Springer Theses

Recognizing Outstanding Ph.D. Research

Rico Berner

Patterns
of Synchrony in
Complex Networks
of Adaptively
Coupled Oscillators
Springer Theses

Recognizing Outstanding Ph.D. Research

Aims and Scope

The series “Springer Theses” brings together a selection of the very best Ph.D.
theses from around the world and across the physical sciences. Nominated and
endorsed by two recognized specialists, each published volume has been selected
for its scientific excellence and the high impact of its contents for the pertinent field
of research. For greater accessibility to non-specialists, the published versions
include an extended introduction, as well as a foreword by the student’s supervisor
explaining the special relevance of the work for the field. As a whole, the series will
provide a valuable resource both for newcomers to the research fields described,
and for other scientists seeking detailed background information on special
questions. Finally, it provides an accredited documentation of the valuable
contributions made by today’s younger generation of scientists.

Theses may be nominated for publication in this series by heads

of department at internationally leading universities or institutes
and should fulfill all of the following criteria
• They must be written in good English.
• The topic should fall within the confines of Chemistry, Physics, Earth Sciences,
Engineering and related interdisciplinary fields such as Materials, Nanoscience,
Chemical Engineering, Complex Systems and Biophysics.
• The work reported in the thesis must represent a significant scientific advance.
• If the thesis includes previously published material, permission to reproduce this
must be gained from the respective copyright holder (a maximum 30% of the thesis
should be a verbatim reproduction from the author’s previous publications).
• They must have been examined and passed during the 12 months prior to
nomination.
• Each thesis should include a foreword by the supervisor outlining the signifi-
cance of its content.
• The theses should have a clearly defined structure including an introduction
accessible to new PhD students and scientists not expert in the relevant field.

Indexed by zbMATH.

More information about this series at http://www.springer.com/series/8790

Rico Berner

Patterns of Synchrony
in Complex Networks
of Adaptively Coupled
Oscillators
Doctoral Thesis accepted by
Technische Universität Berlin, Berlin, Germany
Author Supervisors
Dr. Rico Berner Prof. Eckehard Schöll
Institut für Theoretische Physik Institut für Theoretische Physik
Technische Universität Berlin Technische Universität Berlin
Berlin, Germany Berlin, Germany

PD Dr. Serhiy Yanchuk

Institut für Mathematik
Technische Universität Berlin
Berlin, Germany

ISSN 2190-5053 ISSN 2190-5061 (electronic)

Springer Theses
ISBN 978-3-030-74937-8 ISBN 978-3-030-74938-5 (eBook)
https://doi.org/10.1007/978-3-030-74938-5

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Switzerland AG 2021
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Supervisors’ Foreword

This thesis focuses on the interplay between synchrony and adaptivity in complex
networks. Synchronization is a ubiquitous phenomenon observed in various contexts
in physics, chemistry, biology, neuroscience, medicine, socioeconomic systems, and
engineering. Most prominently, synchronization occurs in the brain, where it is asso-
ciated with cognitive abilities such as learning and memory, but is also a feature of
neurological disorders like Parkinson’s disease and epilepsy. Adaptivity is common
in many networks where connectivity changes over time, i.e., coupling strength
is adjusted depending on the dynamic state of the system, for example, synaptic
neuronal plasticity in the brain. Other examples occur in biological, chemical, ecolog-
ical, epidemiological, communication, and social networks. While nonlinear collec-
tive dynamics on networks with fixed connectivities has been studied extensively,
much less is known about synchronization patterns on adaptive networks.
This work contributes to a fundamental understanding of various synchronization
patterns, including multifrequency clusters, chimeras, and other partial synchroniza-
tion states on adaptive networks. After a concise survey of the fundamentals of
adaptive and complex dynamical networks and synaptic plasticity, the first part of
the thesis discusses the emergence and stability of cluster synchronization in globally
coupled adaptive networks for both paradigmatic phase oscillators as well as a more
realistic neuronal model with spike-timing-dependent plasticity. In the second part of
the thesis, the interplay between adaptivity and connectivity is investigated for more
complex network structures such as nonlocally coupled rings, random networks, and
multilayer systems. Besides presenting a plethora of novel and intriguing patterns
of synchrony, the thesis makes a number of pioneering methodological advances,
some even with rigorous mathematical proofs. These results are of interest not only
from a fundamental perspective, but also with respect to challenging applications
in neuroscience and technological systems. The considered paradigmatic model is a
network of adaptively coupled phase oscillators, where the dynamics on each node
of the network is governed by a single-phase variable. Other, more complex models
such as the Hodgkin-Huxley and the FitzHugh-Nagumo neuronal models are also
considered.

v
vi Supervisors’ Foreword

The first part of the thesis investigates self-organized clustering in globally

coupled adaptive networks. Chapter 3 focuses on multifrequency clusters in the
network of Hodgkin-Huxley neurons with spike-timing-dependent plasticity, as well
as a reduced phase oscillator model. Multifrequency clusters are patterns of partial
synchrony where each cluster has a fixed (averaged) frequency which is distinct
from the other clusters, and intra-cluster coupling is much stronger than inter-
cluster coupling. The sizes and frequencies of the individual clusters are ordered
hierarchically. Chapter 4 considers a generic model of adaptively coupled phase
oscillators with a phase-difference dependent adaptation function that generalizes
the Kuramoto-Sakaguchi model. The occurring one-cluster states associated with
a single frequency are classified as clusters of splay type (phase-locked with fixed
phase differences between any two elements), antipodal (including in-phase and anti-
phase), or double antipodal, depending on their relative phases. Rigorous analytical
as well as numerical results on existence and stability are derived. In Chap. 5, these
one-cluster states are used as building blocks for complex hierarchical multifre-
quency clusters. A systematic analytical and numerical study yields three types of
multifrequency clusters: splay, antipodal, and mixed.
While in the first part of the thesis a globally coupled network is the basis for the
adaptation of the link weights, in the second part the base topology of the network is
varied. In Chap. 6, a nonlocally coupled ring is studied using the Kuramoto-Sakaguchi
model. Besides one-cluster and multicluster states, solitary states are found, where
single nodes are not frequency synchronized with the rest. Strong analytical results on
existence and stability are presented. Chapter 7 deals with general adaptive networks
of diffusively coupled oscillators, and for the first time the concept of the Master
Stability Function is extended to adaptive networks. The master stability approach
splits the linear stability problem on networks into a network part characterized by
the spectrum of the coupling matrix, and a local dynamics part given by the maximum
Lyapunov exponent as a function of a complex parameter µ (i.e., the Master Stability
Function). Thus, the master stability function needs to be computed only once for
all network topologies, and the stability of synchronous states can then be checked
by inserting all eigenvalues of the corresponding coupling matrix for µ. Interest-
ingly, applying these general results to phase oscillators yields stability islands in
the complex µ-plane, which give rise to counter-intuitive desynchronization transi-
tions and the emergence of multicluster states and chimera states as the total coupling
strength increases. In Chap. 8, a two-layer network of adaptively coupled phase oscil-
lators is investigated. It is shown that partial synchronization patterns such as phase
clusters and more complex states, which are unstable or do not exist in singe-layer
networks, can be stabilized or even induced by multiplexing.

Berlin, Germany Eckehard Schöll

February 2021 Serhiy Yanchuk
Abstract

Collective phenomena in systems of interconnected dynamical units are omnipresent

in nature. The swarm behavior in flocks of birds or schools of fish, the synchronous
flashing of fireflies or even coherent spiking of neurons in the human brain are
just a few examples of collective motion. Elucidating the mechanisms that give
rise to synchronization is crucial in order to understand biological self-organization.
For this sake, the theory of dynamical networks has been successfully applied over
the last decades to boil down the complex dynamics from natural systems to their
essentials. In addition, dynamical networks with adaptive couplings and hence non-
constant network structure appear naturally in real-world systems such as power
grid networks, social networks as well as neuronal networks. In this thesis, we study
adaptive networks and their properties which give rise to the emergence of a variety
of synchronization patterns, including complete and cluster synchronization as well
as solitary and chimera-like states. One of the main fields of application that is inves-
tigated in this thesis concerns neuroscience. However, the methods and approaches
developed in this work are not restricted to a specific field of application.
In the first part of this thesis, we report on the phenomenon of frequency clus-
tering in a globally coupled oscillator network with slow adaptation. As a moti-
vating example, we study a network of Hodgkin-Huxley neurons with spike-timing-
dependent plasticity. Here, the clustering leads to a splitting of a neural population
into a few groups synchronized at different frequencies. We propose a phenomenolog-
ical model which describes the dynamics of two clusters taking the adaptive coupling
weights into account. Following the successful application of the phenomenological
model, we investigate a paradigmatic system of adaptively and globally coupled
phase oscillators inspired by neuronal networks with synaptic plasticity. Our numer-
ical as well as analytical study of the phase oscillator model allows for a complete
description of the mechanism behind the emergence of frequency clustering. More-
over, we unveil the role of individual clusters for the shape and stability of frequency
clusters (multiclusters), which explains the high level of multistability found in the
paradigmatic model.

vii
viii Abstract

In the second part, we extend the findings from the first part towards complex
network structures. We observe a variety of partially synchronized states such as
phase-locked, multicluster, solitary, and chimera-like states. Solitary states have been
observed in a plethora of dynamical systems. However, the mechanisms behind their
emergence were largely unaddressed in the literature. Here, we show how solitary
states emerge in the model of phase oscillators due to the adaptive feature of the
network and classify several bifurcation scenarios in which these states are created
and stabilized. In addition, we investigate the stability of synchronous states for
a large class of complex adaptively coupled oscillator networks. In particular, we
generalize the master stability approach beyond the static network paradigm by
taking adaptivity into account. We apply the new method to an adaptive network of
coupled phase oscillators and show how the subtle interplay between adaptivity and
network structure gives rise to the emergence of complex partial synchronization
patterns.
In spite of the lively interest in the topic of adaptive networks, little is known about
the interplay of adaptively coupled groups of networks. Such adaptive multilayer or
multiplex networks appear naturally in neuronal networks. We propose a concept
to generate and stabilize diverse partial synchronization patterns (phase clusters) in
adaptive networks. We show that multiplexing induces various stable phase cluster
states in a situation where they are not stable or do not even exist in the single
layer. Further, we develop a method for the analysis of Laplacian matrices of multi-
plex networks which allows for insight into the spectral structure of these networks
enabling a reduction to the stability problem of single layers. We employ the new
method to provide analytic results for the stability of the multilayer patterns.
Acknowledgements

First of all, I would like to thank Prof. Dr. Dr. h.c. Eckehard Schöll, Ph.D. and PD
Dr. Serhiy Yanchuk for introducing me to the exciting field of adaptive networks and
the theory of nonlinear dynamics. I also thank for their supervision, valuable help,
and support by suggestions and questions regarding my research. I appreciate their
encouragement to frequent visits of international conferences which enabled me to
profit from the exchange of knowledge within the research community from a very
early stage on.
I am very much in debt to Dr. Jakub Sawicki who provided guidance during
the research and writing of this thesis. Furthermore, I would like to thank Prof.
Dr. Vladimir Nekorkin and Dr. Dmitry Kasatkin for the fruitful collaboration which
resulted in some of the results presented in Chaps. 4 and 5. Likewise, I am grateful for
the collaboration with PD Dr. Oleksandr Popovych. I owe thanks to all the current,
former, and visiting members of the Schöll group and the Yanchuk group for the
wonderful working atmosphere. In particular, I thank Dr. Iryna Omelchenko, Prof. Dr.
Anna Zakharova, Ph.D., Prof. Dr. Yuri Maistrenko, Dr. Stefan Ruschel, Dr. Simona
Olmi, Dr. Dmitry Puzyrev, Dr. Ewandson Lameu, Dr. Vander Freitas, Dr. Giulia
Ruzzene, Jan Fialkowski, Florian Stelzer, Nour Eldine Hanbali, Narges Chinichian,
Michael Lindner, Deniz Nikitin, Maria Masoliver, Franziska Beckschulte, Danila
Semenov, my students Jacopo Zurbuch, Sören Nagel, Philippe Lehmann, Simon
Vock, Max Thiele, Amy Searle, Fenja Drauschke, Johanna Czech, Lucas Kluge,
Moritz Gerster, Vera Röhr, and my RISE students Bricker Ostler and Alicja Polanska
for exciting discussions and fruitful collaborations. Additionally, I owe thanks to
Prof. Dr. Sarika Jalan and Dr. Saptarshi Gosh for insightful discussions. Further, I
would like to express my thanks to Andrea Schulze who helped more than once with
her tremendous knowledge on the universities’ bureaucracy to pass administrative
challenges and to Peter Orlowski who could always help me when things went wrong
with the IT.
I would like to thank Prof. Dr. Alessandro Torcini for preparing the third assess-
ment of this thesis and Prof. Dr. Dieter Breitschwerdt for chairing the defense of this
thesis.

ix
x Acknowledgements

This work was supported in the framework of the DFG-RSF (Deutsche

Forschungsgemeinschaft-Russian Science Foundation) project: “Complex dynam-
ical networks: effects of heterogeneous, adaptive and time delayed couplings” under
SCHO 307/15-1 and YA 225/3-1. Further, I would like to thank the DAAD (Deutscher
Akademischer Austauschdienst) which provided funding within the RISE program to
host two students for a summer internship at the AG Schöll. I also thank all members
of the IRTG (International Research Training Group) 1740 and the SFB (Collorative
Research Center) 910 for supporting me to take part in their exciting events and
conferences which were a fruitful ground to develop new ideas and get into contact
with a lot of inspiring people.
Last but certainly not least I thank my family for their constant support during the
whole time of my studies and in particular I thank my dear girlfriend Veronika for
encouraging me to finalize this thesis, for supporting me when things went slowly
or got stuck and for taking me out of my scientific bubble sometimes.
Contents

1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Dynamics on Complex Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Synchronization and Collective Phenomena . . . . . . . . . . . . . . . . . . . . 2
1.3 Dynamics of Complex Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 The Role of Phase Oscillator Models for Complex Dynamical
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2 Fundamentals of Adaptive and Complex Dynamical Networks . . . . . . 23
2.1 Complex Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.1.1 Networks, Subnetworks, and Connectivity . . . . . . . . . . . . . . . 23
2.1.2 Special Network Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
2.1.3 Permutation Symmetries in Networks . . . . . . . . . . . . . . . . . . . 28
2.2 Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
2.2.1 Types of Coupling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.2.2 Kuramoto–Sakaguchi Type Model . . . . . . . . . . . . . . . . . . . . . . 30
2.2.3 Hodgkin–Huxley Model with Chemical Synapses . . . . . . . . . 32
2.3 Adaptive Networks in Neuroscience . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3.1 Spike Timing-Dependent Plasticity . . . . . . . . . . . . . . . . . . . . . 34
2.3.2 Phase Difference-Dependent Plasticity . . . . . . . . . . . . . . . . . . 34
2.3.3 A Network of Adaptively Coupled Phase Oscillator . . . . . . . 35
2.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

Part I Cluster Synchronization in Globally Coupled Adaptive

Networks
3 Population of Hodgkin–Huxley Neurons with Spike
Timing-Dependent Plasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.1 Coupled Hodgkin–Huxley Neurons on a Network with Spike
Timing-Dependent Plasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

xi
xii Contents

3.2 Numerical Observation of Synchrony and Frequency

Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.2.1 Clustering with Independent Random Input . . . . . . . . . . . . . . 48
3.3 Emergence of Two-Cluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.4 Phenomenological Model with Phase Difference-Dependent
Plasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.4.1 Properties of the Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.4.2 Comparison of the Model and Cluster Dynamics
in Hodgkin–Huxley Network . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.4.3 Criteria for the Emergence of Frequency Clusters . . . . . . . . . 58
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4 One-Cluster States in Adaptive Networks of Coupled Phase
Oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
4.1 Classification of One-Cluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
4.2 Stability of One-Cluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
4.3 Adaptation Rate Dependence of One-Cluster Stability . . . . . . . . . . . 74
4.4 Double Antipodal States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
4.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5 Multicluster States in Adaptive Networks of Coupled Phase
Oscillators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.1 Numerical Observation of Multicluster States . . . . . . . . . . . . . . . . . . . 84
5.1.1 Splay Type Cluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.1.2 Antipodal Type Cluster States . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.1.3 Mixed Type Cluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.2 Splay Type Multicluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2.1 Conditions for the Emergence of Splay Type
Multicluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.2.2 Two-Cluster States of Splay Type . . . . . . . . . . . . . . . . . . . . . . 89
5.2.3 Adaptation Rate Dependence for the Emergence
of Two-Cluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3 Conditions for the Emergence of Multicluster
States—A Generalized Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.4 Antipodal Type Multicluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.4.1 Asymptotic Conditions for the Emergence
of Antipodal Type Multicluster States . . . . . . . . . . . . . . . . . . . 96
5.4.2 Two-Cluster States of Antipodal Type . . . . . . . . . . . . . . . . . . . 98
5.5 Mixed Type Pseudo-multicluster States . . . . . . . . . . . . . . . . . . . . . . . . 100
5.5.1 Asymptotic Conditions for the Emergence of Mixed
Type Pseudo-multicluster States . . . . . . . . . . . . . . . . . . . . . . . . 101
5.5.2 Pseudo-two-cluster States of Mixed Type . . . . . . . . . . . . . . . . 102
5.6 Stability of Multicluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
Contents xiii

5.6.1 On the Stability of Multicluster States with Evenly

Sized Clusters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
5.6.2 An Effective Approach for the Stability of Multicluster
States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
5.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

Part II Interplay of Adaptivity and Connectivity

6 Adaptation on Nonlocally Coupled Ring Networks . . . . . . . . . . . . . . . . 113
6.1 Multicluster and Solitary States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.1.1 One-Cluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.1.2 Multicluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.1.3 Solitary States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.2 One-Cluster States: Local Versus Global Features . . . . . . . . . . . . . . . 120
6.2.1 Classification of One-Cluster States . . . . . . . . . . . . . . . . . . . . . 120
6.2.2 Stability of One-Cluster States . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.3 The Emergence of Solitary States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
6.4 Adaptive Networks with Global Base Topology Versus Ring
Base Topology: The Differences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
6.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
7 Synchronization on Adaptive Complex Network Structures . . . . . . . . 133
7.1 The Master Stability Function for Adaptive Complex
Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7.2 Stability Islands in the Presence of Adaptation . . . . . . . . . . . . . . . . . . 139
7.3 Stability Islands and Implications for the Emergence
of Multicluster States . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
8 Multilayered Adaptive Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.1 Lifted States in Multiplex Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
8.2 Birth and Robustness of Phase Clusters . . . . . . . . . . . . . . . . . . . . . . . . 152
8.3 Multiplex Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.4 Stabilizing Through Multiplexing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
8.5 Applications for the Multiplex Decomposition . . . . . . . . . . . . . . . . . . 162
8.5.1 The Master Stability Approach for Multiplex Networks . . . . 162
8.5.2 Analytic Treatment of Diffusive Dynamics
on Multiplex Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
8.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
xiv Contents

9 Conclusion and Outlook . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

Appendix A: Proofs of Results from the Main Text

and Supplemental Material . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
About the Author

Rico Berner is a mathematician and physicist. In his

research, he combines ideas and techniques from both
disciplines to provide a fundamental understanding of
complex dynamical systems. He studied physics and
mathematics at TU Berlin. Apart from his studies, Rico
Berner has worked with Siemens AG on applications
of machine learning algorithms and has taught mathe-
matics and physics to students in several courses. Before
starting his doctoral studies, he has been the coordi-
nator of school activities at the Matheon (TU Berlin) and
has engaged in public events of Stiftung Rechnen. Rico
Berner received the Dr. rer. nat. degree from TU Berlin.
His work was jointly supervised by Prof. Dr. Dr. h.c. Eckehard Schöll, Ph.D., at the
Institute of Theoretical Physics and PD Dr. Serhiy Yanchuk, at the Institute of Math-
ematics. His research interests include the analysis of nonlinear dynamical systems,
synchronization phenomena in complex networks and the modeling of neuronal and
technological systems. He has contributed to the following publications:

• R. Berner, S. Vock, E. Schöll, and S. Yanchuk: Desynchronization transitions in

adaptive networks, Phys. Rev. Lett. 126, 028301 (2021), Editor’s Suggestion.
• M. Gerster, R. Berner, J. Sawicki, A. Zakharova, A. Skoch, J. Hlinka, K. Lehn-
ertz, and E. Schöll: FitzHugh-Nagumo oscillators on complex networks mimic
epileptic-seizure-related synchronization phenomena, Chaos 30, 123130 (2020).
• R. Berner, A. Polanska, E. Schöll, and S. Yanchuk: Solitary states in adaptive
nonlocal oscillator networks, Eur. Phys. J. Spec. Top. 229, 2183 (2020).
• F. Drauschke, J. Sawicki, R. Berner, I. Omelchenko, and E. Schöll: Effect of
Topology upon Relay Synchronization in Triplex Neuronal Networks, Chaos 30,
051104 (2020).
• R. Berner, J. Sawicki, and E. Schöll: Birth and stabilization of phase clusters by
multiplexing of adaptive networks, Phys. Rev. Lett. 124, 088301 (2020).

xv
xvi About the Author

• V. Röhr, R. Berner, E. L. Lameu, O. V. Popovych, and S. Yanchuk: Frequency

cluster formation and slow oscillations in neural populations with plasticity, PLoS
ONE 14, e0225094 (2019).
• R. Berner, E. Schöll, and S. Yanchuk: Multiclusters in networks of adaptively
coupled phase oscillators, SIAM J. Appl. Dyn. Syst. 18, 2227 (2019).
• R. Berner, J. Fialkowski, D. V. Kasatkin, V. I. Nekorkin, S. Yanchuk, and E.
Schöll: Hierarchical frequency clusters in adaptive networks of phase oscillators,
Chaos 29, 103134 (2019).
Chapter 1
Introduction

1.1 Dynamics on Complex Networks

Complex networks are an ubiquitous paradigm in nature and technology, with a

wide field of applications from physics, chemistry, biology, neuroscience, as well
as engineering and socio-economic systems [1]. A lot of work has been devoted to
understand the statistical and topological properties of complex connectivity struc-
tures [2, 3] ranging from random [4, 5] and scale-free networks [6] to small-world
structures [7–9] and even simplicial complexes that are topological structures used
to model many body interactions [10, 11].
The field of complex dynamical networks studies the interplay of dynamics and
(static) network structures [12, 13]. Powerful methodologies as the Master Stability
Function have been suggested to provide a unified approach to study synchronization
of networks with arbitrary structure [14–19].
The investigation of complex networks and their dynamical features has become a
major research branch in neuroscience. Here, network neuroscience [20–22] and neu-
ronal dynamics [23, 24] complement each other in order to build up an understanding
for functional and structural properties of the human brain. Within this context, var-
ious network structures equipped with different neuronal models have been studied
over the last decades. For instance, ring-like structures are important motifs in neural
networks [25–28]. Specifically, nonlocally coupled rings where each node is cou-
pled to all nodes within a certain coupling range, are known to be important systems
appearing in many applied problems and theoretical studies [29–43].
In addition to the nonlocally coupled networks, other coupling structures and
motifs such as modular, scale-free, and small-world structures have been observed
in the structural and functional connectivity of human brains [44–50]. The interplay
between neuronal dynamics and these structures already revealed important mech-
anisms in the functioning of dynamical brain networks [21, 51–57]. Despite this
development and the flourishing interest in network neuroscience many questions
are still unanswered or have not yet been asked [21, 58].

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1

R. Berner, Patterns of Synchrony in Complex Networks of Adaptively Coupled Oscillators,
Springer Theses, https://doi.org/10.1007/978-3-030-74938-5_1
2 1 Introduction

Another focus of recent research in network science are multilayer networks,

which are systems interconnected through different types of links [59–63]. A promi-
nent example are social networks which can be described as groups of people with
different patterns of contacts or interactions between them [64–66]. Other applica-
tions are communication, supply, and transportation networks, for instance power
grids, subway networks, or airtraffic networks [67]. In neuroscience, multilayer net-
works represent for instance neurons in different areas of the brain, neurons con-
nected either by a chemical link or by an electrical synapses [68–70], or the modular
connectivity structure of brain regions [71–73]. A special case of multilayer net-
works are multiplex topologies, where each layer contains the same set of nodes, and
only pairwise connections between corresponding nodes from neighboring layers
exist [74–97].

1.2 Synchronization and Collective Phenomena

Collective behavior in networks of coupled oscillatory systems has attracted a lot of

attention over the last decades. Depending on the network and the specific dynami-
cal system, various synchronization patterns of increasing complexity were explored.
Even in simple models of coupled oscillators, patterns such as complete synchro-
nization [98, 99], cluster synchronization where the network splits into groups of
synchronous elements [17, 100–106], and various forms of partial synchronization
patterns have been found [107]. The investigation of synchronization phenomena
on complex networks [13, 99, 108–116] has become fruitful for applications in
many areas ranging from physics and chemistry to biology, neuroscience, physiol-
ogy, ecology, socio-economic systems, computer science and engineering. For exam-
ple, synchronization of neurons is believed to play a crucial role in brain networks
under normal conditions, for instance in the context of cognition and learning [117,
118], and under pathological conditions such as Parkinson’s disease [119, 120],
epilepsy [57, 121–125], tinnitus [126, 127], schizophrenia, or autism, to name a
few [128]. Another example is the synchronized flashing in groups of fireflies [129].
Also in power grid networks synchronization is essential for their operation [130–
137]. The control of dynamics on networks is another important issue with a lot of
applications [138, 139].
One of the most important partial synchronization patterns are chimera states.
As they have been theoretically predicted [140, 141], these states are character-
ized by a spatial coexistence of coherence (synchronized domains) and incoherence
(desynchronized domains). Numerous works have been devoted to this phenomenon,
theoretically as well as experimentally, showing that chimera states play a crucial
role for the understanding of real-world dynamical systems [36, 38, 40, 43, 69, 142–
149]. Apart from the theoretical importance, complex dynamical scenarios such as
chimera states are found in a wide range of experimental systems, e.g. optical light
modulators [150] and chemical oscillators [151–153], mechanical or optomechan-
ical [154–157], electronic or optoelectronic oscillators [158, 159], electrochemical
1.2 Synchronization and Collective Phenomena 3

systems [160–163], and electronic circuits [164, 165]. Further, it has been shown
that chimera states are crucial for the understanding of epileptic seizures [53, 124]
or unihemispheric sleep [56, 166, 167]. Due to their importance, several approaches
have been developed to control chimera states and other partial synchronization pat-
terns [168–172].
Beyond chimera states various other types of partial synchronization scenarios
have been found [74, 173–178]. Solitary [175] or Bellerophon states [178] are just
two examples of the plethora of synchronization patterns discovered in the transition
from incoherence to coherence or as a result of the destabilization of the coherent
state.
Recently, solitary states have been found in diverse dynamical systems. Soli-
tary states are described as states for which only one single element (or a small
group of elements compared to the total number of all dynamical elements in a net-
works) behaves differently compared with the behavior of the background group,
i.e., the neighboring elements. These kinds of states have been observed in gen-
eralized Kuramoto–Sakaguchi models [175, 179–181], the Kuramoto model with
inertia [182–184], adaptive networks [185], models of power grids [135, 186], the
Stuart-Landau model [187], the FitzHugh–Nagumo model [188, 189], systems of
excitable units [190] as well as in Lozi maps [191] and even in experimental setups of
coupled pendula [155]. Solitary states are considered as important states in the transi-
tion from coherent to incoherent dynamics [86, 182, 192]. Despite their appearance
in many well-studied models, the mechanisms of their emergence are less under-
stood. Until now, only a few works could shed some light on the details behind their
formation [175, 180, 183, 193].

1.3 Dynamics of Complex Networks

Besides the paradigm of static networks the analysis of interconnected dynamical

system on temporally evolving connectivity structures has become a future challenge
in nonlinear science [194]. Here, two approaches have to be distinguished. The first
which uses a prescribed temporal evolution of the network structure [195, 196],
called temporal networks, and a second where the temporal evolution of the network
depends upon on the dynamical state of the network nodes [197], called adaptive net-
works. In this thesis, we focus on the latter approach and consider adaptive networks
throughout.
With regard to this thesis, one of the main motivations for studying adaptive
dynamical networks comes from the field of neuroscience. Here, a possible mecha-
nisms that causes adaptation, which may lead to persistent changes in neural connec-
tions and relates to learning and memory, is synaptic plasticity [198]. The efficiency
of synapses to transmit the electrical potential between neurons may increase or
decrease depending on the mutual neural activity, which results in short- or long-
term potentiation or, respectively, depression of synapses [199, 200]. An example is
spike timing-dependent plasticity (STDP) which describes the change of the synaptic
4 1 Introduction

weight as a function of the difference of spiking times between pre- and post-synaptic
neurons [201–208]. Thus the network structure reorganizes adaptively in response
to the neuronal dynamics. Similarly, chemical systems have been reported [209],
where the reaction rates adapt dynamically depending on the variables of the system.
Activity-dependent plasticity is also common in epidemics [210] and in biological
or social systems [211, 212]. Many studies are devoted to the mechanism of plastic-
ity in chemical synapses, see e.g. [204, 213], however, it has been shown that also
electrical synapses undergo plastic changes [68, 214].
A famous example for a plasticity rule is the Hebbian rule which assumes that
the modifications of the synaptic weights are driven by correlations in the firing
activity of pre- and post-synaptic neurons. More specifically, the rule assumes that
those connections are potentiated, for which one neuron contributes to the firing
of another [198]. Nevertheless, in many publications, the Hebbian rule is consid-
ered in a narrower sense of closeness between the spiking times: the smaller the
distances between the spikes are the higher is the potentiation of the corresponding
synapse [215, 216]. In this work, we are dealing with spike-based learning rules
rather than rate-based.
Previous studies of neural networks with STDP (or STDP-like) showed that such
networks can evolve and create various coupling and spiking patterns [217–219]. For
instance, the coupling weights can exhibit stable localized spatial structures, which
can be interpreted as receptive fields [220]. These structures can be either unidi-
rectionally of bidirectionally coupled, depending on the plasticity rule or external
input properties. The STDP mechanism plays an important role in temporal cod-
ing of information by spikes [201, 220]. On the one hand, synchronized firing in
neural ensembles with STDP can be stabilized through potentiation of synaptic cou-
pling by stimulation-induced transient synchronization of neurons [127, 208, 221,
222]. On the other hand, a desynchronized state can lead to a depression of synaptic
weights [127, 221]. Thus, neural networks with plasticity are prone to co-existence of
different stable dynamical and structural states, which may be realized by choosing
appropriate initial conditions or stimulation procedures.
In addition to the localized spatial structures, the emergence of modular and
scale-free networks has been reported [223–236] in networks with STDP. This fact
underlines the potential importance of adaptive mechanism in the formation of con-
nectivity structures as they have been experimentally found in brain networks [47,
50]. Furthermore, activity based adaptive rewiring has even been shown to enhance
modularity [237].
Besides direct application to neuroscience, spike timing-dependent plasticity rules
have been discussed for neurocomputing [238] and have recently also been imple-
mented into memristive devices [239, 240]. For early work on electronic switching
and S-shaped current-voltage characteristics see [241–243]. Memristors and mem-
ristor arrays play an important role in the development of neuromorphic comput-
ing [244–248] which is believed to be the future of artificial intelligence. An exhaus-
tive survey on approaches to neuromorphic computing is presented in Ref. [249]
showing the plethora of work devoted to this questions. The huge common interest
of the scientific community in developing brain-inspired technologies is not least
1.3 Dynamics of Complex Networks 5

expressed by the flagship project “The human brain project” funded by the European
Union [250, 251] or the NIH BRAIN Initiative [252]. Therefore, there is clearly
an urge for the analysis of adaptive neural networks and the development of novel
methods to understand these systems.
Next to the active interest in adaptive networks with regards to neuroscience,
motivation for systems with adaptive connectivity structure has been found in other
dynamical systems from biology [253] and social science [254–256], in systems
that describe swarming behavior of coupled agents [257, 258], or in the dynamics of
communication networks [259]. Furthermore, it has been shown that synchronization
in complex networks is enhanced by an adaptation of the coupling structure [260,
261]. Due to this, adaptive network structures have been also successfully applied to
control synchronization in complex and even time-evolving networks [262–272].

1.4 The Role of Phase Oscillator Models for Complex

Dynamical Networks

Models of coupled phase oscillators, i.e., models where the dynamical variable is
simply given by an angle, are nowadays a well-known paradigm in order to study
collective behavior of dynamical oscillatory units on complex networks [99, 273]. In
this context, a particularly important feature is that coupled nonlinear oscillators can
be reduced to phase oscillator networks in case of weak interactions [99, 274–276].
The reduction of complex dynamical systems to networks of coupled phase oscillators
is well-known and there exist exhaustive reviews highlighting the importance of
phase oscillator models [99, 276, 277]. Recent studies, in addition, aim to make
phase oscillators models even more powerful by generalizing the conditions under
which the reduction techniques are valid [278–281].
As a representative from the class of phase oscillator models, the Kuramoto model
where all oscillators are coupled in an all-to-all manner, particularly, has attracted a
lot of attention due to its simple form and mathematical tractability [98, 108]. The
Kuramoto model has gained additional popularity due to its application for real-world
problems [99, 111, 282–284]. Despite the simple structure of the Kuramoto model
many different dynamical regimes have been observed. In order to get a hand on
the dynamics of the Kuramoto model sophisticated methods have been developed.
In particular, sinusoidally and globally coupled phase oscillators have been shown
to be partially integrable. The Watanabe–Strogatz theory, which was introduced to
show the integrability, allows for a reduction of a system with finite size to only three
dimensions and can be applied to even more general classes of phase oscillator mod-
els [285–287]. Another approach which has been developed to understand coupled
oscillators in the continuum limit is the Ott-Antonson ansatz [288]. The Ott-Antonson
theory allows for a reduction to a two dimensional dynamical system in the case of an
infinite number of oscillators. The theory has been successfully applied to describe
the emergence of chimera states even beyond phase oscillator models [38, 43, 146,
6 1 Introduction

289–291]. Remarkably for both reduction techniques, the Watanabe–Strogatz and

the Ott-Antonson theory, the reduced systems possess a clear physical interpreta-
tion, and both approaches are intimately related with each other [292, 293]. Just
recently, another very promising reduction technique for finite systems has been
introduced which make use of so-called collective coordinates [294–296].
Beyond the classical Kuramoto model various generalization have been proposed.
Starting from generalization to complex networks [297, 298], the theory of phase
oscillators has been further developed to study phenomena of phase transitions [299–
301], network symmetries [104], the impact of inertia [175, 302–305] and other
forms of frequency adaptation [306], delayed coupling [307], or the effect of time-
depending parameters [308] to name just a few.
Another generalization which gained a lot of attention over the last years concerns
the effects and phenomena in networks of phase oscillators with adaptive coupling.
Various models have been suggested and studied to gain insights into the interplay
of collective dynamics and adaptivity [309]. A lot of works are inspired by recent
findings in neuroscience and implement synaptic plasticity into their models [137,
216, 226, 227, 310–339]. Despite the rich interest in neuroscientific applications,
also other forms of adaptation and other applications have been considered. For
example, effects of adaptive rewiring, an adaptive network growth, or evolutionary
edge-snapping are studied in Refs. [340–344]. Also the interplay of adaptivity and
complex network structure such as multilayers have been only recently numerically
investigated [95, 345–347].

1.5 Outline

This thesis is devoted to the analysis of partial synchronization patterns in complex

oscillator networks with adaptive coupling strength. This analysis is organized in two
parts. After two introductory chapters, in the first part, the Hodgkin-Huxley neuron
model and the phase oscillator model on a globally coupled network with synaptic
plasticity are investigated. Here, we observe the emergence of multiclusters due to
plasticity and underline the importance of phase oscillator models in order to under-
stand these complex synchronization pattern. By rigorous mathematical methods,
we provide a comprehensive understanding of the cluster structures. In part II, the
paradigmatic model of adaptively coupled phase oscillators, introduced in part I, is
used in order to study the interplay of adaptivity and complex connectivity structures.
The basic formalism and the standard notation which is used through out part
I and II is presented in Chap. 2. Here, we provide graph theoretical preliminaries
and introduce all classes of networks studied in this thesis. Furthermore, a brief
introduction into the dynamics on complex networks is given where we discuss
different types of coupling and models. In the last section, we show how adaptivity
is modeled in the context of synaptic plasticity. In addition, the reduction to phase
oscillator models with phase difference-dependent plasticity is given.
1.5 Outline 7

Part I starts with Chap. 3. In this chapter, we analyze a population of Hodgkin-

Huxley neurons with spike timing-dependent plasticity. The numerical observation
of multiclusters is shown and the detailed mechanisms of the stability of frequency
clusters is explained by using the simplest case of two clusters. Here, we propose
a phenomenological model, which describes the dynamics of two clusters taking
into account the adaptation of the weights. The model is shown to reflect not only
qualitative, but also some basic quantitative properties of the two-cluster formation.
We also determine the set of plasticity functions (update rules), which lead to the
clustering.
In Chap. 4, the existence and stability of one-cluster states is rigorously analyzed
in the phase oscillator model. We provide a full classification of one-cluster states and
investigate the impact of the timescale separation on their stability. Different types are
characterized by different phase relations of the nodes. Finally, we discuss the role of
a special type of one-clusters, called double antipodal states. Although such states are
unstable for all parameter values, they appear as saddles connecting synchronous and
splay states. As a result, double antipodal states are observed during a “meta-stable”
transition between the phase-synchronous and non-phase-synchronous state.
Chapter 5 is devoted to the phenomenon of multicluster states which are distin-
guished by different frequencies of the individual clusters. The multiclusters are,
first, illustrated using numerical integration, and then classified. Complementing
earlier work, we find a mixed multicluster state. These new multiclusters consist
simultaneously of splay and (in-) anti-phase synchronous states. It results from this
that all one-cluster states can serve as building blocks for a multicluster. Further,
we provide an analytic description of all types of multicluster states using explicit
methods as well as methods from the theory of asymptotic expansions. We show that
multiclusters corresponding to certain fixed frequency ratio appear in continuous
families, and, moreover, multiclusters with different frequency ratios coexist for the
same parameter values. We derive algebraic equations for the frequencies of coex-
isting multicluster solutions. In a particular case of two clusters, these equations can
be explicitly solved. In the end, the stability of multicluster states is discussed and
connected to the stability of their individual building blocks, i.e., one-cluster states.
The second part starts with Chap. 6. In this chapter, we analyze the emergence
of solitary states on a nonlocally coupled ring in the presence of plastic coupling
weights. For this, numerical results and a rigorous definition for multicluster and
solitary states on complex networks are presented. We further provide a detailed
analysis of one-cluster states and derive explicit relations between local and global
properties. A reduced model for two-clusters is derived which allows us to unveil the
bifurcation scenarios in which solitary states are formed and (de)stabilized.
Complex networks of adaptively coupled oscillators are studied in Chap. 7. Here,
we consider the stability of synchronous states and derive the master stability function
for a large class of adaptive networks. The new method is applied to a system of
adaptively coupled phase oscillators with complex connectivity structure. We observe
the existence of stability islands in the master stability function and connect these
islands to the emergence of multicluster as well as chimera-like states.
8 1 Introduction

In Chap. 8 we show that a plethora of novel patterns can be generated by mul-

tiplexing adaptive networks in a multilayer configuration. In particular, partial syn-
chronization patterns like phase clusters and more complex cluster states which are
unstable in the corresponding monoplex network can be stabilized, or even states
which do not exist in the single-layer case for the parameters chosen, can be born
by multiplexing. We elucidate the delicate balance of adaptation and multiplexing
which is a feature of many real-world networks even beyond neuroscience.
The thesis concludes with Chap. 9 where the obtained results are summarized and
discussed. Furthermore, an outlook on possible continuations of this work is given.

References

1. Newman MEJ (2003) The structure and function of complex networks. SIAM Rev 45:167
2. Albert R, Barabási AL (2002) Statistical mechanics of complex networks. Rev Mod Phys
74:47
3. Costa LDF, Rodrigues FA, Travieso G, Villas Boas PR (2007) Characterization of complex
networks: a survey of measurements. Adv Phys 56:167
4. Erdös P, Rényi A (1959) On random graphs. Publ Math Debrecen 6:290
5. Erdös P, Rényi A (1960) On the evolution of random graphs. Publ Math Inst Hung Acad Sci
5:17
6. Barabási AL, Albert R (1999) Emergence of scaling in random networks. Science 286:509
7. Watts DJ, Strogatz SH (1998) Collective dynamics of“small-world” networks. Nature 393:440
8. Newman MEJ, Moore C, Watts DJ (2000) Mean-field solution of the small-world network
model. Phys Rev Lett 84:3201
9. Maier BF (2019) Generalization of the small-world effect on a model approaching the Erdös-
Rényi random graph. Sci Rep 9:9268
10. Giusti C, Ghrist R, Bassett DS (2016) Two’s company, three (or more) is a simplex. J Comp
Neurosci 41:1
11. Sizemore AE, Phillips-Cremins JE, Ghrist R, Bassett DS (2019) The importance of the whole:
topological data analysis for the network neuroscientist. Netw Neurosci 3:656
12. Porter MA, Gleeson JP (2016) Dynamical systems on networks. Frontiers in applied dynam-
ical systems: reviews and tutorials, vol. 4. Springer International Publishing
13. Boccaletti S, Pisarchik AN, del Genio CI, Amann A (2018) Synchronization: from coupled
systems to complex networks. Cambridge University Press, Cambridge
14. Pecora LM, Carroll TL (1998) Master stability functions for synchronized coupled systems.
Phys Rev Lett 80:2109
15. Choe CU, Dahms T, Hövel P, Schöll E (2010) Controlling synchrony by delay coupling in
networks: from in-phase to splay and cluster states. Phys Rev E 81:025205(R)
16. Kyrychko YN, Blyuss KB, Schöll E (2014) Synchronization of networks of oscillators with
distributed-delay coupling. Chaos 24:043117
17. Pecora LM, Sorrentino F, Hagerstrom AM, Murphy TE, Roy R (2014) Symmetries, cluster
synchronization, and isolated desynchronization in complex networks. Nat Commun 5:4079
18. Wille C, Lehnert J, Schöll E (2014) Synchronization-desynchronization transitions in complex
networks: an interplay of distributed time delay and inhibitory nodes. Phys Rev E 90:032908
19. Lehnert J (2016) Controlling synchronization patterns in complex networks, Springer Theses.
Springer, Heidelberg
20. Bullmore ET, Sporns O (2009) Complex brain networks: graph theoretical analysis of struc-
tural and functional systems. Nat Rev Neurosci 10:186
21. Bassett DS, Sporns O (2017) Network neuroscience. Nat. Neurosci. 20:353 EP (review Arti-
cle)
References 9

22. Bertolero M, Bassett DS (2019) How matter becomes mind. Sci. Am. pp 18–25
23. Gerstner W, Kistler WM, Naud R, Paninski L (2014) Neuronal dynamics: from single neurons
to networks and models of cognition. Cambridge University Press, Cambridge
24. Börgers C (2017) An introduction to modeling neuronal dynamics. Springer, Cham
25. Compte A, Sanchez-Vives MV, McCormick DA, Wang XJ (2003) Cellular and network mech-
anisms of slow oscillatory activity (<1 Hz) and wave propagations in a cortical network model.
J Neurophys 89:2707
26. Sporns O (2011) Networks of the brain. MIT Press, Cambridge
27. Popovych OV, Yanchuk S, Tass P (2011) Delay- and coupling-induced firing patterns in
oscillatory neural loops. Phys Rev Lett 107:228102
28. Yanchuk S, Perlikowski P, Popovych OV, Tass P (2011) Variability of spatio-temporal patterns
in non-homogeneous rings of spiking neurons. Chaos 21:047511
29. Pasemann F (1995) Characterization of periodic attractors in neural ring networks. Neural
Netw 8:421
30. Bressloff PC, Coombes S, de Souza B (1997) Dynamics of a ring of pulse-coupled oscillators:
group-theoretic approach. Phys Rev Lett 79:2791
31. Yanchuk S, Wolfrum M (2008) Destabilization patterns in chains of coupled oscillators. Phys
Rev E 77:26212
32. Bonnin M (2009) Waves and patterns in ring lattices with delays. Phys D 238:77
33. Zou W, Zhan M (2009) Splay states in a ring of coupled oscillators: from local to global
coupling. SIAM J Appl Dyn Syst 8:1324
34. Horikawa Y, Kitajima H (2009) Duration of transient oscillations in ring networks of unidi-
rectionally coupled neurons. Phys D 238:216
35. Perlikowski P, Yanchuk S, Popovych OV, Tass P (2010) Periodic patterns in a ring of delay-
coupled oscillators. Phys Rev E 82:036208
36. Omelchenko I, Maistrenko Y, Hövel P, Schöll E (2011) Loss of coherence in dynamical
networks: spatial chaos and chimera states. Phys Rev Lett 106:234102
37. Kantner M, Yanchuk S (2013) Bifurcation analysis of delay-induced patterns in a ring of
Hodgkin-Huxley neurons. Phil Trans R Soc A 371:20120470
38. Omelchenko I, Omel’chenko OE, Hövel P, Schöll E (2013) When nonlocal coupling between
oscillators becomes stronger: patched synchrony or multichimera states. Phys Rev Lett
110:224101
39. Yanchuk S, Perlikowski P, Wolfrum M, Stefanski A, Kapitaniak T (2015) Amplitude equations
for collective spatio-temporal dynamics in arrays of coupled systems. Chaos 25:033113
40. Schöll E (2016) Synchronization patterns and chimera states in complex networks: interplay
of topology and dynamics. Eur Phys J Spec Top 225:891
41. Klinshov V, Shchapin D, Yanchuk S, Wolfrum M, D’Huys O, Nekorkin VI (2017) Embedding
the dynamics of a single delay system into a feed-forward ring. Phys Rev E 96:042217
42. Burylko O, Mielke A, Wolfrum M, Yanchuk S (2018) Coexistence of hamiltonian-like and
dissipative dynamics in rings of coupled phase oscillators with skew-symmetric coupling.
SIAM J Appl Dyn Syst 17:2076
43. Omel’chenko OE (2018) The mathematics behind chimera states. Nonlinearity 31:R121
44. Eguiluz VM, Chialvo DR, Cecchi GA, Baliki M, Apkarian AV (2005) Scale-free brain func-
tional networks. Phys Rev Lett 94:4
45. Humphries MD, Gurney K, Prescott TJ (2006) The brainstem reticular formation is a small-
world, not scale-free, network. Proc R Soc B: Biol Sci 273:503
46. Bassett DS, Bullmore ET (2006) Small-world brain networks. Neuroscientist 12:512
47. Meunier D, Lambiotte R, Bullmore ET (2010) Modular and hierarchically modular organi-
zation of brain networks. Front Neurosci 4:200
48. Wildie M, Shanahan M (2012) Hierarchical clustering identifies hub nodes in a model of
resting-state brain activity. In: The 2012 international joint conference on neural networks
(IJCNN). IEEE, pp 1–6
49. Rieubland S, Roth A, Häusser M (2014) Structured connectivity in cerebellar inhibitory
networks. Neuron 81:913
10 1 Introduction

50. Ashourvan A, Telesford QK, Verstynen T, Vettel JM, Bassett DS (2019) Multi-scale detection
of hierarchical community architecture in structural and functional brain networks. PLoS ONE
14:e0215520
51. Zhou C, Zemanová L, Zamora G, Hilgetag CC, Kurths J (2006) Hierarchical organization
unveiled by functional connectivity in complex brain networks. Phys Rev Lett 97:238103
52. Zhou C, Zemanová L, Zamora-López G, Hilgetag CC, Kurths J (2007) Structure-function
relationship in complex brain networks expressed by hierarchical synchronization. New J
Phys 9:178
53. Chouzouris T, Omelchenko I, Zakharova A, Hlinka J, Jiruska P, Schöll E (2018) Chimera
states in brain networks: empirical neural vs. modular fractal connectivity. Chaos 28:045112
54. Hövel P, Viol A, Loske P, Merfort L, Vuksanović V (2018) Synchronization in functional
networks of the human brain. J Nonlinear Sci
55. Wang R, Lin P, Liu M, Wu Y, Zhou T, Zhou C (2019) Hierarchical connectome modes and
critical state jointly maximize human brain functional diversity. Phys Rev Lett 123:038301
56. Ramlow L, Sawicki J, Zakharova A, Hlinka J, Claussen JC, Schöll E (2019) Partial synchro-
nization in empirical brain networks as a model for unihemispheric sleep. EPL 126:50007
57. Gerster M, Berner R, Sawicki J, Zakharova A, Skoch A, Hlinka J, Lehnertz K, Schöll E
(2020) FitzHugh-Nagumo oscillators on complex networks mimic epileptic-seizure-related
synchronization phenomena. Chaos 30:123130
58. Bassett DS, Zurn P, Gold JI (2018) On the nature and use of models in network neuroscience.
Nat Rev Neurosci 19:566
59. De Domenico M, Solé-Ribalta A, Cozzo E, Kivelä M, Moreno Y, Porter MA, Gómez S,
Arenas A (2013) Mathematical formulation of multilayer networks. Phys Rev X 3:041022
60. Boccaletti S, Bianconi G, Criado R, del Genio CI, Gómez-Gardeñes J, Romance M, Sendiña
Nadal I, Wang Z, Zanin M (2014) The structure and dynamics of multilayer networks. Phys
Rep 544:1
61. Kivelä M, Arenas A, Barthélemy M, Gleeson JP, Moreno Y, Porter MA (2014) Multilayer
networks. J Complex Netw 2:203
62. De Domenico M, Nicosia V, Arenas A, Latora V (2015) Structural reducibility of multilayer
networks. Nat Commun 6:6864
63. Belykh IV, Carter D, Jeter R (2019) Synchronization in multilayer networks: when good links
go bad. SIAM J Appl Dyn Syst 18:2267
64. Girvan M, Newman MEJ (2002) Community structure in social and biological networks. Proc
Natl Acad Sci USA 99:7821
65. Amato R, Díaz-Guilera A, Kleineberg KK (2017) Interplay between social influence and
competitive strategical games in multiplex networks. Sci Rep 7:7087
66. Amato R, Kouvaris NE, San Miguel M, Díaz-Guilera A (2017) Opinion competition dynamics
on multiplex networks. New J Phys 19:123019
67. Cardillo A, Zanin M, Gòmez Gardeñes J, Romance M, del Amo AG, Boccaletti S (2013) Mod-
eling the multi-layer nature of the european air transport network: resilience and passengers
re-scheduling under random failures. Eur Phys J ST 215:23
68. Pereda AE (2014) Electrical synapses and their functional interactions with chemical
synapses. Nat Rev Neurosci 15:250
69. Majhi S, Bera BK, Ghosh D, Perc M (2019) Chimera states in neuronal networks: a review.
Phys Life Rev 28:100
70. Bera BK, Rakshit S, Ghosh D (2019) Intralayer synchronization in neuronal multiplex net-
work. Eur Phys J Spec Top 228:2441
71. Bentley B, Branicky R, Barnes CL, Chew YL, Yemini E, Bullmore ET, Vétes PE, Schafer
WR (2016) The multilayer connectome of caenorhabditis elegans. PLoS Comput Biol
12:e1005283
72. Battiston F, Nicosia V, Chavez M, Latora V (2017) Multilayer motif analysis of brain networks.
Chaos 27:047404
73. Vaiana M, Muldoon SF (2018) Multilayer brain networks. J Nonlinear Sci, pp 1–23
References 11

74. Zhang X, Boccaletti S, Guan S, Liu Z (2015) Explosive synchronization in adaptive and
multilayer networks. Phys Rev Lett 114:038701
75. Maksimenko VA, Makarov VV, Bera BK, Ghosh D, Dana SK, Goremyko MV, Frolov NS,
Koronovskii AA, Hramov AE (2016) Excitation and suppression of chimera states by multi-
plexing. Phys Rev E 94:052205
76. Sevilla-Escoboza R, Sendiña-Nadal I, Leyva I, Gutiérrez R, Buldú JM, Boccaletti S (2016)
Inter-layer synchronization in multiplex networks of identical layers. Chaos 26:065304
77. Jalan S, Singh A (2016) Cluster synchronization in multiplex networks. Europhys Lett
113:30002
78. Requejo RJ, Díaz-Guilera A (2016) Replicator dynamics with diffusion on multiplex net-
works. Phys Rev E 94:022301
79. Ghosh S, Kumar A, Zakharova A, Jalan S (2016) Birth and death of chimera: interplay of
delay and multiplexing. Europhys Lett 115:60005
80. Leyva I, Sevilla-Escoboza R, Sendiña-Nadal I, Gutiérrez R, Buldú JM, Boccaletti S (2017)
Inter-layer synchronization in non-identical multi-layer networks. Sci Rep 7:45475
81. Andrzejak RG, Ruzzene G, Malvestio I (2017) Generalized synchronization between chimera
states. Chaos 27:053114
82. Frolov NS, Maksimenko VA, Makarov VV, Kirsanov D, Hramov AE, Kurths J (2018) Macro-
scopic chimeralike behavior in a multiplex network. Phys Rev E 98:022320
83. Pitsik E, Makarov V, Kirsanov D, Frolov NS, Goremyko M, Li X, Wang Z, Hramov AE, Boc-
caletti S (2018) Inter-layer competition in adaptive multiplex network. New J Phys 20:075004
84. Leyva I, Sendiña-Nadal I, Sevilla-Escoboza R, Vera-Avila VP, Chholak P, Boccaletti S (2018)
Relay synchronization in multiplex networks. Sci Rep 8:8629
85. Ghosh S, Zakharova A, Jalan S (2018) Non-identical multiplexing promotes chimera states.
Chaos, Solitons Fractals 106:56
86. Mikhaylenko M, Ramlow L, Jalan S, Zakharova A (2019) Weak multiplexing in neural net-
works: Switching between chimera and solitary states. Chaos 29:023122
87. Sawicki J, Omelchenko I, Zakharova A, Schöll E (2018) Synchronization scenarios of
chimeras in multiplex networks. Eur Phys J Spec Top 227:1161
88. Sawicki J, Omelchenko I, Zakharova A, Schöll E (2018) Delay controls chimera relay syn-
chronization in multiplex networks. Phys Rev E 98:062224
89. Semenova N, Zakharova A (2018) Weak multiplexing induces coherence resonance. Chaos
28:051104
90. Omelchenko I, Hülser T, Zakharova A, Schöll E (2019) Control of chimera states in multilayer
networks. Front Appl Math Stat 4:67
91. Rybalova E, Vadivasova T, Strelkova G, Anishchenko V, Zakharova A (2019) Forced synchro-
nization of a multilayer heterogeneous network of chaotic maps in the chimera state mode.
Chaos 29:033134
92. Nikitin D, Omelchenko I, Zakharova A, Avetyan M, Fradkov AL, Schöll E (2019) Complex
partial synchronization patterns in networks of delay-coupled neurons. Phil Trans R Soc A
377:20180128
93. Blaha KA, Huang K, Della Rossa F, Pecora LM, Hossein-Zadeh M, Sorrentino F (2019)
Cluster synchronization in multilayer networks: a fully analog experiment with lc oscillators
with physically dissimilar coupling. Phys Rev Lett 122:014101
94. Jalan S, Kumar A, Leyva I (2019) Explosive synchronization in frequency displaced multiplex
networks. Chaos 29:041102
95. Berner R, Sawicki J, Schöll E (2020) Birth and stabilization of phase clusters by multiplexing
of adaptive networks. Phys Rev Lett 124:088301
96. Drauschke F, Sawicki J, Berner R, Omelchenko I, Schöll E (2020) Effect of topology upon
relay synchronization in triplex neuronal networks. Chaos 30:051104
97. Yamakou ME, Hjorth PG, Martens EA (2020) Optimal self-induced stochastic resonance in
multiplex neural networks: electrical vs. chemical synapses. Front Comput Neurosci 14:62
98. Kuramoto Y (1984) Chemical oscillations, waves and turbulence. Springer, Berlin
12 1 Introduction

99. Pikovsky A, Rosenblum M, Kurths J (2001) Synchronization: a universal concept in nonlinear

sciences, 1st edn. Cambridge University Press, Cambridge
100. Yanchuk S, Maistrenko Y, Mosekilde E (2001) Partial synchronization and clustering in a
system of diffusively coupled chaotic oscillators. Math Comp Simul 54:491
101. Sorrentino F, Ott E (2007) Network synchronization of groups. Phys Rev E 76:056114
102. Belykh IV, Hasler M (2011) Mesoscale and clusters of synchrony in networks of bursting
neurons. Chaos 21:016106
103. Dahms T, Lehnert J, Schöll E (2012) Cluster and group synchronization in delay-coupled
networks. Phys Rev E 86:016202
104. Nicosia V, Valencia M, Chavez M, Díaz-Guilera A, Latora V (2013) Remote synchronization
reveals network symmetries and functional modules. Phys Rev Lett 110:174102
105. Golubitsky M, Stewart I (2016) Rigid patterns of synchrony for equilibria and periodic cycles
in network dynamics. Chaos 26:094803
106. Zhang Y, Motter AE (2020) Symmetry-independent stability analysis of synchronization
patterns. SIAM Rev 62:817
107. Bick C, Goodfellow M, Laing CR, Martens EA (2020) Understanding the dynamics of bio-
logical and neural oscillator networks through exact mean-field reductions: a review. J Math
Neurosci 10:9
108. Strogatz SH (2000) From Kuramoto to Crawford: exploring the onset of synchronization in
populations of coupled oscillators. Phys D 143:1
109. Strogatz SH (2001) Exploring complex networks. Nature 410:268
110. Boccaletti S, Kurths J, Osipov G, Valladares DL, Zhou CS (2002) The synchronization of
chaotic systems. Phys Rep 366:1
111. Strogatz SH (2003) Sync: how order emerges from chaos in the universe, nature, and daily
life. Hyperion, New York
112. Nishikawa T, Motter AE (2006) Synchronization is optimal in nondiagonalizable networks.
Phys Rev E 73:065106
113. Arenas A, Díaz-Guilera A, Kurths J, Moreno Y, Zhou C (2008) Synchronization in complex
networks. Phys Rep 469:93
114. Balanov AG, Janson NB, Postnov DE, Sosnovtseva OV (2009) Synchronization: from simple
to complex. Springer, Berlin
115. Nekorkin VI (2015) Introduction to nonlinear oscillations. Wiley, Weinheim
116. Maia DMN, Macau EEN, Pereira T, Yanchuk S (2018) Synchronization in networks with
strongly delayed couplings. Discr Cont Dyn Syst B 23:3461
117. Singer W (1999) Neuronal synchrony: a versatile code review for the definition of relations?
Neuron 24:49
118. Fell J, Axmacher N (2011) The role of phase synchronization in memory processes. Nat Rev
Neurosci 12:105
119. Hammond C, Bergman H, Brown P (2007) Pathological synchronization in Parkinson’s dis-
ease: networks, models and treatments. Trends Neurosci 30:357
120. Adamchic I, Hauptmann C, Barnikol UB, Pawelczyk N, Popovych OV, Barnikol TT, Silchenko
AN, Volkmann J, Deuschl G, Meissner WG, Maarouf M, Sturm V, Freund HJ, Tass PA
(2014) Coordinated reset neuromodulation for Parkinson’s disease: proof-of-concept study.
Movement Disord 29:1679
121. Jiruska P, de Curtis M, Jefferys JGR, Schevon CA, Schiff SJ, Schindler K (2013) Synchro-
nization and desynchronization in epilepsy: controversies and hypotheses. J Phys 591(4):787
122. Jirsa VK, Stacey WC, Quilichini PP, Ivanov AI, Bernard C (2014) On the nature of seizure
dynamics. Brain 137:2210
123. Rothkegel A, Lehnertz K (2014) Irregular macroscopic dynamics due to chimera states in
small-world networks of pulse-coupled oscillators. New J Phys 16:055006
124. Andrzejak RG, Rummel C, Mormann F, Schindler K (2016) All together now: analogies
between chimera state collapses and epileptic seizures. Sci Rep 6:23000
125. Olmi S, Petkoski S, Guye M, Bartolomei F, Jirsa VK (2019) Controlling seizure propagation
in large-scale brain networks. PLoS Comp Biol 15:e1006805
References 13

