Mathematical and Computational Approaches in Advancing Modern Science and Engineering

Jacques
Bélair · Ian A. Frigaard
Herb Kunze · Roman Makarov
Roderick Melnik · Raymond J. Spiteri Editors
Mathematical
and Computational
Approaches
in Advancing
Modern Science
and Engineering
Mathematical and Computational Approaches in
Advancing Modern Science and Engineering
Jacques Bélair • Ian A. Frigaard • Herb Kunze
Roman Makarov • Roderick Melnik
Raymond J. Spiteri
Editors
Mathematical and
Computational Approaches
in Advancing Modern
Science and Engineering
123
Editors
Jacques Bélair Ian A. Frigaard
Department of Mathematics and Statistics Department of Mathematics
University of Montreal University of British Columbia
Montreal, QC Vancouver, BC
Canada Canada
Herb Kunze Roman Makarov

Department of Mathematics and Statistics Department of Mathematics
University of Guelph Wilfrid Laurier University
Guelph, ON Waterloo, ON
Canada Canada
Roderick Melnik Raymond J. Spiteri

MS2Discovery Institute Department of Computer Science
Wilfrid Laurier University University of Saskatchewan
Waterloo, ON Saskatoon, SK
Canada Canada
ISBN 978-3-319-30377-2 ISBN 978-3-319-30379-6 (eBook)

DOI 10.1007/978-3-319-30379-6
Library of Congress Control Number: 2016943639
Mathematics Subject Classification (2010): 00A69, 00A71, 00A79, 92-XX, 35Qxx, 81T80, 97M10,
47N60, 49-xx, 91Axx, 62Pxx, 97Pxx, 70-xx
© Springer International Publishing Switzerland 2016

This work is subject to copyright. All rights are reserved 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, express or implied, with respect to the material contained herein or for any
errors or omissions that may have been made.
Printed on acid-free paper
This Springer imprint is published by Springer Nature

The registered company is Springer International Publishing AG Switzerland
Preface
This book consists of five parts covering a wide range of topics in applied
mathematics, modeling, and computational science (AMMCS). It resulted from two
highly successful meetings held jointly in Waterloo (Canada) on the main campus
of Wilfrid Laurier University. It is the oldest university in the Cambridge-Kitchener-
Waterloo-Guelph area, a beautiful part of Canada, just west of the city of Toronto.
The main campus of the university is located in a comfortable driving distance
from some of North America’s most spectacular tourist destinations, including the
Niagara Escarpment, a UNESCO World Biosphere Reserve. Over the years, this uni-
versity has become a traditional venue for the International Conference on Applied
Mathematics, Modeling and Computational Science, and in 2015 it was held jointly
with the annual meeting of the Canadian Applied and Industrial Mathematics
(CAIMS) from June 7–12, 2015. The AMMCS interdisciplinary conference series
runs biannually. Focusing on recent advances in applied mathematics, modeling,
and computational science, the 2015 AMMCS-CAIMS Congress drew some of the
top scientists, mathematicians, engineers, and industrialists from all over the world
and was a true celebration of interdisciplinary research and collaboration involving
mathematical, statistical, and computational sciences within a larger international
community.
The book clearly demonstrates the importance of interdisciplinary interactions
between mathematicians, scientists, engineers, and representatives from other dis-
ciplines. It is a valuable source of the methods, ideas, and tools of mathematical
modeling, computational science, and applied mathematics developed for a variety
of disciplines, including natural and social sciences, medicine, engineering, and
technology. Original results are presented here on both fundamental and applied
levels, with an ample number of examples emphasizing the interdisciplinary nature
and universality of mathematical modeling.
The book contains 70 articles, arranged according to the following topics
represented by five parts:
• Theory and Applications of Mathematical Models in Physical and Chemical
Sciences
v
vi Preface
Fig. 1 Participants of the 2015 International AMMCS-CAIMS Congress, Canada (Photo taken by
Tomasz Adamski on the Waterloo Campus at Wilfrid Laurier University)
• Mathematical and Computational Methods in Life Sciences and Medicine

• Computational Engineering and Mathematical Foundation, Numerical Methods,
and Algorithms
• Mathematics and Computation in Finance, Economics, and Social Sciences
• New Challenges in Mathematical Modeling for Scientific and Engineering
Applications
These chapters are based on selected refereed contributions made by the
participants of both meetings. The AMMCS-CAIMS Congress featured over 30
special and contributed sessions with mini-symposia ranging from mathematical
models in nanoscience and nanotechnology to statistical equilibrium in economics
and to mathematical neuroscience, the embedded Conference of the Computational
Fluid Dynamics Society of Canada, and the 2nd Canadian Symposium on Scientific
Computing and Numerical Analysis, as well as larger sessions around such sci-
entific themes as applied analysis and dynamical systems, industrial mathematics,
mathematical biology, financial mathematics, and much more. Over 600 participants
from all continents attended the Congress and shared the latest achievements,
ideas, insights, and theories about modern problems in science, engineering, and
society that can be approached with new advances in mathematical modeling and
mathematical, computational, and statistical methods.
This book presents a selected sample of the above topics and can serve as a
reference to some of the state-of-the-art original works on a range of such topics. It
Preface vii
Fig. 2 Members of the local organizing committee and student volunteers (Photo taken by Dr.
Shyam Badu on the Waterloo Campus at Wilfrid Laurier University)
has a strong multidisciplinary focus, supported by fundamental theories, rigorous

procedures, and examples from applications. Furthermore, the book provides a
multitude of examples accessible to graduate students and can serve as a source
for graduate student projects.
Taking this opportunity, we would like to thank our colleagues on the AMMCS-
CAIMS Congress organizing team, as well as our sponsors and partners, in particu-
lar the Fields Institute and PIMS, and the Centre de Recherches Mathématiques, as
well as Wilfrid Laurier University, NSERC, and the Government of Ontario. Among
others, traditional supporters of the AMMCS Interdisciplinary Conference series
were Maplesoft and SHARCNET, as well as Springer, De Gruyter, and CRC Press.
The Congress was held under the auspices of the MS2Discovery Interdisciplinary
Research Institute based at Wilfrid Laurier University and in cooperation with
the Society of Industrial and Applied Mathematics and the American Institute of
Mathematical Sciences.
The Congress scientific committee included 15 internationally known
researchers. We would like to thank them, as well as the Congress referees whose
help in the refereeing process was invaluable. Among them we had some of the
leading researchers from all parts of the world, and their assistance was decisive in
completing this project. Our technical support committee and students’ team were
exemplary, and we are truly grateful for their efforts. Last but not least, we are
viii Preface
also grateful to the editorial team at Springer, in particular Martin Peters and Ruth
Allewelt, whose continuous support during the entire process was at the highest
professional level.
We believe that the book will be a valuable addition to the libraries, as well
as to private collections of university researchers and industrialists, scientists and
engineers, graduate students, and all of those who are interested in the recent
progress in mathematical modeling and mathematical, computational, and statistical
methods applied in interdisciplinary settings.
Montreal, Canada Jacques Bélair

Vancouver, Canada Ian Frigaard
Guelph, Canada Herb Kunze
Waterloo, Canada Roman Makarov
Waterloo, Canada Roderick Melnik
Saskatoon, Canada Raymond Spiteri
Contents
Part I Theory and Applications of Mathematical Models

in Physical and Chemical Sciences
Compressibility Coefficients in Nonlinear Transport Models
in Unconventional Gas Reservoirs . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 3
Iftikhar Ali, Bilal Chanane, and Nadeem A. Malik
Solutions of Time-Fractional Diffusion Equation
with Reflecting and Absorbing Boundary Conditions Using Matlab . . . . . . 15
Iftikhar Ali, Nadeem A. Malik, and Bilal Chanane
Homoclinic Structure for a Generalized Davey-Stewartson System . . . . . . . 27
Ceni Babaoglu and Irma Hacinliyan
Numerical Simulations of the Dynamics of Vortex Rossby
Waves on a Beta-Plane .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 35
L.J. Campbell
On the Problem of Similar Motions of a Chain of Coupled
Heavy Rigid Bodies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 47
Dmitriy Chebanov
On Stabilization of an Unbalanced Lagrange Gyrostat... . . . . . . . . . . . . . . . . . . . 59
Dmitriy Chebanov, Natalia Mosina, and Jose Salas
Approximate Solution of Some Boundary Value Problems
of Coupled Thermo-Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 71
Manana Chumburidze
Symmetry-Breaking Bifurcations in Laser Systems
with All-to-All Coupling .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 81
Juancho A. Collera
ix
x Contents
Effect of Jet Impingement on Nano-aerosol Soot Formation

in a Paraffin-Oil Flame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 89
Masoud Darbandi, Majid Ghafourizadeh,
and Mahmud Ashrafizaadeh
Normalization of Eigenvectors and Certain Properties
of Parameter Matrices Associated with The Inverse Problem
for Vibrating Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 101
Mohamed El-Gebeily and Yehia Khulief
Computational Aspects of Solving Inverse Problems for Elliptic
PDEs on Perforated Domains Using the Collage Method . . . . . . . . . . . . . . . . . . . 113
H. Kunze and D. La Torre
Dynamic Boundary Stabilization of a Schrödinger Equation
Through a Kelvin-Voigt Damped Wave Equation . . . . . . . .. . . . . . . . . . . . . . . . . . . . 121
Lu Lu and Jun-Min Wang
Molecular-Dynamics Simulations Using Spatial Decomposition
and Task-Based Parallelism .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 133
Chris M. Mangiardi and R. Meyer
Modelling of Local Length-Scale Dynamics and Isotropizing
Deformations: Formulation in Natural Coordinate System .. . . . . . . . . . . . . . . . 141
O. Pannekoucke, E. Emili, and O. Thual
Post-Newtonian Gravitation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 153
Erik I. Verriest
Part II Mathematical and Computational Methods in Life

Sciences and Medicine
A Quantitative Model of Cutaneous Melanoma Diagnosis
Using Thermography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 167
Ephraim Agyingi, Tamas Wiandt, and Sophia Maggelakis
Time-Dependent Casual Encounters Games and HIV Spread.. . . . . . . . . . . . . 177
Safia Athar and Monica Gabriela Cojocaru
Modelling an Aquaponic Ecosystem Using Ordinary
Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 189
C. Bobak and H. Kunze
A New Measure of Robust Stablity for Linear Ordinary
Impulsive Differential Equations . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 197
Kevin E.M. Church
Coupled Lattice Boltzmann Modeling of Bidomain Type
Models in Cardiac Electrophysiology .. . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 209
S. Corre and A. Belmiloudi
Contents xi
Dynamics and Bifurcations in Low-Dimensional Models

of Intracranial Pressure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 223
D. Evans, C. Drapaca, and J.P. Cusumano
Persistent Homology for Analyzing Environmental Lake
Monitoring Data.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 233
Benjamin A. Fraser, Mark P. Wachowiak,
and Renata Wachowiak-Smolíková
Estimating Escherichia coli Contamination Spread in Ground
Beef Production Using a Discrete Probability Model . . . .. . . . . . . . . . . . . . . . . . . . 245
Petko M. Kitanov and Allan R. Willms
The Impact of Movement on Disease Dynamics in a Multi-city
Compartmental Model Including Residency Patch . . . . . .. . . . . . . . . . . . . . . . . . . . 255
Diána Knipl
A Chemostat Model with Wall Attachment: The Effect
of Biofilm Detachment Rates on Predicted Reactor Performance .. . . . . . . . . 267
Alma Mašić and Hermann J. Eberl
Application of CFD Modelling to the Restoration of Eutrophic Lakes . . . . 277
A. Najafi-Nejad-Nasser, S.S. Li, and C.N. Mulligan
On the Co-infection of Malaria and Schistosomiasis . . . . .. . . . . . . . . . . . . . . . . . . . 289
Kazeem O. Okosun and Robert Smith?
A Discrete-Continuous Modeling Framework to Study the Role
of Swarming in a Honeybee-Varroa destrutor-Virus System.. . . . . . . . . . . . . . . . 299
Vardayani Ratti, Peter G. Kevan, and Hermann J. Eberl
To a Predictive Model of Pathogen Die-off in Soil Following
Manure Application.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 309
Andrew Skelton and Allan R. Willms
Mathematical Modeling of VEGF Binding, Production,
and Release in Angiogenesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 319
Nicoleta Tarfulea
A Mathematical Model of Cytokine Dynamics During
a Cytokine Storm.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 331
Marianne Waito, Scott R. Walsh, Alexandra Rasiuk,
Byram W. Bridle, and Allan R. Willms
Examining the Role of Social Feedbacks and Misperception
in a Model of Fish-Borne Pollution Illness . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 341
Michael Yodzis, Chris T. Bauch, and Madhur Anand
xii Contents
Part III Computational Engineering and Mathematical

Foundation, Numerical Methods and Algorithms
Stability Properties of Switched Singular Systems Subject
to Impulsive Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 355
Mohamad S. Alwan, Humeyra Kiyak, and Xinzhi Liu
Input-to-State Stability and H1 Performance for Stochastic
Control Systems with Piecewise Constant Arguments . . .. . . . . . . . . . . . . . . . . . . . 367
Mohamad S. Alwan and Xinzhi Liu
Switched Singularly Perturbed Systems with Reliable Controllers . . . . . . . . 379
Mohamad S. Alwan, Xinzhi Liu, and Taghreed G. Sugati
Application of an Optimized SLW Model in CFD Simulation
of a Furnace .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 389
Masoud Darbandi, Bagher Abrar, and Gerry E. Schneider
Numerical Investigation on Periodic Simulation of Flow
Through Ducted Axial Fan .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 401
Seyedali Sabzpoushan, Masoud Darbandi, Mohsen Mohammadi,
and Gerry E. Schneider
Numerical Analysis of Turbulent Convective Heat Transfer
in a Rotor-Stator Configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 413
D.-D. Dang and X.-T. Pham
Determining Sparse Jacobian Matrices Using Two-Sided
Compression: An Algorithm and Lower Bounds . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 425
Daya R. Gaur, Shahadat Hossain, and Anik Saha
An h-Adaptive Implementation of the Discontinuous Galerkin
Method for Nonlinear Hyperbolic Conservation Laws
on Unstructured Meshes for Graphics Processing Units . . . . . . . . . . . . . . . . . . . . 435
Andrew Giuliani and Lilia Krivodonova
Extending BACOLI to Solve the Monodomain Model. . . .. . . . . . . . . . . . . . . . . . . . 447
Elham Mirshekari and Raymond J. Spiteri
An Analysis of the Reliability of Error Control B-Spline
Gaussian Collocation PDE Software .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 459
Paul Muir and Jack Pew
On the Simulation of Porous Media Flow Using a New
Meshless Lattice Boltzmann Method . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 469
S. Hossein Musavi and Mahmud Ashrafizaadeh
A Comparison Between Two and Three-Dimensional
Simulations of Finite Amplitude Sound Waves in a Trumpet . . . . . . . . . . . . . . . 481
Janelle Resch, Lilia Krivodonova, and John Vanderkooy
Contents xiii
A Dual-Rotor Horizontal Axis Wind Turbine In-House Code

(DR_HAWT) .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 493
K. Lee Slew, M. Miller, A. Fereidooni, P. Tawagi, G. El-Hage,
M. Hou, and E. Matida
Numerical Study of the Installed Controlled Diffusion Airfoil
at Transitional Reynolds Number . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 505
Hao Wu, Paul Laffay, Alexandre Idier, Prateek Jaiswal,
Marlène Sanjosé, and Stéphane Moreau
Part IV Mathematics and Computation in Finance,

Economics, and Social Sciences
Financial Markets in the Context of the General Theory
of Optional Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 519
M.N. Abdelghani and A.V. Melnikov
A Sufficient Condition for Continuous-Time Finite Skip-Free
Markov Chains to Have Real Eigenvalues . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 529
Michael C.H. Choi and Pierre Patie
Bifurcations in the Solution Structure of Market Equilibrium
Problems . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 537
F. Etbaigha and M. Cojocaru
Pricing Options with Hybrid Stochastic Volatility Models . . . . . . . . . . . . . . . . . . 549
Glynis Jones and Roman Makarov
Delay Stochastic Models in Finance .. . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 561
Anatoliy Swishchuk
Semi-parametric Time Series Modelling with Autocopulas . . . . . . . . . . . . . . . . . 573
Antony Ware and Ilnaz Asadzadeh
Optimal Robust Designs of Step-Stress Accelerated Life
Testing Experiments for Proportional Hazards Models .. . . . . . . . . . . . . . . . . . . . 585
Xaiojian Xu and Wan Yi Huang
Detecting Coalition Frauds in Online-Advertising . . . . . . .. . . . . . . . . . . . . . . . . . . . 595
Qinglei Zhang and Wenying Feng
Part V New Challenges in Mathematical Modeling

for Scientific and Engineering Applications
Circle Inversion Fractals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 609
B. Boreland and H. Kunze
Computation of Galois Groups in magma . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 621
Andreas-Stephan Elsenhans
xiv Contents
Global Dynamics and Periodic Solutions in a Singular

Differential Delay Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 629
Anatoli F. Ivanov and Zari A. Dzalilov
Localized Spot Patterns on the Sphere for Reaction-Diffusion
Systems: Theory and Open Problems. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 641
Alastair Jamieson-Lane, Philippe H. Trinh, and Michael J. Ward
Continuous Dependence on Modeling in Banach Space Using
a Logarithmic Approximation .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 653
Matthew Fury, Beth Campbell Hetrick, and Walter Huddell
Solving Differential-Algebraic Equations by Selecting
Universal Dummy Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 665
Ross McKenzie and John D. Pryce
On a Topological Obstruction in the Reach Control Problem . . . . . . . . . . . . . . 677
Melkior Ornik and Mireille E. Broucke
Continuous Approaches to the Unconstrained Binary
Quadratic Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 689
Oksana Pichugina and Sergey Yakovlev
Fixed Point Techniques in Analog Systems. . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 701
Diogo Poças and Jeffery Zucker
A New Look at Dummy Derivatives for Differential-Algebraic
Equations . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 713
John D. Pryce and Ross McKenzie
New Master-Slave Synchronization Criteria of Chaotic Lur’e
Systems with Time-Varying-Delay Feedback Control . . .. . . . . . . . . . . . . . . . . . . . 725
Kaibo Shi, Xinzhi Liu, Hong Zhu, and Shouming Zhong
Robust Synchronization of Distributed-Delay Systems
via Hybrid Control .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 737
Peter Stechlinski and Xinzhi Liu
Regularization and Numerical Integration of DAEs Based
on the Signature Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 749
Andreas Steinbrecher
Symbolic-Numeric Methods for Improving Structural Analysis
of Differential-Algebraic Equation Systems. . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 763
Guangning Tan, Nedialko S. Nedialkov, and John D. Pryce
Pinning Stabilization of Cellular Neural Networks
with Time-Delay Via Delayed Impulses. . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 775
Kexue Zhang, Xinzhi Liu, and Wei-Chau Xie
Contents xv
Convergence Analysis of the Spectral Expansion of Stable

Related Semigroups .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 787
Yixuan Zhao and Pierre Patie
Index . . . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . 799
Part I
Theory and Applications of Mathematical
Models in Physical and Chemical Sciences
Compressibility Coefficients in Nonlinear
Transport Models in Unconventional Gas
Reservoirs
Iftikhar Ali, Bilal Chanane, and Nadeem A. Malik
Abstract Transport models for gas flow in unconventional hydrocarbon reservoirs

possess several model parameters such as the density ./, the permeability .K/,
the Knudsen number .Kn /, that are strongly dependent upon the pressure p. Each
physical parameter, say , in the system has an associated compressibility factor
D . p/ (which is the relative rate of change of the parameter with respect to
changes in the pressure, Ali I et al. (2014, Time-fractional nonlinear gas transport
equation in tight porous media: an application in unconventional gas reservoirs.
In: 2014 international conference on fractional differentiation and its applications
(ICFDA), Catania, pp 1–6, IEEE)). Previous models have often assumed that
D Const, such as Cui (Geofluids 9(3):208–223, 2009), and Civan (Transp
Porous Media 86(3):925–944, 2011). Here, we investigate the effect of selected
compressibility factors (real gas deviation factor .Z /, gas density . /, gas viscosity
. /, permeability .K /, and the porosity . / of the source rock) as functions of
the pressure upon rock properties such as K and . We also carry out a sensitivity
analysis to estimate the importance of each model parameter. The results are
compared to available data.
1 Introduction
Unconventional gas reservoirs include tight gas, coalbed methane, and shale gas.
Shale gas is distributed over large areas and is found in discrete largely unconnected
gas pockets. Different methods are applied to induce fractures inside the rocks to
release the gas, such as hydraulic fracturing, but this is very expensive. Hence,
an initial guess is required before drilling. Reservoir simulations can be crucial in
I. Ali () • B. Chanane • N.A. Malik

Department of Mathematics & Statistics, King Fahd University of Petroleum and Minerals,
P. O. Box 5046, Dhahran 31261, Saudi Arabia
e-mail: [email protected]; [email protected]; [email protected];
[email protected]
© Springer International Publishing Switzerland 2016 3

J. Bélair et al. (eds.), Mathematical and Computational Approaches in Advancing
Modern Science and Engineering, DOI 10.1007/978-3-319-30379-6_1
4 I. Ali et al.
assisting this process for economical recovery. This requires accurate determination
of fluid and rock properties, and a realistic transport model, [2, 5, 11, 15].
Unconventional gas reservoirs are characterized by extremely low permeability,
in the nano- to micro-Darcy range, and low porosity, in the 4 %–15 % range. The
gas extraction process is very complex and involves new technologies, and takes
a lot of time, money and human resources, [18]. The science and technology of
tight gas transport and extraction is still in its infancy, and field data urgently
required especially from shale gas reservoirs in order to test the newly emerging
theories.
Reservoir simulations typically solve model transport equations in the form of
advection-diffusion partial differential equations (PDE). Some of the latest models
are highly non-linear, where the apparent diffusivity D. p/ and the apparent velocity
U. p; px/ are strongly non-linear functions of the pressure and its derivative, [7]. D
and U involve compressibility factors of various physical parameters,
@ ln 1 @
D D : (1)
@p @p
and these must be known as functions of p and px . However, most applications to

date have been simplified by assuming constant compressibility factors. The impact
of this important assumption has not been assessed to date.
The aim here is to assess the importance of using fully pressure dependent model
parameters. This is done through numerical simulations of the transport equation
and matching the results against the data from Pong et al. [17]. A sensitivity analysis
is also carried out to assess the importance of each physical parameter in the
system.
2 Physical Properties of Shale Gas Reservoirs
Various flow regimes occur in the gas transport process through tight shale rock
formations [10]. They are classified by a Knudsen number, see Table 1 and [17, 19],
which is the ratio of mean free path of gas molecules
q () to the radius (R) of the
flow channels, Kn D =R. is given by [13], D 2Rg T , where is gas density,
T is temperature, Rqg is universal gas constant, and is gas viscosity. R is given by,
p
[4, 6], R D 2 2 K , where is the tortuosity and is the porosity of porous
media and K is intrinsic permeability. Several recent works have focused transport
on the so-called four flow regimes, Table 1.
Nonlinear Transport Models in Unconventional Gas Reservoirs 5
Table 1 Classification of Knudsen number Flow regimes

flow regimes based on
Knudsen number, [19] Kn < 0:01 Continuous flow
0:01 < Kn < 0:1 Surface diffusion or slip flow
0:1 < Kn < 10 Transition flow
Kn > 10 Knudsen diffusion or free molecular flow
The correlation between porosity and intrinsic permeability is given by the

Kozeny-Carman equation [8]
s ˇKC
K
D
KC ; (2)
˛KC
where < ˛KC 1, 0 ˇKC < 1 and

KC 0. ˛KC ; ˇKC , and
KC are empirical
constants which must be determined, or estimated, before hand.
For the simulation purposes, we use the following porosity-pressure correlation,
D a exp.b pc /; (3)
where a , b and c are model constants. Tortuosity is related to porosity by,
D 1 C a .1 /; (4)
where a is also a model constant.

There is a difference between the intrinsic permeability, K, and the apparent
permeability, Ka . K is the measured permeability from rock samples, but due
to various physical effects such as slip flow, the quantity appearing in transport
equations is Ka . Beskok [3] has derived an formula that relates the two quantities,
Ka D Kf .Kn / (5)
where f .Kn / is the flow condition function given by
f .Kn / D .1 C Kn / .1 C .4 bSF /Kn / .1 bSF Kn /1 ; (6)
where is called the Rarefaction Coefficient Correlation [6] given by

1
D o 1 C A KnB ; (7)
where A and B are empirical constants and bSF in Eq. 6 is the slip factor.
6 I. Ali et al.
Some of the gas adheres (clings) to pore surfaces due to the diffusion of gas
molecules. Cui [9] and Civan [7] developed a formula for estimating the amount of
adsorbed gas based on Langmuir isotherms and is given by
s Mg s Mg qL p
qD qa D ; (8)
Vstd Vstd pL C p
where s (kg/m3 ) denotes the material density of the porous sample, q (kg/m3) is
the mass of gas adsorbed per solid volume, qa (std m3 /kg) is the standard volume
of gas adsorbed per solid mass, qL (std m3 /kg) is the Langmuir gas volume, Vstd
(std m3 /kmol) is the molar volume of gas at standard temperature (273.15 K) and
pressure (101,325 Pa), p (Pa) is the gas pressure, pL (Pa) is the Langmuir gas
pressure, and Mg (kg/kmol) is the molecular weight of gas.
Gas density (kg/m3 ) is given by the real-gas equation of state,
Mg p
D (9)
ZRg T
where Z (dimensionless) is the real gas deviation factor [12] and it can be found by
using the correlation developed by Mahmoud [14] and it is given by
Z D ap2r C bpr C c (10)

a D 0:702 exp.2:5tr / (11)
b D 5:524 exp.2:5tr / (12)
c D 0:044Tr2 0:164tr C 1:15 (13)
where pc is the critical pressure and tc is the critical temperature, and pr D p=pc and
tr D t=tc are the reduced pressure and temperature respectively.
Mahmoud [14] also gave correlations for determining the gas viscosity,
D Sc exp.AB / (14)

A D 3:47 C 1588T 1 C 0:0009Mg
B D 1:66378 0:04679A
3 4 1=6
1 M pc
Sc D
.10:5/4 Tc

0:807Tr0:618 0:357 exp.0:449Tr / C 0:34 exp.4:058Tr / C 0:018
3 Mathematical Formulation
The ultra low permeability and the occurrence of various flow regimes are key
features of unconventional gas reservoirs (UGR). The PDE’s that are used to
describe transport process in conventional gas reservoirs (CGR) are based on
Darcy’s law u D .K=/dp=dx and continuity equation .u/x D 0, where K, ,
and are constants, but such models do not produce satisfactory results in UGRs.
Civan [7] has proposed a transport model for gas flow through tight porous media
which incorporates all flow regimes that occur in the reservoirs. Civan’s model is a
non-linear advection-diffusion PDE for the pressure field p.x; t/, which is given by,
@p @p @2 p
C U. p; px / D D. p/ 2 : (15)
@t @x @x
The apparent diffusivity D (m2 /s) is given by,
Ka
DD f1 . p/ C .1 /q2 . p/g1 ; (16)

and the apparent convective flux (velocity) U (m/s) is given by,
@p
U D 3 . p/D : (17)
@x
where the 1 , 2 and 3 appearing in D and U are given by
1 . p/ D . p/ C . p/; (18)

2 . p/ D q . p/ . p/; (19)
1
3 . p/ D Œ . p/ C Ka . p/ . p/: (20)
where Ka D K C f which is obtained from Eq. (5).

A numerical solver for the system equations (15), (16), (17), (18), (19), and (20)
has been developed. We use a finite volume implicit method with constant grid size
and constant time step. The system is linearised and iterated to convergence before
advancing to the next time step. The implicit nature of the solver gives stability to
the solver which is essential for such a highly non-linear system. The solver can also
be applied to the steady state system, see below.
8 I. Ali et al.
4 Model Validation Under Steady State Conditions
The steady state solution for the pressure field is obtained by solving, (see [1, 7]),

@p @2 p
La D ; 0 x L; (21)
@x @x2
where
@p
La D . p/ C K . p/ C f . p/ . p/ ; (22)
@x
with boundary conditions, p.0/ D pL and p.L/ D pR ; pL and pR assumed known.

Sixteen different models were considered, Table 2. An entry of ‘0’ means that
the compressibility factor is zero, D 0; an entry of ‘p’ means that 6D 0 and the
associated physical parameter is a function of pressure, D . p/. The final column
shows the relative error between the simulated values and the experimental values
of Pong et al. [16], given by,
N cal
X 2
p pmeas
Relative Error D i i
: (23)
iD1
pcal
i
where the summation is over the N D 30 data-points in [16]. Case 1 in Table 2

corresponds to the Darcy law where all the physical parameters are constant and
Table 2 List of models Cases K f Error

considered. In columns 2–5,
an entry of 0 means that the 1 0 0 0 0 2.69e02
compressibility factor is zero; 2 p 0 0 0 2.68e02
an entry of p means that it is 3 0 p 0 0 4.05e03
nonzero and the associated 4 0 0 p 0 3.16e01
physical parameter is function 5 0 0 0 p 2.69e02
of pressure p. The final
column shows the relative 6 p p 0 0 1.17e01
error from simulations using 7 p 0 p 0 3.19e02
Eq. (23) 8 p 0 0 p 2.68e02
9 0 p p 0 1.84eC00
10 0 p 0 p 4.05e03
11 0 0 p p 3.17e01
12 p p p 0 1.37e04
13 p p 0 p 1.17e01
14 0 p p p 3.19e02
15 p 0 p p 1.84eC00
16 p p p p 1.36e04
D 0. Case 16 is the fully pressure-dependent case. An additional case, from

Civan [8] with constant factors for K , , , and , was also carried out.
Figure 1 shows the comparisons of the simulated results (solid lines) for the
pressure against the distance for selected models (see captions) against the data
of Pong et al. [16] (symbols). The inlet pressures for the different simulations
are, respectively from bottom to top, 135, 170, 205, 240, and 275 kPa. Figure 1a
compares with Darcy’s Law Case 1, and it shows significant errors. Figure 1b
compares with Civan’s case, and Fig. 1c compares with Case 16, which is the best
fit to the data. Figure 1d shows the relative error on log-scale for the 16 cases in
Table 2. We refer to Case 16 as the Base Case henceforth (Table 3).
It is important to note that although the Civan case Fig. 1b and the Base Case
Fig. 1c appear to yield similar results, the rock properties obtained in the two cases
are quite different. Civan used D 0:2 independent of pressure and he predicted
a Darcy’s Law b Civan (2011)

280 280
Pin Pin
135 kPa 135 kPa
240 240
170 kPa 170 kPa
Pressure [kPa]
Pressure [kPa]
205 kPa 205 kPa

200 240 kPa 200 240 kPa
275 kPa 275 kPa
160 160
120 120
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3

−3 −3
Distance [m] x 10 x 10
Distance [m]
New Approach
c 280
Pin d 5
10
135 kPa
240
170 kPa
Pressure [kPa]
205 kPa
Relative Error
200 240 kPa

0
275 kPa 10
160
120 −5
10
0 0.5 1 1.5 2 2.5 3 0 5 10 15

−3
Distance [m] x 10 Cases
Fig. 1 Pressure, p against the distance along the core sample, x, from numerical solutions of the
Steady State Model, Eqs. (21), (22), for different inlet pressures, Pin , as indicated by color. Solids
lines are from the simulations, and symbols are data from Pong et al. [16]. (a) Darcy’s law, Case 1
in Table 2, with compressibility factors, D 0. (b) Civan’s model with constant compressibility
factors, D Const, for some parameters, see [7]. (c) Case 16 in Table 2 (new model) with pressure
dependent parameters and non-constant compressibility factors, . p/. (d) Relative errors for the
16 cases in Table 2
10 I. Ali et al.
Table 3 Reservoir model Parameter Parameter

parameters used in the Base
Case, Case 16 in Table 2 L (m) 0:003 ˛KC 1
Nx 100 ˇKC 1
Rg (J kmol1 K1 ) 8314:4
KC 1
Mg (kg kmol1 K1 ) 16 a 0:2
T (K) 350 b 1 106
pc (kPa) 3:1 103 c 1:96
tc (K) 125 0 10
bSF 1 A 0:2
a 1:5 B 0:4
Table 4 The range of Cases ˛KC ˇKC

KC a b c
parameters that are used to
determine the values of 1 1:0 1:0 1.0 0:20 1e6 1:96
permeability, and porosity in 2 1:0 0:65 1e7 0:08 1e6 2:09
Fig. 3, from Eqs. (2) and (3) 3 0:75 0:66 1e7 0:15 1e6 2:09
4 0:25 0:4 1e8 0:15 1e6 1:96
5 0:5 0:5 0.1 0:10 1e6 2:1112
6 0:5 1:5 1.0 0:05 1e8 2:90
7 0:45 0:65 1e6 0:01 1e8 2:88
K D 1 1015 m2 . From the present calculations the porosity is pressure dependent

and in the range 0:01 0:2, and the permeability is also pressure dependent
and in the range 1020 K 103 m2 , which are more realistic (Table 4).
5 Sensitivity Analysis and Estimation of Model Parameters
It is important to determine how much the results and predicted rock properties
change due to small changes in model parameters. A sensitivity analysis was carried
out by adjusting one model parameter at a time by factors of 2 and 1=2, starting
with Case 16 as the base case – One-at-a-Time (OAT) methodology. Sensitivity is
measured by monitoring the changes in the model output.
Figure 2 shows sensitivity to selected parameters: (a) pc (critical pressure), (b)
T (temperature), (c) a (constant in the tortuosity model), (d) a (constant in the
porosity model). Except for the temperature, Fig. 2b, all results show significant
sensitivity to changes in the selected parameter especially at higher inlet pressures.
Figure 3 illustrates the sensitivity of the calculated permeability, and porosity
against the pressure, for different combinations of ˛KC , ˇKC , and
KC , Eq. (2).
a 300 b 300
135 kPa 135 kPa
170 kPa 170 kPa
250 250
205 kPa 205 kPa
240 kPa 240 kPa
p, [kPa]
p, [kPa]
275 kPa 275 kPa

200 200
150 150
100 100
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
−3 −3
x, [m] x 10 x, [m] x 10
c 300 d 300
135 kPa 135 kPa
170 kPa 170 kPa
250 250
205 kPa 205 kPa
240 kPa 240 kPa
p, [kPa]
p, [kPa]
275 kPa 275 kPa

200 200
150 150
100 100
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
−3 −3
x, [m] x 10 x, [m] x 10
Fig. 2 OAT sensitivity analysis of the new model. Symbols are the data from Pong et al. [16] (see
Fig. 2 for details). Sensitivity to the following parameters: (a) Critical pressure pc , (b) Temperature
T, (c) Tortuosity parameter a in Eq. (3), (d) porosity parameter a in Eq. (4). Red lines are the
Base Case parameter values in Table 3. Blue lines: the specific parameter is divided by 2. Green
lines: the specific parameter is multiplied by 2
0
10 0.25
1
6 1
−5 0.20
10 5
3
4
0.15
K, [m ]
2
−10
10
φ
5
7 0.10 2
10
−15 2
6
4 0.05
3 7
−20
10 0
100 150 200 250 300 100 150 200 250 300
p, [kPa] p, [kPa]
(a) Permeability (b) Porosity
Fig. 3 Permeability (K), and porosity (), against the pressure p, based upon the parameter values
in Table 3. Seven cases for each parameter, shown in Table 4, are considered and are indicated
on the plots. (Case 1 is the Base Case, shown as black dashed line.) (a) permeability curves are
obtained using data from columns 2 to 4 of Table 4. (b) Porosity curves are obtained using data
from columns 5 to 7
12 I. Ali et al.
6 Summary
The Base Case 16 in Table 2, which is the fully pressure-dependent non-linear

model, performs better than other models giving the smallest error against available
data, Fig. 1c. Darcy’s Law performs the worst illustrating its limitations for gas
transport in tight porous media. A OAT sensitivity analysis shows that rock
properties such as the porosity, and permeability, are very sensitive to most of the
model parameters, Figs. 2 and 3. In the future, the sensitivity analysis for all of the
model parameters will be completed.
Acknowledgements The authors would like to acknowledge the support provided by King
Abdulaziz City for Science and Technology (KACST) through the Science Technology Unit at
King Fahd University of Petroleum and Minerals (KFUPM) for funding this work through project
No. 14-OIL280-04.
References
1. Ali, I., Malik, N.A., Chanane, B.: Time-fractional nonlinear gas transport equation in tight
porous media: an application in unconventional gas reservoirs. In: 2014 International Con-
ference on Fractional Differentiation and Its Applications (ICFDA), Catania, pp. 1–6. IEEE
(2014)
2. Aziz, K., Settari, A.: Petroleum Reservoir Simulation, vol. 476. Applied Science Publishers,
London (1979)
3. Beskok, A., Karniadakis, G.E.: Report: a model for flows in channels, pipes, and ducts at micro
and nano scales. Microsc. Thermophys. Eng. 3(1), 43–77 (1999)
4. Carman, P.C., Carman, P.C.: Flow of Gases Through Porous Media. Butterworths Scientific
Publications, London (1956)
5. Chen, Z.: Reservoir Simulation: Mathematical Techniques in Oil Recovery. CBMS-NSF
Regional Conference Series in Applied Mathematics, vol. 77. SIAM, Philadelphia (2007)
6. Civan, F.: Effective correlation of apparent gas permeability in tight porous media. Transp.
Porous Media 82(2), 375–384 (2010)
7. Civan, F., Rai, C.S., Sondergeld, C.H.: Shale-gas permeability and diffusivity inferred by
improved formulation of relevant retention and transport mechanisms. Transp. Porous Media
86(3), 925–944 (2011)
8. Civan, F., et al.: Improved permeability equation from the bundle-of-leaky-capillary-tubes
model. In: SPE Production Operations Symposium, Oklahoma City. Society of Petroleum
Engineers (2005)
9. Cui, X., Bustin, A., Bustin, R.M.: Measurements of gas permeability and diffusivity of tight
reservoir rocks: different approaches and their applications. Geofluids 9(3), 208–223 (2009)
10. Cussler, E.L.: Diffusion: Mass Transfer in Fluid Systems. Cambridge University Press,
Cambridge/New York (2009)
11. Darishchev, A., Rouvroy, P., Lemouzy, P.: On simulation of flow in tight and shale gas
reservoirs. In: 2013 SPE Middle East Unconventional Gas Conference & Exhibition, Muscat
(2013)
12. Kumar, N.: Compressibility factors for natural and sour reservoir gases by correlations and
cubic equations of state. M.Sc. Thesis, Texas Tech University (2004)
13. Loeb, L.B.: The Kinetic Theory of Gases. Courier Dover Publications (2004)
14. Mahmoud, M.: Development of a new correlation of gas compressibility factor (z-factor) for
high pressure gas reservoirs. J. Energy Resour. Technol. 136(1), 012903 (2014)
15. Peaceman, D.W.: Fundamentals of Numerical Reservoir Simulation. Elsevier, New York
(1977)
16. Pong, K.C., Ho, C.M., Liu, J., Tai, Y.C.: Non-linear pressure distribution in uniform microchan-
nels. ASME public. FED 197, 51–51 (1994)
17. Rathakrishnan, E.: Gas Dynamics. PHI Learning, New Delhi (2013)
18. Wang, Z., Krupnick, A.: A retrospective review of shale gas development in the United States.
What led to the boom? Pub. Resources, Washington (2013)
19. Ziarani, A.S., Aguilera, R.: Knudsen’s permeability correction for tight porous media. Transp.
Porous Media 91(1), 239–260 (2012)
Solutions of Time-Fractional Diffusion Equation
with Reflecting and Absorbing Boundary
Conditions Using Matlab
Iftikhar Ali, Nadeem A. Malik, and Bilal Chanane
Abstract The main objective of this work is to develop Matlab programs for
solving the time-fractional diffusion equation (TFDE) with reflecting and absorbing
boundary conditions on finite and infinite domains. Essentially, there are three
major codes, one for finding the exact solution of the TFDE and other two are
for finding the numerical solution of the TFDE. The code for finding the exact
solutions is based on the fundamental solution of the TFDE, whereas the codes
for finding the numerical solutions are based on the explicit and the implicit finite
difference schemes, respectively. Finally, we illustrate the effectiveness of the codes
by applying them to TFDEs with sharp initial data and for various reflecting and
absorbing boundary conditions both on finite and infinite domains. The results show
the difference of solutions between the standard diffusion equation and the time-
fractional diffusion equation.
1 Introduction
Many physical processes evolve in spaces that are heterogeneous in nature, such as,
crowded system, protein diffusion within cells, anomalous diffusion through porous
media, see [3, 4, 6, 17]. Mathematical models, based on standard calculus, have
failed to describe such intricate processes whereas mathematical models, based on
fractional calculus techniques, have proven their effectiveness in explaining such
complex processes, [1, 2, 5, 10, 15].
Time-fractional diffusion equation have been derived in the framework of Con-
tinuous Time Random Walk (CTRW) model. It is based on the idea of considering
the transport processes as the flow of particles in the form of packets and then
assigning a probability of locating a packet at position x at time t. Law of Total
Probability is used to determine probability P.x; t/. Luchko has derived the time-
fractional diffusion equation by using these concept, see the details in [8, 9]. The
I. Ali () • N.A. Malik • B. Chanane

King Fahd University of Petroleum and Minerals, Dhahran, Saudi Arabia
e-mail: [email protected]

16 I Ali et al.
equation is given by,
@t P.x; t/ D @t1˛ Œv@x P.x; t/ C k2 @2x P.x; t/ (1)
as t ! 1 and jxj ! 1. Equation (1) is called time-fractional advection-diffusion

equation and in the case v D 0 it reduces to time-fractional diffusion equation. For
more detailed derivation, see [9, 11].
In this work, we develop Matlab programs for finding exact and numerical
solutions of the time fractional diffusion equation (TFDE) on finite and infinite
domains, and also with various boundary conditions. The manuscript is organized
as follows; in Sect. 2, procedure for finding the fundamental solution of the TFDE is
explained; in Sect. 3, the numerical schemes are discussed; in Sect. 4, Matlab codes
are provided; in Sect. 5, several examples are given to illustrate the effectiveness of
Matlab programs; finally, in Sect. 6, conclusions are given.
2 Fundamental Solution of Time Fractional Diffusion

Equation
Consider the time fractional diffusion equation, in Caputo form, over the whole real
line with given initial data,
@˛ @2
u.x; t/ D u.x; t/; 0<˛1 (2)
@t˛ @x2
u.x; 0/ D f .x/: (3)
Equation (2) can be written in the integral form as follows,

Z
1 t
u.x; t/ D f .x/ C .t /˛1 uxx .x; /d : (4)

.˛/ 0
Application of Laplace transform yields a second order linear differential equation
uQ xx .x; p/ p˛ uQ .x; p/ D f .x/p˛1 : (5)
The solution of Eq. (5) is given by

Z 1
uQ .x; p/ D Q yj; p˛=2 /p˛1 f .y/dy;
k.jx (6)
1
Time-Fractional Diffusion Equations 17
where k.jxj; / D p 1 jxj1=2 k1=2 .jxj/ is modified Bessel function of second kind
2
[16]. Furthermore, Eq. (6) can be expressed as
Z 1
uQ .x; p/ D Q ˛ .jx yj; p/f .y/dy;
G (7)
1
where G Q ˛ .jxj; p/ D k.jxj;

Q p˛=2 /p˛1 :
Note that directly taking the inverse Laplace transform is not feasible, so we use
the relationship between the Laplace and Mellin transforms to obtain
Z 1
Q ˛ .jxj; s/ D 1
G ps G
Q ˛ .jxj; p/dp

.1 s/ 0
Z 1
jxj1=2
D p p3˛=4s1 kQ 1=2 .jxjp˛=2 /dp: (8)
2
.1 s/ 0
Using the results,

1 sC sC
M Œx f .ax / D a b fQ
b
;
b b
hs i s C
kQ .s/ D 2s2

;
2 2
Equation (8) becomes
Q ˛ .jxj; s/ D 1
Œ1 s=˛
Œ1=2 s=˛
G p 22s=˛ jxj2s=˛1 : (9)
˛
Œ1 s
Taking the inverse Mellin transform and using Fox function, we obtain
2=˛
˛ 1 1 20 jxj .1; 1/
G .jxj; t/ D p jxj H12 : (10)
˛ 22=˛ t .1=2; 1=˛/; .1; 1=˛/
The general solution of the time fractional diffusion equation is given by

Z 1
u.x; t/ D G˛ .jx yj; t/f .y/dy: (11)
1
If the initial data is given as delta potential, that is, u.x; 0/ D ı.x/, then the
solution (11) becomes
2=˛
1 1 20 jxj .1; 1/
u.x; t/ D p jxj H12 : (12)
˛ 22=˛ t .1=2; 1=˛/; .1; 1=˛/
For more details, readers are referred to Wyss [18] and Schneider & Wyss [14].
18 I Ali et al.
3 Numerical Solutions
In this section, we provide two numerical schemes based on finite difference

methods, one is explicit and other one is implicit. First, we give an explicit finite
difference scheme which was derived by Yuste [19]. A uniform grid is placed on the
space-time domain, that is, xj D j x and tm D m t. The numerical approximation
of the unknown function u.x; t/ at the point .xj ; tm / is denoted by Ujm and it is
obtained by Ujm umj u.xj ; tm /. The time-fractional Riemann-Liouville derivative
is discretized by Grunwald-Letnikov derivative [13]. The explicit finite difference
scheme is given by the following formula
X
m
.1 /
umC1
j D um
j C S !k mk
uj1 2ujmk C ujC1
mk
; (13)
kD0
where

.1 / .1 / k2C .1 /
!0 D 1; !k D !k1 : (14)
k
The numerical scheme (13) is stable on the interval 0 S 1=2.2 /, where

S D k t = x2 , for more details about the stability condition, see [19]. Note that
in this work, we only develop Matlab codes for uniform grid; for the case of non-
uniform grid, see for example [20].
We use the implicit finite difference scheme derived by Langlands and Henry [7].
The numerical scheme is given by
j1 C .1 C 2S /uj S ujC1 D uj

S um m m m1
X
m
.1 / mk
C S !k uj1 2ujmk C ujC1
mk
: (15)
kD1
The above scheme (15) is unconditionally stable, see Theorem 2.1 in [12]. For more
technical details, readers are referred to [7, 12].
3.1 Discretization of Boundary Conditions
Dirichlet boundary conditions, u.x; t/ D 0, are discretized in the standard way

and implemented naturally, where x represents left or right boundary. However,
Neumann boundary conditions, ux .x; t/ D 0, are discretized by the second order
difference formula, that is, .um
jC1 uj1 /=2 x, where x represents left or right
m
boundary.
In the case of left Neumann boundary condition, we have um m

1 D u1 and letting
j D 0 in the explicit scheme (13), we obtain
X
m
.1 / mk
umC1
0 0 C 2S
D um !k u1 u0mk : (16)
kD0
On the other hand, the implicit scheme (15) yields
X
m
.1 / mk
.1 C 2S /um
0 2S u1 D u0
m m1
C 2S !k u1 u0mk : (17)
kD1
Similarly, we can obtain the schemes at the right boundary.
4 Matlab Codes
The following Matlab code is used for the computation of the exact solution (12).
Listing 1 Matlab code for exact solution

1 clc, clear all
2 format long e
3 syms x xx c_n
4 n = 55; L = 10; T = [0.1 1 5];
5 dx = 0.1; x1 = (0:dx:L);
6 x0 = 1; gma = 0.5; K_gma = 1; v = 1;
7 ExactSol = zeros(length(x1), length(T));
8 for k = 1:length(T)
9 a1 = 1/sqrt(4*K_gma*T(k)^gma);
10 gama = (-1).^((1:n)-1) ...
11 ./(factorial((1:n)-1).*gamma(1-0.5*gma*(1:n)));
12 xx = ((x-x0).^2/(K_gma*T(k)^gma)).^(0.5*((1:n)-1)) ...
13 - ((x+x0).^2/(K_gma*T(k)^gma)).^(0.5*((1:n)-1));
14 c_n = a1*sum(gama.*xx);
15 ExactSol(:,k) = subs(c_n, x1);
16 end
17 plot(x1,ExactSol),
18 axis([0 5 0 0.72])
The following Matlab code is used for the computation of the numerical solutions
which is based on the explicit finite difference scheme (13). The code can be easily
modified for various initial and boundary conditions.
20 I Ali et al.
Listing 2 Matlab code for explicit finite difference scheme

1 clc, clear all
2 gma = 0.5; S_gma = 0.33;
3 K_gma = 1; L = 10;
4 dt = 0.001; T = 5;
5 t = (0:dt:T); M = length(t);
6 dx = sqrt(K_gma*dt.^gma./S_gma);
7 x1 = (0:dx:1); x2 = (1+dx:dx:L);
8 x = [x1 1 x2]; J = length(x);
9 for a = 1:J
10 if x(a)==1; aa = a; break, end
11 end
12 w = ones(1,M);
13 for k = 2:M
14 w(k) = (1-(2-gma)/(k-1))*w(k-1);
15 end
16 V = zeros(J,M); V(aa,1) = 1/dx;
17 for m = 2:M
18 for j = 2:J-1
19 V(j,m) = V(j,m-1) + ...
20 S_gma*sum(flipud(w(1:m-1)')'.*(V(j-1, 1:m-1)...
21 - 2*V(j,1:m-1) + V(j+1,1:m-1)));
22 end
23 end
24 plot(x,V(:,t==0.1),'o', x, V(:,t==1),'o', x, V(:,t==5),'o')
25 axis([0 5 0 0.72])
The following Matlab code is used for the computation of the numerical solutions
which is based on the implicit finite difference scheme (15). The code can be easily
modified for various initial and boundary conditions.
Listing 3 Matlab code for implicit finite difference scheme

1 function ImplicitSolution
2 clc, clear all
3 alfa = 0.5; K_alfa = 1; L = 10;
4 dt = 0.01; T = 5;
5 t = (0:dt:T); M = length(t);
6 dx = 0.1; x = (0:dx:L); J = length(x);
7 S_alfa = K_alfa*dt^alfa/dx^2;
8 for a = 1:J
9 if x(a)==1; aa = a; break, end
10 end
11 w = ones(1,M);
12 for k = 2:M
13 w(k) = (1-(2-alfa)/(k-1))*w(k-1);
14 end
15 V = zeros(J,M); V(aa,1) = 1/dx;
16 d = S_alfa*ones(1,J-2);
17 E = diag(1+2*d) + diag(-d(1:J-3),1) + diag(-d(2:J-2),-1);
18 for m = 2:M
19 V(2:J-1, m) = BTCS(E, S_alfa, V(:, 1:m), w(2:m));

20 end
21 plot(x,V(:,t==0.1), x, V(:,t==1), x, V(:,t==5))
22 axis([0 5 0 1])
23 end
24 %########################################################
25 function vnew = BTCS(E, s, U, v)
26 v = flipud(v'); [J, m] = size(U);
27 uold = U(2:J-1,m-1);
28 V1 = U(1:J-2, 1:m-1);
29 V2 = U(2:J-1, 1:m-1);
30 V3 = U(3:J, 1:m-1);
31 V4 = V1 - 2*V2 + V3;
32 lhs = uold + s*(V4*v);
33 vnew = E\lhs;
34 end
5 Numerical Experiments
In this section, we provide several examples which are solved by using above Matlab
codes on an Intel Core-i7 machine. The computation time is given for each problem.
Example 1 Consider the TFDE

@u 1 @2 u
D K 0 Dt ; (18)
@t @x2
on the domain x > 0 and t > 0 with the initial data is taken as a delta function at
x D 1. Absorbing and reflecting boundary conditions are taken (one by one) at left
boundary x D 0, where as u.x; t/ ! 0 as x ! 1. Exact solution is given by
u.x; t/ D W.x x0 ; t/ ˙ W.x C x0 ; t/; (19)
where W.x; t/ is the solution of TFDE (18) over the whole real line with decaying
boundary conditions when jxj becomes large. In series form, u.x; t/ is expressed as
1
1 X .1/n .x x0 / n=2 .x C x0 / n=2
u.x; t/ D ˙
4K t nD0 nŠ
.1 .1 C n/=2/ K t K t
(20)
Note the minus sign is taken in the case of absorbing BC and plus sign is taken in the
case of reflecting BC. Figure 1 shows the numerical solutions at times t D 0:1; 1; 5.
The data used for numerical computation is t D 0:001, T D 5, D 0:5, S D
0:33; and the computational time is 9:91 s.
22 I Ali et al.
Fig. 1 Solutions of time fractional diffusion equation 18 with absorbing and reflecting boundary
conditions at the left boundary. Initial condition is taken as delta function at x D 1, where D 0:5.
(a) Absorbing boundary condition. (b) Reflecting boundary condition
Fig. 2 Solutions of time fractional diffusion equation 18 with absorbing boundary conditions at
left and right boundaries. Initial condition is taken as delta function at x D 0. Cusp shape is the
distinct feature of the curve in the case D 0:5. (a) Standard diffusion D 1:0. (b) Anomalous
diffusion D 0:5
Example 2 Consider the TFDE (18) on a box 1 x 1 and t > 0 with absorbing
boundaries at x D 1 and 1. Initial data is given by delta function at x D 0. Exact
solution of the problem is given by
1
X
u.x; t/ D ŒW.x C 4n; t/ W.4n x C 2; t/ : (21)
nD1
Figure 2 shows the numerical solutions at times t D 0:05; 0:1 for D 1 and
at times t D 0:006; 0:1 for D 0:5. The data used for numerical computation is
Fig. 3 Solutions of time fractional diffusion equation 18 with reflecting boundary conditions at
left and right boundaries. Initial condition is taken as delta function at x D 0. Cusp shape is the
distinct feature of the curve in the case D 0:5. (a) Standard diffusion D 1:0. (b) Anomalous
diffusion D 0:5
x D 0:1, T D 0:1, D 1:0, S D 0:49; and the computational time is 9:5 s. In the
case of D 0:5, we choose S D 0:33, and the computational time is 14:71 s.
Example 3 We solve TFDE (18) on a box 1 < x < 1 and t > 0 with reflecting
boundaries at x D 1 and 1. Initial data is given by delta function at x D 0. Exact
solution of the problem is given by
1
X
u.x; t/ D ŒW.x C 4n; t/ C W.4n x C 2; t/ : (22)
nD1
Figure 3 shows the numerical solutions at times t D 0:05; 0:1; 0:2 for D 1 and
at times t D 0:006; 0:1; 0:2 for D 0:5. The data used for numerical computation is
x D 0:01, T D 0:2, D 1:0, S D 0:49; and the computational time is 50:12ṡ. In
the case of D 0:5, we choose S D 0:33, x D 0:1, and the computational time
is 99:97 s.
6 Conclusions
Time fractional advection-diffusion equation is derived under the CTRW framework

[8, 9]. The fundamental solution of the time-fractional diffusion equation is obtained
by applying Fourier and Laplace transforms and which is finally expressed in terms
of Fox’s function by using Mellin transform [14, 18].
Matlab codes are developed for obtaining the exact and the numerical solutions
of time-fractional diffusion equation. The code for the exact solution is based on the
24 I Ali et al.
work of Wyss [18] and Schneider & Wyss [14], whereas the code for finding the
numerical solution is based on the work of Yuste & Acedo (explicit case) [19] and
Langlands & Henry (implicit case) [7].
We have given several examples that illustrate the effectiveness of the codes.
Essentially, numerical solution are found both on finite and infinite domains
with absorbing and reflecting boundary conditions. Initial conditions are taken
as nonsmooth functions (delta functions) and it is observed that in the case of
anomalous diffusion the cusp shape remains in the solutions compared to standard
diffusion where the solution smooths up as time increases.
Acknowledgements The author would like to acknowledge the support provided by King
Abdulaziz City for Science and Technology (KACST) through the Science Technology Unit at
King Fahd University of Petroleum and Minerals (KFUPM) for funding this work through project
No. 14-OIL280-04.
References
1. Ali, I., Malik, N.A.: Hilfer fractional advection–diffusion equations with power-law initial
condition; a numerical study using variational iteration method. Comput. Math. Appl. 68(10),
1161–1179 (2014)
2. Ali, I., Malik, N.A., Chanane, B.: Time-fractional nonlinear gas transport equation in tight
porous media: an application in unconventional gas reservoirs. In: 2014 International Con-
ference on Fractional Differentiation and Its Applications (ICFDA), Catania, pp. 1–6. IEEE
(2014)
3. Chen, W., Sun, H., Zhang, X., Korošak, D.: Anomalous diffusion modeling by fractal and
fractional derivatives. Comput. Math. Appl. 59(5), 1754–1758 (2010)
4. Fedotov, S., Iomin, A.: Migration and proliferation dichotomy in tumor-cell invasion. Phys.
Rev. Lett. 98(11), 118,101 (2007)
5. Jeon, J., Tejedor, V., Burov, S., Barkai, E., Selhuber-Unkel, C., Berg-Sørensen, K., Oddershede,
L., Metzler, R.: In vivo anomalous diffusion and weak ergodicity breaking of lipid granules.
Phys. Rev. Lett. 106(4), 048,103 (2011)
6. Koch, D., Brady, J.: Anomalous diffusion in heterogeneous porous media. Phys. Fluids (1958–
1988) 31(5), 965–973 (1988)
7. Langlands, T., Henry, B.: The accuracy and stability of an implicit solution method for the
fractional diffusion equation. J. Comput. Phys. 205(2), 719–736 (2005)
8. Luchko, Y.: Anomalous diffusion: models, their analysis, and interpretation. Adv. Appl. Anal.
115–145 (2012)
9. Luchko, Y., Punzi, A.: Modeling anomalous heat transport in geothermal reservoirs via
fractional diffusion equations. GEM Int. J. Geomath. 1(2), 257–276 (2011)
10. Malik, N., Ali, I., Chanane, B.: Numerical solutions of non-linear fractional transport
models in unconventional hydrocarbon reservoirs using variational iteration method. In: 5th
International Conference on Porous Media and Their Applications in Science (Engineering
and Industry, Eds, ECI Symposium Series, Volume) (Hawaii, 2014). http://dc.engconfintl.org/
porous_media_V/43
11. Metzler, R., Klafter, J.: The random walk’s guide to anomalous diffusion: a fractional dynamics
approach. Phys. Rep. 339(1), 1–77 (2000)
12. Murio, D.A.: Implicit finite difference approximation for time fractional diffusion equations.
Comput. Math. Appl. 56(4), 1138–1145 (2008)
13. Podlubny, I.: Fractional Differential Equations: An Introduction to Fractional Derivatives, Frac-
tional Differential Equations, to Methods of Their Solution and Some of Their Applications,
vol. 198. Academic, New York (1998)
14. Schneider, W., Wyss, W.: Fractional diffusion and wave equations. J. Math. Phys. 30(1), 134–
144 (1989)
15. Tabei, S., Burov, S., Kim, H., Kuznetsov, A., Huynh, T., Jureller, J., Philipson, L., Dinner, A.,
Scherer, N.: Intracellular transport of insulin granules is a subordinated random walk. Proc.
Natl. Acad. Sci. 110(13), 4911–4916 (2013)
16. Watson, G.N.: A Treatise on the Theory of Bessel Functions. Cambridge University Press,
Cambridge/New York (1995)
17. Weiss, M., Elsner, M., Kartberg, F., Nilsson, T.: Anomalous subdiffusion is a measure for
cytoplasmic crowding in living cells. Biophys. J. 87(5), 3518–3524 (2004)
18. Wyss, W.: The fractional diffusion equation. J. Math. Phys. 27(11), 2782–2785 (1986)
19. Yuste, S., Acedo, L.: An explicit finite difference method and a new von Neumann-type
stability analysis for fractional diffusion equations. SIAM J. Numer. Anal. 42(5), 1862–1874
(2005)
20. Yuste, S.B., Quintana-Murillo, J.: A finite difference method with non-uniform timesteps for
fractional diffusion equations. Comput. Phys. Commun. 183(12), 2594–2600 (2012)
Homoclinic Structure for a Generalized
Davey-Stewartson System
Ceni Babaoglu and Irma Hacinliyan
Abstract In this study, we analyze the homoclinic structure for the generalized
Davey-Stewartson system with periodic boundary conditions. This system involves
three coupled nonlinear equations and describes (2 C 1) dimensional wave prop-
agation in a bulk medium composed of an elastic material with coupled stresses.
We first provide linearized stability analysis of the plane wave solutions of the
generalized Davey-Stewartson system. Then, give an analytic description of the
characteristics of homoclinic orbits near the fixed point by finding soliton type
solutions. These solutions are derived via Hirota’s bilinear method. We also show
that two of these solutions form a pair of symmetric homoclinic orbits and all these
symmetric homoclinic orbit pairs construct the homoclinic tubes.
1 Introduction
Homoclinic orbits are important for the study of chaos in deterministic nonlinear
dynamics. In a neighborhood of such an orbit, an extended knowledge of geometric
structures lead one to a better understanding of chaotic dynamics. The homoclinic
structure of nonlinear Schrodinger (NLS) equation is considered by Ablowitz and
Herbst in [1]. They have shown how this structure associated with the cubic NLS
equation may be obtained from the N-soliton solutions of the defocusing NLS
equation. We follow a similar approach and first observe that the fixed point in
the GDS system is hyperbolic. Then, we analyze the homoclinic structure for the
generalized Davey-Stewartson (GDS) equations.
C. Babaoglu () • I. Hacinliyan

Faculty of Science and Letters, Department of Mathematics, Istanbul Technical University, 34469
Maslak-Istanbul, Turkey
e-mail: [email protected]; [email protected]

28 C. Babaoglu and I. Hacinliyan
The generalized Davey-Stewartson system has been introduced in [2] to study

(2 C 1) dimensional wave propagation in a bulk medium composed of an elastic
material with coupled stresses. They are given by
Q 2A C k2
iAQ C pAQ C rAQ D qjAj .3 Q 1; C 1 Q2; /A;
Q
2!
.c2g c21 /Q 1; c22 Q 1; .c21 c22 /Q 2; D 3 k2 .jAj
Q 2 / ;
.c2g c22 /Q 2; c21 Q2; .c21 c22 /Q 1; D 1 k2 .jAj
Q 2 / ; (1)
where and are spatial coordinates and is time; AQ is the complex amplitude of the
short transverse wave mode, and Q1 and Q2 are long longitudinal and long transverse
wave modes, respectively. The coefficients that appear in (1) can be given as follows
C 2
c21 D ; c22 D ; cg D c22 .k C 8m2 k3 /=!;
0 0
B A C 2B
1 D c21 2c22 C ; 3 D c21 C ;
0 20
1 2 c2 k6 32
pD .cg c22 24m2 c22 k2 /; r D 2 .1 C 8m2 k2 /; q D ;
2! 2! !D1 .2k; 2!/
where k is the wave number, ! is the frequency, m, , , A and B are material

constants, cg is the group speed of transverse waves whereas c1 and c2 are phase
speeds of longitudinal and transverse waves, respectively. Also the frequency and
the dispersion relation is as follows
1
! D c2 k.1 C 4m2 k2 / 2 ; D1 .k; !/ D ! 2 c21 k2 4.1 C /c22 m2 k4 : (2)
A simple algebra shows that the coefficients p, q and r are all positive. In terms of
dimensionless variables, AQ D u; Q1 D '1 ; Q2 D '2 ; and
12 .c2g c21 /2 c2g c21 1 .c2g c21 /

D t; D x; D y; (3)
34 k4 r 3 k 2 32 k2
the three coupled evolution equations take the form
iut C ıuxx C uyy D juj2 u C b.'1;x C '2;y /u;

'1;xx C m2 '1;yy C n'2;xy D .juj2 /x ;
'2;xx C m1 '2;yy C n'1;xy D .juj2 /y ; (4)
Homoclinic Structure for a Generalized Davey-Stewartson System 29
where the non-dimensional coefficients are
p 1 2 q12 .c2g c21 /2 c2g c21 1 2

ıD . / ; D ; bD . /;
r 3 r34 k4 2!r 3
c21 3 2 c22 3 2 c22 c2g c21 c22 3
m1 D . / ; m 2 D . / ; D ; n D . /;
c21 c2g 1 c21 c2g 1 c21 c2g c21 c2g 1
so .1 /.m1 m2 / D n2 , m1 > m2 and < 1. It should be also noted that
c2g c22 D 4m2 c42 k4 .3 C 16m2 k2 /=! 2 (5)
is always positive. There exists a critical wave number

1
kc2 D Œc21 4c22 C c1 .c21 C 8c22 / 2 =.32c22 m2 / (6)
such that c21 c2g < 0 if k > kc and c21 c2g > 0 if k < kc because c21 > c22 . (The
case where k D kc corresponds to long-wave short-wave resonance since the phase
speed of longitudinal wave, c1 , is equal to the group speed of the transverse wave,
cg .) Thus, depending on the wave number k chosen, the coefficients of the second
and third equations of the GDS system may change their sign. For example, the
respective sign of .m1 ; m2 ; / is .; ; C/ if k > kc and is .C; C; / if k < kc .
Now the GDS system will be classified according to the values of parameters. In
fact, since ı > 0 the first equation is always elliptic. The classification of the last
two coupled equations of (4) is based on eigenvalues of the coefficient matrix of a
first order linear system with four equations equivalent to the second-order linear
system, (4)2 and (4)3 . Therefore, system (4) can be classified as elliptic-elliptic-
elliptic, elliptic-elliptic-hyperbolic, and elliptic-hyperbolic -hyperbolic according to
the respective sign of .m1 ; m2 ; /: .C; C; C/, .C; C; /, and .; ; C/ [3]. The
descriptions given above lead that m1 , m2 and cannot be positive at the same time.
Thus, the last two cases correspond to physical cases.
2 Linearized Stability Analysis
In this section, the GDS system (4) is considered with the following boundary
u.x; y; t/ D u.x C l1 ; y C l2 ; t/;

'i .x; y; t/ D 'i .x C l1 ; y C l2 ; t/; i D 1; 2; (7)
and initial conditions
u.x; y; 0/ D u0 .x; y/; u0 .x; y/ D u0 .x; y/;

'i .x; y; 0/ D 'i 0 .x; y/; 'i 0 .x; y/ D 'i 0 .x; y/; i D 1; 2; x; y 2 ˝: (8)
The starting point for developing homoclinic-type solutions is to find a suitable fixed
2
point. In this case, for .u; '1 ; '2 /, the fixed point will be chosen as .a eijaj t ; 0; 0/,
where a is any complex number.
Next, we investigate the stability of fixed point by considering small perturba-
tions of the form,
2
u D a eijaj t .1 ".x; y; t//; '1;x D '1" .x; y; t/; '2;y D '2" .x; y; t/: (9)
Substituting (9) into (4) and keeping linear terms leads one to
i"t C "xx C "yy D jaj2 ." C " / b.'1" C '2" /;

" " "
'1;xx C m2 '1;yy C n'2;xx D jaj2 ." C " /xx ;
" " "
'2;xx C m1 '2;yy C n'1;yy D jaj2 ." C " /yy : (10)
Assuming the following form of solutions for the linearized system we obtain
".x; y; t/ D AŒei.n xCNn y/Cn t C BŒei.n xCNn y/Cn t ;

'1" .x; y; t/ D CŒei.n xCNn y/Cn t C ei.n xCNn y/Cn t ; (11)
'1" .x; y; t/ D DŒei.n xCNn y/Cn t C ei.n xCNn y/Cn t ;
where the growth rate n of the nth mode is
2jaj2 .m1 1/

n˙ D ˙jı2n C Nn 2 jŒ 11=2 : (12)
.ı 1/.2n C m1 Nn 2 /
For the proper choice of the parameters we see that the fixed point is hyperbolic.
3 Homoclinic Structure of the GDS System
Homoclinic structure of the GDS system will be studied in this section. One of the
two physical cases, i.e. the respective sign of .m1 ; m2 ; / is .C; C; / with m1 > m2 ,
will be given in detail, since they have similar approach and result.
In order to write the GDS system in Hirota bilinear form we need to make the
following restrictions on the parameter values
.1 ım1 /
m2 D m1 n; D 1 n; bD : (13)
ı1
We have to note that under these conditions imposed on the parameters generalized
Davey-Stewartson equations are isomorphic to that of the standard integrable
Davey-Stewartson equations.
Since we are interested in a homoclinic orbit of the fixed point, the following
substitutions will be made for (4)
G ia2 t 2.1 ı/ 2.1 ı/

uDa e ; '1 D .log F/x ; '2 D .log F/y ; (14)
F .1 m1 / .1 m1 /
where G is complex, a and F are real. After straightforward calculations we write

the system in Hirota bilinear form
.iD t C ıD 2x C D 2y C a2 /G F D ˇG F;
2ˇ
.D 2x C m1 D 2y /F F 2a2 jGj2 D F F: (15)

Now, Eqs. (15) are assumed to possess the following solution functions:
G D 1 C Œb1 ei. p1 xCp2 y/ C b2 ei. p1 xCp2 y/ e˝tC C b3 e2.˝tC / ;

F D 1 C b4 Œei. p1 xCp2 y/ C ei. p1 xCp2 y/ e˝tC C b5 e2.˝tC / ; (16)
where p1 ; p2 ; ˝; ; b4 ; b5 are real and b1 ; b2 ; b3 are complex. Collecting coefficients

leads to the following relations among the constants
ıp21 C p22 C i˝
b1 D b2 D b4 ; ˇ D a2
ıp21 C p22 i˝
.ıp21 C p22 C i˝/3 2 ıp21 C p22 2 2
b3 D b ; b5 D Œ1 C . / b4 ;
˝ 2 .ıp21 C p22 i˝/ 4 ˝
2a2 .m1 1/
˝˙ D ˙jıp21 C p22 jŒ 11=2 : (17)
.ı 1/. p21 C m1 p22 /
Here the result obtained for ˝ coincides with the result in the linearized stability
analysis. At this step, p1 and p2 will be chosen as
s s
.m1 1/ .m1 1/
p1 D a sin ; p2 D a sin ; (18)
.ı 1/ .ı 1/m1
where 0 is a constant. Then, we get the following solutions for the GDS
system (4) as
2t 1 C 2b˙ cos. p1 x C p2 y/e˝˙ tC C b2˙ e2.˝˙ tC / sec2

u˙ D aeia ;
1 C 2 cos. p1 x C p2 y/e˝˙ tC C e2.˝˙ tC / sec2
˙ 4.ı 1/p21 2e˝˙ tC C cos. p1 x C p2 y/.1 C e2.˝˙ tC / sec2 /

'1;x D e˝˙ tC ;
.1 m1 / .1 C 2 cos. p1 x C p2 y/e˝˙ tC C e2.˝˙ tC / sec2 /2
˙ 4.ı 1/p22 2e˝˙ tC C cos. p1 x C p2 y/.1 C e2.˝˙ tC / sec2 /

'2;y D e˝˙ tC ;
.1 m1 / .1 C 2 cos. p1 x C p2 y/e˝˙ tC C e2.˝˙ tC / sec2 /2
(19)
where b˙ D cos.2/ sin.2/.
4 Conclusions
We will conclude by discussing the homoclinic structure of the GDS system (4).
The solutions (19) have the characteristics of a homoclinic orbit having spatially
periodic fixed points with periods p1 D 2m=L1 and p2 D 2m=L2 . The solution
C C 2
.uC ; '1;x ; '2;y / leaves the ring of fixed points .a eia t ; 0; 0/ as t ! 1
2
and returns to the ring .a eia t b2 ; 0; 0/ as t ! 1. On the other hand,
2
the solution .u ; '1;x ; '2;y / leaves the ring of fixed points .a eia t ; 0; 0/ as
2
t ! 1 and returns to the ring .a eia t b2C ; 0; 0/ as t ! 1. Moreover,
C C
phase shift is seen between the solutions .uC ; '1;x ; '2;y / and .u ; '1;x ; '2;y /.
C C C C
Besides, while .uC ; '1;x ; '2;y /.x0 ; y0 ; t/ forms homoclinic orbit .uC ; '1;x ; '2;y /.x0 C
2m=p1 ; y0 C 2m=p2 ; t/ forms homoclinic orbit as well and this result also
C C
holds for .u ; '1;x ; '2;y /. Finally, we can say that the solutions .uC ; '1;x ; '2;y /

and .u ; '1;x ; '2;y / form a pair of symmetric homoclinic orbits and all of these
orbit pairs construct homoclinic tubes. As an illustration of the dynamical behavior
separated by the homoclinic orbits, the amplitude C D juC j for D 10 and
D ju j for D 10 are shown in Fig. 1.
Fig. 1 The solution of (4) for (a) ı D 0:5, D =3, D 0:5, D 1, a D 1, m1 D 1:75,
m2 D 0:25 and n D 1:5 for (b) ı D 1:5, D =3, D 0:1, D 1, a D 1, m1 D 1:2, m2 D 0:1
and n D 1:1
References
1. Ablowitz, M.J., Herbst, B.M.: On homoclinic structure and numerically induced chaos for the
nonlinear Schrodinger equation. SIAM J. Appl. Math. 50, 339–351 (1990)
2. Babaoglu, C., Erbay, S.: Two-dimensional wave packets in an elastic solid with couple stresses.
Int. J. Non-Linear Mech. 39, 941–949 (2004)
3. Babaoglu, C., Eden, A., Erbay, S.: Global existence and nonexistence results for a generalized
Davey-Stewartson system. J. Phys. A Math. Gen. 37, 11531–11546 (2004)
Numerical Simulations of the Dynamics
of Vortex Rossby Waves on a Beta-Plane
L.J. Campbell
Abstract Observational analyses of hurricanes in the tropical atmosphere indicate

the existence of spiral rainbands which propagate outwards from the eye and affect
the structure and intensity of the hurricane. These disturbances may be described as
vortex Rossby waves. It has been suggested that vortex Rossby waves may play
a role in the eyewall replacement cycle observed in tropical cyclones in which
concentric rings of high-intensity wind develop and propagate in towards the centre
of the cyclone. In previous work with Nikitina, we investigated the dynamics of
vortex Rossby waves in a cyclonic vortex in a two-dimensional configuration on
a beta-plane and derived analytical solutions. In the current work, some results of
numerical simulations are presented.
1 Introduction
A tropical cyclone is an axisymmetric vortex of rotating fluid flow that originates

over the tropical ocean and is driven by heat transfer from the ocean. An intense
tropical cyclone with high sustained wind speed tends to develop an eye, an
area of relative calm at the centre of the circulation surrounded by a ring of
intense rainbands with maximum tangential wind speed called the eyewall. A large
percentage of tropical cyclones with wind speeds of at least 60 ms1 undergo
an eyewall replacement cycle [8] in which an outer ring of rainbands called a
secondary eyewall develops outside the original eyewall [25], strengthens and
contracts inwards, while the original inner eyewall weakens and is replaced by the
secondary eyewall. This process repeats over a time frame of up to 2 days [20].
Vortex Rossby waves are oscillations that occur within cyclonic vortices as a
result of the radial gradient of cyclonic vorticity. In the context of hurricanes or
tropical cyclones, vortex Rossby waves have been observed as spiral rainbands
propagating outwards from the eye. It has been suggested that vortex Rossby
wave mean-flow interactions may play a role in the secondary eyewall formation
[3, 13, 14], although this role appears to be less significant [1, 10] to that played
L.J. Campbell ()

Carleton University, Ottawa, ON, Canada

36 L.J. Campbell
by the dominant system-scale processes such as surface friction and boundary layer
effects [15, 21].
Vortex Rossby wave mean-flow interactions take place primarily in the vicinity
of the critical radius where the mean angular velocity of the vortex is equal to
the phase speed of the waves. This suggests that vortex Rossby wave critical layer
theory could be helpful in advancing our understanding of the mechanisms by which
these waves could contribute to the development of secondary eyewalls in tropical
cyclones. The theory for the analogous problem of barotropic planetary Rossby
waves in a rectangular domain on a beta-plane is well-developed and that problem
has been studied extensively using analytical and numerical methods, e.g. [2, 4–
6, 22–24]. Recent analytical investigations with Nikitina [16–18] of the barotropic
vortex Rossby wave configuration indicate several qualitative similarities between
the two problems.
Nikitina and Campbell [17] presented analytical solutions for a configuration
involving a cyclonic vortex with angular velocity ˝.r/
N on a horizontal plane defined
in terms of polar coordinates r and . The f -plane approximation was made in
this preliminary investigation, i.e., the Coriolis parameter was approximated by a
constant. A wave of the form cos .k !t/ was forced at a fixed radius r D r1
representing the location of the primary eyewall and the linearized barotropic
vorticity equation was solved to determine the amplitude of the forced wave as
a function of the radial variable r and time t. For a special quadratic profile of
˝.r/,
N exact analytical solutions were obtained in terms of hypergeometric functions
for waves with steady amplitude and these solutions were then used to find late-
time asymptotic solutions for waves with time-dependent amplitude, first in the
outer region away from the critical radius and then in the inner region in the
vicinity of the critical radius. The solutions obtained show that the wave amplitude
is greatly attenuated at the critical radius as the wave propagates outwards from
the eyewall. This can be interpreted as wave absorption by the mean flow. In the
limit of infinite time, the time-dependent solution in the outer region approaches
the corresponding steady solution, but the inner solution grows with time. These
conclusions are consistent with the situation that is attained in the case of forced
planetary Rossby waves in a rectangular domain on a ˇ-plane where the Coriolis
parameter is considered to be a linear function of latitude.
Nikitina and Campbell [18] extended the investigation of [17] to include the
nonlinear terms in the governing equations, as well as the terms arising from the
latitudinal gradient of the Coriolis parameter which give the so-called ˇ-effect.
The wave amplitude was then considered as an expansion in powers of two small
parameters representing nonlinearity and the ˇ-effect with the leading-order terms
in the expansion being given by the outer and inner solutions derived in [17]. It was
found that nonlinearity gives rise to higher wavenumbers in multiples of the forced
wavenumber k, a zero wavenumber component which represents a divergence of
momentum flux into the mean flow, and an inward displacement of the instantaneous
critical radius. The variation of the Coriolis force gives rise to wave modes with
wavenumbers .k ˙1/. In the case where the forced wavenumber k D 1, the variation
Numerical Simulations of the Dynamics of Vortex Rossby Waves on a Beta-Plane 37
of the Coriolis force thus introduces an additional zero wavenumber component and
hence contributes to the evolution of the mean flow.
The model employed by Nikitina and Campbell [17, 18] is highly idealized
considering that it is two-dimensional and does not take into account effects
such as diabatic heating and boundary layer friction which form the basis of the
conventional theory of tropical cyclone secondary eyewall generation. But on the
other hand, the simplicity of the model allowed us to examine Rossby wave mean-
flow interaction mechanisms in isolation of other effects and the analytical solutions
obtained gave us some insight into the temporal evolution of the solution. The
main features of the solutions, namely the critical layer absorption of the waves,
the development of concentric rings of high wave activity and the changes in the
location of these rings with time, are consistent with the hypothesis that vortex
Rossby waves contribute to the secondary eyewall replacement cycle. However, the
weakly-nonlinear analysis of [18] is valid only for finite time since the higher-order
terms in the perturbation expansion for the solution grow with time. Multiple-time-
scale asymptotic analyses would be needed to continue the analytical investigation
to later time and to higher orders in the expansion parameters. In analogy with the
classical rectangular configuration [22, 24], it can be anticipated that there would be
wave reflection at the critical radius at late time, possibly leading to eventual wave
breaking and instabilities.
The purpose of the current investigation is to use numerical methods to further
elucidate the critical layer behaviour of forced vortex Rossby waves propagating
outwards in a cyclonic vortex and investigate the effects of the critical layer
interaction on the evolution of the mean vortex. The model used is described in
Sect. 2 and some preliminary numerical results are presented in Sect. 3.
2 Configuration and Equations
The numerical simulations presented here make use of a two-dimensional model for
barotropic vortex Rossby waves in a cyclonic vortex on a ˇ-plane, a horizontal plane
on which the Coriolis parameter is approximated by a linear function of latitude.
The flow is described in terms of a streamfunction and is represented by the
nondimensional barotropic vorticity equation (see, e.g., [9]). This can be written in
polar coordinates r and as [17]
1 1 ˇ
r 2 t r 2 r C r r 2 sin C ˇr cos D 0; (1)
r r r
where the subscripts denote partial differentiation with respect to r and . The non-
dimensional parameter ˇ is the latitudinal gradient of the Coriolis parameter f . The
corresponding dimensional quantity is [9]
2˝Earth
ˇ D cos 0 ; (2)
REarth
38 L.J. Campbell
where 0 is the latitude of the centre of the vortex, ˝Earth 7 105 s1 is the
angular velocity of the Earth’s rotation, and REarth 6:3106 km is the radius of the
Earth. Near the equator where 0 is close to zero, ˇ is close to 2.2 1011 m1 s1 .
For a typical tropical cyclonic vortex radius L of about 2 3 105 m, and a typical
tangential wind speed U of about 30-100 ms1 , ˇ D L2 ˇ =U
102 and can
thus be considered as a small parameter in the nondimensional problem. The limit
of ˇ ! 0 gives the f -plane approximation, in which the Coriolis parameter is
approximated by a constant.
The vortex wave is represented as a small-amplitude perturbation to the basic
flow of the cyclonic vortex. The streamfunction N .r/, the angular velocity ˝.r/
N and
the azimuthal component v.r/
N of the velocity of the basic flow are related by
v.r/
N D N 0 .r/; v.r/
N D r˝.r/;
N (3)
where the prime denotes differentiation with respect to r. The total streamfunction
is written as
.r; ; t/ D N .r/ C " .r; ; t/; (4)
where the parameter " is the ratio of the dimensional magnitude of the perturbation
to that of the basic flow. Observations of the asymmetric spiral bands that are
described as vortex Rossby waves in hurricanes indicate small deviations from
circular symmetry. For example, analyses [7, 19] of Doppler wind and Omega
dropwindsonde data from Hurricane Gloria (1985) show that within 500 km of the
hurricane centre, the asymmetric components of the tangential wind corresponding
to azimuthal wavenumbers 1, 2, 3 and 4 are much smaller than the symmetric
tangential wind. This would suggest that " can be considered as a small parameter
in the nondimensional problem and thus justifies the use of linear [14, 17] and
weakly-nonlinear analyses [18] as a means to provide insight into a fully nonlinear
configuration that can only be examined numerically.
Substituting (4) into (1) gives a nonlinear equation for the perturbation

@ vN @ 1 d vN
C r2 vN 0
C
@t r @ r dr r
ˇ ˇ " 2 2
sin C ˇ r cos C vN cos D . rr r r /: (5)
r " r
This equation is examined in the annular region r1 r < 1; 0 < 2 shown

in Fig. 1 for t 0. The waves are forced at r D r1 by a boundary condition of the
form .r1 ; ; t/ D cos .k !t/, where k is the azimuthal wave number and ! is
the circular frequency of the wave.
2
The mean angular velocity profile is taken to be either ˝.r/N D ˝0 e˛r , as
suggested by Martinez et al. [11, 12] as a reasonable representation of typical flows
Fig. 1 Configuration: the waves are forced by a sinusoidal boundary condition at r D r1 and
propagate outwards [17]
–
Ω
Ω0
ω/k
–
Ω=Ω0 exp(–αr2)
–
Ω=Ω0 (1–αr2)
r1 rc √2rc r
Fig. 2 The mean flow angular velocity profiles used in the numerical simulations and the intervals
in which the analytical solutions obtained in [17, 18] are valid
observed in tropical cyclones, or ˝.r/

N D ˝0 .1 ˛r2 /, as suggested by Brunet
and Montgomery [3] as an approximation for the exponential profile of [11, 12].
These functions are shown in Fig. 2. The positive constants ˛ and ˝0 are chosen
so that there is a critical radius within the domain at some value r D rc where
˝.r
N c / D !=k, as shown in Fig. 2. The numerical method used for the solution of (5)
involves a pseudo-spectral approximation and is based on that described in [4].
40 L.J. Campbell
3 Results
The results presented here are for the quadratic profile shown in Fig. 2. The waves
are forced at r D r1 D 2 with a frequency of ! D 1 and a wavenumber of k D 2
or k D 1. The nonlinear parameter is set to " D 0:05 and the gradient of planetary
vorticity is ˇ D 0 or ˇ D 0:05. In each case, the nondimensional parameters ˛ and
˝0 are chosen so that the critical radius is at r D rc D 12.
Three configurations are examined here: ˇ D 0, k D 2 ( f -plane); ˇ D 0:05,
k D 2 (ˇ-plane); ˇ D 0:05, k D 1 (ˇ-plane). In each case, contour plots of
the perturbation streamfunction and the perturbation vorticity and a graph of the
Fourier or wavenumber spectrum of the perturbation streamfunction are presented
at nondimensional time t D 100. The perturbation streamfunction is written as
1
X
.r; ; t/ D .r; ; t/ei (6)
D1
and j.r; ; t/j is shown as a function of wavenumber at r D rc D 12 and t D 100.

The zero wavenumber component in the sum (6) gives the change in the total mean
flow. The total mean angular velocity at location r and time t is ˝.r; t/ D ˝.r/N C
˝.r; t/, where the change in angular velocity is ˝.r; t/ D "r .r; 0; t/=r. A plot
of ˝.r; t/ at r D rc D 12 and t D 100 is presented for each configuration. From
the asymptotic analysis of [18], it is expected that, at least within the early time
frame in which the analysis is valid, the total mean angular velocity decreases as
a result of nonlinearity and, in the case where k D 1, the ˇ-effect also affects the
mean angular velocity.
Figures 3, 4, 5, and 6 show results obtained with ˇ D 0 and k D 2. In each of
the contour plots shown, there are 20 equally-spaced levels. The analytical results
of [17, 18] tell us that the steady term in the solution varies like rk D r2 between
the inner boundary and the critical radius. This attenuation of the waves and their
absorption by the mean flow at the critical radius is seen in the contour plot of the
wave streamfunction in Fig. 3a. A ring of high vorticity is seen in the vicinity of the
critical radius in Fig. 3b. Figure 4 shows the Fourier spectrum of the perturbation
streamfunction at the critical radius and time t D 100. The black circles indicate
the forced wave modes and the unfilled circles indicate the modes generated by the
nonlinear effects. There is a zero wavenumber component of the solution ( D 0)
as well as higher harmonics ( D ˙4; ˙6; : : :), which are nonzero but of small
amplitude. According to Fig. 5, the mean flow decreases initially and then starts to
increase until the change becomes positive indicating a decrease in the extent of
wave absorption at late time and an evolution towards a reflecting state. This is an
analogous result to that obtained in the classical case of forced planetary Rossby
waves in a rectangular domain on a ˇ-plane [2, 4].
Fig. 3 Nonlinear numerical simulations on an f -plane with k D 2: (a) Wave streamfunction; (b)
Wave vorticity
|φ |
κ
Fig. 4 Nonlinear numerical simulations on an f -plane with k D 2: Fourier spectrum of the
streamfunction amplitude at r D rc D 12
Figures 6, 7, and 8 show results obtained with ˇ D 0:05 and k D 2. With

the addition of the ˇ-effect, Fig. 6a, b indicate that there is a greater degree of
transmission beyond the critical radius than is seen in the corresponding plots
for the f -plane configuration. The graph of the Fourier spectrum given in Fig. 7
shows that there are modes corresponding to D ˙1; ˙3; : : : (indicated by
42 L.J. Campbell
ΔΩ
Fig. 5 Nonlinear numerical simulations on an f -plane with k D 2: change in the mean flow as a
function of time at r D rc D 12
Fig. 6 Nonlinear numerical simulations on a ˇ-plane with k D 2: (a) Wave streamfunction; (b)
Wave vorticity
asterisks), which are generated by the ˇ-effect, in addition to those corresponding

to D 0; ˙2; ˙4; : : :. The presence of these additional modes results in a larger
wave amplitude at the critical radius than that in the f -plane case shown in Fig. 4.
Figure 8 indicates again a decrease in the extent of wave absorption at late time.
|φ |
κ
Fig. 7 Nonlinear numerical simulations on an ˇ-plane with k D 2: Fourier spectrum of the
ΔΩ
Fig. 8 Nonlinear numerical simulations on a ˇ-plane with k D 2: change inthe mean flow as a
Figures 9, 10, and 11 show results obtained with ˇ D 0:05 and k D 1. With
this smaller value of k D 1, the wave amplitude decreases less rapidly with radial
distance from the centre followed by a more rapid attenuation at the critical radius.
This is seen in Fig. 9a, b. From Fig. 10 it is also seen that the ˇ-effect generates
modes corresponding to D ˙.k 1/; ˙.k C 1/ D 0; ˙2; : : :, which coincide
44 L.J. Campbell
Fig. 9 Nonlinear numerical simulations on a ˇ-plane with k D 1: (a) Wave streamfunction; (b)
Wave vorticity
|φ |
κ
Fig. 10 Nonlinear numerical simulations on an ˇ-plane with k D 1: Fourier spectrum of the
with the modes generated by the effect of nonlinearity to produce terms with greater
magnitude than those shown in Fig. 7. Figure 11 shows that the mean flow evolution
takes the form of oscillations with an amplitude that increases slowly with time.
ΔΩ
Fig. 11 Nonlinear numerical simulations on a ˇ-plane with k D 1: change in the mean flow as a
4 Concluding Remarks
Numerical methods were used to investigate vortex Rossby wave mean-flow inter-
actions in a barotropic model of a tropical cyclone. Consistent with the analytical
solutions presented in [17, 18], it was found that nonlinearity gives rise to higher
wavenumbers, momentum flux divergence into the mean flow and an (initial) inward
displacement of the critical radius. The ring of high vorticity around the critical
radius resembles a developing secondary wall. The variation of the Coriolis force
gives rise to wave modes with wavenumbers ˙.k ˙ 1/ and affects the mean flow
if k D 1. In each case, the dominant contributions to the vortex perturbation come
from the zero wavenumber and low wavenumber components, at least within the
time frame shown. This is consistent with observations of Hurricane Gloria (1998)
[7] analyzed and discussed by Shapiro and Montgomery [19].
These numerical results are preliminary; further numerical experimentation and
analyses are needed. Ultimately, vertical variation, diabatic heating and other effects
need to be included for a more realistic representation, but the observations from
the results obtained with this simplified barotropic configuration can be used as a
starting point for further studies.
Acknowledgements The author would like to thank Dr. L. Nikitina for drawing the schematic
diagrams shown in Figs. 1 and 2 and the anonymous reviewer for helpful comments.
46 L.J. Campbell
References
1. Abarca, S.F., Montgomery, M.T.: Essential dynamics of secondary eyewall formation. J.

