [565]. See also Appendix D.
One should also mention an alternative approach to the vertex con-
ditions, where the corresponding singularity is included into the defini-
tion of the operator itself, see e.g., [43,44,602] and [614, last Chapter]
See also a brief discussion of further topics, including the non-self-
adjoint vertex conditions, in the last section of Chapter 2.
Quadratic forms
A brief discussion of closed quadratic forms and their relations to
self-adjoint operators can be found in Appendix C, while the detailed
theory is contained in [436, Ch. VI].
Infinite graphs
Infinite combinatorial and quantum graphs arise naturally in many
applications inside mathematics, as well as in physics and chemistry.
The most popular types of such graphs are regular trees (also called
Bethe lattices), which are nice to study, due to their self-similar
structure. They have been used, for instance, for obtaining new results
on Anderson localization [10–12, 14, 15, 446–448]. Analysis on (com-
binatorial, rather than quantum) trees has been developing for some
time [297, 298]. It was understood that analysis on regular trees dis-
plays strong parallels with analysis on the hyperbolic plane. This line of
research, not intersecting with quantum graph theory so far, has been
flourishing recently, leading to advances at the junction of several areas:
infinite discrete groups, automata theory, amenability theory, complex
dynamics, and analysis on fractals (see, e.g., [372, 374, 566, 567]). It
 1.5. FURTHER REMARKS AND REFERENCES 35

also works with more general objects, such as Cayley graphs of finitely
generated groups (in the particular case of a free group one obtains a
regular tree) and Schreier graphs (analogs of Cayley graphs for homo-
geneous spaces of discrete groups) (see Appendix A).
Another class of infinite graphs, popular in physics and chemistry
(solid state physics, photonic crystal theory, carbon nano-structures,
etc.), consists of periodic graphs. It is discussed in detail in Chapter 4.
Another extremely interesting class, which we cannot address in
any detail, consists of fractals (Sierpinski gasket, Koch snowflake, etc.).
One is referred to the books [442, 566, 677] and papers [59, 60, 256,
567] for details and further references. One can notice that this an-
alytic development could be interesting for physics and material sci-
ence, since fractal composites are becoming very popular materials
(e.g., [651]).
Discussion of various other problems concerning infinite quantum
graphs were discussed in [168, 268, 481, 491, 588, 603, 673, 674, 732].
 CHAPTER 2

Quantum Graph Operators. Special Topics

Considering the graph as a system of 1D homogeneous “wires”

with junctions at vertices and looking at waves of a certain frequency
propagating along the wires, one arrives at the notion of scattering
at a vertex, which is discussed in Section 2.1. From the scattering
point of view, second order operators are too restrictive, and first order
operators offer much more flexibility, as shown in Section 2.2. Then
the natural question of the possibility of factorization of a second order
Hamiltonian (analogous to the formula Δ = ∇ · ∇ in the continuous
case) is resolved in Section 2.3. A formula for the Fredholm index of
any (elliptic) differential operator on a graph is obtained in Section
2.4. Dependence of the operator on the vertex conditions and edge
lengths is discussed in Section 2.5 and in Section 2.6 it is shown how
the magnetic Schrödinger operator can be reduced to a non-magnetic
one but with changed vertex conditions. The chapter ends with the
section devoted to further remarks and references.

2.1. Quantum graphs and scattering matrices

There are several important ways the word “scattering” can be and
has been applied to quantum graphs. We describe scattering on vertices
and the “bond scattering matrix” in this section. Both of these notions
describe the way the waves scatter within the quantum graph. Scat-
tering of “external” waves on a graph — the more traditional meaning
of the word “scattering” — is reviewed in Section 5.4.

2.1.1. Scattering on vertices. We introduce here the matrix σ (v)

that describes the scattering at the vertex v of the plane wave e−ikx
coming towards the vertex along one of the edges. More precisely,
consider the scattering problem for the equation

d2 f
(2.1.1) − = k 2 f (x)
dx2
37
38 2. GRAPH OPERATORS. SPECIAL TOPICS

at the vertex v equipped with conditions of the form (1.4.12) (or any
equivalent ones). Since all solutions to equation (2.1.1) can be repre-
sented as superpositions of e−ikx and eikx , we write
⎧
⎪ (e)
⎨f (xe ) = e
−ikx (v)
+ σe,e eikx on e ,
f (e) (xe ) = σe ,e eikx on e = e ,
(v)
(2.1.2)
⎪
⎩
f satisfies the vertex conditions at v,
where x coordinates on each edge are measured from the vertex v.
(v)
We can interpret σe,e as the reflection coefficient along the edge e,
while σe ,e for e = e is the scattering coefficient from the edge e to e .
(v)

(v)
Note that the coefficients σe e normally depend on k.
Definition 2.1.1. The dv × dv matrix σ (v) (k) with the entries
(v)
σe ,e (k) is the scattering matrix at the vertex v.
Remark 2.1.2.
(1) We will see soon that the matrix σ (v) (k) is unitary for any real
k.
(2) It is also quite natural to consider the incoming wave e−ikx as
living on the incoming bond b (corresponding to the edge e),
whereas all other components live on the outgoing bonds. If
(v)
this view is adopted, the matrix elements σb1 ,b2 are indexed by
the incoming bonds b2 and outgoing bonds b1 . Thus σ (v) acts
from a space indexed by incoming bonds to the space indexed
by the outgoing bonds at the vertex v.
The scattering matrix can be directly computed from the vertex
conditions at the vertex v. The corresponding formulas and the uni-
tarity of σ (v) (k) are established in the following lemma.
Lemma 2.1.3.
(1) If the vertex conditions are realized as (1.4.12), then
(2.1.3) σ (v) (k) = −(Av + ikBv )−1 (Av − ikBv ),
(2) If the vertex conditions are realized as (1.4.13), then
 −1  
(2.1.4) σ (v) (k) = − Uv − I + k(Uv + I) Uv − I − k(Uv + I)
 −1  
(2.1.5) = − (1 + k)Uv − (1 − k) (1 − k)Uv − (1 + k) .
(3) If the vertex conditions are realized as (1.4.14), then
(2.1.6) σ (v) (k) = −PD,v + PN,v − (Λv − ik)−1 (Λv + ik)PR,v .
(4) The matrix σ (v) (k) is unitary.
 2.1. QUANTUM GRAPHS AND SCATTERING MATRICES 39

Remark 2.1.4. As it was seen in the proof of Theorem 1.4.4, Uv =

(v)
σ (−1). This can also be verified directly from (2.1.4).
Proof. We will start by deriving equation (2.1.3) for the scattering
matrix. Since the vertex v will be fixed, the sub- and super-scripts v
are omitted. For a fixed edge e the solution of (2.1.2) satisfies
Fe = {δe ,e + σe ,e } , Fe = ik {−δe ,e + σe ,e } ,
where Fe is the vector of values of the solution at the vertex v, Fe is
the vector of derivatives, and δe ,e is the Kronecker delta. Combining
the column-vectors Fe and Fe for all incident edges e into two matrices
F and F  , we obtain
F = I + σ, F  = ik(−I + σ).
We substitute these matrices into the vertex conditions given by equa-
tion (1.4.12) and obtain
A(I + σ) + ikB(−I + σ) = 0,
and, after elementary manipulations, arrive to (2.1.3).
The unitarity of the matrix σ, defined by (2.1.3), was established in
Lemma 1.4.7. Furthermore, equation (2.1.4) is obtained from (2.1.3)
by the correspondence between the vertex conditions, namely
A = i(U − I), B = U + I.
Finally, if the vertex conditions are given in the projector form, with
A = PD + ΛPR and B = PN − PR , direct computation shows that
1
(A+ikB)−1 = (PD +ikPN +(Λ−ik)PR )−1 = PD −i PN +(Λ−ik)−1 PR .
k
Here we have used the relation PD +PN +PR = I. One can now expand
the product in (2.1.3) to obtain (2.1.6). 
Corollary 2.1.5.
(1) The scattering matrix σ (v) can be written in the following form:
(2.1.7) σ = −PD + PN − PR − 2ik(Λ − ikI)−1 PR
(2.1.8) = −PD + PN + PR − 2(Λ − ikI)−1 ΛPR
(sub- and super-scripts v are omitted).
(2) For k → 0, we have
(2.1.9) σ = −PD + PN − PR + O(k).
(3) For k → ∞,
(2.1.10) σ = −PD + PN + PR + O(k −1 ).
40 2. GRAPH OPERATORS. SPECIAL TOPICS

Proof. We can rewrite the third term in equation (2.1.6) as

−(Λ − ikI)−1 (Λ + ikI)PR = −(Λ − ikI)−1 ((Λ − ikI) + 2ik)PR
= −PR − 2ik(Λ − ikI)−1 PR ,
leading to (2.1.7). On the other hand,
−(Λ − ikI)−1 (Λ + ik)PR = −(Λ − ikI)−1 (2Λ − (Λ − ikI))PR
= −2(Λ − ikI)−1 ΛPR + PR ,
leading to (2.1.8). Asymptotic forms (2.1.9) and (2.1.10) follow imme-
diately from (2.1.7) and (2.1.8). 
As we will see, it is often required that the scattering matrix is
k-independent, and so the following result happens to be handy:
Theorem 2.1.6. The following conditions at any vertex v are equiv-
alent:
(1) σ (v) (k) is independent of k.
(2) There is a value k = 0 such that (σ (v) (k))2 = I.
(3) (σ (v) (k))2 = I for all k.
(4) σ (v) has the form I − 2Qv for some orthogonal projection Qv .
(5) There is no Robin part in the vertex conditions: PR,v = 0.
Proof. We shall show the following implications: (1) ⇒ (5) ⇒ (1)
and (5) ⇒ (4) ⇒ (3) ⇒ (2) ⇒ (5).
To show (1) ⇒ (5) we observe from Corollary 2.1.5 that
lim σ (v) = −PD + PN − PR ,
k→0
while
lim σ (v) = −PD + PN + PR .
k→∞
If σ (v) = const, the two limits must agree, leading to PR = 0.
The implication (5) ⇒ (1) is obvious from (2.1.6).
To prove (5) ⇒ (4), we recall that PD + PN + PR = I. Since PR = 0,
we have from (2.1.6)
σ (v) = −PD,v + PN,v = (PD,v + PN,v ) − 2PD,v = I − 2PD,v .
To show (4) ⇒ (3) we square I − 2Qv obtaining I − 4Qv + 4Q2v = I since
Qv is a projector. The implication (3) ⇒ (2) is obvious.
Finally, to show (2) ⇒ (5) we square (2.1.7) and obtain
 2
0 = σ 2 − I = 4ik(Λ − ikI)−1 PR + 4 ik(Λ − ikI)−1 PR
 
= 4ik(Λ − ikI)−1 I + ik(Λ − ikI)−1 PR
= 4ik(Λ − ikI)−1 Λ(Λ − ikI)−1 PR .
 2.1. QUANTUM GRAPHS AND SCATTERING MATRICES 41

Since the matrices Λ and (Λ − ikI)−1 are invertible, we conclude that

PR = 0. 
Example 2.1.7. Revisiting the example of Neumann-Kirchhoff ver-
tex conditions, we find that in this case Rv = 0 and Qv = I − Pv is a
projector onto the vectors with all equal components (all entries of Qv
are equal to 1/dv ). Then the scattering matrix is
σ (v) = −Pv + Qv = 2Qv − I = I − 2Pv .
In this case, the matrix σ (v) is independent of k.
2.1.2. Bond scattering matrix and the secular equation. A
large part of the literature on quantum graphs is devoted to the study of
the spectral properties of the quantum graph Hamiltonian H (see, e.g.,
Chapters 3-5). The Hamiltonian is usually taken to act as −d2 /dx2 ,
with the domain described by vertex conditions described in detail in
Section 1.4. The goal of this section is to relate the spectrum to the
scattering matrices.
We thus consider the eigenvalue equation Hf = λf , or, after de-
noting the spectral parameter by λ = k 2 ,
d2 f
(2.1.11) − 2
= k 2 f (x),
dx
on the whole graph Γ. We look for the values of k for which one can find
a non-zero solution satisfying the vertex conditions at all vertices of the
graph and belonging to an appropriate functional space. In this section,
we will be dealing with compact quantum graphs, and thus solutions we
are looking for belong to L2 (Γ) (see Chapter 3, in particular Sections
3.2 and 4.9 for the spectral structure of non-compact graphs).
In a slight abuse of notation, the spectrum σ(H) of H is sometimes
called the spectrum of the quantum graph Γ and denoted σ(Γ).
In such a case it is tacitly assumed that the above Hamiltonian is used.
Let e be an edge incident to a vertex v. As long as k = 0, the
general solution of equation (2.1.11) can be written as
(2.1.12) f (xe ) = C1 eikxe + C2 e−ikxe ,
where xe is measured from the vertex v. The edge e corresponds to two
directed bonds, the outgoing bond b and the incoming bond b. From
now on we will adopt the point of view that the incoming plane wave
e−ikxe “lives” on the incoming bond b and the outgoing plane wave eikxe
lives on b. Equation (2.1.11) can be re-written as
(2.1.13) f (xb ) = ab eikxb + ab eikxb = ab eikxb + eikLb ab e−ikxb ,
42 2. GRAPH OPERATORS. SPECIAL TOPICS

where xb is measured from the other end of the edge e (i.e. from the
start of the bond b); it is related to xb by xb = Lb − xb .
An eigenfunction of (2.1.11) with given boundary conditions can be
represented locally around a vertex v as a sum of functions fe (x). The
outgoing amplitudes ab are related to the incoming amplitudes
ab through the scattering matrix σ (v) , which is easy to verify by a direct
computation. Indeed, let us denote by α the vector {ab } and by β the
vector {eikLb ab }. Then the values of the solution and its derivative at
the vertex v are F (v) = α + β and F  (v) = ik(α − β). The vertex
conditions then become
A(α + β) + ikB(α − β) = 0,
or, after short transformation,
α = −(A + ikB)−1 (A − ikB)β = σ (v) (k)β.
We use here Lemma 1.4.7 that guarantees invertibility of A + ikB for
real k = 0.
Noting that an outgoing amplitude at one vertex is an incoming
at another, we combine all vertex conditions into a 2E × 2E system
(where E = |E|, see the notational agreements in Section 1.1)
(2.1.14) γ = S(k)eikL γ,
where γ is the vector of all 2E incoming and outgoing amplitudes ab
and the elements of the 2E × 2E matrix S are given by
(v)
(2.1.15) (S)b b = δt(b),v δo(b ),v σb b .
and L is the diagonal matrix
⎛ ⎞
L1
⎜ ... ⎟
⎜ ⎟
⎜ ⎟
⎜ LE ⎟
(2.1.16) L := ⎜ ⎟.
⎜ L1 ⎟
⎜ ... ⎟
⎝ ⎠
LE
The matrix S(k) is called the bond scattering matrix. While
it is not coming from a scattering problem in the traditional sense,
it does describe the scattering of waves at the vertices of the graph,
thus justifying the name. Equation (2.1.14) signifies that the scattering
waves have to be in a steady state to form an eigenfunction of (2.1.11).
Note that the dimension of all matrices is double the total number of
edges, which reflects the transition from undirected edges to pairs of
directed bonds.
 2.1. QUANTUM GRAPHS AND SCATTERING MATRICES 43

System (2.1.14) has a non-zero solution if and only if

 
(2.1.17) det I − S(k)eikL = 0.
This equation is known as the secular equation of the graph. We can
summarize this discussion as follows.
Theorem 2.1.8. Consider the eigenvalue problem (2.1.11) on a
graph with the vertex conditions given in one of the equivalent forms of
Theorem 1.4.4. Let k ∈ C \ {0} be such that S(k) exists. Then λ = k 2
is an eigenvalue of H if and only if k satisfies the secular equation
(2.1.17).
Remark 2.1.9. It is obvious from the representation(2.1.6) that
the scattering matrix S(k) is well defined outside the set v∈V iσ(Λv ),
where σ(Λv ) ⊂ R is the spectrum of the self-adjoint operator Λv that
enters the vertex conditions in the form (1.4.14). In general, we cannot
rule out the possibility that one of the negative eigenvalues k 2 will come
from this problematic set. However, if we assume that the scattering
matrix S is independent of k (see Theorem 2.1.6), then this set is clearly
empty. A special care should still be taken of the point k = 0, where
the solution representation (2.1.12) is not valid.
Remark 2.1.10. The secular equation, (2.1.17), is an efficient way
to find the spectrum of a quantum graph numerically. While the left-
hand side is, in general, complex-valued, it can be made real by a
simple multiplication by a phase. Indeed, it can be easily shown by
conjugation (see the proof of Lemma 3.7.3) that the function
  
(2.1.18) η(k) = det I − S(k)eikL / det (S(k)eikL )
is real-valued.
Remark 2.1.11. The structure of matrix S(k) is “almost” block-
diagonal. More precisely, let the 2E × 2E matrix J be defined by
 
0 IE
J= .
IE 0
This is the unitary permutation matrix that reverses the directions of
the bonds. Now the matrix Σ = JS(k) is block-diagonal with blocks
given by the vertex-scattering matrices σ (v) .
The matrix Σ has properties similar to those of the matrices σ, in
particular, when all σ (v) are k-independent, Σ∗ = Σ (see Theorem 2.1.6
for further details).
Example 2.1.12 (Neumann star graph revisited). Looking again at
the compact star graph with E edges and the central vertex of degree
44 2. GRAPH OPERATORS. SPECIAL TOPICS

d = E, see Example 1.4.3, we would like to write down the secular

equation. Scattering matrices at peripheral vertices are 1 × 1 matrices
I. The σ matrix at the central vertex, according to Example 2.1.7, is
2
given by E − I, where E is the d × d matrix of ones. Let L denote
d
the diagonal E × E matrix of lengths of the edges of the graph, L =
diag(L1 , L2 , . . . LE ) (in particular, L is the principal sub-block of the
matrix L of half the dimension). Then the secular equation can be
written as
       
I I − d2 E eikL I I− d2 E eikL
det = det
−eikL I 0 I + I − d2 E e2ikL
   
2
= det I + I − E e 2ikL
.
d
Multiplying by the non-zero factor det e−ikL , we get
   
−ikL 2 ikL eikL
det e + e − Ee
ikL
= det I − E det(2 cos(kL)),
d d cos(kΛ)
where cos(kL) is the diagonal matrix of cos(kLj ). Using the identity
det(I−ED) = 1−Tr D and simplifying, we reduce the secular equation
to
 d  d

(2.1.19) tan(kLj ) cos(kLj ) = 0.
j=1 j=1

This agrees with the previously derived condition (1.4.6), but makes
more explicit the assertion that if n tan’s have poles at a certain k,
then k 2 is an eigenvalue of multiplicity n − 1.

2.2. First order operators and scattering matrices

It is useful to look also at first order differential operators on metric
graphs, in particular the self-adjoint ones. The standard example of a
∂
symmetric first order operator on the real line is i ∂x . In order to define
such an operator on a metric graph, we need to assume that the graph
is directed, i.e. consists of oriented bonds linking the vertices. We
introduce, as before, the coordinate x along each bond, which increases
from 0 to Lb in the direction of the bond b.
At each vertex v, there is a certain number div of incoming bonds
that end at v and dov of outgoing bonds that start at v. Clearly,
div + dov = dv .
 2.2. FIRST ORDER OPERATORS 45

Given a function f ∈ H  1 (Γ), one can consider at each vertex v the

vector F (v) of its values attained at v. We will split it into two parts:
(2.2.1) F (v) = (F i (v), F o (v)),
where the sub-vectors F i (v) and F o (v) of dimensions div and dov are com-
posed from the values attained by f from the incoming and outgoing
bonds respectively.
We would like now to impose some conditions at vertices that would
∂
make the operator D := i ∂x self-adjoint. Let us consider the most
general case: that at each vertex v the conditions are
(2.2.2) Cv F (v) = 0,
where Cv is a linear operator in Cdv .
∂
Theorem 2.2.1. Let the graph Γ be compact and D := i ∂x be de-
fined on the domain
(2.2.3)  1 (Γ) | Cv F (v) = 0 for all vertices v}.
D(D) := {f ∈ H
Then
(1) Vertex condition matrices Cv that make the operator D self-
adjoint exist if and only if for each vertex v the numbers of
incoming and outgoing bonds are equal, i.e. div = dov = dv :=
dv /2.
(2) If div = dov = dv for every vertex v, the conditions
(2.2.4) F o (v) = σ (v) F i (v),
where all σ (v) are unitary matrices, make the operator D self-
adjoint.
(3) Conversely, any conditions that make D self-adjoint can be
written in the form given by (2.2.4).
 1 (Γ). Consider the expression
Proof. Let f, g ∈ H
(2.2.5)
    Lb  
dg df df dg(x)
f, i − i , g = −i (x)g(x) + f (x) dx.
dx dx b∈B
dx dx
0

Integration by parts reduces it to the boundary form F, G of the

operator D:
!
(2.2.6) F, G := i F o (v)Go (v) − F i (v)Gi (v) .
v∈V
46 2. GRAPH OPERATORS. SPECIAL TOPICS

This is clearly a Hermitian symplectic form on C|B| (see Appendix D).

Moreover, it has the canonical form (D.5):
 
F, G = i (F o , Go ) − (F i , Gi ) = (F, ωG),
where (·, ·) is the standard inner product in Ck and
 
iIn+ 0
(2.2.7) ω= .
0 −iIn−
 
Here n+ = v div and n− = v dov .
Any set of homogeneous linear vertex conditions defines a linear
subspace L in C|B| . Moreover, if the vertex conditions are local, i.e.
involve only values at a single vertex at a time, this linear subspace
is the direct sum of individual subspaces Lv ⊂ Cdv defined for each
vertex.
The calculation of the boundary form that we have just done shows
the validity of the following straightforward statement (See Appendix
D for the definitions of the terms used).
Lemma 2.2.2. Operator D with the conditions imposed is symmetric
if and only if L is an isotropic subspace in C|B| with the boundary form
as the Hermitian symplectic form. It is self-adjoint if and only if L is
Lagrangian. In the case of local conditions, these properties are satisfied
for the subspaces Lv for each vertex v.
Now Theorem D.4 applies. Its first statement shows that a La-
grangian subspace (and thus self-adjoint vertex conditions) exist if and
only if n+ = n− (and in particular, the total number of bonds is even).
Moreover, for local conditions to exist, each Lv must be Lagrangian,
which leads to the necessary and sufficient condition that div = dov for
any vertex v. This proves the first statement of the theorem.
To prove the remaining two statements, we invoke the 3rd statement
of Theorem D.4 and the details of its proof. They show that if div =
dov for all vertices, the conditions on F i (v) and F o (v) will produce
Lagrangian subspaces Lv if and only if they can be written in the form
(2.2.4). 
Let us now compare the vertex scattering matrices studied in Sec-
tion 2.1.1 with the ones we encounter as describing self-adjoint vertex
conditions for the first order operator D = id/dx on a directed metric
graph Γ (defined under the condition that at every vertex the numbers
of incoming and outgoing bonds are the same). Both are unitary ma-
trices. They are, however, of different size: the scattering matrices of
the previous sections are of dimension dv × dv , while the ones we en-
countered for the operator D are of twice smaller size dv × dv = d2v × d2v .
 2.2. FIRST ORDER OPERATORS 47

Moreover, the k-independent vertex scattering matrices for the 2nd or-
der operator −d2 /dx2 are severely restricted: according to Theorem
2.1.6, they are allowed to have only the eigenvalues ±1. At the same
time, k-independent vertex scattering matrices for the first order op-
erator D are arbitrary unitary matrices. Let us try to sort out the
situation. The sizes issue is clear: there are twice as many boundary
conditions for a second order operator as for the first order one. Now,
the square of the first order operator D is a second order self-adjoint
operator that acts as −d2 /dx2 along each bond. Let us check what will
be the vertex conditions and the scattering matrix at a vertex v for this
operator. A function f that belongs to the domain of D2 must belong
to the domain D(D) of D and Df = idf /dx must belong to D(D) as
well. This means that the vectors F = (F i , F o ) of vertex values and
F  = (F i , F o ) of derivatives (taken in the bonds’ directions) at v must
satisfy the equalities
F o = σ (v) F i , F o = σ (v) F i .
In terms of the vectors F = (F1 , F2 ) := (F i , F o ) of vertex values and

Fout = (F1 , F2 ) := (−F i , F o ) of outgoing derivatives, the vertex con-
ditions for the operator D2 become
(2.2.8) F2 = σ (v) F1 , F2 = −σ (v) F1 .
Thus, in conditions in the form (1.4.12), one has the matrices
 (v)   
σ −I 0 0
A= ,B = .
0 0 −σ (v) −I
Now, calculation according to the formula (2.1.3) shows that the vertex
scattering matrix for D2 , which we temporarily denote by Σv , is
 
0 −(σ (v) )−1
(2.2.9) Σv = .
−σ (v) 0
This matrix clearly satisfies the equality (Σv )2 = I and thus has its
spectrum consisting of ±1 only, independently on the unitary matrix
σ (v) .
We see that in the case of a directed graph Γ with the ingoing
and outgoing degrees equal at each vertex (div = dov = dv ), prescrib-
ing arbitrary dv × dv unitary scattering matrix at each vertex v leads
to a self-adjoint first order operator D and then to the very special
scattering matrices for the second order Hamiltonian D2 .
Let us look now at the case of an undirected graph Γ. As described
in Section 1.1, one can replace each edge with a pair of bonds going
in opposite directions (later on, the two bonds will be identified), as
48 2. GRAPH OPERATORS. SPECIAL TOPICS

Figure 1. An undirected graph Γ (left) and the corre-

 obtained by duplicating bonds
sponding directed one Γ
(right).

→
Figure 2. Projection π : Γ  Γ illustrated on a single edge.

shown in Figure 1. The resulting directed graph is denoted Γ. Graph

 automatically satisfies the condition on the numbers of incoming and
Γ
outgoing bonds being equal at each vertex.

 → Γ that maps the points

There exists a natural projection π : Γ
x1 and x2 of the two bonds arising from the same point x of an edge
to the point x (see Fig. 2).

This projection allows for the push forward π∗ of a function f on Γ

to the function π∗ f on Γ defined as follows (use Fig. 2 as an illustra-
tion):
π∗ f (x) = f (x1 ) + f (x2 ).
Let now k be an eigenvalue of the operator D on Γ.  This means
that there is a function f such that on each bond it is proportional to
e−ikx and satisfies the vertex conditions (2.2.4) with some given unitary
scattering matrices σ (v) . Then the push forward function g := π∗ f on
Γ is equal on each edge to a combination of e−ikx and eikx , and thus
satisfies the second order equation −d2 g/dx2 = k 2 g. Let us check now
whether the vertex conditions imposed on f can be pushed forward to
some self-adjoint vertex conditions on Γ and thus correspond to a self-
adjoint Hamiltonian. One expects that this is impossible in most cases,
since the variety of arbitrary unitary scattering matrices σ (v) is much
wider than what Theorem 2.1.6 allows. Indeed, consider a vicinity of a
 2.2. FIRST ORDER OPERATORS 49

single vertex, which is a star graph. We pick the coordinate x starting

from the vertex and increasing along the star’s edges. The structure of
the function g is
g = eikx a + e−ikx σ (v) a,
where a is a vector in Cdv , and thus the vector of the vertex values is
G = (I + σ (v) )a.
Then  
g  = ik eikx a − e−ikx σ (v) a
and
G = ik(I − σ (v) )a.
Suppose that there are vertex conditions (1.4.12) that are satisfied for
any k in the spectrum:
" #
AG + BG = A(I + σ (v) ) + ikB(I − σ (v) ) a = 0.
This implies that A(I + σ (v) ) = 0 and B(I − σ (v) ) = 0. Unless σ (v) has
only ±1 in its spectrum, there is a non-zero subspace on which both
matrices A and B vanish, which, according to Theorem 1.4.4 (or, even
more pointedly, Lemma 1.4.7), makes them unsuitable for self-adjoint
vertex conditions.
So, the conclusion is that although choosing arbitrary unitary ver-
tex scattering matrices allows one to define a self-adjoint first order
operator id/dx on the graph Γ,  in most cases (unless the spectra of
all chosen matrices consist of ±1) this does not lead to a definition
of a self-adjoint Schrödinger operator on Γ. This contradicts claims
sometimes made in the literature.
One can, however, find energy dependent vertex conditions (i.e.,
involving matrices A and B that depend on k) that the function g
above will satisfy. As we will see later (e.g., in Sections 3.5, 5.4, 7.2.6,
and 7.5), energy dependent vertex conditions arise naturally in various
situations. Such vertex conditions, though, need to be handled care-
fully, since they result in spectra of operator functions (often called
operator pencils) H(k) more general than the standard in spectral
theory H − k 2 I. Such spectra are known to behave in more complex
ways than spectra of single operators (e.g., [528]).
The alternative view of quantum graphs through scattering ma-
trices rather than self-adjoint second order Hamiltonians is frequently
used in quantum chaos studies (see Chapter 6). It amounts to directly
prescribing general scattering matrices σ (v) (k), instead of starting with
vertex conditions to define the operator H, as we did before. If the
graph is undirected, the waves can travel in both directions on every
50 2. GRAPH OPERATORS. SPECIAL TOPICS

edge. In the more general situation of a directed graph, waves are

allowed to travel only in the specified direction.
Thus, the system is fully specified by a vector of wave amplitudes
a, indexed by the (directed) bonds. The free propagation along the
bond b results in the amplitude ab being multiplied by the phase factor
eikLb . Scattering at vertices is described by a unitary1 matrix T with
the property that
Tb1 b2 = 0 if t(b2 ) = o(b1 ).
Putting these things together, we arrive to the quantum evolution
operator U (k) = T eikL and the same secular equation as in (2.1.17)
for the eigenvalues,
 
(2.2.10) det I − T eikL = 0.
One is allowed to prescribe arbitrary (in particular, k-dependent)
matrices, although one should be aware that sometimes the resulting
spectra might be non-real. In practice, only k-independent matrices
are considered, in which case, since this is equivalent to the spectral
problem of a self-adjoint first order operator, the spectrum remains
real.
One of the popular choices of a scattering matrix is the Discrete
Fourier Transform (DFT) matrix,
exp(2πijm/d)
(2.2.11) σbk bj (v) = √ ,
d
where j = 1 . . . d indexes the incoming bonds at the vertex v and
m = 1 . . . d indexes outgoing ones. This matrix is “democratic” in as-
2
signing equal classical probability σbk bj to each transition. Another
interesting choice is to use the equitransmitting matrices [404],
which prohibit back-scattering and are intimately connected with com-
binatorial Ihara-Selberg zeta functions.
In general, the above choices of σ (e.g., DFT) do not satisfy the
condition of Theorem 2.1.6 and therefore do not correspond to any self-
adjoint second order Hamiltonian. There is an easy-to-spot difference
in the “spectra” {kn } as well: the k-spectrum of (2.1.11) is obviously
symmetric with respect to 0. However, the solutions of (2.2.10) do not
have to have this symmetry. A simple example is a graph consisting of
one edge with L1 = 1 and
 
1 0
T = .
0 −i
1Aswe have discussed before, the unitarity of this matrix requires that at each
vertex of the graph the numbers of outgoing and incoming bonds are equal.
 2.3. FACTORIZATION OF QUANTUM GRAPH HAMILTONIANS 51

In this case the solutions of (2.2.10) are k = 0, π/2 + 2πj, j ∈ Z and

are obviously not symmetric.
Note that the symmetry of the spectrum alone does not imply com-
pliance with Theorem 2.1.6. An example to this effect is
 
sin φ i cos φ
T =σ= ,
i cos φ sin φ
which, for general values of φ, does not satisfy σ 2 = I.
Despite these differences, some results in this book (in particular,
most of the results of Chapter 6) are based on the secular equation
alone (with k-independent scattering matrices) and do not distinguish
between the two methods of defining a quantum graph.

2.3. Factorization of quantum graph Hamiltonians

Both the continuous Laplace operator and (some) its discrete ver-
sions can be factored as
−Δ = A∗ A,
where A is a first order differential operator (e.g., −Δ = (∇)∗ ∇ =
−∇ · ∇ in the continuous case). One wonders under which vertex
conditions the negative second derivative operator H on a quantum
graph can be factored in a similar manner. It is clear that if such a
factorization is possible, the graph must be directed and A must act
as the derivative d/dxb along the bonds. Let us consider a differential
operator that acts this way on a (compact for simplicity) quantum
graph Γ. The operator A should have the domain consisting of H 1 (Γ)
functions which satisfy some zero order conditions Cv F (v) = 0 at each
vertex v. Here Cv is a linear operator in Cdv , which without loss of
generality can be assumed to be an orthogonal projector.
Theorem 2.3.1. Let H be the negative second derivative operator
on Γ equipped with conditions (1.4.14) and defined accordingly to The-
orem 1.4.4. It is factorable as H = A∗ A with a first order operator
A, if and only if the Robin part of the vertex conditions (1.4.14) is
absent, i.e. PR,v = 0 for any vertex v. In this case, PD,v = Cv and
PN,v = I − Cv .
Proof. Assume first that a factorization exists. It is easy to see
that A∗ acts as −d/dxb on the domain f ∈ H 1 (Γ) with additional vertex
conditions (I − Cv )F (v) = 0. Hence, if the factorization is possible, it
leads to the representation of H as −d2 /dx2 acting on H 2 (Γ) with the
conditions Cv F (v) = 0, (I− Cv )F  (v) = 0. This provides the conditions
(1.4.14) with PD,v = Cv , PN,v = I − Cv , and PR,v = 0.
52 2. GRAPH OPERATORS. SPECIAL TOPICS

The proof of the converse statement is similar. 

We see here that the Robin part of the vertex conditions is the ob-
stacle not only to scale invariance of the domain (see Definition 1.4.9),
but also to factorization of the quantum graph Hamiltonian.

2.4. Index of quantum graph operators

In this section we establish a simple formula for the index of a
differential operator on a compact quantum graph Γ.
We start with reviewing briefly the basic notions concerning Fred-
holm operators and their indexes (see Appendix B for more detailed
information and references). Recall that the codimension of a (closed)
subspace E ⊂ H is defined as the dimension of the quotient space H/E,
or, equivalently (in a Hilbert space), the dimension of an orthogonal
complement of E.
Definition 2.4.1. A bounded operator T : H1 → H2 between two
Hilbert (or Banach) spaces is said to be Fredholm, if it has a closed
range and the dimension of its kernel Ker T and the codimension of
its range Ran T are finite. The index of a Fredholm operator T is
defined as
ind T = dim Ker T − codim Ran T.
The following statement is well known (see Appendix B and refer-
ences therein):
Proposition 2.4.2. If operator T is Fredholm and operator K is
compact (in particular, of finite rank), then T + K is also Fredholm
and ind (T + K) = ind T .
Since we want to formulate an index result for a differential operator
of arbitrary (not necessarily even) order, we fix an orientation for each
edge, so that the arc length coordinate xe is unambiguous.
Theorem 2.4.3. Consider the operator
$
T :H m (Γ) → L (Γ)
2

on Γ, defined by the differential expression of order m

m
dm−j
(2.4.1) T = cj (xe )
j=0
dxm−j
e

with the coefficient c0 (x) continuous on each closed edge (but not nec-
essarily on the whole graph) and never equal to zero and all other cj s
measurable and bounded.
 2.4. INDEX OF QUANTUM GRAPH OPERATORS 53

$
Let T1 be the restriction of T as an operator from H m (Γ) into L2 (Γ)

to a subspace of codimension p (e.g., by imposing p vertex conditions

sustainable by H m , i.e., involving derivatives up to the order m − 1).
Then
(1) The operators so defined are Fredholm.
(2) Their indexes are given by the formulas
(2.4.2) ind T = mE, ind T1 = mE − p.
Here, as before, we use the shorthand notation E = |E|.
Remark 2.4.4. The condition that the leading coefficient c0 (x)
is separated from zero will be referred to as the ellipticity of the
operator, since it implies that the operator is Fredholm.
Proof. Consider T naturally acting as a bounded operator from
$
H m (Γ) into L2 (Γ). All terms in T that involve derivatives of order less

than m are compact operators and thus do not influence the Fredholm
dm
property or the index. We thus assume that T = c0 (xe ) dx m . Now T is
e
m
the composition of d m acting from H $ m (Γ) to L2 (Γ) with the invertible
dxe
operator of multiplication by c0 (x) in L2 (Γ). Hence, the proof reduces
to the case of the mth derivative alone. It is easy to show that it is a
surjective operator from H m (e) onto the whole L2 (e) (and thus from
$
H m (Γ) onto L2 (Γ)). On each edge, it has the m-dimensional kernel

consisting of polynomials of degree less than m. Thus, T is Fredholm

and
ind T = (m − 0) = mE.
e

By definition, T1 is the restriction of T onto a subspace M of

codimension p. Consider any (p-dimensional) complement N to M
$
in H m (Γ) and the extension T of T1 from M to the whole H$ m (Γ) that

acts as the zero operator on N . Then the difference T − T vanishes on

M and thus is a finite-dimensional operator. Hence, T is Fredholm of
the same index mE as T .
It is also clear that the ranges of T and T1 are the same and the
kernel of T is p dimensions larger than the kernel of T . Hence, T is
Fredholm and ind T = ind T − p = mE − p. 
This implies in particular
Corollary 2.4.5. Let A be the first derivative d/dx acting as a
1 (Γ) into
bounded operator from a subspace M of codimension p in H
54 2. GRAPH OPERATORS. SPECIAL TOPICS

L2 (Γ). This operator is Fredholm of index E − p. Therefore,

(2.4.3) ind A = E − p = E − dim Pv ,
v

where Pv are the orthogonal projectors describing the vertex conditions

for A. Thus, in particular,
(1) Without any vertex conditions, one has ind A = E.
(2) With continuity condition imposed on functions at all vertices,
one has
ind A = E − (dv − 1) = E − (2E − V ) = V − E.
v

(3) With the condition that the sum of the values of a function at
each vertex is equal to zero:
fe (v) = 0,
{e|v∈e}

one has
ind A = E − 1 = E − V.
v

Notice that v dim Pv arising in this corollary is just the number
of vertex conditions defining H that contain only the values of the
function and no derivatives (Dirichlet part of the conditions).

2.5. Dependence on vertex conditions

In this section, we discuss the dependence of the quantum graph
Hamiltonian H on the vertex conditions. This issue happens to be
important in many circumstances, e.g., when considering dependence
of the spectrum on the vertex conditions or the lengths of the bonds.
The graph Γ is assumed to be compact (i.e., it has finitely many
vertices and edges of the finite length).
It will be convenient here to consider the vertex conditions in the
form (1.4.12)
(2.5.1) Av F (v) + Bv F  (v) = 0,
with the restriction (1.4.10) that the dv × 2dv -matrix
Av := (Av Bv )
has the maximal possible rank dv . The self-adjointness condition on
Av Bv∗ (see equation (1.4.11)) will not be assumed, unless specifically
noted.
 2.5. DEPENDENCE ON VERTEX CONDITIONS 55

Given a set of matrices Av = (Av Bv )|v∈V , we denote by A the

collection of Av for all v ∈ V:
(2.5.2) A := {Av }|v∈V .
Then A can be considered as a block-diagonal matrix of the size
   
dv ×2 dv ,
v v

with individual blocks of sizes dv × 2d . The space of such complex

v 2
matrices A can be identified with C2 v dv , or just C2q , where we will
use the shorthand notation
(2.5.3) q := d2v .
v

We now define the sub-set U of C2q that consists of matrices A satis-

fying the maximal rank condition for each v ∈ V:
(2.5.4) U := {A | (Av Bv ) has maximal rank for any v ∈ V}.
We denote by Us the subset of U consisting of matrices satisfying the
self-adjointness condition (1.4.11):

(2.5.5) Us := {A ∈ U | Av Bv∗ is self-adjoint for any v ∈ V}.

The following statement describes some simple properties of these
sets:
Lemma 2.5.1.
(1) The complement C2q \ U of U in C2q is algebraic.
(2) U is an open and everywhere dense Stein domain (see Ap-
pendix B) in C2q .
Proof. Algebraicity of the set C2q \ U is clear, since its elements
have to satisfy algebraic relations forcing some of the minors to vanish.
This proves the first statement of the Lemma. The second claim is an
immediate corollary of the first one. 
Let us now return to the quantum graph Hamiltonian. Since we are
going to look into its dependence on vertex conditions, we introduce
the corresponding notation:
Definition 2.5.2. The operator −d2 /dx2 defined on functions from
 2 (Γ) satisfying (2.5.1) at each vertex v is denoted by HA .
H
56 2. GRAPH OPERATORS. SPECIAL TOPICS

As we have already seen, the domain D(HA ) of the operator HA is

a closed subspace DA of H  2 (Γ) that can be described as follows:
(2.5.6) DA := {f ∈ H  2 (Γ) | Av F (v) + Bv F  (v) = 0 for all v ∈ V}.
If we define the continuous operator
KA : H 2 (Γ) → C2E ,
where
(2.5.7) KA : f → {Av F (v) + Bv F  (v)} ∈ Cdv = C2E ,
v∈V
then DA is given by
(2.5.8) DA = ker KA .
This simple relation allows us to derive nice behavior properties of the
domain of HA with respect to the vertex conditions matrix A. In order
to do this, let us start with a simple lemma:
Lemma 2.5.3.
(1) The operator function A → KA is analytic in C2q with values
in the space L(H  2 (Γ), C2E ) of bounded linear operators from
H 2 (Γ) to C2E .
(2) For any A ∈ U , the operator KA is surjective.
Proof. Indeed, according to (2.5.7), the function is in fact lin-
ear, and thus analytic with respect to A. Also, the set of vectors
 2 (Γ) is clearly arbi-
{F (v), F  (v)}|v∈V achievable from elements f of H
trary. Then, if the maximal rank condition is satisfied, this implies the
surjectivity of KA . 
Let us consider the trivial vector bundle
U ×H  2 (Γ) → U
over U with fibers equal to H 2 (Γ). Consider the sub-set
%
D := {(A, DA )} ⊂ U × H  2 (Γ).
A∈U
In other words, we look at the domain DA of the operator HA as a
subspace of H 2 (Γ) “rotating” with A. The next results shows that this
domain rotates “nicely” (analytically) with A and is topologically and
analytically trivial as a vector bundle.
Theorem 2.5.4.
(1) D is an analytic sub-bundle of co-dimension 2E of the ambient
trivial bundle.
 2.5. DEPENDENCE ON VERTEX CONDITIONS 57

(2) The bundle D is trivializable. In other words, there exists a

trivialization, i.e., an analytic operator-function T (A) on U
with values in linear bounded operators from a Hilbert space H
into H 2 (Γ), such that ker T (A) = 0 and ran T (A) = DA for
any A ∈ U .
(3) After the trivialization, the values of the resulting analytic in
U operator-function
A → HA T (A)
are Fredholm operators of index zero from H to L2 (Γ).
Proof. The first statement of the Theorem is local. So, let us
pick a matrix A0 ∈ U . According to Lemma 2.5.3, the operator KA0 is
surjective, and thus has a continuous right inverse R. Thus, KA0 R = I.
Then KA R is invertible for A close to A0 . This implies that for such
A, one has KA R(KA R)−1 = I. This implies that
P (A) := I − R(KA R)−1 KA
is a projector onto the kernel of KA , i.e. on DA . Since this projector,
by construction, is analytic with respect to A in a neighborhood of A0 ,
this proves a part of the first claim of the theorem. It only remains to
notice that the co-dimension of the kernel of KA is the dimension of its
range, i.e., 2E.
To prove the second claim, we notice that D is an infinite dimen-
sional analytic Hilbert bundle. Kuiper’s theorem on contractibility of
the general linear group of any infinite-dimensional Hilbert space [495]
implies that all such bundles are topologically trivial. Since the base U
is holomorphically convex, Bungart’s theorem [156] says that the same
holds in the analytic category (see further discussion of the technique
in Appendix B).
Let us prove the third claim. Since HA is the restriction to DA of
the fixed operator −d2 /dx2 (with no vertex conditions attached), acting
continuously from H  2 (Γ) to L2 (Γ), the operator-function in question
can be written as (−d2 /dx2 )T (A) and thus is analytic. The Fredholm
property and the zero value of the index follow from Theorem 2.4.3. 
We now consider a question about the resolvent of HA . In order
to do so, we need to consider the operator family HA − iI, where I
denotes the identity operator in L2 (Γ). The previous theorem, together
with Sobolev’s compact embedding theorem shows that this is also an
analytic family of Fredholm operators.
As the analytic Fredholm theory shows (see Appendix B), the set
Σ of matrices A for which the operator HA − iI : DA → L2 (Γ) is not
58 2. GRAPH OPERATORS. SPECIAL TOPICS

continuously invertible is analytic (i.e., can be given by an equation

φ(A) = 0, where the function φ is analytic). Since this set Σ does not
include any points A of Us (because in this case the operator HA is
self-adjoint), this singular set is nowhere dense.
Theorem 2.5.5. The resolvent (HA − iI)−1 , defined in U \ Σ, is
analytic and has values that are compact operators in L2 (Γ).
Proof. Analyticity of the inverse to an analytic family of bounded
invertible operators is well known (e.g., [436,730]. Thus, (HA −iI)−1 is
an analytic family of operators from L2 (Γ) to H2 (Γ), and thus also as a
family of operators acting in L2 (Γ). However, considered as operators
in L2 , they get multiplied by the compact embedding of DA into L2 (Γ),
and thus are compact. 
2.5.1. Variations in the edge lengths. Sometimes (e.g., in Sec-
tion 3.1.5), one needs to consider the quantum graph’s dependence on
the variations in the edge length parameters {Le } (without changing
graph’s topology). We thus extend the previous considerations to in-
clude the dependence on the vector
L := {Le }|e∈E ∈ (R+ )E .
In fact, we allow complex values of L, i.e. L ∈ (C \ {0})E .
Let Γ be a quantum graph with the Hamiltonian H = −d2 /dx2
equipped with vertex conditions described by a matrix
A = {(Av Bv )|v∈V }.
We would like to vary the lengths of the edges (independently from
each other), without changing topology or vertex conditions. Let us
consider the vector ξ = {ξe } ∈ (R+ )E of dilation factors along each
edge. It is clear that such a dilation is equivalent to keeping the metric
graph structure the same, while replacing H = −d2 /dx2 with Hξ =
{−ξe−2 d2 /dx2 and (Av Bv ) with (Av Bv Ξv ), where Ξv is the diagonal
dv × dv matrix having ξe as its diagonal entries, where e denotes the
edges incident to the vertex v.
Definition 2.5.6. For any A from the previously described set
U of complex matrices A = {(Av Bv )} satisfying the maximal rank
condition and for any vector ξ ∈ (C \ {0})E , we denote by HA,ξ the
Hamiltonian −ξe−2 d2 /dx2e on Γ with the vertex conditions provided by
the matrices Aξ := {(Av Bv Ξv )}.
Notice that this rescaling allows for “complex lengths” of the edges.
This is useful, when considering analytic properties of the Hamiltonian
with respect to the parameters.
 2.6. MAGNETIC SCHRÖDINGER OPERATOR 59

The same arguments as in proving Theorem 2.5.4 lead to the fol-

lowing statement:
Theorem 2.5.7.
(1) The bundle D over U × (C \ {0})E with the fiber over (A, ξ)
defined by the vertex condition matrix Aξ is an analytic sub-
bundle of co-dimension 2E of the ambient trivial bundle
   
U × (C \ {0})E × H  2 (Γ) → U × (C \ {0})E .
(2) The bundle D is trivializable. In other words, there exists a
trivialization, i.e., an analytic operator-function T (A, ξ) on
U × (C \ {0})E with values in linear bounded operators from
a Hilbert space H into H  2 (Γ), such that ker T (A, ξ) = 0 and
ran T (A, ξ) = DAξ for any (A, ξ) ∈ U × (C \ {0})E .
(3) After the trivialization, the values of the resulting analytic in
U × (C \ {0})E operator-function
A → HA,ξ T (A, ξ)
are Fredholm operators of index zero from H to L2 (Γ).

2.6. Magnetic Schrödinger operator

As mentioned earlier, the magnetic Schrödinger operator is
 2
d
(2.6.1) H =−
A
− iA(x) + V (x),
dx
where A(x) is a vector potential, i.e. it changes sign if the orientation of
the variable x is reversed. We assume that A(x) is piecewise continuous
on the graph.
Denote
d
(2.6.2) D= − iA(x).
dx
Then, the operator HA is self-adjoint on the domain consisting of func-
tions that satisfy local vertex conditions of any of the form prescribed
by Theorem 1.4.4 with the change that DF (v) be used instead of F  (v)
everywhere. For example, self-adjoint conditions can be written as
(2.6.3) Av F (v) + Bv DF (v) = 0,
where the matrices Av and Bv satisfy the usual conditions of maximal
rank and Av Bv∗ being Hermitian. Note that to compute DF (v), we
evaluate the derivative and the magnetic potential in the direction into
the edge.
60 2. GRAPH OPERATORS. SPECIAL TOPICS

It is well known that in one dimension the magnetic potential could

be removed by a gauge change. On a graph the situation is similar,
but not as trivial. Consider first a local gauge change,
(2.6.4) G : f (x) → f (x)eiχ(x) ,
where
(2.6.5) χ (x) = A(x),
and χ(x) is continuous on vertices. Then it is easy to check that
(2.6.6) DGf (x) = Gf  (x).
Using the gauge transform on the operator HA results in
d2
G−1 (−D2 + V (x))G = − + V (x) =: H
dx2
with the domain given by the conditions that Gf must satisfy (2.6.3).
Because of (2.6.6) and continuity of χ(x), this is equivalent to f satis-
fying
(2.6.7) Av F (v) + Bv F  (v) = 0.
To summarize, the gauge transformation would show the unitary equiv-
alence of the magnetic operator to a non-magnetic one, if it were pos-
sible to define the function χ(x). Since a graph, in general, is not
simply-connected, defining χ(x) globally is not possible. However, we
can get close to it.
Choose an arbitrary spanning tree for our graph Γ and then choose
an arbitrary point on every edge not in the spanning tree. According to
the definition of the Betti number β(Γ) (see Section 1.1 and Appendix
A), we will have chosen β points. They will be named cut-points
and denoted cj . We can view cj as vertices of degree 2, imposing the
conditions
f (cj −) = f (cj +), −Df (cj −) = Df (cj +).
Note that the latter condition is equivalent to −f  (cj −) = f  (cj +) if
the potential A(x) is continuous at cj , which can always be arranged.
If we cut the graph Γ at the points cj (as their name suggests), we
get a tree, that is a simply-connected graph. We can now define χ(x)
as an integral of A from an arbitrary fixed starting point on Γ along
the unique path to x that does not pass through any cut-points. A
gauge transformation with such χ will succeed in removing the mag-
netic potential everywhere except for the cut-points, which will receive
the quasiperiodic conditions
eiχ(cj −) f (cj −) = eiχ(cj +) f (cj +), −eiχ(cj −) f  (cj −) = eiχ(cj +) f  (cj +).
 2.6. MAGNETIC SCHRÖDINGER OPERATOR 61

Figure 3. A graph with two cut-points and the paths

along which the fluxes α1 and α2 are calculated.

To summarize, we have established the following.

Theorem 2.6.1. Let C = {cj }βj=1 be the set of cut-points. Let
cj +
αj = A(x),
cj −

where the integral is taken over the unique path that does not pass
through any cut-points (see Fig. 3). Then the magnetic Schrödinger
operator HA with vertex conditions specified by {Av , Bv } and equa-
tion (2.6.3) is unitarily equivalent to the non-magnetic operator H with
vertex conditions (2.6.7) and, in addition, the quasiperiodic condi-
tions
f (cj −) = eiαj f (cj +), −f  (cj −) = eiαj f  (cj +).
Definition 2.6.2. The flux of the magnetic field given by A(x)
through a oriented cycle of Γ is the integral of A(x) over the cycle.
With this definition we immediately have a corollary.
Corollary 2.6.3. Let HA1 and HA2 be two magnetic operators
satisfying vertex conditions with the same parameters {Av , Bv }. Then
they are unitarily equivalent if their fluxes through every cycle on Γ are
equal modulo 2π. In fact, it is enough to compute the fluxes through a
fundamental set of β cycles. A magnetic perturbation of a Schrödinger
operator on a graph Γ is thus fully specified by a set of β numbers
between 0 and 2π.
Note that the set of all magnetic operators can be poorer than
[0, 2π)β if some vertex conditions are decoupling, since it can reduce the
effective value of β. In conclusion, we mention that studying Floquet
theory (the subject of Chapter 4) on the maximal abelian covering of
a graph is equivalent to studying all possible magnetic operators: each
magnetic operator corresponds to a single choice of quasi-momentum
in Floquet theory.
62 2. GRAPH OPERATORS. SPECIAL TOPICS

2.7. Further remarks and references

Scattering approach and secular equation
This approach to quantum graphs first appears in [466] (see also
[467] for a more detailed derivation) for a special set of vertex condi-
tions. Fully general conditions were first treated in [459]. Theorems
2.1.6 and 2.1.8 and Lemma 2.1.3 also appeared in this series of papers.
A different determinant condition for the same model was obtained
in [70] and the same condition for a slightly different setup appears
in [164]. The secular equation can be used to compute the eigenval-
ues of the graph numerically and, for some graphs, study the spectral
statistics analytically [56, 97, 102, 103, 439].
One of the earliest appearances of this construction was in a study
of the quantum Hall effect [179]. Within the context of quantum chaos,
the earliest references for the scattering approach are [468, 642, 684].
First order operators
Our discussion here follows [166, 328, 333, 621].
We would like to emphasize again that although choosing arbitrary
unitary vertex scattering matrices allows one to define a self-adjoint
first order operator id/dx on the graph Γ,  in most cases (unless the
spectra of all chosen matrices consist of ±1) this does not lead to a
definition of a self-adjoint Schrödinger operator on Γ. This contradicts
to the claims sometimes made in the literature.
Factorization and index
This part follows essentially the considerations of [328] (see also
related [72, 333, 621]).
Required information about Fredholm operators and their indexes
can be found in Appendix B.
Dependence on vertex conditions and edge lengths
The results of this section come from [104]. The relevant complex
analysis notions and techniques can be found in Appendix B, as well
as in [474, Ch. 1] and [730].
Magnetic Schrödinger operator
Thorough discussions can be found, for example, in [461, 635].
Ours is a simplified version of the exposition. On discrete graphs,
the corresponding result is discussed in [518].
The Outer Space
In some areas of algebra, what we call a metric graph is called an
R-graph (e.g., [698]). An interesting moduli space of metric graphs
with a fixed fundamental group was introduced quite some time ago
in [210] (see also [140, 473] and the surveys [698, 699]). The name
“Outer space” is used for (a version of) this space. The reason for
 2.7. FURTHER REMARKS AND REFERENCES 63

the name is that this moduli space has been used for studying outer
automorphisms of free groups. Various compactifications of this space
have been introduced and studied. This space could possibly turn out
to be useful for some quantum graph studies, but describing it goes
well beyond what can be done in this text.
Non-selfadjoint conditions
It is sometimes needed to consider more general Hamiltonians than
the self-adjoint ones. For instance, one might be interested in accretive
or dissipative Hamiltonians. The paper [457] provides criteria under
which the vertex conditions lead to such operators.
Conditions involving spectral parameter
Sometimes vertex conditions arise that involve the spectral pa-
rameter λ. This usually happens when the ideal vertex corresponds
to an object with some internal structure in a real world problem,
see [277, 278, 311, 476, 482, 483, 493] and Sections 3.5, 5.4, 7.2.6, and
7.5 for various examples.
 CHAPTER 3

Spectra of Quantum Graphs

In this chapter, we discuss a variety of basic issues related to the

spectral structure of quantum graph Hamiltonians H. The first section
contains the simple result showing that if the graph is compact (i.e.,
has finitely many edges and vertices and all edges are of finite length),
the spectrum of the corresponding Hamiltonian is discrete and of fi-
nite multiplicity. It also addresses other relevant issues, e.g. generic
simplicity of the spectrum and eigenvalue bracketing.
While the spectrum of a compact graph can be detected by pres-
ence of L2 -eigenfunctions (called bound states in physics), in the case
of non-compact graphs, this test would only determine what is called
the point spectrum σp (H) (see Appendix C), which could be a small
(or indeed even empty) part of the whole spectrum σ(H). Other parts
of the spectrum (see Appendix C) cannot be found this way. This also
implies that such true eigenfunctions might not (and most frequently
do not) constitute a sufficiently large family for decomposing arbitrary
functions. This is why the next two sections are devoted to the ways
one can detect the spectrum by presence of generalized eigenfunctions.
Here “generalized” means that the functions that solve the correspond-
ing eigenvalue problem do not necessarily belong to the ambient Hilbert
space. Namely, if A is a differential operator (for example, the quantum
graph Hamiltonian), it can be applied, as a differential expression, to
functions that might grow at infinity, and thus do not belong to the am-
bient L2 -space. Such a function might satisfy the eigenvalue equation
and thus be a “generalized eigenfunction”. For instance, the function
d2
v = ex on the real line satisfies the equation Av = v for A = dx 2 . On

the other hand, the natural self-adjoint realization of the second deriv-
ative as an operator in L2 (R) is a negative operator, and thus cannot
contain 1 in its spectrum. Thus, two questions arise:

• Does the existence for some λ ∈ R of a generalized eigenfunc-

tion of a certain growth imply that λ ∈ σ(A)?
• Is there a sufficiently large set of generalized eigenfunctions of
some controlled growth that allow for spectral decomposition?
65
66 3. SPECTRA OF QUANTUM GRAPHS

The answers to both questions under appropriate conditions are affir-

mative. A theorem that answers the first question is usually called a
Shnol’ type theorem, and a quantum graph version of such a result is
presented in Section 3.2. The second type of theorem is called gener-
alized eigenfunction decomposition, which is addressed in Section 3.3.
Although quantum graph Hamiltonians inherit many properties of
second order ODEs and PDEs, there are some significant differences.
One of the most important differences is the failure of the unique con-
tinuation property, considered in Section 3.4, which has far-reaching
consequences described in other parts of the book. The Dirichlet-to-
Neumann map, which is also frequently used, is considered in Sec-
tion 3.5. This technique is applied in the next Section, which addresses
an important relation between the spectra of quantum and discrete
graph operators. Section 3.7 is devoted to trace formulas, which are
essential for quantum graph theory (as well as for general spectral
theory). It is followed by Section, 3.8, devoted to final remarks and
references.
As before, we will adhere to the notations λ = k 2 for the spectral
parameter. Since the notion of a quantum graph Γ (see 1.4.1) includes a
Hamiltonian H, we will interchangeably use the expressions spectrum
σ(H) of operator H and spectrum σ(Γ) of quantum graph Γ.
We introduce now a set of values of the spectral parameter, which
will often play a special role.
Definition 3.0.1. We denote by σD (H) (or σD (Γ)) the union of the
spectra of H on individual edges with Dirichlet conditions at vertices.
In other words:
σD (H) := {λ ∈ C| on an edge e there exists a non-zero
(3.0.1)
solution of Hu = λu vanishing at the edge’s ends.}
2
More explicitly, in the case H = − dx
d
2,

 &
%  πn 2
(3.0.2) σD (H) = |n ∈ N
e∈E
Le

3.1. Basic spectral properties of compact quantum graphs

In this section, we review basic spectral properties of compact quan-
tum graphs.

3.1.1. Discreteness of the spectrum. The main result of this

section is the following
 3.1. BASIC SPECTRAL PROPERTIES 67

Theorem 3.1.1. Let Γ be a compact quantum graph equipped with

the Hamiltonian H acting as −d2 /dx2 (or a more general Schrödinger
operator) on the edges and with vertex conditions (1.4.12). Then the
resolvent (H − iI)−1 is a compact operator in L2 (Γ). Consequently, the
spectrum σ(H) consists exclusively of isolated eigenvalues λj of finite
multiplicity, with λj → ∞.
j→∞

Proof. Due to the self-adjointness of H, the resolvent (H − iI)−1

 2 (Γ). The standard
exists and maps continuously L2 (Γ) into D(H) ⊂ H
Sobolev embedding theorem [7] claims that the natural embedding
H 2 (Γ) → L2 (Γ)
is compact. This implies compactness of the resolvent and thus the
theorem’s claim about the spectrum. 
This theorem enables one to introduce the spectral counting
function
(3.1.1) NΓ (λ) := number of eigenvalues λj (Γ) < λ.
(When no error can arise, we will use the notation N (λ), abandoning
the subscript Γ.) A Weyl-type estimate for N (λ) will be obtained in
the Section 3.7.2.
3.1.2. Dependence on the vertex conditions. In this section,
we will derive some basic facts about the dependence of the quan-
tum graph eigenvalues on the continuous graph parameters, such as
matrices A = {(Av Bv )}|v∈V of vertex conditions and edges’ lengths
{Le }|e∈E . We will adhere to the notations of Section 2.5. In particular,
we denote by HA the operator −d2 /dx2 on a compact graph Γ with the
domain described by the vertex conditions (1.4.12) that correspond  to
the matrix A = {(Av Bv )}|v∈V ⊂ U . Here U ⊂ C (with q =
2q
d2v )
consists of such complex A that the rank of the dv ×2dv -matrix (Av Bv )
is maximal for any vertex v. We will also use notation HA,ξ for the
re-scaled version of HA that acts as −ξ −2 d2 /dx2e on the domain de-
scribed by Aξ = (Av Bv Ξv ). Here ξ = {ξe } ∈ (C \ {0})E is the vector
of (non-zero) scaling factors applied on each edge.
We are interested in the dependence of the spectrum of HA,ξ on the
parameters (A, ξ) ⊂ U × (C \ {0})E . Thus, we consider the operator
pencil HA,ξ − λI of unbounded operators in L2 (Γ).
The following result, which addresses analytic behavior of the spec-
trum, follows from Theorem 2.5.7 and the analytic Fredholm theo-
rem, see Appendix B) and [730, Theorem 4.11].
68 3. SPECTRA OF QUANTUM GRAPHS

Theorem 3.1.2.
(1) The set S of all vectors
(A, ξ, λ) ⊂ U × (C \ {0})E × C
such that HA,ξ − λI does not have a bounded inverse in L2 (Γ),
is principal analytic (see Appendix B). Namely, there exists
a function Φ(A, ξ, λ), analytic in U × (C \ {0})E × C, such that
S coincides with the set of its zeros.
(2) For any integer k ≥ 1, the set Sk of all vectors
(A, ξ, λ) ⊂ U × (C \ {0})E × C
such that dim Ker (HA,ξ − λI) ≥ k is analytic.
Remark 3.1.3. • The set S is the graph of the multiple-
valued function
(A, ξ) → σ(HA,ξ ),
and can be considered as a kind of “dispersion relation”
(see Section 4.1).
• The sets Sk are clearly nested: Sk+1 ⊂ Sk . Moreover, S1 = S.
• For A ∈ Us (i.e., satisfying the condition guaranteeing self-
adjointness of HA ) and ξ ∈ (R \ {0})E , the operator HA,ξ is
selfadjoint and its spectrum is real and discrete.
3.1.3. Eigenfunction dependence. We present here, for com-
pleteness, a simple and well known consequence of perturbation the-
ory [436].
Theorem 3.1.4. The kernel Ker (HA,ξ − λI) forms a holomorphic
k-dimensional vector bundle over the set Sk \ Sk+1 . In other words, if
one has an local analytic eigenvalue branch λ(A, ξ) of constant multi-
plicity k, then, at least locally, one can choose an analytic basis of k
eigenfunctions.
Proof. The spectral projector onto the corresponding spectral
subspace is clearly analytic with respect to the parameter. 