126. Tass PA, Adamchic I, Freund HJ, von Stackelberg T, Hauptmann C (2012) Counteracting
tinnitus by acoustic coordinated reset neuromodulation. Restor Neurol Neurosci 30:137
127. Tass PA, Popovych OV (2012) Unlearning tinnitus-related cerebral synchrony with acoustic
coordinated reset stimulation: theoretical concept and modelling. Biol Cybern 106:27
128. Uhlhaas PJ, Singer W (2006) Neural synchrony in brain disorders: relevance for cognitive
dysfunctions and pathophysiology. Neuron 52:155
129. Buck J, Buck E (1968) Mechanism of rhythmic synchronous flashing of fireflies: fireflies of
southeast asia may use anticipatory time-measuring in synchronizing their flashing. Science
159:1319
130. Rohden M, Sorge A, Timme M, Witthaut D (2012) Self-organized synchronization in decen-
tralized power grids. Phys Rev Lett 109:064101
131. Motter AE, Myers SA, Anghel M, Nishikawa T (2013) Spontaneous synchrony in power-grid
networks. Nat Phys 9:191
132. Menck PJ, Heitzig J, Kurths J, Schellnhuber HJ (2014) How dead ends undermine power grid
stability. Nat Commun 5:3969
133. Schäfer B, Witthaut D, Timme M, Latora V (2018) Dynamically induced cascading failures
in power grids. Nat Commun 9:1975
134. Kuehn C, Throm S (2019) Power network dynamics on graphons. SIAM J Appl Dyn Syst
79:1271
135. Taher H, Olmi S, Schöll E (2019) Enhancing power grid synchronization and stability through
time delayed feedback control. Phys Rev E 100:062306
136. Totz CH, Olmi S, Schöll E (2020) Control of synchronization in two-layer power grids. Phys
Rev E 102:022311
137. Berner R, Yanchuk S, Schöll E (2021) What adaptive neuronal networks teach us about power
grids. Phys Rev E 103:042315
138. Schöll E, Schuster HG (eds) (2008) Handbook of chaos control. Wiley-VCH, Weinheim
(second completely revised and enlarged edition)
139. Schöll E, Klapp SHL, Hövel P (2016) Control of self-organizing nonlinear systems. Springer,
Berlin
140. Kuramoto Y, Battogtokh D (2002) Coexistence of coherence and incoherence in nonlocally
coupled phase oscillators. Nonlin Phen Complex Sys 5:380
141. Abrams DM, Strogatz SH (2004) Chimera states for coupled oscillators. Phys Rev Lett
93:174102
142. Motter AE (2010) Nonlinear dynamics: spontaneous synchrony breaking. Nat Phys 6:164
143. Panaggio MJ, Abrams DM (2015) Chimera states: coexistence of coherence and incoherence
in networks of coupled oscillators. Nonlinearity 28:R67
144. Yao N, Zheng Z (2016) Chimera states in spatiotemporal systems: theory and applications.
Int J Mod Phys B 30:1630002
145. Schöll E (2016) Chimera states and excitation waves in networks with complex topologies.
AIP Conf Proc 1738:210012
146. Omel’chenko OE, Knobloch E (2019) Chimerapedia: coherence-incoherence patterns in one,
two and three dimensions. New J Phys 21:093034
147. Schöll E, Zakharova A, Andrzejak RG (2019) Editorial on the research topic: chimera states
in complex networks. Front Appl Math Stat 5:62. https://doi.org/10.3389/fams.2019.00062
148. Zakharova A (2020) Chimera patterns in networks: interplay between dynamics, structure,
noise, and delay, understanding complex systems. Springer, Berlin
149. Zhang Y, Nicolaou ZG, Hart JD, Roy R, Motter AE (2020) Critical switching in globally
attractive chimeras. Phys Rev X 10:011044
150. Hagerstrom AM, Murphy TE, Roy R, Hövel P, Omelchenko I, Schöll E (2012) Experimental
observation of chimeras in coupled-map lattices. Nat Phys 8:658
151. Tinsley MR, Nkomo S, Showalter K (2012) Chimera and phase cluster states in populations
of coupled chemical oscillators. Nat Phys 8:662
152. Totz J, Snari R, Yengi D, Tinsley MR, Engel H, Showalter K (2015) Phase-lag synchronization
in networks of coupled chemical oscillators. Phys Rev E 92:022819
14 1 Introduction

153. Totz J, Rode J, Tinsley MR, Showalter K, Engel H (2018) Spiral wave chimera states in large
populations of coupled chemical oscillators. Nat Phys 14:282
154. Martens EA, Thutupalli S, Fourriere A, Hallatschek O (2013) Chimera states in mechanical
oscillator networks. Proc Natl Acad Sci USA 110:10563
155. Kapitaniak T, Kuzma P, Wojewoda J, Czolczynski K, Maistrenko Y (2014) Imperfect chimera
states for coupled pendula. Sci Rep 4:6379
156. Olmi S, Martens EA, Thutupalli S, Torcini A (2015) Intermittent chaotic chimeras for coupled
rotators. Phys Rev E 92:030901(R)
157. Pelka K, Peano V, Xuereb A (2020) Chimera states in small optomechanical arrays. Phys Rev
Res 2:013201
158. Larger L, Penkovsky B, Maistrenko Y (2013) Virtual chimera states for delayed-feedback
systems. Phys Rev Lett 111:054103
159. Larger L, Penkovsky B, Maistrenko Y (2015) Laser chimeras as a paradigm for multistable
patterns in complex systems. Nat Commun 6:7752
160. Wickramasinghe M, Kiss IZ (2013) Spatially organized dynamical states in chemical oscil-
lator networks: synchronization, dynamical differentiation, and chimera patterns. PLoS ONE
8:e80586
161. Wickramasinghe M, Kiss IZ (2014) Spatially organized partial synchronization through the
chimera mechanism in a network of electrochemical reactions. Phys Chem Chem Phys
16:18360
162. Schmidt L, Schönleber K, Krischer K, García-Morales V (2014) Coexistence of synchrony
and incoherence in oscillatory media under nonlinear global coupling. Chaos 24:013102
163. Ocampo-Espindola JL, Bick C, Kiss IZ (2019) Weak chimeras in modular electrochemical
oscillator networks. Front Appl Math Stat 5:38
164. Gambuzza LV, Buscarino A, Chessari S, Fortuna L, Meucci R, Frasca M (2014) Experimental
investigation of chimera states with quiescent and synchronous domains in coupled electronic
oscillators. Phys Rev E 90:032905
165. Rosin DP, Rontani D, Haynes N, Schöll E, Gauthier DJ (2014) Transient scaling and resur-
gence of chimera states in coupled Boolean phase oscillators. Phys Rev E 90:030902(R)
166. Rattenborg NC, Amlaner CJ, Lima SL (2000) Behavioral, neurophysiological and evolution-
ary perspectives on unihemispheric sleep. Neurosci Biobehav Rev 24:817
167. Rattenborg NC, Voirin B, Cruz SM, Tisdale R, Dell’Omo G, Lipp HP, Wikelski M, Vyssotski
AL (2016) Evidence that birds sleep in mid-flight. Nat Commun 7:12468
168. Bick C, Martens EA (2015) Controlling chimeras. New J Phys 17:033030
169. Omelchenko I, Omel’chenko OE, Zakharova A, Wolfrum M, Schöll E (2016) Tweezers for
chimeras in small networks. Phys Rev Lett 116:114101
170. Omelchenko I, Omel’chenko OE, Zakharova A, Schöll E (2018) Optimal design of tweezer
control for chimera states. Phys Rev E 97:012216
171. Ruzzene G, Omelchenko I, Schöll E, Zakharova A, Andrzejak RG (2019) Controlling chimera
states via minimal coupling modification. Chaos 29:051103
172. Sawicki J (2019) Delay controlled partial synchronization in complex networks, Springer
Theses. Springer, Heidelberg
173. Pazó D, Deza RR, Pérez-Muñuzuri V (2005) Parity-breaking front bifurcation in bistable
media: link between discrete and continuous versions. Phys Lett A 340:132
174. Zakharova A, Kapeller M, Schöll E (2014) Chimera death: Symmetry breaking in dynamical
networks. Phys Rev Lett 112:154101
175. Maistrenko Y, Penkovsky B, Rosenblum M (2014) Solitary state at the edge of synchrony in
ensembles with attractive and repulsive interactions. Phys Rev E 89:060901
176. Ashwin P, Burylko O (2015) Weak chimeras in minimal networks of coupled phase oscillators.
Chaos 25:013106
177. Klinshov V, Lücken L, Shchapin D, Nekorkin VI, Yanchuk S (2015) Multistable jittering in
oscillators with pulsatile delayed feedback. Phys Rev Lett 114:178103
178. Bi H, Hu X, Boccaletti S, Wang X, Zou Y, Liu Z, Guan S (2016) Coexistence of quantized,
time dependent, clusters in globally coupled oscillators. Phys Rev Lett 117:204101
References 15

179. Wu H, Dhamala M (2018) Dynamics of kuramoto oscillators with time-delayed positive and
negative couplings. Phys Rev E 98:032221
180. Teichmann E, Rosenblum M (2019) Solitary states and partial synchrony in oscillatory ensem-
bles with attractive and repulsive interactions. Chaos 29:093124
181. Chen B, Engelbrecht JR, Mirollo RE (2019) Dynamics of the Kuramoto-Sakaguchi oscillator
network with asymmetric order parameter. Chaos 29:013126
182. Jaros P, Maistrenko Y, Kapitaniak T (2015) Chimera states on the route from coherence to
rotating waves. Phys Rev E 91:022907
183. Jaros P, Brezetsky S, Levchenko R, Dudkowski D, Kapitaniak T, Maistrenko Y (2018) Solitary
states for coupled oscillators with inertia. Chaos 28:011103
184. Kruk N, Maistrenko Y, Koeppl H (2020) Solitary states in the mean-field limit. Chaos
30:111104
185. Berner R, Polanska A, Schöll E, Yanchuk S (2020) Solitary states in adaptive nonlocal oscil-
lator networks. Eur Phys J Spec Top 229:2183
186. Hellmann F, Schultz P, Jaros P, Levchenko R, Kapitaniak T, Kurths J, Maistrenko Y (2020)
Network-induced multistability through lossy coupling and exotic solitary states. Nat Com-
mun 11:592
187. Sathiyadevi K, Chandrasekar VK, Senthilkumar DV, Lakshmanan M (2019) Long-range inter-
action induced collective dynamical behaviors. J Phys A: Math Theor 52:184001
188. Rybalova E, Anishchenko VS, Strelkova GI, Zakharova A (2019) Solitary states and solitary
state chimera in neural networks. Chaos 29:071106
189. Schülen L, Ghosh S, Kachhvah AD, Zakharova A, Jalan S (2019) Delay engineered solitary
states in complex networks. Chaos, Solitons Fractals 128:290
190. Zaks MA, Tomov P (2016) Onset of time dependence in ensembles of excitable elements with
global repulsive coupling. Phys Rev E 93:020201
191. Rybalova E, Semenova N, Strelkova G, Anishchenko V (2017) Transition from complete
synchronization to spatio-temporal chaos in coupled chaotic systems with nonhyperbolic and
hyperbolic attractors. Eur Phys J Spec Top 226:1857
192. Semenov V, Zakharova A, Maistrenko Y, Schöll E (2016) Delayed-feedback chimera states:
forced multiclusters and stochastic resonance. Europhys Lett 115:10005
193. Semenova N, Vadivasova T, Anishchenko V (2018) Mechanism of solitary state appearance
in an ensemble of nonlocally coupled Lozi maps. Eur Phys J Spec Top 227:1173
194. Porter MA (2020) Nonlinearity + Networks: A 2020 Vision, pp 131–159 chapter 6, Springer
International Publishing, ISBN 978-3-030-44992-6
195. Holme P, Saramäki J (2012) Temporal networks. Phys Rep 519:97
196. Holme P (2015) Modern temporal network theory: a colloquium. Eur Phys J B 88:1
197. Gross T, Sayama H (2009) Adaptive networks. Springer, Berlin
198. Hebb D (1949) The organization of behavior: a neuropsychological theory. Wiley, New York,
new edition ed
199. Brown TH, Chapman PF, Kairiss EW, Keenan CL (1988) Long-term synaptic potentiation.
Science 242:724
200. Bliss TVP, Collingridge GL (1993) A synaptic model of memory: long-term potentiation in
the hippocampus. Nature 361:31
201. Gerstner W, Kempter R, von Hemmen JL, Wagner H (1996) A neuronal learning rule for
sub-millisecond temporal coding. Nature 383:76
202. Markram H, Lübke J, Frotscher M, Sakmann B (1997) Regulation of synaptic efficacy by
coincidence of postsynaptic APs and EPSPs. Science 275:213
203. Bi GQ, Poo MM (1998) Synaptic modifications in cultured hippocampal neurons: dependence
on spike timing, synaptic strength, and postsynaptic cell type. J Neurosci 18:10464
204. Abbott LF, Nelson S (2000) Synaptic plasticity: taming the beast. Nat Neurosci 3:1178
205. Bi GQ, Poo MM (2001) Synaptic modification by correlated activity: Hebb’s postulate revis-
ited. Annu Rev Neurosci 24:139
206. Caporale N, Dan Y (2008) Spike timing-dependent plasticity: a Hebbian learning rule. Annu
Rev Neurosci 31:25
16 1 Introduction

207. Meisel C, Gross T (2009) Adaptive self-organization in a realistic neural network model. Phys
Rev E 80:061917
208. Lücken L, Popovych OV, Tass P, Yanchuk S (2016) Noise-enhanced coupling between two
oscillators with long-term plasticity. Phys Rev E 93:032210
209. Jain S, Krishna S (2001) A model for the emergence of cooperation, interdependence, and
structure in evolving networks. Proc Natl Acad Sci 98:543
210. Gross T, D’Lima CJD, Blasius B (2006) Epidemic dynamics on an adaptive network. Phys
Rev Lett 96:208701
211. Gross T, Blasius B (2008) Adaptive coevolutionary networks: a review. J R Soc Interface
5:259
212. Horstmeyer L, Kuehn C (2020) Adaptive voter model on simplicial complexes. Phys Rev E
101:022305
213. Markram H, Gerstner W, Sjöström PJ (2011) A history of spike-timing-dependent plasticity.
Front Synaptic Neurosci 3
214. Mercier E, Wolfersberger D, Sciamanna M (2014) Bifurcation to chaotic low-frequency fluc-
tuations in a laser diode with phase-conjugate feedback. Opt Lett 39:4021
215. Hoppensteadt FC, Izhikevich EM (1996) Synaptic organizations and dynamical properties of
weakly connected neural oscillators ii. learning phase information. Biol Cybern 75:129
216. Seliger P, Young SC, Tsimring LS (2002) Plasticity and learning in a network of coupled
phase oscillators. Phys Rev E 65:041906
217. Câteau H, Kitano K, Fukai T (2008) Interplay between a phase response curve and spike-
timing-dependent plasticity leading to wireless clustering. Phys Rev E 77:051909
218. Miller A, Jin DZ (2013) Potentiation decay of synapses and length distributions of synfire
chains self-organized in recurrent neural networks. Phys Rev E 88:062716
219. Mikkelsen K, Imparato A, Torcini A (2014) Sisyphus effect in pulse-coupled excitatory neural
networks with spike-timing-dependent plasticity. Phys Rev E 89:062701
220. Clopath C, Büsing L, Vasilaki E, Gerstner W (2010) Connectivity reflects coding: a model of
voltage-based STDP with homeostasis. Nat Neurosci 13:344
221. Tass PA, Majtanik M (2006) Long-term anti-kindling effects of desynchronizing brain stim-
ulation: a theoretical study. Biol Cybern 94:58
222. Popovych OV, Yanchuk S, Tass P (2013) Self-organized noise resistance of oscillatory neural
networks with spike timing-dependent plasticity. Sci Rep 3:2926
223. Ito J, Kaneko K (2001) Spontaneous structure formation in a network of chaotic units with
variable connection strengths. Phys Rev Lett 88:028701
224. Ito J, Kaneko K (2003) Spontaneous structure formation in a network of dynamic elements.
Phys Rev E 67:046226
225. Stam CJ, Hillebrand A, Wang H, Van Mieghem P (2010) Emergence of modular structure
in a large-scale brain network with interactions between dynamics and connectivity. Front
Comput Neurosci 4:133
226. Gutiérrez R, Amann A, Assenza S, Gómez-Gardeñes J, Latora V, Boccaletti S (2011) Emerg-
ing meso- and macroscales from synchronization of adaptive networks. Phys Rev Lett
107:234103
227. Assenza S, Gutiérrez R, Gómez-Gardeñes J, Latora V, Boccaletti S (2011) Emergence of
structural patterns out of synchronization in networks with competitive interactions. Sci Rep
1:99
228. Yuan WJ, Zhou C (2011) Interplay between structure and dynamics in adaptive complex
networks: Emergence and amplification of modularity by adaptive dynamics. Phys Rev E
84:016116
229. Aoki T, Aoyagi T (2012) Scale-free structures emerging from co-evolution of a network and
the distribution of a diffusive resource on it. Phys Rev Lett 109:208702
230. Winkler M, Butscher S, Kinzel W (2012) Pulsed chaos synchronization in networks with
adaptive couplings. Phys Rev E 86:016203
231. Aoki T, Yawata K, Aoyagi T (2015) Self-organization of complex networks as a dynamical
system. Phys Rev E 91:012908
References 17

232. Botella-Soler V, Glendinning P (2012) Emergence of hierarchical networks and polysyn-

chronous behaviour in simple adaptive systems. Europhys Lett 97:50004
233. Botella-Soler V, Glendinning P (2014) Hierarchy and polysynchrony in an adaptive network.
Phys Rev E 89:062809
234. Popovych OV, Xenakis MN, Tass PA (2015) The spacing principle for unlearning abnormal
neuronal synchrony. PLoS ONE 10:e0117205
235. Chakravartula S, Indic P, Sundaram B, Killingback T (2017) Emergence of local synchro-
nization in neuronal networks with adaptive couplings. PLoS ONE 12:e0178975
236. Röhr V, Berner R, Lameu EL, Popovych OV, Yanchuk S (2019) Frequency cluster formation
and slow oscillations in neural populations with plasticity. PLoS ONE 14:e0225094
237. Rubinov M, Sporns O, Van Leeuwen C, Breakspear M (2009) Symbiotic relationship between
brain structure and dynamics. BMC Neurosci 10:55
238. Hoppensteadt FC, Izhikevich EM (1999) Oscillatory neurocomputers with dynamic connec-
tivity. Phys Rev Lett 82:2983
239. Du C, Ma W, Chang T, Sheridan P, Lu WD (2015) Biorealistic implementation of synaptic
functions with oxide memristors through internal ionic dynamics. Adv Funct Mater 25:4290
240. John RA, Liu F, Chien NA, Kulkarni MR, Zhu C, Fu QD, Basu A, Liu Z, Mathews N (2018)
Synergistic gating of electro-iono-photoactive 2d chalcogenide neuristors: coexistence of
hebbian and homeostatic synaptic metaplasticity. Adv Mater 30:1800220
241. Schöll E (1987) Nonequilibrium phase transitions in semiconductors. Springer, Berlin
242. Shaw MP, Mitin VV, Schöll E, Grubin HL (1992) The physics of instabilities in solid state
electron devices. Plenum Press, New York
243. Schöll E (2001) Nonlinear spatio-temporal dynamics and chaos in semiconductors, Nonlinear
science series, vol 10. Cambridge University Press, Cambridge
244. Pickett MD, Medeiros-Ribeiro G, Williams RS (2013) A scalable neuristor built with mott
memristors. Nat Mater 12:114
245. Waldrop MM (2013) Neuroelectronics: smart connections. Nature 503:22
246. Ignatov M, Ziegler M, Hansen M, Petraru A, Kohlstedt H (2015) A memristive spiking neuron
with firing rate coding. Front Neurosci 9:376
247. Hansen M, Zahari F, Ziegler M, Kohlstedt H (2017) Double-barrier memristive devices for
unsupervised learning and pattern recognition. Front Neurol Front Neurosci 11:91
248. Birkoben T, Drangmeister M, Zahari F, Yanchuk S, Hövel P, Kohlstedt H (2020) Slow-Fast
Dynamics in a Chaotic System with Strongly Asymmetric Memristive Element. Int J Bifurc
Chaos 30:08, 2050125
249. Schuman CD, Potok TE, Patton RM, Birdwell JD, Dean ME, Rose GS, Plank JS (2017) A
survey of neuromorphic computing and neural networks in hardware. arXiv:1705.06963
250. Markram H (2012) The human brain project. Sci Am 306:50
251. Amunts K, Knoll AC, Lippert T, Pennartz CMA, Ryvlin P, Destexhe A, Jirsa VK, D’Angelo
E, Bjaalie JG (2019) The human brain project-synergy between neuroscience, computing,
informatics, and brain-inspired technologies. PLoS Biol 17:e3000344
252. Koroshetz W, Gordon J, Adams A, Beckel-Mitchener A, Churchill J, Farber G, Freund M,
Gnadt J, Hsu NS, Langhals N, Lisanby S, Liu G, Peng GCY, Steinmetz M, Talley E, White S
(2018) The state of the NIH BRAIN initiative. J Neurosci 38:6427
253. Pais D, Leonard NE (2014) Adaptive network dynamics and evolution of leadership in col-
lective migration. Phys D 267:81
254. Sayama H, Pestov I, Schmidt J, Bush BJ, Wong C, Yamanoi J, Gross T (2013) Modeling
complex systems with adaptive networks. Comput Math Appl 65:1645
255. Sayama H, Sinatra R (2015) Social diffusion and global drift on networks. Phys Rev E
91:032809
256. Aoki T, Rocha LEC, Gross T (2016) Temporal and structural heterogeneities emerging in
adaptive temporal networks. Phys Rev E 93:040301
257. Iwasa M, Tanaka D (2010) Dimensionality of clusters in a swarm oscillator model. Phys Rev
E 81:066214
18 1 Introduction

258. Iwasa M, Iida K, Tanaka D (2010) Hierarchical cluster structures in a one-dimensional swarm
oscillator model. Phys Rev E 81:046220
259. Gavalda A, Duch J, Gómez-Gardeñes J (2012) Reciprocal interactions out of congestion-free
adaptive networks. Phys Rev E 85:026112
260. Zhou C, Kurths J (2006) Dynamical weights and enhanced synchronization in adaptive com-
plex networks. Phys Rev Lett 96:164102
261. Zhu JF, Zhao M, Yu W, Zhou C, Wang BH (2010) Better synchronizability in generalized
adaptive networks. Phys Rev E 81:026201
262. De Lellis P, Bernardo M, Garofalo F (2008) Synchronization of complex networks through
local adaptive coupling. Chaos 18:037110
263. Sorrentino F, Ott E (2008) Adaptive synchronization of dynamics on evolving complex net-
works. Phys Rev Lett 100:114101
264. Wang L, Dai HP, Dong H, Cao YY, Sun YX (2008) Adaptive synchronization of weighted
complex dynamical networks through pinning. Eur Phys J B 61:335
265. De Lellis P, di Bernardo M, Garofalo F (2009) Decentralized adaptive control for synchro-
nization and consensus of complex networks. In: Chiuso A, Fortuna L, Frasca M, Rizzo
A, Schenato L, Zampieri S (eds) Modelling, estimation and control of networked complex
systems. Springer, Berlin, pp 27–42
266. De Lellis P, di Bernardo M, Garofalo F, Porfiri M (2010) Evolution of complex networks via
edge snapping. IEEE Trans Circuits Syst I 57:2132
267. De Lellis P, Bernardo M, Russo G (2010) On QUAD, Lipschitz, and contracting vector fields
for consensus and synchronization of networks. IEEE Trans Circuits Syst I 58:576
268. Schöll E, Selivanov AA, Lehnert J, Dahms T, Hövel P, Fradkov AL (2012) Control of syn-
chronization in delay-coupled networks. Int J Mod Phys B 26:1246007
269. Selivanov AA, Lehnert J, Dahms T, Hövel P, Fradkov AL, Schöll E (2012) Adaptive synchro-
nization in delay-coupled networks of Stuart-Landau oscillators. Phys Rev E 85:016201
270. Guzenko PY, Lehnert J, Schöll E (2013) Application of adaptive methods to chaos control of
networks of Rössler systems. Cybern Phys 2:15
271. Lehnert J, Hövel P, Selivanov AA, Fradkov AL, Schöll E (2014) Controlling cluster synchro-
nization by adapting the topology. Phys Rev E 90:042914
272. Plotnikov SA, Lehnert J, Fradkov AL, Schöll E (2016) Adaptive control of synchronization
in delay-coupled heterogeneous networks of FitzHugh-Nagumo nodes. Int J Bifurc Chaos
26:1650058
273. Acebrón JA, Bonilla LL, Pérez Vicente CJ, Ritort F, Spigler R (2005) The Kuramoto model:
a simple paradigm for synchronization phenomena. Rev Mod Phys 77:137
274. Winfree AT (1980) The geometry of biological time. Springer, New York
275. Hoppensteadt FC, Izhikevich EM (1997) Weakly connected neural networks. Springer, New
York
276. Pietras B, Daffertshofer A (2019) Network dynamics of coupled oscillators and phase reduc-
tion techniques. Phys Rep 819:1
277. Ashwin P, Coombes S, Nicks R (2016) Mathematical frameworks for oscillatory network
dynamics in neuroscience. J Math Neurosci 6(2): 2 (2016)
278. Klinshov V, Yanchuk S, Stephan A, Nekorkin VI (2017) Phase response function for oscillators
with strong forcing or coupling. Europhys Lett 118:50006
279. Rosenblum M, Pikovsky A (2019) Numerical phase reduction beyond the first order approx-
imation. Chaos 29:011105
280. Rosenblum M, Pikovsky A (2019) Nonlinear phase coupling functions: a numerical study.
Philos Trans Royal Soc A 377:20190093
281. Ermentrout GB, Park Y, Wilson D (2019) Recent advances in coupled oscillator theory. Philos
Trans Royal Soc A 377:20190092
282. Strogatz SH, Stewart I (1993) Coupled oscillators and biological synchronization. Sci Am
269:102
283. Strogatz SH, Abraham D, McRobbie AD, Eckhardt B, Ott E (2005) Crowd synchrony on the
millennium bridge. Nature 438:43
References 19

284. Rodrigues FA, Peron TKDM, Ji P, Kurths J (2016) The Kuramoto model in complex networks.
Phys Rep 610:1
285. Watanabe S, Strogatz SH (1993) Integrability of a globally coupled oscillator array. Phys Rev
Lett 70:2391
286. Watanabe S, Strogatz SH (1994) Constants of motion for superconducting Josephson arrays.
Phys D 74:197
287. Stewart I (2011) Phase oscillators with sinusoidal coupling interpreted in terms of projective
geometry. Int J Bifurc Chaos 21:1795
288. Ott E, Antonsen TM (2008) Low dimensional behavior of large systems of globally coupled
oscillators. Chaos 18:037113
289. Omel’chenko OE, Maistrenko Y, Tass P (2008) Chimera states: the natural link between
coherence and incoherence. Phys Rev Lett 100:044105
290. Abrams DM, Mirollo RE, Strogatz SH, Wiley DA (2008) Solvable model for chimera states
of coupled oscillators. Phys Rev Lett 101:084103
291. Laing CR (2009) The dynamics of chimera states in heterogeneous Kuramoto networks. Phys
D 238:1569
292. Marvel SA, Mirollo RE, Strogatz SH (2009) Identical phase oscillators with global sinusoidal
coupling evolve by möbius group action. Chaos 19:043104
293. Pikovsky A, Rosenblum M (2015) Dynamics of globally coupled oscillators: progress and
perspectives. Chaos 25:097616
294. Hancock EJ, Gottwald GA (2018) Model reduction for kuramoto models with complex topolo-
gies. Phys Rev E 98:012307
295. Smith LD, Gottwald GA (2019) Chaos in networks of coupled oscillators with multimodal
natural frequency distributions. Chaos 29:093127
296. Smith LD, Gottwald GA (2020) Chaos in networks of coupled oscillators with multimodal
natural frequency distributions, Chaos 29:093127
297. Gómez-Gardeñes J, Moreno Y, Arenas A (2007) Paths to synchronization on complex net-
works. Phys Rev Lett 98:034101
298. Dörfler F, Bullo F (2014) Synchronization in complex networks of phase oscillators: a survey.
Automatica 50:1539
299. Pazó D (2005) Thermodynamic limit of the first-order phase transition in the kuramoto model.
Phys Rev E 72:046211
300. Gómez-Gardeñes J, Gómez S, Arenas A, Moreno Y (2011) Explosive synchronization tran-
sitions in scale-free networks. Phys Rev Lett 106:128701
301. Boccaletti S, Almendral JA, Guan S, Leyva I, Liu Z, Sendiña-Nadal I, Wang Z, Zou Y
(2016) Explosive transitions in complex networks’ structure and dynamics: percolation and
synchronization. Phys Rep 660
302. Ermentrout GB (1991) An adaptive model for synchrony in the firefly pteroptyx malaccae. J
Math Biol 29:571
303. Filatrella G, Nielsen AH, Pedersen NF (2008) Analysis of a power grid using a Kuramoto-like
model. Eur Phys J B 61:485
304. Schmietendorf K, Peinke J, Friedrich R, Kamps O (2014) Self-organized synchronization and
voltage stability in networks of synchronous machines. Eur Phys J Spec Top 223:2577
305. Olmi S (2015) Chimera states in coupled Kuramoto oscillators with inertia. Chaos 25:123125
306. Taylor D, Ott E, Restrepo JG (2010) Spontaneous synchronization of coupled oscillator sys-
tems with frequency adaptation. Phys Rev E 81:046214
307. Yeung MKS, Strogatz SH (1999) Time delay in the kuramoto model of coupled oscillators.
Phys Rev Lett 82:648
308. Petkoski S, Stefanovska A (2012) Kuramoto model with time-varying parameters. Phys Rev
E 86:046212
309. Maslennikov OV, Nekorkin VI (2017) Adaptive dynamical networks. Phys Usp 60:694
310. Ren Q, Zhao J (2007) Adaptive coupling and enhanced synchronization in coupled phase
oscillators. Phys Rev E 76:016207
20 1 Introduction

311. Maistrenko Y, Lysyansky B, Hauptmann C, Burylko O, Tass PA (2007) Multistability in the

kuramoto model with synaptic plasticity. Phys Rev E 75:066207
312. Masuda N, Kori H (2007) Formation of feedforward networks and frequency synchrony by
spike-timing-dependent plasticity. J Comp Neurosci 22:327
313. Aoki T, Aoyagi T (2009) Co-evolution of phases and connection strengths in a network of
phase oscillators. Phys Rev Lett 102:034101
314. Niyogi RK, English LQ (2009) Learning-rate-dependent clustering and self-development in
a network of coupled phase oscillators. Phys Rev E 80:066213
315. Takahashi YK, Kori H, Masuda N (2009) Self-organization of feed-forward structure and
entrainment in excitatory neural networks with spike-timing-dependent plasticity. Phys Rev
E 79:051904
316. Li M, Guan S, Lai CH (2010) Spontaneous formation of dynamical groups in an adaptive
networked system. New J Phys 12:103032
317. Aoki T, Aoyagi T (2011) Self-organized network of phase oscillators coupled by activity-
dependent interactions. Phys Rev E 84:066109
318. Skardal PS, Taylor D, Restrepo JG (2013) Complex macroscopic behavior in systems of phase
oscillators with adaptive coupling. Phys D 267:27
319. Chandrasekar VK, Sheeba JH, Subash B, Lakshmanan M, Kurths J (2014) Adaptive coupling
induced multi-stable states in complex networks. Phys D 267:36
320. Ren Q, He M, Yu X, Long Q, Zhao J (2014) The adaptive coupling scheme and the hetero-
geneity in intrinsic frequency and degree distributions of the complex networks. Phys Lett A
378:139
321. Timms L, English LQ (2014) Synchronization in phase-coupled Kuramoto oscillator networks
with axonal delay and synaptic plasticity. Phys Rev E 89:032906
322. Aoki T (2015) Self-organization of a recurrent network under ongoing synaptic plasticity.
Neural Netw 62:11
323. Ha SY, Noh SE, Park J (2016) Synchronization of kuramoto oscillators with adaptive cou-
plings. SIAM J Appl Dyn Syst 15:162
324. Kasatkin DV, Nekorkin VI (2016) Dynamics of the phase oscillators with plastic couplings.
Radiophys Quantum Electron 58:877
325. Nekorkin VI, Kasatkin DV (2016) Dynamics of a network of phase oscillators with plastic
couplings. AIP Conf Proc 1738:210010
326. Asl MM, Valizadeh A, Tass PA (2017) Dendritic and axonal propagation delays determine
emergent structures of neuronal networks with plastic synapses. Sci Rep 7:39682
327. Kasatkin DV, Yanchuk S, Schöll E, Nekorkin VI (2017) Self-organized emergence of multi-
layer structure and chimera states in dynamical networks with adaptive couplings. Phys Rev
E 96:062211
328. Asl MM, Valizadeh A, Tass PA (2018) Delay-induced multistability and loop formation in
neuronal networks with spike-timing-dependent plasticity. Sci Rep 8:12068
329. Asl MM, Valizadeh A, Tass PA (2018) Dendritic and axonal propagation delays may shape
neuronal networks with plastic synapses. Front Phys 9:1849
330. Bacic I, Klinshov V, Nekorkin VI, Perc M, Franović I (2018) Inverse stochastic resonance in
a system of excitable active rotators with adaptive coupling. EPL 124:40004
331. Bacic I, Yanchuk S, Wolfrum M, Franović I (2018) Noise-induced switching in two adaptively
coupled excitable systems. Eur Phys J Spec Top 227:1077
332. Kasatkin DV, Nekorkin VI (2018) The effect of topology on organization of synchronous
behavior in dynamical networks with adaptive couplings. Eur Phys J Spec Top 227:1051
333. Karimian M, Dibenedetto D, Moerel M, Burwick T, Westra RL, De Weerd P, Senden M
(2019) Effects of synaptic and myelin plasticity on learning in a network of kuramoto phase
oscillators. Chaos 29:083122
334. Berner R, Schöll E, Yanchuk S (2019) Multiclusters in networks of adaptively coupled phase
oscillators. SIAM J Appl Dyn Syst 18:2227
335. Berner R, Fialkowski J, Kasatkin DV, Nekorkin VI, Yanchuk S, Schöll E (2019) Hierarchical
frequency clusters in adaptive networks of phase oscillators. Chaos 29:103134
References 21

336. Berner R, Vock S, Schöll E, Yanchuk S (2021) Desynchronization transitions in adaptive

networks. Phys Rev Lett 126:028301
337. Feketa P, Schaum A, Meurer T (2019) Synchronization and multi-cluster capabilities of oscil-
latory networks with adaptive coupling. IEEE Trans Autom Control
338. Franović I, Yanchuk S, Eydam S, Bacic I, Wolfrum M (2020) Dynamics of a stochastic
excitable system with slowly adapting feedback. Chaos 30:083109
339. Vock S, Berner R, Yanchuk S, Schöll E (2021) Effect of diluted connectivities on cluster
synchronization of adaptively coupled oscillator networks. arXiv:2101.05601
340. Gleiser PM, Zanette DH (2006) Synchronization and structure in an adaptive oscillator net-
work. Eur Phys J B 53:233
341. Li MH, Guan SG, Lai CH (2011) Formation of modularity in a model of evolving networks.
Europhys Lett 95:58004
342. Scafuti F, Aoki T, di Bernardo M (2015) Heterogeneity induces emergent functional networks
for synchronization. Phys Rev E 91:062913
343. Papadopoulos L, Kim JZ, Kurths J, Bassett DS (2017) Development of structural correla-
tions and synchronization from adaptive rewiring in networks of kuramoto oscillators. Chaos
27:073115
344. Damicelli F, Hilgetag CC, Hütt MT, Messé A (2019) Topological reinforcement as a principle
of modularity emergence in brain networks. Netw Neurosci 3:589
345. Makarov VV, Koronovskii AA, Maksimenko VA, Hramov AE, Moskalenko OI, Buldú JM,
Boccaletti S (2016) Emergence of a multilayer structure in adaptive networks of phase oscil-
lators. Chaos, Solitons Fractals 84:23
346. Kasatkin DV, Nekorkin VI (2018) Synchronization of chimera states in a multiplex system
of phase oscillators with adaptive couplings. Chaos 28:093115
347. Kasatkin DV, Klinshov V, Nekorkin VI (2019) Itinerant chimeras in an adaptive network of
pulse-coupled oscillators. Phys Rev E 99:022203
Chapter 2
Fundamentals of Adaptive and Complex
Dynamical Networks

Throughout this thesis well-established mathematical and physical concepts are used
and extended in order to describe dynamical systems with complex and adaptive cou-
pling structure. In this chapter, we fix the notation and introduce the basic formalism
that is used in the subsequent considerations. For a comprehensive introduction to
nonlinear dynamics and collective phenomena on complex networks, we refer the
reader to the textbooks [1–3].
The chapter is organized as follows. In Sect. 2.1, we provide graph theoretical
preliminaries and introduce all classes of networks studied in this thesis. Furthermore,
a brief introduction into dynamics on complex networks is given in Sect. 2.2 where
we discuss different types of coupling and models. In the Sect. 2.3, we show how
adaptivity is modeled in the context of synaptic plasticity. In addition, the reduction
to phase oscillator models with phase difference-depend plasticity is given.

2.1 Complex Networks

This section is devoted to the introduction of graph theoretical concepts for networks
and the description of special classes of networks that will be considered in this
thesis. Here, we follow standard textbooks and literature on graph theory and complex
networks [4–6].

2.1.1 Networks, Subnetworks, and Connectivity

In network science the notions graph and network, vertex and node, and edge and
link, respectively, are used interchangeably. For the sake of clarity, we use the notions
network, node, and link. In general we define a directed network N as a triple
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 23
R. Berner, Patterns of Synchrony in Complex Networks of Adaptively Coupled Oscillators,
Springer Theses, https://doi.org/10.1007/978-3-030-74938-5_2
24 2 Fundamentals of Adaptive and Complex Dynamical Networks

N = (V, E, ) of finite sets V and E, and  : E → {(v, w) ∈ V × V }. Elements

from the sets V and E are called nodes and links, respectively. The total number of
nodes of a network is denote by N = |V |. Usually the range of the map ψ is restricted
to {(v, w) ∈ V × V : v = w} in order to avoid nodes to be connected to themselves.
However, in this thesis self-coupling is allowed for all further consideration. Note
that for an undirected network  : E → {X ⊆ V : |X | ∈ {1, 2}} which means that
ψ maps the edge e ∈ E to X = v if the mapping ψ(e) refers to self-coupling or it
maps the edge to X = {v, w} if ψ(e) refers to an undirected link between node v
and w (v = w).
By definition, two links can be mapped through  to the same element of V × V .
If this is the case, i.e., there exist e, e ∈ E such that ψ(e) = ψ(e ), the network is
called a multinetwork (multigraph is more commonly used). If on the contrary all
links can be uniquely identified with their images under the map , the network is
called a simple network. This latter type of networks build the backbone of hundreds
of studies on complex networks and forms the basis of the subsequent chapters. Note
that for simple networks with V = {v1 , . . . , v N } every link e can be assigned to its
image (e) which allows for the shorthand notation ei j as the link connecting two
nodes starting at node v j and ending at node vi (i, j = 1, . . . , N ). In this case the
map  can be omitted in the definition of a network. For the sake of simplicity, we
refer to simple networks as networks unless stated differently. Further, we associate
an undirected network N = (V  , E  ) to a directed network N = (V, E) by V  =
V , E  = {ei j : ei j ∈ E or e ji ∈ E, i, j = 1, . . . , N }. Note that it is also possible to
assign a directed network to an undirected network by introducing an orientation of
the links [4].
Another important notion which will be used in this thesis is the induced subnet-
work. A subnetwork of a network N is given by N = (V  , E  , ) if V  ⊆ V and
E  ⊆ E. In addition, a subnetwork is denoted by induced if E  = {ei j ∈ E : vi , v j ∈
V  }.
The connectivity structure for a simple network is often represented by an N × N
matrix A, called adjacency matrix, which possesses the entries

1, if ei j ∈ E,
ai j = (2.1)
0, otherwise.

This algebraic view on networks by representing their structure in form of a matrix

is the one we use throughout this work [4, 6, 7]. This viewpoint allows one to define
the in-degree d(i) of node vi by considering the row sum of the adjacency matrix,
i.e.,


N
d(i) = ai j . (2.2)
j=1

Further, we define the Laplacian matrix L of a network as

2.1 Complex Networks 25
⎛ ⎞
d(1) 0 ··· 0
⎜ . ⎟
⎜ 0 ... ..
. .. ⎟
L=⎜
⎜ .. . .
⎟ − A.
⎟ (2.3)
⎝ . ..
. . 0 ⎠
0 ··· 0 d(N )

Note that this matrix is a discrete version of the well-known Laplacian operator
known from, e.g., the theory of waves. In particular, for certain dynamical networks
such as nonlocally coupled rings with diffusive coupling, as they are defined later
on in Sect. 2.2, the relation between the discrete and the continuum version of the
Laplacian can be shown explicitly [8–10].
A vi -v j walk of length L ∈ N on a network is a sequence vi , ek1 i , vk1 , . . . , e jk L−1 , v j
for vi , v j ∈ V and ekm km−1 ∈ E for all m = 1, . . . , L where k0 = i, and k L = j. With
this, a network is said to be connected (or weakly connected [11]) if for any two nodes
vi , v j ∈ V (vi = v j ) exists at least one vi -v j walk on the associated undirected net-
work. Note that according to this definition, two nodes vi and v j are connected if
either ei j ∈ E or e ji ∈ E. We can utilize the adjacency matrix in order to analyse
whether there is a vi − v j walk of length L between two nodes. For this, we consider
the Lth power of the adjacency matrix for the associated undirected network and
find that there exists a vi − v j walk if (A L ) ji = 0. Therefore, a network is connected
N
if the matrix l=1 Al possesses no vanishing entry. The spectrum of the Laplacian
matrix, also called graph spectrum, allows for another way to find whether a network
is connected. Note that by definition any Laplacian matrix possesses a zero eigen-
value which belongs to the eigenvector (1, . . . , 1)T . Further, for undirected networks
the Laplacian is symmetric and therefore its spectrum consist of eigenvalues such
that 0 = λ1 (L) ≤ λ2 (L) ≤ · · · ≤ λ|V | (L). The second smallest eigenvalues is called
algebraic connectivity (sometimes denoted as Fiedler number) and vanishes if and
only if the network is unconnected [12, 13].
Lastly, for each network we can define a link weighting  : E → R which assigns
real numbers (weights) to each link of the network. According to this map , the
weight matrix κ with entries

(ei j ), if ei j ∈ E,
κi j = (2.4)
0, otherwise.

With regards to dynamical networks, the weight matrix is also called coupling matrix
and its entries coupling weights.
In the following we describe several types of networks with special structural
properties.
26 2 Fundamentals of Adaptive and Complex Dynamical Networks

2.1.2 Special Network Types

2.1.2.1 Globally Coupled Network

The simplest network is the globally coupled network (also complete or all-to-all
network) which consist of N nodes where each node is connected to every other
node. Hence, the entries of the adjacency are given by

1, if i = j,
ai j = (2.5)
0, otherwise,

if there is no self-coupling, otherwise ai j = 1 for all i, j = 1, . . . , N . An illustration

is provided in Fig. 2.1a.
The spectral properties of this type of network are easily determined. The Lapla-
cian for a globally coupled network possesses, besides the single zero eigenvalue,
the N − 1-times degenerate eigenvalue N .

2.1.2.2 Circulant Networks and Nonlocally Coupled Rings

Another important and already well-studied network structure is represented by cir-

culant adjacency matrices, i.e., where each row vector ai = (ai1 , . . . , ai N ) is rotated
one element to the right relative to the preceding row vector [14, 15]. These adjacency
matrices correspond to a ring network where each element has the same coupling
structure. For this type of networks the spectrum can be determined by using the dis-
crete Fourier transformation. In particular, the discrete Fourier vectors (l = 1, . . . , N )

1 2πl 2π(N −1)l T

ζl = √ ei N , . . . , ei N , 1 (2.6)
N

are the N eigenvectors for the N corresponding eigenvalues


N
2πml
λl = a1m ei N . (2.7)
m=1

A particular case of circulant structures are nonlocally coupled ring networks. These
networks are given by

1, if (i − j) mod N ≤ P,
ai j = (2.8)
0, otherwise,

where P denotes the so-called coupling range. Illustrations for these networks are
shown in Fig. 2.1b, c. In Fig. 2.1b, we show a nonlocally coupled ring network for
2.1 Complex Networks 27

(a) (b) (c) (d)

P =3 P =1

Fig. 2.1 Illustration of different networks with a total number of 12 nodes: a globally coupled
network, b nonlocally coupled network (coupling range P = 3), c locally coupled network (coupling
range P = 1), and d multiplex network consisting of two layers for which one layer is globally
coupled and the other layer has a nonlocally coupled network structure

P = 3. Figure 2.1c displays a locally coupled ring network with coupling range
P = 1 where each node is only coupled to its direct neighbor. By changing the
coupling range between P = 1 and P = N /2 (if N is even, else P = (N + 1)/2)
the corresponding network structure range from the locally coupled to the globally
coupled network, respectively.
Another intriguing circulant network structure arises in the context of fractal con-
nectivities. Here, a Cantor construction algorithm has been used to derive iteratively
a vector a from a base pattern [16–22].

2.1.2.3 Multilayer and Multiplex Networks

Multilayer networks are networks where the whole set of nodes is divided into subsets
of nodes which are said to belong together. The induced subnetworks for the individ-
ual subsets are then called layers. From the mathematical point of view, multilayer
networks are just networks. However, the special structure of a multilayer network,
i.e., the partition into several subsets, has recently been considered very important in
order to describe the dynamics on real-world networks. For exhaustive reviews on
multilayer networks and their mathematical description, we refer to Refs. [23–25].
In the following, we discuss some features of multilayer networks by considering
the special class of multiplex networks.
Multiplex networks are a particular kind of multilayer networks. These networks
consist of L ∈ N layers of N nodes where the connectivity structure within the layer
is given by L, generally different, adjacency matrices Aμ (μ = 1, . . . , L). These
adjacency matrices determine the intra-layer networks structure. Let Per(N ) denote
the set of all N × N permutation matrices [26]. Then, the connectivity between
the layers is given by (L 2 − L) matrices I μν ∈ Per(N ) 0̂ where 0̂ is the zero
matrix. Hence, all nodes of one layer are either uniquely identified with all nodes
of another layer via the inter-layer structure I μν or there are no connection between
the layers. An illustration of a multiplex networks with two layers is provided in
Fig. 2.1d. Here, the two layers are bidirectionally connected with I 12 = I 21 = I N ,N .
Within the layers, however, the coupling structure is different. One layer possesses a
globally coupled while the other layer possesses a nonlocally coupled ring intra-layer
structure.
28 2 Fundamentals of Adaptive and Complex Dynamical Networks

An algebraic representation of multilayer networks is achieved by multilinear

forms [23]. By flattening of these tensor structures, however, a representation via
an adjacency matrix is obtained. Here flattening means that one can relate a finite
dimensional tensor space to another finite dimensional vector space via an isomor-
phism, see [25]. In this context the representation is called supra-adjacency matrix
and takes the following block matrix form
⎛ ⎞
A1 I 12 · · · I 1L
⎜ 21 . . .. .. ⎟
⎜I . . . ⎟
A=⎜
⎜ .. . .
⎟.
⎟ (2.9)
⎝ . ..
. . I (L−1)L ⎠
I L1 · · · I L(L−1) A L

Usually the inter-layer structures are assumed to be given by the identity matrix I N ,N .
The spectral properties of multiplex networks have been firstly discussed in [27,
28]. Here, results based on perturbation for weighted multiplex networks in case of
weak and strong inter-layer coupling have been derived. As a part of the present
thesis, novel results on the spectrum of multiplex networks are provided in Chap. 8.

2.1.3 Permutation Symmetries in Networks

Here, we introduce the notion of symmetry for networks. Symmetries are always
described as special transformations leaving certain structures of interest invariant.
For this, in case of networks, permutations of nodes are considered as symmetry
transformation. A permutation of nodes of a network is given by the endomorphism
: V → V, vi → vπ(i) with π ∈ S N and i = 1, . . . , N . By these means, we call
such a transformation a symmetry if the following holds: ei j ∈ E if and only if
eπ(i)π( j) ∈ E. Transformations with this property, also called automorphisms of a
network, form a group, so-called automorphism group or symmetry group, under the
composition [29]. The symmetry group of a network N is denoted by Aut(N).
Symmetries appear in many real-world systems [30]. Utilizing these symmetry
properties the analysis of the spectral properties [31] of complex networks can be
simplified, complex networks can be reduce to simpler structure [32]. Moreover,
certain scenarios in the spatial and temporal dynamics of systems are found to be
induced by network symmetry [33–35].

2.2 Dynamics

In dynamical networks all nodes and links of a weighted network N represent dynam-
ical entities and their weighted interaction with each other, respectively. Formally,
a dynamical network of N nodes equipped with dynamical d-dimensional variables
x i ∈ Cd (i = 1, . . . , N ), as they are analyzed throughout this thesis, may be written
2.2 Dynamics 29

as follows:


N
ẋ i = f i (xi (t)) − σ ai j κi j G i j (x i , x j ), (2.10)
j=1

where f i ∈ C 1 (Cd , Cd ) describes the local dynamics of node i, the functions

G i j ∈ C 1 (Cd , Cd ) determine the pairwise coupling between the nodes i and j. Here,
C 1 (Cd , Cd ) denotes the set of differentiable functions f : Cd → Cd with continuous
derivative. The basic coupling structure is given by the matrix entries ai j ∈ {0, 1} of
the N × N adjacency matrix A. Each existing link is additionally weighted by the
coupling weights κi j . The parameter σ ∈ R is the overall coupling constant.
In the previously described framework of dynamical systems only pairwise inter-
actions are considered. Note that there exist generalizations in order to describe many-
body interaction by using, for instance, simplicial complexes which are topological
generalization of “classical” networks [36–40]. In generalizations using simplicial
complexes, interactions between nodes are not only described by links (1-simplex)
but also by higher n-dimensional n-simplices that can be thought of as the convex
hull of (n + 1) vertices, see [41] for an introduction. However, these models are
beyond the scope of this work.

2.2.1 Types of Coupling

In real world systems, single dynamical units are not isolated but interact by in some
cases very complicated interaction mechanism. For the purpose of describing these
interactions between the units, simplified interaction schemes given by the coupling
function G have been developed. In the following, we introduce the most important
schemes which are used in the subsequent chapters
In the context of diffusive systems on complex networks, linear diffusive pro-
cesses were considered in order to study the dynamics on social as well as transport
networks [27, 42]. Moreover, also in neuronal populations the coupling through
electrical synapses (gap-junctional connections) is modeled by a linear diffusive
map [43–48]. The described form of coupling can be formally written as

G i j (x i , x j ) = G i j (x i − x j ) (2.11)

where the coupling functions becomes d × d matrices G i j . Usually the coupling

scheme is assumed to be identical for all nodes, i.e., G i j = G for all i, j = 1, . . . , N .
In the case of an identical coupling scheme, the input which the ith node receives
via the interaction with the other nodes of the network is given by the sum
s(i) = Nj=1 ai j G(x i − x j ). Here, ai j are the entries of the adjacency matrix of
the network. Remarkably, this sum can be written as s(i) = Nj=1 li j Gx j where li j
are the entries of the Laplacian matrix L, see (2.3). This fact allows for the relation
30 2 Fundamentals of Adaptive and Complex Dynamical Networks

between dynamical properties and the Laplacian structure of a network. It has been
used, for instance, to introduce the powerful methodology of the master stability
function [49, 50] .
Linear coupling is in some sense the simplest type of coupling. More complex
forms of diffusive couplings can be written as

G i j (x i , x j ) = G i j (x i − x j ), (2.12)

where G i j are nonlinear functions depending on the difference x i − x j . For example,

in the well-known Kuramoto model G i j is given by a sin function [51]. Other forms of
nonlinear coupling schemes, not necessarily diffusive, are used to model interactions
in various contexts, see for instance Refs. [52–54]. Note that despite the nonlinear
coupling scheme the corresponding variational equations around each state are still
linear and diffusive. Hence, stability features depend on the Laplacian structure of
the network, as well.
Very elaborate types of coupling schemes are used in models of neurons that
exhibit inhibitory coupling through chemical synapses. Here, the coupling is usually
assumed to be nonlinear with a slow synaptic decay. The coupling through chemical
synapses is modeled by a first order kinetics of the form

G i j (Vi , V j ) = s j · (Vi − Vr )
αs (2.13)
ṡ j = (1 − s j ) − s j /τs
1 + e−(V j − s )/σs

where Vi is the membrane potential of neuron i, Vr is the reverse potential, s j is the

gating variable describing the synaptic transmission from the jth neuron, τs is the
synaptic decay time constant. Further, the parameter αs , s , and σs characterize the
sigmoidal part in the dynamics of the gating variables [55–59]. Depending on which
process is causing the interaction of neurons, other models for chemical coupling
can be found similar to the one described [60, 61].
Building on these different types of coupling, we introduce the dynamical network
models which are used in this thesis.

2.2.2 Kuramoto–Sakaguchi Type Model

A rather simple model, known as Kuramoto–Sakaguchi type model, describes a

dynamical network of N coupled phase oscillators

1 
N
dφi
= ωi − ai j κi j sin(φi − φ j + α), (2.14)
dt N j=1
2.2 Dynamics 31

where φi ∈ T N , T N denoting the N -torus, represents the phase of the ith oscillator
(i = 1, . . . , N ) and ωi its natural frequency. The interaction is the same between each
pair of phase oscillators and given by a sinusoidal coupling kernel. The weighted
network is represented by the entries of the adjacency matrix ai j and the coupling
weights κi j . The parameter α can be considered as a phase-lag of the interaction [62].
System (2.14) has attracted a lot of attention over the last three decades since it is
a first choice paradigmatic model for the modeling of synchronization and partial
synchronization scenarios on complex dynamical networks [51, 63–66].
Coherence (phase synchronization) in a network of phase oscillators can be mea-
sured by the so-called Kuramoto-Daido order parameter [51, 67, 68]. The nth order
parameter for the state φ ∈ T N is defined as

1  inφi
N
Z n (φ) = Rn (φ)eiθn (φ) = e , (2.15)
N j=1

where n ∈ N. The symbols Rn > 0 and θn ∈ T1 denote the modulus and the phase of
the complex order parameter, respectively. The number n can be also called moment
of the order parameter. Note that the symbol i denotes the imaginary unit to distinguish
it from the index i. In case of full phase synchronization, i.e., φ = (a, . . . , a)T for any
a ∈ T , the modulus of the order parameter Rn = 1. More generally, if for any n ∈ N
the phase of the ith oscillator is given by φi ∈ {a, a + 2π/n, a + 4π/n, . . . , a +
2(n − 1)π/n} for some a ∈ T then Rn = 1. Thus, Rn is a measure for a particular
type of coherence with respect to a certain discrete phase distribution. Hence, the
classical Kuramoto order parameter Z 1 measures in-phase synchrony. If Rn = 0 we
consider the phase distribution as incoherent with respect to the nth moment.
A measure that can be used in order to detect frequency synchronization between
two oscillators relies on the average frequency of each phase oscillator

1
i = lim (φi (t0 + T ) − φ(t0 )) (2.16)
T →∞ T

and the frequency synchronization measure is given by


1, if i −  j = 0,
i j = (2.17)
0, otherwise,

see e.g. [69]. Numerically the limit is approximated by very long average window. In
addition, we use a sufficiently small threshold  in order to detect frequency synchro-
nization numerically, i.e., i j = 1 if i −  j <  . In the subsequenty analysis we
used  = 0.001. Using the measure i j , we may define the following value which
we call cluster parameter
32 2 Fundamentals of Adaptive and Complex Dynamical Networks

1 
N
RC = i j . (2.18)
N 2 i, j=1

The cluster parameter measures the following. First, for each frequency cluster,
the total number of links connecting nodes of the same cluster is computed. In
other word, how much space is occupied by a single cluster. Second, all ratios are
summed up and normalized by the number of all possible links. In case of full
frequency synchronization, i.e., i j = 1 for all i, j = 1, . . . , N then RC = 1 because
the single cluster “occupies” the whole network. A similar measure be found in
Ref. [70]. In the context of phase oscillator many different measures were introduced
to characterize the dynamical states. For an overview and a comparison of different
coherence measures for complex networks of phase oscillators see [71]; see also [72]
where a similar generalized order parameter has been suggested independently.

2.2.3 Hodgkin–Huxley Model with Chemical Synapses

A more realistic model for neuronal dynamics is given by the Hodgkin–Huxley

model [73, 74]. Nowadays there exists a zoo of neuron models describing different
features of neural cells. Due to its rich dynamics the Hodgkin–Huxley model is,
however, one of the most prominent neuron models but also one of the most compu-
tationally expensive [75]. In this work we are interested in neurons with tonic spiking
dynamics. These kinds of neurons are excitable neurons which show a periodic spik-
ing pattern in case of a constant synaptic input [75]. A network of N excitatory
Hodgkin–Huxley neurons with inhibitory coupling is described by the following
system [58, 73, 76, 77]

(Vi − Vr ) 
N
C V̇i = Ii − g N a m i3 h i (Vi − E N a ) − gk n i4 (Vi − E K ) − g L (Vi − Vr ) − κi j s j ,
N
j=1

ṁ i = αm (Vi )(1 − m i ) − βm (Vi )m i ,

ḣ i = αh (Vi )(1 − h i ) − βh (Vi )h i ,
ṅ i = αn (Vi )(1 − n i ) − βn (Vi )n i ,
5(1 − si )
ṡi = −Vi +3
− si .
1+e 8

(2.19)
Here Vi is the membrane potential of the ith neuron with the corresponding equilib-
rium potentials E N a = 50 mV, E K = −77 mV, and El = −54.4 mV. C = 1µF/cm2 .
Our choice of the reverse potential Vr = 20 mV corresponds to the excitatory neu-
rons. The three other variables m, h, and n are gating variables in natrium and
potassium, and their dynamics depends on the opening and closing rates
2.2 Dynamics 33

0.1V + 4
αm (V ) = ,
1 − e(−0.1V −4)
−V −65
βm (V ) = 4e( 18 ) ,
( −V20−65 )
αh (V ) = 0.07e ,
1 (2.20)
βh (V ) = ,
1+ e(−0.2V −3.5)
0.01V + 0.55
αn (V ) = ,
1 − e(−0.1V −5.5)
−V −65
βn (V ) = 0.125e( 80 ) .

The parameters are g N a = 120mS/cm2 , g K = 36 mS/cm2 , and gl = 0.3 mS/cm2 . The

constant current Ii is set to 9µA/cm2 so that the individual neurons are identical and
oscillatory. Note that we usually assume no autapses and set κii = 0.

2.3 Adaptive Networks in Neuroscience

A dynamical network is called an adaptive network if its topological structure is

non-static and depends on the state of the dynamical nodes. In this sense, the topo-
logical structure is determined by either the coupling matrix κ(t) or the adjacency
matrix A(t). We note at this point that adaptive networks are to distinguish from
temporal networks which possess a non-static connectivity structure, as well, but
whose temporal dynamics is independent from the state of the nodes [78, 79].
Adaptive networks appear in many real-world systems [80–82]. In the framework
of this thesis, the most important example of this class of dynamical systems are given
by (tonically) spiking neurons with synaptic plasticity. Synaptic plasticity, in neural
systems, is induced by various different interdependent mechanisms [83]. We focus
on the plasticity which is induced by the timing of spikes, so called spike timing-
dependent synaptic plasticity (STDP). In addition, we consider plasticity of chemical
synapses only. Note, however, that plastic changes in the connectivty through elec-
trical synapses have been also reported [84, 85]. Note that in the description of the
large-scale structure of the human brain exist completely other models of plasticity.
These models, for instance, assimilate probabilistic forms of neuronal plasticity and
do not rely on the precise timing of the neuron [74, 86]. The interplay of all forms of
plasticity plays certainly a crucial role in structure formation, however, the analysis
of this interplay is beyond the scope of this thesis. In the following, we briefly intro-
duce the models used in order to describe STDP. For more details on the modeling
and the physiological properties, we refer the reader to the reviews [87–93].
34 2 Fundamentals of Adaptive and Complex Dynamical Networks

2.3.1 Spike Timing-Dependent Plasticity

The synaptic input current from jth presynaptic to ith postsynaptic neuron is scaled
by the synaptic weight κi j , see for example (3.1). The coupling weights κi j have been
shown to be a variable entity. Adaptation of κi j due to the spike timings of neurons
occurs discontinuously whenever one of the neurons i or j spikes. More specifically,
the discontinuous change is formally given by the following plasticity function
⎧
⎪
⎨0, if κi j + δW (ti j ) < 0,
κi j → κi j + W (ti j ), if 0 ≤ κi j + δW (ti j ) ≤ κmax , (2.21)
⎪
⎩
κmax , if κi j + δW (ti j ) > κmax ,

where ti j = ti − t j is the spike time difference between the postsynaptic and presy-
naptic neurons;  > 0 is a small parameter determining the size of the single update;
and κmax > 0 is the maximal coupling [94].
In neuronal systems various forms of spike timing-depending plasticity exist
depending for example on the type of synapse or the type of the postsynaptic neu-
ron [95]. An overview on the diversity of plasticity rules has been presented in [96].
In the subsequent Chap. 3, we particularly consider the plasticity function [94, 97]

−
|ti j | −
|ti j |
W (ti j ) = c p e τp
− cd e τd
(2.22)

with positive parameters c p , τ p , cd , and τd .