Atmos. Sci. 70, 3216–3230 (2013)
2. Béland, M.: Numerical study of the nonlinear Rossby wave critical level development in a
barotropic zonal flow. J. Atmos. Sci. 33, 2066–2078 (1976)
3. Brunet, G., Montgomery, M.T.: Vortex Rossby waves on smooth circular vortices: part 1.
Theory. Dyn. Atmos. Oceans 35, 135–177 (2002)
4. Campbell, L.J.: Wave mean-flow interactions in a forced Rossby wave packet. Stud. Appl.
Math. 112, 39–85 (2004)
5. Campbell, L.J., Maslowe, S.A.: Forced Rossby wave packets in barotropic shear flows with
critical layers. Dyn. Atmos. Oceans 28, 9–37 (1998)
6. Dickinson, R.E.: Development of a Rossby wave critical level. J. Atmos. Sci. 27, 627–633
(1970)
7. Franklin, J.L., Lord, S.J., Feuer, S.E., Marks, Jr., F.D.: The kinematic structure of Hurricane
Gloria (1985) determined from nested analyses of dropwindsonde and Doppler radar data.
Mon. Weather Rev. 121, 2433–2451 (1993)
8. Hawkins, J.D., Helveston, M.: Tropical cyclone multiple eyewall characteristics. In: 28th
Conference of Hurricanes and Tropical Meteorology, Orlando, 28 April – 2 May. American
Meteorological Society (2008)
9. Holton, J.R.: An Introduction to Dynamic Meteorology, 3rd edn. Academic, San Diego (1992)
10. Huang, Y.-H., Montgomery, M.T., Wu, C.-C.: Concentric eyewall formation in Typhoon
Sinlaku (2008). Part II: axisymmetric dynamical processes. J. Atmos. Sci. 69, 662–674, (2012)
11. Martinez, Y., Brunet, G., Yau, M.K.: On the dynamics of two-dimensional hurricane-like vortex
symmetrization. J. Atmos. Sci. 67, 3559–3580 (2010)
12. Martinez, Y., Brunet, G., Yau, M.K.: On the dynamics of two-dimensional hurricane-like
concentric rings vortex formation. J. Atmos. Sci. 67, 3253–3268 (2010)
13. Martinez, Y., Brunet, G., Yau, M.K., Wang, X.: On the dynamics of concentric eyewall genesis:
space-time empirical normal modes diagnosis. J. Atmos. Sci. 68, 457–476 (2011)
14. Montgomery, M.T., Kallenbach, R.J.: A theory of vortex Rossby-waves and its application to
spiral bands and intensity changes in hurricanes. Q. J. R. Meteorol. Soc. 123, 435–465 (1997)
15. Montgomery, M.T., Smith, R.K.: Paradigms for tropical cyclone intensification. Aust. Meteo-
rol. Oceanograph. J. 64, 37–66 (2014)
16. Nikitina, L.: Dynamics of vortex Rossby waves in tropical cyclones. Ph.D. thesis, Carleton
University (2013)
17. Nikitina, L.V., Campbell, L.J.: Dynamics of vortex Rossby waves in tropical cyclones, part 1:
linear steady-state exact solutions on an f -plane. Stud. Appl. Math. 135, 377–421 (2015)
18. Nikitina, L.V., Campbell, L.J.: Dynamics of vortex Rossby waves in tropical cyclones, part 2:
nonlinear time-dependent asymptotic analysis on a ˇ-plane. Stud. Appl. Math. 135, 422–446
(2015)
19. Shapiro, L.J., Montgomery, M.T.: A three-dimensional balance theory for rapidly rotating
vortices. J. Atmos. Sci. 50, 3322–3335 (1993)
20. Sitkowski, M., Kossin, J.P., Rozoff, C.M.: Intensity and structure changes during hurricane
eyewall replacement cycles. Mon. Weather Rev. 139, 3829–3847 (2011)
21. Smith, R.K., Montgomery, M.T., Nguyen, V.S.: Tropical cyclone spin up revisited. Q. J. R.
Meteorol. Soc. 135, 1321–1335 (2009)
22. Stewartson, K.: The evolution of the critical layer of a Rossby wave. Geophys. Astro. Fluid
Dyn. 9, 185–200 (1978)
23. Warn, T., Warn, H.: On the development of a Rossby wave critical level. J. Atmos. Sci. 33,
2021–2024 (1976)
24. Warn, T., Warn, H.: The evolution of a nonlinear critical level. Stud. Appl. Math. 59, 37–71
(1978)
25. Willoughby, H.E., Clos, J.A., Shoreibah, M.G.: Concentric eyewalls, secondary wind maxima,
and the evolution of the hurricane vortex. J. Atmos. Sci. 39, 395–411 (1982)
On the Problem of Similar Motions of a Chain
of Coupled Heavy Rigid Bodies
Dmitriy Chebanov
Abstract This chapter contributes to the study of dynamic properties of a chain of

n heavy Hess tops coupled by ideal spherical joints by constructing a new class of
particular solutions of the equations of chain’s motion. The new class describes the
chain’s motions under which all bodies move similar to each other. We establish
conditions for existence of the new solutions and show how the equations of motion
can be reduced to quadratures in the case when these conditions are fulfilled.
1 Introduction
One of the classic directions of investigation in the theory of motion of a system

of several coupled rigid bodies is concerned with particular cases of integrability
of the equations of system’s motion [2, 9, 16]. In comparison with the Euler and
Poisson equations describing the dynamics of a single rigid body about a fixed point,
the analytical study of mathematical models for a system of hinge-connected rigid
bodies is a much more complicated problem; as a consequence, there are very few
known cases of integrability in the dynamics of many-body systems [1, 15].
In this chapter we study dynamic properties of a mechanical system consisting of
an arbitrary number of heavy rigid bodies coupled by ideal spherical joints so that
the system constitutes a chain of rigid bodies. The chain rotates about a fixed point.
Under assumption that, for every chain’s link, the axis connecting its attachment
points passes through its center of mass, we construct a class of particular solutions
of the equations of chain’s motion that describes a rotational motion of the chain
with the following property: while the chain is in motion, its skeleton composed of
the segments of the abovementioned axes bounded by the corresponding attachment
points belongs to a vertical plane rotating about the vertical line while the skeleton’s
segments change their position with respect to the plane identically in time, i.e., all
the bodies move similarly. The similar motions were first discovered for a chain
of n heavy Lagrange tops [14, 18]. Some properties of these motions and their
generalizations can be found in the works [4, 6, 7, 14]. Recently, paper [8] suggested
D. Chebanov ()
City University of New York – LaGCC, Long Island City, NY, USA

48 D. Chebanov
a generic approach for establishing the existence of such motions under no initial
assumption regarding the mass distribution in the bodies.
We present a new particular solution of the problem described above that is
analogous to the solution of the Euler and Poisson equations found by Hess [10].
We describe the structure of the new solution, establish conditions for its existence,
prove compatibility of these conditions, and, in the case when the conditions are
fulfilled, reduce the equations of motion to quadratures. Then, we explore some
aspects of the geometry of the chain’s motion described by the new solution.
2 Description of the Model
In this section, we recall the terminology and notations of [8] complementing them
with the ones needed for the purpose of this contribution.
We consider a mechanical system S consisting of n heavy rigid bodies B1 ; B2 ; : : : ;
Bn . The bodies Bi and BiC1 .i D 1; 2; : : : ; n 1/ are coupled by an ideal spherical
joint at a common point OiC1 so that the system S constitutes a chain of rigid bodies.
One of the chain’s end links, B1 , is absolutely fixed at one of its points O1 .¤ O2 /.
It is assumed that the line li .i D 1; 2; : : : ; n 1/ connecting the attachment points
Oi and OiC1 of the body Bi passes through its center of mass Ci . For the body Bn ,
ln denotes the line passing through the body’s attachment point On and its center of
mass Cn . If the position of the points Ci and OiC1 relative to Oi are determined by
the vectors ci and si , respectively, then, due to the above assumptions, ci D ci ei and
si D si ei , where ei is a unit vector directed along li .
While studying the motion of system S, it is convenient to deal with mass
characteristics of so-called augmented body Bi instead of mass characteristics of Bi .
Let us denote the mass of Bi by mi . By definition, the body Bi consists of the body Bi
Pn
and the point mass mi D mj which is rigidly attached to Bi at point OiC1 . Since
jDiC1
it has been assumed that the points Oi ; Ci , and OiC1 lie on li .i D 1; 2; : : : ; n1/, this
definition implies that the mass center Ci of body Bi lies on li as well. It also follows
from the definition of an augmented body that the absolute angular velocities of
bodies Bi and Bi are equal, while the first-order mass momentum ai and the tensor
of inertia Ii of Bi at Oi can be expressed as
ai D .mi C mi /ci D mi ci C mi si D ai ei ; (1)

Ii D Ii C mi s2i ı si ˝ si D Ii C mi s2i .ı ei ˝ ei / ; (2)
where ci D Oi Ci , ai D mi ci C mi si , Ii is the tensor of inertia of Bi at Oi , ı is the

3 3 identity matrix, and ˝ denotes a dyadic product of two tensors. Since the body
Bn has no descendant chain links connected to it, we have mn D 0 implying that the
bodies Bn and Bn coincide. Therefore, an D mn cn and In D In .
On the Problem of Similar Motions of a Chain of Coupled Heavy Rigid Bodies 49
The vector equations of motion for the system S under consideration can be
written as follows [5, 13]:
X
i1
X
n

.Ii !i / C ai sj ei !j ej C si aj ei !j ej C ai gei D 0;
jD1 jDiC1
(3)
where !i is the absolute angular velocity of body Bi , is the upward vertical unit
vector, and the dot denotes absolute derivative.
Let ˙ D fO1 ; 1 2 3 g be a Cartesian reference frame whose vectors are fixed
.i/ .i/ .i/
in inertial space so that 3 D . Let also ė i D fOi ; eQ 1 eQ 2 eQ 3 g be an orthonormal
moving frame whose axes are the principal axes xof yBi at Oi . Then,x the inertia matrix
e
Ii of Bi in this frame takes the form e Ii D diag eI i ;e
I i ;e
I zi , where e I i ;e
I i , and e
y
I zi are the
principal moments of inertia of Bi at Oi .
In the rest of the chapter, we study a case when, for every i, the center of mass of
Bi belongs to one of its principal planes. In this case, without loss of generality, we
.i/ .i/
assume that Ci is located in the plane formed by the vectors eQ 1 and eQ 3 , implying
.i/ .i/ .i/
that ei D .exi ; 0; ezi / in the principal axes frame, and choose ˙i D fOi ; e1 e2 e3 g to
be a so-called “special” frame [11, 12] whose base vectors are given by

.i/
.i/ ei ei eQ 3
.i/ .i/ eQ ei
e3 D ei ; e2 D ˇ 3 ˇ D eQ 2.i/ ; .i/ .i/
e1 D e2 e3
.i/
D ˇ ˇ ;
ˇ .i/ ˇ ˇ .i/ ˇ
ˇei eQ 3 ˇ ˇei eQ 3 ˇ
i.e., the special frame ˙i can be obtained from the principal axes frame ė i by
.i/
rotating the latter about its second axis. Hence, e2 is a unit vector of a principal axis
of Bi at Oi and the rotation matrix Ri describing the above frame transformation is
given by
0 1
ezi 0 exi
Ri D @ 0 1 0 A :
exi 0 ezi
Then, in the special frame, ei D .0; 0; 1/ and the inertia tensor Ii can be represented
as
0 2 2 x 1 0 1
I xi ezi C e
e I zi exi 0 eIi e I zi exi ezi Iix 0 Iixz
B C @
Ii D Rie
Ii RTi D @ 0 ey
Ii 0 A D 0 Ii 0 A :
y
x 2 2
Ii e
e I zi exi ezi 0 e I xi exi C e I zi ezi Ii 0 Iiz
xz
In order to determine the position of body Bi with respect to the reference frame,
we use Euler angles i ; i ; and 'i , where i .0 i < / is the angle of nutation,
50 D. Chebanov
i .0 i < 2/ is the angle of precession, and 'i .0 'i < 2/ is the angle
of proper rotation. The motion of system S is a superposition of the motion of its
skeleton O1 O2 : : : On Cn , that is composed of the segments of axes li bounded by the
corresponding attachment points, and the pure rotation of each body about li . The
former motion is completely determined by all angles i ; i , while the rotation of Bi
about li is described by the angle 'i .
3 Formulation of the Problem
We will say that system S performs similar motions if it moves so that its skeleton
belongs to a vertical plane ˘ rotating about the vertical axis defined by in
accordance with a non-stationary law .t/ while the skeleton’s segments change
their position with respect to ˘ identically in time, i.e., all the bodies move
similarly. For such motions, it is fulfilled that
i D .t/; i D .t/ C ıi ; (4)
where .t/ and .t/ are functions of time to be determined, ıi 2 f1; 0; 1g, and
i D 1; 2; : : : ; n.
The similar motions (4) were first discovered in [14, 18] for a chain of n heavy
Lagrange tops. Below we follow the strategy for analyzing the problem on similar
motions suggested in [8].
Projecting equations (3) onto the axes of the corresponding body-fixed frames
and substituting (4) into the equations so obtained yields the following overde-
termined system of 3n second-order differential equations with respect to n C 2
unknowns .t/; .t/, and 'i .t/:
h i
y
Jix pP i C Iixz rPi C Iiz Ji qi ri C Iixz pi qi ai g C Hi .cos /R sin cos 'i D 0;
h i
Ji qP i C Jix Iiz pi ri C Iixz ri2 p2i C ai g C Hi .cos /R sin sin 'i D 0;
y
y
Iixz pP i C Iiz rPi C Ii Iix pi qi Iixz qi ri D 0; (5)
where pi D P cos 'i C P sin sin 'i ; qi D P sin 'i C P sin cos 'i ; ri D 'Pi C
P cos ;
Ji˛ D Ii˛ C mi s2i C Li .˛ D x; y/; (6)

X
i1 X
n X
i1 X
n
Li D ai sj "ij C si aj "ij ; Hi D a i sj ij C si aj ij ; (7)
jD1 jDiC1 jD1 jDiC1
"ij is equal to 1, if i D j, and 1, otherwise, and ij D 1 "ij .

In what follows, we construct a new class of particular solutions of the system of

equations (5).
4 Structure of the Solution
Let us assume that each body Bi is a Hess top [10], i.e. a rigid body whose center of
mass lies on the perpendicular to the circular cross-section of the gyration ellipsoid.
In the principal axes frame ė i , the condition on the mass parameters of a Hess top
Bi is given by
x 2 x y 2 z x y
ei e Ii eIi e Ii e
I zi D ezi e Ii e
Ii : (8)
When expressed in ˙i , it can be written as

2 y
Iixz D Iiz Iix Ii : (9)
Below we construct a solution of (5) that, as the Hess solution [10] of the Euler and
Poisson equations, is characterized by the invariant relation
Iixz pi C Iiz ri D 0: (10)
By virtue of (9), (10), equations (5) reduce to the system of 2n equations

h 1 i h i
pi qi ai g C Hi .cos /R sin cos 'i D 0;
y
Ji pP i C Iixz Iiz (11)
h 1 2 i h i
pi C ai g C Hi .cos /R sin sin 'i D 0;
y
Ji qP i Iixz Iiz (12)
which possesses the following first integrals

Ji P 2 C P 2 sin2 C Hi P 2 sin2 C 2ai g cos D hi ; Ji P sin2 D ki ;
y y
(13)
where hi and ki are integration constants.

Solving (13) for P 2 and P yields
P 2 D i ./; P D i ./; (14)

y y y
where i ./ D J .h 2ai g cos / sin2 ki2 = Ji Ji CHi sin2 sin2 ;
y 2 i i
i ./ D ki = Ji sin :
To ensure that, for every i D 1; 2; : : : ; n, equations (11), (12) reduce to the set of
integrals (14) with the same right-hand sides, we require
1 ./ 2 ./ : : : n ./; 1 ./ 2 ./ : : : n ./: (15)

52 D. Chebanov
Each of the relations i ./ 1 ./ and i ./ 1 ./ .i D 2; 3; : : : ; n/ can be

transformed to the form P.cos / 0; where P.cos / is a polynomial in cos
which needs to be satisfied identically in cos . Equating the coefficients of all
powers of cos to zero in the identities obtained from (15) leads to the conditions
y
Ji Hi ai hi ki
y D D D D : (16)
J1 H1 a1 h1 k1
Proposition 1 If the conditions (9) and (16) are fulfilled, the system of equations (3)
has a class of exact solutions with properties (4).
Indeed, we infer from the previous discussion that, under the assumptions of this
proposition, the system (5) is compatible. To find the dependence of the variables
; , and 'i on time, one can proceed as follows. From the first equation in (14), we
find
Z cos s y
J i C Hi Hi 2
t D t0 C y d: (17)
cos 0 .hi 2ai g/ .1 2 / ki2 =Ji
Thus, cos can be obtained as the inverse of the hyperelliptic integral (17), in the
form of an hyperelliptic function. Let cos D F.t t0 /. We can get .t/ by solving
the last equation for . Then, we find .t/ from (14):
Z t
ki d
.t/ D 0 C y :
Ji t0 1 F 2 . t0 /
Finally, we observe that, for each i, the relation (10) can be transformed, by virtue
of (14), (17), into a Ricatti equation with respect to the variable ui .t/ D tan.'i .t/=2/:

.1/ .2/ .3/ .1/ .2/
uP i D i .t/ i .t/ u2i i .t/ui i .t/ i .t/; (18)
where
s
.hi 2ai gF.t t0 // .1 F 2 .t t0 // ki2 =Ji
y
.1/ I xz
i .t/ D iz y ;
2Ii Ji C Hi Hi F 2 .t t0 / .1 F 2 .t t0 //
.2/ ki F.t t0 / .3/ I xz ki

i .t/ D y ; i .t/ D p i :
2Ji .1 F 2 .t t0 // y
Iiz Ji 1 F 2 .t t0 /
h
i
.2/ .1/
Using the change of variables ui .t/ D wP i .t/= i .t/ i .t/ wi .t/ , equa-
tion (18) can be reduced to the second order linear differential equation
.1/ .2/
R i Pi .t/wP i C Pi .t/wi D 0;
w (19)
where
.1/
P i P i
.2/
2
2
.1/ .3/ .2/ .2/ .1/
Pi .t/ D .1/ .2/
i ; Pi .t/ D i i :
i i
Given a solution wi .t/ of (19), one can find 'i .t/ from
n h
io
.2/ .1/
'i .t/ D 2 tan1 wP i .t/= i .t/ i .t/ wi .t/ :
Note, however, that, in a general case, equation (19) cannot be integrated in

quadratures.
5 On Compatibility of the Conditions (9) and (16)
In this section we show that there exist physically meaningful values of the
multibody chain parameters making the conditions (9), (16) compatible in the case
when si ¤ 0 and ai ¤ 0: For the sake of brevity, we consider a simplest case of
a two-body system assuming that 2 D 1 C in (4). Then, due to (2), (6), (7),
"12 D 1; 12 D 2; J1 D I1 C m2 s21 C L1 ; J2 D I2 C L2 ; L1 D L2 D a2 s1 ; H1 D
y y y y
H2 D 2a2 s1 , the conditions (9) are equivalent to (8), and the conditions (16) become
y y
J2 D J1 ; a2 D a1 ; h2 D h1 ; k2 D k1 : (20)
Taking into account (1) and (6), we observe that the relations (8), (20) form an
j 2
algebraic system of 6 equations with respect to 19 parameters e
j
I i .> 0/; ei . exi C
z 2 yc yc
ei D 1/; Ii .> 0/; mi .> 0/; ci ; hi ; ki .i D 1; 2I j D x; z/; s1 . Here Ii is the central
.i/
moment of inertia of body Bi with respect to the axis defined by vector eQ 2 , and
hence
Ii D e
I i D Ii C mi c2i :
y y yc
(21)
In the rest of this section we solve the following problem: if the parameters defining
the mass distribution in the bodies are known, find possible ways for coupling the
bodies as well as the initial conditions of their motion.
From the second equation in (20), we obtain that
m1
c2 D c1 C s1 : (22)
m2
54 D. Chebanov
Substituting the expression for c2 into the first equation in (20) and solving the
equation so obtained for s1 yields
yc yc
I1 I2 m2 m1
s1 D C c1 : (23)
2m1 c1 2m2
Using (23), the formula (22) for determining c2 becomes

yc yc
I1 I2 m2 C m1
c2 D C c1 : (24)
2m1 c1 2m2
In order to satisfy the system of relations (9), (16), one can now select the
parameters of system S as follows. Assuming that the masses mi and parameter c1
have been chosen arbitrarily, the inertia moments e I xi ;e
yc
I zi ; Ii and parameters exi ; ezi can
be selected to comply with (8). Then, the values for s1 and c2 can be found from (23)
and (24), respectively. Finally, the values of hi and ki are to be determined to make
the last two relations in (20) true. Knowing the values of the integration constants,
one can obtain the initial conditions of motion, using (4) and (14).
As follows from the above analysis, in the case under consideration, there is a
unique way of coupling the tops B1 and B2 that guarantees the existence of the
yc yc
motion of interest. If c1 > 0; m2 > m1 ; and I1 I2 , then s1 > 0 and c2 > 0, i.e.
the conditions (9), (16) can
q be satisfied by positive values of c1 ; c2 ; and s1 . Moreover,
1
yc yc
s1 > c1 , when 0 < c1 < m1 m2 .m1 C m2 / 1 I1 I2 .
Thus, we have proven that it is possible to select physically meaningful values of
parameters characterizing the chain of rigid bodies under consideration so that the
conditions (9), (16) are fulfilled. This completes our proof of the existence of the
similar motions for the chain S of Hess tops.
6 Geometry of the Motion of the Chain’s Skeleton
The class of particular solutions of equations (3) that is constructed in the previous
sections of this chapter describes a relatively complex motion of system S; therefore,
a complete analysis of geometry of the system’s motion in this case is a quite
complicated problem. Since the motion of S is a superposition of the motion of its
skeleton O1 O2 : : : On Cn and the pure rotation of each body about li , then, in order
to understand how the system moves, one can study the rotational motion of each
of the above components of its motion and then put them together to get a complete
picture. In this section, we give some properties of the skeleton’s motion.
As was noted before, the skeleton’s motion is completely determined by the
angles i ; i . For the similar motions (4), however, it is sufficient to analyze the
rotation of any of the skeleton’s segments Oi OiC1 about Oi in order to get an
understanding of the nature of the skeleton’s motion. Once such an analysis is
complete, one can get a clear idea of how the plane ˘ rotates about the vertical line
through O1 as well as how the barycentric axes li move relative to ˘ . A general
description of the motion of any segment can be obtained from the properties
of differential equations (14) without integration, employing the methods that are
usually used for studying the rotation of the symmetry axis of a symmetric top about
a fixed point [17]. In what follows, we analyze the rotation of O1 O2 about O1 .
y
Equations (14) can be somewhat simplified by taking D cos ; a D ai g=Ji ; h D
y y y
hi =Ji ; H D Hi =Ji ; and k D ki =Ji . Then, the equations become

2 .h 2a / 1 2 k2 2af1 . / k
P D 2
D ; P D : (25)
1 C H H f2 . / 1 2
We leave it to the reader to verify that, due to (6), (7), (16), (21), f2 . / ¤ 0 at
any moment of time, i.e. there are no singularities in the right-hand side of the first
equation in (25) and thus .t/P is a bounded function of time.
The polynomial f1 . / in (25) is negative for D 1; 1, and C1, and positive
for D C1. Since D cos and for real motion is real, there should be two real
roots, 2 and 3 , of f1 . / between 1 and C1, and a third root, 1 , is greater than
C1. The former can be achieved, for example, by requiring h > k2 . We therefore
conclude that oscillates between the values 2 and 3 .
From the second equation in (25), we observe that P has same sign as k at
any moment of time. Hence, when k ¤ 0, the angle of precession monotonically
increases/decreases over time implying that, while the system S performs the similar
motions, the plane ˘ always rotates in the same direction.
Let us consider a unit sphere drawn at O1 as a center. The l1 -axis intersects the
unit sphere in a point P. As the body B1 moves, this point describes a curve C on
the sphere. Let 2 and 3 be the angles corresponding to 2 and 3 , respectively.
If two cones are constructed so that the apex of each cone is at O1 and the cones’
generating angles are 2 and 3 , respectively, these two cones intersect the sphere
in two circles, C2 and C3 . Since 3 < 2 , then 3 > 2 and therefore the circle C2
lies above the circle C3 . The curve C lies on the sphere between these two circles,
touching the first one and then the other.
Let O1 X; O1 Y, and O1 Z be the coordinate axes of ˙ and let O1 x; O1 y, and O1 z
be the coordinate axes of ˙1 . Following [17], we represent the position of the point
P on the unit sphere by the arc vector ZP D and the angle XZP, which is the
longitude, , of the pole of the xy-plane. Then, P D P and
p
d k f2 . /
D p ; (26)
d .1 2 / 2af1 . /
which is the differential equation of the curve C.

56 D. Chebanov
Fig. 1 Trajectory of the point

P of the barycentric axis l1 on
the unit sphere
Let ˛ be the angle which the curve C makes with the arc vector ZP. Using the
results of [17] and (26), we derive that
p
d d k f2 . /
tan ˛ D sin D 1 2 D p : (27)
d d 2af1 . /
As follows from (26) and (27), tan ˛ is infinite only when is equal to 2 or 3 and
therefore the curve C is tangent to the circle C2 or C3 whenever takes on one of
these values. A typical curve C is shown in Fig. 1. By analogy with [3], one can
show that the time it takes for P to reach C2 starting from C3 is equal to the time it
takes to get back from C3 to C2 .
Thus, while chain S of Hess tops rotates about O1 so that its bodies move similar
to each other, the plane ˘ rotates about the vertical line through O1 in the same
direction and the barycentric axis of each body moves along a spherical curve C
oscillating between C2 and C3 .
Acknowledgements Support for this project was provided by a PSC-CUNY Award, jointly
funded by The Professional Staff Congress and The City University of New York (Award # 67306-
00 45).
References
1. Bolgrabskaya, I., Lesina, M., Chebanov, D.: Dynamics of Systems of Coupled Rigid Bodies.
Naukova Dumka, Kyiv (2012)
2. Borisov, A., Mamaev, I.: Dynamics of a Rigid Body. Hamiltonian Methods, Integrability,
Chaos. Institute of Computer Science, Moscow/Izhevsk (2005)
3. Bukhgolts, N.: An Elementary Course in Theoretical Mechanics, vol. 2. Nauka, Moscow
(1969)
4. Chebanov, D.: On a generalization of the problem of similar motions of a system of Lagrange

gyroscopes. Mekh. Tverd. Tela 27, 57–63 (1995)
5. Chebanov, D.: New dynamical properties of a system of Lagrange gyroscopes. Proc. Inst. Appl.
Math. Mech. 5, 172–182 (2000)
6. Chebanov, D.: Exact solutions for motion equations of symmetric gyros system. Multibody
Syst. Dyn. 6(1), 30–57 (2001)
7. Chebanov, D.: A new class of nonstationary motions of a system of heavy Lagrange tops with
a non-planar configuration of the system’s skeleton. Mekh. Tverd. Tela 41, 244–254 (2011)
8. Chebanov, D.: New class of exact solutions for the equations of motion of a chain of n rigid
bodies. Discret. Cont. Dyn. Syst. Supplement, 105–113 (2013)
9. Gashenenko, I., Gorr, G., Kovalev, A.: Classical Problems of the Rigid Body Dynamics.
Naukova Dumka, Kyiv (2012)
10. Hess, W.: Über die Eulerschen Bewegungsgleichungen und über eine neue particulare Lösung
des Problems der Bewegung eines starren Körpers un einen festen Punkt. Math. Ann. 37(2),
178–180 (1890)
11. Kharlamov, P.: On the Equations of Motion for a Heavy Rigid Body with a Fixed Point. J.
Appl. Math. Mech. 27(4), 1070–1078 (1963)
12. Kharlamov, P.: Lectures on Rigid Body Dynamics. Novosibirsk University, Novosibirsk (1965)
13. Kharlamov, P.: The equations of motion of a system of rigid bodies. Mekh. Tverd. Tela 4,
52–73 (1972)
14. Kharlamov, P.: Some classes of exact solutions of the problem of the motion of a system of
Lagrange gyroscopes. Mat. Fiz. 32, 63–76 (1982)
15. Kharlamova, E.: A survey of exact solutions of problems of the motion of systems of coupled
rigid bodies. Mekh. Tverd. Tela 26(II), 125–138 (1998)
16. Leimanis, E.: The General Problem of the Motion of Coupled Rigid Bodies About Fixed Point.
Springer, Berlin (1965)
17. MacMillan, W.: Dynamics of Rigid Bodies. Dover Publications, New York (1960)
18. Savchenko, A., Lesina, M.: A particular solution for motion equations of Lagrange gyroscopes
system. Mekh. Tverd. Tela 5, 27–30 (1973)
On Stabilization of an Unbalanced Lagrange
Gyrostat
Dmitriy Chebanov, Natalia Mosina, and Jose Salas
Abstract In this contribution we consider a chain of two coupled Lagrange

gyrostats moving about a fixed point in a non-homogeneous gravitational field and
show the existence of the following stabilization effect: for a gyrostat that is in its
unstable equilibrium position, there is another gyrostat such that, when the two of
them are coupled to form this chain, the rotation of the latter gyrostat can be used to
stabilize the equilibrium of the former one. We establish and analyze stabilization
conditions in the space of mechanical parameters characterizing the chain.
1 Introduction
In this contribution we investigate dynamic properties of a chain of two coupled

gyrostats rotating about a fixed point in a gravitational field. By definition, a gyrostat
is a mechanical system consisting of a rigid carrier and other bodies connected to
it such that their motion relative to the carrier does not alter the distribution of
masses of the mechanical system. Examples of such systems include a rigid body to
which axes of several symmetric rotors are connected, or a rigid body with cavities
completely filled with a homogeneous fluid [12, 18]. Mechanical models involving
gyrostats have a number of applications; for instance, they are used for controlling
the attitude dynamics of spacecraft and for stabilizing its rotations [1, 8], play an
important role in studying fluid dynamical systems [4, 5], can be employed in the
analysis of the mechanical models of DNA molecules [16], open-loop dynamic
characteristics of smooth air gap brushless dc motors [7], instabilities associated
with the operation of lasers [6], microscopic dynamics of deterministic chemical
chaos [10], etc.
Assuming that, for each gyrostat, its mass distribution is analogous to the one
of a Lagrange top, its attachment point(s) lie on its dynamic symmetry axis, and
its gyrostatic moment is directed along this axis, we study a problem of stability
of the state of chain’s motion with the following property: one of the gyrostats
permanently rotates about a vertical axis, whereas the other one is at rest. While
D. Chebanov () • N. Mosina • J. Salas

City University of New York – LaGCC, Long Island City, NY, USA
e-mail: [email protected]; [email protected]; [email protected]

60 D. Chebanov et al.
studying this problem, we show the existence of an interesting stabilization effect:

for any Lagrange gyrostat that is in its unstable equilibrium position, there is another
Lagrange gyrostat such that, when the two gyrostats are coupled to form a chain,
the rotation of the latter gyrostat can be used to stabilize the former one. A similar
stabilization effect was discovered for a system of two heavy Lagrange tops moving
about a fixed point [9, 15, 17]. This contribution extends the results of [9, 15, 17]
to a case of a system of two gyrostats moving in a non-homogeneous gravitational
field.
2 Description of the Model
We consider a mechanical system S consisting of two gyrostats G1 and G2 . Each

gyrostat Gk is composed by a rigid body Bk and other (variable or rigid) bodies B ek
which are connected to it. The motion of these bodies relative to Bk does not change
the distribution of masses of the gyrostat Gk . We assume that the body B1 is attached
to an immovable base at one of its points O1 , while the bodies B1 and B2 are coupled
at a common point O2 .¤ O1 / by an ideal spherical joint.
Let mk denote the mass of Gk . Let also B1 denote an augmented body [2]
consisting of B1 and the mass point m2 located at O2 . Similarly, the augmented
gyrostat G1 consists of G1 and the mass point m2 located at O2 . For our convenience,
we use the notation B2 D B2 and G2 D G2 . Note that the absolute angular velocities
of bodies Bk and Bk are equal, while the first-order mass momentum ak and the
tensor of inertia Ik of Gk at Ok can be expressed as
a1 D .m1 C m2 /b D m1 c1 C m2 s; a 2 D m2 c 2 ; (1)

I1 D J1 C m2 s2 ı s ˝ s ; I2 D J2 ; (2)
where b D O1 C1 ; C1 is the barycenter of G1 , ck D Ok Ck , Ck is the center of mass

of Gk ; s D O1 O2 , Jk is the tensor of inertia of Gk at point Ok , ı is the 3 3 identity
matrix, and ˝ denotes a dyadic product of two tensors.
We further assume that each gyrostat Gk has the mass distribution analogous to
the one of a Lagrange top, the attachment point(s) of Bk lie on the axis of dynamic
symmetry of Gk , and the gyrostatic moment k of Gk is constant relative to Bk and
directed along the Gk ’s symmetry axis, i.e., O1 ; O2 ; C1 ; C1 2 l1 ; O2 ; C2 2 l2 , and
k jjlk , where lk denotes the dynamic symmetry axis of Gk .
Suppose that system S moves under the gravitational attraction of a point mass
located at a fixed point P whose position with respect to O1 is defined by the vector
R where is the unit vector in the direction of O1 P and R is the distance between
O1 and P.
We introduce a Cartesian reference frame fO1 ; 1 2 3 g whose axes are fixed
.k/ .k/ .k/
in inertial space such that 3 D and a Cartesian frame fOk ; e1 e2 e3 g that is
rigidly embedded in body Bk such that its coordinate axes are the principal axes of
On Stabilization of an Unbalanced Lagrange Gyrostat 61
.k/
inertia of Gk and e3 jj lk . We also assume that, in the corresponding moving frame,
.1/ .k/ .k/
Ik D diagfIk; Ik ; Ikz g; Jk D diagfJk ; Jk ; Jkz g; s D se3 ; ak D ak e3 , and k D k e3 .
We determine the position of gyrostat Gk with respect to the reference frame by
Bryan-Krylov angles ˛k ; ˇk ; and k [11, 18]. The equations of motion of system S
in terms of the Bryan-Krylov angles can be written in the form [3]
.m/ .m/ .m/

Fk C a2 sGkj C "a2 sHkj D 0; Rk ˛R k sin ˇk ˛P k ˇPk cos ˇk D 0; (3)
where " D g=R; m D 1; 2; j; k D 1; 2 with j ¤ k;

.1/
Fk D Ik ˇRk C ˛P k2 sin ˇk cos ˇk C Ikz .Pk ˛P k sin ˇk / ˛P k cos ˇk C k ˛P k cos ˇk

Cak g cos ˛k sin ˇk C 3" Ik Ikz cos2 ˛k sin ˇk cos ˇk ;

.1/
Gkj D ˇRj cos ˇj ˇPj2 sin ˇj cos ˇk

C 2˛P j ˇPj sin ˇj ˛R j cos ˇj sin ˇk sin.˛k ˛j /

C ˇRj sin ˇj C ˛P j2 C ˇPj2 cos ˇj sin ˇk cos.˛k ˛j /;
.1/
Hkj D 2 cos ˛k cos ˛j sin ˛k sin ˛j sin ˇk cos ˇj C sin ˇj cos ˇk ;

.2/
Fk D Ik ˛R k cos ˇk 2˛P k ˇPk sin ˇk Ikz .Pk ˛P k sin ˇk / ˇPk
k ˇPk C ak g sin ˛k C 3".Ik Ikz / sin ˛k cos ˛k cos ˇk ;

.2/
Gkj D ˛R j cos ˇj C 2˛P j ˇPj sin ˇj cos.˛k ˛j /

C ˇRj sin ˇj C ˛P j2 C ˇPj2 cos ˇj sin.˛k ˛j /;
.2/
Hkj D cos ˇj 2 sin ˛k cos ˛j C sin ˛j cos ˛k :
3 The Problem of Interest
The motion of a gyrostat is a permanent rotation if, while the gyrostat is in motion,
its angular velocity vector is constant. This vector defines an axis of permanent
rotation which is fixed in both an inertial space and the gyrostat’s carrier. When the
permanent axis is a vertical line, the gyrostat permanently rotates about a vertical
axis.
In this contribution we investigate a problem of the stability of the state of motion
of system S with the following property: G1 permanently rotates about its dynamic
symmetry axis coinciding with the vertical axis passing through O1 , whereas G2 is
.2/
at rest so that e3 jj. One can check that equations (3) have a particular solution
˛k 0; ˇk 0; 1 D !t; 2 0; ! D const (4)
that describes the motion of interest; the angular velocity ! can be chosen

in (4)
.2/
arbitrarily. Below we discuss a case when the parameters c1 ; s, and c2 D c2 e3
are negative. In this case, a1 < 0 and a2 < 0.
The solution (4) is a special case of a more general solution (˛k 0; ˇk
0; k D !k t; !k D const) of equations (3) that describes permanent rotations
of system S about a vertical axis. The necessary conditions for stability of such
rotations have been recently established in [3]. For the motion of interest, these
conditions assume the form

2

22 1 3 > 0; 12 22 1 3 21 1 5 42 4 C 323 > 0; (5)
3
2
1 5 42 4 C 323 27 1 3 5 C 22 3 4 1 24 22 5 33 > 0;

where 1 D I1 I2 a22 s2 ; 2 D .I1 2 C I2 w/ =4; 3 D w2 I1 d2 I2 d1 C 2"a22 s2
=6; 4 D .wd2 C 2 d1 / =4; 5 D d1 d2 "2 a22 s2 ; w D I1z ! C 1 ; dk D 2"a2s C
ak g C 3" Ik Ikz :
While analyzing the conditions (5) in the rest of the contribution, we aim
to establish the existence of the following stabilization effect. Suppose that the
parameters of G2 are selected so that, if G2 is decoupled from the chain S and
then coupled to an immovable base at point O2 , its equilibrium position would be
unstable. (Using Lyapunov’s
indirect
method,
it is possible to show that this is the
case when 22 C 4J2 m2 c2 g C 3" J2 J2z < 0 or, by virtue of (1),

22 C 4I2 a2 g C 3" I2 I2z < 0: (6)
Special cases of (6) are given in [1, 13, 14].) We seek to justify that the parameters
of G1 can be chosen so that, when both gyrostats form chain S, the motion of S
described by (4) is stable, i.e., the permanent rotation of G1 stabilizes G2 .
4 Analysis of Necessary Stability Conditions
The conditions (5) depend on 12 parameters: angular velocity !, mass characteris-

tics Ik ; Ikz ; ak , the magnitudes of the gyrostatic moments k , parameters " and g, and
the distance s between the gyrostats’ attachment points. Recall that ak < 0; s < 0
and assume additionally that Ik < Ikz . Then, for sufficiently small ", we have dk < 0.
In order to reduce the number of parameters in the problem of study, we introduce
dimensionless parameters x; 1 ; 2 ; 3 ; 4 by the formulas:

s
jd2 j p
w D I1 x ; I2 d1 D 1 I1 d2 ; a22 s2 D 2 I1 I2 ; "I2 D 3 d2 ; 2 D 4 I2 jd2 j:
I2
(7)
The conditions (5) now deduce to the form
j > 0; j D 1; 2; 3; (8)
where
1 D 3x2 C 2 .42 1/ 4 x C 342 8 .1 2 / .1 22 3 C 1/ ;

2 D 3x4 C 4 .42 1/ 4 x3 C 2 822 C 1 42 16 .1 2 / .1 22 3 C 2 / x2

C4 .42 1/ 42 4 .1 2 / .1 C 1/ .32 1/ 422 3 4 x
C344 16 .1 2 / .1 2 22 3 C 1/ 42

C16 .1 2 /2 .1 1 /2 C 42 .1 3 / .1 3 / ;
3 D 6 x6 C 5 x5 C 4 x4 C 3 x3 C 2 x2 C 1 x C 0 :
The expressions for 6 ; 5 ; : : : ; 0 in terms of the parameters 1 ; 2 ; 3 ; 4 are given

in Appendix.
The conditions (8) form a system of three algebraic inequalities (with respect to
x) whose left-hand sides, 1 .x/; 2 .x/; and 3 .x/, are polynomials of the second,
fourth, and six order in x, respectively. The coefficients of these polynomials depend
on four parameters 1 ; 2 ; 3 , and 4 . It is quite difficult to conduct a thorough
analytical study of the conditions (8) due to their high nonlinearity and tediousness.
Therefore, we have studied them numerically, seeking to find the values of the
parameters under which the system (8) is compatible.
Note that due to the assumptions stated at the beginning of this section and (7),
the parameters 1 ; 2 are positive, while 3 is negative and j3 j is relatively small.
Moreover, 2 < 1. (Indeed, by (7), this inequality is equivalent to I1 I2 a22 s2 > 0.
In order to prove the latter, we observe that, due to (1), (2), a2 D m2 c2 ; I1 D J1 C
ms2 ; I2 D J2 . Furthermore, J2 D J2c C m2 c22 , where J2c is the central equatorial
moment ofinertia of G2. Then, I1 I2 a22 s2 D J1 C m2 s2 J2c C m2 c22 m22 c22 s2 D
J1 J2c C m2 J1 c22 C J2c s2 > 0:) Since each of the polynomials j is invariant with
respect to the simultaneous replacement of x and 4 with x and 4 , respectively,
it is enough to study only a case when 4 > 0.
We also note that the inequality (6) can be written in terms of 2 ; 3 , and 4 as
follows
4
q < 2: (9)
p p
1 23 2 I1 =I2
Since 3 is p negative and relatively small and 0 < 2 < 1, the expression
p
1 23 2 I1 =I2 is positive for all possible values of the parameters it contains
and its second term is negligible. Moreover, 4 provides an upper bound for the
left-hand side in (9). Hence, requiring
4 < 2 (10)
guarantees the fulfillment of (9).

Thus, based on the above discussion, the suitable values of 1 ; 2 ; and 4 belong
to the set D D f.1 ; 2 ; 4 /j1 > 0; 0 < 2 < 1; 0 < 4 < 2g. Setting 3 D 0:01,
we have conducted a numerical analysis of the conditions (8) for the members of
D. Our analysis reveals that there exists a simply connected region D D in the
parameter space where these conditions are fulfilled.
The existence of the interval(s) of x for which the conditions (8) are fulfilled
depends on the number of real zeros of j .x/ and their relative position along the
x-axis. Our study indicates that, when parameters 1 ; 2 ; and 4 are selected from
D , the number of zeros of 1 .x/ can vary from 0 to 2, while 2 .x/ and 3 .x/
have ether two zeros or no zeros at all. (The number of zeros of 1 .x/ is completely
determined by the sign of the expression L D .22 C 1/42 3.1 22 3 C 1/. In
more details, 1 .x/ has two zeros, when L < 0, no zeros, when L > 0, and one zero,
otherwise). We also observe that the leading coefficients of 1 .x/; 2 .x/, and 3 .x/
take on positive, positive, and negative values, respectively, for any .1 ; 2 ; 4 / 2 D.
Figures 1 and 2 demonstrate various cases of the relative position of the intervals
of positiveness of i .x/ that have been encountered in our analysis. It is interesting
to note that for any set of parameters from D , we have been able to find only one
stabilization interval Œx ; x whose endpoints are always the zeros of 3 .x/.
60
Λ(x) Λ(x)
60
40 40
20
20
x* x
–5 –4 –3 x * –2 –1 0 1 2 :
x* x*
–4 –3 –2 –1 0 x 1
–20
–40 –20
–60
–40
Fig. 1 The graphs of 1 .x/ (dashed), 2 .x/ (dotted), 3 .x/ (dash-dotted), and the interval
Œx ; x : Left: 1 D 2; 2 D 0:6; 3 D 0:01; 4 D 1:9; x D 3:6608; x D 2:4547. Right:
1 D 1; 2 D 0:5; 3 D 0:01; 4 D 1:9; x D 3:4564; x D 1:8527
3 1
Λ(x ) Λ(x )
2
0.5
1
x* x
0
–1.4 –1.2 x –1.0 –0.8 –0.6
x x *
*
–2 –1 x* 0
–0.5
–1
2 –1
Fig. 2 The graphs of 1 .x/ (dashed), 2 .x/ (dotted), 3 .x/ (dash-dotted), and the interval
Œx ; x D Œ1:0816; 0:6153 for 1 D 0:1; 2 D 0:9; 3 D 0:01, and 4 D 1:35. Left:
zoom on 2 x 1. Right: zoom on 1:5 x 0:4
Table 1 Numerical 4 1 2 Œx ; x

estimates for 1 ; 2 ; x , and
x for some values of 4 1.40 0.11 0.75 Œ0:6668; 0:6473
1.40 0.17 0.80 Œ0:7914; 0:8154
1.40 0.25 0.85 Œ0:9535; 0:9378
1.65 0.20 0.60 Œ0:9651; 0:8665
1.65 0.40 0.70 Œ1:2547; 1:1815
1.65 0.70 0.80 Œ1:4966; 1:4746
1.90 2.00 0.50 Œ2:5674; 2:5324
1.90 2.50 0.60 Œ3:2193; 2:7011
1.90 3.80 0.70 Œ3:1720; 3:0646
We have also observed that, for a fixed 4 , there are some values k such that
only the values k satisfying 0 < 1 < 1 and 2 < 2 < 1 belong to D ; moreover,
k depends on both 4 and i .i; k D 1; 2; i ¤ k/. Numerical estimates for k as
well as for the interval Œx ; x of the fulfillment of (8) are given in Table 1. When
4 is close to its upper bound (10), there is a relatively wide range of values for 1
and 2 in D . As 4 starts getting smaller, this range narrows down significantly at a
high rate. In particular, we have observed that in this case 1 tends to drop its value
significantly for the same 2 .
Thus, it is possible to select the dimensionless parameters x and 1 through 4
to achieve the desired stabilization effect. Once the mass characteristics and the
gyrostatic moment of G2 are given, and the dimensionless parameters are chosen to
comply with (8), the distance s, the angular velocity, mass parameters, and gyrostatic
moment of G1 can be determined from (7). A similar analysis can be conducted to
justify that it is possible to stabilize an unstable equilibrium position of G1 by a
permanent rotation of G2 .
5 Conclusion
In this contribution, we have conducted a dynamic analysis of a multibody chain

consisting of two Lagrange gyrostats and moving about a fixed point in a non-
homogeneous gravitational field. We have established and analyzed the conditions
for stability of the state of chain’s motion when the gyrostat connected to the fixed
point permanently rotates about a vertical axis while the other one is at rest. It
follows from our analysis that that permanent rotation of the former gyrostat can
be used to stabilize an unstable equilibrium position of the latter one.
Acknowledgements Support for this project was provided by a PSC-CUNY Award, jointly
funded by The Professional Staff Congress and The City University of New York (Award # 68091-
00 46). The work of the second and third authors was supported by the NASA New York Space
Grant CCPP Program.
Appendix
The coefficients 6 ; 5 ; : : : ; 0 of 3 in (8) admit the following representation in

terms of the dimensionless parameters 1 ; 2 ; 3 ; 4 .
6 D 42 4;

5 D 24 1 22 1 C 32 C 1 42 C 92 1 C 32 2 .21 2 3 C 2/ ;

4 D 82 32 .22 1/ C 1 .1 82 C 4/ C 1 44 2 1222 33 C 31 2 32
C4022 32 81 2 3 C 1222 3 192 32 C 12 191 2 82 3

C121 C 32 C 1 42 2722 34 C 3622 33 C 361 2 32 222 32
4012 3 C 3622 3 122 32 812 121 2 2722 402 3
C321 C 362 8;

3 D 24 12 C 1 212 2 22 32 44 C 6423 33 C 81 22 32 4022 33
C212 2 3 40122 3 C 71 2 32 C 822 32 C 13 C 712 2 C 241 2 3

C72 32 912 C 71 2 C 22 3 91 C 1 42 C 4823 34 C 61 22 33
1602333 1822 34 C 312 2 32 381 22 32 C 4823 32 C 1302233
212 2 3 C 130122 3 261 2 32 3822 32 413 C 712 2
181 22 921 2 3 C 622 3 C 72 32 C 2012 261 2 22 3 C 201

C32 4 ;

2 D 12 46 2 1212 22 3 C 401 22 32 C 1222 33 C 313 2 812 2 3

191 2 32 C 13 C 3812 2 81 2 3 C 32 32 C 1212 C 1 44

C 25624 34 641 23 33 3202334 6612 22 32 320123 32
C761 22 33 6423 33 222 34 413 2 3 C 7612 22 3 C 4012 2 32
C696122 32 C 762233 C 14 C 4013 2 212 22 10812 2 3 C 761 22 3
19212 32 6622 32 413 19212 2 10812 3 C 40232 C 10212

C401 2 42 3 41 C 1 42 C 4 6424 35 241 23 34 12824 34
7223 35 612 22 33 561 23 33 C 361 22 34 C 6424 33 C 15223 34
C13 2 32 C 4012 22 32 C 1521 23 32 C 521 22 33 562333 1222 34
C613 2 3 1412 22 3 4912 2 32 721 23 3 1921 22 32 2423 32
1422 33 14 1513 2 1212 22 C 26122 3 C 521 22 3 C 311 2 32
C4022 32 C 1213 C 3112 2 C 361 22 C 261 2 3 622 3 152 32

2212 491 2 C 62 3 C 121 C 2 1 ;

1 D 24 913 2 C 212 2 3 C 91 2 32 413 412 44 C 4812 23 32
160123 33 C 4823 34 C 613 22 3 3812 22 32 C 1301 22 33 1822 34
C314 2 1813 22 213 2 3 C 1301222 3 C 712 2 32 381 22 32
C622 33 414 2613 2 9212 2 3 261 2 32 C 2013 C 712 2 C 32 32

21 2 3 C 2012 41 42 C 4 812 23 33 321 23 34 C 823 35 C 213 22 32
3212 23 32 212 22 33 C 96123 33 C 201 22 34 3223 34 131 2 32
213 22 3 313 2 32 C 812 23 3 C 5412 22 32 321 23 32 921 22 33
C823 33 C 2022 34 314 2 C 2013 22 413 2 3 9212 22 3 1312 2 32
C541 22 32 222 33 C 414 1313 2 C 201222 C 7212 2 3 21 22 3

C222 32 413 1312 2 41 2 3 32 32 412 31 2 C 41 ;

0 D 413 46 C 2714 22 C 3613 22 3 212 22 32 C 36122 33 2722 34
C3614 2 40132 3 1212 2 32 814 1213 2 4012 2 3

C361 2 32 C 3213 812 44 C 2561224 33 5121 24 34 C 2562435
961323 32 22412 23 33 C 608123 34 28823 35 2414 22 3
2881323 3 C 1601322 32 C 60812 23 32 5612 22 33 2241 23 33
481 22 34 9623 34 C 415 2 C 14414 22 C 2414 2 3 C 20813 22 3
60132 32 76812 22 32 C 2081 22 33 C 14422 34 415 196142
4813 22 C 10413 2 3 5612 22 3 C 12412 2 32 C 1601 22 32 2422 33
C4814 C 12413 2 C 104122 3 19612 32 8813 6012 2

C241 2 3 C 42 32 C 4812 41 42 C 2561224 34 5121 24 35
C2562436 12813 23 33 5121224 33 1281223 34 C 10241 24 34
C5121 23 35 5122435 2562336 C 1614 22 32 1281323 32
C12813 22 33 C 2561224 32 C 115212 23 33 1281222 34 5121 24 33
1536123 34 C 2562434 C 51223 35 C 12814 22 3 1614 2 32
C51213 23 3 6413 22 32 15361223 32 6401222 33 C 11521 23 33
C5121 22 34 1282334 1615 2 12814 22 128142 3 25613 23
6401322 3 C 19213 2 32 C 51212 23 3 C 16321222 32 1281 23 32
640122 33 12823 33 1282234 C 1615 C 192142 C 51213 22
C12813 2 3 6401222 3 35212 2 32 641 22 32 C 12822 33 6414
352132 12812 22 C 12812 2 3 C 1281 22 3 C 1921 2 32 C 1622 32
C9613 C 19212 2 12812 3 162 32 6412 161 2 C 161:
References
1. Beletsky, V.: The Motion of a Satellite About Its Centre of Mass in a Gravitational Field.
Moscow University Press, Moscow (1975)
2. Chebanov, D.: Precessional motions of a chain of coupled gyrostats in a central Newtonian
field. In: Proceedings of the ASME 2014 International Design and Engineering Technical
Conferences (DETC2014), Buffalo (2014). Paper DETC2014-35590
3. Chebanov, D., Salas, J.: On permanent rotations of a system of two coupled gyrostats in a
central Newtonian force field. In: Proceedings of the ASME 2015 International Design and
Engineering Technical Conferences (DETC2015), Boston (2015). Paper DETC2015-47434
4. Gluhovsky, A.: Modeling turbulence by systems of coupled gyrostats. In: Fitzmaurice, N.,
Gurarie, D., McCaughan, F., Woyczynski, W. (eds.) Nonlinear Waves and Weak Turbulence.
Progress in Nonlinear Differential Equations and Their Applications, vol. 11, pp. 179–197.
Birkhäuser, Boston (1993)
5. Gluhovsky, A.: Energy-conserving and Hamiltonian low-order models in geophysical fluid
dynamics. Nonlinear Proc. Geoph. 13(2), 125–133 (2006)
6. Haken, H.: Analogy between higher instabilities in fluids and lasers. Phys. Lett. A 53(1), 77–78
(1975)
7. Hemati, N.: Strange attractors in brushless DC motors. IEEE Trans. Circuits Syst. I Fundam.
Theory Appl. 41(1), 40–45 (1994)
8. Hughes, P.: Spacecraft Attitude Dynamics. Wiley, New York (1986)
9. Lesina, M.: Stabilization of an unbalanced Lagrange gyroscope at rest. Mekh. Tverd. Tela 11,
88–92 (1979)
10. Li, Q., Wang, H.: Has Chaos implied by macrovariable equations been justified? Phys. Rev. E
58, 1191–1194 (1998)
11. Lurie, A.: Analytical Mechanics. Springer, Berlin (2002)

12. Moiseyev, N., Rumyantsev, V.: Dynamic Stability of Bodies Containing Fluid. Springer,
New York (1968)
13. Rubanovskii, V.: On bifurcation and stability of stationary motions in certain problems of
dynamics of a solid body. J. Appl. Math. Mech. 38(4), 573–584 (1974)
14. Rumyantsev, V.: On the stability of motion of gyrostats. J. Appl. Math. Mech. 25(1), 9–19
(1961)
15. Savchenko, A., Bolgrabskaya, I., Kononyhin, G.: Stability of Motion of Systems of Coupled
Rigid Bodies. Naukova Dumka, Kyiv (1991)
16. Starostin, E.: Three-dimensional shapes of looped DNA. Meccanica 31(3), 235–271 (1996)
17. Varkhalev, I., Savchenko, A., Svetlichnaya, N.: On the stabilization of an unbalanced Lagrange
gyroscope at rest. Mekh. Tverd. Tela 14, 105–109 (1982)
18. Wittenburg, J.: Dynamics of Multibody Systems. Springer, Berlin (2008)
Approximate Solution of Some Boundary Value
Problems of Coupled Thermo-Elasticity
Manana Chumburidze
Abstract We consider a non-classical model of a pseudo oscillation system of

partial differential equations of coupled thermo-elasticity in the Green-Lindsay
formulation. The matrices of fundamental and singular solutions for isotropic
homogeneous elastic materials have been obtained. We propose and justify a
technique of approximate method for the solution of boundary value problems with
mixed boundary conditions. The tools applied in this development are based on
singular integral equations, the potential method and the generalized Fourier series
analysis.
Mathematics Subject Classifications (2010) 26A33 60G22 35R60 34K37
1 Introduction
As we know, non-classical theories of thermo-elasticity have been developed in

order to remove the paradox of physically impossible phenomenon of infinite
velocity of thermal signals in the conventional coupled thermo-elasticity. Lord-
Shulman [17, 23] theory and Green-Lindsay [14] theory are important generalized
theories of thermo-elasticity that become centre of interest of recent research in this
area. In Lord-Shulman theory, a flux rate term into the Fourier’s law [2] of heat
conduction is incorporated (with one relaxation time) and formulated a generalized
theory admitting finite speed for thermal signals. Green-Lindsay theory called as
temperature rate-dependent is included among the constitutive variables with two
constants that act as two relaxation times, which does not violate the classical
Fourier law of heat conduction when the body under consideration has a center of
symmetry. There are special classes of thermo-elasticity problems such as coupled
thermo-elasticity, which require entirely different mathematical approaches and
means of analysis [11, 20–22].
M. Chumburidze ()
Ak.Tsereteli State University, Kutaisi, Georgia

72 M. Chumburidze
The dynamical problems for a conjugated system of two-dimensional linear

theory coupled thermo-elasticity (CPTE) [4, 12, 16] in the Green-Lindsay formu-
lation for isotropic homogeneous elastic materials with a center of symmetry are
investigated in work [4]. Green’s tensors for basic boundary value problems are
derived. The boundary integral method in combination with the harmonic potentials
theory and Laplace transform [3, 7, 16] are applied to solve the problems.
This paper is devoted to the development of approximate methods for the
construction of solutions in the form that admit efficient numerical evaluation [3,
6, 8, 16]. In particular, we consider two-dimensional stationary problems of CPTE
in Green-Lindsay formulation. The formulation is given for isotropic homogeneous
elastic materials with a center of symmetry. Boundary value problems for a finite
domain have been derived, when the couple-stresses components, displacement
components, rotation, heat flux and temperature are given on the surface of Holder
class [19].
Introduce the notations: R2 is two-dimensional Euclidean space, x D xj . j D
1; 2/ is point of this space. ˝ R2 is finite domain, bounded by the closed surface
@˝ of Holder class.
We have [4]:
. C ˛/ u .x; / C . C ˛/ graddivu C 2˛rotu3 gradu4

2 u D h.1/ .x/
(1)
. C ˇ/ u3 .x; / C 2˛rotu 4˛u3 I 2 u3 D h3 .x/
u4 .x; / < u4 divu D h4 .x/
where u D .u1 ; u2 / is the displacement vector, u3 is a characteristic of the rotation,

1
u4 is the temperature variation, D .1 C 2 /; D <1 .1 C 1 /, 2 > 1 > 0
<
are constants of relaxation [3], > 0; ˛ > 0; > 0; 3 C 2 > 0; >
0; ˇ > 0; > 0; < > 0; I > 0 are constants of elasticity [4, 16, 20], is two-
dimensional Laplacian operator [13], D C iq; > 0 corresponds to the general
C
dynamic problems [4], H D h.1/ ; h3 ; h4 D .h1 ; h2 ; h3 ; h4 / 2 C0;˛ .˝ /rotu3 D
T
@u3 @u3
@x2 ; @x1 ; rotu D @u @u1
@x1 @x2 :
2
Let us construct the matrix of differential operators:

ˇ ˇ
ˇ kT.@x; n.x//k3x3 k N k3x1 ˇ
P.q/ .@x; n.x// D ˇˇ ˇ ; q D 0; 3
kı4k Œ.ı1q C ı3q / C .ı0q C ı3q / @n k1x4 ˇ
@
ˇ ˇ
ˇ kıjk k3x4 ˇ
ˇ
Q.q/ .@x; n.x// D ˇ ˇ ; q D 0; 3
kı4k Œ.ı0q C ı3q / C .ı1q C ı3q / @n k1x4 ˇ
@
Approximate Solution of Some Boundary Value Problems 73
where N.x/ D .n; 0/; n D .n1 ; n2 /; T.@x; n.x// D kTjk .@x; n.x//k3x3 is the matrix of
stress operator on the plain of couple-stress elasticity:
@ @ @
Tjk .@x; n.x// D nj .x/ C . ˛/nk .x/ . C ˛/ıkj ; j; k D 1; 2
@xk @xj @nx
2
X
Tjk .@x; n.x// D 2˛ jkp np .x/; j D 1; 2; k D 3
pD1
@
Tjk .@x; n.x// D . C ˇ/ıkj ; j D 3; k D 1; 3
@nx
Differential equations seem to be well suited as models for systems. Thus an

understanding of differential equations are at least as important as an understanding
of matrix equations.
Allow us introduce the matrix L.@x; / D jLjk .@x; /j4x4 of differential operator
of pseudo oscillation:
ˇ ˇ
ˇ kL.1/ k2x2 kL.2/ k2x1 k GT .@x/k2x1 ˇˇ
ˇ
ˇ ˇ
L.@x; / D ˇ kL.3/ k1x2 L.4/ 0 ˇ
ˇ ˇ
ˇk G.@x/k1x2 0 < ˇ
where
.1/ @2
Lij .@x/ D ıij . C ˛/ 2 C . C ˛/ ; i; j D 1; 2
@xi @xj
2
X
.2/ .3/ @
Lij .@x/ D Lij .@x/ D 2˛ ijp .x/; i D 1; 2j D 3I j D 1; 2; i D 3
pD1
@xp
L.4/ .@x/ D . C ˇ/ 4˛ I 2 ; G.@x/ D .@x1 ; @x2 /
where ıij is Kronecker’s symbol, ijp is Levi-Chivita’ s symbol.

Now the system (1) can be written in the form:
L.@x; /U.x; / D H.x/ (2)
In view of this the conjugated with L.@x; / operator will be considered:

ˇ ˇ
ˇ kL.1/ k2x2 kL.2/ k2x1 k GT .@x/k2x1 ˇˇ
ˇ
b ˇ ˇ
L.@x; / D ˇ kL.3/ k1x2 L.4/ 0 ˇ
ˇ ˇ
ˇk G.@x/k1x2 0 < ˇ
74 M. Chumburidze
The matrices L.@x; / and b L.@x; / we use in technical point of view efficiently
solving the system of partial differential equations of CPTE in Green-Lindsay
formulation. In next section we will solve the matrix equations (2).
2 Fundamental and Singular Solutions
In our investigations the fundamental [16] and singular solutions together with basic
potentials becomes an useful tool for the development of approximate methods
solution of stationary problems.
Let us construct the matrix of fundamental solutions [4, 16] of the operator (2):
˚.x; / D j˚ij .x; /j4x4 D L.@x; /'.x; / (3)
where L is the associated with L.@x; / matrix:
L.@x; /L.@x; /'.x; / D L.@x; /'.x; /L.@x; / D IdetL.@x; / (4)
where I is the 4 4 dimensional unit matrix. According to (2) and (3) we have:
4
Y
detL.@x; / D . C 2/. C ˛/. C ˇ/ . C k2 /'.x; / D 0 (5)
kD1
From (5)we shall get:
4
X .1/
'.x; / D ak H0 .k jxj/ (6)
kD1
.1/
where H0 .k jxj/ is Hankel Function of the first kind (the zero order) [18], jxj D
q
x21 C x22 ; k .k D 1; 4/ are parameters of thermo-elasticity [4, 16], '.x; / is
unknown scalar, ak .k D 1; 4/ are constants, they are sought in such manner that
partial derivatives of the eight order of function '.x; / has an isolated singularity
of the kind lnjxj. If the matrix of fundamental solutions of the operator L.@x; / we
denote as b̊ .x; / by a direct check we can make sure that: b̊ .x; / D ˚ T .x; /:
Let us construct the basic potentials:
Z
V.x; '/ D ˚.y x; /'.y/dy l
l
Z
.1/
M .x; '/ D P.q/ .@y; n/ b̊ .y x; /T '.y/dy l
Œb
l
Z
M .2/ .x; '/ D Q.q/ .@y; n/ b̊ .y x; /T '.y/dy l
Œb
l
Z
U.x; '/ D ˚.y x; /'.y/dy
˝
where l 2 L2 .˛/ ; ˛ > 0; V.x; '/ is the single-layer potential, M .1/ .x; '/; M .2/ .x; '/
are the mixed-type potentials, U.x; '/ is the volume potential. Conjugated operators
b
P.q/ ; b
Q.q/ can be obtained from corresponding operators P.q/ ; Q.q/ replacing and
.
3 Approximate Solutions
The aim of this section is to obtain approximate solutions of stationary problems

of CPTE in Green-Lindsay formulation. Similar methods of these techniques which
are extended to certain classes of problems are developed in the recent papers: [3, 10,
16]. In particular, we formulate basic boundary value problems for two-dimensional
isotropic homogeneous elastic materials with a center of symmetry. It is assumed
that surfaces are sufficiently smooth.
Problem P.q/ .ı/. It is required to find regular solution U D .u; u3 ; u4 / – with
the following conditions:
8x 2 ˝ W L.@x; /U.x; / D H.x/

8z 2 @˝ W P.q/ .@z; n.z//U.z/ D F .1/ .z/
Problem Q.q/ .ı/. It is required to find regular solution U D .u; u3 ; u4 / – with

the following conditions:
8x 2 ˝ W L.@x; /U.x; / D H.x/

8z 2 @˝ W Q.q/ .@z; n.z//U.z/ D F .2/ .z/
The existence and uniqueness of this solution has been proved in [4].
Solution of the Problem P.q/ .ı/ will be found by the formula:
Z Z
1
U.x/ D ˚.y x; /H.y/dy C Œb
Q.q/ .@y; n/˚ T .y x; /T '.y/dy l (7)
2 ˝ l
where '.y/ is solution of the singular integral equations of normal type with zero
(total) index:
Z
'.z/ C P.q/ .@z; n/Œb
Q.q/ .@y; n/˚ T .y z; /T '.y/dy l D F .1/ .z/
l
Z
1
P.q/ .@z; n/˚.y z; /H.y/dy (8)
2 ˝
76 M. Chumburidze
Equation (8) is the singular integral equations with a Cauchy kernel [10] of
normal type, which have an index equal to zero (total) and in this case the Fredholm
theorems hold [4, 5, 15].
Let us represent solutions in the following form:
Z
1
U.x; / D GTP.q/ .x; yI ; ˝/H.y/dy C
2 ˝
Z
C Œb
Q.q/ .@y; n/GTP.q/ .x; yI ; ˝/T F .1/ .y/dy l (9)
l
where GP.q/ .x; yI ; ˝/ is the tensor of Green of Problem P.q/ .ı/ [4, 16].
Analogically we will get:
Z
1
U.x; / D GT .x; yI ; ˝/H.y/dy C
2 ˝ Q.q/
Z
C Œb P.q/ .@y; n/GTQ.q/ .x; yI ; ˝/T F .2/ .y/dy l (10)
l
where GQ.q/ .x; yI ; ˝/ is the tensor of Green of Problem Q.q/ .ı/

Formulas (9),(10) have an essential meaning to verify the generalized Fourier
series method [8, 16] for Problem P.q/ .ı/ and Problem Q.q/ .ı/.
Let us construct auxiliary domains and surfaces. Let us consider b̋ ˝ finite
domain,bounded by the closed surfaces of Holder class :@ b̋ be sufficiently smooth
surface – boundary of b̋ , fxk g1kD1 @
b̋ be everywhere accounted set of points.
The next theorems are proved there [3, 4, 6, 16]:
˚ 1
Theorem 1 Accounted set of the vectors P.q/ .@y; n/˚ .j/ .y xk / kD1 ; j D
1; 2; 3; 4; y 2 @˝, is linearly independent and full in the Hilbert space L2 .@˝/.
˚ 1
Theorem 2 Accounted set of the vectors Q.q/ .@y; n/˚ .j/ .y xk / kD1 ; j D
1; 2; 3; 4; y 2 @˝, is linearly independent and full in the Hilbert space [1] L2 .@˝/.
Let us prove the following theorem:
Theorem 3 For any " > 0 can be found the natural number N0 such, that when
N > N0 in any domain ˝ ˝, uniformly holds the inequality
jU.x/ U N .x/j < "; x 2 ˝;
where
X
N X
k Z
.1/ j Œ
jC3 1
U .x/ D
N
Xk ak ˚ ej .y x 4 ; i / ˚.y x; /H.y/dy
kD1 jD1
2 ˝
Proof Introduce the following notations:

kC3
k
.y/ D P.q/ .@y; n/˚ ek .y xŒ 4 /; k D 1; 1
where

k1
ek D k 4
4
Let us consider that ' k .y/ is orthonormal system of vectors j

.y/ on @˝, then
X
k
j
' k .y/ D ak j
.y/; k D 1; 1; y 2 @˝ (11)
jD1
j
where ak are coefficients of the orthonormalization.
According to (3), we have:
X
k
jC3
ak P.q/ .@y; n/˚ ej .y xŒ 4
j
' k .y/ D /; k D 1; 1; y 2 @˝ (12)
jD1
Let us consider that U.x/ be the direct solution of the Problem P.q/ .ı/
(existences this solutions are proved [4]). Let us consider the vector:
Z
1
V.x/ D U.x/ ˚.y x; /H.y/dy
2 ˝
Obviously V.x/ is the regular solution of the following boundary value problem:
8x 2 ˝ W L.@x; /V.x; / D 0; x 2 ˝
8z 2 @˝ W P.q/ .@z; n.z//V.z/ D X .1/ .z/ (13)
where
Z
1
X .1/ .z/ D F .1/ .z/ P.q/ .@z; n/˚.y z; /H.y/dy
2 ˝
According to the above X .1/ .z/ 2 C.0;˛/ .@˝/; ˛ > 0 we have:
X
n
.1/
X .1/.y/ Xk ' .k/ .y/; y 2 @˝
kD1
78 M. Chumburidze
where
Z
.1/
Xk D X .1/ .y/' .k/.y/dl
@˝
Allow us introduce the vectors:
X
N
.1/
X
N
.1/
X
k X
N X
k
.1/ j jC3
Xk ' .k/ .y/ D Xk ak ˚ ej .y xŒ 4
j
V N .x/ D Xk ak j
.y/ D ; i /
kD1 kD1 jD1 kD1 jD1
Let us show that it is approximation solution of Problem P.q/ .ı/:

Obviously V N .x/ is the regular solution of the following boundary value problem:
8x 2 ˝ W L.@x; /V N .x; / D 0; x 2 ˝
P .1/
8z 2 @˝ W P.q/ .@z; n.z//V N .z/ D NkD1 Xk ' .k/ .z/ (14)
But, according to (9) we have:

Z X
N
1 .1/
V N .x; / D Œb
Q.q/ .@y; n/GTP.q/ .x; yI ; ˝/T Xk ' .k/ .y/dy l (15)
2 l kD1
and also we have:

Z
1
V.x; / D Œb
Q.q/ .@y; n/GTP.q/ .x; yI ; ˝/T X .1/ .y/dy l (16)
2 l
Hence, from the formulas (15) and (16) and according to the assumptions of
Green tensor [4] and applying the Cauchy-Bunyakovski inequality [9], we have:
V.x/ D lim V N .x/

N!1
Analogically we can show that approximate solution of Problem Q.q/ .ı/ has
following form:
X
N X
k Z
.2/ j Œ
jC3 1
U .x/ D
N
Xk bk ˚ ej .y x 4 ; i / ˚.y x; /H.y/dy
kD1 jD1
2 ˝
where
Z
.2/
Xk D X .2/ .y/! .k/ .y/dl;
@˝
Z
1
X .2/ .z/ D F .2/ .z/ Q.q/ .@z; n/˚.y z; /H.y/dy
2 ˝
X
k
! .k/ .y/ D
j
bk j .y/;
jD1
jC3
j .y/ D Q.q/ .@y; n/˚ ej .y xŒ 4 /
4 Conclusion
Thus, two-dimensional stationary problems of CPTE in the Green-Lindsay for-

mulation have been investigated by applying the boundary integral method in
combination with the harmonic potentials theory and generalized Fourier series
analysis. Approximate solutions of boundary value problems for a finite domain
bounded by closed surface of Holder class have been constructed.
As a result, it is shown that the approximate method is reliable for obtaining
effective solutions (explicitly) of boundary value problems for isotropic homoge-
neous elastic materials with a center of symmetry.
Acknowledgements I would like to express my special gratitude to memories of my professors

Tengis Burchuladze and Davit Gelashvili.
References
1. Bachman, G., Narici, L.: Fourier and Wavelet Analysis. Universitext. Springer, Berlin/New
York (2000)
2. Bonetto, F., Lebowitz, J.L., Rey-Bellet, L.: Fourier’s Law: A Challenge to Theorists. Mathe-
matical Physics. Imperial College Press, London (2000)
3. Burchuladze, T., Gegelia, T.: Development of the Potential Method in Elasticity Theory.
Mecniereba, Tbilisi (1985)
4. Chumburidze, M.: Non-classical Models of the Some Theory of Boundary Value Problems.
LAP LAMBERT Academic Publishing, Saarbrucken (2014)
5. Chumburidze, M., Lekveishvili, D.: Effective Solution of Boundary Value Problems of the
Theory of Thermopiezoelasticity for a Half-Plane. American Institute of Physics (2013).
doi:10.1063/1.485476
6. Chumburidze, M. Lekveishvili, D.: Approximate solution of some mixed boundary value
problems of the generalized theory of couple-stress thermo-elasticity. Int. J. Math. Comput.
Nat. Phys. Eng. (2014). WASET. http://waset.org/Publication/9998377
80 M. Chumburidze
7. Chumburidze, M., Lekveishvili, D., Khurcia, Z.: Solutions Of boundary-value problems of the
generalized theory of couple-stress thermo-diffusion. J. Math. Syst. Sci. 3(7), 365–370 (2013).
David Publishing, New York
8. Constanda, C.: Generalized Fourier Series. Mathematical Methods for Elastic Plates. Springer,
London (2014)
9. Dragomir, S.: A survey on Cauchy–Bunyakovsky–Schwarz type discrete inequalities. J.
Inequal. Pure Appl. Math. (JIPAM) 4(3), 222–226 (2003)
10. Eshkuvatov, Z., Long, N.: Approximate solution of singular integral equations of the first kind
with Cauchy kernel. Appl. Math. Lett. (2009). doi:10.1016/j.aml.2008.08.001
11. Ezzat, M., Zakaria, M.: Generalized thermoelasticity with temperature dependent modulus of
elasticity under three theories. J. Appl. Math. Comput. 14, 193–212 (2004)
12. Gelashvili, D.: To Ward the Theory of Dynamic Problems of Couple-Stress Thermodiffusion
of Deformable Solid Micropolar Elastic Bodies. North-Holland, Amsterdam (1979)
13. Gilbarg, D., Trudinger, N.: Elliptic Partial Differential Equations of Second Order. Springer,
Berlin (2001)
14. Green, E., Lindsay, K.A.: Thermoelasticity. J. Elast. 2, 1–7 (1972)
15. Hakl, R., Zamora, M.: Fredholm-type theorem for boundary value problems for systems of
nonlinear functional differential equations. Bound. Value Probl. 2014(1), 113 (2014). Springer
16. Kupradze, V. et al.: Three-Dimensional Problems of the Mathematical Theory of Elasticity and
Thermo Elasticity. North-Holland, Amsterdam-New York (1983)
17. Lord, H., Shulman, Y.: A generalized dynamical theory of thermoelasticity. J. Mech. Phys.
Solids 15, 299–309 (1967)
18. Orlando, F.: Hankel Functions. Mathematical Methods for Physicists, 3rd edn. Academic,
Orlando (1985)
19. Qing, H., Fanghua, L.: Elliptic Partial Differential Equations. Courant Institute of Mathemati-
cal Science, New York (1997)
20. Rezazadeh, G., Vahdat, A.: Thermoelastic damping in a micro-beam resonator using modified
couple stress theory. Acta Mechanica 223(6), 1137–1152 (2012). Springer
21. Sherief, H., Hamza, F., Saleh, H.: The theory of generalized thermoelastic diffusion. Int. J. Eng
5, 591–608 (2004). Elsevier
22. Tripathi, J., Kedar, G., Deshmukh, K.: Dynamic problem of generalized thermoelasticity for a
semi-infinite cylinder with heat sources. J. Thermoelast. 2(1), 01–08 (2014)
23. Youssef, HM.: Theory of two-temperature-generalized thermoelasticity. J. Appl. Math. IMA
71(3), 383–390 (2006)
Symmetry-Breaking Bifurcations in Laser
Systems with All-to-All Coupling
Juancho A. Collera
Abstract We consider a system of n semiconductor lasers with all-to-all coupling

that is described using the Lang-Kobayashi rate equations. The lasers are coupled
through their optical fields with delay arising from the finite propagation time of
the light from one laser to another. As a consequence of the coupling structure, the
resulting system of delay differential equations is equivariant under the symmetry
group Sn S1 . Since symmetry gives rise to eigenvalues of higher multiplicity,
implementing a numerical bifurcation analysis to our laser system is not straightfor-
ward. Our results include the use of the equivariance property of the laser system to
find symmetric solutions, and to correctly locate steady-state and Hopf bifurcations.
Additionally, this method identifies symmetry-breaking bifurcations where new
branches of solutions emerge.
1 Introduction
Semiconductor lasers are highly sensitive to optical feedback that even a small
amount of optical feedback is enough to produce chaotic instabilities [7, 8]. In 1980,
Lang and Kobayashi [9] examined the influences of external optical feedback on
semiconductor laser properties. A single mode laser was examined where a portion
of the laser output is reflected back to the laser cavity from an external mirror. The
dimensionless form of the Lang-Kobayashi (LK) rate equations derived in [2] are
given by the following system of differential equations
P D .1 C i˛/N.t/E.t/ C ei˝ E.t /,

E.t/
(1)
P
T N.t/ D P N.t/ .1 C 2N.t//jE.t/j2 ,
J.A. Collera ()

Department of Mathematics and Computer Science, University of the Philippines Baguio, Gov.
Pack Road, Baguio City 2600, Philippines

82 J.A. Collera
where E.t/ is the complex electric field, N.t/ is the excess carrier number,
and the fixed delay time represents the external cavity roundtrip time of the
feedback light. The parameters ˛, , ˝, T and P correspond to the linewidth
enhancement factor, feedback strength, angular frequency of the solitary laser,
electron decay rate, and pump parameter, respectively. System (1) was shown to
correctly describe the dominant effects observed experimentally, see for example
[10–12].
In 2006, Erzgraber et al. [4] studied a model of two mutually delay-coupled
semiconductor lasers in a face-to-face configuration. The two lasers are coupled
through their optical field, and the finite propagation time of the light from one laser
to the other constitute the time delay. A special case with zero detuning given by the
following LK-type rate equations
EP 1 .t/ D .1 C i˛/N1 .t/E1 .t/ C eiCp E2 .t /,

EP 2 .t/ D .1 C i˛/N2 .t/E2 .t/ C eiCp E1 .t /,
(2)
T NP 1 .t/ D P N1 .t/ .1 C 2N1 .t//jE1 .t/j2 ,
T NP 2 .t/ D P N2 .t/ .1 C 2N2 .t//jE2 .t/j2 ,
where Cp is called the coupling phase parameter, was shown to have great
importance as this case organizes the dynamics for small non-zero detuning. Notice
that (2) is in fact an extension of the LK equations in (1) into the case of two
mutually delay-coupled semiconductor lasers.
In this paper, we consider a system of n semiconductor lasers with all-to-
all coupling. The lasers are coupled through their optical fields with time delay
arising from the finite propagation time of the light from one laser to another. This
generalizes the one-laser case in (1) and the zero-detuning two delay-coupled lasers
case in (2). As a consequence of the coupling structure, the resulting system of delay
differential equations (DDEs) is equivariant under a symmetry group. Symmetric
systems are known to give rise to eigenvalues with higher multiplicity [6] and this
makes numerical bifurcation analysis of such systems harder to implement. Our
results include employing a group-theoretic approach from [6] to overcome this
problem. This method not only locates steady-state and Hopf bifurcations correctly
but also identifies symmetry-breaking (SB) bifurcations where branches of new
solutions emerge.
The paper is organized as follows. In Sect. 2, we introduce our model, its
symmetry properties and basic solutions. Then, in Sect. 3 we show how to find
symmetric solutions and use numerical continuation to find branches of solutions
and their stability. In Sect. 4, we give our main result which is a method of finding
and classifying steady-state and Hopf bifurcations. We also look at the symmetry
group of bifurcating branches of solutions from SB bifurcations. Then, lastly we
give our conclusions.
Symmetry-Breaking Bifurcations in Laser Systems with All-to-All Coupling 83
2 Laser Systems with All-to-All Coupling
We now consider the n-laser system with all-to-all coupling as described by the
following LK rate equations, for j D 1; : : : ; n,
X
n
EP j .t/ D .1 C i˛/Nj .t/Ej .t/ C eiCp Ek .t /;
kD1; k¤j
(3)
T NP j .t/ D P Nj .t/ .1 C 2Nj .t//jEj .t/j2 :
To determine the symmetry group of system (3), we first define the action of the
permutation group Sn and the circle group S1 to the state variables .Ej .t/; Nj .t//,
j D 1; : : : ; n, as follows:
.Ej .t/; Nj .t// D .E1 . j/ ; N1 . j/ /; and # .Ej .t/; Nj .t// D .Ej ei# ; Nj /;
where 2 Sn and # 2 S1 . Notice that the action of Sn permutes the position of the
lasers, while S1 acts only on the optical fields Ej .t/ for all j. Moreover, observe that
if .Ej .t/; Nj .t// is a solution to (3) for all j, then .Ej .t/; Nj .t// and # .Ej .t/; Nj .t//
are also solutions to (3) for all j. Hence, system (3) is equivariant under the group
Sn S1 .
Basic solutions to (3), called compound laser modes (CLMs), are of the form
Ej .t/ D Rj ei!tCij and Nj .t/ D Nj for j D 1; : : : ; n, where !, Rj > 0, j , and Nj are
all real-valued for all j, and with 1 D 0. We refer to CLMs with Rj D R, j D 0,
and Nj D N for all j, as the fully symmetric CLMs and are given by
Ej .t/ D Rei!t ; and Nj .t/ D N; (4)
for j D 1; : : : ; n. These fully symmetric CLMs are fixed by elements .; 0/ of Sn S1
for all 2 Sn . We use the symbol S0n to denote this symmetry group of CLMs in (4).
Note that this subgroup S0n of Sn S1 is isomorphic to Sn and is generated by all
elements of the form .; 0/ where 2 Sn . We call such symmetry group of CLMs
as isotropy subgroup [5].
3 Symmetric CLMs
In this section, we provide a method of finding symmetric CLMs, that is, CLMs
fixed by an isotropy subgroup of Sn S1 . In particular, we use the case when the
isotropy subgroup is S0n to find the fully symmetric CLMs in (4).
84 J.A. Collera
Substituting the ansatz in (4) to system (3) and then following a similar
computation in [4], we obtain the following transcendental equation in !
p
! C .n 1/ 1 C ˛ 2 sin.Cp C ! C tan1 ˛/ D 0: (5)
Once a value for ! is found by solving (5), we can thenp compute for corresponding
values of N D !=.˛ C tan.Cp C ! // and R D .P N/=.1 C 2N/. Fully
symmetric CLMs in (4) are obtained using these values of R, !, and N. It is
worth noting that the isotropy subgroup of the CLMs that we seek provides
relations amongst Rj , Nj , and j , and this to some extent simplifies the form of the
transcendental equation, such as in (5), which is key in finding symmetric CLMs.
A branch of CLMs can be obtained through numerical continuation by varying
a single parameter. We use DDE-Biftool [3] to obtain such branch of solutions. We
employ the same technique as in [4] to follow a CLM as an equilibrium in DDE-
Biftool. We choose to vary the coupling phase parameter Cp for two reasons. First,
because system (3) has a 2-translational symmetry in Cp , and secondly, because
the coupling phase can be changed accurately in experiments [1].
Example 1 Consider system (3) with n D 4, and parameters ˛ D 2:5, T D 392,
P D 0:3, D 20, D 0:1, and Cp D 10. Solving for ! in (5) gives 11 CLMs which
are shown in the left panel of Fig. 1 as dots. By following any of these CLMs in
DDE-Biftool, a branch of fully symmetric CLMs is obtained which is the ellipse in
the left panel of Fig. 1. The stability of this branch is also determined using DDE-
Biftool. The right panel of Fig. 1 shows the stable and unstable parts of the branch
in dashed line and solid line, respectively.
0.2 0.2
N 0 N 0
−0.2 −0.2
−0.8 −0.4 0 0.4 0.8 −0.8 −0.4 0 0.4 0.8
ω ω
Fig. 1 (Left) A branch of fully symmetric CLMs obtained in DDE-Biftool by varying the coupling
phase parameter Cp . (Right) Stable part of the branch is shown in dashed line while unstable part
of the branch is shown in solid line. Regular steady-state and Hopf bifurcations are marked with
(˘) and (ı), respectively. For SB bifurcations, we use (4) and () to marked pitchfork and SB
Hopf bifurcations, respectively
Symmetric systems, such as (3), are known to have eigenvalues of higher

multiplicity. Hence, implementing a numerical bifurcation analysis on such system
is not straight-forward. In the section that follows, we develop a method to locate
steady-state and Hopf bifurcations along a branch of solutions and classify these
bifurcations into regular and symmetry-breaking.
4 Symmetry-Breaking Bifurcations
We now give our main result which is a method of finding and classifying steady-
state and Hopf bifurcations.
We first determine the linearized system corresponding to (3) around the fully
symmetric CLM (4). We follow a similar computation done in [13] for the one-laser
model in (1). In polar form, the LK rate equations in (3) are given by
X
n

RP j .t/ D Nj .t/Rj .t/ C Rk .t / cos Cp C 'k .t / 'j .t/ ,
kD1; k¤j
X
n
Rk .t /
'Pj .t/ D ˛Nj .t/ C sin Cp C 'k .t / 'j .t/ , (6)
Rj .t/
kD1; k¤j
1 h ˇ ˇ2 i
NP j .t/ D P Nj .t/ .1 C 2Nj .t// ˇRj .t/ˇ ;
T
T
for j D 1; : : : ; n. If we let Xj .t/ D Rj .t/; 'j .t/; Nj .t/ and Yj .t/ D Xj .t /, then (6)
can be written in the form XP j .t/ D f .Xj .t/; Y1 .t/; : : : ; Yj1 .t/; YjC1 .t/; : : : ; Yn .t//; for
j D 1; : : : ; n. Now, let X.t/ D ŒX1 .t/; : : : ; Xn .t/T and Y.t/ D ŒY1 .t/; : : : ; Yn .t/T so
that the full system can be written as X.t/ P D F.X.t/; Y.t//. The fully symmetric
CLM in (4) written in polar form is given by X D ŒX1 .t/; : : : ; Xn .t/T where
Xj .t/ D ŒR; !t; NT for all j. To obtain the linear variational equation around X , we
first compute for A WD dXj .t/ f .Xj / and B WD dYj .t/ f .Xj /. Now, let M1 be the block-
diagonal matrix with the block A on the main diagonal and the block 0 elsewhere,
and let M2 be the block matrix with the block 0 on the main diagonal and the block
B on all off main diagonal entries. Then, the Jacobian matrix evaluated at X is
dF.X / D ŒM1 j M2 . The linear variational equation around the fully symmetric
CLM is given by X.t/ P D M1 X.t/ C M2 X.t /, and its corresponding characteristic
equation is det ./ D 0 where ./ D In M1 e M2 . Notice that if we let
A WD I3 A and B WD e B, then
2 3
A B B
6B A B7
6 7
L WD ./ D 6 : : : 7
4 :: : : :: 5
B B A
where blocks A and B are 3 3 matrices.