3.1.4. An Hadamard-type formula. Variation of the spectral

data with respect to the domain perturbation is usually studied using
the so-called Hadamard’s variational formulas [332,389,427,600].
For simplicity, we will consider the case of variation of the eigenvalues
with respect to a change in a “loose” edge’s length. Namely, the end
of the edge is assumed to be a vertex of degree 1 and the length of the
edge will be denoted by s. We will also use s to denote the coordinate
 3.1. BASIC SPECTRAL PROPERTIES 69

of the vertex of degree 1. The proof follows the well established pattern
(see, e.g. [381]) and can be easily generalized.
Proposition 3.1.5. Let λ = λ(s) be a simple eigenvalue of a graph
Γ with a loose edge of length s and the Dirichlet condition imposed
at the end-vertex s. Let f = fs (x) be the corresponding normalized
eigenfunction. Then
dλ
= − |f  (s)| ,
2
(3.1.2)
ds

where f (s) is the value of the derivative of the eigenfunction at the
end-vertex.
Proof. First we remark that by preceding theorems we can dif-
ferentiate both the eigenvalue and the eigenfunction. We will omit the
subscript s of f unless we want to highlight the dependence of the
eigenfunction on s.
The Dirichlet condition at the vertex s has the form
fs (s) = 0.
Differentiating it with respect to s we get
∂f
(3.1.3) + f  (s) = 0.
∂s
On the other hand, the eigenfunction is L2 -normalized. Differentiating
the normalization condition (f, f ) = 1 we get
 
∂f
2 Re f, = 0,
∂s
where we used the fact that the loose edge’s contribution to the deriv-
ative is
s s
∂ 2 ∂f
(3.1.4) |fs (x)| dx = |f (s)| + 2 Re
2
f dx
∂s 0 0 ∂s
and fs (s) = 0. Finally, representing the eigenvalue as λ = h[f, f ],
where h is the quadratic form (1.4.20), we get
' (
dλ  2 2 ∂f
(3.1.5) = |f (s)| + V (s) |f (s)| + 2 Re h f, .
ds ∂s
To evaluate the sesqui-linear form h we note that the derivative ∂f /∂s
satisfies the same vertex conditions as f everywhere apart from the
point s. Integrating by parts we get
' (  
∂f ∂f ∂f
h f, = Hf, + f  (s) (s).
∂s ∂s ∂s
70 3. SPECTRA OF QUANTUM GRAPHS

Now we use that Hf = λf and equations (3.1.3) and (3.1.4) to get

' (  
∂f ∂f
2 Re h f, = 2λ Re f, − 2|f  (s)|2 = −2|f  (s)|2 .
∂s ∂s
Substituting the last equation and the Dirichlet condition into (3.1.5)
yields the desired result. 
One can obtain similar results when varying vertex conditions rather
than the length. For simplicity we only consider the δ-type vertex con-
ditions.
Proposition 3.1.6. Let λ = λ(α) be a simple eigenvalue of a graph
Γ which satisfies δ-type vertex condition at v with the parameter α = ∞.
Then
dλ
(3.1.6) = |f (v)|2 .
dα
If we re-parameterize the conditions at v as
df
(3.1.7) ζ (v) = −f (v),
e∈E
dxe
v

now allowing Dirichlet (ζ = 0) and excluding Neumann (ζ = ∞) con-

ditions, the derivative is
2
dλ df
(3.1.8) = (v) .
dζ e∈Ev
dxe

Proof. The proof follows the pattern of the proof of Proposi-

tion 3.1.5 with minor modifications. We deal with the α-derivatives
first. The derivative of the normalization condition is now
 
∂f
2 Re f, = 0.
∂α
Taking the derivative of the quadratic form, see (1.4.27), yields
' (
dλ ∂f
= |f (v)| + 2 Re h f,
2
.
dα ∂α
Integrating by parts inside the sesqui-linear form h and collecting to-
gether all the “boundary” terms at v, we get
' (  
∂f ∂f ∂f df ∂f
h f, = α (v)f (v) − (v) (v) + Hf,
∂α ∂α ∂α e∈Ev
dxe ∂α
 
∂f
= Hf, ,
∂α
 3.1. BASIC SPECTRAL PROPERTIES 71

where we used the δ-type condition at the vertex v. Now we use Hf =

λf and the derivative of the normalization condition to get
' (
∂f
2 Re h f, = 0,
∂α
and, therefore,
dλ
= |f (v)|2 .
dα
To deal with the Dirichlet case, we calculate, for α = 0 and ζ = 0,
2
dλ 1 dλ |f (v)|2 df
= 2 = = (v) ,
dζ ζ dα ζ2 e∈Ev
dxe

using condition (3.1.7) in the last step. Since λ(ζ) is an analytic func-
tion, the value of the derivative at ζ = 0 now follows by continuity. 

3.1.5. Generic simplicity of the spectrum. There are well

known results about generic (with respect to varying the coefficients of
the operator) simplicity of spectra of elliptic self-adjoint operators
on compact manifolds [19, 691]. It is believed that in “most situa-
tions” the spectrum of a compact quantum graph is generically (with
respect to varying the lengths of the edges) simple. the restriction to
“most situations” is due to a simple counterexample — a graph that is
topologically a circle, equipped with Neumann conditions. In the case
of Neumann vertex conditions, it happens to be the only exception,
as the following result of Friedlander [319], provided here without a
proof, shows:
Theorem 3.1.7. Let Γ be a compact graph with at least one ver-
tex of degree different from 2 and with the Hamiltonian H = −d2 /dx2
equipped with Neumann conditions (1.4.4): the functions are continu-
ous at vertices and
df
(v) = 0
e∈E
dxe
v

at any vertex v. Then, for any ε > 0 one can change the edge lengths
Le of Γ by δLe < ε to obtain a graph with simple spectrum for H.
Moreover, the set of lengths vectors (Le1 , Le2 , . . . , LeE ) for which the
spectrum is simple is a subset of the 2nd Baire category in RE +.

An analog of this result has not been established for more general
self-adjoint vertex conditions. One has to be careful here, though, since
allowing general conditions, one can have the situations where the circle
graph with Neumann conditions is hidden. Consider, for instance, the
72 3. SPECTRA OF QUANTUM GRAPHS

“figure eight” graph in Fig. 1 (left), where four functions fj converge

along four edges at a single vertex v.

f4 f1 f3 f1
v v1
v2
f3 f2 f4 f2
Figure 1. A figure of 8 graph (left) and its equivalent
loop (right).

Suppose that the vertex conditions are

f1 (v) = f3 (v), f2 (v) = f4 (v), f1 (v) + f3 (v) = 0, f2 (v) + f4 (v) = 0.
It is clear that the Hamiltonian H in this case is unitarily equivalent to
the one on a loop with two vertices, equipped with Neumann conditions
(Fig. 1, right). Thus, the spectrum is multiple for any set of edge
lengths. We conjecture, however, that generic simplicity holds for δ-
type conditions.
Another interesting and important question is whether the eigen-
functions generically do not vanish on entire edges. More restrictively,
one can ask if the eigenfunctions generically do not vanish on vertices.
One motivation for this question the reader will encounter in Assump-
tion 5.2.2 which underlies most of section 5.2.
An interesting counter-example to a blanket positive statement is
provided by the “figure 8” graph of Fig. 1, now with Neumann con-
ditions at v. Indeed, let L1 denote the length of the left loop. When
λ = (π/L1 )2 there is an eigenfunction that is equal to sin(x/L1 ) on the
left loop (here, x = 0 and x = L1 both correspond to the vertex v)
and identically zero on the right loop. Obviously, changing the edge
lengths does not qualitatively affect the situation.
We therefore conjecture that, if Γ is a compact connected graph
with no loops,1 the eigenfunctions generically — with respect to the
edge lengths — do not vanish on vertices of Γ. Note that the “no loop”
condition excludes the circle graph as well.

3.1.6. Eigenvalue bracketing. We will be assuming that the

Hamiltonian H is −d2 /dx2 + V (x) with the vertex conditions specified
in the results. The eigenvalues λn of H (also referred to as the eigen-
values of the graph and denoted λn (Γ)) are labeled in non-decreasing
order, counting their multiplicity.
1Modulo “dissolving” the Neumann vertices of degree 2 into their edge.
 3.1. BASIC SPECTRAL PROPERTIES 73

The first theorem of this section describes the effect of modifying

the vertex condition at a single vertex. We denote by Γα a compact
(not necessarily connected) quantum graph with a distinguished vertex
v. Arbitrary self-adjoint conditions are fixed at all vertices other than
v, while v is endowed with the δ-type condition with coefficient α:

f is continuous at v and
 df
dxe
(v) = αf (v).
e∈Ev

Theorem 3.1.8. Let Γα be the graph obtained from the graph Γα
by changing the coefficient of the condition at vertex v from α to α .
If −∞ < α < α ≤ ∞ (where α = ∞ corresponds to the Dirichlet
condition, see section 1.4.4), then
(3.1.9) λn (Γα ) ≤ λn (Γα ) ≤ λn+1 (Γα ).

λn(Γα ) is simple and its eigenfunction f is such that

If the eigenvalue
either f (v) or f (v) is non-zero, then the inequalities can be made
strict,
(3.1.10) λn (Γα ) < λn (Γα ) < λn+1 (Γα ).
Proof. The case of strict inequalities follows simply from the pos-
itivity of the derivative of λn (Γα ) with respect to the parameter of
the vertex condition, Proposition 3.1.6. For the possibly degenerate
case we directly use the monotonicity of the quadratic form and the
rank-one nature of the perturbation.
Denoting by Γ∞ the graph with the Dirichlet condition at the vertex
v, we will actually prove the chain of inequalities
(3.1.11) λn (Γα ) ≤ λn (Γα ) ≤ λn (Γ∞ ) ≤ λn+1 (Γα ),
which is obviously equivalent to inequality (3.1.9). Since we are now
considering the Dirichlet case separately, we will assume that α = ∞.
Consider the quadratic forms hα , hα and h∞ of the corresponding
Hamiltonians. According to the discussion in section 1.4.4, we have
(3.1.12) hα [f, f ] = h∞ [f, f ] + α|f (v)|2
and
hα [f, f ] = h∞ [f, f ] + α |f (v)|2
on the appropriate subspaces of H  1 (Γ). In fact, D(hα ) = D(hα ) and
D(h∞ ) = {f ∈ D(hα ) : f (v) = 0}.
All inequalities follow from the min-max description of the eigen-
values, namely
(3.1.13) λn = min max h[f, f ].
X⊂D: dim(X)=n f ∈X: f =1
74 3. SPECTRA OF QUANTUM GRAPHS

The first inequality in (3.1.11) follows immediately from the obser-

vation that hα ≥ hα for all f .
The domain D(h∞ ) is smaller than D(hα ) and the forms h∞ and
hα agree on D(h∞ ). Minimization over a smaller space results in a
larger result, implying the second inequality in (3.1.11).
The last inequality follows from the fact that D(h∞ ) is a subspace of
D(hα ) of co-dimension one. To provide more detail, let the minimum
for λn+1 (Γα ) be achieved on the subspace Y (which is the span of the
first n + 1 eigenvectors) of dimension n + 1. Then there is a subspace
Y∞ of dimension n, such that Y∞ ⊂ Y and Y∞ ⊂ D(h∞ ). Then
λn (Γ∞ ) = min max h∞ ≤ max h∞
X: dim(X)=n f ∈X f ∈Y∞

= max hα ≤ max hα = λn+1 (Γα ).

f ∈Y∞ f ∈Y

This is precisely the last needed inequality. 

This theorem allows us to prove a simple but useful criterion for
the simplicity of the spectrum of a tree.
Corollary 3.1.9. Let T be a tree with a δ-type condition at every
internal vertex and an extended δ-type condition at every vertex of de-
gree 1. If the eigenvalue λ of T has an eigenfunction that is non-zero
on all internal vertices of T , then λ is simple.
Equivalently, if an eigenvalue λ of the tree T is multiple, there is
an internal vertex v such that all functions from the eigenspace of λ
vanish on v.
Proof. The two statements are almost contrapositives of each
other, modulo the following observation: if for every internal vertex
there is an eigenfunction that is non-zero on it, one can construct an
eigenfunction which is non-zero on all internal vertices. Indeed, if m is
the dimension of the eigenspace of λ, then the subspace of the eigen-
functions vanishing on any given v is at most m − 1. A finite union of
subspaces of dimension m − 1 cannot cover the entire eigenspace.
We will work by induction on the number of internal vertices. If a
tree has no internal vertices, it is an interval and there is nothing to
prove since all eigenvalues are simple.
Assume the contrary: there is an eigenfunction f which is not zero
on all internal vertices of the tree, but λ is not simple. Take an arbitrary
internal vertex v and another eigenfunction g. Cutting the tree at the
vertex v we obtain dv subtrees. On at least one of them the function
g is not identically zero and not a multiple of f . Let T  be a such a
subtree.
 3.1. BASIC SPECTRAL PROPERTIES 75

Then there is an α < ∞ such that (λ, f ) is an eigenpair of the

tree Tα endowed with the condition f  (v) = αf (v). Similarly, there is
an α ≤ ∞ such that (λ, g) is an eigenpair of the tree Tα  . By induc-
tive hypothesis, λ is simple on Tα , therefore α = α. This, however,
contradicts inequality (3.1.10). 
The next theorem deals with the modification of the structure of
the graph by gluing a pair of vertices together.
Theorem 3.1.10. Let Γ be a compact (not necessarily connected)
graph. Let v0 and v1 be vertices of the graph Γ endowed with the δ-type
conditions, i.e.
⎧
⎨f is continuous at vj and
 df
⎩ (v ) = αj f (vj ),
dxe j
j = 0, 1.
e∈Evj

Arbitrary self-adjoint conditions are allowed at all other vertices of Γ.

Let Γ be the graph obtained from Γ by gluing the vertices v0 and
v1 together into a single vertex v, so that Ev = Ev0 ∪ Ev1 , and endowed
with the δ-type condition
df
(3.1.14) (v) = (α0 + α1 )f (v).
e∈Ev
dxe

Then the eigenvalues of the two graphs satisfy the inequalities

(3.1.15) λn (Γ) ≤ λn (Γ ) ≤ λn+1 (Γ).
Proof. Similarly to the proof of Theorem 3.1.8, we consider the
quadratic forms of the two graphs and observe that they are defined
by exactly the same expression (see section 1.4.4). However, joining
the vertices together imposes an additional restriction on the domain
of the quadratic form of the graph Γ , namely
D(h ) = {f ∈ D(h) : f (v0 ) = f (v1 )} .
Thus, the domain D(h ) is a co-dimension one subspace of D(h) and
the rest of the proof is identical to the proofs of the second and third
inequalities in (3.1.11). 
Notice that if the domain of h had co-dimension k, then one would
have obtained by applying the same argument the inequality
λn (Γ) ≤ λn (Γ ) ≤ λn+k (Γ).
This observation immediately leads to the following generalization of
Theorem 3.1.10:
76 3. SPECTRA OF QUANTUM GRAPHS

Theorem 3.1.11. Let the graph Γ be obtained from Γ by k iden-

tifications, for example by gluing vertices v0 , v1 , . . . vk into one, or
pairwise gluing of k pairs of vertices. Each identification results also
in adding parameters αj in the vertex δ-type conditions, as in (3.1.14).
Then
(3.1.16) λn (Γ) ≤ λn (Γ ) ≤ λn+k (Γ).
This statement can be also proved by the repeated application of
Theorem 3.1.10.
Remark 3.1.12. Theorem 3.1.8 applied to the case α = ∞ is also
a result about joining dv vertices together, since the vertex Dirichlet
condition has the effect of disconnecting the edges at the vertex. Note
the difference with the result in Theorem 3.1.11: the eigenvalues of
the joined graph are now shifted down, but not further than the next
eigenvalue of Γ, which contrasts with the weaker “not further than k-th
next eigenvalue” result of the Theorem 3.1.11.
3.1.7. Dependence on the coupling constant at a vertex.
As in section 3.1.6 we consider the family of compact graphs Γα with
a distinguished vertex v. Arbitrary self-adjoint conditions are fixed at
all vertices other than v, while v is endowed with the δ-type condition
with coefficient α:

f is continuous at v and
 df
dxe
(v) = αf (v).
e∈Ev

We will be using the second condition in the form

df
(3.1.17) (z + 1) (v) = i(z − 1)f (v),
e∈Ev
dxe

where |z| = 1. The Dirichlet condition corresponds to z = −1, while

the Neumann-Kirchhoff corresponds to z = 1.
We denote by H(z) the operator with the above condition at v
and the previously fixed set of conditions at all other vertices. The
main result of this section is a representation of the spectrum of this
operator as the range at z of an irreducible multiple valued analytic
function defined near the unit circle, plus a fixed discrete set, see Fig. 2.
Theorem 3.1.13. There exist a bounded from below discrete set
Δ ∈ R and a multiple valued function Λ(z) (“dispersion relation”)
analytic in a neighborhood of the unit circle S 1 and real on S 1 , such
that:
 3.1. BASIC SPECTRAL PROPERTIES 77

Figure 2. Spectrum of a graph as a function of the

parameter z, which controls the matching condition of
the form (3.1.17) at one of the vertices of the graph. The
figure shows the non-decreasing function Λ(z) spiraling
upwards and one of the constant branches from the set
Δ.

(1) For any z ∈ S 1 , one has

%
(3.1.18) σ(H(z)) = Λ(z) Δ,

where σ(A) denotes the spectrum of the operator A.

(2) The function Λ is irreducible, i.e. any branch determines by
analytic continuation the whole function Λ.
(3) Each analytic branch of Λ is monotonically increasing in the
counter-clockwise direction of S 1 . The branches do not inter-
sect.
78 3. SPECTRA OF QUANTUM GRAPHS

Proof. Notice that, as we already know, for each z ∈ S 1 , H(z) is a

bounded from below self-adjoint operator with a discrete spectrum. Let
λn (z) be its nth eigenvalue, counted with multiplicity in non-decreasing
order. Then, according to the perturbation theory, it is continuous on
the unit circle cut at z = −1, where one can observe that the ground
state λ1 has a one-sided limit equal to −∞.
Lemma 3.1.14.
(1) The function λn (z) is non-decreasing in the counter-clockwise
direction on the unit circle.
(2) There exists a function F (λ, z) analytic in a neighborhood of
the unit circle S 1 such that λ ∈ σ(H(z)) if and only if F (λ, z) =
0.
Indeed, the first claim of the Lemma is a rephrasing of a part of
Theorem 3.1.8 (and, essentially, a consequence of monotonicity of the
quadratic form (3.1.12) with respect to α). The second claim is a direct
consequence of Theorem 3.1.2.
Lemma 3.1.15. If either
(1) λ is a multiple eigenvalue of H(z), for some z ∈ S 1 , or
(2) λ belongs to the spectra of H(z1 ) and H(z2 ) for two different
points z1 , z2 ∈ S 1 ,
then λ ∈ σ(H(z)) for any z ∈ S 1 .
Indeed, if the eigenspace of H(z) corresponding to λ is more than
one-dimensional, then it contains an eigenfunction vanishing at the
vertex v. Then, assuming z = −1, we conclude from (3.1.17) that the
sum of derivatives is zero as well. Therefore this eigenfunction works
for any point z ∈ S 1 . In the case z = −1 we select an eigenfunction
with vanishing sum of derivatives at v and it automatically vanishes at
v since it satisfies the Dirichlet
) condition.
Let now λ ∈ σ(H(z1 )) σ(H(z2 )) with the corresponding eigen-
functions f1 and f2 . Without loss of generality, z1 = −1. First con-
sider the case f1 (v) = 0. Then, by Theorem 3.1.8, λ as an eigenvalue
of H(z1 ) lies strictly between consecutive eigenvalues of H(z2 ) which
contradicts our assumption. Thus f1 (v) = 0 and the sum of derivatives
of f1 is zero by (3.1.17). We conclude that f1 is an eigenfunction for
any z ∈ S 1 thus completing the proof of the Lemma.
We now define the set Δ as the set of all λ ∈ R such that λ ∈
σ(H(z)) for all z ∈ S 1 . As the previous
) lemma shows, one can de-
scribe Δ as the intersection σ(H(−1)) σ(H(1)) of the Dirichlet and
Neumann spectra and thus it is discrete and bounded from below.
 3.2. THE SHNOL’ THEOREM 79

Consider now a neighborhood U in C of the circle S 1 and the set Σ

that is the closure of the set
 
% !
(3.1.19) Σ := {z} × σ(H(z)) \ (C × Δ) .
z∈U

To put things simply, we remove from the union of the spectra all
“horizontal” branches (z, λ) with λ ∈ Δ. The second statement of
Lemma 3.1.14 then implies that the set Σ is analytic. Moreover, for
the neighborhood U being sufficiently small, it is a smooth complex
analytic curve. Indeed, Lemma 3.1.15 implies that for all λ ∈ / Δ the
eigenvalues are simple, and thus the eigenvalue branch is analytic. Let
now λ ∈ Δ and (z0 , λ) ∈ Σ for some z0 ∈ S 1 . Then Lemma 3.1.15 says
that the eigenvalue λ near z0 either stays constant, or splits into the
constant and possibly one or more increasing branches. The constant
branches are excluded in the definition of Σ, equation (3.1.19). Also,
there can be no more than one increasing branch, otherwise one would
find two values of z with the same values of λ in Σ, which the lemma
allows to happen only on horizontal branches. Moreover, as we have
already concluded, the eigenvalue on such a branch must be simple.
Then Rellich’s theorem (e.g., [626, v.4, Theorem XII.3]) says that the
increasing branch is analytic. We thus conclude that Σ consists of one
or more non-intersecting analytic curves.
We will show now that there is only one component, if the neighbor-
hood )U is sufficiently small. First of all, the projection onto the λ-axis
of Sj (S 1 × R), where Sj is any of the components of Σ is the whole
real axis. Otherwise the component whose projection does not cover
the whole real axis would create an accumulation point, which would
contradict the discreteness of the spectrum for each z ∈ S 1 . Then exis-
tence of two or more components would create equal eigenvalues for at
least two distinct values of z, which would contradict Lemma 3.1.15.
Thus, Σ is an irreducible analytic curve that intersects each line
{z} × R on the cylinder along the discrete spectrum σ(H(z)) and such
that it is monotonically increasing counterclockwise. Hence, it forms a
kind of a spiral winding around the cylinder infinitely many times when
λ → ∞ and having a vertical asymptote when λ → −∞ (otherwise one
would get a contradiction with the boundedness of each of the spectra
from below). This proves the statement of the theorem. 

3.2. The Shnol’ theorem

Let Γ be a connected quantum graph. If it is compact, the spectrum
is pure point, as shown in Section 3.1.1, and is determined by the
80 3. SPECTRA OF QUANTUM GRAPHS

presence of L2 -eigenfunctions (bound states). In the non-compact

case, though, other types of spectrum appear and one can not expect
to detect it by looking for L2 -eigenfunctions. However, it turns out
that the presence of the spectrum is still signaled by the “generalized
eigenfunctions” — solutions to the eigenvalue equation that do not
grow too fast. This is the subject of the so-called Shnol’ type theorems.
Assumption 3.2.1. We will assume that Γ satisfies the following
conditions:
(1) The degree of every vertex is finite and the lengths Le of the
edges are bounded away from zero and infinity: 0 < Lmin ≤
Le ≤ Lmax < ∞ for any edge e.
d2
(2) The Hamiltonian is the Schrödinger operator − 2 +V (x) with
dx
an L∞ -potential2 V and any self-adjoint vertex conditions.
We will assume that a “root” vertex o is singled out (the result will
not depend on the choice of the root). Then we can define a “norm”
ρ(x) of a point x as
ρ(x) = dist(o, x).
This now allows us to define for any r ≥ 0 the “ball” Br :
Br = {x ∈ Γ | ρ(x) ≤ r}.
Theorem 3.2.2 (A Shnol’ type theorem). Let the graph Γ sat-
isfy Assumption 3.2.1 and λ ∈ R. If for any  > 0 there exists a
function φ(x) = φ (x) on Γ that belongs to the Sobolev space H 2 on
each edge, satisfies all vertex conditions, satisfies the equation
d2 φ
(3.2.1) − 2 + V (x)φ = λφ for a.e. x ∈ Γ,
dx
and the sub-exponential growth condition

(3.2.2) |φ(x)|2 dx ≤ C er ,

Br

then λ ∈ σ(H).
This theorem implies in particular the following
Corollary 3.2.3 (A Bloch type theorem). Let the graph Γ
satisfy Assumption 3.2.1 and be of a sub-exponential growth (i.e., the
volume of Br grows sub-exponentially with the radius r). If there exists
a bounded solution of the equation (3.2.1), then λ ∈ σ(H).
2In
fact, weaker conditions on the potential, e.g. local square integrability and
boundedness from below, suffice.
 3.2. THE SHNOL’ THEOREM 81

Simple examples show that existence of a bounded solution does not

guarantee that λ ∈ σ(H), if the graph is of exponential growth (e.g., a
regular3 tree of degree 3 or higher). In fact, there are bounded harmonic
functions (e.g., a constant) on such a graph, while the spectrum σ(H)
does not contain zero.
Proof of Theorem 3.2.2. Let us define for any r > 0 the fol-
lowing compact subset Γr of the graph: it consists of all edges whose
both ends belong to Br . The following inclusions are obvious:
(3.2.3) Γr−Lmax ⊂ Br ⊂ Γr+Lmax .
We hence conclude that in the integral sub-exponential growth condi-
tion (3.2.2) one can replace Br by Γr . We introduce the function

(3.2.4) J(r) := |φ(x)|2 dx.

Γr

Given an  > 0, one can find a sequence rk → ∞ such that

J(rk + Lmax ) ≤ eLmax J(rk ),
otherwise one gets a contradiction with the sub-exponential growth
condition. We remind to the reader that each set Γr consists of com-
plete edges only.
Let θ(x) be any smooth function on [0, Lmin /4] that is identically
equal to 1 in a neighborhood of 0 and identically equal to zero close to
Lmin /4. Here Lmin is the lower bound for the lengths of all edges of Γ,
which was assumed to be strictly positive. We define a cut-off function
θk on Γ. It is equal to 1 on Γrk and to 0 on all edges which do not
have vertices in Γrk . We only need to define it along the edges that
have exactly one vertex in Γrk . Let e be an edge whose one vertex v is
contained in Γrk . The function θk is defined to be equal to 1 along e
starting from v till the middle of the edge, then it is continued by an
appropriately shifted copy of θ(x) (which by construction will become
zero at least at the distance le /4 from the end of the edge), and stays
zero after that. Notice that due to the construction, any derivative of
the functions θk (x) is uniformly bounded with respect to k and x ∈ Γ.
Besides, these functions are identically equal to 1 or 0 around any
vertex.
We can now construct a sequence of approximate eigenfunctions
φk (x) of the operator H as follows:
φk (x) = θk (x)φ(x).
3A graph is regular, if all vertex degrees are equal. See Appendix A.
82 3. SPECTRA OF QUANTUM GRAPHS

One can notice that the functions φk (x) satisfy the same vertex con-
ditions that φ did, since the factors θk are identically equal to 1 or 0
around the vertices. This implies that φk (x) belongs to the domain of
H in L2 (Γ). Besides, we clearly have
(3.2.5) φk 2
≥ J(rk ).
Let us now apply H − λ to these test functions:
(H − λ)φk = θk (−φ + V (x)φ − λφ) − 2θk φ − θk φ
(3.2.6)
= −2θk φ − θk φ.
We have used here that φ satisfies (3.2.1).
Using the properties of the cut-off functions θk , one gets
 
(3.2.7) (H − λ)φk 2 ≤ C |φ(x)|2 + |φ (x)|2 dx.
x∈ supp θk

Since the potential V (x) is uniformly bounded from below, and since
the supports of the derivatives θk belong to the interiors of the edges
and at a qualified distance from the vertices, we have standard Schauder
estimates for
|φ (x)|2 dx
x∈ supp θk

by for instance the integral

|φ(x)|2 dx ≤ J(rk + Lmax ) − J(rk ).

Lmin L
ρ(x)∈[rk + 4
,rk +Lmax − min
4 ]
This leads to the estimate
(H − λ)φk 2 ≤ C(J(rk + Lmax ) − J(rk )) ≤ C(e − 1)J(rk )
(3.2.8)
≤ C(e − 1) φk 2 ,
where the constant C does not depend on k, . Since  > 0 was arbi-
trary, we conclude that λ ∈ σ(H). 

3.3. Generalized eigenfunctions

In this section we describe briefly the situation with the generalized
eigenfunction expansion on quantum graphs. Since the detailed con-
sideration would become longer and more technical than this book’s
aim would allow, we refer the reader to the original works listed in the
corresponding part of Section 3.8.
Let Γ be a quantum graph. In order to make the presentation sim-
pler, we consider the (negative) second derivative Hamiltonian H =
 3.3. GENERALIZED EIGENFUNCTIONS 83

−d2 /dx2 on Γ with some self-adjoint vertex conditions, which we as-

sume to be written in the form (1.4.14):
PD,v F (v) = 0, PN,v F  (v) = 0, PR,v F  (v) = Λv PR,v F (v)
for each vertex v. Here, as before, PD,v , PN,v , and PR,v are complemen-
tary (“Dirichlet, Neumann, and Robin”) orthogonal projectors in Cdv
and Λv is a self-adjoint invertible operator on the range of PR,v .
The quantum graph can be very general, including being infinite
(i.e., having infinite numbers of edges and vertices), non-compact (e.g.,
infinite edges are allowed), having multiple edges and loops. The only
restrictions that we impose are:
(1) The degree of each vertex is finite.
(2) The lengths le of the edges are bounded from below: Le ≥
Lmin > 0.
(3) The “Robin” operators Λv are uniformly bounded:
Λv ≤ C < ∞.
Because of the locality of the operator, one can define the values
of Hφ for any function φ(x) ∈ H 2 loc (Γ) satisfying the imposed vertex
conditions. Notice that we impose no control on behavior of the func-
tion φ(x) at infinity. In particular, it does not have to belong to the
domain of the Hamiltonian, or even to the ambient space L2 (Γ).

Definition 3.3.1. A function φ(x) ∈ H 2 loc (Γ) satisfying the im-

posed vertex conditions is said to be a generalized eigenfunction
with the spectral parameter λ of the Hamiltonian H, if
Hφ(x) = λφ(x) a. e. on Γ.
The considerations of Section 3.2 suggest that generalized eigen-
functions (i.e. the solutions to the eigenvalue equation that are not
necessarily from L2 ) have spectral meaning only if one imposes some
reasonable bounds on their growth. The next result shows that un-
der similar growth conditions one can expand functions from L2 (Γ) in
terms of the generalized eigenfunction of H.
Theorem 3.3.2 ([508], see also previous related results in [138,
413]). Under the imposed assumptions on the quantum graph Γ and
Hamiltonian H, let μ be the spectral measure for H. Let also w : Γ →
[1, ∞) be continuous and such that 1/w ∈ L2 (Γ). Then there exists a
“spectral decomposition” into functions φλ such that for μ-almost all
λ ∈ σ(H), φλ is a generalized eigenfunction of H with the spectral
parameter λ and w−1 φλ ∈ L2 (Γ).
84 3. SPECTRA OF QUANTUM GRAPHS

The details concerning the “spectral expansion” and the proof of

this theorem can be found in the referenced work [508].
One should notice that in order to have a meaningful application
of this theorem, one needs to choose the weight function w as slowly
growing as possible. Then one obtains the expansion into generalized
eigenfunctions of moderate growth. Picking a weight w that increases
too fast renders the theorem’s conclusion trivial.

3.4. Failure of the unique continuation property. Scars

The standard uniqueness theorem claims that a solution u(x) of a
second-order ODE with a non-vanishing coefficient at the second deriv-
ative that has a degenerate zero at a point x0 (i.e., u(x0 ) = u (x0 ) = 0),
vanishes identically. There exist analogs of this statement for second
order elliptic PDEs [549], which claim that a solution that vanishes
to the infinite order at a point, is identically equal to zero (unique
continuation property). A weaker formulation (weak unique con-
tinuation property) says that a solution vanishing on an open do-
main vanishes identically4. The uniqueness of continuation property
is extremely significant in many issues, e.g. in spectral theory and in-
verse problems. It is thus important to mention that, although graph
Hamiltonians are similar to ordinary differential operators, as well as
to elliptic partial differential operators, uniqueness of continuation
in general fails for quantum graphs.
Proposition 3.4.1. There exist connected quantum graphs such
that some of the eigenfunctions of the corresponding Hamiltonian van-
ish identically on an edge. In particular, there exist infinite periodic
quantum graphs that have some compactly supported eigenfunctions.
It is indeed easy to construct such eigenfunctions. We have seen
one example in Section 3.1.5. For a further example, consider the circle
graph Γ with no vertices, equipped with the natural angle coordinate
x. Let u(x) be a non-constant real-valued eigenfunction of −d2 /dx2 .
It has certain number of zeros x1 , . . . , xn ∈ Γ. Let us declare these
points the vertices and attach any additional graph structures to the
circle at these vertices, see Fig. 3. We impose Neumann conditions at
the vertices x1 , . . . , xn and any homogeneous vertex conditions (e.g.,
Dirichlet, Neumann, Robin, etc.) throughout the rest of the graph.
Extending u(x) as zero to the rest of the graph, we clearly obtain an
eigenfunction of the new quantum graph, which vanishes everywhere
4The weak uniqueness of continuation does not necessarily hold for higher order
elliptic equations (e.g., [613]).
 3.5. DIRICHLET-TO-NEUMANN MAP 85

x1 x3
x2

Figure 3. Creation of a compactly supported eigen-

function (a scar). It demonstrates the failure of the
uniqueness continuation property.

apart from the original circle. Further examples can be found, for
instance, in Section 7.6.3. It is also clear that in this manner one
can also construct a periodic quantum graph for which the compactly
supported function u is an eigenfunction.
Definition 3.4.2. In some cases (e.g., in quantum chaos studies)
eigenfunctions that vanish on an entire edge are called perfect scars,
or simply scars. In particular, compactly supported eigenfunctions on
infinite graphs are scars.

3.5. The ubiquitous Dirichlet-to-Neumann map

In this section, we will address an object (or rather approach) that
has appeared under different guises in a variety of areas of mathematics
and mathematical physics. In inverse problems for PDEs it is given the
long, but revealing name Dirichlet-to-Neumann map (DtN map),
which we will adopt here as well. Names like Schur complement in
matrix algebra, Feshbach formula or admittance (or Poincare-Steklov)
operator in mathematical physics, and Weyl’s m-function in spectral
theory essentially address the same object. The celebrated M. Krein’s
resolvent formula is also closely related. The DtN map has recently
played an important role in analysis on graphs and fractals.
The idea of the DtN map is rather simple: If one is interested in the
spectrum of a system that consists of two sub-systems communicating
through a common boundary, one can eliminate one of the sub-systems,
replacing it with a (usually eigenvalue-dependent) boundary condition
for the second sub-system. We will see and use several incarnations of
this idea in the text.

3.5.1. DtN map for a single edge. We will consider here the
simplest case of a quantum graph, consisting of a single finite edge. In
86 3. SPECTRA OF QUANTUM GRAPHS

spite of its simplicity, this case is important, for instance, for establish-
ing relations between the spectra of discrete and quantum graphs in
Section 3.6.
We consider the graph consisting of two vertices v and w and a
single edge of the length L ∈ (0, ∞), see Fig. 4. The DtN works as
v w

Figure 4.

follows. We designate the vertices v and w as the “boundary” of Γ,

prescribe the boundary values of a function φ(x) as a vector (a, b)t ,
and solve with these (inhomogeneous) Dirichlet boundary conditions
the eigenvalue problem with the spectral parameter λ = k 2 :
 2
− ddxφ2 = k 2 φ,
(3.5.1)
φ(v) = a, φ(w) = b.
Notice that such a solution might not exist for certain values of k,
d2
namely, when λ = k 2 belongs to the spectrum σD of the operator − dx 2

with zero Dirichlet conditions on the segment [0, L]. This spectrum is
clearly equal to {(nπ/L)2 }, where n takes all non-zero integer values.
Assuming that the x-coordinate runs from 0 at v to L at w, one
gets (outside the indicated discrete set) the explicit formula
1
(3.5.2) φ(x) = (a sin(k(L − x)) + b sin(kx)) .
sin(kL)
Let us now compute the Neumann data (φ (0), −φ (L))t for this solution
(the choice of sign corresponds to taking the derivatives in the “internal
normal” directions, i.e., into the segment). This leads to
    
φ (0) k −a cos kL + b
=
−φ (L) sin kL a − b cos kL
(3.5.3)   
λ − cos kL 1 a
= .
sin kL 1 − cos kL b
Thus, the matrix
 
k − cos kL 1
(3.5.4) Λ(k) :=
sin kL 1 − cos kL
transforms the Dirichlet data into Neumann data and thus deserves
the name of a Dirichlet-to-Neumann (or DtN) map.
Notice that the DtN map, as a matrix function of the spectral
parameter λ, has poles at the excluded Dirichlet spectrum. This feature
 3.5. DIRICHLET-TO-NEUMANN MAP 87

will play an important role in some future considerations, e.g., in the

spectral gap opening procedures.
3.5.2. DtN map for a compact graph with a “boundary”.
The procedure described above can now be generalized to an arbitrary
compact quantum graph Γ with a subset B ⊂ V of vertices designated
as a “boundary” (Fig. 5). In the preceding example, the boundary
consisted of both vertices of the single-edge graph. The vertices that

Figure 5. A compact graph with a sub-set B of vertices

designated as a “boundary.”

are not designated as the boundary will be called interior.

Let us now have a Hamiltonian H = −d2 /dx2 (or a more general
Schrödinger operator) on Γ with certain vertex conditions at the inte-
rior points and zero Dirichlet conditions at the boundary (i.e., φ|B = 0).
Let us pick a vector a ∈ C|B| of boundary values on B and consider the
following problem:
⎧ 2
⎪
⎨− dx2 = k φ on Γ,
d φ 2

(3.5.5) prescribed conditions at interior vertices,

⎪
⎩φ| = a.
B

When λ = k 2 is outside the discrete spectrum of H, this problem has

a unique solution. Indeed, uniqueness is guaranteed by the exclusion
of the Dirichlet spectrum. Existence also follows from this condition.
Indeed, one can find an extension of a to a function ψ on Γ that sat-
isfies all vertex conditions (e.g., the extension can be made vanishing
everywhere except a small neighborhood of B). Introducing the new
unknown function φ1 := φ − ψ, one obtains an inhomogeneous equa-
tion with zero Dirichlet conditions. Since the Dirichlet spectrum is
excluded, there exists a solution.
Now, the Dirichlet-to-Neumann map is defined as the linear
operator Λ(k) in C|B| that maps the vector a of Dirichlet values of
the solution φ into the vector of its “Neumann data” on B. We here
understand the “Neumann data” of φ at some boundary vertex v ∈ B
as the sum of all outgoing (i.e., into the edges) derivatives of φ at v.
We also need to use the mapping
N :H 2 (Γ) → C|B|
88 3. SPECTRA OF QUANTUM GRAPHS

2 (Γ) and any boundary vertex v ∈ B

that assigns to any function f ∈ H
its “Neumann data” at v. More precisely,
df
(3.5.6) N f (v) := (v),
e∼v
dxe

where the sum is extended over all edges incident to v and the deriva-
tives are taken in the directions away from the vertex.
It is clear that the DtN map is closely related to the resolvent of the
operator H. It is easy to make this relation explicit. Let E : C|B| →
2 (Γ) be a linear (and thus continuous) extension mapping, such that
H
for any a ∈ C|B| the function ψ := Ea ∈ H 2 (Γ) satisfies the following
three conditions:
(1) ψ|B = a,
(2) ψ satisfies at all interior vertices the vertex conditions associ-
ated with the operator H,
(3) N ψ = 0.
Note that ψ is not required to satisfy any differential equations on
the graph. It is clear that such a mapping exists (and is manifestly
non-unique).
One can now use this function to re-write problem (3.5.5) in terms
of the operator H. This is done by representing φ as the sum f + ψ =
f + Ea in (3.5.5). Here f clearly belongs to the domain of the operator
H and one can re-write (3.5.5) as follows:
 ! !
d2 d2
(H − λ)f = dx 2 + λ ψ = dx2
+ λ Ea on Γ,
(3.5.7)
f |B = 0.

One thus obtains the following useful formula that expresses the DtN
map as a block of the resolvent:
 2 
d
(3.5.8) Λ(λ) = N R(λ, H) + λ E.
dx2

(We employed the notation Λ(λ) for Λ(k) with k 2 = λ.)

One again notices that this matrix function is meromorphic with
possible poles at the excluded Dirichlet spectrum σD (H). However,
there could be situations (e.g., in presence of scars, see Section 4.5)
when some of these possible poles do not materialize. We will provide
in the next section, in a special case of a single point boundary, a
sufficient condition for the corresponding pole to appear.
 3.5. DIRICHLET-TO-NEUMANN MAP 89

3.5.3. DtN map for a single vertex boundary. The simplest

(and useful) special case is of a graph with a single vertex boundary
B = {v}. Here one expects the DtN map to be a scalar meromorphic
function with poles possible at the Dirichlet spectrum σ(H).
The following theorem can be derived easily from (3.5.8):
Theorem 3.5.1. Let λ0 = k02 ∈ σ(H) be a simple Dirichlet eigen-
value of H and f (x) be the corresponding eigenfunction. Then the
principal part of the DtN map Λ(k) near k0 is
N f (v)
(3.5.9) λ0 .
λ0 − λ
In particular, if λ0 = 0 and N f (v) = 0, there is indeed a pole at this
location.
If a “scar” exists (Section 3.4), i.e., there is an eigenfunction ψ
compactly supported away from the boundary B (situation impossible
for elliptic second order equations in Rn ), the formula (3.5.9) shows
that the corresponding pole is masked in the DtN map.
3.5.4. DtN map and the secular equation. Consider the sit-
uation opposite to the one discussed in Section 3.5.3: every vertex is
the boundary. We can apply then the DtN map to each edge (Section
3.5.1).
More precisely, consider a compact graph Γ and let F = F (v, e) de-
note the 2E-dimensional vector of the values of the function f attained
at the edges’ ends. Similarly, let F  = F  (v, e) be the vector of the
function’s derivatives. We denote by ΛE (k) the DtN map from F to
F  , which treats every edge as being disjoint from the others. In other
words, ΛE (k) is an amalgamation of the DtN maps of the individual
d2
edges. For example, when the operator is − dx 2 (with no potential), the

results of Section 3.5.1 imply that

 
− cot(kL) csc(kL)
(3.5.10) ΛE (k) = k ,
csc(kL) − cot(kL)
where L = diag(L1 , . . . , Le ) and the edge-ends have been arranged
appropriately.
Suppose now that the vertex conditions on the graph Γ are written
in the form
AF + BF  = 0,
where the matrices A and B are appropriate amalgamations of the
matrices Av and Bv , satisfying the usual conditions, see Theorem 1.4.4.
The following version of the secular equation is now immediate.
90 3. SPECTRA OF QUANTUM GRAPHS

Theorem 3.5.2. Let k 2 ∈ σD (Γ) (see Definition 3.0.1). Then k 2 ∈

σ(Γ) if and only if

(3.5.11) F (k) := det(A + BΛE (k)) = 0.

3.5.5. DtN map and number of negative eigenvalues. Using

the DtN map and the secular equation of Section 3.5.4 one can get a
beautifully simple characterization of the number of negative eigenval-
ues of the quantum graph.

Theorem 3.5.3 ([62]). The number of negative eigenvalues of the

d2 
operator − dx 2 on the graph with self-adjoint conditions AF + BF = 0

is given by the number of positive eigenvalues of the matrix

AB ∗ + BΛE (0)B ∗ ,

where
 
−L−1 L−1
ΛE (0) = lim ΛE (k) = .
k→0 L−1 −L−1
The idea of the proof is to use a symmetrized version of the secular
determinant (3.5.11) and to observe that as k goes from i∞ to i0, the
eigenvalues of the matrix inside the determinant move monotonically.
The number of times k 2 was an eigenvalue is equal to the number of
eigenvalues of the matrix that moved across zero and became positive.

3.6. Relations between quantum and discrete graph spectra

One can wonder whether there is any relation between the spectra
of quantum and discrete operators (Hamiltonians) on a graph. As it
happens, a simple and useful relation does exist, if one assumes that
the graph is equilateral. So, in this section we will consider a count-
able (i.e., with countably many vertices and edges) and equilateral
d2
quantum graph Γ with the Hamiltonian H = − dx 2 and Neumann-

Kirchhoff vertex conditions (1.4.4). The common length of the edges

will be denoted l. It will be also assumed that the degrees of vertices
are bounded from above by a constant d.
We will use here the technique analogous to the one of the preceding
Section 3.5.4.
We will assume that the unbounded self-adjoint operator H is de-
fined according to Section 1.4. We will be considering its spectrum
σ(H) and comparing it to the spectrum of the (harmonic) Laplacian
Δ := ΔΓ on the discrete graph Γ defined as in Section 1.2.
 3.6. QUANTUM VS DISCRETE SPECTRA 91

We remind the reader that Δ acts on a function f (v) defined on

the set V of vertices of Γ as follows:
1
(Δf )(v) = f (v) − f (w).
dv w∼v
We will also be interested in various parts σi of the spectra, where
i = p, pp, disc, ess (see Appendix C for the notation).
At the first glance, one might get the wrong impression that the
two operators are too different to be related. Indeed, H is unbounded
and Δ is bounded. A quick formal consideration, however, fixes this.
Indeed, let u be a (generalized) eigenfunction on the quantum graph Γ
and f := u|V be its restriction to the vertices of Γ. Solving the equation
d2 u
− 2 = k2u
dx
on each edge (which can be identified with the segment [0, L]) in terms
of the values f (0), f (L) of u at the end points gives
1
(3.6.1) u(x) = (f (0) sin k(L − x) + f (L) sin kx) .
sin kL
This is possible, unless λ = k 2 is an eigenvalue of the Dirichlet problem
for the second derivative, i.e. unless sin kL = 0, or λ ∈ σD (Γ), see
Definition 3.0.1.
Now the only thing that remains to satisfy is the Neumann ver-
tex conditions. In fact, when we used the vertex values of u, it was
already tacitly assumed that the continuity condition at vertices is sat-
isfied. Thus, we only need to make sure that the sum of the outgoing
derivatives at each vertex is equal to zero. Using (3.6.1), one can easily
compute the derivative of u at x = 0:
k
u (0) = (f (L) − f (0) cos kL) .
sin kL
Substituting this into the Kirchhoff condition at a vertex v, one gets it
in the form
f (v)dv cos kL − f (w) = 0,
w∼v
or
(3.6.2) Δf (v) = (1 − cos kL)f (v).
This suggests that unless λ = k 2 ∈ σD (Γ), λ ∈ σ(H) if and only if
(1 − cos kL) ∈ σ(Δ).
Indeed, our calculation in fact proves this for σp . Using a little bit
more elaborate technique based on M. Krein’s resolvent formula, one
can prove the following result, cited from [588]:
92 3. SPECTRA OF QUANTUM GRAPHS

Theorem 3.6.1. Unless λ = k 2 ∈ D, the following relation holds:

λ ∈ σ(H)
if and only if
(1 − cos kL) ∈ σ(Δ).
The same is true for the parts σi of the corresponding spectra, where
i = p, pp, disc, ess.
One should treat the points of the excluded Dirichlet spectrum with
care, since the situation here can look different for different quantum
graphs. For example, at these points scars could appear (see, e.g.,
Section 7.6.3).
One can notice that the described procedure that relates the quan-
tum and discrete graph spectra amounts to replacing each edge with
its DtN map (see Section 3.5.1). We will also see other examples where
DtN maps for more general graphs than a single edge will be useful.
More general operators that involve electric and magnetic poten-
tials, as well as somewhat more general vertex conditions, can be
treated similarly, see Section 3.8.

3.7. Trace formulas

In this section we explore the connection between the spectrum of a
quantum graph and the set of all periodic orbits on it. This connection,
known as the trace formula, first appeared in the quantum graph case
in the work of Roth [630, 631] and was later rediscovered by Kottos
and Smilansky [466, 467].
We will assume that the boundary conditions have no Robin part,
i.e. PR,v = 0 for all vertices v (see Theorem 1.4.4). In particular, this
implies that the eigenvalues of the graph are non-negative and that the
bond scattering matrix S(k) is k-independent.
To count the eigenvalues of the graph we shall count the roots of
the secular equation
 
(3.7.1) det I − S(k)eikL = 0,
where the unitary bond scattering matrix S(k) is composed of the
scattering matrices σ (v) (see Lemma 2.1.3) in a manner prescribed by
(2.1.15). We will show that the roots k correspond to the eigenvalues
λ = k 2 up to the multiplicity with one exception: eigenvalue zero.
Let N sp (K) be the spectral counting function, i.e. the number of
eigenvalues λ = k 2 of the graph with k < K. Further, let N alg (K1 , K2 )
be the number of roots of (3.7.1) in the interval (K1 , K2 ). Both eigen-
values and roots are counted with their multiplicity. Then the two
 3.7. TRACE FORMULAS 93

counting functions are related by

1
(3.7.2) #{λ < K 2 } =: N sp (K) = N alg (K, Kxs) − C,
2
where C is a constant accounting for the wrong multiplicity of the
eigenvalue zero5. Namely, C = N0alg /2 − N0sp .
Due to the simple connection between the two functions, we will
slightly abuse the terminology and call N alg (K1 , K2 ) the algebraic
counting function. We will also omit the superscript in the notation.

3.7.1. Secular equation. Under the assumption S(k) = S(0) we

can re-state Theorem 2.1.8 in a more powerful form, which makes sure
equation (3.7.1) gives the eigenvalues with their correct multiplicities.
Theorem 3.7.1. Consider the eigenvalue problem for the operator
H described in Theorem 1.4.4 with the vertex conditions that have no
Robin part (PR,v ≡ 0 on all vertices v). Then k 2 ∈ C \ {0} is an
eigenvalue with multiplicity g if and only if k is a root of (3.7.1) with
the same multiplicity.
Proof. There are two differences between this result and Theo-
rem 2.1.8. The first is that the spectrum of operators Λv is now empty,
so the only restriction on k 2 is k 2 = 0. The second difference is that now
we are able to predict the multiplicities of the eigenvalue from equa-
tion (3.7.1). It is easy to see from the construction of Section 2.1.2
that the multiplicity of the eigenvalue k 2 is the same as the multiplic-
ity of 1 as the eigenvalue of SeikL . The fact that the multiplicity of 1
is the same as the multiplicity of the root of (3.7.1) follows from the
representation
   2E
det I − Se ikL
= (1 − eiθj (k) ),
j=1

where eiθj (k) are eigenvalues of SeikL , and the inequality dθj (k)/dk > 0;
see the following Lemma. 
Lemma 3.7.2. Let S be a 2E × 2E unitary matrix, and let eiθ(k) be
an eigenvalue of the unitary matrix SeikL and u(k) be the corresponding
normalized eigenvector. Then,
dθ
(3.7.3) = u(k), Lu(k),
dk
5One can find a discussion of the relation between this value and index theory
in [328].
94 3. SPECTRA OF QUANTUM GRAPHS

where L is the matrix of bond lengths defined by (2.1.16). In particular,

dθ
(3.7.4) Lmin ≤ ≤ Lmax .
dk
Proof. Since U = SeikL is unitary and analytic, the function θ(k)
can be chosen real analytic. Thus we can differentiate the eigenvalue
equation,

(3.7.5) SeikLu(k) = eiθ(k)u(k),

with respect to k to obtain (we suppress the k-dependence of u and θ)

dθ iθ
(3.7.6) SeikL iLu + SeikLu = i e u + eiθ u .
dk
We now compute the inner product of this expression with (3.7.5).
After cancellations due to unitarity we obtain

dθ
u, iLu + u, u  = u, i u + u, u ,
dk
leading to (3.7.3) because of normalization of u. Inequality (3.7.4)
follows from the normalization too. 

The determinant of I−SeikL is in general a complex-valued function,

but a standard trick can be used to make it real-valued. This makes
easier both numerical search for the roots of (3.7.1) and the derivation
of a trace formula for it.
We define
 !  
ζ(k) = det(S ∗ ) det e−i 2 L det I − SeikL .
k
(3.7.7)

 !
Since the pre-factor det(S ∗ ) det e−i 2 L is never zero, the zeros of
k

ζ(k) are exactly the same as the zeros of the original secular equa-
tion (3.7.1); in particular they give the eigenvalues of the correspond-
ing spectral problem with correct multiplicities (apart from the special
case k = 0). The function ζ(k) is analytic and we will now prove that
it is real-valued on the real line.

Lemma 3.7.3. The analytic function ζ(k) defined by equation (3.7.7)

is real on the real line. Therefore, it satisfies ζ(k) = ζ(k) for k ∈ C.
 3.7. TRACE FORMULAS 95

Proof. Using the unitarity of the matrices involved in the defini-

tion of ζ(k), we evaluate for real k
!  
ζ(k) = det(S ∗ )−1/2 det ei 2 L det I − e−ikL S ∗
k

!    
= det(S ∗ )−1/2 det ei 2 L det SeikL − I det e−ikL S ∗
k

= ζ(k).

3.7.2. Weyl’s law. Using Lemma 3.7.2, we can obtain an estimate
on the number of eigenvalues in a given interval. This type of result
is often referred to as the “Weyl’s law”. We note that because graphs
are essentially one-dimensional, the error term is of order 1.
Lemma 3.7.4. Let N (K1 , K2 ) be the algebraic counting function of
a quantum graph with non-Robin boundary conditions. Then,
2L
N (K1 , K2 ) = (K2 − K1 ) + R,
2π
where L = L1 + · · · + LE and R is bounded uniformly in K1,2 .
Proof. Assume that 0 ∈ (K1 , K2 ). Then each eigenvalue of the
quantum graph corresponds to one of the phases θj (k) being a multiple
of 2π. We know that θj are increasing functions and can estimate
Nj (K1 , K2 ), the number of eigenvalues in the interval (K1 , K2 ) due to
j-th phase only, by
1 1
(θj (K2 ) − θj (K1 )) − 1 ≤ Nj (K1 , K2 ) ≤ (θj (K2 ) − θj (K1 )) + 1.
2π 2π
Representing the difference of θ’s by an integral and summing them
over j we obtain
K2 2E
1 dθj
N (K1 , K2 ) = dk + R,
2π K1 j=1
dk

where the error term R is bounded by |R| ≤ 2E. Using (3.7.3) we

obtain
K2 2E
1
N (K1 , K2 ) = uj , Luj  dk + R
2π K1 j=1
K2
1 2L
= Tr(L) dk + R = (K2 − K1 ) + R,
2π K1 2π
as required.
96 3. SPECTRA OF QUANTUM GRAPHS

If 0 ∈ (K1 , K2 ), there is a possible discrepancy between the mul-

tiplicity of 0 as a graph eigenvalue and as a solution of the secular
equation. But this discrepancy can be shown to be smaller than 2E in
absolute value. 
3.7.3. Derivation of the trace formula. We derive the trace
formula in two steps: first an expansion in terms of traces and then
evaluation of the traces in terms of periodic orbits.
Theorem 3.7.5. Let N (−K, K) be the algebraic counting function
of a graph with non-Robin boundary conditions. Then
∞
L 2 1  i(K+iε)L n
(3.7.8) N (−K, K) = 2K + lim Im Tr Se ,
π ε→0 π n=1
n
where the limit is taken pointwise for all K that are not the eigenvalues
of the graph (and therefore points of discontinuity of the function N ).
Proof. The trace formula derivation is based on the Argument
Principle: if f (z) is a meromorphic function having no zeros or poles
on a simple positively oriented contour C, then
* 
1 f (z) 1
(3.7.9) N −P = dz = ΔC arg f (z),
2πi C f (z) 2π
where N and P denote respectively the number of zeros and poles of
f (z) inside the contour C counted together with their multiplicities (or-
ders). The second form uses ΔC , the change in the continuous branch
of the argument over the contour C.
To find an expansion for N (−K, K) we will integrate ζ  (k)/ζ(k)
over a rectangular contour with corners at ±K ± iε. We will then take
the limit ε → 0 while K remains fixed and satisfies ζ(K) = 0. This
restriction makes sure that the contour does not pass through a zero
of ζ(k) and is quite natural because N (−K, K) is undefined (has a
discontinuity) when K is an eigenvalue.
First we estimate the change in the argument of ζ(k) over the ver-
tical parts of the contour. Since the function ζ(k) is continuous,
max |ζ(K + iκ) − ζ(K)| = δ(ε) → 0 as ε → 0.
κ∈[−ε,ε]

A simple geometric reasoning then shows that the maximum change in

the argument of ζ(K + iκ) is 2 arcsin (δ(ε)/ζ(K)), therefore it also goes
to zero. We stress again that this argument is valid only for a fixed
value of K that is not an eigenvalue of the graph, i.e. ζ(K) = 0.
Thus we have a pointwise limit N (−K, K) = limε→0 Nε (−K, K),
where Nε is the sum of the integrals over the horizontal parts of the
 3.7. TRACE FORMULAS 97

contour. These integrals are evaluated as follows

−K
1 K
ζ  (s − iε) 1 ζ  (s + iε)
Nε (−K, K) = ds + ds
2πi −K ζ(s − iε) 2πi K ζ(s + iε)

1 K
ζ (s + iε) ζ  (s + iε)
=− − ds
2πi −K ζ(s + iε) ζ(s + iε)
1 ζ  (s + iε)
K
= − Im ds
π −K ζ(s + iε)
1 K+iε
= − Im ln ζ(k) ,
π −K+iε

where the logarithm is chosen to be continuous over the part of the

contour from −K + iε to K + iε. Substituting the definition of the
function ζ, equation (3.7.7), we obtain
1 "  # K+iε
Nε (−K, K) = − Im ln e−ikL det I − SeikL
π −K+iε
L 1   K+iε
= 2K − Im Tr ln I − SeikL ,
π π −K+iε

where L = Tr(L)/2 is the sum of lengths of the edges of the graph.

Since the matrix S is unitary, the eigenvalues of Sei(x+iε)L with x ∈ R
and ε > 0 satisfy
(3.7.10) |μ| ≤ e−εLmin < 1,
where Lmin is the minimal of the edge lengths. Indeed, e−εLmin is the
maximal singular value of Sei(x+iε)L , i.e. the square root of the maximal
eigenvalue of
 i(x+iε)L ∗ i(x+iε)L
Se Se = e−2εL .
and the inequality follows from the relationship between eigenvalues
and singular values. Due to estimate (3.7.10) we can expand the Tr ln
into an absolutely convergent series
+∞ ,K+iε
L 1 1  iλL n
Nε (−K, K) = 2K + Im Tr Se .
π π n
n=1 −K+iε

The series converges uniformly in k (since estimate (3.7.10) does not

depend on k) and is thus continuous in k. Therefore the sum of the
series automatically satisfies the restriction on the continuity of the
logarithm along the corresponding contour.
Using k-independence of the matrix S and Remark 2.1.11 together
with some simple matrix manipulations we can show that the traces
98 3. SPECTRA OF QUANTUM GRAPHS
 n
evaluated at the endpoints are complex conjugate, Tr Sei(−K+iε)L =
n
Tr (Sei(K+iε)L ) . This leads to
∞
L 2 1  i(K+iε)L n
(3.7.11) Nε (−K, K) = 2K + Im Tr Se .
π π n=1
n
Taking the limit pointwise completes the proof. 
Let us now analyze the convergence issues of the series in (3.7.8).
First of all, each term of the series is continuous in K and, for ε > 0,
the series is absolutely convergent due to estimate (3.7.10). Moreover,
since the estimate is uniform in K, the sum of the series is continuous
in K, acquiring discontinuities only in the limit ε → 0.
Furthermore, fix K which is not an eigenvalue of the graph. De-
noting by μj (ε) the j-th eigenvalue of the matrix Sei(K+iε)L (which can
be chosen to be continuous in ε), we notice that the series in (3.7.11)
can be represented as
2E ∞
1 n
(3.7.12) Im μ (ε).
j=1 n=1
n j
For ε = 0 the eigenvalues lie on the unit circle, but are not equal
to 1, since this would mean that ζ(K) = 0. Thus, the series is also
convergent for ε = 0 although not absolutely. We can take termwise
limit ε → 0 in (3.7.11) if we show that series converges uniformly in ε.
To show this, we observe that μj (ε) are bounded away from 1 and
lie in the closed unit circle. Let δ be such that |μj (ε) − 1| ≥ δ. We use
summation by parts to estimate the partial sum
k+m n k+m k+m−1   n
μ 1 1 1
= j
μ + − μj
n=k
n k + m n=k n=k
n n + 1 j=k
k+m−1   k
1 μk − μk+m+1 1 1 μ − μn+1
= + −
k+m 1−μ n=k
n n+1 1−μ
k+m−1  
1 2 1 1 2 2
≤ + − = .
k+mδ n=k
n n+1 δ kδ
Uniform convergence follows immediately.
Another consequence of the representation (3.7.12) is that, since we
are dealing with 2E expansions of the principal branches of Logμj(ε),
the entire sum is bounded in absolute value by 2Eπ 1 − e−εLmin ≤
2Eπ. This, incidentally, is another proof of the Weyl law, Lemma 3.7.4.
Using dominated convergence we conclude that the limit in (3.7.8)
 3.7. TRACE FORMULAS 99

can be taken in the distributional sense, with sufficiently fast decaying

integrable test functions. We summarize the above observations in the
following remarks.

Remark 3.7.6.
(1) For ε > 0 the series in (3.7.8) converges absolutely. The sum
is a continuous function uniformly bounded by 2Eπ.
(2) The limit ε → 0 can be taken termwise leading to

∞
L 2 1  iKL n
(3.7.13) N (−K, K) = 2K − Im Tr Se ,
π π n=1
n

where the series is conditionally convergent for K’s that are

not eigenvalues.

3.7.4. Expansion in terms of periodic orbits. We will now

proceed to expand the trace formula in terms of periodic orbits on the
graph.

Definition 3.7.7. The ordered set of bonds γ = (b1 , . . . , bn ) is a

path if e(bj ) = o(bj+1 ) for j = 1, . . . , n − 1. The path is closed if, in
addition, e(bn ) = o(b1 ). Here n is called the topological length of
the path. The set of all closed paths of length n will be denoted by Cn .

Let σ be the left shift operation on Cn , σ(b1 , . . . , bn ) = (b2 , . . . , bn , b1 ).


Let n be the smallest integer such that σ n (γ) = γ. A closed path γ is
called primitive if n = n. Otherwise, γ is a repetition of a closed
path γ  of length n . The repetition number of the closed path γ is
rγ = n/n . It is easy to see that rγ is an integer. It follows from the
definition that the repetition number of a primitive path is 1.

Definition 3.7.8. The periodic orbit is an equivalence class of

the closed paths with respect to the shift σ. We write p = [b1 , . . . , bn ].
The set of all periodic orbits of length n is denoted by Pn .

Other terms such as the topological length and the repetition num-
ber are defined for the periodic orbits in the same fashion as for the
closed paths.
We are now ready to formulate the periodic orbit version of the
trace formula.
100 3. SPECTRA OF QUANTUM GRAPHS

Theorem 3.7.9.
∞
L 2 1
N (−K, K) = 2K − Im Aγ eikLγ
π π n=1
n γ∈C n
∞
L 2 1
= 2K − Im Ap eikLp ,
π π n=1 p∈P
r p
n

where, for a closed path γ = (b1 , . . . , bn ) (correspondingly, periodic orbit

p = [b1 , . . . , bn ]),
Aγ = Sb1 ,b2 Sb2 ,b3 · · · Sbn ,b1
is the stability amplitude and
Lγ = Lb1 + . . . + Lbn
is the metric length.
The series are conditionally convergent for all K that are not eigen-
values and the summation over the (finite) sets Cn (correspondingly,
Pn ) must be performed before the summation over n.
Proof. The formula above follows directly from (3.7.13) with the
application of the identity
Tr(M n ) = Mj1 ,j2 Mj2 ,j3 · · · Mjn ,j1 .
j1 ,...,jn


3.7.5. Other formulations of the trace formula. In the physics
literature, the quantity that appears on the left-hand side of the trace
formula is the density of states, which is the derivative of N (K),
∞
d(k) = δ(k − kn ).
n=−∞

When the issues of convergence are given due attention, the density of
states appears either in a mollified form or integrated against a suitably
regular test function h(k),

h(k)dN (k).

Similar to the presentation in this section, most derivations of trace

formulae on graphs concentrated on non-Robin vertex conditions [456,
467, 499, 630, 701]. When a Robin component is present, significant
difficulties appear, such as negative eigenvalues, failure of bound (3.7.4)
etc. These difficulties were overcome by Bolte and Endres in [127] at
the expense of significant restrictions on the test function (analytic in
 3.8. FURTHER REMARKS AND REFERENCES 101

a strip around the real line) and the graph itself (a condition on the
minimal edge length). We refer the interested reader to [126] for a
survey of the history of the question and difficulties involved.