2.3.2 Phase Difference-Dependent Plasticity

Models which show self-sustained limit cycle oscillations are often used to describe
excitatory neurons [75, 98, 99]. In the context of these models a neuron is said to
spike, and thus releases an action potential, if the dynamical variable which describes
the membrane potential reaches its maximum.
It is well-known that networks of such oscillators interacting with each other can
be reduced to dynamical networks of weakly coupled phase oscillators [76, 100,
101]. In this framework a neuron is said to spike whenever the neuron’s phase equals
zero. In order to reduce the complexity of neuronal networks with STDP it is desirable
to establish a reduced model of coupled phase oscillators accompanied by synaptic
plasticity which depends rather on the phase than the spike time difference. For this
model of synaptic plasticity, Lücken et al. introduced the notion “phase difference-
dependent plasticity” (PDDP) and drew a relation between STDP and PDDP [77].
Such a PDDP rule is written as

κ̇i j = h(φi − φ j ). (2.23)

2.3 Adaptive Networks in Neuroscience 35

The relation between STDP and PDDP is derived as follows. Consider the spik-
ing time of a postsynaptic tpost and presynaptic tpre . the spiking time difference
t = tpost − tpre is thus always positive or negative (or zero) if the postsynaptic
or the presynaptic neuron spikes (if the neurons spike synchronously), respectively.
Consider now the phase difference θ = φpre − φpost and assume that the phase differ-
ence changes much slower than the phase dynamics. Then by averaging the synaptic
effect induced by a spike of the post and presynaptic neuron and by the assumption
that the chances due to the plasticity are sufficiently small, we get the following
continuous PDDP
    
 θ θ − 2π
h(θ ) = W+ + W−
2π  

where /2π is the number of spike per unit of time, W+ = W (t) for t > 0,
and W− = W (t) for t ≤ 0 (t ≤ 0 if W is discontinuous) [77]. Note that this
adaptation rule does not guarantee that the values of the coupling weights are bounded
during the course of time. This is why they have to be forced to stay within a bounded
interval [κmin , κmax ] by either by a hard-cut as presented in (2.21) or by a soft-cut
given as

2   
κ̇i j = −   κ − κ κi j − (κmax − κmin ) h(φi − φ j ) + κmax κ − κmin κ
(κmax − κmin ) κ − κ
(2.24)

where κ = maxθ∈[0,2π] h(θ ) and κ = minθ∈[0,2π] h(θ ). This soft-cut rule is obtained
by normalization. Another rule using coupling weight dependent scaling of the plas-
ticity function has been used in [102]. In the following, we use a simplified PDDP
in order to introduce a paradigmatic model for adaptively coupled phase oscillators.

2.3.3 A Network of Adaptively Coupled Phase Oscillator

In the following we describe the model which is in the focus of this thesis. By intro-
ducing a PDDP, we introduced a differential equation which governs the dynamics
of the coupling weights, see Eq. (2.23). In addition, the adaptation function h is a
2π -periodic that can expanded in a Fourier series. Consider an adaptive networks
of N adaptively coupled phase oscillators as given by (2.14) with PDDP of the
form (2.24) where h is approximated by the first Fourier mode and we restrict the
coupling weights to [−1, 1]. Then, the dynamical networks reads

1 
N
dφi
=ω− κi j sin(φi − φ j + α), (2.25)
dt N j=1
dκi j  
= − κi j + sin(φi − φ j + β) , (2.26)
dt
36 2 Fundamentals of Adaptive and Complex Dynamical Networks

(a) (b) (c)

Hebbian-like STDP Anti-Hebbian-like

− sin(Δφ)

Δφ/π Δφ/π Δφ/π

Fig. 2.2 The plasticity function − sin(φ + β) and corresponding plasticity rules. a β = − π2 , b
β = 0, c β = π2 . Figure reproduced from [114]. c 2019 AIP Publishing

2
The phase space of the system (φ(t), κ(t)) ∈ T N × [−1, 1] N is (N + N 2 )-
dimensional. System (2.25)–(2.26) have attracted a lot of attention recently [103–
119], since it is a first choice paradigmatic model for the modeling of the dynam-
ics of adaptive networks. In particular, it generalizes the Kuramoto (or Kuramoto–
Sakaguchi) model with fixed κ.
The matrix κ(t) characterizes the coupling topology of the network at time t.
Assume that  is a small but not vanishing parameter. Then, the dynamical equation
(2.26) describes the adaptation of the network topology depending on the dynamics
of the network nodes. The chosen adaptation function in the form sin(φi − φ j + β)
with control parameter β can take into account different plasticity rules that can
occur in neuronal networks, see Fig. 2.2. For instance, for β = −π/2, the Hebbian
rule is obtained where the coupling κi j is increasing between any two systems with
close phases, i.e., φi − φ j close to zero [120–123]. If β = 0 the link κi j will be
strengthened if the j-th oscillator is advancing the i-th. Such a relationship is typical
for spike-timing-dependent plasticity in neuroscience [58, 77, 90, 102].
Let us mention important properties of the model. Firstly, the parameter  1
separates the time scales of the slow dynamical behavior of the coupling strengths
and the fast dynamics of the oscillatory system [124]. Due to the invariance of sys-
tem (2.25)–(2.26) with respect to the phase-shift φi → φi + ψ for all i = 1, . . . , N
and ψ ∈ T1 , the frequency ω can be set to zero in the co-rotating coordinate
frame φ → φ + ωt. Moreover, one can restrict the consideration of the coupling
weights to theinterval −1 ≤ κi j ≤ 1 due to the existence
 of the attracting region
G := φi , κi j : φi ∈ T1 , |κi j | ≤ 1, i, j = 1, . . . , N [111].
Finally, let us mention the parameter symmetries of the model

(α, β, φi , κi j ) → (−α, π − β, −φi , κi j ),

(α, β, φi , κi j ) → (α + π, β + π, φi , −κi j ).

As a result of these symmetries one can restrict the analysis to the parameter region
α ∈ [0, π/2) and β ∈ [−π, π ).
2.4 Summary 37

2.4 Summary

In this chapter, we have given the theoretical framework which is used throughout the
subsequent chapters. References for further reading are provided and the most impor-
tant results for different topics are highlighted. We have first introduced notations
and concepts from graph theory in order to have a proper description of weighted
networks. Several examples for complex network structures are presented and some
of their topological properties are reviewed. The notion of symmetry in networks is
briefly introduced.
After this, we have used the concept of networks to define systems of coupled
dynamical units. We have described how the links of a network relate to the coupling
dynamics of nodes and what kind of coupling functions can be found in physical
applications. Subsequently, a Kuramoto–Sakaguchi-like and the Hodgkin–Huxley
model are introduced. These models are used throughout this thesis. Particularly,
for the phase oscillator model, we have presented measures which allow for the
quantitative description of synchronization in the system.
Extending the concepts from coupled dynamical units on static networks, we have
described how adaptivity is modeled in the context of neuroscience. We have intro-
duced spike timing-dependent plasticity which is used to describe changes of the
synaptic coupling weights dependent on the spike-timing of the neurons. Under the
assumption of topically spiking neuron, neuronal systems with spike-timing depen-
dent plasticity rules can be reduced to phase oscillators models with phase difference-
dependent plasticity. We have shown how this is done and introduced a paradigmatic
model for adaptively coupled phase oscillators with synaptic plasticity.

References

1. Pikovsky A, Rosenblum M, Kurths J (2001) Synchronization: a universal concept in nonlinear

sciences, 1st edn. Cambridge University Press, Cambridge
2. Strogatz SH (2014) Nonlinear dynamics and chaos. CRC Press, Boca Raton
3. Boccaletti S, Pisarchik AN, del Genio CI, Amann A (2018) Synchronization: from coupled
systems to complex networks. Cambridge University Press, Cambridge
4. Godsil C, Royle G (2001) Algebraic graph theory. Springer, New York
5. Costa LDF, Rodrigues FA, Travieso G, Villas Boas PR (2007) Characterization of complex
networks: a survey of measurements. Adv Phys 56:167
6. Newman MEJ (2010) Networks: an introduction. Oxford University Press Inc, New York
7. Cvetkovic D, Rowlinson P, Simic S (1997) Eigenspaces of graphs. Cambridge University
Press, Cambridge
8. Kouvaris NE, Isele TM, Mikhailov AS, Schöll E (2014) Propagation failure of excitation
waves on trees and random networks. Europhys Lett 106:68001
9. Isele TM, Schöll E (2015) Effect of small-world topology on wave propagation on networks
of excitable elements. New J Phys 17:023058
10. Isele TM, Hartung B, Hövel P, Schöll E (2015) Excitation waves on a minimal small-world
model. Eur Phys J B 88:104
11. Korte B, Vygen J (2018) Combinatorial optimization. Springer, Berlin
12. Fiedler M (1973) Algebraic connectivity of graphs. Czech Math J 23:298
38 2 Fundamentals of Adaptive and Complex Dynamical Networks

13. Fiedler M (1989) Laplacian of graphs and algebraic connectivity. Banach Cent Publ 25:57
14. Davis PJ (1979) Circulant matrices. Wiley, Hoboken
15. Gray RM (2005) Toeplitz and circulant matrices: a review. Found Trends Commun Inf Theory
2
16. Hizanidis J, Panagakou E, Omelchenko I, Schöll E, Hövel P, Provata A (2015) Chimera states
in population dynamics: networks with fragmented and hierarchical connectivities. Phys Rev
E 92:012915
17. Omelchenko I, Provata A, Hizanidis J, Schöll E, Hövel P (2015) Robustness of chimera states
for coupled FitzHugh-Nagumo oscillators. Phys Rev E 91:022917
18. Plotnikov SA, Lehnert J, Fradkov AL, Schöll E (2016) Synchronization in heterogeneous
FitzHugh-Nagumo networks with hierarchical architecture. Phys Rev E 94:012203
19. Tsigkri-DeSmedt ND, Hizanidis J, Hövel P, Provata A (2016) Multi-chimera states and tran-
sitions in the leaky integrate-and-fire model with excitatory coupling and hierarchical con-
nectivity. Eur Phys J ST 225:1149
20. Ulonska S, Omelchenko I, Zakharova A, Schöll E (2016) Chimera states in networks of Van
der Pol oscillators with hierarchical connectivities. Chaos 26:094825
21. Krishnagopal S, Lehnert J, Poel W, Zakharova A, Schöll E (2017) Synchronization patterns:
from network motifs to hierarchical networks. Phil Trans R Soc A 375:20160216
22. Sawicki J, Omelchenko I, Zakharova A, Schöll E (2019) Delay-induced chimeras in neural
networks with fractal topology. Eur Phys J B 92:54
23. De Domenico M, Solé-Ribalta A, Cozzo E, Kivelä M, Moreno Y, Porter MA, Gómez S,
Arenas A (2013) Mathematical formulation of multilayer networks. Phys Rev X 3:041022
24. Boccaletti S, Bianconi G, Criado R, del Genio CI, Gómez-Gardeñes J, Romance M, Sendiña
Nadal I, Wang Z, Zanin M (2014) The structure and dynamics of multilayer networks. Phys
Rep 544:1
25. Kivelä M, Arenas A, Barthélemy M, Gleeson JP, Moreno Y, Porter MA (2014) Multilayer
networks. J Complex Netw 2:203
26. Liesen J, Mehrmann V (2015) Linear algebra. Springer, Cham
27. Gómez S, Díaz-Guilera A, Gómez-Gardeñes J, Pérez Vicente CJ, Moreno Y, Arenas A (2013)
Diffusion dynamics on multiplex networks. Phys Rev Lett 110:028701
28. Solé-Ribalta A, De Domenico M, Kouvaris NE, Díaz-Guilera A, Gómez S, Arenas A (2013)
Spectral properties of the laplacian of multiplex networks. Phys Rev E 88:032807
29. Bollobás B (1998) Modern graph theory. Springer, New York
30. MacArthur BD, Sanchez-Garcia RJ, Anderson JW (2008) Symmetry in complex networks.
Discrete Appl Math 156:3525
31. MacArthur BD, Sanchez-Garcia RJ (2009) Spectral characteristics of network redundancy.
Phys Rev E 80:026117
32. Xiao Y, MacArthur BD, Wang H, Xiong M, Wang W (2008) Network quotients: structural
skeletons of complex systems. Phys Rev E 78:046102
33. Fiedler B (1988) Global bifurcation of periodic solutions with symmetry. Springer, Heidelberg
34. Golubitsky M, Stewart I (1988) Singularities and groups in bifurcation theory. Volume 2,
Applied mathematical sciences, vol 69. Springer, New York
35. Ashwin P, Swift JW (1992) The dynamics of n weakly coupled identical oscillators. J Non-
linear Sci 2:69
36. Tanaka T, Aoyagi T (2011) Multistable attractors in a network of phase oscillators with three-
body interactions. Phys Rev Lett 106:224101
37. Ashwin P, Rodrigues A (2016) Hopf normal form with S N symmetry and reduction to systems
of nonlinearly coupled phase oscillators. Phys D 325:14
38. Bick C, Ashwin P, Rodrigues A (2016) Chaos in generically coupled phase oscillator networks
with nonpairwise interactions. Chaos 26:094814
39. Skardal PS, Arenas A (2019) Abrupt desynchronization and extensive multistability in glob-
ally coupled oscillator simplexes. Phys Rev Lett 122:248301
40. León I, Pazó D (2019) Phase reduction beyond the first order: the case of the mean-field
complex Ginzburg-Landau equation. Phys Rev E 100:012211
References 39

41. Giusti C, Ghrist R, Bassett DS (2016) Two’s company, three (or more) is a simplex. J Comp
Neurosci 41:1
42. Barthélemy M (2011) Spatial networks. Phys Rep 499:1
43. Bennett MV, Zukin RS (2004) Electrical coupling and neuronal synchronization in the mam-
malian brain. Neuron 41:495
44. Kopell N, Ermentrout GB (2004) Chemical and electrical synapses perform complementary
roles in the synchronization of interneuronal networks. Proc Natl Acad Sci USA 101:15482
45. Connors BW, Long MA (2004) Electrical synapses in the mammalian brain. Annu Rev Neu-
rosci 27:393
46. Curti S, O’Brien J (2016) Characteristics and plasticity of electrical synaptic transmission.
BMC Cell Biol. 17
47. Reimbayev R, Daley K, Belykh IV (2017) When two wrongs make a right: synchronized
neuronal bursting from combined electrical and inhibitory coupling. Phil Trans R Soc A
375:20160282
48. Alcami P, Pereda AE (2019) Beyond plasticity: the dynamic impact of electrical synapses on
neural circuits. Nat Rev Neurosci 20:263
49. Pecora LM, Carroll TL (1998) Master stability functions for synchronized coupled systems.
Phys Rev Lett 80:2109
50. Lehnert J (2016) Controlling synchronization patterns in complex networks, Springer Theses.
Springer, Heidelberg
51. Kuramoto Y (1984) Chemical oscillations, waves and turbulence. Springer, Berlin
52. Truitt PA, Hertzberg JB, Altunkaya E, Schwab KC (2013) Linear and nonlinear coupling
between transverse modes of a nanomechanical resonator. J Appl Phys 114:114307
53. Penkovsky B, Porte X, Jacquot M, Larger L, Brunner D (2019) Coupled nonlinear delay
systems as deep convolutional neural networks. Phys Rev Lett 123:054101
54. Baumann F, Lorenz-Spreen P, Sokolov IM, Starnini M (2020) Modeling echo chambers and
polarization dynamics in social networks. Phys Rev Lett 124:048301
55. Wang XJ, Rinzel J (1992) Alternating and synchronous rhythms in reciprocally inhibitory
model neurons. Neural Comput 4:84
56. White JA, Chow CC, Ritt J, Soto-Trevino C, Kopell N (1998) Synchronization and oscillatory
dynamics in heterogeneous, mutually inhibited neurons. J Comp Neurosci 5:5
57. Hauptmann C, Tass PA (2009) Cumulative and after-effects of short and weak coordinated
reset stimulation: a modeling study. J Neural Eng 6:016004
58. Popovych OV, Yanchuk S, Tass P (2013) Self-organized noise resistance of oscillatory neural
networks with spike timing-dependent plasticity. Sci Rep 3:2926
59. Popovych OV, Xenakis MN, Tass PA (2015) The spacing principle for unlearning abnormal
neuronal synchrony. PLoS ONE 10:e0117205
60. Rinzel J (1990) Mechanisms for nonuniform propagation along excitable cables. Ann N Y
Acad Sci 591:51
61. Gillies A, Willshaw D (2004) Models of the subthalamic nucleus: the importance of intranu-
clear connectivity. Med Eng Phys 26:723
62. Sakaguchi H, Kuramoto Y (1986) A soluble active rotater model showing phase transitions
via mutual entertainment. Prog Theor Phys 76:576
63. Strogatz SH (2000) From Kuramoto to Crawford: exploring the onset of synchronization in
populations of coupled oscillators. Phys D 143:1
64. Acebrón JA, Bonilla LL, Pérez Vicente CJ, Ritort F, Spigler R (2005) The Kuramoto model:
a simple paradigm for synchronization phenomena. Rev Mod Phys 77:137
65. Omel’chenko OE, Wolfrum M (2012) Nonuniversal transitions to synchrony in the Sakaguchi-
Kuramoto model. Phys Rev Lett 109:164101
66. Pikovsky A, Rosenblum M (2008) Partially integrable dynamics of hierarchical populations
of coupled oscillators. Phys Rev Lett 101:264103
67. Daido H (1992) Order function and macroscopic mutual entrainment in uniformly coupled
limit-cycle oscillators. Prog Theor Phys 88:1213
40 2 Fundamentals of Adaptive and Complex Dynamical Networks

68. Daido H (1994) Generic scaling at the onset of macroscopic mutual entrainment in limit-cycle
oscillators with uniform all-to-all coupling. Phys Rev Lett 73:760
69. Ashwin P, Burylko O (2015) Weak chimeras in minimal networks of coupled phase oscillators.
Chaos 25:013106
70. Kasatkin DV, Nekorkin VI (2018) The effect of topology on organization of synchronous
behavior in dynamical networks with adaptive couplings. Eur Phys J Spec Top 227:1051
71. Schröder M, Timme M, Witthaut D (2017) A universal order parameter for synchrony in
networks of limit cycle oscillators. Chaos 27:073119
72. Mehrmann V, Morandin R, Olmi S, Schöll E (2018) Qualitative stability and synchronicity
analysis of power network models in port-Hamiltonian form. Chaos 28:101102
73. Hodgkin AL, Huxley AF (1952) A quantitative description of membrane current and its
application to conduction and excitation in nerve. J Phys 117:500
74. Gerstner W, Kistler WM, Naud R, Paninski L (2014) Neuronal dynamics: from single neurons
to networks and models of cognition. Cambridge University Press, Cambridge
75. Izhikevich EM (2004) Which model to use for cortical spiking neurons? IEEE Trans Neural
Netw 15:1063
76. Hansel D, Mato G, Meunier C (1993) Phase dynamics for weakly coupled Hodgkin-Huxley
neurons. Europhys Lett 23:367
77. Lücken L, Popovych OV, Tass P, Yanchuk S (2016) Noise-enhanced coupling between two
oscillators with long-term plasticity. Phys Rev E 93:032210
78. Holme P, Saramäki J (2012) Temporal networks. Phys Rep 519:97
79. Holme P (2015) Modern temporal network theory: a colloquium. Eur Phys J B 88:1
80. Gross T, Blasius B (2008) Adaptive coevolutionary networks: a review. J R Soc Interface
5:259
81. Gross T, Sayama H (2009) Adaptive networks. Springer, Berlin
82. Sayama H, Pestov I, Schmidt J, Bush BJ, Wong C, Yamanoi J, Gross T (2013) Modeling
complex systems with adaptive networks. Comput Math Appl 65:1645
83. Feldman DE (2012) The spike-timing dependence of plasticity. Neuron 75:556
84. Landisman CE, Connors BW (2005) Long-term modulation of electrical synapses in the
mammalian thalamus. Science 310:1809
85. Mathy A, Clark BA, Häusser M (2014) Synaptically induced long-term modulation of elec-
trical coupling in the inferior olive. Neuron 81:1290
86. Breakspear M (2017) Dynamic models of large-scale brain activity. Nat Neurosci 20:340
87. Abbott LF, Nelson S (2000) Synaptic plasticity: taming the beast. Nat Neurosci 3:1178
88. Dan Y, Poo MM (2006) Spike timing-dependent plasticity: from synapse to perception. Phys
Rev 86:1033
89. Letzkus JJ, Kampa BM, Stuart GJ (2007) Does spike timing-dependent synaptic plasticity
underlie memory formation? Clin Exp Pharmacol Phys 34:1070
90. Caporale N, Dan Y (2008) Spike timing-dependent plasticity: a Hebbian learning rule. Annu
Rev Neurosci 31:25
91. Sjöström PJ, Rancz EA, Roth A, Häusser M (2008) Dendritic excitability and synaptic plas-
ticity. Phys Rev 88:769
92. Froemke RC, Letzkus JJ, Kampa BM, Hang GB, Stuart GJ (2010) Dendritic synapse location
and neocortical spike-timing-dependent plasticity. Front. Synaptic Neurosci 2
93. Markram H, Gerstner W, Sjöström PJ (2011) A history of spike-timing-dependent plasticity.
Front Synaptic Neurosci 3
94. Gerstner W, Kempter R, von Hemmen JL, Wagner H (1996) A neuronal learning rule for
sub-millisecond temporal coding. Nature 383:76
95. Bi GQ, Poo MM (1998) Synaptic modifications in cultured hippocampal neurons: dependence
on spike timing, synaptic strength, and postsynaptic cell type. J Neurosci 18:10464
96. Shulz DE, Feldman DE (2013) Chapter 9 - spike timing-dependent plasticity. Neural circuit
development and function in the brain. Academic, Cambridge, pp 155–181
97. Bell CC, Han VZ, Sugawara Y, Grant K (1997) Synaptic plasticity in a cerebellum-like
structure depends on temporal order. Nature 387:278
References 41

98. Izhikevich EM (2003) Simple model of spiking neurons. IEEE Trans Neural Netw 14:1569
99. Ashwin P, Coombes S, Nicks R (2016) Mathematical frameworks for oscillatory network
dynamics in neuroscience. J Math Neurosci 6(2):2
100. Hoppensteadt FC, Izhikevich EM (1996) Synaptic organizations and dynamical properties of
weakly connected neural oscillators. Biol Cybern 75:117
101. Pietras B, Daffertshofer A (2019) Network dynamics of coupled oscillators and phase reduc-
tion techniques. Phys Rep 819:1
102. Maistrenko Y, Lysyansky B, Hauptmann C, Burylko O, Tass PA (2007) Multistability in the
Kuramoto model with synaptic plasticity. Phys Rev E 75:066207
103. Ren Q, Zhao J (2007) Adaptive coupling and enhanced synchronization in coupled phase
oscillators. Phys Rev E 76:016207
104. Aoki T, Aoyagi T (2009) Co-evolution of phases and connection strengths in a network of
phase oscillators. Phys Rev Lett 102:034101
105. Aoki T, Aoyagi T (2011) Self-organized network of phase oscillators coupled by activity-
dependent interactions. Phys Rev E 84:066109
106. Picallo CB, Riecke H (2011) Adaptive oscillator networks with conserved overall coupling:
Sequential firing and near-synchronized states. Phys Rev E 83:036206
107. Timms L, English LQ (2014) Synchronization in phase-coupled Kuramoto oscillator networks
with axonal delay and synaptic plasticity. Phys Rev E 89:032906
108. Gushchin A, Mallada E, Tang A (2015) Synchronization of phase-coupled oscillators with
plastic coupling strength. In: Information theory and applications workshop ITA 2015, CA,
USA. IEEE, San Diego, pp 291–300
109. Kasatkin DV, Nekorkin VI (2016) Dynamics of the phase oscillators with plastic couplings.
Radiophys Quantum Electron 58:877
110. Nekorkin VI, Kasatkin DV (2016) Dynamics of a network of phase oscillators with plastic
couplings. AIP Conf Proc 1738:210010
111. Kasatkin DV, Yanchuk S, Schöll E, Nekorkin VI (2017) Self-organized emergence of multi-
layer structure and chimera states in dynamical networks with adaptive couplings. Phys Rev
E 96:062211
112. Avalos-Gaytán V, Almendral JA, Leyva I, Battiston F, Nicosia V, Latora V, Boccaletti S (2018)
Emergent explosive synchronization in adaptive complex networks. Phys Rev E 97:042301
113. Berner R, Schöll E, Yanchuk S (2019) Multiclusters in networks of adaptively coupled phase
oscillators. SIAM J Appl Dyn Syst 18:2227
114. Berner R, Fialkowski J, Kasatkin DV, Nekorkin VI, Yanchuk S, Schöll E (2019) Hierarchical
frequency clusters in adaptive networks of phase oscillators. Chaos 29:103134
115. Berner R, Sawicki J, Schöll E (2020) Birth and stabilization of phase clusters by multiplexing
of adaptive networks. Phys Rev Lett 124:088301
116. Berner R, Vock S, Schöll E, Yanchuk S (2021) Desynchronization transitions in adaptive
networks. Phys Rev Lett 126:028301
117. Berner R, Polanska A, Schöll E, Yanchuk S (2020) Solitary states in adaptive nonlocal oscil-
lator networks. Eur Phys J Spec Top 229:2183
118. Berner R, Yanchuk S, Schöll E (2021) What adaptive neuronal networks teach us about power
grids. Phys Rev E 103:042315
119. Vock S, Berner R, Yanchuk S, Schöll E (2021) Effect of diluted connectivities on cluster
synchronization of adaptively coupled oscillator networks. arXiv:2101.05601
120. Hebb D (1949) The organization of behavior: a neuropsychological theory. Wiley, New York
(new edn)
121. Hoppensteadt FC, Izhikevich EM (1996) Synaptic organizations and dynamical properties of
weakly connected neural oscillators ii. learning phase information. Biol Cybern 75:129
122. Seliger P, Young SC, Tsimring LS (2002) Plasticity and learning in a network of coupled
phase oscillators. Phys Rev E 65:041906
123. Aoki T (2015) Self-organization of a recurrent network under ongoing synaptic plasticity.
Neural Netw 62:11
124. Kuehn C (2015) Multiple time scale dynamics. Springer, Cham
 Part I
Cluster Synchronization in Globally
Coupled Adaptive Networks
Chapter 3
Population of Hodgkin–Huxley Neurons
with Spike Timing-Dependent Plasticity

In this chapter we report on the phenomenon of clustering with respect to connectivity

and frequencies in a network of adaptively coupled Hodgkin–Huxley neurons. The
spike timing-dependent plasticity is considered to be symmetric as it was experimen-
tally found for hippocampal synapses [1] and derived from asymmetric STDP for
an “effective time window” [2]. In this case the observed clusters are bidirectionally
coupled [3]. Splitting of a neural population to a few clusters that are synchronized
at different frequencies could lead to a slow waxing and waning of the amplitude of
the mean field, where the clusters transiently gather together and move apart as the
time evolves [3]. The frequency of such a modulation of the mean neural activity
could be much smaller than the firing rate of individual neurons and depends on the
differences between the clusters’ frequencies. The emergence of synchronized clus-
ters could explain the origin of the low-frequency modulation of the power spectral
density of macroscopic brain signals like local field potentials (LFP) or electroen-
cephalographic (EEG) signals in higher frequency bands, which also correlates with
slow oscillations of the blood oxygen level-dependent (BOLD) signal measured by
fMRI [4–7]. Several other modeling studies have also reported on clustering in the
neural populations with plasticity [3, 8, 9]. These clusters have been observed for
different models that ranged from simple phase oscillators to the models of spiking
and bursting neurons and demonstrate stability with respect to heterogeneity of the
interacting neurons and random perturbations [3, 8, 9]. In this chapter we provide
a simple phenomenological model and explain a mechanism governed by synaptic
plasticity for the stabilization of such clusters in a neural population.
This chapter reproduces contents that have been published in [10]. The chapter is
organized as follows. In the first Sect. 3.1 we present the model. The next Sect. 3.2
shows numerically observed multiclusters. The detailed mechanisms of the stability
of frequency clusters is explained afterwards in Sect. 3.3 using the simplest case
of two clusters. Then, in Sect. 3.4, we propose a phenomenological model which
describes the dynamics of two clusters taking into account the adaptation of the
weights. The model is shown to reflect not only qualitative, but also some basic
quantitative properties of the two-cluster formation. We also determine the set of

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 45

R. Berner, Patterns of Synchrony in Complex Networks of Adaptively Coupled Oscillators,
Springer Theses, https://doi.org/10.1007/978-3-030-74938-5_3
46 3 Population of Hodgkin–Huxley Neurons with Spike …

plasticity functions (update rules), which lead to the clustering. The findings of this
chapter are summarized in Sect. 3.5.

3.1 Coupled Hodgkin–Huxley Neurons on a Network with

Spike Timing-Dependent Plasticity

For the present study we consider N coupled Hodgkin–Huxley neurons which are
described by the dynamical equations

(Vi − Vr ) 
N
C V̇i = Ii − g N a m i3 h i (Vi − E N a ) − gk n i4 (Vi − E K ) − g L (Vi − Vr ) − κi j s j ,
N
j=1
ṁ i = αm (Vi )(1 − m i ) − βm (Vi )m i ,
ḣ i = αh (Vi )(1 − h i ) − βh (Vi )h i ,
ṅ i = αn (Vi )(1 − n i ) − βn (Vi )n i ,
5(1 − si )
ṡi = 
−Vi +3
 − si .
1+e 8
(3.1)
Each of the neurons (i = 1, . . . , N ) is described by the membrane potential Vi and
the three gating variables m i , h i , and n i . The parameters are chosen such that each
individual neuron gives rise to self-sustained oscillatory dynamics. For more details,
we refer to (2.19)–(2.20) in Chap. 2.
The interaction between the neurons is given by the synaptic input current s j which
each neuron receives from the jth neuron. the input current is scaled by the synaptic
strength κi j , which we assume to change due to plasticity. The adaptation of κi j
occurs discontinuously whenever one of the neurons i or j spikes. More specifically,
the discontinuous change is given by the following plasticity function, see also (2.21)
⎧
⎪
⎨0, if κi j + δW (ti j ) < 0
κi j → κi j + δW (ti j ), if 0 ≤ κi j + δW (ti j ) ≤ κmax (3.2)
⎪
⎩
κmax , if κi j + δW (ti j ) > κmax

where ti j = ti − t j is the spike time difference between the postsynaptic and presy-
naptic neurons; δ > 0 is a small parameter determining the size of the single update;
κmax > 0 is the maximal coupling; and the plasticity function [11–14] is
|ti j | |ti j |
− −
W (ti j ) = c p e τp
− cd e τd
(3.3)

with positive parameters c p , τ p , cd , and τd . We also assume no autapses and set

κii = 0.
3.1 Coupled Hodgkin–Huxley Neurons on a Network with Spike … 47

Fig. 3.1 Plasticity function

W (ti j ) for τ p = 2, τd = 5,
c p = 2, cd = 1.6. Figure re-
produced from [10]

An example of the considered plasticity function W used in our simulations is

shown in Fig. 3.1. This is a symmetric function, which corresponds to a potentiation
of the coupling weights of the neurons with highly correlated firing. As we will
discuss at the end of this chapter, there is a family of plasticity functions of similar
form that allow for the frequency clustering.

3.2 Numerical Observation of Synchrony and Frequency

Clustering

In order to investigate the dynamics of the network (3.1) with plasticity (3.3), we
initialize the neurons and the coupling randomly and integrate the system numeri-
cally. For the parameter values τ p = 2, τd = 5, c p = 2, cd = 1.6, and κmax = 1.5 we
observe two phenomena: complete synchronization and the emergence of frequency
clusters hierarchical in size, see Figs. 3.2 and 3.3, respectively.
Figure 3.2a shows the initial coupling weights, and Fig. 3.2c illustrates the spike
times of the neurons at the beginning of the simulation. One can observe that while the
neurons start with an incoherent spiking, they enhance the coherence already after a
few spikes due to the interaction between them as well as the plasticity. The plasticity
potentiates the connections of neurons that fire together. A complete synchronization
is established and the coupling weights increase to κmax after a transient (Fig. 3.2b,
d). In a completely synchronized state, the individual neurons spike simultaneously,
hence, the spike time differences ti j = 0.
The emergence of frequency clusters is shown in Fig. 3.3 for two clusters. The
system in Fig. 3.3 possesses the same parameters as in Fig. 3.2, and the difference is
just another realization of random initial conditions. In contrast to the synchronized
state, the final state shown in Fig. 3.3f consist of two groups of synchronized neurons.
These cluster states also manifest themselves as two groups of strongly coupled
elements in the coupling matrix κ (Fig. 3.3c). The coupling weights between the
neurons from the different groups is very small or zero.
We observe that the largest cluster is formed rather quickly as time evolves,
whereas the formation of the small cluster takes much more time. This is illustrated
48 3 Population of Hodgkin–Huxley Neurons with Spike …

Fig. 3.2 Evolution of the coupling matrix κi j (t) starting from random initial conditions and con-
verging to a completely synchronous state. Panel a shows initial coupling matrix, b the coupling
matrix after the transient t = 2000 ms. Raster plot of spiking times at the beginning of simulations
(c) and after the transient (d). The asymptotic state (b, d) is a completely synchronized spiking with
all coupling weights κi j potentiated to kmax . Other parameters N = 200, τ p = 2, τd = 5, c p = 2,
cd = 1.6, and κmax = 1.5. Figure reproduced from [10]

in Fig. 3.4, where the time courses of the mean coupling within each of the two clusters
are shown. The average coupling within the big cluster reaches its maximum fast
(at t ≈ 1000, solid curve in Fig. 3.4), whereas the smaller cluster in Fig. 3.3c, f is
formed through the merging of transient clusters and finally establishes at t ≈ 17000
(dashed curve in Fig. 3.4).
For the states with more clusters, each new formed cluster is significantly smaller
than the previous one, see Fig. 3.5, where three clusters are shown. The spiking period
of the cluster appears to be proportional to its size: the bigger the cluster the larger is
the period. Simulation of the cases with even more clusters becomes computationally
expensive due to large transients.

3.2.1 Clustering with Independent Random Input

To investigate the robustness of our findings, we added an α-train as additional

independent random input to the membrane potential Vi of every neuron:
3.2 Numerical Observation of Synchrony and Frequency Clustering 49

Fig. 3.3 Evolution of the coupling matrix κi j (t) starting from random initial conditions and con-
verging to frequency clusters hierarchical in size. Panel a shows initial coupling matrix, b the
coupling matrix after the transient t = 5600 ms, and c t = 20000 ms. d–f Corresponding raster
plots of spike times. The asymptotic state (c, f) is a hierarchical cluster state with the coupling
weights κi j potentiated to kmax within each cluster and small or zero otherwise. Other parameters
as in Fig. 3.2. The oscillators are ordered accordingly to their mean frequency. Figure reproduced
from [10]

Fig. 3.4 Formation of

individual clusters over time
(corresponds to the
dynamical scenario in
Fig. 3.3). The dashed and
solid curves depict the time
course of the mean coupling
within the small and big
clusters, respectively. Figure
reproduced from [10]


α(t − τi,k )e−α(t−τi,k ) )
input
Ii (t) = I (Vr − Vi (t)) (3.4)
τi,k <t

The Eq. (3.4) models a postsynaptic potential (PSP) that is received by the neuron
at certain random times τi,k . The inter-spike interval is Gaussian distributed τi,k+1 −
τi,k = τi,k ∼ N(14 ms, 4 ms). α is set to 24/τi,k .
50 3 Population of Hodgkin–Huxley Neurons with Spike …

Fig. 3.5 Example of a three-cluster state for N = 500, τ p = 2, τd = 5, c p = 2, cd = 1.6, and

κmax = 1.5 with a random initial distribution of κi j in [0, 0.75]. Figure reproduced from [10]

Fig. 3.6 Coupling matrices for t = 10000 ms and different amplitudes of independent random
input I (see Eq. (3.4)). a I = 0.005, b I = 0.01, c I = 0.02, d I = 0.05 and e I=0.07. All other
parameters as in Fig. 3.5. Figure reproduced from [10]

The numerical simulations Fig. 3.6 show that the clustering is still observed under
input
the influence of random input Ii (t) of intensity I . More specifically, for suffi-
ciently weak perturbations with I < 0.01, all three clusters survive (Fig. 3.6a). With
increasing the amplitude I , the clusters start to decouple. The smaller clusters are
affected first (Fig. 3.6b–d), they start desynchronizing at I = 0.01. The biggest clus-
ter keeps shrinking in size while I is increased and finally for I = 0.07 the whole
network decouples (Fig. 3.6e).

3.3 Emergence of Two-Cluster States

In this section we numerically show that depending on the relative size of the two
clusters, such two-cluster states can be either dynamically stable or transient leading
to complete synchronization. In order to investigate the cluster stability, we initialize
the system in a two cluster state with the number of neurons Ns in the small cluster
and Nb = N − Ns in the big cluster. The total number of neurons is set to N = 50.
The inter-cluster couplings are set to zero initially while the intra-cluster couplings
3.3 Emergence of Two-Cluster States 51

Fig. 3.7 a Difference between synchronization frequencies of the two clusters for different size of
the smaller cluster Ns . b Time until cluster fusion for different initial size of the smaller cluster Ns .
Figure reproduced from [10]

equals κmax . All neurons in the same cluster are initialized with the same initial
conditions, so the clusters are fully synchronized at t = 0.
Figure 3.7a shows frequency difference of two uncoupled clusters as a function of
the size of the small cluster. The frequency difference demonstrates an almost linear
dependence on the cluster size and decays as the size of the smaller cluster increases.
Moreover, we also observe that clusters with sufficiently different sizes are stable
while the clusters of similar sizes, in the considered case with Ns > 8, are transient,
merge into a single cluster and eventually lead to a stable completely synchronous
state, see Fig. 3.8.
Although the threshold of how different the clusters should be in order to be stable
is certainly model dependent, the synchronization of similar clusters is a general
property. The merging of two clusters can be explained qualitatively as follows.
Initially uncoupled clusters evolve each with their natural frequencies s and b . If
their sizes are different, then s = b , and the clusters arrive in-phase periodically
with the “beating” frequency  = b − s . As soon as such an in-phase episode
occurs, the interspike intervals ti j between any two neurons from different clusters
become small and, hence, due to the plasticity rule W (see Eq. (3.2) and Fig. 3.1) the
inter-cluster coupling weights increase. Moreover, the duration of such an in-phase
episode depends on the frequency difference between the clusters. As a result, for a
small frequency difference , the time interval where the clusters are practically in-
phase is sufficiently long in order to potentiate the coupling weights to their maximum
value. This unites the two clusters into one. In contrast, for large , such an
episode is short, and the inter-cluster coupling remains small, which keeps the clusters
oscillating at different frequencies in a stable manner.
Figure 3.8a–c displays an example of two-cluster stable state with Ns = 8. Start-
ing from the two-cluster state, after t = 1500 ms, the coupling between the clusters
increases, see Fig. 3.8b due to the “in-phase episode” when the clusters are syn-
chronous. Afterwards, however, the inter-cluster coupling weights return to their
initial configuration (Fig. 3.8c), since the spike time differences for neurons from
52 3 Population of Hodgkin–Huxley Neurons with Spike …

Fig. 3.8 Evolution of the coupling matrix for N = 50 and the number of neurons Ns = 8 (a–
c) and Ns = 9 (d–f) in the small cluster. In panels (a–c) the clusters are stable, while in (d–f)
they are merging to one synchronous cluster. (g, h) Time courses of the spiking synchronization
frequencies of small (Ns neurons) and large (Nb neurons) clusters depicted by dashed and solid
curves, respectively, for g Ns = 8 and Nb = 42 and h Ns = 9 and Nb = 41. Parameter κmax = 1.0.
Figure reproduced from [10]

different clusters are again far enough apart to cause the depression of the inter-
cluster synapses. Such a process is repeated every time the clusters meet and is
typical for the stable cluster states. A typical case of transient clusters is presented
in Fig. 3.8d–f for Ns = 9. The inter-cluster coupling is again potentiated when the
clusters meet, but it does not decrease again, and the clusters merge in a single clus-
ter of a fully coupled and synchronized regime (Fig. 3.8f). The transient time that
could be elapsed until the cluster fusion depends on the cluster size as illustrated in
Fig. 3.7b.
3.3 Emergence of Two-Cluster States 53

Fig. 3.9 Mean synaptic activity S(t) of the neural population in the case of stable two cluster state.
Panel a shows the dynamics of S(t) on the time interval of 12 s, where modulation of the amplitude
(blue line) is visible, while the fast oscillations are not recognized on this timescale. The maximum
amplitude corresponds to the two clusters being synchronised, while the low amplitude corresponds
to the clusters being out of phase. Panel b shows the zoom of a small time interval. The modulation
takes place on the timescale which is two orders of magnitude larger than the individual spikes of
S(t) as well as individual neural spikes in both clusters. Cluster frequencies ω1 = 0.065012 kHz
and ω2 = 0.065416 kHz. The corresponding period of modulation is T ≈ 2.5 s. Figure reproduced
from [10]

Figures 3.8g, h show how the spiking frequency of the clusters change over time.
During the in-phase episode, the cluster with the higher natural spiking rate slows
down significantly, while the slower cluster (with larger number of neurons Nb )
speeds up a little. For a stable cluster state the cluster frequencies again deviate from
each other (Fig. 3.8g), whereas all neurons fire with the same frequency when the
clusters unite into one (Fig. 3.8h). We found this phenomenon for different numbers
of neurons and different κmax . Increasing κmax increases the initial period difference,
but the behavior in general stays the same.
N
Figure 3.9 shows the dynamics of the mean synaptic activity S(t) = N1 i=1 si (t)
of the network in the case of two stable clusters, which models the dynamics of
LFP. During the in-phase episodes of the two clusters, S(t) has a higher amplitude,
because both clusters spike synchronously. The maximum amplitude is generated by
maximum synchronization in the network. The low amplitude of S(t), on the other
hand, corresponds to the time intervals when the clusters are out of phase. In the latter
case, the mean synaptic activity shows two peaks, the higher peak is generated by
the larger cluster and the lower by the smaller one, see Fig. 3.9b. For the considered
case, the synchronized oscillations of individual neurons in the clusters take place at
a time scale of several milliseconds (period ∼ 15 ms, Fig. 3.9b), see also Fig. 3.8g, h.
The neurons are tonically spiking. The frequency difference  between clusters is,
however, of the order of sub-Hz, because the corresponding cluster frequencies are
close to each other (Fig. 3.8g,h). Then the modulation of S(t) is observed at a much
slower timescale of a few seconds, which is of two orders of magnitude slower than
54 3 Population of Hodgkin–Huxley Neurons with Spike …

the intrinsic neural firing, see Fig. 3.9a, as observed in empirical data of the brain
activity [4, 5].
In the following section, a phenomenological model is introduced in order to
further investigate the dynamics of two clusters.

3.4 Phenomenological Model with Phase

Difference-Dependent Plasticity

In this section we introduce a reduced qualitative model for the coupling and phase
difference of two clusters. The model is based on the assumption that oscillators are
synchronized identically within each cluster and the coupling between the clusters
is weak. As a result, the interaction between oscillatory clusters can be described in
the framework of two coupled phase oscillators that are interacting via their phase
differences [15–18]

ϕ̇1 = ω1 − F1 (ϕ1 − ϕ2 ), (3.5)

ϕ̇2 = ω2 − F2 (ϕ2 − ϕ1 ), (3.6)

where ω1 and ω2 are the natural frequencies of the individual clusters, F1 and F2
are effective interaction functions. For the phase difference ϕ = ϕ1 − ϕ2 , this system
reads
ϕ̇ = ω − F(ϕ) (3.7)

where ω = ω1 − ω2 is the difference of the natural frequencies, and F(ϕ) = F1 (ϕ) −

F2 (−ϕ).
Since the clusters are synchronized for a sufficiently small frequency mismatch
ω, the periodic interaction function F(ϕ) must satisfy F(0) = 0 and F (0) > 0.
The latter means that there is a stable equilibrium ϕ = 0 for small ω. Aiming at a
qualitative insight, we further simplify the model by assuming that F(ϕ) = σ sin(ϕ +
α), where sin ϕ can be viewed as a first Fourier harmonic of the interaction function
and σ as an effective coupling weight. The parameter α = sin−1 (ω/σmax ) is a constant
phase shift assuring that the phase difference of the synchronized cluster is zero. In
fact, for small ω, this parameter is also small and it does not play important role in
the qualitative behavior of the model apart from a small shift of the synchronized
state to ϕ = 0.
Another component of the model is the plasticity-driven changes of the coupling
σ . In order to derive the equation for σ , we consider the STDP update in the case
of a periodic motion of the clusters. We assume that the coupling σ is proportional
to an averaged coupling between the clusters. This is a natural assumption in the
case of weakly coupled systems. Let us find out how the update of the intercluster
coupling depends on the phase difference ϕ. For a given phase difference ϕ and the
frequencies ω1 = ω̄ + ω/2, ω2 = ω̄ − ω/2 (here we introduced the mean frequency
3.4 Phenomenological Model with Phase Difference-Dependent Plasticity 55

Fig. 3.10 a Update function G(ϕ) for τ p = 2, τd = 5, c p = 2, and cd = 1.6. b Schematic spiking
of two oscillators with spike time difference T and periods close to T . Figure reproduced from [10]

ω̄), the spiking period of the both clusters can be approximated as T ≈ 2π/ω̄ up to
small terms of order ω, and the distance T between the spikes of two clusters

ϕ1 ϕ2 ϕ ϕ mod 2π
T = T −T mod T = T mod T ≈ .
2π 2π 2π ω̄
Since the spike time differences T and T − T occur recursively, see Fig. 3.10b,
the updates per unit time sum to the function

δ δ ω̄
(W (T − T ) + W (T )) = G(ϕ), (3.8)
T 2π
where
2π − (ϕ mod 2π ) ϕ mod 2π
G(ϕ) := W +W . (3.9)
ω̄ ω̄

Since the update of σ is proportional to the obtained function, and taking into account
the smallness of δ, this update can be written as σ̇ = εG(ϕ), where ε is a small
parameter of the coupling adaptation that controls the scale separation between the
fast dynamics of the clusters and the slow dynamics of the coupling.
Additionally, the coupling strength σ (t) should be bounded to the interval [0, σmax ]
by imposing cut-off conditions. More specifically, the derivative σ̇ is discontinuous
at the boundaries σ = 0 and σ = σmax , i.e. σ̇ = max{0, εG(ϕ)} for σ = 0 and σ̇ =
min{0, εG(ϕ)} for σ = σmax . The considered cut-off corresponds to “hard” bound
conditions [19]. Another possibility would be “soft” or “multiplicative” bounds [20],
when the update is proportional to the distance to the boundary. We consider here the
hard bound, since it corresponds to the hard bound of the STDP rule for Hodgkin–
Huxley system.
The final phenomenological model reads as follows
56 3 Population of Hodgkin–Huxley Neurons with Spike …

Fig. 3.11 Phase portraits of model (3.10)–(3.11) for a monostable regime of complete synchro-
nization; b co-existence of stable synchronized and clustered states; and c bifurcation moment
of transition between the phase portraits illustrated in (a) and (b). The basins of attraction of the
synchronized regime (point S), clustered state (limit cycle indicated by thick black curve) and the
saddle fixed point (ϕ ∗ , σ ∗ ) are depicted by gray, blue, and white colors, respectively. The nullclines
of the system and stable and unstable manifolds of the saddle point are indicted by the thin gray and
black curves, respectively. Parameters a ω = 0.037 kHz, b ω = 0.06 kHz, c ω ≈ 0.455 Hz, and the
other parameters τ p = 2, τd = 5, c p = 2, cd = 1.6, and ε = 0.08. Figure reproduced from [10]

ϕ̇ = ω − σ sin(ϕ + α), (3.10)

⎧
⎪
⎨G(ϕ) for 0 < σ < σmax ,
σ̇ = ε · max{0, G(ϕ)} for σ = 0, (3.11)
⎪
⎩
min{0, G(ϕ)} for σ = σmax .

with frequency mismatch ω > 0 and α = sin−1 (ω/σmax ).

3.4.1 Properties of the Model

Phase space of system (3.10)–(3.11) is two dimensional with (ϕ, σ ) ∈ S 1 × [0, σmax ].
The nullclines are given by G(ϕ) = 0 for σ̇ = 0 and σ = ω/sin(ϕ + α) for ϕ̇ = 0
in the internal points of the phase space. For the parameter values as in Fig. 3.11,
the ϕ-nullcline corresponds to the two lines ϕ = ϕ ∗ ≈ 0.23 and ϕ = −ϕ ∗ , while the
σ -nullcline to a U-shaped nonlinear curve (grey lines in Fig. 3.11).
There is just one fixed point (ϕ ∗ , σ ∗ = ω/ sin ϕ ∗ ) of saddle-type within the region
σ ∈ (0, σmax ). This point is given by the intersection of the nullclines. Figure 3.11
shows this fixed point and its stable and unstable separatrixes (black lines). An
additional fixed point as well as periodic attractor emerge in system (3.10)–(3.11)
due to the non-smoothness at the boundaries. More specifically, three situations are
observed:
(I): One globally stable fixed point S = (0, σmax ) which corresponds to the fusion
of the two clusters into one. The coupling σ = σmax and the phase difference is zero
at the fixed point, see Fig. 3.11. All orbits are approaching this stable fixed point with
time. This corresponding phase portrait is shown in Fig. 3.11a.
3.4 Phenomenological Model with Phase Difference-Dependent Plasticity 57

(II): Coexistence of the stable fixed point S = (0, σmax ) and a stable periodic
orbit, see Fig. 3.11b. As in the case (I), the fixed point corresponds to the merging of
two clusters. The periodic orbit corresponds to two simultaneously existing clusters.
The clusters possess different frequencies and, as a result, the phase difference is not
bounded and rotate along the circular direction ϕ. Part of the periodic orbit is located
on the boundary σ = 0, i.e. vanishing inter-cluster coupling. The coupling σ (t)
increases between −ϕ ∗ and ϕ ∗ and decreases otherwise. In fact, one can parameterise
the coupling σ by the phase ϕ on the periodic attractor. In the case when σ (ϕ ∗ ) < σ ∗ ,
the solution returns to the boundary σ = 0, moves along it till the orbit reaches the
point (−ϕ ∗ , 0), and the periodic motion repeats.
(III): When maxϕ G(ϕ) < 0 then there exists globally stable periodic solution
ϕ = ωt + ϕ0 , σ = 0. In such a case, the fixed point on the boundary disappears.
Formally, this corresponds to an uncoupling between the clusters. However, in the
original Hodgkin–Huxley system, this parameter regime corresponds to complete
uncoupling of all oscillators because of the depression of all synapses.
In fact, the parameter boundary between the cases (I) and (II) is determined by the
condition σ (ϕ ∗ ) = σ ∗ , which can be interpreted geometrically as hitting the point
(−ϕ ∗ , 0) by the stable manifold of the saddle equilibrium point, see Fig. 3.11c. In this
special case, the saddle equilibrium attracts the whole set of points from the phase
space that is below the stable manifolds, see white area in Fig. 3.11c. In case (II), the
separation between the basins of attraction of the fixed point and the periodic orbit
are given by the saddle equilibrium and its stable manifolds. A sufficient condition
for the case (III) is given by cd ≥ c p and τd ≥ τ p . Under these conditions G(ϕ) ≤ 0
for all ϕ.
Summarizing, the case (II) corresponds to the situation when clusters are stable
and do not merge into one. For this, initial conditions must belong to the basin of
attraction of the periodic solution (Fig. 3.11b, blue domain). The analysis of the
phenomenological model indicates that the cluster case always coexists with stable
complete synchronization.

3.4.2 Comparison of the Model and Cluster Dynamics in

Hodgkin–Huxley Network

In order to compare dynamics of the phenomenological model (3.10)–(3.11) and the

original system (3.1)–(3.2), we ran a series of simulations of the Hodgkin–Huxley
network for parameter values that allow for a stable two-cluster solution. The phases
of the clusters are calculated as ϕ1,2 (t) = 2π tk+1 t−tk
−tk
+ 2π k for t ∈ [tk , tk+1 ), where
{t1 , ..., tn , ...} are spiking times with tk < tk+1 [16]. Correspondingly, the phase dif-
ference is ϕ H H (t) = ϕ1 (t) − ϕ2 (t). The coupling measure σ H H is given by the mean
inter-cluster coupling.
Extracting the quantities σ H H and ϕ H H from the numerically computed solutions
of Hodgkin–Huxley system (3.1)–(3.3) we obtain a two-dimensional projection of
58 3 Population of Hodgkin–Huxley Neurons with Spike …

Fig. 3.12 Dynamics of the phase difference between the clusters ϕ H H and mean inter-cluster
coupling σ H H for the solutions of the Hodgkin–Huxley system (3.1)–(3.3) for different initial
conditions. N = 50 with Ns = 7 neurons in the small cluster and Nb = 43 in the big one. Red
orbits converge to the regime of complete synchronization, and blue trajectories lead to a stable
two-cluster solutions. The nullclines of the phenomenological model are shown in gray. Other
parameters: τ p = 2, τd = 5, c p = 2, cd = 1.6, and κmax = 1.5. Figure reproduced from [10]

the solution to the plane (ϕ H H , σ H H ), see Fig. 3.12. The discontinuities in the orbits
are related to the discrete STDP updates. Additionally, since the phases ϕ1,2 (t) can
be firstly accessed after the both clusters fired, some of the area of the phase diagram
(see white area in Fig. 3.12) was not accessible. This “empty” area corresponds to
anti-phase initial conditions, which are very sensitive, and, after each cluster fires,
they appear immediately either in the red or blue area. Nevertheless, the behavior has
the same qualitative features as in the phenomenological model, compare Figs. 3.11
and 3.12.

3.4.3 Criteria for the Emergence of Frequency Clusters

Model (3.10)–(3.11) can be used to describe plasticity functions, which lead to mul-
tiple clusters. For this, we investigate numerically the condition σ (ϕ ∗ ) = σ ∗ . More
specifically, system (3.10)–(3.11) was initialized at the point (−ϕ ∗ , 0) and numeri-
cally integrated forward in time. If σ (ϕ ∗ ) < σ ∗ , the two clusters are stable and do
not merge. This procedure can be repeated for different parameter values.
In order to restrict the set of plasticity parameters, we fix τ p = 2 and τd = 5 and
vary c p and cd . The results of the simulation are shown in Fig. 3.13a. The white, black
and grey parameter areas correspond to the appearance of stable periodic solution
of (3.10)–(3.11) (case (II)), globally stable fixed point (case (I)) and the case (III),
respectively.
In order to compare the parameter regions obtained for the phenomenological
model (Fig. 3.13a) with those for the original Hodgkin–Huxley system, we ran numer-
ical simulations of system (3.1)–(3.3) with N = 50 neurons and Ns = 7 neurons in
3.4 Phenomenological Model with Phase Difference-Dependent Plasticity 59

Fig. 3.13 Panel a: system (3.10)–(3.11). White region: stable periodic solution coexisting with a
stable fixed point, case II. Black region: globally stable fixed point, case I. Grey region: globally
stable periodic solution with σ = 0. Panel b: original system (3.1)–(3.3). White: stable two-clusters
(white); black: stable synchrony and no stable clusters; grey: decoupling of all neurons. Other
parameters τ p = 2, τd = 5, N = 50, Ns = 7, and κmax = 1. Figure reproduced from [10]

the small cluster. Starting from the two-cluster state, we monitor the dynamics of
the clusters. Figure 3.13b shows the results: the white region corresponds to the case
when the clusters survive and stay apart after the simulation time 3000 ms, black—
when the clusters merge into one synchronous group, and grey—when the clusters
split into uncoupled neurons. This behaviour stays qualitatively the same for different
cluster sizes. However, depending on the frequency difference between the clusters,
the set of parameters allowing stable cluster states may change its size.
Comparison of the results for the phenomenological system and the Hodgkin–
Huxley system in the Figs. 3.13a, b shows that the phenomenological model provides
a reasonable approximation.

3.5 Summary

In summary, our results show that adaptive neural networks are able to generate
self-consistently dynamics with different frequency bands. In our case, each cluster
corresponds to a strongly connected component with a fixed frequency. Due to a
sufficiently large difference of the cluster sizes and frequencies, the inter-cluster
interactions are depreciated, while the intra-cluster interactions are potentiated. In
this study, we have described the mechanisms behind the formation and stabilization
of these clusters. In particular, we have explained why the significant difference
between the cluster sizes is important for the decoupling of the clusters. From a
60 3 Population of Hodgkin–Huxley Neurons with Spike …

larger perspective, the decoupling of the clusters in our case is analogous to the
decoupling of timescales in systems with multiple timescales.
In addition, we have presented a two-dimensional phenomenological model which
allows for a detailed study of the clustering mechanisms. Despite of the approxima-
tions made by the derivation, the model coincides surprisingly well with the adap-
tive Hodgkin–Huxley network. Using the phenomenological model, we have found
parameter regions of the plasticity function, where stable frequency clustering can
be observed.
Clustering behavior also emerges at the brain scale, where synchronized commu-
nities of brain regions constituting large distributed functional networks can inter-
mittently be formed and dissolved [21, 22]. Such clustering dynamics can shape the
structured spontaneous brain activity at rest as measured by fMRI. In this study, we
have shown that slow oscillations based on the modulation of synchronized neural
activity can already be formed at the resolution level of a single neural population
if adaptive synapses are taken into account. These modulations of the amplitude of
the mean field can be generated in a stable manner, see Fig. 3.9 and Ref. [3]. The
mechanism relies on fluctuations of the extent of synchronization of tonically firing
neurons. This is caused by the splitting of the neural population into clusters and the
corresponding cluster dynamics. It might contribute to the emergence of slow brain
rhythms of electrical (LFP, EEG) and metabolic (BOLD) brain activity reported by
Refs. [4–7].
However, other mechanisms for generating slow oscillations are possible. The
papers [23, 24] discussed the emergence of slow oscillatory activity (< 1 Hz) that
can be observed in vivo in the cortex during slow-wave sleep, under anesthesia or
in vitro in neural populations. The suggested mechanism relies on the correspond-
ing modulation of the firing of individual neurons, and the slow oscillation at the
population level was proposed to be the result of very slow bursting of individ-
ual neurons that synchronize across the neural population. In contrast, the present
work shows that the slow oscillations of the population mean field can also emerge
when the firing of individual neurons is not affected. The neurons may tonically fire
at high frequencies. The amplitude of the population mean field then oscillates at
much lower frequencies due to the slow modulation caused by the cluster dynamics.
Additionally, more recent work suggested that slow collective oscillations in neural
networks with spike timing-dependent plasticity are induced by the so-called Sisy-
phus effect unveiled using a free energy approach on the temporal behavior of the
order parameter.