86 J.A. Collera
2
We now use an method from [6] to examine the eigenvalues o of L. Let D e n i
k 2k .n1/k T
3
and define Vk D v; v; v; : : : ; v j v 2 R , for k D 0; 1; 2; : : : ; n 1.
˚
Observe that V0 D Œv; v; v; : : : ; v j v 2 R3 and the action of L on V0 is given
T
by L Œv; : : : ; vT D Œ.A C .n 1/B/v; : : : ; .A C .n 1/B/vT : This means that the

eigenvalues of LjV0 are those of A C .n 1/B.
In general, the eigenvalues of LjVk ,
Pn1 jk
for k D 0; 1; 2; : : : ; n 1, are those of A C jD1 B: When n is odd and
k ¤ 0, the complex numbers k ; 2k ; : : : ; .n1/k can be grouped into pairs of
conjugates. Observe that jk C .nj/k D 2 cos.2kj=n/ since jk and .nj/k are
Pn1 jk P.n1/=2
conjugates. Hence, jD1 D jD1 2 cos.2kj=n/ D 1 using the Dirichlet
PN
kernel identity 1 C jD1 2 cos.j/ D sin.N C 12 /= sin. 21 /: When n is even and
k ¤ 0, we have n=2 D 1. Consequently,
2 1
n
X
n1 X sin.k k=n/
D .1/ C
jk k
2 cos.2kj=n/ D .1/k C 1
jD1 jD1
sin.k=n/
by pairing up conjugates, and then using the Dirichlet kernel identity. Now, since
sin.k k=n/ D .1/kC1 sin.k=n/ and sin.k=n/ ¤ 0 for 1 k n 1, we get
P n1 jk
jD1 D 1. This is essentially the same for the case when n is odd. Therefore,
the eigenvalues of LjVk , for k D 1; 2; : : : ; n1, are those of AB. This means that the
problem of solving the characteristic equation det ./ D 0, reduces to solving the
equations det.A C .n 1/B/ D 0 and det.A B/ D 0. Furthermore, the eigenvalues
of L from A C .n 1/B are simple while those from A B are of multiplicity n 1.
Also, notice that the symmetry group S0n acts trivially on V0 while its action on Vk ,
for k D 1; 2; : : : ; n 1, is non-trivial. This means that SB bifurcations are obtained
from block AB while bifurcations from block AC.n1/B are regular bifurcations.
We now illustrate the above technique in finding ordinary and SB bifurcations.
Example 2 Continuing from Example 1, we first look for SB bifurcations. Pitchfork
bifurcations along the branch of fully symmetric CLMs are found by looking at the
intersections of the curves det.A B/jD0 D 0 and (5), while SB Hopf bifurcations
are found by finding the intersections of the curves det.A B/jDiˇ D 0 with ˇ >
0, and (5). Two pitchfork bifurcations and two SB Hopf bifurcations were found
and are shown in the right panel of Fig. 1 with markers (4) and (), respectively.
Similarly, we can get the regular bifurcations from the block A C 3B. Saddle-node
bifurcations are obtained from the intersections of the curves det.A C 3B/jD0 D
0 and (5), while regular Hopf bifurcations were found by intersecting the curves
det.A C 3B/jDiˇ D 0 with ˇ > 0, and (5). Two saddle-node bifurcations and six
regular Hopf bifurcations were obtained and are shown in the right panel of Fig. 1
using markers (˘) and (ı), respectively.
Identification of SB bifurcations is important because they give rise to new
branches of solutions. We now examine the symmetry group of branches of
solutions that emerge from the SB bifurcations obtained in Example 2. To do this,
0.67
Re(E1(t))
0.66
0.2 0 0.25 0.5 0.75 1
0.67
Re(E2(t))
0.66
N 0 0 0.25 0.5 0.75 1
0.67
Re(E3(t))
0.66
−0.2 1 0.25 0.5 0.75 1
0.67
Re(E4(t))
0.66
−0.8 −0.4 0 0.4 0.8
ω
0 0.25 0.5 0.75 1
t
Fig. 2 (Left) A new branch of CLMs shown in dotted curve emanates from the pitchfork
bifurcations (4) and has symmetry group S03 . (Right) Time-series plots showing that a branch
of periodic solutions emerging from the SB Hopf bifurcations (ı) has spatio-temporal symmetry
Z2 .; /
we use DDE-Biftool to perform numerical continuation, again varying the coupling

phase parameter Cp . A new branch of symmetric CLMs shown as the dotted curve
in the left panel of Fig. 2 is obtained. This new branch of CLMs emanates from the
pitchfork bifurcations (4) and has a smaller symmetry group S03 compared to the
symmetry group S04 of the fully symmetric CLMs.
Similarly, a branch of periodic solutions is obtained by following a SB Hopf
bifurcation () in DDE-Biftool. This branch of periodic solutions has isotropy
subgroup Z2 .; /, which is a subgroup of S4 S1 that is isomorphic to Z2 and
is generated by the order-two element .; /. Here, is an element of S4 whose
action interchanges lasers 1 and 2, and lasers 3 and 4. Hence, the element .; /
interchanges lasers 1 and 2 (respectively, lasers 3 and 4) and then shifts the phase
of each laser by half a period. The right panel of Fig. 2 shows a time series of the
real part of the Ej .t/ for j D 1; 2; 3; 4 where the interval for t is one period. Notice
that interchanging the curves for Re E1 .t/ and Re E2 .t/ (respectively, Re E3 .t/ and
Re E4 .t/) and then shifting them by half a period preserves the plots. Indeed, the
bifurcating branch of periodic solutions emanating from the SB Hopf bifurcations
has isotropy subgroup Z2 .; /.
5 Conclusion
The main contribution of this research is the identification of symmetry-breaking

bifurcations in symmetric systems. The n-laser system (3) with all-to-all coupling
is an example of a symmetric system. Symmetry gives rise to eigenvalues of higher
multiplicity and this makes numerical bifurcation analysis of such system harder
to implement. However, we showed that symmetry actually aids us in finding sym-
metric solutions in Example 1, and in locating steady-state and Hopf bifurcations
88 J.A. Collera
along the branch of fully symmetric CLMs as illustrated in Example 2. The action
of the symmetry group decomposes the physical space into invariant subspaces.
This, in turn, gives us a means to identify symmetry-breaking bifurcations where
new branches of solutions emerge.
Acknowledgements This work was funded by the UP System Emerging Interdisciplinary

Research Program (OVPAA-EIDR-C03-011). The author also acknowledged the support of
the UP Baguio through RLCs during the A.Y. 2014–2015.
References
1. Agrawal, G.P., Dutta, N.K.: Long-Wavelength Semiconductor Lasers. Van Nostrand Reinhold,
New York (1986)
2. Alsing, P.M., Kovanis, V., Gavrielides, A., Erneux, T.: Lang and Kobayashi phase equation.
Phys. Rev. A 53, 4429–4434 (1996)
3. Engelborghs, K., Luzyanina, T., Samaey, G.: DDE-BIFTOOL v. 2.00: A MATLAB package
for bifurcation analysis of delay differential equations. Technical report TW-330. Department
of Computer Science, K.U. Leuven, Leuven (2001)
4. Erzgraber, H., Krauskopf, B., Lenstra, D.: Compound laser modes of mutually delay-coupled
lasers. SIAM J. Appl. Dyn. Syst. 5, 30–65 (2006)
5. Golubitsky, M., Stewart, I.: The Symmetry Perspective: From Equilibrium to Chaos in Phase
Space and Physical Space. Birkhauser Verlag, Basel (2002)
6. Golubitsky, M., Stewart, I., Schaeffer, D.G.: Singularities and Groups in Bifurcation Theory,
vol. II. Springer, New York (1988)
7. Gray, G.R., Ryan, A.T., Agrawal, G.P., Gage, E.C.: Control of optical-feedback-induced laser
intensity noise in optical data recording. Opt. Eng. 32, 739–745 (1993)
8. Gray, G.R., Ryan, A.T., Agrawal, G.P., Gage, E.C.: Optical-feedback-induced chaos and its
control in semiconductor lasers. In: SPIE’s 1993 International Symposium on Optics, Imaging,
and Instrumentation, San Diego, pp. 45–57. International Society for Optics and Photonics
(1993)
9. Lang, R., Kobayashi, K.: External optical feedback effects on semiconductor injection laser
properties. IEEE J. Quantum Electron 16, 347–355 (1980)
10. Mork, J., Tromborg, B., Mark, J.: Chaos in semiconductor lasers with optical feedback: theory
and experiment. IEEE J. Quantum Electron 28, 93–108 (1992)
11. Sano, T.: Antimode dynamics and chaotic itinerancy in the coherence collapse of semiconduc-
tor lasers with optical feedback. Phys. Rev. A 50, 2719–2726 (1994)
12. van Tartwijk, G.H.M., Levine, A.M., Lenstra, D.: Sisyphus effect in semiconductor lasers with
optical feedback. IEEE J. Select. Top. Quantum Electron. 1, 466–472 (1995)
13. Verduyn Lunel, S.M., Krauskopf, B.: The mathematics of delay equations with an application
to the Lang-Kobayashi equation. In: Krauskopf, B., Lenstra, D. (eds.) Fundamental Issues
of Nonlinear Laser Dynamics. AIP Conference Proceedings, vol. 548. American Institute of
Physics, New York (2000)
Effect of Jet Impingement on Nano-aerosol Soot
Formation in a Paraffin-Oil Flame
Masoud Darbandi, Majid Ghafourizadeh, and Mahmud Ashrafizaadeh
Abstract In this paper, the effects of mico-jet impingement on the formation of

soot nano-particles, CO, CO2 , and C6 H6 species in a turbulent paraffin-oil flame
are investigated numerically. In this regard, we use a two-equation - turbulence
model, a PAH-inception two-equation soot model, a detailed chemical kinetic
consisting of 121 species and 2613 elementary reactions, and steady flamelet
combustion model. We take into account the turbulence-chemistry interaction by
using presumed-shape probability density functions PDFs. We also take into account
the radiation heat transfer of soot and gases assuming optically-thin flame. In the
first place, we solve a documented experimental test case and compare the flame
structure to evaluate our numerical results. Then, we embed a micro-scale injector at
the burner wall, split the incoming air-flow between primary and secondary streams,
inject the secondary air into the burner via the embeded injector, and compare
the results for different values of impinging-jet mass flow rate. Our results show
that the mico-jet impingement affects the reactive flow behavior within the burner
and results in a compact-flame near the fuel-injector nozzle. Our calculations also
show that the emission of CO2 is slightly affected by the mico-jet impingement.
They show that increasing the mass flow rate of the micro-impinging-jet effectively
reduces the formation and emission of the soot nano-aerosol and CO, C6 H6
pollutants. It is also found that mico-jet impingement leads to a more uniform
profiles of temperature, soot volume fraction, and mass fractions of CO, CO2 , and
C6 H6 at the burner outlet.
M. Darbandi ()
Department of Aerospace Engineering, Center of Excellence in Aerospace Systems, Institute
for Nanoscience and Nanotechnology, Sharif University of Technology, Tehran, P. O. Box
11365-8639, Iran
M. Ghafourizadeh
Department of Aerospace Engineering, Center of Excellence in Aerospace Systems, Sharif
University of Technology, Tehran, P. O. Box 11365-8639, Iran
M. Ashrafizaadeh
Department of Mechanical Engineering, Isfahan University of Technology, Isfahan, P. O. Box
84156-83111, Iran

90 M. Darbandi et al.
1 Introduction
There are several types of flame holders to maintain the combustion process inside
a burner. Combustion can be stabilized using (1) an external energy supply such
as heated plate, torch, highly reactive chemicals, lasers, electric arcs, chemical
additives, (2) bluff bodies such as V-gutters, cylinders, spheres, flat plates, (3) steps
such as rearward facing steps, forward facing steps, side dumps, cavities, (4) flow
instabilities such as swirls, cyclones, (5) jets such as reverse flow jets, opposed jets,
and fuel jet blockages [1]. The choice of flame holding method depends on the
burner size, weight, pressure loss, flammability limits, material temperature, and
structural limitations. Most of these methods stabilize flames through establishing
recirculation zones which carry high-temperature products upstream to separate
flow region for igniting the incoming flows. Among the above mentioned methods,
the opposed jet, i.e. an impinging jet, is an interesting flame-holding method
regarding the limitations. So, a deep understanding on the behavior of impinging
jets would help to keep a stable combustion in combustors burning reactants and
forming products.
Many researchers have studied the formation of CO and CO2 pollutants in
combustion processes while nano-particulate soot, incorporated with hazardous
aromatics, i.e. C6 H6 , needs special attention and considerations. Even a low
emission of these pollutants can cause serious health problems, i.e. different organs
cancer. On the other hand, C6 H6 has been widely used as an important component
of gasoline to enhance the octane number of fuels. So, this aromatic compound
requires to be understood very well due to vast applications in engines as well as its
carcinogenicity.
Literature shows aerosol modeling of soot formation in laminar flames has
been studied in last decades [2, 3]. Few researchers also studied soot formation in
turbulent flames fueled by simple hydrocarbons [4, 5]. There is a lack of resources
studying the formation of nano-particulate soot aerosol in turbulent flames burning
common fuels such as paraffin-oil, gasoline, jet fuel, or diesel. Brooks and Moss
[4] proposed soot acetylene-inception model and showed that soot formation in
simple fuels is limited by inception rate of particles from single-ring aromatics
formed from acetylene. However, soot formation in practical fuels, which have
higher amounts of carbon atoms as well as aromatics compounds, is limited by
growth rate of aromatics. In this regard, Hall et al. [6] proposed a soot model,
i.e. soot PAH-inception model, with inception rate based on two and three-ringed
aromatics formed from single-ring aromatic species, i.e. C6 H6 and C6 H5 radical.
Back to our past publications, we have already simulated the nano-aerosol soot
formation in turbulent non-premixed flames fed with methane [7, 8], ethylene
[9, 10], and propane [11, 12]. In this paper, we use and extend our previous studies
and simulate the nano-aerosol soot formation in a turbulent non-premixed flame
burning paraffin-oil as its fuel. In this regard, we employ a two-equation soot
model to solve for the soot mass fraction and the number density considering
Effect of Jet Impingement on Nano-aerosol Soot Formation in a Paraffin-Oil Flame 91
the soot formation and its oxidation based on PAHs and OH agents, respectively.
We utilize a steady flamelet model as our combustion model considering a large-
detailed kinetic reaction mechanism with 2613 reversible chemical reactions and
121 chemical species. We use a two-equation - turbulence model with round-jet
corrections and take into account the turbulence-chemistry interaction using some
presumed-shape PDFs. We also take into account the radiation heat transfer of the
soot and gases assuming an optically-thin flame and calculate their radiations locally
only by emissions. To evaluate our numerical solution, we simulate a benchmark
turbulent paraffin-oil non-premixed flame inside a burner and compare the obtained
results with those of experiment as well as another numerical study. The obtained
results indicate our numerical simulation can predict soot volume fraction, mixture
fraction, and temperature distributions of the flame. Then, we embed a micro-scale
injector at burner wall, split incoming air-flow between primary- and secondary-air
streams and inject the secondary-air into burner via the micro-scale injector and
compare the results. We also study mass-flow-rate effect of the micro-scale injector
on nano-aerosol soot formation, the emissions of CO, CO2 , and C6 H6 from turbulent
non-premixed paraffin-oil flame inside burner.
2 The Governing Equations
In the cylindrical coordinates, i.e. r, z, the fluid flow conservation laws consisting of
continuity, r-momentum, and z-momentum are given by
u
r .V/ C D0 (1)
r
@p u e @u
r .Vu/ D C r .e r u/ e 2 C (2)
@r r r @r
@p e @v
r .Vv/ D C r .e r v/ C g (3)
@z r @r
The radial and axial components of the velocity vector are u and v, respectively. The
mixture density, velocity vector, pressure, and effective viscosity are represented by
, V, p, and e , respectively. The transport equations for turbulence quantities, i.e.
turbulence kinetic energy and its dissipation rate , are given by

e e @
r .V/ D r r C C G (4)
r @r

e e @
r .V/ D r r C C .c1 G c2 / (5)
r @r
In a confined jet, the turbulence model constants for Eqs. (4) and (5) are taken from
Ref. [13]. To model combustion in a turbulent diffusion flame, we use the steady
flamelet model. In this study, we choose a detailed kinetic scheme, i.e., 121 chemical
species and 2613 chemical reactions, to perform our simulations. The transport
equations for the first two moments of mixture fraction, i.e., f and f 002 , are given by

e e @f
r .Vf / D r rf C (6)
f f r @r

e 002 e @f 002
r Vf 002 D r rf C C cg e .r f /2 c f 002 (7)
f f r @r
As is known, the results from pre-computed laminar flamelets and turbulent

statistics can be tabulated as a 3D lookup table in terms of mixture fraction,
mixture fraction variance, and the scalar dissipation rate. If so, all thermo-chemical
quantities in the solution domain can be obtained from this 3D lookup table [14].
To simulate the aerosol dynamics, we need to solve two transport equations for
the soot mass fraction m and the soot number density n [7]. These equations are
given by

e e @m
r .Vm / D r r m C C S m (8)
soot soot r @r

e e @n
r .Vn / D r r n C C Sn (9)
nuc nuc r @r
Assuming a unit Lewis number, the total enthalpy equation h is given by

e e @h
r .Vh/ D r rh C C qrad (10)
h h r @r
The radiation source term in the energy conservation law, can be determined locally
only by emission assuming an opticallyP thin flame. Finally, the density is obtained
from the equation of state as p D RT nmD1 Ym =Wm where m counts the number
of chemical species in the mixture from 1 to n total number of species. The
temperature, gas constant, mass fraction, and molecular weight are represented by
T, R, Y, and W, respectively.
3 Computational Method
Back to our past experiences [7–12], we choose a hybrid finite-volume-element

FVE discipline and break the solution domain into a large number of quadrilateral
elements. We use our past experiences in treating non-staggered (collocated) grid
arrangement and implement finite element shape functions and physical influence
upwind scheme PIS for the diffusion and convection terms, respectively. Using PIS
scheme we also handle the pressure-velocity coupling effectively.
4 The Benchmark Test Case and Validation
A paraffin-oil turbulent non-premixed flame is chosen to verify our numerical

solutions. We employ the experimental conditions of Young et al. [15] to perform
our simulations. Figure 1 shows the configuration of the burner consuming paraffin-
oil. Because of the symmetry of problem, we consider a rectangular solution domain
applying the symmetry boundary conditions at the center line. The computational
domain has 0:0775 0:6 m dimensions, i.e. R0 D 0:0775 m and L D 0:6 m, see
Fig. 1. The fuel nozzle diameter is 1.5 mm, i.e. R1 D 0:75 mm. This fuel nozzle
injects the gaseous paraffin-oil, which consists of 80 % n-decane, i.e. C10 H22 , and
20 % toluene, i.e. C7 H8 , as fuel at a speed of 22.28 m/s into the burner. The oxidizer,
i.e. co-flow air stream, which consists of 23.3 % oxygen and 76.7 % nitrogen, enters
the burner at a speed of 0.234 m/s. The initial temperatures of paraffin-oil and air
are 598 K and 288 K, respectively. As understood, the paraffin-oil is evaporated and
delivered through a heated line and injected into the burner. The turbulence intensity
and eddy length scale for both the paraffin-oil and air streams are 3 % and 0.02 m at
their nozzle exits, respectively.
To evaluate the accuracy of our numerical solution in simulating turbulent
reacting flow and aerosol modeling of soot nano-particles, we solve the test case
given in Young et al. [15] and compare the predicted flame structure, i.e. the
distributions of mixture fraction, temperature, and species concentrations, with the
data collected by this reference as well as another available numerical solution [16].
Wen et al. [16] numerically solved this test case considering a detailed chemical
kinetic mechanism with 141 species and 1015 elemental reactions to construct
the flamelet library. Figure 2 presents the axial distributions of mixture fraction,
temperature, and soot volume fraction at centerline, i.e. r D 0. It also shows the
radial distributions of mixture fraction, temperature, and soot volume fraction at
z D 0:3 m above the burner. The figure shows that there are good agreements with
the data reported by Young et al. [15]. Considering the experimental data [15], our
results are more satisfactory than those of obtained by Wen et al. [16] in the context
of same assumptions and models for turbulence, combustion, radiation, and soot.
Our numerical results predict the flame length of 0.3 m, i.e. the stoichiometric value
of mixture fraction for n-decane, i.e. fst D 0:0615, which also corresponds to the
peak of axial temperature distribution. As seen in Fig. 2, in far downstream of burner
exit, the mixture fraction at centerline is lower than its stiochiometric value, i.e. the
fuel-lean region, which indicates the fuel is fully depleted and that the burner is long
enough to result in a full burn of paraffin-oil injected into the burner. As seen, the
peak of axial temperature distribution is lower than the adiabatic flame temperature
of paraffin-oil/air flame, i.e. 2366 K. There is a discrepancy between our results with
Fig. 1 The configuration of

the burner consuming
paraffin-oil [15]
the measured data. It can be attributed to (1) the relatively simple models used in our
calculations, e.g., the turbulence model, the radiation model, and the soot model, (2)
the assumptions made to simplify the problem, e.g., the gas-phase nucleation and
the free-molecular-regime coagulation assumptions used in soot modeling, and the
optically-thin flame assumption used in calculating the radiative heat transfer rate,
which took into account only the most radiating species. Evidently, the use of more
sophisticated radiation and soot models would help to overcome such shortcomings
and predict the soot characteristics and flame structure more accurately, which is
beyond the scope of this paper.
(a) r =0 (b) r =0 (c) r =0
(d) z =0.3 m (e) z =0.3 m (f) z =0.3 m
Fig. 2 The current axial and radial distributions of mixture fraction, temperature, and soot volume
fraction in the flame and comparison with the experimental data [15] and numerical solution [16];
(a) r D 0, (b) r D 0, (c) r D 0, (d) z D 0:3 m, (e) z D 0:3 m, and (f) z D 0:3 m
5 The Results and Discussion
In this section, we study the effect of mico-jet impingement on the reactive flow
behavior and the resulting emissions of CO, CO2 , C6 H6 species and soot nano-
aerosol using the current developed numerical method. We first study the effect of
mico-jet impingement on the above parameters comparing with the benchmark test
case. So, we embed a micro-scale injector at the burner wall, split the incoming air-
flow between the primary- and secondary-air streams and inject the primary air into
the burner via the embedded micro-scale injector. The dimension of micro-scale
injector is 100 m embedded at a height of z D 0:1 m above the fuel nozzle exit.
We inject a mass flow rate of 2.7 g/s of the total incoming-air via the micro-scale
injector while the rest of incoming air would enter the burner as primary-air. The
other parameters, i.e. the geometry and boundary conditions (BCs), are similar to
the benchmark test case.
Figure 3 shows the effect of mico-jet impingement on the distributions of stream
function, temperature, OH and C6 H6 mass fractions in the turbulent paraffin-oil
flame. In each subfigure, the left part depicts the distributions for benchmark
test case, i.e., without any mico-jet impingement, while the right part depicts the
distribution for the burner incorporated with the mico-jet impingement, which
(a) (b) (c) (d)
Fig. 3 The effect of mico-jet impingement on the distributions of (a) stream function (kg/s), (b)
temperature (K), (c) OH mass fraction, and (d) C6 H6 mass fraction in the turbulent paraffin-oil
flame
injects a mass flow rate of 2.7 g/s via the embedded injector inserted at the
burner wall. As seen, the mico-jet impingement causes two recirculation zones
upstream and downstream of the micro-injector, which affect the flame envelop.
The upstream recirculation, which is due to the injected air micro-jet, would cause
the entrance of fresh air to the flame and this results in a well mixing of that
with the gaseous high-temperature products. The downstream recirculation, which
is due to the injected air micro-jet, would cause the return of high-temperature
products at regions downstream of the burner, which is eventually mixed with
the low-temperature gaseous reactants, i.e. the exhaust gas recirculation EGR.
Such enhancement in mixing performance due to the appeared recirculation zones,
which are formed upstream and downstream of the micro-impinging jet, would
result in high-temperature combustion products at the burner outlet and a boost
up in the combustion efficiency. The mixing enhancement also leads to a uniform
temperature of exhaust gases at the burner outlet. So, the mico-jet impingement
improves the combustion efficiency via enhancing the existing mixing. As seen, the
mico-jet impingement also results in a compact distribution of OH mass fraction
near the fuel injector exit. There are several definitions for the flame length. In
the numerical simulations, this parameter is defined as the location of maximum
OH concentration, the stoichiometric line, or the location of maximum temperature
gradient. Considering the aforementioned definition, it is observed that the mico-jet
impingement reduces the flame length and results in a compact flame. The compact
flame together with the resulting EGR phenomenon would cause the by-products to
be fully-burnt at the regions near the fuel-injector exit. So, C6 H6 species would be
(a) (b) (c) (d)
Fig. 4 The effect of mico-jet impingement on the distributions of (a) CO mass fraction, (b) CO2
mass fraction, (c) soot volume fraction, and (d) soot particles diameter (m) in the turbulent paraffin-
oil flame
formed within the burner at regions near fuel-injector exit and so their emissions
would be reduced at the outlet, i.e., a very low concentrations release into the
surrounding ambient.
As seen in Fig. 4, the EGR, due to the appeared two recirculation zones of micro-
impinging-jet, would cause the rebrurn of combustion by-products, i.e. the CO, and
the combustion products, i.e. CO2 , at the regions near the fuel-injector exit. So, the
mico-jet impingement would reduce the emission of CO at the outlet. As seen, the
appeared recirculation zones of micro-impinging-jet would result in a well mixing
of combustion products with the unburnt reactants, which in turn results in a uniform
distribution of combustion products, i.e. the CO2 , within the burner and at the outlet.
As seen also, the nano-aerosol soot formation within the burner is reduced as a result
of EGR, i.e. due to the appearance of two recirculation zones. The soot depletion
within the burner would result in low concentrations of the soot in the exhaust gases
released into the surrounding atmosphere.
Now, we summarize, compare, and study the effect of different values of micro-
impinging-jet mass flow rate on the exhaust-gases temperature and the emissions
of CO, CO2 , C6 H6 , and the soot nano-aerosol released from the current turbulent
paraffin-oil flame. In this regard, we consider different values of micro-impinging-
jet mass flow rate, i.e. 1.8, 2.7, 3.2, and 3.6 g/s. The other parameters, i.e. the
geometry and BCs, are similar to the benchmark test case. As observed in Table 1,
the emissions of CO, C6 H6 , and the soot nano-aerosol would decrease as the micro-
impinging-jet mass flow rate increases. It is also observed that the exhaust-gases
temperature would increase as the micro-impinging-jet mass flow rate increases.
Table 1 The mass-flow-rate effect of mico-jet impingement on temperature and emissions of

pollutants at the burner outlet
Jet mass flow Temperature (K) CO mass CO2 mass Soot volume Soot particles C6 H6 mass
rate (g/s) fraction fraction fraction diameter (m) fraction
– 767 3.60E06 0.0576 1.13E06 1.20E07 5.84E11
1.8 774 3.73E06 0.0573 9.33E07 1.10E07 9.04E11
2.7 747 0 0.0579 5.46E07 9.37E08 0
3.2 787 0 0.0579 3.38E07 7.63E08 0
3.6 807 0 0.0578 1.90E07 5.97E08 0
This indicates the role of mico-jet impingement in the amount of combustion

efficiency enhancement.
6 Conclusion
Using a numerical study, we have investigated the effect of mico-jet impingement

on the formation of soot nano-aerosol and the emissions of CO, CO2 , C6 H6 from
a turbulent paraffin-oil flame. We have first evaluated our numerical simulation by
simulating a benchmark turbulent paraffin-oil non-premixed flame in a burner and
comparing the flame structure with those of measured and another available numer-
ical solution. We have compared the distributions of mixture fraction, temperature,
and soot volume fraction in the flame. The comparison showed a good agreement
with those of experiment. Then, we embedded a micro-scale injector in the burner
wall, split the incoming air-flow between primary and secondary streams, injected
the secondary-air into the burner via the micro-scale injector, and compared the
results obtained for different values of micro-impinging-jet mass flow rates. Our
numerical study have shown that mico-jet impingement changes the reactive flow
pattern within the burner and reduces the flame length. They have also shown that
as the mass flow rate of micro-impinging-jet increases, the emissions of soot nano-
aerosol, CO, and C6 H6 reduce. The profiles of exhaust-gas temperature, soot volume
fraction, mass fractions of CO, CO2 , and C6 H6 in exhaust gases also become more
uniform in the case of a mico-jet impingement.
Acknowledgements The authors would like to thank the financial support received from the
Deputy of Research and Technology in Sharif University of Technology. Their financial support
and help are greatly acknowledged.
References
1. Civilian, D.L.D.: Numerical analysis of two and three dimensional recessed flame holders for
scramjet applications. PhD thesis, Air University, Air Force Institute of Technology (1996)
2. Leung, K.M., Lindstedt, R.P., Jones, W.P.: A simplified reaction mechanism for soot formation
in nonpremixed flames. Combust. Flame 87, 289–305 (1991)
3. Kennedy, I.M., Yam, C., Rapp, D.C.: Modeling and measurements of soot and species in a
laminar diffusion flame. Combust. Flame 107, 368–382 (1996)
4. Brookes, S.J., Moss, J.B.: Predictions of soot and thermal radiation properties in confined
turbulent jet diffusion flames. Combust. Flame 116, 486–503 (1999)
5. Kronenburg, A., Bilger, R.W., Kent, J.H.: Modeling soot formation in turbulent methane-air jet
diffusion flames. Combust. Flame 121, 24–40 (2000)
6. Hall, R.J., Smooke, M.D., Colket, M.B.: Predictions of soot dynamics in opposed jet diffusion
flames, chapter 8. In: Sawyer, R.F., Dryer, F.L. (eds.) Physical and Chemical Aspects of
Combustion: A Tribute to Irvin Glassman. Combustion Science and Technology Book Series,
vol. 189, pp. 189–230. Gordon and Breach, Amsterdam (1997)
7. Darbandi, M., Ghafourizadeh, M., Schneider, G.E.: Aerosol modeling of soot nanoparticles
in a turbulent diffusion flame using an extended detailed kinetic scheme. In: 52nd Aerospace
Sciences Meeting, National Harbor. AIAA 2014-0580 (2014)
8. Darbandi, M., Ghafourizadeh, M., Jafari, S.: Simulation of soot nanoparticles formation and
oxidation in a turbulent nonpremixed methane-air flame at elevated pressure. In: 13th IEEE
International Conference on Nanotechnology, Beijing, pp. 608–613. IEEE-NANO 2013-0382
(2013)
9. Darbandi, M., Ghafourizadeh, M., Schneider, G.E.: Extending a numerical procedure to
simulate the micro/nanoscale soot formation in ethylene-air turbulent flame using acetylene-
route nucleation. In: 11th AIAA/ASME Joint Thermophysics and Heat Transfer Conference,
Atlanta. AIAA 2014-2385 (2014)
10. Darbandi, M., Ghafourizadeh, M., Schneider, G.E.: Numerical study on surface oxidation of
carbonaceous nano- and micro-particles in a heavily sooting ethylene turbulent jet flame. In:
53rd AIAA Aerospace Sciences Meeting, Kissimmee. AIAA 2015-2084 (2015)
11. Darbandi, M., Ghafourizadeh, M., Schneider, G.E.: Prediction of soot nano/micro particles
emission (smoke pollution) in a turbulent propane nonpremixed flame. In: The CSME
International Congress, Toronto (2014)
12. Darbandi, M., Ghafourizadeh, M., Schneider, G.E.: Two-phase flow simulation of nano- and
micro-particulate soot in an air-preheated nonpremixed turbulent propane flame. In: 22nd
Annual Conference of the CFD Society of Canada, Toronto (2014)
13. Kent, J.H., Honnery, D.: Soot and mixture fraction in turbulent diffusion flames. Combust. Sci.
Technol. 54, 383–397 (1987)
14. Sozer, E., Hassan, E.A., Yun, S., Thakur, S., Wright, J., Ihme, M., Shyy, W.: Turbulence-
chemistry interaction and heat transfer modeling of H2 /O2 gaseous injector flows. In: 48th
AIAA Aerospace Sciences Meeting, Orlando. AIAA 2010-1525 (2010)
15. Young, K.J., Stewart, C.D., Moss, J.B.: Soot formation in turbulent nonpremixed kerosine-air
flames burning at elevated pressure: experimental measurement. Proc. Combust. Inst. 25(1),
609–617 (1994)
16. Wen, Z., Yun, S., Thomson, M.J., Lightstone, M.F.: Modeling soot formation in turbulent
kerosene/air jet diffusion flames. Combust. Flame 135, 323–340 (2003)
Normalization of Eigenvectors and Certain
Properties of Parameter Matrices Associated
with The Inverse Problem for Vibrating Systems
Mohamed El-Gebeily and Yehia Khulief
Abstract Solutions of the equation of motion of an n-dimensional vibrating system

M qR C DPq C Kq D 0 can be found by solving the quadratic eigenvalue problem
L./x WD 2 Mx C Dx C Kx D 0. The inverse problem is to identify real
definite matrices M > 0; K > 0 and D 0 from a specified pair .; Xc /
of n-eigenvalues and their corresponding eigenvectors of the eigenvalue problem.
We assume here that D U C iW, where U 0 and W > 0 are diagonal
matrices. The well posedness of the inverse problem requires that the matrix Xc
be specially normalized. It is known that for such specially normalized Xc , there
exist a nonsingular matrix XR and an orthogonal matrix , both real, such that
Xc D XR .I i/. The identified matrices depend on a matrix polynomial Pr ./ D
Ur C Wr T C Wr Wr T ; r D 1; 0; 1, where Ur D <.r / and Wr D =.r /.
In this work we give an explicit characterization of normalizers of Xc , introduce
some new results on the class of admissible orthogonal matrices and characterize
the invertibility of the polynomials Pr ./ in terms of the invertibility of r . For
r D 1; 1 this is equivalent to identifying M > 0; K > 0. For r D 2 but Ur not
strictly negative, we give an example to show that Pr ./ is indefinite for all .
1 Introduction
In this article we are interested in the identification problem associated with a

free vibrating beam or pipe. The discretized equation of motion that represents the
transverse vibration of an Euler-Bernoulli beam with no axial deformations has the
general form [11].
M qR C DPq C Kq D 0; (1)
where M; D; K are called the mass matrix, damping matrix and stiffness matrix,
respectively. For a physical stably vibrating system, M and K are strictly positive
M. El-Gebeily () • Y. Khulief

King Fahd University of Petroleum and Minerals, Dhahran 31261, Saudi Arabia

102 M. El-Gebeily and Y. Khulief
definite (M > 0; K > 0) while D is negative definite (D 0). All these matrices are
assumed to be of dimension n. It is well known [9] that solving (1) is equivalent to
solving the eigenvalue problem
L./x D 0; (2)
where L./ is the second order matrix polynomial
L./ D 2 M C D C K: (3)
Solutions .; x/ of (2) appear in conjugate pairs and for vibrating beams or pipes,
has a non-positive real part: <./ 0 and a nonzero imaginary part: =./ ¤ 0.
Hence, it suffices to consider the case =./ > 0. Let 1 ; 2 ; ; n be eigenvalues
and x1 ; x2 ; ; xn be their corresponding eigenvectors. Following the notation in [8],
we let
U D diag.< .k //nkD1 ; W D diag.=

.k //nkD1 ;; (4)
D U C iW; X c D x1 x2 : : : xn
The inverse problem for (1) is one of identifying the real matrices M > 0; K >
0 and D 0 from a specified pair .; Xc /. There is a great deal of interest and
vast literature devoted to this inverse problem for vibrating beams and other related
vibrating systems including system monitoring and fault detection, inverse Sturm-
Liouville problems, applied physics, and signal processing. The reader is referred to
[1–3, 9, 10, 12] and the references therein.
The well posedness of the inverse problem requires that the matrix Xc be specially
normalized [8] in the sense that there is a complex diagonal matrix such that the
normalized matrix
XQ c WD Xc (5)
satisfies
XQ R XQ TR D XQ I XQ TI ; (6)
where XQ R D <.XQc / and XQ I D =.XQc /. To the best of our knowledge, no formula has
been given for finding a normalization matrix except through solving case by case
a quadratic system of equations involving the elements of , see, e.g. [4–6, 8]. In this
work we give an explicit formula for , which turns out to be also a characterization
of such normalization matrices. This is done in Sect. 3 in which we also give an
illustrative example of an actual vibrating beam problem.
Furthermore, it follows from the polar decomposition of XQ c that its real and
imaginary parts are related by
XQ I D XQ R ;
Eigenvector Normalization of Vibrating Systems 103
where is a real orthogonal matrix. Therefore,
XQ c D XQ R .I i/: (7)
We also show that the spectrum of satisfies
./ fz 2 C W arg.z/ < =2g: (8)
This result is then used to show that a certain matrix Q associated with the so
called Jordan triples is invertible, which eliminates the need to check this condition
from Algorithm 1 in [8]. Other properties of and a certain matrix polynomial
defined in terms of it are discussed in Sect. 4, where we characterize the invertibility
of this polynomial in terms of the given parameter . This has direct consequence
on the identification of positive definite M; K. We also provide an example to show
that the identification of a D 0, may not be possible to in general.
In Sect. 2, we collect some facts about discretized vibrating systems which we
will need in this paper.
2 Preliminaries
Let U; W; be as defined in (4) and, for r 2 Z, let
r D r ; Ur D <.r /; and Wr D =.r /: (9)
Here, 0 D I. We will always assume that U 0; W > 0. It is well known that

.L.// D [ N D .A B/, where

DM K 0
AD ; BD :
M 0 0 M
With Xc ; ; M defined as above, given matrices X 2 Cn2n ; J 2 C2n2n ; Y 2

2nn
C , the triple .X; J; Y/ is called a Jordan triple for L./ if

0 0
X D Xc X c ; J D ; Y D Q1 ; (10)
0 M 1
where

X
QD (11)
XJ
and N denotes complex conjugate. To determine real symmetric matrices K and D

from and Xc , it suffices that .X; J; X T / be a Jordan triple for L./. In this case, we
must have
X eRT D X
eR X eIT ;
eI X (12)
for some normalization X eR C iX

ec D X eI of Xc . When (12) is satisfied, we have the
representations

eI D X
X e
eR ; ec D X
X e
eR I i (13)
e
for some real orthogonal matrix .
3 Eigenvector Normalization
In this section we address the question: which normalizations (if any) of Xc

allow (12) to hold?
Define the complex matrix by
D Xc1 XN c (14)
Proposition 1 Assume Xc D XR C iXI is a given n n matrix whose columns are

eigenvectors of a symmetric system and let be defined by (14). There exists a
complex diagonal n n normalization matrix such that X ec D Xc satisfies (12) if
and only if there exists an n n invertible matrix A such that
1. A is diagonal,
2. AT D A and
3. 2 D A:
ec D Xc satisfies (12). The latter can be written in
Proof Suppose exists. Then X
matrix form as
T
R2 I2 2R I XR
XR XI D 0:
2R I I2 R2 XIT

I I I iI
Observing that D 2I2n , we can rewrite the last equation as
iI iI I iI
" #
0 2
T
Xc
Xc X c D0
2 0 XcT
Multiplying on the left and right by Xc1 and XcT , respectively, and putting WD
Xc1 XN c
we get

0 2 T
I D0 (15)
2 0 I
T
0 2
Clearly, the kernel of the operator 2 W Cn ! C2n is f0g. Furthermore,
0 I

the operator I W C2n ! Cn has rank n, and therefore, its kernel also has rank n:
T
0 2
For (15) to hold, the operator 2 must map onto the kernel of the
0 I

operator I : The latter is spanned by the columns of : Therefore, there
I
exists an n n change of base matrix A such that

0 2 T
D A:
2 0 I I
Hence, we must have
2 D A;
2 D A T :
Since 1 D ,
N we may rewrite the above equations as
2 D A D A
N T: (16)
Furthermore, since the matrices in the last two equalities of (16) must be diagonal
matrices and is invertible, we conclude that AN D AT .
The only if part of the proof can be done by reversing the above steps.
Proposition 1 suggests the following algorithm to find a normalization matrix . The
steps parallel the three parts of the proposition.
Algorithm 1 To compute a normalization diagonal matrix :
1. Perform column reduction operations on to reduce it to a lower triangular
form. (This fixes the upper triangular part of A and, by Proposition 1-2, the lower
triangular part. At this point, the matrix A is actually diagonal.)
2. Ensure that the diagonal part of A is purely imaginary so that Proposition 1-2 is
satisfied.
3. Calculate from Proposition 1-3.
Since column reduction operations are not unique, we expect that the normalization
matrix is not unique. The following example shows that this is the case. It also
explores an alternative approach for finding that works in special situations.
Example 1 The coefficients of a finite element discretization of a vibrating beam
(see [7], Chapter 8) are given by
2 3
:0929 0 0161 :0967
6 0 1:4881 :0967 :558 7
MD6 4 :0161 :0967 :0464 :1637 5 ;
7
:0967 :558 :1637 :7448

2 3
:0052 0 :0026 :0326
6 0 1:088 :0326 :272 7
K D 104 6 7
4 :0026 :0326 :0026 :0326 5
:0326 :272 :0326 :544
and D D 0:2M C :005K: The eigenvalues and ‘un-normalized’ eigenvectors of this

system are
2 3 2 3
128:86 0 0 0 186:81 0 0 0
6 0 15:46 0 0 7 6 0 7
D6 7 C i 6 0 76:84 0 7;
4 0 0 1:43 0 5 4 0 0 23:05 0 5
0 0 0 0:13 0 0 0 3:60
2 3 2 3
0:6 0:3 2:3 20:1 0:9 1:3 30:8 71:8
6 0:3 0:4 0 1:4 7 6 0:4 1:9 0:4 4:9 7
10 Xc D 6
3 7 6 7
4 2:5 2:5 3:2 59:5 5 C i 4 3:6 12:5 42:6 212:4 5
1:0 0:5 0:3 1:6 1:4 2:4 4:1 5:9
The real part of this matrix is invertible, therefore, we will put

2 3
1:4497 0 0 0
6 0 4:9605 0 0 7
D XR1 XI D 6
4
7;
0 0 13:187 0 5
0 0 0 3:5725
which happens to be a diagonal matrix. In this rather simple case, (12) simplifies to
R D I .I C / .I /1 :
This equation shows the arbitrariness in assigning, say I : If we choose I D I, then

2 3
5:4473 0 0 0
6 0 1:5050 0 0 7
R D 6
4
7:
5
0 0 1:1641 0
0 0 0 0:5626
The normalized eigenvectors are

2 3 2 3
2:5 0:9 2:8 83:1 5:6 2:2 38:2 20:3
6 1:0 1:3 3:0 5:7 7 6 2:3 3:3 0:5 1:4 7
10 X c D 6
3e
4 10:0 8:7 38:8
7 C i6 7:
245:9 5 4 22:3 21:3 52:8 60:0 5
3:9 1:7 3:7 6:8 8:6 4:1 5:1 1:7
4 Consequences of Normalization
From now on, we will assume that Xc is properly normalized so that (12) is satisfied.
It was shown in [8] that, to determine a mass matrix M > 0, it is necessary that
C T > 0; (17)
where is a real orthogonal matrix satisfying (13). In this section, we will discuss
some consequences of this condition. Let 1 ; 2 ; ; n be an orthonormal system
of eigenvectors of T corresponding to its eigenvalues ei'1 ; ei'2 ; ; ei'n with
'1 ; '2 ; ; 'n < . Define the matrices

˘ D diag ei'1 ; ei'2 ; ; ei'n and ˚ D Œ1 ; 2 ; ; n : (18)
Notice that T ˚ D ˚˘ and ˚ D ˘ ˚ . If matrix Q of (11) is nonsingular,

then .X; J; X T / is a Jordan triple for L./ and we have
M D .XJX T /1 ; (19)

2 T 1
D D M.XJ X / M; (20)
K D .XJ 1 X T /1 : (21)
Equations (19), (20) and (21) determine a symmetric system whose eigenvalues and
eigenvectors are and Xc , respectively. We shall see later (see Lemma 2 below)
that (17) ensures the invertibility of Q. This will require a characterization of the
eigenvalues of the matrix T , which will be given in Lemma 1.
The set of orthogonal matrices can be parameterized using real skew-symmetric

matrices C as
D .I C/ .I C C/1 (22)
Condition (17) restricts the class of parameters as in the following lemma. We

should observe here that skew-symmetric matrices have pure imaginary eigenvalues.
Lemma 1 Let be a given orthogonal matrix and define C by (22). The following
are equivalent:
1. Equation (17) holds
p p
2. .C/ i.1= 2; 1= 2/, that is, every eigenvalue iı of the parameterization
matrix C satisfies
p
jıj < 1= 2 (23)

3. T CC WD fz 2 C W < .z/ > 0g :
Proof 1 ” 2. Let be an orthogonal matrix and let C be its corresponding
parameter matrix defined by (22). Then
C T D .I C/ .I C C/1 C .I C/1 .I C C/ :
If iı is an eigenvalue of C, then by the spectral mapping theorem,
1 iı 1 C iı
C
1 C iı 1 iı
is a corresponding eigenvalue of C T . Then, by (17),
1 2ı 2
> 0;
1 C ı2
which is equivalent to (23).

1 ” 3. If (17) holds, then

0 < ˚ C T ˚ D ˘ ˚ ˚ C ˚ ˚˘ D 2< .˘ / ;
which implies 3. If 3. holds, then for any y 2 Cn , we may write y D ˚˛ for some
˛ 2 Cn . Then

y C T y D ˛ ˚ C T ˚˛ D 2˛ < .˘ / ˛ > 0;
which implies 1.

Lemma 2 The matrix Q defined by (11) is invertible. Consequently, .X; J; X T / is a

Jordan triple for L./.
Proof It follows from the invertibility of Xc and (13) that both XR and I i are
invertible. Furthermore, I C i is also invertible since its eigenvalues are conjugates
of those of I i: Now,

X XR .I i/ XR .I C i/
QD D
XJ XR .I i/ XR .I C i/

Xc 0 I .I i/1 .I C i/
D :
0 Xc .I i/1 .I C i/
Put

D i .I i/1 .I C i/ : (24)
A straightforward calculation shows that .

/ RC and hence,
has a well
defined square root. Then

iI 0 I .I i/1 .I C i/ iI
D :
i iI .I i/1 .I C i/ 0
C

Therefore, the invertibility of Q is equivalent to the invertibility of

C
,
which, in turn, is equivalent to the invertibility of
1
. To show the
invertibility of the latter, we write

1
D
1=2 I
1=2 1
1=2
1=2
1=2
1=2

1
D
1=2 I
1=2
1=2
1=2
1=2
1=2

D
1=2 I B1 B
1=2 ;
where B D
1=2
1=2 and where
we made use of the fact that
D
1 .
1
Therefore, we have to show that I B B is invertible. If not, then, invoking the
spectral mapping theorem, we have, for some ı 2 .B/ D ./,
ı = .ı/
0D1 D 2i ;
ı ı
which contradicts the assumption that W > 0.
Next, we investigate the invertibility of the matrices M; D; K as identified
from (21), (22) and (23). For r 2 Z, let r ; Ur ; Wr be as defined in (9) and let
be a real orthogonal matrix. Define the matrix function

Pr ./ D < .I i/ r I i T
D Ur C Wr T C Wr Ur T :
Then

0 r I i T
Pr ./ D I C i I i
r 0 I C i T
and we have the following proposition.

Proposition 2 Pr ./ is invertible if and only if r is invertible.
Proof Pr ./ is invertible if and only if ker Pr ./ D f0g : Note that Pr ./ D
0 r I
.I i/ i
I I i T ; where
is given by (24). Therefore,
r 0 i
T

0 r I
ker Pr ./ D ker i
I : Arguing as in Proposition 1, the
r 0 i
T

0 r I
operator W Cn ! C2n has rank n; and the operator i
I W
r 0 i
T
C2n ! Cn has rank n and kernel space of dimension n: It follows that

0 r I
ker i
I D f0g (25)
r 0 i
T

0 r I
if and only if maps into the orthogonal complement of the kernel
r 0 i
T

I
of i
I : The kenel of i
I is spanned by the columns of and it is
i
straightforward
T to see that its orthogonal complement is spanned by the columns of
i
: Therefore, (25) holds if and only if there exists an invertible matrix B such
I
that
T
0 r I i
D B:
r 0 i
T I
The second of the above equations gives
r D B:
Under our assumptions on U and W, P1 ./; P1 ./ are always invertible. This, in
turn is equivalent [8] to the positive definiteness of M; K. However,the following
example shows that P2 ./ may be indefinite if U is not strictly negative definite.
Example 2 Let

20 10
UD ;W D :
00 02
Then

2 3 0 4 0
D Ci :
0 4 0 0
All orthogonal 2 2 matrices have the form

cos sin cos sin
or :
sin cos sin cos
The eigenvalues of the first type are e˙i and by Lemma 1, is admissible only if
2 .0; =2/. The eigenvalues of the second kind are 1; 1 and thus, they do not
correspond to an admissible . For the first kind, we calculate

8 cos 7=2 cos 2 C 7=2 4 sin 7=2 sin 2
P2 ./ D
4 sin 7=2 sin 2 7=2 cos 2 7=2
p
which has eigenvalues 1;2 D 4 cos 1=2 162 98 cos 2: It can be shown
that 1 < 0; 2 > 0 for all 2 .0; =2/. Therefore, P2 ./ is non-definite for any
admissible :
Acknowledgements This work is funded by KACST-NSTIP Project No. 12-ADV3005-04. The

authors acknowledge the support provided by King Abdulaziz City for Science & Technology and
King Fahd University of Petroleum & Minerals.
References
1. Bai, Z.J.: Constructing the physical parameters of a damped vibrating system from eigendata.
Linear Algebra Appl. 428(2–3), 625–656 (2008)
2. Chu, M.T.: Inverse eigenvalue problems. SIAM Rev. 40, 1–39 (1998)
3. Chu, M.T., Golub, G.H.: Inverse Eigenvalue Problems: Theory, Algorithms, and Applications.
Oxford University Press, Oxford (2005)
4. De Angelis, M., Imbimbo, M.: A procedure to identify the modal and physical parameters of a
classically damped system under seismic motions. Advances in Acoustics and Vibration, Hin-
dawi Publishing Corporation, Article ID 975125, 11 pages (2012). doi:10.1155/2012/975125
5. Friswell, M.I., Prells, U.: A measure of non-proportional damping. Mech. Syst. Sig. Process
14(2), 125–137 (2000)
6. Garvey, S.D., Penny, J.E.T.: The relationship between the real and imaginary parts of complex
modes. J. Sound Vib. 212(1), 64–72 (1998)
7. Kwon, Y.W., Bang, H.: The Finite Element Method Using Matlab. CRC Press, Boca Raton
(1997)
8. Lancaster, P., Prells, U.: Inverse problems for damped vibrating systems. J. Sound Vib. 283,
891–914 (2005)
9. Lancaster, P., Tismenetsky, M.: The Theory of Matrices. Academic, Orlando, Academic Press,
New York (1985)
10. Lancaster, P., Zaballa, I.: On the inverse symmetric quadratic eigenvalue problem. SIAM J.
Matrix Anal. Appl. 35(1), 254–278 (2014)
11. Reddy, J.N., Wang, C.M.: Dynamics of fluid-conveying beams, CORE Report No. 2004-03,
Centre for Offshore Research and Engineering, National University of Singapore (2004)
12. Xu, S.F.: An Introduction to Inverse Algebraic Eigenvalue Problems. Peking University
Press/Friedr. Vieweg & Sohn, Beijing/Braunschweig/Wiesbaden (1998)
Computational Aspects of Solving Inverse
Problems for Elliptic PDEs on Perforated
Domains Using the Collage Method
H. Kunze and D. La Torre
Abstract The treatment of an inverse problem on a perforated domain is com-

plicated heavily by the presence of the perforations or holes. We present several
theoretical results that provide relationships between the problem on the perforated
domain and the same problem on the corresponding unperforated/solid domain. The
results establish that we can approximate the solution of the inverse problem on the
perforated domain by instead solving the inverse problem on the associated solid
domain. Examples are provided.
1 Introduction
In this paper, we are interested in the parameter estimation inverse problem for
elliptic PDEs. One physical setting for such problem is the estimation of the thermal
diffusivity in a lamina based on (perhaps noisy) observational data of its equilibrium
temperature. Indeed, the examples in the final section of the paper include images
that can be thought of as isotherm plots in a lamina. In Sect. 2, we present the
Collage Method approach for solving such inverse problems. In the past [5], we have
established that the Collage Method compares favourably with established methods,
such a Tikhonov regularization [9]. The complication in the current work is that we
wish to consider such an inverse problem on a perforated domain.
A perforated domain (or porous medium) is a material characterized by a
partitioning of the total volume into a solid portion often called the “matrix” and
a pore space usually referred to as “holes.” Mathematically speaking, these holes
can be either materials different from that of the matrix or real physical holes. When
H. Kunze ()
Department of Mathematics and Statistics, University of Guelph, Guelph, ON, Canada
D. La Torre
Department of Economics, Management, and Quantitative Methods, University of Milan, Milan,
Italy
Department of Applied Mathematics and Sciences, Khalifa University, Abu Dhabi, UAE

114 H. Kunze and D. La Torre
formulating differential equations over porous media, the term “porous” implies that
the state equation is written in the matrix only while boundary conditions should be
imposed on the whole boundary of the matrix, including the boundary of the holes.
Solving differential equations over a perforated domain is typically a complicated
task because the size and distribution of the holes within the material play an
important role in its characterization. Simulations conducted over a perforated
domain that includes a large number of matrix-hole interfaces present numerical
challenges since a very fine discretization mesh and a large computation time
are required. The direct problem is of great interest in various areas of science,
engineering, and industry (see, for example, [2] and [3] for discussion of real-
world problems). One way to treat these (perhaps idealized) problems with rigorous
mathematics is called homogenization (see [3] and [8]), which takes advantage of
the multiscale nature of the perforated domain.
The inverse problem over a perforated domain inherits all of these challenges.
In this paper, we seek to avoid the complications of working with the inverse
problem solution machinery on the perforated domain. In Sect. 3, we present several
theoretical results that connect the problem on the perforated domain with the same
problem on the corresponding unperforated (solid) domain. To frame the situation,
we set up the two types of problems here.
Let ˝ 2 R2 be compact and convex and ˝B be collection of circular m holes
[iD1 B.xj ; "j /, where xj 2 ˝, the radii "j > 0, and the holes B.xj ; "j / are non-
m
overlapping and strictly inside ˝. We let " D maxj "j , the maximum hole radius.
Let ˝" denote the closure of ˝ n ˝B . Then, we consider the problem .P" / on the
perforated domain

r .K .x; y/ru.x; y// D f .x; y/ in ˝"
(P" )
u.x; y/ D 0 on @˝"
and the associated problem .P/ on the corresponding unperforated domain

r .K .x; y/ru.x; y// D f .x; y/ in ˝
; (P)
u.x; y/ D 0 on @˝
where 2 Rn is a parameter belonging to the compact set . The Dirichlet

boundary conditions above can be replaced by Neumann boundary conditions, in
particular on the boundaries of the holes ( @u
@n D 0 on @˝B ).
In each case, the inverse problem is to estimate the parameter (perhaps
coefficients defining the diffusivity K and/or the source/sink function f ) given
observational data values of the solution u.x; y/.
Given observational data for the solution of problem .P" /, we solve the inverse
problem for .P/; the results we present in Sect. 3 show that the obtained parameter
values solve (approximately) the inverse problem for .P" /.
Computational Aspects of Solving Inverse Problems on Perforated Domains 115
2 The Collage Theorem for Elliptic PDEs
The variational equation associated with an elliptic PDE can be written as
a.u; v/ D .v/; v 2 H; (1)
where .v/ and a.u; v/ are linear and bilinear maps, respectively, both defined on
a Hilbert space H. We denote by h; i the inner product in H, kuk2 D hu; ui and
d.u; v/ D ku vk, for all u; v 2 H.
The inverse problem of interest may now be viewed as follows: Suppose that we
have an observed solution u and a given (restricted) family of bounded, coercive
bilinear functionals a .u; v/, 2 , and a family of bounded linear functionals .
Then, by the Lax-Milgram theorem, for each 2 there exists a unique u 2 H
such that .v/ D a .u ; v/ for all v 2 H. We would like to determine if there
exists a value of the parameter such that u D u or, more realistically, such that
ku uk is small enough. The following theorem is useful for the solution of this
problem.
Theorem 1 (Generalized Collage Theorem [6]) For all 2 , suppose that
a .u; v/ W H H ! R is a family of bilinear forms and W H ! R is
a family of bounded linear functionals. Let u denote the solution of the equation
a .u; v/ D .v/ for all v 2 H, as guaranteed by the Lax-Milgram theorem. Then,
given a target element u 2 H,
1
ku u k F .u/; (2)
m
where
ˇ ˇ
F .u/ D sup ˇa .u; v/ .v/ˇ (3)
v2H; kvkD1
and m > 0 is the coercivity constant of a .

In order to ensure that the approximation u is close to a target element u 2 H,
we can, by the Generalized Collage Theorem, try to make the term F .u/=m as
close to zero as possible. If inf2 m m > 0 then the inverse problem can be
reduced to the minimization of the function F .u/ on the space , that is,
min F .u/: (4)

2
We refer to the minimization of the functional F .u/ as a “generalized collage

method” because of philosophical connection with our earlier approach to inverse
problems for ODEs [7]; in that work, the word “collage” was used due to
connections to ideas in fractal imaging [1]. Such an optimization problem has a
solution that can be approximated with a suitable discrete and quadratic program,
derived from the application of the Generalized Collage Theorem and the use of an
orthonormal basis in the Hilbert space H [6].
3 Theoretical Results
We introduce the Sobolev spaces H D H01 .˝/ and H" D H01 .˝" /. Since any
function in H" can be extended to be zero over the holes, it is trivial to prove that H"
can be embedded in H. We let ˘" u be the projection of u 2 H onto H" ; it follows
that
ku ˘" ukH ! 0 whenever " ! 0:
When Neumann boundary conditions are considered, it is still possible to extend a

function in H" to a function of H: these extension conditions are well studied (see
[8]) and they typically hold when the domain ˝ has a particular structure.
The variational formulations of (P" ) and (P) are
find u 2 H" such that a" .u; v/ D " .v/; 8v 2 H" (P" )
and
find u 2 H such that a .u; v/ D .v/; 8v 2 H: (P)
The generalized collage distance associated to .P/ is stated in (3) in Theorem 1,

while the generalized collage distance associated to .P" / is
ˇ ˇ
F" .u/ D sup ˇa .u; v/ .v/ˇ : (5)
" "
v2H" ; kvkH" D1
In the results that follow, we assume that the continuous and bilinear forms a" and
a are uniformly coercive and bounded with respect to and ", namely there exists
two positive constants m and M such that for all 2
8
ˆ
ˆ a .u; u/ mkuk2 8u 2 H"
ˆ "
ˆ
< a .u; v/ Mkukkvk 8u; v 2 H"
"
(H1)
ˆ a .u; u/ mkuk2
ˆ 8u 2 H
ˆ
:̂
a .u; v/ Mkukkvk 8u; v 2 H
We also assume that the linear functionals " and are uniformly bounded with
respect to and ", namely there exists a positive constant such that
(
" .u/ kuk 8u 2 H"
(H2)
.u/ kuk 8u 2 H
Under the hypotheses (H1) and (H2), (P" ) and (P) have unique solutions u" and u ,
respectively, for each 2 and for each fixed choices of "j , j D 1; : : : ; m.
We now state three results relating (P" ) and (P), first presented with proofs in [4].
Proposition 1 The following estimate holds:
F .u/ M
k˘" u u" kH" C ku ˘" ukH (6)
m m
Proposition 2 There exists a constant C, that does not depend on ", such that the
following estimate holds:
F .˘" u/ F" .˘" u/ C C" (7)
for all 2 , " > 0.

Proposition 3 Suppose that F .u/; F" .v/ W ! RC are continuous for all u 2 H,
v 2 H" , and " > 0. Let " be a sequence of minimizers of F" .u/ over . Then
there exists "n ! 0 and 2 such that "n ! , with a minimizer of F .u/
over .
The practical fallout of these results is that we have some justification in
minimizing F .u/ (the generalized collage distance for the problem on the domain
with no holes) to obtain estimates of the parameter values for .P" / (the problem
on the domain with holes). The quality of the estimate may be affected by the size
of the largest hole; certainly, when all holes are sufficiently small, we expect the
approximation to be good.
4 Examples
Example 1 We place nine holes of assorted sizes inside ˝ D Œ0; 12 , as in Fig. 1 We

choose K.x; y/ D Ktrue .x; y/ D 8 C 3x2 C y2 and consider the problem
8
< r .K.x; y/ru.x; y// D 2x2 C y2 ; in ˝" ;
u.x; y/ D 0; on @˝; (8)
: @u
@n
.x; y/ D 0; on @˝B ;
Fig. 1 The domain and level curves of solutions for Example 1
Table 1 Results for the inverse problem in Example 1. Due to the normalization of 0 , the true
values are .1 ; 2 ; 3 / D .8; 3; 1/
ı (%) 1 2 3
0 7.1935 2.9783 1.2431
1 7.1586 3.0466 1.2573
3 7.0737 3.1945 1.2934
where ˝B is the union of the nine holes. We solve the diffusion problem numerically
and sample the solution u" at 49 uniformly-distributed points strictly inside ˝. The
level curves of the solution are illustrated in Fig. 1. If a sample point lies inside a
hole, we obtain no information at the point.
Now, beginning with the observational data points, we consider the inverse
problem of estimating K and f . To this end, we define K .x; y/ D 0 C 1 x2 C 2 y2 .
Using the data values and f .x; y/ D 4x2 C y2 , we seek to estimate the values of i
in K .x; y/ by applying the generalized collage theorem to solve the related inverse
problem on ˝ with no holes.
The results for various cases, with relative noise of ı% added, are presented in
Table 1. The results worsen as noise is added, but remain reasonably good.
1
Example 2 For " 2 f0:1; 0:025; 0:01g, define N" D 10" and
[
N"
1 1
˝B D B" i "; j " ;
i;jD1
2 2
a domain with N"2 uniformly-distributed holes all of radius ". Choosing K.x; y/ D
Ktrue .x; y/ D 9 C 3x C 2y, we consider the steady-state diffusion problem
8
< r .K.x; y/ru.x; y// D 2x2 C y2 ; in ˝" ;
u.x; y/ D 0; on @˝; (9)
: @u
@n
.x; y/ D 0; on @˝ B :
For each fixed value of ", we solve the diffusion problem numerically and sample
the solution at 49 uniformly-distributed points strictly inside ˝, obtaining no
information at the point if it lies in a hole. Using the data values, with relative noise
of ı% added, we use the generalized collage theorem to solve the related inverse
problem, knowing f .x; y/ D 2x2 C y2 and seeking a diffusivity function of the form
K.x; y/ D 0 C 1 x C 2 y.
The level curves of each solution are illustrated in Fig. 2. When the hole is too
large, as in the N D 1 case, the estimates are very poor. In this case, the hole needs
to be incorporated into the macroscopic-scale model, as it can’t be considered part
of the smaller-scale model. In the other cases of the table, the estimates are good.
We see that as the size of the holes decreases (even while the number increases),
the solution to the inverse problem produces better estimates of the parameters. As
the noise increases, the results worsen but remain good (Table 2).
Fig. 2 Level curves of solutions in Example 2, with " D 0:1, 0:025, and 0:01
Table 2 Results for the Recovered parameters

inverse problem in
" N" ı (%) 0 1 2
Example 2. True values are
.0 ; 1 ; 2 / D .10; 3; 2/ 0.1 1 0 14:4169 10:1594 1:8807
1 14:2991 10:2965 1:8130
3 14:0300 10:5812 1:6543
0.025 4 0 8:7683 4:1071 2:4944
1 8:7188 4:1853 2:5131
3 8:6024 4:3496 2:5587
0.01 10 0 8:8155 3:7757 2:3408
1 8:7645 3:8531 2:3661
3 8:6453 4:0158 2:4252
Acknowledgements H. Kunze thanks the Natural Sciences and Engineering Research Council for
supporting this research.
References
1. Barnsley, M.F.: Fractals Everywhere. Academic, New York (1989)

2. Civan, F.: Porous Media Transport Phenomena. Wiley, Hoboken (2011)
3. Espedal, M.S., Fasano, A., Mikelić, A.: Filtration in Porous Media and Industrial Application.
Lecture Notes in Mathematics, vol. 1734. Springer, New York (1998)
4. Kunze, H., La Torre, D., Collage-type approach to inverse problems for elliptic PDEs on
perforated domains. Electron. J. Differ. Equ. 48, 1–11 (2015)
5. Kunze, H., La Torre, D., Levere, K., Ruiz Galán, M.: Inverse problems via the “generalized
collage theorem” for vector-valued Lax-Milgram-based variational problems. Math. Probl. Eng.
(2015). http://dx.doi.org/10.1155/2015/764643
6. Kunze, H., La Torre, D., Vrscay, E.R.: A generalized collage method based upon the Lax–
Milgram functional for solving boundary value inverse problems. Nonlinear Anal. 71(12), 1337–
1343 (2009)
7. Kunze, H., Vrscay, E.R.: Solving inverse problems for ordinary differential equations using the
Picard contraction mapping. Inverse Probl. 15, 745–770 (1999)
8. Marchenko, V.A., Khruslov, E.Y.: Homogenization of Partial Differential Equations. Birkhauser,
Boston (2006)
9. Tychonoff, A.N.: Solution of incorrectly formulated problems and the regularization method.
Dokl. Akad. Nauk SSSR 151, 501–504 (1963)
Dynamic Boundary Stabilization
of a Schrödinger Equation Through
a Kelvin-Voigt Damped Wave Equation
Lu Lu and Jun-Min Wang
Abstract In this paper, we study an interconnected system of a Schrödinger and a

wave equation with Kelvin-Voigt (K-V) damping, where the K-V damped wave
equation performs as a dynamic feedback controller. We show that the system
operator generates a C0 -semigroup of contractions in the energy state space, and the
system is well-posed. By detailed spectral analysis, we know that the spectral of the
system operator composes of two parts, point spectrum and continuous spectrum.
Moreover, the points in the spectra all have negative real parts. It follows that the
C0 -semigroup generated by the system operator achieves asymptotic stability.
1 Introduction
There has been much interests in the problem of dynamic feedback controller,
which is modeled by coupled partial differential equations (PDEs) or interconnected
PDEs. This kind of problems are quite challenging, yet have became attractive to
researchers these days. Control design and stability analysis for such systems have
become active over the past 5 years, see [8, 9] and the references therein. In [9],
an interconnected system of Schrödinger equation and heat equation is carefully
studied, which replaces the static feedback by dynamic feedback governed by a
heat equation. It shows the exponential stability of the system and the Gevery class
property of the semigroup. The boundary and internal stabilizations for Schrödinger
equation are considered in [6], where the stability of the systems is achieved by
using multiplier techniques. Two kinds of boundary controllers for Schrödinger
equation are concerned in [2], which shows that a simple proportional collocated
boundary controller can exponentially stabilize the system but the decay rate cannot
be prescribed, while the backstepping method can ensure it to have arbitrary decay
rate.
It is also known that viscoelastic materials have been widely used in engineering
and lots of researchers have put much efforts to analyze the dynamic behavior of
L. Lu () • J.-M. Wang

School of Mathematics and Statistics, Beijing Institute of Technology, Beijing 100081, P.R. China

122 L. Lu and J.-M. Wang
vibration for elastic structures with viscoelasticity over the past several decades.
Kelvin-Voigt (K-V) damping is one of those most commonly used viscoelastic
model, due to its easy and huge applications in modern technology. And there has
been an abundance of literature on the study of elastic system with viscoelastic
damping. In [3, 4], it is shown that local K-V damping can ensure the exponential
stability of both a string and an Euler-Bernoulli beam system when the material
parameter is smooth enough at the interface. However, if the smooth condition
cannot be satisfied, the string system is non-exponentially stable even if the material
parameter is a constant. Passive control of a wave equation with internal K-V
damping is studied in [1]. Results there reveal that the spectrum of the system
operator consists of point spectrum and continuous spectrum, due to the fact that the
resolvent of the system operator for a viscoelastic system is not compact anymore.
In this paper, we present a dynamic input/output feedback controller, which feeds
the K-V damped wave equation and Schrödinger equation into each other through
the boundary. The coupled Schrödinger-wave system (as shown in Fig. 1) is written
as follows:
8
ˆ
ˆ yt .x; t/ C iyxx .x; t/ D 0; 0 < x < 1; t > 0;
ˆ
ˆ
ˆ
ˆ ztt .x; t/ zxx .x; t/ ˛zxxt .x; t/ D 0; 1 < x < 2; t > 0;
ˆ
ˆ
ˆ
< y.0; t/ D z.2; t/ D 0; t 0;
y.1; t/ D kzt .1; t/; t 0; (1)
ˆ
ˆ
ˆ
ˆ ˛z .1; t/ C z .1; t/ D iky .1; t/; t 0;
ˆ
ˆ
xt x x
ˆ
ˆ y.x; 0/ D y0 .x/; 0 < x < 1;
:̂
z.x; 0/ D z0 .x/; zt .x; 0/ D z1 .x/; 1 < x < 2;
where ˛ > 0, k ¤ 0. The two equations are coupled at x D 1 with interconnected

conditions and fixed at each end.
y0 (x) y(x, t)
Schrödinger equation
yt (x, t) + iyxx (x, t) = 0
y(1, t) = kzt (1, t)
kzt (1, t) −ikyx (1, t)
Wave equation with K-V damping

ztt (x, t) − zxx (x, t) − αzxxt (x, t) = 0
z0 (x) αzxt (1, t) + zx (1, t) = −ikyx (1, t) z(x, t)
Dynamic feedback controller
Fig. 1 Block diagram for the dynamic boundary feedback of the coupled system
Stabilization of Coupled Schrödinger-Wave System 123
By introducing the following transformation

w.x; t/ D y.1 x; t/; 0 < x < 1; t > 0;
(2)
u.x; t/ D z.x C 1; t/; 0 < x < 1; t > 0;
then (1) becomes

8
ˆ
ˆ wt .x; t/ C iwxx .x; t/ D 0; 0 < x < 1; t > 0;
ˆ
ˆ
ˆ
< tt
u .x; t/ u xx .x; t/ ˛u xxt .x; t/ D 0; 0 < x < 1; t > 0;
w.1; t/ D u.1; t/ D 0; t 0; (3)
ˆ
ˆ
ˆ w.0; t/ D kut .0; t/;
ˆ t 0;
:̂
˛uxt .0; t/ C ux .0; t/ D ikwx .0; t/; t 0;
Accordingly, the initial conditions for system (3) are w.x; 0/ D w0 .x/; u.x; 0/ D
u0 .x/; ut .x; 0/ D u1 .x/; 0 < x < 1.
The energy function for (3) is given by
Z 1
1
E.t/ D jw.x; t/j2 C jux .x; t/j2 C jut .x; t/j2 dx: (4)
2 0
In this paper, we analyze the spectrum of (3) in which the system operator has
no compact resolvent. We first set up the system operator and show it generates a
C0 -semigroup of contractions, and the system is well-posed. By detailed spectral
analysis, we obtain that the residual spectrum is empty and the continuous spectrum
contains only one negative point. Moreover, all the eigenvalues of the system lie in
the open left half plane. Therefore, this controller design moves the eigenvalues of
the Schrödinger and wave equations into the second quadrant. It follows that the
C0 -semigroup generated by the system operator achieves asymptotic stability.
2 Well-Posedness of System (3)
We consider system (3) in the energy space H D L2 .0; 1/ HE1 .0; 1/ L2 .0; 1/,
where HE1 .0; 1/ D fg 2 H 1 .0; 1/jg.1/ D 0g. The norm in H is induced by the inner
product
Z 1 h i
hX1 ; X2 i D f1 .x/f2 .x/ C g01 .x/g02 .x/ C h1 .x/h2 .x/ dx; (5)
0
where Xs D . fs ; gs ; hs / 2 H , s D 1; 2. Define the system operator of (3) by

8
ˆ
ˆ A . f ; g; h/ D .if 00 ; h; .g0 C ˛h0 /0 /; 8 . f ; g; h/ 2 D.A /;
ˆ
ˆ 8 ˇ 0 9
ˆ
< ˇ g C ˛h0 2 H 1 .0; 1/;
ˆ
ˆ ˇ >
>
< ˇ =
ˆ
ˆ D.A / D . f ; g; h/ 2 H ; A . f ; g; h/ 2 H ; ˇ f .1/ D 0; f .0/ D kh.0/; :
ˆ
ˆ ˆ ˇ >
:̂ :̂ ˇ >
ˇ g0 .0/ C ˛h0 .0/ D ikf 0 .0/ ;
(6)
Then (3) can be written as an evolution equation in H :
8
ˆ
< dX.t/ D A X.t/; t > 0;
dt (7)
:̂ X.0/ D X ;
0
where X.t/ D .w.; t/; u.; t/; ut .; t//.

Theorem 1 Let A be given by (6). Then A 1 exists, and hence 0 2 .A /, the
resolvent set of A . Moreover A is dissipative in H and A generates a C0 -
semigroup eA t of contractions in H .
Proof For any given . f1 ; g1 ; h1 / 2 H by,
A . f ; g; h/ D .if 00 ; h; .g0 C ˛h0 /0 / D . f1 ; g1 ; h1 /; (8)
we have
( 00
f .x/ D if1 .x/; h.x/ D g1 .x/; .g0 .x/ C ˛h0 .x//0 D h1 .x/;
f .1/ D g.1/ D 0; f .0/ D kh.0/; g0 .0/ C ˛h0 .0/ D ikf 0 .0/; g1 .1/ D 0;
(9)
and the solution of (9) is given by
8 Z x Z 1
ˆ
ˆ 0
ˆ
ˆ f .x/ D f .0/.x 1/ i .1 x/f1 ./d i .1 /f1 ./d;
ˆ
ˆ 0
ˆ
ˆ Z 1
x
Z x
ˆ
ˆ
ˆ
< g.x/ D ik.x 1/f 0 .0/ ˛g1 .x/ .1 /h1 ./d C .x 1/ h1 ./d;
x 0
ˆ
ˆ
ˆ
ˆ h.x/ D g1 .x/
ˆ
ˆ
ˆ
ˆ Z 1
ˆ
ˆ 0
:̂ f .0/ D i .1 /f1 ./d kg1 .0/:
0
(10)
Hence, we get the unique solution . f ; g; h/ 2 D.A / to equation (8), thus A 1

exists. Now we show that A is dissipative in H . Let X D . f ; g; h/ 2 D.A /. Then
we have
˝ ˛
hA X; Xi D .if 00 ; h; .g0 C ˛h0 /0 /; . f ; g; h/
Z 1 Z 1 Z 1
D .if 00 /f dx C h0 g0 dx C .g0 C ˛h0 /0 hdx
0 0 0
ˇ1 Z 1 Z 1 ˇ1 Z 1
ˇ ˇ
D if 0 f ˇ C i jf 0 j2 dx C h0 g0 dx C .g0 C ˛h0 /hˇ .g0 C ˛h0 /h0 dx
0 0 0 0 0
Z 1 Z 1
D if 0 .1/f .1/ C if 0 .0/f .0/ C i jf 0 j2 dx C h0 g0 dx C .g0 .1/ C ˛h0 .1//h.1/
0 0
Z 1 Z 1
.g0 .0/ C ˛h0 .0//h.0/ h0 g0 dx ˛ jh0 j2 dx
0 0
Z 1 Z 1 Z 1
0 0 2
D if .0/f .0/ C i jf j dx C h0 g0 dx 0
ikf .0/h.0/ ˛ jh0 j2 dx
0 0 0
Z 1
h0 g0 dx
0
Z 1 Z 1 Z 1 Z 1
D ˛ jh0 j2 dx C i jf 0 j2 dx C h0 g0 dx h0 g0 dx (11)
0 0 0 0
and
Z 1
RehA X; Xi D ˛ jh0 j2 dx 0: (12)
0
Hence A is dissipative and A generates a C0 -semigroup eA t of contractions in H

by the Lumer-Phillips theorem [7]. The proof is complete.
3 Spectral Analysis
In this section, we consider the eigenvalue problem of (3). Let A X D X, where
0 ¤ X D . f ; g; h/ 2 D.A /, then f ; g; h satisfy:
8 00
ˆ
ˆ f .x/ if .x/ D 0;
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ h.x/ D g.x/;
<
.1 C ˛/g00 .x/ 2 g.x/ D 0; (13)
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ f .1/ D g.1/ D 0; f .0/ D kh.0/;
ˆ
:̂ 0
˛h .0/ C g0 .0/ D ikf 0 .0/:
1
Let p./ D 1 C ˛, when p./ ¤ 0 i.e. ¤ , (13) changes to
˛
8 00
ˆ
ˆ f .x/ D if .x/;
ˆ
ˆ
ˆ
ˆ 2 2
ˆ 00
< g .x/ D g.x/ D g.x/;
1 C ˛ p./ (14)
ˆ
ˆ
ˆ
ˆ f .1/ D g.1/ D 0; f .0/ D kg.0/;
ˆ
ˆ
:̂
.1 C ˛/g0 .0/ D p./g0 .0/ D ikf 0 .0/:
We can get
p p q q
2 2
ix p./ x p./ x
f .x/ D a1 e ix
C b1 e ; g.x/ D c1 e C d1 e ; (15)
where a1 ; b1 , c1 and d1 are constants. Substituting these into the boundary conditions
of (14), we have
8 p p
ˆ
ˆ a1 e i C b1 e i D 0;
ˆ
ˆ
ˆ
ˆ q q
ˆ
< 2 2
c1 e p./ C d1 e p./ D 0;
(16)
ˆ
ˆ a1 C b1 D k.c1 C d1 /;
ˆ
ˆ
ˆ
ˆ q
:̂ p./ 2 .c d / D ikpi.a b /:
p./ 1 1 1 1
Then (14) has the nontrivial solution if and only if the characteristic equation
det ./ D 0, where
2 p p 3
e i
e i
0 0
6 q q 7
6 2 2 7
6 0 0 e e p./ 7
6 7
p./
./ D 6 7: (17)
6 1 1 k k 7
6 7
4 p p q q 5
2 2
ik i ik i p./ p./ p./ p./
Lemma 1 Let A be defined by (6). Then for each 2 p .A /, we have Re < 0.
Proof By Theorem 1, since A is dissipative, we have for each 2 .A /, Re 0.
So we only need to show there is not any eigenvalue on the imaginary axis. Let
D ˙i2 2 p .A / with 2 RC and X D . f ; g; h/ 2 D.A / be its associated
eigenfunction of A . Then by (12), we have
Z 1
RehA X; Xi D ˛ jh0 j2 dx D 0:
0
Hence h0 .x/ D 0. By the second and third equations of (13), we have

h.x/ D g.x/ D 0. Then by the first equation of (13) and its boundary conditions
we have:
( 00
f .x/ D if .x/;
f .0/ D f 0 .0/ D f .1/ D 0:
A direct computation yields f .x/ D 0. Hence, X D . f ; g; h/ D 0. Therefore, there

is no eigenvalue on the imaginary axis. This completes the proof.
1
Proposition 1 Let A be defined by (6). Then D … p .A /.
˛
1
Proof When D , then p./ D 0. From the third and second equation of (13),
˛
we can easily get g.x/ D h.x/ D 0. Following the same manner as the proof of
1
Lemma 1, we have f .x/ D 0. Hence, X D . f ; g; h/ D 0. So that D …
˛
p .A /.
Theorem 2 Let A be defined by (6). The eigenvalues of A have the following
asymptotic expressions:
1
(i) When ! , the asymptotic eigenvalue is given by:
˛
p
1 1 ˛A
2n D 2 2 3 C 4 4 5 C O.n5 /; (18)
˛ n ˛ n ˛
where
p
p
2 1 B 1 C B 2i i k2 i
AD p Œ p p e ˛ ; B D : (19)
2i i
B.1 C e ˛ / ˛ ˛ ˛2
(ii) When jj ! 1, there are two families of eigenvalues given by:
" #
2 ln2 r
1n D .n C /j ln rj C .n C / i C O.n1 /; (20)
2 2 4
' 2 ˛ 2 1 h ' i
3n D ˛.n C / C ln r C ˛.n C / ln r i C O.n1 /:
2 4 ˛ 2
(21)
Here r; and ' are three constants given by

p
˛ 2 C k8
0<rD p < 1; ln r < 0; ln r1 D ln r > 0; (22)
˛ C 2˛k2 C k4
8 p
ˆ 2˛k2
< arctan ˛k4 ; ˛ k4 > 0;

D 2; p
˛ D k4 ; (23)
:̂ 2˛k2
arctan k4 ˛ ; ˛ k4 < 0;
and
8 p
ˆ 2˛k2
< arctan k4 ˛ ; k4 ˛ > 0;

' D 2; p
˛ D k4 ; (24)
:̂ 2˛k 2
4
arctan ˛k 4 ; k ˛ < 0:
Moreover,
Re1n ; Re3n ! 1; when n ! 1: (25)
Proof Due to
q space limitation, we give the outline of the proof here. From (17), and
2
let s./ WD p./ 2 C, a direct computation gives,
h p p ih p i
det ./ D e i es./ e i es./ p./s./ C ik2 i
h p p ih p i
C e i es./ e i es./ p./s./ ik2 i : (26)
Let det ./ D 0, asymptotic expressions of the eigenvalues can be achieved.

(I) When ! ˛1 , let " D C ˛1 , then " ! 0. We get 2n as shown in (18)
and (19).
(II) When ! 1, we consider WD i2 ; 0 arg 2 . We then get 1n and 3n
(given by (20) and (21)), when belongs to 0 arg 8 and 8 < arg
3
8
, respectively. There is no high frequency solution in 3
8
< arg 2 .
Proposition 2 Let A be defined by (6). Then its adjoint operator A has the
following form:
8
ˆ
ˆ A . f ; g; h/ D .if 00 ; h; .g0 ˛h0 /0 /; 8 . f ; g; h/ 2 D.A /;
ˆ
ˆ 8 ˇ 0 9
ˆ
< ˇ g ˛h0 2 H 1 .0; 1/;
ˆ
ˆ ˇ >
>
< ˇ =
ˆ
ˆ D.A
/ D . f ; g; h/ 2 H ; A
. f ; g; h/ 2 H ˇ f .1/ D 0; f .0/ D kh.0/; :
ˆ
ˆ ˆ ˇ >
:̂ :̂ ˇ >
ˇ g0 .0/ ˛h0 .0/ D ikf 0 .0/ ;
(27)
˚ 1S
Proposition 3 Let A is defined by (6). Then .A / D ˛ p .A /.
˚
Proposition 4 Let A be defined by (6). Then r .A / D ;, and c .A / D ˛1 .
˚ S
Proof From Propositions 1 and 3, we have ˛1 … p .A / and ˛1 p .A / D
.A /. The desired results will be got if ˛1 … r .A /. Now we suppose ˛1 2
r .A /, then ˛1 2 p .A /. By A X D ˛1 X, where X D . f ; g; h/ 2 D.A /, we
get
8
ˆ 00 i
ˆ
ˆ f .x/ ˛ f .x/ D 0
ˆ
ˆ
ˆ
ˆ
ˆ
ˆ 1
ˆ
ˆ h.x/ D g.x/
ˆ
ˆ ˛
<
1 (28)
ˆ
ˆ .g0 .x/ ˛h0 .x//0 D h.x/
ˆ
ˆ ˛
ˆ
ˆ
ˆ
ˆ f .1/ D g.1/ D 0; f .0/ D kh.0/;
ˆ
ˆ
ˆ
ˆ 0
:̂ ˛h .0/ g0 .0/ D ikf 0 .0/:
From the second and third equation of (28), we get g.x/ D h.x/ D 0. Then f .x/
satisfy
8
< f 00 .x/ i f .x/ D 0;
˛ (29)
:
f .0/ D f 0 .0/ D f .1/ D 0:
Simple computation shows that f .x/ D 0. This implies that X D . f ; g; h/ D 0,

which is a contradiction. So ˛1 … r .A /. The proof is complete.
4 Asymptotic Stability of System (3)
Definition 1 A C0 -semigroup T.t/ is called asymptotically (strongly) stable, if
lim jjT.t/xjj D 0:
t!1
Theorem 3 ([5]) Let T.t/ be a uniformly bounded C0 -semigroup on a Banach

space X and let A be its generator. If
\
.A/ iR c .A/;
and c .A/ is countable, then T.t/ is asymptotically stable.

Theorem 4 Let A be defined by (6). Then the system (3) achieves asymptotic
stability.
Proof Since from Lemma 1 and Proposition 3, we know that when 2 .A /,
Re < 0. So we have
\
1
.A / iR D ; c .A / D ;
˛
and c .A / is obviously countable. Then system (3) achieves asymptotic stability by

Theorem 3.
5 Conclusions
In this paper, we use a wave equation with K-V damping to be a dynamic feedback
controller for a Schrödinger equation. We give the asymptotic expression of the
eigenvalues, and also the exact composition of the spectrum of the system operator.
At last, the asymptotic stability of the system was achieved.
Acknowledgements This work was supported by the National Natural Science Foundation of
China.
References
1. Guo, B.Z., Wang, J.M., Zhang, G.D.: Spectral analysis of a wave equation with Kelvin-Voigt
damping. Z. Angew. Math. Mech. 90(4), 323–342 (2010)
2. Krstic, M., Guo, B.Z., Smyshlyaev, A.: Boundary controllers and observers for the linearized
Schrödinger equation. SIAM J. Control Optim. 49(4), 1479–1497 (2011)
3. Liu, K., Liu, Z.: Exponential decay of energy of the Euler-Bernoulli beam with locally
distributed Kelvin-Voigt damping. SIAM J. Control Optim. 36, 1086–1098 (1998)
4. Liu, K., Liu, Z.: Exponential decay of energy of vibrating strings with local viscoelasticity.
Zeitschrift für angewandte Mathematik und Physik ZAMP 53, 265–280 (2002)
5. Luo, Z.H., Guo, B.Z., Morgul, O.: Stability and Stabilization of Infinite Dimensional Systems
with Applications. Communications and Control Engineering Series. Springer, London (1999)
6. Machtyngier, E.: Exact controllability for the Schrödinger equation. SIAM J. Control Optim.
32, 24–34 (1994)
7. Pazy, A.: Semigroups of Linear Operators and Applications to Partial Differential Equations.
Springer, New York (1983)
8. Wang, J.M., Krstic, M.: Stability of an interconnected system of Euler-Bernoulli beam and heat
equation with boundary coupling. ESAIM: Control, Optim. Calc. Var. 21(4), 1029–1052 (2015)
9. Wang, J.M., Ren, B., Krstic, M.: Stabilization and Gevrey regularity of a Schrödinger equation
in boundary feedback with a heat equation. IEEE Trans. Autom. Control 57, 179–185 (2012)
Molecular-Dynamics Simulations Using Spatial
Decomposition and Task-Based Parallelism
Chris M. Mangiardi and R. Meyer
Abstract This article discusses the implementation of a hybrid algorithm for

the parallelization of molecular-dynamics simulations. The hybrid algorithm com-
bines the spatial decomposition method using message passing with a task-based,
threaded approach for the parallelization of the workload. Benchmark simulations
on a multi-core system and an Intel Xeon Phi co-processor show that the hybrid
algorithm provides better performances than the message-passing or threaded
approaches alone.
1 Introduction
Molecular-dynamics (MD) is a computer simulation method that is widely used in

computational physics, chemistry and materials science. The method is described in
detail in Refs. [1, 3]. Since MD is frequently used to perform large-scale simulations
on high-performance computers, it is important to develop MD algorithms that make
the best use of modern computing architectures.
This article discusses a hybrid approach that uses a two-level approach for
the parallelization of MD simulations. The first level is based on the spatial
decomposition method [10] and is implemented with the Message-Passing Interface
(MPI) [6]. The second parallelization level employs the cell-task method for the
parallelization of the work-load within the spatial domains and is implemented with
Intel’s Threading Building Blocks Library [11].
If a MD simulations uses a short-range force model and the simulated system
is sufficiently homogeneous, the well-known spatial decomposition method [10]
provides an effective means for the parallelization of the simulation. In this method,
the system being simulated is divided into equally sized and shaped sub-volumes,
wherein each of these sub-volumes are processed by a processor, in order to
improve the performance and reduce the time required to run the simulation. These
sub-volumes of the system, however, require utilizing message passing in order
to communicate their particles’ data to neighbouring sub-volumes. The spatial
C.M. Mangiardi () • R. Meyer

Laurentian University, Sudbury, Canada

134 C.M. Mangiardi and R. Meyer
decomposition method works best when the particles are evenly distributed between
sub-volumes, as in homogeneous systems. If the system is inhomogeneous, load
balancing problems occur and the parallel efficiency is reduced.
The cell-task method [7–9] is made to work well for inhomogeneous systems,
wherein the work load would not be evenly balanced with the spatial decomposition
method. This method works in a similar manner to the spatial decomposition
method, in that it splits the system into sub-volumes; although for this method
these sub-volumes are the width of the interaction range (Verlet-radius), and there is
typically thousands of these sub-volumes. Each of these sub-volumes are scheduled
to run on a processor core in an order wherein no two of the sub-volumes currently
being simulated interact with the same particles, in order to remove the requirement
of cache blocking. This method, unlike spatial decomposition, does not require the
use of message passing, since the design is for a shared-memory system.
The hybrid method discussed in this article utilizes both of these methods, by
first dividing the simulation into large sub-volumes for distribution to separate
processors, as in the spatial decomposition method. After this point, the cell-task
method is used to create many smaller sub-volumes of the system for dynamic
scheduling across the processor cores. Communication between the larger sub-
volumes is still required for particles near the border with the other subsections.
The primary rationale for the implementation of the hybrid method is that it
extends the range of the cell-task method to more than one compute node. However,
even on a single node, the hybrid approach can be advantageous. While the cell-task
method is more efficient for inhomogeneous systems, the situations is less clear
for homogeneous systems where spatial decomposition works well. In this case,
the situation depends on system details since the overhead of the task management
in the cell-task method competes with the communication overhead of the spatial
decomposition method. Further, the spatial decomposition approach may have a
slight advantage through a more localized memory access pattern. In this situation,
a hybrid approach can lead to performance enhancements as it allows to interpolate
between the cell-task method and spatial decomposition.
This work focuses on the performance of the hybrid model on a single compute
node. In Sect. 3, results from a series of benchmark simulations on two test machines
are presented. The results show that the hybrid method enables performance gains
compared to the pure spatial decomposition or cell-task approaches.
2 Benchmark Procedures
The parallelization methods are tested using a system with multi-core processors
and a separate system with an Intel Xeon Phi co-processor. The multi-core compute
nodes contain two Intel Xeon E5-2680 processors, each with eight cores using the
AVX instruction set architecture. The Intel Xeon Phi co-processors utilized are the
5110P model, with 60 physical cores, using hardware based threading for a total
MD Simulations Using Spatial Decomposition and Task-Based Parallelism 135
of 240 threads (details on the Xeon Phi many-core architecture can be found in
Ref. [4]).
For the purposes of this work, the measurement being utilized is the parallel
speedup factor. This represents the number of times faster than the baseline run the
current run is, and is measure as the baseline time divided by the current run time.
On both systems, a variety of combinations of threads and MPI ranks are used, in
order to gauge and compare performances. Speedups are measured against baseline
runs using a single MPI rank and a single thread. In addition to the simulations
using the hybrid method, all MPI runs were carried out using a single thread per
MPI rank, and all threaded runs on the multi-core system using a single MPI rank.
On the Xeon Phi, all threaded runs were based upon a recompiled version of the
code without the MPI overhead, including its baseline run, as the speedups on this
system are affected by the MPI overhead; conversely, this overhead does not effect
the speedups on the multi-core system. On the multi-core systems, tests were then
done using one, two, four, eight and sixteen MPI ranks, and varying numbers of
threads to the total number of cores available per system. On the Xeon Phi co-
processor, tests employed one, two, four, eight, sixteen, thirty, sixty, one-hundred
twenty, and two-hundred forty MPI ranks, with varying numbers of threads.
All simulation runs, including the baseline, used vectorized implementations of
the potentials [9] for better performance. All tests were run for 1000 time steps (2
femtoseconds per step), with the Verlet-lists (neighbour-lists) regenerated every 10
time steps. Each set of tests were run five times, taking the average time of each, in
order to more accurately measure the time required to perform the simulation on a
given system.
A number of simulation systems are used to ensure results extend beyond a single
system. A bulk copper system with approximately 4.5-million particles, Cu (bulk),
is used for its homogeneity, which works well with spatial decomposition. A porous
copper system with two-million particles, Cu (porous), is used as it highlights
the advantages of the cell-task method. Both these systems use the tight-binding
potential [2], which has moderate force calculations. Additionally, an iron system
with four-million particles, Fe (bulk), using the Mendelev potential [5] is included,
as it has more complex force calculations and is less sensitive to memory access.
Lastly, a liquid silver system with four-million particles, Ag (liquid), using the
Lennard-Jones potential [1] is used due to its fast force calculations which makes it
highly sensitive to memory access speed.
3 Results
As an example for all model systems, Fig. 1 shows the Cu (bulk) system’s speedups
using different numbers of processor cores on the multi-core system. For the hybrid
method, only the data obtained with two MPI ranks is shown. The speedups with
four and eight ranks are nearly identical to those obtained with two ranks and have
been omitted to avoid crowding of the figure.
16
Threaded
MPI
Hybrid (2 MPI ranks)
12
Speedup S(p)
0
0 4 8 12 16
Processing Cores p
Fig. 1 Parallel speedups S in simulations of the Cu (bulk) system on the multi-core system as a
function of the number of processing cores used p. The dashed line indicates the ideal speedup
S.p/ D p
Table 1 Best parallel MPI Threaded Hybrid

speedup factors in speedup speedup speedup
simulations involving System (ranks) (threads) (ranks threads)
different systems on the
Cu (bulk) 13.1 (16) 11.2 (16) 13.5 (2 8)
multi-core processor
Cu (porous) 7.9 (16) 12.6 (16) 13.6 (2 8)
Fe (bulk) 13.1 (16) 10.4 (16) 13.2 (2 8)
Ag (liquid) 7.6 (16) 7.4 (16) 4.7 (8 2)
The best speedup factors, together with the number of ranks or threads where they
were obtained, are shown in Table 1 for all test systems on the multi-core machine.
Data is shown for spatial decomposition method (MPI), the cell-task approach
(threaded) and the hybrid method utilizing the optimal combination of ranks and
threads.
Figure 2 shows the Cu (bulk) system’s speedups on the Intel Xeon Phi system. To
avoid cluttering, only data for the cases of eight and thirty MPI ranks are shown for
the hybrid method. The best results for the hybrid method were attained using thirty
ranks, however, as seen with the eight ranks, the behaviour of the speedups are not
monotonous, but instead are dependent upon the combination of ranks and threads.
The best speedup factors with the corresponding numbers of ranks and threads are
summarized for all test systems in Table 2.
From Fig. 1 and Table 1 it can be seen that when utilizing the tight-binding
potentials on the multi-core system, the hybrid method produces better speedups
than either the spatial decomposition or cell-task methods alone. For the homo-
120
Threaded
MPI
90
Speedup S(p)
60
30
0
0 60 120 180 240
Processing Cores p
Fig. 2 Parallel speedups S in simulations of the Cu (bulk) system on the Intel Xeon Phi system as
a function of the number of processing cores used p
Table 2 Best parallel MPI Threaded Hybrid

speedup factors in speedup speedup speedup
simulations involving System (ranks) (threads) (ranks threads)
different systems on the Intel
Cu (bulk) 98.8 (240) 117.7 (230) 116.1 (30 8)
Xeon Phi co-processor
Cu (porous) 22.8 (240) 86.6 (185) 71.6 (4 40)
Fe (bulk) 100.6 (240) 124.1 (240) 118.5 (16 15)
Ag (liquid) 81.3 (240) 79.2 (240) 88.4 (120 2)
geneous system, the spatial decomposition method works well, allowing particles
to be evenly divided amongst the different cores. The task-based method similarly
works well, but shows a slower performance then the other two methods for this
test system. This is in contrast to previous results [7, 8] where the performance of
spatial decomposition and the cell-task method was very similar for homogeneous
systems. A likely reason for this discrepancy is the larger size of the test systems in
this work. Since the generation of the task schedule for the cell-task method does not
parallelize well, and the number of tasks grows with the system size, larger systems
might shift the balance towards spatial decomposition.
As shown by Fig. 2 and Table 2, the behaviour is different on the Xeon Phi co-
processor where the cell-task method and the hybrid approach outperform spatial
decomposition for the Cu (bulk) system. This is in agreement with the design of this
architecture which favours threaded programs and is not optimal for large numbers
of MPI ranks [4].
When utilizing the Cu (porous) system, the speedups of the hybrid method are
more noticeable as compared to the spatial decomposition method. Due to the
system being inhomogeneous, the spatial decomposition method does not evenly
divide the system’s particles amongst the processor cores. This allows for both
the threaded and hybrid approaches to significantly outperform this method. When
utilizing the Intel Xeon Phi, the system is divided into 240 areas with the spatial
decomposition method, resulting in a large portion of the co-processor cores being
idle, waiting on other cores to finish their work. This is significantly alleviated by
the hybrid method, which divides the system into less areas, and uses the task-
based approach to reduce the number of idle cores. The cell-task method, however,
outperforms the hybrid method on the Xeon Phi, as it does not split the system
into any areas, which allows it to better allocate cores to reduce the effects of the
inhomogeneity.
For the Fe (bulk) system both the hybrid and cell-task methods outperform the
spatial decomposition method. Despite the spatial decomposition method having a
smaller portion the system to work on for each area, as compared to the hybrid
method, the overhead associated with MPI and the inter-process communication
counteracts the speed improvements attained calculating the forces. This effect is
reduced significantly in the hybrid method, as it is uses less MPI ranks therefore
reducing the overhead of the spatial decomposition method. Further, the cell-task
method has larger speedups compared to the other methods for this system. This is
most likely due to the higher complexity of the force calculations for the Mendelev
potential. With more time spent per force calculation, the impact of limiting factors
like memory access time is reduced.
The Ag (liquid) system performs significantly slower using the hybrid method
compared to the other methods on the multi-core system. Due to this being a
liquid system, particles are able to more freely move about the system, and as a
result, there is a high amount of inter-process communication required. Further,
the overhead of creating the schedule utilizing less threads further degrades the
performance of the hybrid method. The Lennard-Jones potential is also simpler than
other potentials, reducing the amount of time spent in the calculations of forces, and
instead emphasizes other sections of the program, such as the Verlet-list generation,
scheduler, and inter-process communication. On the Xeon Phi system, however, the
hybrid method performs significantly better than the other two methods on the Ag
(liquid) system. The different behaviour of this test system on the two machines is
not fully understood and requires more studying.
4 Conclusions
The results of the benchmark simulations shown in the previous section demonstrate
that the hybrid method utilizing both spatial decomposition and task-based paral-
lelism, is able to provide better performance on molecular-dynamics simulations
than either method alone. The hybrid method takes advantage of each method’s
benefits, whilst minimizing their drawbacks.
For homogeneous systems spatial decomposition works very well and often
delivers, in particular on multi-core systems, a better performance than the cell-
task method. The Intel Xeon Phi hardware and system software, however, favours
threaded programming and is not designed to work optimally with 240 MPI
ranks. For simulations of homogeneous systems, the hybrid model provides the
means to optimize the performance by reducing the number of threads and MPI
communications.
By its design, the cell-task method outperforms spatial decomposition for
inhomogeneous systems since it avoids the load balancing problems that occur
with the spatial decomposition method when the number of particles varies between
domains. In principle, the hybrid method suffers from the same issue, however since
the hybrid methods operates with larger domains, inhomogeneities may average
out. By reducing the task management overhead of the cell-task method, the hybrid
approach can therefore in certain cases still improve the total performance.
An advantage of the hybrid parallelization method described in this work is that
it makes the simulation program adaptable. By choosing the combination of MPI
ranks and threads, the speed of the simulation can be optimized for the specific type
of the simulated system and the computer system on which the simulation is run.
Overall, the performance gains of the hybrid method are much more pronounced
on the Xeon Phi co-processor than they are on the multi-core system. This
emphasizes that novel many-core architectures require the development of new
algorithms to maximize the performance.
Acknowledgements This work has been made possible by generous allocation of computer time
on computer systems managed by Calcul Québec, the Shared Hierarchical Academic Research
Network (SHARCNET) and Compute/Calcul Canada. Financial support by Laurentian University
and the Natural Sciences and Engineering Research Council of Canada (NSERC) is gratefully
acknowledged.
References
1. Allen, M.P., Tildesley, D.J.: Computer Simulations of Liquids. Clarendon, Oxford (1987)
2. Cleri, F., Rosato, V.: Tight-binding potentials for transition metals and alloys. Phys. Rev. B
48(1), 22–33 (1993)
3. Frenkel, D., Smit, B.: Understanding Molecular Simulation. Academic, San Diego (2002)
4. Jeffers, J., Reinders, J.: Intel Xeon Phi Coprocessor High Performance Programming. Morgan
Kaufman, New York (2013)
5. Mendelev, M.I., Han, S., Srolovitz, D.J., Ackland, G.J., Sun, D.Y., Asta, M.: Development
of new interatomic potentials appropriate for crystalline and liquid iron. Philos. Mag. 83(35),
3977–3994 (2003)
6. Message Passing Interface Forum: http://www.mpi-forum.org/ (2016)
7. Meyer, R.: Efficient parallelization of short-range molecular dynamics simulations on many-
core systems. Phys. Rev. E 88(5), 053,309 (2013)
8. Meyer, R.: Efficient parallelization of molecular dynamics simulations with short-ranged
forces. J. Phys.: Conf. Ser. 540(1), 012,006 (2014)
9. Meyer, R., Mangiardi, C.M.: Parallelization of molecular-dynamics simulations using tasks.