3.8. Further remarks and references

Basic spectral properties
Discreteness of the spectrum is standard and works the same way
as for elliptic operators on compact manifolds. The consideration here
goes back to the work [343] by Gerasimenko and Pavlov.
The dependence on the vertex conditions is taken from [104].
Hadamard’s variational formulas are useful within various areas of
analysis [331,332,381,389,427,600,648]. We have just adopted these
formulas for the rather simple case of quantum graphs.
Generic simplicity of the spectrum
Statements on “generic” properties of the spectrum and eigenfunc-
tions are well known [19, 691] and their proofs are usually based on
applying some transversality theorems (see [3, 135, 504]). The result
of Theorem 3.1.7 comes from [319]. Regrettably, this result has not
been transferred to the case of general vertex conditions.
Eigenvalue bracketing
The results use the quadratic form only, so they will apply to most
general self-adjoint operators, in particular H = −d2 /dx2 +V (x). Some
of the results parallel the discussion in [644], although we use different
methods of proof which allowed, in some cases, for significant general-
izations.
Dependence on the coupling constant at one vertex
The results presented here are from [104]. Using the technique of
Schatten-von Neumann operator ideals and regularized deter-
minants [246, 360, 663], one can bootstrap the statements on ana-
lyticity of various sets to being the sets of zeros of entire functions of
finite exponential order.
Shnol’ type theorems The Shnol’ theorem of this section was
proved in [481]. See also [137] for a more general Shnol’ type theorem.
Such results are usually stated with point-wise bounds |φ(x)| ≤
C e|x| instead of the integral one (3.2.2) (see [216,348,653]). The two
formulations are equivalent when the volume of the ball Br of radius
r grows polynomially (or at least sub-exponentially). They, however,
are not equivalent otherwise. The statement with point-wise estimates
does not transfer to negatively curved manifolds (it fails, for instance,
102 3. SPECTRA OF QUANTUM GRAPHS

for the hyperbolic plane, where φ = 1 is a bounded generalized eigen-

function for λ = 0, while 0 is not in the spectrum of the correspond-
ing Laplace-Beltrami operator). Similarly, it does not work for many
graphs, e.g. trees. The integral formulation presented in Theorem 3.2.2
and its straightforward reformulation for a manifold case do not have
this restriction.
It is interesting also to notice that the standard proofs (e.g., [348])
in fact use only the integral, rather than point-wise, estimates, al-
though the latter are usually required in the theorem formulations.
This apparently reflects the “Rn mentality”: in the Euclidean setting
the two are equivalent. If one works on Cayley graphs of infinite dis-
crete groups that are not virtually nilpotent and are not of intermediate
(sub-exponential) growth, then the point-wise version fails, while the
integral one survives.
Generalized eigenfunctions
Generalized eigenfunction expansions for PDEs are described in de-
tail in the classical book [88]. One can find a simple introduction to
this topic in an Appendix in [89]. For the quantum graph case, the
corresponding results were obtained in [137, 413, 508, 509].
Notice that in order to use the Theorem of this section in a useful
manner, one needs to find the weight function w(x) as slowly increasing
as possible. Then this will give a meaningful a priori estimate on the
generalized eigenfunctions into which one can expand.
Failure of unique continuation
It has been known for quite some time, both for combinatorial and
quantum graphs, that the unique continuation property, common for
ODEs and 2nd order elliptic PDEs [549], often fails for graphs. Since
this is the property that prevents existence of compactly supported so-
lutions, one can expect such solutions to appear sometimes. And indeed
one can easily construct examples where such effects do occur. In terms
of spectral theory, one can encounter bound states (L2 -eigenfunctions)
in situations where they never appear in the full-dimensional case. For
instance, 2nd order periodic elliptic Schrödinger operators on a quan-
tum graph can have point spectrum (see, e.g., the spectral structure
of graphene and carbon nanotubes described in Chapter 7), while 2nd
order periodic elliptic operators in Rn have absolutely continuous spec-
trum [118, 665, 666, 672, 688].
Since this effect is well known, we are not sure to whom attribute
the folklore result of Proposition in Section 3.4.
DtN maps
The DtN map is a darling of inverse problems theory [693,695]. It
also appears in many guises in various other areas of mathematics, e.g.
 3.8. FURTHER REMARKS AND REFERENCES 103

as Schur complement in matrix theory. In graph terms, it describes

the boundary response of a (combinatorial or quantum) graph. Even
the examples of a single edge with two “boundary” vertices and of a
graph with a one-point boundary, are very useful, as the next Section
and further chapters show. It would be nice to be able to understand
better the properties of the DtN map for rather general graphs. This
would, for instance, help one create quantum graph circuits with pre-
scribed scattering responses [542, 595, 597], or open resonant gaps in
the spectrum of a network (see Section 5.1).
Condition (3.5.11) for k 2 to be an eigenvalue first appeared in [467,
Sec. III.A] under the name “the vertex secular equation” and was given
only for the δ-type vertex conditions. A more general version for the
d2
operator − dx 2 was derived in a somewhat roundabout fashion in [401].

The version for the Schrödinger operator (non-zero potential) was de-
rived in [402]. This representation of the secular equation turns out to
be very useful for studying the spectral determinant and the vacuum
energy (cf. Theorem 7.6.5).
A first estimate of the number of negative eigenvalues of a graph
appeared in [463]. The exact characterization that we gave in Theo-
rem 3.5.3 is due to [62]. We note that Theorem 3.5.3 applies to graphs
with leads as well, with some appropriate modifications, see [62].
Discrete-quantum graph relation
This extremely helpful relation between quantum and combinatorial
graph operators is well known and has been exploited many times (e.g.,
in [21,39,40,177,264,480–482,489,490,588] and references therein).
It is our first application of the DtN map technique. Theorem 3.6.1
was proved in [588]. A more limited version can be found in the earlier
work [177].
Trace formulas
Roth [630] was first to discuss the trace formula for metric graphs.
Kottos and Smilansky [466,467] rediscovered it in the context of quan-
tum chaos on graphs. Kurasov and Nowaczyk [499] observed a problem
with the multiplicity of zero in the trace formula. Kostrykin, Potthoff
and Schrader [456] considered the heat trace on graphs with both fi-
nite and infinite edges. Winn [701] obtained conditional convergence
for sufficiently fast decaying integrable test functions (the decay should
be such that xh(x) is in L1 ). Finally, Bolte and Endres [126,127] over-
came the restriction to non-Robin vertex conditions and extended the
treatment to k-dependent bond scattering matrices.
The elegant proof of the Weyl’s law (Lemma 3.7.4) we gave in sec-
tion 3.7.2 first appeared in [367].
104 3. SPECTRA OF QUANTUM GRAPHS

An interesting finite sum formula can be found in the recent paper

by Band, Harrison, and Joyner [48]. See also [17].
 CHAPTER 4

Spectra of Periodic Graphs

In many applications (for instance, to Carbon nano-structures, see

Section 7.6.3), one deals with a periodic infinite combinatorial or quan-
tum graph Γ. An example of a periodic graph is shown in Fig. 1, where
the dashed arrows indicate two basis vectors of periods. The part W
inside the thin dotted square (a fundamental domain), when shifted
by integer combinations of the two basis vectors, reconstructs the whole
graph.

Figure 1. A Z2 -periodic graph with the basic period

vectors shown by dashed arrows. On of the possible
choices of the fundamental domain W is surrounded
by the dotted line.

The so-called Floquet theory of periodic ordinary differential

equations provides a simple but powerful tool for studying such equa-
tions. Analogs of this theory have been developed also for PDEs,
mostly due to the needs of solid state physics and the recent theory of
photonic crystals. As will be shown in this chapter, the Floquet tech-
nique works for graph operators as well, which turns out to be useful
in many applications, e.g. to photonic crystals, graphene, and carbon
non-tubes (see Chapter 7).

4.1. Periodic graphs

Let us start making the necessary notions precise.
105
106 4. PERIODIC GRAPHS

Definition 4.1.1. An infinite combinatorial, metric, or quan-

tum graph Γ is said to be periodic (or Zn -periodic) if Γ is
equipped with an action of the free abelian group G = Zn , i.e. a map-
ping (g, x) ∈ G × Γ → gx ∈ Γ, such that the following properties are
satisfied:

(1) Group action:

For any g ∈ G, the mapping x → gx is a bijection of Γ;
0x = x for any x ∈ Γ, where 0 ∈ G = Zn is the neutral
element;
(g1 g2 )x = g1 (g2 x) for any gi ∈ G, x ∈ Γ.
(2) Continuity: For any g ∈ G, the mapping x → gx of Γ into
itself is continuous.
(3) Faithfulness: If gx = x for some x ∈ Γ, then g = 0.
(4) Discreteness: For any x ∈ Γ, there is a neighborhood U of x
such that gx ∈ U for g = 0.
(5) Co-compactness: The space of orbits Γ/G is compact. In
other words, the whole graph can be obtained by the G-shifts
of a compact subset.
(6) Structure preservation:
• gu ∼ gv if and only if u ∼ v. In particular, G acts
bijectively on the set of edges.
• In the case of a metric or quantum graph, the action pre-
serves the edges’ lengths: Lge = Le .
• In the case of a quantum graph, the action commutes
with the Hamiltonian H (and in particular, preserves the
vertex conditions).

A simple geometric model is of a graph Γ embedded into Rn in such

a way that it is invariant with respect to the shifts by integer vectors
g ∈ Zn ⊂ Rn , which produce an action of Zn of Γ. In fact, it is not
hard to conclude that when n ≥ 3, one can always achieve such an
embedding (allowing curved edges), so the reader can assume without
any loss of generality that Γ is embedded into Rn and the action is by
integer shifts in Rn . If n = 2, edges of a graph might have to cross when
it is embedded into R2 , and if n = 1, severe overlapping is imminent.
However, if n = 1 or 2, one can embed Γ into R3 so that it is periodic
with respect to the shifts by vectors belonging to an isomorphic image
of the lattice Z or Z2 .

Definition 4.1.2. Due to the co-compactness condition, there ex-

ists a compact (in the discrete case, finite) part W of Γ such that
 4.2. FLOQUET-BLOCH THEORY 107

• The union of all G-shifts of W covers the whole Γ:

%
gW = Γ.
g∈G

• Different shifted copies of W , i.e. g1 W and g2 W with g1 =

g2 ∈ G, have only finitely many points in common, none of
which are vertices.
A compact subset W with these properties is a fundamental domain
for the action of G on Γ. Note that a fundamental domain W is not
uniquely defined.
An example of a fundamental domain is shown in Fig. 1. Another
illustration of these notions is given in Figure 2.

4.2. Floquet-Bloch theory

We will need to use the so called dual lattice G∗ to G = Zn , which
is G∗ := 2πZn , the set of all n-dimensional vectors whose coordinates
are integer multiples of 2π. It can be also described as consisting of
all vectors k such that g · k ∈ 2πZ for any g ∈ G, which allows for
generalization to lattices other than Zn .
Floquet theory can be thought of as a version of Fourier series ex-
pansion. One thus is interested in harmonics into which the expansion
is done. These are the so called characters of the group G of periods:
Definition 4.2.1. A character of the group G is a homomorphism
γ : G → C \ {0}, where the set C \ {0} of non-zero complex numbers is
considered as a group with respect to multiplication. In other words,
γ satisfies the following conditions:
γ(e) = 1, where e is the unit in G,
γ(g1 g2 ) = γ(g1 )γ(g2 ) for any elements g1 , g2 of G.
A unitary character is a character that maps G into the unit circle
S := {z ∈ C | |z| = 1}.
The following statement is well known (and easy to prove):
Lemma 4.2.2.
(1) Every character of G = Zn can be represented by a vector
k ∈ Cn as follows:
(4.2.1) γ(g) = γk (g) := eik·g , g ∈ Zn ,

where k · g = j kj gj .
(2) The character in (4.2.1) is unitary if and only if k ∈ Rn .
108 4. PERIODIC GRAPHS

Figure 2. A Z2 -periodic graph with vertices indicated

by small black disks (other crossings of the edges are ar-
tifacts of the planar embedding). A fundamental domain
W is surrounding by the dotted line. The two curved ar-
rows show the shift of a vertex x by the generators (1, 0)
and (0, 1) of Z2 .

(3) Characters (unitary characters) provide all irreducible (unitary

irreducible) representations of Zn .

Definition 4.2.3. In physics, vectors k in (4.2.1) are called quasi-

momenta.

One notices that the character γk depends on k in a G∗ -periodic

manner. Thus, considering only real quasi-momenta k, one can restrict
the values of k to any fundamental domain B of the action of G∗ =
2πZn on Rn by shifts, since using points k outside B does not increase
 4.2. FLOQUET-BLOCH THEORY 109

the set of characters. A particular choice of B is the cube

B = [−π, π]n .
Definition 4.2.4. We will use for B the physics name Brillouin
zone. In physics it is reserved to a rather specific fundamental domain
[29], although this difference will not matter for our purpose.
In some cases, it is convenient to explicitly factor out the G∗ pe-
riodicity and consider instead of vectors k ∈ Cn the complex vectors
with non-zero components
(4.2.2) z := eik := (eik1 , . . . , eikn ) ∈ (C \ {0})n .
Definition 4.2.5. We will call the vectors z in (4.2.2) Floquet
multipliers (the name comes from the Floquet theory for ODEs, where
n = 1).
When the quasi-momentum k is real, the corresponding Floquet
multiplier belongs to the unit torus
(4.2.3) Tn = {z ∈ Cn | |zj | = 1, j = 1, ..., n} ⊂ Cn .

In terms of Floquet multipliers, the characters can be rewritten as

follows:
(4.2.4) γz (g) = z g := z1g1 z2g2 . . . zngn ∈ C.

4.2.1. Floquet transform on combinatorial periodic graphs.

The standard idea of harmonic analysis suggests that, as long as we
are dealing with a linear problem that commutes with an action of the
abelian group Zn , the Fourier series expansion (i.e., an expansion into
irreducible representations) with respect to this group should simplify
the problem. Its implementation leads more or less straightforwardly
to the so called Floquet transform. To describe this in detail, let us
start with the combinatorial graph case.
Let Γ be a Zn -periodic graph and f be a compactly supported (or
sufficiently fast decaying) function on the set of vertices V of Γ.
Definition 4.2.6. We define Floquet transform of f as
(4.2.5) f (v) → fˆ(v, z) = f (gv)z −g ,
g∈Zn

where gv denotes the action of g ∈ Zn on the vertex v ∈ V and z =

(z1 , ..., zn ) ∈ (C\0)n is Floquet multiplier.
110 4. PERIODIC GRAPHS

If one uses quasi-momenta instead of Floquet multipliers, (4.2.5)

translates into
(4.2.6) f (v) → fˆ(v, eik ) = f (gv)e−ik·g .
g∈Zn

The reader can notice that (4.2.5)-(4.2.6) is just the Fourier trans-
form with respect to the action of G = Zn on the set of vertices V.
Let us formulate some basic properties of the Floquet transform.
The following statement follows by direct inspection of (4.2.5)-(4.2.6):
Lemma 4.2.7. The following identities hold:
(4.2.7)
!
fˆ(gv, z) = z g fˆ(v, z), equivalently, fˆ(gv, eik ) = eik·g fˆ(v, eik )
(4.2.8) fˆ(v, ei(k+γ) ) = fˆ(v, eik ) for γ ∈ G∗ = 2πZn .
Equalities (4.2.7) of the lemma show that, as one would expect,
G-shifts after Floquet transform become multiplications by the corre-
sponding characters. To put it differently, for a fixed z, the function
fˆ(v, z) on Γ is automorphic with the character z g = eik·g . It also
shows that the values of fˆ(v, z) are determined completely if they are
known for v ∈ W only, since then they extend to the whole graph
according to (4.2.7). We thus introduce the following notation:
Definition 4.2.8. Let W be a (finite) fundamental domain of the
action of the group G = Zn on Γ. We will denote fˆ(v, z)|v∈W by fˆ(z),
where the latter expression is considered as a function of z with values
in the space of functions on W . In other words, fˆ(z) takes values in
C|W | .
We also see that the identity (4.2.8) says that the Floquet trans-
formed function is Γ∗ -periodic function of the quasimomentum k.
Let us consider a simple example. Consider the periodic graph
shown in Fig. 1 and a function φ that is zero at all vertices except
the vertices a through h. We consider the horizontal dashed arrow as
the first basis shift vector in Z2 and the vertical one as the second.
The space C|W | = C4 can be identified with the space of vectors la-
beled by the vertices a, b, c, d: (φ(a), φ(b), φ(c), φ(d))t ∈ C4 . Let us also
introduce the shorthand notation
⎛ ⎞ ⎛ ⎞
α φ(a)
⎜ β ⎟ ⎜ ⎟
⎜ ⎟ := ⎜ φ(b) ⎟ .
⎝ γ ⎠ ⎝ φ(c) ⎠
δ φ(d)
 4.2. FLOQUET-BLOCH THEORY 111

- z) is
Then the function φ(v,
(4.2.9)
⎛ ⎞
φ(a) + z2 φ(h)
⎜ ⎟
- = ⎜ φ(b) + z−1
φ(z) 1 φ(e) ⎟
⎝φ(c) + z2 φ(f )⎠
φ(d) + z1−1 φ(g)
⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞
α φ(h) 0 0 0
⎜β ⎟ ⎜ 0 ⎟ ⎜φ(e)⎟ 1 ⎜ 0 ⎟ 1 ⎜ 0 ⎟
=⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟ ⎜ ⎟
⎝ γ ⎠ + z2 ⎝ 0 ⎠ + z1 ⎝ 0 ⎠ + z2 ⎝φ(f )⎠ + z1 ⎝ 0 ⎠ .
δ 0 0 0 φ(g)
Remark 4.2.9. The example we have just considered also reveals
the following interpretation of the Floquet transform: one takes a func-
tion φ on Γ and cuts its into non-overlapping pieces by restricting to
the shifted copies gW of a fundamental domain W . These pieces are
shifted back to W and then are taken as (vector valued) Fourier coeffi-
cients of the Fourier series (4.2.6) that defines the Floquet transform.
This remark enables one to invert the Floquet transform easily.
This inversion can be done in two ways: either using f-(v, eik ) for all
vertices v, or just for those in a fundamental domain W .
Theorem 4.2.10. The following two inversion formulas hold:
(1)
1
(4.2.10) f (v) = f-(v, eik )dk
(2π)n
B

for any v ∈ V, or
(2)
1
(4.2.11) f (gv) = f-(v, eik )eik·g dk
(2π)n
B

for any v ∈ W and g ∈ G.

The proof of this theorem follows from the application of the stan-
dard formulas for the Fourier coefficients of a periodic (vector valued)
function.
We now need to establish how the Floquet transform acts on some
natural function spaces. The next elementary theorem, which we in-
troduce without a proof, provides Plancherel and Paley-Wiener type
results for the Floquet transform. We thus introduce first the space
D(Γ) of all finitely supported functions on Γ.
112 4. PERIODIC GRAPHS

Theorem 4.2.11.
(1) The transform u(v) → (2π)−n/2 û(z) is an isometry between
L2 (Γ) and L2 (Tn , C|W | ), where the latter is the space of L2 -
functions on the unit torus Tn ⊂ Cn with values in C|W | .
(2) The image under the Floquet transform (4.2.5) of the space
D(Γ) coincides with the space of all finite sums
(4.2.12) a(z) = ag z g ,
g∈Zn

where ag ∈ C|W | . We will call such sums Laurent polyno-

mials1.
One can notice that the first statement of the theorem follows from
the standard Plancherel theorem for Fourier series, applied to vector
valued functions. Indeed, (4.2.6) is a standard Fourier series, if its
coefficients are considered as vectors in the finite dimensional space
L2 (W ) ∼
= C|W | . The second statement is straightforward.

4.2.2. Floquet transform of periodic difference operators.

Let A be a periodic difference operator on a graph Γ equipped with a
free action of the group Zn . We assume the action to be co-finite, i.e.
there exists a fundamental domain with finitely many vertices. We will
show how after Floquet transform this operator becomes an operator of
multiplication by a |W | × |W | - matrix A(z), which depends rationally
on the Floquet multiplier z (or analytically on the quasi-momentum
k).
As an example, we consider the Laplace operator on the regular
hexagonal 2D lattice2 Γ (see Fig. 3): The group Z2 acts on Γ by
√ vectors p1 e1 +√p2 e2 , where (p1 , p2 ) ∈ Z and vectors
2
the shifts by
e1 = (3/2, 3/2), e2 = (0, 3) are shown in Figure 3. We choose
as a fundamental domain (Wigner-Seitz cell) of this action the shaded
parallelogram region W . Two black vertices a and b belong to W , while
b , b and b belong to some shifted copies of W . Three edges f, g, h,
directed as shown in the picture, belong to W .
We consider the Laplace operator
Af (v) = f (w) − 3f (v).
w∼v

1An example of such a Laurent polynomial is given in (4.2.9). Notice that the
singularities that can appear are located on the union of coordinate hyperplanes
only, i.e. when at least one of the coordinates of z is equal to zero.
2we will see this lattice again Section 7.6.3 devoted to carbon nano-structures
 4.2. FLOQUET-BLOCH THEORY 113

b
f
 Γ
b a
h
e2
b
W
e1
Γ b
Figure 3. The hexagonal lattice Γ and a fundamental
domain W together with its set of vertices V (W ) = {a, b}
and set of edges E(W ) = {f, g, h}.

The way one finds the “symbol” A(z), is to apply A to functions f

automorphic with the character z = eik . Such a function is uniquely
determined by its values at the vertices a and b. Indeed, since b is
obtained from b by shifting once opposite to the basis vector e1 , one
has f (b ) = z1−1 f (b). Analogously, f (b ) = z2−1 f (b). One also finds
that the values at two neighbors of the vertex b are z1 f (a) and z2 f (a).
Hence,
(Af )(a) = −3f (a) + (z1−1 + z2−1 + 1)f (b),
(Af )(b) = (z1 + z2 + 1)f (a) − 3f (b).
We thus obtain the expression for the “symbol” of A:
 
−3 z1−1 + z2−1 + 1
A(z) = ,
z1 + z2 + 1 −3
or, in terms of the quasi-momenta,
 
−3 e−ik1 + e−ik2 + 1
A(k) = ik1
e + eik2 + 1 −3

4.2.3. Floquet transform on quantum periodic graphs. The

Floquet transform on metric and quantum graphs, at least in the set-
ting of non-smooth (e.g. L2 ) functions, is very much similar to the one
on discrete ones. Namely, one has the natural analog of (4.2.5) and
114 4. PERIODIC GRAPHS

(4.2.6):

(4.2.13) f → fˆ(x, z) = f (gx)z −g , x ∈ Γ,

g∈Zn

(4.2.14) f → fˆ(x, eik ) = f (gx)e−ik·g , x ∈ Γ.

g∈Zn

The same invariance properties as (4.2.7)–(4.2.8) clearly hold. This,

in particular, allows one to use only the points x from a fundamental
domain W (which is assumed to be chosen in such a way that it does not
contain any vertices on its boundary, see Fig. 1). Similarly, the values
of the quasimomentum from a fixed Brillouin zone B are sufficient.
In the classes of smooth functions, however, a new intricacy arises,
requiring one to use sections of vector bundles rather than vector-valued
functions as in the L2 situation above. This issue is addressed in the
next section.

4.2.4. Floquet transform of periodic operators. The Floquet

transform of periodic quantum graph operators works similarly to the
discrete case, which we discussed in Section 4.2.2. Namely, the Hamil-
tonian needs to be applied to functions automorphic with a given mul-
tiplier z (quasi-momentum k), which reduces H to a family of operators
H(z) (respectively, H(k)) on a chosen fundamental domain W . It is
convenient in the quantum graph case to pick a fundamental domain
W that does not contain any vertices on its boundary. Such a choice
is always possible. New vertices of degree one automatically appear on
the intersection of the boundary of W with the edges of the graph.
Suppose, for instance, that we are dealing with the graph shown in
d2
Fig. 1 and the Hamiltonian dx 2 + V (x) with some appropriate vertex

conditions. Choosing W as shown in Fig. 1, one obtains the Hamilton-

ian H(z) = H(z1 , z2 ), that acts on functions defined on W the same
way H did, but its domain is now restricted by the boundary condi-
tions that say that when going from the vertex s− +
j to sj (see Fig. 4),
the value of the function and its derivative multiply by z1 = eik1 . The
derivative needs to be adjusted for direction:
−
(4.2.15) f (s+ ikj
j ) = e f (sj ),

(4.2.16) f  (s+ ikj  −

j ) = −e f (sj ).

These conditions are sometimes called quasiperiodic.

 4.3. BAND-GAP STRUCTURE OF SPECTRUM 115

s-2
a

s+
1 d b s-1

c
s+
2

Figure 4. A fundamental domain W .

4.3. Band-gap structure of spectrum

Periodic operators on combinatorial and quantum graphs display
the same important feature of their spectra as the periodic elliptic coun-
terparts of mathematical physics do: the so called band-gap struc-
ture. Namely, there is a natural way to represent the spectrum as
the (finite in the discrete case and infinite in the quantum one) union
of finite closed intervals, called spectral bands, sometimes allowing
spectral gaps to appear between them.

4.3.1. Discrete case. Let Γ be a Zn -periodic (see Definition 4.1.1)

combinatorial graph and A be a self-adjoint finite order difference op-
erator (Definition 1.2.5) on Γ commuting with the Zn -action.
According to Section 4.2.4, after the Floquet transform, A becomes
the operator of multiplication in L2 (Tn , C|W | ) by a rational |W | × |W |
self-adjoint matrix function A(z).
Let λ1 (z) ≤ λ2 (z) ≤ · · · ≤ λ|W | (z) be the eigenvalues of A(z).
Perturbation theory shows (e.g., [436]) that these eigenvalues are con-
tinuous and piece-wise analytic with respect to z. One notices that
σ(A) = σ(A(z)), which leads to the following standard result:
z

Theorem 4.3.1. The spectrum σ(A) of the operator A in L2 (Γ)

is equal to the union of the closed finite intervals called the spectral
bands
Ij = {λ : λ = λj (z) for some z ∈ Tn }.

It is possible (and indeed happens often) that the spectral bands

leave some intervals between them not covered, which correspondingly
are called spectral gaps. This explains the following customary no-
tion:
116 4. PERIODIC GRAPHS

Definition 4.3.2. The representation

|W |
%
(4.3.1) σ(A) = Ij
j=1

is called the band-gap structure of σ(A).

One can imagine the situation graphically as follows: Consider the
graph of the multiple-valued function z → σ(A(z)) (Fig. 5).

λ4
I4
λ3
I3
I2
λ2
σ(A)

I1
λ1

π 2π k

Figure 5. The dispersion curve and the spectrum of a

periodic operator A.

This function is called the dispersion relation in physics, and its

graph is called either the dispersion curve or the Bloch variety of
the operator A. Now Theorem 4.3.1 can be re-formulated as saying
that the spectrum σ(A) is the range of the dispersion relation. One
sees a close resemblance to (and a generalization of) the spectrum of
the operator of multiplication by a scalar valued function.

4.3.2. Quantum graph case. In the quantum graph case with a

periodic Hamiltonian H, after Floquet transform an analytic family of
operators H(z) arises. This time, though, they are infinite dimensional
self-adjoint operators rather than finite matrices, like in the preceding
section. Being elliptic operators on the compact graph Γ/Zn , they
have discrete spectra in L2 (Γ) (see Theorem 3.1.1). They also act
as self-adjoint Fredholm operators between appropriate Sobolev spaces
(Theorem 2.4.3). Thus, a similar construction of the dispersion relation
 4.4. ABSENCE OF THE SINGULAR CONTINUOUS SPECTRUM 117

is possible and the band-gap structure of the spectrum

%
σ(H) = Ij
j

still applies. Here, however, the operator H is unbounded, and thus

infinitely many spectral bands Ij = [aj , bj ] arise, with lim aj = ∞.
j→∞

4.3.3. Floquet transform in Sobolev classes. In the quantum

graph case, as in the PDE situation [474, Ch. 2], action of the Floquet
transform in Sobolev function classes requires more careful considera-
tion than the L2 -case or the discrete graph situation.
Let is choose a fundamental domain W of the periodic graph Γ,
such that its boundary does not contain any graph vertices (see Defini-
tion 4.1.2). If one deals with the functions of the smooth Sobolev class
$
H m (Γ), then one can see that, although the Floquet transform will

produce a function of the class L2 (Tn , H m (W )), the mapping between

these spaces will not be surjective. Indeed, the inverse transforms takes
the Fourier coefficients, which belong to H $ m (W ) and tiles with them

the whole graph. One notices that at the boundary of adjacent copies
of the fundamental domain W there is no enforcement of smoothness.
This is resolved as follows. Let k ∈ B be a quasi-momentum, We define
the closed subspace H $ m $ m
k (W ) of H (W ) that consists of restrictions to
W of all k-automorphic functions from H $ m
loc (Γ). The same way it was
done in [474, 475], one can prove that the union over k ∈ Tn of these
sub-spaces forms an analytic Hermitian vector bundle H $ m . Then, the

Floquet image of H $ m (Γ) consists of all L -sections of this bundle.

2

4.4. Absence of the singular continuous spectrum

The representation of the operator as the operator of multiplication
by an analytic operator-function H(z) and the consequent expansion
into operators of multiplication by scalar piecewise analytic functions
λj (z) allow one to establish the following result, rather standard in
studies of periodic problems:
Theorem 4.4.1. Let H be a self-adjoint periodic Hamiltonian on
Γ, then its spectrum has no singular continuous part: σsc (H) = ∅.
This conclusion follows from the simple fact that the spectral mea-
sure of the operator of multiplication by a real, continuous, piecewise
analytic function a(z) without constant pieces is absolutely continuous
with respect to Lebesgue measure on R. The flat pieces, on the other
hand, lead to the point spectrum, but not to the singular continuous
118 4. PERIODIC GRAPHS

one (see, e.g., [341,688], or the proof of Theorem 4.5.9 in [474] for the
proof of these statements).

4.5. The point spectrum

According to Theorem 4.4.1, the spectrum of a periodic difference
operator on a combinatorial graph or of a periodic elliptic (i.e., with
non-vanishing coefficient at the highest order term) differential oper-
ator on a quantum graph can contain only absolute continuous and
pure point parts. It has been conjectured that the spectrum of any
self-adjoint elliptic second order periodic differential operator in Rn is
purely absolutely continuous and thus the point spectrum is empty.
To put it differently, such operators do not have any bound states (i.e.,
true L2 -eigenfunctions). In the latter formulation, the statement does
not require self-adjointness. This conjecture has been proven by now
in a wide generality (see the references in Section 3.8). However, the
point spectrum can emerge for periodic elliptic operators of higher order
(see [474]), due to the failure of the uniqueness of continuation for such
equations. Since, as we have seen in Section 3.4, uniqueness of contin-
uation also does not hold on (combinatorial and quantum) graphs, one
should not be surprised that for periodic operators on graphs not only
bound states, but even compactly supported eigenfunctions (“scars”)
might exist.
We now give another example of a compactly supported eigenfunc-
tion on a discrete graph and then explain that appearance of point
spectrum for periodic graph operators can only be due to such states.
Consider the graph in Figure 6 and the function f on it, whose
values are indicated at the vertices. Let us now periodically repeat this
1

0 0 0 0

-1

Figure 6.

graph to the left and right, joining the copies at the dangling left and
right vertices and extending the function f by zero. Then the extended
graph Γ is Z-periodic. The extended function f is clearly compactly
supported and satisfies the eigenvalue equation
LΓ f = 2f.
 4.5. THE POINT SPECTRUM 119

Thus, 2 ∈ σp (LΓ ), where LΓ is the combinatorial Laplacian on Γ (see

(1.2.1)), which is clearly periodic.
We see that one possible reason for appearance of the point spec-
trum on periodic graphs is that, due to the failure of the unique contin-
uation principle, compactly supported eigenfunctions might arise. We
will show now that this in fact is the only reason: if λ ∈ σp (L), then
the corresponding eigenspace in L2 (Γ) is the closed linear span of the
compactly supported eigenfunctions it contains.

Remark 4.5.1. To avoid possible confusion we stress that there are

eigenfunctions with non-compact support (they are easy to construct by
shifting and combining the compactly supported ones), but the ones with
compact supports form a dense linear subspace.

We prove this statement for the combinatorial case first. Let Γ

be a Zn -periodic combinatorial graph and A be a finite order finite
difference operator (not necessarily self-adjoint). We assume that A is
G-periodic, i.e. commutes with Zn -shifts.

Theorem 4.5.2. If the equation Au = 0 has a non-zero solution

in l2 (Γ), then it has a non-zero compactly supported solution as well.
Moreover, the compactly supported solutions form a complete set in the
space of all l2 -solutions.

Proof. After the Floquet transform, A becomes the operator of

multiplication in L2 (Tn , C|W | ) by a rational |W | × |W | matrix function
A(z) analytic in (C \ {0})n . Thus, non-zero l2 -solutions of Au = 0
are in one-to-one correspondence with C|W | -valued L2 -functions û on
Tn such that A(z)û(z) = 0 a.e. on Tn . Since we assumed that u is a
non-zero element of l2 , we conclude that the set of points of the torus
Tn over which the matrix A(z) has a non-trivial kernel, has a positive
measure. On the other hand, this is an algebraic subset in Cn , given
by the equation det A(z) = 0. The only case it can intersect the torus
over a subset of a positive measure is that it coincides with the whole
space (C \ {0})n . Hence, A(z) has a non-zero kernel at any point z ∈
(C \ {0})n and its determinant is identically equal to zero. Considering
this matrix over an appropriate field Q of rational functions, one can
apply the standard linear algebra statement that claims existence of a
non-zero rational solution φ(z) of A(z)φ(z) = 0. As indicated before,
such functions before the Floquet transform were compactly supported
solutions of Au = 0. This proves the first statement of the theorem,
about the existence of compactly supported eigenfunctions.
120 4. PERIODIC GRAPHS

To prove completeness, let Q1 (z), . . . , Qr (z) be a finite set of the

generators of all non-zero polynomial (vector-valued) solutions of
A(z)Q(z) = 0
(it is known to exist, e.g. [417, Lemma 7.6.3, Ch.VII]). Floquet trans-
form reduces the completeness statement we need to prove to the fol-
lowing
Lemma 4.5.3. Combinations
(4.5.1) y(z) = aj (z)Qj (z),
j=1,..,r

where aj (z) are finite Laurent sums, are L2 -dense in the space of all
C|W | -valued L2 -solutions of the equation
(4.5.2) A(z)y(z) = 0.
Proof of the lemma. First of all, any L2 (Tn )-function aj can be
approximated by a finite Laurent sum of powers of z. Indeed, this is
done by taking finite partial sums of the Fourier series of aj on the torus
Tn . So, it is sufficient to approximate any L2 -solution y(z) of (4.5.2)
by sums (4.5.1) with L2 coefficients aj . Let k > 0 be the minimal (over
z ∈ Cn or z ∈ Tn , which is the same) dimension of Ker A(z). The set
B ⊂ Tn of points z where dim Ker A(z) > k is an algebraic variety of
codimension at least 2, and thus has zero measure on Tn . Hence, it is
sufficient to do L2 approximation outside of small neighborhoods of B.
Let z0 ∈ Tn \B and U be a sufficiently small neighborhood of z0 not
intersecting B. Then over (a complex neighborhood of) U the kernels
Ker A(z) form a trivial holomorphic vector bundle [730]. Let fl (z) be
a basis of holomorphic sectionsof this bundle. Then the portion of y
over U can be represented as bl (z)fl (z) with L2 -functions bl . Now,
one uses [417, Lemma 7.6.3, Ch. VII] again to see that sums (4.5.1)
with analytic aj approximate the sections fl . This proves the Lemma
and hence the Theorem. 
Now the case of quantum graphs (at least when the Dirichlet spec-
trum is excluded) can be reduced to the combinatorial one, similarly
to the way described in Section 3.6 (see also [480]).
Theorem 4.5.4. Let Γ be a Zn -periodic (in the meaning already
specified) quantum graph equipped with the second derivative Hamil-
tonian and arbitrary vertex conditions at the vertices. Suppose that λ
does not belong to the Dirichlet spectrum σD (see Definition 3.0.1) of
Γ. Then, existence of a non-zero L2 -eigenfunction corresponding to the
eigenvalue λ implies existence of a compactly supported eigenfunction,
 4.6. WHERE DO THE SPECTRAL EDGES OCCUR? 121

and the set of compactly supported eigenfunctions is complete in the

eigenspace.
Proof. Let F be an L2 -eigenfunction. Since we are away from
the Dirichlet spectrum σD , the standard estimates show that the vec-
tor f = {F (v)} of the vertex values belongs to l2 (V ) if and only if
F ∈ L2 (Γ). Since λ is not in σD , solving the boundary value problem
for the eigenfunction equation HF = λF on each edge separately in
terms of the boundary values of F , we can express the derivatives of
F at each vertex in terms of its vertex values f solely (this is just the
DtN procedure for a single edge, see Section 3.5). Thus, boundary
conditions (which involve the values of F and of its vertex deriva-
tives) lead to a periodic finite order difference equation Af = 0 on
the combinatorial counterpart of the quantum graph (see Section 3.6).
Theorem 4.5.2 claims existence and completeness of combinatorial com-
pactly supported solutions. Reversing the procedure (which is possible
since we are not on the Dirichlet spectrum), we conclude existence and
completeness of compactly supported eigenfunctions of the quantum
graph. 

4.6. Where do the spectral edges occur?

In this section we discuss briefly the following frequently arising
and somewhat controversial question concerning the spectral structure
of periodic operators of mathematical physics:
Question 4.6.1. Are extrema of the dispersion relation always at-
tained on the set of symmetry points of the Brillouin zone? In other
words, is it sufficient to check just the symmetry points in order to find
the edges of the spectrum?
This is well known to be true for periodic 2nd order ODEs (e.g.,
[247, 525, 626, 690]). Namely, in this case the edges of the spectrum
occur at k = 0 or k = π, i.e. at the spectra of the periodic and
anti-periodic problems on the single period. If the similar claim were
correct for PDEs, the computational task of finding the spectrum would
be significantly simplified, due to reduced dimension. An experimental
observation is that in most practical cases this is indeed correct. One
frequently encounters the belief that this is always true (albeit no jus-
tification is ever provided). In any case, computations (e.g., for finding
possible spectral gaps) are often reduced to the symmetry points only.
This does not seem to lead to errors in practice, although it is never
clear why. However, as shown in [403], the answer to the question is
in fact negative. It is also a simple remark that the counterexamples
122 4. PERIODIC GRAPHS

form an open set. We thus are just lucky to live in the world where
the environment is homogeneous enough to make the answer positive.
In particular, the paper [403] contained examples of Z2 -periodic,
discrete and quantum graphs where spectral edges are attained deeply
inside the Brillouin zone, far from any symmetry points. One of the
counter-examples constructed in [403] is the periodic graph with the
fundamental domain shown in Fig. 7, i.e., the full graph Γ is obtained
by tiling the plane with the Z2 -shifted copies of W .
6

1 W

9 4 5 2 7
3

8
Figure 7. The periodic graph Γ with fundamental re-
gion W .

Remarkably, even the case of Z-periodic graphs is qualitatively dif-

ferent from the situation with the ODEs, where the spectral edges
should still be attained at periodic and anti-periodic spectra. It was
shown in [275] that the spectral edges of a Z-periodic graph can still
come from extrema inside the Brillouin zone, unless the graph forms a
chain of graphs connected one to the next by a single edge.
Consider a quantum (or combinatorial) graph that consists of a se-
quence of identical copies {Γj }∞
j=−∞ of a compact graph Γ, with the
consecutive copies Γj and Γj+1 being connected by m edges, thus form-
ing a Z-periodic structure (see Fig. 8).

Γj−1 Γj Γj+1

Figure 8. A “chain” Z-periodic graph with 3 joining edges.

Theorem 4.6.2 ([275, 403]). Consider the operator H acting as

−d2 /dx2 on the edges of a Z-periodic chain graph with m connecting
 4.7. EXISTENCE AND LOCATION OF SPECTRAL GAPS 123

edges between the copies of the fundamental domains, with some self-
adjoint vertex conditions. Then, if m = 1, the endpoints of the spectral
bands of σ(H) coincide with the points of the spectra of the periodic
and anti-periodic problems (i.e., the spectra of the Bloch Hamiltonians
H(0) and H(π)). This claim is in general incorrect for m > 1 and for
Zn -periodic graphs with n > 1.

We will not prove the results of this section, since after graph ex-
amples are provided, checking them is the matter of (often tedious)
combinatorial calculations, which one can find in [275, 403]. After
that, the result is translated to the quantum graph situation using
methods of Section 3.6.
The difference in the case when m = 1 is that then the space of
solutions of a Cauchy problem on the chain is, like in the ODE case,
two-dimensional. This, together with time reversal symmetry, forces
the branches of the dispersion relation to have the same monotonicity
as for ODEs, which implies that the extrema are attained exactly for
periodic and anti-periodic problems.

4.7. Existence and location of spectral gaps

The band-gap structure of the spectra of periodic operators allows
spectral gaps to appear, but does not mandate them to do so. In the
case of elliptic periodic operators, it is known that in dimension 1, for
an operator in “general position,” infinitely many gaps appear [661].
It is actually complicated to create examples of operators with a fi-
nite number of gaps. The situation is essentially opposite in higher
dimensions, when even existence of at least one gap cannot be guaran-
teed. The reason for such difference is that the spectral bands can only
touch in 1D, while in higher dimensions they can, and usually do, over-
lap. For instance, periodic elliptic 2nd order operators with coefficients
(e.g., potential in the Schrödinger operator) close to constants, do not
have any spectral gaps. Moreover, the Bethe-Sommerfeld conjecture
(mostly proven by now) claims that the number of spectral gaps in
dimensions higher than one is always finite.
The situation with graphs is similar: Z-periodic graphs with a single
connecting link (see the previous section) behave like 1D operators,
while in other cases of periodic graphs (with more links or with the
group of periods of a higher rank), the situation resembles the higher
dimensional case. This is not a strict law, since we will see that the
technique of Section 5.1 allows to construct examples of Zn -periodic
graphs with infinitely many spectral gaps. Thus, the natural analog of
124 4. PERIODIC GRAPHS

the Bethe-Sommerfeld conjecture for discrete and quantum graphs is

false.
In general, controlling the existence and location of spectral gaps
arising due to periodicity (sometimes called Bragg gaps) is extremely
hard. Since existence, location, number, and size of spectral gaps are
often crucial (e.g. in solid state physics, optics, expander theory, etc.),
it would have been nice to have a method of spectral gap opening that
does not rely upon periodicity and allows a better control. Such a
method, of opening the so called resonant gaps, has been considered
in the last decades, but its “clean” implementation is known probably
only in the graph case. We postpone discussion of this issue till Section
5.1.

4.8. Impurity spectra

The situation often encountered in the solid state physics, photonic
crystals, and waveguide theory is a periodic medium perturbed by a
localized (compactly supported or sufficiently fast decaying) perturba-
tion, an “impurity”. In the graph case, this might mean a local change
of geometry of the graph, local change in potentials, vertex conditions,
etc.
Standard spectral theory theorems show that this perturbation can
introduce only a discrete set of eigenvalues (and thus bound states) of
finite multiplicity. The “general wisdom” (and a vast literature) says
that, if the decay of the impurity is sufficiently fast (definitely if it
is compactly supported) this impurity spectrum arises only in the
gaps of the original absolutely continuous spectrum and does not get
embedded into the a.c. spectrum. This is a very well studied issue for
localized perturbations of a homogeneous background (e.g., [216,248]).
However, in the periodic situation the absence of embedded eigenvalues
is known only for ODEs (see [628,629]). The case of PDEs is extremely
hard and only some weak results are known (see [490]).
As one can expect, the graph case brings some surprises: embedded
eigenvalues can indeed arise, even under a compactly supported pertur-
bation. However, the impurity eigenfunctions are compactly supported,
and their support can not stray too far away from the perturbation
support, see details in [491].

4.9. Further remarks and references

Periodic media, and in particular periodic combinatorial and quan-
tum graphs have attracted a lot of attention. The reason is that such
structures arise frequently in solid state physics, material science, and
 4.9. FURTHER REMARKS AND REFERENCES 125

nano-technology. It is enough to mention the role played by them

in studying semi-conductors [29, 734], photonic crystals [158, 429,
432, 475, 641], and carbon nano-technology and mesoscopic physics,
e.g. [398, 422, 423, 452, 455, 489, 514, 638].
Floquet-Bloch theory
The basics of Floquet theory [308] can be found in the ODE case in
[247,704], for PDEs in [247,474,626,665], see also the classics [141].
The technique transfers without any difficulty to periodic combinatorial
[494] and quantum [481] graphs. Although we follow [481,494] in our
exposition, the technique has been in the mathematical folklore forever.
No singular continuous spectrum
It is well known to the experts that the standard proofs of absence
of singular continuous spectrum, developed for Schrödinger operators
with periodic potentials (e.g., [626, vol. 4] or [474]) work smoothly
for any periodic “elliptic” problem, be it discrete or continuous [341].
Thus, it is hard for us to attribute this result to any particular author.
The differences in the graph case start to appear when one discusses
the absence or presence of the point spectrum.
Point spectrum
Point spectrum (corresponding to bound states) is empty for peri-
odic second order elliptic operators under mild regularity conditions.
However, bound states (and even compactly supported eigenfunctions)
can arise for elliptic periodic operators of orders higher than two [474].
The reason is the same as for graphs — failure of the uniqueness of con-
tinuation. This is why compactly supported states can arise on periodic
quantum graphs. Theorems 4.5.2 and 4.5.4 show that these compactly
supported eigenfunctions are indeed the only reason for appearance of
bound states.
Proving a continuous analog of Theorems 4.5.2 and 4.5.4 for peri-
odic elliptic second order PDEs would immediately prove absence of
point spectrum and thus the absolute continuity of the whole spectrum.
However, the known analogs of these theorems only show that existence
of an L2 -eigenfunction implies the existence of a super-exponentially
decaying eigenfunction [474]. A uniqueness of continuation result at
infinity should prohibit such decay, but unfortunately, the only known
PDE theorems of this type [323,541] (see also [475] for the discussion)
apply only to the stationary Schrödinger operator with electric poten-
tial in dimension not exceeding 4, where the absolutely continuity of
the spectrum has been known for a long time [665].
In the graph setting, an analog of Theorems 4.5.2 and 4.5.4 in
a much more general than periodic case was proven in [696], using
different (and more elaborated) techniques.
126 4. PERIODIC GRAPHS

Where do the spectral edges occur?

The importance of symmetry points of the Brillouin zone has been
known since at least the paper [134]. It is sometimes incorrectly
claimed that [134] implies the positive answer to Question 4.6.1. The
positive answer would be important, for instance, for optimization pro-
cedures, when one needs to run the spectral computation at each iter-
ation [208, 209].
There is no analytic reason for this property to hold in dimensions
higher than 1. Many researchers are aware that numerics produce coun-
terexamples. Surprisingly, such examples are hard to come by and are
usually not very convincing for an analyst (e.g., the error in computing
the spectrum using only the boundary of the Brillouin zone is usually
very small). The experience is that one needs to make the medium
inside the fundamental domain (Wigner-Seitz cell) highly asymmetric
to achieve such examples.
As we have shown, the answer is in general negative. It seems to
be puzzling, however, that the claim that we show to be incorrect in
general, is still correct (or almost correct) so often. In most practical
cases, computations along the boundary of the reduced Brillouin zone
do provide the correct spectrum. An explanation of this effect was
given in [403]. At the first glance, the property of edges of the spec-
trum being attained on the boundary only can be destroyed by small
perturbation, and thus should not be seen often in practice. However,
the symmetry of the problem influences the validity of this argument.
Namely, it is shown in [403] that “generically” (modulo some generic-
ity conjecture), the occurrence of the spectral edges at the symmetry
points is stable under small perturbations that preserve the symme-
try. In other words, there are open sets of “good” and “bad” periodic
operators. This probably explains the frequent appearance of the ef-
fect in practice: we are lucky to reside mostly in the “good” open set.
Even after appearance of [403] sometimes the illusion persisted that
the claim about band edges being attained at the symmetry points
of the Brillouin zone still holds for Z-periodic graphs (since they are
kind of one-dimensional). This was disproved in [275]. The counterex-
amples of these papers were also bootstrapped to the full-dimensional
model by using “fat graph” convergence theorems, discussed in Section
7.5.
We remark that whenever one needs to find the density of states, a
computation over the full Brillouin zone is always required.
Spectral gaps
The number of spectral gaps for 1D periodic operators is generi-
cally infinite [661]. The so called finite-gap potentials are rare, can all
 4.9. FURTHER REMARKS AND REFERENCES 127

be described, and play important role in the inverse scattering method

(e.g., [239, 240]). The situation changes drastically in higher dimen-
sions, when the so called Bethe-Sommerfeld conjecture [112] claims
that the number of spectral gaps should be finite. This is proved by
now in significant generality (see [474, 475, 590, 591, 665] and refer-
ences therein).
Equations on abelian coverings. Liouville theorems
A periodic graph can be represented as a covering, with the deck
group equal to Zn , of a compact graph. Floquet theory is applicable
to periodic equations on such abelian coverings of manifolds or graphs.
There are some questions that are of interest for equations on cover-
ings. For instance, one is of validity of Liouville type theorem for
solutions on the covering3. This question arises when studying random
walks, group amenability, and other issues. More generally, one would
like to understand whether the space of solutions of a prescribed poly-
nomial grows is finite-dimensional, and if yes, what is its dimension,
see e.g. [150, 151, 193, 252, 433, 515–517, 520, 521, 709, 731]. It was
expected that answers could be obtained at least in the case of nilpo-
tent (in particular, abelian) coverings, and on more general manifolds
with some conditions on growth.
An interesting observation was made by Avellaneda and Lin [37]
and Moser and Struwe [559] that in the case of periodic divergence type
elliptic operators on Rn , the answer about the validity the Liouville
theorem can be given and exact dimensions can be found.
It was shown in [487, 488] that more detailed answers can be ob-
tained for “elliptic” periodic problems on abelian coverings of Rie-
mannian manifolds, complex manifolds, and compact combinatorial or
quantum graphs. This conforms with Gromov’s idea that when look-
ing “from afar” on such a covering, one sees only the covering group.
This agrees also with the idea of homogenization in fast oscillating me-
dia [45,81]. We note that papers [37,559] indeed used homogenization
techniques, while [487, 488] did not, which allowed the authors of the
latter to consider validity of Liouville theorem at more general spec-
tral locations than the bottom of the spectrum, as it was the case
in [37, 559].
Maximal abelian coverings - no gap conjecture
An interesting conjecture by Sunada [681, 682] states that on a
maximal abelian covering of a regular (combinatorial) graph, the spec-
trum of the discrete Laplacian has no gaps. This has been proven for

3Liouville type theorems address existence of bounded or polynomially growing

solutions and finite dimensionality of the spaces of such solutions.
128 4. PERIODIC GRAPHS

the case of a regular graph of an even degree [411], and only for some
special cases, when the degree is odd (see details in [682]).
Other studies of periodic graph and “fattened graph” struc-
tures
There are quite a few other studies of periodic graph structures,
which we cannot address in this introductory book. We will just list
some of these works [39, 42, 72, 82, 119, 153, 168, 264, 267, 274, 290,
291, 299, 300, 302–306, 310, 318, 340, 347, 470, 486, 491, 557, 578,
587, 625, 710, 729].
 CHAPTER 5

Spectra of Quantum Graphs. Special Topics

In this chapter, we consider some more special topics from the spec-
tral theory of quantum graphs. Section 5.1 addresses the problem of
existence of gaps in the spectrum of infinite quantum graphs, which
arises in many applications. Section 5.2 is devoted to discussion of
the number of zeros and nodal domains of eigenfunctions of compact
quantum graphs. On a single segment, this question is answered by
the standard Sturm theorem, while it is known to be highly non-trivial
(and far from being resolved) in higher dimensions. Spectral determi-
nants are addressed in Section 5.3. Scattering on graphs is described in
Section 5.4. As always, the last section contains additional literature
references and remarks.

5.1. Resonant gap opening

Existence of spectral gaps is known to be of high interest in the
various fields ranging from solid-state physics to photonic crystal the-
ory, to waveguides, to theory of discrete groups and graphs (e.g., in
the popular theory of expanders and Ramanujan graphs [219, 627]).
A standard way of trying to create spectral gaps is to make a medium
periodic (see Section 4.3). This is why most photonic crystal materi-
als that are being created are periodic. However, periodicity neither
guarantees the existence of gaps (except in the 1D case, where it is
generic), nor does it allow any easy control over gap locations or sizes,
nor is it a unique way to achieve spectral gaps. It has been noted
by several researchers (the first such references known to the authors
are [598, 599]) that distributing small geometric scatterers through-
out the medium (not necessarily in a periodic fashion) might lead to
spectral gaps. This was also noticed in some specific quantum graph
models in [39,264], and finally made very clear and precise in the case
of combinatorial graphs in [647]. It was proposed in [647] that a sim-
ple “decoration” a graph leads to a very controllable gap structure. We
will describe the decoration procedure of [647] adapted to the quan-
tum graph situation. The reader will see again the appearance of the
Dirichlet-to-Neumann map.
129
130 5. SPECTRA. SPECIAL TOPICS

Assumption 5.1.1. Let Γ0 be a quantum graph satisfying the fol-

lowing condition:
(1) the lengths of all edges and degrees of all vertices are bounded
from below and from above:
(5.1.1) 0 < Lmin ≤ Le ≤ Lmax < ∞ for any edge e,
(5.1.2) 0 < dv ≤ D0 < ∞ for any vertex v,
(2) the Hamiltonian is the negative second derivative along the
edges with the Neumann conditions (1.4.4) at the vertices1.
Let also Γ1 be a compact connected quantum graph with the same
type of the Hamiltonian, with any self-adjoint vertex conditions. The
graph Γ1 will be our “decoration.” We assume that a root vertex v1 is
singled out in Γ1 . The decoration procedure works as follows: The new
graph Γ is obtained by attaching a copy of Γ1 to each vertex v of Γ0
and identifying v1 with v (see Fig. 1). Notice that there is a natural
Γ1

Γ0

Figure 1. Decoration of a quantum graph Γ0 by Γ1 .

embedding Γ0 ⊂ Γ. We will denote by V, V0 , and V1 the vertices sets
of Γ, Γ0 , and Γ1 correspondingly. The Hamiltonian H on Γ is defined
as the negative second derivative on each edge, with the Neumann
conditions at each vertex of Γ0 (including the former vertices v1 of the
decorations) and the initially assumed conditions on V1 \v1 , repeated
on each attached copy of the decoration.
As we have already seen, Dirichlet eigenvalues of the edges (which
are clearly determined by the edge lengths spectrum) often play an
exceptional role in quantum graph considerations. Let {Le } be the
lengths of the edges of the original graph Γ0 . Then the Dirichlet
spectrum σD of Γ0 is the closure of the set
∪ {π 2 n2 /L2e } ⊂ R.
n∈(Z\0),e∈E

1More general conditions can also be considered.

 5.1. RESONANT GAP OPENING 131

Remark 5.1.2. If the set {Le }e∈E is finite, no closure is required.

Let us also define the operator H1 on the decoration graph Γ1 that
acts as the negative second derivative on each edge and satisfies the
self-adjoint conditions assumed before on V1 \ {v1 } and zero Dirichlet
condition at v1 .
We can now state the result of this section. The conditions of the
theorem can be weakened.
Theorem 5.1.3. Let Γ0 satisfy Assumption 5.1.1 and λ0 ∈ R \ σD
be a simple eigenvalue of H1 with the eigenfunction ψ. Suppose that the
sum of the derivatives of ψ at v1 along all outgoing edges is not zero.
Then there is a punctured neighborhood of λ0 that does not belong to
the spectrum σ(H) of Γ.
Remark 5.1.4. (1) This theorem claims that spectral gaps are
guaranteed to arise around the spectrum of the decoration
(with the Dirichlet condition at the attachment point v1 ), un-
less one deals with the Dirichlet spectrum σD of the underlying
graph Γ0 . Simple examples show that points of this spectrum
can indeed belong to the spectrum rather than to a gap. For
instance, if Γ0 contains a cycle consisting of edges of equal (or
commensurate) lengths, then the decoration procedure can-
not remove the eigenvalues that correspond to the sinusoidal
waves running around this loop (see Fig. 2). This is another
influence of scars (Section 3.4).
(2) If the background graph is equipped with a group action, this
action naturally extends to the decorated graph. In particu-
lar, if Γ0 is Zn -periodic, so is the decorated graph Γ. Since the
spectrum of the decoration Γ1 is infinite, the theorem implies
that the spectrum of Γ has infinitely many gaps. Our experi-
ence has already shown that when n > 1, Zn -periodic quantum
graph Hamiltonians in various regards resemble PDEs more
than ODEs. In particular, spectral zones can and often do
overlap. On the other hand, we see that the number of gaps,
in contrast to the case of periodic elliptic partial differential
operators, can be infinite.
Corollary 5.1.5. The Bethe-Sommerfeld conjecture on
finiteness of the number of spectral gaps for periodic elliptic
operators, fails for Zn -periodic quantum graphs.

Proof. The statement is trivial for finite graphs, since their spec-
tra are discrete. Still, we provide a proof for the finite graphs, which
with minor modifications goes through in the infinite case.
132 5. SPECTRA. SPECIAL TOPICS

Figure 2.

The proof consists of removing the decorations and replacing them

by altered vertex conditions. This is done simultaneously and the same
way at each vertex v ∈ V0 ⊂ V, so we will describe it for one vertex v,
which will be identified with v1 ∈ Γ1 .
Let us consider v1 as the one-point boundary of Γ1 and define, as
in Section 3.5.3, the Dirichlet-to-Neumann (DtN) function Λ(λ) for Γ1 .
According to Theorem 3.5.1, it is analytic in a punctured neighborhood
of λ0 not intersecting σD (Γ), with a first order pole (with non-zero
residue) at λ0 .
Let now λ0 be as in the theorem. Suppose that u(x) is an eigen-
function of H corresponding to an eigenvalue λ close to λ0 . For any
vertex v ∈ V0 , we can solve the equation Hu = λu on the decora-
tion attached to v, using u(v) as the Dirichlet data. Then the sum
of outgoing derivatives of u at v along the edges of the decoration is
equal to Λ(λ)u(v). Hence, the eigenfunction equation for u on Γ can
be re-written on Γ0 solely as follows:
⎧
⎪
⎪ −u = λu on each edge of Γ0 ,
⎨
(5.1.3) u is continuous at all vertices v ∈ V0 ,
⎪  du
⎪
⎩ dx (v) = −Λ(λ)u(v).
e
v∈e

We will show now that (5.1.3) is impossible for a non-zero function

u, if λ is close to λ0 . Indeed, with λ being at a positive distance from
the Dirichlet spectrum σD of all edges, standard resolvent estimates
give

(5.1.4) u 2
H 2 (e) ≤ C1 |u(v)|2 .
e∈Γ0 v∈V0
 5.1. RESONANT GAP OPENING 133

Now Sobolev trace theorem (see Lemma 1.3.8) implies

2
du
(5.1.5) (v) ≤ C2 u 2
H 2 (e) ≤ C3 |u(v)|2 .
dxe
{e∈Γ0 ,v∈V0 | v∈e} e∈Γ0 v∈V0

Since Λ(λ) has a pole at λ0 , for λ and λ0 sufficiently close, we get

a contradiction between (5.1.5) and the last equality of (5.1.3).


5.1.1. “Spider” decorations. One can try to create gaps by a

different, often more feasible, decoration procedure. Namely, instead
of attaching sideways little “flowers” (or “kites,” as they were called
in [647]) as in Fig. 1, one could incorporate an internal structure into
each vertex, putting a little “spider” there as shown in Fig. 3 below.
Assuming the original graph being regular (i.e., having all vertex de-
grees equal to the same number d), one can use the same decoration
at each vertex, except that one has the freedom of d! choices of attach-
ments at each vertex.

Figure 3. A “spider” decoration. The broken line circle

shows a former vertex, blown up to accommodate the
“spider.”

The hope is that gaps can be created using this construction. Al-
though this is indeed sometimes possible, a quick look at the last equal-
ity in (5.1.3) shows that things might not be as straightforward as in
the “flower” decoration. Indeed, the whole play was based on the ex-
pression Λ(λ)u(v) blowing up at the spectrum of the decoration. In
the “spider” case, u(v) is the vector of dimension equal to the degree d
of the former vertex, and Λ(λ) is a matrix function. Even if the matrix
function has a pole, this pole can be masked after multiplying by a
vector (in our case, u(v)). Although some positive results on the gap
opening by “spider” decoration exist [579], there is no general analysis
of this effect.
134 5. SPECTRA. SPECIAL TOPICS

5.2. Zeros of eigenfunctions and nodal domains

According to the well known ODE theorem by Sturm [412, 678,
679], the zeros of the n-th eigenfunction of a vibrating string divide the
string into n “nodal intervals”. The Courant’s nodal line theorem [207]
carries over a part of the Sturm’s theorem to the theory of membranes.
Namely, let ψn (x) be the n-th eigenfunction of the Laplace operator
with Dirichlet boundary conditions in a bounded connected domain in
Rd (more general Schrödinger operators can also be considered). The
nodal domains of the eigenfunction are the connected components
of the set of points where the eigenfunction does not vanish. Courant
proved that the number of the nodal domains for the n-th eigenfunc-
tion cannot exceed n. However, in contrast to the 1D Sturm case,
where this number is indeed equal to n, A. Stern, a student of Courant
provided [675]2 an example showing that no non-trivial lower bound
for the number of nodal domains can be hoped for in Rd , d ≥ 2, n > 1
(see also [206]). Namely, for an arbitrarily large n there are examples
of nth eigenfunctions with just two nodal domains. Moreover, Plei-
jel showed that even the Courant’s upper bound can be improved for
large n in dimensions d ≥ 2 [612]. The issues related to the structure
of nodal sets and number of nodal domains of eigenfunctions happen
to be very complex and are still under intensive investigation by distin-
guished mathematicians and physicists (see, e.g. [114, 233, 235, 409]
and references therein). The intriguing and beautiful nodal patterns
(so called “Chladni figures”, although da Vinci [524], Galileo [330]
and Robert Hooke [116, 623] observed them hundreds years before
Ernst Chladni, who provided a more regular study of those [188]) have
fascinated the general public for four hundred years (e.g., [116, 188,
330, 363, 524, 535, 623]). In spite of this long history, many basic is-
sues concerning nodal sets and nodal domains remain unresolved and
attract attention of many mathematics and physics researchers (see,
e.g. [46,123,154,202,233–237,355,392,408,409,434,445,535,562]
and the references therein).
In this section we are interested in studying the number of zeros
and nodal domains of real eigenfunctions of quantum graph Hamilto-
nians, which also presents non-trivial problems. As it often happens
with graph operators, their behavior is intermediate between that of
1-dimensional and multi-dimensional differential operators, and some-
time shows features uncommon to both.

2Seehttps://dmv.mathematik.de/m-die-dmv/m-geschichte/m-dissertationen-
1907-1945.html
 5.2. ZEROS OF EIGENFUNCTIONS AND NODAL DOMAINS 135

Figure 4. Three examples of zeros distribution on a

graph, with μ = 2, ν = 3, with μ = 2, ν = 2, and with
μ = 6, ν = 6.

Let f be a continuous real valued function on a compact connected

metric graph Γ, which has finitely many isolated zeros, none of which
at the vertices of Γ. The number of its zeros will be denoted by
μ(f ). A positive (negative) domain with respect to f is a maximal
connected subset in Γ where f is positive (correspondingly, negative).
The total number of positive and negative domains will be called the
nodal count of f and denoted by ν(f ). The two quantities are closely
related, although, due to the graph topology, the relationship is more
complex than on a line, where ν = μ + 1, as one can see in examples
shown in Figure 4.
Let us first start with a simple estimate for the number of compo-
nents into which a graph can be split by a number of cuts.
Lemma 5.2.1. Let μ distinct points be placed on the edges of a com-
pact connected metric graph Γ in such a manner that none of the points
coincides with any of the vertices. Let ν be the number of connected
components Γj , j = 1, . . . , ν into which the graph splits when cut at
these points. Then
(5.2.1) μ − βΓ + 1 ≤ ν ≤ μ + 1,
where βΓ is the first Betti number of Γ (see Definition 1.1.1).
More precisely,
ν
(5.2.2) ν = μ − βΓ + βΓj + 1.
j=1

In particular, if all connected components are simply connected, then

the left inequality in (5.2.1) becomes an equality:
(5.2.3) μ − βΓ + 1 = ν.