References

1. Wittenberg GM, Wang SS (2006) Malleability of spike-timing-dependent plasticity at the ca3-

ca1 synapse. J Neurosci 26:6610
2. Kepecs A, van Rossum MCW, Song S, Tegner J (2002) Spike-timing-dependent plasticity:
common themes and divergent vistas. Biol Cybern 87:446
References 61

3. Popovych OV, Xenakis MN, Tass PA (2015) The spacing principle for unlearning abnormal
neuronal synchrony. PLoS ONE 10:e0117205
4. Mantini D, Perrucci MG, Del Gratta C, Romani GL, Corbetta M (2007) Electrophysiological
signatures of resting state networks in the human brain. Proc Natl Acad Sci USA 104:13170
5. Magri C, Schridde U, Murayama Y, Panzeri S (2012) The amplitude and timing of the bold
signal reflects the relationship between local field potential power at different frequencies. J
Neurosci 32:1395
6. Monto S, Palva S, Voipio J, Palva JM (2008) Very slow EEG fluctuations predict the dynamics
of stimulus detection and oscillation amplitudes in humans. J Neurosci 28:8268
7. Alvarado-Rojas C, Valderrama M, Fouad-Ahmed A, Feldwisch-Drentrup H, Ihle M, Teixeira
CA, Sales F, Schulze-Bonhage A, Adam C, Dourado A, Charpier S, Navarro V, Le Van Quyen M
(2014) Slow modulations of high-frequency activity (40–140 hz) discriminate preictal changes
in human focal epilepsy. Sci Rep 4:4545
8. Maistrenko Y, Lysyansky B, Hauptmann C, Burylko O, Tass PA (2007) Multistability in the
Kuramoto model with synaptic plasticity. Phys Rev E 75:066207
9. Câteau H, Kitano K, Fukai T (2008) Interplay between a phase response curve and spike-
timing-dependent plasticity leading to wireless clustering. Phys Rev E 77:051909
10. Röhr V, Berner R, Lameu EL, Popovych OV, Yanchuk S (2019) Frequency cluster formation
and slow oscillations in neural populations with plasticity. PLoS ONE 14:e0225094
11. Gerstner W, Kempter R, von Hemmen JL, Wagner H (1996) A neuronal learning rule for
sub-millisecond temporal coding. Nature 383:76
12. Markram H, Lübke J, Frotscher M, Sakmann B (1997) Regulation of synaptic efficacy by
coincidence of postsynaptic APs and EPSPs. Science 275:213
13. Bi GQ, Poo MM (1998) Synaptic modifications in cultured hippocampal neurons: dependence
on spike timing, synaptic strength, and postsynaptic cell type. J Neurosci 18:10464
14. Clopath C, Büsing L, Vasilaki E, Gerstner W (2010) Connectivity reflects coding: a model of
voltage-based STDP with homeostasis. Nat Neurosci 13:344
15. Hoppensteadt FC, Izhikevich EM (1997) Weakly connected neural networks. Springer, New
York
16. Pikovsky A, Rosenblum M, Kurths J (2001) Synchronization: a universal concept in nonlinear
sciences, 1st edn. Cambridge University Press, Cambridge
17. Guckenheimer J (1975) Isochrons and phaseless sets. J Math Biol 1:259
18. Winfree AT (2001) The geometry of biological time. Springer, New York
19. Song S, Miller KD, Abbott LF (2000) Competitive hebbian learning through spike-timing-
dependent synaptic plasticity. Nat Neurosci 3:919
20. Rubin J, Lee DD, Sompolinsky H (2001) Equilibrium properties of temporally asymmetric
hebbian plasticity. Phys Rev Lett 86:364
21. Deco G, Jirsa VK, McIntosh AR, Sporns O, Kötter R (2009) Key role of coupling, delay, and
noise in resting brain fluctuations. Proc Natl Acad Sci USA 106:10302
22. Ponce-Alvarez A, Deco G, Hagmann P, Romani GL, Mantini D, Corbetta M (2015) Resting-
state temporal synchronization networks emerge from connectivity topology and heterogeneity.
PLoS Comp Biol 11:e1004100
23. Bazhenov M, Timofeev I, Steriade M, Sejnowski TJ (2002) Model of thalamocortical slow-
wave sleep oscillations and transitions to activated states. J Neurosci 22:8691
24. Compte A, Sanchez-Vives MV, McCormick DA, Wang XJ (2003) Cellular and network mech-
anisms of slow oscillatory activity (<1 Hz) and wave propagations in a cortical network model.
J Neurophys 89:2707
Chapter 4
One-Cluster States in Adaptive Networks
of Coupled Phase Oscillators

The findings in Chap. 3 suggested that the emergence of frequency clusters in neu-
ronal populations with synaptic plasticity is well described by a phase oscillator
model with phase difference-dependent plasticity. A model of this kind was intro-
duced in Refs. [1, 2], see also (2.25)–(2.26) in Sect. 2.3.3, and reads

1 
N
dφi
=ω− κi j sin(φi − φ j + α), (4.1)
dt N j=1
dκi j  
= − κi j + sin(φi − φ j + β) . (4.2)
dt
System (4.1)–(4.2) has attracted a lot of attention recently [1–11], since it is a
first choice paradigmatic model for the modeling of the dynamics of globally cou-
pled adaptive networks. In particular, it generalizes the Kuramoto (or Kuramoto-
Sakaguchi) model with fixed κ. As it is known from numerical studies performed
in Ref. [2], the system (4.1)–(4.2) reaches different cluster states depending on the
initial conditions. One-cluster states are a particular kind of state characterized by
a frequency synchronized motion of all oscillators. In this chapter, we discuss in
detail the nature and properties of these states. We provide necessary and sufficient
condition for their existence as well stability and provide insights in the impact of
parameter changes.
This chapter reproduces contents that have been published in [12] with the permis-
sion of Society for Industrial and Applied Mathematics (2019,c Society for Indus-
trial and Applied Mathematics, https://epubs.siam.org/doi/10.1137/18M1210150,
all rights reserved) and [13] with the permission of AIP Publishing (2019, c AIP
Publishing, https://doi.org/10.1063/1.5097835). The structure of this chapter is as
follows. In Sect. 4.1 we prove the existence of one-cluster states for the system (4.1)–
(4.2) and provide a complete classification of these states. The stability of one-clusters

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 63

R. Berner, Patterns of Synchrony in Complex Networks of Adaptively Coupled Oscillators,
Springer Theses, https://doi.org/10.1007/978-3-030-74938-5_4
64 4 One-Cluster States in Adaptive Networks of Coupled Phase Oscillators

is rigorously described in the subsequent Sect. 4.2. Building on these result, the
impact of the adaptation rate on the stability is explored, Sect. 4.3. Finally, we dis-
cuss the role of double antipodal states in Sect. 4.4. All findings of this chapter are
summarized in Sect. 4.5.

4.1 Classification of One-Cluster States

We start with the collective dynamics, where all oscillators are synchronous up to
phase shifts, i.e. φi = s(t) + ai . It is easy to see from (4.1), that ds/dt = const in
this case, and, hence, s(t) = t with some constant frequency . Moreover, due to
the symmetry of system (4.1)–(4.2) with respect to the phase-shift φi → φi − a1 one
can consider a1 = 0 without loss of generality. Therefore, we define the following
solutions.
Definition 4.1.1 Phase oscillators φi (t), i = 1, . . . , N are said to be
(i) in-phase synchronous if φi (t) = s(t) for all i;
(ii) antiphase synchronous if φi (t) = s(t) + ai with ai ∈ {0, π } and there are i = j
such that ai = a j ;
(iii) rotating-waves if φi (t) = s(t) + (i − 1)2π k/N , where k ∈ {1, . . . , N } is the
wave number;
(iv) phase-locked if φi = s(t) + ai with arbitrary ai ∈ T1 .
Note that the following implications hold: rotating-waves with k = 0 and k = N /2
are in-phase and anti-phase synchronous, respectively. Rotating-waves, in-phase,
and antiphase solutions are phase-locked. The term “rotating-wave” relates to the
rotating symmetry of the solutions with respect to the spatial coordinate given by the
index i. These solutions are also known as twisted states [14, 15] or splay states [16].
It is straightforward to check the values of the nth order parameter (2.15) for the
special solutions defined above. For instance, for an in-phase synchronous solution
it holds Z n (φ(t)) = eint and, hence Rn = 1 for all n ∈ N and t ∈ R. For the other
solutions the results are summarized in Table 4.1. If φ(t) is a phase-locked solution,
the modulus of the nth order parameter does not depend on t, and hence it will be
often referred to as Rn (a), where a = (a1 , . . . , a N )T is the phase shift vector.
With the help of the order parameter we introduce the following three types of
phase-locked states.
Definition 4.1.2 Phase oscillators φi (t), i = 1, . . . , N are said to form a
(i) (Splay cluster) if R2 (φ) = 0;
(ii) (Antipodal cluster) if R2 (φ) = 1, i.e., φi ∈ {0, π } for all i = 1, . . . , N ;
(iii) (Double antipodal cluster) if φi ∈ {0, π, ψ, ψ + π } for all i = 1, . . . , N with
ψ ∈ (0, π ).
Note that if φ is in-phase or anti-phase synchronous, φ forms an antipodal cluster
as well. Rotating waves from Definition 4.1.1 are also known as splay states or
incoherent clusters [16]. For all theses rotating waves, it holds R1 (φ) = R2 (φ) = 0.
4.1 Classification of One-Cluster States 65

Table 4.1 The table summarizes the values for nth order parameter for the phase-locked solutions
introduced in Definition 4.1.1. Table modified from [12]. 2019
c Society for Industrial and Applied
Mathematics. All rights reserved
State nth order parameter
In-phase: φi = t Rn = 1  
 
Anti-phase: φi = t + ai , ai ∈ {0, π } R2n = 1, R2n+1 = 2 NN1 − 1, where N1 is the
number of ai = 0
Rotating-wave: φi = t + i 2π
N k, k  = 0,
N
2 Rn = 1 if n·k
N ∈ N0 and  Rn = 0
1  N 
Phase-locked: φi = t + ai , ai ∈ T1 Rn = N  j=1 e  ina j

As it will be shown in Proposition 4.1.1, system (4.1)–(4.2) generically possesses

solutions with R2 (φ) = 0 rather than R1 (φ) = 0. Both uniformity criteria are clearly
related since R2 (φ) = R1 (2φ). We use the notion splay cluster in a more general sense
to stress that the phases are uniformly distributed around the unit circle with respect to
the second moment of the order parameter. The following result describes all possible
phase-locked solutions in the system of adaptively coupled oscillators (4.1)–(4.2).
We call these solutions one-cluster solutions, since all oscillators possess the same
frequency.
Proposition 4.1.1 System (4.1)–(4.2) possesses the following phase-locked solu-
tions

φi = t + ai , (4.3)
κi j = − sin(ai − a j + β), i, j = 1, . . . , N (4.4)

if and only if one of the following three conditions is fulfilled:

(i) the phases ai form a splay cluster, i.e., R2 (a) = 0;
(ii) the phases ai form an antipodal cluster, i.e., R2 (a) = 1;
(iii) the phases ai form a double antipodal cluster with m ∈ {1, . . . , N − 1}, ai ∈
{0, π, ψm , ψm + π }, i = 1, . . . , N and ψm being the unique modulo 2π solution to
the following equation

N −m
sin(ψ − α − β) = sin(ψ + α + β), (4.5)
m
and the number of phase shifts ai such that ai ∈ {0, π } equals to m.
The corresponding frequencies are given by
⎧
⎪
⎨cos(α − β)/2 if R2 (a) = 0,
 = sin α sin β if R2 (a) = 1, (4.6)
⎪
⎩1
2 (cos(α
− β) − R 2 (a) cos(ψ)) in case (iii).
66 4 One-Cluster States in Adaptive Networks of Coupled Phase Oscillators

Fig. 4.1 Illustration of the three types of one-cluster solutions given by (4.3)–(4.4) for an ensemble
of 50 oscillators. One-cluster solutions a of splay type (R2 (a) = 0) for α = 0.3π , β = 0.1π , b of
antipodal type (R2 (a) = 1), for α = 0.2π , β = −0.95π and c of double antipodal type satisfying
condition (iii) of Proposition 4.1.1 with m = 30 for α = 0.3π , β = −0.15π . Figure reproduced
from [12]. 2019
c Society for Industrial and Applied Mathematics. All rights reserved

The proof of this and other propositions are given in the Appendix A. Note that
for the special cases α = 0, π together with β = α + π/2, α + 3π/2 solutions with
R2 (a) ∈/ {0, 1} were discussed in [8]. Moreover, similar solutions were found in
experimental settings with delay-coupled chemical oscillators [17].
Note that conditions (i)–(iii) of Proposition 4.1.1 imply that there are three possible
types of one-cluster solutions: splay, antipodal, and double antipodal. We illustrate
these solutions in Fig. 4.1a–c.
In the following, we focus our study mainly on the splay and antipodal clusters
with R2 (a) ∈ {0, 1}, i.e., the phase-locked solutions given by cases (i) and (ii) of
Proposition 4.1.1. The role of double antipodal states is discussed in the last section
of this chapter. We further remark that the phase-locked solutions (4.3)–(4.4) are
relative equilibria with respect to the phase-shift defined in Sect. 2.3.3, i.e., they are
equilibria in the co-rotating frame φ → φ + t.
If R2 (a) = 1, Proposition 4.1.1 implies that ai are either 0 or π . Therefore, there
are 2 N −1 isolated solutions of this kind. Note that the in-phase synchronous solution
is an antipodal one-cluster solution.
The situation is different for the splay cluster. The relation R2 (a) = 0 gives the
N − 2 parametric (N > 2) family
4.1 Classification of One-Cluster States 67

Fig. 4.2 Illustration of the family of solutions S a N = 2, b N = 3, c N = 4. Figure reproduced

from [12]. 2019
c Society for Industrial and Applied Mathematics. All rights reserved


N 
N
S := (φi , κi j ) : φi = t + ai , κi j = − sin(ai − a j + β), sin(2a j ) = cos(2a j ) = 0 ,
j=1 j=1
(4.7)

where  = cos(α − β)/2. Moreover, analogously to [18], one can show that S is the
union of N − 2 dimensional manifolds.
The structure of the solution family (4.7) is illustrated in Fig. 4.2a–c for N =
2, 3, 4. Figure 4.2a shows one of the two disjoint one-dimensional subsets of S
for the case of two adaptively coupled oscillators modulo common rotation of the
phases on the circle. In fact, the oscillators have to have a phase shift of π/2 in
order to meet the condition R2 (a) = 0, i.e., a = (γ , γ + π/2) with γ ∈ [0, 2π ).
The dimension of S for N = 2 is one. For a system consisting of three or four
phase oscillators the dimension of S is either 1 or 2, respectively. For N = 3,
one has a = (γ , γ + π/3, γ + 2π/3), see Fig. 4.2b. For the case N = 4, we have
a = (γ , γ + ξ, γ + π/2, γ + ξ + π/2) with γ , ξ ∈ [0, 2π ), see Fig. 4.2c.
Note that the set of phases satisfying the condition R1 (a) = 0 was described
in [18–20]. Our case of splay clusters can be related to this set using the equality
R2 (a) = R1 (2a) = 0.
Rotating-waves are a particular case of the splay cluster, namely, the following
corollary holds.
Corollary 4.1.2 For any k ∈ {1, . . . , N − 1} , k = N /2 the rotating-wave

2π k
φi = t + (i − 1) ,
N 
2π k
κi j = − sin (i − j) +β ,
N

with  = cos(α − β)/2 is a solution of system (4.1)–(4.2).

Let us make a short remark, which allows for a better understanding and interpretation
of the phase-locked solutions given in Proposition 4.1.1. Assume that the phase
variables are in a phase-locked solution φi = t + ai . Then, the coupling weights
68 4 One-Cluster States in Adaptive Networks of Coupled Phase Oscillators

κi j have to satisfy the linear system

κ̇i j = −κi j −  sin(ai − a j + β). (4.8)

This system has the unique solution κi j = − sin(ai − a j + β) which is constant,

bounded on R, and asymptotically stable as t → ∞. Therefore, the specific network
connectivity κi j = − sin(ai − a j + β) is associated with a given phase-shift a.

4.2 Stability of One-Cluster States

In the previous section the existence of one-cluster were discussed. This section is
devoted to the analysis of their stability.
In order to study the local stability of one-cluster solutions described in Sect. 4.1,
we linearize the system (4.1)–(4.2) around the solutions (4.3)–(4.4). We obtain the
following linearized system

N −1
d 1 
δφi = sin(ai − ai+m + β) cos(ai − ai+m + α) (δφi − δφi+m ) (4.9)
dt N m=0
N −1
1 
− sin(ai − ai+m + α)δκi(i+m) ,
N m=0
d  
δκi(i+m) = − δκi(i+m) + cos(ai − ai+m + β) (δφi − δφi+m ) , (4.10)
dt
where we have introduced the new label m := j − i and the convention i + m =
(i + m) mod N for convenience. Throughout this paragraph we will make use of
the Schur complement [21, 22] in order to simplify characteristic equations. More
precisely, any m × m matrix M in the 2 × 2 block form can be written as
   
A B I p B D −1 A − B D −1 C 0 Ip 0
M= = (4.11)
C D 0 Iq 0 D D −1 C Iq

where A is a p × p matrix and D is an invertible q × q matrix. The matrix A −

B D −1 C is called Schur complement. A simple formula for the determinant of M
can be derived with the decomposition (4.11)

det(M) = det(A − B D −1 C) · det(D).

This result is important for the subsequent stability analysis. Note that in the following
an overline indicates the complex conjugate.
Lemma 4.2.1 Suppose ak = (0, 2π k/N , . . . , (N − 1)2π k/N )T with k ∈ {0, . . . ,
N − 1} and the linear system around the one-cluster solution φ = t · (1, . . . , 1)T +
4.2 Stability of One-Cluster States 69

ak is given by (4.9)–(4.10). Then there exist new coordinates (δψ, δζ ) such that the
linearized system can be decomposed into N linear differential equations of the form
⎛ ⎞ ⎛ ⎞
δ ψ̇l δψl
⎜ δ ζ̇l0 ⎟ ⎜ δζl0 ⎟
⎜ ⎟ ⎜ ⎟
⎜ .. ⎟ = Cl ⎜ .. ⎟ l = 0, . . . , N − 1 (4.12)
⎝ . ⎠ ⎝ . ⎠
δ ζ̇l(N −1) δζl(N −1)

with

λ̂l b
Cl := ,
cl −I N

where, I N is the N -dimensional identity matrix and

1
λ̂l = ((Z 1 (al ) − 1) sin(α − β) − Im(Z 2 (ak )) cos(α + β) + Re(Z 2 (ak )) sin(α + β)) (4.13)
2
1 
+ Z 1 (a2k−l )iei(α+β) − Z 1 (a2k+l )ie−i(α+β) .
4

1 2π
b= sin(−α), . . . , sin((N − 1)k − α) , (4.14)
N N
 T
2π 2π 2π 2π
cl = 0, cos(k − β) 1 − eil N , . . . , cos((N − 1)k − β) 1 − eil(N −1) N (4.15)
N N

with any j ∈ {1, . . . , N }.

Proof Due to the cyclic structure in the Eqs. (4.9) and (4.10) it is possible to decouple
them using a discrete Fourier ansatz [23]

N −1
 2π
δφ j = eil j N δψl ,
l=0
N −1
 2π
δκ j ( j+m) = eil j N δζlm .
l=0

Taking this Fourier ansatz and plugging it into the Eqs. (4.9) and (4.10) we get

 −1 N −1   N
−1 
1 
N
2π 2π 2π 2π 2π
eil j N ˙ l =
δψ sin −mk + β cos −mk +α eil j N 1 − eilm N δψl
N N N
l=0 m=0 l=0
N −1  N
−1
1  2π 2π
− sin −mk +α eil j N δζlm ,
N N
m=0 l=0

N −1 
N −1   
2π 2π 2π 2π 2π
eil j N ˙ lm = −
δζ eil j N δζlm + cos −mk + β eil j N 1 − eilm N δψl .
N
l=0 l=0
70 4 One-Cluster States in Adaptive Networks of Coupled Phase Oscillators

After making use of well known trigonometric identities and using the order param-
eters defined in (2.15) we find

N −1    
1  4π 2π 2π
λ̂l = sin − mk + α + β − sin(α − β) 1 − cos lm − i sin lm
2N N N N
m=0
1
= ((Z 1 (al ) − 1) sin(α − β) − Im(Z 2 (ak )) cos(α + β) + Re(Z 2 (ak )) sin(α + β))
2
1 
+ Z 1 (a2k−l )iei(α+β) − Z 1 (a2k+l )ie−i(α+β) .
4

The row and the column vectors b and cl can directly be read of from the transformed
equation above. 

Note that the values λ̂l are exactly the eigenvalues for the case where no interac-
tion between the oscillators and their coupling are assumed or the dynamics of the
coupling weights are left constant. One might expect that due to the slow-fast dynam-
ics of the system (4.1)–(4.2) a small perturbation in the coupling weights could be
neglected for the analysis of stability [5]. In contrast, we show that the local dynam-
ics of the system around the one-cluster solution depends on the interplay between
phases and couplings.
Proposition 4.2.2 Suppose ak = (0, 2π N
k, . . . , (N − 1) 2π
N
k)T and the linear system
around the one-cluster solution φ = t · (1, . . . , 1) + ak is given by (4.9)–(4.10).
T

Then the Jacobian J of this linearized system possesses the following spectrum
 
σ (J ) = −, (λl;1,2 )l=0,...,N −1

with

λ̂l −  1
λl;1,2 = ± (λ̂l + )2 + 4 (b · c)l (4.16)
2 2

and λ̂l , bl , cl as defined in (4.13), (4.14) and (4.15).

Proof Using Lemma 4.2.1 we decompose the linear system (4.9)–(4.10) into the N
blocks (4.12). Consider now the characteristic polynomial for the (N + 1) × (N + 1)
matrix Cl and assume that λl = − then by (4.11) we obtain

  
λl − λ̂l −bl
det(λl I N +1 − Cl ) = det = ( + λl ) N −1 ( + λl )(λl − λ̂l ) −  (b · c)l = 0.
−cl ( + λl )I N

Thus for each l ∈ 0, . . . , N − 1 there are N − 1 eigenvalues λl = −. For the two
remaining eigenvalues we have to solve the quadratic equation
4.2 Stability of One-Cluster States 71

λl2 + ( − λ̂l )λl −  λ̂l −  (b · c)l = 0. (4.17)

In the case of no weight dynamics or no interdependence between the oscillators

and the weights the eigenvalues would read λl;1 = λ̂l and λl;2 = −. Therefore, the
spectrum would look like σc = {−, (λ̂l )l=0,...,N −1 } with (N − 1)N -fold multiplicity
for the eigenvalue −. In contrast to that, we get in general 2N eigenvalues that are
different from − which stem from the interplay of phases and coupling weights. We
should further mention that λ̂l (α + π2 , β − π2 ) = (b · c)l (α, β). With this we write
Eq. (4.17) as
   π π 
λl2 (α, β) +  − λ̂l (α, β) λl −  λ̂l (α, β) + λ̂l (α − , β + ) = 0.
2 2
The following corollary summarizes the results on the spectrum of the linearized
system (4.9)–(4.10).
Corollary 4.2.3 Suppose we have ak = (0, 2π
N
k, . . . , (N − 1) 2π
N
k)T and the linear
system (4.9)–(4.10) then
1. (in-phase and anti-phase synchrony) if k = 0 or k = N /2, the spectrum is given
by  
σ (J ) = (0)1 , (−)(N −1)N +1 , (λ1 ) N −1 , (λ2 ) N −1

where λ1 and λ2 solve λ2 + ( − cos(α) sin(β)) λ −  sin(α + β) = 0,

2. (incoherent rotating-wave) if k = 0, N /2, N /4, 3N /4, the spectrum is

  
sin(α − β)    
σ (J ) = (0) N −2 , (−)(N −1)N +1 , − − , (ϑ1 )1 , (ϑ2 )1 , ϑ 1 1 , ϑ 2 1
2 N −3

 
where ϑ1 and ϑ2 solve ϑ 2 +  + 21 sin(α − β) − 14 iei(α+β) ϑ − 2 iei(α+β) = 0,
3. (4-rotating-wave solution) if k = N /4, 3N /4, the spectrum is
  
sin(α − β)
σ (J ) = (0) N −1 , (−)(N −1)N +1 , − − , (λ1 )1 , (λ2 )1
2 N −2

where λ1 and λ2 solve λ2 + ( + sin(α) cos(β)) λ +  sin(α + β) = 0.

Here, the multiplicities for each eigenvalue are given as lower case labels.
As we can see from this corollary there exists always at least one zero eigenvalue. This
is due to the phase-shift symmetry of (4.1)–(4.2) we already discussed in Sect. 4.1.
The additional zero eigenvalues for the wave numbers k = 0, N /2 can be explained
with our findings from Proposition 4.1.1 and Corollary 4.1.2. These linear rotating-
wave solutions belong to a N − 2 dimensional family of solutions characterized by
R2 (a) = 0. Thus, around any point of this family the linear equation
72 4 One-Cluster States in Adaptive Networks of Coupled Phase Oscillators

Fig. 4.3 Stability diagrams for rotating-wave clusters depending on the parameters α and β are
shown. The regions are colored according to numerical simulation. Blue regions correspond to stable
solutions while yellow regions correspond to unstable solutions. The black dashed lines show to the
borders of stability determined by Corollary 4.2.3. Parameter  = 0.01 is fixed for all simulations.
a k = 1, b k = N /2, c k = N /4. Figure reproduced from [12]. 2019
c Society for Industrial and
Applied Mathematics. All rights reserved


N
ei2a j δφ j = 0 (4.18)
j=1

 
holds for a certain choice of coordinates δφ, δκi j , and hence there are two lin-
early independent equations for the infinitesimal perturbations δφi . This explains
the appearance of N − 2 zero eigenvalues. They correspond to the variation along
the manifold of solution. In the special case of k = N /4, 3N /4 the two algebraic
Eqs. (4.18) are linear dependent and we are thus left with only one linear equation
which increases the multiplicity of the zero eigenvalue by one.
The results of Corollary 4.2.3 are presented in Fig. 4.3a–c and compared with
numerical simulations. The numerical results are obtained by numerical integration
of system (4.1)–(4.2) with N = 20. The initial conditions for each simulation are set
to the one-cluster solution given in Proposition 4.1.1 with a small perturbation added
to each dynamical variable and randomly chosen from the interval [−0.01, 0.01].
The numerical integration is stopped after t = 5000 time steps. The relative coordi-
nates i := φi − φ1 for i = 1, . . . , N are introduced in order to compare the initial
phase configuration with the distribution of the phases after numerical integration. A
one-cluster is said to be stable if  after numerical integration is closer to the theoret-
ical one-cluster state than  of the initially perturbed phase distribution. Otherwise,
the one-cluster solution is considered as unstable. Closeness is measured by the
Euclidean distance. The parameter regions in the (α, β) plane for stable one-cluster
solutions are colored blue while the regions for unstable one-cluster solutions are
colored yellow. The black dashed lines correspond to the borders of stability deter-
mined with the results in Corollary 4.2.3. In particular, a state is asymptotically stable
if Re(λ) < 0 for all λ ∈ σ (C) except the zero eigenvalues related to the perturbations
4.2 Stability of One-Cluster States 73

(a) (b)

α /π

β/π β/π

Fig. 4.4 Stability diagram for splay and antipodal one-cluster solutions depending on the param-
eters α and β are shown. The regions are colored according numerical eigenvalues of the Jaco-
bian (4.9)–(4.10). Blue areas correspond to stable while yellow areas correspond to unstable regions.
Parameter  = 0.01 is fixed in all simulations. a Splay solution as in Fig. 4.1a, b Anti-phase solution
as in Fig. 4.1b. Figure reproduced from [12]. 2019
c Society for Industrial and Applied Mathemat-
ics. All rights reserved

along the solution families. In all three cases the numerical and analytic results agree
very well.
In addition to the analysis of rotating-wave solutions, we investigate the stability
for the splay solutions characterized by R(a) = 0 and the antipodal solutions charac-
terized by R(a) = 1. For this, the stability is calculated semi-analytically by taking
the solutions displayed in Fig. 4.1a–b, plugging them into the Jacobian matrix given
by the linearized Eqs. (4.9)–(4.10) and determining the eigenvalues of the Jacobian
numerically. The results of this procedure are shown in Fig. 4.4 together with the
borders of stability calculated with Proposition 4.1.1. In comparison with Fig. 4.3,
the analysis yields the same stability regions which are found for the rotating-wave
solutions. The numerical findings indicate that the stability for all splay and antipodal
solutions coincide with the stability of the rotating-waves.
Complementary to the previous stability results, all cases of antipodal, 4-phase,
and double antipodal cluster are analyzed. For these states the analysis is slightly more
subtle and given in the appendix A.2. In the following Proposition, we summarize
the findings.
Proposition 4.2.4 The following statements hold true.
1. The set of eigenvalues of the linearized system (4.9)–(4.10) around all antipodal
states with ai ∈ {0, π } agrees with the set L in Proposition 4.2.2 for rotating-wave
states with k = 0, N /2.
2. The set of eigenvalues to the linearized system (4.9)–(4.10) around all 4-phase-
cluster states with ai ∈ {0, π/2, π, 3π/2} and R2 (a) = 0 agrees with the set L
in Proposition 4.2.2 for 4-rotating-wave states.
3. The double antipodal states are unstable for all α and β.
74 4 One-Cluster States in Adaptive Networks of Coupled Phase Oscillators

As indicated by the numerical analysis in Fig. 4.4, Proposition 4.2.4 shows that in fact
in- as well as anti-phase cluster share the same stability features as all other antipo-
dal clusters. In addition, also general 4-phase cluster possess the same Lyapunov
spectrum as their special 4-rotating wave cluster.
Remarkably, the third type of one-cluster, namely, the double antipodal state is
unstable everywhere in parameter space. In Sect. 4.4, we describe the role of double
antipodal state for the global structure of the phase space.

4.3 Adaptation Rate Dependence of One-Cluster Stability

In the previous sections, the existence and stability of one-cluster solutions are rig-
orously described. Building on this, we analyze dependence of the one-cluster states
on the adaptation rate . In contrast to the existence of one-cluster states, the stability
of those states depend crucially on the time-separation parameter.
The diagram in Fig. 4.5 shows the regions of stability for antipodal and rotating-
wave one-cluster states. The diagram is based on the analytic results presented in
Corollary 4.2.3 and Proposition 4.2.4.
In Fig. 4.5 the regions of stability are presented for several values of the time
separation parameter . The first case in panel (a) assumes  = 0, where the network
structure is non-adaptive but fixed to the values given by the one-cluster states, i.e.,
κi j = − sin(ai − a j + β) as given in Sect. 4.1. The linearized system in this case is
given by

1   
N
dδφi
=− sin(α − β) − sin(2(ai − a j ) + α + β) δφi − δφ j . (4.19)
dt 2N j=1

For the synchronized or antipodal state, the value 2ai mod 2π is the same for
all i. Hence, the term 2(ai − a j ) disappears and the linearized system (4.19) pos-
sesses the same form as the linearized system for the synchronized state of the
Kuramoto-Sakaguchi system [19] with coupling constant σ (β) = − sin(β). As it
follows from Ref. [19], the synchronized as well as all other antipodal states are
stable for σ (β) cos(α) > 0. The region of stability of the rotating-wave cluster has
a more complex shape, see hatched area in Fig. 4.5a. We find large areas where
both types of one-cluster states are stable simultaneously, as well as the regions
where no frequency synchronized state is stable. The results shown in Fig. 4.5a are
in agreement with Ref. [5], where the authors consider the case  = 0 in order to
approximate the limit case of extremely slow adaptation  → 0. However, such an
approach for studying the stability of clusters for small adaptation is not correct in
general. As Figs. 4.5b–d show, the stability of the network with small adaptation
 > 0 is different.
The case  = 0.01 is shown in Fig. 4.5b, where we observe regions for stable
antipodal and rotating-wave states as well. The introduction of a small but non-
4.3 Adaptation Rate Dependence of One-Cluster Stability 75

antipodal states rotating wave states (splay)

(a)
=0
α /π Asynchronous
behaviour

(b)
 = 0.01
α /π

Asynchronous
behaviour

(c)
 = 0.1
α /π

Asynchronous
behaviour

(d)
=1
α/π

β/π

Fig. 4.5 The regions of stability for antipodal and rotating-wave states are presented in (α, β)
parameter space for different values of . Coloured and hatched areas refer to stable regions for
these states as indicated in the legend. White areas refer to region where these one-cluster states
are unstable. a  = 0; b  = 0.01; c  = 0.1; d  = 1. Figure reproduced from [13]. 2019
c AIP
Publishing

vanishing adaptation changes the regions of stability significantly. The diagram in

Fig. 4.5b remains qualitatively the same for smaller values of . This can be read
of from the analytic findings presented in the Corollary 4.2.3 and Proposition 4.2.4.
The changes in the stability areas are due to subtle changes in the equation which
determine the eigenvalue of the corresponding linearized system. In fact, the adapta-
tion introduces the necessary condition sin(α + β) < 0 for the stability of antipodal
states. Additionally, for all rotating-wave states of splay type, the necessary condition
sin(α − β) + 2 > 0 is introduced, see Corollary 4.2.3. This is why, a non-trivial
effect of adaptivity on the stability in Figs. 4.5 is observed. In particular, the parameter
β, which determines the plasticity rule, has now a non-trivial impact on the stability
of antipodal states for any  > 0. As it can be seen in Fig. 4.5b, one-cluster states
of antipodal type are supported by a Hebbian-like adaptation (β ≈ −π/2) while
splay states are supported by causal rules (β ≈ 0). For the asynchronous region, the
76 4 One-Cluster States in Adaptive Networks of Coupled Phase Oscillators

dynamical system (4.1)–(4.2) can exhibit very complex dynamics and show chaotic
motion [9]. This region is supported by an anti-Hebbian-like rule (β ≈ π/2).
By increasing the parameter , see Fig. 4.5c, d, two observations can be made.
First, the region of asynchronous dynamical behavior is shrinking. For  = 1, we
find at least one stable one-cluster state for any choice of the phase lag parameters
α and β. Secondly, the regions where both types of one-cluster states are stable
are shrinking as well. In the limit of instant network adaptation, i.e.,  → ∞, the
stability regions are completely separated. Both types of one-cluster states divide
the whole parameter space into two areas. In this case, the boundaries are described
by α + β = 0 and α + β = π . This division can be seen from the analytic findings
presented in Corollary 4.2.3. In the case of antipodal states, the quadratic equation
which determines the Lyapunov coefficients has negative roots if and only if  +
cos(α) sin(β) > 0 and sin(α + β) < 0. Here, even for  > 1, the condition sin(α +
β) < 0 is the only remaining one. Similarly, we find sin(α + β) > 0 as a condition
for the stability of rotating-wave states for  → ∞.

4.4 Double Antipodal States

Splay and antipodal clusters serve as building blocks for multicluster states. The
third type, the double antipodal clusters, are not of this nature since they appear to
be unstable everywhere. As unstable objects, they can still play an important role
for the dynamics. Here we would like to present an example, where the double
antipodal clusters become part of a simple heteroclinic network. As a result, they
can be observed as metastable states in numerics.
As an example, we first analyze the system of N = 3 adaptively coupled phase
oscillators which is the smallest system with a double antipodal state. According
to the definition of a double antipodal state, laid out in Sect. 4.1, the phases ai of
the oscillators φi are allowed to take values from the set {0, π, ψ, ψ + π } where ψ
uniquely solves Eq. (4.5) for a given m ∈ {1, . . . , N }. Further, at least one oscillator
φi with ai ∈ {0, π } and one oscillator φ j ( j = i) with a j ∈ {ψ, ψ + π } are needed
in order to represent one of the two antipodal groups. Note that for the parameters
given in Fig. 4.6 and N = 3, the Eq. (4.5) yields ψ = 1.602π if a1 , a2 ∈ {0, π } and
a3 ∈ {ψ, ψ + π }.
In Fig. 4.6a we present trajectories which initially start close to antipodal clus-
ters. The trajectories in phase space are represented by the relative coordinates
θ12 = φ1 − φ2 and θ13 = φ1 − φ3 . In particular, the two configurations with (θ12 =
0, θ13 = 0) and (θ12 = 0, θ13 = π ) are considered. The coupling weights are initial-
ized according to Eq. (4.4). With the given parameters, the unstable manifold of the
antipodal state is one-dimensional which can be determined via Corollary 4.2.2. For
the numerical simulation, we perturb the antipodal state in such a way that two dis-
tinct orbits close to the unstable manifold are visible. For both configuration the two
orbits are displayed in Fig. 4.6a. It can be observed that after leaving the antipodal
state the trajectories approach the double antipodal states before leaving it towards
4.4 Double Antipodal States 77

(a)

θ/π

θ12
θ13

(b)
metastable
double antipodal
cluster
R2 (φ(t))

(c)
R2 (φ(t))

time

Fig. 4.6 Heteroclinic orbits between several steady states in a system of 3 and 100 adaptively
coupled phase oscillators. a The time series for the relative phases θ12 (solid lines) and θ13 (dashed
lines) for N = 3 are shown. Lines with the same colour correspond to the same trajectories. Panel b, c
show time series for the second moment order parameter R2 (φ(t)) as well as a schematic illustration
of the observed heteroclinic connections (right) for b N = 3 and c N = 100. Parameter values:
 = 0.01, α = 0.4π , and β = −0.15π . Figure reproduced from [13]. 2019 c AIP Publishing
78 4 One-Cluster States in Adaptive Networks of Coupled Phase Oscillators

the direction of a splay state. With this we numerically find orbits close to “hetero-
clinic”, which connect antipodal, double antipodal, and splay clusters, see schematic
picture on the right in Fig. 4.6b. The phase differences θ12 and θ13 at the double
antipodal state agree with the solution ψ of Eq. (4.5) or ψ + π .
Figure 4.6b further justifies our statements on the heteroclinic contours. Here,
we see the time series for the second moment order parameter for all trajecto-
ries in Fig. 4.6b. It can be seen that in all cases we start at an antipodal cluster
(R2 (φ) = 1) from which the double antipodal state (R2 (φ) ≈ 0.447, theoretical) is
quickly approached. The trajectories stay close to the double antipodal cluster for
approximately 2000 time units (shaded area) before leaving the invariant set towards
the splay state (R2 (φ) = 0).
As a second example we analyze a system of N = 100 adaptively coupled phase
oscillators. Here, we choose two particular antipodal states as initial condition and
add a small perturbation to both. One of the states is chosen as an in-phase syn-
chronous cluster. In both cases, the couplings weights are initialized in accordance
with Eq. (4.4). For Fig. 4.6c we depict the trajectories which show a clear hetero-
clinic contour between antipodal, double antipodal, and splay state as in the example
of three phase oscillators. We illustrate the heteroclinic connections and present
R2 (t) for the corresponding trajectories in Fig. 4.6c. Here, the zoomed view clearly
shows that the trajectories for both initial conditions starting at an antipodal cluster
(R2 (φ) = 1) again first approach a double antipodal state (R2 (φ) ≈ 0.990, theo-
retical) before leaving it towards a splay state (R2 (φ) = 0). More precisely, each
trajectory comes close to a particular double antipodal state for which only one
oscillator has a phase in {ψ, ψ + π }. Remarkably, these states, also known as soli-
tary states, have been found in a range of other systems of coupled oscillators, as
well [24]. With the results in Proposition A.2.3 in the appendix, one can show that
this double antipodal states have a stable manifold with co-dimension one which
thus divides the phase space. Next to this fact, numerical evidence for the existence
of heteroclinic connections between antipodal and double antipodal states as well
as between double antipodal and the family of splay states is provided in Fig. 4.6c.
With this, double antipodal states play an important role for the organization of the
dynamics in system (4.1)–(4.2).

4.5 Summary

The numerical analysis in Ref. [2] suggested that one-cluster states in system (4.1)–
(4.2) may serve as building blocks for multicluster states. In this chapter, we have
exhaustively analyzed the properties of one-cluster states in a network of adaptively
coupled phase oscillators. In particular, we have found that there are only three types
of one-cluster states: splay, antipodal, and double antipodal. It has been shown that all
one-cluster solutions of splay type form an N − 2 dimensional family and thus give
rise to infinitely many solutions the system can achieve. In order to understand the
stability of these cluster states, we have performed a linear stability analysis. We have
4.5 Summary 79

provided analytic results for the stability of antipodal as well as of special types of
splay states, namely, rotating-wave states. The stability of these states is rigorously
described, and the impact of all parameters has been shown. Note that due to the
N − 2 dimensional family of splay type cluster, the rotating-wave states possess
N − 2 neutrally stable directions. Remarkably, this property of splay states has been
also found in networks of pulse-coupled rotators [25] and excitable neurons [26].
In this chapter, we have seen that ring-like structures are dynamically associated
with rotating-wave and splay states. Furthermore, ring networks are important motifs
in neural networks [27–29] and have been the basis for many interesting dynamical
regimes found for different systems, e.g. [30–33]. Moreover, ring-structures were
recently found in fly brains [34]. In our study, we have presented how ring-like
structures could emerge due to adaptation. In particular, we have shown that causal
plasticity rules, see also Fig. 3.1, as they are found in many experimental studies [35–
37] support the appearance of ring-like networks.
While the time-scale separation has no influence upon the existence of the one-
cluster states found in this study, it plays an important role for the stability of the
one-clusters. The regions of stability in parameter space have been presented for
different choices of the time-scale separation parameter. The singular limit ( → 0)
and the limit of instantaneous adaptation have been analyzed. The latter shows that
the stability regions of the splay and the antipodal states divide the whole space
into two equally sized regions without intersection. Instantaneous adaptation cancels
multistability of these states. The consideration of the singular limit shows that it
differs from the case of no adaptation. Therefore, even for very slow adaptation, the
oscillatory dynamics alone is not sufficient to describe the stability of the system.
Additionally to the analysis of splay and antipodal clusters, we have proved that
the double antipodal states are unstable in the whole parameter range. In fact, they
appear to be saddle-points in the phase space. We have found that in a system of 3
oscillators, double antipodal states are transient states in a small heteroclinic network
between antipodal and splay states. They appear to be metastable, i.e., observable
for a relatively long time and therefore are physically important transient states.
Moreover, an additional analysis for an ensemble of 100 phase oscillators has revealed
the importance of the double antipodal states for the global dynamics of the whole
system.

References

1. Aoki T, Aoyagi T (2009) Co-evolution of phases and connection strengths in a network of

phase oscillators. Phys Rev Lett 102:034101
2. Kasatkin DV, Yanchuk S, Schöll E, Nekorkin VI (2017) Self-organized emergence of multi-
layer structure and chimera states in dynamical networks with adaptive couplings. Phys Rev E
96:062211
3. Seliger P, Young SC, Tsimring LS (2002) Plasticity and learning in a network of coupled phase
oscillators. Phys Rev E 65:041906
80 4 One-Cluster States in Adaptive Networks of Coupled Phase Oscillators

4. Ren Q, Zhao J (2007) Adaptive coupling and enhanced synchronization in coupled phase
oscillators. Phys Rev E 76:016207
5. Aoki T, Aoyagi T (2011) Self-organized network of phase oscillators coupled by activity-
dependent interactions. Phys Rev E 84:066109
6. Picallo CB, Riecke H (2011) Adaptive oscillator networks with conserved overall coupling:
sequential firing and near-synchronized states. Phys Rev E 83:036206
7. Timms L, English LQ (2014) Synchronization in phase-coupled Kuramoto oscillator networks
with axonal delay and synaptic plasticity. Phys Rev E 89:032906
8. Gushchin A, Mallada E, Tang A (2015) Synchronization of phase-coupled oscillators with
plastic coupling strength. In: Information theory and applications workshop ITA 2015, CA,
USA. IEEE, San Diego, pp 291–300
9. Kasatkin DV, Nekorkin VI (2016) Dynamics of the phase oscillators with plastic couplings.
Radiophys Quantum Electron 58:877
10. Nekorkin VI, Kasatkin DV (2016) Dynamics of a network of phase oscillators with plastic
couplings. AIP Conf Proc 1738:210010
11. Avalos-Gaytán V, Almendral JA, Leyva I, Battiston F, Nicosia V, Latora V, Boccaletti S (2018)
Emergent explosive synchronization in adaptive complex networks. Phys Rev E 97:042301
12. Berner R, Schöll E, Yanchuk S (2019) Multiclusters in networks of adaptively coupled phase
oscillators. SIAM J Appl Dyn Syst 18:2227
13. Berner R, Fialkowski J, Kasatkin DV, Nekorkin VI, Yanchuk S, Schöll E (2019) Hierarchical
frequency clusters in adaptive networks of phase oscillators. Chaos 29:103134
14. Wiley DA, Strogatz SH, Girvan M (2006) The size of the sync basin. Chaos 16:015103
15. Girnyk T, Hasler M, Maistrenko Y (2012) Multistability of twisted states in non-locally coupled
Kuramoto-type models. Chaos 22:013114
16. Choe CU, Dahms T, Hövel P, Schöll E (2010) Controlling synchrony by delay coupling in
networks: from in-phase to splay and cluster states. Phys Rev E 81:025205(R)
17. Blaha K, Lehnert J, Keane A, Dahms T, Hövel P, Schöll E, Hudson JL (2013) Clustering in
delay-coupled smooth and relaxational chemical oscillators. Phys Rev E 88:062915
18. Ashwin P, Burylko O, Maistrenko Y (2008) Bifurcation to heteroclinic cycles and sensitivity
in three and four coupled phase oscillators. Phys D 237:454
19. Burylko O, Pikovsky A (2011) Desynchronization transitions in nonlinearly coupled phase
oscillators. Phys D 240:1352
20. Ashwin P, Bick C, Burylko O (2016) Identical phase oscillator networks: bifurcations, sym-
metry and reversibility for generalized coupling. Front Appl Math Stat 2
21. Boyd S, Vandenberghe L (2004) Convex optimization. Cambridge University Press, Cambridge
22. Liesen J, Mehrmann V (2015) Linear algebra. Springer, Cham
23. Perlikowski P, Yanchuk S, Popovych OV, Tass P (2010) Periodic patterns in a ring of delay-
coupled oscillators. Phys Rev E 82:036208
24. Maistrenko Y, Penkovsky B, Rosenblum M (2014) Solitary state at the edge of synchrony in
ensembles with attractive and repulsive interactions. Phys Rev E 89:060901
25. Calamai M, Politi A, Torcini A (2009) Stability of splay states in globally coupled rotators.
Phys Rev E 80:036209
26. Dipoppa M, Krupa M, Torcini A, Gutkin BS (2012) Splay states in finite pulse-coupled networks
of excitable neurons. SIAM J Appl Dyn Syst 11:864
27. Compte A, Sanchez-Vives MV, McCormick DA, Wang XJ (2003) Cellular and network mech-
anisms of slow oscillatory activity (<1 Hz) and wave propagations in a cortical network model.
J Neurophysiol 89:2707
28. Sporns O (2011) Networks of the brain. MIT Press, Cambridge
29. Popovych OV, Yanchuk S, Tass P (2011) Delay- and coupling-induced firing patterns in oscil-
latory neural loops. Phys Rev Lett 107:228102
30. Abrams DM, Strogatz SH (2004) Chimera states for coupled oscillators. Phys Rev Lett
93:174102
31. Klinshov V, Lücken L, Shchapin D, Nekorkin VI, Yanchuk S (2015) Multistable jittering in
oscillators with pulsatile delayed feedback. Phys Rev Lett 114:178103
References 81

32. Jaros P, Brezetsky S, Levchenko R, Dudkowski D, Kapitaniak T, Maistrenko Y (2018) Solitary

states for coupled oscillators with inertia. Chaos 28:011103
33. Omel’chenko OE, Knobloch E (2019) Chimerapedia: coherence-incoherence patterns in one,
two and three dimensions. New J Phys 21:093034
34. Kim SS, Rouault H, Druckmann S, Jayaraman V (2017) Ring attractor dynamics in the
drosophila central brain. Science 356:849
35. Abbott LF, Nelson S (2000) Synaptic plasticity: taming the beast. Nat Neurosci 3:1178
36. Bi GQ, Poo MM (2001) Synaptic modification by correlated activity: Hebb’s postulate revisited.
Annu Rev Neurosci 24:139
37. Caporale N, Dan Y (2008) Spike timing-dependent plasticity: a Hebbian learning rule. Annu
Rev Neurosci 31:25
Chapter 5
Multicluster States in Adaptive Networks
of Coupled Phase Oscillators

In this chapter we investigate the shape and properties of frequency-cluster (mul-

ticluster) states as they have been found in Chap. 3, i.e., where each cluster has a
different frequency. Inspired by the reduced phenomenological model presented in
Chap. 3, we have studied a paradigmatic model of adaptively and globally coupled
phase oscillator in Chap. 4 which reads

1 
N
dφi
=ω− κi j sin(φi − φ j + α), (5.1)
dt N j=1
dκi j  
= − κi j + sin(φi − φ j + β) , (5.2)
dt
see also Sect. 2.3.3. In Chap. 4, the properties of one-cluster state have been exhaus-
tively analyzed. In this chapter we study numerically as well as analytically multi-
cluster states which we observe by solving the Eqs. (5.1)–(5.2) From the numerical
study in Ref. [1], we know that certain phase-locked solutions may serve as building
blocks for (hierarchical) multicluster solutions. Here, we show that indeed multiclus-
ters are build-up by one-cluster states and even may inherit their dynamical features.
Multicluster states are formally defined as follows.
Definition 5.0.1 Phase oscillators φi (t) form a multicluster if they can be separated
into M groups of phase-locked oscillators (clusters), i.e., for all μ ∈ {1, . . . , M}
the phase oscillators φi,μ , i ∈ {1, . . . , Nμ }, from each group μ satisfy φi,μ (t) =
sμ (t) + ai,μ .
The appearance of multiclusters is interesting and nontrivial, since such solutions,
in contrast to one-clusters, are no more relative equilibria of (5.1)–(5.2), but are
periodic or quasi-periodic solutions, which appear due to the special structure of
the equation and adaptive nature of the coupling. The oscillators within one cluster
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 83
R. Berner, Patterns of Synchrony in Complex Networks of Adaptively Coupled Oscillators,
Springer Theses, https://doi.org/10.1007/978-3-030-74938-5_5
84 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

possess a synchronized temporal dynamics with possible phase lags. In a multicluster,

the coupling matrix κ is divided into different blocks according to the division by
clusters: ki j,μν will refer to the coupling weight between the ith oscillator of cluster
μ to the jth oscillator of cluster ν.
This chapter reproduces contents that have been published in [2] with the permis-
sion of Society for Industrial and Applied Mathematics (2019,
c Society for Indus-
trial and Applied Mathematics, https://epubs.siam.org/doi/10.1137/18M1210150,
all rights reserved) and [3] with the permission of AIP Publishing (2019, c AIP
Publishing, https://doi.org/10.1063/1.5097835). The chapter is organized as follows.
First, in Sect. 5.1, system (5.1)–(5.2) is numerically analyzed and all cluster mul-
ticluster states are qualitatively described and classified. Subsequently, we derive
analytic expression for all types of multicluster states. The results are given and
exhaustively discussed in Sects. 5.2, 5.4, and 5.5 for splay, antipodal, and mixed
type multicluster states, respectively. Additionally, the main existence theorem for
multicluster states, particularly for antipodal and mixed type, is given in Sect. 5.3. In
Sect. 5.6, we analyze numerically as well as analytically the stability of multiclus-
ter states and discuss the hierarchical structure of cluster sizes. All findings of this
chapter are summarized in Sect. 5.7.

5.1 Numerical Observation of Multicluster States

Figure 5.1 shows a hierarchical multicluster state. The solution was obtained by
integrating the system (5.1)–(5.2) numerically starting from uniformly distributed
random initial conditions. The self-couplings κii are set to zero in numerical sim-
ulations, since they do not influence the relative dynamics of the system. This is
due to the fact that all variables κii converge to − sin β, which leads to the same
constant term sin(β) sin(α)/N in the right hand side of Eq. (5.1). The latter term
can be absorbed in the co-rotating coordinate frame. We reorder the oscillators (after
sufficiently long transient time) by first sorting the oscillators with respect to their
average frequencies. After that the oscillators with the same frequency are sorted
by their phases. Figure 5.1a displays the coupling matrix (right) of the multicluster
state and a representation of the coupling structure as a network graph (left). The
coupling matrix demonstrates a clear splitting into three groups. This splitting is also
visible in the graph representation of the coupling network. The coupling weights
between oscillators of the same group vary in a larger range than between those of
different groups which are generally smaller in magnitude. The splitting into three
groups is manifested in the behaviour of the phase oscillators, as well. We find that
the oscillators of the same group possess the same constant frequency with possible
phase lags, Fig. 5.1b. We call the groups of oscillators (frequency) clusters and the
corresponding dynamical states multi(frequency)cluster states.
In a multicluster, the coupling matrix κ is divided into different blocks. We observe
that the behavior for each oscillator in a M-cluster state takes the form
5.1 Numerical Observation of Multicluster States 85

Fig. 5.1 Three-frequencycluster of splay type at t = 10000. a Coupling weights represented as a

graph (left) and as a coupling matrix (right). In the graph representation, the dynamical nodes are
represented by red nodes and the edges are coloured with respect to the coupling weight. Red and
blue refer to positive and negative coupling weights, respectively. Light and dark colors refer to
weak and strong coupling weights, respectively. b Distribution of the phases φi for each of the three
clusters. Each node represents one oscillator and is coloured with respect to the cluster to which it
belongs. Parameter values:  = 0.01, α = 0.3π , β = 0.23π , ω = 0, and N = 100. Figure modified
from [3]. 2019
c AIP Publishing

μ = 1, . . . , M
φi,μ (t) = μt + ai,μ + si,μ (t) (5.3)
i = 1, . . . , Nμ

where M is the number of clusters, Nμ is the number of oscillators in the μth cluster,
ai,μ ∈ [0, 2π ) are phase lags, and μ ∈ R is the collective frequency of the oscillators
in the μth cluster. The functions si,μ are bounded.
The numerical analysis of system (5.1)–(5.2) shows the appearance of different
multicluster states depending on particular choices of the phase lag parameters α and
β as well as on initial conditions. Starting from uniformly distributed random initial
conditions the system end up in several states such as multiclusters and chimera-like
states [1]. Figure 5.2 shows examples for the three types of multicluster states which
appear dynamically in (5.1)–(5.2).

5.1.1 Splay Type Cluster States

The first type is called splay type multicluster state, see Fig. 5.2a. The separation
into three clusters is clearly visible in the coupling matrix, as well as a hierarchical
structure in the cluster sizes. Regarding the distribution of the phases, we notice that
the oscillators from each group are almost homogeneously dispersed on the circle.
In fact, the phases from each cluster fulfill the condition R2 (φ μ ) = 0 (μ = 1, 2, 3)
with order parameter (2.15). Note that splay states as they are defined in several other
works [4–6] share the property R1 (φ) = 0. This property can be seen as a measure of
86 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

κij

(a)

φ1,µ − Ωµ t
index i

(b)
Cluster 1
Cluster 2
Cluster 3

φ1,µ − Ωµ t
index i

(c)
Cluster 1 (i = 1)
Cluster 2 (i = 1)
Cluster 2 (i = 21)
φi,µ − Ωµ t
index i

time
index j

Fig. 5.2 Three different types of multicluster states at t = 10000 with N = 100 and  = 0.01.
For all types, the coupling matrix (left), distribution of the phases (middle), and time series of
representative phase oscillators from each cluster (right) are presented. In the plot of the phase
distribution, each node represents one oscillator and is colored with respect to the cluster to which
it belongs. The time series are shown after subtracting the average linear growth φi,μ (t) −  μ t.
The colouring of the time series (shaded for visibility) of a representative phase oscillator from one
cluster is in accordance with the pictures in the middle panel. a Splay type 3-cluster for α = 0.3π ,
β = 0.23π ; b Antipodal type 3-cluster for α = 0.3π , β = −0.53π ; c Mixed type 2-cluster for
α = 0.3π , β = −0.4π . Figure reproduced from [3]. 2019c AIP Publishing

incoherence for the oscillator phases, as well. In fact, it was shown that splay states
are part of a whole family of solutions [7, 8] given by exactly R1 (φ) = 0. Further,
R2 (φ) = R1 (2φ) relates the two measures of incoherence. These facts motivate the
definition of those clusters with R2 (φ) = 0 as splay type clusters.
The temporal behavior for all phase oscillators in the splay multicluster state
is characterized by a constant frequency which differs for the different clusters,
i.e., according to (5.3), φi,μ (t) = μ t + ai,μ with R2 (aμ ) = 0 for all μ = 1, 2, 3
and i = 1, . . . , Nμ . In addition, the hierarchical cluster sizes are reflected in the
frequencies. Oscillators of a big cluster have a higher frequency than those of smaller
clusters. The coupling weights between the phase oscillators are fixed or change
periodically with time depending on whether the oscillators belong to the same or
different clusters, respectively. Moreover, the amplitude of coupling weights between
5.1 Numerical Observation of Multicluster States 87

clusters depends on the frequency difference of the corresponding clusters. The

higher the frequency difference, the smaller is the amplitude. The periodic behavior
of the coupling weights between clusters is present in all types of multicluster states
(Fig. 5.2a–c).

5.1.2 Antipodal Type Cluster States

Figure 5.2b shows another possible multicluster state. As in Fig. 5.2a the clusters are
clearly visible and their oscillators show frequency synchronized temporal behavior.
In addition, the time series for the oscillators show periodic modulations on top of
the linear growth. This additional dynamics is the same for all oscillators of the
same cluster, and hence they are still temporally synchronized. We have φi,μ (t) =
μ t + ai,μ + sμ (t). In analogy to the coupling weights between the clusters, the
amplitudes of the bounded function sμ (t) depend on the differences of the cluster
frequencies.
In contrast to the splay states, the phase distribution fulfills R2 (aμ ) = 1 for all
μ = 1, 2, 3, see Fig. 5.2b, middle panel. Hence, all oscillators of a cluster have either
the same phase aμ ∈ [0, 2π ) or the antipodal phase aμ + π such that 2ai,μ = 2aμ
modulo 2π for all i = 1, . . . , Nμ . Therefore, the clusters represented in Fig. 5.2b are
called antipodal type clusters. Note that with this formal definition of an antipodal
state, in-phase clusters belong to the class of antipodal clusters.

5.1.3 Mixed Type Cluster States

The third type of multicluster states combines the previous two types. The 2-cluster
state shown in Fig. 5.2c consists of one splay cluster and one antipodal cluster. We
call these states mixed type multicluster. Despite using several different uniformly
distributed initial conditions, we have not found multi cluster by numerical sim-
ulations. Note that uniformly distributed phases are closer to splay states than to
antipodal states. In fact, due to the co-stability with splay type multiclusters, mix
type multicluster states are very unlikely to find from uniformly distributed random
initial conditions. In previous studies mix type multiclusters have not even been men-
tioned [9]. In order to find these particular states, we used specially prepared initial
condition motivated by the results described in Sect. 5.5.
As we have seen before, the interaction of a cluster with an antipodal cluster
induces a modulation s(t) additional to the linear growth of the oscillator’s phase.
In contrast, the interaction with a splay cluster does not introduce any modulation.
Thus, the temporal dynamics of the oscillators in the antipodal cluster (μ = 1) have
si,1 (t) ≡ 0 while the oscillators in the splay cluster (μ = 2) show additional bounded
modulations si,2 (t), see Fig. 5.2c. For the oscillators of the splay cluster we plot the
time series of two representatives. We notice the temporal shift in the dynamics of
88 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

the two representatives of the splay cluster. The oscillators in the splay cluster are not
completely temporally synchronized. More specifically we have φi,1 (t) = 1 t + ai,1
with R2 (a1 ) = 1 for i = 1, . . . , N1 and φi,2 (t) = 2 t + ai,2 + si,2 (t) with R2 (a2 ) =
0 for i = 1, . . . , N2 .
Despite the complexity of the three types of multiclusters states, the structures can
be broken down into simple blocks. In fact, one-cluster states of splay and antipodal
type serve as building blocks in order to create more complex multicluster structures.
In the following sections we provide an in-depth analysis of these blocks.

5.2 Splay Type Multicluster States

As seen in Fig. 5.2, one-cluster of splay type serve as building blocks for multicluster
states. In the following, we describe this build-up explicitly.

5.2.1 Conditions for the Emergence of Splay Type

Multicluster States

The multicluster solutions of splay type are composed by the clusters from the con-
tinuous family S of phase-locked solutions with R2 (a) = 0 and different frequencies.
The following proposition describes them.
Proposition 5.2.1 System 5.1–(5.2) possesses the multicluster solution

i = 1, . . . , Nμ
φi,μ (t) = μt + ai,μ , (5.4)
μ = 1, . . . , M
j = 1, . . . , Nν
κi j,μν (t) = −ρμν sin( μν t + ai,μ − a j,ν + β − ψμν ), (5.5)
ν = 1, . . . , M
 
with pairwise different frequencies μ , μν := μ − ν , ρμν := 1 + μν /
 1
2 −2
) and ψμν := arctan( μν /) if and only if
R2 (aμ ) = 0 for all μ = 1, . . . , M and the frequencies ( 1 , . . . , M ) solve the
following system of equations

1 
M

μ = ρμν Nν cos(α − β + ψμν ), μ = 1, . . . , M. (5.6)

2N ν=1

Note that ρμν and ψμν are functions of μ − ν.

5.2 Splay Type Multicluster States 89

The proof of Proposition 5.2.1 is to be found in the Appendix A.3. Similarly to the
one-cluster case, multicluster solutions of splay type give rise to a (N − 2M − 1)-
dimensional manifold of solutions
 
S M := (φi,μ , κi j,μν ) : (φi,μ , κi j,μν ) as in(5.4)−(5.5), R2 (aμ ) = 0 for all μ = 1, . . . , M .

Let us remark that the collective frequencies (5.6) are only defined up to a con-
stant due to the phase-shift symmetry of system (5.1)–(5.2) while the frequency
difference is unaffected. An example of a 3-cluster solution of splay type is shown
in Fig. 5.3. The solution was obtained by integrating system (5.1)–(5.2) numerically
starting from random initial conditions. After sufficiently long transient time, the
order of the oscillators is given by first sorting the oscillators with respect to their
average frequencies. After that the oscillators with the same frequency are sorted by
their phases. It can be seen from the pictures that the sizes of the three clusters Nμ
(μ = 1, 2, 3) possesses a hierarchical structure, i.e., N3 < N2 < N1 . The coupling
strengths between oscillators of the same cluster vary in a larger range than between
those of different clusters.The coupling between different clusters scales with 
 3 
since ρμν = / μν + O / μν and is thus close to zero (uncoupled). The
oscillators of the same cluster evolve in time with the same frequencies φ̇i,μ = μ,
i = 1, . . . , Nμ .