MRS Proc. 1753, mrsf14–1753–nn10–09 (2015)
10. Plimpton, S.: Fast parallel algorithms for short-range molecular dynamics. J. Comput. Phys.
117(1), 1–19 (1995)
11. TBB official website: http://threadingbuildingblocks.org/ (2016)
Modelling of Local Length-Scale Dynamics
and Isotropizing Deformations: Formulation
in Natural Coordinate System
O. Pannekoucke, E. Emili, and O. Thual
Abstract We propose an algorithm to model anisotropic correlation functions using

an approach based on the deformation of locally isotropic ones. In a previous work,
a set of equations that allow to calculate the desired deformation was derived for a
flat coordinates system. However this strategy is not adapted for curved geometry
as the sphere (regional and global atmospheric models), where it is suitable to state
the local isotropy in terms of the local Riemannian metric.
This paper introduces the theoretical background to deal explicitly with natural
coordinate systems leading to a formulation adapted with the Riemannian metric. It
results that the isotropizing deformation is obtained from the resolution of coupled
non-linear equations depending on the geometry. This procedure is illustrated within
a 2D setting.
1 Introduction
In variational data assimilation, the analysis state minimizes the cost function
J.X / D .X X b /T B1 .X X b / C .Y o HX /R1 .Y o HX /; (1)

T
where X 2 Rn is the state vector, Y o 2 Rp the observational vector, B D EŒ"b "b
denotes the covariance matrix of the background error "b D X b X t , X b (X t )
O. Pannekoucke ()
CERFACS/CNRS URA 1875, Toulouse, France
CNRM/GAME, Météo-France/CNRS UMR 3589, Toulouse, France
INPT-ENM, Toulouse, France
E. Emili
O. Thual
Université de Toulouse; INPT, CNRS; IMFT; F-31400 Toulouse, France

142 O. Pannekoucke et al.
is the background state (the true state), R D EŒ"o "oT denotes the covariance matrix
of the observational error "o D Y o HX t , and H is the observation operator
(assumed linear here) that maps the model state to observation locations. For large
dimensional problems, like meteorological applications, the covariance matrix B
cannot be explicitly represented and is often modelled.
Preliminary works have shown that it is possible to estimate and to use
the correlation length-scale, to produce a coordinate change that facilitates the
modelling of the anisotropic covariance matrix B. The approach, developed in [1],
relies on the local metric tensor gx defined from the Taylor expansion of a correlation
function .x; x C ıx/ D 1 12 jjıxjj2gx C O.jıxj3 /, with jjxjj2E D xT Ex (this
formalism can be extended in the particular case where the gradient of is not
zero at x D y). This local definition for gx serves to define a metric field g over a
domain (thereafter the two notations Gx and G.x/ are used to denote the value of a
quantity G at point x). In particular, isotropic correlation functions in Rn takes the
form iso .u1 ; u2 / D iso .jju1 u2 jjRn /, where jj jjRn denotes the Euclidian norm
and where .u1 ; u2 / are two points in Rn . Considering a Riemannian manifold M
embedded in Rn and denoting by ! the metric induced by the Euclidian metric of
Rn , then the restriction of iso to the manifold defines correlation functions on M
[2]. The local behavior of these functions is given by .x; x C ıx/ D iso jjıxjj!x .
Then, the local metric tensor gx of at point x is often expressed in terms of the
Riemannian metric !x as gx D L12 !x , where Lh is the constant length-scale, as
h
r2

2L2h
encountered for, e.g. iso .r/ D e . By analogy with correlation modelling on the
sphere, we say that the correlation functions on M are isotropic.
As an example of Riemann manifold, we can consider the sphere of radius
a, parametrized thanks to the longitude/co-latitude coordinate system x D .; /
and equipped with the natural metric ! defined from the square arc-length ds2 D
!ij dxi dxj (in Einstein’s summation convention) with ds2 D a2 .sin /2 d 2 C a2 d 2 .
The local metric tensor is a very attractive quantity. It can be diagnosed from
ensemble estimation [3], associated with a filtering step to damp spurious sampling
noise [4]. This local metric is often associated with the local diffusion tensor x [5],
defined by ij D 12 gij where gij denotes the inverse tensor of g, i.e. gij gjl D ıil where
ıij denotes the Kronecker symbol.
In data assimilation it is usual to find anisotropic correlation functions which
require sophisticated covariance models, e.g. the wavelets formulation [6], the
recursive filter formulation [7] the diffusion equation formulation [8, 9] or the
coordinate change of isotropic correlations [10]. This coordinate change is a tricky
way to reduce the numerical cost or facilitate the parallelization of the algorithm
[1]: ones the computation of the isotropizing coordinate change has been done,
anisotropic correlations simply result from low cost interpolations between the two
coordinate systems where efficient algorithms exist. For example a locally isotropic
correlation field can be efficiently obtained by alternate applications of diffusion
equation method, parallelized in the space dimension. Thereafter the application of
the coordinate change consists in introducing a coordinate change, able to transform
Modelling of Local Length-Scale Dynamics and Isotropizing Deformations:. . . 143
isotropic correlation into the desired anisotropic correlation. The coordinate change
can be estimated from ensemble method, e.g. [11] has proposed a procedure that
relies on the wavelet estimation of the gradient of the deformation [12]. Legrand
and Michel [13] illustrated the feasibility of the isotropization procedure for real
data, and the potential for data assimilation. Until now, the general framework that
leads to a local isotropy for the curved spaces has not been addressed.
In [1] the coordinate change is obtained from the local metric tensor g.x/: it
consists in finding a differential map x.Qx/ that transforms a coordinate system xQ ,
where the local metric tensor gQ .Qx/ is isotropic, into a coordinate system, where the
local metric g.x/ is the anisotropic diagnosed one, and respecting the chain rule
gQ ˛ˇ D @xQ˛ xi @xQˇ xj gij : (2)
In particular, it has been shown possible to construct a coordinate change leading

to a local isotropic tensor gQ ˛ˇ .Qx/ D L21.Qx/ ı˛ˇ .Qx/. This coordinate change is found in
two steps:
(a) first one has to find x.Qx/, solution of the coupled non-linear Poisson’s like
equations
xi C
jki @xQ˛ xj @xQ˛ xk D 0; (3)
where is the Laplacian operator, and

jki D gil
jkl denotes the Christoffel
symbols of the second kind associated with the metric g defined from the
Christoffel symbols of the first kind
ijk D .@i gkj C @j gkj @k gij /=2;
(b) in a second step, the appropriate coordinate change xQ .x/ is obtained as the
inverse of the differential map x.Qx/.
For general manifold, the metric ıij is not the one found in natural coordinates.
As mentioned above, on a part of the sphere, the metric ıij must be replaced by the
natural metric !ij and a global isotropic correlation function is locally expressed in
terms of !ij so that gQ ij .Qx/ D L12 !ij .Qx/ where Lh is a constant length-scale. The local
h
isotropic version is thus gQ ij .Qx/ D L21.Qx/ !ij .Qx/, where L.Qx/ is a length-scale field. It
follows that the metric !ij has to be taken into account in the isotropizing process.
The aim of the present contribution is to extend the results from [1], by taking
into account natural coordinate systems. We first introduce a geometrical framework
for the isotropization issue in Sect. 2, where the isotropizing deformation appears as
an harmonic map, solution of a system of non-linear partial differential equations.
The isotropization procedure proposed is then tested into a simplified 2D setting in
Sect. 3. The conclusions are reported in Sect. 4.
2 General Formalism for the Isotropization Procedure
In this section we describe the theoretical background for constructing the isotropiz-
ing coordinate transform. This relies on Riemannian geometry [14, 15] and in
particular on the properties of harmonic maps [16].
2.1 Coordinate Change Equation Set
The geographical domain considered here is assumed to be represented by a first

Riemannian manifold M of dimension d and equipped with the metric !, it is
denoted by .M; !/. The diagnostic of the local metric tensor field g endows M
with a second Riemannian structure, .M; g/. A third Riemannian structure .M; gQ /
is considered, where the metric gQ takes the form gQ D L12 ! with L is an unknown
length-scale field. The isotropizing coordinate change takes the form of a differential
map x.Qx/ from .M; gQ / to .M; g/ so that gQ and g are related according to Eq. (2). From
a theoretical point of view, x.Qx/ can be considered as an isometric diffeomorphism
from .M; Q gQ / to .M; g/, where M Q is isomorphic to M. Note that when considering a
differential map f from M Q to .M; g/, it is possible to define a third metric f g on MQ
by . f g/˛ˇ D @xQ˛ f i @xQˇ f j gij . Then, Eq. (2) also reads gQ D x g.
It can be shown from transformation rules of the Christoffel symbols of the

second kind [15]
Q˛ˇ D @xQ˛ xj @xQˇ xk @xi xQ
jki C @2xQ˛ Qxˇ xi @xi xQ , that the coordinate
functions xi are solutions the system of coupled non-linear equations
gQ xi C
jki @xQ˛ xj @xQˇ xk gQ ˛ˇ D 0; (4)
where gQ denotes the Beltrami’s Laplacian with

jki (respectively
Qjki ) are the
Christoffel symbols of the second kind associated with the metric field g (respec-
tively gQ ).
Hence, Eqs. (4) provides a necessary condition verified by the isotropizing
coordinate change, and can be used to find the unknown coordinate change.
However multiple solutions can exist, which means that several length-scale fields
could exist for a given metric g. A major difficulty is that these equations directly
make use of the unknown metric gQ .
To better clarify these two statements we introduce a functional which is related
to the energy of the deformation x.Qx/.
A differential map, solution of Eq. (4), is an harmonic map [14, 16], that is a
stationary point (critical point), for variations R of the differential mappx.Qx/, of the
energy functional [17], EŒx.Qx/; gQ ; g D MQ 21 jjdxjj2 d M, Q where dM Q D gQ dQx1 dQxd
is the invariant volume element of M Q and jj jj denotes the Hilbert-Schmidt norm.
This implies that the coordinate notation for the energy density exQ Œx.Qx/; gQ ; g D
1 2
x/; gQ ; g D 12 gQ ˛ˇ @xQ˛ xi @xQˇ xj gij . In particular, Eq. (4) are the Euler-
2 jjdxjj reads exQ Œx.Q
Lagrange of the energy E for variation of the differential map x.Qx/ [14]. Note that
EŒx.Qx/; gQ ; g is a quadratic functional but since it is not convex in x it may have
multiple critical points.
2.2 Conformal Equivalence Classes
For covariance modelling in data assimilation, a simplified decomposition of linear

operators can be introduced [6, 8, 18, 19]. For instance the background error
covariance matrix B as specified in the non-separable spectral decomposition
1=2
takes the form B D ˙ S1 Dh Ehv Dh ST ˙ T where linear operators are as
T=2
follows: Ehv encodes the horizontal-vertical correlations, Dh encodes the horizontal

correlations, S is the grid point to spectral transform operator and ˙ encodes
the grid point standard-deviation [18]. The separable formulation based on the
diffusion equation [8, 9, 20] takes the form B D ˙ S1 Lh Lv W1 LTv LTh ˙ T where
the linear operator W is the metric tensor, Lh and Lv encode the horizontal and the
vertical diffusion time integration with the diffusion tensor deduced from objective
estimation [3, 9, 20]. Note that, [21] has proposed a non-separable formulation of the
diffusion formulation that relies on the wavelets. It results, from the linear operator
decomposition described above, that the horizontal correlation can be treated with a
linear operator that is 2D for a given level.
A particular property of the 2D case is that the energy functional is conformaly
invariant in gQ , i.e. it remains constant when replacing gQ by L.Q1x/2 !, i.e. EŒx; gQ ; g D
EŒx; !; g (Weyl transformation and invariance, see [15]). Thanks to this invariance,
the unknown length-scale field L can be eliminated by solving the problem within
the conformal equivalent class of !. It results that, for two dimensional problems,
one can replace Eq. (4) by equations
! xi C
jki @xQ˛ xj @xQˇ xk ! ˛ˇ D 0; (5)
this can also be directly obtained from Eq. (4) when replacing gQ by L2 ! in 2D. In
the particular case where !ij D ıij , Eq. (5) leads to Eq. (3) as previously found in
[1].
The unknown metric gQ can be explicitly deduced from the solution x.Qx/ of Eq. (5)
thanks to the metric change Eq. (2) that reads gQ D x g, and then provides the length-

1
1
1=2
scale field L.Qx/ D 2 Trace x gxQ !xQ .
We are now able to describe the isotropizing procedure.
2.3 Isotropizing Procedure
The two step isotropizing procedure, detailed in Algorithm 1, is similar to the one
described in [1], using a finite difference scheme for spatial derivative, a forward
Euler time scheme, and a spline bi-cubic interpolation. First it consists in computing
the inverse isotropization transformation, obtained by solving Eq. (5). For that
purpose the pseudo-time diffusion scheme (the heat flow) [16]
@ xi D ! xi C
jki @xQ˛ xj @xQˇ xk ! ˛ˇ ; (6)
can be employed to find the stationary state, solution of Eq. (5), where the initial
condition is the identity map id.Qx/ D xQ . This defines a family of differential
map x indexed by 2 Œ0; 1/, continuously dependent of . For diffeomorphic
compact manifold without boundary, if the algorithm converges toward a stationary
differential solution, then this solution is an harmonic map, and this harmonic map
is continuously obtained from the initial condition as the limit lim !1 x [22].
Note that the dynamics Eq. (6) is the gradient flow associated with the energy. As
/
a consequence, the tendency must be non-positive, dE.x d 0, so that the energy
EŒx .Qx/; !; g is minimized along the path x .
Algorithm 1 Two-steps isotropizing procedure in curved space

Require: The metric fieldg is assumed to be known (e.g. estimated from an ensemble)
1: Compute
ijk .x/ D 12 @j gik .x/ C @k gji .x/ @i gjk .x/
2: Compute
jki .x/ D gil .x/
ljk .x/
3: Set q1 D 500 and k1 D 2
4: Compute
jki .x/ D gil .x/
ljk .x/
5: # Step 1: Pseudo-diffusion
6: xi1 .Qx/ D xQi
7: for q from 1 to q1 do
8: for i from 1 to d do
9: jki .Qx/ D
jki Œx1 .Qx/ (Spline interpolation)

xi2 .Qx/ D xi1 .Qx/ C d ! xi1 C jki @Qx˛ x1 @Qxˇ xk1 ! ˛ˇ .Qx/
j
10:
11: end for
12: xi1 .Qx/ D xi2 .Qx/
13: end for
14: x.Qx/ D x1 .Qx/
15: # Step 2: Inverse transform (fixed-point’s iterations)

16: ı1 .Qx/ D 0
17: for k from 1 to k1 do
18: ı2 .Qx/ D 12 fxŒQx ı1 .Qx/ xQg
19: ı1 .Qx/ D ı2 .Qx/
20: end for
21: xQ .x/ D x 2ı1 .Qx/
The isotropizing coordinate change xQ .x/ is obtained as the inverse of the

ı
differential map x.Qx/. The differential map xQ .x/ is denoted D.
The isotropization procedure is illustrated within a 2D setting in the next section.
Anisotropic correlations are numerically constructed in 2D from globally isotropic

ones with a deformation derived from an arbitrary stream function. This construc-
tion will be used to test the skill of local isotropization methods. The 2D setting
is specified in order to mimic a portion of the sphere, which is often the case of
horizontal coordinates employed in regional models.
3.1 Experimental Set-Up
To validate the isotropizing process, the following route is considered: a Riemannian

manifold .M; !/ is first introduced, that mimics the situation encountered in
atmospheric modelling where the local metric ! is not flat ; a local isotropic metric
gQ , with respect to the metric !, is then produced from a given length-scale field L ;
a deformation D is constructed and used to generate an anisotropic metric g D D gQ
(the pushforward metric of gQ by D). Using the isotropizing process, the challenge
is to find from g and ! – the only information available in real applications – a
coordinate change that transforms g into a local isotropic metric with respect to !.
As a first step,a bi-periodic domain M is considered,
parametrized with a single
bi-periodic chart U D Œ0; 1Œ0; 1; x D .x; y/ discretized in n D 81 points along
the two directions x and y. This coordinate system is similar to an angular chart on
the sphere, and is equipped with the metric !ij defined from the arc-length square

ds D R ˛ C .1 ˛/.sin y/ dx2 C R2 dy2 ;
2 2 2
with ˛ D 0:2 and R D 1000 km, such that the resolution at x D .0:5; 0:5/ is
isotropic with ds 12 km, the length of the domain along x for y Dp0:5 is equal to
R D 1000 km, while the length for y D 0 or for y D 1 is equal to R ˛ 447 km.
The length along y does not depends on the position x and is equal to R.
A deformation D.x/ D x C d.x/ is constructed as follows. First, a wind field
u0 .x/ D k r is introduced, where is a stream function, and it is normalized
so that u D .u; v/ D .u0 =max.ju0 j/; v0 =max.jv0 j//. Then, for each position x, the
geodesic curve x .t/ is computed, starting at x with the velocity Px .0/ D u.x/dt
where dt D 0:05 is a magnitude factor. Denoting by V.t/ the velocity Px .t/,
Fig. 1 Stream function (shading) and displacement field resulting from the geodesic time
integration with initial velocity deduced from the stream current. The arrows indicate the direction
and the intensity of the associated displacement
the geodesic curve is obtained as the time integration

of the geodesic system

n i i
o
dx
dt D V ; dt D ˝jk V V , with initial condition x .0/ D x; V.0/ D u.x/dt
i dV i j k
and where ˝jki are the Christoffel symbols of the second kind associated with the
metric field !. The displacement field is then defined as d.x/ D x .1/ x.
The stream function used for numerical experiment is shown in Fig. 1, with the
displacement field deduced from the velocity field.
3.2 Anisotropic Correlations and Isotropizing Procedure
In a data assimilation framework, the metric g is computed from ensemble esti-

mation, and it can be used within the isotropizing procedure we propose. In this
work, a simulated experiment is designed to validate the isotropizing procedure and
a synthetic anisotropic metric field is generated from the deformation of a locally
isotropic metric field.
For this purpose, locally isotropic correlation functions are considered, specified
by their local metrics gQ .Qx/ D L.Q1x/2 !.Qx/ where the varying length-scale field L
has been arbitrarily fixed, with average value Lh D 50 km. The diffusion tensor
Fig. 2 The initial locally isotropic metric, represented on a regular coordinate system (a), is
deformed under the action of the deformation D leading to anisotropic metric field (b). Then,
the anisotropic metric field represented within the natural coordinate system (c) is diagnosed and
ı
employed in the isotropizing process which provides the isotropizing deformation D (correspond-
ı
ing theoretically to D1 ). The action of D on the anisotropic metric field is the nearly locally
isotropic metric field (d). The coordinate systems are reproduced in gray
associated with gQ is illustrated in Fig. 2a The coordinate change D is then applied

on these functions to obtain anisotropic correlations functions diagnosed by their
diffusion tensor, Fig. 2b.
The isotropization procedure is considered on the metric g (diagnosed within a
natural coordinate system) as illustrated in Fig. 2c. In a first step, the Christoffel
symbols of second kind
jki are computed from g. Then time integration of Eq. (6),
with initial condition x D xQ and time step ı D ıx2 =4 (see [1] for the details), is
achieved to obtain a solution of Eq. (5). In a second step, the inverse transformation
is constructed following the fixed-point procedure ı D dŒx Q ı (valid in curved
geometry), computed similarily to a second-order semi-Lagrangian scheme: the
fixed-point problem ı D 12 dŒxı

Q
is iteratively first solved, then the displacement

field is deduced as as ı D 2ı (see Algorithm 1).
ı
The diffusion tensor of the isotropization of the metric g by D is given in Fig. 2d,
and appears close to the theoretical one in Fig. 2a. Hence, in this experiment, we
have shown possible to find a coordinate system in which the isotropizing metric is
conformal to the natural metric !.
4 Conclusion
We have extended the preliminary work of [1] within a geometrical background

that offers a new insight on the isotropizing issue. Here, we have shown how the
formalism of the isotropization should be stated to incorporate the natural metric on
a geographical domain, as those encountered in atmospheric data assimilation.
In particular, we have shown that, in 2D, the diagnosis of the length-scale (i.e.
the local metric tensor) can be used to find a coordinate change, solution of a
pseudo-diffusion scheme, that provides isotropic local metric. The pseudo-diffusion
scheme used to solve the non-linear system of coupled Poisson’s equations, can be
considered for real application since it relies on classical numerical components
of Numerical Weather Prediction Systems or Chemical Transport Model (semi-
Lagrangian scheme, first-order derivative and Beltrami’s Laplacian).
Further work is to propose and to test more efficient algorithms than the pseudo-
diffusion scheme.
Acknowledgements OP would like thanks Joseph Tapia, Jean-Pierre Otal and Marina Ville, John
Harlim, Tyrus Berry and Dimitrios Giannakis for interesting discussions ; This work was supported
by the French LEFE INSU program and the MACC2 project within the FP7 E.U. reasearch
program.
References
1. Pannekoucke, O., Emili, E., Thual, O.: Modeling of local length-scale dynamics and isotropiz-
ing deformations. Q. J. R. Meteorol. Soc. 140, 1387 (2014)
2. Gaspari, G., Cohn, S.: Construction of correlation functions in two and three dimensions. Q. J.
R. Meteorol. Soc. 125, 723 (1999)
3. Pannekoucke, O., Berre, L., Desroziers, G.: Background error correlation length-scale esti-
mates and their sampling statistics. Q. J. R. Meteorol. Soc. 134, 497 (2008)
4. Raynaud, L., Pannekoucke, O.: Sampling properties and spatial filtering of ensemble
background-error lengthscales. Q. J. R. Meteorol. Soc. 139, 784 (2013)
5. Pannekoucke, O., Massart, S.: Estimation of the local diffusion tensor and normalization for
heterogeneous correlation modelling using a diffusion equation. Q. J. R. Meteorol. Soc. 134,
1425 (2008)
6. Fisher, M.: Background error covariance modeling. In: ECMWF (ed.) Proceedings. ECMWF
Seminar on Recent Developments in Data Assimilation for Atmosphere and Ocean, pp. 45–63
(2003)
7. Purser, R., Wu, W.S., Parrish, D., Roberts, N.: Numerical aspects of the application of recursive
filters to variational statistical analysis. Part I: Spatially homogeneous and isotropic Gaussian
covariances. Mon. Weather Rev. 131, 1524 (2003)
8. Weaver, A., Courtier, P.: Correlation modelling on the sphere using a generalized diffusion
equation (Tech. Memo. ECMWF, num. 306). Q. J. R. Meteorol. Soc. 127, 1815 (2001)
9. Weaver, A.T., Mirouze, I.: On the diffusion equation and its application to isotropic and
anisotropic correlation modelling in variational assimilation. Q. J. R. Meteorol. Soc. 139(670),
242 (2013)
10. Desroziers, G.: A coordinate change for data assimilation in spherical geometry of frontal
structures. Mon. Weather Rev. 125, 3030 (1997)
11. Michel, Y.: Estimating deformations of random processes for correlation modelling in a limited
area model. Q. J. R. Meteorol. Soc. 139, 534 (2013)
12. Clerc, M., Mallat, S.: The texture gradient equation for recovering shape from texture. IEEE
Trans. Pattern Anal. Mach. Intell. 24, 536 (2002)
13. Legrand, R., Michel, Y.: Modelling background error correlations with spatial deformations: a
case study. Tellus 66, 23984 (2014)
14. Jost, J.: Riemannian Geometry and Geometric Analysis. Springer, Berlin (2005)
15. Nakahara, M.: Geometry, Topology and Physics, 2nd edn. Taylor & Francis, New York (2003)
16. Eells, J., Sampson, J.H.: Harmonic mappings of Riemannian manifolds. Am. J. Math. 86, 109
(1964)
17. Eells, J., Lemaire, L.: A report on harmonic maps. Bull. Lond. Math. Soc. 10, 1 (1978)
18. Courtier, P., Andersson, E., Heckley, W., Pailleux, J., Vasiljević, D., Hamrud, M.,
Hollingsworth, A., Rabier, F., Fisher, M.: The ECMWF implementation of three-dimensional
variational assimilation (3D-Var). I: formulation. Q. J. R. Meteorol. Soc. 124, 1783 (1998)
19. Derber, J., Bouttier, F.: A reformulation of the background error covariance in the ECMWF
global data assimilation system. Tellus A 51, 195 (1999)
20. Massart, S., Piacentini, A., Pannekoucke, O.: Importance of using ensemble estimated back-
ground error covariances for the quality of atmospheric ozone analyses. Q. J. R. Meteorol. Soc.
138, 889 (2012)
21. Pannekoucke, O.: Heterogeneous correlation modeling based on the wavelet diagonal assump-
tion and on the diffusion operator. Mon. Weather Rev. 137, 2995 (2009)
22. Jost, J.: Harmonic mapping between Riemannian manifolds. In: Proceedings of the Centre for
Mathematical Analysis, Australian National University (1983)
Post-Newtonian Gravitation
Erik I. Verriest
Abstract Einstein’s field equations relate space-time geometry to matter and

energy distribution. These tensorial equations are so unwieldy that solutions are only
known in some very specific cases. A semi-relativistic approximation is desirable:
One where space-time may still be considered as flat, but where Newton’s equations
(where gravity acts instantaneously) are replaced by a post-Newtonian theory,
involving propagation of gravity at the speed of light. As this retardation depends
on the geometry of the point masses, a dynamical system with state dependent
delay results, where delay and state are implicitly related. We investigate several
cases with Lagrange’s inversion technique and perturbation expansions. Interesting
phenomena (entrainment, dynamic friction, fission and orbital speeds) emerge.
1 Introduction
A post-Newtonian (pN) gravitation (see [7]) is developed, with the main assumption
that gravity cannot act instantaneously at a distance, but has an interaction speed
that is limited by the speed of light, otherwise one would violate relativity [4].
Section 2 establishes the formulas for the retarded potentials in analogy to the
Liénard-Wiechert potentials in electromagnetic theory. From it, the retarded field,
which determines the motion, is obtained by the gradient of the potential in Sect. 3.
This results in a functional differential equation (FDE) with state dependent delay
(See [3]). In particular, interesting new phenomena (entrainment, dynamic friction,
orbital speeds) emerge from this pN theory. In Sect. 4 we discuss the dynamics for
two masses flying apart in a rectilinear way, a (1-D) toy model for a supernova. The
main problem is the implicit relation between state and delay occurring in the FDE
describing the motion. We discuss several cases where the Lagrange’s inversion
technique [5] can be applied to render such a relation explicit [6].
E.I. Verriest ()

Georgia Institute of Technology, Atlanta, GA 30306, USA

154 E.I. Verriest
2 Retarded Gravitational Potential
Let at time t0 the mass distribution be represented by the scalar function .r0 ; t0 /,
the density field. This assumption requires already a notion of simultaneity and
therefore cannot be consistent with the theory of relativity. We postulate that this
density field generates a gravitational field, which assumes at a fixed position,
r, and time t the value due to the superposition of the fields of all infinitesimal
contributions. However the contribution from r0 must have traveled a (straight
line) distance jr r0 j, which takes a propagation time .r0 / D jr r0 j=c. Hence
only the past value, .r0 ; t .r0 //, contributes to the field at r at time t. Again,
general relativity implies the warping of space and time, and hence geodesics (along
which signals propagate) are not straight lines in general. The following will be
approximately valid for weak fields and small speeds relative to c, the speed of
light.
Assumption Given the time varying mass density .r0 ; t0 /, the gravitational poten-
tial at a fixed point r and time t is given by the linear superposition
Z
.r0 ; t .r0 //
.r; t/ D G dV.r0 /; (1)
R3 jr r0 j
where G is the gravitational constant, and V.r0 / denotes the volume element at
position r0 . This equation is coupled with an implicit delay-state relation
c .r0 / D jr r0 j: (2)
If is continuous in its arguments, the field at a point r where .r; / ¤ 0 can be

obtained by taking the integral over an -small ball around r away, since that part
will not contribute to the field by virtue of the asymptotic isotropy.
Let us now assume that the mass distribution is due to a unit point mass in motion,
prescribed by its trajectory, rp .t/. This means that the mass density is given by a
three-dimensional Dirac delta .r0 ; t0 / D ı.r0 rp .t0 //. The gravitational potential
at the fixed point and time .r; t/ is then
Z
ı.r0 rp .tb //
.r; t/ D G dV.r0 /; (3)
R3 jr r0 j
where tb D t .r0 / is the retarded time satisfying (2). We evaluate this integral by
first principles. First, a Dirac delta peaks where its argument equals zero. However,
this information is not sufficient to describe this behavior in its entirety. The rate at
which the zero is approached is also required. To figure this out, let first r00 be the
Post-Newtonian Gravitation 155
solution to r00 D rp .t .r00 // and set r0 D r00 C rQ 0 : Substituting these expressions

in the Dirac delta, we get (I stands for the identity matrix)
1
ı.r00 C r0 rp .t .r00 C rQ 0 /// D ı.Qr0 /: (4)
j det ŒI C rP p .t .r00 //rr0 .r00 /j
Using the identity detŒI C pq D detŒI C qp, and substituting (4) in (3), yields:
Z
ı.Qr0 /
.r; t/ D G dV.r0 /:
R3 jr r0 j j det.1 C rr .r00 /Prp .t .r00 ///j
Taking the gradient w.r.t. r0 of (2) and evaluating at r0 D rp gives
.r rp .tb //>
crr0 D ŒI ;
jr rp .tb /j
so that finally
G
.r; t/ D ˇ ˇ: (5)
ˇ .rrp .tb //> rP p .tb / ˇ
jr rp .tb /j ˇ1 cjrr p .tb /j
ˇ
This result could also have been obtained using the standard trick to derive
the Liénard-Wiechert potentials in some physics texts, by reducing the three
dimensional Dirac to a one-dimensional time Dirac (see [5]).
3 Gravitational Field Due to a Moving Point Mass
The gravitational field is the gradient of the gravitational potential, F.r; t/ D

rr .r; t/, obtained in Sect. 2. Thus
G 1
F.r; t/ D ˇ ˇ rr jr rp .tb /j C
ˇ .rrp .tb //> rP p .tb / ˇ jr rp .tb /j2
ˇ1 cjrrp .tb /j ˇ
ˇ ˇ
G 1 ˇ .r rp .tb //> rP p .tb / ˇˇ
ˇ ˇ rˇ
r ˇ 1 : (6)
jr rp .tb /j ˇ .rrp .tb //> rPp .tb / ˇ2 cjr rp .tb /j ˇ
ˇ1 cjrrp .tb /j ˇ
From rr jr r0 j2 D 2jr r0 j rr jr r0 j and noting that the left hand side is equal to
rr .r r0 /> .r r0 / D 2.r r0 /> rr .r r0 / D 2.r r0 /> ;

„ ƒ‚ …
DI
156 E.I. Verriest
it follows that
.r r0 />
rr jr r0 j D :
jr r0 j
Note that the gradient with respect to a column vector is represented by a row vector.
Similarly, if jv0 j < c
ˇ ˇ
ˇ .r r0 /> v0 ˇˇ .r r0 /> v0 v> r r0
ˇ
rr ˇ1 D rr 1 D 0
rr :
cjr r0 j ˇ cjr r0 j c jr r0 j
Also,

r r0 1 .r r0 /.r r0 />
rr D I :
jr r0 j jr r0 j jr r0 j2
The matrix between the square brackets is a projection operator (projection onto the
plane perpendicular to the vector r r0 ). Putting this all together, the gravitational
field due to a moving mass is
.r rp .tb //> G
F.r; t/ D ˇ ˇC (7)
jr rp .tb /j3 ˇˇ1 .rrp .tb //> rP p .tb / ˇ
ˇ
cjrrp .tb /j

G 1 rP p .tb /> .r rp .tb //.r rp .tb //>
C ˇ ˇ I :
jr rp .tb /j2 ˇ .rrp .tb //> rPp .tb / ˇ2 c jr rp .tb /j2
ˇ1 cjrrp .tb /j ˇ
If the gravitating mass moves in such a way that rP p is parallel to .r rp .tb //, then
the second term in (7) is zero and the field is directed along r rp .tb /. Aligning the
x-axis with this direction, force, position and velocities can be expressed in the x-
coordinate. In particular, the x-component of the force acting on a unit mass particle
at x is
sgn.x xp .tb //G

Fx D ˇ ˇ; (8)
ˇ xP .t / ˇ
jx xp .tb /j2 ˇ1 sgn.x xp .tb // p c b ˇ
where as always, tb denotes the retarded time.

This form displays two deviations from Newton’s law: First the delayed position
of the gravitating particle appears, and second, there is an aberration which depends
on the (delayed) velocity.
It should also be observed that a naive generalization to Newton’s law in the form
sgn.x xp .tb //G

Fx D ;
jx xp .tb /j2
instead of (8) cannot be correct.
3.1 Expression the Field in Terms of the Predicted Position
First, reorganize (7) in the form
G 1 rPp .tb />

F.r; t/ D ˇ ˇ C
jr rp .tb /j2 ˇ .rrp .tb //> rPp .tb / ˇ2 c
ˇ1 jrrp .tb /j c ˇ
G.r rp .tb //>

ˇ ˇ
ˇ .rr .t //> rP .t / ˇ2
jr rp .tb /j3 ˇ1 jrrpp .tbb /j p c b ˇ
ˇ ˇ
ˇ .r rp .tb //> rPp .tb / ˇˇ .r rp .tb //> rPp .tb /
ˇ
ˇ1 C : (9)
jr rp .tb /j c ˇ jr rp .tb /j c
Introducing rbp D rp .tb /C .tb /Prp .tb /, which is the expected position of the particle in
motion at time t, if the velocity of the mass were fixed at the constant vp .s/ D rP p .tb /
for tb .t/ D t .tb / s t, it follows that
ˇ ˇ
ˇ > ˇ
ˇ1 .r rp .tb // rP p .tb / ˇ D ˇ 1
ˇ j1> .r rbp /j;
ˇ jr rp .tb /j c ˇ ˇr rp .tb /ˇ rrp
.rr .t //>
where 1rrp D jrrpp .tb b /j denotes the unit vector in the direction of rrp . Relativity
imposes jPrp j < c, which with the Cauchy-Schwarz inequality reduces (9) to
G >
rr .r; t/ D ˇ ˇ2 .r rbp / : (10)
ˇ > ˇ
jr rp .tb /j ˇ1rrp .r rbp /ˇ
Thus the gravitational force exerted by a particle in motion is directed towards the
predicted position based on a uniform motion given the delayed information (i.e.
delayed position and velocity).
158 E.I. Verriest
v0 x(t − τ ) x(t)
d cτ
x
Fig. 1 Uniform motion
3.2 Field Due to Particle with Uniform Velocity
For a particle moving with uniform velocity, rbp .t/ is the position of the particle at
time t, rbp .t/ D rp .t/. Without loss of generality, we fix r, the observation point, at
the origin, and let the particle move along a line parallel to the x-axis, at a distance d
with uniform velocity v0 . For notational simplicity, the x-coordinate at t D 0 is taken
to be zero. From the geometry of the problem (See Fig. 1): a quadratic equation for
results: .c2 v02 / 2 C 2v02 t .v02 t2 C d 2 / D 0: The field magnitude follows
G..d2 C v02 t.t //2 C d2 v02 2 /

FD q : (11)
.d2 C v02 t.t //2 .d2 C v02 .t /2 /.d2 C v02 t2 /
In Fig. 2 the magnitude is shown as function of time for d D 1, v0 D 1 and c D

1:1; 2 and the Newtonian case (c D 1). As the mass closes in on the origin for
t < 0, the delayed effect is quite pronounced as the speed c is increased. When
d D 0, the delayed gravitational force at time t exceeds Newton’s law by a factor
1 C vc0 if the particle is moving towards the origin, and a factor 1 vc0 when it is
moving away from the origin. The same holds for all d in the limits respectively for
t ! 1 and t ! C1.
3.3 Gravitational Current and Its Entrainment
Let the origin coincide with the observation point, r, and consider an infinitely long
straight line mass moving along its axis, parallel to the x-axis, at a distance d with
uniform velocity v0 . Thus the field at the origin due to a gravitational current is
studied. Consider first a mass-element at p position .x0 ; d/ at time t D 0. From the
geometry
p of the problem (Fig. 1): c .t/ D .x0 C v0 .t .t///2 C d2 : Letting D
2 2
.x0 C v0 tb / C d ; the x-component of the field at the origin at time 0 due this
Fig. 2 Field at the origin due to particle with uniform motion. The symmetric curve corresponds
to the Newtonian case (c D 1)
elementary current element is

.x0 Cv0 tb /2
x0 C v0 tb v0 1 2
Fx .x0 ; v0 / D

2 ; (12)
3 v0 .x0 Cv0 tb / c
1C c
2
1C v0 0 Cv0 tb /
.x
c
where tb is the backward time for t D 0, i.e., tb D .0/: Likewise, the y-component
d v0 d 1
Fy .x0 ; v0 / D

2 : (13)
v0 .x0 Cv0 tb / c
3 1C c 4 1 C v0 .x0 Cv0 tb /
c
Integrating over x0 from 1 to 1 gives the total gravitational force of the mass
current. The Figs. 3 and 4 show for d D 1 respectively the x and y components of
the field at the origin as function of ˇ D v=c. In the static case (v0 D 0) this is
Z 1
Gd 2G
Fx .0/ D 0; Fy .0/ D dx D 2 :
1 .x2 C d2 /3=2 d
This static field is directed perpendicular to the current direction, and points towards
the linear mass. With v0 ¤ 0 it is augmented by a velocity dependent term. For the
160 E.I. Verriest
Fig. 3 Parallel field component as function of v=c for G D d D 1
Fig. 4 Orthogonal field component as function of v=c for G D d D 1
parallel (to the current) component, this additive term behaves linearly near v D 0,
and quadratically for the component directed towards the current line. The effect
of the current is to entrain the surrounding mass. The large scale structure of the
universe is made up of large filaments of galaxy superclusters that may be modelled
as gravitational currents.
Remark One may be tempted to approximate r.t / by its truncated Taylor
expansion in terms of the instantaneous delay. It was found that such a naive use
of Taylor expansions can not yield solutions that are consistent in the Newtonian
limit.
3.4 Dynamic Friction
The results of the previous section imply a phenomenon analogous to dynamic

friction which was analyzed by Chandrasekhar in a classical setting [1]. A lone
star colliding with a galaxy (moving through a uniformly distributed star field) will
slow down. As only relative motion is significant, we may consider the problem as
that of a stationary mass point (say located at the origin) embedded in a steady flow
of mass in the direction of the x-axis. Locally, at O, the problem is isotropic in the
(y,z) plane orthogonal to the motion. Hence Fy and Fz experienced by the star are
zero. There is however a resulting force in the x-direction. As seen in Sect. 3.3, the
star will be entrained by the moving field of stars and follows a law of the form
xR D k.v xP /; for some k which is computable by a more detailed analysis and must
depend on the density of the star field. As before, v is the initial relative speed of
the star field with respect to the star. But this means that the galactic medium acts
as a viscous medium on the star. At equilibrium, the star moves in unison with the
medium.
4 Escape from Gravity
Consider the motion of two equal masses, m, moving symmetrically in one

dimension, hence having the same speed with opposite directions (gravitational
fission, supernova). Let r.t/ and l.t/ respectively be the position of the masses
moving towards the right and the left. The symmetry imposes l.t/ D r.t/. See
Fig. 5.
Assumption Newton’s law Fr .t/ D mRr.t/ holds, where Fr .t/ is force acting on the
mass moving towards the right at time t. The gravitational force felt by the particle
cτ
m m
O
(t) (t − τ ) r(t)
Fig. 5 Gravitational Fission
162 E.I. Verriest
at r.t/ due to the particle moving towards the left leads to the the coupled set
Gm
rR .t/ D
(14)
rP.t .t//
1C c c2 2 .t/
c .t/ D r.t/ C r.t .t//: (15)
We present two solution methods: Lagrange inversion and perturbation expansion.
4.1 Lagrange Inversion
The explicit form of .t/ can be obtained from c .t/ D r.t/ C r.t .t// about a
point r.t0 / D r0 , .t0 / D 0 , by Lagrange inversion [2] provided that r is analytic,
and jPrj < c at time t D t0 , where it holds that c 0 D r0 C r.t0 0 /;
1
X .c.t t0 / C r0 r.t//i
.t/ D .t t0 C 0 / i : (16)
i1
iŠ
With D D d=ds, the coefficients of the series expansion are expressed as

8 9
ˆ
< >
=
1
i D Di1
i > : (17)
:̂ cC r.t0 0 Cs/c 0 Cr0 ;
s sD0
4.2 Solution via Perturbation Expansion
Postulate a solution to equations (14) and (15) as a series in D 1=c.
r.t/ D r0 .t/ C r1 .t/ C 2 r2 .t/ C : : : (18)

2
.t/ D 1 .t/ C 2 .t/ C : : : : (19)

r.tb /
The velocity factor 1 C c
expands to 1 C Pr0 C 2 ŒPr1 rR0 1 C 3 ŒPr2 rR0 2 C
1 .3/ 2
2 r0 1 rR1 1 C Matching the expansions in for the state dependent delay
differential equation and eliminating k , k D 1; 2; : : : gives

rR0 D (20)
4r02
Fig. 6 Position r.t/ as function of t for sub-escape with D 1 (r0 D 1; v0 D 1). The curves
correspond to the Newtonian case (marginal escape), and first and second order perturbations
(which return eventually to r0 )

rR1 D r1 2 rP0 (21)
2r03 4r0
g
rR2 D 3 4 r12 C 2 3 .r1 rP0 C r2 / 2 rP1 (22)
4r0 r0 4r0
::
:
The solutions of order 0 (Newtonian case), and uppto order 1 and 2 are shown in
Fig. 6 for D 4, r0 D 1 and v0 D vesc jNewton D 2=r0 , the escape velocity in
the Newtonian case. In pN theory v0 D vesc jNewton is insufficient for escape. For
D 1, r0 D 1 and v0 D 1, the post-Newtonian escape velocity was computed
p found that for small 1=c the
(up to second order for several values of c and it was
adjusted quantity r vesc2.c/ is close to vesc .1/ D 2. This led to the conjecture:
vesc .1/
1C
c2
vesc .c/ D r vesc .1/ .

2 .1/
vesc
1
c2
5 Conclusions and Beyond
The physical aspect of the paper established post-Newtonian gravitation as a system

with state dependent delay. The mathematics centers around an implicit relation
of delay and state, which is resolved by Lagrange inversion and perturbation
expansions. The field of mass in uniform motion was described in detail. Its
164 E.I. Verriest
integrated form gives rise to the consideration of gravitational currents and their
ensuing entrainment. This led to the emergence of dynamic friction. Finally we
discussed the escape velocity for splitting masses (fission). Lack of space did not
allow the discussion of the Kepler problem in pN theory. As expected, stable orbits
exist, which is not the case had one considered the naive generalization (delayed
Newton’s law) of the gravitational field mentioned in Sect. 3. It should also be
pointed out that naive Taylor expansions lead to results that are inconsistent in the
Newtonian limit.
References
1. Chandrasekhar, S.: Dynamical friction. I. General considerations: the coefficient of dynamical

friction. Astrophys. J. 97, 255–262 (1943)
2. Good, I.J.: Generalization to several variables of Lagrange’s expansion, with applications to
stochastic processes. Proc. Camb. Philos. Soc. 56, 367–380 (1960)
3. Hartung, F., Krisztin, T., Walther, H.O., Wu, J.: Functional differential equations with state-
dependent delays: theory and applications. In: Canada, A., Drábek, P., Fonda, A. (eds.)
Handbook of Differential Equations, vol. 3, pp. 435–545. Elsevier, Amsterdam (2006)
4. Misner, C.W., Thorne, K.S., Wheeler, J.A.: Gravitation. Freeman, San Francisco (1973)
5. Smith, G.: Classical Electromagnetic Radiation. Cambridge University Press, Cambridge/
New York/Melbourne (1997)
6. Verriest, E.I.: Inversion of state-dependent delay. In: Karafyllis, I., Malisoff, M., Mazenc, F.,
Pepe, P. (eds.) Recent Results on Nonlinear Delay Control Systems. Springer International
Publishing Switzerland, 327–346 (2016)
7. Will, C.: The renaissance of general relativity. In: Davies, P. (ed.) The New Physics, pp. 7–33.
Cambridge University Press, Cambridge (1989)
Part II
Mathematical and Computational Methods
in Life Sciences and Medicine
A Quantitative Model of Cutaneous Melanoma
Diagnosis Using Thermography
Ephraim Agyingi, Tamas Wiandt, and Sophia Maggelakis
Abstract Cutaneous melanoma is the most commonly diagnosed cancer and its
incidence is on the rise worldwide. Early detection and differentiation of a malignant
melanoma from benign cutaneous lesions provides an excellent chance for treating
the disease. Thermography is a non-invasive tool that can be used to detect
and monitor skin lesions. We model heat transfer in a skin region containing a
lesion. The model which is governed by the Pennes equation uses the steady state
temperature at the skin surface to determine whether there is an underlying lesion.
Numerical simulations from the model ascertain whether the lesion is malignant or
benign.
1 Introduction
The skin is the largest organ in the body and has the most exposure to the external
environment. Thus, it is prone to lesions attributed to both internal and external
factors. Skin lesions begin when alterations in cellular metabolism allow cells to
grow without restriction. Skin lesions do constitute the majority of all cancers
and their incidence is on the rise [1]. The majority of skin lesions are benign and
harmless. Melanoma is a malignant skin cancer that can easily undergo metastasis
and consequently lead to death if not detected early. Melanoma accounts for the
most cancer deaths in the United States compared to other cancers [2].
One of the many tools for diagnosing skin cancers is thermography [3]. The
procedure is based on established observations that the temperature of the skin
directly above a tumor is significantly higher than the one in the absence of a tumor.
Thermography uses an infrared camera to map the temperature distribution over the
desired skin surface. The FDA in the United States approved thermography as an
adjunct tool for diagnosing breast cancers in 1982 [4]. Technological advances in the
past decades have improved thermographic imaging so that temperature differences
of about 0:025 ıC can be detected [5].
E. Agyingi () • T. Wiandt • S. Maggelakis

Rochester Institute of Technology, Rochester, NY 14623, USA

168 E. Agyingi et al.
The published literature contains mathematical models and simulations illustrat-

ing thermography. The majority of the studies are for breast cancer [6–12] and only
a few are dedicated to skin cancer [13–16]. Most of these papers model tumors that
are vascularized and governed by some version of the Pennes bio-heat equation.
The metabolic heat generating rate of tumors are known to range from 0.205–
1.645 W/cm3 , which is about 20–200 times the metabolic heat generation rates of
normal tissue [6]. Lin et al. [8] considered a tumor metabolic rate of 24:156 W/cm3 ,
which is about 2500–5000 times that of normal tissue. In a more recent paper [17],
the authors studied a prevascular breast tumor, and related the temperature of the
skin surface to the tumor size, tumor depth and the tumor metabolic heat generating
rate.
The most effective treatment for melanoma is surgical incision before metastasis.
Almost all tumors start up as prevascular tumors. The tumor has a single location,
and has no vasculature connecting it to the host capillary network. The aim of this
paper is to investigate whether melanoma can be distinguished from a benign skin
lesion at the prevascular level. To achieve this, we model the tumor as a spheroid,
consisting of a necrotic core and surrounded by a viable region of proliferating
cells. The work reported here builds on the one-dimensional model reported by
[18], which was recently extended in [17]. In [17], it was assumed that the heat
diffusion was the dominant effect and the net effect of perfusion was zero in the
healthy region. This paper differs from [17] in that the boundary conditions take
into account the effect of blood perfusion in the healthy region.
2 Mathematical Model
We present a model of heat transfer in prevascular skin tumors. We assume that the
tumors are spheroids and in the interest of simplicity, we shall present the model
as a spatially two-dimensional model. We consider a two-dimensional cross section
of the skin tissue, containing a circular shaped tumor as depicted in Fig. 1. The
temperature of the cross-sectional domain is given by T.x; y/, where the x coordinate
is the horizontal direction and y is the depth. The two dimensional cut is placed in
the .x; y/-plane so that the y D 0 level corresponds to the bottom layer of the skin,
and y D d corresponds to the surface. The origin is located so that a x a,
with the center of the tumor at .0; d=2/. The radius of the tumor is R.
The equations governing heat flow in each portion of the entire region are derived
from the Pennes equation [19]:
@T
Nc D r .KrT/ C mb cb .TA T/ C S; (1)
@t
where is the tissue’s density, cN is the tissue’s specific heat, K is the tissue’s thermal
conductivity, mb is the mass flow rate of blood, cb is the blood’s specific heat, TA is
the arterial blood temperature, and S is the metabolic heat generation rate.
A Quantitative Model of Cutaneous Melanoma Diagnosis Using Thermography 169
y=d
healthy region viable region
necrotic
x = −a x=a
core
y=0
Fig. 1 Cross-sectional sketch of the skin containing a prevascular tumor
We note that at steady state, the time derivative is zero. Therefore the equation
describing the temperature T.x; y/ in the healthy region is given by
Sh mb cb .TA Th /
Th D ;
Kh Kh
where Sh is the metabolic heat generation rate of the healthy tissue and Kh is its
thermal conductivity.
For the viable tumor region, we remark that the tumor has no vasculature and
therefore the perfusion term in equation (1) vanishes. Consequently, at steady state,
the equation describing the temperature Tt .x; y/ of the viable region of the tumor is
given by
St
Tt D ;
Kt
where St is the metabolic heat generation rate of the tumor and Kt is its thermal
conductivity.
Finally, we note that the necrotic core is comprised of dead cells and generates
no heat. This portion is simply described by the Laplace equation Tc D 0.
The boundary conditions of the various parts of the entire domain are listed
below. The outer boundary conditions of the healthy region are provided by items
(i)(iii). Items (iv)(v) provide inner boundary conditions for the healthy region
and outer boundary conditions for the viable tumor region, and item (vi) provides
the boundary conditions between the viable region and the necrotic core.
(i) At the bottom layer, y D 0, Th .x; 0/ D Tb isˇ the temperature of the body.
hˇ
(ii) At the skin surface, y D d. Hence, K @T @y ˇ
D .Th Ta /, where Ta is the
yDd
ambient temperature and is the surface heat transfer coefficient.
(iii) At the boundaries x D ˙a, Th .a; y/ D Th .a; y/ is the temperature distribution

of healthy tissue. To compute this value, we assume homogeneity of the tissue
and therefore reduce the problem into a one-dimensional equation of the form
Sh mb cb .TA Th /
Th D Th00 .y/ D
Kh Kh
and therefore,
mb cb Th Sh mb cb TA
Th00 .y/ D ;
Kh Kh Kh
where y is the tissue depth, Th .0/ D Tb , and Th0 .d/ D .Th Ta /=Kh . The
solution of this equation is given by
Th .y/ D Tb C cosh.! y/

.Ta Tb / K! sinh.! d/ cosh.! d/
C sinh.! y/
sinh.! d/ C Kh ! cosh.! d/
q
where D Tb mbScb TA and ! D mKbhcb .
(iv) We assume that the temperature is continuous across the interface of the
healthy tissue and the viable region of the tumor, i.e. Th .x; y/ D Tt .x; y/ when
x2 C .y d=2/2 D R2 .
(v) The heat flux is continuous across the interface of the healthy tissue and the
viable region of the tumor, i.e. Kt rTt D Kh rTh when x2 C .y d=2/2 D R2 .
(vi) On the interface between the viable region of the tumor and the necrotic core,
we assume continuous temperature and flux as well.
3 Results and Discussion
In this section, we present numerical simulations of the model presented above. The
results are for a single tumor and multiple tumors for a given skin cross section. The
calculations were performed using a MATLAB finite element solver. All thermo-
physical parameter values used were chosen within the range of published data. We
set Sh D 0:009 W/cm3 , Kh D Kt D 0:0042 W/((cm)ı C), D 0:0005 W/((cm2 )ı C),
mb D 0:0005 g/(ml s) and cb D 4:2 J/gı C. We also set the arterial blood temperature
TA D 37 ı C and the body temperature, at the level y D 0, to be Tb D 37 ıC. Other
parameters used were chosen to investigate the behavior of the model.
The first results presented in Fig. 2 are for a single tumor of radius 1 mm and
center located at .0; 0:25/. Figure 2a provides a contour map and temperature
distribution over the entire domain. As expected, we observe that heat diffuses away
from the tumor towards the cooler surrounding region. In Fig. 2b we examine the
a b 36.85
m = 0.001
b
m = 0.0005
b
36.8 mb= 0.00025
mb= 0.0
36.75
36.7
Temperature(T)
36.65
36.6
36.55
36.5
36.45
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Axial Distance (x)
Fig. 2 Numerical simulations of one tumor with radius R D 1 mm and St D 20Sh . The plot in
(a) is the temperature distribution over the cross section and (b) is the skin surface temperature for
different perfusion rates (mb )
steady state temperature at the skin surface for different perfusion rates. We note that
the results in Fig. 2b are almost identical even when the perfusion rate was double
(mb D 0:001) or neglected (mb D 0:0). This suggests that the effect of perfusion is
negligible for prevascular tumors, that is, very small tumors that are close to the skin
surface. This may be attributed to the fact that there are no large blood vessels in
the skin region. The average temperature increase at the skin surface, caused by the
presence of the tumor, was observed to be about 0:33 ı C, which is a very significant
number that can be easily detected using an infrared camera.
Next, we consider two tumors of the same radii 1 mm, with centers located at
.0:4; 0:25/ and .0:0; 0:25/. We investigate two cases; firstly, the tumors have the
same St values, and secondly, one of the values is altered. The results are given in
Fig. 3. Figure 3a represents a 3D steady state temperature profile for tumors with
the same metabolic heat generating rates St D 20Sh . In Fig. 3b the metabolic heat
generating rate of the right tumor was reduced to St D 10Sh. Figure 3c gives the
steady state temperature at the skin surface for the two cases. The results affirm that
a tumor with a higher metabolic heat generating rate will produce more heat and
consequently a better temperature profile at the skin surface.
The next results as presented in Fig. 4 are for two tumors with the same metabolic
heat generating rates (St D 20Sh ) and different radii. The tumor on the left has a
radius of 1 mm with center located at .0:4; 0:25/, while the tumor on the right has
a radius of 0:75 mm with center located at .0:4; 0:25/. A 3D steady state temperature
profile for tumors is given in Fig. 4a and the steady state temperature at the skin
surface is given in Fig. 4b. Here we see that a bigger tumor will produce more heat
compared to a smaller tumor having the same metabolic heat generating rate.
Finally, Fig. 5 illustrates two tumors with different metabolic heat generating
rates and different radii. The bigger tumor (i.e. left tumor) with radius 1 mm and
center located at .0:4; 0:25/ was given a smaller metabolic heat generating rate
St D 5Sh . The smaller tumor (i.e. right tumor) with radius 0:75 mm and center
located at .0:4; 0:25/ was given a higher metabolic heat generating rate St D 20Sh .
Figure 5a shows a 3D steady state temperature profile for tumors over the entire
37.2
37.2
37.1
37.2 37.1 37.2
37 37 37
37
Temperature(T)
Temperature(T)
36.8 36.8
36.9 36.9
36.6 36.6
36.8 36.8
36.4 36.4
36.7 36.7
0.5 0.5
0.4 1 0.4 1
36.6 36.6
0.3 0.5 0.3 0.5
0.2 0 y 0.2 0
y
0.1 −0.5 x 36.5 0.1 −0.5 x 36.5
0 −1 0 −1
(a) Case I (b) Case II

36.85
Same S
h
Different Sh
36.8
36.75
36.7
Temperature(T)
36.65
36.6
36.55
36.5
36.45
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Axial distance (x)
(c)
Fig. 3 Numerical simulation of two tumors of the same radii. The results in (a) for the same St ,
(b) for different St and (c) the temperature profile at the skin surface for cases I and II
b 36.85
37.2 36.8
a
Temperature(T)
37.2 37.1 36.75

Temperature(T)
37 37 36.7
36.8 36.9 36.65
36.6 36.8 36.6
36.4
36.7 36.55
0.5
0.4 1 36.6 36.5
0.3 0.5
y 0.2 0
0.1 −0.5 36.5 36.45
−1 x −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
0
Axial Distance (x)
Fig. 4 Numerical simulations of two tumors with same St D 20Sh and different radii. The plot in
(a) is a 3D temperature distribution over the cross section and (b) is the skin surface temperature
domain, while Fig. 5b provides the steady state temperature at the skin surface. The
results show that a smaller tumor with a high metabolic heat generating rate will
36.7
a 37.1
b
37.2 37
36.65
37
36.9
36.8
36.6
Temperature(T)
36.6
36.8
36.4
36.55
36.2 36.7
0.5
0.4
36.5
36.6
0.3 1
0.2 0.5
0
0.1 36.5
−0.5 36.45
y −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
0 −1 Axial Distance (x)
x
Fig. 5 Numerical simulations of two tumors with different St and different radii. The plot in (a) is
a 3D temperature distribution over the cross section and (b) is the skin surface temperature
produce more heat compared to a bigger tumor with a low metabolic heat generating
rate.
An analysis of how the results of a variant of the model for breast tumors depend
on the underlying parameters was carried out in [17]. Among others, different
values of the depth of the tumor, the ambient temperature, and the internal body
temperature were investigated. A similar analysis yields analogous results for the
current model for melanomas.
3.1 Future Work
There are several directions in which the model presented in this paper can be
extended. The case of multiple tumors can be investigated by considering different
size tumors at multiple, randomly chosen locations. This would give a better
understanding of possible heat signatures. The case of non-spherical tumors can
be explored as well. However, this requires a slight modification of the current
numerical methods. Another aspect of interest is how the sizes of the blood vessels
in the neighborhood of the tumor modify the temperature profile of the tumor at the
skin surface.
The extension of the model from the 2-dimensional domain to 3 dimensions
is under way. The numerical methods have to be adjusted to deal with the added
dimension, but the extended model might discover some new phenomena not
present in the 2-dimensional investigations.
The inverse problem requires a different approach to the model. One possible
extension of our investigation is to set up an optimization problem to identify the
location and size of the tumors for given heat signatures.
4 Conclusion
We have presented a mathematical model for cutaneous melanoma at the prevascular

level based on the Pennes bio-heat transfer equation. We studied the effect of blood
perfusion rate, tumor size and tumor metabolic heat generating rate. The model
suggests that the perfusion rate does not influence the steady state temperature
at the skin surface for prevascular tumors. The size and tumor metabolic heat
generating rate are shown to be very important factors in determining the steady
state temperature distribution of the region under consideration. These two factors
are therefore useful bio-markers that can be used to diagnose whether a skin lesion
is dormant and harmless or malignant. The results of the model associate melanoma
with a tumor that has a high metabolic heat generating rate and to one that is
increasing in size.
References
1. Stern, R.S.: Prevalence of a history of skin cancer in 2007: results of an incidence-based model.
Arch. Dermatol. 146, 279–282 (2010)
2. American Cancer Society: Cancer Facts & Figures. http://www.cancer.org/acs/groups/content/
@editorial/documents/document/acspc-044552.pdf (2015)
3. Estee, L., Psaty, B.A., Allan, C., Halpern, M.D.: Current and emerging technologies in
melanoma diagnosis: the state of the art. Clin. Dermatol. 27, 35–45 (2009)
4. Arora, N., Martins, D., Ruggerio, D., Tousimis, E., Swistel, A.J., Osborne, M.P., Simmons,
R.M.: Effectiveness of a noninvasive digital infrared thermal imaging system in the detection
of breast cancer. Am. J. Surg. 196, 523–526 (2008)
5. Bronzino, J.D.: Medical Devices and Systems. CRC/Taylor & Francis, Boca Raton (2006)
6. González, F.J.: Thermal simulation of breast tumors. Revista Mexicana de Fisica, 53, 323–326
(2007)
7. González, F.J.: Non-invasive estimation of the metabolic heat production of breast tumors using
digital infrared imaging. QIRT J. 8, 139–148 (2011)
8. Lin, Q.Y., Yang, H.Q., Xie, S.S., Wang, Y.H., Ye, Z., Chen, S.Q.: Detecting early breast tumour
by finite element thermal analysis. J. Med. Eng. Technol. 33, 274–280 (2009)
9. Sudharsan, N.M., Ng, E.Y.K., Teh, S.L.: Surface temperature distribution of a breast with and
without tumour. Comput. Methods Biomech. Biomed. Eng. 2, 187–199 (1999)
10. Agnelli, J.P., Barrea, A.A., Turner, C.V.: Tumor location and parameter estimation by
thermography. Math. Comput. Model.: Int. J. 53, 1527–1534 (2011)
11. Mital, M., Scott, E.P.: Thermal detection of embedded tumors using infrared imaging. J.
Biomech. Eng. 129, 33–39 (2007)
12. Paruch, M., Majchrzak, E.: Identification of tumor region parameters using evolutionary
algorithm and multiple reciprocity boundary element method. Eng. Appl. Artif. Int. 20, 647–
655 (2007)
13. Deng, Z., Liu, J.: Mathematical modeling of temperature mapping over skin surface and its
implementation in thermal disease diagnostics. Comput. Biol. Med. 34, 495–521 (2004)
14. Pirtini Cetingül, M., Herman, C.: A heat transfer model of skin tissue for the detection of
lesions: sensitivity analysis. Phys. Med. Biol. 55, 5933–5951 (2010)
15. Pirtini Cetingül, M., Herman, C.: Quantification of the thermal signature of a melanoma lesion.
Int. J. Therm. Sci. 50, 421–431 (2011)
16. Bhowmik, A., Repaka, R., Mishra, S.C.: Thermographic evaluation of early melanoma within
the vascularized skin using combined non-Newtonian blood flow and bioheat models. Comput.
Biol. Med. 53, 206–219 (2014)
17. Agyingi, E., Wiandt, T., Maggelakis, S.: Thermal detection of a prevascular tumor embedded
in breast tissue. Math. Biosci. Eng. 12, 907–915 (2015)
18. Maggelakis, S.A., Savakis, A.E.: Heat transfer in tissue containing a prevascular tumor. Appl.
Math. Lett. 8, 7–10 (1995)
19. Pennes, H.H.: Analysis of tissue and arterial blood temperatures in the resting forearm. J. Appl.
Physiol. 1, 93–122 (1948)
Time-Dependent Casual Encounters Games
and HIV Spread
Safia Athar and Monica Gabriela Cojocaru
Abstract In Tully et al. (Math Biosci Eng AIMS, 2015, to submitted) the authors
model and investigate casual sexual encounters between two members of a pop-
ulation with two possible HIV states: positive and negative, using a Nash game
framework in which players try to maximize their expected payoff resulting out of
a possible encounter. Each player knows their own HIV status, but do not know the
HIV status of a potential partner. They do however have a personal assessment of the
risk that the potential partner may be HIV positive. Last but not least, each player
has a ranked list of preferences of potential types of sexual outcomes: unprotected,
protected, or no sexual outcome. In Tully et al. (Math Biosci Eng AIMS, 2015, to
submitted), the game model is studied via 1- and 2-dimensional sensitivity analyses
on parameters such as the utility values of unprotected sex of an HIV negative
individual with an HIV positive, and values of personal risk (of encountering an
HIV positive partner) perception.
In this work, we introduce time as a variable which affects players’ risk
perceptions, and thus their strategies. Given that HIV transmission happens when
an HIV positive player has a non-zero probability (strategy) of having unprotected
sex with a HIV negative player, we are also able to keep track of the time evolution
of the overall fraction of HIV positive individuals in the population, as reflected
as an outcome of repeated casual encounters. We model a continuous time dynamic
game (as in Cojocaru et al. (J Optim Theory Appl 127(3):549–563, 2005)) where we
compute the stable strategies of each player based on a dynamical system defined
on a set of functions. We observe that with change in choices the HIV prevalence in
the population increases.
1 Introduction
HIV/AIDS (human immunodeficiency virus/acquired immunodeficiency syndrome)

was first observed in California, 1980 [6]. Patients, at that time, were treated by the
local physicians for intense fever, diarrhoea, weight loss and swollen lymph nodes.
S. Athar () • M.G. Cojocaru

University of Guelph, 50 Stone Rd E, Guelph, ON N1G 2W1, Canada

178 S. Athar and M.G. Cojocaru
In 1982, 600 cases were reported by CDC, out of which 75 % were identified as
homosexual or bisexual males. Therefore, the early name given to these symptoms
was ‘GRID’ (gay related immune deficiency). In late 1982, it was changed to AIDS
by CDC after determining the fact that this disease is not exclusive only to the gay
population. About 1500 lives were claimed by this disease in 1984 [5]. A poll in
1985 indicated that nearly half of Canadians were concerned about getting infected
by this disease [9].
HIV is a unique virus among others as it incorporates its own DNA into the
host’s cellular DNA. The virus also contains a protein that takes over the host cell’s
reproduction ability; is then using it as an aid for self replication. This virus impacts
the immune system resulting in life threatening infections. It can be transferred
through blood, semen,vaginal fluid or breast milk [12].
Nearly 15,000 cases of HIV were discovered in United States in 1985, whereas in
Africa the estimated number was approximately half a million. The latest research
proved that HIV first appeared in humans in central Africa. First evidence of the
evolution of HIV emerged in 1985, when scientists found a virus in Macaque
monkeys that was closely related to HIV virus [11, 12].
HIV was classified as a pandemic by the World Health Organization (WHO).
According to estimates by WHO and UNAIDS, 35 million people were living with
HIV globally at the end of 2013. In the same year, 2.1 million people became newly
infected, and 1.5 million died of AIDS-related causes. Although medical treatments
have reduced the annual rate of HIV, the drop in new HIV infections is still not
significant. The primary mode of transmission, for this disease, is sexual encounters
in many countries. For instance, 80 % of the cases in the United States are due to
unprotected sexual encounters [13].
According to [11], the spread of HIV infection is largely influenced by people’s
way of thinking about their sexual encounters. Transmission of HIV in a population
may increase if the individuals have unprotected sex. To better understand the spread
of HIV, population models have been used. These models are helpful in observing
the change in infected populations caused by different parameters. Usually, the
concept of probability is used to understand the process of decision-making in
relation to unprotected sex. Game theory is an important mathematical tool that
has been extensively used to describe the decision-making of individuals in certain
non cooperative situations. Different classes for the games are being used to model
individual’s decision depending upon for example: linear payoff with zero or non-
zero sum game as in a 2-player game; or non-linear payoff as in a multi-player
game [11].
In this paper, we are interested in studying a model of two players engaged
in finding a casual sexual partner. The game here is dynamic, i.e., we consider
time-dependency of equilibrium(Nash) strategies under time evolution of utilities
and player’s risk assessment. The main purpose is to investigate the influence
of different parameters involved in the game upon the infected population over
a certain period of time. This paper has following structure: Sect. 2 describes
briefly the one-shot 2-player game; Sect. 3 presents the dynamic game using the
frame work of evolutionary variational inequalities; Sect. 4 discusses the results
Time-Dependent Casual Encounters Games and HIV Spread 179
of the computational work along with the analysis of parameters. We close with
conclusions and a few ideas for future work.
2 2-Player Game: A Brief Introduction
Tully et al. [11] described a casual sexual encounter between two individuals as a
game. The status of the two players are known only to themselves, while players
are aged 15 years and above. Players are denoted by P1 and P2 whose HIV status is
positive and negative respectively. We denote by C the proportion of HIV positives
(HIVC ) in total population, and by the proportion of HIV negatives (HIV ) in
the total population with the condition that C C D 1.
The authors [11] used game models to find Nash equilibrium probabilities of
having unprotected sex (US) in casual encounters. The probabilities of unprotected
sex US for the two players are denoted by:
xi 2 Œ0; 12 I xi D .xi ; xiC /; i 2 f1; 2g
where xi represents the probability of Pi having US with an HIV individual and xiC
represents the probability of Pi having US with an HIVC individual. The expected
utility for Pi ; i 2 f1; 2g, when interacting with HIV and HIVC individuals, is given
by:
i
E D Œxi U.US; C; / C .1 xi /U.notUS; C; /
i
EC D ŒxiC U.US; C; C/ C .1 xiC /U.notUS; C; C/
where i 2 f1; 2g
By the term notUS means either the players have protected sex or no sex at all.
Therefore, the overall expected utility for Pi ; i 2 f1; 2g is given by:
Ei .xi ; xiC / D .1 bC /E

i i
C b C EC
where represents the activity parameter for player Pi ; i 2 f1; 2g and bC , b

represent the risk assessments of P1 and P2 respectively, regarding the individual
they are engaging in a casual encounter with. The risk parameters bC and b are
defined as follows:
b D ˇ bself
C .1 ˇ /C
self
bC D ˇC bC C .1 ˇC /C : (1)
self
The parameters bC and bself represent the personal belief of player P1 , respectively
P2 about the level of HIV infection in the population. Their values are between 0
and 1, with 1 representing the belief that everyone else in the population is infected.
The parameters ˇ and ˇC represent the weight a player places upon personal
assumptions of HIV prevalences. Their values range again between 0 and 1. With
these in mind, C is defined as:
C D C .0/ C Œx1 C .0/ C x2C .0/ ;
where C .0/ is the initial fraction of HIVC in the population. By initial we mean
here: before the encounter, where as .0/ represents the initial fraction of HIV in
the population. We let D 0:02 be the transmission rate of HIV in the population
[2].
Both players want to optimize their expected payoffs subject to constraints
defined as follows: for each i 2 f1; 2g,Pi solves the optimization problem,
(
max Ei WD Ei .x1 ; x2 /
s.t xi 2 Ki WD Œ0; 12 \ fxi C xiC D 1g
Definition 1 Assume each player is rational and wants to maximize their payoff.
Then the Nash equilibrium is a vector
x 2 K WD K1 K2 which satisfies the inequalities: For all i
fi .xi ; xi / fi .xi ; xi /, 8xi 2 Ki where xi D .x1 ; : : :; xi1 ; xiC1 ; : : :; x2 /.
The authors used variational analysis to find the Nash equilibria for the above game
as in Definition 1.
Definition 2 Given a set K Rn , closed, convex, non-empty and given F W K !
Rn is a function, the variational inequality(VI) problem is to find a vector x 2 K
such that
hF.x /; y x i 0; 8y 2 K (2)
Specifically, the Nash game is reformulated into a variational inequality problem as

follows [3, 10]:
Theorem 1 Provided the utility functions ui are of class C1 and concave (meaning
ui is convex) with respect to the variables xi , then x 2 K is a Nash equilibrium if
and only if it satisfies the VI:
hF.x /; x x i 0 8x 2 K (3)
where F.x/ D .rx1 u1 .x/; : : ::: rxn un .x// and rxi ui .x/ D . @u@x1 .x/
1
; @u@x2 .x/
2
: : ::
@un .x/
@xN
/
The variational inequality problem has at least one solution if the constraint set K
is closed, bounded and convex and F W K ! Rn is continuous. In other words, K
is a compact and convex set then solutions for variational inequality problems exist.
Further, solution may be unique if the function F is strictly monotone on K [8].
In our case we have: F.x1 ; x2 / D . 5 Ei .x1 ; x2 // D . 5x1 E1 ; 5x2 E2 /.
Here, K is closed, bounded and convex and F.x1 ; x2 /, being linear, is continuous;
therefore, the existence of solution for the problem is proved. However, F.x1 ; x2 / is
not strictly monotone, therefore the strategy applied here, by the authors, is to look
at the Nash points .x1 ; x2 / 2 K as critical points of a set of differential equations
driven by the vector field F and constraint set K:
dx
D PTK .x. // .F.x. ///I x.0/ D .x1 .0/; x1C .0/; x2 .0/; x2C .0// 2 K D K1 K2
d
(4)
2.1 One-Shot Game: Example
We develop a base case for our further investigations. The parameters of the model
are set in such a way that we can compute the Nash equilibrium vector for an ideal
situation. Ideal situation is developed to restrict the players to their own groups, i.e.,
P1 interacts with HIV and P2 interacts with HIVC player. We solve VI problem
attached to the game, using the following values in the Table 1 for the parameters
involved. The utilities for US and notUS are also given in the following Table 1.
We start with uniformly distributed initial conditions and find the unique Nash
equilibrium for the 2-player game.The value for this equilibrium is (0,1,1,0). This
shows that the two players are careful regarding the choice of their partner for casual
encounters.
The strategies of the two players are plotted to see the evolution towards the
equilibrium point when different initial values have been considered. The plot in
Fig. 1 shows the strategy of P1 and P2 , when selecting their partners for a casual sex
encounter.
Table 1 Parameters used in ˇ ˇC bself

self
bC

base case
0.5 0.5 0.6 0.3
.0/ C .0/
0.95 0.05 0.02 1
US.; / US.; C/ US.C; / US.C; C/
1 0.5 0.5 1
notUS.; / notUS.; C/ notUS.C; / notUS.C; C/
0.5 0.5 0.5 0.5
Fig. 1 Phase portrait for the strategy selections of P1 and P2 ; when P1 being HIVC interact with
HIV , whereas P2 interact with HIVC . We started with 150 uniformly distributed initial conditions
to solve (4) using Matlab
3 Time-Dependent 2-Player Encounter Game
We convert the one-shot game, as described in Sect. 2, into a time-dependent game.

The purpose behind this extension is to see the effect of time on the decision-
making process of the two players. Moreover, an additional purpose is to see
how this decision-making can affect the change in HIV prevalence (C ), if the
2-player game is played repeatedly. We use the combined theories of projected
dynamical system (PDS) and evolutionary variational inequality (EVI), in order to
investigate the dynamics of the 2-player game [1, 4]. These two theories have been
applied to a number of other problems in different disciplines, such as operational
research, economics and finance etc. [1, 4]. For this purpose, we reformulate
the variational inequality problem of our 2-player game into a time-dependent
variational inequality, (EVI). We then measure the effects of time dependency on
the HIV prevalence and on the strategies used by the players.
First, we start by defining constraint set K, as in Sect. 2, and rewrite it for the
EVI.
Our feasible vector u.t/ D .x1 .t/; x2 .t// D .x1 .t/; x1C .t/; x2 .t/; x2C .t// has to
satisfy the time-dependent constraints on xi ; i 2 f1; 2g, and it belongs to the set of
functions given by:
K D fxi .t/ 2 L2 .Œ0; T; R4 / W 0 xiC .t/; xi .t/ 1; xiC .t/ C xi .t/ D 1; i 2
f1; 2g a:e 2 Œ0; Tg
Since F.x1 ; x2 / D .rE.x1 ; x2 //, we write F.u.t// D .rE.u.t/// where F W

K ! L2 .Œ0; T; R4 /. Then EVI is defined as:
Definition 3 Find u 2 K such that
hF.u.t//; v.t/ u.t/i 0 8 v.t/ 2 K.t/ for a.a t
where
K D fxi .t/ 2 L2 .Œ0; T; R4 / W 0 xiC .t/; xi .t/ 1;

xiC .t/ C xi .t/ D 1; i 2 f1; 2g a.e in Œ0; Tg
This is called point-wise form for EVI.

Theorem 2 Let H be Hilbert space and let K H be non-empty, closed and
convex subset. Let F W K ! H be the Lipschitz continuous vector field with
Lipschitz constant b. Then the solutions of the time dependent variational inequality
Definition 3 are the same as the critical points of the projected differential equation:
du.t; /
D PTK.x/ .u.t; /; F.u.t; ///
d
That is the point x 2 K such that PTK.x/ .u.t/; F.u.t/// 0, and the converse also
holds.
If we choose H D L2 .Œ0; T; R4 / then the solution of EVI in Definition 3 and the
critical points of the equation in (2) are the same.
The important point in (2) is the difference between the two times, t; . The time
Œ0; T represents the time interval the game is considered to be played over. As t
varies over the interval [0,T], we can obtain one or more curves representing how
Nash strategies of both players may change. The time is the simulated time of
evolving from an initial point of the differential equation (2) towards one of the Nash
equilibria on this curve(s). For the computational purpose, a sequence of partition
ftn0 ; tn1 ; : : ::; tnN g 2 Œ0; T is defined such that
0 D tn0 < tn1 <; : : :: < tnN D T:
Then for each tnj , we solve the time dependent variational inequality:
hF.u.tnj //; v u.tnj /i 0 for all v 2 K.tnj /;
where
K.tnj / D fxi .tnj / 2 L2 .Œ0; T; R4 / W 0 xiC .tnj /; xi .tnj / 1; xiC .tnj / C xi .tnj / D
1; i 2 f1; 2g a.e in Œ0; Tg.
We compute all solutions of this finite dimensional variational inequality, by

finding the critical point of the projected dynamical system [4]:
PTK.x/ .u.tnj ; /; F.u.tnj ; /// D 0
4 Computation of Time-Dependent Nash Strategies and

Their Effects on the HIV Prevalence
4.1 Base Case
We use the same values for the parameters in the base case as in the one-shot game.
The only difference in this base case is that both the strategies and C are time
dependent. The parameters used for the base case are described in Table 1. We run
this case for T D 10 periods, and reach the same equilibrium point as in the one-shot
game. No change in the value of C .t/ is observed (Fig. 2).
Fig. 2 Increase in HIV prevalence over a period of 10 periods

4.1.1 Influence of US.; C/ and US.C; / on HIV Prevalence
We vary US.; C/ & US.C; / as a function of time, whereas the other parameters
are kept constant. The change in choices about having US over the given period
of time can be considered the result of emotional decision-making and/or the
level of information (Fig. 3). Gutnik et al. in [7] discussed the role of emotions
in decision-making. The role was considered as negative hindrance in the rational
decision-making. According to new research, people use analytic and experimental
systems to understand and assess the factors of risk. However, the emotions are
the most common factor that works in the experimental system. These emotions
depend on past experiences or the perceived risks [7]. The players can change their
preferences over the given period under the influence of emotion. These decisions
under emotional circumstances perhaps occur due to lack of adequate knowledge or
due to the “heat of moment” (Table 2).
Fig. 3 3-D plots shows the change in HIV prevalence, when the US.; C/ and US.C; / are
made time dependent. Upper left shows the change in US.; C/ and US.C; / are linear function
of time; upper right shows US.; C/ and US.C; / are quadratic function of time. Lower left
shows US.; C/ and US.C; / are quadratic function as well but in different formation and lower
right shows US.; C/ and US.C; / are cubic function of time. The increase is much higher in
linear case, US.; C/ and US.C; / are linear function of time, as compared to the case when
these choices are quadratic or cubic function of time
Table 2 Effect of US.; C/ and US.C; / on HIV prevalence C .t/

Cases US.; C/ US.C; / .t/
Case 1 0:5 C 0:5 .t=10/ 0:5 C 0:5 .t=10/ 5 % to 32 %
Case 2 0:5 C 0:5 .t=10/2 0:5.t/ C 0:5 .t=10/2 5 % to 24 %
Case 3 0:5 C 0:5 .t=10/3 0:5 C 0:5 .t=10/3 5 % to 21 %
The time is taken as 10 periods and Matlab is used for simulations. We observe an
increase in HIV prevalence from 5 % to 32 % when US.C; / is linear function of
time; whereas the increase is 24 % when US.C; / in quadratic in time. The impact
of US.C; / on increase in HIV prevalence is perhaps due to change in preferences
of player 1(HIVC) over the time.
5 Conclusion and Future Work
In this work we converted a discrete time 2-player game into a time-dependent

game. We investigated the effects of changes in utilities for having unprotected sex
between HIVC and HIV players on the HIV prevalence rate over a given period of
time. Also, we examined the influence of individual’s personal risk perception about
their partners in casual encounters.
In the near future we are interested in formulating a multi-player time-dependent
game, when two different types of HIV viruses, i.e., HIV 1 and HIV 2 are
involved. The players with these viruses interact with HIV players from same
age groups or different age groups. Further, we want to investigate the impact of
prophylactic vaccination against HIV in younger population groups in multi-players
time-dependent games.
References
1. Barbagallo, A., Cojocaru, M.-G.: Dynamic vaccination games and variational inequalities on
time-dependent sets. J. Biol. Dyn. 4(6), 539–558 (2010)
2. Basar, T., Olsder, G.J.: Dynamic Noncooperative Game Theory, vol. 200. SIAM, Philadelphia,
(1995)
3. Cojocaru, M.-G., Greenhalgh, S.: Dynamic games and hybrid dynamical systems. Optim. Eng.
Appl. Var. Inequal. Issue 13(3), 505–517 (2011)
4. Cojocaru, M., Daniele, P., Nagurney, A.: Projected dynamical systems, evolutionary variational
inequalities, applications, and a computational procedure. In: Pareto Optimality, Game Theory
and Equilibria, pp. 387–406. Springer, New York (2008)
5. Engel, J.: The Epidemic: A History of AIDS. HarperCollins, New York (2009)
6. Grmek, M.D.: History of AIDS: Emergence and Origin of a Modern Pandemic. Princeton
University Press, Princeton, NJ (1993)
7. Gutnik, L.A., et al.: The role of emotion in decision-making: a cognitive neuroeconomic

approach towards understanding sexual risk behavior. J. Biomed. Inform. 39(6), 720–736
(2006)
8. Kinderlehrer, D., Stampacchia, G.: An Introduction to Variational Inequalities and Their
Applications. Academic, New York (1980)
9. Levon, B.: Facing a fatal disease. Maclean’s 1(6), 48 (1986)
10. Moulin, H.: On the Uniqueness and Stability of Nash Equilibrium in Non-cooperative Games,
vol. 130. North-Holland Publishing Company, Amsterdam (1980)
11. Tully, S., Cojocaru, M.-G., Bauch, C.: Multiplayer games and HIV transmission via casual
encounters. Math. Biosci. Eng. AIMS (2015, to submitted)
12. World Health Organization: Health topics: HIV/AIDS. http://www.who.int/hiv/pub/guidelines/
en. Accessed 6 Mar 2013
13. World Health Organization: Health topics: HIV/AIDS. http://www.who.int/features/qa/71/en/
http://www.who.int/bulletin/volumes/85/11/06-033779/en/index.html. Accessed 6 Mar 2013
Modelling an Aquaponic Ecosystem Using
Ordinary Differential Equations
C. Bobak and H. Kunze
Abstract Aquaponic agriculture is a sustainable system which uses interdependent

processes and has been growing in popularity. However, relatively little mathemat-
ical and other academic research has been conducted in the practice. In this paper,
we develop a system of ordinary differential equations to model the population
and concentration dynamics of the environment. Our model has an asymptotically
stable non-trivial equilibrium, representing the inherent symbiotic relationship of
the variables. Values of the nine parameters in the system are estimated from the
research literature. We provide simulated results simulated results to illustrate the
nature of solutions to the system, and we present and discuss a sensitivity analysis.
1 Introduction
Aquaponics is a closed-loop agricultural system which uses a symbiotic relationship

between aquatic organisms and aquatic macrophytes. The system recirculates water
through an aquaculture environment (fish in a designated body of water) and
a hydroponic structure (aquatic plants in soilless water) to create a sustainable
environment which fully conserves water and nutrients. The key motivation behind
aquaponics is using waste produced by fish in the system as a nutrient source
for the plants. This process not only allows the fish waste to act as a natural
fertilizer for the plants but also inherently cleans the water to be returned to the
fish [6]. Thus the environment is in a natural state of stability, which we seek to
model using a system of differential equations. This paper discusses background
theory in aquaponics, develops a compatible aquaponic model in the form of a
system of differential equations, discusses equilibria of the model and their stability,
presents specific solution simulations for parameter values derived from the research
literature, performs a sensitivity analysis, and suggests future research direction.
C. Bobak () • H. Kunze


190 C. Bobak and H. Kunze
2 Background: Research of Aquaponic Environments
Aquaponic agriculture systems have been growing in popularity due to their robust
economic and ecological benefits. Both hydroponics and aquaculture have proven
to have a detrimental effect on the environment. Some studies have suggested that
aquaponics has a 75 % smaller carbon footprint when compared to traditional farm-
ing methods [7] and is able to satisfy human demand long-term [3]. Importantly,
aquaponics achieves all of these benefits while maintaining economic viability for
farmers. The lower cost of resources and lower spatial requirements has led to
suggestions that aquaponic agriculture may be a viable solution in densely populated
urban regions such as Pakistan and India [3].
The basic aquaponic system consists of two main components: one for fish
and one for plants. The system relies heavily on food input as many studies have
demonstrated that factors like protein content of food and feeding frequency have
the largest effect on the efficiency of the system. These factors highly contribute
to fish growth and are also directly related to the amount of fish waste in the
environment [6].
Organic nitrogen in fish waste naturally converts to ammonia through biological
degradation [4]. Ammonia is highly toxic to fish, and is an inefficient nutrient source
for plants [6]. In order for ammonia to be used as fertilizer for the plants, it must
go through a natural microbial process called the Nitrogen Cycle, which causes it to
convert into nitrate. Nitrate is a nutrient rich food source for plants [4] and studies
have shown that plants’ uptake efficiency of nitrates ranges from 86 % to 98 %.
Whatever concentration of nitrate is left in the water is not harmful to the fish, so
the water can be recirculated for fish use [6].
3 Development of the Model
The overarching goal of the model is to capture the symbiotic relationship between
the fish and the macrophytes. However, as evidenced by background research on
aquaponic agriculture, the aerobic microbial process which converts ammonia to
nitrate needs to be considered. The assumpions made are as follows:
i. The aquaponic ecosystem is a closed environment.
ii. The fish population increases at some natural survival rate . Deaths
Births
/, hindered
by a carrying capacity due to the limited tank space.
iii. There is additional fish decay due to increased ammonia presence in the water
until it reaches a critical ammonia level where no fish survive. This can be
reasonably modelled using a linear constant of . Toxic Ammonia
Ammonia Level /.
iv. Ammonia is present in the system exclusively due to fish waste and hence
grows at a rate proportional to the fish population. It decays due to its
conversion to nitrate.
Aquaponic Ecosystems Using ODEs 191
v. Nitrate grows at a rate proporational to the level of ammonia, and decays due
to plant uptake.
vi. Plants grow at a constant rate hindered by a carrying capacity indicative of the
limited surface area of the system.
vii. Modelling the concentrations of ammonia and nitrate in the system will capture
any other relationships between other variables in the nitrogen cycle.
viii. The system is well mixed so the nitrogen cycle occurs naturally and plants
have even access to nitrate.
The proposed model is below, with variables F, A, N, and P representing the
population of fish, ammonia (in mg), nitrate (in mg) and population of plants
respectively:

PF D a1 1 F FP A F (1)
KF KA
AP D a2 F a3 A (2)
NP D a4 A a5 NP (3)

P
PP D a6 1 PN (4)
KP
where ai 0 8i are growth and decay rates, and KF ; KA ; KP > 0 are the carrying
capacities of fish, ammonia, and plants respectively.
Equation (1) models the evolution of the fish population, using assumptions ii and
iii. Equation (2) models the evolution of the ammonia concentration in the system
using assumption iv. Equation (3) captures the growth rate of nitrate concentration
as it relates to the conversion from ammonia and the decay rate due to plant uptake
using assumption v. The final equation (4) captures the growth rate of the plants
using assumption vi.
4 Analyzing the Equilibria
Our aquaponic environment model (1), (2), (3) and (4) consists of four equations
with nine unknown parameters. The Jacobian matrix of this model is as follows:
2 a1 FP

3
KF Ca1 1 KF P KA KF 0 a1 1 KF F
6 7
F A A F
a3 0 0
Df .F; A; N; P/ D 4 a2
0 a 5 (5)
a4
5P
a5 N
a6 NP
0 0 a6 P 1 KP a6 N 1 KP K
P P P
The system (1) (2), (3) and (4) has three equilibria:
fF D 0; A D 0; N D 0; P D Pg (6)
fF D 0; A D 0; N D N; P D 0g (7)

a1 a3 KF KA KP a1 a2 KF KA KP
FD ;AD ;
a1 a3 KP KA C a2 KF a1 a3 KA KP C a2 KF

a1 a2 a4 KF KA
ND ; P D KP (8)
a5 .a1 a3 KA KP C a2 KF /
of which the third, in equation (8), since it represents all variables surviving in the
environment. We focus on it in the next section.
4.1 Equilibrium (8): Coexistence
We analyzed the stability of the nontrivial equilibrium in equation (8) using the
research literature to provide estimates for the nine parameters: a1 D 0:0124, a2 D
0:1, a3 D 0:94, a4 D 3:6, a5 D 0:92, a6 D 0:056, KF D 250, KP D 300 and
KA D 20 [1–4, 6]. The values selected for our carrying capacities were based on an
arbitrary initial tank size of 10L, however, this can easily be scaled up or down to
accomodate systems of various sizes. Other initial values were fF.0/ D 10; A.0/ D
0; N.0/ D 0; P.0/ D 0:5g, where P.0/ D 0:5 was used to represent plants which
had not yet reached maturity in the system. Substituting the estimated parameter
values in the Jacobian gives:
2 3
2:77 9:31 0 0:59
6 0:1 0:98 0 0 7
A D 6
4 0
7 (9)
3:6 276 0:2285
0 0 0 0:014
The eigenvalues of A in this particular simulation are:
31 D 1:84 C 0:332i (10)

32 D 0:014 (11)
33 D 276 (12)
34 D 1:84 0:332i (13)
Notably, the real parts of (10) and (13) are negative, and values (11) and (12)
are negative, thus an asymptotically stable equilibrium is achieved in this case.
Because (10) and (13) are complex, some spiralling behaviour is present in the
system.
Note that 32 is very close to 0, so the stability status of this equilibrium
point may be sensitive to the parameter values chosen. Some experimentation was
done within ranges of realistic values for the estimated parameters. In every case,
all eigenvalues have negative real parts suggesting with some generality that the
equilibrium case where all four variables are present in the system is asymptotically
stable. We discuss sensitivity in general in the next section.
Substituting these values into the non-trivial equilibrium and subsequent solution
given in the previous section provides the following equilibrium points:
fF D 186:18; A D 19; N D 0:25; P D 300g (14)
The real-world context of these equilibrium points is exciting; it suggests that

fish grow to a point well bounded by their carrying capacity, (KF D 250). Ammonia
tends towards an amount just below the threshold to which all the fish would die
(KA D 20). This suggests that the system is stabilizing in a way which approaches
the maximum level of food production for the plants. Nitrate levels appear to be
quite low, which is likely due to the superior nutritional uptake of the plants which
has been noted in the literature. The plants themselves reach equilibrium at their
carrying capacity (KP D 300). This is exactly as anticipated based on the model
structure.
Figure 1 shows two plots of the modelled variables over time given the parameter
assumptions. As evidenced here, all the variables are approaching their above
equilibrium values.The fish and plant follow a logistic trend tending to their carrying
capacities. Ammonia also seems to be following this trend, which is due to its
Fig. 1 The population and concentration dynamics of the model over time (a) The solid line
represents the fish population and the dotted line represents the plant population (b) The solid
line represents the nitrate population and the dotted line represents the ammonia population
growth dependence on fish. Nitrate also shows an interesting trend; as the system
is being established, nitrate experiences a brief spike. However, once the plant
growth reaches a level which requires significant nitrate sustenance, this spike
reverses and the nitrate concentration levels off. These curves appear to loosely
represent expected trends based on the aquaponics literature, suggesting the model
is reasonably capturing aquaponic behaviour.
5 Sensitivity Analysis
A sensitivity analysis for the parameters of this model was performed with the goals
of identifying any unexpected behaviour, guiding any data collection efforts, and
most importantly, to give an indication of the importance of accurately estimating
the parameter values.
A software toolbox in Matlab was used to perform sensitivity analysis of
biological models. SensSB is freely available for academic purposes, and combines
a variety of local and global sensitivity methods, both using relative and absolute
measures to achieve many of these goals [8, 9].
While locally analysing the sensitivity of parameters is a useful exercise, it has
an inherent reliance on the initial numerical estimation of the parameter. Global
methods, which test the effect of a parameter while other parameters are varied
simultaneously, help avoid this stipulation [9]. Since the initial estimation of the
variables from the literature are considered weak points of the model, global
sensitivity measures were analysed.
SensSB analyses global sensitivity through three main methods, all of which
are discussed in the software documentation (see [9]). In this study, Derivative
Based Global Sensitivity Measures (DBGSM), which were introduced in 2009 [5],
were selected to optimize accuracy with minimal loss of computational efficiency
[9]. DBGSM uses Monte Carlo sampling methods to average local derivatives to a
measure M N ij which averages sensitivity measures Sij over the parameter space.
% change in Parameters
Sij D (15)
% change in Variables
Z
N ij D
M Sij dp (16)
Hnp
As seen in Fig. 2, parameter 3 .a3 /, or the conversion rate of ammonia to nitrate,

has high global absolute sensitivity. In fact, according to the output from SensSB,
it accounts for 73.26 % of the total sensitivity in the model. Notably, the relative
sensitivity by variable, which shows how sensitive each variable is to the parameters
in the model, shows high relative sensitivity for parameters a2 , a3 , a6 , and KF . In
the model, these parameters represent the growth rate of ammonia from fish waste,
the conversion rate of ammonia to nitrate, the growth rate of plants and the carrying
Fig. 2 (a) shows the global absolute sensitivity of each of the estimated parameters while (b)
shows the global relative sensitivity by variable. Note parameters 1–6 are a1 through a6 , parameter
7 is the carrying capacity KF , parameter 8 is the carrying capacity KA , and parameter 9 is the
carrying capacity KP
capacity of fish. This suggests that some more care should be taken in estimating
values for these parameters, particularly a2 and a3 , which do not have reliable values
in the literature.
Using values from the previous simulation, both a2 and a3 were varied (individ-
ually and consecutively) between values of 0 and 100 without loss of stability in the
coexistence equilibrium. However, exploring data from an established environment
may help clear potential biases associated with this parameter estimation.
6 Future Research Direction
Future models of aquaponic systems should incorporate additional complexity.

In particular, removing the assumption that the environment is closed to allow
additional harvesting cycle terms for fish and plants would add realism and require
additional analysis. Here, the term harvesting refers to the biological process of
removing some fish and/or plants, creating a discontinuity in the solution.
As well, there is some indication in the literature that the assumption that the
fish population decays linearly with increases in ammonia concentraion is invalid.
Instead, it is expected that the fish population would not be affected by the presence
of ammonia in the system until the ammonia begins to approach a critical level [1].
Instead, a non-linear polynomial relationship could be used, and may be a better
approximation of the real world relationship.
Building from the sensitivity analysis, real data sets could be used to solve the
parameter estimation inverse problem to see how well this model, and more complex
variations, can accommodate real-world data as an approximation solution.
7 Conclusion
The developed four equation model appears to reasonably simulate biological

behaviour. Simulated solutions lead to a few conclusions which may or may not
hold in the general case. The most important of these is that the model admits
an asymptotically stable equilibrium where fish, ammonia, nitrate, and plants
interdependently coexist in the environment. An exploration of parameter space
suggests that the stability of such an equilibrium point is robust to changes in the
parameters.
Aquaponics shows considerable promise as an ecological agriculture solution.
It is the hope that this research can be used to further test the ability of aquaponic
solutions as viable practices in both developed and undeveloped countries to aid in
concerns over the sustainability of current food practices.
References
1. Wheaton, F.: Recirculating aquaculture systems: an overview of waste management. In:

Proceedings of the 4th international conference on recirculating aquaculture, Roanoke (2002)
2. Bhujel, R.C., Little, D.C., Hossain, M.A.: Reproductive performance and growth of stunted and
normal Nile Tilapia (Oreochromis niloticus) broodfish at varying feeding rates. IEEE Aquac.
273, 71–79 (2007)
3. Blidariu, F., Grozea, A.: Increasing the economical efficiency and sustainability of indoor fish
farming by means of aquaponics – review. Anim. Sci. Biotechnol. 44(2), 1–8 (2011)
4. Endut, A., Jusoh, A., Ali, N.: Nitrogen budget and effluent nitrogen components in aquaponics
circulation systems. Desalinations Water Treat. 52, 744–752 (2014)
5. Kucherenko, S., Rodriguez-Fernadez, M., Pantelides, C., Shah, N.: Monte carlo evaluation of
derivative based global sensitivity measures. Reliab. Eng. Syst. Saf. 94, 1135–1148 (2009)
6. Liang, J.L., Chien, Y.H.: Effects of feeding frequency and photoperiod on water quality and crop
production in a tilapia-water spinache raft aquaponics system. Int. Biodeterior. Biodegrad. 83,
693–700 (2013)
7. Pattillo, A.: Aquaponic system design and management. Aquaculture Extension, Iowa State
University (2014)
8. Rodriguez-Fernandez, M., Banga, J.R.: SensSB: a software toolbox for the development and
sensitivity of systems biology models. Bioinformatics 26(13), 1, 2 (2009)
9. Rodriguez-Fernandez, M., Banga, J.R.: SensSB – a software toolbox for sensitivity analysis in
systems biology models. In: International Conference on Systems Biology (ICSB 2009) Aug
30–Sept 4, Stanford University, Palo Alto (2010)
A New Measure of Robust Stablity for Linear
Ordinary Impulsive Differential Equations
Kevin E.M. Church
Abstract A new measure of robust stability for linear ordinary impulsive dif-
ferential equations with periodic structure is introduced, based on the impulse
extension concept. This new stability measure reflects the sensitivity of the model to
uncertainty in what we see as the fundamental hypothesis of impulsive models: that
the impulse effect occurs quicky enough that its duration can be entirely neglected.
The measure, that we call the time-scale tolerance, E t , has the property that, if the
vector of durations of impulse effect, a, satisfies jjajj < E t , then both the impulsive
model and a family of continuous impulse extension equations (a specific functional
differential equation) to which it is related, will all be asymptotically stable.
We review linear impulse extension equations, state theorems that describe the
convergence of their solutions to the associated impulsive solutions, and introduce
all the machinery necessary in the development of the time-scale tolerance, stating
theoretical results on its existence and how it can be computed in practice. We
conclude with two illustrative examples and a discussion of the limitations of the
techniques presented, as well as elaborate on the ways they can be improved.
1 Introuction
Impulsive differential equations are frequently employed as models of biological,

chemical, and physical systems, among others. The property of these equations
that makes them attractive in applications is the impulse effect, which allows for
the inclusion of fast dynamics that might otherwise complicate the analysis of the
model, were they to be included in a continuous, as opposed to discrete, manner.
Most monographs on impulsive differential equations—for example, [1, 5, 6]—
state that these equations are reasonable approximations of continuous models with
perturbations, if the perturbations themselves occur quickly, relative to the overall
dynamics. In practice, this is often taken as an assumption of the model in question.
Methods to determine how quickly these perturbations must occur for the model
K.E.M. Church ()

Department of Mathematics and Statistics, University of Ottawa, 585 King Edward, Ottawa,
ON K1N6N5, Canada

198 K.E.M. Church
to be a good “fit” to the associated continuous models have yet to be seen in the
literature.
In these proceedings, we introduce a quantity that we call the time-scale
tolerance, denote E t , for a linear, periodic impulsive differential equation that is
asymptotically stable. This quantity has the property, that, if the vector of durations
of impulse effect, a, satisfies jjajj < E t , then both the impulsive model and a
family of continuous impulse extension equations (a specific functional differential
equation) to which it is related, will all be asymptotically stable.
We review linear impulse extension equations, which were first introduced in [2–
4], state theorems that describe the convergence of their solutions to the associated
impulsive solutions, and introduce all the machinery necessary in the development
of the time-scale tolerance, stating theoretical results on its existence and how it
can be computed in practice. We conclude with two illustrative examples and a
discussion of the limitations of the present techniques, and how they can be extended
to accomodate a larger class of problems.
2 Linear Impulse Extension Equations
To begin, we introduce some notation that will be present throughout this chapter.
If x D fxk W k 2 Zg is a real-valued sequence, we denote xk D xkC1 xk . The
kth element of a real-valued sequence x will always be denoted xk , and we may
abuse notation and identify the sequence x with the symbol xk . Indexed families of
sequences, such as, fx j W j 2 Ug, will always have their index appear in the exponent.
j
In this context, xk denotes the kth element of sequence j from the family U. Finally,
our sequences will usually be bi-infinite; that is, indexed by the integers. The symbol
jj jj will denote a (fixed) Euclidean norm, whenever there is no ambiguity, and if A
is a set, its closure will be denoted A.
Consider a linear, impulsive differential equation with impulses at fixed times
dx
D A.t/x C g.t/; t ¤ k
dt (1)
x D Bk x C hk ; t D k :
with t 2 R, phase space ˝ Rn , A W R ! Rnn , Bk 2 Rnn , hk 2 Rn , and sequence

of impulses k for k 2 Z. We assume that A and g are locally integrable, and that the
sequence of impulse times, k , is monotone increasing and unbounded.
An impulse extension equation for (1), as described in Church and Smith? [3, 4]
and Church [2], is defined as follows.
Definition 1 Consider a linear impulsive differential equation (1).
• A step sequence over k is sequence of positive real numbers a D fak W k 2 Zg
such that ak 2 .0; k / for all k 2 Z. We denote Sj D Sj .a/ Œ j ; j C aj / and
A New Measure of Robust Stablity for Linear Ordinary Impulsive Differential. . . 199
S
S D S.a/ j2Z Sj . The set of all step sequences will be denoted S , and is
defined by
S WD fa W Z ! R ; ak 2 .0; k /g:
• The pair .'kB ; 'kh /, with sequences of functions 'kB W R RC ! Rnn and 'kh W
R RC ! Rn , is a family of impulse extension for (1) if for all a 2 S and all
k 2 Z, the functions 'kB .; ak / and 'kh .; ak / are integrable on Sk .a/ and satisfy the
equalities
Z Z
'kB .t; ak /dt D Bk ; 'kh .t; ak /dt D hk : (2)
Sk .a/ Sk .a/
• Given a step sequence a 2 S and a family of impulse extensions ' D .'kB ; 'kh /
for (1), the impulse extension equation associated to (1) and induced by .'; a/ is
the (functional) differential equation

dx A.t/x C g.t/; t … S.a/;
D (3)
dt A.t/x C g.t/ C 'kB .t; ak /x. k / C 'kh .t; ak /; t 2 Sk .a/:
Definition 2 Let a family of impulse extensions, ' D .'kB ; 'kh /, and a step sequence
a 2 S be given. A function y W I ! Rn defined on an interval I R is a classical
solution of the impulse extension equation (3) induced by .'; a/, if y is continuous,
the sets I \ Sk .a/ are either empty or contain k , and y satisfies the differential
equation (3) almost everywhere on I. Given an initial condition
x.t0 / D x0 ; (4)
with .t0 ; x0 / 2 RRn , the function y.t/ is a solution of the initial-value problem (3)–
(4) if, in addition, y.t0 / D x0 .
Remark 1 Definitions 1 and 2 can be readily modified to accomodate nonlinear
ordinary impulsive differential equations; see Church and Smith? [4] and Church
[2].
Definition 3 The predictable set of the impulse extension equation (3) induced by
.'; a/ is
( Z ! )
t
P D R n t 2 S.a/ W det I C X 1 .s; k /'kB .s; ak /ds D 0 ;
max k f k tg
where X.t; s/ is the Cauchy matrix of the homogeneous system z0 D A.t/z.