Proof. Since every cut increases the number of the components

by at most one, the upper bound on ν given in (5.2.1) is obvious.
The lower bound clearly follows from (5.2.2), since βΓj ≥ 0. We
thus concentrate on proving (5.2.2). In order to do so, we observe that
136 5. SPECTRA. SPECIAL TOPICS

each cut introduces two new vertices and splits an existing edge into
two, thus increasing the number of edges by one. If Γ is the graph
obtained as the result of the cuts, then
VΓ − EΓ = VΓ − EΓ + μ.
Applying the explicit expression (1.1.5) for the Betti number, we get
(5.2.4) kΓ − βΓ = kΓ − βΓ + μ,
where k is the number of components of the corresponding graph and β
is the Betti number. To complete the proof of (5.2.2), we observe that
the number of resulting connecting components is ν = kΓ , the number
of starting 
connected components is kΓ = 1, and the Betti number βΓ
is equal to j βΓj .
If each connected component is simply connected, (5.2.2) clearly
implies (5.2.3). 
We will be interested in the nodal counts of the eigenfunctions of
d2
the Hamiltonian H = − dx 2 subject to δ-type conditions at the vertices:

the functions are continuous and

d
(5.2.5) fe (v) = αv fe (v) for all v ∈ V.
e∈E
dx
v

At any vertex v of degree 1 we will allow an extended δ-type condition,

i.e., in addition to the Robin condition f  (v) = αv f (v), we will allow
the Dirichlet condition f (v) = 0, which corresponds to αv = ∞. A
Dirichlet vertex is exempted from the requirement that the eigenfunc-
tion should have no zeros on vertices. On the other hand, they do not
contribute to the count μn .
The operator H can also be defined via its quadratic form
Le
(5.2.6) h[f, f ] = |f  (xe )|2 dxe + αv f 2 (v),
e∈E 0 v:αv =∞

where for the Dirichlet vertices (αv = ∞) the condition f (v) = 0 is

enforced for the functions in the domain of h.
d2
Most of our results also apply3 to Schrödinger operators H = − dx 2+

V (x) with a potential V (x), which is piecewise continuous on every edge

of the graph, see references in Section 3.8.
We assume that λn is the n-th eigenvalue of H and f (n) (x) is the
corresponding eigenfunction. We will abbreviate μ(fn ) and ν(fn ) as μn
and νn correspondingly. Unless stated otherwise, we will be working
under the following assumption.
3Theorem 5.2.4 is a notable exception
 5.2. ZEROS OF EIGENFUNCTIONS AND NODAL DOMAINS 137

Assumption 5.2.2. The eigenvalue λn is simple and the corre-

sponding eigenfunction f (n) does not vanish at any vertex (except the
Dirichlet ones).
As discussed in Section 3.1.5, in the case of Neumann-Kirchhoff ver-
tex conditions the eigenvalues are simple generically. It is also expected
that, in the finite-dimensional space of all edge lengths {Le }, the set
An on which (λn , f (n) ) satisfy the Assumption is open and dense, un-
less the graph contains loops. Moreover, on each connected component
of the set An , the nodal count of f (n) remains the same. Indeed, the
zeros cannot pass through vertices and cannot undergo a bifurcation
(i.e. appear or disappear), otherwise at the bifurcation point the eigen-
function and its derivative are both zero. By uniqueness theorem for
second order ODEs, this would mean that f (n) is identically zero on the
whole edge, contradicting the assumption. The properties described in
the Assumption 5.2.2 are conjectured to be generically satisfied also
in the case of general δ-type vertex conditions, not only the Neumann
ones.
5.2.1. Some basic results. The following statement is trivial but
turns out to be useful. We note that it applies to the eigenfunctions of
the graph Hamiltonian, since their zeros are isolated (under Assump-
tion 5.2.2) and simple (due to the uniqueness theorem for 2nd order
ODEs).
Lemma 5.2.3. Let f be a continuous function on a connected graph
Γ, which is smooth on each edge, does not vanish at vertices, and has
only isolated and simple zeros. Then the number of zeros on each cycle
of the graph is even.
Proof. Since the zeros are simple, the function f changes sign at
each zero. Going around a cycle, the function must change sign an
even number of times. 
(e)
The following theorem gives an explicit formula for the number μn
d2
of zeros of n-th eigenfunction of H = − dx 2 on an edge e, given the
2
eigenvalue λn = kn and the signs of the eigenfunction on the vertices.
Theorem 5.2.4. Let λn = kn2 be the n-th eigenvalue and f be the
d2
corresponding n-th eigenfunction of the operator H = − dx 2 . Suppose

that (λn , f ) satisfy Assumption 5.2.2. Then on each edge e = (u, v) the
number of zeros is
. / !
kn Le 1
1 − (−1) π  sign(f (u)f (v))
kn Le
(e)
(5.2.7) μn = +
π 2
138 5. SPECTRA. SPECIAL TOPICS

and
(5.2.8) νn = μn − E + V = μn − βΓ + 1
for sufficiently large n (e.g., such n that π/kn becomes smaller than the
length of the shortest edge of Γ).
Here . . .  denotes the standard floor function and βΓ = E − V + 1
is the Betti number of Γ.
Proof. The eigenfunction with the eigenvalue λn has the general
form fe (x) = Ae sin(kn x + φe ) on every edge. Thus, the distance be-
tween its two consecutive zeros on e is exactly π/kn . The maximal
number of full segments of this length that can fit inside the segment
of length Le is kn Le /π. If this number materializes, then the number
of zeros will be kn Le /π + 1. However, one can move these segments
in such a way that one of the nodes will disappear, without new one
appearing, which will reduce the number of zeros by one. One can
easily check now that this correction to the main term kn Le /π is
1 if sign(f (u)f (v)) = (−1) kn Le /π and 0 otherwise, which proves the
formula (5.2.7).
Now, if n is so large that π/kn is smaller than the length of the
shortest edge of Γ, then every edge contains at least one zero of the
eigenfunction fn . Thus, all the nodal domains are simply-connected
(since all cycles are cut). Now the last statement of Lemma 5.2.1
implies (5.2.8). 
(e)
0 k LRemark
1 5.2.5. If one of the vertices is Dirichlet, then μn =
n e
π
.
5.2.2. Bounds on the nodal count. We start by proving the
Courant-type bound on the number of nodal domains.
2
Theorem 5.2.6. Let λn be an eigenvalue of H = − dx
d
2 + V (x) on a
(n)
connected graph Γ and its eigenfunction f be non-zero at all vertices
of Γ that have degree larger than 1. Then
(5.2.9) νn ≤ n.
Proof. The proof follows the standard route (see Section 3.8 for
references) and we omit some details. We assume the contrary and let
φ1 , . . . φn+1 be the pieces of the function f (n) restricted to (the first
n + 1 of) the nodal domains. More precisely, φj (x) = f (n) (x) on j-th
nodal domain and φj (x) = 0 otherwise.
There exists a linear combination φ of n of the functions φj which is
orthogonal to first n−1 eigenfunctions of H. Since each of the functions
 5.2. ZEROS OF EIGENFUNCTIONS AND NODAL DOMAINS 139

φj locally satisfies the eigenvalue equation Hφj = λn φj , we find that

h[φ, φ] h[f, f ]
= λn = min .
φ f ⊥f1 ,...fn−1 f
This implies that φ is an eigenfunction. However, it is identically zero
on the (n + 1)-th nodal domain of f (n) . Since all boundaries of the
nodal domains lie inside the edges, by uniqueness of the solution to
an ODE, φ has to be zero on all the domains that border (n + 1)-th
domain. Continuing by induction, we conclude that the function φ
must be zero, which is a contradiction. 
Remark 5.2.7. The Courant theorem on graphs is weaker than that
in Rn since we impose an additional restriction that the eigenfunction
has no zeros on internal vertices. It is easy to see that this is an essential
requirement. Indeed, consider a star graph with Dirichlet conditions
on the degree 1 vertices, Neumann condition at the central vertex and
2v edges of the same length π. Then the second eigenvalue is 1 with
multiplicity 2n−1. An eigenfunction which is equal to sin(x) on v edges
and − sin(x) on the other v edges will have 2v > 2 nodal domains.
The results of [225] on discrete graphs suggest that for degenerate
eigenvalues the upper bound should be amended to
(5.2.10) νn ≤ n + mn − 1,
where mn is the multiplicity of the n-th eigenvalue. However, this
question has not been addressed on quantum graphs yet.
Interestingly, on graphs we can also obtain a lower bound on the
number of nodal domains. We formulate the result together with the
upper bound for easy reference.
2
Theorem 5.2.8. Let λn be a simple eigenvalue of H = − dx d
2 +V (x)
(n)
on a graph Γ and its eigenfunction f be non-zero at all vertices of
Γ. Then
(5.2.11) n − βΓ ≤ νn ≤ n,
(5.2.12) n − 1 ≤ μn ≤ n − 1 + βΓ ,
where βΓ = E − V + 1 is the Betti number of Γ.
Proof. The lower bound for the number of nodal domains νn fol-
lows from the corresponding result for the number of zeros μn and
Lemma 5.2.1. The upper bound has already been proven in Theo-
rem 5.2.6. Thus we will only prove inequality (5.2.12). Its proof con-
sists of two parts. First, the theorem is shown for the tree case, βΓ = 0,
for which (5.2.12) (and (5.2.11)) becomes an equality. Then the bound
140 5. SPECTRA. SPECIAL TOPICS

is proved for general graphs. The upper bound for μn also follows from
Theorem 5.2.6 and Lemma 5.2.1, but we will obtain it here indepen-
dently with no extra work.
Let us start with βΓ = 0. To simplify notation we will drop the
script n when talking about the eigen-pair (λn , f (n) ). We will prove
the result by induction on the number of internal vertices of the tree
Γ. If there are no internal vertices, Γ is simply an interval and the
statement reduces to the classical Sturm’s Oscillation Theorem (see,
e.g. [207]).
Let v be a vertex of degree dv . Then v separates Γ into dv sub-
trees Γi , each with a strictly smaller number of internal vertices. For
each subtree, the vertex v is a vertex of degree 1 with, so far, no
vertex condition. We will impose δ-type condition with the parameter
αi < ∞ chosen in such a way that the function f restricted to Γi
is still an eigenfunction. This value is simply αi = fi (v)/fi (v). The
eigenfunction f still corresponds to the eigenvalue λn which is still
simple (otherwise we can construct another eigenfunction for the entire
tree Γ, contradicting the simplicity of λ). It is an ni -th eigenvalue of Γi
and, applying the inductive hypothesis, we conclude that the function
f has ni − 1 zeros on the subtree Γi . Thus we need to understand the
relationship between the numbers ni and the number n.
To do it, consider the subtree Γi,∞ which now has Dirichlet con-
dition on the vertex v. By Theorem 3.1.8, there are exactly ni − 1
eigenvalues of Γi,∞ that are smaller than λ. Now consider the tree Γ∞ ,
which is the original tree but with the Dirichlet condition at the vertex
v. On one hand it has exactly n − 1 eigenvalues that are smaller than
λ. On the other hand Γ∞ is just a disjoint collection of subtrees Γi,∞
and its spectrum is the superposition of the spectra of Γi,∞ . Therefore
the number of eigenvalues that are smaller than λ is
dv
n−1= (ni − 1).
i=1

Since we already proved that the number of zeros of f is equal to the

sum on the right-hand side, we can conclude that μ(f ) = n − 1.
We will now proceed to derive the lower bound for graphs with
cycles by cutting the cycles and using the result already obtained for
trees.
Assume that cutting the edges e1 , . . . , eβ turns the graph Γ into a
tree. We cut each of these edges at a point xj ∈ ej such that f (n) (xj ) =
0. We thus obtain a tree with EΓ +β edges and VΓ +2β vertices. Denote
this tree by T . There is a natural mapping from the functions on the
 5.2. ZEROS OF EIGENFUNCTIONS AND NODAL DOMAINS 141

graph Γ to the functions on the tree T . In particular, we can think of

f (n) as a function on the tree.
We would like to consider the same eigenproblem Hf = λf on the
tree now. The vertex conditions on the vertices common to T and Γ
will be inherited from the eigenproblem on Γ. But we need to choose
the boundary conditions at the 2β new vertices. Each cut-point xj
gives rise to two vertices, which we will denote by uj+ and uj− . Define
d (n) d (n)
dx
f (uj+ ) dx
f (uj− )
aj+ = , aj− = , j = 1, . . . , β,
f (n) (uj+ ) f (n) (uj− )
where the derivatives are taken into the corresponding edges of T . Since
f (n) , as an eigenfunction, is continuously differentiable and f (n) (uj+ ) =
f (n) (uj− ), we have aj+ = −aj− .
Now we set the boundary conditions on the new vertices of T to be
d d
ψ(uj+ ) = aj+ ψ(uj+ ), ψ(uj− ) = aj− ψ(uj− ), j = 1, . . . , β,
dx dx
where the derivatives, as always, are taken into the edges. By definition
of the coefficients aj± , the function f (n) satisfies the above boundary
conditions. It also satisfies the equation Hf = λf and the vertex
conditions throughout the rest of the tree. Thus, f (n) is also an eigen-
function on T and λn (Γ) is the corresponding eigenvalue, simple by
Corollary 3.1.9. If we denote the ordered eigenvalues of T by λm (T ),
then λn (Γ) = λm (T ) for some m. It is important to note that m is in
general different from n.
We notice that the original graph Γ can be obtained from the tree T
by gluing the pairs of vertices in the sense of Theorem 3.1.10. Since in
our case α0 = −α1 (in the notation of Theorem 3.1.10), the new vertex
has, according to equation (3.1.14), the Neumann vertex condition.
This, for a vertex of degree 2, is equivalent to there being no vertex at
all. Thus we can apply Theorem 3.1.11 with k = β and obtain
λn (T ) ≤ λn (Γ) ≤ λn+β (T ).
This means that m, which is the index of the eigenvalue λn (Γ) in the
spectrum of the tree T satisfies
(5.2.13) n ≤ m ≤ n + β.
Since the cutting did not introduce any new zeros,
μn (Γ) = μm (T ) = m − 1,
where we used the already proved part of the theorem. Combining
with inequality (5.2.13) we obtain the desired result
n − 1 ≤ μn (Γ) ≤ n + β − 1.
142 5. SPECTRA. SPECIAL TOPICS

Figure 5. An example of a graph with a single cycle

(βΓ = 1) and a graph obtained from it by removing the
cycle.


Remark 5.2.9. Theorem 5.2.8 implies that, on a tree graph, an
eigenfunction that has μ zeros must be the eigenfunction number μ+ 1.
We can strength this useful result. Let f be an eigenfunction of a graph
Γ. Assume f is non-zero on vertices of Γ. If the function f has μ
internal zeros and μ + 1 nodal domains (i.e. each zero of f disconnects
Γ into two parts), then it is eigenfunction number μ + 1.
5.2.3. Nodal count for special types of graphs. Theorem 5.2.8
can be used to obtain some (almost) explicit formulae for the nodal
count on a special family of graphs. First we consider the nodal count
on a graph with βΓ = 1, i.e. a graph with a single cycle, see Figure 5
for an example.
Theorem 5.2.10 ([47]). For a graph consisting of a single loop with
trees grafted on it,
2 ! 3
μn = n − 1 + N∅ (λn ) + n + 1 mod 2 ,
where N∅ is the spectral counting function of the disconnected graph
obtained by removing the loop and imposing Dirichlet conditions at the
new vertices.
The above result can be generalized to graphs containing several
disjoint cycles, see Figure 6 for an example.
Theorem 5.2.11 ([47]). For a graph containing several disjoint
cycles,
βΓ 2 ! 3
μn = n − 1 + Nj,∅ (kn ) + n + 1 mod 2 ,
j=1
where Nj,∅ is the spectral counting function of the disconnected graph
obtained by removing the j-th loop and imposing Dirichlet conditions
at the new vertices.
 5.2. ZEROS OF EIGENFUNCTIONS AND NODAL DOMAINS 143

Figure 6. An example of a graph with several disjoint cycles.

5.2.4. Nodal deficiency and Morse indices. A set of interest-

ing results about the nodal count can be obtained by shifting the point
of view: we will ask ourselves what distinguishes the set of points on
the graph that is the zero set of an eigenfunction.
Definition 5.2.12. Let Γ be a quantum graph.
(1) A partition vertex on Γ is a new vertex introduced on an
edge of Γ. The partition vertex is called proper if it is located
in the interior of an edge, i.e. that is not at an existing vertex
of Γ. Otherwise, we call it an improper partition vertex.
(2) An m-partition of Γ is a set of m partition vertices on the
graph. The partition is proper if all of its vertices are proper.
Otherwise, we call it an improper partition. The set of all
proper m-partitions of Γ is denoted by Pm (Γ).
The partition P ∈ P should be understood as a candidate for the
zero set of an eigenfunction. As everywhere in this section, the eigen-
function is required to satisfy Assumption 5.2.2, so the restriction to
proper partitions is natural.
As mentioned in section 1.4.4, imposing vertex Dirichlet conditions
at the partition vertices of P separates Γ into several subgraphs, which
we denote by {Γj }, and call the partition’s subgraphs or connected
components. The number of partition components is denoted by ν(P )
and is related to the number of partition points μ(P ) ≡ m via equa-
tion (5.2.2). We chose the number of points m to act as the size of
the partition to simplify the subsequent notation. Making the other
possible choice, m := ν(P ), would result in only minor changes in the
forthcoming results. We note that in dimensions higher than 1, the
“number of zeros” concept is no longer available, and the “number of
components” therefore acts as the size of the partition.
A partition that is obtained by taking the nodal set of an eigen-
function is called the nodal partition. The following definition is
motivated by the observation that an eigenfunction changes sign when
crossing a zero.
144 5. SPECTRA. SPECIAL TOPICS

(a) (b) (c)

Figure 7. (a) A proper partition P ∈ P3 (Γ) and (b) its

two subgraphs; (c) a non-bipartite partition with βΓ\P =
1

Definition 5.2.13. The partition P ∈ P is called bipartite if

there exists a map from its subgraphs to signs {Γj } → {+, −}, such
that neighboring subgraphs are mapped to different signs. Equiva-
lently, a partition is bipartite if it has an even number of points on
every cycle of Γ.
This and previous definitions are illustrated in Fig. 7.
Further, suppose that a partition is generated by the nodal points of
the n-th eigenfunction f (n) . Restricting this function to the partition’s
subgraph Γj , we obtain an eigenfunction of Γj (the partition points
get the Dirichlet condition). Since it does not have any internal zeros,
the restriction of f (n) is proportional to the ground state of Γj . In
particular, we have
λ1 (Γj ) = λn (Γ) .
This motivates the next definition.
Definition 5.2.14. An m-partition is an equipartition if all of
its subgraphs {Γj } share the same first eigenvalue:
(5.2.14) λ1 (Γj1 ) = λ1 (Γj2 ),
for all j1 and j2 . The set of all proper equipartitions of Γ of size m is
denoted Qm (Γ).
It is obvious that for a partition to be a nodal partition (i.e. to be
generated by an eigenfunction of a specified operator) it is necessary
to be a bipartite equipartition. The next natural question is what
is the sufficient condition to be a nodal partition. To give a quick
preview, the only extra condition is for the eigenvalue λ1 (Γj ) to have
zero first variation with respect to perturbations of the partition within
 5.2. ZEROS OF EIGENFUNCTIONS AND NODAL DOMAINS 145

the set Qm (Γ). In other words, the nodal partitions are the critical
points of the functional λ1 (Γj ). Note, that since we are dealing with
equipartitions, we do not need to specify j in λ1 (Γj ). Extra information
can be gleaned by looking at the second variations of λ1 (Γj ). Let d
denote the Morse index of the critical point P ∈ Qm (Γ) (the number
of negative eigenvalues of the matrix of second derivatives). Then the
partition P corresponds to the eigenfunction number m + d. In other
words, the nodal deficiency n− νn determines the index of instability
of the nodal partition of the corresponding eigenvalue. We will now
make this discussion more precise. We will omit the proofs, referring
the interested reader to the original article [46].
Given a proper partition P ∈ Pm we can parameterize nearby par-
titions by small open neighborhood of zero in Rm . For each partition
point we fix, arbitrarily, the positive direction along the edge. Then,
the j-th entry of p ∈ Rm is the shift of the j-th partition point in
the specified direction. Since P is proper, for a small enough  > 0,
the perturbation by p with |pj | <  results in a proper partition. This
parameterization gives Pm the structure of an m-dimensional manifold.
Starting with a proper equipartition P ∈ Qm we will now try to
understand the structure of nearby equipartitions. To simplify exposi-
tion, we will assume that there is at least one partition point on every
cycle of the graph Γ. In other words, all partition subgraphs Γj are
trees.
From the set P of m points we select a maximal set S of points
that do not disconnect the graph. It is easy to see that the number of
points in S is β independently of how S is chosen. We claim that all
other points in P are uniquely determined by the points in S. Indeed,
imposing the Dirichlet conditions at these points (and, correspondingly,
cutting the graph Γ there), we obtain a tree. On this tree the set P \ S
is also an equipartition, which is unique according to the following
lemma.

Lemma 5.2.15. Let T be a tree graph. For any given m the set
Qm (T ) contains at most one point, which is the nodal partition of the
eigenfunction f (m+1) .

Sketch of the proof. If the eigenfunction f (m+1) is non-zero on

vertices, it generates an equipartition with m points by Theorem 5.2.8.
Conversely, given an equipartition with m points, we can construct
an eigenfunction out of groundstates of Γj . Again by Theorem 5.2.8 it
must be the eigenfunction number m + 1. To construct an eigenstate,
we need to match the groundstates to be continuously differentiable
146 5. SPECTRA. SPECIAL TOPICS

across the partition points (they are continuous automatically). It is

possible to do this recursively on a tree graph. 
This observation suggests a way to parameterize equipartitions that
are near to a given P ∈ Qm : instead of moving all points in P , we
move those in S. The other points in the perturbed partition are
obtained as the nodal points of the (m − β + 1)-th eigenfunction of
the corresponding (perturbed) tree graph. Thus the equipartitions —
at least those with all cycles containing a partition point — live on a
β-dimensional manifold.
The parameterization above is natural, but involves changing edge
lengths in the tree graph. It is usually more convenient to have a fixed
graph but change vertex conditions instead. This construction is in-
spired by the second part of the proof of Theorem 5.2.8. We choose β
cut-points, to have at least one on each cycle of Γ but without discon-
necting the graph. The set of the cut-points will be denoted by C. If
the cut-point x belongs to the partition subgraph Γj , we let
d
f (xj )

aj = dx
(n)
.
f (xj )
We now cut the graph at the points C, we obtain a tree graph T , with
each cut-point xj giving rise to two vertices uj− and uj+ (on the left and
on the right of the cut, respectively, relative to the “default” direction
of the edge that is being cut). The new vertices get the conditions
d d
ψ(uj+ ) = aj ψ(uj+ ), ψ(uj− ) = −aj ψ(uj− ), j = 1, . . . , β.
dx dx
It was explained in the proof of Theorem 5.2.8 that for aj = 
aj the orig-
inal equipartition P is the nodal partition of the eigenfunction number
(m+1) of the tree graph. Perturbing the β parameters aj we get nearby
equipartitions. Furthermore, the two parameterizations are locally dif-
feomorphic. Now we can formulate the main result of this subsections.
Theorem 5.2.16. Consider the functional Λ : Qm (Γ) → R, defined
by
Λ(P ) = λ1 (Γj ), 1 ≤ j ≤ m.
(1) Λ is smooth with respect to either parameterization described
above at any point P ∈ Qm (Γ).
(2) If a bipartite partition P ∈ Qm is a critical point of Λ, it
corresponds to an eigenfunction of Γ.
(3) Conversely, if (λn , f (n) ) satisfy Assumption 5.2.2, the nodal
partition of f (n) is a non-degenerate critical point of Λ.
 5.2. ZEROS OF EIGENFUNCTIONS AND NODAL DOMAINS 147

(4) The Morse index of the critical point that is the nodal partition
of f (n) is equal to the nodal deficiency of f (n) , namely n − νn .
For completeness we remind the reader the definition of the Morse
index. A critical point of a smooth function F : Rβ → R is a point x0
where the differential of F is zero. In other words, the Taylor expansion
of F (x0 + δx) − F (x0 ) starts with quadratic terms,
F (x0 + δx) = F (x0 ) + H(δP, δP ) + higher order terms,
where H is a quadratic form (the Hessian). The Morse index of
x0 is the number of negative eigenvalues of H. The critical point is
non-degenerate if all eigenvalues of the Hessian are non-zero.

Ideas behind the proof. Smoothness of Λ is a consequence of

the general theorems on variation of vertex conditions or edge lengths,
discussed in section 2.5.
The fact that a critical point corresponds to a nodal partition can
be obtained from the Hadamard-type formulae, Proposition 3.1.5 and
Proposition 3.1.6 (depending on the parameterization chosen). In ei-
ther case, if the derivative is zero, we find that the eigenfunction on the
tree T matches smoothly at the cut-points to produce an eigenfunction
on the original graph Γ.
As for the Morse index, we first give a heuristic explanation of the
result and then an indication of one possible proof. Heuristically, the
functional Λ is closely related to the quadratic form on the graph. The
quadratic form, considered on the unit circle in the space H  1 (Γ), has
(n)
a critical point at the eigenfunction f . This critical point has index
n − 1: the negative directions are given by the lower eigenfunction, the
first to (n − 1)-th. On the other hand, the “equipartition condition”
imposes ν − 1 restrictions of the type λ1 (Γj ) = λ1 (Γ1 ), j = 2, . . . , ν.
Thus, the dimension of the negative subspace of the Hessian of Λ is
reduced to (n − 1) − (ν − 1) = n − ν. Of course, this “dimensional
arithmetic” is deceptively simple. For example, we skipped from the
space of functions to a manifold of partitions; subtraction of dimensions
requires verification of transversality, and so on.
A more basic approach is to re-visit the last part of the proof of
Theorem 5.2.8. In it we apply Theorem 3.1.10 repeatedly to glue one
cut-point at a time. Every time we get that λn is equal to either the
minimum or the maximum of the cut-graph eigenvalue with respect to
the parameter aj− . At the end of the process we find that the nodal
deficiency n − νn is equal to the number of maximums we had taken.
This, in turn, is precisely the Morse index of the critical point. 
148 5. SPECTRA. SPECIAL TOPICS

Remark 5.2.17. One can consider the functional Λ extended to all

partitions as
(5.2.15) Λ(P ) = max λ1 (Γj ).
1≤j≤ν

It is obviously an extension of the functional we considered on equipar-

titions. Due to the presence of a maximum, this functional is no longer
smooth. But the nodal deficiency is still the “number of directions
in which Λ(P ) decreases” for the following reason: when we leave the
submanifold of equipartition, the functional Λ goes up because of the
maximum in its definition.

5.3. Spectral determinants of quantum graphs

Spectral determinants of quantum graph Hamiltonians have been
considered in various mathematical and physical publications. We
present here a sample result and refer the reader to the literature men-
tioned in Section 3.8 for details and proofs.
Let Γ be a compact quantum graph and A = (aij ) be a V × V -
matrix4. We consider the Hamiltonian HA = −d2 /dx2 + V (x), where
the potential q(x) is continuous on Γ and smooth on each edge. The
2 (Γ) that are continuous on
domain of HA consists of all functions f ∈ H
Γ and such that the following generalized δ-type condition is satisfied
at any vertex vi :
V
df
(5.3.1) (vi ) = aij f (vj ).
e∼vi
dxe j=1

The spectrum of such defined (in general, non-self-adjoint) operator

is discrete (the proof is analogous to the one of the Theorem 3.1.1):
σ(HA ) = {λk }. There exists a ray
R = {z = teiθ , t > 0}
in the complex plane that does not contain any eigenvalues λk of HA .
The ζ-function of HA is defined as follows:
(5.3.2) ζR (z) = λ−z
k ,
λk =0

where the complex powers are defined with a cut along the ray R (which
explains the subscript R in the ζ-function notation).
This function is analytic in the half-plane Re z > 0.5 and allows an-
alytic continuation to the whole complex plane, except possible simple
poles at the points 0.5 − n for non-negative integers n.
4As before, V := |V| denotes the number of vertices in Γ.
 5.3. SPECTRAL DETERMINANTS OF QUANTUM GRAPHS 149

Definition 5.3.1. The determinant of HA is:


0 if 0 ∈ σ(HA ),
(5.3.3) det HA :=  (0)
−ζR
e otherwise.
In order to describe one of the results on evaluation of the determi-
nant, we need to introduce more notation.
We will denote by HD the operator analogous to HA , but with the
Dirichlet vertex conditions: f (v) = 0 for all vertices v ∈ V. As we have
discussed already (see Section 1.4.4), this operator splits into the direct
sum of single edge operators −d2 /dx2 + V (x) with Dirichlet conditions
at the ends.
Let us also designate the whole set of vertices V of Γ as its bound-
ary B. Then one can define, as in the Section 3.5.2, the Dirichlet-to-
Neumann map Λ(λ) : CV → CV . As we have seen, it is defined when
λ is not in the spectrum of HD .
Theorem 5.3.2 ([320]). The determinant of the operator HA + λ
can be computed as follows:
1
(5.3.4) det(HA + λ) := 4 det (Λ(λ) + A) det (HD + λ) ,
dv
v∈V

where, as before dv denotes the degree of the vertex v.

Remark 5.3.3. (1) Although the expression in the right hand
side of (5.3.4) is defined a priori only outside the Dirichlet
spectrum, one can show that in fact it is an entire function.
(2) The formula enables one to reduce computation of the quan-
tum graph determinant to solving ODEs (for computing the
DtN map Λ(λ) and the determinants of the operators with
Dirichlet conditions on individual edges) and computing finite
determinants det (Λ(λ) + A).
A slightly different formula for the spectral determinant was derived
in [402]. The formula is valid for all self-adjoint vertex conditions. We
cite a simplified version with no potential (V (x) = 0). To highlight the
similarities with Theorem 5.3.2 we re-cast the result of [402] in terms
of a Dirichlet-to-Neumann map we saw in Section 3.5.4.
Theorem 5.3.4. Consider the Hamiltonian H = d2 /dx2 on a graph
Γ with the vertex conditions specified in terms of the global matrices A
and B (cf. Section 1.4.1) and let
 
− cot(kL) csc(kL)
ΛE (k) = k ,
csc(kL) − cot(kL)
150 5. SPECTRA. SPECIAL TOPICS

be the edge-wise Dirichlet-to-Neumann map (see Section 3.5.4). Then

the spectral determinant is
E
2 √ !
(5.3.5) det(H + λ) := √ √ ! det A + BΛE (i λ) .
e=1 λ sinh λ

5.4. Scattering on quantum graphs

A compact graph Γ = (V, E) can be turned into a scattering sys-
tem by attaching infinite edges (usually termed leads) to some of the
graph’s vertices (termed boundary vertices). We denote by E ext
the set of leads (“external edges”) and by Γ  the new graph. At
the boundary vertices some self-adjoint vertex conditions (see Theo-
rem 1.4.4) are imposed.
For a given k = 0 we are looking for a solution of the equation
d2 f
(5.4.1) − = k 2 f (x)
dx2
that, on the leads, has the form
−ikx
(5.4.2) f (xe ) = cin
e e + cout
e e
ikx
,
where x is the coordinate measuring the distance from the attachment
point of the lead. We will see that every set of incoming amplitudes
e }e∈E ext gives rise to a solution of required form. The mapping from
{cin
e }e∈E ext to the outgoing amplitudes {ce }e∈E ext is called the scat-
{cin out

tering matrix.
Remark 5.4.1. We have encountered several “scattering matrices”
so far: the vertex scattering matrix σ (v) of Section 2.1.1 and the bond
scattering matrix of Section 2.1.2. Of those, the current one is what
is traditionally understood to be a scattering matrix: a system (the
graph) is opened up (by attaching leads) and the scattering of incoming
waves is investigated.
To derive an explicit expression for the scattering matrix, we ob-
serve that the solution of (5.4.1) on internal edges will have the form
f (xb ) = ab eikxb + eikLb ab e−ikxb ,
where xb is the distance to a given vertex v along the edge e which
gives rise to the incoming bond b and the outgoing bond b. Now we
follow the derivation of Section 2.1.2. At any vertex v there are local
incoming amplitudes ab , multiplied by the corresponding phases eikLb
and, possibly, incoming amplitudes cin e from the leads which are not
multiplied by a phase. The vertex scattering matrix σ (v) , computed
 5.4. SCATTERING ON QUANTUM GRAPHS 151

from the vertex conditions using Lemma 2.1.3, maps the local incoming
amplitudes to the local outgoing amplitudes (possibly including the
lead amplitudes). But the local outgoing amplitudes on the internal
edges are in fact the incoming amplitudes at the other end of the edge.
This can be summarized as
 out   in 
c ikL c
(5.4.3) = Se .
a a
Here the vectors c in and c out collect all incoming (correspondingly,
outgoing) amplitudes on the leads, a collects all the amplitudes on the
internal edges (two per edge). The matrix S collects coefficients from
the vertex scattering matrices σ (v) . The diagonal lengths matrix L has
effective length 0 associated to the leads. All vectors and matrices have
the dimension |E ext | + 2 E int .
The matrix SeikL can be split into blocks corresponding to internal
edges and leads,
 out     in 
c R(k) To (k) c
(5.4.4) = .
a Ti (k) Ũ (k) a
The matrix Ũ (k) describes the evolution of the waves inside the com-
pact graph and has eigenvalues that can now lie inside the unit circle
due to the “leaking” of the waves into the leads. The matrix R(k) de-
scribes the immediate reflection of waves by the graph (from leads into
leads). The matrices To and Ti describe transmission from the com-
pact part out and from the leads into the compact part correspondingly.
From equation (5.4.4) one can eliminate the variables a and find the
scattering matrix Σ(k).
Theorem 5.4.2. Define the matrix Σ(k) by
!−1
(5.4.5)  (k)
Σ(k) = R(k) + To (k) I − U Ti (k).
Then Σ(k) is a meromorphic function of k which is analytic on the real
line. In particular, the set of points
5 ! 6
(5.4.6) Δ = k ∈ R : det I − U (k) = 0

is discrete and the corresponding singularities of Σ(k) are removable.

For any k ∈ R \ {0} and c in there is a solution to (5.4.1) of the
form (5.4.2) with c out given by
(5.4.7) c out = Σ(k)c in .
Remark 5.4.3. If k ∈ Δ then k 2 is an eigenvalue embedded into

the continuous spectrum of the graph Γ.
152 5. SPECTRA. SPECIAL TOPICS

Remark 5.4.4. Expanding the inverse matrix in (5.4.5) into a geo-

metric series we get
!j

Σ = R + To Ti + . . . + To U Ti + . . . ,

which has a nice meaning: the term R corresponds to the wave reflect-
 j Ti correspond to the
ing off the graph, while the terms of the form To U
wave going into the graph, bouncing j times inside the graph before
finally leaving it.
Example 5.4.5. Consider a simple example of a lasso graph: a
loop of length L with a lead attached. At the attachment point the
standard Neumann conditions are enforced. Such conditions at a vertex
of degree 3 assign the weight −1/3 to backscattering and 2/3 to forward
scattering (see Example 2.1.7). Denoting by a1 and a1 the amplitudes
on the loop, we have
⎛ out ⎞ ⎛ ⎞ ⎛ in ⎞
c −1 2eikL 2eikL c
⎝ a1 ⎠ = 1 ⎝ 2 2eikL −eikL ⎠ ⎝ a1 ⎠ .
a 3 2 −eikL 2eikL a
1 1

Note the reshuffling inside the matrix, due to the fact that backscat-
tering from 1 goes into 1 and vice versa. According to (5.4.5), we have
 −1  
1 4 ikL   3 − 2eikL eikL 1
Σ=− + e 1 1
3 3 e ikL
3 − 2e ikL
1
1 4 2eikL 1 − 3eikL
=− + =− .
3 3 3 − eikL 3 − eikL
Observe that despite the matrix I− U being degenerate whenever eikL =
1, the final expression is well-defined and finite, as promised.
Given the scattering matrix, one can find the spectrum of the graph
using what is sometimes termed the “inside-outside duality”. We state
the theorem in a simplified setting. For a proof, one can consult, e.g.,
[47].

Theorem 5.4.6. Suppose the graph Γ  has at most one lead attached
to any given vertex with the Neumann conditions enforced at the bound-
ary vertices. Suppose that the graph Γ also has Neumann condition at
the same vertices. Then the spectrum of Γ can be expressed as
7 8
σ(Γ) = k 2 : k ∈ Δ or det (I − Σ(k)) = 0 ,
where Δ is the set defined in (5.4.6).
 5.5. FURTHER REMARKS AND REFERENCES 153

Finally, we would like to discuss the intimate connection between

the scattering matrix and the Dirichlet-to-Neumann map which was
described in Section 3.5. Let Γ be a graph with the set B ⊂ V of
boundary vertices. Let Λ(k 2 ) be the corresponding DtN map. Consider
the scattering matrix obtained by attaching one lead to each of the
boundary vertices and enforcing the Neumann conditions there.
Because of the Neumann conditions (more precisely, the continuity
condition), the Dirichlet data at a boundary vertex is given by the value
of the scattering solution at the zero position of the corresponding lead.
Furthermore, because of the current conservation, the Neumann data
(see equation (3.5.6)) is equal to the derivative of the scattering solution
at the boundary vertex.
Using the solution form (5.4.2) on the leads, we can put the above
as
Df (v) = cin out
e + ce ,
 
N f (v) = ik −cin out
e + ce ,
where v is a boundary vertex and e is the lead attached to it. The
vectors cin and cout are connected by equation (5.4.7), therefore
Df = (I + Σ(k)) cin ,
N f = ik (−I + Σ(k)) cin ,
leading to
N f = ik (−I + Σ(k)) (I + Σ(k))−1 Df.
To summarize, we obtain
I − Σ(k)
(5.4.8) Λ(k 2 ) = −ik .
I + Σ(k)
Thus, the Dirichlet-to-Neumann map is the Cayley transform of the
scattering matrix, up to the factor of −k.

5.5. Further remarks and references

Gaps via decorations
The decoration procedure of Section 5.1 was explicitly described by
Aizenman and Schenker [647] for the discrete case. The construction is
easily transferable to quantum graphs [481]. Various indications of this
gap opening effect had been noticed in less explicit form, e.g. in [599].
The “spider” graph decoration procedure was probably used explic-
itly for the first time in [39] (see also [269]) for the same purpose of
gap creation. It is also one of the steps used in the so called Zig-Zag
construction of expanders [627].
154 5. SPECTRA. SPECIAL TOPICS

Nodal count
The formula in Theorem 5.2.4 originally appeared in [356] and is
given here with minor modifications.
The upper bound on the number of nodal domains is a result with
a long history going back to Courant [205,207]. The original proof for
domains in Rd was adapted to metric graphs by Gnutzmann, Smilansky,
and Weber [356], who used the Rd proof from Pleijel [612].
The lower bound for the nodal domains on quantum trees (i.e., when
Γ = 0) was shown by Al-Obeid, Pokornyi and Pryadiev [18, 615, 616]
and by Schapotschnikow [644]. For non-tree graphs ( = 0), the result
was proved in [95].
Similar results are available for nodal count on combinatorial graphs,
see Davies, Gladwell, Leydold and Stadler [225] for the upper bound,
[120] for the tree case and [95] for the non-tree case. A somewhat
outdated survey can be found in [121].
The results on the nodal deficiency and critical partitions were de-
rived in [46]. They were inspired by the result of Helffer, Hoffmann-
Ostenhof and Terracini [409] that the minima of the functional (5.2.15)
on domains in R2 are the nodal partitions of eigenfunctions with zero
nodal deficiency (so-called Courant-sharp eigenfunctions). After the
discovery of the connection between the Morse index and the nodal de-
ficiency on quantum graphs [46], analogs of Theorem 5.2.16 were also
established on domains (billiards) in Rd [105] and discrete graphs [106].
Magnetic perturbations and nodal count
Instead of varying the partition, one can perturb the operator by
magnetic potentials. It is shown in [96, 198] in the discrete graph case
how to relate the corresponding Morse index to the nodal count.
Scattering on quantum graphs
The literature on scattering problems on graphs is vast and growing
and we mention just a selection of works.
Some of the earliest references investigating scattering on graphs
are by Gerasimenko and Pavlov [343], works by Exner and co-authors
(for example, [287], where dependence of resonances on the vertex
condition on a simple graph was investigated), by Kottos and Smilan-
sky [468], where the comparison with predictions of quantum chaos
were made and by Texier and co-workers (e.g. [687]) who discussed
the scattering on graphs from the point of view of mesoscopic theory.
More recently, equivalence of two definitions of resonances, as poles
of the resolvent and poles of the scattering matrix, was established on
graphs in [276]. The anomalous asymptotics of the resonance counting
function were investigated in [221, 224] — the anomaly can be traced
 5.5. FURTHER REMARKS AND REFERENCES 155

to the failure of the unique continuation property discussed in Sec-

tion 3.4. In [47] the scattering approach, via attaching infinite leads
to a compact graph, was used to investigate behavior of zeros of the
graph’s eigenfunctions. The question whether the scattering resonances
can be used to distinguish isospectral graphs was discussed in [51] and
some counter-examples were constructed.
The exposition in section 5.4, especially Theorems 5.4.2 and 5.4.6
follow [47], where complete proofs can be found.
Zeta - functions
Zeta-functions, which we encountered in this chapter, have been
darlings of both mathematics and physics for quite a while. Their
study is very active currently for both combinatorial (see the surveys
[418, 419, 685] and papers [61, 375, 686] for further references) and
quantum graph [250, 320, 401, 402] cases.
 CHAPTER 6

Quantum Chaos on Graphs

Quantum chaos (or “quantum chaology”) is a branch of science that

concerns itself with the signatures of classical chaos in the quantum
world. More precisely, it considers quantum systems whose classical
counterparts are chaotic and asks what features of their spectrum are
different from those of a classically regular (integrable) system. A
standard example of such a classical-quantum pair of systems is, on the
classical side, a billiard in Ω ⊂ R2 where the dynamics is the motion in
a straight line with specular reflections off the hard boundary ∂Ω, and,
on the quantum side, the eigenvalue problem (Helmholtz equation) for
the Laplace operator on Ω with Dirichlet boundary conditions. How
are the dynamical properties of the billiard reflected in the properties
of the eigenvalues of the Laplacian? What about the eigenfunctions?
It has been observed that traces of the classical chaos are found
not in the behavior of individual eigenvalues or eigenfunctions but in
the statistical quantities associated with the spectrum of the quantum
system. Examples of these properties are the empirical distribution
of the eigenvalue spacings and the weak limits of subsequences of the
measures |ψn |2 , where ψn is the n-th eigenfunction.
Since mathematical study of the observed phenomena presented
significant difficulties, quantum graphs were suggested as models that
both display the features of quantum chaos and could be amenable to
complete analytical treatment. This chapter aims to give a taste of the
problems of quantum chaos on graphs, highlighting rigorous results
when they are available. For another review, geared more towards
physicists, we refer the reader to [353].
We will start by describing what is understood as the “classical
motion” on a graph and why it is reasonable to expect a connection
between the properties of the said “motion” and the behavior of the
spectrum. We will then define spectral statistics on graphs and prove
that they exist (even this is hard for billiards). An extremely short
review of two classical random matrix ensembles will be given. Then
we will concentrate on one of the main topics of quantum chaology on
graphs: convergence of the spectral statistics to their predicted form.

157
158 6. QUANTUM CHAOS ON GRAPHS

We will describe one of the two tools that are used to study the above
questions — the periodic orbit expansions (i.e. the trace formula) —
and derive two first terms in the expansion of the form factor.
Of the topics not covered in this chapter, we would like to men-
tion the other tool for deriving spectral statistics, the supersymmetry
integrals. This method is thoroughly reviewed in [353]. Also, we do
not review the eigenfunction statistics, an interesting and still develop-
ing area. More detailed descriptions and references are postponed to
Section 6.5.

6.1. Classical “motion” on graphs

There is no “quantization” procedure for graphs that starts with
a Hamiltonian formulation of the classical dynamics and constructs a
self-adjoint operator corresponding to it. Instead, the classical motion
on graphs is introduced retrospectively and by analogy, to fit with
already described quantum results.
The analogy is rooted in the trace formula, Theorem 3.7.9. The
trace formula involves the stability amplitude (or probability ampli-
tude) Aγ to stay on a given orbit γ. The corresponding probability
is then |Aγ |2 . It is natural to extend this notion from the closed paths
(periodic orbits) to all paths. Thus the probability to stay on the path
γ = (b1 , . . . , bn ) is
Pγ = |Sb1 ,b2 |2 |Sb2 ,b3 |2 · · · |Sbn−1 ,bn |2 .
The product structure means that we can interpret the probability as a
series of independent transitions, or a Markov chain with the transition
matrix
Mb,b = |Sb,b |2 .
Here S is the bond scattering matrix, introduced in section 2.1.2. Be-
cause the matrix S is unitary, the matrix M is stochastic,
Mb,b = 1, for any b.
b
The states of the Markov chain are the bonds of the graph, which can
be interpreted as the phase space: the edge is the position and the
direction of the edge is the momentum.
The described setting allows one to pose questions of the type: If
the classical motion on the graph is “chaotic,” what can be said about
the spectrum of the Schrödinger operator on the graph? Before we
can even attempt to answer it, we need to make precise several points,
including the definition of “chaoticity” and the precise features of the
spectrum one investigates.
 6.2. SPECTRAL STATISTICS AND RANDOM MATRIX THEORY 159

6.2. Spectral statistics and random matrix theory

While individual eigenvalues are usually very sensitive to the pa-
rameters of the problem (such as a quantum graph) in question, it has
been observed that some statistical features are universal. An exam-
ple of such a feature is the nearest-neighbor spacing distribution. Let
{an }∞
n=1 be an ordered sequence with density equal to one, namely

#{an : an < K}
(6.2.1) lim = 1.
K→∞ K
The nearest-neighbor spacing distribution of this sequence is the weak
limit (if it exists)
N
1
(6.2.2) p(x) = lim δ(x − (an+1 − an )).
N →∞ N
n=1

Example 6.2.1. Consider the star graph with d edges of length Lj ,

j = 1, . . . , d. Impose Dirichlet conditions on the vertices of degree 1
and the Neumann-Kirchhoff conditions at the central vertex. Then (cf.
Example 1.4.3) the eigenvalues λn = kn2 can be found as the solutions
of
d
(6.2.3) cot(kLj ) = 0.
j=1

Here we assume that Lj are incommensurate; see Examples 1.4.3 and

2.1.12 for a discussion. While the sequence {λn } has density zero
(since λn ∼ n2 ), the sequence {kn } has density L/π, according to
Lemma 3.7.4. Here L is the total edge length of the graph, L =
L1 + . . . LE . Taking an = Lkn /π, we plotted a numerical approxi-
mation to the nearest-neighbor spacing distribution in Fig. 1. Also in
the same figure, we plotted the distribution for the modification of the
above graph, changing the matching condition at the central vertex to
Dirichlet, in effect decoupling the graph into several disjoint edges.
The two distributions are drastically different. In particular, the
distribution for the Neumann-Kirchhoff graph has level repulsion,
a decreased probability to have small spacing. It appears as if the
eigenvalues prefer to be evenly spaced, without either very large or very
small gaps. The distribution for the Dirichlet graph, on the contrary,
appears exponential, as if {kn } are events in a Poisson process.
The reader familiar with quantum chaos should be warned that de-
spite visual similarity to the spacing distribution of the Gaussian Or-
thogonal Ensemble (see Section 6.2.2 for explanations), the star graphs
160 6. QUANTUM CHAOS ON GRAPHS

0.9

0.8

0.7

0.6

0.5
p(k)

0.4

0.3

0.2

0.1

0
0 1 2 3 4 5 6 7
k (density 1)

Figure 1. Nearest-neighbor spacing distribution com-

puted numerically for two versions of a star graph with
8 edges. One star has Neumann-Kirchhoff conditions
at the central vertex (solid line), the other has Dirich-
let condition (dashed line), which decouples the edges.
Both graphs have Dirichlet conditions at the peripheral
vertices. To produce the plot, 106 eigenvalues were com-
puted for each graph.

with Neumann condition at the center do not have “universal” or “ran-

dom matrix” statistics [97, 100].
Another frequently used statistic is the two-point correlation
function. Informally speaking, it is the frequency of finding two values
an and an a distance x apart. We will define it as the weak limit (if it
exists)
N ∞
1
(6.2.4) R2 (x) = lim δ (x − (an − an )) .
N →∞ N
n=1 n =1

Note that the summation is over two indices but the normalization is
1/N ; the resulting function is not a probability distribution. We can
also express the function R2 (x) using the density of states of the
 6.2. SPECTRAL STATISTICS AND RANDOM MATRIX THEORY 161

sequence {an },
∞
(6.2.5) d(x) = δ(x − an ).
n=1

If the sequence {an } has density 1, then R2 (x) is given by a convolution-

like integral,
T
(6.2.6) R2 (x) = lim d(y)d(x − y)dy.
T →∞ 0

The Fourier transform of the correlation function is known as the form

factor,

(6.2.7) K(τ ) = Fx→τ {R2 } = R2 (x)e−2πixτ dx,

where the second equality is formal, as R2 (x) is a distribution. Using

the trace formula, one can derive a nice expression for the form factor
in terms of the periodic orbits, a basis for the semiclassical methods of
evaluation.

6.2.1. Form factor of a unitary matrix. For future use, we

will study in detail an example where the sequence {an } comes from
eigenphases of a unitary matrix. More precisely, let U be a B × B
unitary matrix with eigenvalues eiθj , j = 1, . . . , B. We would like to
define the two-point correlation function of the numbers θj . To use
definition (6.2.4), we cyclically “unfold” the phases,
B
(a1 , a2 , . . .) =(θ1 , . . . , θB , θ1 + 2π, . . . , θB + 2π, θ1 + 4π . . .).
2π
Note that we included the density factor B/2π to ensure that the re-
sulting sequence {an } is density one. Heuristically, the density of {θj }
is B eigenphases in an interval of length 2π. Upon unfolding, this
heuristic explanation fits with the result of definition (6.2.1).
Applying definition (6.2.4) to the sequence {an } and taking the
limit we get
∞  
B 
B
1
(6.2.8) R2 (x) = δ x− θj − θj  + 2πn ,
B  n=−∞ 2π
j,j =1

or, using the Dirac comb function

∞
(6.2.9) ΔB (x) = δ(x − Bn),
n=−∞
162 6. QUANTUM CHAOS ON GRAPHS

we can write R2 (x) as

 
B 
B
1
(6.2.10) R2 (x) = ΔB x− θj − θj  .
B j,j  =1
2π
The Fourier transform of the Dirac comb function is the subject of
the Poisson summation formula,
1
(6.2.11) Fx→τ {ΔB } = Δ1/B (τ ).
B
Applying this formula to equation (6.2.10) we get
B
1 
K(τ ) = 2 Δ1/B (τ ) eiBτ (θj −θj )
B j,j  =1

n!
∞ B
1 
= 2
δ τ− ein(θj −θj ) .
n=−∞
B B j,j  =1

Since einθj is the n-th power of the j-th eigenvalue of U , we finally have
n! 1
∞
1
(6.2.12) K(τ ) = δ τ− |Tr U n |2 .
n=−∞
B B B
This result is sometimes written as
1
(6.2.13) K(n/B) = |Tr U n |2 .
B
Note that the factor of 1/B in front of the delta-function in (6.2.12)
allows the result to converge weakly if the limit B → ∞ is taken (the
other ingredient is the convergence of K(n/B) to a limit as n → ∞,
B → ∞, n/B = const).
6.2.2. Random matrices. It has been a long-standing conjec-
ture, usually attributed to Bohigas, Giannoni and Schmit [125], that
the spectrum of a typical classically chaotic system displays correla-
tions identical to the correlations of the eigenvalue of a large random
matrix with appropriate symmetries. In particular, after appropriate
rescaling, the nearest-neighbor spacing distribution, the two-point cor-
relation function and the form factor are to coincide with the N → ∞
forms of the corresponding function from GOE (assuming time-reversal
symmetry).
Definition 6.2.2. Gaussian Orthogonal Ensemble (GOE) is
a space of random real symmetric N × N matrices H with independent
entries hj,k , j ≤ k,√ distributed normally according to hj,j ∼ N (0, 1)
and hj,k ∼ N (0, 1/ 2) (see Remark 6.2.4 for explicit formulae).
 6.2. SPECTRAL STATISTICS AND RANDOM MATRIX THEORY 163

Remark 6.2.3. Only the entries on and above the diagonal are
independent, the entries below the diagonal being determined by the
requirement that H is symmetric.
Remark 6.2.4. The probability density functions are
1 1
p(hj,j ) = √ e−hj,j /2 , p(hj,k ) = √ e−hj,k .
2 2
(6.2.14)
2π π
An easy calculation shows that the joint p.d.f. of the entries of the
matrix is
 
N
e−hj,k /2 = AN e− Tr(H )/2 ,
2 2
(6.2.15) p(H) = p(hj,k ) = AN
1≤j≤k≤N j,k=1

where
AN = 2−N/2 π −(N
2 +N )/4
(6.2.16) .
Remark 6.2.5. The probability distribution of H is invariant with
respect to the orthogonal changes of basis, H → RT HR, where R is a
real orthogonal matrix. This symmetry is the origin of “orthogonal” in
the name of the ensemble. This property can also be used as the defi-
nition instead of directly specifying the distribution (the independence
still has to be postulated).
The eigenvalue statistics of the GOE matrices model the energy
statistics of complex systems that are time-reversal invariant. A similar
role for systems with broken time-reversal symmetry is played by the
Gaussian Unitary Ensemble.
Definition 6.2.6. Gaussian Unitary Ensemble is a space of
random Hermitian N ×N matrices H with √ independent entries hj,k , j ≤
k, distributed according to hj,j ∼ N (0, 1/ 2) and hj,k ∼ N (0, 1/2) +
iN (0, 1/2).
Remark 6.2.7. The joint p.d.f. of the matrices from GUE is
p(H) = AN e− Tr(H ) , AN = π −N /2 2(N −N )/2 .
2 2 2
(6.2.17)
This density is invariant under a unitary change of basis, H → U ∗ HU .
The spectral statistics described at the beginning of this section can
be suitably defined and computed for the random matrix ensembles.
Namely, we first rescale the spectrum to have density 1 (the “semi-
circle law” is to be used here). Then the desired function, such as the
two-point correlator, is computed and averaged over the corresponding
probability ensemble. This yields an answer which, in general, is depen-
dent on N . However, the N → ∞ limit of the function is what interests
164 6. QUANTUM CHAOS ON GRAPHS

us. The answers for the form factor, which we will study below, are as
follows (in the limit N → ∞):

|τ |, |τ | ≤ 1,
(6.2.18) K(τ )GUE =
1, |τ | ≥ 1.
  
|τ | 2 − ln(2|τ | + 1) , |τ | ≤ 1,
(6.2.19) K(τ )GOE = |+1
2 − |τ | ln 2|τ
2|τ |−1
, |τ | ≥ 1.
Derivations and other relevant formulas are available in the standard
references [388,537]; a concise list of formulas for the nearest-neighbor
spacing distribution and the two-point correlation function is given
in [353, Appendix B].

6.3. Spectral statistics of graphs

In their ground-breaking article, Kottos and Smilansky [466] pro-
posed to study the spectral statistics of large quantum graphs. They ar-
gued that, if the graph is essentially disconnected (as in Example 6.2.1
in the case of Dirichlet condition at the central vertex), the spectrum
is a superposition of “independent” spectra of constituent parts (edges
in Example 6.2.1) and thus has characteristics of events in a Poisson
process. If the graph is highly connected, such as a complete graph, the
eigenvalue statistics were numerically shown to follow the correspond-
ing random matrix prediction: GUE with the magnetic field (which
breaks the time-reversal symmetry) and GOE without.
Remark 6.3.1. It was later shown that the star graph with N edges
and Kirchhoff-Neumann condition at the central vertices, considered in
Example 6.2.1, does not display random matrix-like statistics.
For a compact quantum graph, the natural√ sequence to calculate
statistics of is the rescaled levels kn = λn . More precisely, the
sequence an = Lkn /π has density one according to the Weyl Law,
Lemma 3.7.4. However, it was observed in [467] that one can instead
look at the spectrum of the matrix S(k)eikL (see section 2.1.2), where
S(k) is the bond scattering matrix and L is the diagonal matrix of edge
lengths.
For simplicity we will assume that the matrix S is k-independent.
The matrix SeikL is unitary of size B = 2E, with eigenvalues on the
unit circle. Denoting the eigenvalues of SeikL by eiθj , j = 1, . . . , B such
that
(6.3.1) 0 ≤ θ1 ≤ . . . ≤ θB ≤ 2π,
 6.3. SPECTRAL STATISTICS OF GRAPHS 165

we can look, for example, at the nearest neighbor spacing distribution

 
B 
B
1
(6.3.2) pθ (x, k) = δ x− θj (k) − θj−1 (k) ,
B j=1 2π
with the convention θ0 := θB − 2π. Note the rescaling factor B/2π to
normalize the “density” to 1. Equation (6.3.2) is in fact the result of
application of definition (6.2.2) to the unfolded sequence of eigenphases
from Section 6.2.1.
We can now average the density p(x, k) with respect to k, over a
large discrete or continuous set, for example
1 K
(6.3.3) pθ (x) = lim pθ (x, k)dk.
K→∞ K 0

The quantity pθ (x) is much more efficient in numerical computation

than finding the eigenvalues kn2 of the graph. It turns out that in the
appropriate limit it is equivalent to computing p(x), equation (6.2.2),
with an = Lkn /π.
Theorem 6.3.2 ([56, 109]). Suppose the graph Γ is such that the
bond scattering matrix S is k-independent. Let h be a continuous func-
tion. If the edge lengths (L1 , . . . , LE ) are linearly independent over Q,
the limits
N  
1 L
(6.3.4) pk [h] := lim h (kn − kn−1 )
N →∞ N π
n=1
and
K B  
1 1 B
(6.3.5) pθ [h] = lim h (θj (k) − θj−1 (k)) dk
K→∞ K 0 B 2π
j=1

exist and define bounded linear functionals of h. If we take the limit

(6.3.6) (L1 , . . . , LE ) → (0 , 0 , . . . , 0 )
for some 0 > 0, while keeping the lengths linearly independent over Q,
the two functionals coincide. Namely,
(6.3.7) lim pk [h] = lim pθ [h],
ΔL→0 ΔL→0
where ΔL := Lmax − Lmin is the spread of the distribution of the edge
lengths.
Remark 6.3.3. If the edge lengths are not linearly independent
over Q, the limits (6.3.4) and (6.3.5) still exist but are highly sen-
sitive to the changes in individual bond length. Thus the limits in
equation (6.3.7) are, in general, singular. In particular, when all bond
166 6. QUANTUM CHAOS ON GRAPHS

k2
Σ

k3 L2
k1
k4
L1

Figure 2. A schematic depiction of the flow piercing

the surface Σ. The direction of increasing k is indicated
by an arrow, and eigenvalues are indicated by little cir-
cles.

lengths are equal (ΔL = 0), the k-spectrum is periodic and the spac-
ings in the θ-spectrum are independent of k. The nearest-neighbor
distribution in this case is highly degenerate.
We will now describe the idea of the proof of Theorem 6.3.2. To
show the existence of the limiting density pk (x), we define the function
 
F (φ1 , . . . , φE ) = det I − SeiΦ ,
where
Φ = diag(φ1 , . . . , φE , φ1 , . . . , φE )
and observe that the secular equation, whose roots are kn (see Sec-
tion 2.1.2, Eq. (2.1.17)), can be written as
F (kL1 , . . . , kLE ) = 0.
The function F is defined on the torus T = [0, 2π]E . Let Σ be the
subset of the torus where F = 0. Then the roots kn are the “times” of
piercing of the surface Σ by the flow
(6.3.8) φ1 = kL1 , ..., φE = kLE ,
see Figure 2.
We can define a function d(φ) on the surface Σ which gives the
“time” to the next piercing. Then, by ergodicity of the flow (6.3.8),
the distribution pk [h] is obtained by integrating h (Ld(φ)/π) over the
surface Σ with the appropriate measure.
 6.4. PERIODIC ORBIT EXPANSIONS 167

Similarly, we can define the function on the whole torus,

 
B 
B
1
h θj (φ) − θj−1 (φ) , φ ∈ T,
B j=1 2π

where θj are the phases of the eigenvalues of the matrix SeiΦ . We again
use the ergodicity of the flow to show that pθ [h] is equal to the integral
of this function over the whole torus.
Finally, to show that the two distributions agree in the limit (6.3.6),
we use the bounds provided by Lemma 3.7.2. We conclude that, at
any point φ ∈ T, there is a relationship between the time to the next
piercing and a θ-spacing, namely
θr+1 (φ) − θr (φ)
≈ d(φ),
l0
where r is the index of the phase that is equal to 0 (existence of one
follows from the definition of F and Σ). Then in the limit (6.3.6),
Ld(φ)/π = B θr+1 (φ) − θr (φ) /2π.
Remark 6.3.4. The result of Theorem 6.3.2 can be extended to
r-th nearest neighbor spacing distributions, i.e. the distributions of
an − an−r , with suitable an . Moreover, when all r-th nearest neighbor
spacing distributions coincide, so do other statistical functions such as
the n-point correlation functions. (See, for example, equation (6.1.39)
of [537].)
Since the nearest neighbor spacing distribution is determined by the
surface Σ which depends, in particular, on the graph’s topology, one
cannot expect agreement with the random-matrix prediction for any
given graph. Instead, the agreement is expected in the limit, as the
size of the graphs increases. This raises an important question: what
conditions on the sequence of graphs should be imposed. A conjecture
by Tanner [684], which will be reviewed below, involved a condition
on the gap in the spectrum of the “classical” matrix M , defined in
Section 6.1. Supersymmetric evaluation by Gnutzmann and Altland
[349, 350] further strengthened Tanner’s condition on the gap.
6.4. Periodic orbit expansions
Given a compact quantum graph with edge lengths independent
over Q, one can combine the trace formula, Theorem 3.7.9, with defi-
nitions (6.2.6) and (6.2.7) to obtain an expression for the form-factor,
∞  
1 Lp Lq Lp
(6.4.1) K(τ ) = Ap Aq δ τ − δLp ,Lq .
4L2 n=1 p,q∈P rp rq 2L
n
168 6. QUANTUM CHAOS ON GRAPHS

Here the sum is over pairs of periodic orbits p and q, Ap stands for
the stability amplitude of p and Lp for the length (the definitions were
given in Section 3.7.4). Note that there are finitely many orbits with
τ = Lp /2L in any given interval, therefore K(τ ) is well-defined as a
distribution.
A notable feature of (6.4.1) is that the double sum runs only over
the pairs of orbits of the same length.

Definition 6.4.1. Periodic orbits p and q are said to belong to

the same degeneracy class if, for any choice of edge lengths of the
graph, Lp = Lq . In other words, q traverses the same edges as p the
same number of times (although, possibly, in a different order and in
different directions). The degeneracy class of the orbit p will be denoted
by {p}.

We will not prove equation (6.4.1) (see Section 6.5 for references),
but will compare it with the form factor of the eigenphases of the matrix
SeikL in the limit when all edge-lengths become equal.

Definition 6.4.2. The reversal of a periodic orbit p = [b1 , . . . , bn ]

is the orbit p̄ = [b¯n , . . . , b¯1 ], where b̄ = (e, o) is the reversal of the bond
b = (o, e).

The properties of the reversal of an orbit that we will use are

(1) Lp̄ = Lp ,
(2) If the bond scattering matrix S is k-independent, for any orbit
p,

(6.4.2) Ap̄ = Ap ,

where on the right-hand side we have complex conjugation.