5.2.2 Two-Cluster States of Splay Type

Let us consider the case of two-clusters in more details. Let φi,μ (μ = 1, 2) with
N1 and N2 being the numbers of oscillators in cluster 1 and 2, respectively. The
following result follows from the Proposition 5.2.1.
Corollary 5.2.2 Suppose R2 (aμ ) = 0 for μ = 1, 2, then

φi,1 = 1t + ai,1 , i = 1, . . . , N1
φi,2 = 2 t + ai,2 , i = 1, . . . , N2
κi j,μμ = − sin(ai,μ − a j,μ + β), μ = 1, 2
κi j,μν = −ρμν sin( μν t + ai,μ − a j,ν + β − ψμν ), μ, ν = 1, 2; μ = ν

is a two-cluster solution of system (5.1)–(5.2) with

1 1 1 1 2
( 12 )1,2 = n1 − cos(α − β) ± n1 − cos2 (α − β) − 2(2 + sin(α − β)),
2 2 2 2
(5.7)
90 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

Fig. 5.3 Three-cluster of splay type. a Coupling weights at t = 10000 showing three clusters; b
Distribution of the phases within each cluster; space-time raster plot; c Average frequency of oscil-
lators; each plateau corresponds to one cluster; d Oscillator phases φi (t) at fixed time t = 10000.
Parameter values:  = 0.01, α = 0.3π , β = 0.23π , and N = 100. Figure reproduced from [2].
2019
c Society for Industrial and Applied Mathematics. All rights reserved

1 
μ = n μ cos(α − β) + ρμν n ν cos(α − β + ψμν ) , (μ, ν = 1, 2; μ = ν),
2
(5.8)
where n μ = Nμ /N and ψμν , ρμν as in Proposition 5.2.1.
The explicit expressions for the frequencies 12 , 1,2 and other parameters of the
solutions follow from the system of equations (5.6), which can be solved explicitly
leading to (5.7)–(5.8) for M = 2. For any given relative cluster size n μ , Eqs. (5.7)
and (5.8) provide either two, one, or no solutions corresponding to the two-cluster
solution. Hence, for each fixed set of parameters, there may be up to 2(N − 4) such
solutions.
5.2 Splay Type Multicluster States 91

Fig. 5.4 The figures show all one- and two-cluster solutions of splay type for the system (5.1)–
(5.2). For this, the frequency differences 12 are displayed corresponding to the Eqs. (4.6) and
(5.7). The dotted lines (black) indicate unstable solutions while the solid lines (blue) indicate
stable solutions. Here, every second solution is plotted for the sake of visibility. Parameter values:
a N = 20,  = 0.01; b N = 50,  = 0.01; c N = 50,  = 0.001; d N = 50,  = 0.1; α = 0.3π
is fixed for all panels. Figure reproduced from [2]. 2019
c Society for Industrial and Applied
Mathematics. All rights reserved

Figure 5.4 shows the frequency differences 12 of these solutions as functions

of parameter β for different number of oscillations N and adaptation parameters .
Interestingly, the frequencies of the solutions depend only on the difference α − β,
see (5.7)–(5.8). By increasing the number of oscillators N in the system, the num-
ber of solutions increases accordingly. This can be seen from Fig. 5.4a–b where we
increase the number of oscillators from N = 20 to N = 50 with all other parameters
fixed. The set of 2-cluster solutions is represented by all 12 (β) for a given param-
eter β. In accordance with (5.7) the number of solutions increases with increasing
N . The region of non-existence of the multicluster solutions corresponds to the cases
where the argument beneath the root in (5.7) becomes negative. The size of the exis-
tence gap depends furthermore on the choice of the time separation parameter . This
can be seen by comparing Fig. 5.4b–d where we vary the value for .
92 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

5.2.3 Adaptation Rate Dependence for the Emergence of

Two-Cluster States

In this section we show the importance of the time-separation parameter  for the
appearance of multicluster states. In particular, we obtain the critical value c above
which the multicluster states cease to exist.
It is quite remarkable that in case of splay-type clusters the multicluster solution
can be explicitly given as it was shown in the Proposition 5.2.1 and Corollary 5.2.2.
With this, we can study directly the role of several parameters for the existence of
multicluster states. In Fig. 5.4b–d, solutions for Eq. (5.7) are presented depending
on the parameter β. The number of oscillators in the system is chosen as N = 50.
Each line 12 (β) in Fig. 5.4b–d represents a frequency difference of two clusters
for which the two-cluster state of splay type exists with fixed relative number n 1
of oscillators in the first cluster. Note that the number of possible two-cluster states
increases proportionally to the total number of oscillators N . Different panels show
solutions for different values of . We note that the existence of those two-cluster
states depends only on the difference of γ := α − β, see Eq. (5.7). The necessary
condition for the existence of a two-cluster state reads
2
1
n1 − cos2 γ > 2(2 + sin γ ). (5.9)
2

From Eq. (5.9) we immediately see that the value of the time separation parameter
epsilon in system (5.1)–(5.2) is important for the existence of the multicluster states.
This dependence is in contrast to the findings for one-cluster states. First of all, note
that the left hand side of condition (5.9) is positive for γ = ±π/2. Hence, for all
parameters, there is a critical value c such that there exists no two-cluster state for
 > c . Explicitly, we have

2
1 1 1 2 1
c = − sin γ + sin γ + n 1 − cos2 γ , (5.10)
4 2 4 2

which is illustrated in Fig. 5.5. The figure shows the critical value c depending on
the parameter γ for different values of n 1 . The function possesses a global maximum
with c = 0.5. This means that there is a particular requirement on the time separation
in order to have two-cluster states of splay type. Indeed, the adaptation of the network
has to be at most half as fast as the dynamics of the oscillatory system.
Further let us remark that the two-cluster state with equally sized clusters n 1 = 0.5
exists only for α − β ∈ (π, 2π ), i.e., c = 0 for all α − β ∈ [0, π ].
In the subsequent sections we discus that the combination of one-cluster states to
a multicluster state can result in modulated dynamics of the oscillators additional to
the linear growth. In fact, this additional temporal behavior is due to the interaction
of the clusters. As we see in Fig. 5.2a, oscillators interacting with a splay-type cluster
will not be forced to perform additional dynamics. This is the reason why we are
5.2 Splay Type Multicluster States 93

n1 = 0.9
n1 = 0.7
n1 = 0.5

c

γ/π

Fig. 5.5 For the case of two-cluster states of splay type, the critical value c of time-separation
parameter  is plotted as a function of γ = α − β for different cluster sizes n 1 = N1 /N . The
function is given explicitly by Eq. (5.10). Figure reproduced from [3]. 2019
c AIP Publishing

able to derive a closed analytic expression for multicluster states of splay type. It is
possible to determine the frequencies explicitly as in Eq. (5.7). Therefore, here the
interaction between cluster causes only changes in the collective frequencies which
are small whenever  is small.
In contrast to splay clusters, the interaction with antipodal cluster leads to bounded
modulation of the oscillator dynamics besides the constant-frequency motion. The
modulations scale with  and hence depend on the time-separation parameter. In
case of a mixed type two-cluster state, both interaction phenomena are present.
The oscillators in the antipodal cluster interact with the splay cluster leading to no
additional modulation. On the contrary, the phase oscillators of the splay cluster get
additional modulation via the interaction with the antipodal cluster.
In the following Sect. 5.3, we provide a general result on the asymptotic existence
of splay, antipodal, and mixed type multicluster states. In the Sects. 5.4 and 5.5, this
general result will be applied to give explicit expression for antipodal as well as
mixed type multiclusters.

5.3 Conditions for the Emergence of Multicluster

States—A Generalized Approach

In this section we give an analytic description of multicluster solutions in terms of

an asymptotic expansion. We consider therefore the expansion of r th order together
with a multi-time scale ansatz [10]
94 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

(r )

r
(r )
φi,μ (, t) := μ (τ0 , . . . , τr ) + ai,μ +  l pi,μ;l (t)
l=1 μ, ν = 1, . . . , M

r i, j = 1, . . . , Nμ
κi(rj,μν
)
(, t) :=  l ki j,μν;l (t)
l=0
(5.11)

where (r )
μ ∈ C (R
1 r +1
) is a function depending on the multi-time scales τl :=  l t.
We show under under which conditions this expansion describes the time evolution
for the system (5.1)–(5.2). The main result on the asymptotic expansion for (pseudo)
multicluster solutions reads as follows.
Proposition 5.3.1 Let r ∈ N. Suppose the system (5.1)–(5.2) possesses a (pseudo)
multicluster solution (φi,μ , κi j,μν ) with φi,μ (, t) = μ ()t + ai,μ + si,μ (, t) where
ai,μ ∈ T1 and the coupling matrix κi j,μν (, t) is given as the parametrization of
the pullback attractor defined in (A.14). Assume further that M1 clusters are of
antipodal type (2ai,μ = aμ ) and M2 are of splay type (R2 (aμ ) = 0). Then, the r th
order asymptotic expansion of φi,μ (, t) for t ∈ O(1/ r ) as  → 0 is given by


r  
(r ) (r ) (r )
φi,μ (, t) := μ,0 t + ai,μ + l μ,l t + pi,μ;l (t) μ, ν = 1, . . . , M
l=1
(5.12)

r
i = 1, . . . , Nμ
κi(rj,μν
)
(, t) :=  l ki j,μν;l (t), (5.13)
j = 1, . . . , Nν
l=0

where
(i) all coefficients of the expansion can be found inductively;
(ii) the first order approximation can be written as
⎛ ⎞
(1) nμ 
M
nν
φi,μ =⎝ (cos(α − β) − cos(α + β)) −  (1)
sin(α − β)⎠ t + ai,μ + pμ;1
2 2 μν
ν=M1 +1

for μ = 1, . . . , M1 , and
⎛ ⎞
⎜ nμ M
nν ⎟
(1)
φi,μ =⎜
⎝ 2 cos(α − β) −  (1)
sin(α − β)⎟
⎠ t + ai,μ + pi,μ;1 (t)
ν=M +1
2 μν 1
ν=μ

for μ = M1 + 1, . . . , M with
5.3 Conditions for the Emergence of Multicluster States—A Generalized Approach 95


M1
nν (1)
pμ;1 = −   cos(2 μν t + aμ − aν + α + β) μ = 1, . . . , M1
(1) 2
ν=1 4 μν
ν=μ


M1
nν (1)
pi,μ;1 = −   cos(2 μν t + 2ai,μ − aν + α + β); μ = M1 + 1, . . . , M
(1) 2
ν=1 4 μν

the coupling weights are given by

κi(1)
j,μμ = − sin(ai,μ − a j,μ + β),

κi(1)
j,μν = (1)
cos( (1)
μν t + ai,μ − a j,ν + β);
μν

(1)
and the cluster frequencies μ solve the system of equations
⎛ ⎞
⎜ 
M
nν ⎟
(1)
μ =⎝ μ;0 − (1)
sin(α − β)⎠
ν=1
2 μν
ν=μ

with

μ;0 = n μ sin(α) sin(β) μ = 1, . . . , M1

nμ
μ;0 = cos(α − β). μ = M1 + 1, . . . , M
2
(1) (1) (1)
where μ = 1, . . . , M, i, j = 1, . . . , Nμ and μν := μ − ν .

The proof makes use of several lemmas and is presented in the Appendix A.4. Overall,
we aim to describe the following particular form for the dynamical behaviour of the
phase oscillators. The phases of the oscillators φi,μ form a pseudo multicluster, c.f.
Definition 5.5.1. Further, the bounded modulations for the phases of each oscillator
are given as Taylor expansions in  with periodic coefficients that can be expressed
as Fourier sums with even modes. The strategy for the proof of the main result is as
follows.
1. Assume that the phases of the oscillators are given as finite Taylor sums in  with
periodic coefficients, which are represented as finite Fourier sums. With this, the
Eqs. (5.2) can be explicitly solved. Introducing the pull-back attractor provides
us with a unique expression for the asymptotic solutions (t → ∞). An explicit
form for the expansion in  of these solutions of the coupling weights κ is given
in Lemma A.4.2.
2. The solutions of the coupling weights depend on the Fourier modes of the periodic
expansion coefficients of the oscillators. An statement on their explicit dependence
is provided by Lemma A.4.3. More specifically, the expansion coefficients of the
96 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

couplings weights only consist of even modes whenever the expansion coefficients
for the phases of the oscillators consist only of odd modes.
3. The expressions for the coupling weights in Lemma A.4.2 are used to derive the
explicit form for the phase oscillators. More specifically, the expansions coeffi-
cients are derived such that they satisfy the Eqs. (5.1). Higher order terms which
contribute to a linear growth are absorbed in an expansion for the oscillator fre-
quencies.
4. Finally, we find an iterative scheme to determine all expansion coefficients of
the phases and coupling weights up to any order. Moreover, it is shown that the
coefficients provided by the iterative scheme are consistent with the assumption
on the expansion coefficients given in the beginning of the proof.
In the following this general result is applied to antipodal and mixed multicluster
in order to characterize their properties. Note that if we apply the expansion result
to splay type multicluster, i.e. M1 = 0, the expressions would agree with the Taylor
expansion in  of the explicit formulas presented in Proposition 5.2.1.

5.4 Antipodal Type Multicluster States

In this section we apply the results presented in Proposition 5.2.1 to antipodal type
multicluster states.

5.4.1 Asymptotic Conditions for the Emergence of Antipodal

Type Multicluster States

In the case when the oscillators are phase synchronized or in an antiphase relation
within each cluster, the situation is different to what was described for the splay type
multiclusters. Particularly, the linear growth of the oscillator phases within each clus-
ter is modulated by periodic or quasi-periodic terms of order . That is, the clusters
possess the form φi,μ (t) = μ t + ai,μ + pμ (t, ). Here, we give important neces-
sary conditions for the existence of such solutions and their asymptotic expansion in
. Additionally, we provide numerical results showing these solutions. In particular,
we present a system of equations for the cluster frequencies μ .
Proposition 5.4.1 Suppose 2ai,μ = aμ mod 2π for all μ=1, . . . , M and i = 1, . . . ,
Nμ . If system (5.1)–(5.2) possesses antipodal multicluster solution (φi,μ , κi j,μν ) then
its first asymptotic expansion in  is given by
5.4 Antipodal Type Multicluster States 97

(1)

M
nν
(1) (1)
φi,μ = μ t + ai,μ −   2 cos(2 μν t + aμ − aν + α + β),
(1)
ν=1
ν=μ
4 μν

κi(1)
j,μμ = − sin(ai,μ − a j,μ + β),
(1)  (1)
κi j,μν = (1)
cos( μν t + ai,μ − a j,ν + β), μ  = ν
μν

(1)
with the cluster frequencies μ up to first order in  whenever the following implicit
equation can be solved
⎛ ⎞
⎜ 
M
nν ⎟
(1)
μ = ⎝n μ sin(α) sin(β) −  (1)
sin(α − β)⎠ . (5.14)
ν=1
2 μν
ν=μ

(1) (1)
Here, μ = 1, . . . , M, i, j = 1, . . . , Nμ and μν := μ − ν .

This first order perturbation reveals a nonlinear modulation pμ , which is periodic or

(1)
quasi-periodic with the frequencies μν given by the differences in the frequencies
of the clusters.
Figure 5.6 shows the numerically obtained 3-cluster solution of antipodal type.
The dynamics of system (5.1)–(5.2) is shown after a sufficiently long transient so
that it represents dynamically stable solution (more on stability in Sect. 5.6). One
can clearly observe three clusters in the coupling matrix. Similarly to the previous
multicluster cases, we first sort the oscillators with respect to their average frequency
and subsequently by their phases. In contrast to the splayed distribution of the phases
described in Sect. 5.2, the oscillators within the clusters additionally form two groups,
in which the phases differ by π .
In order to observe the modulation of the cluster frequencies, time series for three
representative oscillators from each cluster are shown in Fig. 5.7a. The averaged
linear growth of the phases due to  μ t has been subtracted to show the modu-
lation. Such small but non-vanishing oscillations do not exist in the case of splay
type multiclusters. In addition, the black dashed lines show the modulation given
by Proposition 5.4.1 confirming that the asymptotic expansion describes the whole
temporal behaviour very well. Furthermore, Proposition 5.4.1 implies that the ampli-
(1) 2
tudes of the modulations are proportional to n ν /( μν ) . Thus, if the difference in
the frequencies is high the amplitude is small and vice versa. This relation is also
reflected by the power spectrum, see Fig. 5.7b. Figure 5.7b confirms that the fre-
quencies of the modulation oscillations correspond to the differences of the average
frequencies.
98 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

Fig. 5.6 Three-cluster of antipodal type. a Coupling weights at t = 10000 showing three clusters;
b Distribution of the phases within each cluster; space-time raster plot; c Average frequency of
oscillators; each plateau corresponds to one cluster; d Oscillator phases φi (t) at fixed time t =
10000. Parameter values:  = 0.01, α = 0.3π , β = −0.53π , and N = 100. Figure reproduced
from [2]. 2019
c Society for Industrial and Applied Mathematics. All rights reserved

5.4.2 Two-Cluster States of Antipodal Type

Let us consider the case of two clusters in more details. Let N1 and N2 = N − N1
be the numbers of oscillators in group 1 and 2, respectively. The following result
follows from Proposition 5.4.1.
Corollary 5.4.2 Suppose 2ai,μ = aμ for all μ = 1, 2 and i = 1, . . . , Nμ . If system
(5.1)–(5.2) possesses an antipodal multicluster solution (φi,μ , κi j,μν ) then its first
order asymptotic expansion in  is given by
5.4 Antipodal Type Multicluster States 99

(a) (b)
Cluster 1
Cluster 2 S(ω̄)
Cluster 3
φi,µ (t) − Ωµ t

time ω̄

Fig. 5.7 For 3-cluster solution from Fig. 5.6, panel a shows time series of an oscillator from one
of the clusters after subtracting the average linear growth φμ,i (t) −  μ t. The black dashed lines
show the corresponding analytic results from the asymptotic expansion in Proposition 5.4.1. b Power
spectrum of the time series given in (a). Figure modified from [2]. 2019
c Society for Industrial
and Applied Mathematics. All rights reserved

(1) (1) n2 (1)

φi,1 = 1 t + ai,1 −   2 cos(2 12 t + a1 − a2 + α + β),
(1)
4 12

(1) (1) n1 (1)

φi,2 = 2 t + ai,2 −   2 cos(2 12 t + a1 − a2 − α − β),
(1)
4 12

κi(1)
j,μμ = − sin(ai,μ − a j,μ + β),

κi(1)
j,μν = (1)
cos( (1)
μν t + ai,μ − a j,ν + β),
μν

with
 
(1) nν
μ = n μ sin(α) sin(β) −  (1)
sin(α − β)
2 μν

(1) (1) (1)

for μ = 1, 2, ν = μ, i, j = 1, . . . , Nμ , μν := μ − ν ,

  1 1 2 2 
(1)
12 1,2 = n1 −
2
sin(α) sin(β) ± n1 −
2
sin α sin2 β − sin(α − β).
2
(5.15)

This result follows directly from Proposition 5.4.1. It shows, in particular, that the
system of Eqs. (5.14) can be solved explicitly by (5.15) in case of two clusters.
Similarly to the splay multiclusters, for any fixed set of parameters and each n 1 ,
Eq. (5.15) can lead to two antipodal multiclusters with two different frequency dif-
ferences. Hence, a large number of antipodal two-clusters can coexist for the same
100 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

(a)
Δ Ω 12 (b)

β /π β /π
Fig. 5.8 Two-cluster solutions (upper panels) and one-cluster solutions (lower panels) of antipodal
type given by the asymptotic expansion in Corollary 5.4.2 and Proposition 4.1.1, respectively. For
(1)
this, the difference of the frequencies 12 is displayed corresponding to (5.15) and (4.6). The
dotted lines (black) indicate unstable solutions while the solid lines (blue) indicate stable solutions.
Here, every second solution is plotted for the sake of visibility. The insets show a blow-up of the
interval [−, ]. Parameter values: a N = 20,  = 0.01; b N = 50,  = 0.01; α = 0.3π is fixed for
all panels. Figure reproduced from [2]. 2019
c Society for Industrial and Applied Mathematics.
All rights reserved

parameter values. Figure 5.8 illustrates such a coexistence, where we present the one-
cluster solutions given by (4.6) and the solutions to the Eq. (5.15). Blue solid lines
represent those solutions for which the asymptotic expansion led to an existing and
stable two-cluster solutions of antipodal type. Note further that for two-cluster solu-
tions of antipodal type, the asymptotic expansion presented in Proposition 5.4.1 turns
into a formal expansion whenever | | > , i.e.,  is not assumed to be infinitesimal
( → 0). The interval [−, ] is therefore highlighted in Fig. 5.8.

5.5 Mixed Type Pseudo-multicluster States

In this section we apply the results presented in Proposition 5.2.1 to mixed type
multicluster states. In order to this the notion of multiclusters has to be slightly
adjusted which leads to the definition of pseudo multicluster states.
5.5 Mixed Type Pseudo-multicluster States 101

5.5.1 Asymptotic Conditions for the Emergence of Mixed

Type Pseudo-multicluster States

We have seen how clusters are described consisting of oscillator groups of splay type
(Sect. 5.2) as well as clusters consisting of oscillator groups with in- and anti-phase
relation (Sect. 5.4). It is therefore reasonable to ask for multicluster solutions that
consist of both of these types. In order to describe these solutions we have to loosen
the definition of a multicluster solution.
Definition 5.5.1 Phase oscillators φi (t) form a pseudo multicluster if they can be
separated into M groups such that for all μ ∈ {1, . . . , M} the phase oscillators φi,μ ,
i ∈ {1, . . . , Nμ }, from each group μ satisfy φi,μ (t) = μ t + si,μ (t) with bounded
functions si,μ .
Note that every multicluster solution is by definition already a pseudo multicluster
solution.
Proposition 5.5.2 Suppose 2ai,μ = aμ for all μ = 1, . . . , M1 , and R2 (aμ ) = 0 for
all μ = M1 + 1, . . . , M, i = 1, . . . , Nμ where M1 is the number of antipodal type
clusters. The mixed pseudo multicluster solutions of (5.1)–(5.2) with φi,μ (t) =
μ ()t + si,μ (t) + ai,μ possess the following first order asymptotic expansion in

(1) (1)
φi,μ = μ t + ai,μ + pi,μ;1 (t),
κi(1)
j,μμ = − sin(ai,μ − a j,μ + β),

κi(1)
j,μν = (1)
cos( (1)
μν t + ai,μ − a j,ν + β),
μν

with


M1
nν (1)
pi,μ;1 (t) = pμ;1 = −  2 cos(2 μν t + aμ − aν + α + β)
(1)
ν=1
ν=μ
4 μν

for μ = 1, . . . , M1 ,


M
nν (1)
pi,μ;1 (t) = −  2 cos(2 μν t + ai,μ − aν + α + β)
(1)
ν=μ
4
ν=1 μν

for μ = M1 + 1, . . . , M, and the cluster frequencies (1)

μ up to second order in 
whenever the following system of equations can be solved
102 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators
⎛ ⎞
⎜ 
M
nν ⎟
(1)
μ =⎝ μ;0 − (1)
sin(α − β)⎠ (5.16)
ν=1
2 μν
ν=μ

with

μ;0 = n μ sin(α) sin(β) μ = 1, . . . , M1

nμ
μ;0 = cos(α − β). μ = M1 + 1, . . . , M
2
(1) (1) (1)
Here, μ = 1, . . . , M, i, j = 1, . . . , Nμ and μν := μ − ν .

5.5.2 Pseudo-two-cluster States of Mixed Type

As in the previous sections, we are going to show that Eq. (5.16) possesses solutions.
For this, we consider the case of two clusters φi,μ (μ = 1, 2).
Corollary 5.5.3 Suppose 2ai,1 = a1 for all i = 1, . . . , N1 and R(a2 ) = 0. The
mixed pseudo multiclusters of system (5.1)–(5.2) possess the following first order
asymptotic expansion in 
(1) (1)
φi,1 = 1 t + ai,1 ,
(1) (1) n1 (1)
φi,2 = 2 t + ai,2 −   2 cos(2 12 t + a1 − ai,2 − α − β),
(1)
4 12

κi(1)
j,μμ = − sin(ai,μ − a j,μ + β),

κi(1)
j,μν = (1)
cos( (1)
μν t + ai,μ − a j,ν + β),
μν

where  
(1) nν
μ = μ;0 − (1)
sin(α − β) ,
2 μν

1;0 = n 1 sin(α) sin(β),

n2
2;0 = cos(α − β),
2
5.5 Mixed Type Pseudo-multicluster States 103

Fig. 5.9 2-Cluster solution of mixed type. a Coupling weights at t = 10000 showing two clusters,
b Distribution of the phases within each cluster, space-time representation. c Average frequency
of each oscillator, d Oscillator phases φi for fixed time t = 10000. Parameter values:  = 0.01,
α = 0.3π , β = −0.4π , N = 100. Figure reproduced from [2]. 2019
c Society for Industrial and
Applied Mathematics. All rights reserved

   
(1) 1 1 n1
12 = n1 − cos(α − β) − cos(α + β)
1,2 2 2 2
 2
1 1 n1
± n1 − cos(α − β) − cos(α + β) − 2 sin(α − β)
2 2 2
(5.17)

for μ = 1, 2, ν = μ and i, j = 1, . . . , Nμ .
Illustration of the mixed 2-clusters is shown in Fig. 5.9.
Moreover, we performed a Fourier analysis of the temporal behaviour of the
oscillators, see Fig. 5.10. First, it can be observed that the oscillators representing the
104 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

(a) (b)

Cluster 1 S(ω̄)
Cluster 2 (i = 10)
φi,µ (t) − Ωµ t

Cluster 2 (i = 30)

time ω̄

Fig. 5.10 For mixed type 2-cluster solution from Fig. 5.9, panel a shows time series of an oscilla-
tor from one of the clusters after subtracting the average linear growth φμ,i (t) −  μ t. The black
dashed lines show the corresponding analytic results from the asymptotic expansion in Proposi-
tion 5.5.2. b Power spectrum of the time series given in (a). Figure modified from [2]. 2019 c
Society for Industrial and Applied Mathematics. All rights reserved

second cluster (i = 10, 30) show the same evolution in time but with a phase lag due
to the spatial dependency described above. In order to show the agreement with the
asymptotic expansion presented in Corollary 5.5.3, the analytic results are displayed
with black dashed lines. Furthermore, the power spectrum shows a prominent peak
at 2 12 for both oscillators of the second cluster and a flat curve for the repre-
sentative of the first cluster. These numerical results are in complete agreement with
the analytic findings.
Analogously, to the antipodal two-clusters, for any fixed set of parameters and
each n 1 , Eq. (5.17) can lead to two multiclusters of mixed type with two different
frequency differences. Hence, a large number of those clusters can coexist for the
same parameter values. Figure 5.11 illustrates such a coexistence, where we present
the solutions to the Eq. (5.17). Again, blue solid lines represent those solutions for
which the asymptotic expansion led to an existing and stable two-cluster solutions
of mixed type. Additionally, Fig. 5.11 shows the one-cluster solutions of splay and
antipodal type (in both cases 12 = 0) together with their common regions of
stability. As in the case of two-clusters of antipodal type, the asymptotic expansion
presented in Proposition 5.4.1 turns into a formal expansion whenever | | > .
The interval [−, ] is therefore highlighted in Fig. 5.11.

5.6 Stability of Multicluster States

In Sect. 5.2 we discussed multicluster solutions of splay type and showed under
which condition they exist. The solutions for two-cluster solutions of splay type and
their stability are presented in Fig. 5.4. In Fig. 5.4b the solution for the case of 50
oscillators is shown. The solid lines (blue) correspond to solutions that are stable. It
5.6 Stability of Multicluster States 105

Fig. 5.11 Two-cluster solutions of mixed type (upper panels) and one-cluster solutions (lower
panels) of either splay or antipodal type given by the asymptotic expansion in Corollary 5.5.3
(1)
and Proposition 4.1.1, respectively. For this, the difference of the frequencies 12 is displayed
corresponding to (5.17) and (4.6). The dotted lines (black) indicate unstable solutions while the
solid lines (blue) indicate stable solutions. Here, every second solution is plotted for the sake of
visibility. The insets show a blow-up of the interval [−, ]. Parameter values: a N = 20,  = 0.01;
b N = 50,  = 0.01; α = 0.3π is fixed for all panels. Figure reproduced from [2]. 2019
c Society
for Industrial and Applied Mathematics. All rights reserved

can be seen that whenever a 2-cluster solution is stable the one-cluster solution (with
12 = 0) is also stable.
A more detailed validation of this statement is presented in Fig. 5.12, where we
show the stability regions of both one- and two-cluster solutions in the (α, β) plane.
The stability for each type of cluster solution is determined numerically. The numer-
ical approach was already introduced in Sect. 4.2. For the two-cluster solutions the
norm for the phase configuration is calculated in the relative coordinates given by
i,μ = φi,μ − φ1,μ with μ = 1, 2. Additionally, we calculated the maximal value of
all inter-cluster connections and compared it to the theoretical maximum given by
ρ12 . If after numerical integration the maximal inter-cluster coupling is bigger than
ρ12 + 0.01, the two-cluster is considered as unstable. Here, region where the both
types of solutions are stable are colored in dark blue. Regions of only stable one-
cluster solutions are colored in light blue. Since two-cluster solutions do not exist for
certain values of α and β, we can find a light blue stripe in the middle of Fig. 5.12.
Further, we have not found any configuration of α and β for which two-cluster solu-
tions are stable and one-cluster are not. This supports the claim that the stability of a
one-cluster solution is necessary condition for the stability of a two-cluster solution.
This can be explained by the fact that for the stability of the multiclusters, it is neces-
106 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

α /π

β/π

Fig. 5.12 Stability diagram for the one-cluster and two-cluster solution of the splay type depending
on the parameters α and β. Yellow region corresponds to the instability of both solutions, dark blue
to the stability of both solutions, and light-blue to the stability of only the one-cluster solution.
Parameter  = 0.01 is fixed in all simulations. Figure reproduced from [2]. 2019 c Society for
Industrial and Applied Mathematics. All rights reserved

sary that its one-cluster components are each stable with respect to the perturbations
that disturb the structure of just one cluster (see similarly in [11]). A more rigorous
formulation of this issue is beyond the scope of this thesis.
Figure 5.4b further provides us with information about the stability of two-cluster
solutions depending on the ratio between cluster sizes. First, due to (5.7) there exist
two branches of two-cluster solution of splay type. Only solutions with higher fre-
quency difference are stable which can be seen in the inset of Fig. 5.4b. For an
increasing number of oscillators in the second cluster of relative size n 2 = 1 − n 1
the stability changes. Above a certain value of n 2 both branches are unstable. This
observation explains why only multicluster solutions with unequal as well as hierar-
chical cluster sizes were found in simulations, see Fig. (5.3) and [1].

5.6.1 On the Stability of Multicluster States with Evenly Sized

Clusters

In the following, we give a necessary condition for the stability of all one-cluster
states of splay type complementing the result on rotating-wave states given in Propo-
sition 4.2.2. This extension enables us to determine the stability of evenly sized
multiclusters of splay type.
In general all splay one-cluster states have the property R2 (a) = 0 for the phase
given by the vector a. Therefore, the splay states form N − 2 dimensional family of
solution. Hence, around each splay states there are N − 2 neutral variational direc-
tions (δφ, δκ)T which are determined by the condition Nj=1 ei2a j δφ j = 0. Note,
 
δκi j = − cos(ai − a j + β) δφi − φ j in neutral direction.
5.6 Stability of Multicluster States 107

Proposition 5.6.1 Consider an asymptotically stable one-cluster state of splay type.

Then,  + sin(α − β)/2 > 0.

Proof For some notation we refer the reader to Appendix A.2. Due to the block form
of the linearised Eq. (A.6) and the Schur decomposition (4.11), any eigenvalue comes
with a second. We have already seen this in Lemma A.2.1 and Proposition A.2.3.
Variation along the neutral
 direction gives N − 2 times the eigenvalue 0. Suppose

we have δφ such that Nj=1 ei2a j δφ j = 0 and δκi j = − cos(ai − a j + β) δφi − φ j .
Applying Schur decomposition (4.11), we get

δφ I N −( + λ)B (A − λI N ) + 1
+λ BC 0 δφ
(M − λI) N 2 +N = = 0.
δκ 0 IN 2 0 −( + λ) − +λ
1
Cδφ + δκ
(5.18)

With this, we have to find λ such that the last equality in (5.18) is fulfilled. This
is equivalent to solving ((A − λI N )( + λ) + BC) δφ = 0 of which in general only
N − 2 equations are linearly independent. The equivalence can be seen by multiply-
ing  + λ from both sides and keeping in mind that δκ is already determined by δφ.
Using the definition of δφ the matrices A and BC can be effectively reduced in such
a way that they are independent of the actual values for the phases a j . In fact,

−1
− N2N sin(α − β) i = j
ai j = 1
2N
sin(α − β), i = j

−1
 N2N sin(α − β), i = j
(bc)i j = 
− 2N sin(α − β). i = j

In turn, this gives ((A − λI N )( + λ) + BC) a circulant structure which can be used
to diagonalise the matrix, in analogy to Proposition A.2.3. For circulant matrices we
immediately know the eigenvalues. They are

⎛ ⎛ ⎞ ⎞

N −1
N −1 1
μl = −λ2 − ⎝ sin(α − β) − sin(α − β) ⎝ ei2π kl/N − 1⎠ +  ⎠ λ
2N 2N
k=0

with l = 0, . . . , N − 1 and det ((A − λI N )( + λ) + BC) = μ0 (λ) · · · μ N −1 (λ).

Remember we have in general N − 2 independent equations. Thus, solving μl (λ) =
0 for λ results in N − 2 eigenvalues λ = 0, 1 eigenvalue λ = − and N − 3 eigen-
values λ = − − sin(α − β)/2. Note that for 4-phase-cluster states, as considered in
Corollary A.2.5, the number of independent equations is N − 1. This  is due to the fact
that in this case the equations for the imaginary and real part from Nj=1 ei2a j δφ j = 0
agree. 

Finally, we note why the equally-sized splay-clusters are not found to be stable.
Indeed, from Eq. (5.7) we know that 2 + sin(α − β) < 0 is a necessary condition
108 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

to have such equally-sized (n 1 = 1/2) clusters. However, any one-cluster splay state
is unstable for 2 + sin(α − β) < 0 by Proposition 5.6.1.

5.6.2 An Effective Approach for the Stability of Multicluster

States

As mentioned above, the stability of one-cluster states is important for the stability
of multicluster states. For two weakly coupled clusters, the stability of one-clusters
serves as a necessary condition for the stability of the two-cluster state. In Fig. 5.4
the possible two-clusters of splay type are plotted. We notice that for  = 0.001 a
stable two-cluster state exists for almost every relative cluster size n 1 , while this is
not true for  = 0.01 and even more so for  = 0.1.
Another observation from Fig. 5.4b, c is that the possible β-values where the two-
cluster states can be stable mainly correspond to the β values where the one-cluster
state is stable. This is true for small values of , however, a careful inspection of
Fig. 5.4d for the case of larger , here  = 0.1, shows that some two-cluster states
appear to be stable for a parameter region where the corresponding one-cluster state
is unstable. This can be explained as follows. According to (5.1), in the case of
one-cluster states, the inter-cluster interactions are summed over all N oscillators
of the whole system. Additionally, the interactions are scaled with the factor 1/N .
Therefore, the total interaction scales with 1. For two-cluster states, the inter-cluster
interactions for each individual cluster are only a sum over the Nμ (μ = 1, 2) oscilla-
tors whereas the scaling remains 1/N . Hence, the total inter-cluster interaction scales
with n μ = Nμ /N , the relative size of the cluster. Therefore, the effective oscillatory
system, when neglecting the interaction to the other cluster, reads

nμ 
Nμ
dφi,μ
=− κi j,μμ sin(φi,μ − φ j,μ + α),
dt Nμ j=1
dκi j,μμ  
= − κi j,μμ + sin(φi,μ − φ j,μ + β) .
dt
This system is equivalent to (5.1)- (5.2) with Nμ oscillators by rescaling  → /n μ .
Thus, the stability of the inter-cluster system has to be evaluated with respect to the
rescaled effective parameter eff := /n μ . Since n μ < 1 for μ = 1, 2, we have eff >
. As we have discussed, the stability for the one-cluster changes with increasing .
With this, the influence of the cluster size as well as the slight boundary shift in the
regions of stability, see Fig. 5.4, can be explained.
5.7 Summary 109

5.7 Summary

In this chapter, we have focused on multicluster solutions in a network of adap-

tively coupled phase oscillators. The multicluster states are composed of several
one-clusters with distinct frequencies. Starting from random initial conditions, our
numerical simulations show two different types of states. These are the splay and
the antipodal type multicluster states. A third mixed type multicluster state is found
by specially prepared initial conditions. For all these states the collective motion
of oscillators, the shape of the network, and the interaction between the frequency
clusters is presented in detail. It turns out that the oscillators are able to form groups
of strongly connected units. The interaction between the groups is weak compared
to the interaction within the groups. The numerical analysis of multicluster states
reveals the building blocks for these states.
While the one-cluster solutions are relative equilibria of our system due to the
phase-shift symmetry, the multicluster solutions contain components with different
frequencies, and, hence, they cannot be reduced to an equilibrium by transform-
ing into another co-rotating frame. As a result, the study of multiclusters is more
involved. However, to our surprise, we have still been able to find an explicit form
of multiclusters with the components of the splay type. Remarkably, in addition to
its ring-like spatial structure that dynamically emerges, the network behaves in such
a case (quasi-)periodically in time such that the whole solution can be interpreted as
a spatial-temporal wave.
The analysis of multicluster solutions of antipodal type is more subtle due to the
modulation of the frequency. More specifically, we look at multiclusters with bounded
frequency modulation. For these types of multiclusters, we derive an asymptotic
expansion in the parameter  that gives explicit existence conditions. In addition,
we have shown the existence of mixed multiclusters, which consist of clusters of
splay type and clusters of antipodal type. For the mixed multiclusters, the temporal
behavior within one cluster has been shown to be slightly non-identical, namely, the
oscillators possess the same averaged frequency, but they still can have a bounded
quasi-periodically modulated phase difference.
For the splay clusters we analytically show the existence of two-cluster states.
Remarkably, while the existence of the one-cluster states does not depend on the
time-scale separation parameter , the multicluster states crucially depend on the
time-scale separation. In fact, we provide an analysis showing that there exists a
critical value for the time-scale separation. Moreover, we show that in the case of
two-cluster states of splay type the adaptation of the coupling weights must be at most
half as fast as the dynamics of the oscillators. This fact is of crucial importance for
comparing dynamical scenarios induced by short-term or long-term plasticity [12].
The stability of two-cluster states is analyzed numerically and presented for dif-
ferent values of the time-scale separation parameter. By assuming weakly interacting
clusters, we describe the stability of the two-cluster with the help of the analysis of
one-cluster states. The main messages from this analysis are as follows: there is a
high degree of coexistence of stable multiclusters that can be reached from differ-
110 5 Multicluster States in Adaptive Networks of Coupled Phase Oscillators

ent initial conditions; in particular, a certain amount of imbalance in the number of

oscillators within the clusters is needed to achieve stability. This explains the appear-
ance of only hierarchical structures in numerical simulations. In fact, the simulations
indicate that there are no stable two-cluster states with clusters of the same size. We
provide a rigorous argument to understand this property of the system.
Moreover, the findings on multicluster solutions as they are reported in this chapter
are in very good agreement with previous results on adaptive neural networks [13–
15]. Here, stable multicluster solutions of coherently spiking neurons with weak but
time-dependent inter-cluster coupling are reported. With this work we shed some
light on these generic time-dependent network patterns.

References

1. Kasatkin DV, Yanchuk S, Schöll E, Nekorkin VI (2017) Self-organized emergence of multi-

layer structure and chimera states in dynamical networks with adaptive couplings. Phys Rev E
96:062211
2. Berner R, Schöll E, Yanchuk S (2019) Multiclusters in networks of adaptively coupled phase
oscillators. SIAM J Appl Dyn Syst 18:2227
3. Berner R, Fialkowski J, Kasatkin DV, Nekorkin VI, Yanchuk S, Schöll E (2019) Hierarchical
frequency clusters in adaptive networks of phase oscillators. Chaos 29:103134
4. Nichols S, Wiesenfeld K (1992) Ubiquitous neutral stability of splay-phase states. Phys Rev
A 45:8430
5. Strogatz SH, Mirollo RE (1993) Splay states in globally coupled Josephson arrays: analytical
prediction of Floquet multipliers. Phys Rev E 47:220
6. Choe CU, Dahms T, Hövel P, Schöll E (2010) Controlling synchrony by delay coupling in
networks: from in-phase to splay and cluster states. Phys Rev E 81:025205(R)
7. Burylko O, Pikovsky A (2011) Desynchronization transitions in nonlinearly coupled phase
oscillators. Phys D 240:1352
8. Ashwin P, Bick C, Burylko O (2016) Identical phase oscillator networks: bifurcations, sym-
metry and reversibility for generalized coupling. Front Appl Math Stat 2
9. Kasatkin DV, Klinshov V, Nekorkin VI (2019) Itinerant chimeras in an adaptive network of
pulse-coupled oscillators. Phys Rev E 99:022203
10. Verhulst F (2006) Methods and applications of singular perturbations: boundary layers and
multiple timescale dynamics. Springer, Berlin
11. Lücken L, Yanchuk S (2012) Two-cluster bifurcations in systems of globally pulse-coupled
oscillators. Phys D 241:350
12. Fröhlich F (2016) Network neuroscience. Academic, Cambridge
13. Popovych OV, Xenakis MN, Tass PA (2015) The spacing principle for unlearning abnormal
neuronal synchrony. PLoS ONE 10:e0117205
14. Chakravartula S, Indic P, Sundaram B, Killingback T (2017) Emergence of local synchroniza-
tion in neuronal networks with adaptive couplings. PLoS ONE 12:e0178975
15. Röhr V, Berner R, Lameu EL, Popovych OV, Yanchuk S (2019) Frequency cluster formation
and slow oscillations in neural populations with plasticity. PLoS ONE 14:e0225094
 Part II
Interplay of Adaptivity and Connectivity
Chapter 6
Adaptation on Nonlocally Coupled Ring
Networks

In this chapter, we go beyond the model of globally coupled oscillators which were
in the focus of the Chaps. 3–5. In particular, here, we study the interplay of adaptivity
with a non-locally coupled ring structure. Ring-like structures are important motifs
in neural networks [1–4]. Specifically, nonlocally coupled rings where each node is
coupled to all nodes within a certain coupling range, are known to be important for
systems appearing in many applied problems and theoretical studies [5–18].
In the subsequent sections, we consider the following system of N adaptively
coupled identical phase oscillators [19–27]

dφi 1 N
= 1 − N ai j κi j sin(φi − φ j + α) (6.1)
dt j=1 ai j j=1
dκi j  
= − κi j + sin(φi − φ j + β) (6.2)
dt

where i, j = 1, . . . , N , φi ∈ S 1 are the phases of the oscillators, ai j ∈ {0, 1} are the

entries of the adjacency matrix A determining the base topology, κi j ∈ [−1, 1] are
slowly changing adaptive coupling strengths, 0 <   1 is the rate of the adaptation,
and α, β are coupling and adaptation phase lags. Note that the natural oscillation
frequency has been normalized to 1 by the rotating coordinate frame.
The base topology given by the adjacency matrix A determines the structure of
the network, on which the adaptation takes place. Equation (6.2) for the adaptation
is used only for “active” weights κi j corresponding to ai j = 1. Similarly, the sum in
(6.1) goes over these links. Here we consider the topology of a nonlocally coupled
ring given by 
1 for 0 < (i − j) mod N ≤ P,
ai j = (6.3)
0 otherwise.

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 113
R. Berner, Patterns of Synchrony in Complex Networks of Adaptively Coupled Oscillators,
Springer Theses, https://doi.org/10.1007/978-3-030-74938-5_6
114 6 Adaptation on Nonlocally Coupled Ring Networks

This means that any two oscillators are coupled if their indices i and j are separated
at most by the coupling radius P. The coupling Eq. (6.3) defines a nonlocal ring
structure with coupling range P to each side and two special limiting cases: local
ring for P = 1 and globally coupled network for P = N /2 (if N is even, else P =
(N + 1)/2).
 The matrix of the form (6.3) is circulant [28] and has constant row
sum, i.e., Nj=1 ai j = 2P for all i = 1, . . . , N . Due to the same symmetries as they
are present in (2.25)–(2.26), the analysis can be restricted to the parameter regions
α ∈ [0, π/2) and β ∈ [−π, π ).
Two measures of coherence are used in this chapter: first, the nth moment of the
ith (i = 1, . . . , N ) complex local order parameter as given by

1 
N
(n)
Z i(n) (φ) := ai j einφ j = Ri(n) (φ)eiϑi (φ) (6.4)
2P j=1

where φ = (φ1 , . . . , φ N )T , Ri(n) (φ) is the nth local order parameter and ϑi(n) (φ) the
nth local mean-phase; second, the complex (global) order parameter

1  inφ j
N
(n)
Z (n) (φ) := e = R (n) (φ)eiϑ (φ) (6.5)
N j=1

where R (n) (φ) is the nth (global) order parameter and ϑ (n) (φ) the nth (global) mean-
phase, see also Sect. 2.2. Both measures are used throughout the chapter to charac-
terize asymptotic states of (6.1)–(6.2).
This chapter reproduces contents that have been published in [29]. The chapter is
organized as follows. Numerical results and a rigorous definition for multicluster and
solitary states on complex networks are presented in Sect. 6.1. In Sect. 6.2 we provide
a more detailed analysis of one-cluster states. Here, relations between local and global
properties are derived. The salient role of rotating-wave clusters are underlined and
the crucial dependence of their stability on the coupling range and the wavenumber
are rigorously described. Some of the proofs are presented in the Appendixes A.5
and A.6. After this, we focus on the analysis of solitary states in Sect. 6.3. A reduced
model for two-clusters is derived and a variety of bifurcations in which solitary states
are born and stabilized are presented. Section 6.5 summarizes our findings.

6.1 Multicluster and Solitary States

This section is devoted to the numerical analysis of system (6.1)–(6.2) and the
description of several dynamical states which occur for this system. More specifically,
we report one-cluster, multicluster, and solitary states. While this section describes
the states in a phenomenological fashion, more rigorous results are presented in the
subsequent Sects. 6.2–6.3.
6.1 Multicluster and Solitary States 115

Note that the observed one-cluster and multicluster states are similar to those
reported in Chaps. 4 and 5 for the all-to-all coupling base topology. However, there
are important differences due to the ring structure of our system, which will be
discussed in detail.
For the numerical simulations in this section a system of N = 100 oscillators
with coupling range P = 20 is studied. The value of  is set to 0.01 and the system
parameter α and β are varied in the ranges [0, π/2] and [−π, π ], respectively. All
results are obtained starting from uniformly distributed random initial conditions and
a simulation time of t = 20000. Several types of synchronization patterns are found
in the numerical simulations, depending on the values of α and β.

6.1.1 One-Cluster States

A one-cluster is defined as a frequency synchronized state

φi = t + χi , i = 1, . . . , N

with a collective frequency ∈ R and individual phase shifts χi ∈ [0, 2π ), see

Sect. 4.1. The two types of one-cluster states found in the numerical simulations
are either of antipodal or splay type whose asymptotic configurations are displayed
in Fig. 6.1a–d, respectively. The antipodal and splay-type clusters have been intro-
duced previously in Chap. 4. In the antipodal cluster, all phases φi are either in-phase
or in anti-phase, i.e., χi ∈ {χ , χ + π } with χ ∈ [0, 2π ) and hence R (2) (φ) = 1. In
the splay cluster the phases are distributed across the interval [0, 2π ) such that the
global second order parameter, as defined in Eq. (6.5), vanishes, i.e., R (2) (φ) = 0.
In Fig. 6.1a, c the coupling structures corresponding to the two types of one-clusters
are displayed. Note that the coupling weights are solely described by the phase dif-
ferences of the oscillators and are given by

κi j = − sin(χi − χ j + β).

The one-cluster states, which exist in our ring case and the all-to-all base topology
case from Chap. 4 have the same representation except for the fact that some of the
coupling weights are absent in the case of the ring, see empty entries in Fig. 6.1a, c.

6.1.2 Multicluster States

As described in Chap. 5, one-clusters can serve as building blocks for multi-frequency

clustered states where the phase dynamics and the coupling matrix κ are divided into
different groups; κi j,μν refers to the coupling weight for the connection from the ith
oscillator of the μth cluster to the jth oscillator of the νth cluster. Analogously, φi,μ
116 6 Adaptation on Nonlocally Coupled Ring Networks

κij
−1.0 −0.5 0 0.5 1.0
(a) (b)
100 2π

80

60
π

φi
j

40

20

0
(c) (d)
100 2π

80

60
π
φi
j

40

20

0
20 40 60 80 100 20 40 60 80 100
i i

Fig. 6.1 Illustration for two types of one-cluster states. The panels a, c show the asymptotic
coupling matrices and b, d snapshots of the phases at a fixed time. Results for the one-cluster
states of antipodal type are presented in (a, b) where α = 0.19π , β = −0.66π and of splay type in
(c, d) where α = 0.35π , β = 0.01π . Parameters: N = 100, P = 20,  = 0.01. Figure reproduced
from [29]

denotes the ith phase oscillator in the μth cluster. The temporal behavior for each
oscillator in an M-cluster state takes the form, see also Definition 5.0.1

μ = 1, . . . , M
φi,μ (t) = μt + χi,μ + si,μ (t)
i = 1, . . . , Nμ

where M is the number of clusters, Nμ is the number of oscillators in the μth cluster,
χi,μ ∈ [0, 2π ) are phase lags, and μ ∈ R is the collective frequency of the oscillators
in the μth cluster. The functions si,μ (t) are assumed to be bounded.
Both types of one-clusters give rise to multi-frequency cluster states, see Fig. 6.1,
similarly to the case of all-to-all base coupling. However, in case of more com-
plex network structures the definition of a frequency-cluster has to be refined to
account for the connectedness of the individual building blocks. Therefore, a multi-
frequency-cluster (shortly: multicluster) consists of groups of frequency synchronized
oscillators for which the subnetwork (or subgraph), induced by the individual groups
6.1 Multicluster and Solitary States 117

II

Fig. 6.2 Schematic figure illustrating the definition of multicluster and subnetworks induced by
groups of nodes with the same average frequency. The full network (left) consists of N = 20
nodes and has a nonlocal ring structure with P = 4. The colors of the nodes indicate their average
frequencies. Clusters are shown by the equally colored nodes that form connected sub-networks.
Even though the two blue groups I and II possess the same averaged frequencies, they form two
different clusters, since they are not connected. Figure reproduced from [29]

of nodes, is connected. Here, we say a network is connected if there is directed path

from each node to every other node of the network. In case of a directed graph, we
require the property of weak connectedness for the induced subgraph for the cluster.
For an introduction to the terminology we refer the reader to [30].
Let us illustrate the above definition. Consider a nonlocal ring network of N = 20
oscillators with coupling range P = 4 as presented in Fig. 6.2. Suppose that for each
node of the network we have a certain average frequency which is indicated by the
color. In Fig. 6.2, we have three different average frequencies denoted by the green,
blue, and red colors. The individual clusters are given by the connected subnetworks
induced by equally colored nodes. Note that even though the blue nodes have the same
average frequency, they are forming two different clusters (I, II) corresponding to two
connected components. Note that the induced subnetworks are not necessarily regular
even if the base topology is regular, see for instance the red or green subnetworks.
Figure 6.3 shows two-cluster states of antipodal (Fig. 6.3a–c) and splay type
(Fig. 6.3d–f). In both cases, Fig. 6.3c, f show that there are two distinct groups of
oscillators with different averaged frequencies (green, red). It is easily verified that
these groups form connected subnetworks and, hence, they form a two-cluster state.
Due to the frequency difference between the individual clusters, the groups of oscil-
lators decouple effectively. This can be seen in Fig. 6.3a, d where only oscillators of
the same cluster are strongly coupled compared to the coupling between the clusters
given by the respective coupling weights κi j,μμ and κi j,μν ≈ 0 (μ = ν).
118 6 Adaptation on Nonlocally Coupled Ring Networks

Fig. 6.3 Illustration of the different types of multicluster states. The panels a, d, g show the
coupling matrix, b, e, h phase snapshots and c, f, i average frequencies. a–c: antipodal two-cluster
for α = 0.23π , β = −0.56π ; d–f: splay two-cluster for α = 0.19π , β = −0.45π ; g–i: antipodal
five-cluster (I, II denote the two connected components of the red and the blue clusters) for α = 0.3π ,
β = −0.53π . Parameters: N = 100, P = 20,  = 0.01. Figure reproduced from [29]

Snapshots of the phase distributions are presented in Fig. 6.3b, e. Figure 6.3b is
showing the phase distribution of an antipodal cluster. In contrast to the case of
global coupling, see Chaps. 4 and 5, the phases do not possess the exact antipodal
property anymore. In fact, the scattering of the antipodal phase distribution is caused
by the structure of the induced subnetwork which is not regular anymore for the
in-degree of each node, i.e., the subnetwork does not have constant row sum, see e.g.
Figure 6.2. Note that in case of a global base topology, all induced subnetworks are
global. The phase snapshot for the splay multicluster is displayed in Fig. 6.3e. Here,
as in the case with global base structure, the phase distribution possess the property
that R (2) (φ) = 0.
6.1 Multicluster and Solitary States 119

Fig. 6.4 Map of regimes for one- and multicluster states of antipodal and splay type in (α, β)
parameter space. Parameters: N = 100, P = 20,  = 0.01. The horizontal black line at α = 0.1
shows the location for the parameter β where the emergence of solitary states is analyzed, see
Fig. 6.7 in Sect. 6.3. Figure reproduced from [29]

Another, more complex, antipodal five-cluster is presented in Fig. 6.3g–i. In

Fig. 6.3i, we observe three groups, a big one (green) and two smaller ones (red,
blue), with different frequencies. Moreover, in accordance with the definition of a
frequency cluster and the illustration in Fig. 6.2, the blue and the red groups pos-
sesses two connected components (I, II) each. This fact implies the presence of five
individual frequency clusters in Fig. 6.3g–i. Remarkably, the red clusters I and II as
well as the blue clusters I and II are of the same size. This observation is in contrast
to the hierarchical structures discovered and analyzed in [22] and Chap. 5. Hence, in
the case of the ring base structure, the evenly sized clusters can appear, which was
not possible in the case of global base structure [22] and identical oscillators, see
also Chap. 5.
While several examples for one- and multicluster states have been described
above, Fig. 6.4 shows that these states are observable in a wide range in the (α, β)
parameter space. The diagram in Fig. 6.4 is produced by running simulations of (6.1)–
(6.2) from random initial conditions. In case a one-cluster or multicluster is found, the
region is colored or hatched, respectively, in accordance with the legend in Fig. 6.4.
We further used a continuation method in the (α, β) parameter space to show the
full extent where the various types of multiclusters can be observed. In Sects. 6.2.1
and 6.2.2, we provide a more rigorous description for the existence and stability
properties of the one-clusters.

6.1.3 Solitary States

For systems with global base coupling, the clusters in the multicluster states were
found to be hierarchical in nature, i.e. the clusters varied in size significantly [22],
see also Chap. 5. As we have mentioned above, in the nonlocal base coupling case,
multicluster states with one large and many smaller, similar in size, clusters have
been observed. Figure 6.5 shows a particular example of this phenomenon, called
solitary states, where either one single oscillator (upper panels) or three single oscil-
120 6 Adaptation on Nonlocally Coupled Ring Networks

Fig. 6.5 Illustration of solitary states. The panels a, d show coupling matrix, b, e phase snapshots,
and c, f average frequencies. a–c: single solitary state for α = 0.1π , β = −0.3π ; d–f: three uncou-
pled solitary states for α = 0.15π , β = −0.41π . Parameters: N = 100, P = 20,  = 0.01. Figure
reproduced from [29]

lators (lower panels) decouple from a large cluster. The solitary states are particular
examples of multiclusters with a large group of frequency synchronized oscillators
(background cluster) and individual solitary nodes with different frequency, i.e.,
clusters consisting of only one oscillator. These special kind of states, for which we
provide an analysis of their emergence in Sect. 6.3, are of particular interest as they
are found in various dynamical systems [31–39].

6.2 One-Cluster States: Local Versus Global Features

6.2.1 Classification of One-Cluster States

In this section we study the antipodal and splay one-cluster states in more details.
Due to the S 1 symmetry of system (6.1)–(6.2), the following phase-locked solutions
appear generically

φi (t) = t + χi , i = 1, . . . , N , (6.6)
6.2 One-Cluster States: Local Versus Global Features 121

where χi ∈ [0, 2π ) are fixed phase lags and the cluster frequency. It is clear that
such solutions describe a one-cluster state since the frequencies are the same. By
substituting (6.6) into (6.1)–(6.2), we obtain

κi j = − sin(χi − χ j + β), (6.7)

1 
i+P
1
= cos(α − β) − cos(2χi − 2χ j + α + β). (6.8)
2 4P j=i−P

The Eq. (6.8) implies that the one-cluster state exists only if the following expression
is independent of the index i

1 
i+P  (2)

cos(2χi − 2χ j + α + β) = Re Ri(2) ei(ϑi −2χi −α−β) . (6.9)
2P j=i−P

Equation (6.9) allows for the distinction of two types of distributions of the phase-lags
for which it is independent of i. We call a cluster of
(i) Antipodal type, if χi ∈ {0, π } for all i = 1, . . . , N . In this case, the sum in (6.9)
equals cos(α + β), and of
(ii) Local splay type, if χi ∈ [0, 2π ) are such that the second order parameter Z i(2) =
(2)
Ri(2) (χ)eiϑi satisfies

Ri(2) (χ) = Rc(2) (χ) (6.10)

ϑi(2) = 2χi + χ0 (6.11)

with 0 ≤ Rc(2) (χ ) < 1 and χ0 ∈ [0, 2π ) independent on i.

Note that the additional degree of freedom for clusters of local splay type, i.e. χ0 ∈
[0, 2π ) arbitrary, is due to the S 1 symmetry. The constant χ0 can be set to 0 without
loss of generality. The both types of phase distribution lead to one-cluster states for
the system (6.1)–(6.2) with frequency

sin α sin β antipodal type,
=  (2)
 (6.12)
1
2
cos(α − β) − Rc cos(α + β) local splay type.

In the context of globally coupled base topologies, antipodal and splay type phase
distribution have been extensively discussed [40] where the (global) splay clusters,
also called fuzzy clusters [31], are defined by the global condition Z (2) (χ ) = 0,
see also Chap. 4. Remarkably, if a phase distribution is of the local splay type (as
described by (6.10)–(6.11)) then it is of splay type as well, see Appendix A.5 for
more details. The converse is not true in general. Hence, the class of local splay
clusters is “smaller” than the class of global splay clusters. In addition, local splay
122 6 Adaptation on Nonlocally Coupled Ring Networks

clusters do not necessarily form families of solutions. According to the definition

of local splay cluster, generically N complex algebraic equations have to be solved
for N unknown phase-lags χi . Therefore, the set of equations for the phase-lags is
overdetermined and the set of local splay states might be empty. However, it is not
the case due to the symmetries of the system and the base coupling structure. The
symmetry of the nonlocal ring structure allows for constructing explicit, symmetric
examples for the clusters of local splay type. These are clusters of the rotating-wave
type.

(ii’) The clusters are of rotating-wave type, if χi = ik 2π

N
, where k = 1, . . . , N is
the wavenumber. In the literature, the notion “splay state” is often restricted to
this definition.

Let us show that the rotating wave clusters (ii’) are the local splay states (ii). For this
we write the phase distribution as χ k = (2π k/N , . . . 2π k(N − 1)/N , 0)T . Then, we
have

1  ink j 2π
i+P
Z i(n) (χ k ) = (n)
2π
N = e inki N R
e N (χ k ), (6.13)
2P j=i−P

where
⎛ ⎞
1 ⎝
P
2π ⎠
R (n)
N (χ k ) = cos(nk j ) . (6.14)
P j=1 N

we conclude that all rotating-wave states with k = 0, N /2 are local splay states and
thus solutions to (6.1)–(6.2). Rotating-wave clusters with k = 0, N /2 are of antipodal
type. The nth moment local order parameter Z i(n) (χ k ) = R (n) N (χ k ) is constant for
all i = 1, . . . , N and its value depends on the wavenumber k.
Note further that Z i(n) (χ k ) = Z i(nk) (χ 1 ), which connects the moment of the order
parameters with the wavenumber of the rotating-wave states. For globally coupled
base structures, the rotating-wave states are found to be very important in describing
the main features of antipodal and global splay type clusters such as stability. The
next section is devoted to the description of the stability condition for rotating-wave
states.

6.2.2 Stability of One-Cluster States

In the following, the stability of one-clusters is analyzed. In order to study the local
stability of one-cluster solutions described in Sect. 6.2.1, we linearize the system of
differential equations (6.1)–(6.2) around the phase-locked states
6.2 One-Cluster States: Local Versus Global Features 123

φi (t) = t + ai ,
κi j = − sin(ai − a j + β).