The following theorem states the mode in which solutions of the impulse
extension equation, (3), converge to those of the impulsive differential equation, (1).
200 K.E.M. Church
Theorem 1 Suppose det.I C Bk / ¤ 0 for all k 2 Z, and let ' D .'kB ; 'kh / be
a given family of impulse extensions for (1). There exists a positive sequence of
real numbers k , depending only on A.t/ and sequence of impulse times k , with

the following property. Suppose, for 2 fB; hg and each k 2 Z, there exists wk W
Œ k ; kC1 Œ0; k / ! R that is continuous and vanishing at . k ; 0/, for which

1 1
'k .t; s/ k D O wk .t; s/ s (5)
s e k 1
for t 2 Œ k ; k C s/ as s ! 0. The following are true:

• For all t0 2 R, there exists ı > 0, such that, for a 2 S with jjajj1 < ı and all
x0 2 Rn , the impulse extension equation (3) induced by .'; a/ possesses a unique
classical solution, x.tI a/, satisfying the initial condition x.t0 / D x0 .
• The function x.tI a/ converges pointwise to x.tI 0/, the solution the initial value
problem x.t0 / D x0 for the impulsive differential equation, (1), as a ! 0.
• If N R is bounded and no strictly decreasing sequence in N has an impulse
time k as its limit, the above convergence is uniform for t 2 N.
Proof (Outline) The existence of the sequence k follows from the Generalized
k .t k /
Gronwall’sR kC1inequality: we have jjX.tI k /jj e for t 2 Œ k ; kC1 , where
k D k jjA.s/jjds, and X.tI s/ is the Cauchy matrix of x0 D A.t/x. When
jjajj1 < t0 maxf k W k < t0 g, the solution of (3)–(4) induced by .'; a/ satisfying
x.t0 I a/ D x0 , can be written as
x.tI a/ D U.tI a/x0 C xp .tI a/;
for a matrix function t 7! U.tI a/ satisfying U.t0 I a/ D I, and xp .t0 I a/ D 0 (this

follows by Proposition 4.2 of [4]). If U.tI 0/ denotes the fundamental matrix solution
of the homogeneous equation associated to (1), one can show that the inequality
ˇˇ !
ˇˇ Y0
ˇˇ
jjU.t; a/ U.t; 0/jj jjX.tI k /jj ˇˇ La .tI k / X. rC1 I r /La . r C ar I r /
ˇˇ
rDk1
!ˇˇ
Y0 ˇˇ
ˇˇ
.I C Bk / X. rC1 I r /.I C Br / ˇˇ
ˇˇ
rDk1
holds, where we have assumed t0 D 0 for ease of presentation (other cases follow
by similar reasoning, by results from [4]), and
Z minft; k Cak g
La .tI k / D I C X 1 .sI k /'kB .s; ak /ds:
k
It can be shown that the right-hand side of the upper bound converges to zero
as a ! 0 pointwise, as jjajj1 ! 0 (see the proof of Theorem 3.5.5. from
[2] for the main idea; condition (5) is needed). A similar inequality holds for
jjxp .tI a/ xp .tI 0/jj, where xp .tI 0/ is the solution of (1) satisfying xp .t0 I 0/ D 0,
and the convergence result holds for that piece of the solution as well. For uniform
convergence, it suffices to consider N to be a finite union of closed intervals with
xn # x 2 N ) x … f k g.
The hypotheses of Theorem 1 are simplified if the equations (1) and (3) are
periodic.
Definition 4 The linear impulsive differential equation (1) is T-periodic with c
impulses per period if A.t C T/ D A.t/ and g.t C T/ D g.t/ for all t 2 R, and
kCc D k C T, BkCc D Bk and hkCc D hk for all k 2 Z. The step sequence a 2 S is
c-periodic, and we write a 2 Sc , if akCc D ak for all k 2 Z. The family of impulse

extensions ' D .'kB ; 'kh / is .T; c/-periodic if 'kCc .t C T; s/ D 'k .t; s/ for all t 2 R,
all k 2 Z, all s 2 .0; k / and 2 fB; hg.
Corollary 1 Suppose the impulsive differential equation (1) is T-periodic with c
impulses per period. Let ' D .'kB ; 'kh / be a .T; c/-periodic family of impulse
extensions for (1). Suppose det.I C Bk / ¤ 0 for k D 0; : : : ; c 1: Let the
impulsive differential equation (1) have a fundamental matrix X.t/ with Floquet
decomposition X.t/ D P.t/et satisfying X. 0 / D I. The conclusions of Theorem 1
hold for step sequences a 2 Sc , with k jjjj.
The proof of the above corollary is omitted, since it is simple to prove using
Theorem 1. For periodic impulse extension equations, we have an asymptotic
Floquet theorem. A proof is available in [2], where it is listed as Theorem 3.6.16.
Theorem 2 Suppose the impulsive differential equation (1) is T-periodic with c
impulses per period. Let ' D .'kB ; 'kh / be a .T; c/-periodic family of impulse
extensions for (1). Let det.I C Bk / ¤ 0 for k D 0; : : : ; c 1. Then, under the
hypotheses of Corollary 1 on the asymptotic criterion (5), there exists ı > 0 such
that, if a 2 Sc satisfies jjajj < ı, any solution x.t/ of the homogeneous impulse
extension equation induced by .'; a/,

dx A.t/x; t … S.a/
D (6)
dt A.t/x C 'kB .t; ak /x. k /; t 2 Sk .a/:
can be written as a product,
x.t/ D Ua .t/x0 D Pa .t/ea t x0 ; (7)
for some x0 2 Rn , T-periodic matrix Pa .t/, and nonsingular a . Ua can be

normalized so that Ua . 0 / D I, and in this case, we have a ! 0 as a ! 0, where
0 is the matrix appearing in the Floquet decomposition, X.t/ D U0 .t/ D P0 .t/e0 t ,
of the homogeneous equation associated to the periodic impulsive differential
equation, (1), with U0 . 0 / D I.
202 K.E.M. Church
The stability of the periodic linear impulse extension equation induced by some
.'; a/ is determined by the spectrum of a , just as with ordinary and impulsive
differential equations. The main difference is that stability (and uniform stability)
only holds for initial conditions in particular subsets of the predictable set, P, and
such restrictions are in fact, optimal. For details, see [4].
3 The Time-Scale Tolerance for Linear, Periodic Impulsive

Differential Equations
Stability of (3) is completely determined by the associated homogeneous equa-

tion (6); see [4]. As such, our investigation will now shift to homogeneous,
.T; c/-periodic impulsive differential equations,
dx
D A.t/x; t ¤ k
dt (8)
x D Bk x; t D k :
and impulse extension equations for (8), induced by .'; a/ D .' B ; a/, with a 2 Sc ,

dx A.t/x; t … S.a/
D (9)
dt A.t/x C 'kB .t; ak /x. k /; t 2 Sk .a/:
From here onward, M0 will denote the monodromy matrix for (8) satisfying
M0 D X. 0 CT; 0 /, where X.t; s/ is the Cauchy matrix for (8). We assume M0 < 1
from here onward. We will comment in Sect. 3.3 on what can be done if M0 1.
Definition 5 Consider a periodic homogeneous impulsive differential equation, (8).
Let D fk g be a c-element sequence of positive real numbers and w D fwk g be
a c-element sequence of functions wk W Œ k ; kC1 Sc ! RC that are continuous
and vanishing at . k ; 0/ and such that wk .; a/ is integrable on Sk .a/. A family of
periodic impulse extensions, ' D f'k g, is uniformly exponentially .; w/-regulated
in the mean or simply .; w/-regulated if the inequality
ˇˇ ˇˇ
ˇˇ ˇˇ
ˇˇ'k .s; a/ 1 Bk ˇˇ wk .s; a/ (10)
ˇˇ ak ˇˇ ek ak 1
is satisfied, for all s 2 Sk .a/ and k D 0; : : : ; c 1. A pair .; w/ that satisfies

the above criteria will be referred to as a uniform exponential regulator, or simply
exponential regulator. If ' is uniformly .; w/-regulated, we will write ' 2 .; w/.
Definition 6 If R D .; w/ is an exponential regulator, the .R; a/-pseudospectral

radius of (8), denoted .R; a/, is defined by
.R; a/ D sup M.'; a/; (11)

'2R
and M.'; a/ denotes the monodromy matrix of the impulse extension equation
for (8) induced by .'; a/.
Definition 7 Suppose (8) is asymptotically stable. Let R be an exponential regula-
tor. The R-stable set, denoted E s .R/, is defined as follows.
˚
E s .R/ D a 2 Sc W 8' 2 .; w/, M.'; a/ < 1 (12)
The R-time-scale tolerance is the number
E t .R/ D supf W 9a 2 E s .R/; jjajj D ; B .0/ \ Sc E s .R/g: (13)
The time-scale tolerance is defined precisely so that we have the following

elementary property, whose proof we omit.
Proposition 1 Given an exponential regulator R D .; w/, the time-scale tolerance
behaves as a robust stability threshold for the impulsive system (1); if jjajj < E t .R/,
then .R; a/ < 1. In other words, systems (8) and the impulse extension equation (6)
induced by .'; a/ are both stable, for all ' 2 R.
Theorem 3 Suppose D jjjj, as in Corollary 1. If R D .; w/ is an exponential
regulator and (8) is asymptotically stable, then E t .R / is nonzero and the map a 7!
.R ; a/ satisfies
lim .R ; a/ D M0 ;

a!0
where the limit is for a 2 Sc .

Proof (Outline) The monodromy matrix for (6) induced by .'; a/ for ' 2 R can
be written as
0
Y Z
M.'; a/ D X. kC1 I k / X 1 .sI k /k .s; a/ds
kDc1 Sk .a/
Z
1
C I C X 1 .sI k /Bk ds ;
ak Sk .a/
with k .s; a/ D 'k .s; a/ a1k Bk . Taking norms, each of the k terms can be bounded
by inequality (10), and the upper bound is independent on the explicit choice of ',
depending only on the regulator R . With the choice of given in the theorem, each
204 K.E.M. Church
intergral involving K converges to zero, while the other clearly converges to I C Bk .

Therefore, M.'; a/ ! M0 uniformly for ' 2 R . The result follows.
In practice, computing the time-scale tolerance is difficult. We can, thankfully,
resort to conservative estimates.
Theorem 4 Let R be an exponential regulator for (8). Suppose there exists a
continuous function n W Sc ! RC such that n.0/ D 0 and
jjM.'; a/ M0 jj n.a/
for all a 2 Sc and ' 2 R.

1. Let M denote the -pseudospectral radius of the matrix M. The following
inclusion is valid:
Ebs .R/ WD fa 2 Sc W n.a/ M0 < 1g E s .R/:
2. Let h > 0 denote the unique solution of the equation h M0 D 1. The inequality
Ebt .R/ WD supfjjajj W a 2 Bjjajj .0/ \ Ebs .R/ ¤ ;g E t .R/
is valid, and if n is monotone increasing, we have Ebt .R/ D minfjjajj W n.a/ D

h; a 2 Sc g:
Proof (Outline) By definition of the pseudospectral radius, we have
n.a/ M0 D supf.Z/ W Z 2 Rnn ; jjZ M0 jj n.a/g

supfM.'; a/ W ' 2 R, jjM.'; a/ M0 jj n.a/g
D supfM.'; a/ W ' 2 Rg D .R; a/;
which demonstrates the set inclusion. As for the inequality, that the supremum term
is bounded by E t .R/ is obvious from the set inclusion. That Ebt .R/ is achieved at
some a for which n.a/ D h can be seen by noticing that, as n is continuous and
increasing, the set Ebs .R/ is star convex with basepoint 0. Consequently, maximizing
the radius of a ball in the positive orthant within this set is equivalent to minimizing
the distance to the boundary, and the latter is is precisely the level set n.a/ D h.
Corollary 2 Denote X.t/ D X.tI 0 /. If c D 1, the following inequality holds for
all ' 2 R D .; w/.
Z
w0 .s; a/
jjM.'; a/ M0 jj jjX. 1 /jj jjX 1 .s/jj ds
S0 .a/ e a0 1
ˇˇ Z ˇˇ
ˇˇ 1 ˇˇ
C ˇˇˇˇ .X .s/ I/dsB0 ˇˇˇˇ :
1
(14)
a
0 S0 .a/
Proof (Outline)
Z
1 1
M.'; a/ M0 D X. 1 / I C X 1 .s/ '.s; a/ B0 C B0 ds
S0 .a/ a0 a0
X. 1 /ŒI C B0 :
Re-arranging the above, taking norms and using inequality (10) provides the result.
If c ¤ 1, a similar estimate to the above holds. However, it is rather cumbersome,
and the associated proof is a notationally difficult inductive argument. It is omitted
for brevity.
3.1 Example: An Exact Computation for a Scalar Equation
Consider the following scalar impulsive differential equation
x0 D x; t ¤ kT
(15)
x D bx; t D kT;
with parameters > 0, b > 0 and T > 0. Assume M0 D .1 b/e T < 1; so that the
1
trivial solution is asymptotically stable. We choose w.t; a/ D c Ta p for parameters
c and p > 0. The bound on the right-hand side of (14), denote nQ .a/, itself has an
upper bound:

e T
a 1p T 1 e a
nQ .a/ n.a; p/ WD c Ce b 1 ;
T a
and n is strictly increasing in both a and the parameter p. Therefore,

c b
n.a;p/ n.a;1/ D lim n.a;p/ M0 D e T 1 C .1 e a / (16)
p!1 a
for each finite p > 0. Solving the equation n.a ;1/ M0 D 1 for a and applying
Theorem 4, the following theorem is proven.

Theorem 5 Consider the impulsive system (15). Define u WD 1b e T 1 c . If
M0 WD .1 b/e T < 1 and c < .1 M0 /e T , then, for all a > 0 satisfying the
inequality

1 1 1 1
a< W e
u WD a ;
u u
206 K.E.M. Church
1
we have M.'; a/ < 1, for all ' 2 .; w/, with w.t; a/ D c Ta p , for any p > 0,
where W is the principal branch of the Lambert W function, or product logarithm
function (i.e. the inverse of the map x 7! xex ).
3.2 Example: Control of a Pest with Age Structure
Consider the following system of impulsive differential equations.

0 10=21 5=7
X D X; t ¤ k (17)
1=4 1=7

0:7 0
X D X; t D 2k (18)
0 0:4

0 0
X D X; t D 2kC1 ; (19)
0 0:7
with sequence of impulses k D 7bk=2c C .k mod 2/ and t in units of days. The

continuous dynamics, (17), could describe, for example, the population of some
pest organism, X D .X1 ; X2 / 0, with juvenile (X1 ) and adult (X2 ) life stages. With
both impulsive controls included, (17), (18) and (19) is asymptotically stable, with
dominant Floquet multiplier equal to 0:4200. If one or both controls are neglected,
however, the trivial solution is unstable, so both controls are necessary to control the
population.
Figure 1 provides visualizations of the subsets Ebs E s described in Theorem 4,
of the R-stable sets for two uniform exponential regulators for the system (17), (18)
and (19). The first regulator, which generates the smaller of the two stable sets (red
p
in the figure), is R D .; ak /. The second regulator is R D .; ak /. For both
regulators, D jjjj, as in Corollary 1.
3.3 Limitations
There are two main limitations of the techniques described in these proceedings.
First and foremost, only linear systems are treated. The time-scale tolerance can
indeed be defined for nonlinear systems of impulsive differential equations in more
abstract settings, although the definitions must all be localized around periodic orbits
or other stable objects. Some of our current research concerns these problems.
Fig. 1 Conservative approximations, Ebs , of the R-stable sets for two uniform exponential regula-
tors. Arrows indicate the associated lower bounds for time-scale tolerances
Second, we treated only impulsive systems that are asymptotically stable. These
techniques generally fail in the presence of a center subspace; see Example 3.5.6
of [2]. However, a similar approach does work if there is an unstable subspace. For
example, if M0 > 1, one might want to know conditions on a 2 Sc under which
M.'; a/ > 1; for all ' 2 R, with R some suitable set of impulse extensions. If, for
some continuous function n satisfying n.0/ D 0, we have jjM.'; a/ M0 jj n.a/
for all ' 2 R, one can verify the string of inequalities
˚
inf M.'; a/ inf M W jjM M0 jj sup jjM.'; a/ M0 jj
'2R '2R
D inffM W jjM M0 jj inffx W jjM.'; a/ M0 jj x; 8' 2 Rgg

inffM W jjM M0 jj n.a/g WD n.a/ M0
holds. Therefore, an appropriate lower estimate for the analoguous time-scale

tolerance can be found by solving the optimization problem
Ebt .R/ WD supfjjajj W a 2 Bjjajj .0/ fa 2 Sc W n.a/

M0 > 1g ¤ ;g:
208 K.E.M. Church
References
1. Bainov, D.D., Simeonov, P.S.: Impulsive Differential Equations: Periodic Solutions and Appli-
cations. Longman Scientific & Technical, Burnt Mill (1993)
2. Church, K.: Applications of impulsive differential equations to the control of malaria outbreaks
and introduction to impulse extension equations: a general framework to study the validity of
ordinary differential equation models with discontinuities in state. M.Sc Thesis, University of
Ottawa (2014)
3. Church, K.E.M., Smith?, R.J.: Analysis of piecewise-continuous extensions of periodic linear
impulsive differential equations with fixed, strictly inhomogeneous impulses. Dyn. Contin.
Discret. Impuls. Syst. Ser. B: Appl. Algorithms 21, 101–119 (2014)
4. Church, K.E.M., Smith?, R.J.: Existence and uniqueness of solutions of general impulse
extension equations with specification to linear equations. Dyn. Contin. Discret. Impuls. Syst.
Ser. B: Appl. Algorithms 22, 163–197 (2015)
5. Lashmikantham, V., Bainov, D.D., Simeonov, P.S.: Theory of Impulsive Differential Equations.
World Scientific Publishing, Singapore (1989)
6. Samoilenko, A.M., Perestyuk, N.A.: Impulsive Differential Equations. World Scientific Publish-
ing, Singapore (1995)
Coupled Lattice Boltzmann Modeling
of Bidomain Type Models in Cardiac
Electrophysiology
S. Corre and A. Belmiloudi
Abstract In this work, a modified coupling Lattice Boltzmann Model (LBM) in

simulation of cardiac electrophysiology is developed in order to capture the detailed
activities of macro- to micro-scale transport processes. The propagation of electrical
activity in the human heart is mathematically modelled by bidomain type systems.
As transmembrane potential evolves, we take into account domain anisotropical
properties using intracellular and extracellular conductivity, such as in a pacemaker
or an electrocardiogram, in both parallel and perpendicular directions to the
fibers. The bidomain system represents multi-scale, stiff and strongly nonlinear
coupled reaction-diffusion models that consists of a set of ordinary differential
equations coupled with a set of partial differential equations. Due to dynamic
and geometry complexity, numerical simulation and implementation of bidomain
type systems are extremely challenging conceptual and computational problems
but are very important in many real-life and biomedical applications. This paper
suggests a modified LBM scheme, reliable, efficient, stable and easy to implement
in the context of such bidomain systems. The numerical results demonstrate the
effectiveness and accuracy of our approach using general methods for bidomain type
systems and show good agreement with analytical solutions and numerical results
reported in the literature.
1 Introduction
Computational cardiac electrophysiological modeling is now an important field in

applied mathematics. Indeed, nowadays, heart and cardiovascular diseases are still
the leading cause of death and disability all over the world. That is why we need
to improve our knowledge about heart behavior, and more particularly about its
electrical behavior. Consequently we want strong methods to compute electrical
fluctuations in the myocardium to prevent cardiac disorders (as arrhythmias), or
to study interactions between brain and heart. We modelize electrical behavior
S. Corre • A. Belmiloudi ()

UEB-IRMAR, Rennes, France

210 S. Corre and A. Belmiloudi
of the heart in the myocardium with the bidomain system, derived from Ohm’s
law. This biophysical model of electrical cardiac activity links electrophysiological
cell models, at small scales, and myocardial tissue mechanics, metabolism and
blood flow at large scales. In mathematical viewpoint, bidomain system leads us
to compute intracellular and extracellular electrical potentials i and e with taking
into account the cellular membrane dynamics U. This is a system of non-linear
partial differential equations (PDEs) coupled with ordinary differential equations
(ODEs). The PDEs describe the propagation of the electrical potentials and ODEs
describe the electrochemical processes. During last years a lot of studies about
bidomain models have led to results about well-posedness, existence and uniqueness
of solutions (see e.g., [2, 3] and the references therein), and several numerical
methods based on methods as finite difference method or finite element method
are used to solve these models (see e.g., [8] and the references therein). In this
paper, we propose a modified Lattice Boltzmann Method (LBM) which is simple
to implement, effective, accurate and well suited to bidomain systems which is a
coupled nonlinear parabolic/elliptic PDEs. LBM is based on microscopic models
and mesoscopic kinetic equations. Indeed, traditional numerical methods as finite
difference method or finite element method directly solve governing equations
for deriving macroscopic variable, whereas LBM is based on the particle (the
discrete) distribution function and numerical solving the continuous Boltzmann
transport equation. Then the macroscopic variables of the bidomain system can be
recovered from the discrete equations through the multi-scaling Chapman-Enskog
expansion procedure. LBM was originated from Boltzmann’s kinetic theory of
gases (1970s), and attracts more and more attentions for simulating complex fluid
flows since 1990s. More recently, LBM has been extended successfully to simulate
different types of parabolic reaction-diffusion equation as Keller-Segel chemotaxis
model [10] and monodomain model in cardiac electrophysiology [4], or Poisson
equation [5].
This paper is organized as follows: in Sect. 2 we recall briefly the derivation
of the bidomain model. In Sect. 3 we present and describe the modified LBM
method. In Sect. 4, the validity of this method is demonstrated by comparing
the numerical solution to the exact solution of bidomain model with a classical
FitzHugh-Nagumo model (FHN), and convergence of solution is established. Some
interesting numerical simulations to analyze the influence of some parameters on
electrical wave propagation, including the bidomain model with a modified FHN
model, are also carried out in this section. This paper is ended by a conclusion and
some further works.
2 The Bidomain Model
The bidomain model of cardiac tissue is expressed mathematically by the following

steady-state of coupled partial differential equations governing the electrical poten-
tials (in the physical region ˝ occupied by the excitable cardiac tissue, which is an
LBM for Bidomain Type Models 211
open, bounded, and connected subset of Rd , d 3 and during a time interval .0; T/)
div.K i ri / D Im fis ; div.K e re / D Im fes ; (1)
where i and e are the intracellular and extracellular potentials, respectively; K i .x/
and K e .x/ are the conductivity tensors describing the anisotropic intracellular and
extracellular conductive media; fis .x; t/ and fes .x; t/ are the respective externally
applied current sources. The transmembrane current density is described by Im and
is given by the following expression:
@
Im D .cm C I ion /; (2)
@t
where is the transmembrane potential, which is defined as D i e , is the
ratio of the membrane surface area to the volume occupied by the tissue, cm term is
the transmembrane capacitance time unit area. The tissue is assumed to be passive,
so the capacitance cm can be assumed to be not a function of the state variables.
The nonlinear operator I ion .x; tI ; U/ describes the sum of transmembrane ionic
currents across the cell membrane with U the electrophysiological ionic state
variables (which describe e.g., the dynamics of ion-channel and ion concentrations
in different cellular compartments). These variables satisfy the following ODE (with
H a nonlinear operator)
@U
D H.x; tI ; U/: (3)
@t
From (1), (2) and (3), the bidomain model can be formulated in terms of the state
variables , e and U as follows (in Q D ˝ .0; T/)
@
.cm C I ion .:I ; U// div.K i r/ D div.K i re / C fis ;
@t
div..K e C K i /re / D div.K i r/ C . fes C fis /; (4)
@U
D H.:I ; U/:
@t
The operators I ion and H which describe electrophysiological behavior of the
system have usually the following form (affine functions with respect to U)
I ion .:I ; U/ D I1 .:I / C I2 .:I /U; H.:I ; U/ D H0 .:I / C .:/U: (5)
To close the system we impose the following boundary conditions
.K i r. C e //:n D i ; .K e re /:n D e on ˙ D @˝ .0; T/; (6)
where n being the outward normal to

D @˝ and i and e are the intra- and extra-
cellular currents per unit area applied across the boundary, and the following initial
conditions (in ˝)
.t D 0/ D 0 ; U.t D 0/ D U0 : (7)
Such problems have compatibility conditions determining whether there are any
solutions to the PDEs. This is easily found by integrating the second equation of (4)
over the domain and using the divergence theorem with the boundary conditions (6)
(a.e. in (0, T)). Then (for compatibility reasons), we require the following condition
Z Z
.i C e /d
C . fes C fis /dx D 0: (8)

˝
Moreover, the function e is defined within a class of equivalence, regardless of

a time-dependent function. This function can be fixed, for example by setting the
Gauge condition (a.e. in (0, T))
Z
e dx D 0: (9)
˝
Under some hypotheses for the data and parameters of the system and some regular-
ity of operators I ion and H, system (4) with (6)–(7) and under the conditions (8)–(9)
is a well-posed problem (for more details see [2]).
3 Numerical Method and Algorithm
In this section, a numerical method is presented for the bidomain system (4) in two
space dimensions. For this, we introduce a coupled modified LBM for solving the
coupled system of nonlinear parabolic and elliptic equations (i.e. the first and the
second equations of (4)). Then we treat the ODE satisfied by ionic state by applying
Gronwall Lemma to obtain an integral formulation, and by using a quadrature rule
to approximate the obtained integral. In the sequel, without loss of generality, we
assume cm D 1 and D 1. Moreover we assume K i D Ki Id , K e D Ke Id , with Id
identity matrix and Ki , Ke constants.
Remark 1 The developed LBM method has been constructed to take into account
the case in which K i D K i .; e / and K e D K e .; e /. In order to simplify
the presentation, we have assumed in this paper that these operators are constant
matrices.
3.1 LBM for Coupled Parabolic and Elliptic Equations
In this first part, we develop and describe the modified LBM to solve the following
system (which corresponds to two first parts of (4))
@
div.Ki r. C e // D F.:I ; e /;
@t (10)
div.Ki r C .Ki C Ke /re / D G.:I ; e /;
where F and G are non linear operators.

3.1.1 LBE for General Reaction-Diffusion Equations
To begin, we introduce the LBM to solve the following reaction-diffusion equation

with the macroscopic variable ˚
@
˚.x; t/ div.Kr˚.x; t// D H.x; tI ˚/: (11)
@t
The evolution equation of the LBM for (11) is given by
@
h.x; tI e/ C e rh.x; tI e/ D Q.h.x; tI e// C P.x; tI e/; (12)
@t
where h.x; tI e/ is the distribution function of particle moving with velocity e at
position x and time t, P is the distribution type function of particle of macroscopic
external force H moving with velocity e and Q is the Bhatnagar-Gross-Krook (BGK)
collision operator defined by Q.h/ D 1 .h.x; tI e/ heq .x; tI e// ; where heq is
the equilibrium distribution function and is the dimensionless relaxation time.
LBM leads us to approximate (12) to recover reaction-diffusion equation (11) with
Chapman-Enskog expansion. For that, we discretize Q in time and space. Then
we introduce a lattice size cell x and a lattice time step size t, and we define
streaming lattice speed c D x= t.
The D2Q9 lattice, which involves 9 velocity vectors, is considered for applied
lattice scheme, which is the most used scheme for two-dimensional model. Figure 1
shows a typical lattice node of D2Q9 model with velocities ei for various directions
defined by

0 cos .i 1/ 2 p cos .i 92 / 2
e0 D ; eiD1;2;3I4 D c ; eiD5;6;7;8 D 2c :
0 sin .i 1/ 2 sin .i 92 / 2
Fig. 1 Particle velocities for

D2Q9 LBM
Fig. 2 streaming process of a lattice node
For each particle on the lattice, we associate the discrete distribution functions hi
eq
and hi (in the mesoscopic level), and the discrete operator Hi of H for i D 0; : : : ; 8.
Then, the form of the Lattice Boltzmann Equation (LBE) with an external force by
introducing BGK approximations can be written as follows
1
eq
hi .x C ei t; t C t/ D hi .x; t/
hi .x; t/ hi .x; t/
t2 @
C tHi .x; t/ C H .x; t/:
2 @t i
(13)
The key steps in LBM, which is directly derived from LBE (13), are the collision
and streaming processes (shown on Fig. 2) which are given by
1
eq t2 @
i .x; t/ D hi .x; t/ hi .x; t/hi .x; t/ C tHi .x; t/C 2 @t Hi .x; t/;
hcol (14)
hi .x C ei t; t C t/ D hcol
i .x; t/: (15)
From Chapman-Enskog expansion analysis, the above LBM can recover to the
eq
reaction-diffusion equation (11) if we take hi D wi ˚, Hi D wi H and the initial
eq
distribution at t D 0: hi .x; 0/ D hi .x; 0/. This analysis is based on the following
properties:
8
X 8
X
eq
hi .x; t/ D hi .x; t/ D ˚.x; t/.macroscopic variable/;
iD0 iD0
X8 8
X 8
X
eq
Hi .x; t/ D H.x; t/; ei hi .x; t/ D 0; ei Hi .x; t/ D 0; (16)
iD0 iD0 iD0
X8 X8
eq c2 c2
ei ei hi .x; t/ D ˚.x; t/Id ; ei ei Hi .x; t/ D H.x; t/Id :
iD0
3 iD0
3
Remark 2 During streaming and collision processes, in order to satisfy boundary

conditions, the boundary nodes need special treatments on distribution functions,
which are essential to stability and accuracy of the method.
3.1.2 Coupled Modified LBM
To introduce our modified LBM, we have to take into account the coupled terms
which link reaction-diffusion equation and elliptic equation in the system (10). As
in [10], in order to take into account this coupling, we introduce two correction terms

Si D wi .1 .Ki r/ C 2 .Ki re // and Sie D wi 1e .Ki r/ C 2e ..Ki C Ke /re / ,
where the functions 1 ; 2 ; 1e and 2e are determined by Chapman-Enskog expan-
sions. Then, we can solve the reaction-diffusion equation with a first LBE where the
distribution function f leads to recover . We construct exactly the same LBM than
developed in Sect. 3.1.1. So we choose fi D wi , Fi D wi F and D c3K 1
eq
2 t C 2 to
i
satisfy previous properties (16). Finally, we add the corrector term Si as follows
1 eq
fi .x C ei t; t C t/ D fi .x; t/ fi .x; t/ fi .x; t/ C tFi .x; t/

t2 @
C Fi .x; t/ C tSi .x; t/: (17)
2 @t
P
8
Hence, the macroscopic variable , defined as: fi .x; t/ D .x; t/:
iD0
For the elliptic equation, the LBM developed is based on the LBM employed in
[5]. The first step is to introduce a new time variable r as lim Qe .x; rI t/ D e .x; t/.
r!1
Then, the equilibrium distribution function is defined as

eq wi Qe .x; rI t/ for i ¤ 0;
gi .x; r/ D
.w0 1/Qe .x; rI t/ for i D 0
P
8
eq 1
P
8
eq
and we can deduce that gi .x; r/ D 0; 1w0
gi .x; r/ D Qe .x; rI t/: Finally, as for
iD0 iD1
previous LBE (17), we add the corrector term Sie and we obtain the following LBE
(to recover e )
1 eq
gi .x C ei r; r C rI t/ D gi .x; rI t/ gi .x; rI t/ gi .x; rI t/
e
C tGi .x; rI t/ C tSie .x; rI t/: (18)
8
1 X
Hence, the macroscopic variable Qe , defined as: gi .x; rI t/ D Qe .x; rI t/:
1 w0 iD0
3.2 Treatment of ODE
Now, we present briefly the method to solve ODE satisfy by ionic state U, with
initial condition U.x; 0/ D U0 .x/. According to (4) and the form of H given in (5)
by H.x; tI ; U/ D H0 .x; tI / C U.x; t/, with assumed to be a constant, and by
using Gronwall Lemma we can deduce:
Z
tC t
t t
U.x; t C t/ D U.x; t/e Ce H0 .x; sI .s//es ds:
t
Then, according to approximation of derived integral by trapezoidal method

between t and t C t we obtain the following approximation of U denoted also
by U
t
U.:; t C t/ D U.:; t/e t C H0 .:; t C tI .t C t//e t CH0 .:; tI .t// :
2
(19)
Finally, after non-dimentionalization, mesh definition and initialization of initial
conditions, parameters and data, the proposed algorithm to solve the bidomain
system can be summarized as follows
1. Initialization: tD0.
2. LBE according to time t by using (17) to compute .x; t C t/.
3. Trapezoidal method by using (19) to compute U.x; t C t/.
4. Loop on new time variable r:
a. LBE according to time r by using (18) to compute Qe .x; r C rI t C t/ an
approximation of e .x; t C t/ .
b. If convergence criteria is not reached, set r D r C r and go back to 4a.
5. Set e .x; t C t/ WD Qe .x; rI t C t/.
6. If t ¤ T, set t WD t C t and go back to 2.
4 Numerical Simulation and Applications
To validate the capacity of our modified coupled LBM to deal with 2D bidomain
systems, several situations are numerically simulated. In the first study, we consider
the bidomain system with homogeneous Neumann boundary conditions and with a
classical FitzHugh-Nagumo model (FHN), in which non linear operators are defined
as
I D 1=˛1 . 3 =3 U/; H D ˛2 . ˇ1 U C ˇ2 /; (20)

where ˛1 ; ˛2 ; ˇ1 ; ˇ2 are positive constant parameters respectively called excitation

rate constant, recovery rate constant, recovery decay constant and excitation decay
constant. First, we consider the system with known analytical solution in order
to validate and verify the accuracy and stability of the method. And second, we
consider the system as in [8] and we estimate the influence of both constants ˛1 and
˛2 on the behaviour of the system. In the last application we solve and analyze the
system with the following modified FHN model (see e.g., [8])
I D 0:0004. C 85/ .U C . C 70/. 40// ; H D 0:63. C 85/ 0:013U:

(21)
Nota Bene: If the exact solution sol is known, we can measure the efficiency of
ksol kL2 .˝/
method with the following L2 relative error: Err D ksol kL2 .˝/
:
For simplicity, we assume that the domain is a square region ˝ D Œ0I 1 Œ0I 1,
the final time is fixed T D 1, and the cell surface to volume ratio D 1.
4.1 Benchmark Problem and Validation
In this first analysis, we investigate the accuracy and spatial convergence rate
of the proposed modified LBM for which we postulate that the error estimates
of the method is of order 2 in space and of order 1 in time (for sufficiently
regular solution). We perform a convergence study on cartesian grid, by taking
˛1 D 1, ˛2 D 1, ˇ1 D 1, Ki D 1 and Ke D 1, and by setting ˇ2 .x; y; t/ D
sol .x; y; t/ C 2et cos..x C y//: The initial conditions are .x; 0/ D 0; e .x; 0/ D
0; U.x; 0/ D cos..x C y// and we choose I , Ii and Ie to close the problem
according to solution and compatibility conditions. The exact solution is given by:
sol .x; y; t/ D tx2 .x 1/2 y2 .y 1/2 ; esol .x; y; t/ D t.cos.x/ C cos.y//; and
U sol .x; y; t/ D et cos..x C y//. To study the convergence, we have constructed
a sequence of meshes with decreasing spatial step x between 1=30 and 1=200 and
t D x2 . We just care about relative error on and U at t D 0:5 and t D 1
(see Table 1). Indeed, the chosen convergence criteria (for the iterative method to
Table 1 Relation between relative error and lattice spacing for and U
t = 0.5 x Err ErrU t=1 x Err ErrU
1=30 0.1914 0.0005 1=30 1.7618 0.0009
1=50 0.0389 0.0002 1=50 0.4559 0.0003
1=70 0.0116 7.97e5 1=70 0.1654 0.0001
1=100 0.0050 4.48e5 1=100 0.0790 7.92e5
1=150 0.0016 1.99e5 1=150 0.0284 3.52e5
1=200 0.0007 1.12e5 1=200 0.0138 1.99e5
Fig. 3 Error curves for and U
approach e ) involves a constant error because this criteria is not defined in function
of lattice size cell. We present on Fig. 3 (at t D 0:5 and t D 1) the convergence
curves, log.Error/ versus x, for and U. We observe that the slope of error
curves for passes approximately from 3 to 2:5 and the slope of error curves for
U is approximately equal to 2. This shows that our numerical error estimates are
agree with the postulated error estimates, and indicate the good performance of our
proposed method.
4.2 Influence of Some Parameters on Electrical Wave

Propagation
This second computations have been made to test performance of our method
by analyzing the influence of some parameters on electrical wave propagation
according to FHN models: a classical model (20) and a modified model (21). Here,
we take x D 1=50 and t D x2 .
4.2.1 First Data Setting
In this first application, we consider the classical FHN model (20) with fixed ˇ1 D
0:5 and ˇ2 D 1 and we study the effect of parameters ˛1 and ˛2 . We assume Ki D
Ke D 1 and we consider the following initial conditions .x; y; 0/ D 1:28791 C
sin.x/; e .x; y; 0/ D 0; U.x; y; 0/ D 0:5758:
The parameters ˛1 and ˛2 control the dynamic between transmembrane potential
and the ionic state U. In order to study the propagation of an electrical wave
through the cardiac tissue, we analyze the relaxation time of . In Table 2, we
observe that ˛2 has negligible impact on this relaxation time. Conversely, Fig. 4 and
Table 2 show that tiny ˛1 value leads to near-infinite slope values when potential
peaks and relaxes.
Table 2 Relaxation time of according to ˛1 and ˛2 variations

˛1 ˛2 Relaxation time ˛2 ˛1 Relaxation time
0.01 0.4700 0.01 0.0396
0.2 0.05 0.4704 0.2 0.05 0.1568
0.1 0.4716 0.1 0.2764
0.5 0.4988 0.5 0.9168
Fig. 4 .0:5; 0:5; t/ with ˛1 D 0:01 to 0:5, ˛2 D 0:2

60
4 −10
40
−20
3 20
−30
2 0
ρ(x, y, 1)
ρe (x, y, 1)
U (x, y, 1)
−40
−20 −50
1
−40 −60
0
−60 −70
−1 −80
−80
1 0 1
1 1
0.5 0 0.5
0.5
0.5 0.5
0.5
y 0 0 x 1 1 y 0 0 x
y x
Fig. 5 Evolution of e , and U at t D 1 with x D 1=50
4.2.2 Second Data Setting
In this final application, we consider the modified FHN model (21) which is known
to be more adapted than the classical model (20) because of a product between
and U. We assume Ki D 1:75 and Ke D 7 and we consider the following initial
conditions: .x; y; 0/ D 85 C 100.1 sin.xy//; e .x; y; 0/ D 0; U.x; y; 0/ D 0:
Figure 5 shows a color visualization of the simulation for , e and U at final time
t D T. The analyze we have done before is still valid.
Remark 3 It is known that the action potential duration (APD) and conduction
velocity (CV) could be significantly affected by variation in the values of FHN
parameters. So, it will be interesting to perform comparison of the APD and CV for
the various FHM ionic-type models (and others) coupling the bidomain model, but
also to understand how sensitive APD and CV are to variability in these parameters.
5 Conclusion and Commentary
An efficient and stable coupled LBM to solve a two-dimensional bidomain model

of cardiac tissue is developed. From the Chapman-Enskog expansion analysis,
the bidomain system which is a coupled of reaction-diffusion, elliptic and EDO
equations, can be correctly recovered by our modified LBM. This method is easy to
implement and easy to parallelize. It is clear that, due to the multi-scale nature of the
system, the cartesian grid used in our preliminary simulations is not very sufficient
to compute in a computationally efficient manner real life clinical situation with
complex geometry which is in general computationally expensive. Therefore, it is
expected to solve the Lattice Boltzmann system on adapted cartesian or triangular
unstructured grids as e.g., in [6, 9] and the references therein. Moreover, in order
to overcome the limitations of the constraint CFL stability condition, we extend
the method to implicit or semi-implicit time schemes, e.g., by using the -method
(with 2 Œ0; 1) or Runge-kutta methods, coupled with adaptive time stepping
strategies, as e.g. in [7] and the references therein. This coupled LBM method will
be shown in a forthcoming paper for more general coupled models with realistic
complex geometries. It would be interesting to use this developed method with
observations coming from experimental data and a more complete description of
the biophysical model of electrical cardiac activity. In order to get even closer to a
more realistic calculation, it is necessary to study in the future this method coupled
with optimization technique and robust control problems by using the approach
developed in [1].
Acknowledgements The authors are grateful to the referee for many constructive comments and
suggestions which have improved the presentation of this manuscript.
References
1. Belmiloudi, A.: Stabilization, Optimal and Robust Control. Theory and Applications in
Biological and Physical Sciences. Springer, London (2008)
2. Belmiloudi, A.: Robust control problem of uncertain bidomain models in cardiac electrophisi-
ology. J. Coupled Syst. Multiscale Dyn. 19, 332–350 (2013)
3. Bourgault, Y., et al.: Existence and uniqueness of the solution for the bidomain model used in
cardiac electrophysiology. Nonlinear Anal-Real 10, 458–482 (2009)
4. Campos, J.O., et al.: Lattice Boltzmann method for parallel simulations of cardiac electrophys-
iology using GPUs. J. Comput. Appl. Math. 295, 70–82 (2016)
5. Chai, Z., Shi, B.C.: Novel Boltzmann model for the Poisson equation. Appl. Math. Model. 32,
2050–2058 (2008)
6. Fan, Z., et al.: Adapted unstructured LBM for flow simulation on curved surfaces. In: ACM
SIGGRAPH, Los Angeles, pp. 245–254 (2005)
7. Huang, J., et al.: A fully implicit method for lattice Boltzmann equations. SIAM J. Sci. Comput.
37(5), Special Section, S291–S313 (2015)
8. Sharomi, O., Spiteri, R.: Convergence order vs. parallelism in the numerical simulation of the
bidomain equations. J. Phys.: Conf. Ser. 385, 1–6 (2012)
9. Valero-Lara, P., Jansson, J.: A non-uniform Staggered Cartesian grid approach for Lattice-
Boltzmann method. Procedia Comput. Sci. 51, 296–305 (2015)
10. Yang, X., et al.: Coupled lattice Boltzmann method for generalized Keller-Segel chemotaxis
model. Comput. Math. Appl. 12, 1653–1670 (2014)
Dynamics and Bifurcations in Low-Dimensional
Models of Intracranial Pressure
D. Evans, C. Drapaca, and J.P. Cusumano
Abstract Intracranial Pressure (ICP) is a physiological parameter of the brain

which plays an important role in the diagnosis and treatment of pathologies such
as hydrocephalus and traumatic brain injury. Currently, all reliable methods for ICP
monitoring involve drilling through the skull to place a pressure probe inside the
brain. As a result, ICP is only measured in the most critical cases which require
neurosurgical intervention. Mathematical models that relate ICP to physiological
parameters whose measurements are minimally invasive could contribute to better
diagnostic and treatment protocols for brain disorders. Ideally, such mathematical
models should have the capability to predict ICP in real time from non-invasive
measurements of other clinically relevant parameters without the need for high-
risk procedures. In this paper, we examine in detail the dynamics and stability of a
mathematical model proposed by Ursino and Lodi in (J Appl Physiol 82(4):1256–
1269, 1997) which predicts ICP from measurements of arterial blood pressure. We
study how the equilibria vary with the model parameters and, aided by numerical
simulations, we obtain bifurcation diagrams for the system of non-linear ordinary
differential equations. Expanding upon the work of Ursino and Lodi, we show that
the model exhibits not only one Hopf bifurcation but also a reverse Hopf bifurcation
in certain parameter regimes. In addition, we present global phase portraits of the
system in interesting parameter configurations.
1 Introduction
Intracranial pressure (ICP) monitoring provides critical physiological information

on brain damage in patients suffering from neurological disorders such as hydro-
cephalus and traumatic brain injury (TBI). The only reliable method of measuring
ICP involves drilling a hole directly through the skull, and inserting a catheter with
a pressure probe into the brain [5]. Given the invasiveness of this approach, ICP
monitoring is only performed on patients in very critical conditions during life-
saving neurosurgical interventions. Mathematical models that couple ICP dynamics
D. Evans () • C. Drapaca • J.P. Cusumano

The Pennsylvania State University, University Park, PA, USA

224 D. Evans et al.
and other physiological parameters whose monitoring is less invasive provide a

powerful alternative to the direct measurement of ICP. In addition, mathematical
models that also incorporate brain’s regulatory mechanisms of ICP could prove
essential in designing better diagnostic and therapeutic procedures for a wide range
of neurological diseases. Although the potential exists to create a model that could
be used to estimate ICP in real time, the current state of the art is still far from
this goal. A comprehensive review of mathematical models of ICP dynamics can be
found by Wakeland and Goldstein [8].
One particular model proposed by Ursino and Lodi [6], predicts ICP dynamics
from non-invasive measurements of arterial blood pressure, which appears to be
in agreement with the ICP measured in patients with TBI. The model, which we
will refer to as the Ursino-Lodi model, is a generalization of the one-compartment
mathematical model of ICP dynamics pioneered by Marmarou [4] that includes
several blood compartments, each with variable conductances and compliances, and
ICP auto-regulation due to blood vessel wall tension, and viscous forces. Since the
focus of the Ursino-Lodi model was its clinical applicability, a full mathematical
analysis of the model is still missing. Therefore, the aim of this paper is to examine
in detail the dynamics and stability of the Ursino-Lodi model. We use numerical
simulations to study how the equilibria vary with the model parameters and obtain
bifurcation diagrams for the system of non-linear ordinary differential equations.
Our in-depth analysis shows that the model exhibits not only one Hopf bifurcation,
as reported in [6], but also a reverse Hopf bifurcation in certain parameter regimes.
We also present global phase portraits of the system in interesting parameter
configurations.
2 Physiological Considerations
The human brain is a multi-phase mixture of many different materials, including

cerebrospinal fluid (CSF), brain cells, and cerebral vasculature. The majority of
what is considered brain tissue consists of white matter and grey matter, which are
both soft and mildy compressible materials [3]. CSF is a colorless liquid made of
99 % water that fills the brain’s ventricles, the subarachnoid space (space between
the two deeper meninges that envelop the brain tissue: pia mater and the arachnoid
mater), and intracellular space. The presence of CSF introduces damping to the
brain motion, which cushions and protects it from injury [2]. CSF is constantly
formed by the choroid plexuses of the brain’s ventricles from the cerebral blood,
circulates through the brain tissue and spinal cord, and is constantly reabsorbed into
the cerebral veins located in dura mater (above the arachnoid mater) [5]. There exists
an approximately constant pressure gradient from the inside of the brain (ventricles
containing CSF) to the outside (dura matter). CSF is then forced over this pressure
gradient over time, allowing it to be reabsorbed by the venous system. The rate of
absorption is linked to the intracranial pressure as follows. Above a certain value of
ICP, the relationship is nearly linear while at very low values of ICP, the absorption
Dynamics and Bifurcations in Low-Dimensional Models of Intracranial Pressure 225
rate is negligible [1]. The absorption of CSF back into the bloodstream, and hence,
its removal from the intracranial compartment, is the primary natural mechanism for
lowering the intracranial pressure. As a result, abnormally high intracranial pressure
is sometimes caused by high resistances to resorption.
In addition to the brain naturally regulating pressure via the volume of CSF,
the brain also contains autoregulatory mechanisms to control the flow of blood
in the cerebral arteries. In order to ensure a constant artery-to-vein difference in
oxygen concentration, the brain automatically regulates the cerebral blood flow
(CBF) in a process known as cerebral autoregulation. As oxygen consumption,
blood pressure, and blood viscosity change, the cerebral arteries dilate or constrict
to regulate the flow of blood to the brain. The volume of the cerebral arteries, the
compression of the brain tissue, and the volume of CSF within the skull are all
controlled by ICP. The linked mechanisms of feedback and control through all of
these processes have the potential to create instability in the steady-state behavior
of the ICP, most notably in the case of Lundberg A waves. This pathological
phenomenon is characterized by long, sustained increases in the ICP. The ICP, which
is normally approximately constant, increases dramatically in oscillations with
periods of approximately 5–20 min, and amplitudes of about 50–100 mmHg [5].
3 Ursino-Lodi Model
The Ursino-Lodi mathematical model [6] includes both the ventricular CSF and the
cerebral vasculature. The model is made of five compartments: one compartment
for the ventricular CSF and four compartments for the blood in the cerebral arteries,
capillaries, veins, and venous sinus, respectively. In addition, the model provides
a relationship between pressures and volumes of cerebral blood and the pressure
and volume of the ventricular CSF. Lastly, the model includes equations which
dynamically control the flow of cerebral blood, which account for the effects of
cerebral autoregulation. By modeling the effect of cerebral autoregulation, the
Ursino-Lodi model has the capability to predict the oscillatory Lundberg A Waves.
Applying the mass conservation principle to the ventricular component yields
dPic dVa Pc Pic Pic Pvs
Cic D C C Ii ; (1)
dt dt Rf Ro
where Cic is the intracranial “compliance”, Va is the blood volume in the cerebral
arteries and arterioles, Pc ; Pic ; and Pvs are the capillary, intracranial, and venous
sinus pressures, respectively, Rf and Ro are the resistances to CSF formation and,
respectively, CSF outflow, and Ii is the rate of externally injected of extracted CSF
volume. As in [4], the nonlinear compliance term Cic is assumed to be related to Pic
by the following formula:
1
Cic D ; (2)
kE Pic
226 D. Evans et al.
where kE is brain’s elastance coefficient (a stiffness-like modulus introduced by [4]).

The mass preservation principle for the compartment of cerebral capillaries yields
the following equation:
Pa Pc Pc Pic Pc Pic Pc Pic

D C ; (3)
Ra Rf Rpv Rpv
where Ra and Rpv denote the resistances of the arterial-arteriolar and respectively
venular vessels. The approximation made in Eq. (3) is based on the assumption that
the CSF production rate is much less than the cerebral blood flow rate.
By assuming a linear blood volume-pressure relationship Va D Ca .Pa Pic /,
where Ca denotes the arterial-arteriolar compliance, the following equation can be
obtained through differentiation:

dVa dPa dPic dCa
D Ca C .Pa Pic /: (4)
dt dt dt dt
By replacing Eqs. (2) and (4) into Eq. (1), we obtain:

dPic kE Pic dPa dCa Pc Pic Pic Pvs
D Ca C .Pa Pic / C C Ii :
dt 1 C Ca kE Pic dt dt Rf Ro
(5)
Lastly, the model is completed by assuming that the cerebral autoregulation modifies
the compliance Ca as follows:
dCa 1
D .Ca C .Gx//; (6)
dt
where is a time constant, G is the maximum autoregulation gain, and
.Can C Ca =2/ C .Can Ca =2/ exp.Gx=k /

.Gx/ D ; (7)
1 C exp.Gx=k /
with k D Ca =4.
Pa Pc
By definition, the cerebral blood flow q is q D and thus the normalized
Ra
deviation of q from its normal value qn is:
q qn
xD : (8)
qn
The parameter Ca represents the maximum allowed change in the arterial com-
pliance Ca from its basal value Can . Furthermore, it depends on whether the
arterial-arteriolar vessels are contracting (x > 0) or dilating (x < 0) as follows:

Ca1 ; if x < 0
Ca D (9)
Ca2 ; if x > 0 :
By assuming that the arteries are circular cylinders of radius r and the blood is
a viscous Newtonian fluid, the Hagen-Poiseuille equation which states that the
arterial resistance is inversely proportional to r4 can be used to obtain the following
expression for Ra :
2
kR Can
Ra D 2
: (10)
Va
The system of non-linear ordinary differential equations (5) and (6) is the Ursino-
Lodi model written in terms of the state variables Pic and Ca .
4 ICP Dynamics Predicted by Ursino-Lodi Model
The clinically measured ICP is a pulsatile periodic waveform which has a period
equal to one heartbeat. However, the Lundberg A waves happen on a time scale
that is much longer than one heartbeat, so in order to observe them we replace
the heartbeat with an averaged pressure over one period which greatly simplifies
the analysis without loss of applicability. Notably, it ensures that the system of
differential equations we work with remains autonomous. The parameters of the
Ursino-Lodi model are the variables that are treated as constants with respect to
time. A point in the parameter space representing basal parameters accompanies the
model, which is presented as Table 1. In our numerical simulations, all parameters
are set to these values unless otherwise indicated. Variations from these parameters
Table 1 Ursino-Lodi Model Parameter Value

Basal Parameters [6]
Ro 526:3 mmHg s ml1
Rpv 1:24 mmHg s ml1
Rf 2:38 103 mmHg s ml1
Ca1 0:75 ml=mmHg
Ca2 0:075 ml=mmHg
Can 0:15 ml=mmHg
kE 0:11 ml1
kR 4:91 104 mmHg3
20 s
qn 12:5 ml=s
G 1:5 ml mmHg1 100 % CBF change1
Pa 100 mmHg
Pic 9:5 mmHg
Pc 25 mmHg
Pvs 6:0 mmHg
Ca 0:15 ml=mmHg
228 D. Evans et al.
will be represented as unitless normalized values with respect to the values given in
Table 1.
The equilibrium points in the model are found at locations where PP ic D CP a D
0, where the dot operator indicates the time derivative. Given positive real input
parameters, this system has consistently two equilibrium points in the domain of
Pic 0, Ca 0. These two equilibria are marked with stars in Fig. 2b. What we
will refer to as the “primary” equilibrium point in this work is marked “SN” here, for
stable node, and what we refer to as the “secondary” equilibrium point is marked
“SP” here, for saddle point, which we will discuss shortly. Although the primary
equilibrium point is a stable node for the values of Table 1, we will soon discuss
how deviations from these can change its stability.
Nevertheless, we treat both equilibria equally when employing the following
techniques. By slowly varying the initial state of the system by one model parameter
(here, either Ro , kE , or G), we obtain a locus of points that correspond to the steady
state value of the system. Using the basal case as an initial guess, we use a root-
solver (the MATLAB built-in function fzero) to track the equilibrium values of
Pic and Ca as a function of three of the model parameters: Ro , kE , and G, for
both equilibrium points. Because there is more than one root for the equilibrium
state, we kept each solution branch separate by, upon increase of the bifurcation
parameter, using the previous solution as the subsequent initial guess for the root-
solver. Figures 1 and 2a show how the locations of both equilibria vary as a function
of both Ro and G. We found that varying the intracranial elastance, kE , does not
affect the equilibrium values predicted by the model for either equilibrium point.
In addition, the location of the secondary equilibrium point does not change as a
function of Ro .
We would like to describe how the stability of these equilibria change with
respect to the model parameters. We therefore linearize the system of equations (5)
and (6) about the equilibria to obtain the linear system xP D Jx. Here, x D
Œ.Pic Pic /; .Ca Ca /T is the perturbation from the equilibrium value, .Pic ; Ca /,
and J is the Jacobian matrix of the two-dimensional vector field defined by the right
hand sides of Eqs. (5) and (6). Then, the eigenstructure of J determines the stability
of the equilibrium point [7].
Confirming what was previously inferred from the phase portrait, the primary
equilibrium point is a stable node, having two negative real eigenvalues in the basal
condition. Since there are no other stable equilibria in the domain, the basin of
attraction is the entire domain Pic > 0, Ca > 0. In addition, we find that the
secondary equilibrium point always occurs at a value of Pic D 0 regardless of other
input parameters. The eigenvalues of the linearized system near this equilibrium are
real with opposite signs. Therefore, this equilibrium can be classified as a saddle
point.
It is clinically valid that the primary equilibrium is a stable node in the basal
condition. That is, given an initial state (Pic , Ca ), the system will autonomously
regulate itself to one steady-state value. However, as previously alluded to, there
will be instances where the system no longer converges to a constant steady-state
value, but instead converges to a stable limit cycle in the phase space, representing
(a) 70 (b) 0.5

60 0.45
Ca Root value (ml/mmHg)

50 0.4
P Root (mmHg)
0.35
40
0.3
30
0.25
ic
20
0.2
10 0.15
0 0.1
0 5 10 15 20 25 0 5 10 15 20 25
Ro (normalized) R (normalized)
o
(c) 34 (d) 0.25

33 Ca Root (ml/mmHg) 0.24
32 0.23
P Root (mmHg)
31 0.22
30 0.21
ic
29 0.2
28 0.19
27 0.18
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
G (normalized) G (normalized)
Fig. 1 Location of the primary equilibrium point. (a) P

ic versus Ro , (b) Ca versus Ro , (c) Pic
versus G, (d) Ca versus G
(a) 0.137 (b) 90 Equilibria

Trajectories
80
0.136
70
ic
0.135
Intracranial Pressure: P
0.134 60
0.133 50
0.132 40
0.131 30
0.13 20
0.129 10
0.128 0
0 1 2 3 4 5 6 7 0 0.1 0.2 0.3 0.4 0.5 0.6
G (normalized) Arterial Compliance: Ca
Fig. 2 (a) Location of Ca on the secondary equilibrium point as a function of G (P ic is unaffected).
(b) Phase portrait of the system in the basal condition as per Table 1. Here, the primary equilibrium
point is a stable node (SN), while the secondary equilibrium point is a saddle point (SP)
230 D. Evans et al.
the clinically observed case of Lundberg A Waves. Mathematically, the transition

between the two cases implies the occurrence of a Hopf bifurcation, where a stable
limit cycle develops around the equilibrium point, which has then lost its stability.
The point at which this transition occurs is known as the Hopf bifurcation point,
which occurs where the pair of eigenvalues are pure imaginary with zero real part.
Therefore, in order to find where this bifurcation takes place, we numerically search
for cases where the eigenvalues are pure imaginary. We use a fourth-order finite
difference approximation of the Jacobian evaluated at the equilibrium point, and
evaluate the eigenvalues of the resulting matrix. These relationships are shown in
Figs. 3a–c.
If we follow the eigenstructure of the primary equilibrium point as Ro increases
(as shown in Fig. 3a), we notice that the equilibrium passes through several stability
(a) (b)
0.15
0.04
0.1
0.02
0.05
0
0
−0.02
λ
λ
−0.04 −0.05
−0.06 Real λ1
−0.1 Real λ1
Imag λ1 Imag λ1
−0.08 Real λ2 −0.15 Real λ2
Imag λ2 Imag λ2
−0.1 −0.2
0 5 10 15 20 25 1 2 3 4 5 6
Ro (normalized) kE
(c) (d) 1
0.25
Real λ1
0.2 Imag λ1 0
Real λ2
Imag λ2
0.15 −1
Real λ1
0.1 Imag λ1
−2
λ
Real λ2
λ
0.05 Imag λ2
−3
0
−4
−0.05
−0.1 −5
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
G (normalized) G (normalized)
Fig. 3 Eigenvalue plots of equilibrium solution branches, showing changes in stability with
parameter variations. The Hopf bifurcation points are marked with stars. (a) Primary equilibrium:
versus Ro . Hopf bifurcation occurs at Ro 9:793, and reverse Hopf bifurcation occurs at
Ro 17:430. (b) Primary equilibrium: versus kE (with increased Ro D 8) Hopf bifurcation
occurs at kE 2:615. (c) Primary equilibrium: versus G (with increased Ro D 8 and kE D 2:1).
Hopf bifurcation occurs at G 2:072. (d) Secondary equilibrium: versus G. The imaginary
parts are not visible in the plot because they are nearly coincident with the real part of 1 , which
takes very small positive values
classifications. The equilibrium begins as a stable node when Ro is unchanged.

As Ro increases, the eigenvalues transition from negative and real, to a complex
conjugate pair with negative real part, changing the equilibrium to a stable focus.
Increase Ro further, and we reach the Hopf bifurcation point, where at this instant,
the eigenvalues are a pure imaginary complex conjugate pair. The Hopf bifurcation
also marks the creation of a limit cycle around the equilibrium point. After the
bifurcation point, the eigenvalues are still a complex conjugate pair, but this time
they have a positive real part, which means the equilibrium is an unstable node
in this region. A further increase in Ro causes the reverse to take place, taking
the equilibrium back through a reverse Hopf bifurcation, extinguishing the limit
cycle, and returning it to a stable focus, and eventually a stable node. It is in the
region between the forward and reverse Hopf bifurcation points that we observe a
limit cycle in the system. A slightly similar situation occurs when the bifurcation
parameters are kE and G (Fig. 3b, c). While increasing these bifurcation parameters
produces the forward Hopf bifurcation in the same way, further increases do not
cause a reverse Hopf bifurcation. Instead, the eigenvalues of the unstable focus meet
and become a pair of positive real eigenvalues, indicating the equilibrium’s change
to an unstable node (Fig. 4).
For the sake of completeness, we have also analyzed the stability of the saddle
point equilibrium. However, we have not observed any changes in the stability of
this equilibrium point. Increasing Ro and kE do not change the eigenvalues at all,
and while increasing G changes the values of the eigenvalues, it preserves the overall
structure of two real eigenvalues with differing signs (see Fig. 3d).
(a) 90 Equilibria
(b) 110
Equilibria
Trajectories Trajectories
80 100
90
ic
ic
70
80
60
70
50 60
40 50
40
30
30
UN
20
20
10 10
0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
Arterial Compliance: Ca Arterial Compliance: C
a
Fig. 4 Phase portraits exhibiting limit cycles, where the primary equilibrium is an unstable node
(UN), and the secondary equilibrium is a saddle point (SP). (a) Normalized kE D 2:1, Ro D 10,
and G D 4. (b) Normalized kE D 8 and Ro D 8
232 D. Evans et al.
5 Conclusions and Further Work
Through the use of numerical simulations, we have shown how varying the
parameters G, kE , and Ro in the Ursino-Lodi model changes the value of the
equilibrium intracranial pressure and arterial compliance. We have also shown how
varying these parameters can force the primary equilibrium to undergo a Hopf
bifurcation, creating a stable limit cycle around the now unstable equilibrium point.
In addition to confirming the claims made by Ursino and Lodi, we have also shown
the existence of a reverse Hopf bifurcation at very high levels of Ro To the authors’
knowledge, no interesting bifurcation phenomena is witnessed by varying the model
parameters not mentioned here, hence their exclusion from the work. We intend to
use the work presented here when developing new models of intracranial pressure
dynamics, to offer additional insight on the stability structure of the steady-state
solutions.
References
1. Cutler, R.W.P., Page, L., Galicich, J., Watters, G.V.: Formation and absorption of cerebrospinal
fluid in man. Brain 91(4), 707–720 (1968)
2. Goldsmith, W.: The state of head injury biomechanics: past, present, and future: part 1. Crit.
Rev. Biomed. Eng. 29(5&6), 441–600 (2001)
3. Kyriacou, S.K., Mohamed, A., Miller, K., Neff, S.: Brain mechanics for neurosurgery: modeling
issues. Biomech. Model. Mechanobiol. 1(2), 151–164 (2002)
4. Marmarou, A., Shulman, K., Rosende, R.M.: A nonlinear analysis of the cerebrospinal fluid
system and intracranial pressure dynamics. J. Neurosurg. 48(3), 332–344 (1978)
5. Suarez, J.I.: Critical Care Neurology and Neurosurgery. Springer, New York (2004)
6. Ursino, M., Lodi, C.A.: A simple mathematical model of the interaction between intracranial
pressure and cerebral hemodynamics. J. Appl. Physiol. 82(4), 1256–1269 (1997)
7. Verhulst, F.: Nonlinear Differential Equations and Dynamical Systems, 2nd edn. Springer,
New York (1996)
8. Wakeland, W., Goldstein, B.: A review of physiological simulation models of intracranial
pressure dynamics. Comput. Biol. Med. 38(9), 1024–1041 (2008)
Persistent Homology for Analyzing
Environmental Lake Monitoring Data
Benjamin A. Fraser, Mark P. Wachowiak, and Renata Wachowiak-Smolíková
Abstract Topological data analysis (TDA) is a new method for analyzing large,
high-dimensional, heterogeneous, and noisy data that are characteristic of modern
scientific and engineering applications. One major tool in TDA is persistent homol-
ogy, wherein a filtration of a simplicial complex is generated from point clouds
and subsequently analyzed for topological features. Betti numbers are computed
across varying spatial resolutions, based on a proximity parameter R, where the n-th
Betti number equals the rank of the n-th homology group. In this paper, persistent
homology is applied to lake environmental monitoring data collected from a sonde
sensor attached to a commercial cruise vessel, and to weather station observations.
A modified form of the witness complex described by de Silva is used in an attempt
to eliminate the need for persistence and thus to reduce computation time. From
preliminary results, witness complexes are very promising in capturing the shape
of the data and for detecting patterns. It is therefore proposed that TDA, combined
with standard statistical techniques and interactive visualizations, enable insights
into observations collected from environmental monitoring sensors.
1 Introduction
Improvements in sensor network technology and sensor mechanisms have facilitated

acquisition of vasts amount of data. In the case of biomedical and environmental
monitoring, these data tend to be complex, heterogeneous, noisy, unstructured, and
high-dimensional [3]. Furthermore, data are frequently acquired as long vectors,
where only a small number of coordinates are meaningful to specific questions
at hand, but it is often not known a priori which coordinates are relevant [1].
Discovering meaningful structural or temporal patterns in these data – that is, to
obtain knowledge about the data’s large-scale organization – is very difficult, even
for domain experts. New, innovative computational methods and visual data mining
tools are therefore needed [3]. Among recently proposed techniques, topological
data analysis (TDA) is very promising. TDA provides robust feature definitions,
B.A. Fraser () • M.P. Wachowiak • R. Wachowiak-Smolíková

Nipissing University, North Bay, ON, Canada

234 B.A. Fraser et al.
emphasizing the global aspects of the data and shape recognition in point clouds
of high dimensions. Specifically, homology provides the basis for computing
simplicial complexes (a mesh that represents a space in specific ways), in which
the topology and geometry of the point clouds are separated [10]. Topological
methods for data analysis were described in the seminal paper by Carlsson [1] and
subsequently expanded. They are useful in studying connectivity information, such
as the classification of loops and high-dimensional surfaces that may implicitly exist
in the data. Unlike geometric methods, TDA is not sensitive to properties such as
curvature, and, because the focus is on the global structure of the data, numerical
values of distance functions are less important. These features allow researchers to
study complex data where there is an unclear or incomplete idea of what specific
metrics are to be extracted [1]. TDA has been applied to a variety of scientific
problems, most notably in biomedicine, including the organization of genomic data
from many different sources in the study of breast cancer [6], and for computer-
aided diagnosis of pulmonary embolism [7]. TDA is also used for business and
social analysis [4].
Analytics and advanced computational methods are increasingly important in
environmental data analysis [2, 5]. New institutes and research initiatives focus
exclusively on this topic (e.g. the Institute for Environmental Analytics at the
University of Reading – www.the-iea.org). In the current paper, we adapt TDA
to the analysis of environmental sensor data. Specifically, we modify the witness
complex filtration described by de Silva [8] to eliminate the need for persistence,
and consequently, to reduce computation time. We analyze a large number of data
vectors in R6 representing various properties (e.g. time, temperature, air pressure,
etc.) collected from sensors attached to a commercial cruise vessel, with the
goal of obtaining insights into the complex physical phenomena characterizing
dynamic lake conditions. In another example, data from weather station sensors
are analyzed. After normalization of a Euclidean distance matrix for these vectors,
a mesh or triangulation is constructed which accurately represents the topology
of this data set. Although the data cannot be characterized as “big data” (very
large and heterogeneous data sets that generally cannot be processed or analyzed
using standard computational or statistical techniques), the readings are sufficiently
numerous and high-dimensional that constructing such a mesh on the full set of
points would be prohibitively costly. We must therefore choose a subset that is
representative of the full set and retains its topology. The resulting analyses are
expected to complement visual and statistical methods to enrich insights and to
facilitate discovery of various lake phenomena. The next section describes the TDA
algorithm as well as our adaptations, followed by a preliminary investigation of
applying TDA to environmental lake monitoring data.
Persistent Homology for Analyzing Environmental Lake Monitoring Data 235
2 Algorithm
We construct a simplicial complex on a representative subset (landmarks) of the full

data set using the remaining points as witnesses to decide whether edges will be
added between them or not. A different choice of filtration of the witness complex
is used than that described by de Silva [8], and the two approaches are compared
for their reliance on persistent versus simplicial homology to accurately recover the
Betti number profile of the data.
2.1 Simplicial Complexes
A k-simplex is a complete graph on k 1 vertices. Every clique that is a subgraph

of a simplex is a face of that simplex.
Definition 2.1 A simplicial complex is a collection K of simplices such that:
• Every face of a simplex in K is in K
• The intersection of any two simplices in K is a face of each of them
In particular, we are interested in the Vietoris-Rips complex [11], the largest
simplicial complex on a given set of vertices and edges.
Definition 2.2 Given a vertex set V, the Vietoris-Rips complex VR.V; R/ contains
the p-simplex Œa0 ; : : : ; ap iff for every edge Œej ; ek , 0 j k p, jej ek j R
Constructing a VR complex for given V and R then amounts to inserting edges
between vertices that are within R of each other, and subsequently determining all
cliques in the graph, and including them as simplices.
2.2 Landmarks and Witnesses
Given a set X of N data points, we need to choose a subset of those points consisting
of n landmarks that accurately represents the shape of the point cloud. More
complex shapes may require more landmarks for accurate representation, whereas a
circle could be represented by only four points. Since the remaining N n points will
act as witnesses, it is important not to select too large a proportion of the original N.
Ideally, N should be quite large, so that a large number of landmarks can be selected.
In practice, the ratio N=n should range from 30 to 100, depending on the size of the
original data.
Fig. 1 The red, green and

blue points are weak
witnesses to the triangles of
their respective colour, and to
the edges of the inner
triangle, but there is no
witness to the inner triangle
itself
The minmax [8] method of landmark selection is employed, which attempts to

choose landmarks as evenly spaced as possible throughout the data. We proceed as
follows:
• Choose a random point to be first in a set of landmarks L
• Add each subsequent point x 2 X to L to be such that minfdist.x; p/ 8 p 2 Lg is
maximized
Let the points in P D X n L be the set of potential witness points.
Definition 2.3 ([8]) A point w 2 P is a weak witness to a p-simplex A D
.a0 ; a1 ; : : : ap / in L if jw aj < jw bj 8 a 2 A and b 2 L n A. It is a strong
witness if jw a0 j D jw a1 j D : : : D jw ap j as well.
Example 2.1 Take three vertices of an equilateral triangle, as well as the three
midpoints of its edges, as landmarks. Consider the three witnesses which lie just
outside the central triangle formed by the midpoints (Fig. 1). These are weak
witnesses to its edges, but none of them is a weak witness to the triangle itself;
they are weak witnesses to the outer triangles corresponding to their colour.
We may wish to include a simplex, such as the triangle in the preceding example,
in the case where all of its subsimplices are included. This motivates the following
definition of a witness complex that will be employed, where witnesses are used
to determine the 1-skeleton, or neighbourhood graph of edges, and then the higher
dimensional simplices will be added by Vietoris-Rips expansion [11].
2.3 Witness Complexes
Two witness complex filtrations are described below. The first is from De Silva
[8], the second is the proposed adaptation of this construction. A filtration of a
witness complex is a nested sequence of increasing subsets of a simplicial complex.
There are three inputs: a parameter v that determines how many landmarks can
be witnessed simultaneously by a single point, the distance matrices between data
points, and a value R which ranges across an interval and creates the filtration.
Take D to be the n N Euclidean distance matrix (on the n landmarks and N full
data set). Take E to be the n n distance matrix between landmarks, and E0 to be
the n jPj distance matrix (between the n landmarks and jPj potential witnesses).
Both witness complexes have the n landmarks as their 0-simplices.
For W.DI R; v/ (De Silva, [8]), R 2 Œ0; 1/, and for W .E; E0 I R; v/, R 2
Œ0; 1:
• if v D 0, then for i D 1; 2; : : : ; N define mi D 0 and for i D 1; 2; : : : ; jPj
define ni D 0
• if v > 0, then for i D 1; 2; : : : ; N define mi to be the v-th smallest entry
of the i-th column of D, and for i D 1; 2; : : : ; jPj define ni to be the v-th
smallest entry of the i-th column of E0
– the edge D Œab belongs to W.DI R; v/ iff there exists a witness i 2
f1; 2; : : : ; Ng such that max.D.a; i/; D.b; i// R C mi
– the edge D Œab belongs to W .E; E0 I R; v/ iff there exists a witness
i 2 f1; 2; : : : ; jPjg such that max.E0 .a; i/; E0 .b; i// ni and E.a; b/ R
• the p simplex D Œa0 a1 : : : ap belongs to the witness complex iff all of its
edges also belong to the witness complex
Example 2.2 The reason for splitting the distance matrix D into E and E0 in our
construction is that landmarks should not act as witnesses. Consider a point cloud
where nearly all the vectors are tightly clustered, and there are a few widely spaced
outliers (Fig. 2). All of these widely spaced outliers are likely to be chosen as
landmarks by minmax, and if they can act as witnesses to each other, they can all be
connected to each other by edges, resulting in a complex which does not accurately
represent the data’s shape. But if landmarks and witnesses are mutually exclusive,
then the problem of outlier selection will be mitigated (albeit, each outlier selected
as a landmark will still affect b0 ).
Fig. 2 An example of a
witness complex W.DI 0; 2/
where outliers were selected
as landmarks (red), and then
acted as witnesses to each
other. Features
unrepresentative of the data
are found
If landmarks are not taken as witnesses, observe the following equality:

W.DI 0; v/ D W .E; E0 I 1; v/. Further, W.DI R; v/ is the same as the Vietoris-
Rips complex for v D 0, and W .E; E0 I R; v/ is the same for v n. Below
we explore these relations, and why appropriately chosen values of v can make
persistent homology unnecessary.
At the point in the filtrations where W.DI 0; v/ D W .E; E0 I 1; v/, each witness
point is the center of a smallest possible closed ball which contains at least v
landmarks. Any landmarks that all fall within one such ball are joined by edges.
This is the intermediate stage, say, W , between completely disconnected vertices
and a fully connected graph in which we are interested.
• W.DI R; v/; R 2 Œ0; 1/ describes the sequence of simplicial complexes from W
up to the complete graph on n vertices. It does this by expanding the closed balls
around each witness point until all of the landmarks are contained in one.
• W .E; E0 I R; v/; R 2 Œ0; 1 describes the sequence of simplicial complexes from
the set of disconnected 0-simplices to W . It does this by expanding closed balls
centered on the landmarks as an additional condition for connection by an edge.
The rationale behind our definition of W .E; E0 I R; v/ is that we find joining
landmarks to each other by considering their own distances from each other rather
than from a witness to be a more natural progression, and that we consider W to
be a suitable stopping point, where the simplicial complex accurately represents the
shape of the data. (In fact, as will be evident from the examples that follow, there
appears to be no need to compute simplicial complexes that are more connected
than W .)
Example 2.3 One reason that persistence is necessary for W.DI R; v/ is that only v
values up to v D 2 are considered. Consider the following example: the four corners
of a square centered on the origin are chosen as landmark points, and infinitely
many witness points are taken uniformly from the interior of the square, excluding
the origin (Fig. 3). Then at R D 0, W.D; R; 2/ is nearly correct, but there exists no
witness which has opposite corners across the diagonal as its two nearest landmarks:
Fig. 3 The 1-skeleton

resulting from W.D; 0; 2/,
with the corners of a square
as landmarks. Without a
witness precisely at the
origin, a 2-dimensional hole
is detected. At R D 0:1, the
hole is filled
as such, they are not joined by edges, and the Betti number b1 is found to be 1 rather
than 0, as it should be. As such, it is necessary to increment R by at least a small
amount to fill the hole.
This issue may have been avoided by choosing v D 3. In fact, with data of
dimension d, choosing v D d C 1 generally suffices. Witness complexes are used to
reduce the number of cells: in 3 dimensions, a clique on 10 vertices is superfluous.
However, a clique on 4 vertices is not. We can recover an accurate Betti profile of
our data by increasing v slightly and still keep the number of cells relatively low.
And by eliminating the need for persistence (since homology need be computed
only once), more landmarks can be selected.
Finally, it remains to be explained why the parameter R should be used to create
a filtration at all, if persistence is not used. The answer is that outliers can still be a
problem, and W may not actually be correct.
Example 2.4 Take as our data set a large number of points from 23 of a circle,
and two additional points which are very near to each other from the center of
the missing arc – these are the outliers. Minmax selects one of these outliers as
a landmark, as well as some evenly distributed ones along the 23 arc. An ideal
simplicial complex should only connect the landmarks along the 23 arc, but with
v D 3, the other outlier which was not taken as a landmark witnesses the two ends
of that arc and connects them, completing the circle (Fig. 4).
Clearly, this problem could have been avoided by taking the finished complex to
be at some R < 1, as the edges in the 23 arc would have been formed by R 0:1,
but the green edges needed R 0:3. In general, the more landmarks that are taken,
the closer they will be to each other, and the complex should be complete at a lower
value of R. Any edges which emerge at higher values are more likely to be artifacts
of outliers being witnessed, as in the above example. Therefore, we propose that
the filtration be calculated normally, but when the rate of edge addition (relative to
Fig. 4 For W .E; E0 I 1; 3/,

the black point witnesses the
green edges, which connect
an outlier point and bridge a
chasm in the data, which
should not have been done
the increase in R) begins to fall off dramatically, that the complex be considered
complete and its Betti profile found.
3 Application
The proposed methods are applied to three types of data. The first is an artificially
generated sphere point cloud, to mirror the experiments in [8]. The second is lake
monitoring data, and the last are vectors of various observed weather properties.
3.1 Sphere Data
The first artificial data set consists of N D 3200 points chosen uniformly from
the surface of S2 . Using minmax selection, 64 landmarks were chosen. Both
W .E; E0 I 1; 3/ and W .E; E0 I 1; 4/ (Fig. 5) recovered the correct Betti profile
.b0 ; b1 ; b2 / D .1; 0; 1/ in each of 100 trials.
Then 12,500 points were chosen uniformly from within S2 having four interior
spherical voids, ranging in diameter from 0.2 to 0.5. Again, using minmax, 125
landmarks were chosen. W .E; E0 I 1; 3/ (Fig. 5) recovered the correct Betti profile
.b0 ; b1 ; b2 / D .1; 0; 4/ in each of 10 trials.
Fig. 5 Simplicial complexes constructed on points taken from S2 (left) and D2 with 4 spherical
voids in it (right)
3.2 Lake Monitoring Data
The TDA approach developed in this paper was applied to lake monitoring
data to complement existing statistical and visualization analysis tools [9] (see
visual.nipissingu.ca/Commanda2). Data were collected from Lake Nipissing (46ı
160 1200 N 79ı 470 2400 W) via a sonde sensor attached to a commercial cruise
vessel (the Chief Commanda II). The vectors are in six dimensions: temperature
(ı C), specific conductivity ( s/cm), dissolved oxygen concentration (mg/L), pH,
chlorophyll RFU (relative fluorescence units), and total algae RFU. A 3-dimensional
witness complex W .E; E0 I R; 4/ was constructed on a subseet of 36 landmarks of
216 data vectors from Sept. 4, 2011.
Betti profiles of .5; 1; 0/ or .5; 0; 1/ were found, but are considered to be
unreliable due to the random nature of landmark selection. Removing various
dimensions can simplify locating the source of the feature by constructing the same
complex on the same data set, minus one of its coordinates. For instance, eliminating
chlorophyll RFU made the detection of these features rarer, so we expect it is a
contributing factor, while eliminating total algae RFU resulted in these features
being detected more frequently. Consequently, the total algae property appears to
be irrelevant, as it interferes with the topological feature.
In this case, reducing the coordinates to specific conductivity, pH, and chloro-
phyll RFU resulted in reliably reproducible 3-dimensional holes appearing in the
witness complex W .E; E0 I 1; 4/ (Fig. 6a). Further research is needed to understand
these results, and to determine whether these features indicate some interaction
between these properties of the lake.
Fig. 6 (a) Reliably detected 3-dim hole on dimension-reduced lake data; (b) b1 increasing by 2 at
R D 0:76 in weather data, a result of outliers
3.3 Weather Data
TDA was applied to data collected from a Nipissing University weather sta-
tion to complement existing web-based analysis and interactive visualizations
(geovisage.nipissingu.ca). Six-dimensional weather data (photosynthetically active
radiation (PAR) (mol/J), rainfall (mm), relative humidity (%), soil moisture (%),
soil temperature (ı C), and wind speed (m/s)) for Temiskaming Shores, Ontario
(approximately 47ı 310 N 79ı 410 W) were sampled at 5-min intervals, resulting
in 8640 vectors for the month of April, 2013. A 2-complex W .E; E0 I 1; 3/ was
constructed on a subset of 86 landmarks, and during every such trial there were
edges added for large values of R, some even at R > 0:9. These edge additions
resulted in an increase of the found value of b1 . This is an instance of outliers
affecting the result (Fig. 6b), as the majority of the complex was already complete
by R D 0:35, and the Betti profile found at that point is considered more accurate.
Another strategy for isolating features is to build the complex on different time
scales. For instance, a Betti profile of .b0 ; b1 / D .1; 5/ was most often found for the
month of April. It may be helpful to perform the same analysis on individual weeks,
or days, to determine whether the appearance of certain features can be related to
weather events, such as a thunderstorm, at those particular times.
4 Conclusions
This paper presents an improvement to the persistent homology paradigm for TDA
that adapts and builds on the work of de Silva. By employing witness complexes,
and with less dependence upon persistence, we are better able to represent the shape
of complex data, with improved computational efficiency. Based on its success with
artificial data sets and preliminary results from real environmental data, we propose
that simplicial homology on this modified version of a witness complex is effective

in approximating the shape of complex data, and is potentially useful as an adjunct
tool for identifying patterns in these data. Further investigation with methods such
as dimension reduction, and determining the parts of the data that give rise to a
topological feature would aid in understanding how properties interact. Future work
will concentrate on improving algorithm efficiency, and, as described above, in more
detailed temporal analysis, and in temporal and spatial comparisons of the features
obtained from TDA to determine how these features reflect physical phenomena.
Finally, TDA and traditional computational and statistical techniques, as well as
interactive visualizations, will be integrated to gain a more complete picture of
dynamic environmental conditions.
Acknowledgements M. P. Wachowiak is supported by NSERC Grant #386586-2011.
References
1. Carlsson, G.: Topology and data. Bull. Am. Math. Soc. 46, 255–308 (2009)
2. Dutta, R., Li, C., Smith, D., Das, A., Aryal, J.: Big data architecture for environmental
analytics. In: Denzer, R., Argent, R.M., Schimak, G., Hr̆ebíc̆ek, J. (eds.) Environmental
Software Systems. Infrastructures, Services and Applications, vol. 448, pp. 578–588. Springer
International Publishing, Cham (2015)
3. Holzinger, A.: Extravaganza tutorial on hot ideas for interactive knowledge discovery and data
mining in biomedical informatics. In: Ślȩzak, D., Tan, A.-H., Peters, J.F., Schwabe, L. (eds.)
Brain Informatics and Health. LNCS, vol. 8609, pp. 502–515. Springer, Heidelberg (2014)
4. Lum, P.Y., Singh, G., Lehman, A., Ishkanov, T., Vejdemo-Johansson, M., Alagappan, M.,
Carlsson, J., Carlsson, G.: Extracting insights from the shape of complex data using topology.
Nature 3 (2013). doi:10.1038/srep01236
5. Namieśnik, J., Wardencki, W.: Monitoring and analytics of atmospheric air pollution. Pol. J.
Environ. Stud. 11(3), 211–216 (2002)
6. Nicolau, M., Levine, A.J., Carlsson, G.: Topology based data analysis identifies a subgroup of
breast cancers with a unique mutational profile and excellent survival. Proc. Natl. Acad. Sci.
108(17), 7265–7270 (2011)
7. Rucco, M., Falsetti, L., Herman, D., Petrossian, T., Merelli, E., Nitti, C., Salvi, A.: Using
Topological Data Analysis for diagnosis pulmonary embolism. http://arxiv.org/abs/1409.5020
[physics.med-ph]
8. de Silva, V., Carlsson, G.: Topological estimation using witness complexes. In: Proceedings
of the Symposium on Point-Based Graphics, pp. 157–166. Eurographics Association, Aire-la-
Ville (2004)
9. Wachowiak, M.P., Wachowiak-Smolikova, R., Dobbs, B.T., Abbott, J., Walters, D.: Interactive
web-based visualization for lake monitoring in community-based participatory research.
Environ. Pollut. 4(2), 42–54 (2015)
10. Zomorodian, A.: Computational topology. In: Algorithms and Theory of Computation Hand-
book. Applied Algorithms and Data Structures Series, p. 3. Chapman & Hall/CRC, Boca Raton
(2010)
11. Zomorodian, A.: Fast construction of the Vietoris-Rips complex. Comput. Graph. 34, 263–271
(2010)
Estimating Escherichia coli Contamination
Spread in Ground Beef Production Using
a Discrete Probability Model
Petko M. Kitanov and Allan R. Willms
Abstract Human illness due to contamination of food by pathogenic strains of

Escherichia coli is a serious public health concern and can cause significant
economic losses in the food industry. Undercooked ground beef is the primary
means of transmission of pathogenic E. coli to humans. In the Western world, most
ground beef is produced in large facilities where many carcasses are butchered and
various pieces of them are ground together in large batches. Assuming that the
source of contamination is a single carcass, the primary determinant of how many
batches of ground beef from a particular production cycle are affected is the manner
in which pieces of that carcass are spread about in the raw sources that contribute to
the ground beef batches. Assuming that ground beef from a particular batch has been
identified at the consumer end as contaminated by E. coli, we model the probability
that previous and subsequent batches generated in the same production cycle are
also contaminated. This model may help the beef industry to identify the likelihood
of contamination in other batches and potentially save money by not needing to
cook or recall unaffected batches of ground beef.
1 Introduction
Pathogenic Escherichia coli strains, particularly O157:H7, cause illness in humans.

Scallan et al. [8] indicate that a significant part of all cases of acquired food-borne
illness in the U.S.A. is caused by the pathogenic strains of E. coli. These strains can
be transmitted to humans through various food products including produce, dairy,
and meat. The primary transmission of these strains from cattle to humans is by
consumption of beef, especially under-cooked ground beef.
P.M. Kitanov
University of Ottawa, Ottawa, ON, K1N 6N5, Canada,
A.R. Willms ()
University of Guelph, Guelph, ON, N1G 2W1, Canada,

246 P.M. Kitanov and A.R. Willms
In the western world, since ground beef production is concentrated in large

meat processing plants, an E. coli contamination event at one of these plants can
potentially affect many people over a wide area. Food safety regulations in North
America require the removal of all production and raw sources associated with
an identified contamination event. This leads to the recalling of large amounts of
beef, much of it likely uncontaminated, and consequently large economic losses for
the beef industry along with damage to reputation. This paper presents a discrete
probability model for estimation of the likelihood that sequential batches of ground
beef produced in a large plant are contaminated with pathogenic E. coli given that
one batch is contaminated.
There have been many studies regarding the transmission of pathogenic E.
coli and its impact on human health. For example, risk assessment for E. coli in
ground beef and burgers in different countries has been quantified [4–7, 9]. A risk
assessment model for E. coli O157:H7 in ground beef and beef cuts in Canada is
presented in [10]. One study, [11], used data from a large E.coli outbreak in the
U.S.A. to identify possible sources of contamination. The differing conditions in
production and meat processing plants, distribution networks, and cooking methods,
etc., make risk assessment of contamination a very difficult task. The primary
method used in the above studies is statistical analysis of empirical data. Stochastic
models for outbreak and transmission of E. coli O157:H7 infection in cattle, are
presented in [12] and [15].
Aslam et al. [1, 2] and Bell [3] have studied the sources of contamination by
pathogenic E. coli in a beef-packing plant. E. coli in meat products, originates
mainly from the hides of the incoming animals and is transferred to the trimmings,
and subsequently the ground beef, during the dressing of carcasses and carcass
breaking. Various decontamination treatments, such as antimicrobial solutions and
pasteurization, have been implemented at some beef processing plants. Although it
was found in [14] that these treatments significantly reduce the risk of contamina-
tion, the hazard can not be completely removed. Also, there are still no effective
methods for quickly screening large amounts of ground beef in big production
facilities. All of these uncertainties justify using probabilistic methods for control
and estimation of E. coli contamination in the production of ground beef.
2 The Model
Consider a large ground beef production facility. The ground beef is produced in
batches of several tonnes each. The input to these batches are several raw sources,
typically a “lean” fresh source and a “fat” fresh source, but also often a frozen source
and other sources such as BLBT (Boneless Lean Beef Trimmings). The batch is
well-mixed, so that if any contamination is present on any of the raw source material
that is input into the batch, the entire batch is deemed to be contaminated.
Estimating E. coli Spread in Beef Production 247
2.1 Basic Assumptions
The primary assumption is that the contamination is due to a single carcass,

often associated with the fat layer on that carcass. Spread of the contaminant in
the production process is assumed to be due to division and dispersion of the
contaminated carcass portions; transfer via physical contact with other pieces and
machinery surfaces is assumed to be negligible. Further, due to the temperature
at which production occurs, it is assumed that the contaminant does not grow
appreciably.
The manner in which a carcass is spread over some region in a raw source is
highly dependent on the production process—the way in which trim is added to the
raw source bins, etc. We make a number of simplifying assumptions:
• for each raw source, the number and mass of all pieces from all carcasses is the
same, and the manner of spread of the pieces from each carcass throughout the
raw source is the same
• material in a raw source bin is not mixed, or what mixing occurs is captured by
the carcass distribution function,
• the material in the raw source bins can be ordered and is used as input to the
ground beef production batches in that order
This last assumption allows us to define a “mass location” in each raw source; mass
from a particular raw source used in batch number b comes from locations just
prior to those for the mass from the same source used in batch number b C 1. This
location is a discrete variable increasing in increments of the mass of a piece, which
by the first assumption is always the same in a given raw source. Although these
assumptions are all invalid to some degree, they are useful for making a tractable
model that we believe to have relevancy to the problem. Further, replacing them
with more realistic assumptions would require considerably more information on
the production process.
Suppose there are B batches of ground beef produced in a production cycle and
S raw sources used as input to these batches, not all of which need be used in
any particular batch. It is assumed that a particular batch of ground beef has been
identified as being contaminated. This batch is referred to as the “hot” batch, and it
is identified as batch number h. The origin of this contamination is due to a single
hot raw source and in turn, to a single hot carcass in this raw source. This model
computes probabilities of contamination due to the same origin for all other batches
in the production cycle.
2.2 Carcass Spread in a Raw Source
Let ps be the number of pieces, each of mass as contributed by each carcass in raw
source s. Let Cs be the total number of carcasses in the raw source, then the total
number of pieces in the raw source is Ns D Cs ps . Since we are assuming the raw
source material is ordered, we assign piece position numbers 1 to Ns for this source.
The total mass in the raw source is Ms D Cs ps as .
The pieces from each carcass are distributed throughout this raw source in some
manner, the same for each carcass. Here we consider a piece-wise linear distribution
function that is even around its centre:
8 C
ˆ .jnjLi1 /Hi C.Li jnj/Hi1
ˆ
< Li Li1 if Li1 < jnj < Li , 1 i K;
C
F.n/ D H CH (1)
ˆ 2
i i
if jnj D Li , 0 1 K;
:̂
0 if jnj > LK :
This function’s domain has 2K pieces, distributed symmetrically around zero,

divided by the constants 0 D L0 < L1 < L2 < < LK , where each Li , 1 i K,
is a positive multiple of ps , and Hi˙ 0, 0 i K, are the limiting values of F at
Li from the left () and right (C), respectively. Necessarily, H0 D H0C , HKC D 0,
and
X
K
C
.Li Li1 /.Hi C Hi1 / D 1:
iD1
This class of functions includes the uniform distribution, obtained with K D 1, and
H0˙ D H1 D 2L1 1 . If the parameters are chosen such that Hi D HiC , 1 i K, then
the function F is continuous. Each raw source s may have a different distribution
function Fs , and thus a different set of parameters K, L, and H. We assume that the
distribution centres, c of each carcass are uniformly spread through the source:

1 Ns Ms
c D c ps ; c 2 Z; 1 c Cs D D : (2)
2 ps ps as
Here c is the carcass number and the discrete probability density function for carcass
c in this raw source is Fs .n c /. The expected fraction of a carcass c present in
any set of piece positions R f1; 2; : : : ; Ns g is
X
Qsc .R/ D Fs .n c /: (3)
n2R
Equivalently, Fs .n c / may be interpreted as the density function for a particular

piece of this carcass, and Qsc .R/ as the probability that this piece is located in the
set R. The piece-wise linear nature of this density function allows it to satisfy the
requirement that the fraction of all carcasses at a particular piece position must add
to p1s so that the expected number of pieces at that location will be one.
2.3 Batch Contamination Probability
Let msb be the mass input from source s to batch b, and let Msb be the mass input
from source s to batches prior to batch b (both assumed to be multiples of as ). Let nsb
and Nsb be the number of pieces from source s input to batch b and input to batches
prior to b, respectively, that is, msb D nsb as and Msb D Nsb as . The sequential piece
positions in raw source s that are input to batch b are then the integers in the interval
Bsb D ŒNsb C 1; Nsb C nsb :
For source s, denote the probability of carcass c being absent from the set Bsb as
Asc .Bsb /. Modelling the selection of pieces as being independent of the other pieces
that have already been selected from this carcass and other carcasses, this absence
probability is
Asc .Bsb / D .1 Qsc .Bsb //ps ; (4)
where Qsc is defined by (3).