The first property follows from the fact that edge lengths are in-
dependent of direction, Lb = Lb̄ . The second property follows from
Theorem 2.1.6. Indeed, at every vertex v, the scattering matrix σ (v)
is k-independent and unitary. By Theorem 2.1.6, part 3, σ (v) is also
Hermitian. Finally, according to the definition of S, equation (2.1.15),

(v) (v)
Sb̄,b¯ = σb̄,b¯ = σb ,b = Sb ,b ,

where v = t(b) = t(b̄ ) = o(b ) = o(b̄). We refer the reader to Re-

mark 2.1.2 for information regarding indexing of the matrix σ (v) .
 6.4. PERIODIC ORBIT EXPANSIONS 169

Now we can change the orbit q to q̄ in formula (6.4.1) to obtain

∞  
1 Lp Lq̄ Lp
K(τ ) = Ap Aq̄ δ τ − δLp ,Lq̄
4L2 n=1 p,q∈P rp rq̄ 2L
n

∞   2
1 L Lp
= δ τ− Ap .
4L2 n=1 L
2L p∈Pn , Lp =L
rp

In terms of the closed paths the sum is slightly different,

∞   2
1 L Lγ
(6.4.3) K(τ ) = δ τ− Aγ .
4L2 n=1 L
2L γ∈Cn , Lγ =L
n

On the other hand, the form factor of the eigenphases of the matrix
ikL
Se , according to (6.2.12), is equal to
n!9  ikL n 2 :
∞
1
Kθ (τ ) = 2
δ τ − Tr Se .
n=−∞
B B k

We now expand the trace and perform the average with respect to k,
9  n 2 : K
2

Tr SeikL = lim Aγ e ikLγ

dk
k K→∞ 0 γ∈Cn
; <
= Aγ Aγ  eik(Lγ −Lγ  ) k = Aγ Aγ  δLγ ,Lγ  .
γ,γ  ∈Cn γ,γ  ∈Cn

Note that we can take the limit term-wise as the sums are all finite.
Finally, we get
2
n!
∞
1
(6.4.4) Kθ (τ ) = δ τ − Aγ .
n=−∞
B2 B L γ∈Cn , Lγ =L

Now we can compare equations (6.4.3) and (6.4.4) in the limit when
all lengths tend to the same value l0 . We observe that Lγ → l0 n,
2L → l0 B and the two distributions obviously converge, in agreement
with Theorem 6.3.2.
We can now make the form-factor into a step function with steps of
width 1/B centered at τ = n/B, such that the weak limit is unaffected.
The value around point τ = n/B is then
2
1
(6.4.5) K(n/B) = Aγ ,
B L γ∈Cn , Lγ =L
170 6. QUANTUM CHAOS ON GRAPHS

with a slight abuse of notation. We will now explain the ideas behind
the periodic orbit evaluation of the form factor starting from equa-
tion (6.4.5).
6.4.1. On time-reversal invariance.
Definition 6.4.3. A quantum graph is time-reversal invariant
if, for any path γ = (b1 , . . . , bn ), the stability amplitudes of γ and its
reversal γ̄ = (b¯n , . . . b¯1 ) coincide,
(6.4.6) Aγ = Aγ̄ .
A standard way to break time-reversal invariance is to introduce
a magnetic field. It can be easily shown that this results in multi-
plying the bond scattering matrix on the right by a diagonal matrix
diag(eiφb1 , . . .), where φb is the integral of the magnetic potential Ab (x),
see (1.4.3), over the edge b. Note that φb = −φb̄ , so that the phase de-
pends on the direction of traversal of an edge. However, not for all
paths, closed or open, is equation (6.4.6) broken. For example the
closed path γ = (b1 , b¯1 , b2 , b¯2 ) satisfies γ = γ̄ and yet Aγ = Aγ̄ for all
choices of the phases φb . The same applies to all tree-like paths. In
particular, the magnetic field has no effect on the closed paths on a tree
graph (see Section 2.6). However, it can be argued that for a typical
well-connected graph with rationally independent magnetic phases φb ,
a typical closed path is not time-reversal invariant.
6.4.2. Diagonal approximation. The philosophy of evaluating
a periodic orbit (or closed path) expansion like
2

Aγ = Aγ Aγ  δLγ ,Lγ  ,
γ∈Cn , Lγ =L γ,γ  ∈Cn

see (6.4.4), is to try to select dominant contributions by choosing “fami-

lies” of pairs of paths γ and γ  that are related to each other by a simple
transformation. This ensures that the phases of Aγ Aγ  are aligned (for
example, if all Aγ Aγ  are positive), reducing cancellation. One strives
to make this classification into families as exhaustive as possible. How-
ever, even in the simplest cases, there are still no mathematical proofs
that the overlaps between families or the omissions are insignificant.
The diagonal approximation is the first step in this classification.
Assuming the presence of TR-invariance, we first evaluate the contri-
bution of the terms with p = p or p = p̄, p ∈ Pn . In terms of closed
paths (which, unlike orbits, have a definite starting point), we are look-
ing at γ  = σ k (γ) or γ  = σ k (γ) for some k, where σ is the cyclic shift
 6.4. PERIODIC ORBIT EXPANSIONS 171

operator. To facilitate future notation, we define the set of diagonal

transformations,
Td = {σ k , σ k : k = 0 . . . n − 1}.
Now, if γ  = τd (γ), τd ∈ Td , then Aγ Aγ  = |Aγ |2 and, of course, Lγ =
Lγ  . The total contribution of such pairs to the form-factor can be
written as
1
K d (n/B) = |Aγ |2 |Td (γ)| ,
B γ∈C
n

where Td (γ) is the set of (distinct) paths that result from applying
transformations from Td to γ. In other words, Td (γ) is the orbit of γ
under Td .
By the orbit-stabilizer theorem,
|Td | 2n
|Td (γ)| = = .
|{τ ∈ Td : τ (γ) = γ}| |StabTd (γ)|
There are two reasons why there could be more than one element in
the stabilizer of γ:
(1) the path γ is not primitive, i.e., there is an integer r that
divides n and σ n/r (γ) = γ.
(2) the path γ is a (shift of a) palindrome: for some k, σ k (γ) =
(b1 , b2 , . . . , b2 , b1 ). In this case, σ k (γ) = σ k (γ).
The number of closed paths of repetition r is equal to the number of
primitive closed paths of length n/r. In the case of the palindromes,
r is essentially 2: the number of palindromes is equal to the number
of paths (not necessarily closed) of length n/2. Combined with simple
exponential estimates of the number of paths of length n, one can esti-
mate the number of γ with non-trivial stabilizer as being exponentially
smaller than the number of all closed paths of length n. Furthermore,
if n is prime, there are no such γ. Altogether, it can be shown that
1  
K d (n/B) = 1 + O(e−αn ) |Aγ |2 |Td |,
B γ∈C
n

for some α > 0. Since |Td | = 2n, we now need to sum |Aγ |2 over all
closed paths γ. These quantities are what we called the “probabilities”
to stay on the path γ is Section 6.1. Defining the B × B matrix M by
Mb ,b = |Sb ,b |2 ,
we find
|Aγ |2 = Tr M n .
γ∈Cn
172 6. QUANTUM CHAOS ON GRAPHS

We now impose some natural conditions on the graph:

(1) the vertex conditions are strongly coupling:
Sb ,b = 0 if t(b) = o(b ),
(2) the graph is aperiodic: there is no integer k > 1 such that
periods of all closed paths are divisible by k.
Then the Markov chain defined by M is ergodic and Perron-Frobenius
theorem implies that M has a single eigenvalue on the unit circle. This
eigenvalue is equal to 1 (since M is stochastic); all other eigenvalues
are strictly less than 1 in absolute value. Therefore, as n → ∞ we have
Tr M n → 1.
However, if we are to expect agreement with RMT predictions we
need to take the limit n → ∞ together with the limit B → ∞, so that
τ = n/B is fixed. The matrix M depends on B and we need some
type of “uniform ergodicity” condition. We formulate one sufficient
condition as a lemma.
Lemma 6.4.4. Let {M (B)} be a sequence of stochastic matrices
of increasing size B. Order the spectrum of each of the matrices by
decreasing absolute value, that is
1 = λ1 ≥ |λ2 | ≥ . . . ≥ |λB |.
Then,
B
(6.4.7) (1 − |λ2 (B)|) → ∞
ln B
implies Tr M n → 1 in the limit n, B → ∞, n/B → τ > 0.
Proof. The proof is a simple exercise in inequalities. We expand
the trace in terms of the eigenvalues,
B
|Tr M n − 1| ≤ |λj |n ≤ (B − 1) exp(n ln |λ2 |) = exp(−R),
j=2

where the exponent R can be written as

 
n B ln(B − 1)
R= (− ln |λ2 |) − ln B.
B ln B ln B
Since − ln |λ2 | > 1 − |λ2 |, condition (6.4.7) implies that R → ∞ in the
limit B → ∞, n/B → τ > 0. 
Assuming (6.4.7) we now get
2n
(6.4.8) K d (n/B) → K d (τ ) = lim = 2τ.
B
 6.4. PERIODIC ORBIT EXPANSIONS 173

This result agrees with the first term in the expansion of the GOE form
factor, equation (6.2.19)
Finally, when the TR-invariance is broken, the group of transfor-
mations Td does not include reversals, Td = {σ k : k = 0 . . . n − 1}. The
rest of the derivation is still valid, leading to the diagonal term
n
(6.4.9) K d,NTR (n/B) → K d,NTR (τ ) = lim = τ.
B
Again the answer agrees with the leading term of the small τ expansion
of the corresponding result from the Random Matrix Theory, (6.2.18).

6.4.3. The simplest example of an off-diagonal term. The

main difficulty in evaluating further terms in the expansion of the form
factor is that expression (6.4.5) is absolutely divergent. More precisely,
⎛ ⎞2
1 ⎝
(6.4.10) |Aγ |⎠ → ∞,
B L γ∈C , L =L
n γ

as B → ∞, n/B → τ . There is no known method to evaluate the

contribution of a degeneracy class {γ ∈ Cn , Lγ = L} for graphs that
exhibit universal statistics.1 The progress is achieved by expanding
the square in (6.4.10) and selecting families of pairs of paths that have
equal length and are related to each other by a specified transformation
(the simplest case being the diagonal approximation). Then the limit
B → ∞ is taken for the contribution of each family. However, doing
such re-grouping of terms in a conditionally convergent series is fraught
with danger, and careful bookkeeping of paths is required to avoid
getting infinities. There are different methods of bookkeeping; none
yield mathematically rigorous results. Here we follow the outline of
[94], which is the closest to the method used in general chaotic systems
[561], which is now the industry standard (see Section 6.5 for further
references).
We will consider the first off-diagonal correction to the form factor
in the TR-invariant case. It is somewhat easier to argue that there
are no further corrections to the leading term (6.4.9) in the NTR case,
see [108, Sec. 4], but we feel it is more instructive to derive a non-zero
result in the TR case.
The first off-diagonal correction comes from the Sieber-Richter pairs,
also called “figure of eight” paths. We will say that a path γ =
(b1 , . . . , bn ) has a self-intersection at vertex v if there are indices k
1It has been done for star graphs, but they have so-called “intermediate”
statistics.
174 6. QUANTUM CHAOS ON GRAPHS

f1 s2
k-1 k

arc 1 arc 2
v
j+1 j
s1 f2

Figure 3. Notation around a self-intersection. The

partner orbit is obtained by reversing the directions of
the bonds number k and j and also reversing arc 2.

and j such that k = j + 1 and o(bk ) = t(bj ), see Fig. 3. Then the two
partner paths are
γ = (b1 , . . . , bk−1 , bk , bk+1 , . . . , bj−1 , bj , bj+1 , . . . , bn ),
(6.4.11) γ  = (b1 , . . . , bk−1 , bj , bj−1 . . . , bk+1 , bk , bj+1 , . . . , bn ).
It is easy to verify that γ  is a legitimate path and has the same length
as γ. Note that we also allow k > j; the general rule is that we reverse
all the bonds starting with bk , cycling from bn to b1 if necessary, and
until we have reversed bj .
The outline of the summation is as follows. In analogy to Sec-
tion 6.4.2 we define a family of transformations T8 = {τk,j }, parame-
terized by k and j, k = j + 1. The action of the transformation τk,j on
a closed path is described above. Of course it only acts on the paths
satisfying o(bk ) = t(bj ). For each k and j we will sum over all paths
satisfying this condition. Then we will sum over all choices of k and j.
In addition, each cyclic shift of γ  is a suitable partner for γ; we will
account for this fact by an addition factor of n.
However, there is a complication. If, in addition to o(bk ) = t(bj ), it
happens that bk = bj then the action of τk+1,j−1 on γ yields
γ  = (b1 , . . . , bk−1 , bk , bj−1 . . . , bk+1 , bj , bj+1 , . . . , bn ),
which coincides with γ  . We are thus in danger of over-counting pairs
of orbits. We will see later that this would produce infinity in the
limit B → ∞. An analogous problem arises if bk−1 = bj+1 . To avoid
over-counting, we will first count the orbits with
(6.4.12) ( = 0) bk−1 = bj+1 , bk = bj ,
then with
( = 1) bk−1 = bj+1 , bk = bj , bk+1 = bj−1 ,
( = 2) bk−1 = bj+1 , bk = bj , bk+1 = bj−1 , bk+2 = bj−2 ,
 6.4. PERIODIC ORBIT EXPANSIONS 175

...

Figure 4. The “figure of eight” orbits with encounter

length 0, 1 and 2. To ensure that the length is well
defined, we impose restrictions on both sides of the in-
tersection. The restrictions are drawn as double-ended
arrows.

and so on. The structure of these classes of orbits is shown in Fig. 4.

The parameter  is the length of the “extended” intersection, or en-
counter.
We start with the case  = 0, which happens to be the trickiest.
First we simplify the amplitude product Aγ Aγ  . We have
Aγ = · · · Sbk−2 ,bk−1 Sbk−1 ,bk Sbk ,bk+1 · · · Sbj−1 ,bj Sbj ,bj+1 Sbj+1 ,bj+2 · · · .
Using TR-invariance of the S-matrix, Sb,b = Sb ,b , we can write the
amplitude Aγ  as
Aγ  = · · · Sbk−2 ,bk−1 Sbk−1 ,bj Sbk ,bk+1 · · · Sbj−1 ,bj Sbk ,bj+1 Sbj+1 ,bj+2 · · · .
This suggests to introduce a special notation for the common parts
(called arcs),
A1 = Sbj+1 ,bj+2 · · · Sbk−2 ,bk−1 , A2 = Sbk ,bk+1 · · · Sbj−1 ,bj ,
which allow us to write
Aγ Aγ  = |A1 |2 |A2 |2 Sbk−1 ,bk Sbk−1 ,bj Sbj ,bj+1 Sbk ,bj+1
= |A1 |2 |A2 |2 σf1 ,s2 σf1 ,f2 σf2 ,s1 σs2 ,s1 ,
where we used the notation (see Fig. 3)
bk−1 = (f1 , v), bk = (v, s2 ), bj = (f2 , v), bj+1 = (v, s1 ),
and σ = σ (v) is the vertex scattering matrix of the vertex v.
Note that the restrictions of equation (6.4.12) impose some condi-
tions f1,2 and s1,2 , namely that s1 = f1 and s2 = f2 . This implies that
there is at least one bond between bj and bk , and at least one bond be-
tween bj+1 and bk−1 . We can write it symmetrically as |k − j − 1| ≥ 3,
where the distance | · | is understood to be on a cycle 1, . . . , n. Thus,
the sum over all possibilities is realized by summing
(1) over k, j ∈ {1, . . . , n}, |k − j − 1| ≥ 3,
(2) v ∈ V,
(3) s1 , s2 , f1 , f2 ∈ V, adjacent to v and s1 = f1 , s2 = f2 ,
176 6. QUANTUM CHAOS ON GRAPHS

(4) path 1 from bj+1 to bk−1 ,

(5) path 2 from bk to bj .
Summation over paths 1 and 2 are to be done first. We observe that
the corresponding sums of |A1 |2 and |A2 |2 are nothing else but classical
probabilities, in the sense of Section 6.1, to go from bj+1 to bk−1 in
k − j − 2 steps and from bk to bj in j − k steps.2 At this point we make
a simplifying assumption.
Assumption 6.4.5 (Almost-instant equilibration). We will as-
sume that the classical probability to go from b to b in t ≥ 2 steps is
already at its ergodic limit 1/B. Examples of graphs possessing such
properties are DFT star graphs and DFT complete graphs (see Sec-
tion 2.2, especially the discussion around equation (2.2.11)).
Analogous assumptions are made (implicitly) in the literature about
general quantum chaotic systems. An enthusiastic reader can actually
adapt the calculations below to require much less restrictive conditions
on the classical evolution on the graph (or see [93, 107] for alternative
methods). The trick is to exhibit exact cancellations between subfam-
ilies of path pairs. However, calculations under less restrictive condi-
tions become much harder when one tries to evaluate further terms in
the form factor expansion [93].
With the above assumption, we get for the contribution of the  = 0
family,
n2 (n − 5)
C0 = σf1 ,s2 σf1 ,f2 σf2 ,s1 σs2 ,s1 (1 − δs1 ,f1 ) (1 − δs2 ,f2 ) ,
B2 v,s ,s ,f ,f
1 2 1 2

where the factor n(n − 5) counts the number of possible values of j and
k. Indeed, j can take any value between 1 and n, k can take any value
apart from those that result in k − j − 1 being −2, −1, 0, 1 or 2. Also,
we have an extra factor of n accounting for possible shifts of γ  and we
added Kronecker deltas to enforce the condition si = fi .
We now sum over, for example, f2 . Due to unitarity and symmetry
of σ,
σf2 ,s1 σf1 ,f2 = δs1 ,f1 ,
f2
which, multiplied by (1 − δs1 ,f1 ), is zero. Thus we get
n2 (n − 5)
C0 = − σf1 ,s2 σf1 ,f2 σf2 ,s1 σs2 ,s1 (1 − δs1 ,f1 ) δs2 ,f2 .
B2 v,s 1 ,s2 ,f1 f2

2When the number of steps is calculated, n should be added to a negative

answer.
 6.4. PERIODIC ORBIT EXPANSIONS 177

f1 s2

v1 v2
s1 f2

Figure 5. Notation around an encounter of length l = 1.

Because of the Kronecker delta, the sum over f2 is now trivial: we

substitute s2 for f2 everywhere. In fact, for future comparisons, we
will replace both by v2 and rename v to v1 .
n2 (n − 5) 2
C0 = − S(f1 ,v1 )(v1 ,v2 ) |S(v2 ,v1 )(v1 ,s1 ) |2 (1 − δs1 ,f1 ) .
B2 v 1 ,v2 s1 ,f1

Before we simplify this expression further, let us look at the case

 = 1. The first difference is in the possible values of j and k. There is
still at least one bond between bj+1 and bk−1 but on the other side the
same now applies to bk+1 and bj−1 . To put it symmetrically, |k −j| ≥ 4.
There are now n(n − 7) possibilities for the choice of k and j. The sum
is now
n2 (n − 7)
C1 = (1 − δs1 ,f1 ) (1 − δs2 ,f2 )
B2 v 1 ,v2 s1 ,s2 ,f1 ,f2

× |S(f1 ,v1 ),(v1 ,v2 ) |2 |S(v1 ,v2 ),(v2 ,s2 ) |2 |S(f2 ,v2 ),(v2 ,v1 ) |2 |S(v2 ,v1 ),(v1 ,s1 ) |2 ,
see Fig. 5 for notation. In a similar fashion we multiply out (1 − δs2 ,f2 )
and sum over f2 and s2 . In the first term we use

(6.4.13) |S(v1 ,v2 ),(v2 ,s2 ) |2 = |S(f2 ,v2 ),(v2 ,v1 ) |2 = 1.

s2 f2

In the second term we substitute both s2 and f2 with v3 ,

n2 (n − 7)
C1 = (1 − δs1 ,f1 ) |S(f1 ,v1 ),(v1 ,v2 ) |2 |S(v2 ,v1 ),(v1 ,s1 ) |2
B2 v 1 ,v2 s1 ,f1

n (n − 7)
2
− (1 − δs1 ,f1 )
B2 v 1 ,v2 ,v3 s1 ,f1 ,

× |S(f1 ,v1 ),(v1 ,v2 ) | |S(v1 ,v2 ),(v2 ,v3 ) |2 |S(v3 ,v2 ),(v2 ,v1 ) |2 |S(v2 ,v1 ),(v1 ,s1 ) |2 .
2
178 6. QUANTUM CHAOS ON GRAPHS

...

Figure 6. A schematic depiction of the edge configura-

tions entering the sum in equation (6.4.14). The dotted
arrows denote the transitions whose amplitudes enter the
sum.

Remarkably, the first term of C1 is identical, apart from the prefactor,

to the last expression for C0 .
Evaluating C2 and so on, in a similar fashion, and adding them up
we get
(6.4.14)

C0 + C1 + C2 + . . .

2n2
=− 2 (1 − δs1 ,f1 ) |S(f1 ,v1 ),(v1 ,v2 ) |2 |S(v2 ,v1 ),(v1 ,s1 ) |2
B s1 ,f1 v1 ,v2

+ |S(f1 ,v1 ),(v1 ,v2 ) | |S(v1 ,v2 ),(v2 ,v3 ) | |S(v2 ,v1 ),(v1 ,s1 ) | + . . . .
2 4 2

v1 ,v2 ,v3

Now we perform the summation over s1 and f1 . We can do it as

before, using the unitarity of S (namely equation (6.4.13)) and get a
telescoping sum. Alternatively, we can take a pictorial shortcut, noting
that the sum above, schematically drawn in Fig. 6, is an inclusion-
exclusion principle resulting in
2n2
C0 + C1 + C2 + . . . = − 2 |S(f1 ,v1 ),(v1 ,v2 ) |2 |S(v2 ,v1 ),(v1 ,s1 ) |2
B v ,v s ,f
1 2 1 1
2
2n 2n2
=− 2 1=−
B B
(v1 ,v2 )∈E

Together with the prefactor 1/B, equation (6.4.5), we get

2n2
(6.4.15) K 8 (n/B) = −
or K 8 (τ ) = −2τ 2 .
B2
This coincides with the next term in the expansion of the random-
matrix form factor (6.2.19) around τ = 0.
We mention that we implicitly assumed the sum over C to be
infinite. Of course,  is bounded by n/2. However, by relatively simple
 6.5. FURTHER REMARKS AND REFERENCES 179

order counting it can be shown that C decays exponentially with , so

terminating the sum early does not present any difficulties.
Finally, we would like to mention briefly the difficulties involved
in recovering further terms in the expansion of the form factor. The
first step is to identify the families of transformations relating the pairs
of paths; those involve more (and higher-order) self-intersections and
are known as diagrams. A contribution of each diagram is calculated
in terms of the number and type of self-intersections in the diagram.
Then a combinatorial summation is performed over all possible di-
agrams. Each of these steps is of substantial difficulty and involve
assumptions that are not justified rigorously. In particular, it is not
checked whether each pair of contributing paths is counted once and
only once. The basic method, however, is well-accepted in the physics
community, since in all contexts it produces results agreeing with ran-
dom matrix predictions.

6.5. Further remarks and references

Quantum graphs as models for quantum chaos were first sug-
gested in the pioneering articles by Kottos and Smilansky [466, 467].
In particular, they established the trace formula (Section 3.7) and
used it to represent the form factor as a sum over periodic orbits
(cf. equation (6.4.1)). The formula most similar to equation (6.4.1)
was derived in hand-written notes by Smilansky (see also [353]). Some
convergence issues were investigated in [92].
Further results appeared for specific families of quantum graphs.
In particular, in [97, 100], the complete expansion of the form fac-
tor was obtained for star graphs (it exhibited non-universal behavior,
sometimes referred to as “intermediate statistics”). Binary graphs were
considered in [683], graphs related to 1D maps in [586] and line-graphs
in [585].
An approach to the nearest neighbor spacing distribution described
in the second part of Section 6.3 was put forward by Barra and Gaspard
[56]. It was extended to eigenfunction statistics in [109].
Universality of the form factor
While the diagonal approximation (for general quantum-chaotic
systems) was pioneered by Berry [111] as far back as 1985, it was
not until 2001 that further progress was made in the evaluation of the
spectral form factor using periodic orbit expansions, when Sieber and
Richter [659, 660] identified and evaluated the pairs of orbits produc-
ing the next-order contribution. The corresponding results on graphs
followed soon after [107]. The expansion was then pushed to the third
180 6. QUANTUM CHAOS ON GRAPHS

order, first on graphs [93, 108] and then on general systems [410].
Finally, a complete expansion was accomplished in [560, 561]. An in-
terpretation of their method in the context of quantum graphs was
described in [94]. The presentation in Section 6.4.3 is based on the
latter article.
The third universality class of Wigner — symplectic — has been
covered by Bolte and Harrison [128, 129], who considered the spectral
statistics of the Dirac operator on graphs. The intermediate statistics
of the Dirac operator on rose graphs were studied in [405].
An alternative approach to studying spectral statistics is by use of
supersymmetric integrals. This has been accomplished on graphs
by Gnutzmann and Altland [349, 350]. We do not discuss this very
important subject here, as it was thoroughly covered in the review
[353].
Other topics
An important branch of quantum chaos is the study of eigenfunc-
tion statistics. A classical result is “quantum ergodicity”, whereby
a density-one subsequence of eigenfunctions |fnk |2 converges weakly to
the uniform distribution on the whole accessible space. On graphs, such
questions were studied in [102, 103, 439] on star graphs, in [101] for
quantum graphs constructed from interval maps and in [352] on gen-
eral quantum graphs using methods of field theory (supersymmetric
integrals).
Quantum graphs can also be used as models for chaotic open
systems; see [57, 58, 468, 469, 687].
 CHAPTER 7

Some Applications and Generalizations

In this chapter we provide a brief survey of various applications and

generalizations of quantum graphs. It is impossible to review in any
depth this bristling with activity area. This would take another book,
which would become obsolete by its publication date. Our task is much
more modest: to indicate (some of the) applications and to direct the
reader to some relevant literature. Due to the review nature of this
chapter, most literature references are provided in its text, rather than
in the last section, as we did this in the previous chapters.
Section 7.1 addresses the actively developing field of inverse prob-
lems for quantum graphs. Sections 7.2, 7.3, and 7.4 provide references
to works devoted to studying various differential equations on quan-
tum graphs, fractals, and multi-structures. In Section 7.5 we discuss a
rather tricky issue of approximating wave propagation in thin branch-
ing structures by suitable quantum graph models. Finally, Section 7.6
provides a list of various physics applications and models.

7.1. Inverse problems

In this section, we survey various actively studied inverse problems
for quantum graphs. Due to the high level of activity in this exciting
area, the brief description and bibliography provided below can not
possibly be complete.
As we have seen in Chapter 1, the notion of a quantum graph
involves several types of structures and parameters:
(1) the topology of the underlying graph Γ,
(2) the edge lengths le that arise when one introduces the structure
of a metric graph on Γ,
(3) the potentials V (x) involved into the definition of the action
of the quantum graph Hamiltonian along the edges:
d2 f (x)
Hf = − + V (x)f (x),
dx2
and finally,
(4) the vertex condition that describe the domain of the operator.
181
182 7. APPLICATIONS AND GENERALIZATIONS

Remark 7.1.1. It has been noted already that the distinction be-
tween the items 1) and 4) on this list is far from being clear cut. Indeed,
as we explained in Section 1.4, any quantum graph can be reduced to
the bouquet of several loops attached to a single vertex. On the other
hand, if a vertex condition does not connect some of the edges entering
the vertex, this effectively changes the graph’s topology. A way to ex-
clude this possibility is to assume that the pair of matrices (A, B) defin-
ing the vertex conditions cannot be simultaneously block-diagonalized
to the form
   
A1 0 B1 0
A= , B= .
0 A2 0 B2

Another way to achieve the same result is to assume that the vertex
conditions are strongly coupling, in the sense that the corresponding
scattering matrices σv do not have zero entries.

One can ask questions about the feasibility of recovering all, or

some (with others given) of the above parameters from some “exter-
nal” information, such as the spectral data, the scattering operator
or the boundary Dirichlet-to-Neumann map. The analogous issue of
recovering the potential of one-dimensional Strum–Liouville operators
has been (and still is) studied intensively, with plenty of books and ar-
ticles published (see, e.g., [1, 2, 242, 295, 315, 344–346, 511–513, 527,
534, 570, 571, 617, 662] for details and further references). Similar
questions for PDEs, which happen to be much harder, are also in-
tensively studied, due to their practical importance for physics, meta-
materials (“invisibility cloaks”), geophysics, medical imaging and so
on [200, 426, 435, 444, 692, 693, 695]
In the combinatorial case, various inverse (boundary value, spectral,
and scattering) problems for graphs have also been considered [85–87,
143, 144, 192, 203, 211, 212, 357, 415, 424, 425, 650]. This interest,
besides already familiar applications, was also triggered by the arrival
of the so-called Internet (or network) tomography [85, 87, 190,
335, 507].
The quantum graph inverse problems are usually (but not always)
more tractable than the PDE ones and harder than the ODE problems.
Although there are positive results parallel to the cases of ordinary and
partial differential equations, the non-trivial graph topology can mask
some elements of its structure. One of the incarnations of this effect
we have already seen in Section 3.4 when we discussed the absence of
unique continuation property and the appearance of compactly sup-
ported eigenfunctions (scars).
 7.1. INVERSE PROBLEMS 183

In Section 7.1.1, we describe the crucial result of Gutkin and Smi-

lansky that answers affirmatively some of these questions of recovery of
a “generic” (in the sense of rational independence of the edge lengths)
quantum graph from its spectrum. There exist, however, isospectral
quantum graphs, a subject which is addressed in Section 7.1.2. Re-
construction from scattering data is considered in Section 7.1.4. Sec-
tion 7.1.5 contains a discussion of discrete versions of the electrical
impedance tomography (or inverse conductivity problem).
As in most remaining sections, we do not include proofs (and in
most cases even exact formulations) here; the reader is invited to con-
sult the provided references for the details.

7.1.1. Can one hear the shape of a quantum graph?

Examples of isospectral1 (i.e., non-isomorphic, while having the
same spectrum) combinatorial graphs have been known for quite a
while (e.g., see [115, p. 12] on Collatz and Sinogowitz 1957 example
and Brooks [145] on a nice introduction to isospectrality). Due to
the relation between the spectra of equilateral quantum graphs and
combinatorial graphs that we explored in Section 3.6, one should expect
to find isospectral quantum graph examples, and this indeed happens,
as shown by von Below [73] and Gutkin and Smilansky [387].
Gutkin and Smilansky [387] have also shown that if one stays suf-
ficiently far away from having an equilateral graph, e.g., assuming that
the lengths of edges are rationally independent, one essentially (up to
few other reasonable conditions) gets unique reconstruction. Namely,
they considered compact quantum graph satisfying the following con-
ditions:
(1) The vertex conditions are strongly coupling (i.e., the scattering
matrices σv have no zero entries).
(2) The graph has no loops and multiple edges.
(3) The lengths of bonds are linearly independent over rational
numbers Q.
Under these conditions, it was proven in [387] (see also a later pa-
per [499]) that the spectrum of the Laplacian (i.e., the negative sec-
ond derivative) on Γ uniquely determines the topology of the graph and
the vertex conditions. The proof is based upon the exact trace formula
found by Roth [631] and rediscovered by Kottos and Smilansky [466]
(see Section 3.7). The authors of [387] also show some simple exam-
ples of isospectral quantum graphs (see also [130,502]). Nowaczyk has
shown [574, 575] that the condition of rational independence can be
1Also called co-spectral graphs.
184 7. APPLICATIONS AND GENERALIZATIONS

c c c c
N D N D
a 2b a b 2a b

Figure 1. A pair of simple isospectral graphs. Edge

lengths are indicated next to the edges. N and D stand
for Neumann and Dirichlet conditions correspondingly.
The conditions at the other peripheral vertices must be
the same, while the vertices of degree 3 should have
Neumann-Kirchhoff conditions.

somewhat weakened, allowing some edges to have equal length, pro-

vided they are “sufficiently far apart”. Finally, the knowledge of the
spectrum as a function of the magnetic field (the so-called “Floquet-
Bloch” spectrum) allows for recovery of many graph features regardless
of the lengths of the edges [635,636] (analogous studies for PDEs were
performed in [253, 254]).

7.1.2. Quantum graph isospectrality. The examples of isospec-

tral quantum graphs triggered an active study of graph isospectrality.
Several goals are being pursued here. On one hand, one can attempt
to construct new isospectral families and to understand their “size.”
On the other hand, one can try to prove absence of isospectral graph
pairs in some special graph families. Finally, one can ask what kind of
additional information, besides the spectrum, could resolve the isospec-
trality (see, for example, Section 7.1.3 below).
A significant extension of the famous Sunada’s method [680] of
constructing isospectral manifolds, graphs, and quantum graphs was
developed in [50,592]. A simple pair of isospectral graphs, constructed
using this method is shown in Fig. 1.
In spite of examples of isospectrality, one can extract some informa-
tion about the graph from its spectrum, for example the total length,
Euler characteristic, the number of connected components [73, 496],
and possibly some information about the cycles being present [68].
The problem of recovery of the edge potentials (even given the
topology and vertex conditions) is much harder. It was treated in the
case of star graphs and trees, where various analogs of Ambartzumyan-
Borg-Levenson theorems for ODEs were obtained (see, e.g., [148, 149,
606, 697, 707, 711, 718, 719, 721, 722] and references therein). Some
cases of trees were handled by Belishev [64, 65, 67], using the bound-
ary control method [63, 66]. Compactness of the set of isospec-
tral potentials was shown by Carlson [166], using ζ-function approach.
 7.1. INVERSE PROBLEMS 185

Kurasov [498] considered using magnetic potentials in presence of cy-

cles.
An interesting general inverse spectral theorem, applicable in par-
ticular to combinatorial and quantum graphs, was obtained recently
by Davies [220].

7.1.3. Can one count the shape of a graph? An interest-

ing variation of the inverse problem discussed in Section 7.1.1 is the
question “Can one count the shape of a drum?”, posed by Smilan-
sky and co-authors [355]. The idea is that the eigenfunction nodal
count (see Section 5.2) can contain additional information about the
domain, possibly even resolving isospectrality in both the billiard and
quantum/combinatorial graph situations. Several encouraging positive
results have been obtained [49, 53, 154, 351, 355, 434, 581]. However,
counter-examples were found recently both in graph [582] and bil-
liard [152] cases. Note that the pair of Fig. 1 will also have identical
number of nodal domains as a consequence of Theorem 5.2.8, if the
domains are counted in the natural way. An alternative way to count
was suggested in [49], but [582] provides a counter-example that works
for all types of count.
Still, the question of the type of information that can be extracted
from the nodal count is very promising. A taster of the possible results
is provided by a recent conjecture by Smilansky, proved by Band (as yet
unpublished) on discrete graphs using the magnetic characterization
of [96]: the only graphs with the “exact” nodal sequence μn = n are
trees.

7.1.4. Inverse scattering. The closely related inverse scatter-

ing problem was for the first time considered by Gerasimenko and
Pavlov [342, 343]. Unique solvability for star graphs was shown in
[459] and for general graphs under the rational independence condi-
tion in [387]. A “spectral surgery” technique is considered in [131]
for computing scattering responses and solving the inverse scattering
problem on compact graphs (with the goal of applying it to self-similar
fractal structures). For restricted scattering data, the tree case was
handled in [33,34]. See also [51,52,185,394,397,523] and references
therein for various inverse scattering results (and counter-examples) for
quantum graphs.

7.1.5. Discrete “electrical impedance” problem. The elec-

trical impedance tomography (EIT) attempts to recover the internal
186 7. APPLICATIONS AND GENERALIZATIONS

conductivity of a body using boundary measurements. It is an impor-

tant method of industrial non-destructive testing and medical imag-
ing (see, e.g., the surveys [133, 181, 692, 693, 695] and references
therein). Regretfully, it suffers from high instability and low reso-
lution. The mathematical problems arising there are difficult, and
various attempts have been made of studying discrete/graph analogs
of EIT [192, 211, 212, 424, 425], which amount to recovering electri-
cal conductivity of a network from its boundary Dirichlet-to-Neumann
map.

7.2. Other types of equations on metric graphs

We have concentrated so far on the stationary Schrödinger-type
equations on graphs. In fact, it is what gave origin to the name “quan-
tum graph”. However, due to a wide variety of applications, analogs of
other equations of mathematical physics, in particular the heat equa-
tion and the wave equation, have been studied on metric graphs as well.
Significant attention has been also paid lately to control type problems
on graphs. Besides direct applications, control theory techniques are
also known to be instrumental in addressing inverse problems.

7.2.1. Heat equation. Various properties (estimates and asymp-

totics of the heat kernel, vacuum energy, etc.) of heat equations
and diffusion processes have been intensively studied on combinato-
rial and metric graphs, multistructures, and more general spaces. See
[30, 55, 71, 80, 132, 187, 311–313, 324, 325, 328, 333, 334, 376–380,
456, 464, 465, 605, 655] for further information and references.

7.2.2. Wave equation. Spectral theory of quantum graphs nat-

urally leads (although the historic sequence is sometime reversed) to
studying the wave equation on such structures. Quite a few such stud-
ies have been conducted, which covered unique solvability, d’Alembert
type solution formulas, estimates of solutions and so on, see, for in-
stance [65, 217, 340, 556, 614] for more detailed information and ref-
erences. In particular, an important part of studying wave equations
on graphs has been the development of boundary control theory.

7.2.3. Control theory. Boundary control, besides being an im-

portant applied technique (e.g., for controlling a network of elastic
strings [31, 36, 217, and references therein]), happens to be, the same
way as in the Rn case [63, 66], a major approach to inverse problems
on graphs [32, 33, 35, 64, 65, 67, 68, 217].
 7.2. OTHER EQUATIONS 187

7.2.4. Reaction-diffusion equations. Reaction-diffusion and

convection-diffusion equations on metric graphs have been considered
in [580, 705, 706]. Not surprisingly, the topology of the graph plays
an important role, e.g. in stability of steady states [706].

7.2.5. Dirac and Rashba operators. Dirac operators (see

[513] for the 1D model) have attracted, unsurprisingly, a lot of at-
tention in the area of quantum graphs. Regretfully, the size lim-
itations and the introductory nature of this text do not let us go
into any details of these studies. The interested reader is referred
to [129, 155, 399, 405, 620, 621] for details and further references.
Dirac operators on discrete graphs and fractal sets are discussed
in [441] and [189] correspondingly.
Rashba Hamiltonian is a two-dimensional version of the Dirac
Hamiltonian (with a 90o rotation of the spins), which plays an impor-
tant role in the recently very active field of spintronics [54]. One can
find studies of this operator on quantum graphs in [83, 84, 395, 399].

7.2.6. Pseudo-differential Hamiltonians. Pseudo-differential

operators on graphs arise naturally in the studies of photonic crys-
tals [41, 304, 475, 476, 483, 484] and leaky wires [266, and references
therein]. In both cases, one deals with the following two-spectra prob-
lem on the plane:
−Δu = λδΓ u + cu,
where Γ is a planar graph in R2 , δΓ is the delta-function on R2 sup-
ported on Γ, and c, λ are constants. It is interesting to note that while
in photonic crystal theory λ is a spectral parameter and c is a coupling
constant, in the leaky wire case they play just the opposite roles.
When c = 0, the problem reduces to the spectral problem for a (non-
local) “pseudo-differential” Dirichlet-to-Neumann operator acting on Γ,
see [41, 304, 475, 476, 483, 484] and Section 7.5.3.2. This reduction is
also possible in the general case, but one ends up with a two-parametric
pencil of operators on Γ.
It would have been interesting to be learn how to use algebras of
symbols of pseudo-differential operators on singular varieties, such as
graphs, to derive, for instance, some high energy spectral asymptotics
and approximations obtained in [483, 484]. This study has barely
started [576,625] and has not reached the level of justifying the heuris-
tic results of [483].
In the discrete case, an important technique of pseudo-difference
operators was developed by Shubin in [656, 657]. Related techniques
188 7. APPLICATIONS AND GENERALIZATIONS

can be found in books by Rabinovich, Roch, and Silberman [624] and

Chandler-Wilde and Lindner [180].
7.2.7. Non-linear Schrödinger equation (NLS). Studies of
NLS on graphs have started appearing recently, see [6, 354, 416, 671]
and references therein. In particular, [354] studied the amplification
of the non-linear effects in the vicinity of a (linear) scattering reso-
nance and [6] investigated an unusual behavior of cubic focusing NLS
equation on star graphs.

7.3. Analysis on fractals

An amazing blend of graph theory, fractals, discrete groups, har-
monic analysis, spectral theory, mathematical physics and complex dy-
namics has been emerging recently. We refer the interested readers to
books and surveys [59, 60, 272, 372, 442, 566, 567, 677].

7.4. Equations on multistructures

Differential equations also arise on “multistructures”, i.e. com-
plexes that involve cells of mixed dimensions (not just dimension one,
like in quantum graphs). In particular, “open book” structures arise
in fluid dynamics [314] and other areas [22,25,80,142,284,533,543–
545, 605]. The study of such objects, as well as justifying them as
models for thin structures (e.g., photonic crystals), is largely in its
infancy.
One can also consider as examples of such multi-structures the elec-
tromagnetic circuits recently introduced by Milton and Seppecher
[544, 545].

7.5. Graph models of thin structures

In many current applications one has to deal with wave propaga-
tion in very thin branching structures. Such examples are abundant,
for example, in chemistry [637], physics and engineering (supercon-
ducting structures [21, 226, 633], meso- and nano-technology [689],
waveguides (acoustic, electric, quantum) [161–163, 282, 548]), proba-
bility and dynamical systems [311, 313, 314]. Since both analytic and
numerical approaches in such cases are usually extremely difficult, the
temptation is to approximate the structure by a graph and the govern-
ing PDE by a differential equation on this graph. Issues of feasibility
of such graph approximations are non-trivial and some had waited for
a rather long time (or are still waiting) for their resolution. The results
also depend on the type of conditions imposed at the boundary of the
 7.5. GRAPH MODELS OF THIN STRUCTURES 189

thin graph-like domain. We briefly review some of the available re-

sults, grouping them not according to their applied origins, but rather
according to the similarities of the mathematical models.
In all cases below, one deals with a PDE problem in a “fattened
graph” domain Ωd of thickness d  1 surrounding a graph Γ (see Fig.
2). Wave propagation in Ωd is assumed to be governed by the Laplace

Ωd

Γ
Figure 2. A graph Γ (black) surrounded by a “fattened
graph” domain Ωd .
 2
operator −Δu = − ∂dxu2 (or a more general Schrödinger operator)
j
with either Dirichlet or Neumann conditions on ∂Ωd .
Here the wave is confined to the thin branching domain by the “hard
wall” boundary conditions. On the other hand, in the case of “leaky
structures”, e.g., photonic crystals and quantum wires, the waves are
not strictly confined to the thin domain, but “leak” into surrounding
volume (being evanescent there). Then the equation has to be consid-
ered in the whole space, but has coefficients with high contrast near
the graph.
As it has already been mentioned, such models arise in studies of
quantum, electromagnetic, and acoustic waveguides, as well as thin
super-conducting structures and photonic crystals (e.g., [21, 91, 163,
163, 226, 266, 271, 282, 283, 286, 304, 475, 476, 548, 632–634]).
In order to establish feasibility of quantum graph models of thin
structures, one should answer the following question: How does the
relevant differential operator/equation in the “fattened graph” domain
Ωd behave in the limit when the width d is tending to zero? In par-
ticular, do the spectra converge to the spectrum of a graph operator?
Do the solutions of the underlying “fattened graph” PDE converge to
solutions of an appropriate equation on the limit graph? In such a wide
generality, there is no universal answer to these questions. The answer
depends on the specifics of the boundary value PDE problem in the
thin domain, geometry of the domain, etc.
We thus address the major cases of interest, where some under-
standing has been achieved. The case of the Neumann conditions on
190 7. APPLICATIONS AND GENERALIZATIONS

the boundary of the thin domain Ωd , which arises in superconductiv-

ity [632–634] and in dynamical systems [311,313,314], is the simplest.

7.5.1. Neumann tubes. Suppose that a compact graph Γ is C 2 -

embedded into R2 (natural multidimensional generalizations also hold).
The “fattened graph” domain Ωd consists of narrow tubes along the
edges joined by some neighborhoods of vertices. The tubes have width
d × p(x), where p(x) > 0 is a C 1 function on the edge and x is the arc
length coordinate. Notice that the width function p is allowed to be
discontinuous at the vertices. Each vertex neighborhood is contained
in a ball of radius of order ∼ d and is star-shaped with respect to a
smaller ball of a radius of the same order of smallness2. Consider the
magnetic Schrödinger operator Hd in Ωd
Hd (A, q) = −(∇ − iA(x))2 + q(x),
where the scalar electric q(x) and vector magnetic A(x) potentials are
defined in a fixed neighborhood of Γ, q is of the Lipschitz class, and A
belongs to C 1 . We impose Neumann conditions on ∂Ωd .
We also introduce the following quantum graph operator H(A, q)
on Γ (skipping the exact definition):
   
1 d d
H(A, q)f = − − iAj (x) p(x)
τ
− iAj (x) f + qf,
τ
p(x) dxj dxj
where q(x) denotes the restriction of the potential q to Γ, p(x), as
before, is the width function, and Ajτ is the tangential component of
the field A to the edge ej of Γ. The description of the operator is
completed by imposing the following vertex conditions:
1. f is continuous through each vertex.
2. at each vertex v,
 
dfj
pj (v) − iAj fj (v) = 0.
τ
dxj
{j| v∈ej }

Here pj denotes the width function of the tube along ej , and thus the
values pj (v) at the same vertex v can be different for different edges ej
adjacent to v.
It was shown by Freidlin and Wentzel [311, 313] that solutions of
the diffusion equation generated by Hd in the tubular domain Ωd con-
verge to the solutions of the corresponding “quantum graph” diffusion
2Geometricassumptions imposed in [492, 493, 632–634, 733] on the fattened
graph domain near the junctions are slightly different, although all of them have
the same goal of achieving a uniform Poincare inequality.
 7.5. GRAPH MODELS OF THIN STRUCTURES 191

equation. A lot of effort has been devoted to obtaining spectral conver-

gence results. The next theorem summarizes some of the results of Ru-
binstein and Schatzman, Kuchment and Zeng [492,493,632–634,646],
and Carlson (unpublished):

Theorem 7.5.1. For any n ∈ N

lim λn (Hd (A, q)) = λn (H(A, q)),

d→0

where λn is the n-th eigenvalue counted in increasing order (taking into

account multiplicities).

A kind of resolvent convergence in the case of trees was also shown

in [639, 640].
Using specifically developed operator theory techniques, this result
was significantly generalized and strengthened by Exner and Post [277–
279, 619, 622] to the case of “sleeve” manifolds converging to a graph,
as well as to convergence of resonances.
Notice that the limiting graph operator “forgets” most of the ge-
ometry of the graph, e.g., curvature of the edges, angles at the vertices,
etc. A crude explanation of this effect can be obtained by looking at
the case of a thin rectangular domain Ωd , which is just a strip of width
d along a straight segment. By separation of variables, the spectrum
on Ωd is the sum of the axial spectrum, which does not depend on d,
with the transversal spectrum, which consists of zero (due to Neumann
nπ !2
conditions) and a sequence of eigenvalues , all of which tend to
d
infinity when d → 0. Thus on any finite part of the spectrum, for small
values of d, only the transversal (constant) groundstate influences the
result. This groundstate clearly does not contain any geometric infor-
mation.
7.5.1.1. Large protrusions at the vertices.
An interesting effect was discovered in [311,313,492,493]. Namely,
let us assume that the domain Ωd is m-dimensional. The result of Theo-
rem 7.5.1 and its proof remain correct if the interior and exterior radii
of junctions are of the order dα with α ∈ (1/m, 1] (in the Theorem,
α = 1 is assumed). However, the answer changes when α < 1/m,
and thus the vertex junctions are too large. One can expect that hav-
ing a large junction with narrow rods attached, one could see some
waves “stuck” in the large chamber at the junction. In other words,
in the limit one can expect some modes attached to vertices to ap-
pear. This is indeed what happens. The limiting operator (different
for α < 1/m and the borderline case α = 1/m) was found in [493,733]
192 7. APPLICATIONS AND GENERALIZATIONS

(see also [476] for more details and [277–279, 619, 622] for general-
izations). It is interesting to note that when α = 1/m, this leads to
a problem with energy-dependent vertex conditions (see also [483] for
further discussion of such conditions). It is known that such conditions
usually reflect an internal structure being present in the vertices [598].
7.5.2. Dirichlet tubes. The situation becomes drastically differ-
ent (and much more complex) in the Dirichlet case. The Dirichlet
condition at the boundary of the “fattened graph” domain arises in
many applications. Historically, the first application of this kind ap-
peared in chemistry in the free-electron theory of conjugated molecules,
such as the molecule of naphthalene (see Fig. 3) (see [370, 371, 594,
594, 611, 637] for chemistry background and [476] for a mathematical
overview). In this molecule, each atom contributes three electrons for

Figure 3. Naphthalene molecule.

chemical binding. Two σ-electrons are responsible for forming bonds

that maintain the “skeleton” of the molecule, i.e. the graph Γ obtained
by eliminating the doubling of bonds (Fig. 4). The skeleton creates

Figure 4. The “skeleton” Γ of the naphthalene molecule.

a potential in which the remaining, π-electrons (one per each atom),

move through the entire structure, being confined to the skeleton graph
by the potential. It was suggested in [637] that a simplified approxi-
mate model for studying the motion of the π-electrons can be obtained
by using a second order differential Hamiltonian on the graph Γ, anal-
ogously to the case of Neumann tubes. They tested this conjecture on
the case of a circle (which can be viewed as a graph without vertices
and with constant curvature of the edges). However, a close inspection
of the argument shows that it implicitly assumes absence of the states
concentrating around the vertices. As we will see soon, such states
 7.5. GRAPH MODELS OF THIN STRUCTURES 193

in fact do occur, and thus the asymptotic (with respect to the tube’s
width going to zero) behavior of the spectrum is much more complex.
Exactly the same “Dirichlet tube”problem arises in studies of quan-
tum wires (or quantum waveguides). These are circuits of quasi-
one-dimensional semi-conductor or metallic thin wires, where “thin”
might mean a few nanometers across. One can again think of a graph
Γ within a ”fattened graph” domain Ωd of thickness d  1 around Γ
(see Fig. 2). Such models also arise in studies of electromagnetic and
acoustic waveguides (e.g., [163,163,226,243,244,255,282,286,548]).
Consider the simplest case of the domain Ωd which is the paral-
lel strip of width d along a smooth curve Γ (see the picture below).
The curve can be understood as a connected graph without vertices of
degree higher than 1.

d
Ωd

Figure 5. A strip of width d.

In this case, due to the transversal modes, the spectrum of −ΔD,d

blows up (i.e., its bottom goes to infinity) when d → 0. It is natural
then to shift the spectrum down by the first transversal eigenvalue
2
λ1 = πd . Changing the coordinates appropriately in order to flatten
the strip, one can derive the following result (the reader should see the
precise formulation and a survey in [244]):
Theorem 7.5.2 ([244]). The spectrum of −ΔD,d − λ1 converges to
the spectrum of the operator
d2 γ(x)2
(7.5.1) − − ,
dx2 4
where x is the arc length coordinate on Γ and γ(x) is the curvature of
Γ.
This result is one of the reasons for studying Schrödinger operators
with potentials involving curvature (e.g., [270]).
One sees that bending a quantum wire has an effect of an attracting
potential. This should, and indeed does lead to existence of trapped
states (e.g., [162, 163, 243–245, 255, 280, 366, 472]). Furthermore, if
instead of a quantum wire one bends a “quantum sheet” (a quasi-two-
dimensional structure), forming an edge, this can create an electron
194 7. APPLICATIONS AND GENERALIZATIONS

current along this edge. This suggests a wealth of promising applica-

tions.3
However, when trying to carry this consideration (i.e., presence of
branch points and corners) over to the case of a graph Γ, one faces big
difficulties. Indeed, if we approximate the simplest graph — a broken
segment (an angle) — with smooth curves and then try to see where
the square curvature potentials leads us in the limit, it tries to be
equal to the “square of the delta function” at the vertex. Moreover,
one discovers two parts of the spectrum, which behave differently when
d → 0. Namely, unlike the Neumann case, the transversal groundstate
is not a constant anymore and the lowest (we will call it threshold)
transversal eigenvalue, (π/d)2 , tends to infinity when d → 0. Again,
unlike the Neumann case, bound states arise near vertices (e.g., [38,
285, 649]). They correspond to the spectrum below the threshold,
while there is certainly a massive part of the spectrum above it. The
previously discussed source [637] suggests that there is still convergence
to a graph operator. This conclusion, however, is incorrect, and no
choice of boundary conditions at vertices can result in the convergence
results analogous to Theorems 7.5.1 or 7.5.2. The heuristic arguments
of [637] implicitly assumed the absence of bound states concentrated
near vertices, which is exactly what does occur.
It had been expected that the limit operator would live on an ex-
tension of the space L2 (Γ) rather than on Γ itself, the additional com-
ponents being responsible for the vertex states (as, for instance, in the
case of the large protrusions in the Neumann tubes). The convergence
would have different meaning for different components of the extension.
Progress in implementing this program and in related directions (such
as the spectral effects of curvature vs. bending of thin tubes) has been
made by many researchers (e.g., Duclos, Exner, Krejčiřı́k, Molchanov,
Šeba, Vainberg, and many others [160, 231, 232, 243–245, 278, 280,
281, 285, 285, 286, 472, 553–555, 557]). It was completed recently
by Grieser (see [368, 369]). We refer the reader to these papers and
references therein.
7.5.3. “Leaky” structures. Besides imposing the “hard wall”
Dirichlet or Neumann conditions on the boundary of a narrow strip,
one can think of “soft” confining to a curve (a graph) Γ. For example,
one can add to the governing Hamiltonian a potential cV (x) that grows
with the distance from the curve and study the large coupling constant
limit c → ∞. This was done for a smooth Γ in [230, 322].
3
It is interesting to notice that twisting, rather than bending a thin waveguide
has an effect of introducing a repulsive potential, see [472] and references therein.
 7.5. GRAPH MODELS OF THIN STRUCTURES 195

7.5.3.1. “Leaky” wires. The differential quantum graph models for

thin waveguides (quantum wires) that we discussed above are based
on the assumption of the strict confinement (due to Dirichlet or Neu-
mann conditions) of the wave to the guide. In particular, quantum me-
chanical tunneling is mostly suppressed. Thus, more realistic “leaky
wire” models have been introduced and studied (see the recent survey
in [266]), which allow tunneling. Namely, one can consider a deep po-
tential well along Γ being modeled as −cδΓ (x), where δΓ is the delta
function supported on Γ. The spectral problem for such a “leaky
wire” looks as follows:
(7.5.2) − Δu = αδΓ u + λu.
It can be considered as a two-parametric (α and λ) spectral problem.
In the case of a smooth curve (or even a higher dimensional mani-
fold) Γ, justification and study of such a model was done in [271, 273,
291, 292]. The graph case (i.e., in presence of vertices), has not been
explored well.
7.5.3.2. Photonic crystals. A photonic crystal, first suggested in
[430, 702] (see also [158, 161, 238, 429, 475, 641] for surveys of this
topic), is a periodic dielectric medium, whose properties with regard to
light propagation resemble properties of semi-conductors with respect
to electric currents. The governing equation is the spectral problem for
the stationary Maxwell system in a periodic medium:

1
μ(x)
∇ × ε(x)
1
∇F = λF,
(7.5.3)
∇ · μ(x)F = 0,
which serves here as an analog of the Schrödinger operator with pe-
riodic potential in the solid state theory. The periodic functions μ, ε
characterize the magnetic and electric properties of the medium [428].
One of the issues of particular interest is the spectrum of this prob-
lem, in particular existence of spectral gaps (which are the frequency
regions in which electromagnetic waves are not allowed to propagate
in the medium). One can read in [429] about many marvelous ex-
isting and projected applications of such materials to efficient light
sources, lasers, optical and quantum computing, mirrors, slowing down
light, super-lenses, etc. The hot topic of “invisibility cloaking”
(see [365, 384, 510, 694] and references therein) also has one of its
origins in the photonic crystals (or more general, meta-materials) re-
search.
In the studies of photonic crystals, one often deals with a thin
domain asymptotics. Here one assumes that the domain Ωd is filled
with an optically dense dielectric, while the rest is filled with air. One
196 7. APPLICATIONS AND GENERALIZATIONS

is interested in the limiting (d → 0) behavior of the frequency spectra

of electromagnetic (EM) waves in such a medium, in particular whether
this behavior is governed by an operator on the skeleton graph Γ. It
is rather clear that a differential quantum graph Hamiltonian most
probably will not work. The reason is that EM waves are not confined
to the domain Ωd , but also penetrate the air, although some special
“dielectric” modes might stay mostly inside the narrow dielectric tubes.
Thus, one expects that the limiting equation (if it exists) should still
hold in the whole space. Trying to reduce it to a graph problem,
one arrives to a non-local (“pseudo-differential”) operator on Γ, see
[41, 301, 302, 302, 304, 483, 484].
Let us concentrate on the case of 2D photonic crystals. These are,
in spite of the name, three-dimensional media that are periodic with re-
spect to two variables and homogeneous with respect to the third one.
The Figure 6 shows an example of the cross-section of such a medium.
Here the dark areas are assumed to be filled with an optically dense
dielectric, while the rest is filled with air (or another dielectric of low
optical density). The dielectric constant is assumed to be ε(x) = ε0 > 1

Figure 6. The cross-section of a 2D photonic crystal.

in the dark domains of thickness d and ε = 1 (air) in the white ones.

The material is assumed to have no magnetic properties, so the mag-
netic permeability μ equals a constant, which we will assume to be
equal to 1 (in appropriate units). We are interested here in thin high
contrast structures, which have small width d of the dielectric “walls”
and large dielectric constant ε0 . Even though neither of these two
conditions can be easily satisfied with currently available optical mate-
rials, in some instances the thin high contrast approximation hints to
interesting properties of more realistic media (see, e.g., [41,483,484]).
We also mention that the acoustic analogs of photonic crystals, which
enjoy many similar properties, allow for very high contrast materials.
 7.5. GRAPH MODELS OF THIN STRUCTURES 197

In such 2D media there are two polarizations of the electromagnetic

waves [428, 429]: the magnetic field can be orthogonal or parallel to
the plane of periodicity (i.e., the plane of Fig. 6). For monochromatic
waves of frequency ω the spectral problem for the Maxwell operator
reduces to two scalar spectral problems that correspond to these po-
larizations:
1
(7.5.4) − ∇ · ∇H = λH
ε
and

(7.5.5) − ΔE = λεE,

where λ = (ω/c)2 and c is the speed of light.

High contrast and thin structure asymptotics d → 0, ε0 d → ∞ were
considered in [41, 302, 302, 304, 311, 483, 484, 652] and more realistic
cases of the finite limit of ε0 d were treated in [311, 313, 484]. It was
discovered that for the H-mode (7.5.4) the waves tend to concentrate
overwhelmingly in the air, and correspondingly the spectrum of (7.5.4)
asymptotically concentrates in a small vicinity of the spectrum of the
Neumann Laplacian on a single “air bubble,” thus opening large gaps.
On the other hand, the E-mode (7.5.5) leads to two distinct types of
waves: air waves that behave in a manner essentially similar to the one
we have just described, and dielectric “guided” modes that prefer to
stay (due to the total internal reflection) inside the narrow dielectric
tubes and are evanescent into the air. The latter provide a much more
complicated spectrum with a very narrow bands separated by narrow
gaps of approximately the same size. This suggests that the dielectric
waves could be governed by an operator living on the graph Γ obtained
when the dielectric tubes shrink. This happens to be true.
The fine structure of the spectrum we have just mentioned, requires
rescaling (“zooming in” in order to make the small bands and gaps
observable) by multiplying the spectral parameter λ by ε0 d.
Theorem 7.5.3 ([304]). After the rescaling, the spectrum of the
dielectric modes converges to the spectrum of the problem

(7.5.6) − Δu = λδΓ u,

where δΓ is the Dirac’s delta-function of the graph Γ, i.e.

δΓ , φ = φdx.
Γ
198 7. APPLICATIONS AND GENERALIZATIONS

Although the problem (7.5.6) seems to involve the whole plane, its
spectrum in fact can be described as the spectrum of the Dirichlet-
to-Neumann operator ΛΓ on the graph Γ. Let us define this version
of the DtN operator. Given a “sufficiently nice” function φ(x) on Γ,
one uses it as the Dirichlet boundary value to find a harmonic function
u on each face of the planar graph Γ (see Fig. 7). When such a

Figure 7. Dirichlet-to-Neumann map.

function u is found, it is automatically continuous through Γ, while its

normal derivatives are not. Now one takes the jump across Γ of the
normal derivatives of u to get a function ψ(x) on Γ. The “jump” here
means the sum of outward normal derivatives of u from the two faces
adjacent to a given edge. Now the Dirichlet-to-Neumann operator ΛΓ
is the operator transforming φ into ψ:
(7.5.7) Λ Γ φ = ψ.
Remark 7.5.4.
(1) One can weaken the asymptotic conditions imposed on the way
the width d of the dielectric tends to zero and the dielectric
constant ε to infinity [304]. This leads to the problem of the
type
(7.5.8) − Δu = λδΓ u + cu,
where c is a constant that reflects the strength of coupling
between the air and dielectric modes. This problem can also
be reduced to the spectral problem of a pseudo-differential
operator on Γ, which is now the DtN map for the operator
−Δ − c on the air-bubble.
(2) As a two-parametric (c and λ) spectral problem, (7.5.8) is
identical to (7.5.2). However, the meaning of the parameters
is different: their roles are switched.
(3) Thus, one is lead to computing the spectrum of the pseudo-
differential DtN operator on Γ, which can become a formidable
task by itself. One can hope, however, that for high values of
 7.5. GRAPH MODELS OF THIN STRUCTURES 199

λ an approximation by a differential operator on Γ might be

possible. We address this issue in the next section.
7.5.3.3. Differential approximations. In the case when Γ is smooth
(i.e., when no vertices are present),√the full symbol of the DtN opera-
tor Λ Γ coincides with the one of 2 −Δ Γ , where −Δ Γ is the Laplace-
Beltrami operator on Γ (i.e., just the second
√ arc length derivative). In
particular, the spectra σ(Λ Γ ) and σ(2 −Δ Γ ) coincide asymptotically
for high energies. This also holds in higher dimensions [484]. In the
case of a sufficiently smooth curve Γ, the spectra converge very rapidly
(see
√ [484] for this and other related discussions). Since the spectrum of
2 −Δ Γ can be immediately calculated, this gives a fast method of ap-
proximating the spectrum of σ(Λ Γ ) and hence the frequency spectrum
of the dielectric modes.
The question arises: is there any analog of this approximation of the
pseudo-differential DtN operator by a root of a differential operator on
Γ in the more realistic case when the vertices are present? Analogously
to the smooth case, can one think of ΛΓ as a “pseudo-differential”
operator of first order on Γ (whatever this could mean)? Rephrasing
the question, one can ask whether there exists a differential spectral
problem A of order 2m on the graph such that the 2mth power of ΛΓ
is in some sense close to A, and hence
√
σ(ΛΓ ) ≈ σ( A).
2m

There are not many indications that this should be true, except that
it would be helpful to have such a relation. This study was attempted
in [483] with some heuristic analysis. It was discovered that in the case
of symmetric junctions at vertices one can sometimes write reason-
able differential operators as candidates for A. Although no theorem
about comparison of spectra was proven, the numerical experiments
conducted showed strong agreement of the spectra. Let us take, for
instance, the case of symmetric triple junctions at the vertices (e.g.,
the honeycomb lattice). Then the heuristic analysis of [483] shows
that the spectrum of the following problem is a good candidate for the
approximation to the spectrum of ΛΓ :
⎧ 2
⎪
⎪
d u
− 2j = λ2 u, on each edge ej ,
⎪
⎨ dxj
(7.5.9) u is continuous, at each vertex v,
⎪
⎪  du
⎪
⎩ dxj
(v) = −( 2 ) cot 3 u(v), at each vertex v.
3λ π
j∈J(v)

In the case of the honeycomb lattice, this problem can be solved explic-
itly [483], and the numerical results show amazing agreement between
200 7. APPLICATIONS AND GENERALIZATIONS

the spectra of the two problems. This is especially interesting since the
nature of the two spectra can be significantly different: the spectrum
of the problem (7.5.9) has a pure point part of infinite multiplicity
(see [483, 484]), while the spectrum of ΛΓ was conjectured [475, 484],
and then proven [119] to be absolutely continuous.
Another observation of [483] was that in the case of a symmetric
quadruple junction at each vertex (square lattice) one needs to employ
a fourth order differential problem on the graph Γ, which is also re-
sponsible for some observed differences between the spectral behavior
of the square and honeycomb Dirichlet-to-Neumann operators [484].
It is not clear at the moment how to make the analysis of [483]
rigorous and whether it has an analog for asymmetric vertex junctions.
Some initial analysis in terms of algebra of pseudo-differential operators
on singular manifolds was conducted in [576, 625].

7.6. Quantum graph modeling of various physical

phenomena
We provide here a brief list of various physics applications of the
quantum graph techniques.

7.6.1. Simulation of quantum graphs by microwave net-

works. One of the reason for quantum graph modeling was simulation
of thin branching waveguides. The authors of [420] (see also the refer-
ences in that paper) take a converse approach: trying to experimentally
simulate quantum graphs by microwave networks. They successfully
constructed quantum graphs with time reversal symmetry and studied
their spectral statistics.