These solutions are equilibria relative to the S 1 symmetry [41], therefore the lineariza-
tion around such solutions leads to a linear system with constant coefficients, despite
the time dependency of φ(t). Practically, one can first move to the co-rotating coordi-
nate system by introducing the new variable φ(t) − t and then linearize around the
equilibrium in the new coordinates. As a result, we obtain the following linearized
system for the perturbations δφi and δκi j :

1    
i+P
d
δφi = sin(β − α) + sin(2(ai − a j ) + α + β) δφi − δφ j
dt 4P j=i−P
(6.15)
1 
i+P
− sin(ai − a j + α)δκi j ,
2P j=i−P

and
d   
δκi j = −ai j δκi j + cos(ai − a j + β) δφi − δφ j . (6.16)
dt

System (6.15)–(6.16) is a (N + N 2 )-dimensional linear system of ordinary differ-

ential equations, which can be written in the form x  = L x, x ∈ R N +N , and the
2

stability of which is determined by the eigenvalues of the matrix L. For all antipodal
and rotating-wave states the stability analysis can be done explicitly. However, the
calculations are quite lengthy (see Appendix A.6). Summarizing the results of these
calculations, the spectrum S of the eigenvalues, corresponding to the rotating-wave
one-clusters, is given by
  N  N 
S = 0, −, λl,1 l=1 , λl,2 l=1 . (6.17)

Here λl,1 and λl,2 are the solutions of the quadratic equation

λl  
λl2 − L(α, β, l, k) + (R (l)
N (χ 1 ) − 1) sin(α − β) − 2 −  L(α, β, l, k) = 0,
2
(6.18)

where the complex function L as defined in Eq. (A.27) in Appendix A.6, k is the
wavenumber.
In Fig. 6.6 we show the stability of rotating-wave one-clusters in an ensemble
of N = 50 oscillators. The figures demonstrate regions of stability in the (α, β)
parameter plane for different wavenumber k and coupling range P. The stability is
obtained by numerical simulations as well as by the Lyapunov spectrum Eq. (6.17).
The borders of stability, as they are provided by the analytical results, are displayed
124 6 Adaptation on Nonlocally Coupled Ring Networks

Fig. 6.6 Stability of one-cluster states for different wavenumbers k and coupling ranges P. Regions
of stability for the one-cluster states are colored in blue, while instability in yellow. The borders of
stability (black dashed lines) are obtained from the eigenvalues (6.17). Parameters are as follows:
a P = 10, k = 1; b P = 10, k = 4; c P = 10, k = 25; d P = 5, k = 1; e P = 20, k = 1; and f
P = 25, k = 1. The other parameters are N = 50 and  = 0.01. Figure reproduced from [29]

with a dashed black line. Numerically the stability was computed as follows: (i) the
theoretical shape of the one-cluster given by (6.6)–(6.7) is used as initial conditions
with a small random perturbation in the range of [−0.01, 0.01]; (ii) then we solve
the system numerically for t = 20000 time units; (iii) compute the euclidean norm
between the initially perturbed state and the theoretical one as well as between the
final state after t = 20000 and the theoretical one; (iv) in case the second norm is
smaller than the first, meaning that the trajectory approaches the theoretical one-
cluster state, we consider the one-cluster state as stable and color the corresponding
region in blue, otherwise the state is considered as unstable and the corresponding
region is colored in yellow.
The diagrams in the first row of Fig. 6.6 show the influence of the wavenumber
on the stability of one-clusters. Here, the coupling range is fixed to P = 10. For
adaptively coupled phase oscillators with a global base topology, it has been shown
that the stability of rotating-waves of local splay type, i.e., k = 0, N /2, does not
depend on the wavenumber, see Chap. 4. However. we observe that in case of a
nonlocal base structure the shape of the stability regions crucially depends on the
wavenumber k, see Fig. 6.6a–c. Moreover, there are no common regions of stability
for the one-cluster states in Fig. 6.6b, c, i.e., the region of stability in both figures
have an empty intersection.
The diagrams in Fig. 6.6d–f exemplify the influence of the coupling range on
the stability of the rotating-wave cluster with k = 1. We see that in comparison with
Fig. 6.6a, the regions of stability change significantly. Note further that for P = 25 the
stability regions resemble the results known for the globally coupled base topology,
see Chap. 4.
In contrast to the local splay type clusters, the stability regions for the antipodal
one-cluster states are the same which can be derived from the following. For antipodal
states, the quadratic Eq. (6.18) simplifies to
6.2 One-Cluster States: Local Versus Global Features 125

λl  
λl2 − (1 − R (l)
N (χ 1 )) sin(β) cos(α) − 2 − (1 − R (l)
N (χ 1 )) sin(α + β) = 0.
2
(6.19)

since L(α, β, l, k) = sin(α + β)(1 − R (l) N (χ 1 )) for k = 0, N /2.

The regions of stability in Fig. 6.6c for a nonlocal coupling structure are in
agreement with the regions of stability found for a global coupling structure, see
Chap. 4. However, note that Eq. (6.19) differs analytically from the expression found
in Corr. 4.2.3. The similarity in the stability regions is only due to the small value of 
and the differences would be more pronounced in the presence of larger . Note fur-
ther, that in case of P = N /2, which is equivalent to global coupling, the local order
parameter R (l) is either 1 for l = 0 and 0 otherwise. This agrees with the findings in
Chap. 4.
Our stability analysis shows how strongly the ring network structure and confined
coupling range alter the stability properties of the clusters. Since the analysis in
Appendix A.6 is not restricted to nonlocal coupling structures, it provides the analytic
tools to study the influence of more general complex base topologies on the stability
of rotating-wave states.

6.3 The Emergence of Solitary States

In this section, we unveil the mechanism behind the formation of solitary states
as they are illustrated in Fig. 6.5. As solitary state we define the state where all
oscillators in the system are frequency synchronized except one single oscillator, or
several oscillators which do not share their local neighborhood. The majority cluster
of the synchronized oscillators is also called background cluster.
Let us restrict ourselves to the analysis of a solitary cluster interacting with an
in-phase synchronous cluster (see (6.6) and (6.7) with χi = 0). Imposing the assump-
tions, we end-up with the following 4-dimensional model where φ and ψ describe the
dynamics of the solitary cluster and the background in-phase synchronized cluster,
respectively, which are dynamically coupled through κ1 and κ2 :

N −1
φ̇ = 1 − κ1 sin(φ − ψ + α),
N
N −2 1
ψ̇ = 1 + sin α sin β − κ2 sin(φ − ψ − α),
N N
κ̇1 = − (κ1 + sin(φ − ψ + β)) ,
κ̇2 = − (κ2 − sin(φ − ψ − β)) .

The latter equations can be simplified by introducing the phase difference θ = φ − ψ

as well as by considering a large ensemble of oscillators (N → ∞). We obtain the
126 6 Adaptation on Nonlocally Coupled Ring Networks

Fig. 6.7 Phase portraits for two-dimensional system (6.20)–(6.21). The graphics show the two
classes of asymptotic states that are equilibria (colored nodes) and periodic solutions (colored lines).
The stability properties of the individual asymptotic states are indicated by the coloring where the
blue refers to stable and the red (dashed) to unstable states. In addition, several trajectories are plotted
in black including those close to the stable and unstable manifold of the equilibria. The nullclines
are displayed as gray lines. For the different panels parameter β is varied as shown in Fig. 6.4: a
β = −0.601π ; b β = −0.599π ; c β = −0.58π ; d β = −0.5515π ; e β = −0.5π ; f β = −0.08π ;
g β = −0.0563π ; and h β = 0.05π . The other parameters are α = 0.1π and  = 0.01. Figure
reproduced from [29]

following two-dimensional system for the dynamics of two clusters, one of which is
solitary, in the large ensemble limit:

θ̇ = − sin α sin β − κ sin(θ + α), (6.20)

κ̇ = −(κ + sin(θ + β)), (6.21)

where we denote κ = κ1 .
In the following, we study the structure of the phase space of (6.20)–(6.21) for
fixed α = 0.1π and different values of parameter β. Several bifurcation scenarios are
discovered which give rise to the birth and stability changes of the solitary states. We
observe how the stable solitary state emerges in a subcritical pitchfork bifurcation of
periodic orbits and disappears in a homoclinic bifurcation with increasing β. Note
that solitary states are given by periodic solutions, where the phase difference θ (t)
rotates. Equilibria of this system describe one-cluster states. Figure 6.7 shows several
characteristic phase portraits of (6.20)–(6.21) illustrating the bifurcation scenarios
with increasing β (from (a) to (h)), see also Fig. 6.4.
In Fig. 6.7a, we observe four equilibria which correspond to certain one-cluster
solution. The stable equilibria at θ = 0 and θ = π correspond to in-phase syn-
chronous and antipodal where a1 = π and ai=1 = 0, respectively. The other two sad-
6.3 The Emergence of Solitary States 127

dle equilibria correspond to the special class of double antipodal states, see Chap. 4,
and describe therefore phase-cluster similar to those described in [38, 42]. While
these equilibria can be stable for the reduced system (6.20)–(6.21), they are always
unstable for (6.1)–(6.2) in case of global coupling, see Proposition 4.2.4. Addition-
ally, to these four equilibria we find an unstable periodic orbit which corresponds to
an unstable solitary state.
With increasing β, we observe a subcritical pitchfork bifurcation of periodic orbits
at α + β = −π/2 in which the unstable periodic orbit is stabilized and two additional
periodic orbits are created. Figure 6.7b shows the phase portrait directly after the
pitchfork bifurcation. Therefore, we conclude that there exist three solitary states,
two of which are unstable and one stable. It is worth to remark that the stability for
the reduced system is only necessary but not sufficient to be a stable asymptotic state
for the network (6.1)–(6.2).
By increasing β even further the basin of attraction of the stable periodic orbit
increases and its boundaries are given by the unstable periodic orbits, see Fig. 6.7c.
For β = −0.5515π , the trajectories of the unstable solitary states merge with the
equilibria and become homoclinic orbits of the saddle equilibria (Fig. 6.7d). The
phase portrait after this homoclinic bifurcation is shown in Fig. 6.7e.
After the homoclinic bifurcation, with increasing β, the equilibria are moving
towards each other in phase space and exchange their stability in a transcritical
bifurcation. This can seen analytically by considering the determining equations for
the equilibria (θ, κ) of (6.20)–(6.21)

0 = cos(α + β) − cos(2θ + α + β), (6.22)

κ = − sin(θ + β).

In general, Eq. (6.22) possesses two solutions for 2θ . At α + β = 0, π , however, these

two solutions coincide which describes the point of the transcritical bifurcation. The
stability of the equilibria can be further computed by considering the two-dimensional
system linearized around the equilibria. In a more general setup this has be done
Sect. 6.2.2. Figure 6.7f displays the phase portrait after the transcritical bifurcation.
Remember that although the double antipodal clusters are stable for the reduced
system, they are always unstable for the full system, see Proposition 4.2.4.
In Fig. 6.7g, h another homoclinic bifurcation is presented in which the stable
solitary state becomes a homoclinic orbit of the in-phase and antipodal cluster. The
phase portrait close to the homoclinic bifurcation is presented in Fig. 6.7h. After the
homoclinic bifurcation the phase space is divided into the basins of attraction of the
two double antipodal states and no more solitary state exist.
128 6 Adaptation on Nonlocally Coupled Ring Networks

6.4 Adaptive Networks with Global Base Topology Versus

Ring Base Topology: The Differences

Adaptive networks of coupled phase oscillators have been extensively studied on

an all-to-all base structure [19, 20, 22], see also Chaps. 4 and 5. Here, we extend
previous work towards more complex base topologies by considering a nonlocal ring
base topology on which adaptation takes place. In this section we briefly summarize
the main differences resulting from the different base topologies.
For the global base topology, all links between the nodes with the same fre-
quency become active leading to the all-to-all structures within each cluster. There-
fore, strongly connected components can be equivalently described by the frequency
synchronization of nodes. In contrast to this, for ring networks (also more complex
base structures), the frequency synchronization does not necessarily imply connec-
tivity, see Fig. 6.2. As a result, we have adapted the definition of the frequency cluster
on a complex base topology as a connected subnetwork with frequency synchronized
oscillators.
Another effect induced by the ring base topology concerns the hierarchical order-
ing of cluster sizes. For global base structures is has been found that a sufficiently large
difference of the cluster sizes is necessary for the appearance of multicluster states.
In case of a ring structure, the hierarchy is not necessary anymore. In Fig. 6.3g–i, we
present a five-cluster states that possesses two clusters each of size 7 and additionally
two clusters each of size 2. For solitary states this nonhierarchical clustering implies
that on a ring structure there can be several solitary nodes (Fig. 6.5d–f). A simple
explanation for the appearance of the clusters of a similar size is based on the fact
that such clusters can be uncoupled in the base coupling structure and, hence, not
synchronized. In contrast, in networks with the global base structure, similar clusters
tend to be synchronized and merge into one larger cluster.
Regarding the stability of rotating-wave states another striking difference between
global and ring base topology is observed. Here the differences are twofold. For
a global base topology rotating-waves constitute a N − 2-dimensional family of
solutions with the same collective frequency. On a nonlocal ring, this invariant family
is not present anymore and all rotating-wave are different from each other including
their frequencies. The same holds true for their stability. While the stability features of
all rotating-waves agree on global structures, the stability properties depend crucially
on the wavenumber (see Fig. 6.6a–c).

6.5 Summary

In summary, a model of adaptively coupled identical phase oscillators on a nonlocal

ring has been studied. Various frequency synchronized states are observed including
one-cluster, multicluster, and solitary states. Those states are similar to those found
for a global base topology [22], also Chaps. 4 and 5. However, to account for the
6.5 Summary 129

complex base topology, we have introduced a new definition of one-clusters by means

of connected induced subnetworks. This definition allows furthermore to distinguish
between multicluster and solitary states in a more strict way than it was done before.
Since one-cluster states form building blocks for multicluster states, Chap. 5, we
have first investigated the existence and stability properties of one-cluster states.
Here, we have introduced a novel type of phase-locked states for complex networks,
namely local splay states, and have shown that this class of states is nonempty for
any nonlocal ring base topology. In particular, we have proved that rotating-wave as
well as antipodal states are always phase-locked solutions. Compared with the case
of a global base topology, the different clusters of local splay type on a nonlocal ring
structure can possess different collective frequencies. In addition, we have proved that
local splay cluster are always global splay cluster. This statement relates, therefore,
local with global (with respect to “spatial” structures in the network) properties of
solutions.
The stability features of rotating-wave states have been studied numerically and
analytically. The comparison of both approaches results in a very good agreement.
Due to the analytic findings for rotating-wave states on a nonlocal ring, we are able
to describe their stability depending on the coupling range P and the wavenumber
k. The limiting case of global coupling, i.e. P = N /2, is shown to be in agreement
with the results presented in Chap. 4.
An interesting feature of the system’s behavior are solitary states. They have been
previously found to emerge in the Kuramoto–Sakaguchi model with inertia [37]. In
this chapter, we show that solitary states are born in a homoclinic bifurcation and
can be (de)stabilized in a pitchfork bifurcation of periodic orbits. In order to show
this, a two-dimensional effective model is derived governing the dynamics of solitary
states. In contrast to the Kuramoto–Sakaguchi model with inertia, we observe a much
more complicated bifurcation behavior. In particular, three different solitary states
are created due to two individual homoclinic bifurcations. Two of these three solitary
states, however, are unstable and bifurcate together with stable solitary states in a
subcritical pitchfork bifurcation of periodic orbits.
Our results highlight the delicate interplay between adaptivity and complexity of
the network structure. Since this interplay has been rarely investigated from the math-
ematical viewpoint, so far, this work raises many questions for future research which
could be conducted for different network structures beyond nonlocal rings, other
dynamical models for the local dynamics, nonidentical units or different adaptation
rules.

References

1. Compte A, Sanchez-Vives MV, McCormick DA, Wang XJ (2003) Cellular and network mech-
anisms of slow oscillatory activity (<1 Hz) and wave propagations in a cortical network model.
J Neurophys 89:2707
2. Sporns O (2011) Networks of the brain. MIT Press, Cambridge
130 6 Adaptation on Nonlocally Coupled Ring Networks

3. Popovych OV, Yanchuk S, Tass P (2011) Delay- and coupling-induced firing patterns in oscil-
latory neural loops. Phys Rev Lett 107:228102
4. Yanchuk S, Perlikowski P, Popovych OV, Tass P (2011) Variability of spatio-temporal patterns
in non-homogeneous rings of spiking neurons. Chaos 21:047511
5. Pasemann F (1995) Characterization of periodic attractors in neural ring networks. Neural
Netw 8:421
6. Bressloff PC, Coombes S, de Souza B (1997) Dynamics of a ring of pulse-coupled oscillators:
group-theoretic approach. Phys Rev Lett 79:2791
7. Yanchuk S, Wolfrum M (2008) Destabilization patterns in chains of coupled oscillators. Phys
Rev E 77:26212
8. Bonnin M (2009) Waves and patterns in ring lattices with delays. Phys D 238:77
9. Zou W, Zhan M (2009) Splay states in a ring of coupled oscillators: from local to global
coupling. SIAM J Appl Dyn Syst 8:1324
10. Horikawa Y, Kitajima H (2009) Duration of transient oscillations in ring networks of unidirec-
tionally coupled neurons. Phys D 238:216
11. Perlikowski P, Yanchuk S, Popovych OV, Tass P (2010) Periodic patterns in a ring of delay-
coupled oscillators. Phys Rev E 82:036208
12. Omelchenko I, Maistrenko Y, Hövel P, Schöll E (2011) Loss of coherence in dynamical net-
works: spatial chaos and chimera states. Phys Rev Lett 106:234102
13. Kantner M, Yanchuk S (2013) Bifurcation analysis of delay-induced patterns in a ring of
Hodgkin-Huxley neurons. Phil Trans R Soc A 371:20120470
14. Omelchenko I, Omel’chenko OE, Hövel P, Schöll E (2013) When nonlocal coupling between
oscillators becomes stronger: patched synchrony or multichimera states. Phys Rev Lett
110:224101
15. Yanchuk S, Perlikowski P, Wolfrum M, Stefanski A, Kapitaniak T (2015) Amplitude equations
for collective spatio-temporal dynamics in arrays of coupled systems. Chaos 25:033113
16. Klinshov V, Shchapin D, Yanchuk S, Wolfrum M, D’Huys O, Nekorkin VI (2017) Embedding
the dynamics of a single delay system into a feed-forward ring. Phys Rev E 96:042217
17. Burylko O, Mielke A, Wolfrum M, Yanchuk S (2018) Coexistence of Hamiltonian-like and
dissipative dynamics in rings of coupled phase oscillators with skew-symmetric coupling.
SIAM J Appl Dyn Syst 17:2076
18. Omel’chenko OE (2018) The mathematics behind chimera states. Nonlinearity 31:R121
19. Aoki T, Aoyagi T (2009) Co-evolution of phases and connection strengths in a network of
phase oscillators. Phys Rev Lett 102:034101
20. Aoki T, Aoyagi T (2011) Self-organized network of phase oscillators coupled by activity-
dependent interactions. Phys Rev E 84:066109
21. Nekorkin VI, Kasatkin DV (2016) Dynamics of a network of phase oscillators with plastic
couplings. AIP Conf Proc 1738:210010
22. Kasatkin DV, Yanchuk S, Schöll E, Nekorkin VI (2017) Self-organized emergence of multi-
layer structure and chimera states in dynamical networks with adaptive couplings. Phys Rev E
96:062211
23. Kasatkin DV, Nekorkin VI (2018) The effect of topology on organization of synchronous
behavior in dynamical networks with adaptive couplings. Eur Phys J Spec Top 227:1051
24. Berner R, Schöll E, Yanchuk S (2019) Multiclusters in networks of adaptively coupled phase
oscillators. SIAM J Appl Dyn Syst 18:2227
25. Berner R, Fialkowski J, Kasatkin DV, Nekorkin VI, Yanchuk S, Schöll E (2019) Hierarchical
frequency clusters in adaptive networks of phase oscillators. Chaos 29:103134
26. Berner R, Vock S, Schöll E, Yanchuk S (2021) Desynchronization transitions in adaptive
networks. Phys Rev Lett 126:028301
27. Vock S, Berner R, Yanchuk S, Schöll E (2021) Effect of diluted connectivities on cluster
synchronization of adaptively coupled oscillator networks. arXiv:2101.05601
28. Gray RM (2006) Toeplitz and circulant matrices: a review, Foundations and Trends in Com-
munications and Information Theory, vol 2. Now Publishers Inc., Hanover, pp 155–239
References 131

29. Berner R, Polanska A, Schöll E, Yanchuk S (2020) Solitary states in adaptive nonlocal oscillator
networks. Eur Phys J Spec Top 229:2183
30. Korte B, Vygen J (2018) Combinatorial optimization. Springer, Berlin
31. Maistrenko Y, Penkovsky B, Rosenblum M (2014) Solitary state at the edge of synchrony in
ensembles with attractive and repulsive interactions. Phys Rev E 89:060901
32. Ashwin P, Burylko O (2015) Weak chimeras in minimal networks of coupled phase oscillators.
Chaos 25:013106
33. Semenov V, Zakharova A, Maistrenko Y, Schöll E (2016) Delayed-feedback chimera states:
forced multiclusters and stochastic resonance. Europhys Lett 115:10005
34. Wojewoda J, Czolczynski K, Maistrenko Y, Kapitaniak T (2016) The smallest chimera state
for coupled pendula. Sci Rep 6:34329
35. Premalatha K, Chandrasekar VK, Senthilvelan M, Lakshmanan M (2016) Imperfectly synchro-
nized states and chimera states in two interacting populations of nonlocally coupled Stuart-
Landau oscillators. Phys Rev E 94:012311
36. Maistrenko Y, Brezetsky S, Jaros P, Levchenko R, Kapitaniak T (2017) Smallest chimera states.
Phys Rev E 95:010203(R)
37. Jaros P, Brezetsky S, Levchenko R, Dudkowski D, Kapitaniak T, Maistrenko Y (2018) Solitary
states for coupled oscillators with inertia. Chaos 28:011103
38. Teichmann E, Rosenblum M (2019) Solitary states and partial synchrony in oscillatory ensem-
bles with attractive and repulsive interactions. Chaos 29:093124
39. Taher H, Olmi S, Schöll E (2019) Enhancing power grid synchronization and stability through
time delayed feedback control. Phys Rev E 100:062306
40. Ashwin P, Burylko O, Maistrenko Y (2008) Bifurcation to heteroclinic cycles and sensitivity
in three and four coupled phase oscillators. Phys D 237:454
41. Golubitsky M, Stewart I (1988) Singularities and groups in bifurcation theory. Volume 2,
Applied mathematical sciences, vol 69. Springer, New York
42. Maistrenko Y, Vasylenko A, Sudakov O, Levchenko R, Maistrenko VL (2014) Cascades of
multi-headed chimera states for coupled phase oscillators. Int J Bifur Chaos 24:1440014
Chapter 7
Synchronization on Adaptive Complex
Network Structures

In this chapter, we extend the master stability approach to complex dynamical net-
work of N diffusively [1–3] and adaptively coupled oscillators. Moreover, we apply
the result to the paradigmatic model of adaptively coupled phase oscillators and pro-
vide a novel description for the emergence of multicluster due to the presence of
stability islands.
In the following, we consider a network of N diffusively [1–5] and adaptively
coupled oscillators of the form


N
ẋ i = f (x i (t)) − σ ai j κi j G(x i − x j ), (7.1)
j=1
 
κ̇i j = − κi j + ai j ρ H (x i − x j ) , (7.2)

where x i is a d-dimensional state vector of the ith node with x i ∈ Cd , f ∈

C 1 (Cd , Cd ) describes the local dynamics of node i, the function G ∈ C 1 (Cd , Cd )
determines the coupling between nodes. The coupling between the nodes is weighted
by the dynamical variables κi j ∈ R which are adapted according to the func-
tion H ∈ C 1 (Cd , R). The parameter σ ∈ R is the coupling constant and ρ ∈ R is
called the adaptation strength. The basic coupling structure is given by the matrix
entries ai j ∈ {0, 1} of the N × N adjacency matrix A where a constant row sum

r = Nj=1 ai j for all i = 1, . . . , N is assumed.
Let s(t) be the synchronized solution, meaning that x i = s(t) for all i = 1, . . . , N .
Hence the fully synchronized solution for the Eqs. (7.1)–(7.2) is given by the solution
of the following equations

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 133
R. Berner, Patterns of Synchrony in Complex Networks of Adaptively Coupled Oscillators,
Springer Theses, https://doi.org/10.1007/978-3-030-74938-5_7
134 7 Synchronization on Adaptive Complex Network Structures

ṡ = f (s) + r σρ H (0)G(0), (7.3)


−ρ H (0), if ai j = 1,
κ sync
= (7.4)
0, otherwise.

Note that the synchronous solution depends on the row sum if neither G(0) = 0 nor
H (0) = 0. By scaling σ with 1/r this dependency can be omitted.
The chapter is structured as follows. In Sect. 7.1, we establish the master stabil-
ity approach for complex adaptive networks of coupled oscillators and prove the
main Proposition 7.1.3. This result is then applied to a complex network of adap-
tively coupled phase oscillators for which the master stability function is completely
described in Sect. 7.2. Here, we further show the emergence of stability islands. In
Sect. 7.3, we use the master stability functions to show how multicluster states as
well as chimera-like states emergence due to the stability islands. The results of this
chapter are summarized in Sect. 7.4.

7.1 The Master Stability Function for Adaptive Complex

Networks

In Ref. [1], the master stability function for dynamical systems of the form (7.1) was
introduced. Here, the master stability function is defined as the largest Lyapunov
exponent depending on a complex parameter, the master function parameter, for
which eventually the eigenvalues of the Laplacian matrix will be inserted. Once the
master stability function is determined, the stability of the synchronous solution can
be directly deduced for any topology.
In the following, we follow and extend the approach of Refs. [1, 3] and present a
master stability function which takes adaptation of the coupling weights into account.
For this we have to consider the variational equations along the synchronous state.
For convenience we define the following notation
⎛ ⎞
ai1
⎜ .. ⎟
ai = (ai1 , . . . , ai N ), diag(ai ) = ⎝ . ⎠
ai N

and the N × N 2 , N 2 × N , and N 2 × N matrices

⎛ ⎞ ⎛ ⎞
a1 diag(a1 )
⎜ .. ⎟ ⎜ .. ⎟
B=⎝ . ⎠ , C = B − D,
T
D=⎝ . ⎠,
aN diag(a N )

respectively.
7.1 The Master Stability Function for Adaptive Complex Networks 135

Then, the variational equations for system (7.1)–(7.2) along the synchronous
solution s(t) can be written in the following vectorized form
  
ξ̇ S −σ B ⊗ G(0) ξ
= , (7.5)
χ̇ −ρC ⊗ D H (0) −I N 2 χ

where S = I N ⊗ D f (s) + σρ H (0) (L(A) ⊗ DG(0)), ξ = x − I N ⊗ s, and χ =

κ − κ sync with Kronecker product denoted by ⊗. Further, x ∈ C N ·d is the system state
vector where all individual nodal states are stacked on each other ordered by the node
index, i.e., x = (x 1 , . . . , x N )T . Analogously, κ = (κ11 , . . . , κ1N , . . . , κ N N ). Further,
the Laplacian L(A) for the corresponding adjacency matrix A is given by

L(A) = D(A) − A,

where
⎛ N ⎞
j=1 a1 j
⎜ .. ⎟
D(A) = ⎝ . ⎠, (7.6)
N
j=1 a N j

see (2.3).
In the following Lemma, we summarize a few simple relations which are used to
simplify calculations subsequently.
Lemma 7.1.1 Let B, C, and D given as above. Then the following relations hold
true:
1. B · B T = r IN ,
2. B · D = A,
3. B · C = L(A),
4. C T C = L(A) + L(A T ).

Proof The results are proved by direct calculation.

1.
⎛ ⎞⎛ ⎞ ⎛ N ⎞
a1 a1T j=1 a1 j a1 j
⎜ .. ⎟⎜ .. ⎟ ⎜ .. ⎟
B · BT = ⎝ . ⎠⎝ . ⎠=⎝ . ⎠.
N
aN a TN j=1 a N j a N j

2.
⎛ ⎞⎛ ⎞ ⎛ ⎞ ⎛ ⎞
a1 diag(a1 ) a1 diag(a1 ) a11 a11 · · · a1N a1N
⎜ .. ⎟⎜ . ⎟ ⎜ . ⎟ ⎜ . .. . ⎟
B·D=⎜
⎝ .
⎟⎜
⎠⎝ .
.
⎟ ⎜
⎠=⎝ .
.
⎟ ⎜
⎠=⎝ .
. . .
.
⎟.
⎠
aN diag(a N ) a N diag(a N ) aN 1aN 1 · · · aN N aN N
136 7 Synchronization on Adaptive Complex Network Structures

3. Using the results from 1. and 2., we get

B · C = B · B T − B · D = L(A).

4. By applying 1. and 2., we find

  T
  
N
C · C = B − D · B − D = r I N − A − (B · D) +
T T T
diag(a j )2
j=1

= r I N − A − A + D(A ).
T T


It was shown in Chaps. 4 and 6 that due to the structure of the variational equations,
there are locally always degrees of freedom corresponding to the N 2 − N eigenval-
ues λ = −. In order to reduce the Eqs. (7.5), we are studying the corresponding
eigenspace. Consider the equation
 
S − I N d −σ B ⊗ G(0) ξ
= 0.
−ρC ⊗ D H (0) 0 χ

From C ⊗ D H (0)ξ = 0, it can be immediately deduced that ξ i = ξ̄ ∈ Cd for all i =

1, . . . , N . Assuming that ξ = 0 would lead to Bχ = 0 as the defining equation for the
eigenvectors. In fact, the following results show that Bχ = 0 suffices to find N 2 −
N many linearly independent vectors spanning the eigenspace. The eigenvectors
themselves are additionally independent of time.
Lemma 7.1.2 Let χ = (χ 1T , . . . , χ TN )T with χ iT = ((χ i )1 , . . . , (χ i ) N ) for all i ∈
{1, . . . , N }, then the following two relations hold true
  
1. ker(B) = χ ∈ R N : Nj=1 ai j (χ i ) j = 0 ∀i ∈ {1, . . . , N }
2

  
2. ker(D T ) = χ ∈ R N : Nj=1 a ji (χ j )i = 0 ∀i ∈ {1, . . . , N } .
2

Further, dim(ker(B)) = dim(ker(D T )) = N 2 − N .

Proof 1. We find by direct calculation
⎛ ⎞ ⎛ N ⎞
a1 χ 1 j=1 a1 j (χ 1 ) j
⎜ ⎟ ⎜ ⎟
Bχ = ⎝ ... ⎠ = ⎝ ..
. ⎠
N
aN χ N a
j=1 N j (χ N j)

2.
⎛ ⎞
diag(a1 )
 
⎟  T 
N N
⎜ ..  
χ T D = χ 1T , . . . , χ TN ⎝ . ⎠ = χ a
j j = (χ j )1 a j1 , . . . , (χ j ) N a j N
diag(a N ) j=1 j=1
7.1 The Master Stability Function for Adaptive Complex Networks 137

It is easy to verify that rank(B) = rank(D T ) = N and dim(ker(B)) =

dim(ker(D T )) = N 2 − N . 

With this result we are able to find a proper transformation in order to determine the
master stability function for adaptively coupled dynamical systems.
Proposition 7.1.3 Let (7.1)–(7.2) possess a synchronous state given by (7.3)–(7.4).
Further, let (7.5) be the variational equations around this synchronous solution and
assume that the Laplacian matrix L(A) is diagonalizable. Then, the synchronous
solution is stable if and only if for all eigenvalues μ ∈ C of the Laplacian matrix
the largest Lyapunov exponent of the following differential equations is smaller than
zero.

dζ
= (D f (s) + σρμH (0)DG(0)) ζ − σ G(0)κ (7.7)
dt
dκ
= − (ρμDH (0)ζ + κ) . (7.8)
dt

Here, ζ ∈ Cd and κ ∈ C.

Proof Due to Lemma 7.1.2 there are N 2 − N independent vectors wl (l = 1, . . . ,

N 2 − N ) spanning the kernel of B. Using the Gram-Schmidt procedure we find
an ortho-normal basis for ker(B) = span{v 1 , . . . , v N 2 −N }. With this we define the
N 2 × (N 2 − N ) matrix Q = (v 1 , . . . , v N 2 −N ). Consider now the (N 2 + N d) ×
(N 2 + N d) matrix

I 0 0
R = Nd
0 (1/r )B T Q

with left inverse

⎛ ⎞
IN d 0
R −1 = ⎝ 0 B ⎠,
0 QT
 
ξ̂ ξ
i.e., R −1 R = I N 2 +N d . Introduce the new coordinates given by R = for
χ̂ χ
which the variational Eqs. (7.5) read
  
d ξ̂ S −σ B ⊗ G(0) ξ̂
= R −1 R .
dt χ̂ −ρC ⊗ D H (0) −I N 2 χ̂

We further observe
138 7 Synchronization on Adaptive Complex Network Structures
 
S −σ B ⊗ G(0) S −σ I N ⊗ G(0) 0
R −1 R = R −1
−ρC ⊗ D H (0) −I N 2 −ρC ⊗ D H (0) −/r B T − Q
⎛ ⎞
S −σ I N ⊗ G(0) 0
= ⎝ −ρ L(A) ⊗ D H (0) −I N 0 ⎠.
−ρ Q T C ⊗ D H (0) 0 −I N 2 −N

These equations yield that there are the N d + N many coupled differential equations
left
  
d ξ̂ S −σ I N ⊗ G(0) ξ̂
= (7.9)
dt χ̃ −ρ L(A) ⊗ D H (0) −I N χ̃

with χ̃ = χ̂ 1 that determine the stability for the synchronous state and N 2 − N
equations
⎛ ⎞
d   ξ̂
χ̄ = −ρ Q T C ⊗ D H (0) 0 −I N 2 −N ⎝χ̃ ⎠
dt
χ̄

with χ̄ = (χ̂ 2T , . . . , χ̂ TN )T which are unidirectionally coupled to the variables ξ̂ and

χ̃ and can be solved explicitly once the latter once are known. By assumption, there
is a unitary matrix D L(A) = U H L(A)U where D L(A) is the diagonalization of the
Laplacian matrix L(A). Transforming the differential equations (7.9) by using the
unitary transformation U , we get
    
d ζ I N ⊗ D f (s) + σρ H (0) D L(A) ⊗ DG(0) −σ I N ⊗ G(0) ζ
=
dt κ −ρ D L(A) ⊗ D H (0) −I N κ
  
U ⊗ Id 0 ξ̂ ζ
where = . 
0 U χ̃ κ

By using the result in Proposition 7.1.3, the master stability function can be deter-
mined for any diffusive dynamical system on an adaptive network. Note that Eq. (7.8)
is explicitly solvable. the solutions reads
 t

κ = κ0 e−(t−t0 ) − ρμDH (0) e−(t−t ) ζ (t  ) dt  .
t0

Here, the first term vanishes as t → ∞. Hence, we can neglect it in order to study the
asymptotic dynamics of (7.7)–(7.8) which are rewritten in integro-differential form
(as t → ∞)
 t 
dζ −(t−t  )  
= D f (s)ζ + μ̃ H (0)DG(0)ζ + DH (0)G(0) e ζ (t ) dt . (7.10)
dt t0
7.1 The Master Stability Function for Adaptive Complex Networks 139

with master function parameter μ̃ = ρσ μ. We apply this result in the subsequent

sections and connect it to our understanding of the stability of multicluster states.

7.2 Stability Islands in the Presence of Adaptation

In this section, we apply the general result on the master stability function for com-
plex adaptive networks, as derived in the previous section. For this, we use a phase
oscillator model. Note, however, that the master stability approach presented above
is not restricted to this kind of networks.
In analogy with Chap. 6, we consider the following adaptive network of N coupled
phase oscillators

dφi  N
=1−σ ai j κi j sin(φi − φ j + α) (7.11)
dt j=1
dκi j  
= − κi j + ai j sin(φi − φ j + β) . (7.12)
dt

Let us further consider all networks with constant row-sum, i.e., r = Nj=1 ai j for
all i = 1, . . . , N . Then the synchronous solution of (7.11)–(7.12) is given by

φ sync = (1 − r σ sin(α)) t (7.13)

κ sync
= − sin(β). (7.14)

Note that this sort of solutions have been already exhaustively analyzed for the case
of a globally coupled network in Chap. 4. The synchronous state (7.13)–(7.14) is an
equilibrium relative to the shift symmetry of (7.11)–(7.12). Thus, the synchronous
state is independent of the value r .
In order to analyze the stability of the synchronous solution (7.13)–(7.14), we
have to linearize the Eqs. (7.11)–(7.12) around these solution. Using the result in
Proposition 7.1.3, the stability of the synchronous solution is governed by the two
dimensional set of equations
  
d ζ μσ cos(α) sin(β) −σ sin(α) ζ
= ,
dt κ −μ cos(β) − κ

where μ ∈ C is an eigenvalue of the Laplacian matrix L corresponding to the base

network described by the adjacency matrix A. The characteristic polynomial in λ of
the latter system is of degree two and reads

λ2 + ( − σ μ cos(α) sin(β)) λ − σ μ sin(α + β) = 0. (7.15)

140 7 Synchronization on Adaptive Complex Network Structures

Note that for μ = 0 the eigenvalues are λ = − and λ = 0 which corresponds to the
shift symmetry. With this, the master stability function for the Eqs. (7.11)–(7.12) is
given by max(Re(λ1 (μ̃))), Re(λ2 (μ̃))) where λ1,2 solve Eq. (7.15) for a given set of
parameters α, β and . Here, the master function parameter is given by μ̃ = σ μ.
Corollary 7.2.1 Let (7.13)–(7.14) be the synchronous solution for the Eqs. (7.11)–
(7.12) with σ > 0. Further, assume the Laplacian matrix L(A) is symmetric, i.e., the
network is undirected. Then the synchronous solution is stable if the parameters α
and β are from following set
 π π π 3π

{−π < α+β < 0} − <α < , −π < β < 0 <α< ,0 < β < π mod 2π.
2 2 2 2
(7.16)

Proof Since L(A) is symmetric, it is diagonalizable. Further, by definition, L(A)

is positive definite. Hence, L(A) possesses only positive real eigenvalues. As
shown above, we are allowed to use Proposition 7.1.3. From Eq. (7.15), we find
that the eigenvalues λ are strictly negative if and only if sin(α + β) < 0 and
( − σ μ cos(α) sin(β)) > 0. The first of the latter inequalities implies the neccessary
condition for stability −π < α + β < 0. The other inequality yields

> cos(α) sin(β).
σμ

Since , μ, σ > 0, the latter inequality implies the statement of the corollary. 

In corollary 7.2.1, a relation between stability and the parameters α and β is estab-
lished for undirected networks. In case of directed networks, the Laplacian eigenvalue
might be complex, and hence the stability cannot be immediately deduced from the
Eq. (7.15). Figure 7.1 displays the master stability function determined for different
values for the parameter of β. Here, we fix α and . The blue colored areas correspond
to master function parameter μ̃ = σ μ, where μ is an eigenvalues o the Laplacian,
that would lead to stable dynamics. By changing the control parameter β various
shapes of the stable regions are visible.
We, first, observe that the master stability function in all cases is symmetric with
respect to the real axis (Im(μ̃) = 0). This symmetry stems from the fact that if λ1,2
solve (7.15) for μ then λ1,2 solve (7.15) for μ where the overline indicates the complex
conjugate. Secondly, for some parameter, e.g. Fig. 7.1a, d, e, almost the whole half-
space from the left or right of the imaginary axis belong to the stable regime. Note that
in case of no adaptation the stability of the synchronous solution is solely described
by the sign of μ̃ sin(β) cos(α). As a third observation, we find parameters where
almost all of the μ̃ parameter space correspond to unstable dynamics, e.g. Fig. 7.1b,
c, f. However, there exist small islands, i.e., bounded regions in μ̃ parameter space,
that correspond to stable dynamics. To understand the emergence of these islands,
we further analyze the boundary that separates the stable (Re(λ1,2 ) > 0) from the
unstable region (Re(λ1,2 ) > 0).
7.2 Stability Islands in the Presence of Adaptation 141

(a) (b) (c)

max(Re(λ))
Im(μ̃)

(d) (e) (f)

Im(μ̃)

0
Re(μ̃) Re(μ̃) Re(μ̃)

Fig. 7.1 The figure shows the master stability function for the Eqs. (7.11)–(7.12). Regions belong-
ing to negative Lyapunov exponents are colored blue. The one dimensional curve where at least
one eigenvalue of (7.15) has zero real part is given as a black dotted line. Parameter: a β = −0.8π ,
b β = −0.2π , c β = −0.02π , d β = 0.05π , e β = 0.1π , and f β = 0.98π . The other parameters
are α = 0.3π and  = 0.01

In order to describe the boundaries between regions in μ̃ parameter space that

would lead to stable local dynamics or regions that would lead to unstable local
dynamics, we consider λ = iγ . Plugging this into the Eq. (7.15) we obtain the fol-
lowing parameterized expression for the boundary
 
− sin(α) cos(β)γ̃ 2 + i sin(α + β)γ̃ + cos(α) sin(β)γ̃ 3 )
μ̃(γ̃ ) = 
sin2 (α + β) + cos2 (α) sin2 (β)γ̃ 2

where we use γ̃ = γ /. The curves given by the latter parametrization of the bound-
ary are displayed in Fig. 7.1 as dotted lines and agree very well with the actual
borders. We further notice that the emergence of the stability islands is a direct
consequence of the presence of adaptation. This can be seen by using λ = iγ and
Eq. (7.15) where we put  = 0. This ansatz results in the simple linear boundary
equation μ̃ = iγ /(cos(α) sin(β)) which does not give rise to stability islands. Due
to the symmetry of the master stability function, a necessary condition to observe
a stability island is that the curve μ̃(γ̃ ) possesses two crossings with the real axis.
From Eq. 7.15, we thus deduce the following condition

Corollary 7.2.2 The master stability function of (7.11)–(7.12) possesses stability

islands if and only if
142 7 Synchronization on Adaptive Complex Network Structures

sin(α + β)
< 0. (7.17)
cos(α) sin(β)

The presence of stability islands is a very intriguing and unexpected effect introduced
by adaptation. In the following section the impact of stability islands on the formation
of multicluster is described.

7.3 Stability Islands and Implications for the Emergence of

Multicluster States

In this section, we explore the relation between the stability islands and the formation
of multiclusters. In order to do this, we perform the following numerical analysis.
We choose parameters α and β such that there exists a stability island. Then we
prepare the initial condition to be an in-phase synchronous state, i.e., φi = 0 and
κi j = − sin(β) and integrate numerically the system (7.11)–(7.12) with N = 200
oscillators for t = 30000. Thus, we perform an adiabatic continuation starting from
σ = 0.001 and ending at σ = 0.03. Note that σ = 1/N = 0.005 lies within this
range. In order to illustrate the impact of the stability island on the dynamics, we
perform this analysis for a globally coupled as well as for a complex network. The
complex network which we have chosen for the subsequent analysis is connected,
directed and has row sum r = 50. An illustration of the adjacency matrix can be
found in the Appendix A.8.
Figure 7.2 shows the cluster parameter RC , see (2.18) in Chap. 2, for different
values of the coupling constant σ . Note that RC = 1 refers to full in-phase synchrony
of the oscillators. We observe that for small σ the synchronous solution is stable,
see Fig. 7.2a. For the sake of simplicity, the coupling matrix κi j is not displayed
for all examples Fig. 7.2a–c. Here, the stability of the synchronous state is directly
implied by the master stability function. We note that all Laplacian eigenvalues μ of
a globally coupled network are given by μ = 0 and μ = N . In Fig. 7.2a, all master
function parameters σ μ lie within the stability islands.
By increasing the coupling constant, the master function parameter gets pushed
out of the regions of stability and the synchronous solution becomes unstable. For
intermediate values of σ the emergence of multiclusters with hierachical structure in
the cluster size is observed, see also Chap. 5. In Fig. 7.2b a multicluster states is shown
with three clusters. For numerical convenience, we only count groups of oscillators
with more than or equal to 10 members as a cluster. Note that for the systems (7.11)–
(7.12) in-phase synchronous and antipodal clusters have the same properties. In
Chap. 5, the role of the hierarchical structure of the cluster sizes have been discussed.
We have found that due to the frequency difference the coupling between the clusters
vanish on average. Hence we argued that the stability is effectively described by the
stability of the one-clusters of which the multicluster consists, see Sect. 5.6. Here, we
follow this argument and consider the subnetworks induced by the groups of nodes
(oscillators) with the same frequency. In the case of global coupling each of the
7.3 Stability Islands and Implications for the Emergence of Multicluster States 143

(a) (b) (c)

max(Re(λ))

Im(μ̃)
Re(μ̃)

(a)

φi
RC

Ωi
(b)

index i
(c)

Fig. 7.2 Results for an adiabatic continuation in the coupling constant σ of the full synchronous
solution φi = 0 and κi j = − sin(β) for a globally coupled network of N = 200 oscillators (7.11)–
(7.12). The cluster parameter RC for different values of σ is presented. As an inset, for the three
values a σ = 0.002, b σ = 0.006, and c σ = 0.025, the master stability function, the phases φi
of the final state, and the frequencies of the oscillators i are plotted in the respective column.
The oscillators are sorted as in Fig. 5.2. If more than or equal to ten (for numerical convenience)
oscillators have the same frequency (coherent groups) all nodes of this group are plotted as circles
and with a respective color. All other oscillators are plotted as an asterisk. The master function
parameter μ̃ = σ μ for the subnetworks induced by the coherent groups are plotted together with
the master stability function. The colors of each coherent group agree in all three plots. Parameters:
α = 0.49π , β = 0.88π ,  = 0.01

subnetworks is globally coupled, as well. The individual master function parameter

for the induced subnetworks of the three clusters are plotted together with the master
stability function. For all three subnetworks the master function parameters lie in
the stable region and thus the corresponding one-clusters are stable and hence the
multicluster. This formation of a multicluster is remarkable and is well explained
by the master stability function. In addition it is in very good agreement with the
numerical analysis in Ref. [6].
Increasing the coupling constant further shows the emergence of incoherence. In
Fig. 7.2c, we show the coexistence of a coherent and an incoherent cluster. These
states, also called chimera-like states, have been numerically analyzed in Refs. [6–8].
Surprisingly, also for these states, the stability of the coherent cluster is determined
by the master function parameter corresponding to the subnetwork induced by the
coherent nodes. This fact underlines, once more, the generality of the building block
approach outlined in Chaps. 4 and 5.
144 7 Synchronization on Adaptive Complex Network Structures

(a) (b) (c)

max(Re(λ))

Im(μ̃)
Re(μ̃)

(a)

φi
(b)
RC

Ωi

index i
(c)

Fig. 7.3 Results for an adiabatic continuation in the coupling constant σ of the full synchronous
solution φi = 0 and κi j = − sin(β) for a complex network (see Appendix A.8) of N = 200 oscil-
lators (7.11)–(7.12). The cluster parameter RC for different values of σ is presented. As an inset,
for the three values a σ = 0.003, b σ = 0.007, and c σ = 0.019, the master stability function, the
phases φi of the final state, and the frequencies of the oscillators i are plotted in the respective col-
umn. If more than or equal to ten (for numerical convenience) oscillators have the same frequency
(coherent groups) all nodes of this group are plotted as circles and with a respective color. All other
oscillators are plotted as an asterisk. The master function parameter μ̃ = σ μ for the subnetworks
induced by the coherent groups are plotted together with the master stability function. The colors
of each coherent group agree in all three plots. Parameters: α = 0.49π , β = 0.88π ,  = 0.01

In the following we show that the results obtained for the global network translate
to set-ups with complex network structure. In Fig. 7.3, again, the cluster parameter
RC for different values of the coupling constant σ is presented. As in the case of
global coupling, we observe that the synchronous state is stable for small values of the
coupling constant. After the destabilization of the synchronous state, the emergence
of a multicluster is shown, see Fig. 7.3b. In contrast to the globally coupled case,
the coherent groups do not necessarily induce subnetworks with a constant row
sum. Hence, clusters with slightly perturbed phase distribution occur. However, the
stability of the single clusters is again well described by the master stability function
since all master function parameters for the individual clusters lie within the stability
region. The same holds true for the state presented in Fig. 7.3c. Here, we observe the
coexistence of an incoherent and a coherent cluster. The stability of the latter can be
verified by the master stability function and the building block approach.
7.4 Summary 145

7.4 Summary

In this chapter, we have studied the stability of the synchronous solution for a complex
network of adaptively coupled oscillators. For this, the well-known master stability
approach is generalized to networks with adaptive and hence time evolving coupling
weights. The master stability approach is already well established and was general-
ized for various applications, e.g., to understand the stability of cluster synchronous
states [9–11], to study systems with single or distributed delays [3, 12–17], or even to
allow for discontinuous dynamical systems [18, 19]. In addition, the master stability
function has been used to understand effects in temporal [20, 21] as well as adaptive
networks [22, 23] within a static formalism. Beyond the local stability described
by the master stability function, Belykh et al. have developed the connection graph
stability method to provide analytic bounds for the global asymptotic stability of syn-
chronized states [24–29]. So far, however, all generalizations of the master stability
approach were either introduced for a static topological structure or the interdepen-
dence between node and coupling dynamics was not taken into account.
In the case of an adaptive network the topological structure is non-constant in time
and depends on the state of the network nodes. In turn, the nodal dynamics depends on
the coupling structure given by the coupling adaptive weights. This subtle interplay
changes the theoretical approach entirely that has to be taken in order to derive a
master stability function for adaptive networks. In this chapter, we have developed
a master stability approach and thus provided a novel extension towards adaptive
networks that can be regarded as non-standard. Studying network synchronization
by adapting the coupling weights [30–33] is only one field of application where our
new approach could be applied.
The master stability approach has been further applied to the paradigmatic model
of adaptively coupled phase oscillators. We have illustrated several forms for the
master stability function with respect to different adaptation rules, i.e., different val-
ues for the control parameter β. Remarkably, the emergence of bounded regions that
would lead to stable synchronous dynamics has been observed in the master stability
function. These bounded regions are called stability islands. We have described the
structure and formation of stability islands via a cubic curve and derived an exis-
tence criterion. Additionally, we have analytically provided sufficient conditions for
the system parameters which imply the stability of the synchronous solution on a
undirected background network structure.
In addition, the implications of the stability islands have been further explored.
Using an adiabatic continuation of the synchronous state with respect to the coupling
constant σ , we have obtained that the presence of a stability islands gives rise to
the emergence of multicluster states and chimera-like states. The stable existence
of certain multiclusters and chimera-like states have been shown numerically and
analytically explained using the master stability function. Thus, the findings do not
only complement the results obtained in Chaps. 5 and 6. They also underline the subtle
interplay between cluster size and network structure for the stability of complex
synchronization patterns [4, 5].
146 7 Synchronization on Adaptive Complex Network Structures

References

1. Pecora LM, Carroll TL (1998) Master stability functions for synchronized coupled systems.
Phys Rev Lett 80:2109
2. Keane A, Dahms T, Lehnert J, Suryanarayana SA, Hövel P, Schöll E (2012) Synchronisation
in networks of delay-coupled type-I excitable systems. Eur Phys J B 85:407
3. Lehnert J (2016) Controlling synchronization patterns in complex networks, Springer Theses.
Springer, Heidelberg
4. Berner R, Vock S, Schöll E, Yanchuk S (2021) Desynchronization transitions in adaptive
networks. Phys Rev Lett 126:028301
5. Vock S, Berner R, Yanchuk S, Schöll E (2021) Effect of diluted connectivities on cluster
synchronization of adaptively coupled oscillator networks. arXiv:2101.05601
6. Kasatkin DV, Yanchuk S, Schöll E, Nekorkin VI (2017) Self-organized emergence of multi-
layer structure and chimera states in dynamical networks with adaptive couplings. Phys Rev E
96
7. Kasatkin DV, Nekorkin VI (2018) Synchronization of chimera states in a multiplex system of
phase oscillators with adaptive couplings. Chaos 28
8. Kasatkin DV, Nekorkin VI (2018) The effect of topology on organization of synchronous
behavior in dynamical networks with adaptive couplings. Eur Phys J Spec Top 227:1051
9. Dahms T, Lehnert J, Schöll E (2012) Cluster and group synchronization in delay-coupled
networks. Phys Rev E 86
10. Pecora LM, Sorrentino F, Hagerstrom AM, Murphy TE, Roy R (2014) Symmetries, cluster
synchronization, and isolated desynchronization in complex networks. Nat Commun 5:4079
11. Sorrentino F, Pecora LM, Hagerstrom AM, Murphy TE, Roy R (2016) Complete character-
ization of the stability of cluster synchronization in complex dynamical networks. Sci Adv
2
12. Sorrentino F, Ott E (2007) Network synchronization of groups. Phys Rev E 76
13. Flunkert V, Yanchuk S, Dahms T, Schöll E (2010) Synchronizing distant nodes: a universal
classification of networks. Phys Rev Lett 105
14. Dahms T (2011) Synchronization in delay-coupled laser networks. Ph.D. thesis, Technische
Universität Berlin
15. Heiligenthal S, Dahms T, Yanchuk S, Jüngling T, Flunkert V, Kanter I, Schöll E, Kinzel W
(2011) Strong and weak chaos in nonlinear networks with time-delayed couplings. Phys Rev
Lett 107
16. Kyrychko YN, Blyuss KB, Schöll E (2014) Synchronization of networks of oscillators with
distributed-delay coupling. Chaos 24
17. Wille C, Lehnert J, Schöll E (2014) Synchronization-desynchronization transitions in complex
networks: an interplay of distributed time delay and inhibitory nodes. Phys Rev E 90
18. Ladenbauer J, Lehnert J, Rankoohi H, Dahms T, Schöll E, Obermayer K (2013) Adaptation
controls synchrony and cluster states of coupled threshold-model neurons. Phys Rev E 88
19. Coombes S, Thul R (2016) Synchrony in networks of coupled non-smooth dynamical systems:
extending the master stability function. Eur J Appl Math 27:904
20. Stilwell DJ, Bollt EM, Roberson DG (2006) Sufficient conditions for fast switching synchro-
nization in time-varying network topologies. SIAM J Appl Dyn Syst 5:140
21. Kohar V, Ji P, Choudhary A, Sinha S, Kurths J (2014) Synchronization in time-varying networks.
Phys Rev E 90
22. Zhou C, Kurths J (2006) Dynamical weights and enhanced synchronization in adaptive complex
networks. Phys Rev Lett 96
23. Sorrentino F, Ott E (2008) Adaptive synchronization of dynamics on evolving complex net-
works. Phys Rev Lett 100
24. Belykh VN, Belykh IV, Hasler M (2004) Connection graph stability method for synchronized
coupled chaotic systems. Phys D 195:159
25. Belykh IV, Belykh VN, Hasler M (2004) Blinking model and synchronization in small-world
networks with a time-varying coupling. Phys D 195:188
References 147

26. Belykh IV, de Lange E, Hasler M (2005) Synchronization of bursting neurons: what matters
in the network topology. Phys Rev Lett 94
27. Belykh IV, Belykh VN, Hasler M (2006) Generalized connection graph method for synchro-
nization in asymmetrical networks. Phys D 224:42
28. Belykh IV, Belykh VN, Hasler M (2006) Synchronization in asymmetrically coupled networks
with node balance. Chaos 16
29. Daley K, Zhao K, Belykh IV (2020) Synchronizability of directed networks: the power of
non-existent ties. Chaos 30
30. Yu W, DeLellis P, Chen G, di Bernardo M, Kurths J (2012) Distributed adaptive control of
synchronization in complex networks. IEEE Trans Autom Control 57:2153
31. De Lellis P, di Bernardo M, Garofalo F, Porfiri M (2010) Evolution of complex networks via
edge snapping. IEEE Trans Circuits Syst I 57:2132
32. Lehnert J, Hövel P, Selivanov AA, Fradkov AL, Schöll E (2014) Controlling cluster synchro-
nization by adapting the topology. Phys Rev E 90
33. Hövel P, Lehnert J, Selivanov A, Fradkov AL, Schöll E (2016) Adaptively controlled syn-
chronization of delay-coupled networks. In: Schöll E, Klapp SHL, Hövel P (eds) Control of
self-organizing nonlinear systems. Springer, Berlin, pp 47–63
Chapter 8
Multilayered Adaptive Networks

In this chapter, we show that a plethora of novel patterns can be generated by multi-
plexing adaptive networks. In particular, partial synchronization patterns like phase
clusters and more complex cluster states which are unstable in the corresponding
monoplex network can be stabilized, or even states which do not exist in the single-
layer case for the parameters chosen, can be born by multiplexing. Thus our aim is
to provide fundamental insight into the combined action of adaptivity and multiplex
topologies. Hereby we elucidate the delicate balance of adaptation and multiplexing
which is a feature of many real-world networks even beyond neuroscience [1–4]. As
local dynamics we use the paradigmatic Kuramoto phase oscillator model, which is
a simple generic model and has been successfully applied in the modeling of syn-
chronization phenomena in a wide range of natural and technological systems [5].
A general multiplex network with L layers each consisting of N identical adap-
tively coupled phase oscillators is described by

1  μ 
N L
μ μ μ μ
φ̇i = ω − κi j sin(φi − φ j + α μμ ) − σ μν sin(φi − φiν + α μν ),
N j=1 ν=1,ν=μ
 
μ μ μ μ
κ̇i j = − κi j + sin(φi − φ j + β μ ) ,
(8.1)

where φiμ ∈ [0, 2π ) represents the phase of the ith oscillator (i = 1, . . . , N ) in the
μth layer (μ = 1, . . . , L), and ω is the natural frequency. The interaction between the
μ
phase oscillators within each layer is described by the coupling matrix elements κi j ∈
[−1, 1]. The intra-layer coupling weights κiμj are determined adaptively, whereas the
inter-layer coupling weights σ μν ≥ 0 are fixed. The parameters α μν are the phase
lags of the interaction [6]. The adaptation rate 0 <   1 is assumed to be a small
parameter separating the time scales of the slow dynamics of the coupling weights

© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 149
R. Berner, Patterns of Synchrony in Complex Networks of Adaptively Coupled Oscillators,
Springer Theses, https://doi.org/10.1007/978-3-030-74938-5_8
150 8 Multilayered Adaptive Networks

and the fast dynamics of the oscillatory system. The adaptation function in each layer
is chosen as in the Chaps. 4 and 5.
Let us note important properties of the model. First, ω can be set to zero with-
out loss of generality due to the shift-symmetry of Eq. (8.1), i.e., considering the
co-rotating frame
 φ→  φ + ωt. Moreover, due to the existence of the attracting
region G ≡ φi , κi j : φiμ ∈ (0, 2π ], |κiμj | ≤ 1, i, j = 1, . . . , N , μ = 1, . . . , L},
μ μ

one can restrict the range of the coupling weights to the interval −1 ≤ κi j ≤ 1 [7].
Finally, based on the parameter symmetries of the model

(α, β, φ, κ) → (−α, π − β, −φ, κ),

μμ μ μ μ μ
(α , β μ , φi , κi j ) → (α μμ + π, β μ + π, φi , −κi j ),

where α, β, φ, κ abbreviate the whole set of variables and parameters, it is sufficient

to analyze the system within the parameter region α 11 ∈ [0, π/2), α μμ ∈ [0, π ) (μ =
1), α μν ∈ [0, 2π ) (μ = ν) and β μ ∈ [−π, π ).
This chapter reproduces contents that have been published in [8] with the permis-
sion of American Physical Society (APS) ( c 2020 American Physical Society (APS),
https://doi.org/10.1103/PhysRevLett.124.088301). The chapter is organized as fol-
lows. In Sect. 8.1, we introduce the notion of lifted states and show how states from
one-layer can be used to find states for multiplex systems. Additionally, the existence
of new states induced by the muliplex structure is shown in Sect. 8.2. We describe the
properties of these states and study numerically their robustness against heterogene-
ity ind the local dynamics. In order to study the stability of lifted states, we develop
the new methodology of the multiplex decomposition in Sect. 8.3. Subsequently in
Sect. 8.4, the method is applied to lifted states to prove the stabilizing features of the
multiplex set-up. In Sect. 8.5, we provide a brief outlook on further applications of
the multiplex decomposition. All findings are summarized in Sect. 8.6.