Consider two batches, h and j, which receive input from source s, and a particular
carcass with distribution centred at c . The probability that this carcass has at least
one piece that is input to batch h and at least one that is input to batch j is

Prob.c from s in h & j/ D 1 Asc .Bsh / C Asc .Bsj / Asc .Bsh [ Bsj / : (5)
The last term is present because it is included in both the previous terms but should
only be counted once.
Given that batch h is the hot batch and assuming the contamination is due to
a hot carcass in raw source s, then the probability that a particular carcass c from
source s is the hot carcass is a uniform probability depending on the total number of
carcasses, Csh from source s input to h:
1 ps
Prob.c from s is hot/ D : (6)
Csh nsh
The right hand expression in (6) is an upper bound on the probability since the
number of distinct carcasses present, Csh , will likely be more than nsh =ps , especially
if the spread of each carcass is large. A reasonable estimate for Csh might be to
include all carcasses c whose expected number of pieces in batch h is at least one,
that is Qsc .Bsh / 1=ps.
The contamination in batch h may be due to any of the raw sources that were
input for this batch and these raw sources may have varying degrees of relative
susceptibility to being contaminated. For example, the susceptibility factor, gs , for
raw source number s, might be near zero if the source is frozen material, and might
be one if the source is fresh material from the plant. Let fs be the fraction of fat in
raw source s. The probability that raw source s is the origin of the contamination in
batch h is the weighted fraction of the input mass, the weights being the product of
the susceptibility factors and the fat fractions:
gs fs msh
Prob.s is hot/ D PS : (7)
kD1 gk fk mkh
The probability that batch j is also contaminated given that batch h is contami-
nated is then found by summing over all raw sources:
Prob. j hot j h/ D
X
S X
Cs
Prob.s is hot/ Prob.c from s is hot/ Prob.c from s in h & j/: (8)
sD1 cD1
This expression can be evaluated using (4), (5), (6), and (7).
One aspect that has not been accounted for is the possibility that one carcass
may be present in more than one raw source. This could occur, for example, in a
production facility where one fresh lean bin is filled and removed from the trimming
line while a particular carcass has only partially been trimmed. In this case pieces
from one carcass will occur at the “top” of one raw source bin and the “bottom”
of the next, thus potentially spreading the contamination over a much wider area, if
this happens to be the hot carcass. The model could be modified to account for this
possibility if desired, but would require information about the order and means of
filling of the various raw sources.
3 Example
To illustrate the above model we present one example. These data are fictitious
but based on typical values one might encounter in a large ground beef production
facility. In this example, there are a total of S D 7 sources (three frozen lean, two
fresh lean, and two fresh fat), and a total of B D 14 one-tonne batches of ground
beef are produced. Relevant information for the sources is provided in Table 1. The
uniform distribution for F (K D 1, H0˙ D H1 D 1=.2L1/) was used for all raw
sources in this example. The mass from each source used in each batch is provided
in Table 2.
Using the above model, and letting each batch be the hot batch in turn,
probabilities of contamination for each batch were computed. The probability
that a particular carcass c in source s is the hot carcass, Prob.c from s is hot/,
was computed using the suggestion following (6), that is, counting the number
of carcasses in source s that are present in batch h as all those carcasses c with
Table 1 Model parameters for the raw sources

Source gs fs ps as (kg) L1s Cs Ms (kg)
I (frozen lean) 0.2 0.05 25 0.6 400 134 2000
II (frozen lean) 0.2 0.09 25 0.6 400 134 2000
III (frozen lean) 0.2 0.07 25 0.6 400 134 2000
IV (fresh lean) 0.8 0.10 20 0.7 350 179 2500
V (fresh lean) 0.8 0.08 20 0.7 350 215 3000
VI (fresh fat) 1.0 0.40 50 0.5 500 80 2000
VII (fresh fat) 1.0 0.45 50 0.5 500 80 2000
Table 2 Source input mass, msb , (kg) and total fat percentage for each batch
Source
Frozen lean Fresh lean Fresh fat
Batch I II III IV V VI VII Fat %
1 312 136 552 25
2 384 52 564 25
3 114 404 260 222 25
4 262 239 231 268 25
5 201 205 89 293 212 25
6 320 180 292 100 108 15
7 407 105 284 204 15
8 390 456 154 15
9 300 205 325 170 15
10 209 211 543 37 10
11 293 132 536 39 10
12 318 94 540 48 10
13 479 454 67 10
14 701 226 73 10
Qsc .Bsh / 1=ps. Complete results are given in Table 3 and a selection of these are
plotted in Fig. 1.
The likelihood of other batches being contaminated is highly dependent on the
source input. In the example, if one of the early batches is the hot batch, then
the likelihood of contamination is only nonzero for batches near the hot batch.
Conversely, if one of the last batches is the hot batch, then a larger number of
other batches have a nonzero probability of being contaminated. This is due to
the source configuration shown in Tables 1 and 2. Since the contamination is most
likely to be carried in the fatty sources (VI and VII) followed by the lean fresh
source (IV and V), the distribution of these sources across the batches is a primary
contributor to the contamination probability distribution. The other factor we found
to be very important was the values of L1s . If these spread indicators were small,
then the contamination was much more confined to nearby batches to the hot batch.
Table 3 Probability of contamination for each batch in percent. Each column corresponds to a
different hot batch
Hot batch
Batch 1 2 3 4 5 6 7 8 9 10 11 12 13 14
1 99 47 6 0 0 0 0 0 0 0 0 0 0 0
2 62 99 34 13 0 0 0 0 0 0 0 0 0 0
3 5 46 100 65 35 3 0 0 0 0 0 0 0 0
4 0 21 79 100 72 47 19 0 0 0 0 0 0 0
5 0 0 38 66 100 81 47 18 2 0 0 0 0 0
6 0 0 3 34 63 98 69 34 12 0 0 0 0 0
7 0 0 0 12 24 48 100 66 32 6 6 5 2 0
8 0 0 0 0 11 24 66 100 60 24 14 15 15 13
9 0 0 0 0 1 8 37 66 100 63 34 26 29 30
10 0 0 0 0 0 0 17 39 72 100 66 30 32 34
11 0 0 0 0 0 0 13 28 49 60 100 62 35 38
12 0 0 0 0 0 0 8 23 39 20 50 100 64 43
13 0 0 0 0 0 0 3 18 34 17 21 57 100 74
14 0 0 0 0 0 0 0 12 28 14 18 27 67 100
100
80
Probability %
60
40
20
0
2 5 8 11 14
Batch
Fig. 1 Probability of contamination for each batch given that a fixed batch is contaminated (hot).
The five separate curves correspond to batches 2,5,8,11, and 14 being the hot batch
The overall result then, is the obvious observation that if one wishes to restrict
contamination, then one should restrict the spread of each carcass within the raw
source, and restrict the spread of the raw source across batches. In other words, try
to make it so that one carcass is only present in one batch.
4 Conclusion
We believe that the proposed model may help to reduce economic losses in the beef
industry. To our knowledge, this is a novel method for estimating of the probability
of E. coli contamination in the production of ground beef. Due to the necessarily
cautious food safety regulations, in order for this model to be adopted by regulatory
agencies, it would need to be further refined to account for plant-specific processes,
and its predictions thoroughly tested.
Most of the input to the model is easily obtained from production records, or
easily estimated (such as the average size of pieces). The most difficult input values
to estimate are the spread parameters, K, L, and H ˙ , however, some knowledge
of these values could be obtained either through detailed observations of the plant
processes or by using genetic sampling experiments in the raw sources as well as
the final ground beef product. A useful next step would be to use some type of
innocuous surrogate for E. coli that could be applied to pieces from a particular
carcass and then identified in the final product batches. This would allow for better
determination of carcass spread within a raw source and would provide data for
direct validation of the model.
Further development of the model, allowing for each carcass to have different
size pieces, and allowing for carcasses to be present in more than one raw source, is
currently being completed [13]. This extended work also incorporates genetic typing
data from a production facility to help inform the choice of carcass distribution
function.
Acknowledgements The second author was supported by a grant from the Applied Livestock
Genomics Program of Genome Alberta.
References
1. Aslam, M., Greer, G.G., Nattress, F.M., Gill, C.O., McMullen, L.M.: Genotypic analysis of
Escherichia coli recovered from product and equipment at a beef-packing plant. J. Appl.
Microbiol. 97, 78–86 (2004)
2. Aslam, M., Nattress, F.M., Greer, G.G., Yost, C., Gill, C.O., McMullen, L.: Origin of
contamination and genetic diversity of Eschericia coli in beef cattle. Appl. Environ. Microbiol.
69(5), 2794–2799 (2003)
3. Bell, R.G.: Distribution and sources of microbial contamination on beef carcasses. J. Appl.
Microbiol. 82, 292–300 (1997)
4. Cassin, M.H., Lammerding, A.M., Todd, E.C.D., Ross, W., McColl, R.S.: Quantitative risk
assessment for Escherichia coli O157:H7 in ground beef hamburgers. Int. J. Food Microbiol.
41(1), 21–44 (1998)
5. Duffy, G., Butler, F., Cummins, E., O’Brien, S., Nally, P., Carney, E., Henchion, M.,
Mahone, D., Cowan, C.: E. coli O157:H7 in beef burgers produced in the Republic of Ireland:
a quantitative microbial risk assessment. Tech. rep., Ashtown Food Research Centre, Teagasc,
Dublin (2006)
6. Duffy, G., Cummins, E., Nally, P., O’Brien, S., Butler, F.: A review of quantitative microbial
risk assessment in the management of Escherichia coli O157:H7 on beef. Meat Sci. 74(1),
76–88 (2006)
7. Ebel, E., Schlosser, W., Kause, J., Orloski, K., Roberts, T., Narrod, C., Malcolm, S.,
Coleman, M., Powell, M.: Draft risk assessment of the public health impact of Escherichia
coli O157:H7 in ground beef. J. Food Prot. 67(9), 1991–1999 (2004)
8. Scallan, E., Hoekstra, R.M., Angulo, F.J., Tauxe, R.V., Widdowson, M.A., Roy, S.L.,
Jones, J.L., Griffin, P.M.: Foodborne illness acquired in the United States – major pathogens.
Emerg. Infect. Dis. 17(1), 7–15 (2011)
9. Signorini, M., Tarabla, H.: Quantitative risk assessment for verocytotoxigenic Escherichia coli
in ground beef hamburgers in Argentina. Int. J. Food Microbiol. 132(2–3), 153–161 (2009)
10. Smith, B.A., Fazil, A., Lammerding, A.M.: A risk assessment model for Escherichia coli
O157:H7 in ground beef and beef cuts in canada: evaluating the effects of interventions. Food
Control 29, 364–381 (2013)
11. Tuttle, J., Gomez, T., Doyle, M.P., Wells, J.G., Zhao, T., Tauxe, R.V., Griffin, P.M.: Lessons
from a large outbreak of Escherichia coli O157:H7 infections: insights into the infectious dose
and method of widespread contamination of hamburger patties. Epidemiol. Infect. 122(2), 185–
192 (1999)
12. Wang, X., Gautam, R., Pinedo, P.J., Allen, L.J.S., Ivanek, R.: A stochastic model for
transmission, extinction and outbreak of Escherichia coli O157:H7 in cattle as affected by
ambient temperature and cleaning practices. J. Math. Biol. (2013). doi:10.1007/s00285-013-
0707-1
13. Willms, A.R., Kitanov, P.M.: Probability of Escherichia coli contamination spread in ground
beef production (2016, submitted)
14. Yang, X., Badoni, M., Youssef, M.K., Gill, C.O.: Enhanced control of microbiological
contamination of product at a large beef packing plant. J. Food Prot. 75(1), 144–149 (2012)
15. Zhang, X.S., Chase-Topping, M.E., McKendrick, I.J., Savill, N.J., Woolhouse, M.E.J.: Spread
of Escherichia coli O157:H7 infection among Scottish cattle farms: Stochastic models and
model selection. Epidemics 2, 11–20 (2010)
The Impact of Movement on Disease Dynamics
in a Multi-city Compartmental Model Including
Residency Patch
Diána Knipl
Abstract The impact of population dispersal between two cities on the spread of a
disease is investigated analytically. A general SIRS model is presented that tracks
the place of residence of individuals, allowing for different movement rates of local
residents and visitors in a city. Provided the basic reproduction number is greater
than one, we demonstrate in our model that increasing the travel volumes of some
infected groups may result in the extinction of a disease, even though the disease
cannot be eliminated in each city when the cities are isolated.
1 Introduction
The spatial spread of infectious diseases has been observed many times in history.
Most recent examples include the 2002–2003 SARS epidemic in Asia and the global
spread of the 2009 pandemic influenza A(H1N1). The Middle East Respiratory
Syndrome coronavirus (MERS-CoV) outbreak emerged in 2012, and West Africa is
currently witnessing the extensive Ebola virus (EBOV) outbreak, that pose a global
threat. There is an increasing interest in the mathematical modelling literature for
the spread of epidemics between discrete geographical locations (patches, or cities).
Such metapopulation models incorporate single or multiple species occupying
multiple spatial patches that are connected by movement dependent or independent
of disease status. Such models have been discussed for an array of infectious
diseases including measles and influenza, by Arino and coauthors [2–5], Sattenspiel
D. Knipl ()
Agent-Based Modelling Laboratory, 331 Lumbers, York University, 4700 Keele St., Toronto, ON,
M3J 1P3, Canada
MTA–SZTE Analysis and Stochastic Research Group, University of Szeged, Aradi vértanúk
tere 1, Szeged, H-6720, Hungary
Current Institution: Department of Mathematics, University College London, Gower Street,
London, WC1E 6BT, UK

256 D. Knipl
and coauthors [14, 15], Wang and coauthors [9, 13, 19, 20]. The work of Arino [1],
and Arino and van den Driessche [6] provide a thorough review of the literature.
When considering intervention strategies for epidemic models, our attention is
focused on the basic reproduction number R 0 , which is the expected number of
secondary cases generated by a typical infected host introduced into a susceptible
population. This quantity serves as a threshold parameter for disease elimination; if
R 0 < 1 then the disease dies out when a small number of infected individuals is
introduced whereas if R 0 > 1 then the disease can persist in the population. The
above mentioned works illustrate that in metapopulation models R 0 often arises
as a complicated formula of the model parameters. Such models include multiple
infected classes, and individuals’ movement makes it challenging to compute the
number of new infections generated by an infected case, and to understand the
dependence of R 0 on the movement rates. To calculate R 0 in metapopulation
models, the next-generation method is used (see Diekmann et al. [7]).
The models in [4, 5, 14, 15] include residency patch, that is, these models keep
track of the patch of origin of an individual as well as where an individual is
at a given time (either as resident, or as visitor). There are many reasons why
individuals should be distinguished in an epidemic model by their residential
statuses; visitors and local residents may have very different contact rates and
mixing patterns, but more significantly, these groups are different in their travel rates
because in reality, a large part of outbound travels from a city are return trips. In the
above works, the basic reproduction number was calculated and its dependence on
the movement rates was studied numerically. Some complicated behavior of R 0
in these parameters was highlighted in [1, 2, 4]: numerical simulations suggest
that when the infection is present in same patches but absent in others without
movement, then travel with small rates can allow for disease persistence in the
metapopulation although higher travel rates can drive the disease to extinction.
In this work we present a demographic SIRS epidemic metapopulation model
in two cities, and analytically investigate the impact of individuals’ movement
between the two cities on the disease dynamics. In each city we distinguish residents
from visitors, and consider the general situation when individuals with different
disease statuses and residential statuses have different movement rates. In our
analysis we utilize the concept of the target reproduction number, developed by
Shuai et al. in [16, 17]. This quantity measures the effort required to eliminate
infectious diseases, when an intervention strategy is targeted at single entries
or sets of such entries of a next-generation matrix. Focusing on the control of
infected individuals’ movement between the two cities—an intervention strategy
often applied in pandemic situations—we give conditions and describe how the
travel rate of a specific group or some of these groups should be changed to prevent
an outbreak.
The Impact of Movement in an Epidemic Model with Residency Patch 257
2 Model Formulation
We formulate a dynamical model to describe the spread of an infectious disease

among two cities. We divide the entire populations of the two cities into the disjoint
classes Sjm , Ijm , Rmj , j 2 f1; 2g, m 2 fr; vg, where the letters S, I, and R represent
the compartments of susceptible, infected, and recovered individuals, respectively.
Lower index j 2 f1; 2g specifies the current city, upper index m 2 fr; vg denotes
the residential status of the individual in the current city (resident or visitor). An
individual who is currently in city j and has residential status v, originally belongs
to city k hence we say that this individual has origin in city k (k 2 f1; 2g; k ¤ j). Let
Sjm .t/, Ijm .t/, Rm
j .t/, j 2 f1; 2g, m 2 fr; vg be the number of individuals belonging
to Sjm , Ijm , Rmj respectively, at time t. The transmission rate between a susceptible
individual with residential status m and an infected individual with residential status
n in region j ( j 2 f1; 2g; m; n 2 fr; vg) is denoted by ˇjmn , and disease transmission is
modelled by standard incidence. Model parameter j is the recovery rate of infected
individuals in city j, and dj is the natural mortality rate of all individuals with origin
in city j. Recovered individuals with residential status m in city j lose disease-
induced immunity by rate jm . For the total population of residents and visitors
currently being in city j we use the notations Njr and Njv , and let Njo denote the
total population with origin in j. It holds that
Njr D Sjr C Ijr C Rrj ; Njv D Sjv C Ijv C Rvj ;

Njo D Sjr C Ijr C Rrj C Skv C Ikv C Rvk ; k ¤ j:
For the recruitment term j into the susceptible resident population we assume that
j is a function of the populations Njr and Nkv (k ¤ j), that is, the populations with
origin in city j. We denote by mSm Im Rm
kj , mkj , and mkj the travel rate of susceptible,
infected, and recovered individuals, respectively, with residential status m in city
j travelling to city k. Based on the assumptions formulated above, we obtain the
following system of differential equations for the disease transmission in city j:
dSrj Sr I r Sr I v
dt D j .Njr ; Nkv / ˇjrr N r CN
j j
v ˇj
rv j j r r r Sr r Sv v
N r CN v dj Sj C j Rj mkj Sj C mjk Sk ;
j j j j
dIjr Sr I r Sr I v Iv v
dt D ˇjrr N r CN
j j
v C ˇj
rv
N r CN v .j C dj /Ij mkj Ij C mjk Ik ;
j j r Ir r
j j j j
dRrj Rv v
dt D j Ijr .jr C dj /Rrj mRr
kj Rj C mjk Rk ;
r
dSvj Sv I r Sv I v
dt
D ˇjvr N r CN
j j vv j j
v ˇj
N r CN v
dk Sjv C jv Rvj mSv v
kj Sj C mjk Sk ;
Sr r
j j j j
dIjv Sv I r Svj Ijv
dt
D ˇjvr N r CN
j j vv
v C ˇj
Njr CNjv
.j C dk /Ijv mIv v
kj Ij C mjk Ik ;
Ir r
j j
dRvj
dt D j Ijv .jv C dk /Rrj mRv v
kj Rj C mjk Rk :
Rr r
(1)
258 D. Knipl
Standard arguments from the theory of differential equations guarantee that the
system (1) is well posed. The function forming the right hand side of the system
is Lipschitz continuous, which implies the existence of a unique solution. The
derivative of each system variable is nonnegative when the variable is zero, hence
solutions remain nonnegative for nonnegative initial data. For the dynamics of the
total population with origin in city j, we obtain the equation
dNjo
D j .Njr ; Nkv / dj .Njr C Nkv /; k ¤ j:
dt
If j .Njr ; Nkv / D dj .Njr C Nkv / then the population with origin in j is constant. For
constant recruitment term j it is easy to derive that NO oj D j =dj gives the unique
equilibrium of Njo . With fixed N1o and N2o it is obvious from nonnegativity that the
solutions of the system (1) are bounded. The model is at an equilibrium if the time
derivatives in the system (1) are zero. At a disease-free equilibrium it holds that
I1r D I1v D I2r D I2v D 0 that implies Rr1 D Rv1 D Rr2 D Rv2 D 0. Thus at a DFE S1r ,
S1v , S2r , S2v satisfy
1 .N1r ; N2v / d1 S1r mSr r Sv v

21 S1 C m12 S2 D 0;
d2 S1v m21 S1v C mSr
Sv r
12 S2 D 0;
v Sv v
2 .N2 ; N1 / d2 S2 m12 S2 C m21 S1
r r Sr r
D 0;
d1 S2v mSv v Sr r
12 S2 C m21 S1 D 0:
Hence if N1o and N2o are fixed then using that 1 D d1 N1o and 2 D d2 N2o , it follows
that
r 1
S1 d1 C mSr21 mSv
12 d1 N1o
D ;
S2v mSr
21 d1 C mSv12 0
r 1
S2 d2 C mSr12 mSv
21 d2 N2o
D :
S1v mSr
12 d2 C mSv21 0
The following result is proved.

Proposition 1 Assume that the total populations with origin in city 1 and with
origin in city 2 are constant. Then there is a unique DFE in the model (1) where
.dj CmSv jk /dj Nj

o
Sjr D .dj Cmkj /.dj Cmjk /mSv
Sr Sv Sr ;
jk mkj
mSr d N o
Sjv D ; j; k 2 f1; 2g; k ¤ j;
jk k k
.dk CmSr
jk /.dk Cmkj /mkj mjk
Sv Sv Sr
Ijr D Ijv D 0; Rrj D Rvj D 0; and Njr D Sjr ; Njv D Sjv :
For the stability of the DFE in the full model (1) we linearize the subsystem of (1)
that consists of the equations for I1r , I1v , I2r , and I2v —the infected subsystem—about
the DFE, and give the Jacobian J, as
J D B G M:
2 ˇ1rr N1r ˇ1rv N1r
3
N1r CN1v N1r CN1v 0 0 2 Ir 3
6 7 m21 0 0 mIv 12
6 ˇ1vr N1v ˇ1vv N1v 7
6 0 0 7 6 0 mIv Ir
0 7
N1r CN1v N1r CN1v 6 21 m12 7;
BD6 ˇ2rv N2r 7 ; M D 4
6 0 0
ˇ2rr N2r
7 0 m21 m12
Iv Ir
0 5
4 N2r CN2v N2r CN2v 5
0 0
ˇ2vr N2v ˇ2vv N2v mIr
21 0 0 mIv
12
N2r CN2v N2r CN2v
and G D diag.1 C d1 ; 1 C d2 ; 2 C d2 ; 2 C d1 / DW diag.gr1 ; gv1 ; gr2 ; gv2 /. Let s.A/

denote the maximum real part of all eigenvalues of any square matrix A, and .A/
denote the dominant eigenvalue of any square matrix A. We say that a square matrix
A has the Z-sign pattern if all entries of A are nonpositive except possibly those in
the diagonal. If A1 0 holds then A is a non-singular M-matrix (several definitions
exist for M-matrices, see [8, Theorem 5.1]). By [18, Lemma 1] the stability of
the DFE is determined by the eigenvalues of J; more precisely, the DFE is locally
asymptotically stable if s.J/ < 0, meaning that all eigenvalues have negative real
part, and the DFE is unstable if s.J/ > 0, when there is an eigenvalue with positive
real part. The proof of the next proposition follows by similar arguments as those in
the proof of [18, Theorem 2].
Proposition 2 Consider a splitting F V of the Jacobian of the infected subsystem
about the DFE, where F is a nonnegative matrix and V is a non-singular M-matrix.
Then, it holds that s.J/ < 0 if and only if .FV 1 / < 1, s.J/ D 0 if and only if
.FV 1 / D 1, and s.J/ > 0 if and only if .FV 1 / > 1.
The stability of the DFE is often characterized through the basic reproduction
number R 0 , that is defined as the dominant eigenvalue of the next-generation
matrix (NGM). The concept of the NGM was initially introduced by Diekmann
et al. [7]. This matrix is computed as K0 WD F0 V01 , where F0 equals B, the
transmission matrix describing new infections, and V0 is defined as G C M, the
transition matrix for the transitions between and out of infected classes. F0 and
V0 satisfy the conditions of Proposition 2, hence R 0 D .K0 / D .F0 V01 / is a
threshold quantity for the stability of the DFE. We obtain the following corollary
from Proposition 2.
Corollary 1 Consider a splitting F V of the Jacobian of the infected subsystem
about the DFE, where F is a nonnegative matrix and V is a non-singular M-matrix.
Then, it holds that .FV 1 / < 1 if and only if R 0 < 1, .FV 1 / D 1 if and only if
R 0 D 1, and .FV 1 / > 1 if and only if R 0 > 1.
With other words, for any splitting F V of the Jacobian where F is nonnegative
and V is a non-singular M-matrix, there arises an alternative NGM by FV 1 ;
moreover, .FV 1 / and R 0 agree at the threshold value for the stability of the
DFE. In the next section we will investigate the impact of movement on the disease
260 D. Knipl
dynamics by constructing some alternative next-generation matrices and utilizing

the method of Shuai et al. [16] to measure the effort required to control the disease.
3 Main Results
Using the definition of G and the transmission matrix F0 , we introduce the quantities
ˇ1rr N1r Cˇ1vr N1v ˇ1rv N1r Cˇ1vv N1v

R r1 D gr1 .N1r CN1v / ; R v1 D gv1 .N1r CN1v / ;
ˇ2rr N2r Cˇ2vr N2v ˇ2rv N2r Cˇ2vv N2v
R r2 D gr2 .N2r CN2v / ; R v2 D gv2 .N2r CN2v / ;
where R mj denotes the expected number of new cases in city j when a single infected
individual with residential status m who doesn’t travel is introduced into city j.
Consider the matrices F1 D B M C diag.mIr 21 ; m21 ; m12 ; m12 / and V1 D
Iv Ir Iv
G C diag.m21 ; m21 ; m12 ; m12 /. Then J D F1 V1 gives another splitting of the

Ir Iv Ir Iv
Jacobian, moreover F1 0 and V1 is a non-singular M-matrix. We obtain the

following theorem.
Theorem 1 If R r1 > 1, R v1 > 1, R r2 > 1, and R v2 > 1 then the DFE is
unstable when the cities are isolated, and movement cannot stabilize the DFE. If
the inequalities are reversed then the DFE is stable when the cities are isolated, and
movement cannot destabilize the DFE.
Proof Consider the splitting J D F1 V1 . As V1 is a diagonal matrix, one easily
computes the alternative NGM
2 ˇ1rr N1r ˇ1rv N1r mIv
3
.mIr v v v 0 12
v
21 Cg1 /.N1 CN1 / .mIv
21 Cg1 /.N1 CN1 / .mIv
12 Cg2 /
r r r
6 7
6 ˇ1vr N1v ˇ1vv N1v mIr 7
6 v v v
12
0 7
D6 .mIr
21 Cg1 /.N1 CN1 / .mIv
21 Cg1 /.N1 CN1 / .mIr
12 Cg2 / 7:
r r r r
K1 D F1 V11 6 mIv ˇ2 N2
rr r
ˇ2 N2
rv r 7
6 0 21
v v v v 7
4 .mIv
21 Cg1 / .m12 Cg2 /.N2 CN2 / .m12 Cg2 /.N2 CN2 / 5
Ir r r Iv r
mIr ˇ2vr N2v ˇ2vv N2v

.mIr
21
0 v v v
21 Cg1 / .mIr
12 Cg2 /.N2 CN2 / .m12 Cg2 /.N2 CN2 /
r r r Iv r
A standard result for nonnegative matrices (see, e.g., [12, Theorem 1.1]) says that
the dominant eigenvalue of a nonnegative matrix is bounded below and above by
the minimum and maximum of its column sums. We look at the column sums of K1
to give upper and lower bounds on the dominant eigenvalue. The column sum in the
ˇ1rr N1r Cˇ1vr N1v CmIr
21
first column is .mIr v , and using basic calculus we derive that
21 Cg1 /.N1 CN1 /
r r
ˇ1rr N1r Cˇ1vr N1v CmIr ˇ1rr N1r Cˇ1vr N1v ˇ1rr N1r Cˇ1vr N1v
1< .mIr v
21
gr1 .N1r CN1v / if .N1r CN1v / gr1 > 0 , R r1 > 1;
21 Cg1 /.N1 CN1 /
r r
ˇ1rr N1r Cˇ1vr N1v ˇ1rr N1r Cˇ1vr N1v CmIr ˇ1rr N1r Cˇ1vr N1v
gr1 .N1r CN1v /
.mIr v
21
<1 if .N1r CN1v /
gr1 < 0 , R r1 < 1:
21 Cg1 /.N1 CN1 /
r r
Similar results follow for the second, third, and fourth columns. Thus if R r1 > 1,
R v1 > 1, R r2 > 1, and R v2 > 1 hold then all column sums are greater than 1 for any
mIr Iv Ir Iv
21 , m21 , m12 , and m12 , that implies by Proposition 2 that the dominant eigenvalue
of K1 is greater than 1 and the DFE is unstable. On the other hand, if the above
inequalities are reversed then the column sums are less than 1 for any movement
rates and the DFE is stable by .K1 / < 1. t
u
Next, we investigate some cases when changing the movement rates of some
groups can stabilize the DFE. We construct the matrix K2 WD F2 V21 , where F2 is
formed as we let ŒF2 1;1 D ŒF1 1;1 gr1 and ŒF2 i;j D ŒF1 i;j if .i; j/ ¤ .1; 1/, and
v v
V2 D diag.mIr21 ; g1 C rrmIv ; gr2 C mIr
12 ; g2 C m12 /. V2 is a non-singular M-matrix and
Iv
21
ˇ N r
F2 is nonnegative if N r1CN1v > gr1 . This condition is equivalent to when the number of
1 1
new infections amongst residents of city 1 is less than 1, when an infected resident
who doesn’t travel is introduced into city 1. The alternative NGM is computed as
2 3
ˇ1rr N1r gr1 ˇ1rv N1r mIv
v v v 0 12
v
6 mIr 7
r Ir
.N
21 1 CN 1 / m21 .mIv
21 Cg r
1 /.N1 CN1 / .mIv
12 Cg2 /
6 ˇ1vr N1v ˇ1vv N1v mIr 7
6 v v v
12
0 7
K2 D F2 V2 D 6
1 7;
Ir r
m21 .N1 CN1 / .mIv r
21 Cg1 /.N1 CN1 / .mIr
12 Cg2 /
r
6 mIv rr r
ˇ2 N2 ˇ2rv N2r 7
6 0 21
v 7
4 .mIv
21 Cg1 /
v
.mIr r r v
12 Cg2vr/.Nv2 CN2 / .m12 Cg2 /.N2 CN2 / 5
Iv v
vv v
r
ˇ2 N2 ˇ2 N2
1 0 .mIr r r v
.mIv v r v
12 Cg2 /.N2 CN2 / 12 Cg2 /.N2 CN2 /
which is irreducible. Denote by L2 the matrix that is formed by replacing ŒK2 1;1 and
ŒK2 2;1 in K2 by 0. Observe that K2 converges to L2 as mIr
21 goes to infinity. We show
that the disease can be eliminated by controlling only the travel rate of the residents
of a single city.
ˇ1rr N1r
Theorem 2 Assume that R 0 > 1, that is, the DFE is unstable. If N1r CN1v
> gr1 and
ˇ N
rr r
.L2 / < 1 then increasing mIr
21 can stabilize the DFE. In particular, if N r1CN1v > gr1
1 1
and R v1 < 1, R r2 < 1, and R v2 < 1, then increasing mIr
21 can stabilize the DFE.
Proof We utilize some terminology and results from [16, 17]. Let S D
f.1; 1/; .2; 1/g, and define the 4 4 matrix K2S as ŒK2S i;j D ŒK2 i;j if .i; j/ 2 S and 0
S
otherwise. Note that S identifies the set of elements in K2 that depend on mIr 21 , and K2
Ir
contains elements of K2 that are subject to change when m21 is targeted. Following
the terminology of [16] it is thus meaningful to refer to S as the target set and to K2S
as the target matrix. Note that L2 D K2 K2S , hence the condition .L2 / < 1 implies
that .K2 K2S / < 1, that is, the controllability condition holds and it is possible to
stabilize the DFE by controlling only the elements in S [16].
We compute T S D .K2S .I K2 C K2S /1 /, the number referred to as the target
reproduction number in [16]. Here I denotes the 4 4 identity matrix. Let .mIr 21 / D
c
m21 T S , where we denote by .m21 / the controlled travel rate of infected residents
Ir Ir c
of city 1 travelling to city 2. It follows from Corollary 1 and [16, Theorem 2.1] by
R 0 > 1 that .mIr 21 / > m21 . The matrix K2 , constructed as we replace m21 in K2 by
c Ir c Ir
262 D. Knipl
.mIr
21 / , satisfies .K2 / D 1 by [16, Theorem 2.2], which means that the disease can
c c
be eradicated by increasing mIr 21 .

Note that the conditions R v1 < 1, R r2 < 1, and R v2 < 1 ensure that .L2 / <
1. Indeed, it is easy to see that the column sums of the second, third, and fourth
columns in L2 are less than 1 for any travel rates, and the column sum in the first
column is 1. We now show that 1 is not an eigenvalue of L2 , which together with
[12, Theorem 1.1] implies that the dominant eigenvalue of L2 is less than 1. Assume
that 1 is an eigenvalue of L2 , and consider a left eigenvector v D Œv1 ; v2 ; v3 ; v4
corresponding to 1. It holds that
v L2 D 1 v;
and we deduce that
v4 D v1 ; max.v1 ; v2 ; v3 / > v2 ; max.v2 ; v3 ; v4 / > v3 ; max.v1 ; v3 ; v4 / > v4 :
From the fourth inequality and v4 D v1 it follows that v3 > v4 , which together with
the third inequality implies v2 > v3 > v4 , but v1 > v2 by the second inequality, a
contradiction to v1 D v4 . The proof is complete. t
u
To reveal the impact of visitors’ travel, a result analogous to Theorem 2 can be
formulated. The proof of the following theorem follows by similar arguments to
those in Theorem 2.
ˇ1vv N1v
Theorem 3 Assume that R 0 > 1, that is, the DFE is unstable. If N1r CN1v
> gv1 , and
R r1 < 1, R r2 < 1, and R v2 < 1, then increasing mIv
21 can stabilize the DFE.
Lastly, we give conditions under which controlling outbound travel from one
city is sufficient for disease elimination. Consider two matrices F3 and V3 , defined
as ŒF3 1;1 D ŒF1 1;1 gr1 , ŒF3 2;2 D ŒF1 2;2 gv1 , and ŒF3 i;j D ŒF1 i;j otherwise, and
v
V3 D diag.mIr 21 ; m 21 ; g2 C m12 ; g2rrCr m12 /. V3 is a non-singular M-matrix and F3 is
Iv r Ir Iv
vv v
ˇ N ˇ N
nonnegative if N r1CN1v > gv1 and N r1CN1v > gr1 . The following theorem concerns about
1 1 1 1
whether changing the movement rates of the current population of one city can lead
to disease eradication.
ˇ1rr N1r
Theorem 4 Assume that R 0 > 1, that is, the DFE is unstable. If N1r CN1v > g1 and
r
ˇ1vv N1v
> gv1 but R r2 < 1 and R v2 < 1, then increasing mIr
N1r CN1v 21 and mIv
21 can stabilize
the DFE.
Proof The proof is similar to the proof of Theorem 2. We compute the alternative
NGM
2 3
ˇ1rr N1r gr1 ˇ1rv N1r mIv
v v 0 12
v
6 21 1 vr 1 v
mIr
.N r
CN / mIr
21 m Iv r
21 .N1 CN1 / .mIv
12 Cg2 / 7
6 ˇ1vv N1v gv1 mIr 7
6 mIr ˇ.N1 rNCN 1
12
0 7
K3 D F3 V31 D6 6 21 1
v
1/ mIv r
21 .N1 CN1 /
v
mIv
21 .mIr
12rrCg
r
2/ 7;
ˇ2 N2r ˇ2rv N2r 7
6 0 1 .m12 Cg2 /.N2 CN2v / v v 7
4 Ir r r
.m12 Cg2 /.N2 CN2 / 5
Iv r
ˇ2vr N2v vv v
ˇ2 N2
1 0 .mIr r r v
.mIv v r v
12 Cg2 /.N2 CN2 / 12 Cg2 /.N2 CN2 /
which is irreducible, and define the target set U by identifying the entries of K3 that
depend on mIr 21 and/or m21 . We let U D f.1; 1/; .1; 2/; .2; 1/; .2; 2/g, and define the
Iv
4 4 target matrix K3 as ŒK3U i;j D ŒK3 i;j if .i; j/ 2 U and 0 otherwise. Note that
U
.K3 / > 1 holds by R 0 > 1. However, the result in [12, Theorem 1.1] on the upper
bound of the dominant eigenvalue implies that .K3 K3U / 1.
Assume that .K3 K3U / D 1, that is, 1 is an eigenvalue of K3 K3U . Then there
is a left eigenvector v D Œv1 ; v2 ; v3 ; v4 such that
v .K3 K3U / D 1 v
holds. We derive that
v4 D v1 ; v3 D v2 ; max.v2 ; v3 ; v4 / > v3 ; max.v1 ; v3 ; v4 / > v4 ;
so v4 > v3 and v3 > v4 hold by the third and fourth inequalities, a contradiction. We
showed that .K3 K3U / < 1, which means that there is a potential to control mIr 21 and
mIv
21 in a way such that the dominant eigenvalue of the controlled matrix drops below
1 (by decreasing targeted entries of K3 to values close to 0). This condition also
allows us to compute the target reproduction number T U D .K3U .I K3 CK3U /1 /.
By [16, Theorem 2.2], the controlled matrix K3c satisfies .K3c / D 1 where K3c is
formed by replacing ŒK3 i;j by ŒK3 i;j =T U if .i; j/ 2 U, that is achieved by replacing
21 by .m21 / D m21 T U , and m21 by .m21 / D m21 T U . Note that T U > 1 by [16,
mIr Ir c Ir Iv Iv c Ir
Theorem 2.2], which means that the disease can be eradicated by increasing mIr 21 and
mIv
21 . t
u
In the case of transmission coefficients equal for all populations present in a city,
recovery rates equal for all populations and death rates equal for all populations,
R r1 and R v1 reduce to ˇ1 =. C d/, and R r2 and R v2 reduce to ˇ2 =. C d/. Note
that these quantities give the expected number of secondary infections generated by
a single infected case in city 1 and city 2, respectively, in the absence of movement
between the cities. Hence the local reproduction numbers in city 1 and city 2 can be
defined as we consider our model without dispersal:
ˇ1 ˇ2
R loc
1 D ; R loc
2 D :
Cd Cd
We derive the following results from Theorems 1 and 4.

Corollary 2 Suppose that ˇ1mn D ˇ1 and ˇ2mn D ˇ2 for all m; n 2 fr; vg, j D
and dj D d for all j 2 f1; 2g. Then, the DFE is unstable when the cities are isolated
and R loc
1 > 1, R 2 > 1, and movement cannot stabilize the DFE. In the case
loc
when R 1 < 1 and R loc

loc
2 < 1, the DFE is stable when the cities are risolated, and
ˇ N
movement cannot destabilize the DFE. If the DFE is unstable and N r 1CN1 v > . C d/,
1 1
ˇ1 N1v
N1r CN1v> . C d/ but R loc2 < 1, then increasing the movement rates of individuals
in city 1 can stabilize the DFE.
264 D. Knipl
4 Discussion
A two-city compartmental epidemic model was considered to reveal the impact of

population dispersal on disease persistence. This general SIRS model is applicable
for an array of infectious diseases, and it can also be reduced to simpler models (SIS,
SIR models) by setting parameters (or their inverses) to zero. In the model setup we
distinguish local residents from temporary visitors in each city, that results in four
infected classes in the model. We demonstrated that controlling the movement of
one or two infected groups can be sufficient for preventing a disease outbreak. It
was discussed in [11] that the role of different inflow rates of residents and visitors
into a city is not necessarily significant in regards of the total epidemic burden, but
it is of particular importance for pandemic preparedness, when it comes to assessing
the risk for each group to import the infection to a disease-free city.
Modelling the spatial spread of diseases in metapopulations remains a complex
task. This paper does not concern with models that include multiple species, hence
more analysis is needed to quantify the effect of movement between patches in such
models, which are useful in investigating vector-borne diseases and their control
strategies. Combining some intervention measures—like the mutual control of
dispersal rates and transmission rates—requires less effort for disease elimination,
hence there is a potential to incorporate the results of this work into systematic risk
assessment analyses, as described in [10].
Acknowledgements D. Knipl acknowledges the support by the Cimplex project funded by

the European Commission in the area “FET Proactive: Global Systems Science” (GSS), as a
Research and Innovation Action, under the H2020 Framework programme, Grant agreement
number 641191.
References
1. Arino, J.: Diseases in metapopulations. In: Modeling and Dynamics of Infectious Diseases.
Series in Contemporary Applied Mathematics, vol. 11, pp. 65–123. Higher Education Press,
Beijing (2009)
2. Arino, J., Davis, J.R., Hartley, D., Jordan, R., Miller, J.M., van den Driessche, P.: A multi-
species epidemic model with spatial dynamics. Math. Med. Biol. 22(2), 129–142 (2005)
3. Arino, J., Jordan, R., van den Driessche, P.: Quarantine in a multi-species epidemic model with
spatial dynamics. Math. Biosci. 206(1), 46–60 (2007)
4. Arino, J., van den Driessche, P.: A multi-city epidemic model. Math. Popul. Stud. 10(3), 175–
193 (2003)
5. Arino, J., van den Driessche, P.: The basic reproduction number in a multi-city compartmental
epidemic model. In: Positive Systems, pp. 135–142 Springer, Berlin/Heidelberg (2003)
6. Arino, J., van den Driessche, P.: Disease spread in metapopulations. In: Nonlinear Dynamics
and Evolution Equations, vol. 48, pp. 1–13. American Mathematical Society, Providence
(2006)
7. Diekmann, O., Heesterbeek, J.A.P., Metz, J.A.J.: On the definition and computation of the basic
reproduction ratio R0 in models for infectious diseases in heterogeneous populations. J. Math.
Biol. 28, 365–382 (1990)
8. Fiedler, M.: Special Matrices and Their Applications in Numerical Mathematics. Martinus
Nijhoff Publishers, Dodrecht (1986)
9. Jin, Y., Wang, W.: The effect of population dispersal on the spread of a disease. J. Math. Anal.
Appl. 308(1), 343–364 (2005)
10. Knipl, D.: A new approach for designing disease intervention strategies in metapopulation
models. J. Biol. Dyn. 10(1), 71–94 (2016). http://dx.doi.org/10.1080/17513758.2015.1107140
11. Knipl, D.H., Röst, G., Wu, J.: Epidemic spread and variation of peak times in connected regions
due to travel related infections—dynamics of an anti-gravity type delay differential model.
SIAM J. Appl. Dyn. Syst. 12(4), 1722–1762 (2013)
12. Minc, H.: Nonnegative Matrices. Wiley Interscience, New York (1988)
13. Ruan, S., Wang, W., Levin, S.A.: The effect of global travel on the spread of SARS. Math.
Biosci. Eng. 3(1), 205 (2006)
14. Sattenspiel, L., Dietz, K.: A structured epidemic model incorporating geographic mobility
among regions. Math. Biosci. 128(1), 71–91 (1995)
15. Sattenspiel, L., Herring, D.A.: Simulating the effect of quarantine on the spread of the 1918–19
flu in central Canada. Bull. Math. Biol. 65(1), 1–26 (2003)
16. Shuai, Z., Heesterbeek, J.A.P., van den Driessche, P.: Extending the type reproduction number
to infectious disease control targeting contacts between types. J. Math. Biol. 67(5), 1067–1082
(2013)
17. Shuai, Z., Heesterbeek, J.A.P., van den Driessche, P.: Erratum to: extending the type repro-
duction number to infectious disease control targeting contacts between types. J. Math. Biol.
71(1), 1–3 (2015)
18. Van den Driessche, P., Watmough, J.: Reproduction numbers and sub-threshold endemic
equilibria for compartmental models of disease transmission. Math. Biosci. 180(1), 29–48
(2002)
19. Wang, W., Mulone, G.: Threshold of disease transmission in a patch environment. J. Math.
Anal. Appl. 285(1), 321–335 (2003)
20. Wang, W., Zhao, X.Q.: An epidemic model in a patchy environment. Math. Biosci. 190(1),
97–112 (2004)
A Chemostat Model with Wall Attachment:
The Effect of Biofilm Detachment Rates
on Predicted Reactor Performance
Alma Mašić and Hermann J. Eberl
Abstract We consider a previously introduced mathematical model of chemostat

with suspended and wall attached growth and exchange of biomass via biofilm
detachment and reattachment. In this study we investigate the role of the specific
choice of a biomass detachment criterion. We find that this choice does greatly affect
output parameters such as biomass in the system, but it does not affect strongly
effluent concentration and hence the prediction of reactor performance.
1 Introduction
Bacterial biofilms are layers on immersed surfaces that form wherever environ-
mental conditions sustain microbial growth. They play an important role in several
environmental engineering applications, most notably in wastewater treatment [12],
where several technologies have been designed based on biofilm properties. An
important part of the biofilm life cycle is detachment, or dispersal, the transfer of
biomass from the biofilm into the aqueous phase [3, 10]. Several triggers for this
phenomenon have been identified, including external factors, such as shear forces, or
internal factors, such as weakening of the EPS matrix and quorum sensing signaling
mechanisms [11, 13]. Several attempts to incorporate detachment into mesoscale
mathematical models of biofilms have been suggested in the literature, both in the
traditional 1D Wanner-Gujer biofilm model and its simplification, e.g. [2, 8, 11, 16],
and in two- and three-dimensional biofilm models, e.g. [4–6, 18].
It is clear that the mesoscopic description of biofilm detachment strongly affects
the mesoscopic biofilm structure predicted by these models. What is less clear is
to which extent the mesoscopic description of biofilm detachment affects global
parameters that assess reactor performance. For the setting of biofilms in a porous
A. Masic
EAWAG, Überlandstrasse 133, 8600 Dübendorf, Switzerland
H.J. Eberl ()
University of Guelph, 50 Stone Rd E, Guelph, ON N1G2W1, Canada

268 A. Mašić and H.J. Eberl
medium, a partial answer was given in [1], where a multi-scale model was obtained
by upscaling a Wanner-Gujer type biofilm model with four different detachment
criteria to the reactor scale; it was found that the particular choice of the mesoscopic
detachment rate does not affect the macroscopic description. Here we are interested
in the question whether for a continuous stirred tank reactor (CSTR) with wall
attached and suspended growth the particular description of biofilm detachment
affects the effluent substrate concentration predicted by the model. Models of this
type arise in the modeling of wastewater treatment processes where both biofilms
and nonsessile bacteria are present and contribute to the substrate degradation
process. In the focus of our interest is the case where the reactor per se is designed
as a biofilm reactor, in which suspended biomass occurs as a side effect through the
exchange of biomass between biofilm and aqueous phase.
We cast a simple model for a continuous stirred tank reactor with suspended and
wall attached growth in terms of the dependent variables substrate concentration
S [gm3 ], suspended biomass u [g] and biofilm thickness [m]. Following [9], it
reads
uu .S/ J.S; /

SP D D.Sin S/ ; (1)
V V
uP D u.u .S/ D ku / C Ad./ ˛u; (2)
˛u
P D v.; t/ C d./: (3)
A
Here, D [d1 ] is the dilution rate, Sin [gm3 ] the inflow substrate concentration,
[-] the yield coefficient, ku [d1 ] the cell death rate for suspended bacteria, and ˛
[d1 ] is the rate at which suspended bacteria attach to the biofilm, V [m3 ] is the
reactor volume and A [m2 ] the colonizable surface area. The biomass density in the
biofilm is [gm3 ].
In (3), the function v D v.z; t/ [md1 ] denotes the growth induced velocity of the
biomass at a location z in the biofilm. Due to the incompressibility assumption that
the biomass density is constant across the biofilm, biofilm expansion is essentially
equivalent to biomass growth. Velocity v is obtained as the integral of the biomass
production rate
Z z
v.z; t/ D . .C.// k /d; (4)
0
where k [d1 ] is the cell death rate for biofilm bacteria.

Biofilm Detachment in a Chemostat with Wall Attachment 269
The substrate dependent bacterial growth rates are defined via Monod kinet-
ics, i.e.
max
u S max
C.z/
u .S/ D ; .C.z// D ; (5)
Ku C S K C C.z/
where max
u ;
max
[d1 ] are the maximum specific growth rates, Ku ; K [gm3 ] the
half-saturation concentrations and C.z/ [gm3 ] denotes the substrate concentration
in the biofilm at thickness z [m] from the substratum. It is obtained as the solution
of the two-point boundary value problem

Dc C00 .z/ D .C.z//; C0 .0/ D 0; C./ D S: (6)

Here Dc [m2 d1 ] is the diffusion coefficient. The boundary condition at the
substratum, z D 0, describes that substrate does not leave the reactor through the
walls, while the boundary condition at z D implies that external mass transfer
resistance at the biofilm/water interface is neglected. In (6) we used that substrate
diffusion is a much faster process than biofilm growth, i.e. that (6) can be considered
in a quasi-steady state.
In (1), the sink J [gd1 ] denotes the substrate flux from the aqueous phase into
the biofilm, i.e.
dC
J.S; / D ADc ./: (7)
dz
Detachment of biomass from the biofilm is described by the volumetric detach-

ment rate d./ [d1 ], which we assume to be differentiable.
A frequently used detachment rate expression in biofilm modeling is to assume
that d is proportional to ,
d./ D E1 ; (8)
leading to a quadratic sink term in (3); E1 [d1 m1 ] is the erosion or detachment
parameter. This traditional detachment model was the only one used in [9]. Another
detachment rate function that is found in the literature is to assume a constant rate, i.e
d./ D E2 ; (9)
which leads to a first order sink term with erosion parameter E2 [d1 ].
It is often assumed that the detachment rate depends also on the hydraulic
conditions in the reactor, which determine shear forces acting on the biofilm.
Therefore, one can correlate E1;2 with D, see below in Sect. 3.
We assume all model parameters to be positive.
First we formally re-write our model as an ordinary initial value problem. Note
that integrating (6) once and using the boundary conditions gives
Z
dC
./ D .C.z//dz: (10)
dz Dc 0
We define
( R

Dc 0 .C.z//dz; >0
j.; S/ WD (11)
0 D 0:
Note that C.z/ is indirectly a function of S due to the boundary condition in (6),
therefore also j is a function of S. Then (1), (2), (3), (4), (5), (6), and (7) becomes
1 uu .S/
SP D D.Sin S/ C ADc j.; S/ (12)
V
uP D u.u .S/ D ku / C Ad./ ˛u (13)
Dc ˛u
P D j.; S/ k C d./: (14)
A
While the function j.; S/ is not known explicitly in terms of elementary

functions, important qualitative properties are known, which we repeat here from
[9]:
Lemma 1 The function j.; S/ for 0, S 0 is well-defined, non-negative and
differentiable. It has the following properties:
(a) j.; 0/ D 0, j.0; / D 0,
@j
(b) @S .0; S/ D 0,
(c)
S @j S
.0; S/ ; (15)
K C S @ K
max
where D
Dc and K is the half-saturation coefficient from (5).
Proposition 1 The initial value problem of (12), (13) and (14) with non-negative
initial data possesses a unique non-negative solution for all t > 0. There is no time
interval .t1 ; t2 /, 0 < t1 < t2 , over which a non-trivial solution exists with either
u 0 or 0.
Proof (sketch) The function j.; S/, and thus the system (12), (13) and (14) is well-
defined. The tangent criterion of [15] in the usual way can be used to confirm the
positive invariance of the non-negative cone. In the non-negative cone the system
satisfies a Lipschitz condition, which implies existence and uniqueness. An upper
estimate for the solutions can be constructed by the differential inequality techniques
from a linear combination of the model equations, more specifically, an upper bound
can be derived for VS.t/ C u.t/ C A.t/ which then, using non-negativity, implies
upper estimates on each of S.t/; u.t/; .t/. The last statement in the assertion follows
directly from (13) and (14). t
u
Proposition 2 The washout equilibrium .Sin ; 0; 0/ always exists. It is unstable if
at least one of u .Sin / D ku ˛ and Dc @ @j
.0; Sin / d.0/ k is positive;
if both expressions are
negative then the washoutequilibrium is stable if WD
@j
u .Sin / D ku ˛ Dc @ .0; Sin / d.0/ k ˛d.0/ > 0 and unstable if
the inequality is reversed.
Proof It is easily verified that the trivial equilibrium E0 D .Sin ; 0; 0/ always exists.
To determine the stability of the equilibrium we calculate the Jacobian J.Sin ; 0; 0/
of the right hand side of (12), (13) and (14). Using Lemma 1(a),(b) we find
0 1
uV
.S / c @j
in
V @ .0; S /
AD in
D
B C
J.Sin ; 0; 0/ D @ 0 u .Sin / D ku ˛ Ad.0/ A (16)
˛ Dc @j
0 A @ .0; S in
/ d.0/ k
The eigenvalues are D < 0 as well as the eigenvalues of the 2 2 sub-matrix

!
j22 j23 u .Sin / D ku ˛ Ad.0/
WD ˛ Dc @j (17)
j32 j33 A @ .0; Sin / d.0/ k
We distinguish now between three cases: (i) if j22 > 0; j33 > 0 then the trace of this
sub-matrix is positive, hence at least one eigenvalue has positive real part and the
equilibrium is unstable. (ii) if j22 < 0; j33 > 0 or vice versa, then the sub-matrix has a
negative determinant and hence one positive and one negative eigenvalue, implying
instability. (iii) if j22 < 0; j33 < 0 then the trace of the sub-matrix is negative and its
determinant is obtained as

Dc @j
D u .S / D ku ˛
in
.0; S / d.0/ k ˛d.0/;
in
(18)
@
i.e. depends on parameters. Positive implies stability, negative implies insta-

bility. While for d.0/ D 0, as in detachment rate function (8), stability is automatic,
this is not necessarily true if d.0/ > 0, as in detachment rate function (9). t
u
Remark 1 Weaker, but easier to evaluate and to apply stability criteria can be
@j
obtained by replacing @ .0; S/ in Proposition 2 with the estimates in Lemma 1(c): [i]
max
S
in
The inequality K CSin
d.0/ k > 0 is sufficient (but not necessary) for instability
max
S
in
of the washout equilibrium. [ii] If u .Sin /Dku ˛ < 0 and K d.0/k < 0
max Sin
then the inequality .u .S in
/ D ku ˛/. KCSin d.0/ k / > ˛d.0/ is sufficient

(but not necessary) for stability of the washout equilibrium, whereas the inequality
max
S
in
.u .Sin / D ku ˛/. K
d.0/ k / < ˛d.0/ is sufficient for its instability.
Remark 2 For detachment rate functions with d.0/ D 0, such as (8), the above
analysis simplifies. The eigenvalues of J.Sin ; 0; 0/ are then its diagonal elements and
case (iii) above will always be stable. The stability results for cases (i), (ii) are the
same but follow immediately from the sign of the diagonal entries. This implies that
the answer to the question whether a biofilm can be established in the CSTR with
wall attached and suspended growth can depend on the detachment criterion used
in the modeling study. Whereas in models with d.0/ D 0 the specific detachment
rate coefficient does not affect the outcome, it does so in detachment models with
d.0/ > 0. This is also consistent with the results of the upscaling study of [1] for
porous medium systems.
Non-trivial steady states, in which both wall attached and suspended biomass are
present are much more difficult to analyse, even in the algebraically much simpler
Freter model, in which no substrate gradients in the wall depositions are accounted
for [7, 14]. Therefore, we do not expect any insightful results in pursuing this line
of investigation and turn to a numerical study instead.
3 Simulation Results
Typical simulation results are shown in Fig. 1, using the parameters in Table 1.
In one case the bulk substrate concentration Sin is chosen low enough so that the
bacteria cannot be sustained; in the other case it is high enough such that both a
biofilm and a suspended population attain a positive equilibrium value. The vast
majority of biomass in the system is sessile. These simulations were conducted with
the detachment rate function (8).
In Fig. 2 we show for both detachment rate function (8) and (9) the steady state
values for ; u; S for different detachment parameters E1 and E2 , which have been
correlated with the dilution rate. Motivated by [11] we chose the relationship as
0:58
D
EQ i D Ei ; (19)
D0
where by D0 we denote a reference dilution rate such that for D D D0 this flow rate
dependent criterion is equivalent to a detachment with a given rate constant Ei and
for D > D0 we have EQ i > Ei , while for D < D0 we have EQ i < Ei . Note that an
increased dilution rate implies a more plentiful substrate supply.
The main observation is that the choice of detachment rate function and erosion
constant does not have an effect on the substrate concentration in outflow, i.e. the
substrate removal performance of the reactor: For smaller dilution rates D < 50, the
effluent substrate concentration increases almost linearly with D, for larger D > 50,
−6
x 10
1
Aρλ
0.1 u
0.8
suspended and biofilm biomass (g)

substrate concentration (g/m )
3
0.6
0.0999 0.4
0.2
0.0998 0
0 5 10 15 20 25 0 5 10 15 20 25
t (days) t (days)
10
Aρλ
0.08 u
substrate concentration in reactor (g/m )
3
8
suspended and biofilm biomass (g)
0.06
6
0.04
4
0.02
2
0 0
0 5 10 15 0 5 10 15
t (days) t (days)
Fig. 1 A typical simulation of model (12), (13) and (14). Case 1 (top): The washout equilibrium
is stable, case 2 (bottom): the washout equilibrium is unstable. In both simulations the detachment
criterion (8) was used
Table 1 Model parameters used in the simulations

Symbol Parameter Value Reference
˛ Attachment rate 1/day Assumed
Dc Diffusion coefficient 104 m2 /day [17]
E1 Erosion parameter 1000=.mday) Assumed
Yield of biomass from substrate 0:63 - [17]
K , Ku Half-saturation coefficients 4 g/m3 [17]
k , ku Death rates 0:4/day [17]
max
, u
max
Maximum specific growth rates 6/day [17]
Biofilm biomass density 10;000 g/m3 [17]
the increase becomes sublinear. The results obtained for different base dilution
rates D0 , hence different detachment rates Ei , appear indistinguishable for each
detachment criterion (8) and (9). Also comparing the results between the simulation
experiments for (8) and (9), respectively, shows that for both detachment criteria the
same values of S are found.
The biofilm thickness, however, depends strongly on the detachment rate func-
tion, and so does the suspended biomass density and by extension the amount of
biomass in the effluent of the reactor.
An explanation for this phenomenon is that in a thick enough biofilm, as is typical
for many biofilm based wastewater treatment applications, only a small part of the
biofilm is active whereas a major part does not contribute to reactor performance
due to substrate limitations in the inner layers of the film. An increased detachment
force decreases the biofilm thickness but as long as it remains above the thickness
of the active layer this does not have an effect on performance. If the detachment is
strong enough such that the resulting biofilm becomes too thin, this statement does
not hold anymore.
4 Conclusion
We have investigated the question whether the choice of a mesoscale biofilm

detachment model affects the prediction of reactor performance in a macroscale
biofilm reactor model.
We have found that for the question whether a biofilm can be established or
whether biomass is washed out the choice of detachment rate can matter: The
detachment coefficient does not affect stability of the washout equilibrium in case of
a detachment rate function that vanishes in the absence of biofilm, i.e. if d.0/ D 0.
However, if detachment rate functions with d.0/ > 0 are used, then the coefficients
of the detachment model can affect stability of the washout equilibrium.
In biofilm based reactor technology, the washout equilibrium is of minor
importance, because reactors are designed such that sufficiently thick biofilms are
−3
x 10
25 0.015 1
D =1
0
D =10
0
D0=40
D0=100
20
E=1000
0.01
substrate in reactor (g/m3)
suspended biomass (g)
biofilm thickness (m)

15
0.5
10
0.005
D =1 D0=1
5 0
D0=10 D =10
0
D0=40 D0=40
D0=100 D =100
0
E=1000 E=1000
0 0 0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
D (1/day) D (1/day) D (1/day)
× 10 -3
25 0.008
D 0=1
D 0=10
D 0=40
20 D 0=100 2
E=0.05
substrate in reactor (g/m3)
suspended biomass (g)
biofilm thickness (m)
15
0.004
10 1
D 0=1 D 0=1
5 D 0=10 D 0=10
D 0=40 D 0=40
D 0=100 D 0=100
E=0.05 E=0.05
0 0 0
0 20 40 60 80 100 0 20 40 60 80 100 0 20 40 60 80 100
D (1/day) D (1/day) D (1/day)
Fig. 2 Steady states attained in dependence of dilution rate D. Top: detachment criterion (8),
reprinted from [9] with permission; bottom: detachment criterion (9)
obtained. In our simulations we show that in such situations with well developed,
sufficiently thick biofilms the particular choice of a mesoscopic detachment rate
function does not quantitatively affect macroscopic reactor performance as mea-
sured in terms of effluent substrate concentration, but it does affect the predicted
biofilm thickness.
Acknowledgements HJE was supported by NSERC Canada through the Discovery Grant Pro-
gram and through the Canada Research Chairs Program.
References
1. Abbas, F., Eberl, H.J.: Investigation of the role of mesoscale detachment rate expressions in a
macroscale model of a porous medium biofilm Reactor, Int. J. Biomath. Biostat. 2, 123–143
(2013)
2. Abbas, F., Sudarsan, R., Eberl, H.J.: Longtime behaviour of one-dimensional biofilm models
with shear dependent detachment rates. Math. Biosci. Eng. 9, 215–239 (2012)
3. Bester, E., Edwards, E.A., Wolfaardt, G.M.: Planktonic cell yield is linked to biofilm
development. Can. J. Microbiol. 55, 1195–1206 (2009)
4. Duddu, R., Chopp, D.L., Moran, B: A two-dimensional continuum model of biofilm growth
incorporating fluid flow and shear stress based detachment. Biotechnol. Bioeng. 103, 92–104
(2008)
5. Emerenini, B.O., Hense, B.A., Kuttler, C., Eberl, H.J.: A mathematical model of quorum
sensing induced biofilm detachment. PLoS ONE 10(7), e0132385 (2015)
6. Hunt, S.M., Hamilton, M.A., Sears, J.T., Harkin, G., Reno, J: A computer investigation of
chemically mediated detachment in bacterial biofilms. J. Microbiol. 149, 1155–1163 (2003)
7. Jones, D., Kojouharov, H.V., Le, D., Smith, H.L.: The Freter model: a simple model of biofilm
formation. Math. Biol. 47, 137–152 (2003)
8. Kommendal, R., Bakke, R.: Modeling Pseudomonas aeruginosa biofilm detachment. HIT
Working Paper no 3/2003 (2003)
9. Masic, A., Eberl, H.J.: Persistence in a single species CSTR model with suspended flocs and
wall attached biofilms. Bull. Math. Biol. 74, 1001–1024 (2012)
10. Morgenroth, E.: Detachment: an often-overlooked phenomenon in biofilm research and
modeling. In: Wuertz, S., et al. (eds.) Biofilms in Wastewater Treatment, pp 246–290. IWA
Publishing, London (2003)
11. Rittmann, B.E.: The effect of shear stress on biofilm loss rate. Biotechnol. Bioeng. 24, 501–506
(1982)
12. Rittmann, B.E, McCarty, P.L.: Environmental Biotechnology. McGraw-Hill, Boston (2001)
13. Solano, C., Echeverz, M., Lasa, I.: Biofilm dispersion and quorum sensing. Curr. Opin.
Microbiol. 18, 96–104 (2014)
14. Stemmons, E.D., Smith, H.L.: Competition in a chemostat with wall attachment. SIAM J. Appl.
Math. 61, 567–595 (2000)
15. Walter, W.: Gewöhnliche Differentialgleichungen, 7th edn. Springer-Verlag, Berlin (2000)
16. Wanner, O., Gujer, W.: A multispecies biofilm model. Biotechnol. Bioeng. 28, 314–328 (1986)
17. Wanner, O., Eberl, H., Morgenroth, E., Noguera, D.R., Picioreanu, C., Rittmann, B., van
Loosdrecht, M.: Mathematical modeling of biofilms. Scientific and Technical Report No.18.
IWA Publishing, London (2006)
18. Xavier, J.B., Picioreanu, C., van Loosdrecht, M.C.M: A general description of detachment for
multidimensional modeling of biofilms. Biotechnol. Bioeng. 91, 651–669 (2005)
Application of CFD Modelling to the Restoration
of Eutrophic Lakes
A. Najafi-Nejad-Nasser, S.S. Li, and C.N. Mulligan
Abstract Eutrophication has been a worldwide lake pollution problem, caused by

the presence of excessive nutrients in lakes. The nutrients can come from an external
or internal source. The release of phosphorous (P) from resuspended sediments
from the lake bottom is a significant internal source. This paper discusses how
to effectively control such a release. We considered using artificial circulation
technique, and carried out CFD modelling of circulation triggered by air-bubble
injection into the lake water. The simulations are based on the RANS equations. We
predicted distributed water and air-bubble velocities, as well as air volume fraction.
The predictions compare well with experimental data. Turbulent eddy motions cause
oxygenated surface water to flow downward and effectively mix with bottom water.
The air bubbles directly enhance the dissolved oxygen level. Both mechanisms
would inhibit the release of P from bottom sediments. Using proper methods for
interphasal forces and turbulence closure is the key to success.
1 Introduction
Lake eutrophication is a nutrient-enrichment scenario where an excess amount of

nutrients enters a lake and causes a drastic growth of algae. The subsequent death
of algae typically forms a thin greenish layer on the lake surface. Eutrophication
reduces light penetration into the lower water column and re-oxygenation of water
through air circulation [2]. Dead algae would become food for bacteria. The food
consumption process uses oxygen, and thus causes the dissolved oxygen level of the
lake water to drop, leading to fish kills [2]. Also, eutrophication causes undesirable
odour and taste [6, 9].
Lake eutrophication occurred worldwide. It has been a major environmental
issue [10]. Thus, it is worth a while to develop strategies for the effective control of
trophic state related to nutrient loading. The aim of this paper is: (1) to characterise
water circulation, turbulent mixing, and oxygen concentration, induced by air
bubble injection; (2) to determine the accuracy of interphase-force and turbulence
A. Najafi-Nejad-Nasser () • S.S. Li • C.N. Mulligan

Concordia University, 1455 de Maisonneuve Blvd. W., Montreal, QC, H3G 1M8, Canada

278 A. Najafi-Nejad-Nasser et al.
closure methods through a comparison between computational results and available

laboratory data.
In phytoplankton growth and dynamics, P is a key element. It is the limiting
nutrient in fresh water systems. Correspondently, existing strategies for eutrophi-
cation control and/or remediation have mostly focused on reducing P load [5, 17].
Various efforts were made to reduce external sources of nutrients loading [12, 21].
However, in the 1970s, studies [13, 20] revealed that P released from lake bottom
sediments could continue as internal loading. Internal loading significantly delayed
the recovery of eutrophic lakes even after the control of external loading [14]. Under
specific conditions, internal loading contributed up to 80 % of total P in lakes [11].
For example, in Lake Ockeechobe in Florida, internal loading was in the same order
of magnitude as external sources [19]. Thus, it is of equal importance to be able to
effectively control the release of P from bottom sediments to the lake water.
One plausible technique to control internal loading is to introduce bubble
plumes to the lake water. Previous studies have identified some relevant parameters
associated with this technique [8, 9, 15, 16, 24]. Kim et al. [9] investigated
experimentally the effects of bubble size and diffusing area (geometric parameters)
on lake destratification. Destratification tends to promote air circulation and improve
the dissolved oxygen level. Kim et al. [9] suggested that destratification efficiency
is proportional to the bubble diffusing area and inversely proportional to the bubble
diameter and overall tank area. Rensen and Roig [15] studied the non-stationary
behaviour of flow by conducting experiments of 2D bubble plumes in a confined
tank. According to Imteaz and Aseada [8], the number of ports, air flow rate, and
bubbler starting time were important parameters for optimal bubbling operation.
Using the modelling approach, Sahoo and Luktenia [16] studied 1-D bubble
plumes. They reported that bubbles of close to 1 mm in radius gave a higher oxygen
transfer rate and mechanical efficiency than larger bubbles. Yum et al. [24] simulated
two-phase bubble plumes, using experimental data for model calibration/verifi-
cation. Their work led to the derivation of relationships between stratification
efficiency, plume spacing, and destratification number. Since the calibration data
are from a small tank under controlled environment, uncertainties about the
relationships’ suitability for application to field conditions exist. More advanced
modelling studies are needed to quantify the beneficial effects of bubble plumes on
water quality.
This paper tackles the problem by numerically solving the RANS equations on
unstructured finite volume mesh. The solution methods entail the use of suitable
models for turbulence closure. Previously, a number of investigators have assessed
the suitability of turbulence closure models for different applications [7, 18, 22].
Application of CFD Modelling to the Restoration of Eutrophic Lakes 279
2 Computational Model
2.1 Model Domain
The model domain is a cylinder (Fig. 1a) with a height h D 40 cm, and a diameter
FG D 50 cm. The top of the cylinder is open to air. The bottom has a circular hole
at the middle. This hole has a diameter D D 6 cm. Through this hole (inlet), air
bubbles of a given diameter are injected into the otherwise stagnant water contained
in the cylinder. This model domain is used because experimental results are available
to allow a direct comparison. For efficient computation of water and air bubble
motions, we consider three geometric configurations of computational domain: (1)
Computational domain A: a simple 2D plane cutting through the centreline of the
cylinder (Fig. 1c); (2) Computational domain B: an axis-symmetrical domain cutting
through the centreline (Fig. 1b); (3) Computational domain C: a 10ı wedge of the
cylinder (Fig. 1d). Domains A and B are covered by quadrilateral mesh generated
using a cell size of 1 mm. The mesh contains 401,802 and 199,562 computing nodes,
respectively. Domain C is covered by quadrilateral mesh created using a cell size of
5 mm. The total computing nodes are 33,580.
Fig. 1 A diagram of the model cylinder containing water. Water and air-bubble motions are
induced by the injection of air bubbles
2.2 Solution Method-Eulerian Approach
The turbulent motions of the liquid phase (water) and gas phase (air bubble) were
simulated with the Euler-Euler method using ANSYS Fluent. The motions are
described using separate momentum equations. The two phases are related through
a momentum exchange term. The interphase momentum transfer is due to interfacial
forces acting and interactions between water and air bubbles (such as lift force and
drag force) [3, 4, 23]. Hence, a proper solution for the bubble columns dependents on
the correct modelling of interphase forces and turbulence models. In this paper, we
compare the performance of different interphase force models and two turbulence
closure models (the k- model and the SST k-! model). All the simulations use
unsteady formulation. Phase Coupled SIMPLE (PC-SIMPLE) algorithm is used for
pressure velocity coupling. This algorithm is an extension of the SIMPLE algorithm
to multiphase problems. The velocity solutions are obtained in a segregated fashion,
and coupled by the liquid and gas phases [4].
2.3 Initial and Boundary Conditions
At time t D 0, imposed initial conditions are as follows: The free water surface is
located at the equilibrium position. The volume fraction of water ˛w is equal to one
below the free surface (for x2 D h). Water is stagnant or the velocity components
u1 and u2 are zero in the entire computational domain. Kinematic and dynamic
conditions are imposed at the boundaries of the computational domains (Fig. 1),
including
(1) Inlet (at the bottom of the cylinder)
(2) Outlet (on the top of the cylinder)
(3) Solid side walls of the cylinder
(4) Axis , OM in Fig. 1b (Runs 2–10, listed in Table 2)
(5) Symmetry (for 3D simulation, Run 11)
At the inlet (D=2 < x1 < D=2 and x2 = 0, Fig. 1), air bubbles of a given
diameter d (d < D) enter the domain continuously during the simulation time period
T. The direction of the entering velocity is upward, and the magnitude is uo (or u2 D
uo ; u1 D 0). At the inlet, the volume fraction of water ˛w is taken as zero.
At the outlet (at x2 D h, Fig. 1), fluids are exposed to the atmosphere.
Accordingly, the pressure equals to the atmospheric pressure. The volume fraction
of water ˛w is set to zero.
At the solid walls, no-slip condition is applied for computational domains A,
B, and C. For the three cases, the wall distance of the first cell from the walls is
yC D 0:33, yC D 0:45, and yC D 0:81, respectively. Thus, no-slip condition is
valid. This condition means that both the tangential and normal components of the
fluid velocity are set to zero.
2.4 Simulations
Model parameters and their values used in simulations are summarised in Table 1. A
total of 11 runs (Runs 1–11, Table 2) were carried out using the Eulerian approach.
Runs 1–6 and Runs 10–11 differ from each other in the choice of turbulence
dispersion force and lift force; these runs use the k- model for turbulence closure.
Runs 7–9 use the SSTk-! model.
The time period of all the simulations was t = 10.7 s. This is long enough since it
is more than two times of the advection time scale.
All the 11 runs produced finite volume solutions to the RANS equations. The
results are presented and discussed in the next section, along with comparisons with
available experimental data.
Under given conditions (Tables 1 and 2), Runs 1–11 produce water velocities.
As an example, velocity vectors and corresponding flow streamlines at a state of
equilibrium for Run 3 are plotted in Figs. 2 and 3, respectively. A strong jet is
seen to occur in the central region (Fig. 2), as a direct response to bubble injection.
Water motions are visible in the entire domain. The jet flow entrains water from both
Table 1 Control parameters and their values

Parameter Value Parameter Value
Time step (s) 0.001 Initial time (s) 0
Cell size for domains A and B (mm) 1 Simulation time period (s) 10.7
Air velocity at the inlet (m/s) 0.085 Number of time step 11,000
Water velocity at the inlet (m/s) 0 Pressure at outlet (Pa) 0
Flow rate at the inlet (m3 /s) 2:4 104 Surface diameter (m) 0.5
Bubble size at the inlet (mm) 3 Depth (m) 0.4
Air volume fraction at the inlet 1 Depth of water (m) 0.4
Convergence criteria 106 Inlet pipe diameter (m) 0.06
Table 2 Solution domain, water-bubble interaction method, and turbulence closure model
Run 1 2 3 4 5 6 7 8 9 10 11
Domain A B B B B B B B B B C
Drag SNa SN SN SN SN SN SN SN SN SN SN
Lift (–) (–) La L Ta T (–) T T T (–)
Dispersion (–) (–) LDBa Sa Ba LDB (–) LDB S S (–)
Closure k k k k k k SST SST SST k k
a
SST SSTk !, SN Schiller-Nauman, L Legendre, LDB Lopez-de-Bertodano, S Simonin, T
Tomiyama, B Burns et al. [4]
Fig. 2 Water velocity vectors for Run 3 (Table 2)
Fig. 3 Water flow streamlines for Run 3
sides in the lower water column and creates eddies (Fig. 3). These eddies produce
diverging flows from the centre in the upper water column. Water flow converges to
compensate the upward motion at the centre. These flow features are realistic. Also,
there are upward and downward motions on both the left and right sides of the water
body. These flow patterns would enhance renewal of bottom water with oxygenated
surface water.
For Run 3 (Fig. 2), the maximum velocity has a magnitude of nearly 5.6 times of
the initial velocity, uo , of bubbles entering the water column. The maximum velocity
has a magnitude of 5.4uo for Run 5, being the lowest among Runs 1–11. For the
other runs, the maximum velocity has a magnitude ranging from 5.5uo (for Run 4)
to 8.5uo (for Run 7).
The flow streamlines (Fig. 3) show clockwise and counterclockwise eddy

motions on the right and left sides of the water body. These eddy motions penetrate
the entire water depth, meaning that aeration is effective in producing exchange of
water masses. For Runs 1–11, the radius of significant influence is larger than eight
times the inlet radius D. Water circulation occurs over virtually the whole width of
the water body. In other words, artificial circulation can effectively be created by
injecting air bubbles.
The vertical component of water velocities at three selected heights (0.05, 0.1 and
0.2 m) above the bottom varies with radial distance, r, from the centre (Fig. 4a–c).
The velocity component has peak values at the centre (or r D 0), decreasing rapidly
with r. It drops to zero at r D 0:1 m for Runs 1, 2, 7 (not shown) and 11 (Fig. 4a–
c). The velocity component drops to zero at r D 0:12 m, and the flow direction
changes from upward to downward for Runs 3–6 and 8–10 (Fig. 4a–c, not shown
for Runs 1–7, 9, 10). In summary, Runs 3–6 and 8–10 predict upward flow within
0 < r < 0:12 m; Runs 1–2, 7 and 11 predict upward flow within 0 < r < 0:1; this is
true at different heights above the bottom. These predictions indicate that the results
are sensitive to model setup (Table 2). Experimental data [1] show similar features;
water velocity intensifies with increasing height above the inlet but weakens with
increasing radial distance.
Bubble rising velocity, ua , varies with r (Fig. 5a–c). The values are extracted from
the model results at model time of t D 10:7 s for the same heights as in Fig. 4a–c. ua
decreases from peak values at the centre with increasing r; this feature was observed
from experiments [1].
In Table 3, bubble rising velocities at different radial distances are compared with
experimental data [1]. Runs 11 produces the smallest relative error.
The results of air and water velocities for Run 3 are better than those for Run
2, especially at larger r distances (not shown). The reason is that the former run
considers the effect of interfacial forces. On average (at five different heights above
the bottom and different radial distances from the centreline), this consideration has
improved the predictions of ua by 7.3 %, uw by 6.8 %, and VFa by 62 %. At the
specific height z = 0.3 m, the consideration has reduced the average relative error in
the predictions of ua from 18.8 % to 3.2 %.
a b c
uw (m/s)
0.4 0.4 0.4
0 0 0
0 0.1 0.2 r (m) 0 0.1 0.2 r (m) 0 0.1 0.2

r (m)
Run8 EXP Run11
Fig. 4 Water velocity .uw / distribution with radial distance, r. (a) z D 0.05 (m). (b) z D 0.1 (m).
(c) z D 0.2 (m)
a b c
1 1 1
ua (m/s)
0.5 0.5 0.5
0 0 0
0 0.1 0.2 r (m) 0 0.1 0.2 r (m) 0 0.1 0.2
r (m)
Run8 EXP Run11
Fig. 5 Distribution of air velocity .ua / with radial distance, r. (a) z D 0.05 (m). (b) z D 0.1 (m).
(c) z D 0.2 (m)
Table 3 Relative errors of computed air velocity (ua ), and average of water turbulent kinetic
energy (K)
Run 1 2 3 4 5 6 7 8 9 10 11
ua (%) 35:3 23:6 19:6 20:1 19:8 17:8 25:9 14:4 20:1 16:2 12:0
K(J/kg) 0:019 0:007 0:006 0:006 0:006 0:006 0:005 0:004 0:004 0:006 0:002
a b c
0.04
K (J/kg)
0.01 0.02
0.02
0.005 0.01
0 0 0
0 0.1 0.2 r (m) 0 0.1 0.2 r (m) 0 0.1 0.2
r (m) Run8 Run11
Fig. 6 Distribution of water turbulent kinetic energy .K/ with r. (a) z D 0.05 (m). (b) z D 0.1 (m).
(c) z D 0.2 (m)
Variations in turbulent kinetic energy, K, of water flow with radial distance r at

three different heights from the bottom are illustrated in Fig. 6a–c. As bubbles rise,
part of their kinetic energy is converted into turbulent kinetic energy. The maximum
K values occur slightly off the centreline. K increases with increasing height. At
a given height, it decreases with increasing r. Overall, air bubbles entering into
the water body cause water motions and turbulence. Water and bubble motions
help prevent sedentary condition, which is known to contribute to water-quality
degradation. The averages (over the whole water body) of K values are listed in
Table 3.
Distributions of predicted air volume fraction, VFa , with r are shown in Fig. 7a–
c. Values of VFa are extracted from the model results for Runs 1 to 11 at model time
of t D 10:7s for the same heights as in Fig. 4a–c. The snapshots show that at a given
height, VFa decreases from peak values with increasing r. The peak values occur
slightly off the centre for Runs 1–11 (Fig. 7a–c, not shown for Runs 1–7, 9, 10).
a b c
0.3
VF3 (m/s)
0.2 0.2
0.2
0.1 0.1
0.1
0 0 0
0 0.05 r (m) 0 0.05 r (m) 0 0.05
r (m)
Run8 EXP Run11
Fig. 7 Distribution of air volume fraction, VFa , with radial distance r. (a) z D 0.05 (m). (b) z D 0.1
(m). (c) z D 0.2 (m)
A strong lake stratifications known to prohibit fluid motions and mixing in the
vertical. However, the consideration of density stratification effects is beyond the
scope of this paper.
4 Conclusions
This paper discusses artificial circulation in lakes, induced by injecting air bubbles,
for the control of eutrophication. We have simulated artificial circulation using CFD
modelling techniques, and reached the following conclusions: (1) The injection
triggers turbulent motions of water and bubbles, which feature a strong upward
flow above the injection location and energetic turbulent eddies on both sides of
the upward flow. (2) These large scale eddies enhance renewal of bottom water with
oxygenated surface water, which helps improve the dissolved oxygen level in the
lower water column. (3) Air bubbles entering the water column produce turbulent
kinetic energy; this source of energy will maintain small scale eddy motions in the
lake water, with beneficial mixing effects. (4) The dissolved oxygen level in the lake
water is improved as a direct response to air bubbles entering the water column. (5)
From the computational perspective, a 2D axisymmetric computational domain is
recommended for two reasons: (a) it offers high computational efficiency, relative to
a 3D domain; and (b) it appears to be sufficient to capture measured characteristics
of air and water velocities. (6) Model predictions of water velocity, air velocity, air
volume fraction agree well with experimental data. The best agreement is obtained
with the use of the Schiller-Nauman model for drag, the Tomiyama model for lift,
the Lopez-de-Bertodano model for turbulent dispersion, and the k model for
turbulence closure.
Acknowledgements This study received financial support through Discovery Grants held by S.S.
Li and C.N. Mulligan
References
1. Anagbo, P.E., Brimacombe, J.K.: Plume characteristics and liquid circulation in gas injection
through a porous plug. Metall. Trans. B 21B, 637–648 (1990)
2. Ansari, A.A., Gill, S.S., Lanza, G.R., Rast, W.: Eutrophication: Causes, Consequences and
Control. Springer, Dordrecht/Heidelberg/London/New York (2011)
3. Azzopardi, B., Zhao, D., Yan, Y., Morvan, H., Mudde, R.F., Lo, S.: Hydrodynamics of Gas-
liquid Reactors: Normal Operation and Upset Conditions. Wiley. Chichester, United Kingdom,
(2011)
4. ANSYS: Fluent 6.3 User’s Guide (2006)
5. Dai, L., Pan, G.: The effects of red soil in removing P from water column and reducing P
release from sediment in Lake Taihu. Water Sci. Technol. 65(5), 1052–1058 (2014)
6. Environment Canada: Canadian guidance framework for the management of P in freshwater
systems. Ecosyst. Health: Sci.-Based Solut. 1–8, 1–114 (2004)
7. Hjarne, J., Chernoray, V., Larsson, J., Lofdahl, L.: Numerical validations of secondary flows
and loss development downstream of a highly loaded low pressure turbine outlet guide Vane
cascade. ASME Turbo Expo 2007, 723–733 (2007)
8. Imteaz, M.A., Asaeda, T.: Artificial mixing of lake water by bubble plume and effects of
bubbling operations on algal bloom. Water Res. 34(6),1919–1929 (2000)
9. Kim, S.H., Kim, J.Y., Park, H., Park, N.S.: Effects of bubble size and diffusing area on
destrafication efficiency in bubble plumes of two-layer stratification. ASCE J. Hydraul. Eng.
136, 106–115 (2010)
10. Matsui, S., Ide, S., Ando, M.: Lake reservoirs: reflecting waters of sustainable use. Water Sci.
Technol. 32(7),221–224 (1955)
11. Penn, M.R., Auer, M.T., Doerr, S.M., Driscoll, C.T., Brooks, C.M., Effler, S.W.: Seasonality
in P release rates from the sediments of a hypereutrophic lake under a matrix Of Ph and redox
conditions. Can. J. Fish. Aquat. Sci. 57, 1033–104 (2000)
12. Phillips, G., Bramwell, A., Pitt, J., Stansfield, J., Perrow, M.: Practical application of 25 Years’
research into the management of shallow lakes. Hydrobiologia 395/396 61–76 (1999)
13. Qunhe, W., Renduo, Z., Shan, H., Hengjun, Z.: Effects of bacteria on nitrogen and P release
from river sediment. Environ. Sci. 20, 404–412 (2008)
14. Reddy, K.R., Q’connor, G.A., Schelske, C.L.: P Biogeochemistry of Subtropical Ecosystems,
1st edn., p. 101. CRC Press, Boca Raton (1999)
15. Rensen, J., Roig, V.: Experimental study of the unsteady structure of a confined bubble plume.
Int. J. Multiph. Flow 27, 1431–1449 (2001)
16. Sahoo, G.B., Luketina, D.: Modeling of bubble plume design and oxygen transfer for reservoir
restoration. Water Res. 37, 393–401 (2003)
17. Schauser, I., Chorus, I.: Assessment of internal and external lake restoration measures for two
Berlin lakes. Lake Reserv. Manag. 23, 366–376 (2007)
18. Schuler, M., Zehnder, F., Weigand, B., von Wolfersdorf, J., Neumann, S.O.: The effect of
turning vanes on pressure loss and heat transfer of a Ribbed rectangular two-pass internal
cooling channel. ASME J. Turbomach. 133(2), 0211017–(1–10) (2011)
19. Sharpley, A.N., Chapra, S.C., Wedepohl, R., Sims, J.T., Daniel, T.C., Reddy, K.R.: Managing
agricultural P for protecting of surface waters: issues and options. J. Environ. Qual. 23(3),
437–451 (1994)
20. Sondergaard, M., Jensen, J.P., Jappesen, E.: Role of sediment and internal loading of P in
shallow lakes. Hydrobiologia 506–509, 135–145 (2003)
21. Van Der Molen, D.T., Boers, P.C.M.: Eutrophication Control in the Netherlands. Hydrobiologia
136, 403–409 (1999)
22. Wang, X., Naji, H., Mezrhab, A.: Computational investigation of different models when
predicting airflow in an enclosure. ASME Press. Vessels Piping Div. Conf. 4, 179–188 (2008)
23. Yeoh, G.H., Tu, J.: Computational Techniques for Multiphase Flows: basics and applications.
Elsevier, Butterworth-Heinmann, United Kingdom, (2009)
24. Yum, K., Kim, S.H., Park, H.: Effects of plume spacing and flowrate on destrafication efficiency
of air diffusers. Water Res. 42, 3249–3262 (2008)
On the Co-infection of Malaria
and Schistosomiasis
Kazeem O. Okosun and Robert Smith?
Abstract Mathematical models for co-infection of diseases (that is, the simulta-
neous infection of an individual by multiple diseases) are sorely lacking in the
literature. Here we present a mathematical model for the co-infection of malaria and
schistosomiasis. We derive reproduction numbers for malaria and schistosomiasis
independently, then combine these to determine the effects of disease interactions.
Sensitivity indices show that malaria infection may be associated with an increased
rate of schistosomiasis infection. However, schistosomiasis infection is not asso-
ciated with an increased rate of malaria infection. Therefore, whenever there is
co-infection of malaria and schistosomiasis in the community, our model suggests
that control measures for each disease should be administered concurrently for
effective control.
1 Introduction
Malaria and schistosomiasis often overlap in tropical and subtropical countries,

imposing tremendous disease burdens [4, 8, 14]. The substantial epidemiological
overlap of these two parasitic infections invariably results in frequent co-infections
[7, 18]. The challenges facing the development of a highly effective malaria vaccine
have generated interest in understanding the interactions between malaria and co-
endemic helminth infections, such as those caused by Schistosoma, that could
impair vaccine efficacy by modulating host-immune responses to Plasmodium
infection and treatment [13, 14]. Both malaria and schistosomiasis are endemic to
most African nations. However, the extent to which schistosomiasis modifies the
rate of febrile malaria remains unclear.
K.O. Okosun
Department of Mathematics, Vaal University of Technology, Vanderbijlpark, South Africa
R. Smith? ()
University of Ottawa, Ottawa, ON, Canada

290 K.O. Okosun and R. Smith?
Mathematical modelling has been an important tool in understanding the dynam-

ics of disease transmission and also in the decision-making processes regarding
intervention mechanisms for disease control. For example, Ross [12] developed
the first mathematical models of malaria transmission. His focus was on mosquito
control, and he showed that, for the disease to be eliminated, the mosquito
population should be brought below a certain threshold. Another classical result
is due to Anderson and May [1], who derived a malaria model with the assumption
that acquired immunity in malaria is independent of exposure duration.
There is an urgent need for co-infection models for infectious diseases, particu-
larly those that mix neglected tropical diseases with “the big three” (HIV, TB and
malaria) [8]. Recently, the authors in [10] proposed a model for schistosomiasis and
HIV/AIDS co-dynamics, while the co-infection dynamics of malaria and cholera
were studied in [11]. However, few studies have been carried out on the co-
infection of schistosomiasis. To the best of our knowledge, no work has been done
to investigate the malaria–schistosomiasis co-infection dynamics.
In this paper, we formulate and analyse an SIR (susceptible, infected and
recovered) model for malaria–schistosomiasis co-infection, in order to understand
the effect that controlling for one disease may have on the other.
2 Model Formulation
Our model subdivides the total human population, denoted by Nh , into subpop-
ulations of susceptible humans Sh , individuals infected only with malaria Im ,
individuals infected with only schistosomiasis Isc , individuals infected with both
malaria and schistosomiasis Cms , individuals who have recovered from malaria Rm
and individuals who have recovered from schistosomiasis Rs . The total mosquito
vector population, denoted by Nv , is subdivided into susceptible mosquitoes Sv and
mosquitoes infected with malaria Iv . Similarly, the total snail vector population,
denoted by Nsv , is subdivided into susceptible snails Ssv and snails infected with
schistosomiasis Isv . Thus Nh D Sh C Im C Is C Cms C Rs C Rm , Nv D Sv C Iv and
Nsv D Ssv C Isv :
The model is given by the following system of ordinary differential equations.
Sh0 D h C "Rs C ˛Rm ˇ1 Sh 1 Sh h Sh

Im0 D ˇ1 Sh 1 Im . C h C /Im
0
Isc D 1 Sh ˇ1 Isc .! C h C /Isc
0
Cms D ˇ1 Isc C 1 Im .ı C h C C /Cms
R0m D Im .˛ C h /Rm C ıCms
R0s D !Isc ." C h /Rs C .1 /ıCms
On the Co-infection of Malaria and Schistosomiasis 291
Sv0 D v ˇ2 Sv v Sv
Iv0 D ˇ2 Sv v Iv
0
(1)
Ssv D s 2 Ssv sv Ssv
0
Isv D 2 Ssv sv Isv ;
with the transmission rates given by
ˇh Iv Isv ˇv .Im C Cms / s .Isc C Cms /

ˇ1 D ; 1 D ; ˇ2 D ; 2 D :
Nh Nh Nh Nh
Birth rates for humans, mosquitoes and snails are, respectively, h , v and
sv , while the corresponding mortality rates are h , v and sv . Here is the
schistosomiasis-related death rate and is the malaria-related death rate. The
immunity-waning rates for malaria and schistosomiasis are ˛ and " respectively,
while the recovery rates from malaria, schistosomiasis and co-infection are , !
and ı respectively. The term ı accounts for the portion of co-infected individuals
who recover from malaria, while .1 /ı accounts for co-infected individuals
who recover from schistosomiasis; thus (satisfying 0 1) represents the
likelihood of individuals to recover from malaria first. Note that all parameters
might in practice vary with time; however, we shall take variations in our critical
parameters into account with a sensitivity analysis.
3 Analysis of Malaria–Schistosomiasis Co-infection Model
The malaria–schistosomiasis model (1) has a disease-free equilibrium, given by

!
h v s
E 0 D .Sh ; Im ; Isc

; Cms ; Rm ; Rs ; Sv ; Iv ; Ssv

; Isv /D ; 0; 0; 0; 0; 0; ; 0; ;0 :
h v sv
The linear stability of E 0 can be established using the next-generation method

[17] on the system (1). It follows that the reproduction number of the malaria–
schistosomiasis model (1), denoted by R msc , is given by
R msc D maxfR sc ; R 0m g ;
where
s
v ˇh ˇv h
R 0m D
h 2v . C C h /
s
s s h
R sc D :
h .m C ! C h /2sv
Note that the reproduction number produced by the next-generation method pro-
duces a threshold quantity and not necessarily the average number of secondary
infections [9]. We thus have the following theorem.
Theorem 1 The disease-free equilibrium E 0 is locally asymptotically stable when-
ever R msc < 1 and unstable otherwise.
3.1 Impact of Disease Interactions
To analyse the effects of schistosomiasis on malaria and vice versa, we begin by

expressing R sc in terms of R 0m . We solve for h to get
D1 R 20m
h D ;
D2 D3 R 20m
where
D1 D h 2v . C / ; D2 D v ˇh ˇv ; D3 D h 2v :
Substituting into the expression for R sc , we obtain

s
s s D1 R 20m
R sc D : (2)
Œ. C !/D2 C .D1 . C !/D3 /R 20m h 2sv
Differentiating R sc with respect to R 0m leads to

r
s 1 0m D R2
. C !/D2 Œ.C!/D C.D s.C!/D
@R sc 2 1
2 2
3 /R0m h sv
D : (3)
@R 0m Œ. C !/D2 R 0m C .D1 . C !/D3 /R 30m
Similarly, expressing h in terms of R sc , we get
D4 R 2sc
h D ; (4)
D5 D6 R 2sc
where
D4 D h 2sv . C !/ ; D5 D s s ; D6 D h 2sv :

Substituting into the expression for R 0m , we obtain

s
D4 ˇh ˇv v R 2sc
R 0m D : (5)
Œ. C /D5 C .D4 . C /D6 /R 2sc h 2sv
3.2 Sensitivity Indices of Rsc when Expressed in Terms of R0m
We next derive the sensitivity of Rsc in (2) (i.e., when expressed in terms of R0m )
to each of the 13 different parameters. However, the expression for the sensitivity
indices for some of the parameters are complex, so we evaluate the sensitivity
indices of these parameters at the baseline parameter values as given in Table 1.
Since the effect of immunity in the control of re-infection is not entirely known
[6], we have assumed the schistosomiasis immunity waning rate. Due to a lack of
data in the literature, assumptions were made for the recovery rate of co-infected
individuals, ı, recovery rate of schistosomiasis-infected individuals, !, and the rate
of recovery from malaria for co-infected individuals, .
Table 1 Parameters in the co-infection model

Parameter Description value Ref
Malaria-induced death 0.05–0.1 day1 [16]
ˇh Malaria transmissibility to humans 0.034 day1 [2]
ˇv Malaria transmissibility to mosquitoes 0.09 day1 [2]
Schistosomiasis transmissibility to 0.406 day1 [15]
humans
s Schistosomiasis transmissibility to 0.615 day1 [3]
snails
h Natural death rate in humans 0.00004 day1 [2]
v Natural death rate in mosquitoes 1/15–0.143 day1 [2]
sv Natural death rate in snails 0.000569 day1 [3, 15]
˛ Malaria immunity waning rate 1/(60*365) day1 [2]
" Schistosomiasis immunity waning rate 0.013 day1 Assumed
h Human birth rate 800 people/day [3]
v Mosquitoes birth rate 1000 mosquitoes/day [2]
s Snail birth rate 100 snails/day [5]
ı Recovery rate of co-infected individual 0.35 day1 Assumed
! Recovery rate of 0.0181 day1 Assumed
schistosomiasis-infected individual
Recovery rate of malaria-infected 1/(2*365) day1 [2]
individual
Co-infected proportion who recover 0.1 Assumed
from malaria only
Schistosomiasis-induced death 0.0039 day1 [3]
Table 2 Sensitivity indices of Rsc expressed in terms of R0m