7.6.2. Realizability questions. When modeling thin branching

structures by quantum graphs, one faces two realizability questions
which are important for applications. We address them in the next
subsections.
7.6.2.1. Realizability of vertex conditions. We described all possi-
ble self-adjoint vertex conditions in Section 1.4.1. Manipulating those,
one can effectively influence the transmission and scattering properties
of the network (see, e.g., [595, 597]), thus adjusting them to a task
in hand. It is not obvious, however, whether all conditions that we
have described can actually be achieved in practice. In other words,
what vertex conditions are realizable as the limits of thin branching
structures, where one is allowed to manipulate the topology, shape,
thickness, and other geometric parameters, as well as possibly creating
additional potentials? This issue has attracted significant attention in
 7.6. MODELING OF PHYSICAL PHENOMENA 201

the recent decade, and a lot of understanding has been achieved. Being
limited by the scope of this text, we address the interested reader to the
papers [182–184,186,279,288,289]. There are still unanswered ques-
tions here, e.g., which conditions can be obtained just by manipulating
geometry of the branching tube system.
7.6.2.2. Realizability of the network response. One of the origins of
the quantum graph theory (often overlooked), is what is called the
system theory (see [8, 23, 522] and literature cited there). One of
the main issues there is to describe what boundary responses (in our
terms, Dirichlet-to-Neumann maps) can be obtained from quantum
graphs. This is an extremely hard problem, even in the case of discrete
model of electrical impedance tomography, where it was resolved by
Curtis, Ingerman, and Morrow in [211,424,425] using rather sophisti-
cated techniques. In the case of a network of elastic strings, as well as
acoustic and elastodynamic networks, this problem has been completely
resolved by Milton and co-workers in [383, 543]. At this junction, the
realizability of the system’s response happens to be related to complex
operator theoretical techniques, going back to M. Krein and M. Livsic
(see [26, 78, 79]).
The issue of designing appropriate boundary responses for creating
networks performing special tasks (switches, etc.) has been intensively
studied by Pavlov et al. [124, 310, 396, 539, 540, 542, 595, 597].

7.6.3. Spectra of graphene and carbon nanotubes. Carbon

nano-structures, in particular fullerenes (buckyballs), carbon nan-
otubes, graphene, and most recently various graphynes have at-
tracted a lot of attention, due to their peculiar properties and existing
or expected applications (e.g., [174, 398, 443, 638]). Such structures
have in particular been modeled by quantum graphs (e.g., [24, 454,
455, 489, 514]).
One can get rather complete understanding of the dispersion rela-
tions for Schrödinger operators on the graphene and carbon nanotube
structures. The approach uses the idea that spectral analysis of an
equilateral quantum graph splits into two essentially independent parts:
analysis of the Dirichlet-to-Neumann operator on a single edge (see Sec-
tion 3.5), and then spectral analysis on the combinatorial graph (us-
ing the correspondence between the quantum and combinatorial graph
spectra, as described in Section 3.6). The former is independent of
the graph structure, while the latter is independent of the potential.
The analysis on a single edge leads to using the monodromy matrix
(or rather its trace, also called the Hill discriminant, or Lyapunov
function [247], a well studied object) of the 1D potential on one edge
202 7. APPLICATIONS AND GENERALIZATIONS

of the graphene lattice. Then this dispersion relation, just by simple

restriction procedure, gives spectra for all carbon nanotubes: zig-zag,
armchair, or chiral. One can extract explicitly (modulo the discrim-
inant) the dispersion relation, which in turn contains all the spectral
information.
Although historically carbon nanotubes have been known for much
longer time than graphene, mathematically it is advantageous to start
with graphene first.
7.6.3.1. Graphene. Graphene is a recently created, and already fa-
mous, allotrope of carbon (see [337–339,573]), in which carbon atoms
are located at the vertices of a regular (one atom thick) hexagonal 2D
lattice Γ (see Fig. 8). We assume that all edges of Γ have length equal

b
f
 Γ
b a
h
e2
b
W
e1
Γ b
Figure 8. The hexagonal lattice Γ and a fundamental
domain W together with its set of vertices V (W ) = {a, b}
and set of edges E(W ) = {f, g, h}.
to 1. The lattice group Z2 acts on Γ by the shifts
√ by vectors p1 e√1 +p2 e2 ,
where (p1 , p2 ) ∈ Z2 and vectors e1 = (3/2, 3/2), e2 = (0, 3) are
shown in Figure 8. We choose as a fundamental domain (Wigner-Seitz
cell) of this action the shaded parallelogram region W . Two black ver-
tices a and b belong to W , while b , b and b belong to some shifted
copies of W . Three edges f, g, h, directed as shown in the picture,
belong to W .
In addition to Z2 shifts, there is also the three-fold rotational sym-
metry, which happens to be crucial for the spectral (and thus physical)
properties of graphene.
We consider Γ as a metric graph with respect to the length structure
coming from the natural embedding of Γ into the Euclidean plane.
 7.6. MODELING OF PHYSICAL PHENOMENA 203

Each edge e in Γ can be identified with the segment [0, 1]. Under this
identification, the endpoints of an edge correspond to the points 0 and
1.
We also allow for a potential q along the edges. Let q0 (x) be an
even L2 -function on [0, 1]:
(7.6.1) q0 (x) = q0 (1 − x).
Using the fixed identification of the edges e ∈ E with [0, 1], we lift the
function q0 (x) to a (potential) q(x) on Γ. Due to evenness of q0 (x),
the definition of the potential q does not depend on the orientations
chosen along the edges.
Remark 7.6.1. The potential q, as defined above, is invariant with
respect to the full symmetry group of the honeycomb lattice Γ. More-
over, all invariant potentials from L2,loc (G) are obtainable by this pro-
cedure.
We now define the Hamiltonian H that acts along each edge e as
d2 u(x)
(7.6.2) Hu(x) = − + q(x)u(x),
dx2
where, as usual, x stands for the coordinate xe along e.
We define the domain D(H) of the operator H as consisting of the
functions u such that
(1) u belongs to the space H  2 (Γ),
(2) u satisfies the Kirchhoff-Neumann vertex conditions at each
vertex v, see Section 1.4.4.
Thus defined operator H is unbounded and self-adjoint in the Hilbert
space L2 (G) (see Section 1.4.5). Due to the condition on the potential,
operator H commutes with all symmetries of the hexagonal lattice Γ,
in particular with respect to the Z2 -shifts, which will play a crucial role
in our considerations.
As it was already mentioned in Section 7.5, such Hamiltonians have
been used for modeling electronic band-gap structure of various mate-
rials, including graphene. The complete understanding of the disper-
sion relation, band-gap spectral structure, bound states, and “diabolic”
Dirac conical points for any such operator was achieved in [489].
7.6.3.2. Carbon nanotubes. We provide here a very brief introduc-
tion to (single-wall) carbon nanotubes. One can find more detailed
discussion and classification of nanotubes, for instance, in [398, 638].
In a nutshell, a carbon nanotube is a graphene sheet rolled into a cylin-
der in such a way that overlapping patterns match perfectly (see Fig.
9). In other words, to get a nanotube, the elements of the graphene
204 7. APPLICATIONS AND GENERALIZATIONS

Figure 9. A zig-zag nanotube.

lattice that differ by an integer multiple of a fixed period vector p, are

identified. Thus, spectral analysis of Schrödinger operators on nan-
otube graphs can be deduced from the one for graphene by enforcing
an additional periodicity. Namely, let p ∈ R2 be a vector that belongs
to the lattice of translation symmetries of the honeycomb structure Γ.
In other words, Γ + p = Γ. We denote by ιp the equivalence relation
that identifies vectors z1 , z2 ∈ Γ such that z2 − z1 is an integer multiple
of the vector p. A nanotube Tp is the quotient of Γ with respect to
this equivalence relation:
(7.6.3) Tp := Γ/ιp .
Tp is naturally isometrically embedded into the cylinder R2 /ιp .
If p = p1 e1 + p2 e2 , we will denote Tp by T(p1 ,p2 ) . For example, T(0,N )
is the so-called zig-zag nanotube, while T(N,N ) is called armchair
nanotube. The names correspond to the shapes of the boundaries of
the corresponding natural fundamental domain (cf. Figure 10). The
zig-zag nanotube T(0,1) consists of a sequence of loops (“beads”) joined
by edges into a 1D-periodic necklace structure (Figure 10). Other
types of nanotubes are usually called chiral.
Remark 7.6.2. Different vectors p can produce the same (isomor-
phic) nanotubes Tp . For instance, T(m,n) = T(n,m) . Zig-zag tubes
T(0,N ) , T(N,0) , and T(N,−N ) are the same (see [398] and references therein
for a detailed classification of nanotubes).
The Hamiltonian Hp on T = Tp is defined as for the graphene
lattice Γ. One can think of Hp as of the graphene operator H acting
on functions on Γ that are periodic with the period vector p (this
definition requires some precision, since such functions do not belong
to L2 (Γ)).
The complete analysis of spectra of all such nanotube operators
was provided in [489]. Some special classes of nanotubes (sometimes
 7.6. MODELING OF PHYSICAL PHENOMENA 205

T(0,2) T(1,1)

T(0,1)

Figure 10. A zig-zag (left) and armchair (right) nan-

otube T(0,2) and T(1,1) , respectively. The vectors show the
translation vector p. The name-giving fundamental do-
main of each of the nanotubes is shaded in dark grey.
The dashed lines have to be identified. Below, the (de-
generate) zig-zag nanotube T(0,1) is shown.

in a more general situation, e.g. in presence of external electric and

magnetic fields) were considered in [422, 451–455].
7.6.4. Vacuum energy and Casimir effect. The Casimir ef-
fect (also Casimir-Polder force) (see [218,326,431,503,546,547])
is a force arising in quantum field theory due to different density of vir-
tual photons in different parts of the space. In the simplest setting it is
the attractive force between two uncharged metallic plates in vacuum.
Since its theoretical discovery in 1948, [175, 176], the Casimir effect
has been studied theoretically and verified experimentally in various
geometries.
Some geometries, especially so-called “pistons” lead to analysis on
what is effectively a quantum star graph [327]. Here we discuss results
on general graphs. The vacuum energy is formally defined as
1
(7.6.4) Ec = kn ,
2 n
where kn2 are the eigenvalues of a Hamiltonian, H, and we have taken
 = 1 = c. A short review of the quantum field-theoretic analysis lead-
ing to (7.6.4) can be found in [329]. The above expression is divergent
and to get a meaningful result, the vacuum energies for two different
configurations are subtracted from one another. To accomplish this in
a systematic way, a regularization procedure is employed, for example,
206 7. APPLICATIONS AND GENERALIZATIONS

defining the energy as the regular part of

1
(7.6.5) E(t) = kn e−kn t ,
2 n
as t → 0 (so-called ultra-violet cutoff) or letting
 
1 1
(7.6.6) Ec = FPζ − ,
2 2
where FP stands for the finite part and ζ(s) is the spectral zeta func-
tion, equation (5.3.2).
One of the methods of calculating the vacuum energy is to expand it
as a sum over periodic orbits, although usually only the shortest orbits
can be taken into account. On graphs a full expansion is available, due
to the exact trace formula (Section 3.7).
Theorem 7.6.3 ([99]). Let Γ be a graph with k-independent scat-
tering matrices. Then its vacuum energy, defined as
+∞ ,
1 L
(7.6.7) Ec = lim kn e−tkn − 2 ,
2 t→0+ n=1 πt
is given by
∞
1 1 Aγ
(7.6.8) Ec = − Re ,
2π n=1
n γ∈C
L γ
n

where Cn denotes the set of all closed paths of period n, Aγ is the

amplitude and Lγ the metric length of the path. The vacuum energy is
C ∞ -smooth as a function of bond lengths on the set {Lb > 0}.
Remark 7.6.4. The regularization in equation (7.6.7) is done by
subtracting the “Weyl term” πtL2 , which is due to the average density of
states, see Section 3.7.2. Here, as before, L stands for the total length
of the graph.
While the periodic orbit expansion is good for comparing the con-
tribution of the shortest orbits with the full sum, an approximation
widely (and often uncontrollably) used in more complicated geome-
tries, on graph one can actually get a remarkable closed-form answer,
Theorem 7.6.5 ([401]). Consider the Laplace operator on a graph
whose vertex matching conditions are defined by a global pair of matri-
ces (A, B), see Theorem 1.4.4 and Section 1.4.6. Then, the Casimir
force on the edge e is
∂Ec π 1 ∞ ∂
Fce := − = + log fˆ(t) dt,
∂Le 24L2e π 0 ∂Le
 7.6. MODELING OF PHYSICAL PHENOMENA 207

where
  
ˆ coth(tL) − csch(tL)
f (t) = det A − tB ,
− csch(tL) coth(tL)
and
coth(tL) = diag{coth(tL1 ), . . . , coth(tLE )},
csch(tL) = diag{csch(tL1 ), . . . , csch(tLE )},

provided the graph is generic: the poles of f (z) = fˆ(−iz) are the whole
of the set {mπ/Le : m ∈ Z, e ∈ E}.
Note that the function f (z) is the secular function of Section 3.5.4
(see also Section 5.3).

7.6.5. Anderson localization. The effect of the so called Ander-

son localization4 is a developed area of physics (see, e.g., [519, 735]),
which still presents numerous challenges to rigorous analysis [4, 136,
421, 519, 593, 676]. While appearance of localized states is by now
sufficiently well established (see the references above), preservation of
some of the ac spectrum (and thus of extended states) is still an ac-
tively investigated problem. Graphs, and nowadays quantum graphs,
have been useful simpler models, e.g. [413, 421, 449, 450, 587, 589].
In particular, existence of extended states was shown on discrete trees
(Bethe lattices) [446–448], as well on a variety of quantum graphs
[10, 13–15].

7.6.6. Bose-Einstein condensates. The amazing state of some

cold gases, the Bose-Einstein condensate (BEC) (see [604]) is a topic
of active research. In particular, models of BEC on graphs/networks
have been studied (see, e.g., [5, 157, 293, 296, 532]). There are various
reasons to do so. On one hand, such models seem to be more amenable
to study. On the other hand, it has been discovered that BEC theory
can be applied to studying behavior of some complex random networks,
such as Internet [113].

7.6.7. Quantum Hall effect. There have been some mathemat-

ical physics studies of the quantum Hall effect on combinatorial and
quantum graphs, see [159, 179, 361, 421].

4Localized states arising, due to random perturbations, at the edges of abso-

lutely continuous (ac) spectra of Schrödinger operators.
208 7. APPLICATIONS AND GENERALIZATIONS

7.6.8. Flat band phenomena and slowing down light. Pe-

riodic structures described by second order “elliptic” operators (e.g.,
quantum graphs) have band-gap structure of their spectra, see Chap-
ter 3.2.3. Appearance of flat branches in the dispersion relation of
such structures is equivalent to existence of bound states. In the
case of continuous media (e.g., periodic Schrödinger operators), it is
known in most cases that bound states do not exist (see, e.g., [688]
[118, 119, 317, 318, 485, 672]). However, one can create very thin
(sometimes numerically rather flat) bands of the spectrum. This can
be achieved, for instance, by periodic placement of thin graph-like di-
electric resonators [484]. The corresponding states are not localized, as
it was conjectured [475,484] and then proven [119]. However, they cor-
respond to very slowly moving (hopping from a resonator to resonator)
wave packets. Thus, such structures, CROWs (Coupled-resonator op-
tical waveguides) were proposed [708] as possible devices for slowing
down light (similar results were obtained in [484]). The problem with
such devices based upon flat bands is that a wave of the correspond-
ing frequency, shined upon a slab of such material, might be reflected
significantly.
Another proposal is based on resonant gaps, due to an array of
resonators side coupled to a waveguide, the so called SCISSORs (Side-
Coupled Integrated Spaced-Sequences of Resonators) [406, 407]. One
recognizes here the “decoration procedure” of Section 5.1.
A quantum graph approach to slowing down wave packets, while
increasing the transmission was studied intensively by Molchanov and
Vainberg [550–552].
 APPENDIX A

Some Notions of Graph Theory

Notions from the graph theory naturally play an important role in

this book. In this Appendix we collect the notions that are used in
the text, as well as some that might be useful in the future quantum
graph studies. Most of them can be found in essentially any book
on graph theory, especially in algebraic graph theory books, such as
[90, 115, 358]. Spectral graph theory has been developed extensively,
with applications both inside and outside the graph theory itself. The
books [191, 196, 215], devoted solely to the spectra of graphs and
their application, can provide the reader with a broad overview of this
fascinating area.

A.1. Graph, edge, vertex, degree

A graph Γ consists of a finite or countably infinite set of vertices
V = {vi } and a set E = {ej } of edges connecting the vertices. We
use the notation E := |E| and V := |V| for the number of edges and
vertices correspondingly. Two vertices u and v are called adjacent
(denoted u ∼ v) if there is an edge connecting them.
A graph Γ is fully specified by its |V| × |V| adjacency matrix AΓ .
In the simplest case when there are no loops or multiple edges, the
elements of the adjacency matrix are given by

1 if u ∼ v,
Au,v =
0 otherwise.

The degree  du of a vertex u is the number of edges emanating

from it, du = v∈V Au,v . All degrees are assumed to be finite (local
finiteness of the graph).
We will denote by DΓ the degree matrix, i.e. the diagonal |V|×|V|
matrix with the diagonal entries dv :

Du,v = dv δu,v ,
209
210 A. SOME NOTIONS OF GRAPH THEORY

where δu,v is the Kronecker delta


1 if u = v,
δu,v =
0 otherwise.
A graph is regular, if the degrees of all its vertices are equal.

A.2. Some special graphs

A graph is called bipartite, if its set of vertices can be split into two
complementary subsets in such a way that edges connect only vertices
from different subsets. To put it differently, the vertices of a bipartite
graph can be colored in two colors in such a way that the colors of the
ends of any edge are different.
The star graph is a bipartite graph with one vertex in one set and
n vertices in the other. There are n edges connecting the vertex from
the first set to every vertex from the second.
The complete graph Kn has n vertices and one edge connecting
every pair of vertices.

A.3. Graphs and digraphs

A graph is called directed graph or digraph, if each of its edges
is assigned a direction. In other words, each edge has one origin and
one terminal vertex. Directed edges are called bonds in this book.
The set of all bonds is denoted by B. We use the shorthand notation
B := |B| for the total number of bonds in a directed graph Γ. The
origin and terminal points of a bond are specified via the functions
o : B → V and t : B → V, i.e. a bond b begins at vertex o(b) and ends
at t(b).
We define the set of incoming bonds at a vertex v as the set of
bonds satisfying t(b) = v. If o(b) = v, the bond b is called outgoing at
vertex v. The number of incoming (resp. outgoing) bonds at a vertex
v is called incoming (resp. outgoing) degree of v and denoted div
(resp. dov ). Clearly, dov + div = dv .
For a digraph without multiple edges, one can define a directed
adjacency matrix as follows:
⎧
⎪
⎨1 if there is a bond b with o(b) = u, t(b) = v,
Au,v = −1 if there is a bond b with o(b) = v, t(b) = u,
⎪
⎩0 otherwise.
In some cases it is beneficial to consider a “directified” version of a
non-directed graph, substituting two bonds running in opposite direc-
tions for each edge of the original graph.
 A.5. PERIODIC GRAPH 211

A.4. Paths, closed paths, Betti number

A path is an ordered sequence of bonds, such that the end of the
(n − 1)-st bond in the sequence (except the last one) coincides with the
beginning of the n-th one. A closed path is such that its last edge
ends at the beginning of first one.
A closed path is a simple cycle (sometimes simply cycle) if it does
not pass through the same edge or vertex more than once. A connected
graph without simple cycles is called a tree. A disconnected union of
several trees is called a forest.
Let Γ be a connected graph. The cyclomatic number βΓ (or the
1st Betti number or simply Betti number) of Γ is the number of
edges that have to be removed from E (without reducing the number
of vertices) to turn Γ into a connected tree [90].
It is immediate from the definition that βΓ = 0 if and only if Γ is
a tree. Since the number of vertices is preserved and a connected tree
on |V| vertices has |V| − 1 edges, we see that to turn the graph Γ into
a tree we need to remove
(A.4.1) βΓ = |E| − |V| + 1
edges. This explicit expression implies that the Betti number is well-
defined, in that it does not depend on which set of edges is removed.
And alternative description of the Betti number is the rank of the
fundamental group of Γ considered as a 1D complex.
In the case of a graph with k connected components, one can gen-
eralize the concept of the Betti number to mean the number of edges
to be removed in order to turn Γ into a forest (a disjoint union of trees)
with the same number of components. It is easy to see that the Betti
number of the disconnected graph is equal to the sum of the Betti
numbers of its components. Formula (A.4.1) is thus replaced with
βΓ = |E| − |V| + k.

A.5. Periodic graph

An infinite graph Γ is said to be periodic (or Zn -periodic) if Γ
is equipped with an action of the free abelian group G = Zn , i.e. a
mapping (g, x) ∈ G × Γ → gx ∈ Γ, such that the following properties
are satisfied:
(1) Group action:
For any g ∈ Γ, the mapping x → gx is a bijection of Γ;
0x = x for any x ∈ Γ, where 0 ∈ G = Zn is the neutral
element;
(g1 g2 )x = g1 (g2 x) for any gi ∈ G, x ∈ Γ.
212 A. SOME NOTIONS OF GRAPH THEORY

(2) Continuity: For any g ∈ G, the mapping x → gx of Γ into

itself is continuous.
(3) Faithfulness: If gx = x for some x ∈ Γ, then g = 0.
(4) Discreteness: For any x ∈ Γ, there is a neighborhood U of x
such that gx ∈ U for g = 0.
(5) Co-compactness: The space of orbits Γ/G is compact. In
other words, the whole graph can be obtained by the G-shifts
of a compact subset.
(6) Structure preservation:
• gu ∼ gv if and only if u ∼ v. In particular, G acts
bijectively on the set of edges.
• In the case of a metric or quantum graph, the action pre-
serves the edges’ lengths: Lge = Le .
• In the case of a quantum graph, the action commutes
with the Hamiltonian H (and in particular, preserves the
vertex conditions).
A.6. Cayley graphs and Schreier graphs
Let G be a discrete group with finitely many generators. Let us
choose a set S of generators. The Cayley graph [178, 219, 538] of
the group G has the group G as the set of vertices. Two vertices
a ∈ G, b ∈ G are connected by an edge if and only if a = gb with either
g ∈ S or g −1 ∈ S.
Let X be a homogeneous space of the finitely generated group G
with a chosen set of generators S. Then the set X as a set of vertices
can be equipped with a set of edges, exactly like in the Cayley graph
case. The resulting graph is a Schreier graph [178, 685] of group G
(or rather of the homogeneous space X).
 APPENDIX B

Linear Operators and Operator-Functions

In this Appendix, we provide some notations and facts about some

classes of linear operators, including in particular Fredholm operators,
their indexes, and analytic families of such operators. More detailed
information concerning Fredholm theory can be found in [359, 436,
471]. A survey on properties of analytic Fredholm operator-valued
functions can be found in [730].

B.1. Some notation concerning linear operators

Let E and F be Banach spaces. We denote by L(E, F ) the space
of all bounded linear operators A : E → F equipped with the operator
norm. When E = F , we use the notation L(E) := L(E, E). The
identity operator in E is denoted as IE (we also use the notations id
or I when the domain of the mapping is clear from the context).
We denote by GL(E) ⊂ L(E) the group of invertible bounded linear
operators acting in E and by C(E, F ) ⊂ L(E, F ) the space of compact
linear operators acting from E to F (C(E) is used instead of C(E, E)).
For a linear operator A acting from E to F we denote by Ran A,
Ker A, coKer A and rank A correspondingly the range, the kernel, the
co-kernel and the rank of A:
Ran A = {y ∈ F | y = Ax, for some x ∈ E} ⊂ F,
Ker A = {x ∈ E | Ax = 0} ⊂ E,
coKer A = {f ∈ F ∗ | f (Ax) = 0 for all x ∈ E} ⊂ F ∗ ,
rank A = dim Ran A.

B.2. Fredholm and semi-Fredholm operators. Fredholm

index
Details of what is presented in this section can be found, for in-
stance, in [359, 436, 471, 730].
Definition B.2.1. An operator A ∈ L(E, F ) is said to be a Fred-
holm operator if
• dim Ker A < ∞;
• Ran A is closed;
213
214 B. LINEAR OPERATORS

• dim coKer A < ∞.

In this case
ind(A) := dim Ker A − dim coKer A
is said to be the index of A.
The set of Fredholm operators acting from E into F will be denoted
Φ(E, F ) ⊂ L(E, F ) (Φ(E) := Φ(E, E)). The set Φn (E, F ) consists of
Fredholm operators of index n. So, Φ = ∪ Φn . We also use the
n∈Z
notations
(B.2.1) Sp Φ(E) := {A ∈ Φ(E) | dim Ker A ≥ p}
and
(B.2.2) CSp Φ(E) = {A ∈ Φ(E) | dim coKer A ≥ p}
The next result collects some standard properties of the Fredholm op-
erators and their indexes:
Theorem B.2.2.
(1) Each of the sets Φn (and thus the whole Φ) is open in the
operator norm topology.
(2) The index ind is a continuous (and thus locally constant) func-
tion on Φ(E, F ).
(3) ind(AB) = ind(A) + ind(B).
(4) A ∈ Φ(E) if and only if A∗ ∈ Φ(E ∗ ); besides, ind(A∗ ) =
− ind(A).
(5) If A ∈ Φ(E, F ) and B ∈ C(E, F ), then (A + B) ∈ Φ(E, F )
and ind(A + B) = ind(A).
(6) The set Φ0 (E, F ) consists of the sums A + B, where A ∈
L(E, F ) is invertible and B ∈ C(E, F )).
(7) An operator A ∈ L(E, F ) is Fredholm if and only if there exists
an operator B ∈ L(F, E) (called regularizer of A) such that
AB − I and BA − I are compact operators.
(8) An operator A ∈ L(E, F ) is a Fredholm operator of index zero
(i.e., A ∈ Φ0 (E, F )) if and only if it has an invertible regu-
larizer.
Theorem B.2.2 in particular claims that any Fredholm operator of
zero index can be perturbed by a compact (or even finite-dimensional)
operator to an invertible one.
The following facts are also often useful:
Lemma B.2.3.
(1) The functions dim Ker A and dim coKer A are upper semi-con-
tinuous on Φ(E, F ).
 B.3. ANALYTIC OPERATOR FUNCTIONS 215

(2) Let A0 ∈ Φ(E, F ) and M be a closed subspace of finite codi-

mension in E such that M ∩ Ker A = {0}. Then for all oper-
ators A in a neighborhood of A0 one has M ∩ Ker A = {0}.
A wider class is formed by the so called semi-Fredholm opera-
tors.
Definition B.2.4. The class Φ+ (E, F ) (correspondingly Φ− (E, F ))
consists of all continuous linear operators from E to F with closed range
and finite-dimensional kernel (correspondingly, co-kernel).
Under the additional condition that the range (correspondingly, the
kernel) is a complemented subspace in F (in E), one obtains the class
Φl (E, F ) (respectively, Φr (E, F )).
In the Hilbert case Φ+ (E, F ) = Φl (E, F ) and Φ− (E, F ) = Φr (E, F )
The names used are correspondingly left semi-Fredholm and
right semi-Fredholm operators. The names chosen are explained
by the following result:
Lemma B.2.5. An operator A ∈ L(E, F ) belongs to Φl (E, F ) if and
only if there exists an operator B ∈ L(F, E) (called left regularizer
of A) such that the operator BA − I is compact.
Similarly, an operator A ∈ L(E, F ) belongs to Φr (E, F ) if and only
if there exists an operator B ∈ L(F, E) (called right regularizer of
A) such that the operator AB − I is compact.

B.3. Analytic Fredholm operator functions

In this section, we present briefly some basic facts about holomor-
phic Fredholm operator functions. Most of the proofs and relevant
references can be found in [730], and partly in [474, Ch. 1].

B.3.1. Some notions from the several complex variables

theory. In some instances of this book we needed to look at opera-
tor functions analytically depending on a multi-dimensional (or even
infinitely-dimensional) complex parameter. We thus introduce first a
class of complex domains/manifolds/analytic sets that is the easiest to
handle and which we encounter in this text. These are the so called
Stein manifolds. In particular, the so called Oka’s principle holds
for these manifolds, which, to put it crudely, says that everything that
can be done topologically can be done analytically (e.g., topologically
trivial vector bundles are automatically analytically trivial). One can
find this and other basic notions and results of the several complex vari-
ables theory in many books, for instance in [321, 364, 386, 417, 526].
216 B. LINEAR OPERATORS

Definition B.3.1. A complex analytic countable at infinity man-

ifold X is called a Stein manifold if:
(1) it is holomorphically convex (i.e., the holomorphic convex hull
of any compact subset of X is compact),
(2) it is holomorphically separable (i.e., for any two distinct points
x, y ∈ X there is a holomorphic function f on X with the
property that f (x) = f (y),
(3) for any point x0 ∈ X there is a set of holomorphic functions
on X forming local coordinates in a neighborhood of x0 .
There are several frequently encountered cases, when X is Stein,
e.g.:
• Any one-dimensional domain X ∈ C is Stein. Thus, when
dealing with functions of one variable, one should not worry
about satisfying the Stein properties.
• Any direct product of Stein domains U1 × U2 × · · · × Uk ⊂ Cn
is Stein.
• Any analytic submanifold of Cn (or even a finite-dimensional
analytic sub-manifold of an infinite-dimensional complex Ba-
nach space) is Stein.
Definition B.3.2. A subset X of an analytic manifold M is said
to be analytic set, if it can be locally described as the set of com-
mon zeros of several analytic functions. If (locally) one such function
suffices, the set is principal analytic.
B.3.2. Analytic Fredholm operator functions. Let A(z), z ∈
X be an analytic function on analytic manifold X with values in the
set Φ(E) of Fredholm operators in a complex Banach space E. We
are interested in the sets of points z ∈ X, where the kernel of A(z)
is non-trivial (the spectrum of the operator function A(·)) and
where the co-kernel is non-trivial (co-spectrum). In order to do this,
instead of working with particular operator functions, let us turn to
the sets Sp Φ(E) and CSp Φ(E) (see (B.2.1) and (B.2.2)).
Theorem B.3.3.
(1) The set Sp Φ(E) is an analytic sub-set of Φ(E).
(2) The same is true for CSp Φ(E).
(3) In particular, the set Φ(0) (E) \ GL(E) is a principal analytic
subset in Φ(0) (E).
Let now A(z), z ∈ X be a Fredholm-valued analytic function on
analytic manifold X.
We define a scale of its spectra (co-spectra) as follows:
 B.3. ANALYTIC OPERATOR FUNCTIONS 217

Definition B.3.4.
• Sp (A(·)) := {z ∈ X | A(z) ∈ Sp Φ(E)} = A−1 (Sp Φ(E)) = {z ∈
X | dim Ker A(z) ≥ p}.
• CSp (A(·)) := {z ∈ X | A(z) ∈ CSp Φ(E)} = A−1 (CSp Φ(E)) =
{z ∈ X | dim coKer A(z) ≥ p}.
Corollary B.3.5.
• Sp (A(·)) and CSp (A(·)) are analytic sub-sets of X.
• If the values A(z) are of zero index, then S1 (A(·)) = CS1 (A(·))
is a principal analytic sub-set of X.
Here is another often useful fact about analytic Fredholm operator
functions:
Theorem B.3.6. Let Ω be a Stein manifold and A(z), z ∈ Ω be an
analytic operator function with values in the set of Fredholm operators
in a Banach space E. If the minimum dimension of the kernel Ker A(z)
is not zero, then there exists an analytic vector-function F : Ω → E,
which is not identically equal to zero and such that A(z)F (z) = 0 for
all z ∈ Ω.
 APPENDIX C

Structure of Spectra

We briefly list here some basic notions from the spectral theory
that are used in this text. The reader can find detailed discussion of
spectral theory in many a good book, such as [117,216,222,223,246,
414, 436, 565, 626, 703].
For our purposes, it is sufficient to consider the operators acting in
a Hilbert space H.
Let A be a linear operator in H with domain D(A).
Definition C.0.7.
• The resolvent set ρ(A) ⊂ C consists of all complex λ ∈ C
such that the operator A−λIH has a bounded inverse operator
(A − λI)−1 , called the resolvent of A: R(λ, A) := (A − λI)−1 .
• The spectrum σ(A) ⊂ C consists of all complex λ ∈ C such
that the operator A − λIH does not have a bounded inverse.
It is clear that
ρ(A) ∪ σ(A) = C, ρ(A) ∩ σ(A) = ∅.
Let Q(x, y) be a sesqui-linear Hermitian closed form defined
on a dense sub-space D ⊂ H and bounded from below. In other
words,
(1) (sesqui-linear) Q(f, g) is linear w.r.t. f and conjugate linear
w.r.t. g,
(2) (Hermitian) Q(f, g) = Q(g, f ),
(3) (bounded from below) Q(x, x) ≥ −C x 2 for some C > 0,
(4) (closed) the domain D of Q is complete in the norm

Q(x, x) + (C + 1) x 2H .
Let A be a non-negative self-adjoint operator in H. Then it defines
on its domain D(A) a quadratic form Q(x, x) := (Ax, x). Extending it
to a closed form produces the quadratic form of the operator. Vice-
versa, if a form Q satisfies the conditions formulated above, it defines
a unique self-adjoint non-negative operator A, whose quadratic from it
is.
219
220 C. STRUCTURE OF SPECTRA

C.1. Classification of the points of the spectrum

There are several possible classifications of the types of spectrum;
we review here the classes that are most important to us.

Definition C.1.1.
• A point λ ∈ C is said to be an eigenvalue of A, if the kernel
of the operator A − λI is non-trivial. The elements of this
kernel are said to be the eigenfunctions corresponding to
the eigenvalue λ.
• The set of all eigenvalues is called the point spectrum of the
operator A and is denoted σp (A).
• The set of all isolated eigenvalues of finite (algebraic) multi-
plicity is the discrete spectrum σd (A).
• The complement to the discrete spectrum in the whole spec-
trum is called the essential spectrum σess (A) := σ(A) \
σd (A).
• If A − λI is injective, but its range Ran(A − λI) is not dense
in H, then λ is in the residual spectrum σr (A).

If the operator is self-adjoint, there are several simplifications. In

particular, the spectrum is real, the residual spectrum is empty and
the essential spectrum can be characterized as the set of all complex
numbers λ such that A − λI is not a Fredholm operator.

C.2. Spectral theorem and spectrum classification

Although one can find this information in many books on spectral
theory (including those referenced above), a good reference for this part
is [626, vol. 1, Ch. VII and VIII].
We will assume now that A is a (bounded or unbounded) self-adjoint
operator in H. Here is one version of the spectral theorem for self-
adjoint operators.

Theorem C.2.1 (see, e.g. [222, 626]). Let A be a self-adjoint op-

erator in H. There exists a unique map φ from the space C0 (σ(A))
of continuous functions on the spectrum to the space L(H) with the
following properties:
(1) φ is an algebraic ∗-homomorphism of ∗-algebras.
(2) f (A) ≤ f L∞ , where we denote f (A) := φ(f ) ∈ L(H).
(3) / R and rλ (z) := (λ − z)−1 , then rλ (A) = (λI − A)−1 .
Let λ ∈
(4) If suppf ∩ σ(A) = ∅, then f (A) = 0.
 C.2. SPECTRAL THEOREM AND SPECTRUM CLASSIFICATION 221

Let now w ∈ H. Then one can define a measure μw on the real line
as follows:
f (z)dμw (z) := (w, f (A)w).
R
The standard decomposition of measures allows one to split this mea-
sure into three parts: pure point, absolutely continuous, and singular
continuous. One can define now the following invariant subspaces for
the operator A:
Hpp := {w|μw is purely point},
Hac := {w|μw is absolutely continuous w.r.t. Lebesgue measure},
Hsc := {w|μw is singular continuous w.r.t. Lebesgue measure}.
We can now introduce the classification of the spectrum of A into pure
point σpp (A), absolutely continuous σac (A), and singular contin-
uous σsc (A) as the spectra of the restrictions of the operator A onto
the invariant subspace Hpp , Hac and Hsc correspondingly.
The reader is referred to [16, 216, 414, 626] concerning deeper de-
tails of spectral theory, including spectral measure and density of
states.
 APPENDIX D

Symplectic Geometry and Extension Theory

The real version of symplectic geometry is well known (e.g.,

see [27, 28, 110, 309, 529]) and extensively used in mechanics, geome-
try, PDEs, and other areas. It has been realized for quite a while and
formulated implicitly [16, 565] or explicitly [572, 599] that the stan-
dard J. von Neumann’s theory of self-adjoint extensions of symmetric
operators [700] (see also [626, vol 2, Ch. X] or [16,565]) could be inter-
preted in terms of a complex (Hermitian) version of symplectic geome-
try. This has lead to the corresponding applications in [261–263,393].
We briefly describe here, following [393], the facts from the (finite-
dimensional) Hermitian symplectic geometry that are used in the book
to describe vertex conditions of quantum graph operators.
Let M be a complex m-dimensional vector space and ·, · be a
sesqui-linear 2-form on M . Namely, the number f, g for f, g ∈ M
depends linearly on the first and anti-linearly on the second argument.
Definition D.1. The form f, g is Hermitian symplectic, if
(D.1) f, g = −g, f ,
where the horizontal bar denotes complex conjugation.
The form is non-degenerate, if for any f ∈ M, f = 0, there exists
g ∈ M such that f, g = 0.
A space M equipped with a non-degenerate Hermitian symplectic
form f, g will be called a Hermitian symplectic space.
If {ej }m
j=1 is a basis of M , then one can define the matrix ω = {ωij },
where
(D.2) ωi,j := ei , ej .
Then, denoting by (·, ·) the standard inner product in M associated
with the basis, we have
f, g = (f, ωg).
In the standard (real) symplectic case, the vector space and hence
the form are real, and thus (D.1) turns into f, g = −g, f .
223
224 D. SYMPLECTIC GEOMETRY

The reader familiar with the real symplectic geometry should be

warned that in spite of clear analogies, there are also significant dif-
ferences between the complex and real symplectic cases. Indeed, it is
well known that in the real case the space must have even dimension
m = 2n and the symplectic structure matrix ω can be reduced to the
canonical form
 
0 In
(D.3) ω= ,
−In 0
where In is the n-dimensional identity matrix.
Neither even dimensionality, nor unique canonical form survive in
the complex case. Indeed, the following result can be found in [393,
Lemma 1]:
Proposition D.2. A Hermitian symplectic space M of dimension
m, up to a change of basis, is uniquely characterized by two non-
negative integers n− , n+ such that n− + n+ = m. The matrix ω can be
reduced to the form
 
iIn+ 0
(D.4) ω= .
0 −iIn−
The numbers n− , n+ correspond to the defect indices in von Neu-
mann’s extension theory of symmetric operators.
In order to formulate the results needed in this book, we require
some notions similar to the ones in the standard symplectic geometry.
Definition D.3. Let M be a Hermitian symplectic space. Vectors
f, g ∈ M are skew orthogonal if f, g = 0.
For a linear subspace N ⊂ M , we denote by N ⊥ the space of all
vectors skew-orthogonal to N .
A linear subspace N ⊂ M is isotropic if N ⊂ N ⊥ and Lagrangian
if N = N ⊥ .
Theorem D.4 ([393, Lemmas 5 – 7]).
(1) A Hermitian symplectic space M contains a Lagrangian sub-
space if and only if its dimension m is even and the indices
n− and n+ are equal (and thus n− = n+ = n := m2 ).
(2) Every Lagrangian subspace has dimension n = m2 .
(3) The Grassmannian Λn of all Lagrangian subspaces in M can
be naturally identified with the unitary group U (n).
We will not prove the full Theorem D.4 (the needed simple argu-
ments can be found in [393]), but will describe the identification of
elements of the Grassmannian with the unitary n × n matrices.
 D. SYMPLECTIC GEOMETRY 225

Let m = 2n and n− = n+ = n (otherwise, according to the first

claim of the theorem, there are no Lagrangian 
subspaces. According to
Proposition D.2, we can assume that M = R n
Rn and thus elements
f of M can be written as pairs f1 ⊕ f2 ∈ Rn Rn . We can also assume
that the symplectic form is given as
(D.5) f1 ⊕ f2 , g1 ⊕ g2  = −i ((f1 , g1 ) − (f2 , g2 )) ,
where (·, ·) denotes the standardinner product in Cn . 
Let us also denote Π+ := R n
{0} and Π− := {0} Rn .
Suppose now that L ⊂ M is an (automatically n-dimensional) La-
grangian subspace in M . If x ⊕ y ∈ L, then x ⊕ y, x ⊕ y = 0 and thus,
according to (D.5),

2
(D.6) x = (x, x) = (y, y) = y 2 .
) )
This equality implies that L Π+ = L Π− = {0}. Hence, L is
the graph of an invertible linear operator U : Π+ → Π− , such that the
elements of L have the form x ⊕ U x. Using (D.6) one concludes that
U is isometric. The correspondence between Lagrangian subspaces L
and unitary operators U is clearly a bijection.
 Bibliography

[1] M. J. Ablowitz and P. A. Clarkson. Solitons, nonlinear evolution equations

and inverse scattering, volume 149 of London Mathematical Society Lecture
Note Series. Cambridge University Press, Cambridge, 1991.
[2] M. J. Ablowitz and H. Segur. Solitons and the inverse scattering transform,
volume 4 of SIAM Studies in Applied Mathematics. Society for Industrial and
Applied Mathematics (SIAM), Philadelphia, Pa., 1981.
[3] R. Abraham. Lectures of Smale on differential topology. Lecture Notes,
Columbia Univ., New York, N.Y., 1962.
[4] E. Abrahams, editor. 50 years of Anderson localization. World Scientific Pub-
lishing Co. Pte. Ltd., Hackensack, NJ, 2010.
[5] L. Accardi, A. Ben Ghorbal, and N. Obata. Monotone independence, comb
graphs and Bose-Einstein condensation. Infin. Dimens. Anal. Quantum
Probab. Relat. Top., 7(3):419–435, 2004.
[6] R. Adami, C. Cacciapuoti, D. Finco, and D. Noja. On the structure of
critical energy levels for the cubic focusing nls on star graphs. J. Phys. A,
45(19):192001, 2012.
[7] R. A. Adams. Sobolev spaces, volume 65 of Pure and Applied Mathematics.
Academic Press [A subsidiary of Harcourt Brace Jovanovich, Publishers], New
York-London, 1975.
[8] V. Adamyan. Scattering matrices for microschemes. In Operator theory and
complex analysis (Sapporo, 1991), volume 59 of Oper. Theory Adv. Appl.,
pages 1–10. Birkhäuser, Basel, 1992.
[9] V. Adamyan and B. Pavlov. Local scattering problem and a solvable model
of quantum network. In Recent advances in operator theory in Hilbert and
Krein spaces, volume 198 of Oper. Theory Adv. Appl., pages 1–10. Birkhäuser
Verlag, Basel, 2010.
[10] M. Aizenman, R. Sims, and S. Warzel. Absolutely continuous spectra of quan-
tum tree graphs with weak disorder. Comm. Math. Phys., 264(2):371–389,
2006.
[11] M. Aizenman, R. Sims, and S. Warzel. Fluctuation-based proof of the stability
of ac spectra of random operators on tree graphs. In Quantum graphs and their
applications, volume 415 of Contemp. Math., pages 1–14. Amer. Math. Soc.,
Providence, RI, 2006.
[12] M. Aizenman, R. Sims, and S. Warzel. Fluctuation based proof of the sta-
bility of ac spectra of random operators on tree graphs. In N. Chernov,
Y. Karpeshina, I. W. Knowles, R. T. Lewis, and R. Weikard, editors, Re-
cent advances in differential equations and mathematical physics, volume 412
of Contemporary Mathematics, pages 1–14. Amer. Math. Soc., Providence,
RI, 2006.

227
228 BIBLIOGRAPHY

[13] M. Aizenman, R. Sims, and S. Warzel. Fluctuation-based proof of the stability

of ac spectra of random operators on tree graphs. In G. Berkolaiko, R. Carlson,
S. Fulling, and P. Kuchment, editors, Quantum graphs and their applications,
volume 415 of Contemp. Math., pages 1–14. Amer. Math. Soc., Providence,
RI, 2006.
[14] M. Aizenman, R. Sims, and S. Warzel. Stability of the absolutely continuous
spectrum of random Schrödinger operators on tree graphs. Probab. Theory
Related Fields, 136(3):363–394, 2006.
[15] M. Aizenman and S. Warzel. Persistence under weak disorder of AC spec-
tra of quasi-periodic Schrödinger operators on trees graphs. Mosc. Math. J.,
5(3):499–506, 742, 2005.
[16] N. I. Akhiezer and I. M. Glazman. Theory of linear operators in Hilbert space.
Dover Publications Inc., New York, 1993. Translated from the Russian and
with a preface by Merlynd Nestell. Reprint of the 1961 and 1963 translations,
two volumes bound as one.
[17] E. Akkermans, A. Comtet, J. Desbois, G. Montambaux, and C. Texier. Spec-
tral determinant on quantum graphs. Ann. Phys., 284(1):10–51, 2000.
[18] O. Al-Obeid. On the number of the constant sign zones of the eigenfunctions
of a dirichlet problem on a network (graph). Technical report, Voronezh:
Voronezh State University, 1992. in Russian, deposited in VINITI 13.04.93,
N 938 – B 93. – 8 p.
[19] J. H. Albert. Generic properties of eigenfunctions of elliptic partial differential
operators. Trans. Amer. Math. Soc., 238:341–354, 1978.
[20] S. Albeverio, F. Gesztesy, R. Høegh-Krohn, and H. Holden. Solvable models
in quantum mechanics. Texts and Monographs in Physics. Springer-Verlag,
New York, 1988.
[21] S. Alexander. Superconductivity of networks. A percolation approach to the
effects of disorder. Phys. Rev. B (3), 27(3):1541–1557, 1983.
[22] F. Ali Mehmeti, J. von Below, and S. Nicaise, editors. Partial differential
equations on multistructures, volume 219 of Lecture Notes in Pure and Applied
Mathematics, New York, 2001. Marcel Dekker Inc.
[23] D. Alpay and V. Vinnikov, editors. Characteristic functions, scattering func-
tions and transfer functions, volume 197 of Operator Theory: Advances and
Applications. Birkhäuser Verlag, Basel, 2010. The Moshe Livsic memorial
volume.
[24] C. Amovilli, F. E. Leys, and N. H. March. Electronic energy spectrum of two-
dimensional solids and a chain of C atoms from a quantum network model.
J. Math. Chem., 36(2):93–112, 2004.
[25] V. Ancona, B. Gaveau, and M. Okada. Harmonic forms and cohomology of
singular stratified spaces. Bull. Sci. Math., 131(5):422–456, 2007.
[26] Y. Arlinskii, S. Belyi, and E. Tsekanovskii. Conservative realizations of
Herglotz-Nevanlinna functions, volume 217 of Operator Theory: Advances
and Applications. Birkhäuser/Springer Basel AG, Basel, 2011.
[27] V. I. Arnold. Mathematical methods of classical mechanics. Springer-Verlag,
New York, 1978. Translated from the Russian by K. Vogtmann and A. We-
instein, Graduate Texts in Mathematics, 60.
 BIBLIOGRAPHY 229

[28] V. I. Arnol d and A. B. Givental . Symplectic geometry. In V. I. Arnol d and

S. P. Novikov, editors, Dynamical systems, volume IV of Encyclopedia Math.
Sci., pages 1–138. Springer, Berlin, 2001.
[29] N. W. Ashcroft and N. D. Mermin. Solid State Physics. Holt, Rinehart and
Winston, New York-London, 1976.
[30] P. Auscher, T. Coulhon, and A. Grigor’yan, editors. Heat kernels and anal-
ysis on manifolds, graphs, and metric spaces, volume 338 of Contemporary
Mathematics. Amer. Math. Soc., Providence, RI, 2003.
[31] S. Avdonin. Control problems on quantum graphs. In Analysis on graphs
and its applications, volume 77 of Proc. Sympos. Pure Math., pages 507–521.
Amer. Math. Soc., Providence, RI, 2008.
[32] S. Avdonin, B. P. Belinskiy, and J. V. Matthews. Dynamical inverse problem
on a metric tree. Inverse Problems, 27(7):075011, 21, 2011.
[33] S. Avdonin and P. Kurasov. Inverse problems for quantum trees. Inverse
Problems and Imaging, 2(1):1–21, 2008.
[34] S. Avdonin, P. Kurasov, and M. Nowaczyk. Inverse problems for quantum
trees II: recovering matching conditions for star graphs. Inverse Probl. Imag-
ing, 4(4):579–598, 2010.
[35] S. Avdonin, G. Leugering, and V. Mikhaylov. On an inverse problem for tree-
like networks of elastic strings. ZAMM Z. Angew. Math. Mech., 90(2):136–
150, 2010.
[36] S. Avdonin and V. Mikhaylov. Controllability of partial differential equations
on graphs. Appl. Math. (Warsaw), 35(4):379–393, 2008.
[37] M. Avellaneda and F.-H. Lin. Un théorème de Liouville pour des équations
elliptiques à coefficients périodiques. C. R. Acad. Sci. Paris Sér. I Math.,
309(5):245–250, 1989.
[38] Y. Avishai, D. Bessis, B. G. Giraud, and G. Mantica. Quantum bound states
in open geometries. Phys. Rev. B, 44(15):8028–8034, 1991.
[39] J. E. Avron, P. Exner, and Y. Last. Periodic Schrödinger operators with large
gaps and Wannier-Stark ladders. Phys. Rev. Lett., 72(6):896–899, 1994.
[40] J. E. Avron, A. Raveh, and B. Zur. Adiabatic quantum transport in multiply
connected systems. Rev. Mod. Phys., 60(4):873–915, 1988.
[41] W. Axmann, P. Kuchment, and L. Kunyansky. Asymptotic methods for
thin high-contrast two-dimensional PBG materials. J. Lightwave Technology,
17(11):1996–2007, 1999.
[42] A. V. Badanin and E. L. Korotyaev. A magnetic Schrödinger operator on a
periodic graph. Mat. Sb., 201(10):3–46, 2010.
[43] M. Baker and X. Faber. Metrized graphs, Laplacian operators, and electrical
networks. In Quantum graphs and their applications, volume 415 of Contemp.
Math., pages 15–33. Amer. Math. Soc., Providence, RI, 2006.
[44] M. Baker and R. Rumely. Harmonic analysis on metrized graphs. Canad. J.
Math., 59(2):225–275, 2007.
[45] N. Bakhvalov and G. Panasenko. Homogenization: averaging processes in pe-
riodic media, volume 36 of Mathematics and its Applications (Soviet Series).
Kluwer Academic Publishers Group, Dordrecht, 1989. Mathematical prob-
lems in the mechanics of composite materials. Translated from the Russian
by D. Leı̆tes.
230 BIBLIOGRAPHY

[46] R. Band, G. Berkolaiko, H. Raz, and U. Smilansky. The number of nodal

domains on quantum graphs as a stability index of graph partitions. Commun.
Math. Phys., 311(3):815–838, 2012.
[47] R. Band, G. Berkolaiko, and U. Smilansky. Dynamics of nodal points and
the nodal count on a family of quantum graphs. Annales Henri Poincare,
13(1):145–184, 2012.
[48] R. Band, J. M. Harrison, and C. H. Joyner. Finite pseudo orbit expansions
for spectral quantities of quantum graphs. preprint arXiv:1205.4214, 2012.
[49] R. Band, I. Oren, and U. Smilansky. Nodal domains on graphs—how to count
them and why? In Analysis on graphs and its applications, volume 77 of Proc.
Sympos. Pure Math., pages 5–27. Amer. Math. Soc., Providence, RI, 2008.
[50] R. Band, O. Parzanchevski, and G. Ben-Shach. The isospectral fruits of rep-
resentation theory: quantum graphs and drums. J. Phys. A, 42(17):175202,
42, 2009.
[51] R. Band, A. Sawicki, and U. Smilansky. Scattering from isospectral quantum
graphs. J. Phys. A, 43(41):415201, 17, 2010.
[52] R. Band, A. Sawicki, and U. Smilansky. Note on the role of symmetry in
scattering from isospectral graphs and drums. preprint arXiv:1110.2475,
2011.
[53] R. Band, T. Shapira, and U. Smilansky. Nodal domains on isospectral quan-
tum graphs: the resolution of isospectrality? J. Phys. A, 39(45):13999–14014,
2006.
[54] S. Bandyopadhyay and M. Cahay. Introduction to Spintronics. CRC Press,
Boca Raton, FL, 2008.
[55] M. Barlow, T. Coulhon, and A. Grigor’yan. Manifolds and graphs with slow
heat kernel decay. Invent. Math., 144(3):609–649, 2001.
[56] F. Barra and P. Gaspard. On the level spacing distribution in quantum
graphs. J. Statist. Phys., 101(1–2):283–319, 2000. Dedicated to Grégoire Nico-
lis on the occasion of his sixtieth birthday (Brussels, 1999).
[57] F. Barra and P. Gaspard. Transport and dynamics on open quantum graphs.
Phys. Rev. E, 65(1):016205, 2002.
[58] F. Barra and T. Gilbert. Algebraic decay in hierarchical graphs. J. Stat. Phys.,
109(3–4):777–801, 2002.
[59] L. Bartholdi, R. Grigorchuk, and V. Nekrashevych. From fractal groups
to fractal sets. In Fractals in Graz 2001, Trends Math., pages 25–118.
Birkhäuser, Basel, 2003.
[60] L. Bartholdi and R. I. Grigorchuk. Spectra of non-commutative dynamical
systems and graphs related to fractal groups. C. R. Acad. Sci. Paris Sér. I
Math., 331(6):429–434, 2000.
[61] H. Bass. The Ihara-Selberg zeta function of a tree lattice. Internat. J. Math.,
3(6):717–797, 1992.
[62] J. Behrndt and A. Luger. On the number of negative eigenvalues of the Lapla-
cian on a metric graph. J. Phys. A, 43(47):474006, 11, 2010.
[63] M. I. Belishev. Boundary control in reconstruction of manifolds and metrics
(the BC method). Inverse Problems, 13(5):R1–R45, 1997.
[64] M. I. Belishev. Boundary spectral inverse problem on a class of graphs (trees)
by the BC method. Inverse Problems, 20(3):647–672, 2004.
 BIBLIOGRAPHY 231

[65] M. I. Belishev. On the boundary controllability of a dynamical system de-

scribed by the wave equation on a class of graphs (on trees). Zap. Nauchn.
Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI), 308(Mat. Vopr. Teor.
Rasprostr. Vol. 33):23–47, 252, 2004.
[66] M. I. Belishev. Recent progress in the boundary control method. Inverse
Problems, 23(5):R1–R67, 2007.
[67] M. I. Belishev and A. F. Vakulenko. Inverse problems on graphs: recovering
the tree of strings by the BC-method. J. Inverse Ill-Posed Probl., 14(1):29–46,
2006.
[68] M. I. Belishev and N. Wada. On revealing graph cycles via boundary mea-
surements. Inverse Problems, 25(10):105011, 21, 2009.
[69] B. Bellazzini, M. Burrello, M. Mintchev, and P. Sorba. Quantum field theory
on star graphs. In Analysis on graphs and its applications, volume 77 of Proc.
Sympos. Pure Math., pages 639–656. Amer. Math. Soc., Providence, RI, 2008.
[70] J. v. Below. A characteristic equation associated to an eigenvalue problem on
c2 -networks. Linear Algebra Appl., 71:309–325, 1985.
[71] J. v. Below. Classical solvability of linear parabolic equations on networks. J.
Differential Equations, 72(2):316–337, 1988.
[72] J. v. Below. The index of a periodic graph. Results Math., 25(3–4):198–223,
1994.
[73] J. v. Below. Can one hear the shape of a network? In Partial differential
equations on multistructures (Luminy, 1999), volume 219 of Lecture Notes in
Pure and Appl. Math., pages 19–36. Dekker, New York, 2001.
[74] J. v. Below. An index theory for uniformly locally finite graphs. Linear Algebra
Appl., 431(1-2):1–19, 2009.
[75] J. v. Below and J. A. Lubary. Harmonic functions on locally finite networks.
Results Math., 45(1-2):1–20, 2004.
[76] J. v. Below and J. A. Lubary. The eigenvalues of the Laplacian on locally
finite networks under generalized node transition. Results Math., 54(1-2):15–
39, 2009.
[77] J. v. Below and J. A. Lubary. Isospectral infinite graphs and networks and
infinite eigenvalue multiplicities. Netw. Heterog. Media, 4(3):453–468, 2009.
[78] S. Belyi, S. Hassi, H. de Snoo, and E. Tsekanovskiı̌. A general realization the-
orem for matrix-valued Herglotz-Nevanlinna functions. Linear Algebra Appl.,
419(2-3):331–358, 2006.
[79] S. V. Belyi and E. R. Tsekanovskii. Inverse Stieltjes-like functions and inverse
problems for systems with Schrödinger operator. In Characteristic functions,
scattering functions and transfer functions, volume 197 of Oper. Theory Adv.
Appl., pages 21–49. Birkhäuser Verlag, Basel, 2010.
[80] A. Bendikov, L. Saloff-Coste, M. Salvatori, and W. Woess. The heat semi-
group and Brownian motion on strip complexes. Adv. Math., 226(1):992–1055,
2011.
[81] A. Bensoussan, J.-L. Lions, and G. Papanicolaou. Asymptotic analysis for
periodic structures, volume 5 of Studies in Mathematics and its Applications.
North-Holland Publishing Co., Amsterdam, 1978.
[82] F. Bentosela, P. Duclos, and P. Exner. Absolute continuity in periodic thin
tubes and strongly coupled leaky wires. Lett. Math. Phys., 65(1):75–82, 2003.
232 BIBLIOGRAPHY

[83] D. Bercioux, M. Governale, V. Cataudella, and V. M. Ramaglia.