8.1 Lifted States in Multiplex Networks

In Chap. 4, we exhaustively described the properties of one-cluster in case of one

layer and provided rigorous existence and stability results. In this section, we lift
these finding to the case of two layers.
Let us now consider one-cluster states in multiplex structures. In general, one-
cluster states are given as relative equilibria
μ μ
φi = t + ai ,
μ μ μ (8.2)
κi j = − sin(ai − a j + β μ ),

μ
with collective frequency and relative phases ai ∈ [0, 2π ). In order to connect
one-cluster states of the single layer case to one-cluster states of the multiple layer
case, we introduce the notion of lifted one-cluster states.
8.1 Lifted States in Multiplex Networks 151

Definition 8.1.1 Let (φ 1 , . . . , φ L , κ 1 , . . . , κ L ) be a one-cluster state (8.2) solving

the Eqs. (8.1). Then (φ μ , κ μ ) is called a lifted one-cluster state if for each layer
μ = 1, . . . , L the state (φ μ , κ μ ) is a monoplex one-cluster, i.e., (φ μ (t), κ μ (t))
solves (6.1)–(6.2) and hence the distribution of phases aμ are of splay, antipodal,
or double antipodal type.
In the following we show that in duplex systems (L = 2) the phase difference of
oscillators between the layers a ≡ ai1 − ai2 takes only two values and solves

= σ 12 sin( a + α 12 ) + σ 21 sin( a − α 12 ), (8.3)

where ≡ (α 11 , β 1 ) − (α 22 , β 2 ) is given in (4.6) for the three different one-

cluster states (splay, antipodal, double antipodal).
To show this, suppose we have two one-cluster states where each is of either
splay, antipodal, or double antipodal type which form a duplex one-cluster (8.2) for
μ μ
L = 2 and φi = (α μμ , β μ )t + ωμ t + ai (μ = 1, 2), where (α μμ , β μ ) is given
by (4.6) and the coupling weights are given by κiμj = − sin(aiμ − a μj + β μ ). We verify
μ μ
by directly inserting that φi and κi j solve Eq. (2.26). For the given ansatz, Eq. (8.1)
reads
1 1  
cos(α 11 − β 1 ) − Re e−i(2ai +α +β ) Z 2 (a1 )
1 11 1
(α 11 , β 1 ) + ω1 =
2 2
− σ 12 sin( t+ ωt + ai1 − ai2 + α 12 ),

and
1 1  
cos(α 22 − β 2 ) − Re e−i(2ai +α +β ) Z 2 (a2 )
1 22 2
(α 22 , β2 ) + ω2 =
2 2
+ σ 21 sin( t+ ωt + ai1 − ai2 − α 21 ).

where = (α 11 , β 1 ) − (α 22 , β 2 ) and ω = ω1 − ω2 , respectively. Thus, φ =

(φ , φ , κ , κ 2 ) is a duplex one-cluster if
1 2 1

+ ω=0

which is equivalent to

= σ 12 sin(ai1 − ai2 + α 12 ) + σ 21 sin(ai1 − ai2 − α 21 )

for all i = 1, . . . , N . Note that is not necessarily zero even if the phase-lag
parameters for both layers agree. They can still differ in the type of one-cluster state.
The former equation can be written as

= sin(ai1 − ai2 + ν) (8.4)

C
with
152 8 Multilayered Adaptive Networks

1  12 
sin(ν) = σ sin(α 12 ) − σ 21 sin(α 21 ) , (8.5)
C
1  12 
cos(ν) = σ cos(α 12 ) + σ 21 cos(α 21 ) ,
C
where

C= (σ 12 )2 + (σ 21 )2 + 2σ 12 σ 21 cos(α 12 + α 21 ).

Whenever (σ 12 )2 + (σ 21 )2 + 2σ 12 σ 21 cos(α 12 + α 21 ) ≥ 0 and

(σ 12 )2 + (σ 21 )2 + 2σ 12 σ 21 cos(α 12 + α 21 ) ≥ 2
, (8.6)

Equation (8.4) has the two solutions ai1 − ai2 = arcsin( /C) − ν and ai1 − ai2 =π −
arcsin( /C) − ν. Considering the inverse function arcsin : [−1, 1] → [−π/2,
π/2] applied to Eq. (8.5) determines ν to be either ν or π − ν , where ν :=
arcsin(sin(ν)) and sin(ν) as given in (8.5). The second equation for cos(ν) then
fixes ν to take one of the values.
The condition (8.6) is a relation between all parameters of the system which
has to be fulfilled for the existence of duplex relative equilibria. Note that for any
given inter-layer coupling σ 12 = 0 and α 12 + α 21 = ±π/2 or ±3π/2 there exists a
minimum coupling weight σ 21 < ∞ such that the lifted one-clusters exist. In case
of unidirectional coupling, i.e., σ 12 = 0, the condition gives the minimum weight
σ 21 ≥ .
In Fig. 8.1 different duplex one-cluster states are presented that are observed by
numerical simulations. Panels (a),(b),(d) in Fig. 8.1 display lifted states of splay,
antipodal, and splay type, respectively. The phase distributions in both layers are the
same but shifted by the constant value a in agreement with the above equation. In
contrast to the lifted states, Fig. 8.1c shows another possible one-cluster for the duplex
network. Due to the interaction of the two layers we can find a phase distribution
which is of double antipodal type in each layer but not a lifted state. This means
that these states are born by the duplex set-up. Moreover, in contrast to the other
examples the phase distribution between the layers does not agree, ψ 1 = ψ 2 . For the
monoplex case, it has been shown that double antipodal states are unstable for any
set of parameters, see Chap. 4. Hence, finding stable double antipodal states which
interact through the duplex structure is unexpected.

8.2 Birth and Robustness of Phase Clusters

For more insight into the birth of phase-locked states by multiplexing, Fig. 8.2 dis-
plays the emergence of double antipodal states in a parameter regime where they
do not exist in single-layer networks. They are characterized by the second moment
order parameter R2 . It is remarkable that the new double antipodal state can be found
8.2 Birth and Robustness of Phase Clusters 153

Layer 1 Layer 2 Layer 1 Layer 2

50
(a) (b)
30
i
10
2π
π
π
φjµ

π Δa
0
50
(c) (d)
30
i

10
2π

π
π ψ1 ψ2
φµj

π Δa
0
10 30 50 10 30 50 10 30 50 10 30 50
j j

Fig. 8.1 Different duplex states of Eq. (4.3) (L = 2) for an ensemble of 50 oscillators in each layer
μ μ
with color-coded coupling weights κi j (upper panels, color code as in Fig. 8.1), phases φ j (lower
μ
panels): Duplex one-cluster states (a) of lifted splay type (R2 (a ) = 0) for α 12/21 = 0.3π , σ 12/21 =
0.07; b of lifted antipodal type (R2 (aμ ) = 1) for α 12 = 0.3π , α 21 = 0.75π , σ 12/21 = 0.62; c of
double antidodal type (not a lifted state) for α 12/21 = 0.05π , σ 12/21 = 0.28; d of lifted splay type
for α 12 = 0.3π , α 21 = 0.4π , σ 12/21 = 0.8, and  = 0.01. In the lower panels phase differences
between the two layers are indicated by a ≡ ai1 − ai2 , and between the two new antipodal states
(c) by ψ 1 , ψ 2 . Figure reproduced from [8]. c 2020 American Physical Society (APS)

for a wide range of the inter-layer coupling strength larger than a certain critical value
σc , and is clearly different from those of the monoplex. Below the critical value σc , the
double antipodal states are no longer stable, and more complex temporal dynamics
occurs which causes temporal changes in R2 . This leads to non-vanishing temporal
variance indicated by the error bars in Fig. 8.2.
So far, we have investigated a system of identical oscillators. There the existence
of particular phase cluster states of double-antipodal type is demonstrated in Figs. 8.1
and 8.2. In order to show that these states are also present in a system of heterogeneous
phase oscillators, we modify the Eqs. (8.1)–(2.26) as follows:
154 8 Multilayered Adaptive Networks

1  μ 
μ N L
dφi
= ωiμ − κi j sin(φiμ − φ μj + α μμ ) − σ μν sin(φiμ − φiν + α μν ),
dt N j=1 ν=1,ν=μ
(8.7)
μ
dκi j  
μ μ
= − κi j + sin(φi − φ νj + β μ ) . (8.8)
dt
For the numerical analysis of system (8.7)–(8.8), we consider randomly uniformly
distributed natural frequencies on the interval [− ω, ω]. To check for the robust-
ness of the phase cluster presented in the inset of Fig. 8.2, the following steps are
performed. We fix a random realization of a uniform distribution. For any inter-
layer coupling strength σ = σ μν we take the final state from the simulation with
ωiμ = ωiν = 0 (i.e., those obtained from Fig. 8.2) as initial condition. We perform
an adiabatic continuation of the state by running the simulation for t = 5000 and
increasing the width of the distribution ω with a stepsize of 0.01. This is done
until ω reaches 0.5. The continuation is performed for each value of σ and for 10
different realizations of the uniform distribution. Afterwards, we first check whether
the final state has still the same form as the one presented in Fig. 8.2. For this, we cal-
culate the second moment order parameter for both states in each layer individually,
determine the difference of the order parameters for both layers, and set the upper
limit to 0.01. States with a difference of less than the limit are considered to possess
the same form. Secondly, we check whether the final state is still a phase-locked
state, i.e., all oscillators are frequency-synchronized. The range and the boundaries
up to which the final state is still a duplex one-cluster state with or without the form

1 10
i

0.8 5

2π
0.6 π
φµj
R2

π π
0.4 ψ1 ψ2
0
0.2 5 10 5 10
j
R2 (φ1 )
0
0 σc 0.5 1 1.5 R2 (φ2 )
σ

Fig. 8.2 Birth of double antipodal state in a duplex network (N = 12) for a wide range of inter-
layer coupling strength σ = σ 12 = σ 21 . The solid lines are the temporal averages for the second
moment order parameter R2 of the individual layers (layer 1: black, layer 2: red). The error bars for
σ < σc denote the standard deviation of the temporal evolution of R2 . The dashed horizontal lines
represent the unique values of R2 for the double antipodal state in a monoplex network. The plot was
obtained by adiabatic continuation of a duplex double antipodal state (see inset) in both directions
starting from σ = 0.5. Parameters: α 11/22 = 0.3π , α 12/21 = 0.05, β 1 = 0.1π , β 2 = −0.95π , and
 = 0.01. Figure reproduced from [8]. c 2020 American Physical Society (APS)
8.2 Birth and Robustness of Phase Clusters 155

10 (a) (b)

i
5

2π
π
π ψ2
φµ
j
π π 1
ψ ψ2 π ψ1
0
5 10 5 10 5 10 5 10
j j
Δω

Fig. 8.3 The figure shows the range ω where duplex one-cluster states in general (gray) and of the
form presented in Fig. 8.2 (red) can be found. For this the system (8.7), (8.8) is integrated numerically
for 10 different random uniform distributions of the natural frequencies in the interval [− ω, ω].
The results are obtained by adiabatic continuation starting with the phase clusters found for ω = 0
(see Fig. 8.2). Duplex one-cluster states of double antipodal type with (a) σ = 0.5, = 0.02 and
b σ = 0.5, = 0.07 are shown as insets. Parameters: α 11/22 = 0.3π , α 12/21 = 0.05, β 1 = 0.1π ,
β 2 = −0.95π ,  = 0.01, and N = 12. Figure reproduced from [8]. c 2020 American Physical
Society (APS)

from Fig. 8.2 are presented in Fig. 8.3. For the boundaries and the range the mean
value over the 10 realizations is determined and the error bars indicate the standard
deviation.
It is clearly visible that duplex one-cluster states of double-antipodal type are still
present for a considerable range of heterogeneity ω of the natural frequencies. In
the inset Fig. 8.3a we present a duplex one-cluster double-antipodal state of the same
form as shown in Fig. 8.2. The phases are distorted slightly due to the frequency
distribution but the double-antipodal configuration is still clearly visible. The inset
Fig. 8.3b shows another one-cluster state of double-antipodal type. However, due to
the frequency mismatch the phase distribution becomes different compared with the
one presented in Fig. 8.2.
A similar result, as it is shown in Fig. 8.3, is obtained if we consider a Gaussian
instead of an uniform distribution of frequencies, see Fig. 8.4.
156 8 Multilayered Adaptive Networks

10

i
5

2π
π ψ2
φµj
π
π
ψ1
0
5 10 5 10
j
ρ

Fig. 8.4 The figure shows the standard deviation ρ where duplex one-cluster states in general (gray)
and of the form presented in Fig. 8.2 (red) can be found. For this the system (8.7), (8.8) is integrated
numerically for 10 different random normal distributions of the natural frequencies with standard
deviation ρ and zero mean. The results are obtained by adiabatic continuation starting with the phase
clusters found for ρ = 0 (see Fig. 8.2). A duplex one-cluster state of double antipodal type with
σ = 0.5, = 0.02 is shown as an inset. Parameters: α 11/22 = 0.3π , α 12/21 = 0.05, β 1 = 0.1π ,
β 2 = −0.95π ,  = 0.01, and N = 12. Figure reproduced from [8]. c 2020 American Physical
Society (APS)

8.3 Multiplex Decomposition

In this section, we provide important tools and theorems to find the spectrum of
multiplex networks. These results are then subsequently used to analyze the stability
of lifted one-cluster states.
Let us start with a general result on the determinant of block matrices.
Theorem 8.3.1 Let R be a commutative subring of C N ×N and let M ∈ R L×L . Then,

detC M = det C (det R M) .

The proof can be found in Refs. [9, 10]. This rather abstract result allows for a very
nice decomposition for pairwise commuting matrices and yields a useful tool to study
the local dynamics in multiplex systems.
Proposition 8.3.2 Let M ∈ C N ×N be a unitary diagonalizable matrix with M =
U D M U H where U , U H and D M are a unitary, its adjoint and a diagonal matrix,
respectively. Let further D M be the set of simultaneously diagonalizable matrices to
8.3 Multiplex Decomposition 157

M, i.e., the set of all matrices which commute pairwise with M. Then,
⎛ ⎞ ⎛ ⎡ ⎤⎞
A11 · · · A1L
⎜ .. . . .. ⎟  L
det ⎝ . . . ⎠ = det ⎝ ⎣sgn(σ ) D Aμ,σ (μ) ⎦⎠ (8.9)
A L1 · · · A L L σ ∈SL μ=1

where Aμν ∈ D M for μ, ν = 1, . . . , L and SL is the set of all permutations of the

numbers 1, . . . , L.
Proof Consider any A, B ∈ D M , then they are simultaneously diagonalizable with
M and hence A = D A U H and B = U D B U H with the same U . Thus, all Aμν can be
diagonalized with the same U . Since U is unitary,i.e. (det U )2 = 1, we find
⎛ ⎞ ⎛ ⎞
A11 · · · A1L D A11 · · · D A1L
⎜ .. . . .. ⎟ ⎜ .. .. . ⎟
det ⎝ . . . ⎠ = det ⎝ . . .. ⎠
A L1 · · · A L L D A L1 · · · D AL L

by applying the block diagonal matrices diag(U, · · · , U ) and diag(U H , · · · , U H )

from the left and right, respectively. The set of diagonal matrices with usual matrix
multiplication and addition form a commutative subring of C N ×N . Applying The-
orem 8.3.1 and using the well-known determinant representation of Leibniz, the
expression (8.9) follows. 
Remark 8.3.3 The set D M consists of all matrices which commute with M and
all the other elements of D M . In particular, the identity matrix I N ∈ D M for any
M ∈ C N ×N .
In the following, we apply the last result to a duplex and triplex system and connect
the local dynamics on the one-layer network to the multiplex case. We specify our
consideration by defining two special multiplex systems.
Definition 8.3.4 Suppose A, B, C ∈ C N ×N and m i j ∈ C (i, j = 1, . . . , 3). Then,
the 2N × 2N block matrix
 
(2) A m 12 I
M = (8.10)
m 21 I B

and the 3N × 3N block matrix

⎛ ⎞
A m 12 I m 13 I
M (3) = ⎝m 21 I B m 23 I⎠ (8.11)
m 31 I m 32 I C

are called (complex) duplex and triplex network, respectively.

Suppose we know how to diagonalize the individual layer topologies. The next result
shows how the eigenvalues of the individual layers are connected to eigenvalues of
158 8 Multilayered Adaptive Networks

the multiplex system. This will be done for the duplex and triplex network. For the
proof of our following statement, we provide two different ways.
The first approach makes use of the Schur decomposition [11, 12], see (4.11) in
Chap. 4, which will be used, later on, in order to derive the characteristic equations.
An extension of the first approach to any number of layers in the network can be
found by induction but is very technical, see [10, 13, 14]. The second approach uses
Proposition 8.3.2 which allows for a straightforward extension to any number of
layers in a multiplex network.
Proposition 8.3.5 Suppose A, B, C ∈ C N ×N , they commute pairwise, and are diag-
onalizable with diagonal matrices D A , D B , DC and unitary matrix U . Then, the
eigenvalues μ for the multiplex networks M (2) and M (3) can be found by solving the
N quadratic

μ2 − ((d A )i + (d B )i ) μ + (d A )i (d B )i − m 12 m 21 = 0 (8.12)

and cubic polynomial equations

μ3 + a2,i μ2 + a1,i μ + a0,i = 0, (8.13)

respectively, with

a2,i = − ((d A )i + (d B )i + (dC )i )

a1,i = (d A )i (d B )i + (d A )i (dC )i + (d B )i (d D )i
− m 12 m 21 − m 13 m 31 − m 23 m 32
a0,i = m 12 m 21 (dC )i + m 13 m 31 (d B )i + m 23 m 32 (d A )i
− (d A )i (d B )i (dC )i − m 12 m 23 m 31 − m 13 m 32 m 21

and (d A )i , (d B )i , and (dC )i being the respective diagonal elements of D A , D B , and

DC .

Proof Since A, B, C are diagonalizable and commute, Proposition 8.3.2 can be

applied to both matrices M (2) , M (3) . Anyhow, for the matrix M (2) we will provide
another proof using Schur’s decomposition.
The determinant is an antisymmetric multilinear form. Thus, we can write
     
A − μI N m 12 · I N m 12 · I N A − μI N
det M (2) − μI2N = det = (−1) N det
m 21 · I N B − μI N m B − μI N m 21 · I N .

By assumption A and B are both diagonalizable with respect to the unitary transfor-
mation matrix U , and so are A − μI and B − μI. This allows us to write
   
m 12 · I N A − μI N m 12 I N D A − μI N
det = det
B − μI N m 21 · I N . D B − μI N m 21 I N
8.3 Multiplex Decomposition 159

by applying the block diagonal matrices diag(U, · · · , U ) and diag(U H , · · · , U H )

from the left and right, respectively. Now, using Schur’s decomposition (4.11) the
determinant can written as
   
m 12 I N D A − μI N 1
det = n N det m − (D A − μI N ) (D B − μI N )
D B − μI N m 21 I N n
= det (m 12 m 21 I N − (D A − μI N ) (D B − μI N )) .
 
The last expression together with det M (2) − μI2N = 0 yields the N quadratic
equations (8.12).
Using that (A − μI), (B − μI), (C − μI) commute pairwise, Proposition 8.3.2
can be applied. We find
 
det M (3) − μI3N = det ((D A − μI N ) [(D B − μI N )(DC − μI N ) − m 23 m 32 I N ]

−m 21 [m 12 (DC − μI N ) − m 13 m 32 I N ] + m 31 [m 12 m 23 I N − m 13 (D B − μI N )])

 
The last expression together with det M (3) − μI3N = 0 yields the N cubic equa-
tions (8.13). 

Let us briefly discuss some special cases for both the duplex and triplex network.
Consider a duplex network with master and slave layer, i.e., either m 12 = 0 or m 21 =
0. Then, the quadratic equations (8.12) yield

(μ − (d A )i ) (μ − (d B )i ) = 0. (8.14)

As shown in Proposition 8.3.2, the eigenvalues for special triplex networks can be
found by solving cubic equations. For the solution even closed forms exist. Despite
this, the explicit form of the solutions is rather tedious, in general. However, if we
consider A = B = C and a ring-like inter-layer connection between the networks,
i.e., m 12 = m 23 = m 31 = 0, then Eq. (8.13) has the following solutions for all j =
1, . . . , N

μ1 = −(d A ) j + (m 13 m 32 m 21 )1/3 ,
1 √
μ2 = −(d A ) j + i(i + 3) (m 13 m 32 m 21 )1/3 ,
2
1 √
μ3 = −(d A ) j − (i + 3) (m 13 m 32 m 21 )1/3 ,
2

where i denotes the imaginary unit. In analogy to Eq. (8.14), a decoupling for the
eigenvalues can be found. Consider three pairwise commuting matrices A, B, C,
and the structure between the layers is a directed chain, i.e., m 12 = m 13 = m 31 =
m 23 = 0, then
160 8 Multilayered Adaptive Networks

(μ − (d A )i ) (μ − (d B )i ) (μ − (dC )i ) = 0. (8.15)

In the following we apply the results of this section to the stability analysis of lifted
duplex one-cluster states. In the last Sect. 8.5 of this chapter, further two different
applications of the multiplex decomposition are briefly discussed.

8.4 Stabilizing Through Multiplexing

In the following we show how the dynamics in a neighborhood of a single layer one-
cluster state can be lifted to the dynamics in a neighborhood of lifted multiplex one-
cluster state, i.e., we investigate and relate their local stability features. Everything
is exemplified for antipodal type states but can be generalized to lifted splay and
lifted double antipodal states in a straightforward manner. To study the dynamics
around the one-cluster states described by Eq. (8.2), we linearize Eq. (8.1) around
these states:

1 
N
˙ iμ =
δφ
μμ
sin( a + β μ ) cos( a + α μμ ) i j δφ
N j=1
μ
M
μν
μμ
− sin( a + α )δκi j − σ μν cos( a + α μν ) i j δφ,
ν=1
 
˙ iμj = − δκiμj + cos( a + β μ )
δκ
μμ
i j δφ (8.16)

μν μ
where i j δφ ≡ δφi − δφ νj .
Consider a duplex antipodal one-cluster state Eq. (8.2) with ai1 ∈ {0, π } and ai2 =
ai − a, Eq. (8.16) can be brought to the duplex form in Definition (8.3.4) and
1

possesses the following set of eigenvalues

 
SDuplex = {−, λi,1 , λi,2 , λi,3 , λi,4 i=1,...,N }

where λi,1,...,4 solve the following N quartic equations

        
(λ + )2 m 1 m 2 − 1
λ − ρi,1 1
· λ − ρi,2 + m 1 (λ + ) 2
λ − ρi,1 2
λ − ρi,2 + m 2 (λ + ) = 0,
(8.17)
μ
with m 1 = σ 12 cos( a + α 12 ), m 2 = σ 21 cos( a − α 21 ) and the eigenvalues ρi,1,2 ≡
ρi,1,2 (α μμ , β μ ) for the monoplex system, see Corollary 4.2.3. Using the real parts
of the eigenvalues of this analysis for the duplex antipodal clusters, the stability for
duplex antipodal states is found. All details and the proof of (8.17) is provided in the
Appendix A.7.
We have seen that the stability analysis of the duplex system can be reduced to
that of the monoplex case. Thus, we are now able to analyze the stabilizing and
8.4 Stabilizing Through Multiplexing 161

σ 12 = 0 σ 12 = 0.3 σ 21
(a) (b)

α22 /π

β 2 /π β 2 /π

Fig. 8.5 Regions of stability (blue) and instability (white) of the lifted antipodal state in the
(α 22 , β 2 ) parameter plane for different values of interlayer coupling (indicated by different blue
shading) σ 21 , where regions of stronger coupling σ 21 (lighter blue) include such of weaker σ 21
(darker blue). Stability regions for single-layer antipodal clusters are indicated by red hatched
areas. The inter-layer coupling is considered as a unidirectional (σ 12 = 0) and b bidirectional
(σ 12 = σ 21 ). Parameters: α 11 = 0.2π , β 1 = −0.8π , α 12 = 0, α 21 = 0.3π , and  = 0.01. Figure
reproduced from [8]. c 2020 American Physical Society (APS)

destabilizing features of a duplex network numerically and analytically. To illustrate

the effect of multiplexing, the interaction between two clusters of antipodal type
is presented in Fig. 8.5. The stability of these states is determined by integrating
Eq. (4.3) numerically starting with a slightly perturbed lifted antipodal state. The
states are stable if the numerical trajectory is approaching the lifted antipodal state.
Otherwise, the state is considered as unstable. The black contour lines in Fig. 8.5
show the borders of the stability regions in dependence of the coupling strength σ 21 ,
as calculated from the Lyapunov exponents. The borders are in remarkable agreement
with the numerical results.
We investigate two different situations in Fig. 8.5: In both panels the parameters for
the first layer α 11 , β 1 are chosen such that the antipodal state is stable without inter-
layer coupling. The stability of the duplex antipodal states is displayed in the (α 22 , β 2 )
parameter plane for several values of the inter-layer coupling σ 21 . To compare the
effects of the duplex network with the mono-layer case, the stability regions for
monoplex antipodals states are displayed, as well, as red hatched areas. They are
markedly different. In Fig. 8.5a, the two layers are connected unidirectionally (σ 12 =
0). It can be seen that with increasing inter-layer coupling weight σ 21 the region of
stability for the lifted antipodal state also grows. Already for small values of the inter-
layer couplings σ 21 , a stabilizing effect of the duplex network can be noticed. For
σ = 0.1 there exist already regions for which the duplex antipodal state is stable but
the corresponding monoplex state would not be stable. The opposite effect is found
as well where the duplex network destabilizes a lifted state. Figure 8.5b shows the
results for two layers with bidirectional coupling. In this case, the duplex structure
can have stabilizing and destabilizing effects, as well. Further, for the bidirectional
coupling we also notice a growth of the stability region with increasing σ 21 similar
to the unidirectional case. However, the regions of stability grow at different rates in
162 8 Multilayered Adaptive Networks

dependence on σ 21 and non-monotonically with respect to the parameters α 22 , β 2 .

Comparing the size of the stability region for both cases, one can see that for small
values of σ 21 the region for bidirectional coupling is larger. In turn, for higher inter-
layer coupling, the regions for the unidirectional case are larger. It is worth noting
that in Fig. 8.5 the stability regions for smaller values of σ 21 are always contained in
the region for larger values of σ 21 .

8.5 Applications for the Multiplex Decomposition

In the previous section, we have shown how the multiplex decomposition can be
applied to understand the stability features of lifted states. Here, we provide two
additional perspective where the newly derived decomposition can be used to gen-
eralize existing results and make others analytically accessible.

8.5.1 The Master Stability Approach for Multiplex Networks

In Ref. [15], the master stability function for dynamical systems on multiplex net-
works was introduced. In their article, the authors considered diffusive systems sim-
ilar to (2.10). Consider now the synchronous solution which solves ṡ = f (s). The
master stability function is then derived from the following variational equations
    
ξ̇ = I N L ⊗ D f (s) − σ Lintra ⊗ H − ρ Linter ⊗ G ξ , (8.18)

where H and G are linear coupling functions. Here, ξ = x − I L ⊗ I N ⊗ s and x ∈

C L·N ·d is the system state vector where all individual nodal states are stacked on
each other ordered by the layer and node index. Further, the intra-layer Laplacian is
defined is defined as
⎛ 1 ⎞
L
L
⎜ ⎟
Lintra = Ll = ⎝ . . . ⎠
l=1 L
L

where
⎛ N k ⎞
j=1 a1 j
⎜ .. ⎟
Lk = ⎝ . ⎠− A .
k
N k
j=1 a N j

The inter-layer Laplacian is defined as Linter = L I ⊗ I N where

8.5 Applications for the Multiplex Decomposition 163
⎛ L 1l
⎞
l=1 m
⎜ .. ⎟
LI = ⎝ . ⎠ − M,
L Ll
l=1 m

where m kl with m ll = 0 are the entries of the L × L matrix M. All further necessary
details on the model equations can be found in Ref. [15].
In Ref. [15], it was shown that if Lintra and Linter commute, a master stability
equation for system (8.18) can be found which reads

ẏ = [D f (s) − α H − βG] y,

where α = σ λ, β = ρμ, λ and μ are the (complex) eigenvalues of Lintra and Linter ,
respectively. Note that if H = G the master stability equation can be reduced to
 
ẏ = D f (s) − γ H y, (8.19)

with γ = α + β and y ∈ Cd . Equation (8.19) is called the master stability equation

for the composite system where a single supra-Laplacian matrix σ Lintra + ρ Linter
described the network topology [16, 17]. 
Direct evaluation shows that Lintra , Linter = 0 is equivalent
 intra  to m kl L l − m lk L k =
0 for all l, k = 1, . . . , L. Thus, to require L , L inter
= 0 yields a linear depen-
dence between the individual layer topologies. Using Proposition 8.3.4 for composite
system, the master stability function can be used under much milder conditions.
Proposition 8.5.1 Let us consider the multiplex dynamical system (8.18), with linear
function H = G. Suppose that further all L l , l = 1, . . . , L, commute pairwise. Then,
the master stability equation is given by

ẏ = [D f (s) − μH ] y, (8.20)

and μ = μ(λi1 , . . . , λli ) being a non-linear function, mapping the ith eigenvalues
λli (i = 1, . . . , N ) of the layer topologies L l , to the master function parameter μ
in (8.20). The non linear mapping is given as the formal solution to the Lth order
polynomial equation
  
 
L
det sgn(σ ) D(Llσsupra
(l) −μI N δlσ (l) )
= 0, (8.21)
σ ∈SL l=1

where Lsupra = σ Lintra + ρ Linter is a L × L block matrix divided into N × N

supra
matrices which we individually refer to with Lkl , k, l = 1, . . . , L, and δkl is the
Kronecker symbol.

Proof Using H = G, the variation equation for (8.18) on the synchronous solution
s(t) is given by
164 8 Multilayered Adaptive Networks
 
ξ̇ = I N L ⊗ D f (s) − Lsupra ⊗ H ξ ,

By assumption all L l , l = 1, . . . , L, commute pairwise and Linter is a L × L block

matrix consisting of N × N identity matrices multiplied by scalar. Hence, Proposi-
tion 8.3.2 can be applied to (Lsupra − μI L·N ) in order to diagonalize Lsupra . 

With this Proposition, we have a very powerful tool in order to investigate not only the
influence of the multiplex structure network on the stability of the synchronous state
but also the impact of different layer topologies which are not necessarily linearly
dependent. As an example, we consider a duplex systems with [L 1 , L 2 ] = 0 and
 1 
σ L + ρm 12 I −ρm 12 I
L supra
= .
−ρm 21 I σ L 2 + ρm 21 I

Knowing the eigenvalues for L 1 and L 2 , the master function parameter μ is deter-
mined using Eq. (8.12). The matrices L 1 and L 2 are Laplacian matrices which
have at least one zero eigenvalue, corresponding to the so-called Goldstone mode
1̂ = (1, . . . N − times · · · , 1)T . As a result there are two parameters μ = 0 and
 T
μ = ρ(m 12 + m 21 ). The first value corresponds to the Goldstone mode 1̂, 1̂ . The
second parameter is exclusively induced by the duplex structure and completely
independent from the individual layer topologies. Further, we find that in case of
m 12 = −m 21 another zero parameter μ exists which corresponds to the eigenmode
 T
1̂, −1̂ . The other eigenvalues of the supra-Laplacian matrix Lsupra are given by
the non-linear mappings

σ (λ1 + λ2 ) + ρ(m 12 + m 21 ) 1  2
μ(λ1 , λ2 ) = ± σ (λ1 − λ2 ) + ρ(m 12 − m 21 ) + 4ρ 2 m 12 m 21
2 2
(8.22)

for which we have formally solved Eq. (8.12) with respect to μ. Let us consider two
special cases.
First, we assume that there is no connection from the second to the first layer. We
have a master-slave set-up which means that m 12 = 0. With this, the master function
parameter is μ(λ1 , λ2 ) = σ λ1 and μ(λ1 , λ2 ) = σ λ2 + ρm 21 . Remarkably, in this
set-up the stability of a synchronous state in a duplex network is reduced to the pure
one-layer system. The stability in the duplex system is determined by the spectrum
of the individual layer topologies where only in the second layer the spectrum is
shifted due to the interaction.
 The second case starts from the consideration in [15]. In particular, we consider
Lintra , Linter = 0 which leads to a pairwise linear dependence of all individual
layer topologies, and hence λli = λi for all l = 1, . . . , L and i = 1, . . . , N with λli ∈
C. Taking this into account,
 the equation
 for the master function parameter yields
μ(λ1 , λ2 ) = σ λ1 + ρ m 12 + m 21 and μ(λ1 , λ2 ) = σ λ1 . In order to see that the
master function parameter agrees with the set for the parameter γ in (8.19), we
determine the eigenvalues of Lintra and Linter individually. Since L (1) and L (2) are
8.5 Applications for the Multiplex Decomposition 165

linearly dependent, the set of eigenvalues for Lintra consists of the eigenvalues of L 1
with double multiplicity. Using Proposition 8.3.5, we find that Linter has eigenvalues
0 and (m 12 + m 21 ) each with multiplicity N . Since both Laplacian matrices commute,
there exists a common set of eigenvectors. We find γ = ρλ(1) for the eigenvector
(1, 1)T ⊗ v and γ = ρλ(1) + σ (m 12 + m 21 ) for the eigenvector (m 12 , −m 21 )T ⊗ v
where L 1 v = λ1 v. Here, again everything boils down to a pure one-layer set up with
an additional shift.

8.5.2 Analytic Treatment of Diffusive Dynamics on Multiplex

Networks

In the previous section we considered the dynamics of linear systems as they are given
by the variational equation (8.18). In the context of diffusive systems on complex
networks, recently linear diffusive processes were considered in order to study the
dynamics on social as well as transport networks [18, 19]. Compared with equation
in [19], the authors investigate a duplex system (L = 2) with f = 0, σ = Dk (k =
1, 2), H, G = I, and ρm kl = Dx . Additionally, they considere weighted networks
for which instead of aikj ∈ {0, 1} coupling weights κikj (k = 1, 2) were taken. Note
that the former results on multiplex matrices still hold true for weighted connections.
Let us assume that the super-Laplacian matrix is given by
 
D1 L 1 + D x I −Dx I
L supra
= .
−Dx I D2 L 2 + D x I

with [L 1 , L 2 ] = 0. Due to the structure of the supra-Laplacian we are allowed

to apply Proposition 8.5.1. In accordance with [19] and our former findings in
Sect. 8.5.1, we have the two eigenvalues μ = 0 and μ = 2Dx corresponding to the
Goldstone mode (1̂, 1̂)T and the vector (1̂, −1̂)T , respectively. The other eigenvalues
are given as solution to the Eqs. (8.12) and read
!
D1 λi1 + D2 λi2 + 2Dx 1  2
μi = ± D1 λi1 − D2 λi2 + 4(Dx )2 ,
2 2

where λi1 and λi2 are the eigenvalues of the Laplacian matrices L 1 and L 2 , respectively.
With this we derived a complete analytic expression for the spectral properties of
the diffusive system which was considered in [19] without using any perturbation
method.
166 8 Multilayered Adaptive Networks

8.6 Summary

In summary, we have proposed a concept to induce diverse partial synchronization

patterns (phase clusters) in adaptively coupled phase oscillator networks. While adap-
tive networks have recently attracted a lot of attention in the fields of neuroscience
and social sciences, biology, engineering, and other disciplines, and multilayer net-
works are a paradigm for real-world complex networks, little has been known about
the interplay of multilayer structures and adaptivity. We have aimed to fill this gap
within a rigorous framework of theoretical analysis and computer simulations. We
have shown that multiplexing in a multi-layer with symmetry can induce various
stable phase cluster states like splay states, antipodal states, and double antipodal
states, in a situation where they are not stable or do not even exist in the single layer.
Further, we have developed a novel method for analysis of Laplacian matrices
of duplex networks which allows for insight into the spectral structure of these net-
works, and can easily be generalized to more than two layers. This new approach of
multiplex decomposition has a broad range of applications to physical, biological,
socio-economic, and technological systems, ranging from plasticity in neurodynam-
ics or the dynamics of linear diffusive systems [19, 20] to generalizations of the
master stability approach [15, 21] for adaptive networks. We have used the mul-
tiplex decomposition to provide analytic results for the stability of lifted states in
the multilayer system. As local dynamics we have used the paradigmatic Kuramoto
phase oscillator model, supplemented by adaptivity of the link strengths with a phase
lag parameter which can model a whole range of adaptivity rules from Hebbian via
spike-timing dependent plasticity to anti-Hebbian.

References

1. Gross T, D’Lima CJD, Blasius B (2006) Epidemic dynamics on an adaptive network. Phys Rev
Lett 96
2. Shai S, Dobson S (2013) Coupled adaptive complex networks. Phys Rev E 87
3. Wardil L, Hauert C (2014) Origin and structure of dynamic cooperative networks. Sci Rep 4
4. Klimek P, Diakonova M, Eguiluz VM, San Miguel M, Thurner S (2016) Dynamical origins of
the community structure of an online multi-layer society. New J Phys 18
5. Boccaletti S, Pisarchik AN, del Genio CI, Amann A (2018) Synchronization: from coupled
systems to complex networks. Cambridge University Press, Cambridge
6. Sakaguchi H, Kuramoto Y (1986) A soluble active rotater model showing phase transitions via
mutual entertainment. Prog Theor Phys 76:576
7. Kasatkin DV, Yanchuk S, Schöll E, Nekorkin VI (2017) Self-organized emergence of multi-
layer structure and chimera states in dynamical networks with adaptive couplings. Phys Rev E
96
8. Berner R, Sawicki J, Schöll E (2020) Birth and stabilization of phase clusters by multiplexing
of adaptive networks. Phys Rev Lett 124
9. Silvester JR (2000) Determinants of block matrices. Math Gaz 84:460
10. Sothanaphan N (2017) Determinants of block matrices with noncommuting blocks. Lin Alg
Applic 512:202
11. Boyd S, Vandenberghe L (2004) Convex optimization. Cambridge University Press, Cambridge
References 167

12. Liesen J, Mehrmann V (2015) Linear algebra. Springer, Cham

13. Kovacs I, Silver DS, Williams SG (1999) Determinants of commuting-block matrices. Am
Math Mon 106:950
14. Powell PD (2011) Calculating determinants of block matrices. arXiv:1112.4379
15. Tang L, Wu X, Lü J, Lu J, D’Souza RM (2019) Master stability functions for complete,
intralayer, and interlayer synchronization in multiplex networks of coupled rössler oscillators.
Phys Rev E 99
16. De Domenico M, Solé-Ribalta A, Cozzo E, Kivelä M, Moreno Y, Porter MA, Gómez S, Arenas
A (2013) Mathematical formulation of multilayer networks. Phys Rev X 3
17. Kivelä M, Arenas A, Barthélemy M, Gleeson JP, Moreno Y, Porter MA (2014) Multilayer
networks. J Complex Netw 2:203
18. Barthélemy M (2011) Spatial networks. Phys Rep 499:1
19. Gómez S, Díaz-Guilera A, Gómez-Gardeñes J, Pérez Vicente CJ, Moreno Y, Arenas A (2013)
Diffusion dynamics on multiplex networks. Phys Rev Lett 110:028701
20. Solé-Ribalta A, De Domenico M, Kouvaris NE, Díaz-Guilera A, Gómez S, Arenas A (2013)
Spectral properties of the Laplacian of multiplex networks. Phys Rev E 88
21. Pecora LM, Carroll TL (1998) Master stability functions for synchronized coupled systems.
Phys Rev Lett 80:2109
Chapter 9
Conclusion and Outlook

Part I of this thesis was devoted to the analysis of cluster states in populations of
globally coupled neurons with synaptic plasticity. Here, we used mathematical as
well as numerical tools in order to understand the mechanism behind the emergence
of frequency clustering.
In Chap. 3, we showed that adaptive neural networks are able to generate self-
consistently dynamics with different frequency bands. Here, each cluster corresponds
to a strongly connected component with a fixed frequency. Due to a sufficiently
large difference of the cluster sizes and frequencies, the inter-cluster interactions
are suppressed, while the intra-cluster interactions are enhanced. In this study, we
described the mechanisms behind the formation and stabilization of these clusters.
In particular, we explained why the significant difference between the cluster sizes
is important for the decoupling of the clusters. From a broader perspective, the
decoupling of the clusters in our case is analogous to the decoupling of timescales
in systems with multiple timescales.
In addition, we presented a two-dimensional phenomenological model which
allows for a detailed study of the clustering mechanisms. Despite of the approxima-
tions made by the derivation, the model coincides surprisingly well with the adaptive
Hodgkin–Huxley network. Using the phenomenological model, we found parameter
regions of the plasticity function, where stable frequency clustering can be observed.
Clustering behavior also emerges at the brain scale, where synchronized commu-
nities of brain regions constituting large distributed functional networks can inter-
mittently be formed and dissolved [1, 2]. Such clustering dynamics can shape the
structured spontaneous brain activity at rest as measured by fMRI. In this study,
we showed that slow oscillations based on the modulation of synchronized neural
activity can already be formed at the resolution level of a single neural population
if adaptive synapses are taken into account. These modulations of the amplitude of
the mean field can be generated in a stable manner, see Fig.3.9 and Ref. [3]. The
mechanism relies on fluctuations of the extent of synchronization of tonically firing
neurons that are neurons which fire periodically provided they receive a constant
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 169
R. Berner, Patterns of Synchrony in Complex Networks of Adaptively Coupled Oscillators,
Springer Theses, https://doi.org/10.1007/978-3-030-74938-5_9
170 9 Conclusion and Outlook

synaptic input. The fluctuations are caused by the splitting of the neural population
into clusters and the corresponding cluster dynamics. Our findings might contribute
to the emergence of slow brain rhythms of electrical (LFP, EEG) and metabolic
(BOLD) brain activity reported by [4–7].
However, other mechanisms for generating slow oscillations are possible. The
papers [8, 9] discussed the emergence of slow oscillatory activity (< 1 Hz) that can
be observed in vivo in the cortex during slow-wave sleep, under anesthesia or in vitro
in neural populations. The suggested mechanism relies on the corresponding modu-
lation of the firing of individual neurons, and the slow oscillation at the population
level was proposed to be the result of very slow bursting of individual neurons that
synchronize across the neural population. In contrast, the presented work shows that
the slow oscillations of the population mean field can also emerge when the firing of
individual neurons is not affected. The neurons may tonically fire at high frequencies.
The amplitude of the population mean field then oscillates at much lower frequencies
due to the slow modulation caused by the cluster dynamics.
Additionally, we would like to mention that the observed frequency clustering
resembles phenomenologically the weak chimera states [10, 11] where clusters with
different frequencies are formed in symmetrically coupled oscillators without adap-
tation. However the properties and mechanisms of the appearance of such clusters
are different from those presented here, which are essentially based on the slow
adaptation.
In Sect. 3.4 the phenomenological phase oscillator model was shown to be a very
powerful tool to predict the clustering dynamics of neuronal population with synaptic
plasticity. It is well-known that various models of weakly coupled oscillatory systems
can be reduced to coupled phase oscillators. For this, we implemented a simplified
phase oscillator model in Chaps. 4–5 which is able to describe the slow adaptive
change of the network depending on the oscillatory states. The slow adaptation is
controlled by a time-scale separation parameter.
In this framework, we developed the analytical approach of building blocks for
multicluster states. We showed that for the case of phase oscillators the multiclus-
ters are composed of certain one-cluster states (building blocks) that are themselves
phase-locked solutions of the adaptive model. Moreover, we provided rigorous con-
ditions under which multiclusters emerge and can be found stable. For the analysis
we divided the problem into two main step. Firstly, we analyzed the single one-cluster
states. Secondly, the multicluster states were analytically described and were shown
to inherit important properties from their building blocks.
In Chap. 4, we exhaustively described the one-cluster states in an adaptive network
of globally coupled phase oscillators. In particular, the following three types of one-
cluster (relative equilibria) are possible building blocks for multicluster states: splay,
antipodal, and double antipodal. In order to understand the stability of these states,
we performed a linear stability analysis for the relative equilibria. The stability of
these states was rigorously described, and the impact of all parameters was shown.
We showed that there exists an N − 2 dimensional family of splay type cluster states
which gives rise to N − 2 neutrally stable directions. Remarkably, this property of
splay states was also found in networks of pulse-coupled rotators [12] and excitable
9 Conclusion and Outlook 171

neurons [13]. We further proved that the double antipodal states are unstable in the
whole parameter range. They appear to be saddle-points in the phase space. While the
time-scale separation has no influence upon their existence, it plays an important role
for the stability the relative equilibria. The regions of stability in parameter space were
presented for different choices of the time-scale separation parameter. The singular
limit ( → 0) and the limit of instantaneous adaptation were analyzed. The latter
shows that the stability region of the splay and the antipodal states divide the whole
space into two equally sized regions without intersection. Instantaneous adaptation
cancels multistability of these states. The consideration of the singular limit showed
that it differs from the case of no adaptation. Therefore, even for very slow adaptation,
the oscillatory dynamics alone is not sufficient to describe the stability of the system.
Subsequently, the role of double antipodal states was discussed. We found that in a
system of 3 oscillators these states are transient states in a small heteroclinic network
between antipodal and splay states. They appear to be metastable, i.e. observable for a
relative long time and therefore are physically important transient states. Moreover an
additional analysis for an ensemble of 100 phase oscillators revealed the importance
of the double antipodal states for the global dynamics of the whole system.
In Chap. 5, we described the appearance of several different frequency cluster
states. Starting from random initial conditions, our numerical simulations showed
two different types of states. These are the splay and the antipodal type multicluster
states. A third mixed type multicluster state was found by using well prepared initial
condition. For all these states the collective motion of oscillators, the shape of the
network, and the interaction between the frequency clusters was presented in detail.
It turned out that the oscillators are able to form groups of strongly connected units.
The interaction between the groups is weak compared to the interaction within the
groups.
While the one-cluster are relative equilibria of the model due to the phase-shift
symmetry, the multicluster states contain components with different frequencies,
and, hence, they cannot be reduced to an equilibrium by transforming into another
co-rotating frame. As a result, the study of multiclusters is more involved. However,
to our surprise, we were still able to find an explicit form of multiclusters with the
components of the splay type. Remarkably, in addition to its ring-like spatial structure
that dynamically emerges, the network behaves in such a case (quasi-)periodically
in time such that the whole solution can be interpreted as a spatio-temporal wave.
Moreover, while the existence of the one-cluster states does not depend on the time-
scale separation parameter, the multicluster states crucially depend on the time-scale
separation. In fact, we provided an analysis showing that there exists a critical value
for the time-scale separation. Moreover, we showed that in the case of two-cluster
states of splay type the adaptation of the coupling weights must be at most half as
fast as the dynamics of the oscillators. For the splay clusters we analytically pro-
vided an explicit existence criterion. This fact is of crucial importance for comparing
dynamical scenarios induced by short-term or long-term plasticity [14].
In contrast to the splay type multicluster, the analysis of multicluster states of
antipodal type is more subtle due to the modulation of the frequency. More specif-
ically, we looked at multiclusters with bounded frequency modulation. For these
172 9 Conclusion and Outlook

types of multiclusters, we derived an asymptotic expansion in the parameter  that

gave explicit existence conditions. In addition, we showed analytically the existence
of mixed multiclusters, which consist of clusters of splay type and clusters of antipo-
dal type. For the mixed multiclusters, the temporal behavior within one cluster was
shown to be slightly non-identical, namely, the oscillators possess the same averaged
frequency, but they still can have a bounded quasi-periodically modulated phase dif-
ference.
The stability of two-cluster states was analyzed numerically and was presented
for different values of the time-scale separation parameter. By assuming weakly
interacting clusters, we described the stability of the two-cluster with the help of the
analysis of one-cluster states. The simulations showed that there are no stable two-
cluster states with clusters of the same size. We provided an argument to understand
this property of the system.
Moreover, the findings on multicluster solutions as they are reported in this work
are in very good agreement with previous results on adaptive neural networks [3,
15]. Here, stable multicluster states of coherently spiking neurons with weak but
time-dependent inter-cluster coupling were reported. With this work we shed some
light on these generic time-dependent network patterns.
In Part II, we extended the findings of the first part towards more complex con-
nectivity structures.
In Chap. 6, a model of adaptively coupled identical phase oscillators on a nonlo-
cally coupled ring was studied. Various frequency synchronized states were observed
including one-cluster, multicluster, and solitary states. Similar states had been found
also for networks with global base topolog [16]. To account for the complex base
topology, we introduced a new definition of one-cluster states by means of connected
induced subnetworks. This definition allows furthermore to distinguish between mul-
ticluster and solitary states in a more rigorous way than it had been done before.
Since one-cluster states form building blocks for multicluster states, see Part I of
this thesis, we first investigated the existence and stability properties of one-cluster
states. Here, we introduced a novel type of phase-locked states for complex networks,
namely local splay states, and showed that this class of states is nonempty for any
nonlocal ring base topology. In particular, we proved that rotating-wave as well as
antipodal states are always phase-locked solutions. Compared with the case of a
global base topology, the different clusters of local splay type on a nonlocal ring
structure can possess different collective frequencies. In addition, we proved that
local splay cluster are always global splay cluster. This statement relates, therefore,
local with global (with respect to “spatial” structures in the network) properties of
solutions.
The stability features of rotating-wave states were studied numerically and ana-
lytically. The comparison of both approaches resulted in a very good agreement. Due
to the analytic findings for rotating-wave states on a nonlocal ring, we were able to
describe their stability depending on the coupling range P and the wavenumber k.
The limiting case of global coupling, i.e. P = N /2, was shown to be in agreement
with the results presented in Chap. 4.
9 Conclusion and Outlook 173

An interesting feature of the system’s behavior are solitary states. They had been
previously found to emerge in the Kuramoto-Sakaguchi model with inertia [17, 18].
In this thesis, we showed that solitary states are born in a homoclinic bifurcation and
can be (de)stabilized in a pitchfork bifurcation of periodic orbits. In order to show this,
a two-dimensional effective model was derived governing the dynamics of solitary
states. In contrast to the Kuramoto-Sakaguchi model with inertia, we observed a
much more complicated bifurcation behavior. In particular, three different solitary
states are created due to two individual homoclinic bifurcations. Two of these three
solitary states, however, are unstable and bifurcate together with the stable solitary
states in a subcritical pitchfork bifurcation of periodic orbits.
Our results highlight the delicate interplay between adaptivity and complexity of
the network structure. Since this interplay has been rarely investigated from the math-
ematical viewpoint, so far, this work raises many questions for future research which
could be conducted for different network structures beyond nonlocal rings, other
dynamical models for the local dynamics, nonidentical units or different adaptation
rules.
In Chap. 7, we studied the stability of the synchronous state for a complex net-
work of adaptively coupled oscillators. For this, the well-known master stability
approach was generalized for networks with adaptive and hence time evolving cou-
pling weights. The master stability approach [19] is already well established and was
generalized for various applications, e.g., to understand the stability of cluster syn-
chronous states [20–22] or to study systems with single or distributed delays [23–30].
Beyond the local stability described by the master stability function, Belykh et al. had
developed the connection graph stability method to provide analytic bounds for the
global asymptotic stability of synchronized states [31–35]. Despite these advances,
the novel extension towards adaptive networks can be regarded as non-standard. So
far all generalizations were introduced for a static topological structure. However,
in the case of an adaptive network the topological structure is non-constant in time
which changes the theoretical approach entirely.
The master stability approach was further applied to the paradigmatic model of
adaptively coupled phase oscillators. We illustrated several forms for the master
stability function with respect to different adaptation rules, i.e., different values for
the control parameter β. Remarkably, the emergence of bounded regions that would
lead to stable synchronous dynamics was observed in the master stability function.
These bounded regions were called stability islands. We described the structure and
formation of stability islands via a cubic curve and derived an existence criterion.
Additionally, we analytically provided sufficient conditions for the system parame-
ters which imply the stability of the synchronous state on a undirected background
network structure.
Afterwards, the implications of the stability islands were further explored. Using
an adiabatic continuation of the synchronous state with respect to the coupling con-
stant σ , we obtained that the presence of a stability island gives rise to the emergence
of multicluster states and chimera-like states. The stable existence of certain multi-
clusters and chimera-like states were shown numerically and analytically explained
using the master stability function. Thus, the findings do not only complement the
174 9 Conclusion and Outlook

results obtained in Chaps. 5 and 6. They also underline the subtle interplay between
cluster size and network structure for the stability of complex synchronization pat-
terns [36, 37].
We showed how the generalized master stability approach is used to understand
the mechanisms behind the formation of frequency clusters. Another field of applica-
tion could be the analysis of synchronous states in real-world neuronal systems with
synaptic plasticity. Such systems have been discussed recently in order to study patho-
logical diseases such as tinnitus or Parkinson and their therapeutic treatment [38–
42]. However, the connectivity structure of the neuronal systems were predefined
and fixed in these studies. Our master stability approach for adaptive networks is not
restricted to a particular complex network. Thus, the new method could be used to
introduce a generalized approach in order to understand unwanted neural synchro-
nization in complex neuronal networks with synaptic plasticity.
In Chap. 8, we proposed a concept to induce diverse partial synchronization pat-
terns (phase clusters) in adaptively coupled multilayer phase oscillator networks.
While adaptive networks have recently attracted a lot of attention in the fields of
neuro- and social sciences, biology, engineering, and other disciplines, and mul-
tilayer networks are a paradigm for real-world complex networks, little had been
known about the interplay of multilayer structures and adaptivity. We aimed to fill
this gap within a rigorous framework of theoretical analysis and computer simula-
tions. We showed that multiplexing in a multilayer with symmetry can induce various
stable phase cluster states like splay states, antipodal states, and double antipodal
states, in a situation where they are not stable or do not even exist in the single
layer. Further, we developed a novel method for the analysis of Laplacian matrices
of duplex networks which allows for insight into the spectral structure of these net-
works, and can easily be generalized to more than two layers. This new approach of
multiplex decomposition has a broad range of applications to physical, biological,
socio-economic, and technological systems, ranging from plasticity in neurodynam-
ics or the dynamics of linear diffusive systems [43, 44] to generalizations of the
master stability approach [19, 45] for adaptive networks. We have used the multi-
plex decomposition to provide analytic results for the stability of lifted states in the
multilayer system. As local dynamics we used the paradigmatic Kuramoto phase
oscillator model, supplemented by adaptivity of the link strengths with a phase lag
parameter which can model a whole range of adaptivity rules from Hebbian via spike
timing-dependent plasticity to anti-Hebbian.
While several applications of the multiplex decomposition were presented in this
thesis, many other possible fields of application remain untouched. An intriguing
opportunity to apply our new methodology arises in the control of power grid net-
works. For these particular real-world networks, it is of great importance to retain
synchronization when single nodes or lines fail [46]. Multiplexing with a control
layer could be used to establish an efficient way to stabilize power grids against
topological changes [47]. A control layer approach might also be used to prevent the
power grids from the cascading of line failures due to overloads [48].
As we saw in this thesis, adaptive networks of phase oscillators give rise to a
plethora of dynamical scenarios. These scenarios include complete and cluster syn-
9 Conclusion and Outlook 175

chronization, frequency clustering, solitary states, and chimera-like states. The reason
behind this variety of states is the adaptivity of the coupling weights. This flexibility
in the coupling structure allows each individual oscillator to adapt indirectly its own
frequency and thus leads to the formation of different frequency clusters. Frequency
adaptation is also a property of Kuramoto phase oscillators with inertia that is a key
model in the theory of power grids. The introduction of inertia enables the oscillators
to adapt their frequencies directly. Remarkably, frequency clustered states are also
widely observed in networks of Kuramoto phase oscillators with inertia [17, 47, 49–
54]. Frequency adaptation is present but implemented differently in both systems.
Therefore, an open and very intriguing question arises as to whether it is possible to
relate the models of phase oscillators on adaptive networks to the models of phase
oscillators with inertia [55].

References

1. Deco G, Jirsa VK, McIntosh AR, Sporns O, Kötter R (2009) Key role of coupling, delay, and
noise in resting brain fluctuations. Proc Natl Acad Sci USA 106:10302
2. Ponce-Alvarez A, Deco G, Hagmann P, Romani GL, Mantini D, Corbetta M (2015) Resting-
state temporal synchronization networks emerge from connectivity topology and heterogeneity.
PLoS Comp Biol 11
3. Popovych OV, Xenakis MN, Tass PA (2015) The spacing principle for unlearning abnormal
neuronal synchrony. PLoS ONE 10
4. Mantini D, Perrucci MG, Del Gratta C, Romani GL, Corbetta M (2007) Electrophysiological
signatures of resting state networks in the human brain. Proc Natl Acad Sci USA 104:13170
5. Magri C, Schridde U, Murayama Y, Panzeri S (2012) The amplitude and timing of the bold
signal reflects the relationship between local field potential power at different frequencies. J
Neurosci 32:1395
6. Monto S, Palva S, Voipio J, Palva JM (2008) Very slow eeg fluctuations predict the dynamics
of stimulus detection and oscillation amplitudes in humans. J Neurosci 28:8268
7. Alvarado-Rojas C, Valderrama M, Fouad-Ahmed A, Feldwisch-Drentrup H, Ihle M, Teixeira
CA, Sales F, Schulze-Bonhage A, Adam C, Dourado A, Charpier S, Navarro V, Le Van Quyen M
(2014) Slow modulations of high-frequency activity (40–140 Hz) discriminate preictal changes
in human focal epilepsy. Sci Rep 4:4545
8. Bazhenov M, Timofeev I, Steriade M, Sejnowski TJ (2002) Model of thalamocortical slow-
wave sleep oscillations and transitions to activated states. J Neurosci 22:8691
9. Compte A, Sanchez-Vives MV, McCormick DA, Wang XJ (2003) Cellular and network mech-
anisms of slow oscillatory activity (<1 Hz) and wave propagations in a cortical network model.
J Neurophys 89:2707
10. Ashwin P, Burylko O (2015) Weak chimeras in minimal networks of coupled phase oscillators.
Chaos 25
11. Bick C, Ashwin P (2016) Chaotic weak chimeras and their persistence in coupled populations
of phase oscillators. Nonlinearity 29:1468
12. Calamai M, Politi A, Torcini A (2009) Stability of splay states in globally coupled rotators.
Phys Rev E 80
13. Dipoppa M, Krupa M, Torcini A, Gutkin BS (2012) Splay states in finite pulse-coupled networks
of excitable neurons. SIAM J Appl Dyn Syst 11:864
14. Fröhlich F (2016) Network neuroscience. Academic, Cambridge
15. Chakravartula S, Indic P, Sundaram B, Killingback T (2017) Emergence of local synchroniza-
tion in neuronal networks with adaptive couplings. PLoS ONE 12
176 9 Conclusion and Outlook

16. Kasatkin DV, Yanchuk S, Schöll E, Nekorkin VI (2017) Self-organized emergence of multi-
layer structure and chimera states in dynamical networks with adaptive couplings. Phys Rev E
96
17. Jaros P, Brezetsky S, Levchenko R, Dudkowski D, Kapitaniak T, Maistrenko Y (2018) Solitary
states for coupled oscillators with inertia. Chaos 28
18. Kruk N, Maistrenko Y, Koeppl H (2020) Solitary states in the mean-field limit. Chaos 30
19. Pecora LM, Carroll TL (1998) Master stability functions for synchronized coupled systems.
Phys Rev Lett 80:2109
20. Dahms T, Lehnert J, Schöll E (2012) Cluster and group synchronization in delay-coupled
networks. Phys Rev E 86
21. Pecora LM, Sorrentino F, Hagerstrom AM, Murphy TE, Roy R (2014) Symmetries, cluster
synchronization, and isolated desynchronization in complex networks. Nat Commun 5:4079
22. Sorrentino F, Pecora LM, Hagerstrom AM, Murphy TE, Roy R (2016) Complete character-
ization of the stability of cluster synchronization in complex dynamical networks. Sci Adv
2
23. Sorrentino F, Ott E (2007) Network synchronization of groups. Phys Rev E 76
24. Choe CU, Dahms T, Hövel P, Schöll E (2010) Controlling synchrony by delay coupling in
networks: from in-phase to splay and cluster states. Phys Rev E 81:025205(R)
25. Flunkert V, Yanchuk S, Dahms T, Schöll E (2010) Synchronizing distant nodes: a universal
classification of networks. Phys Rev Lett 105
26. Dahms T (2011) Synchronization in delay-coupled laser networks. Ph.D. thesis, Technische
Universität Berlin
27. Heiligenthal S, Dahms T, Yanchuk S, Jüngling T, Flunkert V, Kanter I, Schöll E, Kinzel W
(2011) Strong and weak chaos in nonlinear networks with time-delayed couplings. Phys Rev
Lett 107
28. Kyrychko YN, Blyuss KB, Schöll E (2014) Synchronization of networks of oscillators with
distributed-delay coupling. Chaos 24
29. Wille C, Lehnert J, Schöll E (2014) Synchronization-desynchronization transitions in complex
networks: an interplay of distributed time delay and inhibitory nodes. Phys Rev E 90
30. Lehnert J (2016) Controlling synchronization patterns in complex networks, Springer Theses.
Springer, Heidelberg
31. Belykh VN, Belykh IV, Hasler M (2004) Connection graph stability method for synchronized
coupled chaotic systems. Phys D 195:159
32. Belykh IV, de Lange E, Hasler M (2005) Synchronization of bursting neurons: what matters
in the network topology. Phys Rev Lett 94
33. Belykh IV, Belykh VN, Hasler M (2006) Generalized connection graph method for synchro-
nization in asymmetrical networks. Phys D 224:42
34. Belykh IV, Belykh VN, Hasler M (2006) Synchronization in asymmetrically coupled networks
with node balance. Chaos 16
35. Daley K, Zhao K, Belykh IV (2020) Synchronizability of directed networks: the power of
non-existent ties. Chaos 30
36. Berner R, Vock S, Schöll E, Yanchuk S (2021) Desynchronization transitions in adaptive
networks. Phys Rev Lett 126:028301
37. Vock S, Berner R, Yanchuk S, Schöll E (2021) Effect of diluted connectivities on cluster
synchronization of adaptively coupled oscillator networks. arXiv:2101.05601
38. Tass PA, Popovych OV (2012) Unlearning tinnitus-related cerebral synchrony with acoustic
coordinated reset stimulation: theoretical concept and modelling. Biol Cybern 106:27
39. Adamchic I, Hauptmann C, Barnikol UB, Pawelczyk N, Popovych OV, Barnikol TT, Silchenko
AN, Volkmann J, Deuschl G, Meissner WG, Maarouf M, Sturm V, Freund HJ, Tass PA (2014)
Coordinated reset neuromodulation for parkinson’s disease: proof-of-concept study. Movement
Disord 29:1679
40. Kromer JA, Tass PA (2020) Long-lasting desynchronization by decoupling stimulation. Phys
Rev Research 2
References 177

41. Kromer JA, Khaledi-Nasab A, Tass PA (2020) Impact of number of stimulation sites on long-
lasting desynchronization effects of coordinated reset stimulation. Chaos 30
42. Khaledi-Nasab A, Kromer JA, Tass PA (2021) Long-lasting desynchronization of plastic neural
networks by random reset stimulation. Front Phys 11
43. Gómez S, Díaz-Guilera A, Gómez-Gardeñes J, Pérez Vicente CJ, Moreno Y, Arenas A (2013)
Diffusion dynamics on multiplex networks. Phys Rev Lett 110:028701
44. Solé-Ribalta A, De Domenico M, Kouvaris NE, Díaz-Guilera A, Gómez S, Arenas A (2013)
Spectral properties of the Laplacian of multiplex networks. Phys Rev E 88
45. Tang L, Wu X, Lü J, Lu J, D’Souza RM (2019) Master stability functions for complete,
intralayer, and interlayer synchronization in multiplex networks of coupled rössler oscillators.
Phys Rev E 99
46. Witthaut D, Rohden M, Zhang X, Hallerberg S, Timme M (2016) Critical links and nonlocal
rerouting in complex supply networks. Phys Rev Lett 116
47. Totz CH, Olmi S, Schöll E (2020) Control of synchronization in two-layer power grids. Phys
Rev E 102
48. Schäfer B, Witthaut D, Timme M, Latora V (2018) Dynamically induced cascading failures in
power grids. Nat Commun 9:1975
49. Olmi S (2015) Chimera states in coupled Kuramoto oscillators with inertia. Chaos 25
50. Mehrmann V, Morandin R, Olmi S, Schöll E (2018) Qualitative stability and synchronicity
analysis of power network models in port-Hamiltonian form. Chaos 28
51. Tumash L, Olmi S, Schöll E (2018) Effect of disorder and noise in shaping the dynamics of
power grids. Europhys Lett 123:20001
52. Taher H, Olmi S, Schöll E (2019) Enhancing power grid synchronization and stability through
time delayed feedback control. Phys Rev E 100
53. Tumash L, Zakharova A, Panteley E, Schöll E (2019) Synchronization patterns in Stuart-Landau
networks: a reduced system approach. Eur Phys J B 92:100
54. Hellmann F, Schultz P, Jaros P, Levchenko R, Kapitaniak T, Kurths J, Maistrenko Y (2020)
Network-induced multistability through lossy coupling and exotic solitary states. Nat Commun
11:592
55. Berner R, Yanchuk S, Schöll E (2020) What adaptive neuronal networks teach us about power
grids. Phys Rev E 103:042315
Appendix A
Proofs of Results from the Main Text and
Supplemental Material

A.1 One-Cluster States on Globally Coupled Adaptive

Networks

Here we provide a proof of Proposition 4.1.1. We first need a preliminary lemma.