Sensitivity index Sensitivity index
Parameter Description if R0m < 1 if R0m > 1
1 sv Snail mortality 1 1
2 v Mosquito mortality 0:56 0:07
3 s Schistosomiasis transmissibility to snails 0:5 0:5
4 s Snail birth rate 0:5 0:5
5 ˇh Malaria transmissibility to humans 0:28 0:03
6 ˇv Malaria transmissibility to mosquitoes 0:28 0:03
7 v Mosquito birth rate 0:28 0:03
8 h Human birth rate 0:22 0:47
9 Malaria-induced death 0:12 0:31
10 ! Recovery from schistosomiasis 0:10 0:26
11 m Schistosomiasis-induced death 0:02 0:05
12 Recovery from malaria rate 0:003 0:0084
The sensitivity index of Rsc with respect to , for example, is
@Rsc
Rsc D 0:5 : (6)
@ Rsc
The detailed sensitivity indices of Rsc resulting from the evaluation of the other
parameters of the model are shown in Table 2.
Table 2 shows the parameters, arranged from the most sensitive to the least.
For R0m < 1, the most sensitive parameters are the snail mortality rate, the
mosquito mortality rate, the transmissibility of schistosomiasis to snails and the
snail birth rate (sv , v , s and s , respectively). Since Rsvsc D 1, increasing
(or decreasing) the snail mortality rate sv by 10 % decreases (or increases) Rsc by
10 %; similarly, increasing (or decreasing) the mosquito mortality rate, v , by 10 %
increases (or decreases) Rsc by 5:6 %. In the same way, increasing (or decreasing)
the transmissibility of schistosomiasis to snails, s , increases (or decreases) Rsc
by 5 %. As the malaria parameters ˇh , ˇv and v increase/decrease by 10 %, the
reproduction number of schistosomiasis, Rsc , decreases by 2:8 % in all three cases.
For R0m > 1, the most sensitive parameters are the snail mortality rate, the rate
of a snail getting infected with schistosomiasis, the snail birth rate, the human birth
rate, malaria-induced death and recovery from schistosomiasis (sv , s , s , h , ,
!, respectively). Since Rssc D 0:5, increasing (or decreasing) by 10 % increases (or
decreases) Rsc by 5 %; similarly, increasing (or decreasing) the recovery rate, !, by
10 % increases (or decreases) Rsc by 2:6 %. Also, as the malaria parameters ˇh , ˇv
and v increase/decrease by 10 %, the reproduction number of schistosomiasis, Rsc ,
decreases by only 0:3 % in all three cases.
Table 3 Sensitivity indices of R0m expressed in terms of Rsc

Sensitivity index Sensitivity index
Parameter Description if Rsc < 1 if Rsc > 1
1 ˇv Malaria transmissibility to mosquitoes 0:5 0:5
2 v Mosquito birth rate 0:5 0:5
3 Schistosomiasis transmissibility to humans 0:5 0:5
4 s Schistosomiasis transmissibility to snails 0:5 0:5
5 s Snail birth rate 0:5 0:5
6 Malaria-induced death 0:49 0:49
7 ! Recovery from schistosomiasis 0:41 0:41
8 m Schistosomiasis-induced death 0:09 0:09
9 Recovery from malaria 0:01 0:01
10 sv Snail mortality 0:0000002 0:000007
11 h Human birth rate 0:0000001 0:000004
It is clear that Rsc is sensitive to changes in R0m . That is, the sensitivity of Rsc to
parameter variations depends on R0m ; whenever, R0m < 1, Rsc is less sensitive to the
malaria parameters.
3.3 Sensitivity Indices of R0m when Expressed in Terms of Rsc
Similar to the previous subsection, we derive the sensitivity of R0m in (5) (i.e. when
expressed in terms of Rsc ) to each of the different parameters. The sensitivity index
of R0m with respect to ˇh , for example, is
@R0m ˇh
ˇRh0m D 0:5 : (7)
@ˇh R0m
The detailed sensitivity indices of R0m resulting from the evaluation to the other
parameters of the model are shown in Table 3. It is clearly seen from Table 3
that the malaria reproduction number, R0m , is not sensitive to any variation in the
schistosomiasis reproduction number Rsc .
4 Numerical Simulations
Table 1 lists the parameter descriptions and values used in the numerical simulation
of the co-infection model.
(a) (b) = 0.4
3 = 0.4 70
= 0.2
= 0.1
Malaria Infected Individuals
2.5
60
Co Infected Individuals
2
= 0.3 50
1.5
40
= 0.2
1
30
0.5
20
0
0 20 40 60 80 100 0 20 40 60 80 100
Time (days) Time (days)
4
x 10
(c) 2 (d) 600
h
= 0.234
1.8
h
= 0.234
500
Schistosomiasis Infected Individuals
1.6
1.4 = 0.134
h
= 0.134 400
h
1.2
1
300
= 0.034
h
0.8
0.6 200
0.4 = 0.034
h
100
0.2
0 20 40 60 80 100 0 20 40 60 80 100
Fig. 1 Simulations of the malaria–schistosomiasis model showing the effect of varying transmis-
sion rates
Figure 1a,b shows the effect of varying the schistosomiasis transmission param-
eter on the number of individuals infected with malaria, Im , and the number of
co-infected individuals, Cms . This illustrates that effective control of schistosomiasis
would enhance the control of malaria. Conversely, Fig. 1c,d shows the effect of
varying the malaria transmission parameter ˇh on the number of individuals infected
with schistosomiasis, Isc , and the number of co-infected individuals. This illustrates
that effective control of malaria would enhance control of co-infection but have only
minimal effect on schistosomiasis prevalence.
Figure 2 shows the effect of varying the death rate of mosquitoes v (for exam-
ple, through spraying) on the number of individuals infected with schistosomiasis
and the number of co-infected individuals. As the mosquitoes are controlled, the
number of individuals infected with malaria falls dramatically, as does the number
of co-infected individuals, while the number of schistosomiasis-infected individuals
only decreases slightly.
(a) 4.5 (b)

100
4 90
μv = 0.07
80
Malaria Infected Individuals
3.5
70
3
60
2.5 μv = 0.09
μv =0.07
50
2
40
1.5 μ =0.09
v
30
μ = 0.143
1 v
μ =0.143
v 20
0.5
10
0
0 20 40 60 80 100 0 20 40 60 80 100
(c) 11000 μ =0.09

v
10000
μ =0.07
v
Schistosomiasis Infected Individuals
9000
8000
7000
μ =0.143
v
6000
5000
4000
3000
2000
1000
0 20 40 60 80 100
Time (days)
Fig. 2 Simulations of the malaria–schistosomiasis model showing the effect of varying the
mosquito death rate
5 Concluding Remarks
In this paper, we formulated and analysed a deterministic model for the transmission
of malaria–schistosomiasis co-infection. We derived basic reproduction numbers for
each infection and determined the sensitivity of each reproduction number to all
parameters. Our analysis shows that malaria infection may be associated with an
increased rate of schistosomiasis infection. However, in our model, schistosomiasis
infection is not associated with an increased rate of malaria infection. Therefore,
whenever there is co-infection of malaria and schistosomiasis in the community,
our model suggests that control measures for both diseases should be administered
concurrently for effective control.
Acknowledgements The authors are grateful to an anonymous reviewer whose comments

enhanced the manuscript. OKO acknowledges the Vaal University of Technology Research Office
and the National Research Foundation (NRF), South Africa, through the KIC Grant ID 97192 for
the financial support to attend and present this paper at the AMMCS-CAIM 2015 meeting. RS?
is supported by an NSERC Discovery Grant. For citation purposes, please note that the question
mark in “Smith?” is part of the author’s name.
References
1. Anderson, R.M., May, R.M.: Infectious Diseases of Humans: Dynamics and Control. Oxford
University Press, Oxford (1991)
2. Blayneh, K.W., Cao, Y., Kwon, H.D.: Optimal control of vector-borne diseases: treatment and
prevention. Discret. Contin. Dyn. Syst. Ser B 11(3), 587–611 (2009)
3. Chiyaka, E.T., Magombedze, G., Mutimbu, L.: Modelling within host parasite dynamics of
schistosomiasis. Comput. Math Methods Med. 11(3), 255–280 (2010)
4. Doumbo, S., Tran, T.M., Sangala, J., Li, S., Doumtabe, D., et al.: Co-infection of long-term
carriers of plasmodium falciparum with schistosoma haematobium enhances protection from
febrile malaria: a prospective cohort study in Mali. PLoS Negl. Trop. Dis. 8(9), e3154 (2014)
5. Feng, Z., Li, C.-C., Milner, F.A.: Schistosomiasis models with two migrating human groups.
Math. Comput. Model. 41(11–12), 1213–1230 (2005)
6. Fulford, A.J.C., Butterworth, A.E., Dunne, D.W., Strurrock, R.F., Ouma, J.H.: Some mathemat-
ical and statistical issues in assessing the evidence for acquired immunity to schistosomiasis.
In: Isham, V., Medley, G. (eds.) Models for Infectious Human Diseases, pp. 139–159.
Cambridge University Press, Cambridge (1996)
7. Hotez, P.J., Molyneux, D.H., Fenwick, A., Ottesen, E., Ehrlich, S.S., et al.: Incorporating a
rapid-impact package for neglected tropical diseases with programs for HIV/AIDS, tuberculo-
sis, and malaria. PLoS Med. 3, e102 (2006)
8. Kealey, A., Smith?, R.J.: Neglected tropical diseases: infection, modelling and control.
J. Health Care Poor Underserv. 21, 53–69 (2010)
9. Li, J., Blakeley, D., Smith?, R.J.: The Failure of R0 . Comput. Math Methods Med. 2011, Article
ID 527610 (2011)
10. Mushayabasa, S., Bhunu, C.P.: Modeling Schistosomiasis and HIV/AIDS co-dynamics. Com-
put. Math. Methods Med. 2011, Article ID 846174 (2011)
11. Okosun, K.O., Makinde, O.D.: A co-infection model of malaria and cholera diseases with
optimal control. Math. Biosci. 258, 19–32 (2014)
12. Ross, R.: The Prevention of Malaria, 2nd edn. Murray, London (1911)
13. Salgame, P., Yap, G.S., Gause, W.C.: Effect of helminth-induced immunity on infections with
microbial pathogens. Nat. Immunol. 14, 1118–1126 (2013)
14. Semenya, A.A., Sullivan, J.S., Barnwell, J.W., Secor, W.E.: Schistosoma mansoni Infection
impairs antimalaria treatment and immune responses of rhesus macaques infected with
mosquito-borne plasmodium coatneyi. Infect. Immun. 80(11), 3821–3827 (2012)
15. Spear, R.C., Hubbard, A., Liang, S., Seto, E.: Disease transmission models for public health
decision making: toward an approach for designing intervention strategies for Schistosomiasis
japonica. Environ. Health Perspect. 10(9), 907–915 (2002)
16. Smith?, R.J., Hove-Musekwa, S.D.: Determining effective spraying periods to control malaria
via indoor residual spraying in sub-saharan Africa. Hindawi Publishing Corporation. J. Appl.
Math. Dec. Sci. 2008, Article ID 745463, 19p (2008)
17. van den Driessche, P., Watmough, J.: Reproduction numbers and sub-threshold endemic
equilibria for compartmental models of disease transmission. Math. Biosci. 180, 29–48 (2002)
18. Yapi, R.B., HRurlimann, E., Houngbedji, C.A., Ndri, P.B., SiluKe, K.D., et al.: Infection and co-
infection with helminths and plasmodium among school children in Côte d’Ivoire: results from
a national cross-sectional survey. PLoS Negl. Trop. Dis. 8(6), e2913 (2014)
A Discrete-Continuous Modeling Framework
to Study the Role of Swarming in a
Honeybee-Varroa destrutor-Virus System
Vardayani Ratti, Peter G. Kevan, and Hermann J. Eberl
Abstract In this paper, we present a general discrete-continuous modeling frame-

work to study the effect of swarming on the dynamics of a honeybee colony
infested with varroa mite and Acute Bee Paralysis Virus (ABPV) . Two scenarios are
studied under which swarming takes place i.e., swarming due to overcrowding and
swarming at fixed time intervals. For this purpose, we use an existing mathematical
model in the literature. The dependent variables in the model are uninfected bees,
infected bees, virus carrying mites and total mites that infest the colony. The model
is studied in variable coefficients, in particular, step functions with each season as
a constant in time. It is observed that the percentage of healthy bees leaving with
the swarm has a great impact on the strength and survival of the parent colony. A
colony, that otherwise dies off due to virus, survives as a properly working colony
if the percentage of the mites leaving the parent colony is above a critical value.
1 Introduction
A honey bee colony consists of a single reproductive queen, 20,000–60,000 adult

workers, 10,000–30,000 individuals at brood stage (eggs, larvae and pupae), and
several hundred drones [13]. A large population of workers is needed to carry out the
tasks of the bee colony, including cleaning, brood rearing, guarding the hive etc. In
order to maintain a high reproduction level in the colony, the bees start preparing for
swarming. Swarming is the natural method of reproduction of honeybee colonies.
In the process of swarming, the original single colony reproduces to two or more
colonies. During swarming, almost 50–70 % of the worker bees leave the parent
colony with the queen (old mated queen in case of the first swarm) to a new site
[3, 14]. The colony starts preparing for the swarm 1 month before the swarm is
issued. Swarming normally takes place in mid-spring [14]. The timing of swarming
varies with seasons and location of the hives. Since swarming affects pollination
and honey production, it has an impact on agriculture and economy as well. The
V. Ratti () • P.G. Kevan • H.J. Eberl

University of Guelph, 50 Stone Road E, Guelph, ON, N1G 2W1, Canada

300 V. Ratti et al.
main symptom of swarming is the preparation of brood cups and queen rearing.
The causes of swarming could be (i) large colony size, (ii) high proportion of
young worker bees, (iii) reduced queen pheromone due to overcrowding, and (iv)
abundance of pollen and nectar leading to overcrowding. There is a common cause
behind all the symptoms i.e., overcrowding in the colony.
When a swarm issues, the parasites present in the parent colony are divided
among the parent colony and the new colony formed after swarming [3]. This
may reduce the disease infestation in the parent colony. We investigate the effect
of swarming on the colony infested with a parasite Varroa destructor and Acute
Bee Paralysis Virus. Varroa destructor is an ectoparasitic mite that not only feeds
on the bees but also carries and transmits fatal viruses in the colony. The mite feeds
on bees’ haemolymph by piercing their inter-segmental membrane and transmits the
virus while feeding on them. When a virus carrying mite feeds on an infected bee,
it releases the virus into the bee’s haemolymph. Thus, the uninfected bee becomes
infected and mite becomes virus free. When a virus free mite feeds on an infected
bee, it begins to carry virus. There are 20 bee viruses known so far, out of which
at least 14 viruses are reported to be associated with mites [5, 9]. These viruses
have different routes of transmission and different levels of virulence. One of the
most common virus is the Acute Bee Paralysis Virus. It has also been implicated
in colony losses [4, 5, 15]. The bees infected by this virus are unable to fly, loose
their body hair and tremble uncontrollably. Unlike the Deformed Wing Virus, where
brood infested with the virus develop into an adult sick bee, the brood infected with
ABPV does not survive to the adult stage and dies immediately.
There have been several SIR-type mathematical models developed for honeybee-
varroa mite-virus systems [2, 11–13]. There have also been models on some
aspects of swarming such as the use of bee dance, design of nest selection and
decision-making processes [1, 8, 10]. However, none of these models studied the
combined effect of swarming and the varroa-virus infestation in a honeybee colony.
In this paper, we provide a general framework of difference-differential equations
to investigate how swarming affects the fate of the honeybee colony infested with
varroa mites and virus. We also use numerical simulations to study (i) swarming
due to overcrowding (ii) swarming after a fixed time interval mimicking the natural
death cycle of a queen bee. We vary the proportion of healthy bees leaving the
parent colony and observe its effect on the dynamics of the colony. We numerically
calculate a critical value of the proportion of mites leaving the parent colony below
which the parent colony dies off and above which it survives.
2 Model Equations
The underlying model of honeybee-varroa-ABPV disease dynamics from [11] is:
dm y x
D ˇ1 .M m/ ˇ2 m (1)
dt xCy xCy
Swarming in a Honeybee Colony 301
dx x
D g.x/h.m/ ˇ3 m d1 x 1 Mx (2)
dt xCy
dy x
D ˇ3 m d2 y 2 My (3)
dt xCy

dM M
D rM 1 (4)
dt ˛.x C y/
The parameters are assumed to be non-negative. Because the size of the bee
colony and the life span of bees vary drastically with seasons, the parameters are
assumed to be seasonally varying. In particular, we assume the parameters to be
periodic functions of time with a period T; in practice T D 1 year.
The parameter in (2) is the maximum birth rate, specified as the number of
worker bees emerging as adults per day.
The function g.x/ expresses that a sufficiently large number of healthy worker
bees is required to care for the brood. We think of g.x/ as a switch function. If x
falls below a critical value, which may seasonally depend on time, essential work
in the maintenance of the brood cannot be carried out anymore and no new bees are
born. If x is above this value, the birth of bees is not hampered. Thus g.0; / D 0,
dg.0/
dx 0, limx!1 g.x/ D 1. A convenient formulation of such switch like behavior
is given by the sigmoidal Hill function
xn
g.x/ D (5)
K n C xn
where the parameter K is the size of the bee colony at which the birth rate is half
of the maximum possible rate and the integer exponent n > 1. If K D 0 is chosen,
then the bee birth terms of the original model of [13] is recovered. Then the brood is
always reared at maximum capacity, independent of the actual bee population size,
because g.x/ 1.
The function h.m/ in (2) indicates that the birth rate is affected by the presence
of mites that carry the virus. This is in particular important for viruses like ABPV
that kill infected pupae before they develop into bees. The function h.m/ is assumed
to decrease as m increases, h.0/ D 1, dm dh
.m/ < 0 and limm!1 h.m/ D 0; [13]
suggests that this is an exponential function h.m/ emk , where k is non-negative.
We will use this expression in the computer simulations later on.
The parameter ˇ1 in (1) is the rate at which mites that do not carry the virus
acquire it. The rate at which infected mites lose their virus to an uninfected host
is ˇ2 . The rate at which uninfected bees become infected is ˇ3 , in bees per virus
carrying mite and time.
Finally, d1 and d2 are the death rates for uninfected and infected honeybees. We
can assume that infected bees live shorter than healthy bees, thus d2 > d1 .
Equation (4) is a logistic growth model for varroa mites. By r we denote the
maximum mite birth rate. The carrying capacity for the mites changes with the host
population site, x C y, and is characterized by the parameter ˛ which indicates how
302 V. Ratti et al.
many mites can be sustained per bee on average. The parameters 1;2 in Eqs. (2)
and (3) represent the mortality rates of bees due to mites feeding on them.
3 Mathematical Formulation of Discrete Interventions
We formulate the process of swarming as discrete interventions in a continuous

system in the form of a hybrid system. The discrete interventions (ti ’s) may depend
on the time or the state variables. For instance, if we assume that swarming takes
place after fixed time intervals i.e., mimicking the natural life cycle of a queen bee,
the time at which discrete interventions takes place will be given a priori. On the
other hand, if we assume that swarming takes place due to overcrowding in the
hive, the occurrence of discrete interventions will be state dependent.
The mathematical formulation in terms of hybrid dynamical systems is
zP.t/ D f .t; z.t//; t 2 .ti ; tiC1 /; z.t0 / D z0 ; z.ti / D Xi ; (6)

XiC1 D F.ziC1 /; where ziC1 D lim z.t/: (7)
t%tiC1
Here ti are the discrete times at which events take place and t0 < : : : < ti < tiC1 .
Also, z 2 R4 , t 2 RC ; i 2 N, X 2 R4 and f 2 K where
K D f.m; x; y; M/ 2 R4 W m 0; M 0; x 0; y 0; x C y > 0g:
In the case of swarming, we assume that F.X/ D DX, where D 2 R44 is a

diagonal matrix.
Also, z D .m; x; y; M/T , f is the RHS of (1), (2), (3) and (4), and, X and D are
given by
3 2 2 3
m a0 00
6x7 60 b 0 07
XD6 7
4 y 5; DD6
40 0
7:
1 05
M 00 0a
We assume that the infected bees do not leave the parent colony, because the
disease progresses rapidly and sick bees are unable to fly. We also assume that
the percentage of total mites and virus carrying mites leaving the parent colony
is the same. Thus, two new parameters a and b are introduced. The parameter a
is the percentage of mites staying in the parent colony and the parameter b is the
percentage of uninfected bees staying in the parent colony after the swarm leaves.
4 Computer Simulations
In the simulation experiments, we study the process of swarming taking place due to
overcrowding in the hive and due to events like queen supercedure. We investigate
how these two causes of swarming affect the strength and survival of the parent
colony infested with mites and virus. In the case where swarming takes place due to
overcrowding in the hive, discrete interventions take place when
x.t/ D cx where 0 < c 1: (8)
The quantity x is calculated numerically by running the disease free model

(without swarming) until the steady state is reached and then taking the maximum
of the steady state solution over a 2 year period. For the simulation experiments, we
fix c D 0:95. In the case where swarming takes place due to queen failure, the time
at which discrete interventions (ti ’s) take place are fixed a priori as ti D t0 C i. t/,
where t D 2T, T D 365 days and i D 1; 2; : : : ;. We assume that t0 D 45 i.e.,
swarming takes place in mid May [14].
The seasonal averages of the parameters ˇi ; r; d1 ; d2 ; k and are given in Table 1.
Lacking more detailed information about the parameters, we use these values
to construct piecewise constant time varying parameter functions, assuming four
equally long seasons of 91.25 days [11]. Two sets of seasonal averages of the
parameter ˛, [0.4784, 0.5, 0.5, 0.4784] and [0.1 0.1 0.1 0.1], for the spring, summer,
fall and winter respectively, are used depending upon the scenario under study
(see [11] for the choice of these values). In order to better present the simulation
experiments, we introduce two new variables a1 D 1 a and b1 D 1 b. Here,
a1 is the proportion of mites leaving the parent colony and b1 is the proportion of
uninfected bees leaving the parent colony during swarming.
Table 1 Seasonal averages of model parameters, derived from the data presented in the literature
[6, 7, 11, 13]. The parameters included here are kept constant for all simulations; the values of the
parameters that are varied are given in the text
Parameter Spring Summer Autumn Winter Source
ˇ1 0.1593 0.1460 0.1489 0.04226 [13]
ˇ2 0.04959 0.03721 0.04750 0.008460 [13]
ˇ3 0.1984 0.1460 0.1900 0.03384 [13]
d1 0.02272 0.04 0.02272 0.005263 [13]
d2 0.2 0.2 0.2 0.005300 [13]
500 1500 500 0 [13]
k 0.000075 0.00003125 0.000075 N/A [13]
K 8000 12,000 8000 6000 [11]
r 0.0165 0.0165 0.0045 0.0045 [6, 7]
1 D 2 107 107 107 107 [11]
304 V. Ratti et al.
4
(a) x 10
4 (b) 3 x 10 Virus carrying mites(m)
3.5
Virus carrying mites(m) Uninfected bees(x)
Uninfected bees(x) 2.5 Total mites(M)
3
Infected bees (y)
2.5 Total mites(M) 2
Population
Population
2 1.5
1.5
1
1
0.5
0.5
0 0
0 500 1000 1500 2000 2500 3000 0 1000 2000 3000 4000 5000 6000 7000 8000
Fig. 1 (a) Bee-mite-virus system in the absence of swarming. (b) Bee-mite-virus system in the
presence of swarming due to overcrowding. Threshold bee population at which swarming takes
place is 31,342 and we assume a1 D 0:65; b1 D 0:5
Illustrative Simulation: The purpose of this simulation is to (i) show a typical

simulation of the temporal dynamics of a honeybee colony infested with mites and
virus and, (ii) to compare the system with and without the process of swarming. In
Fig. 1a, we consider a honeybee-mite-virus system with no swarming taking place.
We assume the lower value of the parameter ˛ i.e., [0.1 0.1 0.1 0.1] so that the colony
fights off the virus but not the mites [11]. The uninfected bee population increases
from 13,350 in the beginning of spring and reaches its maximum of 31,195 in the
summer, decreases in the fall and reaches its minimum of 13,350 in the winter.
This pattern repeats annually. After an initial transient period of 2 years, the mites
establish themselves in the colony and follow the same pattern as the bee population.
In Fig. 1b, we consider a honeybee colony infested with mites and virus where
swarming takes place due to overcrowding i.e., swarming takes place if the bee
population exceeds a certain threshold value x (see Eq. 8). We assume the higher
value of the parameter ˛ i.e., [0.5 0.4784 0.4784 0.5]. We assume that a1 D 0:65 and
b1 D 0:5 which means 50 % of the uninfected bees and 65 % of the mites leave the
parent colony during swarming. We choose the parameters, a1 ; b1 and ˛, in order to
present a reference case. The bee population follows the same pattern as in Fig. 1a in
the first year. In the second year, the population increases in the spring and summer
and then swarming takes place when the bee population reaches the threshold value
x which in this case is 31,342. The population the drops from 29,670 to 15,207
bees during swarming and then increases to 23,780 due to the model dynamics. The
population starts decreasing again in the fall and winter and reaches its minimum
of 13,180. This pattern repeats itself annually. After a transient period of 1 year, the
mite population also follows the same oscillatory behavior as the bee population.
Simulation Experiment I: In this simulation experiment, we consider the case
where swarming takes place due to overcrowding. As in the illustrative simulation,
we assume that swarming takes place if the bee population exceeds a certain
threshold value, x . We use the value of the parameter ˛ to be [0.5 0.4784 0.4784
0.5] for the spring, summer, fall and winter. In Fig. 2a, we compare a disease free
colony, a mite infested colony in which only 5 % of the mites leave the parent colony
(a1 D 0:05), and, a mite infested colony in which 70 % of the mites leave the parent
4
(a) 2
x 10
Average bee population at the finaltime

Disease free
Mite infested with tm = 5%
1.5
0.5
0
0.5 0.6 0.7 0.8 0.9 1
Percentage of the bees leaving the parent colony
(b)3 x 10 4 (c) x 10
4
3
Virus carrying mites(m) Virus carrying mites(m)
Uninfected bees(x) Uninfected bees(x)
2.5 Total mites(M) 2.5 Total mites(M)
2 2
Population
Population
1.5 1.5
1 1
0.5 0.5
0 0
0 1000 2000 3000 4000 5000 6000 7000 8000 0 1000 2000 3000 4000 5000 6000 7000 8000
(d)3 x 10 4
Virus carrying mites(m)

(e) 3 x 10 4

Uninfected bees(x) Uninfected bees(x)
2.5 Total mites(M) 2.5 Total mites(M)
2 2
Population
Population
1.5 1.5
1 1
0.5 0.5
0 0
0 1000 2000 3000 4000 5000 6000 7000 8000 0 1000 2000 3000 4000 5000 6000 7000 8000
Fig. 2 Swarming due to overcrowding: (a) Effect of the percentage of the healthy bees leaving the
parent colony on the average healthy bee population of the (i) disease free colony, (ii) mite infested
colony with a1 D 0:05, and (iii) mite infested colony with a1 D 0:7. (b) The colony fights off
the virus when b1 D 0:5 and a1 D 0:65. (c) The colony dies off after 7000 days when b1 D 0:6,
a1 D 0:65. (d) The colony dies off after 6600 days when a1 D 0:64, b1 D 0:5. (e) The colony
fights off the virus and survives as a properly working colony when a1 D 0:65, b1 D 0:5
colony (a1 D 0:7). We vary the percentage of healthy bees leaving the parent colony
over 50–100 % which covers the range (50–70 %) given by [3, 14] i.e., we vary b1
from 0.5 to 1. In the case of the disease free colony, the average population starts
from 19,000 and decreases gradually as the parameter b1 increases. The average
population suddenly drops down when b1 D 0:87. In the case of the mite infested
colony with a1 D 0:05, the average bee population starts at a lower level (i.e., at
18,000) than in the disease free case and decreases to 11,000 bees as b1 increases
followed by a sudden drop to 0 when b1 D 0:87. When a1 D 0:7, the average bee
population starts at 18,500 which is between the initial average population in the
disease free case and the case when a1 D 0:05. The critical value for the percentage
of bees leaving the parent colony is the same in all three cases. Figure 2b, c show
the effect of the percentage of uninfected bees leaving the parent colony (b1 ) on the
306 V. Ratti et al.
dynamics of the colony; we fix the parameter a1 D 0:65. In Fig. 2b, the colony fights
of the virus and survives as a properly working colony when b1 D 0:5. In Fig. 2c,
the colony dies off after 7000 days when b1 D 0:6. Figure 2d, e show how the
colony, that otherwise dies off due to virus, survives as a properly working colony if
the percentage of mites leaving the parent colony is above a threshold value; we fix
the parameter b1 D 0:5. In Fig. 2d, the colony dies off due to virus after 6600 days
when a1 D 0:64. Figure 2e shows that when a1 D 0:65, the colony fights off the
virus and works as a properly working colony.
Simulation Experiment II: In this simulation experiment, we assume that swarm-
ing occurs at fixed time intervals of 2 years mimicking the natural death cycle of
the queen bee. We assume that the swarm leaves the colony every 2 years in the
mid of May [14]. We use the value of the parameter ˛ to be [0.1 0.1 0.1 0.1] for
the spring, summer, fall and winter. In Fig. 3a, we compare a disease free colony, a
mite infested colony with a1 D 0:05, and, a mite infested colony with a1 D 0:7. We
investigate the effect of the percentage of bees leaving with the swarm (b1 ) on the
average population of the parent colony. The parameter b1 is varied from 0.5 to 1.
In case of the disease free colony, the average bee population starts from 21,000 and
remains constant when the parameter b1 is varied from 50 % to 76 % and suddenly
drops down to 0 when b1 reaches 0:77. In case of the mite infested colony with
a1 D 0:05, the average bee population starts below the disease free population and
remains constant until the parameter b1 reaches 0:75 when it suddenly drops down
to 0. In case of the mite infested colony where a1 D 0:7, the average bee population
starts at the same level as in case of a1 D 0:05 and remains constant until a1 reaches
0:76 and then it suddenly drops down to 0. It is interesting to note that the threshold
value for the parameter b1 is the maximum in case of the disease free colony which
is followed by the mite infested case with a1 D 0:7 which in turn is followed by the
mite infested case with a1 D 0:05.
Figure 3b, c show the effect of the percentage of uninfected bees leaving the
parent colony (b1 ) on the survival of the colony. The parameter a1 is fixed to be 0.7.
In Fig. 3b, the colony fights off the virus and survives as a properly working colony
when b1 D 0:76. In Fig. 3c, the colony dies off after 1000 days when b1 D 0:77.
Figure 3d, e show how the colony, that otherwise dies off due to virus, survives
as a properly working colony if the parameter a1 is above a threshold value. The
parameter b1 is fixed to be 0.5. In Fig. 3d, the colony dies off due to virus after
6000 days when a1 D 0:91. Figure 3e shows that when a1 D 0:92, the colony fights
off the virus and survives as a properly working colony.
5 Summary and Conclusion
• A general framework is provided to incorporate the discrete interventions into

the model using a discrete-continuous model. This framework is applied to the
process of swarming which involves two types of discrete interventions: (i) due
4
x 10
(a) 2.5
Average bee population at the finaltime

Disease free
2 Mite infested with tm = 70%
1.5
0.5
0
0.5 0.6 0.7 0.8 0.9 1
Percentage of the bees leaving the parent colony
(b) 3.5 x 10 4
(c) 3.5 x 10 4
Virus carrying mites(m) Virus carrying mites(m)

3 Uninfected bees(x) 3 Uninfected bees(x)
Total mites(M) Total mites(M)
2.5 2.5
Population
Population
2 2
1.5 1.5
1 1
0.5 0.5
0 0
0 1000 2000 3000 4000 5000 6000 7000 8000 0 500 1000 1500 2000 2500 3000
(d) 3.5
x 10
4

(e) x 10
4
3.5
Uninfected bees(x) Virus carrying mites(m)
3
Total mites(M) Uninfected bees(x)
3
Total mites(M)
2.5
2.5
Population
Population
2
2
1.5
1.5
1 1
0.5 0.5
0 0
0 1000 2000 3000 4000 5000 6000 7000 8000 0 1000 2000 3000 4000 5000 6000 7000 8000
Fig. 3 When swarming takes place every 2 years: (a) Effect of the percentage of the healthy bees
leaving the parent colony on the average healthy bee population of the colony that is (i) disease
free, (ii) mite infested with a1 D 0:05, and (iii) mite infested with a1 D 0:7. (b) The colony
fights off the virus when b1 D 0:76 and a1 D 0:7. (c) The colony dies off after 1000 days when
b1 D 0:77 and a1 D 0:7. (d) The colony dies off after 6000 days when a1 D 0:91 and b1 D 0:5.
(e) The colony fights off the virus and survives as a properly working colony when a1 D 0:92 and
b1 D 0:5
to queen failure i.e., when discrete events occur at times that are a priori fixed (ii)
due to overcrowding i.e., when occurrence of discrete events depend on the state
of the system.
• In case of a disease free colony and a mite-infested colony, a critical value of
the percentage of bees leaving the colony plays an important role. The colony
survives only if the percentage of bees is below this critical value.
• In case of a colony infested with mites and virus, in addition to the percentage of
the bees leaving the colony, the percentage of virus carrying mites and virus free
mites leaving the colony during swarming also has a huge impact on the survival
of the parent colony. Particularly, a colony, that otherwise dies off due to virus,
can survive if the percentage of mites leaving the parent colony is above a critical
value.
308 V. Ratti et al.
• The critical value of the percentage of bees leaving the colony is lower in case
of swarming due to overcrowding as compared to the case where swarming
takes place after fixed intervals. This difference could be due to the fact that
swarming due to overcrowding takes place every year, however, swarming after
fixed intervals is basically every 2 years. The parent colony is able to tolerate
greater loss of bees because it gets longer time to establish itself before the next
swarming event takes place.
References
1. Britton, N.F., Franks, N.R., Pratt, S. C., Seeley, T.D.: Deciding on a new home: how do
honeybees agree? Proc. R. Soc. Biol. Sci. 269(1498), 1383–1388 (2002)
2. Eberl, H.J., Frederick, M.R., Kevan, P.G.: The importance of brood maintenance terms in
simple models of the honeybee – Varroa destructor – acute bee paralysis virus complex.
Electron. J. Differ. Equ. Conf. Ser. 19, 85–98 (2010)
3. Fries, I., Hansen, H., Imdorf, A., Rosenkranz, P.: Swarming in honey bees (Apis mellifera) and
Varroa destructor population development in Sweden. Apidologie 34, 389–397 (2003)
4. Genersch, E., von der Ohe, W., Kaatz, H., Schroeder, A., Otten, C.,́ Büchler, R., Berg, S.,
Ritter, W., Mühlen, W., Gisder, S., Meixner, M., Liebig, G., Rosenkranz, P.: The German bee
monitoring project: a long term study to understand periodically high winter losses of honey
bee colonies. Apidologie 41, 332–352 (2010)
5. Kevan, P.G., Hannan, M., Ostiguy, N., Guzman-Novoa, E.: A summary of the varroa-virus
disease complex in honey bees. Am. Bee J. 146(8), 694–697 (2006)
6. Martin, S.: A population dynamic model of the mite varroa jacobsoni. Ecol. Model. 109, 267–
281 (1998)
7. Martin, S.J.: Varroa destructor reproduction during the winter in apis mellifera colonies in UK.
Exp. Appl. Acarol. 25(4), 321–325 (2001)
8. Myerscough, M.R.: Dancing for a decision: a matrix model for nest-site choice by honeybees.
Proc. R. Soc. Lond. B: Biol. Sci. 270(1515), 577–582 (2003)
9. Ostiguy, N.: Honey bee viruses: transmission routes and interactions with varroa mites. In: 11
Congreso Internacional De Actualizacion Apicola, vol. 9 al 11De Junio De 2004. Memorias.,
p. 47 (2004)
10. Passino, K.M., Seeley, T.D.: Modeling and analysis of nest-site selection by honeybee swarms:
the speed and accuracy trade-off. Behav. Ecol. Sociobiol. 59(3), 427–442 (2006)
11. Ratti, V., Kevan, P.G., Eberl, H.J.: A mathematical model for population dynamics in honeybee
colonies infested with varroa destructor and the acute bee paralysis virus. Can. Appl. Math. Q.
21(1), 63–93 (2013)
12. Ratti, V., Kevan, P.G., Eberl, H.J.: A mathematical model for population dynamics in honeybee
colonies infested with varroa destructor and the acute bee paralysis virus with seasonal effects.
Bull. Math. Biol. 77(8), 1493–1520 (2015)
13. Sumpter, D.J., Martin, S.J.: The dynamics of virus epidemics in varroa-infested honey bee
colonies. J. Anim. Ecol. 73(1), 51–63 (2004)
14. Winston, M.L.: The biology of the honey bee. Harvard University Press, Cambridge (1991)
15. ZKBS (Zentralkommittee für biologiche Sicherheit des Bundesamts für Verbraucherschutz
und Lebensmittelsicherheit); Empfehlung Az.: 45242.0087 - 45242.0094, 2012, (in German:
Central Committee for Biological Safety of the Federal Agency for Consumer Protection and
Food Safety, Recommendation 45242.0087-45242.0094, 2012)
To a Predictive Model of Pathogen Die-off in Soil
Following Manure Application
Andrew Skelton and Allan R. Willms
Abstract The application of manure is an important component of nutrient man-

agement in the production of field crops. The regulations governing the safe
application of manure are based on laboratory data, which may or may not
accurately reflect the environmental fluctuations seen in field conditions.
This study aims to develop a predictive model for pathogen die-off in soil
following manure application. An ordinary differential equation model is presented
and fit to experimental data. The challenges of modelling field derived data,
including detection thresholds, viable but nonculturable bacteria and difficulties in
winter data collection, are discussed. The capabilities of the model for predictive
purposes and the development of additional experimental trials and data collection
methods are discussed.
1 Background
Field production of fruits and vegetables typically includes application of manure

as a vital component of nutrient management. In organic farming, manure, compost
and other organic materials are the only substances acceptable for use as fertilizer.
The Standards Council of Canada 2008 regulations require that
manure application shall be designed to ensure that manure application a. does not
contribute to the contamination of crops by pathogenic bacteria, b. minimizes the potential
for run-off into ponds, rivers and streams, c. does not significantly contribute to ground
and surface water contamination. The non-composted solid or liquid manure shall be
a. incorporated into the soil at least 90 days before the harvesting of crops for human
consumption that do not come into contact with soil; b. incorporated into the soil at least
120 days before the harvesting of crops having an edible part that is directly in contact with
the surface of the soil or with soil particles.
Pathogen die-off rates depend on environmental and biological conditions and,

as such, will differ from season to season and year to year. A number of researchers
have studied the effect such conditions have on pathogen die-off, such as the effect
A. Skelton () • A.R. Willms


310 A. Skelton and A.R. Willms
of temperature [7, 8, 14], soil and manure type [18], manure application methods
[2], manure storage [9], intervention strategies [5] and genetic factors [11, 15]. A
nice overview of studies conducted on survival of pathogens is given in [17]. A
recent study [6] presented results from a meta regression on a large number of
potential influencing factors and concluded that the three most important factors
affecting pathogen die-off rates were temperature, water and soil types, and whether
or not the study was conducted in the lab or in the field. Temperature and soil
moisture are environmental factors that are relatively easy to record and predict and,
as such, are of use in a predictive model. The third factor is of particular interest
as the waiting times required by the regulations are based on data obtained under
laboratory conditions.
Data obtained from laboratory studies are subject to highly consistent and
controlled environmental conditions and thus do not necessarily accurately reflect
the wide fluctuations seen in field conditions. The fixed waiting times required by
the regulations may either over- or under-estimate the time required to obtain safe
levels of pathogen reduction. It is, therefore, of value to model pathogen die-off
in local climatic conditions. Obtaining data in field conditions, however, brings its
own set of challenges and concerns. The aim of this study is to develop a predictive
model of pathogen die-off and answer questions arising from the study of pathogen
die-off in soil. Given measurements of pathogen levels in soil and some knowledge
of future environmental conditions, can future pathogen levels be predicted? Can
field derived data provide scientific information that can lead to informed policy
development which allows for sustainable food production while guarding against
food contamination?
2 Experiment
A field study was conducted in 2011–2012 by a team led by Dr. Ann Huber from the
Soil Resource Group and Dr. Keith Warriner from the Department of Food Science
at the University of Guelph. The project was funded by the Food Safety Research
Program facilitated by the Ontario Ministry of Agriculture, Food and Rural Affairs
(OMAFRA). A detailed description of the experimental procedure is available in
[10] and illustrations and photographs of the experiment are presented in Fig. 1.
Trials were conducted at a farm northwest of Belwood, Ontario at two sites,
one with Perth Loam soil and the other with Hillsburg Fine Sandy Loam soil. The
two sites were adjacent to each other and thus subject to the same environmental
conditions. Manure (either dairy or swine) was spiked with either E.coli O157,
Salmonella or Listeria, mixed with the soil, and placed in sentinel vials (see
Fig. 1). These vials, which have a membrane impermeable to the contaminant, allow
the monitoring of pathogen populations in a confined environment and provide a
mechanism for separating the environmental effect on pathogen die-off from the
effect due to transport, such as run-off and movement through the soil profile.
The sentinel vials were buried either directly under the surface or at a depth of
To a Predictive Model of Pathogen Die-off in Soil Following Manure Application 311
Fig. 1 Illustrations and photographs of the experimental procedure (Images reproduced with
permission of Dr. Huber at the Soil Resource Group). Top: Introducing the inoculated sentinel
vials into the test plot at different depths. Middle: Assembly of the sentinel vials into which the
pathogen inoculated manure amended soil will be introduced. Bottom: Methodology to be applied
in the study
approximately 15 cm and additional non-spiked manure was hand applied at a rate

equivalent to within the vials. Three trials were conducted: the first from June 16,
2011 to May 29, 2012, the second from September 22, 2011 to June 6, 2012 and the
third from November 2, 2011 to June 6, 2012. Each trial included pre-determined
sampling dates on which three sentinel vials were extracted and pathogen levels
reported. The three trials contained 13, 11 and 8 sampling dates respectively. Our
experimental data therefore consists of 72 (3 trials 3 pathogens 2 depths 2 soil
types 2 manures) time series data sets, with three data replicates at each sampling
date.
A typical data set is found in Table 1. Note that there are no sampling dates
over the winter months due to the soil becoming frozen and too snow covered to
be realistically sampled. When a sentinel vial is extracted from the ground, a soil
sample of approximately 10 g is obtained. The sample is processed, diluted and
bacteria are cultured and counted. If there are too many bacteria to count after being
cultured, the sample is repeatedly diluted by a factor of 10 until the number of
bacteria is countable. The data in row four of Table 1, for example, has been diluted
twice, so the recordings of 7, 5 and 11, denote true counts of 700, 500 and 1100
bacteria respectively. In each dilution, triplicate counts are obtained. For ease of
reading, only the median result is reported in Table 1, but averages are used in
the analysis. If there are no bacteria visible at a given dilution, we conclude only
that there is less than 1 bacteria at that given dilution level. The data in the last
row of Table 1, for example, displays counts of 0 bacteria at a dilution of 0.1. The
conclusion that we may draw from this observation is that there is less than 10
bacteria in the sample. Our time series data will, therefore, contain censored data
that is known only to have a value beneath a detection threshold. The bacteria counts
Table 1 A typical data set. Data collected for over the September 22, 2011 to June 6, 2012 trial,
using dairy manure, in sandy soil, at the surface, measuring levels of E.coli O157. To determine
the data values for model fitting, we take the bacterial count, divide by the dilution, scale by the
sample weight and take the logarithm of the answer to determine the log bacterial count per 10 g
of soil
Sample weights (g) Bacterial counts
Date Sample 1 Sample 2 Sample 3 Sample 1 Sample 2 Sample 3 Dilution
Sep 22/11 10:000 10:000 10:000 1 1 1 0:001
Sep 25/11 10:063 10:019 10:253 75 73 101 0:01
Oct 3/11 10:247 10:154 10:057 1 4 3 0:001
Oct 13/11 10:297 10:024 10:054 7 5 11 0:01
Oct 19/11 10:328 10:035 10:264 1 2 65 0:1
Nov 10/11 10:254 10:106 10:012 3 0 3 0:1
Nov 25/11 10:107 10:057 10:028 1 2 0 0:1
Mar 27/12 10:039 10:041 10:048 0 0 0 0:1
May 1/12 10:059 10:005 10:140 0 0 0 0:1
May 17/12 10:152 10:128 10:020 0 0 0 0:1
Jun 7/12 9:999 10:002 10:006 0 0 0 0:1
30 0.55
Average Daily Air Temperature
0.5
20
0.45
% Soil Moisture
0.4
10
0.35
0.3
0
0.25
−10 0.2
0 50 100 150 200 250 300 350 400 0 50 100 150 200 250 300 350 400
Days since June 10, 2011 Days since June 10, 2011
Fig. 2 Environmental data collected. Recordings from the weather station for average daily air
temperature and soil moisture over the course of the three trials. Missing data were interpolated
from records kept by nearby weather stations
are then scaled by the relative weight of the sample so if, for example, the sample
weight was 10:107 g, we would multiply the count by 10=10:107 to obtain a scaled
count. The data with which we will fit our model will be given as a log count of
bacteria per 10 g of soil.
Environmental data were collected by a weather station installed at the site. The
average daily air temperature and percentage of soil moisture are shown in Fig. 2.
There were two periods of missing environmental data. The first was due to strong
winds knocking over the weather station, and the second was due to a raccoon
consuming the power supply. In both cases, it was possible to obtain replacement
weather data spatially averaged between nearby weather stations using data made
freely available by Environment Canada.
3 Model and Results
We now fit our experimental data using an ordinary differential equation model.
Let x D x.t/ denote the logarithm of the bacteria count per 10 g of soil, and T D
T.t/; M D M.t/ denote the average daily air temperature (degrees Celsius) and soil
moisture (%) linearly interpolated between adjacent data points from the data shown
in Fig. 2.
Pathogen die-off plots typically follow a characteristic logistic curve [2, 3, 7,
11, 14, 18], so we set up a logistic differential equation in which the rate constant
is dependent on both temperature and soil moisture. It is common to discuss
pathogen die-off in terms of numbers of log reductions, so an empirical model of
pathogen die-off in which the dependent variable is given on a logarithmic scale is
appropriate. Using a quadratic dependence on temperature and linear dependence
on soil moisture, we obtain the following ordinary differential equation
x0 D .k1 C k2 T C k3 T 2 C k4 M/.x B/.C x/; (1)

where the four rate constants k1 ; k2 ; k3 ; k4 , the two shape constants B; C (the
baseline and maximum values, respectively, which will depend on the trial, soil type,
manure type and sample depth) and the initial condition x.t0 /, where t0 denotes the
start of the particular trial of interest and B < x.t0 / < C, will be obtained from
finding the best model fit to the data. The initial condition must be estimated due
to the use of lab bacteria in field trials. When the lab bacteria are exposed to field
conditions, a subset of the bacteria will experience shock due to the sudden change
in environmental conditions and enter a viable but nonculturable (VBNC) state.
Bacteria in a VBNC state cannot be cultured on growth media [13], but are still
viable and potentially dangerous, and thus must be accounted for in our analysis.
This phenomenon can be seen in the data in Table 1. The first row of data, obtained
at the end of the first day of the trial, has an average of 1000 bacteria per 10 g of
soil. This observed value is significantly less than the number of bacteria that were
originally placed in the spiked soil. The second row of data, obtained 3 days later,
has an average of 8406 bacteria per 10 g of soil. This discrepancy reflects the number
of bacteria that entered a VBNC state initially, but had recovered after 3 days in
the field. In our analysis, we ignore the initial data value and instead estimate the
number of bacteria present at the start of the experiment (Table 2).
Owing to the fact that some of the data points are censored due to detection
thresholds, we will be required to use the Expectation-Maximization (EM) algo-
rithm to fit our model to the data. The EM algorithm first sets all censored data to
exactly half of the appropriate detection threshold (in our case, to either 5, 50, 500,
or 5000 bacteria). The following steps are then repeated until convergence: a least
squares procedure is used to find the best model fit given the current data values
and then the censored data points are adjusted to fit the current parameter set. The
mathematical details of the EM algorithm can be found in [4, 12] and the algorithm
as applied to censored data at multiple detection thresholds can be found in [1].
Table 2 Best Fit Parameters for the plots shown in Fig. 3. Details of the four trials are as follows.
(a) Dairy manure, loam soil, surface, E.Coli, (b) Dairy manure, sand soil, surface, E.Coli, (c) Dairy
manure, sand soil, depth, Salmonella, (d) Swine manure, loam soil, depth, Listeria
Parameter Figure 3a Figure 3b Figure 3c Figure 3d
k1 5:15 104 3:46 103 1:62 103 7:48 106
k2 6:51 105 3:87 104 4:39 104 4:53 103
k3 2:00 105 2:89 105 1:21 104 1:85 104
k4 1:60 104 2:33 104 9:95 104 3:17 103
B 0:79 0:39 0:92 1:37
C 9:71 5:25 6:56 6:32
x.0/ 4:46 4:22 4:18 4:80
To find the best fit parameter set, we will be required to search in a seven-
dimensional parameter space. Very little a priori information is available as to the
values of many of the parameters. It is therefore useful to reduce the size of the
parameter space we will be required to search. We used the method presented
in [16] to reduce the size of the search space and improve the speed of the
minimization steps of the EM algorithm. This method uses interval analysis and
linear multistep discretizations to remove boxes of parameter space that are deemed
to be inconsistent with the data. The method is able to quickly remove large regions
of parameter space to allow traditional minimization techniques to work more
effectively.
Results for representative data sets are presented in Fig. 3. We were, in general,
able to obtain good fits to our experimental data. In plots (a) and (d), we can see
that there is an expected initial rapid decrease in the bacterial count, followed by
a slower die-off (and in fact a small growth in the case of Listeria) over the winter
season, followed by another period of more rapid die-off during the spring months.
Plots (b) and (c) illustrate the difficulties faced in modelling these data sets. The
only viable data exists before the winter freeze, making any dynamics during or
after the winter break difficult to model. This lack of information in many data sets
has made calibrating a model to the data very difficult. It has, however, directly led
to improved experimental design and data collection in a follow-up trial.
4 Conclusion and Future Work
This paper presents the first step in a process by which we hope to develop a
predictive model for pathogen die-off in soil following manure application. We fit
an ordinary differential equation model to field derived data. Our model was able
to reasonably describe the pathogen die-off using soil moisture and air temperature,
both commonly available environmental data. To establish a predictive, rather than
descriptive, model, we will require additional winter and long-term data with which
we can calibrate our model and then use for testing of predictive capabilities. Such
data is currently being collected and analysis of the data is ongoing and will be
reported elsewhere.
Acknowledgements This work was supported by an NSERC Discovery Grant.

Fig. 3 Best fit plots. The (a) 6

solution to the differential
equation model with best fit
Bacteria Log count

parameters plotted for four 4
representative data sets from
the September 22/11 to June 2
7/12 trial. The open circles
represent data points whose
values are known only to 0
exist beneath the detection
threshold. The closed dots −2
represent the predicted data 100 150 200 250 300 350 400
as outputted from the EM Days since June 10, 2011
algorithm. (a) Dairy manure, (b) 6
loam soil, surface, E.Coli, (b)
Dairy manure, sand soil,
Bacteria Log count
surface, E.Coli, (c) Dairy 4

manure, sand soil, depth,
Salmonella, (d) Swine 2
manure, loam soil, depth,
Listeria
0
−2
100 150 200 250 300 350 400
Days since June 10, 2011
(c) 5
4
Bacteria Log count
0
100 150 200 250 300 350 400
(d) 5
Bacteria Log count
1
100 150 200 250 300 350 400
References
1. Banks, H.T., Davidian, M., Hu, S., Kepler, G.M., Rosenberg, E.S.: Modelling HIV immune
response and validation with clinical data. J. Biol. Dyn. 2(4), 357–385 (2008)
2. Bech, T., Dalsgaard, A., Jacobsen, O.S., Jacobsen, C.S.: Leaching of Salmonella enterica in
clay columns comparing two manure application methods. Ground Water 49(1), 32–42 (2011)
3. Coroller, L., Leguerinel, I., Mettler, E., Savy, N., Mafart, P.: General model, based on two
mixed Weibull distributions of bacterial resistance, for describing various shapes of inactivation
curves. Appl. Environ. Microbiol. 72(10), 6493–6502 (2006)
4. Dempster, A.P., Laird, N.M., Rubin, D.B.: Maximum likelihood from incomplete data via the
EM algorithm. J. R. Stat. Soc.: Ser. B 39, 1–38 (1977)
5. Franz, E., Semenov, A.V., van Bruggen, A.H.C.: Modelling the contamination of lettuce with
Escherichia coli O157:H7 from manure-amended soil and the effect of intervention strategies.
J. Appl. Microbiol. 105, 1569–1584 (2008)
6. Franz, E., Schijven, J., de Roda Husman, A.M., Blaak, H.: Meta-regression analysis of
commensal and pathogenic Escherichia coli survival in soil and water. Environ. Sci. Technol.
48, 6763–6771 (2014)
7. Fremaux, B., Prigent-Combaret, C., Delignette-Muller, M.L., Mallen, B., Dothal, M., Gleizal,
A., Vernozy-Rozand, C.: Persistence of Shiga toxin-producing Escherichia coli O26 in various
manure-amended soil types. J. Appl. Microbiol. 104, 296–304 (2008)
8. Holley, R.A., Arrus, K.M., Ominski, K.H., Tenuta, M., Blank, G.: Salmonella survival in
manure-treated soils during simulated seasonal temperature exposure. J. Environ. Q. 35, 1170–
1180 (2006)
9. Huber, A.: Survival of pathogens during storage of livestock manures. OMAFRA Final Report.
SR9182 (2009)
10. Huber, A., Warriner, K.: Pathogen die-off rates following manure application under Ontario
field conditions – relationship to pre-harvest wait times. OMAFRA Final Report. SF6088
(2012)
11. Ma, J., Ibekwe, A.M., Yi, X., Wang, H., Yamazaki, A., Crowley, D.E., Yang, C.-H.: Persistence
of Escherichia coli O157:H7 and its mutants in soils. PLoS ONE 6(8), 1–9 (2011)
12. McLachlan, G.J., Krishnan, T.: The EM Algorithm and Extensions. Wiley, New York (1997)
13. Oliver, J.D.: The viable but nonculturable state in bacteria. J. Microbiol. 43(S), 93–100 (2005)
14. Ongeng, D., Muyanja, C., Ryckeboer, J., Springael, D., Geeraerd, A.H.: Kinetic model-based
prediction of the persistence of Salmonella enterica serovar Typhimurium under tropical
agricultural field conditions. J. Appl. Microbiol. 110, 995–1006 (2011)
15. Saint-Ruf, C., Garfa-Traore, M., Collin, V., Cordier, C., Franceschi, C., Matica, I.: Massive
diversification in aging colonies of Escherichia coli. J. Bacteriol. 196(17), 3059–3073 (2014)
16. Skelton, A., Willms, A.R.: Parameter range reduction in ordinary differential equation models.
J. Sci. Comput. 62(2), 517–531 (2015)
17. van Elsas, J.D., Semenov, A.V., Costa, R., Trevors, J.T.: Survival of Escherichia coli in the
environment: fundamental and public health aspects. ISME J. 5, 173–183 (2011)
18. You, Y., Rankin, S.C., Aceto, H.W., Benson, C.E., Toth, J.D., Dou, Z.: Survival of Salmonella
enterica Serovar newport in manure and manure-amended soils. Appl. Environ. Microbiol.
72(9), 5777–5783 (2006)
Mathematical Modeling of VEGF Binding,
Production, and Release in Angiogenesis
Nicoleta Tarfulea
Abstract This paper presents a new mathematical model for the transduction
of the extracellular vascular endothelial growth factor (VEGF) signal into the
intracellular VEGF signal. It is based on a signal transduction pathway that accounts
for VEGF binding, production, and release. The resulting mathematical model
for the evolution of the chemical species concentrations is analyzed analytically
and numerically to address qualitative aspects (including positivity, stability and
robustness with respect to variations of the secretion rate parameter). Various
secretion rate functions are investigated as well.
1 Introduction
In recent years, tumor-induced angiogenesis has become an important field of

research because it represents a crucial step in the development of malignant tumors
[18]. When a cancerous tumor initially forms, it acquires the nutrients needed
for survival from the surrounding tissue. When it grows to a large enough size
[17], it needs another source of food and channel for depositing waste. As an
early response of tumor cells to hypoxia, they start releasing molecules called
tumor angiogenic factors, such as VEGF, into the surrounding tissue to initiate a
process called angiogenesis (growth of new blood vessels from preexisting ones).
Secreted VEGF then diffuses into the surrounding tissue and binds to specific
endothelial cell (EC) receptors. As a response, ECs start to secrete proteolytic
enzymes (such as PA-plasmin, MMPs) which dissolve the basal lamina of the
parent vessel and the surrounding extracellular matrix (ECM). ECs respond to
the stimulus chemotactically and begin to migrate toward it, and so small sprouts
which grow toward the tumor are formed [1, 23, 32, 35]. Over the last years it has
become evident that an effective way of treating cancer is inhibition of angiogenesis
[15, 19, 20, 29, 41], which could be done by targeting endothelial cells rather than
tumor cells since: (i) ECs are more accessible than tumor cells; (ii) death of a single
N. Tarfulea ()
Purdue University Calumet, 2200 169th Street, Hammond, IN, 46323, USA

320 N. Tarfulea
cell will cause death of approximate 500 tumor cells; (iii) ECs are relatively stable
and will not mutate easily into drug resistant variants [50].
The existing mathematical models in the literature [35] have been able to repro-
duce some characteristics of angiogenesis. These include continuous approaches
[2–5, 10–12, 24, 27, 28, 31, 38–40, 44, 52–54], random walk models [2, 6, 13, 22,
38, 39, 46, 47, 51], and cell-based formulations [7–9, 25, 30, 33, 36, 43, 50]. Reviews
of such models can be found in [35, 37, 42]. However, it has been discovered that the
activated ECs secrete detectable amounts of VEGF [16, 45]. Thus, in this paper we
will develop a new mathematical model that account for VEGF binding, production,
and release.
2 Description of the Signal Transduction Pathway
The dynamic interaction between angiogenic factors, the uPA-plasmin system, and
the presumed activation of matrix-degrading metalloproteinases during angiogene-
sis is shown in Fig. 1. It can be described as follows [50]: (1) hypoxic tumor cells
secrete VEGF; (2) VEGF is sequestered in the ECM by binding to ECM heparin
binding sites; (3) VEGF binds to VEGFR (VEGF receptor) on EC cell (VEGFR-
2 is considered the main signaling receptor in ECs); (4) inactive uPA (urokinase
plasminogen activator) pro-uPA, binds to specific cell surface receptor uPAR;
(5) plasminogen (pgn) binds to pgnR; (6) uPAR-bound uPA reacts with bound-
plasminogen to generate Plasmin; (7) uPAR occupied by uPA forms complexes with
PAI-1 (plasminogen activator inhibitor); (8) Plasmin degrades ECM; (9) Plasmin
mobilizes VEGF from ECM reservoir; (10) Plasmin induces the activation of LTGF-
(1) (3)
Cell Endothelial cell
receptors
VEGF
(2)
pro−MMPs
pgnR uPAR
(9)
Plasmin pgn uPA pro−uPAR
(5) (4)
(6) (14) (15)
(10)
ECM
(13) MMPs
(8)
LTGF−β TGF− β PAIs
(12)
(16)
(17) TIMPs
Matrix degradation
Fig. 1 The biochemical pathway of angiogenesis

Mathematical Modeling of VEGF Binding, Production, and Release in Angiogenesis 321
ˇ (latent type beta transforming growth factor) from ECM; (11) Plasmin can convert
LTGF-ˇ into TGF-ˇ (type beta transforming growth factor); (12) active TGF-ˇ
induces synthesis of PAI-1; (13) PAI-1 blocks the conversion of plasminogen into
Plasmin; (14) Plasmin, uPA activate some pro-MMPs (matrix metalloproteinase);
(15) some MMPs can activate other MMPs; (16) all MMPs are inhibited by TIMPs
(tissue inhibitor of MMP enzyme) by binding to pro-MMPs or MMPs; (17) MMPs
degrades ECM. The signal transduction part incorporates known biology more
completely than any other model [35] and involves 18 variables and 14 biochemical
reactions. Using the law of mass action, these lead to a system of ten differential
equations and auxiliary algebraic equations for the time evolution of the intracellular
species.
3 The Simplified Pathway and the Corresponding Model
3.1 Assumptions for the Simplified Model Signal Pathway
Since biological and biochemical processes in angiogenesis are very complex, we

seek reductions in the number of variables that do not alter the dynamics of the
model.
After VEGF is been secreted into the surrounding tissue by tumor cells and
has encountered endothelial cells, it binds to specific proteins (called receptors)
that are sitting on the outer surface of the cell. This binding activates a series of
relay proteins which transmit a signal that starts a series of cascades inside the cell
[50]. As a result the motile machinery of the endothelial cell is activated, and also
some VEGF is secreted [26, 45]. Using these facts, we have reduced the scheme
to four primary species for the intracellular dynamics. A schematic diagram of the
interactions in the proposed model is shown in Fig. 2. The definition of the symbols
used in these reactions as well as the mathematical symbol used later are given in
Table 1. Here ki represent the reaction rates, hi represent the degradation rates, and
bsr is the basal secretion rate; all of them are given in Table 2.
Two classes of species are represented in this model: volumetric (growth factor
– VEGF) and surface (receptors and ligand-receptor complexes). No factors in the
Fig. 2 Simplified pathway Ve

for angiogenesis
R Y*
Vi
Y
322 N. Tarfulea
Table 1 The variables, their symbols, and non-dimensional form

Species Conc. Definition Dimensionless form
Ve y1 .t/ Extracellular VEGF u1 D kk15 y1
y2
R y2 .t/ VEGFR u2 D ŒRT
y3
Ve R y3 .t/ VEGFR bound with VEGF u3 D ŒRT
k2 Ck3
Ve RY y4 .t/ Ve RY complex bound with Y u4 D y
k2 ŒRT ŒYT 4
y5
Y y5 .t/ Activated enzyme u5 D ŒYT
k4
S y6 .t/ Substrate u6 D k3 6
y
YS y7 .t/ Activated enzyme Y bound with S u7 D y7
ŒYT
Vi y8 .t/ Intracellular VEGF u8 D k1
k5 8
y
Y z.t/ Free Y –
t – Time D k5 t
Table 2 Estimates for the kinetic parameters – for numerical simulations

Par. Estimated value Source Par. Estimated value Source
k1 3:8 106 M1 s1 [14, 21] k4 104 m2 /.mol s/ [48, 49]
k1 95 106 s1 [14, 21] k4 0 s1 [48, 49]
k2 103 m2 =.mol s/ [14, 21] k5 6 102 s1 [48, 49]
k2 0 s1 [48, 49] k6 10s1 [48, 49]
k3 103 s1 [48, 49] bsr 3 102 s1 [49]
s 3:8 109 mol=.m2 s/ [48, 49] l3 2:8 104 s1 [48, 49]
l2 105 s1 [21, 49] l4 2:8 104 s1 [48, 49]
fluid that could bind the ligands, such as secreted ECM, are included. Also, we
assume that receptor concentration is uniform over the cell surface (with no receptor
clusters being formed). The detailed biochemical reactions in this mechanism are as
follows.
k1
Ve C R ! Ve R (1)
k1
k2
Ve R C Y ! Ve RY (2)
k2
k3
Ve RY Y + Ve R (3)
k4
Y C S !
Y S (4)
k4
k5
Y S Vi + Y (5)
k6
Y Y (6)
bsr
Vi Ve (7)
3.2 The Mathematical Model for the Intracellular Dynamics
We consider the time evolution of species concentrations deduced according to the

mechanism described above. A rectangular box was used having dimensions 1 mm
1 mm 1:5R, where R D 10m is the radius of EC when spherical geometry is
assumed. Furthermore, N represents the number of cells, Ac and Vc represent the
surface area and the volume of each cell. The reactions that are considered in the
mechanism then lead to the following system of differential equations:
dy1
V0 D N Ac k1 y3 N Ac k1 y1 y2 C N Vc sr.y8 / N Vc h1 y1 (8)
dt
dy2
D k1 y3 k1 y1 y2 C s l2 y2 (9)
dt
dy3
D k1 y1 y2 k1 y3 k2 y3 z C .k2 C k3 /y4 l3 y3 (10)
dt
dy4
D k2 y3 z .k2 C k3 /y4 l4 y4 (11)
dt
dy5
D k4 y5 y6 C .k4 C k5 /y7 C k3 y4 k6 y5 (12)
dt
dy6
D k4 y5 y6 C k4 y7 (13)
dt
dy7
D k4 y5 y6 .k4 C k5 /y7 (14)
dt
dy8
Vc D Ac k5 y7 Vc sr.y8 / Vc h2 y8 : (15)
dt
Here s is the insertion rate of surface species into endothelial cell membrane and
l2 , l3 , and l4 represent the internalization rates of surface receptors and complexes.
Their values are listed in Table 2. We introduce the secretion function sr.y8 / which
occurs in the secretion step (7). First, we consider it to be a linear function of the
intracellular VEGF (Vi ) of the form sr.y8 / D bsry8 : This does not affect the validity
of our model since the true secretion rate function is not known. Also, we consider
the case where bsr is piecewise and periodic function. Moreover, the secretion
function can be easily modified and the model amended as new experimental data
becomes available.
324 N. Tarfulea
We proved the nonnegativity and boundedness of the solution to the initial-value

problem corresponding to the system (8), (9), (10), (11), (12), (13), (14) and (15).
The initial data is nonnegative.
3.3 Insertion of Surface Species into EC Membrane
In this section, we consider two separated sub-cases, namely, when insertion

of surface species into endothelial cell membrane and internalization of surface
receptors and complexes are not taken into consideration (Case A), and when they
are considered in the system (Case B). A comparison between the two systems will
be done.
In Case A, we consider the system of differential equations (8), (9), (10), (11),
(12), (13), (14) and (15), with the values of the parameters s, l2 , l3 , and l4 being
zero. The conservation relations are ŒVe R C ŒR C ŒVe RY D ŒRT ; ŒY C ŒY S C
ŒY D ŒYT ; where subscript T denotes the initial value of a quantity. When these
conditions are solved for y2 .t/ and z.t/ z.t/ D ŒYT y5 .t/ y7 .t/ y2 .t/ D
ŒRT y3 .t/ y4 .t/; and the result is used in Eqs. (8), (9), (10), (11), (12), (13),
(14) and (15), we obtain the system of differential equations for y1 ; y3 ; : : : ; y8 . In
the interest of space, this will not be included here. The values used for the kinetic
parameters are in Table 2 and were found from experimentally based estimates
[14, 21, 48, 49]. We also use Ve .0/ D 2:22 nM, RT D 3:3 105 mol/m2 , and
YT D 2:50 104 mol/m2 . In Case B we take into consideration insertion of
surface species into the endothelial cell membrane and internalization of surface
receptors and complexes. The reactions that are considered in the mechanism then
lead to system of differential equations (8), (9), (10), (11), (12), (13), (14) and
(15), which describes the time evolution of species concentrations. The conservation
relation is ŒY C ŒY S C ŒY D ŒYT and the quantity that is determined by it is z.t/
(corresponding to Y – free enzyme) is z.t/ D ŒYT y5 .t/ y7 .t/; where subscript T
denotes the initial value of a quantity. As before, using this conservation condition,
we obtain a system of differential equations for y1 ; : : : ; y8 .
3.4 Nondimensionalization and Analysis of the Systems
We introduce dimensionless variables for the independent concentrations (Table 1).

We find that they are of two categories: those who reach steady-state rapidly, which
are called fast variables and the remaining ones which are called slow variables.
In our case the fast variables are denoted by u4 and u7 . The nondimensionalization
k5 k5
gives rise to two small parameters 1 D and 2 D . On the time scale,
k2 C k3 k3
the singular equations, corresponding to the fast variables u4 and u7 , can be reduced
du4 du7
to the algebraic system of equations D 0 and D 0. By replacing u4 and
d d
u7 and by dropping the O./ terms, after some simplifications, we obtain a system
of differential equations for the species concentrations. In case A, the variables are
u1 ; u3 ; u5 ; u6 , and u8 . In addition, in Case B we have u2 . Having investigated the
dynamics of species concentrations in the described cases, several observations can
be made. The analysis in each case was made using the same set of parameters and
initial conditions. The reduced system in Case A agrees with the full system only on
short time, however we see that the reduced system in Case B and the full system
agree on both short and long time period.
3.5 Further Assumptions for Intracellular Dynamics Model
After VEGF is been secreted and released into nearby tissues by the hypoxic tumor,
it binds to endothelial cell receptors, process that activates a series of relay proteins
which in turn transmit a signal that starts a series of cascades inside the cell. So far
the mechanism was reduced to four primary species for the intracellular dynamics.
Next, we make the biologically realistic assumption that the substrate is not only
consumed, but also produced during the entire process Thus, we can consider that
the rate of change of substrate concentration is approximately zero. Moreover, we
investigate Case A and Case B too. A similar analysis as well as a comparison
between the two systems is performed and we obtain that Case B agrees with the
full system on both short and long time period. Thus, we consider the following
system to represent the governing equations for the intracellular dynamics.
du1
D N1 ˛1 u3 N1 u1 u2 C ˇ1 u8 ˇ2 u1 (16)
d
du2
D ˛1 u3 u1 u2 C ı1 ı2 u2 (17)
d
du3
D u1 u2 .˛1 C ı3 /u3 (18)
d
du5
D ˛4 ˛2 2 .1 .1 C ˛3 u6 /u5 /u3 ˛5 u5 (19)
d
du8 N 1
D ˛3 u5 u6 .ˇ1 C ˇ3 /u8 ; (20)
d 2
where u6 is constant and initial condition
.u1 .0/; u2 .0/; u3 .0/; u5 .0/; u8 .0//T D .u10 ; u20 ; 0; 0; 0/T :

326 N. Tarfulea
We performed a stability and bifurcation analysis for the above system considering
it to be on the form
du
D f .u; p/; u.0/ D u0 (21)
dt
where u 2 M R5 and p 2 I Rm [34]. We consider m D 1 and the p D ˇ1 which

is the nondimensional form of the VEGF secretion rate. Furthermore, the subset
I Œ0; 1, according with data from biological experiments. We showed that the
system has a steady state .0; 1; 0; 0; 0/ and a unique steady state in the interior of R5C .
Moreover, the unique positive steady state is stable for any ˇ1 2 Œ0; 1 and, since the
Jacobian is nonsingular for any ˇ1 2 Œ0; 1, there are no possible bifurcation points.
4 Conclusions
We have reduced the complicated biochemical pathway of angiogenesis (Fig. 1) to

a simplified one (Fig. 2) involving only four primary species for the intracellular
dynamics (namely, VEGF, Receptor, Enzyme, and Substrate). Two separated cases
have been considered for describing this mechanism (rate of change of substrate
concentration is not zero or approximately zero). The behavior of time evolution
of species concentrations have been investigated in each case and compared with
the expected biological phenomenon. As a conclusion, we have been able to
develop a new mathematical model that describes the intracellular dynamics which
incorporates a realistic model for signal transduction, VEGF production and release.
Considering the fact that it has only recently been established that the activated
endothelial cells secrete detectable amounts of VEGF, quantitative data relative to
this fact is almost entirely absent. Therefore, the model was tested for different sets
of parameters. A qualitative analysis revealed the fact that the system has a unique
positive, asymptotically stable steady state solution. Finally, a bifurcation analysis of
the ODE kinetics of the model showed that there are no possible bifurcation points
with respect to the coefficient of the secretion rate function, and that the system is
robust to variations of this parameter.
First, we considered that the VEGF secretion rate function has a linear depen-
dency on the intracellular VEGF. Next, we investigated the cases when bsr is a
piecewise function and a periodic function in time, with the same average as the
constant value. Interestingly, we obtained the same behavior for short and long time
periods for our system (Fig. 3). We note that this analysis could be extended further
by investigating different parameter variations, for example the ratio of the total
cell volume to the extracellular volume. Additionally, the secretion rate function
can be further modified and the model amended as new experimental data becomes
available.
Fig. 3 Piecewise (left) and periodic (right) secretion rate-bsr
References
1. Adams, R.H., Alitalo, K.: Molecular regulation of angiogenesis and lymphangiogenesis. Nat.
Rev. Mol. Cell Biol. 8, 464–478 (2007)
2. Anderson, A.R.A., Chaplain, M.A.J.: Continuous and discrete mathematical models of tumor-
induced angiogenesis. Bull. Math. Biol. 60, 857–900 (1998)
3. Anderson, A.R.A., Chaplain, M.A.J.: A mathematical model for capillary network formation
in the absence of endothelial cell proliferation. Appl. Math. Lett. 11, 109–114 (1998)
4. Arakelyan, L., Vainstein, V., Agur, Z.: A computer algorithm describing the process of vessel
formation and maturation, and its use for predicting the effects of anti-angiogenic and anti-
maturation therapy on vascular tumor growth. Angiogenesis 5, 203–214 (2002)
5. Balding, D., McElwain, D.L.S.: A mathematical model of tumor-induced capillary growth. J.
Theor. Biol. 114, 53–73 (1985)
6. Bartha, K., Rieger, H.: Vascular network remodeling via vessel cooption, regression and growth
in tumors. J. Theor. Biol. 21, 903–918 (2006)
7. Bauer, A., Jackson, T., Jiang, Y.: A cell-based model exhibiting branching and anastomosis
during tumor-induced angiogenesis. Biophys. J. 92, 3105–3121 (2007)
8. Bentley, K., Gerhardt, H., Bates, P.A.: Agent-based simulation of notch-mediated tip cell
selection in angiogenic sprout initialisation. J. Theor. Biol. 250, 25–36 (2008)
9. Bentley, K., Mariggi, G., Gerhardt, H., Bates, P.A.: Tipping the balance: robustness of tip cell
selection, migration and fusion in angiogenesis. PLoS Comput. Biol. 5, e1000549 (2009)
10. Billy, F., Ribba, B., Saut, O., et al.: A pharmacologically based multiscale mathematical model
of angiogenesis and its use in investigating the efficacy of a new cancer treatment strategy. J.
Theor. Biol. 260, 545–562 (2009)
11. Byrne, H.M., Chaplain, M.A.J.: Mathematical models for tumour angiogenesis: numerical
simulations and nonlinear wave solutions. Bull. Math. Biol. 57, 461–486 (1995)
12. Byrne, H.M., Chaplain, M.A.J.: Explicit solutions of a simplified model of capillary sprout
growth during tumor angiogenesis. Appl. Math. Lett. 9, 69–74 (1996)
13. Capasso, V., Morale, D.: Stochastic modelling of tumour-induced angiogenesis. J. Math. Biol.
58, 219–233 (2009)
14. Claesson-Welsh, L.: Signal transduction by vascular endothelial growth factor receptors.
Biochem. Soc. Trans. 31, part 1 (2003)
15. Delgado, M., Morales-Rodrigo, C., SuÃarez, ˛ A.: Anti-angiogenic therapy based on the binding
to receptors. DCDS-A 32(11), 3871–3894 (2012)
16. Di Talia, S., Gamba, A., Lamberti, F., Serini, G.: Role of repulsive factors in vascularization
dynamics. Phys. Rev. E 73, 041917:1–11 (2006)
328 N. Tarfulea
17. Folkman, J.: The vascularization of tumors. In: Cancer Biology: Readings from Scientific
American, pp. 115–124. Freeman, New York (1996)
18. Folkman, J. What is evidence that tumors are angiogenesis dependent? J. Natl. Cancer Inst. 82,
4–6 (1990)
19. Folkman, J.: Clinical applications of research on angiogenesis. N. Engl. J. Med. 333, 1757–
1763 (1995)
20. Folkman, F.: Fighting cancer by attaching its blood supply. Sci. Am. 275, 150–154 (1996)
21. Gabhann, F.M., Popel, A.S.: Model of competitive binding of vascular endothelial growth
factor and placental growth factor to VEGF receptors on endothelial cells. Am. J. Physiol.
Heart Circ. Physiol. 286, H153–H164 (2004)
22. Goede, V., Schmidt, T., Kimmina, S., Kozian, D., Augustin, H.G.: Analysis of blood vessel
maturation processes during cyclic ovarian angiogenesis. Lab. Invest. 78, 1385 (1998)
23. Hegen, A., Koidl, S., Weindel, K., et al.: Expression of angiopoietin-2 in endothelial cells is
controlled by positive and negative regulatory promoter elements. Arterioscler. Thromb. Vasc.
Biol. 24, 1803–1809 (2004)
24. Holmes, M.J., Sleeman, D.: A mathematical model of tumour angiogenesis incorporating
cellular traction and viscoelastic effects. J. Theor. Biol. 202, 95–112, 2000
25. Jackson, T.L., Zheng, X.: A cell-based model of endothelial cell elongation, proliferation and
maturation during corneal angiogenesis. Bull. Math. Biol. 72, 830–868 (2010)
26. Jekunen, A., Kairemo, K.: Inhibition of angiogenesis at endothelial cell level. Microsc. Res.
Tech. 60, 85–97 (2003)
27. Levine, H.A., Nilsen-Hamilton, M.: Angiogenesis – a biochemial/mathematical perspective.
In: Friedman, A. (ed.) Tutorials in Mathematical Biosciences III, chapter 2. Springer (2006)
28. Levine, H.A., Pamuk, S., Sleeman, B.D., Nilsen-Hamilton, M.: Mathematical modeling of
capillary formation and development in tumor angiogenesis: penetration into the stroma. Bull.
Math. Biol. 63, 801–863 (2001)
29. Lignet, F., Benzekry, S., Wilson, S., Billy, F., Saut, O., Tod, M., You, B., Berkane, A.A.,
Kassour, S., Wei, M.X., Grenier, E., Ribba, B.: Theoretical investigation of the efficacy of
antiangiogenic drugs combined to chemotherapy in xenografted mice. J. Theor. Biol. 320, 86–
99 (2013)
30. Liu, G., Qutub, A.A., Vempati, P., Mac Gabhann, F., Popel, A.: Module-based multiscale
simulation of angiogenesis in skeletal muscle. Theor. Biol. Med. Model. 8, 6 (2011)
31. Manoussaki, D.: A mechanochemical model of angiogenesis and vasculogenesis. ESAIM:
Math. Model. Numer. Anal. 37, 581–599 (2003)
32. Matsushita, K., Yamakuchi, M., Morrell, C.N., et al.: Vascular endothelial growth factor
regulation of Weibel-Palade-body exocytosis. Blood 105, 207 (2005)
33. Milde, F., Bergdorf, M., Koumoutsakos, P.: A hybrid model for three-dimensional simulations
of sprouting angiogenesis. Biophys. J. 95, 3146–3160 (2008)
34. Murray, J.D.: Mathematical Biology, 3rd edn. Springer-Verlag, Berlin/Heidelberg/New York
(2004)
35. Othmer, H.G., Mantzaris, N., Webb, S.: Mathematical modeling of tumor-induced angiogene-
sis. J. Math. Biol. 49(2), 111–187 (2004)
36. Peirce, S.M., Van Gieson, E.J., Skalak, T.C.: Multicellular simulation predicts microvascular
patterning and in silico tissue assembly. FASEB J. 18, 731–733 (2004)
37. Peirce, S.M.: Computational and mathematical modeling of angiogenesis. Microcirculation 15,
739–751 (2008)
38. Plank, M.J., Sleeman, B.D.: A reinforced random walk model of tumor angiogenesis and anti-
angiogenesis strategies. IMA J. Math. Med. Biol. 20, 135–181 (2003)
39. Plank, M.J., Sleeman, B.D.: Lattice and non-lattice models of tumour angiogenesis. Bull. Math.
Biol. 66, 1785–1819 (2004)
40. Plank, M.J., Sleeman, B.D., Jones, P.F.: A mathematical model of tumour angiogenesis,
regulated by vascular endothelial growth factor and the angiopoietins. J. Theor. Biol. 229,
435–454 (2004)
41. Poleszczuk, J., Bodnar, M., Forys, U.: New approach to modeling of antiangiogenic treatment
on the basis of Hahnfeldt et al. model. MBE 8(2), 591–603 (2011)
42. Qutub, A.A., Mac Gabhann, F., Karagiannis, E.D., Vempat, P., Popel, A.: Multiscale models of
angiogenesis. IEEE Eng. Med. Biol. Mag. 28, 14–31 (2009)
43. Qutub, A.A., Popel, A.: Elongation, proliferation & migration differentiate endothelial cell
phenotypes and determine capillary sprouting. BMC Syst. Biol. 3, 13 (2009)
44. Schugart, R.C., Friedman, A., Zhao, R., Sen, C.K.: Wound angiogenesis as a function of tissue
oxygen tension: a mathematical model. PNAS 105, 2628–2633 (2008)
45. Serini, G., Ambrosi, D., Giraudo, E., Gamba, A., Preziosi, L., Bussolino, F.: Modeling the early
stages of vascular network assembly. EMBO J. 8, 1771–1779 (2003)
46. Stokes, C.L., Lauffenburger, D.A.: Analysis of the roles of microvessel endothelial cell random
mobility and chemotaxis in angiogenesis. J. Theor. Biol. 152, 377–403 (1991)
47. Sun, S., Wheeler, M.F., Obeyesekere, M., Patrick, C.: A deterministic model of growth factor-
induced angiogenesis. Bull. Math. Biol. 67, 313–337 (2005)
48. Tang, Y., Othmer, H.G.: A G-protein-based model for cAMP dynamics in Dictyostelium
discoideum. Math. Biosci. 120, 25–76 (1994)
49. Tang, Y., Othmer, H.G.: Excitation, oscillations and wave propagation in a G-protein-based
model of signal transduction in Dictyostelium discoideum. Philos. Trans. R. Soc. 349, 179–195
(1996)
50. Tarfulea, N.: Mathematical Modeling of Signal Transduction and Cell Motility in Tumor
Angiogenesis. University of Minnesota Ph.D. thesis, ProQuest (2006)
51. Tong, S., Yuan, F.: Numerical simulations of angiogenesis in the cornea. Microvasc. Res. 61,
14–27 (2001)
52. Travasso, R.D.M., Corvera Poir, E., Castro, M., et al.: Tumor angiogenesis and vascular
patterning: a mathematical model. PLoS ONE 6, e19989 (2011)
53. Xue, C., Friedman, A., Sen, C.K.: A mathematical model of ischemic cutaneous wounds.
PNAS 106, 16782–16787 (2009)
54. Zheng, X., Koh, G.Y., Jackson, T.: A continuous model of angiogenesis: initiation, extension,
and maturation of new blood vessels modulated by vascular endothelial growth factor,
angiopoietins, platelet-derived growth factor-B, and pericytes. DCDS-B 18, 1109–1154 (2013)
A Mathematical Model of Cytokine Dynamics
During a Cytokine Storm
Marianne Waito, Scott R. Walsh, Alexandra Rasiuk, Byram W. Bridle,

and Allan R. Willms
Abstract Cytokine storms are a potentially fatal exaggerated immune response

consisting of an uncontrolled positive feedback loop between immune cells and
cytokines. The dynamics of cytokines are highly complex and little is known about
specific interactions. Researchers at the Ontario Veterinary College have encoun-
tered cytokine storms during virotherapy. Multiple mouse trials were conducted
where a virus was injected into mice whose leukocytes lacked expression of the type
I interferon receptor. In each case a rapid, fatal cytokine storm occurred. A nonlinear
differential equation model of the recorded cytokine amounts was produced to
obtain some information on their mutual interactions. Results provide insight into
the complex mechanism that drives the storm and possible ways to prevent such
immune responses.
1 Introduction
Overly robust cytokine responses are responsible for a broad array of very chal-
lenging and often fatal clinical conditions. These include infectious diseases such
as influenza [16], severe acute respiratory syndrome (SARS) [11], bacteria-induced
toxic shock syndrome [20] and sepsis [9], as well as multiple sclerosis [13], graft-
versus-host disease [7] and sometime adverse effects of therapies [22]. Cytokines
are released by cells to coordinate an immune response to help protect against
foreign and/or dangerous matter. They are produced in response to infection and
inflammation [5]. There are five major classifications of cytokines; interferons
(IFN), interleukins (IL), chemokines, colony-stimulating factors (CSF), and tumour
necrosis factors (TNF) [23]. Together they function in order to stimulate a response
that will control cellular stress as well as minimize the amount of damage to a
particular cell or group of cells [5].
M. Waito () • A.R. Willms

S.R. Walsh • A. Rasiuk • B.W. Bridle
Department of Pathobiology, University of Guelph, Guelph, ON, Canada

332 M. Waito et al.
Although the goal of each cytokine is to ultimately protect the host against
dangerous insults, it is not uncommon for this process to become unbalanced.
For reasons unknown, positive feedback loops can become up-regulated during an
immune response involving complex interaction between cytokines and immune
cells. The fine balance that is typical of healthy individuals is lost and gives rise to a
cascade which is termed a cytokine storm. The dynamics of cytokines during normal
immune responses and even more so during storms are highly complex and little
is known about specific interactions [5, 27]. Cytokine storms not only have very
negative effects on the immune response, they often have life threatening effects
including a decrease in blood pressure, increase in heart rate and can often lead to
death of the host [27].
As more biological data is collected, the immense complexities of the immune
system is becoming increasingly apparent [26]. It is clear that mathematical mod-
elling has become an essential tool to complement experimental techniques both in
vitro and in vivo. Combining these approaches can result in major advancements in
both the understanding of cancer and immune system dynamics including cytokine
interactions, with the potential to identify strategies to control toxic cytokine storms
[26].
Previous mathematical models include at most two cytokines and do not focus
the research on the interactions and function of the cytokines. Typical cytokines
included in models are IL-2, IL-10, IL-12 and TNF. Unfortunately, even the models
that contain cytokines do not investigate the specific interactions that we are
interested in. These models exclude many of the cytokines that are of greatest
interest to us [6, 12, 15, 19].
Since our focus is to model only the dynamics of cytokines, we turn to ordinary
differential equation (ODE) models to provide a good framework for exploring
interactions between different cytokine populations. To keep the model relatively
simple, we ignore the effects of effector cells (activated immune cells) due to the
complexities it brings to the model. To date, the only model that is of a similar
nature is produced by [27] called Dynamics of a Cytokine Storm. In this paper, the
interactions of nine cytokines were modelled using data from a human clinical trial
where six volunteers took part in a study that accidentally led them to undergo the
effects of a cytokine storm [27]. This model was primarily constructed to look at the
effects of the antibody responsible for the cytokine storm using a linear ODE model
where parameter estimation determined coupling parameters between cytokines.
The coupling parameters identified which cytokine was responsible for enhancing
or inhibiting each other cytokine as well as self-regulation [27].
In this paper, we determine how cytokines interact with one another based on a
set of time series data provided by Dr. Byram Bridle and his lab members Dr. Scott
Walsh and Alexandra Rasiuk at the Ontario Veterinary College (OVC). Following
administration of a highly attenuated virus to mice with leukocytes lacking the
type I IFN receptor, a deadly cytokine storm developed leading to death in only
24 h. Down-regulation of anti-viral IFN signaling is a common mechanism used
by viruses during infection, including those associated with cytokine storms (e.g.
influenza virus [14], SARS-coronavirus [3] and Ebola virus [2]). Thus, a natural
A Mathematical Model of Cytokine Dynamics During a Cytokine Storm 333
question arises: what are the specific dynamics of cytokines with respect to both
initiating and exacerbating a cytokine storm when type I interferon signalling is
impaired?
Most biological systems are innately non-linear and thus a non-linear ODE model
is essential to obtain the proper dynamics [26]. The model we use is given by:
Mi
xP i D i xi C for i; j D 1; 2; : : : ; 13: (1)
1 C eyi
The first and most common assumption throughout the literature is that the
concentration of a particular cytokine, xi , continues to decrease linearly when there
is no outside stimulus [10, 17]. Since cytokines are secreted by other cells it makes
sense that in the absence of these producers, the number of cytokines will rapidly
decline at a rate i (>0) [10, 17, 25]. The second assumption is that the rate of
production of cytokines is dependent on interactions with other cytokines and is
sigmoidal in shape of the form
Mi
;
1 C eyi
where Mi (>0) is the maximum production rate. The interaction with other
cytokines, yi , determines the slope of the function, and thus how much of a cytokine
is produced, potentially offsetting some or all of the decay [10, 17, 25]. The
interaction factor yi is given by
X
n
yi D ˛0;i C ˛i;j xj C Si ; (2)
jD1
meaning that yi is composed of effects from the presence of other cytokines, ˛i;j ;
effects from components for which there is no data, ˛0;i ; and the stimulus, Si , by the
virus. The stimulus is of the form
Si D bet= ;
where there is an initial dose of the drug, b, which decays exponentially with
characteristic time .
334 M. Waito et al.
3 Data
Data was collected and provided by Dr. Byram Bridle, Dr. Scott Walsh and
Alexandra Rasiuk who study the role of type I IFN signalling in the regulation
of cytokine responses at the OVC. Chimeric mice were made by lethal irradiation
of the bone marrow of C57BL/6 mice (Charles River Laboratories) followed by
reconstitution with bone marrow from either wild-type or type I IFN receptor
(IFNAR)-knockout donors (the latter provided by Laurel Lenz, University of
Colorado School of Medicine). These mouse-based experiments were approved
by the institutional Animal Care Committee and complied with the standard of
the Canadian Council of Animal Care. These mice were infected intravenously
with recombinant Vesicular Stomatitis Virus with a deletion of methionine at
position 51 of the matrix protein (VSV m51) [21]. The matrix protein of VSV
suppresses antiviral type I IFN responses. Therefore, this mutant virus renders the
already attenuated laboratory strain of VSV even safer and it is being developed
as an oncolytic virus for the treatment of cancers via intravenous infusion [24].
Surprisingly, mice lacking the IFNAR on their leukocytes experienced a profound
cytokine storm, ultimately leading to death in only 24 h. The resulting time series
data provided concentrations of 13 different cytokines in plasma, measured using a
multiplex array (BioRad).
Concentrations of 13 cytokines were recorded at times 0, 2.5, 5, 10 and
24 h for both the wild-type mice and mice lacking the IFNAR on 20% of their
leukocytes. Raw data was then normalized to account for the vast variability in the
concentrations. This was done by dividing the concentration at each time point by
the sum of the concentrations across all time points for a particular cytokine. In order
to produce an accurate fit to the model it was essential to group cytokines to reduce
the number of parameters. Grouping was based on the inflammatory classification
of each cytokine as well as similarity of cytokine profile. The time at which the peak
concentration occurred was recorded along with either the pro- or anti-inflammatory
classification. Groupings can be seen in Table 1. Parameter estimation techniques,
specifically fmincon in MATLAB, were used to fit the model to the data. The cost
Table 1 Grouping of 13 cytokines based on raw data and inflammatory classification

Inflammatory WT data peak KO data peak
Group Cytokines classification (h) (h)
1 IL-13, IL-5 Anti 2:5 24
2 MIP-1ˇ, TNF-˛, IL-12 Pro 2:5 24
3 IL-1ˇ, eotaxin, IFN- Pro 5 24
4 IL-4, IL-10 Anti 5 24
5 IL-6 Pro/Anti 2:5 10
6 MIP-1˛ Pro 5 24
7 Rantes Pro 10 24
function is the typical least squares function with a normalization matrix to offset
discrepancies between groups.
4 Results
Mathematical analysis is an important tool for biological predictions and for

producing more questions and further areas of research both in biology and
mathematics [1]. Model results from parameter optimization of IFNAR-knockout
mice data are shown in Fig. 1. The model provides an accurate fit to the cytokine
time series data for both IFNAR-knockout and wild-type mice (not shown). The
model parameters that provide the most interesting information are the ˛i;j values, as
they relate how each cytokine group interacts with each of the other groups. The sign
and magnitude of these values indicate the type of effect a particular cytokine has; a
negative parameter value indicates an inhibitory effect, a positive value an enhancing
effect and the largest magnitude produces the most significant coupling. The three
largest ˛i;j values for each group of the IFNAR-knockout mice are displayed in
Fig. 2 where the lines indicate either enhancing (solid line with an arrowhead) or
inhibiting (dotted line with a ‘T’) and the line thickness displays the magnitude (the
thicker the line, the greater the magnitude).
Preliminary results from Fig. 2 show that Groups 2, 4, 5 and 6 are the most
significant. Each of them have a large number of interactions, multiple two-way
paths, exhibit definite enhancement or inhibition, and fit well with what is known
biologically.
The most significant group, Group 2, consists of three pro-inflammatory
cytokines: MIP-1ˇ, TNF-˛ and IL-12. Of the three, TNF-˛ is assumed to be the
dominating cytokine. TNF has been thoroughly studied and is one of the most well
known cytokines as it plays an important role in the outcome of immune function
GROUP 1 GROUP 2 GROUP 3 GROUP 4

0.4 0.4 0.5 0.8
0.4
0.3 0.3 0.6
Concentration
Concentration
Concentration
Concentration
0.3
0.2 0.2 0.4
0.2
0.1 0.1 0.2

0.1
0 0 0 0
0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25
Time (hours) Time (hours) Time (hours) Time (hours)
GROUP 5 GROUP 6 GROUP 7

0.8 0.35 0.8
0.6 0.3 0.6

Concentration
Concentration
Concentration
0.4 0.25 0.4
0.2 0.2 0.2
0 0
0 5 10 15 20 25 0 5 10 15 20 25 0 5 10 15 20 25
Time (hours) Time (hours) Time (hours)
Fig. 1 Model results from parameter estimation of IFNAR-knockout mice. Dots represent raw
time series data, lines show the model prediction
336 M. Waito et al.
GROUP
3
GROUP CYTOKINES
GROUP GROUP
1 IL-13, IL-5
2 4
2 MIP- , TNF- , IL-12
3 IL-1 , eotaxin, IFN-