Rashba-effect-induced localization in quantum networks. Phys. Rev. Lett.,
93(5):056802, 2004.
[84] D. Bercioux, M. Governale, V. Cataudella, and V. M. Ramaglia. Rashba effect
in quantum networks. Phys. Rev. B, 72(7):075305, 2005.
[85] C. Berenstein, F. Gavilánez, and J. Baras. Network tomography. In Integral
geometry and tomography, volume 405 of Contemp. Math., pages 11–17. Amer.
Math. Soc., Providence, RI, 2006.
[86] C. A. Berenstein, E. Casadio Tarabusi, J. M. Cohen, and M. A. Picardello.
Integral geometry on trees. Amer. J. Math., 113(3):441–470, 1991.
[87] C. A. Berenstein and F. Gavilánez. Network tomography. Rev. Colombiana
Mat., 41(Numero Especial):143–150, 2007.
[88] J. M. Berezanskii. Expansions in Eigenfunctions of Self-Adjoint Operators,
volume 17. Transl. Amer. Math. Soc., Providence, RI, 1968.
[89] F. A. Berezin and M. A. Shubin. The Schrödinger equation, volume 66 of
Mathematics and its Applications (Soviet Series). Kluwer Academic Publish-
ers Group, Dordrecht, 1991. Translated from the 1983 Russian edition by Yu.
Rajabov, D. A. Leı̆tes and N. A. Sakharova and revised by Shubin. With
contributions by G. L. Litvinov and Leı̆tes.
[90] C. Berge. Graphs and hypergraphs. North-Holland Publishing Co., Amster-
dam, 1973. Translated from the French by Edward Minieka, North-Holland
Mathematical Library, Vol. 6.
[91] J. Berger, J. Rubinstein, and M. Schatzman. Multiply connected mesoscopic
superconducting structures. In Calculus of variations and differential equa-
tions (Haifa, 1998), volume 410 of Chapman & Hall/CRC Res. Notes Math.,
pages 21–40. Chapman & Hall/CRC, Boca Raton, FL, 2000.
[92] G. Berkolaiko. Quantum start graphs and related systems. PhD thesis, Uni-
versity of Bristol, 2000.
[93] G. Berkolaiko. Form factor for large quantum graphs: evaluating orbits with
time reversal. Waves Random Media, 14(1):S7–S27, 2004. Special section on
quantum graphs.
[94] G. Berkolaiko. Form factor expansion for large graphs: a diagrammatic ap-
proach. In Quantum graphs and their applications, volume 415 of Contemp.
Math., pages 35–49. Amer. Math. Soc., Providence, RI, 2006.
[95] G. Berkolaiko. A lower bound for nodal count on discrete and metric graphs.
Comm. Math. Phys., 278(3):803–819, 2008.
[96] G. Berkolaiko. Nodal count of graph eigenfunctions via magnetic perturba-
tion. preprint arXiv:1110.5373, 2011.
[97] G. Berkolaiko, E. B. Bogomolny, and J. P. Keating. Star graphs and Šeba
billiards. J. Phys. A, 34(3):335–350, 2001.
[98] G. Berkolaiko, R. Carlson, S. Fulling, and P. Kuchment, editors. Quantum
graphs and their applications, volume 415 of Contemp. Math., Providence,
RI, 2006. Amer. Math. Soc.
[99] G. Berkolaiko, J. M. Harrison, and J. H. Wilson. Mathematical aspects of
vacuum energy on quantum graphs. J. Phys. A, 42(2):025204, 20, 2009.
[100] G. Berkolaiko and J. P. Keating. Two-point spectral correlations for star
graphs. J. Phys. A, 32(45):7827–7841, 1999.
 BIBLIOGRAPHY 233

[101] G. Berkolaiko, J. P. Keating, and U. Smilansky. Quantum ergodicity for

graphs related to interval maps. Comm. Math. Phys., 273(1):137–159, 2007.
[102] G. Berkolaiko, J. P. Keating, and B. Winn. Intermediate wave function sta-
tistics. Phys. Rev. Lett., 91(13):134103, 2003.
[103] G. Berkolaiko, J. P. Keating, and B. Winn. No quantum ergodicity for star
graphs. Comm. Math. Phys., 250(2):259–285, 2004.
[104] G. Berkolaiko and P. Kuchment. Dependence of the spectrum of a quantum
graph on vertex conditions and edge lengths. In Spectral Geometry. American
Math. Soc., 2012.
[105] G. Berkolaiko, P. Kuchment, and U. Smilansky. Critical partitions and nodal
deficiency of billiard eigenfunctions. to appear in Geom. Funct. Anal., 2012.
preprint arXiv:1107.3489.
[106] G. Berkolaiko, H. Raz, and U. Smilansky. Stability of nodal structures in
graph eigenfunctions and its relation to the nodal domain count. J. Phys. A,
45(16):165203, 2012.
[107] G. Berkolaiko, H. Schanz, and R. S. Whitney. Leading off-diagonal correction
to the form factor of large graphs. Phys. Rev. Lett., 88:104101, 2002.
[108] G. Berkolaiko, H. Schanz, and R. S. Whitney. Form factor for a family of
quantum graphs: an expansion to third order. J. Phys. A, 36(31):8373–8392,
2003.
[109] G. Berkolaiko and B. Winn. Relationship between scattering matrix and spec-
trum of quantum graphs. Trans. Amer. Math. Soc., 362(12):6261–6277, 2010.
[110] R. Berndt. An introduction to symplectic geometry, volume 26 of Graduate
Studies in Mathematics. American Mathematical Society, Providence, RI,
2001. Translated from the 1998 German original by Michael Klucznik.
[111] M. V. Berry. Semiclassical theory of spectral rigidity. Proc. Roy. Soc. London
Ser. A, 400(1819):229–251, 1985.
[112] H. Bethe and A. Sommerfeld. Elektronentheorie der metalle. In H. Geiger
and K. Scheel, editors, Handbuch der Physik, volume 24,2, pages 333–622.
Springer, 1933. This nearly 300-page chapter was later published as a separate
book: Elektronentheorie der Metalle (Springer, 1967).
[113] G. Bianconi and A.-L. Barabási. Bose-Einstein condensation in complex net-
works. Phys. Rev. Lett., 86(5632–5635):229–251, 2001.
[114] W. E. Bies and E. J. Heller. Nodal structure of chaotic eigenfunctions.
J. Phys. A: Math. Gen., 35(27):5673–5685, 2002.
[115] N. Biggs. Algebraic Graph Theory. Cambridge University Press, Cambridge,
1974.
[116] T. Birch. History of the Royal Society of London. A. Millar, London, 1756.
[117] M. S. Birman and M. Z. Solomjak. Spectral theory of selfadjoint operators
in Hilbert space. Mathematics and its Applications (Soviet Series). D. Reidel
Publishing Co., Dordrecht, 1987. Translated from the 1980 Russian original
by S. Khrushchëv and V. Peller.
[118] M. S. Birman and T. A. Suslina. A periodic magnetic Hamiltonian with a
variable metric. The problem of absolute continuity. Algebra i Analiz, 11(2):1–
40, 1999.
[119] M. S. Birman, T. A. Suslina, and R. G. Shterenberg. Absolute continuity of
the two-dimensional Schrödinger operator with delta potential concentrated
on a periodic system of curves. Algebra i Analiz, 12(6):140–177, 2000.
234 BIBLIOGRAPHY

[120] T. Bıyıkoğlu. A discrete nodal domain theorem for trees. Linear Algebra Appl.,
360:197–205, 2003.
[121] T. Bıyıkoğlu, J. Leydold, and P. F. Stadler. Laplacian eigenvectors of graphs,
volume 1915 of Lecture Notes in Mathematics. Springer, Berlin, 2007. Perron-
Frobenius and Faber-Krahn type theorems.
[122] J. Blank, P. Exner, and M. Havlı́ček. Hilbert space operators in quantum
physics. Theoretical and Mathematical Physics. Springer, New York, second
edition, 2008.
[123] G. Blum, S. Gnutzmann, and U. Smilansky. Nodal domains statistics: A
criterion for quantum chaos. Phys. Rev. Lett., 88(11):114101, 2002.
[124] V. B. Bogevolnov, A. B. Mikhailova, B. S. Pavlov, and A. M. Yafyasov. About
scattering on the ring. In Recent advances in operator theory (Groningen,
1998), volume 124 of Oper. Theory Adv. Appl., pages 155–173. Birkhäuser,
Basel, 2001.
[125] O. Bohigas, M. J. Giannoni, and C. Schmit. Characterization of chaotic
quantum spectra and universality of level fluctuation laws. Phys. Rev. Lett.,
52(1):1–4, 1984.
[126] J. Bolte and S. Endres. Trace formulae for quantum graphs. In Analysis on
graphs and its applications, volume 77 of Proc. Sympos. Pure Math., pages
247–259. Amer. Math. Soc., Providence, RI, 2008.
[127] J. Bolte and S. Endres. The trace formula for quantum graphs with general
self adjoint boundary conditions. Ann. Henri Poincaré, 10(1):189–223, 2009.
[128] J. Bolte and J. Harrison. Spectral statistics for the Dirac operator on graphs.
J. Phys. A, 36(11):2747–2769, 2003.
[129] J. Bolte and J. Harrison. The spin contribution to the form factor of quantum
graphs. J. Phys. A, 36(27):L433–L440, 2003.
[130] J. Boman and P. Kurasov. Symmetries of quantum graphs and the inverse
scattering problem. Adv. in Appl. Math., 35(1):58–70, 2005.
[131] A. N. Bondarenko and V. A. Dedok. Spectral surgery of quantum graphs.
Sib. Zh. Ind. Mat., 7(4):16–28, 2004.
[132] J. D. Bondurant and S. A. Fulling. The Dirichlet-to-Robin transform. J. Phys.
A, 38(7):1505–1532, 2005.
[133] L. Borcea. Electrical impedance tomography. Inverse Problems, 18(6):R99–
R136, 2002.
[134] L. P. Bouckaert, R. Smoluchowski, and E. Wigner. Theory of Brillouin zones
and symmetry properties of wave functions in crystals. Phys. Rev., 50(1):58–
67, 1936.
[135] N. Bourbaki. Éléments de mathématique. Fasc. XXXVI. Variétés
différentielles et analytiques. Fascicule de résultats (Paragraphes 8 à 15). Ac-
tualités Scientifiques et Industrielles, No. 1347. Hermann, Paris, 1971.
[136] J. Bourgain. Green’s function estimates for lattice Schrödinger operators and
applications, volume 158 of Annals of Mathematics Studies. Princeton Uni-
versity Press, Princeton, NJ, 2005.
[137] A. Boutet de Monvel, D. Lenz, and P. Stollmann. Schnol’s theorem for
strongly local forms. Israel J. Math., 173(1):189–211, 2009.
[138] A. Boutet de Monvel and P. Stollmann. Eigenfunction expansions for gener-
ators of Dirichlet forms. J. Reine Angew. Math., 561:131–144, 2003.
 BIBLIOGRAPHY 235

[139] O. Boyko and V. Pivovarchik. Inverse spectral problem for a star graph of
Stieltjes strings. Methods Funct. Anal. Topology, 14(2):159–167, 2008.
[140] M. R. Bridson and K. Vogtmann. Automorphisms of automorphism groups
of free groups. J. Algebra, 229(2):785–792, 2000.
[141] L. Brillouin. Wave propagation in periodic structures. Electric filters and crys-
tal lattices. Dover Publications Inc., New York, N. Y., 1953. 2d ed.
[142] M. Brin and Y. Kifer. Brownian motion, harmonic functions and hyperbolicity
for Euclidean complexes. Math. Z., 237(3):421–468, 2001.
[143] R. Brooks. The fundamental group and the spectrum of the Laplacian. Com-
ment. Math. Helv., 56(4):581–598, 1981.
[144] R. Brooks. The spectral geometry of a tower of coverings. J. Differential
Geom., 23(1):97–107, 1986.
[145] R. Brooks. Constructing isospectral manifolds. Amer. Math. Monthly,
95(9):823–839, 1988.
[146] B. M. Brown, M. S. P. Eastham, and I. G. Wood. An example on the dis-
crete spectrum of a star graph. In Analysis on graphs and its applications,
volume 77 of Proc. Sympos. Pure Math., pages 331– 335. Amer. Math. Soc.,
Providence, RI, 2008.
[147] B. M. Brown, M. Langer, and K. M. Schmidt. The HELP inequality on trees.
In Analysis on graphs and its applications, volume 77 of Proc. Sympos. Pure
Math., pages 337–354. Amer. Math. Soc., Providence, RI, 2008.
[148] B. M. Brown and R. Weikard. A Borg-Levinson theorem for trees. Proc. R.
Soc. Lond. Ser. A Math. Phys. Eng. Sci., 461(2062):3231–3243, 2005.
[149] B. M. Brown and R. Weikard. On inverse problems for finite trees. In Methods
of spectral analysis in mathematical physics, volume 186 of Oper. Theory Adv.
Appl., pages 31–48. Birkhäuser Verlag, Basel, 2009.
[150] A. Brudnyi. Holomorphic functions of polynomial growth on abelian coverings
of a compact complex manifold. Comm. Anal. Geom., 6(3):485–510, 1998.
[151] A. Brudnyi. Holomorphic functions of exponential growth on abelian cover-
ings of a projective manifold. Compositio Math., 127(1):69–81, 2001.
[152] J. Brüning and D. Fajman. On the nodal count for flat tori. Communications
in Mathematical Physics, 313:791–813, 2012. 10.1007/s00220-012-1432-0.
[153] J. Brüning, V. Geyler, and K. Pankrashkin. Cantor and band spectra for pe-
riodic quantum graphs with magnetic fields. Comm. Math. Phys., 269(1):87–
105, 2007.
[154] J. Brüning, D. Klawonn, and C. Puhle. Comment on: “Resolving isospec-
tral ‘drums’ by counting nodal domains” [J. Phys. A 38 (2005), no. 41,
8921–8933] by S. Gnutzmann, U. Smilansky and N. Sondergaard. J. Phys.
A, 40(50):15143–15147, 2007.
[155] W. Bulla and T. Trenkler. The free Dirac operator on compact and noncom-
pact graphs. J. Math. Phys., 31(5):1157–1163, 1990.
[156] L. Bungart. On analytic fiber bundles. I. Holomorphic fiber bundles with
infinite dimensional fibers. Topology, 7:55–68, 1967.
[157] R. Burioni, D. Cassi, and A. Vezzani. Topology, hidden spectra and Bose-
Einstein condensation on low-dimensional complex networks. J. Phys. A,
35(5):1245–1252, 2002.
236 BIBLIOGRAPHY

[158] K. Busch, G. v. Freymann, S. Linden, S. F. Mingaleev, L. Tkeshelashvili, and

M. Wegener. Periodic nanostructures for photonics. Physics Reports, 444(3–
6):101–202, 2007.
[159] M. Büttiker. Absence of backscattering in the quantum Hall effect in multi-
probe conductors. Phys. Rev. B, 38(14):9375–9389, 1988.
[160] C. Cacciapuoti and P. Exner. Nontrivial edge coupling from a Dirichlet net-
work squeezing: the case of a bent waveguide. J. Phys. A, 40(26):F511–F523,
2007.
[161] J. Carini, D. Murdock, and J. Londergan. Binding and Scattering in Two-
Dimensional Systems: Applications to Quantum Wires, Waveguides and Pho-
tonic Crystals. Springer Lecture Notes in Physics. Springer-Verlag, Berlin,
1999.
[162] J. P. Carini, J. T. Londergan, and D. P. Murdock. Bound states in waveg-
uides and bent quantum wires. ii. electrons in quantum wires. Phys. Rev. B,
55(15):9852–9859, 1997.
[163] J. P. Carini, J. T. Londergan, D. P. Murdock, D. Trinkle, and C. S. Yung.
Bound states in waveguides and bent quantum wires. I. Applications to waveg-
uide systems. Phys. Rev. B, 55(15):9842–9851, 1997.
[164] R. Carlson. Adjoint and self-adjoint differential operators on graphs. Electron.
J. Differential Equations, pages No. 6, 10 pp. (electronic), 1998.
[165] R. Carlson. Hill’s equation for a regular graph. In Mathematical and numer-
ical aspects of wave propagation (Golden, CO, 1998), pages 227–231. SIAM,
Philadelphia, PA, 1998.
[166] R. Carlson. Inverse eigenvalue problems on directed graphs. Trans. Amer.
Math. Soc., 351(10):4069–4088, 1999.
[167] R. Carlson. Nonclassical Sturm-Liouville problems and Schrödinger opera-
tors on radial trees. Electron. J. Differential Equations, pages No. 71, 24 pp.
(electronic), 2000.
[168] R. Carlson. Spectral theory and spectral gaps for periodic Schrödinger opera-
tors on product graphs. Waves Random Media, 14(1):S29–S45, 2004. Special
section on quantum graphs.
[169] R. Carlson. Linear network models related to blood flow. In Quantum graphs
and their applications, volume 415 of Contemp. Math., pages 65–80. Amer.
Math. Soc., Providence, RI, 2006.
[170] R. Carlson. Boundary value problems for infinite metric graphs. In Analysis
on graphs and its applications, volume 77 of Proc. Sympos. Pure Math., pages
355–368. Amer. Math. Soc., Providence, RI, 2008.
[171] R. Carlson. Harmonic analysis for star graphs and the spherical coordinate
trapezoidal rule. J. Comput. Appl. Math., 235(8):2077–2089, 2011.
[172] R. Carlson. Spectral theory for nonconservative transmission line networks.
Netw. Heterog. Media, 6(2):257–277, 2011.
[173] R. Carlson and V. Pivovarchik. Spectral asymptotics for quantum graphs with
equal edge lengths. J. Phys. A, 41(14):145202, 16, 2008.
[174] J. Cartwright. Graphyne could be better than graphene. Science Now, 2012.
[175] H. Casimir. On the attraction between two perfectly conducting plates. Proc.
K. Ned. Akad. Wet., 51:793–795, 1948.
[176] H. B. G. Casimir and D. Polder. The influence of retardation on the London-
van der Waals forces. Phys. Rev., 73:360–372, Feb 1948.
 BIBLIOGRAPHY 237

[177] C. Cattaneo. The spectrum of the continuous Laplacian on a graph. Monatsh.

Math., 124(3):215–235, 1997.
[178] T. Ceccherini-Silberstein and M. Coornaert. Cellular automata and groups.
Springer Monographs in Mathematics. Springer-Verlag, Berlin, 2010.
[179] J. T. Chalker and P. D. Coddington. Percolation, quantum tunnelling and
the integer Hall effect. Journal of Physics C: Solid State Physics, 21(14):2665,
1988.
[180] S. N. Chandler-Wilde and M. Lindner. Limit operators, collective compact-
ness, and the spectral theory of infinite matrices. Mem. Amer. Math. Soc.,
210(989):viii+111, 2011.
[181] M. Cheney, D. Isaacson, and J. C. Newell. Electrical impedance tomography.
SIAM Rev., 41(1):85–101 (electronic), 1999.
[182] T. Cheon and P. Exner. An approximation to δ  couplings on graphs. J. Phys.
A, 37(29):L329–L335, 2004.
[183] T. Cheon, P. Exner, and O. Turek. Approximation of a general singular vertex
coupling in quantum graphs. Ann. Physics, 325(3):548–578, 2010.
[184] T. Cheon, P. Exner, and O. Turek. Tripartite connection condition for a
quantum graph vertex. Phys. Lett. A, 375(2):113–118, 2010.
[185] T. Cheon, P. Exner, and O. Turek. Inverse scattering problem for quantum
graph vertices. Phys. Rev. A, 83:062715, 2011.
[186] T. Cheon and O. Turek. Fulop-Tsutsui interactions on quantum graphs. Phys.
Lett. A, 374(41):4212–4221, 2010.
[187] V. L. Chernyshev. Time-dependent Schrödinger equation: statistics of the
distribution of Gaussian packets on a metric graph. Proc. Steklov Inst. Math.,
270(1):246–262, 2010.
[188] E. Chladni. Entdeckungen über die Theorie des Klanges. Weidmanns Erben
und Reich, Leipzig, 1787.
[189] E. Christensen, C. Ivan, and M. L. Lapidus. Dirac operators and spectral
triples for some fractal sets built on curves. Adv. Math., 217(1):42–78, 2008.
[190] F. Chung, M. Garrett, R. Graham, and D. Shallcross. Distance realization
problems with applications to internet tomography. J. Comput. System Sci.,
63(3):432–448, 2001.
[191] F. R. K. Chung. Spectral graph theory, volume 92 of CBMS Regional Con-
ference Series in Mathematics. Published for the Conference Board of the
Mathematical Sciences, Washington, DC, 1997.
[192] S.-Y. Chung and C. A. Berenstein. ω-harmonic functions and inverse conduc-
tivity problems on networks. SIAM J. Appl. Math., 65(4):1200–1226 (elec-
tronic), 2005.
[193] T. H. Colding and W. P. Minicozzi, II. Harmonic functions on manifolds.
Ann. of Math. (2), 146(3):725–747, 1997.
[194] Y. Colin de Verdière. Spectres de variétés riemanniennes et spectres de
graphes. In Proceedings of the International Congress of Mathematicians, Vol.
1, 2 (Berkeley, Calif., 1986), pages 522–530, Providence, RI, 1987. Amer.
Math. Soc.
[195] Y. Colin de Verdière. Le spectre du laplacien: survol partiel depuis le Berger-
Gauduchon-Mazet et problèmes. In Actes de la Table Ronde de Géométrie
Différentielle (Luminy, 1992), volume 1 of Sémin. Congr., pages 233–252.
Soc. Math. France, Paris, 1996.
238 BIBLIOGRAPHY

[196] Y. Colin de Verdière. Spectres de graphes, volume 4 of Cours Spécialisés [Spe-

cialized Courses]. Société Mathématique de France, Paris, 1998.
[197] Y. Colin de Verdière. Spectre d’opérateurs différentiels sur les graphes. In
Random walks and discrete potential theory (Cortona, 1997), Sympos. Math.,
XXXIX, pages 139–164. Cambridge Univ. Press, Cambridge, 1999.
[198] Y. Colin de Verdière. Magnetic interpretation of the nodal defect on graphs.
preprint arXiv:1201.1110, 2012.
[199] Y. Colin de Verdière, Y. Pan, and B. Ycart. Singular limits of Schrödinger
operators and Markov processes. J. Operator Theory, 41(1):151–173, 1999.
[200] D. Colton and R. Kress. Inverse acoustic and electromagnetic scattering the-
ory, volume 93 of Applied Mathematical Sciences. Springer-Verlag, Berlin,
second edition, 1998.
[201] V. Coluci, S. Braga, S. Legoas, D. Galvao, and R. Baughman. New Families
of Carbon Nanotubes. eprint arXiv:cond-mat/0207085, July 2002.
[202] M. Conti, S. Terracini, and G. Verzini. On a class of optimal partition prob-
lems related to the Fučı́k spectrum and to the monotonicity formulae. Calc.
Var. Partial Differential Equations, 22(1):45–72, 2005.
[203] J. H. Conway. The sensual (quadratic) form, volume 26 of Carus Mathematical
Monographs. Mathematical Association of America, Washington, DC, 1997.
With the assistance of Francis Y. C. Fung.
[204] H. D. Cornean, P. Duclos, and B. Ricaud. On the skeleton method and an
application to a quantum scissor. In Analysis on graphs and its applications,
volume 77 of Proc. Sympos. Pure Math., pages 657–672. Amer. Math. Soc.,
Providence, RI, 2008.
[205] R. Courant. Ein allgemeiner Satz zur Theorie der Eigenfunktione selbstad-
jungierter Differentialausdrücke. Nach. Ges. Wiss. Göttingen Math.-Phys.
Kl., pages 81–84, July 1923.
[206] R. Courant and D. Hilbert. Methods of mathematical physics. Vol. I. Inter-
science Publishers, Inc., New York, N.Y., 1953.
[207] R. Courant and D. Hilbert. Methods of mathematical physics. Vol. II. Wiley
Classics Library. John Wiley & Sons Inc., New York, 1989. Partial differential
equations. Reprint of the 1962 original. A Wiley-Interscience Publication.
[208] S. J. Cox and D. C. Dobson. Maximizing band gaps in two-dimensional pho-
tonic crystals. SIAM J. Appl. Math., 59(6):2108–2120 (electronic), 1999.
[209] S. J. Cox and D. C. Dobson. Band structure optimization of two-dimensional
photonic crystals in h-polarization. J. Comp. Phys., 158:214–224, 2000.
[210] M. Culler and K. Vogtmann. Moduli of graphs and automorphisms of free
groups. Invent. Math., 84(1):91–119, 1986.
[211] E. B. Curtis, D. Ingerman, and J. A. Morrow. Circular planar graphs and
resistor networks. Linear Algebra Appl., 283(1-3):115–150, 1998.
[212] E. B. Curtis and J. A. Morrow. Inverse Problems for Electrical Networks. On
Applied Mathematics. World Scientific Publishing Company, 2000.
[213] D. Cvetković, P. Rowlinson, and S. Simić. Eigenspaces of graphs, volume 66
of Encyclopedia of Mathematics and its Applications. Cambridge University
Press, Cambridge, 1997.
[214] D. M. Cvetković, M. Doob, I. Gutman, and A. Torgašev. Recent results in
the theory of graph spectra, volume 36 of Annals of Discrete Mathematics.
North-Holland Publishing Co., Amsterdam, 1988.
 BIBLIOGRAPHY 239

[215] D. M. Cvetković, M. Doob, and H. Sachs. Spectra of graphs. VEB Deutscher

Verlag der Wissenschaften, Berlin, second edition, 1982. Theory and applica-
tion.
[216] H. L. Cycon, R. G. Froese, W. Kirsch, and B. Simon. Schrödinger opera-
tors with application to quantum mechanics and global geometry. Texts and
Monographs in Physics. Springer-Verlag, Berlin, study edition, 1987.
[217] R. Dáger and E. Zuazua. Wave propagation, observation and control in 1-d
flexible multi-structures, volume 50 of Mathématiques & Applications (Berlin)
[Mathematics & Applications]. Springer-Verlag, Berlin, 2006.
[218] D. Dalvit, P. Milonni, D. Roberts, and F. da Rosa, editors. Casimir Physics,
volume 834 of Lecture Notes in Physics. Springer Verlag, Berlin, 2011.
[219] G. Davidoff, P. Sarnak, and A. Valette. Elementary number theory, group
theory, and Ramanujan graphs, volume 55 of London Mathematical Society
Student Texts. Cambridge University Press, Cambridge, 2003.
[220] E. Davies. An inverse spectral theorem. arXiv:1010.0094, 2010.
[221] E. Davies and A. Pushnitski. Non-Weyl resonance asymptotics for quantum
graphs. Analysis & PDE, 4:729–756, 2011.
[222] E. B. Davies. Spectral theory and differential operators, volume 42 of Cam-
bridge Studies in Advanced Mathematics. Cambridge University Press, Cam-
bridge, 1995.
[223] E. B. Davies. Linear operators and their spectra, volume 106 of Cambridge
Studies in Advanced Mathematics. Cambridge University Press, Cambridge,
2007.
[224] E. B. Davies, P. Exner, and J. Lipovský. Non-Weyl asymptotics for quantum
graphs with general coupling conditions. J. Phys. A, 43(47):474013, 16, 2010.
[225] E. B. Davies, G. M. L. Gladwell, J. Leydold, and P. F. Stadler. Discrete nodal
domain theorems. Linear Algebra Appl., 336:51–60, 2001.
[226] P.-G. de Gennes. Champ critique d’une boucle supraconductrice ramefiee. C.
R. Acad. Sci. Paris B, 292:279–282, 1981.
[227] B. Dekoninck and S. Nicaise. Spectre des réseaux de poutres. C. R. Acad.
Sci. Paris Sér. I Math., 326(10):1249–1254, 1998.
[228] B. Dekoninck and S. Nicaise. The eigenvalue problem for networks of beams.
In Generalized functions, operator theory, and dynamical systems (Brussels,
1997), volume 399 of Chapman & Hall/CRC Res. Notes Math., pages 335–
344. Chapman & Hall/CRC, Boca Raton, FL, 1999.
[229] B. Dekoninck and S. Nicaise. The eigenvalue problem for networks of beams.
Linear Algebra Appl., 314(1-3):165–189, 2000.
[230] G. Dell’Antonio and L. Tenuta. Quantum graphs as holonomic constraints.
J. Math. Phys., 47(7):072102, 21, 2006.
[231] G. F. Dell’Antonio. Limit motion on metric graphs. Adv. Nonlinear Stud.,
9(4):701–711, 2009.
[232] G. F. Dell’Antonio and E. Costa. Effective Schrödinger dynamics on -thin
Dirichlet waveguides via quantum graphs: I. Star-shaped graphs. J. Phys. A,
43(47):474014, 23, 2010.
[233] H. Donnelly and C. Fefferman. Nodal sets of eigenfunctions on Riemannian
manifolds. Invent. Math., 93(1):161–183, 1988.
[234] H. Donnelly and C. Fefferman. Growth and geometry of eigenfunctions of
the Laplacian. In Analysis and partial differential equations, volume 122 of
240 BIBLIOGRAPHY

Lecture Notes in Pure and Appl. Math., pages 635–655. Dekker, New York,
1990.
[235] H. Donnelly and C. Fefferman. Nodal sets for eigenfunctions of the Laplacian
on surfaces. J. Amer. Math. Soc., 3(2):333–353, 1990.
[236] H. Donnelly and C. Fefferman. Nodal sets of eigenfunctions: Riemannian
manifolds with boundary. In Analysis, et cetera, pages 251–262. Academic
Press, Boston, MA, 1990.
[237] H. Donnelly and C. Fefferman. Nodal domains and growth of harmonic func-
tions on noncompact manifolds. J. Geom. Anal., 2(1):79–93, 1992.
[238] J. P. Dowling. Photonic and acoustic band-gap bibliography,
http://phys.lsu.edu/ jdowling/pbgbib.html.
[239] B. Dubrovin, V. Matveev, and S. Novikov. Non-linear equations of Korteweg-
de Vries type, finite-zone linear operators, and abelian varieties. Russian
Math. Surv., 31(1):59–146, 1976.
[240] B. A. Dubrovin. Geometry of Hamiltonian evolutionary systems, volume 22
of Monographs and Textbooks in Physical Science. Lecture Notes. Bibliopolis,
Naples, 1991.
[241] B. A. Dubrovin, A. T. Fomenko, and S. P. Novikov. Modern geometry—
methods and applications. Part I, volume 93 of Graduate Texts in Mathe-
matics. Springer-Verlag, New York, second edition, 1992. The geometry of
surfaces, transformation groups, and fields. Translated from the Russian by
Robert G. Burns.
[242] B. A. Dubrovin, I. M. Krichever, and S. P. Novikov. Topological and algebraic
geometry methods in contemporary mathematical physics, volume 2 of Clas-
sic Reviews in Mathematics and Mathematical Physics. Cambridge Scientific
Publishers, Cambridge, 2004.
[243] P. Duclos and P. Exner. Curvature vs. thickness in quantum waveguides.
Czech. J. Phys., 441(11):1009–1018, 1991.
[244] P. Duclos and P. Exner. Curvature-induced bound states in quantum waveg-
uides in two and three dimensions. Rev. Math. Phys., 7(1):73–102, 1995.
[245] P. Duclos, P. Exner, and P. Štoviček. Curvature-induced resonances in a two-
dimensional Dirichlet tube. Ann. Inst. H. Poincaré Phys. Théor., 62(1):81–
101, 1995.
[246] N. Dunford and J. T. Schwartz. Linear operators. Part II. Wiley Classics
Library. John Wiley & Sons Inc., New York, 1988. Spectral theory. Selfad-
joint operators in Hilbert space, With the assistance of William G. Bade and
Robert G. Bartle, Reprint of the 1963 original, A Wiley-Interscience Publi-
cation.
[247] M. S. P. Eastham. The Spectral Theory of Periodic Differential Equations.
Scottish Academic Press Ltd., London, 1973.
[248] M. S. P. Eastham and H. Kalf. Schrödinger-type operators with continuous
spectra, volume 65 of Research Notes in Mathematics. Pitman (Advanced
Publishing Program), Boston, Mass., 1982.
[249] D. E. Edmunds and W. D. Evans. Spectral theory and differential operators.
Oxford Mathematical Monographs. The Clarendon Press Oxford University
Press, New York, 1987. Oxford Science Publications.
 BIBLIOGRAPHY 241

[250] S. Egger né Endres and F. Steiner. An exact trace formula and zeta functions
for an infinite quantum graph with a non-standard Weyl asymptotics. J.
Phys. A, 44(18):185202, 44, 2011.
[251] T. Ekholm, R. L. Frank, and H. Kovařı́k. Remarks about Hardy inequalities
on metric trees. In Analysis on graphs and its applications, volume 77 of Proc.
Sympos. Pure Math., pages 369–379. Amer. Math. Soc., Providence, RI, 2008.
[252] C. L. Epstein. Positive harmonic functions on abelian covers. J. Funct. Anal.,
82(2):303–315, 1989.
[253] G. Eskin, J. Ralston, and E. Trubowitz. Isospectral periodic potentials on
Rn . In Inverse problems (New York, 1983), volume 14 of SIAM-AMS Proc.,
pages 91–96. Amer. Math. Soc., Providence, RI, 1984.
[254] G. Eskin, J. Ralston, and E. Trubowitz. On isospectral periodic potentials in
Rn . II. Comm. Pure Appl. Math., 37(6):715–753, 1984.
[255] D. V. Evans, M. Levitin, and D. Vassiliev. Existence theorems or trapped
modes. J. Fluid Mech., 261:21–31, 1994.
[256] W. D. Evans and D. J. Harris. Fractals, trees and the Neumann Laplacian.
Math. Ann., 296(3):493–527, 1993.
[257] W. D. Evans, D. J. Harris, and Y. Saitō. On the approximation numbers of
Sobolev embeddings on singular domains and trees. Q. J. Math., 55(3):267–
302, 2004.
[258] W. D. Evans and Y. Saitō. Neumann Laplacians on domains and operators
on associated trees. Q. J. Math., 51(3):313–342, 2000.
[259] W. N. Everitt and L. Markus. Boundary value problems and symplectic algebra
for ordinary differential and quasi-differential operators, volume 61 of Math-
ematical Surveys and Monographs. American Mathematical Society, Provi-
dence, RI, 1999.
[260] W. N. Everitt and L. Markus. Complex symplectic geometry with applications
to ordinary differential operators. Trans. Amer. Math. Soc., 351(12):4905–
4945, 1999.
[261] W. N. Everitt and L. Markus. Multi-interval linear ordinary boundary
value problems and complex symplectic algebra. Mem. Amer. Math. Soc.,
151(715):viii+64, 2001.
[262] W. N. Everitt and L. Markus. Elliptic partial differential operators and sym-
plectic algebra. Mem. Amer. Math. Soc., 162(770):x+111, 2003.
[263] W. N. Everitt and L. Markus. Infinite dimensional complex symplectic spaces.
Mem. Amer. Math. Soc., 171(810):x+76, 2004.
[264] P. Exner. Lattice Kronig-Penney models. Phys. Rev. Lett., 74(18):3503–3506,
1995.
[265] P. Exner. Contact interactions on graph superlattices. J. Phys., 29(1):87–102,
1996.
[266] P. Exner. Leaky quantum graphs: a review. In Analysis on graphs and its
applications, volume 77 of Proc. Sympos. Pure Math., pages 523–564. Amer.
Math. Soc., Providence, RI, 2008.
[267] P. Exner and R. L. Frank. Absolute continuity of the spectrum for periodically
modulated leaky wires in R3 . Ann. Henri Poincaré, 8(2):241–263, 2007.
[268] P. Exner and R. Gawlista. Band spectra of rectangular graph superlattices.
Phys. Rev. B, 53(11):7275–7286, 1996.
242 BIBLIOGRAPHY

[269] P. Exner and R. Gawlista. Band spectra of rectangular graph superlattices.

Phys. Rev. B, 53(11):7275–7286, 1996.
[270] P. Exner, E. M. Harrell, and M. Loss. Optimal eigenvalues for some Laplacians
and Schrödinger operators depending on curvature. In Mathematical results in
quantum mechanics (Prague, 1998), volume 108 of Oper. Theory Adv. Appl.,
pages 47–58. Birkhäuser, Basel, 1999.
[271] P. Exner and T. Ichinose. Geometrically induced spectrum in curved leaky
wires. J. Phys. A, 34(7):1439–1450, 2001.
[272] P. Exner, J. P. Keating, P. Kuchment, T. Sunada, and A. Teplyaev, editors.
Analysis on graphs and its applications, volume 77 of Proc. Sympos. Pure
Math., Providence, RI, 2008. Amer. Math. Soc.
[273] P. Exner and S. Kondej. Curvature-induced bound states for a δ interaction
supported by a curve in R3 . Ann. Henri Poincaré, 3(5):967–981, 2002.
[274] P. Exner and H. Kovařı́k. Spectrum of the Schrödinger operator in a perturbed
periodically twisted tube. Lett. Math. Phys., 73(3):183–192, 2005.
[275] P. Exner, P. Kuchment, and B. Winn. On the location of spectral edges in
Z-periodic media. J. Phys. A, 43(47):474022, 8, 2010.
[276] P. Exner and J. Lipovský. Equivalence of resolvent and scattering resonances
on quantum graphs. In Adventures in mathematical physics, volume 447 of
Contemp. Math., pages 73–81. Amer. Math. Soc., Providence, RI, 2007.
[277] P. Exner and O. Post. Convergence of spectra of graph-like thin manifolds.
J. Geom. Phys., 54(1):77–115, 2005.
[278] P. Exner and O. Post. Convergence of resonances on thin branched quantum
waveguides. J. Math. Phys., 48(9):092104, 43, 2007.
[279] P. Exner and O. Post. Approximation of quantum graph vertex couplings
by scaled Schrödinger operators on thin branched manifolds. J. Phys. A,
42(41):415305, 22, 2009.
[280] P. Exner and P. Šeba. Trapping modes in a curved electromagnetic wave-
guide with perfectly conducting walls. Soobshcheniya Ob’edinennogo Instituta
Yadernykh Issledovaniı̆. Dubna [Communications of the Joint Institute for
Nuclear Research. Dubna], E5-88-379. Joint Inst. Nuclear Res., Dubna, 1988.
With a Russian summary.
[281] P. Exner and P. Šeba. Bound states in curved quantum waveguides. J. Math.
Phys., 30(11):2574–2580, 1989.
[282] P. Exner and P. Seba. Electrons in semiconductor microstructures: a chal-
lenge to operator theorists. In Proceedings of the Workshop on Schrödinger
Operators, Standard and Nonstandard (Dubna 1988), pages 79–100. World
Scientific, Singapore, 1989.
[283] P. Exner and P. Šeba. Free quantum motion on a branching graph. Rep. Math.
Phys., 28(1):7–26, 1989.
[284] P. Exner and P. Šeba. A “hybrid plane” with spin-orbit interaction. Russ. J.
Math. Phys., 14(4):430–434, 2007.
[285] P. Exner, P. Šeba, and P. Štoviček. On existence of a bound state in an
L-shaped waveguide. Czech. J. Phys., 39(11):1181–1191, 1989.
[286] P. Exner, P. Šeba, and P. Štoviček. Semiconductor edges can bind electrons.
Phys. Lett. A, 150(3–4):179–182, 1990.
[287] P. Exner and E. Šerešová. Appendix resonances on a simple graph. J. Phys.
A, 27(24):8269–8278, 1994.
 BIBLIOGRAPHY 243

[288] P. Exner and O. Turek. Approximations of permutation-symmetric vertex

couplings in quantum graphs. In Quantum graphs and their applications, vol-
ume 415 of Contemp. Math., pages 109–120. Amer. Math. Soc., Providence,
RI, 2006.
[289] P. Exner and O. Turek. Approximations of singular vertex couplings in quan-
tum graphs. Rev. Math. Phys., 19(6):571–606, 2007.
[290] P. Exner and O. Turek. High-energy asymptotics of the spectrum of a periodic
square lattice quantum graph. J. Phys. A, 43(47):474024, 25, 2010.
[291] P. Exner and K. Yoshitomi. Band gap of the Schrödinger operator with a
strong δ-interaction on a periodic curve. Ann. Henri Poincaré, 2(6):1139–
1158, 2001.
[292] P. Exner and K. Yoshitomi. Persistent currents for the 2D Schrödinger op-
erator with a strong δ-interaction on a loop. J. Phys. A, 35(15):3479–3487,
2002.
[293] P. Exner and V. A. Zagrebnov. Bose-Einstein condensation in geometrically
deformed tubes. J. Phys. A, 38(25):L463–L470, 2005.
[294] S. Fedorov and B. Pavlov. A remark on discrete wave scattering. J. Phys. A,
39(11):2657–2671, 2006.
[295] J. Feldman, H. Knörrer, and E. Trubowitz. Riemann surfaces of infinite
genus, volume 20 of CRM Monograph Series. American Mathematical So-
ciety, Providence, RI, 2003.
[296] F. Fidaleo, D. Guido, and T. Isola. Bose-Einstein condensation on inhomo-
geneous amenable graphs. Infin. Dimens. Anal. Quantum Probab. Relat. Top.,
14(2):149–197, 2011.
[297] A. Figà-Talamanca and C. Nebbia. Harmonic analysis and representation
theory for groups acting on homogeneous trees, volume 162 of London Mathe-
matical Society Lecture Note Series. Cambridge University Press, Cambridge,
1991.
[298] A. Figà-Talamanca and M. A. Picardello. Harmonic analysis on free groups,
volume 87 of Lecture Notes in Pure and Applied Mathematics. Marcel Dekker
Inc., New York, 1983.
[299] A. Figotin. Existence of gaps in the spectrum of periodic dielectric structures
on a lattice. J. Statist. Phys., 73(3–4):571–585, 1993.
[300] A. Figotin. Photonic pseudogaps for periodic dielectric structures. J. Statist.
Phys., 74(1–2):433–446, 1994.
[301] A. Figotin and Y. A. Godin. Spectral properties of thin-film photonic crystals.
SIAM J. Appl. Math., 61(6):1959–1979 (electronic), 2001.
[302] A. Figotin and P. Kuchment. Band-gap structure of spectra of periodic dielec-
tric and acoustic media. I. Scalar model. SIAM J. Appl. Math., 56(1):68–88,
1996.
[303] A. Figotin and P. Kuchment. Band-gap structure of spectra of periodic di-
electric and acoustic media. II. Two-dimensional photonic crystals. SIAM J.
Appl. Math., 56(6):1561–1620, 1996.
[304] A. Figotin and P. Kuchment. Spectral properties of classical waves in high-
contrast periodic media. SIAM J. Appl. Math., 58(2):683–702 (electronic),
1998.
[305] A. Figotin and I. Vitebsky. Spectra of periodic nonreciprocal electric circuits.
SIAM J. Appl. Math., 61(6):2008–2035 (electronic), 2001.
244 BIBLIOGRAPHY

[306] N. Filonov and F. Klopp. Absolute continuity of the spectrum of a Schrödinger

operator with a potential which is periodic in some directions and decays in
others. Doc. Math., 9:107–121 (electronic), 2004.
[307] H. Flechsig and S. Gnutzmann. On the spectral gap in Andreev graphs. In
Analysis on graphs and its applications, volume 77 of Proc. Sympos. Pure
Math., pages 381–395. Amer. Math. Soc., Providence, RI, 2008.
[308] G. Floquet. Sur les équations différentielles linéaires à coefficients périodiques.
Ann. Ecole Norm., 12(2):47–89, 1883.
[309] A. T. Fomenko. Symplectic geometry, volume 5 of Advanced Studies in Con-
temporary Mathematics. Gordon and Breach Science Publishers, New York,
1988. Translated from the Russian by R. S. Wadhwa.
[310] C. Fox, V. Oleinik, and B. Pavlov. A Dirichlet-to-Neumann map approach to
resonance gaps and bands of periodic networks. In Recent advances in differ-
ential equations and mathematical physics, volume 412 of Contemp. Math.,
pages 151–169. Amer. Math. Soc., Providence, RI, 2006.
[311] M. Freidlin. Markov processes and differential equations: asymptotic prob-
lems. Lectures in Mathematics ETH Zürich. Birkhäuser Verlag, Basel, 1996.
[312] M. Freidlin and M. Weber. Small diffusion asymptotics for exit problems on
graphs. In Quantum graphs and their applications, volume 415 of Contemp.
Math., pages 137–150. Amer. Math. Soc., Providence, RI, 2006.
[313] M. I. Freidlin and A. D. Wentzell. Diffusion processes on graphs and the
averaging principle. Ann. Probab., 21(4):2215–2245, 1993.
[314] M. I. Freidlin and A. D. Wentzell. Diffusion processes on an open book and
the averaging principle. Stochastic Process. Appl., 113(1):101–126, 2004.
[315] G. Freiling and V. Yurko. Inverse Sturm-Liouville problems and their appli-
cations. Nova Science Publishers Inc., Huntington, NY, 2001.
[316] G. Freiling and V. A. Yurko. Inverse nodal problems for differential operators
on graphs with a cycle. Tamkang J. Math., 41(1):15–24, 2010.
[317] L. Friedlander. On the spectrum of a class of second order periodic elliptic
differential operators. Comm. Math. Phys., 229(1):49–55, 2002.
[318] L. Friedlander. Absolute continuity of the spectra of periodic waveguides. In
Waves in periodic and random media (South Hadley, MA, 2002), volume 339
of Contemp. Math., pages 37–42. Amer. Math. Soc., Providence, RI, 2003.
[319] L. Friedlander. Genericity of simple eigenvalues for a metric graph. Israel J.
Math., 146:149–156, 2005.
[320] L. Friedlander. Determinant of the Schrödinger operator on a metric graph.
In Quantum graphs and their applications, volume 415 of Contemp. Math.,
pages 151–160. Amer. Math. Soc., Providence, RI, 2006.
[321] K. Fritzsche and H. Grauert. From holomorphic functions to complex man-
ifolds, volume 213 of Graduate Texts in Mathematics. Springer-Verlag, New
York, 2002.
[322] R. Froese and I. Herbst. Realizing holonomic constraints in classical and quan-
tum mechanics. Comm. Math. Phys., 220(3):489–535, 2001.
[323] R. Froese, I. Herbst, M. Hoffmann-Ostenhof, and T. Hoffmann-Ostenhof. L2 -
lower bounds to solutions of one-body Schrödinger equations. Proc. Roy. Soc.
Edinburgh Sect. A, 95(1–2):25–38, 1983.
[324] S. A. Fulling. Systematics of the relationship between vacuum energy calcu-
lations and heat-kernel coefficients. J. Phys. A, 36(24):6857–6873, 2003.
 BIBLIOGRAPHY 245

[325] S. A. Fulling. Local spectral density and vacuum energy near a quantum graph
vertex. In Quantum graphs and their applications, volume 415 of Contemp.
Math., pages 161–172. Amer. Math. Soc., Providence, RI, 2006.
[326] S. A. Fulling, L. Kaplan, K. Kirsten, Z. H. Liu, and K. A. Milton. Vacuum
stress and closed paths in rectangles, pistons and pistols. Journal of Physics
A: Mathematical and Theoretical, 42(15):155402, 2009.
[327] S. A. Fulling, L. Kaplan, and J. H. Wilson. Vacuum energy and repulsive
Casimir forces in quantum star graphs. Phys. Rev. A (3), 76(1):012118, 7,
2007.
[328] S. A. Fulling, P. Kuchment, and J. H. Wilson. Index theorems for quantum
graphs. J. Phys. A, 40(47):14165–14180, 2007.
[329] S. A. Fulling and J. H. Wilson. Vacuum energy and closed orbits in quantum
graphs. In Analysis on graphs and its applications, volume 77 of Proc. Sympos.
Pure Math., pages 673–679. Amer. Math. Soc., Providence, RI, 2008.
[330] Galileo. Discorsi e dimostrazioni matematiche, intorno ‘a due nuove scienze.
Louis Elsevier, Leiden, 1638.
[331] P. R. Garabedian. Partial differential equations. Chelsea Publishing Co., New
York, second edition, 1986.
[332] P. R. Garabedian and M. Schiffer. Convexity of domain functionals. J. Anal-
yse Math., 2:281–368, 1953.
[333] B. Gaveau and M. Okada. Differential forms and heat diffusion on one-
dimensional singular varieties. Bull. Sci. Math., 115(1):61–79, 1991.
[334] B. Gaveau, M. Okada, and T. Okada. Explicit heat kernels on graphs and
spectral analysis. In Several complex variables (Stockholm, 1987/1988), vol-
ume 38 of Math. Notes, pages 364–388. Princeton Univ. Press, Princeton, NJ,
1993.
[335] F. H. Gavilanez Alvarez. Network tomography. ProQuest LLC, Ann Arbor,
MI, 2006. Thesis (Ph.D.)–University of Maryland, College Park.
[336] U. Gavish and U. Smilansky. Degeneracies in the length spectra of metric
graphs. J. Phys. A, 40(33):10009–10020, 2007.
[337] A. Geim. Graphene: Status and prospects. Science, 324:1530–1534, 2009.
[338] A. Geim. Nobel lecture: Random walk to graphene. Rev. Mod. Phys., 83:851–
862, 2011.
[339] A. K. Geim and K. S. Novoselov. The rise of graphene. Nature Mater.,
6(3):183–191, 1950.
[340] A. Georgieva, T. Kriecherbauer, and S. Venakides. Wave propagation and
resonance in a one-dimensional nonlinear discrete periodic medium. SIAM J.
Appl. Math., 60(1):272–294 (electronic), 2000.
[341] C. Gérard and F. Nier. The Mourre theory for analytically fibered operators.
J. Funct. Anal., 152(1):202–219, 1998.
[342] N. I. Gerasimenko. The inverse scattering problem on a noncompact graph.
Teoret. Mat. Fiz., 75(2):187–200, 1988.
[343] N. I. Gerasimenko and B. S. Pavlov. A scattering problem on noncompact
graphs. Teoret. Mat. Fiz., 74(3):345–359, 1988.
[344] F. Gesztesy. Inverse spectral theory as influenced by Barry Simon. In Spectral
theory and mathematical physics: a Festschrift in honor of Barry Simon’s
60th birthday, volume 76 of Proc. Sympos. Pure Math., pages 741–820. Amer.
Math. Soc., Providence, RI, 2007.
246 BIBLIOGRAPHY

[345] F. Gesztesy and B. Simon. Inverse spectral analysis with partial information
on the potential. II. The case of discrete spectrum. Trans. Amer. Math. Soc.,
352(6):2765–2787, 2000.
[346] F. Gesztesy and B. Simon. A new approach to inverse spectral theory. II.
General real potentials and the connection to the spectral measure. Ann. of
Math. (2), 152(2):593–643, 2000.
[347] V. A. Geyler and I. Y. Popov. Eigenvalues imbedded in the band spectrum
for a periodic array of quantum dots. Rep. Math. Phys., 39(2):275–281, 1997.
[348] I. M. Glazman. Direct methods of qualitative spectral analysis of singular
differential operators. Translated from the Russian by the IPST staff. Israel
Program for Scientific Translations, Jerusalem, 1965, 1966.
[349] S. Gnutzmann and A. Altland. Universal spectral statistics in quantum
graphs. Phys. Rev. Lett., 93:194101, 2004.
[350] S. Gnutzmann and A. Altland. Spectral correlations of individual quantum
graphs. Phys. Rev. E (3), 72(5):056215, 14, 2005.
[351] S. Gnutzmann, P. D. Karageorge, and U. Smilansky. Can one count the shape
of a drum? Phys. Rev. Lett., 97(9):090201, 4, 2006.
[352] S. Gnutzmann, J. P. Keating, and F. Piotet. Eigenfunction statistics on quan-
tum graphs. Ann. Physics, 325(12):2595–2640, 2010.
[353] S. Gnutzmann and U. Smilansky. Quantum graphs: Applications to quantum
chaos and universal spectral statistics. Adv. Phys., 55(5–6):527–625, 2006.
[354] S. Gnutzmann, U. Smilansky, and S. Derevyanko. Stationary scattering from
a nonlinear network. Phys. Rev. A, 83:033831, Mar 2011.
[355] S. Gnutzmann, U. Smilansky, and N. Sondergaard. Resolving isospectral
‘drums’ by counting nodal domains. J. Phys. A, 38(41):8921–8933, 2005.
[356] S. Gnutzmann, U. Smilansky, and J. Weber. Nodal counting on quantum
graphs. Waves Random Media, 14(1):S61–S73, 2004. Special section on quan-
tum graphs.
[357] Y. A. Godin and B. Vainberg. A simple method for solving the inverse
scattering problem for the difference Helmholtz equation. Inverse Problems,
24(2):025007, 9, 2008.
[358] C. Godsil and G. Royle. Algebraic graph theory, volume 207 of Graduate Texts
in Mathematics. Springer-Verlag, New York, 2001.
[359] I. C. Gohberg and M. G. Kreı̆n. Fundamental aspects of defect numbers,
root numbers and indexes of linear operators. Uspehi Mat. Nauk (N.S.),
12(2(74)):43–118, 1957. English translation: The basic propositions on de-
fect numbers, root numbers and indices of linear operators, Amer. Math.
Soc. Transl. (2) 13 (1960), 185–264.
[360] I. C. Gohberg and M. G. Kreı̆n. Introduction to the theory of linear nonselfad-
joint operators. Translated from the Russian by A. Feinstein. Translations of
Mathematical Monographs, Vol. 18. American Mathematical Society, Provi-
dence, RI, 1969.
[361] N. Goldman and P. Gaspard. Quantum graphs and the integer quantum Hall
effect. Phys. Rev. B, 77:024302, 2008.
[362] A. M. Gomilko and V. N. Pivovarchik. The universe Sturm-Liouville problem
on a figure-eight-type graph. Ukraı̈n. Mat. Zh., 60(9):1168–1188, 2008.
[363] A. Grant. What is this? a psychedelic place mat? Discover, February, 2009.
 BIBLIOGRAPHY 247

[364] H. Grauert and R. Remmert. Theory of Stein spaces. Classics in Mathematics.

Springer-Verlag, Berlin, 2004. Translated from the German by Alan Huckle-
berry, Reprint of the 1979 translation.
[365] A. Greenleaf, Y. Kurylev, M. Lassas, and G. Uhlmann. Invisibility and inverse
problems. Bull. Amer. Math. Soc. (N.S.), 46(1):55–97, 2009.
[366] D. Gridin, R. V. Craster, and A. T. I. Adamou. Trapped modes in curved elas-
tic plates. Proc. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci., 461(2056):1181–
1197, 2005.
[367] D. Grieser. Monotone unitary families. preprint arXiv:0711.2869, 2007.
[368] D. Grieser. Spectra of graph neighborhoods and scattering. Proc. Lond. Math.
Soc. (3), 97(3):718–752, 2008.
[369] D. Grieser. Thin tubes in mathematical physics, global analysis and spectral
geometry. In Analysis on graphs and its applications, volume 77 of Proc.
Sympos. Pure Math., pages 565–593. Amer. Math. Soc., Providence, RI, 2008.
[370] J. S. Griffith. A free-electron theory of conjugated molecules. II. A derived
algebraic scheme. Proc. Camb. Philos. Soc., 49:650–658, 1953.
[371] J. S. Griffith. A free-electron theory of conjugated molecules. Part 1. — Poly-
cyclic hydrocarbons. Trans. Faraday Soc., 49:345–351, 1953.
[372] R. Grigorchuk and V. Nekrashevych. Self-similar groups, operator algebras
and Schur complement. J. Mod. Dyn., 1(3):323–370, 2007.
[373] R. Grigorchuk and Z. Šunić. Schreier spectrum of the Hanoi Towers group
on three pegs. In Analysis on graphs and its applications, volume 77 of Proc.
Sympos. Pure Math., pages 183–198. Amer. Math. Soc., Providence, RI, 2008.
[374] R. Grigorchuk and Z. Šuniḱ. Asymptotic aspects of Schreier graphs and Hanoi
Towers groups. C. R. Math. Acad. Sci. Paris, 342(8):545–550, 2006.
[375] R. I. Grigorchuk and A. Żuk. The Ihara zeta function of infinite graphs, the
KNS spectral measure and integrable maps. In Random walks and geometry,
pages 141–180. Walter de Gruyter GmbH & Co. KG, Berlin, 2004.
[376] A. Grigor yan. Heat kernels on manifolds, graphs and fractals. In European
Congress of Mathematics, Vol. I (Barcelona, 2000), volume 201 of Progr.
Math., pages 393–406. Birkhäuser, Basel, 2001.
[377] A. Grigor’yan. Heat kernels and function theory on metric measure spaces.
In Heat kernels and analysis on manifolds, graphs, and metric spaces (Paris,
2002), volume 338 of Contemp. Math., pages 143–172. Amer. Math. Soc.,
Providence, RI, 2003.
[378] A. Grigor’yan. Heat kernels on metric measure spaces with regular volume
growth. In Handbook of geometric analysis, No. 2, volume 13 of Adv. Lect.
Math. (ALM), pages 1–60. Int. Press, Somerville, MA, 2010.
[379] A. Grigor yan and T. Kumagai. On the dichotomy in the heat kernel two
sided estimates. In Analysis on graphs and its applications, volume 77 of
Proc. Sympos. Pure Math., pages 199–210. Amer. Math. Soc., Providence,
RI, 2008.
[380] A. Grigor yan and A. Telcs. Sub-Gaussian estimates of heat kernels on infinite
graphs. Duke Math. J., 109(3):451–510, 2001.
[381] P. Grinfeld. Hadamard’s formula inside and out. J. Optim. Theory Appl.,
146(3):654–690, 2010.
[382] M. J. Gruber, D. H. Lenz, and I. Veselić. Uniform existence of the integrated
density of states for combinatorial and metric graphs over Zd . In Analysis on
248 BIBLIOGRAPHY

graphs and its applications, volume 77 of Proc. Sympos. Pure Math., pages
87–108. Amer. Math. Soc., Providence, RI, 2008.
[383] F. Guevara Vasquez, G. W. Milton, and D. Onofrei. Complete characteri-
zation and synthesis of the response function of elastodynamic networks. J.
Elasticity, 102(1):31–54, 2011.
[384] F. Guevara Vasquez, G. W. Milton, and D. Onofrei. Exterior cloaking with
active sources in two dimensional acoustics. Wave Motion, 48(6):515–524,
2011.
[385] D. Guido, T. Isola, and M. L. Lapidus. Bartholdi zeta functions for periodic
simple graphs. In Analysis on graphs and its applications, volume 77 of Proc.
Sympos. Pure Math., pages 109–122. Amer. Math. Soc., Providence, RI, 2008.
[386] R. C. Gunning and H. Rossi. Analytic functions of several complex variables.
AMS Chelsea Publishing, Providence, RI, 2009. Reprint of the 1965 original.
[387] B. Gutkin and U. Smilansky. Can one hear the shape of a graph? J. Phys.
A, 34(31):6061–6068, 2001.
[388] F. Haake. Quantum signatures of chaos. Springer Series in Synergetics.
Springer-Verlag, Berlin, 3rd edition, 2010.
[389] J. Hadamard. Mémoire sur le problème d’analyse relatif à l’équilibre des
plaques elastiques encastrées. Mem. Acad. Sci. Inst. de France, 1908.
[390] R. Hadani and A. Singer. Representation theoretic patterns in three dimen-
sional cryo-electron microscopy I: The intrinsic reconstitution algorithm. Ann.
of Math. (2), 174(2):1219–1241, 2011.
[391] R. Hadani and A. Singer. Representation theoretic patterns in three-
dimensional cryo-electron microscopy II—the class averaging problem. Found.
Comput. Math., 11(5):589–616, 2011.
[392] O. H. Hald and J. R. McLaughlin. Inverse nodal problems: finding the po-
tential from nodal lines. Mem. Amer. Math. Soc., 119(572):viii+148, 1996.
[393] M. Harmer. Hermitian symplectic geometry and extension theory. J. Phys.
A, 33(50):9193–9203, 2000.
[394] M. Harmer. Inverse scattering on matrices with boundary conditions. J. Phys.
A, 38(22):4875–4885, 2005.
[395] M. Harmer. Spin filtering on a ring with the Rashba Hamiltonian. J. Phys.
A, 39(46):14329–14341, 2006.
[396] M. Harmer, A. Mikhailova, and B. S. Pavlov. Manipulating the electron cur-
rent through a splitting. In Geometric analysis and applications (Canberra,
2000), volume 39 of Proc. Centre Math. Appl. Austral. Nat. Univ., pages
118–131. Austral. Nat. Univ., Canberra, 2001.
[397] M. S. Harmer. Inverse scattering for the matrix Schrödinger operator and
Schrödinger operator on graphs with general self-adjoint boundary conditions.
ANZIAM J., 44(1):161–168, 2002. Kruskal, 2000 (Adelaide).
[398] P. J. F. Harris. Carbon Nano-tubes and Related Structures. Cambridge Uni-
versity Press, Cambridge, 1999. New Materials for the Twenty-first Century.
[399] J. M. Harrison. Quantum graphs with spin Hamiltonians. In Analysis on
graphs and its applications, volume 77 of Proc. Sympos. Pure Math., pages
261–277. Amer. Math. Soc., Providence, RI, 2008.
[400] J. M. Harrison, J. P. Keating, and J. M. Robbins. Quantum statistics on
graphs. Proc. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci., 467(2125):212–
233, 2011.
 BIBLIOGRAPHY 249

[401] J. M. Harrison and K. Kirsten. Zeta functions of quantum graphs. J. Phys.

A, 44(23):235301, 29, 2011.
[402] J. M. Harrison, K. Kirsten, and C. Texier. Spectral determinants and
zeta functions of Schrödinger operators on metric graphs. J. Phys. A,
45(12):125206, 2012.
[403] J. M. Harrison, P. Kuchment, A. Sobolev, and B. Winn. On occurrence of
spectral edges for periodic operators inside the Brillouin zone. J. Phys. A,
40(27):7597–7618, 2007.
[404] J. M. Harrison, U. Smilansky, and B. Winn. Quantum graphs where back-
scattering is prohibited. J. Phys. A, 40(47):14181–14193, 2007.
[405] J. M. Harrison and B. Winn. Intermediate statistics for a system with sym-
plectic symmetry: the Dirac rose graph. preprint arXiv:1205.6073, 2012.
[406] J. E. Heebner, R. W. Boyd, and Q. Park. Scissor solitons & other propagation
effects in microresonator modified waveguides. JOSA B, 19:722–731, 2002.
[407] P. Hejčı́k and T. Cheon. Enhanced all-optical switching using a nonlinear
fiber ring resonator. Optics Letters, 24:847–849, 1999.
[408] B. Helffer. Domaines nodaux et partitions spectrales minimales (d’après B.
Helffer, T. Hoffmann-Ostenhof et S. Terracini). In Séminaire: Équations aux
Dérivées Partielles. 2006–2007, Sémin. Équ. Dériv. Partielles, pages Exp. No.
VIII, 23. École Polytech., Palaiseau, 2007.
[409] B. Helffer, T. Hoffmann-Ostenhof, and S. Terracini. Nodal domains and
spectral minimal partitions. Ann. Inst. H. Poincaré Anal. Non Linéaire,
26(1):101–138, 2009.
[410] S. Heusler, S. Muller, P. Braun, and F. Haake. Universal spectral form factor
for chaotic dynamics. J. Phys. A-Math. Gen., 37(3):L31–L37, 2004.
[411] Y. Higuchi and T. Shirai. Some spectral and geometric properties for infinite
graphs. In Discrete geometric analysis, volume 347 of Contemp. Math., pages
29–56. Amer. Math. Soc., Providence, RI, 2004.
[412] D. Hinton. Sturm’s 1836 oscillation results: evolution of the theory. In Sturm-
Liouville theory, pages 1–27. Birkhäuser, Basel, 2005.
[413] P. D. Hislop and O. Post. Anderson localization for radial tree-like quantum
graphs. Waves Random Complex Media, 19(2):216–261, 2009.
[414] P. D. Hislop and I. M. Sigal. Introduction to spectral theory, volume 113
of Applied Mathematical Sciences. Springer-Verlag, New York, 1996. With
applications to Schrödinger operators.
[415] L. Hogben. Spectral graph theory and the inverse eigenvalue problem of a
graph. Electron. J. Linear Algebra, 14:12–31 (electronic), 2005.
[416] J. Holmer, J. Marzuola, and M. Zworski. Fast soliton scattering by delta
impurities. Comm. Math. Phys., 274(1):187–216, 2007.
[417] L. Hörmander. An introduction to complex analysis in several variables, vol-
ume 7 of North-Holland Mathematical Library. North-Holland Publishing Co.,
Amsterdam, third edition, 1990.
[418] M. D. Horton, H. M. Stark, and A. A. Terras. What are zeta functions of
graphs and what are they good for? In Quantum graphs and their appli-
cations, volume 415 of Contemp. Math., pages 173–189. Amer. Math. Soc.,
Providence, RI, 2006.
250 BIBLIOGRAPHY

[419] M. D. Horton, H. M. Stark, and A. A. Terras. Zeta functions of weighted

graphs and covering graphs. In Analysis on graphs and its applications, vol-
ume 77 of Proc. Sympos. Pure Math., pages 29–50. Amer. Math. Soc., Prov-
idence, RI, 2008.
[420] O. Hul, M. L awniczak, S. Bauch, and L. Sirko. Simulation of quantum graphs
by microwave networks. In Analysis on graphs and its applications, volume 77
of Proc. Sympos. Pure Math., pages 595–615. Amer. Math. Soc., Providence,
RI, 2008.
[421] N. E. Hurt. Mathematical physics of quantum wires and devices, volume 506 of
Mathematics and its Applications. Kluwer Academic Publishers, Dordrecht,
2000. From spectral resonances to Anderson localization.
[422] A. Iantchenko and E. Korotyaev. Schrödinger operator on the zigzag half-
nanotube in magnetic field. Math. Model. Nat. Phenom., 5(4):175–197, 2010.
[423] Y. Imry. Introduction to Mesoscopic Physics (Mesoscopic Physics and Nan-
otechnology). Oxford University Press, New York, 1997.
[424] D. Ingerman and J. A. Morrow. On a characterization of the kernel of
the Dirichlet-to-Neumann map for a planar region. SIAM J. Math. Anal.,
29(1):106–115 (electronic), 1998.
[425] D. V. Ingerman. Discrete and continuous Dirichlet-to-Neumann maps in the
layered case. SIAM J. Math. Anal., 31(6):1214–1234 (electronic), 2000.
[426] V. Isakov. Inverse problems for partial differential equations, volume 127 of
Applied Mathematical Sciences. Springer, New York, second edition, 2006.
[427] L. Ivanov, L. Kotko, and S. Kreı̆n. Boundary value problems in variable
domains. Differencial nye Uravnenija i Primenen.—Trudy Sem., 19:1–161,
1977.
[428] J. D. Jackson. Classical electrodynamics. John Wiley & Sons Inc., New York,
second edition, 1975.
[429] J. D. Joannopoulos, S. Johnson, R. D. Meade, and J. N. Winn. Photonic
Crystals: Molding the Flow of Light. Princeton University Press, Princeton,
N.J., 2nd edition, 2008.
[430] S. John. Strong localization of photons in certain disordered dielectric super-
lattices. Phys. Rev. Lett., 58(23):2486–2489, 1987.
[431] S. G. Johnson. Numerical methods for computing Casimir interactions. In
D. Dalvit, P. Milonni, D. Roberts, and F. da Rosa, editors, Casimir Physics,
volume 834 of Lecture Notes in Physics, chapter 6, pages 175–218. Springer,
Berlin, June 2011.
[432] S. G. Johnson and J. D. Joannopoulos. Photonic Crystals, The Road from
Theory to Practice. Kluwer Academic Publishers, Boston, 2002.
[433] V. Kaimanovich. The Poisson boundary of covering Markov operators. Israel
J. Math., 89(1–3):77–134, 1995.
[434] P. D. Karageorge and U. Smilansky. Counting nodal domains on surfaces of
revolution. J. Phys. A, 41(20):205102, 2008.
[435] A. Katchalov, Y. Kurylev, and M. Lassas. Inverse boundary spectral problems,
volume 123 of Chapman & Hall/CRC Monographs and Surveys in Pure and
Applied Mathematics. Chapman & Hall/CRC, Boca Raton, FL, 2001.
[436] T. Kato. Perturbation theory for linear operators. Classics in Mathematics.
Springer-Verlag, Berlin, 1995. Reprint of the 1980 edition.
 BIBLIOGRAPHY 251

[437] J. P. Keating. Fluctuation statistics for quantum star graphs. In Quantum

graphs and their applications, volume 415 of Contemp. Math., pages 191–200.
Amer. Math. Soc., Providence, RI, 2006.
[438] J. P. Keating. Quantum graphs and quantum chaos. In Analysis on graphs
and its applications, volume 77 of Proc. Sympos. Pure Math., pages 279–290.
Amer. Math. Soc., Providence, RI, 2008.
[439] J. P. Keating, J. Marklof, and B. Winn. Value distribution of the eigenfunc-
tions and spectral determinants of quantum star graphs. Comm. Math. Phys.,
241(2–3):421–452, 2003.
[440] M. Kelbert and Y. Suhov. Asymptotic properties of Markov processes on
Cayley trees. In Analysis on graphs and its applications, volume 77 of Proc.
Sympos. Pure Math., pages 123–138. Amer. Math. Soc., Providence, RI, 2008.
[441] R. Kenyon. The Laplacian and Dirac operators on critical planar graphs.
Invent. Math., 150(2):409–439, 2002.
[442] J. Kigami. Analysis on fractals, volume 143 of Cambridge Tracts in Mathe-
matics. Cambridge University Press, Cambridge, 2001.
[443] B. G. Kim and H. J. Choi. Graphyne: hexagonal network of carbon with
versatile Dirac cone. ArXiv e-prints, Dec. 2011.
[444] A. Kirsch and N. Grinberg. The factorization method for inverse problems,
volume 36 of Oxford Lecture Series in Mathematics and its Applications. Ox-
ford University Press, Oxford, 2008.
[445] D. Klawonn. Inverse nodal problems. J. Phys. A, 42(17)(11):175209, 2009.
[446] A. Klein. Absolutely continuous spectrum in the Anderson model on the
Bethe lattice. Math. Res. Lett., 1(4):399–407, 1994.
[447] A. Klein. The Anderson metal-insulator transition on the Bethe lattice. In
XIth International Congress of Mathematical Physics (Paris, 1994), pages
383–391. Int. Press, Cambridge, MA, 1995.
[448] A. Klein. Extended states in the Anderson model on the Bethe lattice. Adv.
Math., 133(1):163–184, 1998.
[449] F. Klopp and K. Pankrashkin. Localization on quantum graphs with random
vertex couplings. J. Stat. Phys., 131(4):651–673, 2008.
[450] F. Klopp and K. Pankrashkin. Localization on quantum graphs with random
edge lengths. Lett. Math. Phys., 87(1-2):99–114, 2009.
[451] E. Korotyaev. Effective masses for zigzag nanotubes in magnetic fields. Lett.
Math. Phys., 83(1):83–95, 2008.
[452] E. Korotyaev. Zigzag nanoribbons in external electric and magnetic fields. J.
Comp. Sci. Math., 3(1–2):168–191, 2010.
[453] E. Korotyaev and A. Kutsenko. Zigzag nanoribbons in external electric fields.
Asympt. Anal., 66(3–4):187–206, 2010.
[454] E. Korotyaev and I. Lobanov. Zigzag periodic nanotube in magnetic field.
preprint arXiv:math.SP/0604007, 2006.
[455] E. Korotyaev and I. Lobanov. Schrödinger operators on zigzag nanotubes.
Ann. Henri Poincaré, 8(6):1151–1176, 2007.
[456] V. Kostrykin, J. Potthoff, and R. Schrader. Heat kernels on metric graphs
and a trace formula. In Adventures in mathematical physics, volume 447 of
Contemp. Math., pages 175–198. Amer. Math. Soc., Providence, RI, 2007.
252 BIBLIOGRAPHY