Lemma A.1.1 For a phase-locked solution φ(t), R2 (φ(t)) = 1 for all t if and only
if φ(t) is either an in-phase or an anti-phase synchronous solution.
 
 
Proof As follows from Table 4.1, R2 (a) = N1  Nj=1 ei2a j  = 1 for all in-phase and
anti-phase solutions.
 Let us show
 the opposite. If R2 (a) = 1, then ei2a j = ei2a1 =
 N i2a j   N  i2a j 
1 for all j, since  j=1 e  ≤ j=1 e = N . Hence, a j ∈ {0, π }. The latter
means that the phase-locked solution is either in-phase, if all a j have the same values,
or anti-phase otherwise. 
Now we present the proof of Proposition 4.1.1.
Proof Substituting (4.3)–(4.4) into (2.25)–(2.26) we obtain κ̇i j = 0 and

1 
N
φ̇i (t) =  = sin(ai − a j + β) sin(ai − a j + α)
N j=1
1 1  
= cos(α − β) − Re e−i(2ai +α+β) Z 2 (a) . (A.1)
2 2
Therefore (4.3)–(4.4) are solutions if and only if the expression on the right-hand side
of the Eq. (A.1) is independent of the oscillators index i = 1, . . . , N . In particular, for
any choice of ai the complex second order parameter is either zero or can be written
as Z 2 (a) = R2 (a)e−iγ . Thus, according to (A.1) ai has to be such that R2 (a) = 0 or
cos(2ai + α + β + γ ) is independent of i. For any α, β and γ the latter requirement
is equivalent with 2ai ∈ {0, −2(α + β + γ )}. Here, we made use of the phase-shift
symmetry of (2.25)–(2.26) by setting 2a1 = 0. Due to the definition of the complex
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 179
R. Berner, Patterns of Synchrony in Complex Networks of Adaptively Coupled Oscillators,
Springer Theses, https://doi.org/10.1007/978-3-030-74938-5
180 Appendix A: Proof of Results from the Main Text and Supplemental Material

order parameter the value for γ depends on the choice of the phase lags ai . Assuming
that one fraction q1 = Q 1 /N of the oscillators have 2ai = 0 with Q 1 ∈ {1, . . . , N −
1} and the other fraction of oscillators q2 = 1 − q1 have 2ai = −2(α + β + γ ) one
obtains

q1 + q2 e−i2(α+β) e−i2γ = R2 (a)e−iγ .

which is equivalent to the equations

q1 cos(γ ) + q2 cos(γ + 2ϑ) = R2 (a),

q1 sin(γ ) − q2 sin(γ + 2ϑ) = 0. (A.2)

with ϑ = α + β. Here, the first equation gives the value for the second order param-
eter while the second equation determines γ . A special solution can be given if we
set q2 = 0, equivalently q1 = 1 . Then γ = 0 and γ = π would solve the equation
above and thus 2ai = 0 for all i = 1, . . . , N . Note that, for both values of γ the
value for the complex second order parameter coincide Z 2 (a) = 1. This solution
corresponds to R2 (a) = 1. In any other case the last equation can be written in the
form sin(γ − ν) = 0 which has two solutions γ = ν, ν + π. Both solution would
coincide while determining 2ai . Writing (A.2) as

1
((q1 − (1 − q1 ) cos(2ϑ)) sin(γ ) − (1 − q1 ) sin(2ϑ) cos(γ )) = 0
C
where the normalization constant C is defined as

C = (q1 − (1 − q1 ) cos(2ϑ))2 + (1 − q1 )2 sin2 (2ϑ),

yields the equations

sin(2ϑ)
sin(ν) =  , (A.3)
2
q1
1−q1
+1− q1
2 1−q 1
cos(2ϑ)
q1 − (1 − q1 ) cos(2ϑ)
cos(ν) =  .
(1 − q1 )2 + q12 − 2q1 (1 − q1 ) cos(2ϑ)

Therefore, considering the inverse function arcsin : [−1, 1] → [−π/2, π/2] applied
to (A.3) determines ν to be either ν  or π − ν  , where ν  := arcsin(sin(ν)) and sin(ν)
as given in (A.3). The second equation for cos(ν) then fixes ν to take one of the
values. Thus, γ exists and is unique for every q1 ∈ [0, 1). The Proposition is proved
by taking into account that for a finite number N of oscillators q1 takes values in the
range from 1/N to (N − 1)/N and defining γ := −ψ − ϑ. 
Appendix A: Proof of Results from the Main Text and Supplemental Material 181

A.2 Stability of One-Cluster States on Globally Coupled

Networks

In order to study the local stability of one-cluster solutions described in Sect. 4.1,
we linearise the system of differential equations (2.25)–(2.26) around the phase
locked states described by φi = t + ai and κi j = − sin(ai − a j + β). We obtain
the following linearised system

1  1 
N N
d    
δφi = sin(β − α) δφi − δφ j + cos(2(ai − a j ) + α + β) δφi − δφ j
dt 2N 2N
j=1 j=1

1 
N
− sin(ai − a j + α)δκi j ,
N
j=1
(A.4)

and
d   
δκi j = − δκi j + cos(ai − a j + β) δφi − δφ j . (A.5)
dt
Note that this set of equations can be brought into the following block form

d δφ A B δφ
= (A.6)
dt δκ C − IN 2 δκ

where (δφ)T = (δφ ⎛1 , . .⎞

. , δφ N ), (δκ)T = (δκ11 , . . . , δκ1N , δκ21 , . . . , δκ N N ), B =
C1
  ⎜ .. ⎟
B1 · · · B N , C = ⎝ . ⎠, and A, Bn , Cn are N × N matrices with n = 1, . . . , N .
CN
The elements of the block matrices read
⎧
⎪ 1 1
⎪ − sin(α − β) − sin(β) cos(α)
⎪
⎪
⎪
⎨ 2 N N
i= j
1 
ai j = + sin(2(ai − ak ) + α + β),
⎪
⎪
⎪
⎪
2N k=1
⎪
⎩ 1  
2N
sin(α − β) − sin(2(ai − a j ) + α + β) , i = j

− N1 sin(an − a j + α), i = n
bi j;n =
0, otherwise
182 Appendix A: Proof of Results from the Main Text and Supplemental Material
⎧
⎪
⎪ 0, j = n, i = j
⎪
⎨− cos(a − a + β),
n i j = n, i = j
ci j;n = .
⎪
⎪ cos(a − a + β), j = n, i = j
⎪
⎩
n i
0, otherwise

Throughout this section we will make use of the Schur complement [1, 2] in order
to simplify characteristic equations as it has been introduced Eq. (4.11) in Chap. 4.
A B
In particular, the determinant of the p × p matrix M = can be simplified to
C D
the following expression by using (4.11)

det(M) = det(A − B D −1 C) · det(D).

This result is important for the subsequent stability analysis. So far, we have found the
Lyapunov coefficients for the rotating-wave states, see Corollary 4.2.3. The following
two Lemmata are needed to describe the stability of all antipodal, 4-phase cluster,
and double antipodal states as well.
Lemma A.2.1 Suppose M is a block square matrix of the form

A m 1 1̂ p,q
M=
m 2 1̂q, p B

where A is a circulant p × p matrix, B is a circulant q × q matrix, 1̂ p,q is p × q

where all entries are 1 and m 1 , m 2 ∈ R. Then the eigenvector-eigenvalue pairs are
given by

p−1
T 
p−1
λ0k , . . . , λk , 0, . . . , 0 , μk = a1(1+l) λlk (A.7)
l=0

q−1
T 
q−1
0, . . . , 0, ρl0 , . . . , ρl , νl = b1(1+l) ρll (A.8)
l=0

(1, . . . , 1, a1 , . . . a1 )T , μ̄ = μ0 + m 1 qa1 (A.9)

(1, . . . , 1, a2 , . . . , a2 ) , ν̄ = μ0 + m 1 qa2
T
(A.10)
2π 2π
with λk = ei p k and ρl = ei q l for k = 1, . . . , p − 1 and l = 1, . . . , q − 1, respec-
tively, and a1 and a2 solve the equation

μ0 − ν0 m2 p
a2 + a− =0
m 1q m 1q
p q
with μ0 = j=1 a1 j and ν0 = j=1 b1 j .
Appendix A: Proof of Results from the Main Text and Supplemental Material 183

Proof We can prove the Lemma by direct calculation and find

p−1
T T  p−1
M λ0k , . . . , λk , 0, . . . , 0 = A λ0k , . . . , λk
p−1
m2 l=0 λlk
T
p−1
= μk λ0k , . . . , λk , 0, . . . , 0 .

 p−1
Here, we use that A is a circulant matrix and that l=0 λlk = 0 for all k = 1, . . . , p −
1. Analogous arguments hold for (A.8). The last two eigenvector-eigenvalue pairs
(A.9)–(A.10) can be obtained by

A (1, . . . , 1)T + m 1 qa (1, . . . , 1)T

M (1, . . . , 1, a, . . . , a)T = m2 p
a (1,
. . . , 1)T + a B (1, . . . , 1)T
μ0 + m 1 qa
= m2 p (1, . . . , 1, a, . . . , a)T
a
+ ν0

which solves the eigenvalue problem if a is chosen to be either a1 or a2 . 

Lemma A.2.2 Suppose we have a phase locked state with phases ai ∈ [0, 2π ). Then,
the solution for the characteristic equations corresponding to the linearised sys-
tem (4.9)–(4.10) are given by λ = − with multiplicity N 2 − N and by the solution
of the following set of equations

det ((A − λI N ) ( + λ) + BC) = 0.

Proof Applying Schur’s decomposition (4.11) to the linearised system in the block
form (A.6) yields the result. 

Proposition A.2.3 Suppose we have a state with phases ai ∈ {0, π, ψ, ψ + π }

where i = 1, . . . , N . Further set q1 = Q 1 /N and q2 = Q 2 /N , where Q 1 and Q 2
denote the numbers of phases which are either 0 or π and ψ or ψ + π , respectively.
Then, the linear system (4.9)–(4.10) possesses the following set L of eigenvalues
 
L = (0)1 , (− )(N −1)N +1 , (λ1 ) N1 −1 , (λ2 ) N1 −1 , (ϑ1 ) N2 −1 , (ϑ2 ) N2 −1 , (ρ1 )1 , (ρ2 )1

where λ1 and λ2 solve

1
λ2 + (sin(α − β) − q1 sin(α + β) − q2 sin(−2ψ + α + β) + 2 ) λ
2
− q1 sin(α + β) − q2 sin(−2ψ + α + β) = 0,

ϑ1 and ϑ2 solve
1
ϑ2 + (sin(α − β) − q1 sin(2ψ + α + β) − q2 sin(α + β) + 2 ) ϑ
2
− q1 sin(2ψ + α + β) − q2 sin(α + β) = 0,
184 Appendix A: Proof of Results from the Main Text and Supplemental Material

as well as ρ1 and ρ2 solve

1
ρ2 + (sin(α − β) − q1 sin(2ψ + α + β) − q2 sin(−2ψ + α + β) + 2 ) ρ
2
− q1 sin(2ψ + α + β) − q2 sin(−2ψ + α + β) = 0.

The multiplicities for each eigenvalue are given as subscripts.

Proof For an arbitrary solution of the form φi = t + ai we consider the linearized

system (4.9)–(4.10) in the block form (A.6) and apply Lemma A.2.2. The elements
of the second term D := BC of Schur’s complement are then
 
di j = − sin(α − β) + sin(2(ai − a j ) + α + β)
2N
if i = j and
⎛ ⎞
1 1 1 N
dii = ⎝ sin(α − β) − sin(α) cos(β) + sin(2(ai − a j ) + α + β)⎠ .
2 N 2N j=1

Defining the matrix M := (A − λI N ) ( + λ) + D we get


−λ2 + (aii − )λ + aii + dii i = j
mi j =
λai j + ai j + di j . i = j

Using the assumption for the phases ai , then one group of oscillators (group I )
have ai ∈ {0, π } and the remaining Q 2 oscillators (group I I ) have ai = {ψ, ψ + π }.
Putting this into the definition of m i j , we find that the whole square matrix M can
be written as
⎛ Q 1 ×Q 1
⎞
  
⎜ m m̄ · · · m̄ ⎟
⎜ I ⎟
⎜ . ⎟
⎜ m̄ . . . . . . .. ⎟
⎜ ⎟
⎜ . . . m 1 1̂ Q 1 ,Q 2 ⎟
⎜ . .. .. ⎟
⎜ . m̄ ⎟
⎜ ⎟
⎜ m̄ · · · m̄ m I ⎟
M =⎜ ⎟
⎜ m I I m̄ · · · m̄ ⎟
⎜ . . . ⎟
⎜ ⎟
⎜ m̄ . . . . .. ⎟
⎜ m 2 1̂ Q 2 ,Q 1 .. . . . . ⎟
⎜ ⎟
⎜ . . . m̄ ⎟
⎜ ⎟
⎝ m̄ · · · m̄ m I I ⎠
  
Q 2 ×Q 2
Appendix A: Proof of Results from the Main Text and Supplemental Material 185

where m 1 , m̄, m I , and m I I are real values which depend on all the system parameters
α, β, and additionally on ψ and λ. Note that all diagonal blocks are circulant
matrices. The determinant is invariant under basis transformations which is why we
diagonalize the matrix M and therewith derive equations for the values λ. In order
to do so, we look for the eigenvalues of M determined by the characteristic equation

det (M − μI N ) = 0.

Due to the structure of M we can apply Lemma A.2.1 and find the following set of
eigenvalues

1
μk = −λ2 − (sin(α − β) − q1 sin(α + β) − q2 sin(−2ψ + α+β) + 2 )λ
2
+ q1 sin(α + β) + q2 sin(−2ψ + α + β)

for k = 1, . . . , Q 1 − 1. Analogously, we obtain the equations for νk (k = 1, . . . ,

Q 2 − 1) where m I is substituted with m I I . The two other eigenvalue are given by
μ̄ = μ0 + m 1 Q 2 a1 and ν̄ = μ0 + m 1 Q 2 a2 , respectively, where

μ0 = m I + (Q 1 − 1)m̄

and a1,2 are given by

(m I − m I I ) + (Q 1 − Q 2 ) m̄ m2 Q1
a2 + a− = 0.
m1 Q2 m1 Q2

Considering the row sums of M we find thatall agree with −λ2 − λ and therefore
μ̄ = −λ2 − λ. Resulting from this a1 = − λ2 + λ + μ0 /m 1 Q 2 = 1. Hence,

(m I I − m I ) + (Q 2 − Q 1 ) m̄
a2 = −1
m1 Q2

and we find

ν̄ = m I I + (Q 2 − 1) m̄ − m 1 Q 2
1
= −λ2 − (sin(α − β) − q1 sin(2ψ + α + β) − q2 sin(−2ψ + α + β) + 2 )λ
2
+ q1 sin(2ψ + α + β) + q2 sin(−2ψ + α + β)

After diagonalizing the matrix M the determinant can be easily written as

det(M) = μ̄ · μ1 · . . . · μ N1 −1 · ν̄ · ν1 · . . . · ν N2 −1 .
186 Appendix A: Proof of Results from the Main Text and Supplemental Material

Therewith, finding λ s such that at least one of the eigenvalues of M vanishes solves
the initial eigenvalue problem. 

We will now sum up the results with the following corollaries

Corollary A.2.4 The set of eigenvalues of the linearised system (4.9)–(4.10) around
all antipodal states with ai ∈ {0, π } agrees with the set L in Proposition 4.2.2 for
rotating-wave states with k = 0, N /2.

Proof Put Q 2 = 0 in Proposition A.2.3, then there is only the equation for λ left. 

Corollary A.2.5 The set of eigenvalues to the linearised system (4.9)–(4.10) around
all 4-phase-cluster states with ai ∈ {0, π/2, π, 3π/2} and R2 (a) = 0 agrees with the
set L in Proposition 4.2.2 for 4-rotating-wave states.

Proof The requirement R2 (a) = 0 yields Q 1 = Q 2 . The statement of this proposi-

tion follows by using Proposition A.2.3. 

Corollary A.2.6 For all α and β the double antipodal states are unstable.

Proof Suppose the polynomial equation p(x) = x 2 + ax + b = 0. This equation

has two negative roots if and only if b > 0 and a > 0 meaning that p(0) > 0 and
the vertex of the parabola is at x < 0, respectively. In order to have stable double
antipodal states these two conditions have to be met by all three equations for λ, ϑ
and ρ in Proposition A.2.3. From the condition on the existence of double antipodal
states (4.5) we find q1 sin(2ψ + α + β) + q2 sin(−2ψ + α + β) = − sin(α + β).
With this assumption on the quadratic equation and the latter equation, we find the
following two necessary conditions for the stability of double antipodal states, (1)
q1 sin(2ψ + α + β) + q2 sin(α + β) > 0 and (2) q1 sin(2ψ + α + β) + q2 sin(α +
β) < 0. The two condition cannot be equally fulfilled. 

A.3 Multicluster States of Splay Type

Proof of Proposition 5.2.1 1 We prove by direct substitution. Plugging (5.4) and

(5.5) into (2.26) the identity is obtained. Further, substituting (5.4) and (5.5)
into (2.26) we obtain

Nμ
1  
M
 
μ = ρμν cos(α − β + ψμν ) − cos(2(μν t + ai,μ − a j,ν ) + α + β − ψμν )
2N
ν=1 j=1


M
nν 1
= ρμν cos(α − β + ψμν ) − Re e−i(2μν t+2ai,μ +α+β−ψμν ) Z 2 (aν ) .
2 2
ν=1

If Z 2 (aμ ) = 0 for all μ = 1, . . . , M then

Appendix A: Proof of Results from the Main Text and Supplemental Material 187

1 
M Nμ
μ = ρμν cos(α − β + ψμν )
N ν=1 j=1

which agrees with the system (5.6) for the frequencies μ . On the contrary, assume
that the multicluster phase-locked
M solution
 (5.4) and (5.5) solve the Eq.
 (2.25). Analog
to Proposition 4.1.1 ν=1 ρμν Re e−i(2μν t+2ai,μ +α+β−ψμν ) Z 2 (aν ) has to be inde-
pendent of the oscillator index i = 1, . . . , Nμ and t ∈ R for all μ = 1, . . . , M. Take
any μ = 1, . . . , M and suppose Z 2 (aν ) = 0 for ν ∈ A with A ⊆ {1, . . . , M}. Then


M
 
ρμν Re e−i(2μν t+2ai,μ +α+β−ψμν ) Z 2 (aν ) =
ν=1 (A.11)
  
ρμν R2 (aν ) cos(2μν t + 2ai,μ + α + β − ψμν + γν )
ν∈A

where γν is defined as in Proposition 4.1.1, see type 2 or 3. For fixed μ all frequency
differences μν differ due to the assumption that the frequencies μ are all pairwise
different. This is why only terms with μν and μν  = −μν (ν, ν  ∈ A) are
candidates to compensate each other in the right hand side of (A.11) to give a
constant value for μ . Therefore, the number of clusters with Z 2 (aν ) = 0 excluding
the μth cluster which is under consideration has to be even, i.e., |A \ {μ}| even for
all μ = 1, . . . , M. This already yields that |A| odd and A = {1, . . . , M}. Consider
now μ such that μ = minν∈1,...,M ν . Then for every other ν ∈ {1, . . . , M} with
μν < 0 there has to be ν  ∈ {1, . . . , M} so that −μν = μν  = μ − ν  .
Hence, ν  < μ which contradicts that μ = minν∈1,...,M ν . Therefore, for this
choice of μ ∈ 1, . . . , M the expression in (A.11) cannot be constant contradicting
the assumption made in the beginning.

A.4 Asymptotic Expansions of Multicluster States

In this section we give an analytic description of multicluster solutions in terms of

an asymptotic expansion. We consider therefore the expansion of r th order together
with a multi-time scale ansatz [3] as given in 5.11

(r )

r
φi,μ ( , t) := (r )
μ (τ0 , . . . , τr ) + ai,μ +
l
pi,μ;l (t)
l=1 μ, ν = 1, . . . , M

r
i, j = 1, . . . , Nμ
κi(rj,μν
)
( , t) := l
ki j,μν;l (t)
l=0
188 Appendix A: Proof of Results from the Main Text and Supplemental Material

where (r )
μ ∈ C (R
1 r +1
) is a function depending on the multi-time scales τl := l t.
We show under under which conditions this expansion describes the time evolution
for the system (2.25)–(2.26).
The section is organized as follows. We first introduce some notations and prove
some technical lemmas that will help us to prove the main result in Proposition 5.3.1.
For ease of notation, for the remainder of the section indices are used as follows.
Small Latin letters i, j in the subscript are oscillator indices while Greek letters μ,
ν represent cluster indices. These are separated by a comma. Two further indices,
separated by semicolon, are the coefficient index and the mode index, respectively.
The indices in superscript are either powers or the order for the expansion which are
then written in parenthesis.
The following definition is introduced to handle the order of approximation.
Definition A.4.1 Let f : R × R → R and g : R × R → R real functions. We define
the following notations:
1. f ( , t) ∈ O(g( , t)) as → 0 on the interval I ⊆ R if for any t ∈ I there exist
C(t) > 0 and 0 (t) > 0 such that | f ( , t)| < C(t)|g( , t)| for all < 0 (t),
2. f ( , t) ∈ o(g( , t)) as → 0 on the interval I ⊆ R if for any t ∈ I and all C > 0
there exist 0 (t) > 0 such that | f ( , t)| < C|g( , t)| for all < 0 (t).
Remark A.4.1 If the constants C and 0 can be chosen independently of t ∈ I we
say that f ( , t) ∈ O(g( , t)) (or f ( , t) ∈ o(g( , t))) as → 0 uniformly on I .
In order to find an expressions for the asymptotic expansion of the coupling weights κ,
we use the concept of the pullback attractor. It is defined as an nonempty, compact and
invariant set and well known from the theory of nonautonomous dynamical systems.
For our purposes, suppose we know the functions φi (t) for all i = 1, . . . , N . Then,
the differential equations (2.26) is nonautonomous and can be solved explicitly by
 t
(t−t  )
κi j (t) := κi j,0 e− (t−t0 )
− e− sin(φi (t  ) − φ j (t  ) + β) dt  (A.12)
t0

with κi j,0 ∈ [−1, 1] for all i, j = 1, . . . , N . For this, the pullback attractor A is given
by the set

A := {(κ(t))∗ } (A.13)
t∈R

where
 t
(t−t  )
(κi j (t))∗ := − lim e− sin(φi (t  ) − φ j (t  ) + β) dt  (A.14)
t0 →−∞ t0

for all i, j ∈ {1, . . . , N }. Note that (A.14) is the unique bounded solution of (2.26),
see Theorem IV.1.1 of [4]. We remark the following important properties. For given
functions φi the Eqs. (2.26) possess the compact absorbing set G := {κi j : κi j ∈
Appendix A: Proof of Results from the Main Text and Supplemental Material 189

[−1, 1], i, j = 1, . . . , N }. Hence, the pullback attractor exists, c.f. Theorem 3.18
in [5], and is unique due to Proposition 3.8 [5]. Moreover, (κ(t))∗ is a solution for
the nonautonomous system which can be shown by direct computation. We call
κ(t))∗ the parametrization of the pullback attractor and use it in order to find an
analytic expression for the (pseudo-)multicluster states. Note that we have already
seen such parametrizations explicitly in (4.4) and (5.5). For more details regarding
nonautonomous systems and the pullback attractor we refer the reader to [5, 6].
We use the following notations for the sake of brevity.

M := {m = (m 1 , . . . , m M ) : m 1 , . . . , m M ∈ Z} ,
cm := cm 1 ,...,m M ,

M
(m) := m μ μ .
μ=1

Furthermore, we say that two elements m, n ∈ M are equivalent m ∼ n if and only if

(m) = (n). The corresponding quotient space is denoted by M̃ := M/∼ . If
μ is considered as frequencies the equivalence relation factors out all resonant linear
combinations of those. Let us further define M̃( f ) as the set of all (m 1 , . . . , m M )
such that the function f can be written as f = m∈M( f ) cm ei(m)t for some cm ∈
C. Finally, we introduce the shorthand notion (mμnν) := (0, . . . , 0, m μ , 0, . . . , 0,
m ν , 0, . . . , 0) with m μ = m and m ν = n for further convenience if only frequencies
of two distinguished clusters are considered.
The proof of Proposition 5.3.1 makes use of several lemmas and is presented at
the end of this section. Overall, we aim to describe the following particular form for
the dynamical behavior of the phase oscillators. The phases of the oscillators φi,μ
form a pseudo multicluster, see Definition 5.5.1. Further, the bounded modulations
for the phases of each oscillator are given as Taylor expansions in with periodic
coefficients that can be expressed as Fourier sums with even modes.
To determining the asymptotic expansion explicitly, derivatives of composed func-
tion have to be carried out. The following Lemma provides us with a general form.

Lemma A.4.1 Suppose we have n-times differentiable real functions f and g. Let

Tn := {(k1 , . . . , kn ) : 1k1 + 2k2 + · · · + nkn = n, k1 , . . . , kn ∈ N0 }

denote the partitions of n. The composition ( f ◦ g) is also n-times differentiable and

the nth derivative can be written as

 n!  k1 +...+kn n
Dm g km
Dxn ( f ◦ g)(x0 ) = Dx f ) ◦ g(x0 ) (x0 ).
(k1 , ... ,kn )∈Tn
k1 ! · · · · · kn ! m=1
m!
(A.15)
190 Appendix A: Proof of Results from the Main Text and Supplemental Material

Proof See [7, pp. 95–96]. 

This expression for the nth-derivative is also known as the Faà di Bruno formula.
Lemma A.4.2 Suppose the phase oscillators behave as


r
μ = 1, . . . , M
φi,μ ( , t) := μ t + ai,μ + l
pi,μ;l (t),
l=1
i = 1, . . . , Nμ

with μ ∈ R, ai,μ ∈ T1 and there are m i,μ;l ∈ N0 such that the bounded functions
pi,μ;l : R → R are given by the finite multi-Fourier series

pi,μ;l (t) = ci,μ;l;m ei(m)t
m∈M̃( pi,μ;l )

with |M̃( pi,μ;l )| = m i,μ;l .

Then, for s ≤ r there exist  functions
 ki j,μν;l (t) such that the asymptotic expansion
of the pull-back attractor κi j,μν ∗ defined in (A.14) can be written as

μ, ν = 1, . . . , M
  
s
κi j,μν ∗ = l
ki j,μν;l (t) + R̂i j,μν ( , t) i = 1, . . . , Nμ , (A.16)
l=0 j = 1, . . . , Nν
s
where R̂i j,μν ( , t) ∈ o( s ) uniformly on R as → 0 and κi(s) j,μν (t) :=
l
 l=0

ki j,μν;l (t). The κi(s)
j,μν (t) is called the sth-order asymptotic approximation of κi j,μν ∗ .
Further, all ki j,μν;l can also be written as a Fourier sum.

Proof For fixed functions φi,μ (t) the nonautonomous systems corresponding to the
differential equation (2.26) possess the following parametrization of the pullback
attractor
 t

(κi j,μν (t))∗ := − lim e− (t−t ) sin(φi,μ (t  ) − φ j,ν (t  ) + β) dt  .
t0 →−∞ t0


Using (A.15) with f = sin(φi,μ − φ j,ν +β) and g = μν t + ai j,μν + β + rl=1 l
pi j,μν;l we can perform a Taylor expansion of f ( , t) = sin(φi,μ − φ j,ν + β) around
= 0. Due to Theorem 2.4.15 in [7, pp. 93–94] we get


s 
r
l
ri j,μν;l (β,t) + s
Ri j,μν ( , t) := sin(μν t + ai j,μν + β + l
pi j,μν;l (t))
l=0 l=1

= sin(μν t + ai j,μν + β) + cos(μν t + ai j,μν + β) pi j,μν;1 + · · · + s

Ri j,μν ( , t)
Appendix A: Proof of Results from the Main Text and Supplemental Material 191

where the abbreviations pi j,μν;l := pi,μ;l − p j,ν;l and ai j,μν := ai,μ − a j,ν are used.
Here, Ri j,μν denotes the remainder of the Taylor expansion. The remainder R( , t) →
0 and Ri j,μν ( , t) ∈ o( ) for all t ∈ R as → 0. We get


s  
  t
(t−t  )
t
(t−t  )
κi j,μν (t) ∗ = − l+1
e− ri j,μν;l (β, t  ) dt  + s+1
e− Ri j,μν ( , t  ) dt  .
l=0 −∞ t0
(A.17)
 
In order to derive the expansion for κi j,μν ∗ the integrals of the formula above have
to be investigated. Faà di Bruno’s formula (A.15) provides us with

 D k1 +...+kl sin)(μν t + ai j,μν + β) l km
ri j,μν;l (β, t) := pi j,μν;m .
(k1 , ... ,kl )∈Tl
k 1 ! · · · · · kl ! m=1
(A.18)

First, we conclude that the Taylor coefficients ri j,μν;l (β) can also be written in a
(finite) multi-Fourier sum

ri j,μν;l (β, t) = di j,μν;l;m (β)ei(m)t . (A.19)
m∈M̃(ri j,μν;l (β))

!l km
Second, di j,μν;l;0 = 0 if m=1 pi j,μν,m possess a non vanishing term for eiμν t .
With this, we are able to calculate the integrals in (A.17) and get
 t   t
(t−t  ) (t−t  ) i(m)t 
e− ri j,μν;l (β, t) = di j,μν;l;m (β) e− e dt 
−∞ −∞
m∈M̃(ri j,μν;l (β))
 1
= di j,μν;l;m (β)ei(m)t .
+ i(m)
m∈M̃(ri j,μν;l (β))

The last term in the Eq. (A.17) is in o( s ) which can be seen as follows. Since
Ri j,μν ( , t) ∈
 o( ) as → 0 for all C > 0 there exist an 0 (t) > 0 such that
 Ri j,μν ( , t) < C for all < 0 (t). Due to the boundedness of all pi,μ;l the remain-
der Ri j,μν ( , t) is also bounded by some positive number C̃. Thus, Ri j,μν ( , t) ∈ o( )
uniformly on R as → 0 , hence for all C > 0
 t 
   
 e− (t−t  )
R( , t  ) dt   ≤ C 1 − e (t0 −t) 
.

t0

Finally, we end up with

192 Appendix A: Proof of Results from the Main Text and Supplemental Material

  
s  β 1
κi j,μν ∗ (t) = − l
di j,μν;l;m ei(m)t + R̃i j,μν ( , t)
l=0 β 1 + i (m)
m∈M̃(ri j,μν;l )
(A.20)

"t 
where R̃i j,μν := s+1 t0 e− (t−t ) R( , t  ) dt  ∈ o( s ) uniformly on R as → 0.
∗
By considering the Laurent series
∞
 n
1
(m)
=− in ,
1+i n=1
(m)

which converges whenever < |(m)|, the coefficients of the expansion

 (s)
κi j,μν ∗ =
s
l=0 ki j,μν;l (t) are given by
l

κi j,μν;0 = −di j,μν;0;0 (β),


l−1  di j,μν;n;m (β)
κi j,μν;l>0 = −di j,μν;l;0 (β) + i l−n ei(m)t .
n=0 m∈M̃(ri j,μν;n (β))/{0}
((m))l−n

Note that, can always be chosen such that < |(m)|, since we consider the
asymptotic limit → 0. The coefficients are determined via comparing the terms
of both sides of the Eq. (A.20) with respect to their order in . In the case μ = ν
we get κi j,μν;0 = 0. All terms of order O( s+1 ) and R̃i j,μν ( , t) are summarized in
R̂i j,μν ( , t) ∈ o( s ) as → 0. 
Remark A.4.2 Without considering the asymptotic limit → 0, the sth order for-
mal expansion for κi j,μν would have the form as it is given in Lemma A.4.2 under
#s−1 β
the condition that (m) > for all m ∈ l=1 M̃(ri j,μν;l ).
Lemma A.4.3 Suppose everything is given as in Lemma A.4.2. Then, if for all μ ∈
1, . . . , M, i ∈ {1, . . . , Nμ } and l ∈ 1, . . . , r , pi,μ;l (t) can be written completely in
terms of even modes, i.e., all m 1 , . . . , m M are even for (m 1 , . . . , m M ) ∈ M̃( pi,μ;l ),
then for all (n 1 , . . . , n M ) ∈ M̃(κisj,μν (t)) holds: n λ are even for λ = μ, ν and odd
otherwise.
Proof It is fairly easy to verify that if pi,μ;l (t) can be completely written in
terms of even modes for all i, μ, l, so can pi j,μν;l (t) and moreover, the product
pi j,μν;l · pi j,μν;m . According to (A.18) ri j,μν;l consists only of terms of the form
e±iμν t · ei(m)t and hence m λ are even for every λ = μ, ν and odd otherwise for
all (m 1 , . . . , m M ) ∈ M̃(ri j,μν;l ). Since integration by time (A.17) does not make any
changes in the modes, the same holds for ki j,μν;l (t) and hence κisj,μν (t). 
Now, we have everything which is needed to proof the main result in Proposi-
tion 5.3.1.
Appendix A: Proof of Results from the Main Text and Supplemental Material 193

Proof Note, whenever we write (m), (r ) (m) is meant. We omit the super-
script for the sake of readability. (i) Combing the formal time derivative of the first
equation in (5.11), the system Eqs. (2.25)–(2.26) and Lemma A.4.2 we get

 (r )
 1 
r r M Nν 
r 
r
(r ) ∂μ
φ̇i,μ = l
+ l
ṗi,μ;l (t) = − l+n
ki j,μν;l (t)ri j,μν,n (α, t).
∂τl N
l=0 l=1 ν=1 j=1 l=0 n=0
(A.21)

Assume that pi,μ;l (t) = m∈M̃( pi,μ;l ) ci,μ;l;m e
i(m)t
with |M̃( pi,μ;l )| ∈ N. We get

∂(r )
1  
M Nν
μ
= di j,μν;0;m (α)di j,μν;0;0 (β)ei(m)t ,
∂t N ν=1 j=1
m∈M̃(ri j,μν;0 (α))

∂(r
μ
)
1 
M 
Nν 
l
+ ṗi,μ;l (t) = − ki j,μν;m ri j,μν,l−m (α) (A.22)
∂τl N ν=1 j=1 m=0

by comparing both sides of the Eq. (A.21) with respect to the order of . Due to
Lemma A.4.2

1 
Nμ
∂(r )
μ (τ0 , . . . , τm )
= di j,μμ;0;0 (α)di j,μμ;0;0 (β) =: μ,0 ∈ R.
∂t N j=1

This equation can be solved by  ˜ μ = μ,0 t + 

˜ μ,0 (τ1 , . . . , τm ). Due to our assump-
tions the right hand side of Eq. (A.22) can be written as


M 
Nν 
l
− ki j,μν;m ri j,μν,l−m (α) =
ν=1 j=1 m=0
⎡ ⎤

M 
Nν 
l
⎢ 
m−1  di j,μν;n;m (β) i(m)t ⎥
⎣di j,μν;m;0 (β) − ⎦×
m−n
i e
ν=1 j=1 m=0 n=0 m∈M̃(ri j,μν;n (β))/{0}
((m))m−n
⎡ ⎤
⎢  ⎥
⎣di j,μν;l−m;0 (α) + di j,μν;l−m;m (α)ei(m)t ⎦ .
m∈M̃(ri j,μν;l−m (α))/{0}

By using 1.) and 2.) of Lemma A.4.2 we find

194 Appendix A: Proof of Results from the Main Text and Supplemental Material

(r)
1 
M Nν 
l
∂μ *
= di j,μν;m;0 (β)di j,μν;l−m;0 (α)
∂τl N
ν=1 j=1 m=0
⎤

m−1  di j,μν;n;m (β)di j,μν;l−m;−m (α) ⎥
− i m−n ⎦,
n=0 m∈M̃(ri j,μν;n (β))/{0}
((m))m−n

⎡ ⎛ ⎞
1 
M Nν 
l 
⎢ ⎜ ⎟
ṗi,μ;l = ⎣di j,μν;m;0 (β) ⎝ di j,μν;l−m;m (α)ei(m)t ⎠
N
ν=1 j=1 m=0 m∈M̃(ri j,μν;l−m (β))/{0}
⎛ ⎞
⎜ 
m−1  di j,μν;n;m (β) i(m)t ⎟
−di j,μν;l−m;0 (α) ⎝ i m−n e ⎠
n=0 m∈M̃(ri j,μν;n (β))/{0}
((m))m−n
⎤

  ⎥
di j,μν;n;m (β)di j,μν;l−m;n (α) i(m+n)t ⎥
m−1
− i m−n e ⎥.
((m))m−n ⎥
n=0 ⎦
m∈M̃(ri j,μν;n (β))/{0}
n∈M̃(ri j,μν;l−m (α))/{0,−m}

Note that we use the multi-time scale function μ(r ) to deal with all terms of the
expansion describing a linear growth. All the other terms are considered to determine
the behaviour of pi,μ;l . With this ansatz we are able to maintain the boundedness of
pi,μ;l while letting (r )
μ alone describing unbounded behaviour in t ∈ R. Note further
(r )
that μ can be directly computed if all functions pi,μ;k≤l (t) are known. Thus, we
finally end up with


r
(r
μ =
) l
μ,l t.
l=0

We assume now that for all i, μ and l > 1, pi,μ;l can be written completely in terms of
even modes, c.f., Lemma A.4.3. In particular, (μν) ∈ / M( pi,μ;l ). Thus, di j,μν;l;0 (β) =
0 by (A.18) for all μ, ν = 1, . . . , M, i = 1, . . . , Nμ , j = 1, . . . , Nν and l ≥ 1 and

1   
M Nν l,m−1
di j,μν;n;m (β)di j,μν;l−m;n (α) i(m+n)t
ṗi,μ;l = − i m−n e .
N
ν=1 j=1 m=1
((m))m−n
m∈M̃(ri j,μν;n (β))/{0}
n=0
n∈M̃(ri j,μν;l−m (α))/{0,−m}

Hence, we get an equation to determine the value of pi,μ;l inductively. Due to

Lemma A.4.3, we know that if all pi,μ;l can be written in terms of even modes
then m μ and m ν are odd for all (m 1 , . . . , m M ) ∈ M̃(ri j,μν;l (α)). Therefore pi,μ;l can
be written in terms of even modes. This is consistent with our assumption that pi,μ;l
can be written in terms of even modes. Consider further
Appendix A: Proof of Results from the Main Text and Supplemental Material 195

1  
M Nν
di j,μν;0;m (β)di j,μν;0;n (α) i(m+n)t
ṗi,μ;1 = − i e .
N ν=1 j=1 ((m))m−n
m∈M̃(ri j,μν;0 (β))/{0}
n∈M̃(ri j,μν;0 (α))/{0,−m}

The expression for pi,μ,1 can be found by integration

1  
M Nν
di j,μν;0;m (β)di j,μν;0;n (α) i(m+n)t
pi,μ;1 (t) = − e .
N ν=1 j=1 ((m)) (m + n)
m∈M̃(ri j,μν;0 (β))/{0}
n∈M̃(ri j,μν;0 (α))/{0,−m}

Since pi,μ;1 can be written as a Fourier sum the same holds true for all pi,μ,l by
induction. This is consistent with our assumption for pi,μ,l in the beginning of this
proof. The expressions ki j,μν;l follow from Lemma A.4.2. Furthermore, analog to
Lemma A.4.2, from Theorem 2.4.15 in [7, pp. 93–94] we conclude κi j,μν ( , t) −
(r )
κirj,μν ( , t) ∈ O( r ) and φi,μ ( , t) − φi,μ ( , t) ∈ O( r ) for t ∈ O(1/ r ) as → 0.
(ii) To achieve this result we apply now (i) which allows for iteratively determining
the function appearing in the asymptotic expansion.
0th order: For the expansion of the sine-function from Eq. (A.18) we find
ri j,μμ;0 (β) = sin(ai j,μμ + β) = di j,μμ;0;0 (β) and
(1)
ri j,μν;0 (β) = sin((1)
μν t + ai j,μν + β) = di j,μν;0;(μν) (β)e
iμν t
+ c.c.

where di j,μν;0;(μν) (β) := (1/2i)ei(ai j,μν +β ) and c.c. stands for complex conjugated.
Hence, for the coupling matrix we find

κi j,μμ;0 = − sin(ai j,μμ + β).

Depending on the cluster the zero-th order approximation for the frequencies read

1 
Nμ
nμ
μ,0 = sin(ai j,μμ + β) sin(ai j,μμ + β) = (cos(α − β) − cos(α + β))
N j=1 2

n
for all μ = 1, . . . , M1 and analogously μ,0 = 2μ cos(α − β) for all μ = M1 +
1, . . . , M.
1th order: Since we know the 0th order expansion we are able to calculate the
next order. We get
196 Appendix A: Proof of Results from the Main Text and Supplemental Material

1 
M1  Nν  di j,μν;0;m (β)di j,μν;0;−m (α)
μ,1 = − i
N ν=1 j=1 m∈{(μν),−(μν)} (1) (m)
ν=μ
+ ,
1  ei(ai j,μν +β ) e−i(ai j,μν +α) e−i(ai j,μν +β ) ei(ai j,μν +α)
M Nν
1
= −
2N ν=1 j=1 2i (1)
μν (1)
μν
ν=μ
+ ,
1  
M Nν M
1 ei(β−α) e−i(β−α) nν
= − =− sin(α − β).
2N ν=1 j=1 2i (1)
μν (1)
μν ν=1
2(1)
μν
ν=μ ν=μ

For all μ = 1, . . . , M1 we get

1 
M1  Nν  di j,μν;0;m (β)di j,μν;0;m (α) i2(1) (m)t
ṗμ;1 = − i e
N ν=1 j=1 m∈{(μν),−(μν)} (1) (m)
ν=μ


M 
Nν 
di j,μν;0;m (β)di j,μν;0;m (α) i2(1) (m)t
+ i e
ν=M1 +1 j=1 m∈{(μν),−(μν)}
(1) (m)
+ , + ,
1  n ν ei(aμ −aν +α+β) i2(1)
M −i(aμ −aν +α+β)
n ν e (1)
=− (1)
e μν t −
(1)
e−i2μν t
2 ν=1 2i μν 2i μν
ν=μ

and for all μ = M1 + 1, . . . , M we get

1 
M1 Nν  di j,μν;0;m (β)di j,μν;0;m (α) i2(1) (m)t
ṗi,μ;1 = − i e
N ν=1 j=1 m∈{(μν),−(μν)} (1) (m)


M 
Nν  di j,μν;0;m (β)di j,μν;0;m (α) i2(1) (m)t
+ i e
ν=M1 +1 j=1 m∈{(μν),−(μν)}
(1) (m)
ν=μ
+ , + ,
1  nν
M
ei(ai,μ −aν +α+β) i2(1) nν e−i(ai,μ −aν +α+β) (1)
=− e μν t − e−i2μν t .
2 ν=1 2i (1)
μν 2i (1)
μν

Thus, solving this fairly easy differential equation the following expression is
obtained for μ = 1, . . . , M1
⎛ ⎞ ⎛ ⎞
1  M
nν ⎜ e i(aμ −aν +α+β)
⎟ i2(1) nν ⎜ e −i(aμ −aν +α+β)
⎟ −i2(1)
pμ;1 = ⎝ 2 ⎠
e μν t +
⎝ 2 ⎠
e μν t .
4 ν=1 2 (1) 2 (1)
ν=μ
μν μν
Appendix A: Proof of Results from the Main Text and Supplemental Material 197

Analogously we find the expression for pi,μ;1 with μ = M1 + 1, . . . , M. For the

coupling matrix we get

 di j,μν;0;m (β) i(1) (m)t

κi j,μν;1 = i e
m∈{(μν),−(μν)}
(1) (m)

ei(ai j,μν +β ) (1) e−i(ai j,μν +β ) (1)

= eiμν t + e−iμν t .
2(1)
μν 2(1)
μν

A.5 From Local to Global Order Parameter

From expression (6.9) we derived two types of one-cluster solution namely antipodal
or local splay states. In the following we derive a remarkable relation between local
and global properties on a nonlocal ring which is: If a cluster is of local splay type,
the cluster is also of global splay type. N
In order to show this, we rewrite the sum i=1 Z i(2) in two ways. Firstly, using
the pure definition of the local order parameter (6.4):


N
1 
N N
Z i(2) (χ ) = ai j ei2χ j = ei2χ j = N Z (2) (χ). (A.23)
i=1
2P i, j=1 j=1

Secondly, the sum can be rewritten using the definition of a local splay type cluster
with Ri(2) (χ) = Rc(2) (χ ) and 2χi = ϑi(2) . Then


N 
N
(2) 
N
Z i(2) (χ) = Rc(2) eiϑi = Rc(2) (χ) ei2χi = N Rc(2) (χ)Z (2) (χ) (A.24)
i=1 i=1 i=1

By equating (A.23) and (A.24) we obtain:

(1 − Rc(2) (χ))Z (2) (χ ) = 0

The latter equation yields R (2) (χ ) = 0 for all local splay type clusters, since
Rc(2) (χ) < 1 by definition.
198 Appendix A: Proof of Results from the Main Text and Supplemental Material

A.6 Stability of One-Cluster States on Nonlocally Coupled

Networks

First note that the set of Eqs. (6.15)–(6.16) can be brought into the following block
form

d δφ M B δφ
=
dt δκ C − IN 2 δκ

where (δφ)T = (δφ⎛1 , . . ⎞

. , δφ N ), (δκ)T = (δκ11 , . . . , δκ1N , δκ21 , . . . , δκ N N ), B =
C1
  ⎜ .. ⎟
B1 · · · B N , C = ⎝ . ⎠, and M, Bn , Cn are N × N matrices with n = 1, . . . , N .
CN
The elements of the block matrices read

⎧
⎪
⎪  N
σ 
N
⎨− σ sin(α − β) aik − σ aii sin(β) cos(α) + aik sin(2(ai − ak ) + α + β), i = j
mi j = 2 2
⎪
⎪ σ ai j  k=1
 k=1
⎩ sin(α − β) − sin(2(ai − a j ) + α + β) , i = j
2
⎧
⎪
⎪ N
⎨− σ sin(α − β) + Im(e−i(2ai +α+β) Z (2) (a)) aik − σ aii sin(β) cos(α) i = j
i
= 2
⎪
⎩ σ ai j sin(α − β) − sin(2(a − a ) + α + β) ,
k=1
⎪
2 i j i = j

−σ an j sin(an − a j + α), i = n
bi j;n =
0, otherwise
⎧
⎪
⎪ 0, j = n, i = j
⎪
⎪
⎨− a cos(a − a + β), j = n, i  = j
ni n i
ci j;n = ,
⎪
⎪ ani cos(an − ai + β), j  = n, i = j
⎪
⎪
⎩0, otherwise

where σ = 1/2P. Note that we use the Schur decomposition (4.11) throughout this
stability analysis. Suppose we have a phase locked state with phases ai ∈ [0, 2π ).
Then, the solution for the characteristic equations corresponding to the linearized
system (6.15)–(6.16) are given by λ = − with multiplicity N 2 − N and by the
solution of the following set of equations, see also Lemma A.2.2,

det ((M − λI N ) ( + λ) + BC) = 0. (A.25)

The second term D := BC of the Schur complement are element wise given by
σ  
di j = − ai j sin(α − β) + sin(2(ai − a j ) + α + β)
2
if i = j and
Appendix A: Proof of Results from the Main Text and Supplemental Material 199

σ N
dii = sin(α − β) − Im(e−i(2ai +α+β) Z i(2k) (a)) aik − σ sin(α) cos(β)
2 k=1

Consider that ai (k) = ik 2π and the base topology has constant row sum ρ ∈ N, i.e.
N N
a
k=1 ik = ρ, then the matrix in (A.25) becomes circulant. For a ring structure, as
considered in this thesis, we have ρ = 2P. Hence it can be diagonalized using the
N eigenvectors ζl = exp(i2πl/N ) = exp(ial (1)) and the eigenvalues μl (λ) are

−1
N
  j
μl = ( m N N − λ ( + λ) + d N N ) + (m N j ( + λ) + d N j )ζl
j=1
-
ρσ i i(α+β) (l−2k) (2k) (2k+l) (2k)
= −λ2 + λ e ZN − ZN − e−i(α+β) Z N − ZN
2 2
.
(l) 2 i i(α+β) (l−2k) (2k) (2k+l) (2k)
+(Z N − 1) sin(α − β) − + σρ e ZN − ZN − e−i(α+β) Z N − ZN .
ρσ 2

Here, we use the shorthand form Z (n) (n)

N = Z N (1). If any of the μl vanishes, the deter-
minant in (A.25) vanishes, as well. Thus, μl (λ) = 0 is giving the quadratic equa-
tion which determine the Lyapunov spectrum of the linearized system (6.15)–(6.16)
around a phase-locked solution and with a base topology, given by the adjacency
matrix A, having constant in-degree. Note that with the result in Sect. 6.2.1, the
complex order parameter can be further simplified
⎛ ⎞
1 ⎝
P
2π ⎠
Z (n)
N = R (n)
N = cos(n j ) . (A.26)
P j=1 N

Using the latter equation and well-known trigonometric equations, the following
relations are derived:

⎛ ⎞
(l−2k) (2k) 1 ⎝ 
P
2π 2π
ZN − ZN = cos( l − 2k) j − cos 2k j ⎠,
P N N
j=1
⎛ ⎞
2 P
π π
=− ⎝ sin l j sin( l − 4k) j ⎠,
P N N
j=1
⎛ ⎞
2 P
π π π π π
=− ⎝ sin l j sin l j cos 4k j − cos l j sin 4k j ⎠,
P N N N N N
j=1

and analogously

⎛ ⎞
(l+2k) (2k) 2 
P
π π π π π
ZN − ZN = − ⎝ sin l j sin l j cos 4k j + cos l j sin 4k j ⎠.
P N N N N N
j=1
200 Appendix A: Proof of Results from the Main Text and Supplemental Material

Combining these relations, we find

i i(α+β) (l−2k) (2k) (2k+l) (2k)

L(α, β, l, k) = e ZN − ZN − e−i(α+β) Z N − ZN
2
2 sin(α + β)  2 π
P
π
= sin (l j ) cos(4k j )
P N N
j=1

2 cos(α + β) 
P
π π π
+i sin l j cos l j sin 4k j .
P N N N
j=1
(A.27)

A.7 Stability of Lifted One-Cluster States

μ μ
For an arbitrary duplex equilibrium of the form φi =t + ai with ak1 =(0, 2π N
k, . . . ,
(N − 1) 2π
N
k) T
and a 2
k = ak
1
− a we start with the linearized system (8.16) of the
main text. This can also be written in the block matrix form
⎛ ˙ 1⎞ ⎛ ⎞ ⎛ 1⎞
δφ A1 m 1 I N B1 0 δφ
⎜δφ˙ 2 ⎟ ⎜m 2 I N A2 0 B ⎟ ⎜δφ 2⎟
⎜ ⎟=⎜ 2 ⎟⎜ ⎟
⎝δκ˙ 1 ⎠ ⎝ C1 0 − I N 2 0 ⎠ ⎝δκ ⎠ 1

δκ˙ 2 0 C2 0 − IN 2 δκ 2

 μ μ μ μ μ μ T
with (δφ μ , δκ μ )T = δφ1 , . . . , δφ N , δκ11 , . . . , δκ1N , δκ21 , . . . , δκ N N , the matri-
ces Aμ , B μ , and C μ follow from system (8.16) of the main text, and m 1 , m 2 ∈ R.
With the help of Schur’s decomposition the characteristic equation for the linearized
system takes the form

2
−N )
(λ + )2(N =0
(λI N − A1 ) (λ + ) − B1 C1 −(λ + )m 1 I N
det = 0. (A.28)
−(λ + )m 2 I N (λI N − A2 ) (λ + ) − B2 C2

The second equation has the block matrix form which is required from Proposi-
tion 8.3.5. All blocks can be diagonalized and commute since they all possess a
cyclic structure; compare Lemma 4.2.1. Thus, we are allowed to apply Proposi-
tion 8.3.5 which we use in order to diagonalize the matrix in Eq. (A.28). For the
diagonalized matrix we find the following equations for the diagonal elements μi
Appendix A: Proof of Results from the Main Text and Supplemental Material 201

(λ + )2 m 1 m 2 − ( pi1 (λ; α 11 , β 1 , α 12 , σ 12 ) − μi )( pi2 (λ; α 22 , β 2 , α 21 , σ 21 ) − μi ) = 0

(A.29)
μ
where i = 1, . . . , N , pi (λ; α μμ , β μ ) is a second order polynomial in λ which
depends continuously on α and β as well as functionally on the type of the one-
cluster state. For every i ∈ {1, . . . , N }, these equations will give us two eigenvalues
μi,1 and μi,2 for the matrix in Eq. (A.28) depending on λ and the system parameters.
Thus, we can write Eq. (A.29) as
  
μi − μi,1 (λ; α, β, σ ) μi − μi,2 (λ; α, β, σ ) = 0

where α, β, σ represent all system parameter chosen for (2.25)–(2.26). In order to

find the eigenvalue λ of the linearized system (A.28) one of the eigenvalues μ has to
vanish. This means that we have to find λ such that Eq. (A.29) equals
 
μi μi − μi,2 (λ; α, β, σ ) = 0

which is equivalent to finding λ such that the following quartic equation is solved

pi1 (λ; α 11 , β 1 , α 12 , σ 12 ) pi2 (λ; α 22 , β 2 , α 21 , σ 21 ) − (λ + )2 m 1 m 2 = 0. (A.30)

Note that here the diagonal elements of A1 are slightly different from those in
Proposition 4.2.2 but they do not affect the result, i.e., the diagonal element equals
ρi (α 11 , β 1 ) − m 1 (α12 ). The same holds true for A2 . Thus, with the two possible
eigenvalues ρi,1,2 (α μμ , β μ ) for the monoplex system from Corollary 4.2.3 one finds
the following quartic equation which give the Lyapunov exponents for the lifted
duplex one-cluster

/ 0
λ − ρi,1 (α 11 , β 1 ) · λ − ρi,2 (α 11 , β 1 ) + m 1 (λ + ) ×
/ 0
λ − ρi,1 (α 22 , β 2 ) λ − ρi,2 (α 22 , β 2 ) + m 2 (λ + ) − (λ + )2 m 1 m 2 = 0.
(A.31)

In the case of a duplex antipodal one-cluster state given by Eq. (8.2) of the main
text with ai1 ∈ {0, π } and ai2 = ai1 − a, Eq. (8.16) possesses the following set of
eigenvalues
 
SDuplex = {− , λi,1 , λi,2 , λi,3 , λi,4 i=1,...,N }

where λi,1,...,4 solve the following N quartic equations

*    1 *   1
(λ + )2 m 1 m 2 − λ − ρi,1
1
· λ − ρi,2
1
+ m 1 (λ + ) λ − ρi,1
2
λ − ρi,2
2
+ m 2 (λ + ) = 0,
202 Appendix A: Proof of Results from the Main Text and Supplemental Material

μ
with m 1 = σ 12 cos(a + α 12 ), m 2 = σ 21 cos(a − α 21 ) and the eigenvalues ρi,1,2 ≡
ρi,1,2 (α μμ , β μ ) for the monoplex system, see Corollary 4.2.3.

A.8 Example for a Complex Adjacency Matrix

The adjacency matrix displayed in Fig. A.1 is obtained by the following procedure.
For each node i of the N nodes, r links are randomly (with uniform distribution)
picked from the set consisting of all links from a node j = i to node i. This pro-
cedure results in a random directed network with N nodes and constant row sum
(in-degree) r . After the procedure we test if the resulting network is connected [8],
see also Sect. 2.1.
index i

index j

Fig. A.1 Adjacency matrix of a connected, directed random network of N = 200 nodes with
constant row sum r = 50. The illustration shows the adjacency matrix where black and white refer
to whether a link between two nodes exist or not, respectively
Appendix A: Proof of Results from the Main Text and Supplemental Material 203

References

1. Boyd S, Vandenberghe L (2004) Convex optimization. Cambridge University Press, Cambridge

2. Liesen J, Mehrmann V (2015) Linear algebra. Springer, Cham
3. Verhulst F (2006) Methods and applications of singular perturbations: boundary layers and
multiple timescale dynamics. Springer, Berlin
4. Hale JK (1980) Ordinary differential equations. Krieger Publishing Company, Malabar
5. Kloeden PE, Rasmussen M (2011) Nonautonomous dynamical systems. American Mathematical
Society, Providence
6. Pötzsche C, Rasmussen M (2006) Taylor approximation of integral manifolds. J Dyn Diff Equ
18:427
7. Abraham R, Marsden JE, Ratiu T (1988) Manifolds, tensor analysis, and applications. Springer,
New York
8. Korte B, Vygen J (2018) Combinatorial optimization. Springer, Berlin