[457] V. Kostrykin, J. Potthoff, and R. Schrader. Contraction semigroups on metric

graphs. In Analysis on graphs and its applications, volume 77 of Proc. Sympos.
Pure Math., pages 423–458. Amer. Math. Soc., Providence, RI, 2008.
[458] V. Kostrykin, J. Potthoff, and R. Schrader. Finite propagation speed for
solutions of the wave equation on metric graphs. J. Funct. Anal., 263(5):1198–
1223, 2012.
[459] V. Kostrykin and R. Schrader. Kirchhoff’s rule for quantum wires. J. Phys.
A, 32(4):595–630, 1999.
[460] V. Kostrykin and R. Schrader. The generalized star product and the factor-
ization of scattering matrices on graphs. J. Math. Phys., 42(4):1563–1598,
2001.
[461] V. Kostrykin and R. Schrader. Quantum wires with magnetic fluxes. Comm.
Math. Phys., 237(1-2):161–179, 2003. Dedicated to Rudolf Haag.
[462] V. Kostrykin and R. Schrader. A random necklace model. Waves Random
Media, 14(1):S75–S90, 2004. Special section on quantum graphs.
[463] V. Kostrykin and R. Schrader. Laplacians on metric graphs: eigenvalues,
resolvents and semigroups. In Quantum graphs and their applications, volume
415 of Contemp. Math., pages 201–225. Amer. Math. Soc., Providence, RI,
2006.
[464] M. Kotani and T. Sunada. Albanese maps and off diagonal long time asymp-
totics for the heat kernel. Comm. Math. Phys., 209(3):633–670, 2000.
[465] M. Kotani and T. Sunada. Spectral geometry of crystal lattices. In Heat
kernels and analysis on manifolds, graphs, and metric spaces (Paris, 2002),
volume 338 of Contemp. Math., pages 271–305. Amer. Math. Soc., Providence,
RI, 2003.
[466] T. Kottos and U. Smilansky. Quantum chaos on graphs. Phys. Rev. Lett.,
79(24):4794–4797, 1997.
[467] T. Kottos and U. Smilansky. Periodic orbit theory and spectral statistics for
quantum graphs. Ann. Physics, 274(1):76–124, 1999.
[468] T. Kottos and U. Smilansky. Chaotic scattering on graphs. Phys. Rev. Lett.,
85(5):968–971, 2000.
[469] T. Kottos and U. Smilansky. Quantum graphs: a simple model for chaotic
scattering. J. Phys. A, 36(12):3501–3524, 2003. Random matrix theory.
[470] M. G. Krein and G. I. Liubarskii. On the theory of transmission bands of
periodic waveguides. J. Appl. Math. Mech., 25:29–48, 1961.
[471] S. G. Kreı̆n. Linear equations in Banach spaces. Birkhäuser, Boston, Mass.,
1982. Translated from the Russian by A. Iacob. With an introduction by I.
Gohberg.
[472] D. Krejčiřı́k. Twisting versus bending in quantum waveguides. In Analysis on
graphs and its applications, volume 77 of Proc. Sympos. Pure Math., pages
617–637. Amer. Math. Soc., Providence, RI, 2008.
[473] S. Krstić and K. Vogtmann. Equivariant outer space and automorphisms of
free-by-finite groups. Comment. Math. Helv., 68(2):216–262, 1993.
[474] P. Kuchment. Floquet theory for partial differential equations, volume 60 of
Operator Theory: Advances and Applications. Birkhäuser Verlag, Basel, 1993.
[475] P. Kuchment. The mathematics of photonic crystals. In G. Bao, L. Cowsar,
and W. Masters, editors, Mathematical modeling in optical science, volume 22
of Frontiers Appl. Math., pages 207–272. SIAM, Philadelphia, PA, 2001.
 BIBLIOGRAPHY 253

[476] P. Kuchment. Graph models for waves in thin structures. Waves Random
Media, 12(4):R1–R24, 2002.
[477] P. Kuchment. Differential and pseudo-differential operators on graphs as mod-
els of mesoscopic systems. In H. Begehr, R. Gilbert, and M. W. Wang, edi-
tors, Analysis and applications—ISAAC 2001 (Berlin), volume 10 of Int. Soc.
Anal. Appl. Comput., pages 7–30. Kluwer Acad. Publ., Dordrecht, 2003.
[478] P. Kuchment, editor. Waves in periodic and random media (South Hadley,
MA, 2002), volume 339, Providence, RI, 2003. Amer. Math. Soc.
[479] P. Kuchment, editor. Quantum graphs and their applications, volume 14.
Waves in Random Media, 2004.
[480] P. Kuchment. Quantum graphs. I. Some basic structures. Waves Random
Media, 14(1):S107–S128, 2004. Special section on quantum graphs.
[481] P. Kuchment. Quantum graphs. II. Some spectral properties of quantum and
combinatorial graphs. J. Phys. A, 38(22):4887–4900, 2005.
[482] P. Kuchment. Quantum graphs: an introduction and a brief survey. In Anal-
ysis on graphs and its applications, volume 77 of Proc. Sympos. Pure Math.,
pages 291–312. Amer. Math. Soc., Providence, RI, 2008.
[483] P. Kuchment and L. Kunyansky. Differential operators on graphs and pho-
tonic crystals. Adv. Comput. Math., 16(2–3):263–290, 2002. Modeling and
computation in optics and electromagnetics.
[484] P. Kuchment and L. A. Kunyansky. Spectral properties of high contrast band-
gap materials and operators on graphs. Experiment. Math., 8(1):1–28, 1999.
[485] P. Kuchment and S. Levendorskiı̂. On the structure of spectra of periodic
elliptic operators. Trans. Amer. Math. Soc., 354(2):537–569 (electronic), 2002.
[486] P. Kuchment and B. Ong. On guided waves in photonic crystal waveguides.
In Waves in periodic and random media (South Hadley, MA, 2002), volume
339 of Contemp. Math., pages 105–115. Amer. Math. Soc., Providence, RI,
2003.
[487] P. Kuchment and Y. Pinchover. Integral representations and Liouville theo-
rems for solutions of periodic elliptic equations. J. Funct. Anal., 181(2):402–
446, 2001.
[488] P. Kuchment and Y. Pinchover. Liouville theorems and spectral edge be-
havior on abelian coverings of compact manifolds. Trans. Amer. Math. Soc.,
359(12):5777–5815 (electronic), 2007.
[489] P. Kuchment and O. Post. On the spectra of carbon nano-structures. Comm.
Math. Phys., 275(3):805–826, 2007.
[490] P. Kuchment and B. Vainberg. On absence of embedded eigenvalues for
Schrödinger operators with perturbed periodic potentials. Comm. Partial Dif-
ferential Equations, 25(9–10):1809–1826, 2000.
[491] P. Kuchment and B. Vainberg. On the structure of eigenfunctions correspond-
ing to embedded eigenvalues of locally perturbed periodic graph operators.
Comm. Math. Phys., 268(3):673–686, 2006.
[492] P. Kuchment and H. Zeng. Convergence of spectra of mesoscopic systems
collapsing onto a graph. J. Math. Anal. Appl., 258(2):671–700, 2001.
[493] P. Kuchment and H. Zeng. Asymptotics of spectra of Neumann Laplacians in
thin domains. In Y. Karpeshina, G. Stolz, R. Weikard, and Y. Zeng, editors,
Advances in differential equations and mathematical physics (Birmingham,
254 BIBLIOGRAPHY

AL, 2002), volume 327 of Contemp. Math., pages 199–213. Amer. Math.
Soc., Providence, RI, 2003.
[494] P. A. Kuchment. On the Floquet theory of periodic difference equations.
In Geometrical and algebraical aspects in several complex variables (Cetraro,
1989), volume 8 of Sem. Conf., pages 201–209. EditEl, Rende, 1991.
[495] N. H. Kuiper. The homotopy type of the unitary group of Hilbert space.
Topology, 3:19–30, 1965.
[496] P. Kurasov. Graph Laplacians and topology. Ark. Mat., 46(1):95–111, 2008.
[497] P. Kurasov. Schrödinger operators on graphs and geometry. I. Essentially
bounded potentials. J. Funct. Anal., 254(4):934–953, 2008.
[498] P. Kurasov. Inverse problems for Aharonov-Bohm rings. Math. Proc. Cam-
bridge Philos. Soc., 148(2):331–362, 2010.
[499] P. Kurasov and M. Nowaczyk. Inverse spectral problem for quantum graphs.
J. Phys. A, 38(22):4901–4915, 2005.
[500] P. Kurasov and M. Nowaczyk. Corrigendum: “Inverse spectral problem for
quantum graphs” [J. Phys. A 38 (2005), no. 22, 4901–4915; mr2148632]. J.
Phys. A, 39(4):993, 2006.
[501] P. Kurasov and M. Nowaczyk. Geometric properties of quantum graphs and
vertex scattering matrices. Opuscula Math., 30(3):295–309, 2010.
[502] P. Kurasov and F. Stenberg. On the inverse scattering problem on branching
graphs. J. Phys. A, 35(1):101–121, 2002.
[503] A. Lambrecht. The casimir effect: a force from nothing. Physics World, pages
29–32, September 2002.
[504] S. Lang. Introduction to differentiable manifolds. Universitext. Springer-
Verlag, New York, second edition, 2002.
[505] M. L. Lapidus and E. P. J. Pearse. Tube formulas for self-similar fractals.
In Analysis on graphs and its applications, volume 77 of Proc. Sympos. Pure
Math., pages 211–230. Amer. Math. Soc., Providence, RI, 2008.
[506] C. K. Law and V. Pivovarchik. Characteristic functions of quantum graphs.
J. Phys. A, 42(3):035302, 12, 2009.
[507] E. Lawrence, G. Michailidis, V. N. Nair, and B. Xi. Network tomography:
a review and recent developments. In Frontiers in statistics, pages 345–366.
Imp. Coll. Press, London, 2006.
[508] D. Lenz, C. Schubert, and P. Stollmann. Eigenfunction expansions for
Schrödinger operators on metric graphs. Integral Equations Operator Theory,
62(4):541–553, 2008.
[509] D. Lenz, P. Stollmann, and I. Veselic. Generalized eigenfunctions and spectral
theory for strongly local Dirichlet forms. preprint arXiv:0909.1107, 2009.
[510] U. Leonhardt and T. Philbin. Geometry and light. Dover Publications Inc.,
Mineola, NY, 2010. The science of invisibility.
[511] B. M. Levitan. Inverse Sturm-Liouville problems. VSP, Zeist, 1987. Translated
from the Russian by O. Efimov.
[512] B. M. Levitan and I. S. Sargsjan. Introduction to spectral theory: selfadjoint
ordinary differential operators. American Mathematical Society, Providence,
R.I., 1975. Translated from the Russian by Amiel Feinstein, Translations of
Mathematical Monographs, Vol. 39.
 BIBLIOGRAPHY 255

[513] B. M. Levitan and I. S. Sargsjan. Sturm-Liouville and Dirac operators, vol-

ume 59 of Mathematics and its Applications (Soviet Series). Kluwer Academic
Publishers Group, Dordrecht, 1991. Translated from the Russian.
[514] F. Leys, C. Amovilli, and N. March. Topology, connectivity, and electronic
structure of C and B cages and the corresponding nanotubes. J. Chem. Inf.
Comput. Sci., 44(1):122–135, 2004.
[515] P. Li. Curvature and function theory on Riemannian manifolds. In Surveys
in differential geometry, Surv. Differ. Geom., VII, pages 375–432. Int. Press,
Somerville, MA, 2000.
[516] P. Li and J. Wang. Counting dimensions of L-harmonic functions. Ann. of
Math. (2), 152(2):645–658, 2000.
[517] P. Li and J. Wang. Polynomial growth solutions of uniformly elliptic op-
erators of non-divergence form. Proc. Amer. Math. Soc., 129(12):3691–3699
(electronic), 2001.
[518] E. H. Lieb and M. Loss. Fluxes, Laplacians, and Kasteleyn’s theorem. Duke
Math. J., 71(2):337–363, 1993.
[519] I. M. Lifshits, S. A. Gredeskul, and L. A. Pastur. Introduction to the theory
of disordered systems. A Wiley-Interscience Publication. John Wiley & Sons
Inc., New York, 1988. Translated from the Russian by Eugene Yankovsky [E.
M. Yankovskiı̆].
[520] V. Y. Lin. Liouville coverings of complex spaces, and amenable groups. Mat.
Sb. (N.S.), 132(174)(2):202–224, 1987.
[521] V. Y. Lin and M. Zaidenberg. Liouville and Carathéodory coverings in Rie-
mannian and complex geometry. In Voronezh Winter Mathematical Schools,
volume 184 of Amer. Math. Soc. Transl. Ser. 2, pages 111–130. Amer. Math.
Soc., Providence, RI, 1998.
[522] M. S. Livšic. Operators, oscillations, waves (open systems). American Mathe-
matical Society, Providence, RI, 1973. Translated from the Russian by Scripta
Technica, Ltd. English translation edited by R. Herden. Translations of Math-
ematical Monographs, Vol. 34.
[523] Y. I. Lyubarskiı̆ and V. A. Marchenko. Direct and inverse problems of multi-
channel scattering. Funktsional. Anal. i Prilozhen., 41(2):58–77, 112, 2007.
[524] E. MacCurdy. Notebooks of Leonardo da Vinci. Jonathtan Cape, London,
1938.
[525] W. Magnus and S. Winkler. Hill’s equation. Interscience Tracts in Pure and
Applied Mathematics, No. 20. Interscience Publishers John Wiley & Sons
New York-London-Sydney, 1966.
[526] B. Malgrange. Lectures on the theory of functions of several complex variables,
volume 13 of Tata Institute of Fundamental Research Lectures on Mathemat-
ics and Physics. Distributed for the Tata Institute of Fundamental Research,
Bombay, 1984. Reprint of the 1958 edition, Notes by Raghavan Narasimhan.
[527] V. A. Marchenko. Sturm-Liouville operators and applications. AMS Chelsea
Publishing, Providence, RI, revised edition, 2011.
[528] A. S. Markus. Introduction to the spectral theory of polynomial operator pen-
cils, volume 71 of Translations of Mathematical Monographs. American Math-
ematical Society, Providence, RI, 1988. Translated from the Russian by H.
H. McFaden. Translation edited by Ben Silver. With an appendix by M. V.
Keldysh.
256 BIBLIOGRAPHY

[529] J. E. Marsden and T. S. Ratiu. Introduction to mechanics and symmetry,

volume 17 of Texts in Applied Mathematics. Springer-Verlag, New York, 1994.
A basic exposition of classical mechanical systems.
[530] O. Martinyuk and V. Pivovarchik. Spectral problem for figure-of-eight graph
of Stieltjes strings. Methods Funct. Anal. Topology, 16(4):349–358, 2010.
[531] O. M. Martinyuk and V. M. Pivovarchik. The inverse spectral problem for a
figure-eight Stieltjes string. Ukr. Mat. Visn., 8(1):112–128, 156, 2011.
[532] T. Matsui. BEC of free bosons on networks. Infin. Dimens. Anal. Quantum
Probab. Relat. Top., 9(1):1–26, 2006.
[533] V. Maz ya, S. Nazarov, and B. Plamenevskij. Asymptotic theory of elliptic
boundary value problems in singularly perturbed domains. Vol. I, II, volume
111–112 of Operator Theory: Advances and Applications. Birkhäuser Verlag,
Basel, 2000. Translated from the German by Georg Heinig and Christian
Posthoff (Vol. I). Translated from the German by Plamenevskij (Vol. II).
(English translation of the 1991 German edition by Akademie Verlag).
[534] H. P. McKean and P. van Moerbeke. The spectrum of Hill’s equation. Invent.
Math., 30(3):217–274, 1975.
[535] J. R. McLaughlin. Good vibrations. American Scientist, 86(4):342–349, 1998.
[536] F. Mehmeti, J. v. Below, and S. Nicaise, editors. Partial Differential Equa-
tions on Multistructures. CRC, 2001.
[537] M. L. Mehta. Random matrices, volume 142 of Pure and Applied Mathematics
(Amsterdam). Elsevier/Academic Press, Amsterdam, third edition, 2004.
[538] J. Meier. Groups, graphs and trees, volume 73 of London Mathematical Society
Student Texts. Cambridge University Press, Cambridge, 2008. An introduc-
tion to the geometry of infinite groups.
[539] Y. Melnikov and B. Pavlov. Scattering on graphs and one-dimensional approx-
imations to N -dimensional schrödinger operators. J. Math. Phys., 42(3):1202–
1228, 2001.
[540] Y. B. Melnikov and B. S. Pavlov. Two-body scattering on a graph and ap-
plication to simple nanoelectronic devices. J. Math. Phys., 36(6):2813–2825,
1995.
[541] V. Z. Meshkov. On the possible rate of decrease at infinity of the solutions
of second-order partial differential equations. Mat. Sb., 182(3):364–383, 1991.
(In Russian. English transl. in Math. USSR-Sbornik.).
[542] A. B. Mikhailova, B. Pavlov, and V. I. Ryzhii. Dirichlet-to-Neumann tech-
niques for the plasma-waves in a slot-diode. In Operator theory, analysis and
mathematical physics, volume 174 of Oper. Theory Adv. Appl., pages 77–103.
Birkhäuser, Basel, 2007.
[543] G. W. Milton and P. Seppecher. Realizable response matrices of multi-
terminal electrical, acoustic and elastodynamic networks at a given frequency.
Proc. R. Soc. Lond. Ser. A Math. Phys. Eng. Sci., 464(2092):967–986, 2008.
[544] G. W. Milton and P. Seppecher. Electromagnetic circuits. Netw. Heterog.
Media, 5(2):335–360, 2010.
[545] G. W. Milton and P. Seppecher. Hybrid electromagnetic circuits. Physica B:
Cond. Matter, 405(14):2935–2937, 2010.
[546] K. A. Milton. The Casimir effect. World Scientific Publishing Co. Inc., River
Edge, NJ, 2001. Physical manifestations of zero-point energy.
 BIBLIOGRAPHY 257

[547] K. A. Milton. The Casimir effect: recent controversies and progress. J. Phys.
A, 37(38):R209–R277, 2004.
[548] R. Mittra and S. Lee. Analytical techniques in the theory of guided waves.
Macmillan, London, 1971.
[549] S. Mizohata. The theory of partial differential equations. Cambridge Univer-
sity Press, New York, 1973. Translated from the Japanese by Katsumi Miya-
hara.
[550] S. Molchanov and B. Vainberg. Slowdown of the wave packages in finite slabs
of periodic media. Waves Random Media, 14(3):411–423, 2004.
[551] S. Molchanov and B. Vainberg. Slowing down of the wave packets in quantum
graphs. Waves Random Complex Media, 15(1):101–112, 2005.
[552] S. Molchanov and B. Vainberg. Slowing down and reflection of waves in trun-
cated periodic media. J. Funct. Anal., 231(2):287–311, 2006.
[553] S. Molchanov and B. Vainberg. Transition from a network of thin fibers to the
quantum graph: an explicitly solvable model. In Quantum graphs and their
applications, volume 415 of Contemp. Math., pages 227–239. Amer. Math.
Soc., Providence, RI, 2006.
[554] S. Molchanov and B. Vainberg. Scattering solutions in networks of thin fibers:
small diameter asymptotics. Comm. Math. Phys., 273(2):533–559, 2007.
[555] S. Molchanov and B. Vainberg. Laplace operator in networks of thin fibers:
spectrum near the threshold. In Stochastic analysis in mathematical physics,
pages 69–93. World Sci. Publ., Hackensack, NJ, 2008.
[556] S. Molchanov and B. Vainberg. On general Cwikel-Lieb-Rozenblum and Lieb-
Thirring inequalities. In Around the research of Vladimir Maz’ya. III, vol-
ume 13 of Int. Math. Ser. (N. Y.), pages 201–246. Springer, New York, 2010.
[557] S. Molchanov and B. Vainberg. Wave propagation in periodic networks of
thin fibers. Waves Random Complex Media, 20(2):260–275, 2010.
[558] E. W. Montroll. Quantum theory on a network. I. A solvable model whose
wavefunctions are elementary functions. J. Math. Phys., 11(2):635–648, 1970.
[559] J. Moser and M. Struwe. On a Liouville-type theorem for linear and nonlinear
elliptic differential equations on a torus. Bol. Soc. Brasil. Mat. (N.S.), 23(1–
2):1–20, 1992.
[560] S. Müller, S. Heusler, P. Braun, F. Haake, and A. Altland. Semiclassical
foundation of universality in quantum chaos. Phys. Rev. Lett., 93:014103, Jul
2004.
[561] S. Müller, S. Heusler, P. Braun, F. Haake, and A. Altland. Periodic-orbit
theory of universality in quantum chaos. Phys. Rev. E, 72:046207, 2005.
[562] N. Nadirashvili, D. Tot, and D. Yakobson. Geometric properties of eigenfunc-
tions. Uspekhi Mat. Nauk, 56(6(342)):67–88, 2001.
[563] K. Naimark and M. Solomyak. Eigenvalue estimates for the weighted Lapla-
cian on metric trees. Proc. London Math. Soc. (3), 80(3):690–724, 2000.
[564] K. Naimark and M. Solomyak. Geometry of Sobolev spaces on regular trees
and the Hardy inequalities. Russ. J. Math. Phys., 8(3):322–335, 2001.
[565] M. A. Naimark. Linear differential operators. Part I: Elementary theory of
linear differential operators. Frederick Ungar Publishing Co., New York, 1967.
[566] V. Nekrashevych. Self-similar groups, volume 117 of Mathematical Surveys
and Monographs. American Mathematical Society, Providence, RI, 2005.
258 BIBLIOGRAPHY

[567] V. Nekrashevych and A. Teplyaev. Groups and analysis on fractals. In Anal-

ysis on graphs and its applications, volume 77 of Proc. Sympos. Pure Math.,
pages 143–180. Amer. Math. Soc., Providence, RI, 2008.
[568] S. Nicaise and O. Penkin. Relationship between the lower frequency spectrum
of plates and networks of beams. Math. Methods Appl. Sci., 23(16):1389–1399,
2000.
[569] S. Nicaise and O. Zaı̈r. Identifiability, stability and reconstruction results of
point sources by boundary measurements in heterogeneous trees. Rev. Mat.
Complut., 16(1):151–178, 2003.
[570] S. Novikov, S. V. Manakov, L. P. Pitaevskiı̆, and V. E. Zakharov. Theory of
solitons. Contemporary Soviet Mathematics. Consultants Bureau [Plenum],
New York, 1984. The inverse scattering method, Translated from the Russian.
[571] S. P. Novikov. Solitons and geometry. Lezioni Fermiane. [Fermi Lectures].
Published for the Scuola Normale Superiore, Pisa, 1994.
[572] S. P. Novikov. Schrodinger operators on graphs and symplectic geometry. In
The Arnoldfest (Toronto, ON, 1997), volume 24 of Fields Inst. Commun.,
pages 397–413. Amer. Math. Soc., Providence, RI, 1999.
[573] K. Novoselov. Nobel lecture: Graphene: Materials in the flatland. Rev.
Mod.Phys., 83:837–849, 2011.
[574] M. Nowaczyk. Inverse spectral problem for quantum graphs. Master’s thesis,
Lund Institute of Technology, 2005. Licenciate Thesis.
[575] M. Nowaczyk. Inverse spectral problem for quantum graphs with rationally
dependent edges. In Operator theory, analysis and mathematical physics, vol-
ume 174 of Oper. Theory Adv. Appl., pages 105–116. Birkhäuser, Basel, 2007.
[576] P. Ola and L. Päivärinta. Mellin operators and pseudodifferential operators
on graphs. Waves Random Media, 14(1):S129–S142, 2004. Special section on
quantum graphs.
[577] V. Oleinik and B. Pavlov. Embedded spectrum on a metric graph (an obser-
vation). Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI),
300(Teor. Predst. Din. Sist. Spets. Vyp. 8):215–220, 291, 2003.
[578] V. L. Oleinik, B. S. Pavlov, and N. V. Sibirev. Analysis of the dispersion equa-
tion for the Schrödinger operator on periodic metric graphs. Waves Random
Media, 14(2):157–183, 2004.
[579] B.-S. Ong. Spectral Problems of Optical Waveguides and Quantum Graphs.
PhD thesis, Texas A&M University, 2006.
[580] S. F. Oppenheimer. A convection-diffusion problem in a network. Appl. Math.
Comp., 112:223–240, 2000.
[581] I. Oren. Nodal domain counts and the chromatic number of graphs. J. Phys.
A, 40(32):9825–9832, 2007.
[582] I. Oren and R. Band. Isospectral graphs with identical nodal counts. J. Phys.
A, 45:135203, 2012. preprint arXiv:1110.0158.
[583] I. Oren, A. Godel, and U. Smilansky. Trace formulae and spectral statistics
for discrete Laplacians on regular graphs. I. J. Phys. A, 42(41):415101, 20,
2009.
[584] I. Oren and U. Smilansky. Trace formulas and spectral statistics for discrete
Laplacians on regular graphs (II). J. Phys. A, 43(22):225205, 13, 2010.
[585] P. Pakonski, G. Tanner, and K. Zyczkowski. Families of line-graphs and their
quantization. J. Stat. Phys., 111(5-6):1331–1352, 2003.
 BIBLIOGRAPHY 259

[586] P. Pakoński, K. Życzkowski, and M. Kuś. Classical 1D maps, quantum graphs

and ensembles of unitary matrices. J. Phys. A, 34(43):9303–9317, 2001.
[587] K. Pankrashkin. Localization effects in a periodic quantum graph with mag-
netic field and spin-orbit interaction. J. Math. Phys., 47(11):112105, 17, 2006.
[588] K. Pankrashkin. Spectra of Schrödinger operators on equilateral quantum
graphs. Lett. Math. Phys., 77(2):139–154, 2006.
[589] K. Pankrashkin. Localization in a quasiperiodic model on quantum graphs.
In Analysis on graphs and its applications, volume 77 of Proc. Sympos. Pure
Math., pages 459–467. Amer. Math. Soc., Providence, RI, 2008.
[590] L. Parnovski. Bethe-Sommerfeld conjecture. Ann. Henri Poincaré, 9(3):457–
508, 2008.
[591] L. Parnovski and A. V. Sobolev. Bethe-Sommerfeld conjecture for periodic
operators with strong perturbations. Invent. Math., 181(3):467–540, 2010.
[592] O. Parzanchevski and R. Band. Linear representations and isospectrality with
boundary conditions. J. Geom. Anal., 20(2):439–471, 2010.
[593] L. Pastur and A. Figotin. Spectra of random and almost-periodic operators,
volume 297 of Grundlehren der Mathematischen Wissenschaften [Fundamen-
tal Principles of Mathematical Sciences]. Springer-Verlag, Berlin, 1992.
[594] L. Pauling. The diamagnetic anisotropy of aromatic molecules. J. Chem.
Phys., 4(10):673–677, 1936.
[595] B. Pavlov. Resonance quantum switch: matching domains. In Surveys in
analysis and operator theory (Canberra, 2001), volume 40 of Proc. Centre
Math. Appl. Austral. Nat. Univ., pages 127–156. Austral. Nat. Univ., Can-
berra, 2002.
[596] B. Pavlov. A star-graph model via operator extension. Math. Proc. Cambridge
Philos. Soc., 142(2):365–384, 2007.
[597] B. Pavlov and K. Robert. Resonance optical switch: calculation of resonance
eigenvalues. In Waves in periodic and random media (South Hadley, MA,
2002), volume 339 of Contemp. Math., pages 141–169. Amer. Math. Soc.,
Providence, RI, 2003.
[598] B. S. Pavlov. A model of zero-radius potential with internal structure. Teoret.
Mat. Fiz., 59(3):345–353, 1984.
[599] B. S. Pavlov. The theory of extensions, and explicitly solvable models. Uspekhi
Mat. Nauk, 42(6(258)):99–131, 247, 1987.
[600] J. Peetre. On Hadamard’s variational formula. J. Differential Equations,
36(3):335–346, 1980.
[601] R. Peirone. Existence of eigenforms on nicely separated fractals. In Analysis
on graphs and its applications, volume 77 of Proc. Sympos. Pure Math., pages
231–241. Amer. Math. Soc., Providence, RI, 2008.
[602] O. M. Penkin and Y. V. Pokornyı̆. On a boundary value problem on a graph.
Differentsial nye Uravneniya, 24(4):701–703, 734–735, 1988.
[603] Y. Pinchover, G. Wolansky, and D. Zelig. Spectral properties of Schrödinger
operators on radial N -dimensional infinite trees. Israel J. Math., 165:281–328,
2008.
[604] L. Pitaevskii and S. Stringari. Bose-Einstein condensation, volume 116 of
International Series of Monographs on Physics. The Clarendon Press Oxford
University Press, Oxford, 2003.
260 BIBLIOGRAPHY

[605] M. Pivarski and L. Saloff-Coste. Small time heat kernel behavior on Riemann-
ian complexes. New York J. Math., 14:459–494, 2008.
[606] V. Pivovarchik. Inverse problem for the Sturm-Liouville equation on a simple
graph. SIAM J. Math. Anal., 32(4):801–819 (electronic), 2000.
[607] V. Pivovarchik. Inverse problem for the Sturm-Liouville equation on a star-
shaped graph. Math. Nachr., 280(13-14):1595–1619, 2007.
[608] V. Pivovarchik and H. Woracek. Sums of Nevanlinna functions and differential
equations on star-shaped graphs. Oper. Matrices, 3(4):451–501, 2009.
[609] V. Pivovarchik and H. Woracek. Eigenvalue asymptotics for a star-graph
damped vibrations problem. Asymptot. Anal., 73(3):169–185, 2011.
[610] V. N. Pivovarchik. Ambartsumyan’s theorem for the Sturm-Liouville bound-
ary value problem on a star-shaped graph. Funktsional. Anal. i Prilozhen.,
39(2):78–81, 2005.
[611] J. Platt, K. Ruedenberg, C. W. Scherr, N. S. Ham, H. Labhart, and
W. Lichten. Free-Electron Theory of Conjugated Molecules: Papers of The
Chicago Group 1949-1961. Wiley, 1964.
[612] Å. Pleijel. Remarks on Courant’s nodal line theorem. Comm. Pure Appl.
Math., 9:543–550, 1956.
[613] A. Pliś. Non-uniqueness in Cauchy’s problem for differential equations of el-
liptic type. J. Math. Mech., 9:557–562, 1960.
[614] Y. V. Pokornyı̆, O. M. Penkin, V. L. Pryadiev, A. V. Borovskikh, K. P.
Lazarev, and S. A. Shabrov. Differentsialnye uravneniya na geometricheskikh
grafakh. Fiziko-Matematicheskaya Literatura, Moscow, 2005.
[615] Y. V. Pokornyı̆ and V. L. Pryadiev. Some problems in the qualitative Sturm-
Liouville theory on a spatial network. Uspekhi Mat. Nauk, 59(3(357)):115–150,
2004.
[616] Y. V. Pokornyı̆, V. L. Pryadiev, and A. Al -Obeı̆d. On the oscillation of the
spectrum of a boundary value problem on a graph. Mat. Zametki, 60(3):468–
470, 1996.
[617] J. Pöschel and E. Trubowitz. Inverse spectral theory, volume 130 of Pure and
Applied Mathematics. Academic Press Inc., Boston, MA, 1987.
[618] O. Post. Periodic manifolds with spectral gaps. J. Differential Equations,
187(1):23–45, 2003.
[619] O. Post. Spectral convergence of quasi-one-dimensional spaces. Ann. Henri
Poincaré, 7(5):933–973, 2006.
[620] O. Post. Equilateral quantum graphs and boundary triples. In Analysis on
graphs and its applications, volume 77 of Proc. Sympos. Pure Math., pages
469–490. Amer. Math. Soc., Providence, RI, 2008.
[621] O. Post. First order approach and index theorems for discrete and metric
graphs. Ann. Henri Poincaré, 10(5):823–866, 2009.
[622] O. Post. Spectral Analysis on Graph-like Spaces, volume 2039 of Lecture Notes
in Mathematics. Springer Verlag, Berlin, 2012.
[623] O. U. Press. A Dictionary of Scientists. Oxford Univ. Press, Providence, RI,
1999.
[624] V. Rabinovich, S. Roch, and B. Silbermann. Limit operators and their ap-
plications in operator theory, volume 150 of Operator Theory: Advances and
Applications. Birkhäuser Verlag, Basel, 2004.
 BIBLIOGRAPHY 261

[625] V. S. Rabinovich and S. Roch. Pseudodifferential operators on periodic

graphs. Integral Equations Operator Theory, 72(2):197–217, 2012.
[626] M. Reed and B. Simon. Methods of modern mathematical physics. I–4. Func-
tional analysis. Academic Press, New York, 1972.
[627] O. Reingold, S. Vadhan, and A. Wigderson. Entropy waves, the zig-zag graph
product, and new constant-degree expanders. Ann. of Math. (2), 155(1):157–
187, 2002.
[628] F. S. Rofe-Beketov. A finiteness test for the number of discrete levels which
can be introduced into the gaps of the continuous spectrum by perturbations
of a periodic potential. Dokl. Akad. Nauk SSSR, 156:515–518, 1964.
[629] F. S. Rofe-Beketov. A perturbation of a Hill’s operator, that has a first mo-
ment and a non-zero integral, introduces a single discrete level into each of
the distant spectral lacunae. In Mathematical physics and functional analy-
sis, No. 4 (Russian), pages 158–159, 163. Fiz.-Tehn. Inst. Nizkih Temperatur,
Akad. Nauk Ukrain. SSR, Kharkov, 1973.
[630] J.-P. Roth. Spectre du laplacien sur un graphe. C. R. Acad. Sci. Paris Sér. I
Math., 296(19):793–795, 1983.
[631] J.-P. Roth. Le spectre du laplacien sur un graphe. In Théorie du potentiel (Or-
say, 1983), volume 1096 of Lecture Notes in Math., pages 521–539. Springer,
Berlin, 1984.
[632] J. Rubinstein and M. Schatzman. Spectral and variational problems on mul-
ticonnected strips. C. R. Acad. Sci. Paris Sér. I Math., 325(4):377–382, 1997.
[633] J. Rubinstein and M. Schatzman. Asymptotics for thin superconducting rings.
J. Math. Pures Appl. (9), 77(8):801–820, 1998.
[634] J. Rubinstein and M. Schatzman. Variational problems on multiply connected
thin strips. I. Basic estimates and convergence of the Laplacian spectrum.
Arch. Ration. Mech. Anal., 160(4):271–308, 2001.
[635] R. Rueckriemen. Recovering quantum graphs from their Bloch spectrum.
preprint arXiv:1101.6002, 2011.
[636] R. Rueckriemen and D. C. D. of Mathematics. Quantum Graphs and Their
Spectra, PhD Thesis. Dartmouth College, 2011.
[637] K. Ruedenberg and C. W. Scherr. Free-electron network model for conjugated
systems. i. J. Chem. Phys., 21(9):1565–1581, 1953.
[638] R. Saito, G. Dresselhaus, and M. S. Dresselhaus. Physical Properties of Car-
bon Nanotubes. Imperial College Press, London, 1998.
[639] Y. Saito. The limiting equation for Neumann Laplacians on shrinking do-
mains. Electron. J. Differential Equations, pages No. 31, 25 pp. (electronic),
2000.
[640] Y. Saitō. Convergence of the Neumann Laplacian on shrinking domains. Anal-
ysis (Munich), 21(2):171–204, 2001.
[641] K. Sakoda. Optical Properties of Photonic Crystals. Springer Verlag, New
York, 2001.
[642] H. Schanz and U. Smilansky. Spectral statistics for quantum graphs: periodic
orbits and combinatorics. Phil. Mag. B, 80(12):1999–2021, 2000.
[643] H. Schanz and U. Smilansky. Combinatorial identities from the spectral the-
ory of quantum graphs. Electron. J. Combin., 8(2):Research Paper 16, 16 pp.
(electronic), 2001. In honor of Aviezri Fraenkel on the occasion of his 70th
birthday.
262 BIBLIOGRAPHY

[644] P. Schapotschnikow. Eigenvalue and nodal properties on quantum graph trees.

Waves Random Complex Media, 16(3):167–178, 2006.
[645] P. Schapotschnikow and S. Gnutzmann. Spectra of graphs and semi-
conducting polymers. In Analysis on graphs and its applications, volume 77
of Proc. Sympos. Pure Math., pages 691–705. Amer. Math. Soc., Providence,
RI, 2008.
[646] M. Schatzman. On the eigenvalues of the Laplace operator on a thin set with
Neumann boundary conditions. Appl. Anal., 61(3–4):293–306, 1996.
[647] J. H. Schenker and M. Aizenman. The creation of spectral gaps by graph
decoration. Lett. Math. Phys., 53(3):253–262, 2000.
[648] M. Schiffer. Hadamard’s formula and variation of domain-functions. Amer.
J. Math., 68:417–448, 1946.
[649] R. L. Schult, D. G. Ravenhall, and H. W. Wyld. Quantum bound states in a
classically unbound system of crossed wires. Phys. Rev. B, 39(8):5476–5479,
1989.
[650] W. Shaban and B. Vainberg. Radiation conditions for the difference
Schrödinger operators. Appl. Anal., 80(3-4):525–556, 2001.
[651] V. M. Shalaev. Nonlinear Optics of Random Media: Fractal Composites and
Metal-Dielectric Films, volume 158 of Springer Tracts in Modern Physics.
Springer-Verlag, 1999.
[652] T. J. Shepherd, P. J. Roberts, and R. Loudon. Soluble two-dimensional
photonic-crystal model. Phys. Rev. E, 55(5):6024–6038, 1997.
[653] È. È. Shnol’. On the behavior of the eigenfunctions of Schrödinger’s equation.
Mat. Sb. (N.S.), 42 (84)(3):273–286, 1957.
[654] M. Shubin. Spectral theory of the Schrödinger operators on non-compact
manifolds: qualitative results. In Spectral theory and geometry (Edinburgh,
1998), volume 273 of London Math. Soc. Lecture Note Ser., pages 226–283.
Cambridge Univ. Press, Cambridge, 1999.
[655] M. Shubin and T. Sunada. Geometric theory of lattice vibrations and specific
heat. Pure Appl. Math. Q., 2(3, part 1):745–777, 2006.
[656] M. A. Shubin. Pseudodifference operators and their Green function. Izv. Akad.
Nauk SSSR Ser. Mat., 49(3):652–671, 1985.
[657] M. A. Shubin. Pseudodifferential operators and spectral theory. Springer Series
in Soviet Mathematics. Springer-Verlag, Berlin, 1987. Translated from the
Russian by Stig I. Andersson.
[658] M. A. Shubin. Spectral theory of elliptic operators on noncompact manifolds.
Astérisque, (207):5, 35–108, 1992. Méthodes semi-classiques, Vol. 1 (Nantes,
1991).
[659] M. Sieber. Leading off-diagonal approximation for the spectral form factor
for uniformly hyperbolic systems. J. Phys. A, 35(42):L613–L619, 2002.
[660] M. Sieber and K. Richter. Correlations between periodic orbits and their role
in spectral statistics. Phys. Scr., T90:128–133, 2001. 116th Symposium of the
Nobel on Quantum Chaos Y2K, Sweden, JUN 13-17, 2000.
[661] B. Simon. On the genericity of nonvanishing instability intervals in Hills equa-
tion. Ann. Inst. Henri Poincaré, XXIV(1):91–93, 1976.
[662] B. Simon. A new approach to inverse spectral theory. I. Fundamental formal-
ism. Ann. of Math. (2), 150(3):1029–1057, 1999.
 BIBLIOGRAPHY 263

[663] B. Simon. Trace ideals and their applications, volume 120 of Mathematical
Surveys and Monographs. American Mathematical Society, Providence, RI,
second edition, 2005.
[664] A. Singer, Z. Zhao, Y. Shkolnisky, and R. Hadani. Viewing angle classification
of cryo-electron microscopy images using eigenvectors. SIAM J. Imaging Sci.,
4(2):723–759, 2011.
[665] M. M. Skriganov. Geometric and arithmetic methods in the spectral theory
of multidimensional periodic operators. Trudy Mat. Inst. Steklov., 171:122,
1985.
[666] M. M. Skriganov. The spectrum band structure of the three-dimensional
Schrödinger operator with periodic potential. Invent. Math., 80(1):107–121,
1985.
[667] U. Smilansky. Irreversible quantum graphs. Waves Random Media,
14(1):S143–S153, 2004. Special section on quantum graphs.
[668] U. Smilansky. Quantum chaos on discrete graphs. J. Phys. A, 40(27):F621–
F630, 2007.
[669] U. Smilansky. Exterior-interior duality for discrete graphs. J. Phys. A,
42(3):035101, 13, 2009.
[670] U. Smilansky and M. Solomyak. The quantum graph as a limit of a network
of physical wires. In Quantum graphs and their applications, volume 415 of
Contemp. Math., pages 283–291. Amer. Math. Soc., Providence, RI, 2006.
[671] Z. Sobirov, D. Matrasulov, K. Sabirov, S. Sawada, and K. Nakamura. Inte-
grable nonlinear schrödinger equation on simple networks: Connection for-
mula at vertices. Phys. Rev. E, 81:066602, Jun 2010.
[672] A. V. Sobolev. Absolute continuity of the periodic magnetic Schrödinger op-
erator. Invent. Math., 137(1):85–112, 1999.
[673] A. V. Sobolev and M. Solomyak. Schrödinger operators on homogeneous met-
ric trees: spectrum in gaps. Rev. Math. Phys., 14(5):421–467, 2002.
[674] M. Solomyak. On the spectrum of the Laplacian on regular metric trees.
Waves Random Media, 14(1):S155–S171, 2004. Special section on quantum
graphs.
[675] A. Stern. Bemerkungen ber asymptotisches Verhalten von Eigenwerten und
Eigenfunktionen. PhD thesis, Göttingen University, 1925.
[676] P. Stollmann. Caught by disorder, volume 20 of Progress in Mathematical
Physics. Birkhäuser Boston Inc., Boston, MA, 2001. Bound states in random
media.
[677] R. S. Strichartz. Differential equations on fractals. Princeton University Press,
Princeton, NJ, 2006. A tutorial.
[678] C. Sturm. Mémoire sur les équations différentielles linéaires du second ordre.
J. Math. Pures Appl., 1:106–186, 1836.
[679] C. Sturm. Mémoire sur une classe d’équations à différences partielles. J. Math.
Pures Appl., 1:373–444, 1836.
[680] T. Sunada. Riemannian coverings and isospectral manifolds. Ann. of Math.
(2), 121(1):169–186, 1985.
[681] T. Sunada. A periodic Schrödinger operator on an abelian cover. J. Fac. Sci.
Univ. Tokyo Sect. IA Math., 37(3):575–583, 1990.
264 BIBLIOGRAPHY

[682] T. Sunada. Discrete geometric analysis. In Analysis on graphs and its appli-
cations, volume 77 of Proc. Sympos. Pure Math., pages 51–83. Amer. Math.
Soc., Providence, RI, 2008.
[683] G. Tanner. Spectral statistics for unitary transfer matrices of binary graphs.
J. Phys. A, 33(18):3567–3585, 2000.
[684] G. Tanner. Unitary-stochastic matrix ensembles and spectral statistics. J.
Phys. A, 34(41):8485–8500, 2001.
[685] A. Terras. Zeta functions of graphs, volume 128 of Cambridge Studies in Ad-
vanced Mathematics. Cambridge University Press, Cambridge, 2011. A stroll
through the garden.
[686] A. A. Terras and H. M. Stark. Zeta functions of finite graphs and coverings.
III. Adv. Math., 208(1):467–489, 2007.
[687] C. Texier and G. Montambaux. Scattering theory on graphs. J. Phys. A,
34(47):10307–10326, 2001.
[688] L. E. Thomas. Time dependent approach to scattering from impurities in a
crystal. Comm. Math. Phys., 33:335–343, 1973.
[689] T. J. Thornton. Mesoscopic devices. Rep. Prog. Phys., 58:311–364, 1995.
[690] E. C. Titchmarsh. Eigenfunction expansions associated with second-order dif-
ferential equations. Vol. 2. Oxford, at the Clarendon Press, 1958.
[691] K. Uhlenbeck. Generic properties of eigenfunctions. Amer. J. Math.,
98(4):1059–1078, 1976.
[692] G. Uhlmann. Inverse boundary value problems and applications. Astérisque,
207:153–211, 1996.
[693] G. Uhlmann, editor. Inside out: inverse problems and applications, volume 47
of Mathematical Sciences Research Institute Publications. Cambridge Univer-
sity Press, Cambridge, 2003.
[694] G. Uhlmann. Visibility and invisibility. In ICIAM 07—6th International Con-
gress on Industrial and Applied Mathematics, pages 381–408. Eur. Math. Soc.,
Zürich, 2009.
[695] G. Uhlmann, editor. Inside out: inverse problems and applications, volume 2
of Mathematical Sciences Research Institute Publications. Cambridge Univer-
sity Press, Cambridge, 2012.
[696] I. Veselić. Spectral analysis of percolation Hamiltonians. Math. Ann.,
331(4):841–865, 2005.
[697] F. Visco-Comandini, M. Mirrahimi, and M. Sorine. Some inverse scattering
problems on star-shaped graphs. J. Math. Anal. Appl., 378(1):343–358, 2011.
[698] K. Vogtmann. Automorphisms of free groups and outer space. In Proceedings
of the Conference on Geometric and Combinatorial Group Theory, Part I
(Haifa, 2000), volume 94, pages 1–31, 2002.
[699] K. Vogtmann. What is. . . outer space? Notices Amer. Math. Soc., 55(7):784–
786, 2008.
[700] J. von Neumann. Allgemeine Eigenwerttheorie Hermitescher Funktionaloper-
atoren. Math. Ann., 102:49–131, 1929–1930.
[701] B. Winn. A conditionally convergent trace formula for quantum graphs. In
Analysis on graphs and its applications, volume 77 of Proc. Sympos. Pure
Math., pages 491–501. Amer. Math. Soc., Providence, RI, 2008.
[702] E. Yablonovitch. Inhibited spontaneous emission in solid-state physics and
electronics. Phys. Rev. Lett., 58(20):2059–2062, 1987.
 BIBLIOGRAPHY 265

[703] D. R. Yafaev. Mathematical scattering theory, volume 158 of Mathematical

Surveys and Monographs. American Mathematical Society, Providence, RI,
2010. Analytic theory.
[704] V. A. Yakubovich and V. M. Starzhinskii. Linear differential equations with
periodic coefficients. 1, 2. Halsted Press [John Wiley & Sons] New York-
Toronto, Ont., 1975. Translated from Russian by D. Louvish.
[705] E. Yanagida. Existence of stable stationary solutions of scalar reaction-
diffusion equations in thin tubular domains. Appl. Anal., 36:171–188, 1990.
[706] E. Yanagida. Stability of nonconstant steady states in reaction-diffusion sys-
tems on graphs. Jap. J. Ind. Appl. Math., 18(1):25–42, 2001.
[707] C. F. Yang, V. N. Pivovarchik, and Z. Y. Huang. Ambarzumyan-type theo-
rems on star graphs. Oper. Matrices, 5(1):119–131, 2011.
[708] A. Yariv, Y. Xu, R. K. Lee, and A. Scherer. Coupled-resonator optical wave-
guide: a proposal and analysis. Opt. Lett., 24:711–713, 1999.
[709] S. T. Yau. Harmonic functions on complete Riemannian manifolds. Comm.
Pure Appl. Math., 28:201–228, 1975.
[710] K. Yoshitomi. Band gap of the spectrum in periodically curved quantum
waveguides. J. Differential Equations, 142(1):123–166, 1998.
[711] V. Yurko. Inverse spectral problems for Sturm-Liouville operators on graphs.
Inverse Problems, 21(3):1075–1086, 2005.
[712] V. Yurko. Inverse problems for Sturm-Liouville operators on bush-type
graphs. Inverse Problems, 25(10):105008, 14, 2009.
[713] V. Yurko. Inverse spectral problems for differential operators on a graph with
a rooted cycle. Tamkang J. Math., 40(3):271–286, 2009.
[714] V. Yurko. Uniqueness of recovering differential operators on hedgehog-type
graphs. Adv. Dyn. Syst. Appl., 4(2):231–241, 2009.
[715] V. Yurko. Uniqueness of recovering Sturm-Liouville operators on A-graphs
from spectra. Results Math., 55(1-2):199–207, 2009.
[716] V. Yurko. An inverse problem for Sturm-Liouville differential operators on
A-graphs. Appl. Math. Lett., 23(8):875–879, 2010.
[717] V. Yurko. A remark on inverse problems for Sturm-Liouville operators on
graphs. Results Math., 58(3-4):399–401, 2010.
[718] V. Yurko. Inverse problems for Bessel-type differential equations on noncom-
pact graphs using spectral data. Inverse Problems, 27(4):045002, 17, 2011.
[719] V. A. Yurko. An inverse problem for higher order differential operators on
star-type graphs. Inverse Problems, 23(3):893–903, 2007.
[720] V. A. Yurko. The inverse spectral problem for differential operator pencils on
noncompact spatial networks. Differ. Uravn., 44(12):1658–1666, 2008.
[721] V. A. Yurko. On the reconstruction of higher-order differential operators on
compact graphs. Dokl. Akad. Nauk, 419(5):604–608, 2008.
[722] V. A. Yurko. Recovering higher-order differential operators on star-type
graphs from spectra. Cubo, 10(1):93–102, 2008.
[723] V. A. Yurko. Inverse nodal problems for the Sturm-Liouville differential op-
erators on a star-type graph. Sibirsk. Mat. Zh., 50(2):469–475, 2009.
[724] V. A. Yurko. Inverse spectral problems for Sturm-Liouville differential oper-
ators on a finite interval. J. Inverse Ill-Posed Probl., 17(7):639–694, 2009.
[725] V. A. Yurko. On an inverse spectral problem for differential operators on a
hedgehog-type graph. Dokl. Akad. Nauk, 425(4):466–470, 2009.
266 BIBLIOGRAPHY

[726] V. A. Yurko. Reconstruction of Sturm-Liouville operators from the spectra

on a graph with a cycle. Mat. Sb., 200(9):147–160, 2009.
[727] V. A. Yurko. Inverse spectral problems for differential operators on arbitrary
compact graphs. J. Inverse Ill-Posed Probl., 18(3):245–261, 2010.
[728] V. A. Yurko. Reconstruction of Sturm-Liouville differential operators on A-
graphs. Differ. Uravn., 47(1):50–59, 2011.
[729] M. Zaidenberg. Discrete convolution operators in positive characteristic: a
variation on the Floquet-Bloch theory. In Complex analysis and dynamical
systems IV. Part 2, volume 554 of Contemp. Math., pages 265–284. Amer.
Math. Soc., Providence, RI, 2011.
[730] M. G. Zaı̆denberg, S. G. Kreı̆n, P. A. Kučment, and A. A. Pankov. Banach
bundles and linear operators. Uspehi Mat. Nauk, 30(5(185)):101–157, 1975.
English translation: Russian Math. Surveys 30 (1975), no. 5, 115–175.
[731] M. G. Zaı̆denberg and V. Y. Lin. Finiteness theorems for holomorphic map-
pings. In Current problems in mathematics. Fundamental directions, Vol. 9
(Russian), Itogi Nauki i Tekhniki, pages 127–193, 272. Akad. Nauk SSSR
Vsesoyuz. Inst. Nauchn. i Tekhn. Inform., Moscow, 1986.
[732] D. Zelig. Properties of Solutions of Partial Differential Equations Defined on
Human Lung-shaped Domains. PhD thesis, Technion, Israel, 2005.
[733] H. Zeng. Convergence of Spectra of Mesoscopic Systems Collapsing Onto a
Graph. PhD thesis, Wichita State University, Wichita, KS, 2001.
[734] J. M. Ziman. Principles of the theory of solids. Cambridge University Press,
New York, 1964.
[735] J. M. Ziman. Models of disorder. Cambridge University Press, Cambridge,
1979. The theoretical physics of homogeneously disordered systems.
 Index

1D complex, 3, 4 automorphic, 110

B, 3
E, 2 band-gap structure, 116
N (λ), 67 Bethe lattices, 34, 207
V, 2 Betti number, 4, 211
W , 105, 107 Bloch variety, 116
Φ(E, F ), Fredholm operators, 214 bond, 2, 210
Tn , 109 bond reversal, 3
βΓ , 211 bound state, 80
D(Γ), 111 boundary control method, 184
δ-type condition, 24 boundary form, 20
fˆ, 110 boundary vertices, 150
μ(f ), 135 Brillouin zone, 109
ν(f ), 135
π-electrons, 192 carbon nanotube, 201
ρ(A), 219 Casimir effect, 205
σ-electrons, 192 Casimir-Polder force, 205
σd (A), 220 Cayley graph, 28, 35, 212
σac (A), 221 character, 107
σpp (A), 221 chiral, 204
σr (A), 220 co-boundary operator, 7
σsc (A), 221 co-spectral graphs, 183
σ(A), 219 co-spectrum, 216
σess (A), 220 combinatorial graph, 3
H k (Γ), 11 condition
du , 2 δ-type, 24
div , dov , 3 continuity, 14
1st Betti number, 4, 211 current conservation, 14
decoupling, 26
adjacency matrix, 2, 209 Dirichlet, 25
adjacent vertices, 2 Kirchhoff, 14
algebraic counting function, 93 Neumann, 14
analysis on fractals, 188 Neumann–Kirchhoff, 14
analytic Fredholm theorem, 67 quasiperiodic, 60, 114
analytic set, 216 standard, 14
aperiodic graph, 172 strongly coupling, 172
arcs, 175 continuous functional calculus, 220
267
268 INDEX

critical point, 147 forest, 211

non-degenerate, 147 form
CROWs, 208 closed, 219
current conservation condition, 14 Hermitian, 219
cut-points, 60 Hermitian symplectic, 223
cycle, 211 non-degenerate, 223
cyclomatic number, 4, 211 quadratic, 219
quadratic, of a graph, 22
decoupling condition, 26 sesqui-linear, 219
degeneracy class, 168 form factor, 161
degree matrix, 2, 209 Fredholm operator, 52, 213
degree of a vertex fullerene buckyball, 201
outgoing, 210 function of a discrete graph, 5
degree of a vertex, 2, 209 fundamental domain, 105, 107
incoming, 210
density of states, 100, 160, 221 Gaussian Orthogonal Ensemble
determinant, 149 (GOE), 162
diagrams, 179 Gaussian Unitary Ensemble, 163
digraph, 2, 210 graph, 2, 209
Dirac operators, 187 bipartite, 210
Dirac point, 203 compact, 9
directed adjacency matrix, 210 complete, 210
directed graph, 2, 210 finite, 9
Dirichlet-to-Neumann map, 85, 87, infinite, 9
198, 201 periodic, 105, 211
Discrete Fourier Transform (DFT) regular, 210
matrix, 50 star, 210
discrete graph, 3 tree, 211
dispersion curve, 116 graphene, 201
dispersion relation, 68, 116 graphyne, 201
Grassmannian, 224
eigenfunction, 220
eigenvalue, 220 Hadamard’s variational formulas, 68
electric potential, 12 Hamiltonian, 12, 29
electromagnetic circuits, 188 Hermitian symplectic space, 223
ellipticity, 53 Hill discriminant, 201
encounter, 175
equilateral graph, 8, 90 improper partition vertex, 143
equipartition, 144 impurity, 124
equitransmitting matrices, 50 impurity spectrum, 124
extension theory, 223 incidence matrix, 6
external edges, 150 incoming amplitude, 150
incoming bond, 3, 210
fattened graph, 189 incoming degree, 3
finite ball condition, 9 index of a Fredholm operator, 52,
Floquet multiplier, 109 214
Floquet theory, 105 Internet (or network) tomography,
Floquet transform, 109 182
flux, 61 inverse problem, 181
 INDEX 269

invertible regularizer, 214 nodal, 143

invisibility cloaking, 195 proper, 143
isospectral graphs, 183 partition vertex, 143
isotropic subspace, 224 path, 99, 211
closed, 99, 211
Lagrangian subspace, 224 primitive, 99
Laplacian repetition, 99
combinatorial, 5 perfect scar, 85
discrete, 5 periodic orbit, 99
harmonic, 5 photonic crystal, 195
normalized, 6 positive (negative) domain, 135
weighted, 7 potential, 203
Laurent polynomials, 112 principal analytic, 68
lead, 8, 150 principal analytic set, 216
leaky wire, 195 proper partition vertex, 143
left regularizer, 215
left semi-Fredholm, 215 quadratic form, 5
level repulsion, 159 quantum graph, 3, 13
Liouville type theorem, 127 quantum waveguide, 193
local finiteness, 2, 209 quantum wire, 193
local vertex conditions, 16 quasi-momentum, 108
Lyapunov function, 201 quasiperiodic conditions, 61
magnetic potential, 12 Rashba Hamiltonian, 187
Markov operator, 6 regularized determinant, 101
metric graph, 3, 7 regularizer, 214
metric length, 100 repetition number, 99
Morse index, 147 repetition of a closed path, 99
multistructure, 188 resolvent, 219
resolvent set, 219
naphthalene, 192
right regularizer, 215
necklace structure, 204
right semi-Fredholm, 215
nodal count, 135
nodal deficiency, 145
scale invariance, 22
nodal domain, 134
scar, 85
number of zeros, 135
scattering matrix, 38, 150
Oka’s principle, 215 Schatten-von Neumann operator
operator pencils, 49 ideal, 101
origin vertex, 2, 210 Schreier graph, 28, 35, 212
outgoing, 3 SCISSORs, 208
outgoing amplitudes, 150 semi-Fredholm operators, 215
outgoing bond, 3, 210 simple cycle, 211
outgoing degree, 3 simplicial complex, xi, 8
outgoing directions, 14 simplicity of the spectrum
generic, 71
partition, 143 of a tree, 74
bipartite, 144 skew orthogonal, 224
equipartition, 144 Sobolev space H 1 (Γ), 10
improper, 143 space L2 (Γ), 10
270 INDEX

spectral band, 115

spectral counting function, 67
spectral determinant, 148
spectral gap, 115
spectral measure, 221
spectral theorem, 220
spectrum, 219
absolutely continuous, 221
discrete, 220
essential, 220
point, 220
pure point, 221
residual, 220
singular continuous, 221
spectrum of the operator function,
216
stability amplitude, 100
Stein domain, 55
Stein manifold, 215, 216
strongly coupling vertex conditions,
182
symplectic geometry, 223
system theory, 201

terminal vertex, 2, 210

threshold, 194
time-reversal invariance, 170
topological length, 99
trace formula, 92
tree, 4, 211
two-point correlation function, 160

unique continuation property, 84

unitary character, 107

vertex conditions, 13
vertex Neumann conditions, 27

weak unique continuation property,

84
Selected Published Titles in This Series
186 Gregory Berkolaiko and Peter Kuchment, Introduction to Quantum Graphs, 2013
185 Patrick Iglesias-Zemmour, Diffeology, 2012
184 Frederick W. Gehring and Kari Hag, The Ubiquitous Quasidisk, 2012
183 Gershon Kresin and Vladimir Maz’ya, Maximum Principles and Sharp Constants for
Solutions of Elliptic and Parabolic Systems, 2012
182 Neil A. Watson, Introduction to Heat Potential Theory, 2012
181 Graham J. Leuschke and Roger Wiegand, Cohen-Macaulay Representations, 2012
180 Martin W. Liebeck and Gary M. Seitz, Unipotent and Nilpotent Classes in Simple
Algebraic Groups and Lie Algebras, 2012
179 Stephen D. Smith, Subgroup Complexes, 2011
178 Helmut Brass and Knut Petras, Quadrature Theory, 2011
177 Alexei Myasnikov, Vladimir Shpilrain, and Alexander Ushakov,
Non-commutative Cryptography and Complexity of Group-theoretic Problems, 2011
176 Peter E. Kloeden and Martin Rasmussen, Nonautonomous Dynamical Systems, 2011
175 Warwick de Launey and Dane Flannery, Algebraic Design Theory, 2011
174 Lawrence S. Levy and J. Chris Robson, Hereditary Noetherian Prime Rings and
Idealizers, 2011
173 Sariel Har-Peled, Geometric Approximation Algorithms, 2011
172 Michael Aschbacher, Richard Lyons, Stephen D. Smith, and Ronald Solomon,
The Classification of Finite Simple Groups, 2011
171 Leonid Pastur and Mariya Shcherbina, Eigenvalue Distribution of Large Random
Matrices, 2011
170 Kevin Costello, Renormalization and Effective Field Theory, 2011
169 Robert R. Bruner and J. P. C. Greenlees, Connective Real K-Theory of Finite
Groups, 2010
168 Michiel Hazewinkel, Nadiya Gubareni, and V. V. Kirichenko, Algebras, Rings
and Modules, 2010
167 Michael Gekhtman, Michael Shapiro, and Alek Vainshtein, Cluster Algebras and
Poisson Geometry, 2010
166 Kyung Bai Lee and Frank Raymond, Seifert Fiberings, 2010
165 Fuensanta Andreu-Vaillo, José M. Mazón, Julio D. Rossi, and J. Julián
Toledo-Melero, Nonlocal Diffusion Problems, 2010
164 Vladimir I. Bogachev, Differentiable Measures and the Malliavin Calculus, 2010
163 Bennett Chow, Sun-Chin Chu, David Glickenstein, Christine Guenther, James
Isenberg, Tom Ivey, Dan Knopf, Peng Lu, Feng Luo, and Lei Ni, The Ricci Flow:
Techniques and Applications: Part III: Geometric-Analytic Aspects, 2010
162 Vladimir Mazya and Jürgen Rossmann, Elliptic Equations in Polyhedral Domains,
2010
161 Kanishka Perera, Ravi P. Agarwal, and Donal O’Regan, Morse Theoretic Aspects
of p-Laplacian Type Operators, 2010
160 Alexander S. Kechris, Global Aspects of Ergodic Group Actions, 2010
159 Matthew Baker and Robert Rumely, Potential Theory and Dynamics on the
Berkovich Projective Line, 2010
158 D. R. Yafaev, Mathematical Scattering Theory, 2010
157 Xia Chen, Random Walk Intersections, 2010
156 Jaime Angulo Pava, Nonlinear Dispersive Equations, 2009
155 Yiannis N. Moschovakis, Descriptive Set Theory, Second Edition, 2009

For a complete list of titles in this series, visit the

AMS Bookstore at www.ams.org/bookstore/survseries/.
A “quantum graph” is a graph consid-

Photo courtesy of Vani Subbarao

ered as a one-dimensional complex
and equipped with a differential oper-
ator (“Hamiltonian”). Quantum graphs
arise naturally as simplified models in
mathematics, physics, chemistry, and
engineering when one considers propaga-
tion of waves of various nature through
a quasi-one-dimensional (e.g., “meso-”
or “nano-scale”) system that looks like a thin neighborhood of a graph. Works that
currently would be classified as discussing quantum graphs have been appearing
since at least the 1930s, and since then, quantum graphs techniques have been applied
successfully in various areas of mathematical physics, mathematics in general and its
applications. One can mention, for instance, dynamical systems theory, control theory,
quantum chaos, Anderson localization, microelectronics, photonic crystals, physical
chemistry, nano-sciences, superconductivity theory, etc.
Quantum graphs present many non-trivial mathematical challenges, which makes them
dear to a mathematician’s heart. Work on quantum graphs has brought together tools
and intuition coming from graph theory, combinatorics, mathematical physics, PDEs,
and spectral theory.
This book provides a comprehensive introduction to the topic, collecting the main
notions and techniques. It also contains a survey of the current state of the quantum
graph research and applications.

For additional information

and updates on this book, visit AMS on the Web
www.ams.org/bookpages/surv-186 www.ams.org

SURV/186