GROUP
1 4 IL-4, IL-10
5 IL-6
6 MIP-1
GROUP GROUP
7 5 7 Rantes
GROUP
6
Fig. 2 Largest three interactions for each cytokine group of IFNAR-knockout mice are repre-
sented with a line. Solid lines with an arrow indicate enhancing effects while dotted lines with a
‘T’ indicate inhibitory effects
[18, 23]. Model results for Group 2 show that it is an integral part of six significant
interactions and has multiple two-way paths, one with Group 5 and another with
Group 7. All three interactions that affect Group 2 are enhancing, indicating that as
the storm progresses, the amount of cytokine will continue to rise. TNF-˛, along
with IL-1ˇ (found in Group 3), are considered early-response cytokines, occurring
soon after an immune response is triggered. It is known biologically that TNF-˛
promotes the generation of IL-1ˇ [23]. Although the results are not shown for the
wild-type case, the model results indicate that in fact Group 2 does enhance Group
3. In the case of IFNAR-knockout mice however, Group 2 inhibits Group 3 (the
relatively small interaction is not shown in Fig. 2). This difference begins to shed
light on why a storm occurred in mice lacking IFNAR on leukocytes.
Another central group is Group 4 which is made up of two anti-inflammatory
cytokines, IL-4 and IL-10, with the latter being a very prominent inhibitor. IL-10 is
often produced once a cytokine storm has begun in an attempt to return the balance
that has been lost, termed immunoparalysis. Although overproduction can often
allow the host to survive the cytokine storm longer, it is not likely it will survive long
term [8, 18, 23]. Model results for Group 4 show that there are six interactions that
are significant as well as two primarily enhancing two-way interactions with Groups
3 and 5. The three interactions that affect Group 4 are all enhancing, meaning that
the amount of Group 4 will likely increase as the storm continues. Biologically it is
known that IL-10 plays a role in the down regulation of both TNF-˛ (Group 2) and
IL-1ˇ (Group 3) [18]. Referring to Fig. 2, the model verifies that in fact Group 4
does inhibit Group 2, however Group 4 enhances Group 3. This could be due to the
grouping of cytokines, since Group 2 contains MIP-1ˇ and IL-12 as well. On the
contrary, in the wild-type case, Group 4 inhibits Group 3 while it enhances Group
2.
A well-studied cytokine that is known to be a key component of a cytokine storm
is IL-6, Group 5, which has both pro- and anti-inflammatory properties and is a
central cytokine used to assess cytokine responses in the host [23]. Figure 2 displays
the importance of this group with the eight interactions and two primarily enhancing
two-way paths. It is known biologically that the production of IL-6 is stimulated by
TNF-˛ and IL-1ˇ [23]. Results from the model using wild-type data verify that
indeed both Groups 2 and 3 enhance Group 5, however the interaction between
Group 3 and 5 is relatively small. IFNAR-knockout results shown in Fig. 2 imply
that Groups 2 and 3 inhibit Group 5 and that those interactions are significant. Again,
this can provide some insight into how a storm was able to occur.
Group 6, MIP-1˛, is required for a typical inflammatory response to viruses [4].
It is a pro-inflammatory chemokine that inhibits proliferation of hematopoietic stem
cells in vitro and in vivo [4]. Model results for MIP-1˛ show that six interactions
are significant as well as a two-way interaction with Group 3. The three interactions
that affect MIP-1˛ are both inhibiting and enhancing whereas in the wild-type case
the interactions are purely enhancing, causing an increase in the amount of MIP-1˛
as the storm continues.
It has been noted that the fine balance of pro- and anti-inflammatory mechanisms
is critical in maintaining stability, and if these mechanisms become unbalanced,
the outcome may contribute to a cytokine storm [23]. Groups 1 and 4 are anti-
inflammatory cytokines, while the remainder act primarily as pro-inflammatory
cytokines. For wild-type populations, of the three most significant groups (2, 4, and
6), the anti-inflammatory cytokines are being inhibited, while the pro-inflammatory
cytokines are being enhanced. This balance in this system becomes lost in mice
lacking the IFNAR on their leukocytes, as shown in Fig. 2. Instead of the anti-
inflammatory cytokines (Group 4) being inhibited, they are instead enhanced.
Immediately it becomes apparent that the fine balance that is typical of healthy
individuals has become unstable.
Future work including sensitivity analysis of cytokine groupings could provide
further information on the significance of the groupings and individual cytokines.
In conclusion, cytokines belonging to Groups 2, 4, 5 and 6, particularly TNF-˛,
IL-10, IL-6 and MIP-1ˇ, have the largest effects on the dynamics of this particular
cytokine storm. Changes introduced into the system by knocking out IFNAR cause
key interactions to swap from enhancing to inhibiting and vice versa. It is possible
that reducing the alterations in the effects of Groups 2, 4, 5 and 6 could lead to the
reduction in severity and possibly even the entire storm.
Acknowledgements This research was supported in part by the Government of Ontario and the
University of Guelph.
338 M. Waito et al.
References
1. Bellomo, N., Li, N.K., Maini, P.K.: On the foundations of cancer modelling: selected topics,
speculations, and perspectives. World Sci. 18(4), 593–646 (2008)
2. Chang, T.-H., Kubota, T., Matsuoka, M., Jones, S., Bradfute, S.B., Bray, M., Ozato, K.:
Ebola zaire virus blocks type i interferon production by exploiting the host sumo modification
machinery. PLoS Pathog. 5(6), e1000493 (2009)
3. Chen, X., Yang, X., Zheng, Y., Yang, Y., Xing, Y., Chen, Z.: Sars coronavirus papain-like
protease inhibits the type i interferon signaling pathway through interaction with the sting-
traf3-tbk1 complex. Protein Cell 5(5), 369–381 (2014)
4. Cook, D.N.: The role of mip-1˛ in inflammation and hematopoiesis. J. Leucoc. Biol. 59(1),
61–66 (1996)
5. Dranoff, G.: Cytokines in cancer pathogenesis and cancer therapy. Nature 4(1), 11–22 (2004)
6. Eftimie, R., Bramson, J.L., Earn, D.J.D.: Interactions between the immune system and cancer:
a brief review of non-spatial mathematical models. Bull. Math. Biol. 73(1), 2–32 (2011)
7. Ferrara, J.L.: Cytokine dysregulation as a mechanism of graft versus host disease. Curr. Opin.
Immunol. 5(5), 794–799 (1993)
8. Frazier, W.J., Hall, M.W.: Immunoparalysis and adverse outcomes from critical illness. Pediatr.
Clin. N. Am. 55(3), 647–668 (2008)
9. Harrison, C.: Sepsis: calming the cytokine storm. Nat. Rev. Drug Discov. 9(5), 360–361 (2005)
10. Hone, A., van den Berg, H.: Mathematical analysis of artificial immune system dynamics and
performance. In: Flower, D., Timmis, J. (eds.) In Silico Immunology, pp. 351–374. Springer,
New York (2007)
11. Huang, J., Wu, S., Barrera, J., Matthews, K., Pan, D.: The Hippo signaling pathway coordi-
nately regulates cell proliferation and apoptosis by inactivating Yorkie, the drosophila homolog
of YAP. Cell 122(3), 421–434 (2005)
12. Kirschner, D., Panetta, J.C.: Modeling immunotherapy of the tumor-immune interaction. J.
Math. Biol. 37, 235–252 (1998)
13. Link, H.: The cytokine storm in multiple sclerosis. Mult. Scler. 4(1), 12–15 (1998)
14. Medina, R.A., Garcia-Sastre, A.: Influenza a virus: new research developments. Nat. Rev.
Microbiol. 9, 590–603 (2011)
15. Müller, J., Tjardes, T.: Modeling the cytokine network in vitro and in vivo. J. Theor. Med. 5(2),
93–110 (2003)
16. Pothlichet, J., Meunier, I., Davis, B.K., Ting, J.P.-Y., Skamene, E., von Messling, V.,
Vidal, S.M.: Type I IFN triggers RIG-I/TLR3/NLRP3-dependent inflammasome activation in
influenza A virus infected cells. PLoS Pathog. 9(4), e1003256 (2013)
17. Read, M., Timmis, J., Andrews, P.S.: Empirical investigation of an artificial cytokine network.
In: Bentley, P.J., Lee, D., Jung, S. (eds.) Artifical Immune Systems, pp. 340–351. Springer,
Berlin/Heidelberg (2008)
18. Reis, E.A.G., Hagan, J.E., Ribeiro, G.S., Teixeira-Carvalho, A., Martins-Filho, O.A., Mont-
gomery, R.R., Shaw, A.C., Ko, A.I., Reis, M.G.: Cytokine response signatures in disease
progression and development of severe clinical outcomes for leptospirosis. PLoS Negl. Trop.
Dis. 7(9), e2457 (2013)
19. Rodríguez-Pérez, D., Sotolongo-Grau, O., Riquelme, R.E., Sotolongo-Costa, O., Antonio
Santos Miranda, J., Antoranz, J.C.: Assessment of cancer immunotherapy outcome in terms
of the immune response time features. Math. Med. Biol. 25, 287–300 (2008)
20. Silversides, J.A., Lappin, E., Ferguson, A.J.: Staphylococcal toxic shock syndrome: mecha-
nisms and management. Curr. Infect. Dis. Rep. 12(5), 392–400 (2010)
21. Stojdl, D.F., Lichty, B.D., tenOever, B.R., Paterson, J.M., Power, A.T., Knowles, S., Marius, R.,
Reynard, J., Poliquin, L., Atkins, H., Brown, E.G., Durbin, R.K., Durbin, J.E., Hiscott, J., Bell,
J.C.: Vsv strains with defects in their ability to shutdown innate immunity are potent systemic
anti-cancer agents. Cancer Cell 4(4), 263–275 (2003)
22. Suntharalingam, G., Perry, M.R., Ward, S., Brett, S.J., Castello-Cortes, A., Brunner, M.D.,
Panoskaltsis, N.: Cytokine storm in a phase 1 trial of the anti-cd28 monoclonal antibody
tgn1412. N. Engl. J. Med. 355(10), 1018–1028 (2006)
23. Tisoncik, J.R., Korth, M.J., Simmons, C.P., Farrar, J., Martin, T.R., Katze, M.G.: Into the eye
of the cytokine storm. Microbiol. Mol. Biol. Rev. 76(1), 16–32 (2012)
24. Wollmann, G., Rogulin, V., Simon, I., Rose, J.K., van den Pol, A.N.: Some attenuate variants of
vesicular stomatitis virus show enhanced oncolytic activity against human glioblastoma cells
relative to normal brain cells. J. Virol. 84(3), 1563–1573 (2010)
25. Wu, J.T., Kirn, D.H., Wein, L.M.: Analysis of a three-way race between tumor growth, a
replication-competent virus and an immune response. Bull. Math. Biol. 66, 605–625 (2004)
26. Yates, A., Chan, C.C.W., Callard, R.E., George, A.J.T., Stark, J.: An approach to modelling in
immunology. Brief. Bioinform. 2(3), 245–257 (2001)
27. Yiu, H.H., Graham, A.L., Stengel, R.F.: Dynamics of a cytokine storm. PLOS One 7(10),
e45027 (2012)
Examining the Role of Social Feedbacks
and Misperception in a Model of Fish-Borne
Pollution Illness
Michael Yodzis, Chris T. Bauch, and Madhur Anand
Abstract Pollution-induced illnesses are caused by toxicants that result from

human activity and should be entirely preventable. However, social pressures
and misperceptions can undermine the efforts to limit pollution, and vulnerable
populations can remain exposed for decades. We present a human-environment
system model for the effects of water pollution on the health and livelihood of a
fishing community. The model includes dynamic social feedbacks that determine
how effectively the population recognizes the injured and acts to reduce its pollution
exposure. Our work is motivated by a historical incident from 1949 to 1968 in
Minamata, Japan where methylmercury effluent from a local factory poisoned
fish populations and humans who ate them. We will discuss the conditions that
allow for the outbreak of a pollution-induced epidemic, and explore the role that
misperception plays in allowing it to persist.
1 Introduction
A human-environment system (HES) is characterized by the interaction of human

activities and natural processes [12, 13]. The HES model that we present in this
paper is developed in [20] in the context of the methylmercury-poisoning incident
in Minamata, Japan. It represents the transmission of pollution to humans through
fish-eating, and the social feedbacks to limit exposure through the boycott of fish-
eating and the demand for pollution control.
This model is motivated by several key observations. First is that the epidemio-
logical data from Minamata show that the community was unwilling to recognize
new pollution victims from 1960 to 1971 [7, 8]. Historical accounts describe the
dependence of the community on the polluting industry, and how this dependence
M. Yodzis () • M. Anand

University of Guelph, Guelph, ON, Canada
C.T. Bauch
University of Waterloo, Waterloo, ON, Canada

342 M. Yodzis et al.
fed a widespread stigmatization of the pollution victims and underestimation of the

health damages [5, 9, 17].
Our model depicts the voluntary actions that citizens must take to protect them-
selves when proactive leadership from the government or industry is lacking. Legal
theorists who study the regulatory compliance of polluting industries observe that
government policy may not be enforced effectively without pressure from citizens
[19]. Social misperception can be the single largest obstacle to resolving pollution
problems. Sometimes this is fuelled by an inconclusive scientific understanding,
given that the symptoms of pollution exposure can be slow to appear, and difficult
to separate from other causes or confounding variables [6, 7, 19]. Misperception
may also be fuelled by economic expediency and deliberate foul play. Therefore,
citizen action and changing social perception play a large role in helping to resolve
pollution problems.
Our objectives are to examine the ecological and social conditions that cause the
outbreak of a pollution-induced epidemic, and to study the role of social feedbacks
and misperception effects that allow the epidemic to persist.
2 Model
To focus on numerical simplicity and qualitative behaviour, in [20] we build a

system of normalized equations whose state variables are proportions rather than
absolute quantities. The pollution is driven by an input n.t/, which is the baseline
emissions loading rate from an industry into a water source. Whereas n.t/ is a
prescribed function, the net emissions loading E.t/ is affected by feedback from
the social demand for abatement. The fishery consists of fish F.t/ and boats B.t/,
whose interactions are governed by Lotka-Volterra type equations and emissions
mortality.
Human injuries are driven by the fish-borne pollution exposure. This exposure
depends on the net emissions level E.t/, the fish catch HF.t/B.t/ with harvest rate
H, and a factor of pollution availability to humans, "I .
To model human perception, we introduce P.t/ for the perceived injured and I.t/
for the actual injured. The parameter s is the stigma rate, or the rate at which the
community denies new victims as the perceived injured P.t/ increases. This stigma
effect causes a disparity between the actual and perceived number of injured. If
there is no stigma, s D 0, then I.t/ and P.t/ grow at the same rate. However, when
s > 0, then a distinct negative feedback is introduced in P0 : the rate at which the
community is willing to recognize new victims decreases as the number of perceived
victims increases. Then P.t/ is an underestimate of the actual health damages I.t/.
Observe that in the model, all social feedbacks and decisions are driven by P.t/
but are blind to I.t/. P.t/ affects the level of fish-eating, 1 bP.t/, where b is
the boycott rate. P.t/ also influences the public demand for abatement X.t/, which
grows through social learning in the form of an imitation game [10]. We assume
that citizens sample the preferences of others and aim to minimize a cost function,
Social Feedbacks in a Model of Fish-Borne Pollution Illness 343
hP.t/ ˇX.t/, by choosing either to favour abatement or non-abatement. is a

sampling and imitation rate, h measures the health costs from pollution, and ˇ
reflects costs to productivity and job security that is claimed by the industry and its
supporters in response to growing X. In turn, the population preference X.t/ feeds
back to influence the net emissions E.t/, thus closing the loop.
Altogether, the equations for the human-environment system are:
9
F 0 .t/ D rF.t/.1 F.t// HF.t/B.t/ "F E.t/F.t/ >
>
>
>
B0 .t/ D .1 bP.t//HF.t/B.t/ cB.t/ >
=
I 0 .t/ D "I E.t/HF.t/B.t/ Œ1 bP.t/ I.t/ (1)
>
>
P0 .t/ D "I E.t/HF.t/B.t/ Œ1 bP.t/ sP.t/ P.t/ >
>
>
;
X 0 .t/ D X.t/.1 X.t// ŒhP.t/ ˇX.t/
where the net emissions loading E.t/ is given by:
E.t/ D n.t/ .1 f .X.t/// (2)
Note that the form of the abatement function f .X/ is important. It should be
monotone increasing in X, and have a social concern threshold xM above which
there is complete abatement: i.e. whenever X xM , then f .X/ D 1 and E.t/ D 0.
Table 1 gives a summary of the model variables and parameters, including the values
selected in [20] to simulate the historic pollution epidemic in Minamata. In these
simulations, f is chosen to be a nonlinear function with xM D 0:4,
With the appropriate choice of initial conditions, we can constrain the solutions
of (1) to an invariant region ˝ in the non-negative orthant of the FBIPX state space,
( )
.F; B; I; P; X/ W F; X 2 Œ0; 1; B 0;
˝ WD 1Cs
1
(3)
I 2 0; 1CbCs ; P 2 0; 1CbCs
This result is proved in the supplementary appendix of [20] using standard

invariance theorems for systems of ordinary differential equations.
Pollution Epidemic The system (1) has nine equilibria, with a single nontrivial
equilibrium:
0 1 0 c 1
F H.1bP /
B B C B 1 .r .1 F / " .1 f .X /// C
B C BH F C
B C B 1Cs C
B I C D B 1CbCs C (4)
B C B 1 C
@ P A @ 1CbCs A
X h
ˇ.1CbCs/
This equilibrium corresponds to a pollution epidemic steady-state, as it occurs

when the emissions and fish catch coexist.
Table 1 Model variables and parameters

Variable Meaning Range I.C.a
E Net emissions loading Œ0; 1 0.0
F Fish Œ0; 1 1.0
B Boats 0 0.1
I Cumulative injured Œ0; 1 0.0
P Cumulative perceived injured Œ0; 1 0.0
X Demand for pollution abatement Œ0; 1 0.01
f .X/ Abatement level Œ0; 1 0.0
Parameter Meaning Range Baselineb
H Harvesting rate Œ0; 1 0.7
r Fecundity .0; 1 1.0
c Normalized boat costs .0; 1 0.35
"F Fish pollution mortality Œ0; 1 0.7
"I Pollution availability to humans Œ0; 1 0.009
b Rate of fish boycott per unit injury 0 100
s Rate of stigmatization per unit injury 0 600
Sampling and imitation rate Œ0; 1 0.5
h Rate of health concern per unit injury 0 1000
ˇ Pushback to demand for abatement 0 4.0
xM Social concern threshold Œ0; 1 0.4
a
Initial Condition values selected for simulation in [20]
b
Baseline parameter values selected for simulation in [20]
Existence and Stability Conditions This equilibrium cannot occur in ˝ unless the
following three inequality conditions hold: the social concern is below the threshold
needed to abate the emissions,
h
0 < xM (5)
ˇ.1 C b C s/
the fish reproduction rate exceeds the pollution mortality,
r > "F .1 f .X // (6)
and the returns from the harvest exceed the cost of fishing,

"F .1 f .X //
c < H.1 bP / 1 : (7)
r
Using these same inequalities, it is shown in [20] that when the nontrivial
equilibrium exists in ˝ it is locally attractive there.
Convergence of Solutions to the Pollution Epidemic Equilibrium Even if a pollution
epidemic equilibrium exists and is locally attractive in ˝, it will not necessarily
be reached. The existence and local stability conditions alone are not enough to
guarantee convergence to steady-state, since the system (1) is highly coupled and
nonlinear. The convergence behaviour of this system is sensitive to the relative
growth rates of each variable, particularly those that affect the viability of the fishery.
Solutions converge to the pollution epidemic equilibrium if the injuries P.t/ and
social concern X.t/ grow fast enough that the fish catch F.t/B.t/ attains its steady-
state. Whereas when they grow too slowly, the fish catch F.t/B.t/ will collapse.
These observations are illustrated in the following schematic:
P.t/ ! P P.t/ ! P

FAST SLOW
X.t/ ! X X.t/ ! X
FAST SLOW
F.t/B.t/ ! F B F.t/B.t/ ! 0
converges collapses
Due to the nonlinearity of the system, we are not able to characterize the
convergence behaviour more precisely in terms of the model parameters, as we
were in (5), (6) and (7). However, there are insights to be gained from writing each
variable in terms of its associated integral equation, evaluated up to some time T:
ZT
P.T/ D P.t0 / C "I H F.t/ B.t/ .1 f .X.t/// Œ1 .1 C b C s/ P.t/ dt
t0
ZT
X.T/ D X.t0 / C X.t/ .1 X.t// Œh P.t/ ˇ X.t/ dt
t0
ZT ZT
F.T/ D F.t0 / C r F.t/ .1 F.t// H F.t/ B.t/ dt "F F.t/ .1 f .X.t/// dt
t0 t0
ZT
B.T/ D B.t0 / C .1 b P.t//H F.t/ B.t/ c B.t/ dt
t0
Observe that these integral equations depend on the initial conditions and the
parameter values "I and , whereas the inequalities (5), (6) and (7) do not. We see
that the fishery is sensitive to the time evolution of P.t/ and X.t/, because P.t/
affects both the demand for fish 1 bP.t/ and the demand for abatement X.t/, while
X.t/ affects the emissions. In general, during the time that P.t/ and X.t/ are small,
both the number of boats and the level of emissions are high. If the fish population
is driven below the break-even cost required for the boats, the fish catch collapses
and an epidemic is averted.
3 Analysis
In [20] our model is applied to simulate the pollution epidemic that occurred from
1949 to 1968 in Minamata, Japan. Given that some of the datasets from Minamata
are incomplete or sampled at irregular time intervals, the simulations are meant to
agree qualitatively with the available data, and to identify a set of baseline parameter
values that are physically plausible. These baseline values displayed in Table 1. The
simulations run from 1945 to 1976.
We are interested in examining how the social feedbacks, driven by the parame-
ters s, b and h, affect the resulting steady-state. The system (1) has nine equilibria
that we class into four types: (I) No fish catch, no emissions; (II) Fish catch, no
emissions; (III) No fish catch, emissions; (IV) Fish catch and emissions. Type IV is
a nontrivial equilibrium, a pollution epidemic steady-state.
Figure 1 depicts the steady-states of the system solved in s-b parameter space,
with all other parameters held at their baseline values. We numerically solve F, B
and X for large time t D 1;000;000 to approximate the steady-state, and then plot it
in s-b space coloured according to its equilibrium type.
We are also able to test our theoretical expectations against the numerical results.
In Fig. 1, the hatched region demarcates where there is potential for a pollution
epidemic according to the inequalities (5), (6) and (7). We find that this region
overlaps exactly with the numerically solved steady-state region in dark grey, which
indicates that a pollution epidemic has occurred.
As the parameter plane in Fig. 1 shows us, an epidemic does not occur unless
the stigma s is sufficiently large. It is very important to recognize how unlikely
the conditions for a pollution epidemic (type-IV) equilibrium are without social
stigma/misperception.
Fig. 1 Changing values for the social feedbacks s (stigma) and b (boycott) yield alternative
qualitative outcomes. The parameter plane shows the dynamical outcomes defined by the
equilibrium-types I, II, III, and IV that result for various values of s and b. The hatched region
signifies where a potential epidemic is possible, and the solid point represents the baseline values
used in the simulation of Minamata; s D 600, b D 100
We want to know what parameters are required for us to have a pollution

epidemic when s D 0, and how physically plausible they are compared to the
baseline parameter values used in the simulation of Minamata. Figure 1 depicts the
axis s D 0 in black. To find when a pollution epidemic becomes possible without
stigma, we must vary our parameters so that the inequalities (5), (6) and (7) are
satisfied on the axis s D 0. To find when this happens, we can consider the boundary
case of (5) with X D xM ,
h
D xM (8)
ˇ.1 C b/
so that (6) reduces to r > 0 and (7) reduces to

1
c<H (9)
1Cb
Combining (8) and (9) yields
h HMx
< (10)
ˇ c
Varying each of these parameters individually while holding the others at their
baseline values from Table 1, we find that the potential for a pollution epidemic
exists for some b if either the health concern is decreased to 0 < h < 3:2, the boat
costs are decreased to 0 < c < 0:00112, or the fish harvesting rate is increased to
H > 218:75. Note that in the ratio h=ˇ, it suffices to consider changing h only. As
the numerical results in Fig. 2 show, the time evolution of the solutions is sensitive
to the collapse of the fishery, and they do not necessarily converge to the pollution
epidemic equilibrium.
When the health concern is decreased to h D 3 a pollution epidemic becomes
possible for 0:25 < b < 0:55. However, as Fig. 2a demonstrates, small values of h
and b make the social response to the pollution negligible. The fish catch collapses
before the injuries grow large enough to trigger any abatement, and a pollution
epidemic is averted.
Decreasing the boat costs to c D 0:001, a pollution epidemic becomes possible
for values of b > 624. The numerical results in Fig. 2b indicate that the pollution
epidemic occurs. However, is this plausible? The accompanying plot in Fig. 2b
suggests that this operating cost for the boats is unrealistically low. The boats grow
to an unrealistically high level and push the fish population toward zero. For the
short time-scale we are interested in, the potential for a pollution epidemic is averted
by the decline of the fish catch to near-zero levels. If we solve numerically over a
very large time scale as t ! 1, it turns out that the fish and boats exhibit damped
oscillations, and the injuries grow.
Fig. 2 We seek the conditions for which a pollution epidemic is possible with no stigma. The
panels depict the s-b parameter planes that result when (a) h D 3, (b) c D 0:001 and (c) H D 220.
Each is coloured according to the legend in Fig. 1. The solid dot in each plane represents the values
of s and b used to generate each of the plots below, which show the time evolution of the fish (in
dark grey) and boats (in light grey)
Another option is to increase the harvest rate to H D 220 with b > 624. However,
as the plots in Fig. 2c show, this unsustainable harvest pressure causes the fish catch
to collapse before the variables are able to reach a pollution epidemic steady-state.
To examine the effects of the parameters r and "F on the occurrence of a pollution
epidemic, we return to the inequalities (5), (6) and (7). Allowing r, "F and b to
be free while holding s D 0 and all other parameters at baseline, we find that
the inequalities are satisfied when r > 0 and "F < 0. However, negative values
for the pollution mortality rate of fish violate our parameter requirements and are
biologically implausible.
Instead, we explore the case "F D 0 so that the fish do not die from pollution,
while freeing r and another parameter, h. Holding s D 0 and all other parameters
at baseline, we find that the inequalities (5), (6) and (7) are satisfied if r > 0, 0 <
h < 3:2, and 0 < b < 1. Figure 3a shows that an epidemic does occur for the values
r D 1, h D 1 and b D 0:9. To determine whether this is realistic, we look to the time
series plots in Figs. 3b–e. The emissions and fish catch coexist and the injuries grow
steadily. Since s D 0, all of these injuries are fully perceived. However, the growth
of the social concern X and the reduction of fish-eating bP remain negligible,
because the social feedbacks h and b are very small. Assuming that the nature of the
pollution injuries is serious, like those experienced from methylmercury-poisoning
in Minamata, this situation of full perception without response is socially unrealistic.
Fig. 3 We seek the conditions for which a pollution epidemic is possible with no stigma. Panel
(a) depicts the s-b parameter plane that results when "F D 0 and h D 1. The plane is coloured
according to the legend in Fig. 1, and the black point gives the values s D 0 and b D 1 used to
generate the accompanying time series plots: (b) emissions, (c) social concern, (d) fish (in dark
grey), boats (in light grey), (e) perceived and actual injuries
4 Discussion and Conclusion
The human-environment system model studied in this paper allows us to understand

how social feedbacks that protect people from pollution rely on accurate informa-
tion. Our model is capable of reaching a pollution epidemic steady-state where the
pollution exposure continues indefinitely. We conclude that the pollution epidemic
equilibrium is unlikely to occur in the absence of stigma/misperception, except
under conditions that are socially or biologically implausible, or which violate our
parameter constraints. This highlights the importance of including misperception as
a factor in the model.
Our HES model is based on interactions that are endogenous and localized;
there are no external shocks, and no corrective mechanisms to move the system
away from a pollution equilibrium once it settles there. The very existence of a
pollution epidemic equilibrium, and the role played by misperception in augmenting
its occurrence, reveals that our model can fail to be self-regulating. This reflects
real-world circumstances in which citizen stakeholders have information that is
inaccurate or different to that held by industry, rendering citizens’ voluntary
protective actions inadequate to resolve the problem. Our HES model represents
the conditions that might have persisted for decades in Minamata if a second
methylmercury-poisoning incident had not occurred in Niigata, Japan, to increase
citizens’ awareness and stimulate relief for the victims [1, 17].
In models where human activity interacts with natural processes, we believe
there is more room to study the extent to which human decisions can be made
from inaccurate information, and to study how this affects the qualitative outcome
of the model. For example, although medical researchers recognize the role of
stigma in prolonging epidemics, especially when it shames people from seeking
medical attention [4, 18], it is not clear that these insights have been readily taken
up in epidemiological models. Risk perception and behavioural strategies to avoid
infectious disease are more widely modelled [14], and some agent-based models
explicitly distinguish between injured and perceived injured [15]. Social learning
models are increasingly used in mathematical epidemiology to study vaccine scares
[3], and in ecology to study resource management and conservation [2, 11, 16].
Up to now, these methods have not been widely applied to understand pollution
illnesses, in spite of the fact that pollution is intimately bound to social activity.
We have worked to incorporate these insights into our model, and the feedbacks
involving stigma and fish-eating boycott represent a break from existing approaches.
References
1. Almeida, P., Stearns, L.B.: Political opportunities and local grassroots environmental move-
ments: the case of minamata. Soc. Probl. 45(1), 37–60 (1998)
2. Barlow, L.-A., Cecile, J., Bauch, C.T.. Anand, M.: Modelling interactions between forest pest
invasions and human decisions regarding firewood transport restrictions. PLoS ONE 9(4),
e90511 (2014)
3. Bauch, C.T., Bhattacharyya, S.: Evolutionary game theory and social learning can determine
how vaccine scares unfold. PLoS Comput. Biol. 8(4), 1–12 (2012)
4. Des Jarlais, D.C., et al.: Effects of toxicants on populations: a qualitative approach ii. First
order kinetics. Am. J. Public Health 96(3), 561–567 (2006)
5. George, T.S.: Minamata: Pollution and the Struggle for Democracy in Postwar Japan. Number
194 in Harvard East Asian Monographs. Harvard University Asia Center, Cambridge/London
(2001)
6. Grandjean, P., et al.: Adverse effects of methylmercury: environmental health research
implications. Environ. Health Perspect. 118(8), 1137–1145 (2010)
7. Harada, M.: Sachie, T., George, T.S. (English Trans.) (2004). Minamata Disease. Patients
Alliance, Tokyo, (1972)
8. Harada, M.: Smith, A. (English Trans.). Minamta disease: a medical report. In: Smith, E.,
Smith, A. (eds.) Minamata, pp. 180–192. Holt, Rinehart and Winston, New York (1975)
9. Harada, M.: Minamata disease: methylmercury poisoning in Japan caused by environmental
pollution. Crit. Rev. Toxicol. 25(1), 1–24 (1995)
10. Hofbauer, J., Sigmund, K.: 8.1. Imitation dynamics. In: Evolutionary Games and Population
Dynamics, pp. 86–86. Cambridge University Press, Cambridge (1998)
11. Innes, C., et al.: The impact of human-environmental interactions on the stability of forest-
grassland mosaic ecosystems. Nat. Sci. Rep. 3(2689), 1–10 (2013)
12. Levin, S., Clark, W.C.: Toward a science of sustainability. In: Center for International Develop-
ment Working Papers 196. John F. Kennedy School of Government, Harvard University (2010)
13. Liu, J., et al.: Coupled human and natural systems. AMBIO: J. Hum. Environ. – BioOne 36(8),
639–649 (2007)
14. Perisic, A., Bauch, C.: Social contact networks and disease eradicability under voluntary
vaccination. PLoS Comput. Biol. 5(2), 1–8 (2009)
15. Poletti, P.: Risk perception and effectiveness of uncoordinated behavioral responses in an
emerging epidemic. Math. Biosci. 238, 80–89 (2012)
16. Satake, A., et al.: Synchronized deforestation induced by social learning under uncertainty of
forest-use value. Ecol. Econ. 63(2), 452–462 (2007)
17. Smith, E., Smith, A.: Minamata. Holt, Rinehart and Winston, New York (1975)
18. Van Brakel, W.H.: Measuring health-related stigma – a literature review. Psychol. Health Med.
11(3), 307–334 (2007)
19. Van Rooij, B.: The people vs. pollution: understanding citizen action against pollution in China.
J. Contemp. China 19(63), 56–77 (2010)
20. Yodzis, M., Bauch, C.T., Anand, M.: Coupling fishery dynamics, human health and social
learning in a model of fish-borne pollution exposure. Sustain. Sci. 11, 179–192 (2016)
Part III
Computational Engineering
and Mathematical Foundation,
Numerical Methods and Algorithms
Stability Properties of Switched Singular
Systems Subject to Impulsive Effects
Mohamad S. Alwan, Humeyra Kiyak, and Xinzhi Liu
Abstract This paper addresses impulsive switched singular systems with nonlinear
perturbation term. The main theme is to establish exponential stability of the
systems where the impulses are of fixed time type and treated as perturbation.
We first establish the exponential stability of a single-mode impulsive systems
using the Lyapunov method. We have observed that if the underlying continuous
system is stable and the impulses are applied slowly, then it is guaranteed that
the impulsive system maintains the stability property. Later, a switched system
with impulsive effects is considered. The method of multiple Lyapunov function
and average dwell time switching signal are used. We have noticed that if all
subsystems are exponentially stable and the average dwell time is sufficiently large,
then the impulsive switched system is exponentially stable. Numerical examples
with simulations are given to illustrate the effectiveness of the proposed theoretical
results.
1 Introduction
Singular systems (or differential algebraic systems) have received considerable

attentions due their extensive applications in the areas of control systems, electrical
power systems, mechanical systems, and robotics. Readers may refer to [4, 4–
6, 9, 12]. The theory of singular system is well studied in the literature; see for
instance [3, 4, 6, 7, 12, 15], and [5].
On the other hand, the dynamics of many natural or engineering systems are
subject to abrupt changes, where the durations of these changes are sufficiently
small that can be reasonably approximated by instantaneous changes of states or
impulses. The resulting systems can be suitably modeled as an impulsive systems.
Nowadays, the applications of impulsive system have been found in many different
real world or man-made systems, such as mechanical and engineering systems
including mass-spring systems, robotics, or electrical circuits. It can also be found in
biological systems including the function of the heart and biological neural network,
M.S. Alwan () • H. Kiyak • X. Liu

Department of Applied Mathematics, University of Waterloo, Waterloo, ON, Canada

356 M.S. Alwan et al.
and epidemic disease models including pulse vaccination. The theory of impulsive
system is well presented and more application can be found in [1, 2, 11, 14, 16].
When applying the impulsive effects to the singular system, the resulting system
is called impulsive singular system. Many singular systems exhibit impulsive and
switching behaviors, which are characterized by switches of states and abrupt
changes at switching time; that is, the systems switch with impulse effects [1, 2,
11, 13].
The objective of the paper is to establish the exponential stability of a nonlinear
impulsive switched singular system. The nonlinearity is linearly bounded, and both
the impulsive effects and switching are applied to the system at fixed times. Using
the method of Lyapunov function for impulsive singular system, some sufficient
conditions are given to establish the stability property. On the other hand, multiple
Lyapunov function and the average dwell-time approach are used to determine some
sufficient conditions for impulsive switched singular system.
The rest of the paper is organized as follows: In Sect. 2, the problem formulation
and some background are given. In Sect. 3, we state and prove the main results.
Numerical examples are presented in Sect. 4 to clarify the proposed approach.
Finally, a conclusion of these results is given in Sect. 5.
2 Problem Formulation
Denote by Rn the n-dimensional Euclidean space with the norm k k, Rnxn the set of
all nxn square matrices and N the set of natural numbers. If A 2 Rnxn , denote by min
and max the
p minimum and maximum eigenvalues of A; respectively, and its norm
is kAk D max .AT A/; where the superscript T represents the matrix transpose. If
A is a positive definite matrix, then we write A > 0.
Consider the impulsive switched singular systems of the form
EPx.t/ D A .t/ x.t/ C g .t/ .t; x/; t ¤ tk ;

x.t/ D Bk x.t/; t D tk ; (1)
x.t0 / D x0 ;
where x 2 Rn is the system state variable, and A .t/ ; Bk ; E 2 Rnxn are system
coefficient matrices with E being singular with rank.E/ D r < n, the matrix
pairs .E; A .t/ / being regular, and Bk being constant matrices. The switching signal
W Œt0 ; 1/ ! S is a piecewise constant function taking values in a finite compact
set S D f1; 2; : : : ; Ng for some N 2 N. ftk g1kD1 are the impulsive times that form an
increasing sequence satisfying tk1 < tk and limk!1 tk D 1. x D x.tC / x.t /
where x.t / (and x.tC /) is the state just before (and just after) the impulsive action
with x.tC / D lims!tC x.s/. The solution x is assumed to be left-continuous, i.e.,
x.tk / D x.tk /. For i 2 S , gi .t; x/ W RC Rn ! Rn is piecewise continuous vector-
Stability of Singular Impulsive Systems 357
valued functions with gi .t; 0/ 0; t 2 RC to ensure the existence of solutions for

system (1).
When .t/ D for all t > t0 , system (1) is reduced to the following single-mode
impulsive singular system:
EPx.t/ D Ax.t/ C g.t; x/; t ¤ tk

x.t/ D Bk x.t/; t D tk ; k2N (2)
x.t0 / D x0 ;
where the subscript is dropped for simplicity of notation.

In the following, some definitions and theorems are introduced.
Definition 1 ([7, 8]) Matrix pair .E; A/ is regular if there exists a constant scalar
2 C such that det. E A/ ¤ 0: The matrix pair .E; A/ is said to be impulse free
if deg.det. E A// D rank.E/:
Definition 2 ([7, 17]) Regular system (1) is said to be exponentially stable if there
exist ˛; ˇ > 0 such that its state x.t/ satisfies kx.t/k ˛kx.t0 /keˇt ; for all t > t0 :
Definition 3 ([17]) System (1) is said to be E-exponentially stable if there exist
constants ˛; ˇ > 0 such that kEx.t/k ˛kEx.t0 /keˇt ; for all t > t0 :
Definition 4 ([7]) System (1) is admissible if it is stable and impulse-free.
Theorem 1 ([8]) Matrix pair .E; A/ is regular if and only if there exist two
Q1 1 2 1
2 and P D P P where Q 2 R
rn
nonsingular matrices Q; P such that Q D ,
Q
Q2 2 R.nr/n , P1 2 Rnr , P2 2 Rn.nr/ and the following standard decomposition
holds:

g .t; x1 ; x2 /
QEP D diag .Ir ; N/ ; QAP D diag .A1 ; Inr / ; Qg.t; x/ D 1 ;
g2 .t; x1 ; x2 /
(3)
where r D rank.E/, A1 2 Rrr , N 2 R.nr/.nr/ is a nilpotent matrix with nilpotent

index h.
Theorem 2 ([7]) The singular system EPx.t/ D Ax.t/ is stable if and only if .E; A/
has finite eigenvalues with negative real parts.
Theorem 3 ([7]) If system (1) is admissible, then for any Y > 0 there exists X > 0
satisfying ET XA C AT XE D ET YE:
3 Main Results
In this section, exponential stability of system (1) and (2) are discussed.
3.1 Singular Systems Subject to Impulsive Effects
Theorem 4 Assume that system (2) is impulse free, the eigenvalues of the matrix
pair .E; A/ have negative real parts, the singular matrix E and the matrix I C Bk for
I 2 Rnxn identity matrix are commutative, and kg.t; x/k kExk for < 2kXk
min .Y/
positive constant where X > 0 satisfying
AT XE C ET XA D ET YE (4)
for any Y > 0. Then, the trivial solution of the nonlinear singular impulsive system
(2) is exponentially stable if the following inequality holds:
ln ˛k .tk tk1 / 0; k D 1; 2; : : : (5)
max Œ.ICBk /T X.ICBk / min .Y/2kXk

where ˛k D min .X/
, 0 < < , and D max .X/
.
Proof Let x .t/ D x.t; t0 ; x0 / be the solution of the system (2). Define
$.t/ D V.x.t// D xT .t/ET XEx.t/; t ¤ tk
as a Lyapunov function candidate. Then, along the trajectory of (2),

$.t/
P D xT .t/ AT XE C ET XA x.t/ C gT .t; x/XEx.t/ C xT .t/ET Xg.t; x/
D xT .t/ET YEx.t/ C 2xT .t/ET Xg.t; x/ (6)
where ET YE D AT XE C ET XA: Since X > 0 and Y > 0, we have [10]
min .X/kExk2 xT ET XEx max .X/kExk2 (7)

min .Y/kExk2 xT ET YEx max .Y/kExk2 : (8)
Thus, using (7), (8), and kg.t; x/k kEx.t/k in (6) lead to

P min .Y/ 2kXk kEx.t/k2 $.t/;
$.t/ t 2 .tk1 ; tk ;
min .Y/2kXk min .Y/

where D max .X/ is a positive constant if < 2kXk . Then,
C
$.t/ $.tk1 /e .ttk1 / ; t 2 .tk1 ; tk : (9)
At t D tkC ; we have
$.tkC / D xT .tk /.I C Bk /T ET XE.I C Bk /x.tk / ˛k $.tk /; (10)
max Œ.ICBk /T X.ICBk /

where ˛k D min .X/
: Using (9) and (10), one may get
$.t/ ˛1 ˛2 : : : ˛k $.t0C /e .tt0 /

D $.t0C /˛1 e.t1 t0 / ˛2 e.t2 t1 / : : : ˛k e.tk tk1 / e.ttk / e./.tt0 / :
By assumption (5), we have
$.t/ $.t0C /e./.tt0 / ; t t0 ;
which implies that
min .X/kEx.t/k2 $.t/ $.t0C /e./.tt0 / max .X/kEx.t0C /k2 e./.tt0 /

) kEx.t/k KkEx.t0C /ke./.tt0 /=2 ; t t0 ; (11)
p
where K D max .X/=min .X/. Thus, the trivial solution of the system (2) is
E-exponentially stable.
Let

1 x1 .t/
P x.t/ D : (12)
x2 .t/
Then, it follows from the standard decomposition form that system (2) is equiva-
lent to
xP 1 D A1 x1 C Q1 g.t; x/ (13)
0 D x2 C Q2 g.t; x/ (14)
where x1 2 Rr and x2 2 Rnr . By (12), we have

x1 .t/ Ir 0 x1 .t/ x1 .t/
QEx.t/ D QEP D D : (15)
x2 .t/ 0 0 x2 .t/ 0
From (11) and (15), we have
kx1 k D kQEx.t/k KkQkkEx.t0C /ke./.tt0 /=2 (16)
which implies that x1 is exponentially stable.

We need to show that x2 is also exponentially stable. It follows from (14) and
kg.t; x/k kEx.t/k that
kx2 k D kQ2 g.t; x/k KkQ2 kkEx.t0C /ke./.tt0 /=2 : (17)
This means x2 is exponentially stable. From (16) and (17), the trivial solution of (2)
is exponentially stable. This completes the proof.
Remark 1 The validity of the matrix equation (4) guarantees that the Lyapunov
function is decreasing for all t 6D tk . While the inequality in (5) is made to ensure that
the impulses are applied slowly in order to maintain the stability of the impulsive
system in (2).
3.2 Switched Singular Systems Subject to Impulsive Effects
Theorem 5 For any i 2 S , assume that system (1) is impulse free, the eigenvalues
of the matrix pairs .E; Ai / have negative real parts, and kgi .t; x/k i kExk for i <
min .Yi /
2kXi k positive constant where Xi > 0 satisfying Ai Xi E C E Xi Ai D E Yi E for
T T T
any Yi > 0. Then, the trivial solution of (1) is exponentially stable if the following
assumptions hold:
(i) For any i; j 2 S there exists ˛k > 1 such that
.I C Bk /T ET Xj E.I C Bk / ˛k ET Xi E: (18)
(ii) For any t0 , the switching law satisfies
t t0
N.t; t0 / N0 C
Ta
where N.t; t0 / represents the number of switchings in .t; t0 /, and N0 and Ta are
the chatter bound and average dwell time to be defined, respectively.
Proof Let x .t/ D x.t; t0 ; x0 / be the solution of the system (1). Define
$i .t/ D Vi .x.t// D xT .t/ET Xi Ex.t/; t ¤ tk ;
as a Lyapunov function candidate for ith subsystem. Then, derivative of $i along the
trajectory of (1) is given by
$P i .t/ D xT .t/ET Yi Ex.t/ C 2xT .t/ET Xi gi .t; x/

min .Yi /kEx.t/k2 C 2kxT .t/ET kkXi kkgi .t; x/k
.min .Yi / 2kXi ki / kEx.t/k2 i $i .t/; t 2 .tk1 ; tk ; (19)
min .Yi /2kXi ki min .Yi /

where i D max .Xi /
> 0 if i < 2kXi k
. Then,
C
$i .t/ $i .tk1 /ei .ttk1 / ; t 2 .tk1 ; tk : (20)
At t D tkC , k 2 N, suppose .tkC / D j, it follows from (1) and (18) that
$j .tkC / D xT .tk /.I C Bk /T ET Xj E.I C Bk /x.tk / ˛k $i .tk /: (21)
Using (20) and (21) leads to
$i .t/ ˛1 ˛2 : : : ˛i1 $1 .t0C /e1 .t1 t0 / e2 .t2 t1 / : : : ei .ttk / ; t t0 :
Let D minfi I i 2 S g. Then, we have
$i .t/ ˛1 ˛2 : : : ˛i1 $1 .t0C /e .tt0 / : (22)
To use assumption (ii), let ˛ D maxf˛k I k D 1; 2; : : : ; i 1g. Then, (22) becomes
$i .t/ ˛ i1 $1 .t0C /e .tt0 / D $1 .t0C /e.i1/ ln ˛ .tt0 / :
ln ˛
Applying (ii) with N0 D ln ˛ , where is an arbitrary constant, and Ta D ,
. > / leads to
.tt /
$i .t/ $1 .t0C /e 0
;
which implies that

.tt //=2
kEx.t/k KkEx.t0C /ke. 0
; t t0 ; (23)
p
where K D max .X1 /=min .Xi /. Thus, the trivial solution of the system (1) is
E-exponentially stable.
As done in Theorem 4, we can show that
.tt //=2 .tt //=2
kx1 k KkQi kkEx.t0C /ke. 0
and kx2 k i KkQ2i kkEx.t0C /ke. 0
:
That is, the trivial solution of (1) is exponentially stable.
4 Numerical Example
In this section, we present some examples to illustrate the effectiveness of the

theoretical results.
Example 1 Consider the singular impulsive system in (2) where

2 3 2 3 2 3
1000 0 100 0
60 0 1 07 6 1 0 0 07 6 0 7
ED6 7 6 7 6
40 0 0 05 ; A D 41 0 0 15 ; g.t; x/ D 4
7;
5
0
0000 0 111 x1 .t/ sin.t/
2 3
0:01 0 0 0
6 0 0:01 0 0 7
Bk D 6
4 0
7
0 0:01 0 5
0 0 0 0:01
and the admissible initial value is x.0/ D Œ2 1 0 0T :

Since deg.det.sE A// D rank.E/ D 2, the system ispimpulse free. Also, the
system is stable because the eigenvalues of .E; A/ are 1˙i
2
3
. Given
2 3
2 0 00
60 2 0 07
Y D6
4 0
7;
0 1 05
0 0 0 1
the equation ET XA C AT XE D ET YE is satisfied where

2 3
2 1 2 2
61 3 1 17
XD6
42 1
7:
3 05
2 1 0 8
We also get D 0:1112, D 0:000001 and ˛k D 84:5806, and from (5) tk tk1
39:91. The simulation result is shown in Fig. 1, where tk tk1 D 40:
Example
2 Consider the impulsive switched
singular system given by (1) where
x1 .t/ 40
xD , .t/ 2 S D f1; 2g, E D with rank.E/ D 1, Bk D 0:1I and
x2 .t/ 20

2 1 1 1
T
A1 D ; g1 .t; x/ D 15 tanh.x1 .t// 15 tanh.x2 .t// ;
1 2

4 1 T
A2 D ; g2 .t; x/ D 12 tanh.x1 .t// 1
2
tanh.x2 .t// :
1 4
x1(t)
0
–2
0 20 40 60 80 100 120
t
2
x2(t)
–2
0 20 40 60 80 100 120
t
2
x3(t)
–2
0 20 40 60 80 100 120
t
2
x4(t)
–2
0 20 40 60 80 100 120
t
Fig. 1 Exponential stability result of Example 1
From the Jordan canonical form of . E Ai /1 E for 2 C, we find that

1 1
41
1
1 0 0
Q1 D 5 10
2 4 ; P1 D 4 ; Q2 D 99 ; P2 D 2 ;
5 5 5
12 1 1
12 6 1
3
43
3
10 10 0
such that Q1 EP1 D Q2 EP2 D ;
Q1 A 1 P 1 D
Q2 A 2 P 2 D ;
00 0 1
5
6 0
: From the decomposition form it is clear that the systems are impulse free.
0 1
Moreover,the eigenvalues
of .E; A1 / and .E; A2 / are negative.
5
0
X1 D 3 5 > 0 satisfies AT1 X1 E C ET X1 A1 D ET Y1 E for any Y1 D I > 0.
0 3
Using X1 and Y1 ,we can
2
find that 1 < 0:3 where 1 satisfies kg1 .t; x/k 1 kExk.
0
Similarly, X2 D 3 1 > 0 satisfying AT2 X2 E C ET X2 A2 D ET Y2 E for any Y2 D
0 3
I > 0. Thus, by X2 and Y2 , 2 < 0:75 is found where 2 satisfies kg2 .t; x/k
2 kExk. We also get 1 D 0:62 and 2 D 1:5020. Thus, D 0:62. The inequality
.I CBk /T ET Xj E.I CBk / ˛k ET Xi E holds where ˛k D 1:21: If we choose D 0 and
D 0:02, then the chatter bound N0 D 0, and the average dwell time Ta D 0:3177.
The simulation result is shown in Fig. 2, where tk tk1 D 1.
6
5
4
3
x1(t)
1
0
–1
0 5 10 15 20 25
t
6
5
4
3
x2(t)
1
0
–1
0 5 10 15 20 25
t
Fig. 2 Exponential stability result of Example 2
5 Conclusion
A nonlinear impulsive singular system has been studied. In Sect. 3.1, the exponential
stability has been established for the impulsive system when we have used the
Lyapunov method. It has been noticed that if the continuous system is stable
and the time between every successive impulses is sufficiently large, then the
impulsive system is also exponentially stable. In Sect. 3.2, we have addressed the
impulsive switched system and developed new sufficient condition to guarantee the
exponential stability using multiple Lyapunov function method and average dwell
time switching signal to organize the jumps among the subsystems. It has been
noticed that when all subsystems are stable, if the average dwell time is sufficiently
large, then the entire system is also exponentially stable.
References
1. Bainov, D.D.: Systems with Impulse Effect: Stability, Theory, and Applications. Ellis Hor-
wood, Chichester/Halsted Press, New York, Toronto (1989)
2. Bainov, D.D., Simeonov, P.S.: Impulsive Differential Equations: Periodic Solutions and
Applications. Longman Scientific & Technical, Harlow/Wiley, New York/Burnt Mill, Harlow
(1993)
3. Brenan, K.E., Campbell, S.L., Petzold, L.R.: Numerical Solution of Initial-Value Problems in
Differential-Algebraic Equations. North-Holland, New York (1989)
4. Campbell, S.L.: Singular Systems of Differential Equations. Pitman Advanced Pub. Program,
San Francisco (1980)
5. Campbell, S.L.: Singular Systems of Differential Equations II. Pitman, San Francisco (1982)
6. Dai, L.: Singular Control Systems. Springer-Verlag, Berlin (1989)
7. Duan, G.: Analysis and Design of Descriptor Linear Systems. Springer, New York (2010).
8. Feng, G., Cao, J.: Stability analysis of impulsive switched singular systems. IET Control
Theory Appl. 9(6), 863–870 (2015)
9. Hill, D.J., Mareels, I.M.Y.: Stability theory for differential/algebraic systems with application
to power systems. IEEE Trans. Circuits Syst. 37(11), 1416–1423 (1990)
10. Khalil, H.K.: Nonlinear Systems. Prentice Hall, Upper Saddle River (2002)
11. Lakshmikantham, V.: Theory of Impulsive Differential Equations. World Scientific, Singapore
(1989)
12. Lewis, F.: A survey of linear singular systems. Circuits Syst. Signal Process 5(1), 3–36 (1986)
13. Li, Z.: Switched and Impulsive Systems: Analysis, Design, and Applications. Springer,
Berlin/New York (2005)
14. Sun, J.T., Zhang, Y.P.: Stability analysis of impulsive control systems. IEE Proc. Control
Theory Appl. 150(4), 331–334 (2003)
15. Wang, C.J.: Controllability and observability of linear time varying singular systems. IEEE
Trans. Autom. Control 44(10), 1901–1905 (1999)
16. Yang, T.: Impulsive control. IEEE Trans. Autom. Control 44(5), 1081–1083 (1999)
17. Yao, J., Guan, Z.H., Chen, G., Ho, D.W.C.: Stability, robust stabilization and control of
singular-impulsive systems via switching control. Syst. Control Lett. 55(11), 879–886 (2006)
Input-to-State Stability and H1 Performance
for Stochastic Control Systems with Piecewise
Constant Arguments
Mohamad S. Alwan and Xinzhi Liu
Abstract This paper addresses stochastic control system of differential equations

with piecewise constant arguments (SEPCA). The piecewise constant arguments
are of delay type. The system is viewed as a hybrid (or particularly switched)
system. This approach motivates the applicability of the classical theory of ordinary
differential equations, but not of functional differential equations, and the design
of a switching law. The main theme of this work is to establish the problems of
input-to-state stabilization (ISS), and H1 performance for a class of an uncertain
control SEPCA. To analyze these result, a common Lyapunov function together
with the techniques of differential inequalities and Razumikhin condition is used.
A numerical example with simulations is presented to clarify the validity of the
proposed theoretical approaches.
1 Introduction
A nonlinear deterministic differential equations with piecewise constant arguments

(EPCA) may have the form
xP .t/ D f .t; x.t/; x..t///; (1)
where the argument is a piecewise constant function defined on intervals with a

certain length, and it may be defined by .t/ D Œt; Œt n; t nŒt, or Œt C 1, for all
t and a positive integer n, where Œ is the greatest-integer function [8, 22].
EPCA appeared in some sequential-continuous disease models [6], control
systems [7], population growth models [1, 4, 10], chaotic synchronization–including
delayed secure communication [3], and in approximating the solutions of some
delay differential equations [9, 11, 22]. The dynamics of these equations contain
both continuous and discrete components, which enable this type of systems of
equations to be adequately embedded under the hybrid system umbrella [4]. EPCA
M.S. Alwan () • X. Liu

Department of Applied Mathematics, University of Waterloo, Waterloo, ON, N2L 3G1, Canada

368 M.S. Alwan and X. Liu
can also be a special type of functional differential equations, where the state history
is given at certain individual points, rather than intervals.
The theory of EPCA is well discussed in [8, 22]. Further properties and
applications of these equations were considered in many other works; readers may
refer to [2, 5, 7, 14, 15, 21, 23].
The notion of ISS deals with the response of asymptotically stable systems to
bounded, small input or disturbance regardless of the magnitude of system initial
state. During the last two decays, ISS has become a central foundation of modern
nonlinear feedback and design. It has been now playing a key role in systems with
recursive design and co-prime factorizations, and a connection between the input-
output (or external) stability and state (or internal) stability. For more properties,
applications, and implications of the ISS, readers may consult [13] and [16–20].
In this paper, stochastic EPCA (or SEPCA) is studied. As stated earlier, the
system will be viewed as a switched system. The main objective is to establish
the problem of input-to-state stabilization and H1 performance. To analyze these
results, the Lyapunov function method and Razumikhin condition are used. We also
present exponential stability as a special case of the aforementioned results. Finally,
to enhance the findings of this paper, some numerical examples and simulations
are presented. We believe that these results will have a great contribution to many
research fields in sciences and engineering, particularly in the fields of modern
nonlinear control design and recursive system design and in dynamical systems
subject to external disturbances.
The paper is organized as follows: In Sect. 2, the system is described, and
some notations and definitions that will be used in the sequel are given. The main
contributions of the paper are presented in Sect. 3, and a conclusion is given in
Sect. 4.
2 Problem Formulation and Background
Denote by RC the set of all non-negative qP real numbers, R the n-dimensional

n
n 2
iD1 xi for every x 2 R , R
n nm
Euclidean space and its norm kxk D the set
of all n m real matrices, and ˝ a sample space. Let C .Œa; b; D/ be the space of
continuous functions mapping Œa; b, with a < b for any a; b 2 RC , into D, for
some open set D Rn , and C 1;2 .RC Rn I RC / be the space of all functions V.t; x/
defined on RC Rn such that they are continuously differentiable once in t and
twice in x. For instance, if V.t; x/ 2 C 1;2 .RC Rn I RC /, then we have
@V
@V @V
@2 V
Vt D ; Vx D ; ; ; Vxx D :
@t @x1 @xn @xi @xj nn
Let .˝; F ; fF t gt0 ; P/ be a complete probability space with filtration fF t gt0

satisfying the usual condition (i.e., it is right continuous and F 0 contains all P-
ISS and H1 for Stochastic Control EPCA 369
null set of F ). For all t 2 RC and ! 2 ˝, let W.t; !/ 2 Rm be an m-dimensional

Wiener (or Brownian motion ) process. We also assume that, for a given W and
filtration F t for all t 2 Œa; b RC , the process W is F t -measurable and, for
all s t, the random variable W.t; !/ W.s; !/ is independent of the -algebra
F s . Denote by L ad .˝; Lp Œa; b/ the class of all random processes f which are F t -
adapted and almost all their sample paths are integrable in the Riemann sense. For
simplicity of notation, we will drop ˝ and ! when writing random processes.
A nonlinear system with SEPCA may have the following form

dx.t/ D f t; x.t/; x..t// dt C g t; x.t/; x..t// dW.t/; x.t0 / D x0 ; (2)
where x 2 Rn is the system state, and for all t t0 , the function .t/ is a piecewise
constant function taking values in the set % D fk g1 kD0 . So that one may redefine this
function as follows: for a non-negative integer k, W Œtk ; tkC1 / ! %. This piecewise
constant function represents the switching signal of the system which has the roles
of switching between the values of system state argument x. Accordingly, one may
define system (2) as follows: for all t 2 Œtk ; tkC1 /,

dx.t/ D f t; x.t/; x.k / dt C g t; x.t/; x.k / dW.t/; x.t0 / D x0 ; (3)
or, equivalently,
Z Z
t t
x.t/ D x0 C f s; x.s/; x.k / ds C g s; x.s/; x.k / dW.s/; (4)
t0 t0
where the first integral is a Riemann integral almost surely (a.s.), and the second one
is an Itô integral satisfying
hZ t i
E g s; x.s/; x.k / dW.s/ D 0; and
t0
Z t 2 Z t 2

E g s; x.s/; x.k / dW.s/ D Eg s; x.s/; x.k / ds:
t0 t0
The following definitions and assumption will be needed throughout this paper.
Definition 1 For any ˛; ˇ 2 R, an Rn -valued stochastic process x W .˛; ˇ/ ! R is
said to be a solution of (2) (or (3)) if the following hold:
1. x.t/ is continuous and F t -adapted for all t 2 .˛; ˇ/;
2. f .t; x.t/; x.k // 2 L ad .˝; L1 .˛; ˇ// and g.t; x.t/; x.k // 2 L ad .˝; L2 .˛; ˇ//;
3. the stochastic integral equation (4) holds (a.s.).
Itô Lemma. For any t0 2 RC and t t0 , let x.t/ be an Rn -dimensional Itô

stochastic process, i.e., it is F t -adapted and satisfying,
Z t Z t
x.t/ D x.t0 / C f .s; x.s//ds C .s; x.s//dW.s/; (a.s.):
t0 t0
Suppose that V 2 C 1;2 .RC Rn I R/. Then, for any t t0 , V.t; x.t// is an Itô
stochastic process satisfying
Z t Z t
V.t; x.t// D V.t0 ; x.t0 // C L V.s; x.s//f .s; x.s//ds C Vx .s; x.s// .s; x.s//dW.s/;
t0 t0
(a.s.), where
1
L V.t; x.t// D Vt .t; x.t// C Vx .t; x.t//f .t; x.t// C trŒ T .t; x.t//Vxx .t; x.t//.t; x.t//
2
is the infinitesimal operator acting on the process V.t; x.t// with Vt .t; x.t//,
Vx .t; x.t//, and Vxx .t; x.t// being the partial differentials of the process V.t; x.t//
with respect to t, x, and twice with respect to x, respectively.
Having defined SEPCA, we consider the following uncertain control SEPCA

dx.t/ D .A C A/x.t/ C Bu.t/ C Gw.t/ C f .x / dt C g.x / dW.t/; (5a)
z.t/ D Cx.t/ C Fu.t/; (5b)
x.t0 / D x0 ; (5c)
where x 2 Rn is the system state, u 2 Rp is the control input of the form Kx,
where K 2 Rpn is a control gain matrix, R 1w 2 R 2is a disturbance ninput, whichnm
q
is
2
assumed to be in L2 Œ0; 1/ (i.e., kwk2 D 0 kw.t/k dt < 1), f 2 R and g 2 R
represent lumped uncertainties, z 2 Rr is the controlled measured output, A, B,
G, C, and F are constant matrices that describe the nominal system, and A./ is
real-valued matrix, which is piecewise continuous function representing parameter
uncertainties. We also assume that f .0/ D 0 2 Rn and g.0/ D 0 2 Rnm to ensure
that the system admits a trivial solution. A symmetric matrix P is said to be positive
definite if the scalar xT Px > 0 for all nonzero x 2 Rn and xT Px D 0 for x D 0.
Definition 2 A function a 2 C .RC I RC / is said to belong to class K if a.0/ D 0
and it is strictly increasing [12]; it said to belong to class K 1 if a 2 K and it is
convex; it said to belong to class K 2 if a 2 K and it is concave.
Definition 3 For any t0 2 RC , t t0 , and x0 2 Rn , let x.t/ D x.t; t0 ; x0 / be a
solution of uncertain SEPCA (5). Then, the system is said to be robustly globally
input-to-state stable (ISS) in the m.s. if there exist functions ˇ 2 KL and 2 K
such that, for any input disturbance w, the solution satisfies

EŒkx.t/k2 ˇ EŒkx0 k2 ; t t0 C .kw.t/k/; 8 t t0 ;
for any solution x.t/ D x.t; t0 ; x0 / of (5), and any x0 2 Rn with EŒkx0 k2 < 1.
Particularly, if ˇ.s; t/ D set , for some positive , then the system is said to be
input-to-state exponentially stable in the m.s.
Definition 4 Given a constant N > 0, uncertain SEPCA (5) is said to be input-to-
state stabilizable with an H1 -norm bound N if there exists a state feedback law
u.t/ D Kx.t/;
where K D 12 "BT P for some a constant " > 0 and positive-definite matrix P, such
that, for any admissible parameter uncertainty A./, the corresponding closed-
loop system is uniformly asymptotically (or exponentially) ISS in the (m.s.) and
the controlled output z satisfies
hZ 1 i
kzk2E WD E kz.t/k2 dt N 2 kwk22 C m0 ;
0
for some positive constant m0 .

Assumption A For all t 2 RC , the admissible parameter uncertainty of the system
is defined by A D DU .t/H, where D and H are known real constant matrices
with appropriate dimensions that give the structure of the uncertainty, and U is an
unknown real time-varying matrix containing the uncertain parameters in the linear
part and satisfying kU .t/k 1.
3 Main Results
Theorem 1 For all t 2 Œtk ; tkC1 /, let the controller gain K and disturbance level
N > 0 be given. Assume that Assumption A holds, and there exist positive constants
"1 ; "2 , and a positive-definite matrix P such that the following inequality

T

1
A C BK P C P A C BK C P "1 DDT C I C N 2 GGT P C qN P
"2
1 T
C H H C "2 qN kUk2 I C CcT Cc C ˛P < 0 (6)
"1
holds, where qN > 1, ˛ > 0, > 0 is such that
trŒgT .y/Pg.y/ 2 qN xT Px; (a.s.); (7)

U is an n n matrix such that k f .y/k2 qN kUk2 kxk2 , and Cc D C C FK. Suppose

further that for any k 2 N, the following dwell-time-type condition
1 adk
tkC1 tk ln (8)
˛N bdkC1
holds, where ˛N > 0 and dk is a constant satisfying dkC1 < dk < 1 and limk!1 dk D
0. Then, system (5) is m.s. robustly globally exponentially input-to-state stabilized
by the feedback control u D Kx with an H1 -norm bound N > 0.
Proof For all t 2 Œtk ; tkC1 /, let x.t/ D x.t; t0 ; x0 / be the solution of (5) and define
V.x/ D xT Px as Lyapunov function candidate. Then,
L V.x/ xT ŒAT P C PA C 2K T BT Px C 2xT P. A/x C 2xT Pf .x /

C qN xT Px C 2xT PGw; (a.s.);
where qN is provided by using the Razumikhin technique.

1
Claim There exits 2 > 0 such that 2xT Pf .x / xT "2 qN kUk2 I C P2 x.
"2
Proof of the Claim Since

p 1
p 1 T
0 "2 f T .x / p xT P "2 f T .x / p xT P
"2 "2
1 T 2
D "2 f T .x /f .x / C x P x 2xT Pf .x /
"2
1 T 2
"2 kUk2 kx k2 C x P x 2xT Pf .x /
"2
which implies the desired result. Where qN D q=min .P/ > 1, with q > 1.

By the facts that 2xT P. A/x xT "1 PDDT P C "11 H T H x, we get

1
L V.x/ xT .A C BK/T P C P.A C BK/ x C xT "1 PDDT P C H T H x
"1

1 1
C xT "2 qN kUk2 I C P2 x C qN xT Px C "3 xT PGGT Px C wT w; (a.s.):
"2 "3
Making use of inequality (6) with "3 D N 2 , we get, for all t 2 Œtk ; tkC1 /,
1 T
L V.x/ .˛ /xT Px xT Px C w w
"3
D ˛V.x/;
N ˛N D ˛ ; (a.s.);
q
provided that V.x/ 1"3 kwk2 or kxk a1"3 kwk, for some positive constant <
˛. Applying the Itô formula to process V.x/ and taking the mathematical expectation
give DC m.t/ ˛m.t/,
N which implies that
s
˛.tt
N k/
1
m.t/ m.tk /e ; whenever kxk kwk;
a"3
where m.t/ D EŒV.x.t//; 8t 2 Œtk ; tkC1 /. By the dwell-time condition (8), we get
EŒkx.tkC1 /k2 dkC1 EŒkx.tk /k2 ;
where we have used the fact bkxk2 xT Px akxk2 . Since limk!1 q dk D 0, the
limit of x.tkC1 / will eventually converge to the limit set of radius a1"3 kwk in the
m.s.; that is the solution x is robustly globally input-to-state stabilized in the m.s.
To prove the upper bound on the output magnitude kzk, we introduce the
performance function
Z 1
JDE .zT z N 2 wT w/dt:
t0
It follows that
Z Z
1 1
JDE zT z N 2 wT w dt C E dV.x/dt EŒV.x.1// C EŒV.x0 /
t0 t0
Z h
Z
1 T 1 1
E zT z N 2 wT w dt C E
xT A C BK P C P A C BK C P "1 DDT C I P
t0 t0 "2
1 i
C qN "2 kUk2 I C H T H C qN P N 2 PGGT P C N 2 PGGT P x C 2xT PGw dt C EŒV.x0 /
"1
Z 1
T 1
DE xT A C BK P C P A C BK C P "1 DDT C I C N 2 GGT P
t0 "2
1
C qN "2 kUk2 I C H T H C qN P C CcT Cc x dt
"1
Z 1
T

E w N 2 GT Px N 2 w N 2 GT Px dt C EŒV.x0 /:
t0
By the theorem assumption in (6) and strict negativeness of the second term, we get
kzk2E N 2 kwk22 C m0 ; where m0 D EŒV.x0 / t

u
Remarks
1. The conditions on f and g mean that the system perturbations are bounded by
linear growth bounds.
2. The dwell-time-type condition in (8) is made to generate a sequence of solution

trajectories evaluated at the switching times that is convergent in the m.s. to a
limit set with a radius depending on the system disturbance.
3. If the controlled measured output is available at the time instances , that is to
say z D Cc x , where Cc is as defined before, then assumption (6) becomes

T

1
A C BK P C P A C BK C P "1 DDT C I C N 2 GGT P C qN P
"2
1 T
C H H C "2 qN kUk2 I C qN CcT Cc C ˛P < 0;
"1
Z 1
where in this case J D E .zT z N 2 wT w/dt.
t0
Corollary In Theorem 1, if w.t/ 0 for all t t0 , and the inequality

T

1
A C BK P C P A C BK C P "1 DDT C I P
"2
1 T
C qN P C H H C "2 qN kUk2 I C CcT Cc C ˛P < 0
"1
holds, then x 0 is robustly globally exponentially stable in the m.s.

Example Consider uncertain control SEPCA (5), where
" # " # " # " #
7 0 0:1 0:4 01 1
AD ; BD ; GD ; w.t/ D sin.t/ ; C D I22 ;
0 0:1 0 9 10 1
" # " # " #

0:1 0:1 1 h i 1 x1
FD ; DD ; HD 01 ; U .t/ D sin.t/; f .x / D ;
2 0 0 2 x2
" #
x1 0
g.x / D ; "1 D 1; "2 D 0:1; N D 0:1; ˛ D 2; qN D 2; and D 0:6386:
0 x2
By the theorem assumptions, we have, after some algebraic manipulations, U D

0:5I22 , and from the algebraic Riccati-like inequality (6), we have
" #
0:0593 0:0125
PD ; with b D min .P/ D 0:0579 and a D max .P/ D 0:1762;
0:0125 0:1748
2.5
2
E||x||2 & β(||w||)
1.5
0.5
0
0 2 4 6 8 10 12 14 16
t
Fig. 1 Input-to-state stability in the m.s.
2.5
1.5
E[||(x,y)T||2]
0.5
0
0 1 2 3 4 5 6 7
t
Fig. 2 Exponential stability in the m.s. (w.t/ 0)

0:0009 0:0002
and the control gain matrix is K D . The simulation results of
0:0204 0:2308
EŒkx.t/k2 (above) and ˇ.kw.t/k/ D a1"3 sin2 .t/ (below) is shown in Fig. 1, where
D 1:3, "3 D N 2 D 1=0:01, and, by the dwell-time condition in (8), we have
chosen tkC1 tk D 1, where dk D 1=2k , and k D tk , for all k D 0; 1; . If
w.t/ 0, x.t/ 0 is exponentially stable in the m.s., as shown in Fig. 2.
4 Conclusions
In this paper, we have considered a control system with uncertain SEPCA, which
has been viewed as a switched system. The focus was on establishing input-to-state
stabilization and H1 performance, where the method of Lyapunov function and
Razumikhin techniques have been used to write some sufficient conditions.
Acknowledgements This research was financially supported by Natural Sciences and Engineer-
ing Research Council of Canada (NSERC).
References
1. Akhmet, M.U., Aruğtaslan, D., Liu, X.Z.: Permanence of nonautonomous ratio-dependent

predator-prey systems with piecewise constant argument of generalized type. DCDIS-A: Math.
Anal. 15, 37–51 (2008)
2. Akhmet, M.U., Aruğaslan, D.: Lyapunov-Razumikhin method for differential equations with
piecewise constant argument. DCDS 25(2), 457–466 (2009)
3. Alwan, M.S., Liu, X.Z., Akhmet, M.U.: On chaotic synchronization via impulsive control and
piecewise constant arguments. DCDIS-Ser. B: Appl. Algorithms 22, 53–67 (2015)
4. Alwan, M.S., Liu, X.Z., Xie, W.-C.: Comparison principle and stability results for nonlinear
differential equations with piecewise constant arguments. JFI 350, 211–230 (2013)
5. Bao, G., Wen, S., Zeng, Z.: Robust stability analysis of interval fuzzy Cohen-Grossberg neural
networks with piecewise constant argument of generalized type. Neural Netw. 33, 32–41
(2012)
6. Busenberg, S., Cooke, K.L.: Models of vertically transmitted with sequential-continuous
dynamics. In: Lakshmikantham, V. (ed.) Nonlinear Phenomena: In Mathematical Sciences.
Academic, New York, pp. 179–187 (1982)
7. Chen, L., Leng, Y., Guo, H., Shi, P., Zhang, L.: H1 control of a class of discrete-time Markov
jump linear systems with piecewise-constant TPs subject to average dwell time switching. J.
Frankl. Inst. 349, 1989–2003 (2012)
9. Cooke, K.L., Györi, I.: Numerical approximation of the solutions of delay differential
equations on an infinite interval using piecewise constant arguments. Comput. Math. Appl.
28, 81–92 (1994)
8. Cooke, K.L., Wiener, J.: A survey of differential equations with piecewise constants arguments.
In: Busenberg, S., Martelli, M. (eds.) Delay Differential Equations and Dynamical Systems.
Lecture Notes in Mathematics, vol. 1475, pp. 1–15. Springer-Verlag, Heidelberg/New York
(1991)
10. Gopalsamy, K.: Stability and Oscillations in Delay Differential Equations of Population
Dynamics. Kluwer Academic, Dordrecht (1992)
11. Györi, I., Hartung, F.: On numerical approximation using differential equations with piecewise-
constant arguments. Period. Math. Hung. 56(1), 55–69 (2008)
12. Khalil, H.K.: Nonlinear Systems. Prentice Hall, Upper Saddle River (2002)
13. Kokotovic, P., Arcak, M.: Constructive nonlinear control: Progress in the 900 s. In: Proceedings
of the 14th IFAC World Congress, the Plenary and Index Volume, Beijing, pp. 49–77 (1999)
14. Nieto, J.J., Rodríguez-López, R.: Second-order linear differential equations with piecewise
constant arguments subject to nonlocal boundary conditions. AMC 218, 9647–9656 (2012)
15. Nieto, J.J., Rodríguez-López, R.: Study of solutions to some functional differential equations
with piecewise constant arguments. Abstr. Appl. Anal. art. no. 851691 (2012)
21. Ozturk, I., Bozkurt, F.: Stability analysis of a population model with piecewise constant
arguments. Nonlinear Anal.: Real World Appl. 12, 1532–1545 (2011)
19. Sontag, E.D.: Comments on integral variants of ISS. Syst. Control Lett. 34, 93–100 (1998)
16. Sontag, E.D.: Smooth stabilization implies coprime factorization. IEEE Trans. Autom. Control
34(4), 435–443 (1989)
18. Sontag, E.D., Wang, Y.: New characterization of input-to-state stability property. IEEE Trans.
Autom. Control 34(41) 1283–1294 (1996)
17. Sontag, E.D., Wang, Y.: On characterization of input-to-state stability property. Syst. Control
Lett. 24, 351–359 (1995)
20. Teel, A.R., Moreau, L., Nešic, D.: A unified framework for input-to-state stability in systems
with two time scales. IEEE Trans. Autom. Control 48(9), 1526–1544 (2003)
22. Wiener, J.: Generalized Solutions of Functional Differential Equations. World Scientific,
Singapore (1993)
23. Wu, Z.-G., Park, J.H., Su, H., Chu, J.: Stochastic stability analysis of piecewise homogeneous
Markovian jump neural networks with mixed time-delays. JFI 349, 2136–2150 (2012)
Switched Singularly Perturbed Systems
with Reliable Controllers
Mohamad S. Alwan, Xinzhi Liu, and Taghreed G. Sugati
Abstract This paper addresses the problem of exponential stability for a class
of switched control singularly perturbed systems (SCSPS) not only when all the
control actuators are operational, but also when some of them experience failures.
Multiple Lyapunov functions and average dwell-time switching signal approach
are used to establish the stability criteria for the proposed systems. In this paper,
we assume that a full access to all the system modes is available, though the
mode-dependent, slow-state feedback controllers experience faulty actuators of an
outage type. In the stability analysis, the system under study is viewed as an
interconnected system that has been decomposed into isolated, lower order, slow and
fast subsystems, and the interconnection between them. It has been observed that if
the degree of stability of each isolated mode is greater than the interconnection
between them, the interconnected mode is exponentially stable, and, then, the full
order SCSPS is also exponentially stable for all admissible switching signals with
average dwell-time. A numerical example with simulations is introduced to illustrate
the validity of the proposed theoretical results.
1 Introduction
A switched system is a special class of hybrid systems that consists of a family

of continuous- or discrete-time dynamical subsystems, and a switching rule that
controls the switchings amongst them. In reality, many applications in different
fields such as aircraft, automotive industry, robotics, control systems, biological,
epidemic disease models, etc., are operating based on switchings between multiple
dynamical modes. The stability of such systems has received great attention in the
last few decades. For more readings, one may refer to [3, 9, 10, 14, 17] and the
references therein.
It is well known that even if the individual modes are stable, the switched system
may not be stable [9]. However, it has been shown that the stability of the switched
system can be achieved if the dwell-time, the time between any two consecutive
M.S. Alwan () • X. Liu • T.G. Sugati


switchings, is sufficiently large [12]. Practically, in some situations, such as aging

systems or systems with finite escape time, the dwell-time condition may not hold,
and yet one can get the same stability result if the average dwell-time is satisfied [5].
The reliable control is the controller that tolerates actuator and/or sensor failures.
The failure of control components is frequently encountered in practice, yet the
immediate repair may be impossible in some critical cases. Therefore, designing
a reliable controller to guarantee stability is necessary. This problem has drawn
researcher attention for many years; see for instance [4, 13, 15, 16].
Systems involving multiple time-scale dynamics (known as singularly perturbed
systems) arise in a large class of applications in science and engineering such
as celestial mechanics, many-particle dynamics, and climate systems, etc. They
are characterized by small parameters multiplied by the highest derivatives. The
stability problem of these systems has attracted many researchers; see [1, 2, 6–8, 11]
and some references therein.
In this paper, we address SCSPS where the controllers are subject to faulty
actuators. The continuous states are viewed as an interconnected system with two-
time scale (slow and fast) subsystems. Moreover, due to dominant behaviour of
the reduced systems, the stabilization of the full order systems is achieved through
the controller of the slow reduced order subsystem. This in turn results in lessening
some unnecessary sufficient conditions imposed on the fast subsystem. The stability
analysis is obtained by multiple Lyapunov functions method after decomposing the
system into isolated, lower order, slow and fast subsystems, and the interconnection
between them.
It has been observed that if the degree of stability of each isolated mode is greater
than the interconnection between them, the underlying interconnected mode of the
switched system is exponentially stable. Moreover, if switching among the system
modes follows the average dwell-time rule, then the SCSPS is also exponentially
stable. The relationship between the stability degrees of the isolated subsystems
and the interconnection strength is usually formulated by the so-called M-matrix.
Finally, a numerical example and simulations are provided to justify the proposed
theoretical results.
The rest of this paper is organized as follows. Section 2 involves the problem
description and some definitions. The main results and proofs are stated in Sect. 3.
A numerical example with simulations is presented in Sect. 4. The conclusion is
given in Sect. 5.
2 Problem Formulation and Preliminaries
Throughout this paper, Rn denotes the n-dimensional Euclidean space; RC refers

to the nonnegative real numbers; Rnm is the class of all n m real matrices. A
symmetric matrix P is said to be positive definite if all its eigenvalues are positive.
Moreover, If P 2 Rnn , denote by max .P/(min .P/) the maximum (minimum)
eigenvalue of P. If V.x/ D xT Px, the inequalities min .P/jjxjj2 V.x/
Switched Singularly Perturbed Systems with Reliable Controllers 381
max .P/jjxjj2 are true. If x 2 Rn , then jjxjj refers to the Euclidean vector norm
of x. Consider the following system
xP D A11.t/ x C A12.t/ z C B1.t/ u; (1a)

".t/ zP D A21.t/ x C A22.t/ z C B2.t/ u; (1b)
x.t0 / D x0 ; z.t0 / D z0 ; (1c)
where x 2 Rm ; z 2 Rn are the system slow and fast states respectively, u 2 Rl is the
control input of the form u D Kx for some control gain K 2 Rlm , W Œt0 ; 1/ !
S D f1; 2; ; Ng is a piecewise constant function known as the switching signal
(or law). For each i 2 S ; A11i 2 Rmm ; A12i 2 Rmn ; A21i 2 Rnm ; A22i 2
Rnn ; are known real constant matrices with A22i is a nonsingular Hurwitz matrix,
B1i 2 Rml ; B2i 2 Rnl ; and 0 < "i 1. Setting "i D 0 implies that z D hi .x/ D
A122i ŒA21i x C B2i u: Plug z into (1a) gives the slow reduced subsystem xPs D A0i xs C
B0i u where A0i D A11i A12i A1 1
22i A21i , and B0i D B1i A12i A22i B2i . Choose u D Kxs
such that .A0i ; B0i / is stabilizable.
For simplicity of notation, we use x instead of xs to refer to the slow reduced
system.
Definition 1 The trivial solution of system (1) is said to be globally exponentially
stable (g.e.s.) if there exist positive constants L and such that
jjx.t/jj C jjz.t/jj L.jjx.t0 /jj C jjz.t0 /jj/e.tt0 / ; t t0 2 RC ;
for all x.t/ and z.t/, the solutions of system (1), and any x0 2 Rm ; z0 2 Rn .
Definition 2 An n n matrix M D Œmij with mij 0, for all i 6D j, is said to be an
M-matrix if all its leading successive principle minors are positive, i.e.,
2 3
m11 m12 m1k
6m21 m22 m2k 7
det 6
4
7 > 0; k D 1; 2; ; n:
5
mk1 mk2 mkk
Average dwell-time Condition (ADTC) [5]. The number of switchings N.t0 ; t/ in

the interval .t0 ; t/ for a finite t satisfies N.t0 ; t/ N0 C tt
a ; where N0 is the chatter
0
bound, and a is the average dwell-time.
3 The Main Results
In this section, we present our main results.

3.1 Normal Case
For any i 2 S , the closed-loop system becomes

8
< xP D .A11i C B1i Ki /x C A12i z;
" zP D .A21i C B2i Ki /x C A22i z; (2)
: i
x.t0 / D x0 ; z.t0 / D z0 :
Theorem 1 The trivial solution of system (1) is g.e.s. if ADTC holds, and the
following assumptions hold
(i) ReŒ.A22i / < 0, and .A0i ; B0i / is stabilizable;
(ii) there exist positive constants aji ; j D 1; ; 6 such that
2xT P1i A12i hi .x/ a1i xT x; (3)

2xT P1i A12i .z hi .x// a2i xT x C a3i .z hi .x//T .z hi .x//; (4)
2.z hi .x//T P2i R 1i x a4i xT x C a5i .z hi .x//T .z hi .x//; (5)
.z hi .x//T R 2i .z hi .x// a6i .z hi .x//T .z hi .x// (6)
where hi .x/ D A1 22i .A21i C B2i Ki /x, P2i is the solution of the Lyapunov equa-
tion AT22i P2i CP2i A22i D Iin ; where Iin is an identity matrix, R 1i D A1 22i .A21i C
B2i Ki /ŒA11i C B1i Ki A12i A122i .A 21i C B K
2i i /; and R 2i D 2P A 1
2i 22i .A21i C
B2i Ki /A12i ;
(iii) there exist a positive constant "i such that AN i is an M-matrix where
" max .Ni / a3i
#
max .P1i / min .P2i /
AN i D a4i a6i .1a5i "i / ;
min .P1i / min .P2i / "i max .P2i /
where Ni D Qi C .a1i C a2i /I C M T Pi C Pi M T such that M D A12i A1 22i .A21i C
B2i Ki / and .A0i C B0i Ki /T P1i C P1i .A0i C B0i Ki / D Qi for a given Ki .
Proof Let Vi .x/ D xT P1i x and Wi ..z hi .x//.t// D .z hi .x//T P2i .z hi .x// be
Lyapunov function candidates for the slow and the fast subsystem, respectively.
Then,
VP i .x/ xT .Qi C a1i I C a2i I/x C a3i .z hi .x//T .z hi .x//

max .Ni / a3i
Vi .x/ C Wi ..z hi .x//.t//; (7)
where Ni D Qi C.a1i Ca2i /I CM T Pi CPi M T such that M D A12i A1

22i .A21i CB2i Ki /
is negative definite. We also have
1
WP i ..z hi .x//.t// D .z hi .x//T .z hi .x// 2.z hi .x//T P2i hP i .x/
"i
1
.a5i /.z hi .x//T .z hi .x// C .z hi .x//T R 2i .z hi .x//
"i
C a4i xT x
a4i h a .1 a5i "i / i
6i
Vi .x/ C Wi ..z hi .x//.t//:
min .P1i / min .P2i / "i max .P2i /
(8)
where R 2i D 2P2i A1

22i .A21i C B2i Ki /A12i . Combining (7) and (8), we get
" max .Ni / a3i

#
P i .x/
V max .P1i / min .P2i / Vi .x/
:
P i .z hi .x//.t/
W a4i a6i .1a5i "i /
Wi .z hi .x//.t/
min .P1 / i min .P2i / "i max .P2i /
Then, we have
" max .Ni / a3i
#
AN i D :
a4i a6i
min .P1i / min .P2i /
".1a 5i "i /
i max .P2i /
Then there exists i D max .AN i / > 0 such that for t 2 Œtk1 ; tk /;

Vi .x/ Vi .x.tk1 // C Wi .z hi .x//.tk1 / ei .ttk1 / ;
and

Wi .z hi .x//.t/ Vi .x.tk1 // C Wi .z hi .x//.tk1 / ei .ttk1 / ;
For any i; j 2 S ; M > 1, we have
Vj .x.t// Vi .x.t//;

Wj .z hj .x//.t/ Wi .z hi .x//.t/ :
If the system switches among its modes, one may get for all t t0 ,
Vi .x.t// 2e1 .t1 t0 / 2e2 .t2 t1 / 2ek1 .tk1 tk2 /
h i
V1 .x.t0 // C W1 .z h1 .x//.t0 / ek .ttk1 / :
Let D minfj W j D 1; 2; ; kg: Then

h i
Vi .x.t// .2/k1 V1 .x.t0 // C W1 .z h1 .x//.t0 / e.tt0 /
h i
V1 .x.t0 // C W1 .z h1 .x//.t0 / e.k1/ ln .tt0 / ;

where D 2. Applying the ADTC with N0 D ln , is an arbitrary constant,
ln
a D . /
with > leads to
h i
Vi .x.t// V1 .x.t0 // C W1 .z h1 .x//.t0 / e.tt0 /
and
h i
Wi .z hi .x//.t/ V1 .x.t0 // C W1 .z h1 .x//.t0 / e.tt0 /
This implies that there exists L > 0 such that

.tt /=2
jjx.t/jj C jjz.t/jj L.jjx.t0 /jj C jjz.t0 /jj/e 0
: t
u
3.2 Faulty Case
To analyze the reliable stabilization with respect to actuator failures, for any i 2 S ,
consider the decomposition of the control matrix Bi D Bi˙ C Bi˙N ; where ˙ the
set of actuators that are susceptible to failure, and ˙N f1; 2; : : : ; lg ˙ the
set of actuators which are robust to failures and essential to stabilize the given
system, moreover, the matrices Bi˙ ; Bi˙N are the control matrices associated with
˙; ˙N respectively, and are generated by zeroing out the columns corresponding
to ˙N and ˙, respectively. The pair .Ai ; Bi˙N / is assumed to be stabilizable. For a
fixed i, let ˙ corresponds to some of the actuators that experience failure,
and assume that the output of faulty actuators is zero. Then, the decomposition
becomes Bi D Bi CBiN ; where Bi and BiN have the same definition of Bi˙ and Bi˙N ,
respectively. Since the control input u is applied to the system through the normal
actuators, the closed-loop system becomes
xP D .A11i C B1N i KiN /x C A12i z; (9a)

"i zP D .A21i C B2N i KiN /x C A22i z; (9b)
x.t0 / D x0 ; z.t0 / D z0 : (9c)
where KiN D 12 ˇi BT0iN PiN , with B0iN D B1N i A12i A1

22i B2N i , and PiN is a positive
definite matrix such that .A0i C B0iN KiN / PiN C PiN .A0i C B0iN KiN / D I. Setting
T
"i D 0, one may get z D hiN .x/ D A1

22i .A21i C B2N i KiN /x. In the following theorem,
we assume that N i D ˙i .
N
Theorem 2 The trivial solution of system (9) is g.e.s. if ADTC and the following
assumptions hold for any i 2 S

(i) ReŒ.A22i / < 0, and AT11i P1i C P1i A11i C ˇi P1i A12i A1 T
22i B2˙N i B1˙N i

B1˙N i B1˙N P1i C ˛i I D 0;
T
i
(ii) there exist positive constants aji ; j D 1; ; 6 such that
2xT P1i A12i hi˙N .x/ a1i xT x; (10)

2x P1i A12i .z hi˙N .x// a2i x x C a3i .z hi˙N .x// .z hi˙N .x//;
T T T
(11)
2.z hi˙N .x// P2i R 1i˙N x a4i x x C a5i .z hi˙N .x// .z hi˙N .x//;
T T T
(12)
.z hi˙N .x//T R 2˙N i .z hi˙N .x// a6i .z hi˙N .x//T .z hi˙N .x//; (13)
where hi˙N .x/ D A1 22i .A21i C B2˙N i Ki˙N /x, P2i is the solution of A22i P2i C
T
1 1
P2i A22i D Iin , R 1i Ṅ D A22i .A21i C B2˙N i Ki˙N /ŒA11i C B1i Ki A12i A22i .A21i C
B2˙N i Ki˙N / where Ki˙N D 12 ˇi BT0i˙N Pi˙N , and R 2˙N i D 2P2i A1 22i .A21i C
1 1 T 1
ˇ B B .A12i A22i / P1i /A12i 2 ˇi B2˙N i B1˙N P1i ;
T
2 i 2˙N i 2˙N i
T
i
(iii) there exist a positive constant "i such that AN i˙N is an M-matrix where
" max .N #
i Ṅ / a3i
AN i˙N D a4i "i .a5i Ca6i /1 :
min .P1i / "i max .P2i /

Proof Let Vi .x/ D xT P1i x and Wi .z h˙N i .x//.t/ D .z h˙N i .x//T P2i .z h˙N i .x//
be Lyapunov function candidates. Then, we have
VP i .x/ xT .˛i C a1i C a2i /Ix C a3i .z hi˙N .x//T .z hi˙N .x//
˛i C a1i C a2i a3i
Vi .x/ C Wi .z h˙N i .x//.t/ (14)
We also have
1
WP i .z h˙N i .x//.t/ D .z hi˙N .x//T .z hi˙N .x// 2.z hi˙N .x//T P2i hP i˙N .x/
"i
1
.a5i /.z hi˙N .x//T .z hi˙N .x// C a4i xT x
"i
C .z hi˙N .x//T R 2˙N i .z hi˙N .x//
a4i h " .a C a / 1 i
i 5i 6i
Vi .x/ C Wi ..z hi˙N .x//.t//;
(15)
1 1
where R 2˙N i D 2P2i A1 1 T
22i .A21i 2 ˇi B2˙N i B1˙N i P1i C 2 ˇi B2˙N i B2˙N i .A12i A22i / P1i /A12i .
T T
Combining (14) and (15), we get the M-matrix AN i˙N with
" max .N #
i Ṅ / a3i
AN i˙N D a4i "i .a5i Ca6i /1 :
Proceeding as done in the proof of Theorem 1, we get the desired result. t

u
4 Numerical Example
Example 1 Consider system (1) with S D f1; 2g;

5 0 0:1 2 13 1 2
A111 D ; A121 D ; A211 D ; A221 D ;
0 10 0:1 0 21 3 2

3 1 1 0 23 2 1
A112 D ; A122 D ; A212 D ; A222 D ;
0 6 0:1 0:3 11 1 1

5 0:5 3 1 4 5 2 2
B11 D ; B21 D ; B12 D ; B22 D ;
0:1 0:15 1 4 0:5 1 1 3
"1 D 0:01; ˇ1 D 0:5; a11 D 0:1; a21 D 0:15; a31 D 0:02; a41 D 0:01; a51 D
70; Q1 D 4I;
"2 D 0:02; ˇ2 D 0:25; a12 D 0:3; a22 D 0:2; a32 D 0:2; a42 D 0:02; a52 D
30; and Q2 D I:

0:1025 0:0274
Case 1. When all actuators are operational, we have P11 D ;
0:0274 0:0615

0:1697 0:0616 1:5 1 0:5 0:5
P12 D ; P21 D ; P21 D ; and K1 D
0:0616 0:1322 1 1:75 0:5 1

0:0217 0:0031 0:1638 0:0869
; K2 D : Thus, the matrices A0i C B0i Ki
0:0840 0:0238 0:1449 0:0842

(i D 1; 2) are Hurwitz and a D ˛ln D 1:8330.
Case 2. When there are failures in the first actuator of B1i , and the second actuator
of B2i for both modes, i.e.,

0 0:5 30 05 20
B1˙N 1 D ; B2˙N 1 D ; B1˙N 2 D ; B2˙N 2 D ;
0 0:15 10 01 10
a4 b 3.5
3.5
3
3
2.5
2.5
2
||x||&||z||
||x||&||z||
2
1.5
1.5
1
1
0.5
0.5
0 0
0 0.5 1 1.5 2 2.5 3 0 1 2 3 4 5 6 7 8 9
t t
Fig. 1 Singularly perturbed switched system. (a) Operational actuators. (b) Faulty actuators

0:0993 0:0257 0:2828 0:1570
we have P11 D ; P12 D ; P21 and P22
0:0257 0:0606 0:1570 0:2145

0:1024 0:0278
are the same as for the normal case, and K1 D ; K2 D
0:0134 0:0055

0:1355 0:0991
: Thus, the matrices A0i C B0i Ki (i D 1; 2) are Hurwitz and
0:1964 0:1249

a D ˛ln D 4:1498.
Figure 1a,b show the simulation results of jjxjj (top) and jjzjj (bottom) for the normal
and the faulty cases respectively.
5 Conclusion
This paper has established new sufficient conditions that guaranteed the global
exponential stability of SCSPS. The output of the faulty actuators has been treated
as an outage. So that, as a future work, one may consider nonzero output which
can be viewed as an external disturbance to the system. We have shown that, using
ADTC with multiple Lyapunov functions, the full order switched system has been
exponentially stabilized by u D Ki x where in the faulty case, KiN D 12 ˇi BT0iN PiN .
A numerical example has been introduced to clarify the proposed results.
Acknowledgements This work was partially supported by NSERC Canada. The third author
acknowledges the sponsorship of King Abdulaziz University, Saudi Arabia.
References
1. Alwan, M.S., Liu, X.Z.: Stability of singularly perturbed switched systems with time delay and
impulsive effects. Nonlinear Anal. 17, 4297–4308 (2009)
2. Alwan, M.S., Liu, X.Z., Ingalls, B.: Exponential stability of singularly perturbed switched
systems with time delay. Nonlinear Anal.: Hybrid Syst. 2(3), 913–921 (2008)
3. Branicky, M.S.: Multiple Lyapunov functions and other analysis tools for switched hybrid
systems. IEEE Trans. Autom. Control 43(4), 475–482 (1998)
4. Cheng, X.M, Gui, W.H., Gan, Z.J.: Robust reliable control for a class of time-varying uncertain
impulsive systems. J. Cent. S. Univ. Technol. 12(1), 199–202 (2005)
5. Hespanha, J.P., Morse, A.S.: Stability of switched systems with average dwell-time. In:
Proceedings of the 38th IEEE Conference on Decision and Control, vol. 3, Phoenix, pp. 2655–
2660 (1999)
6. Kang, K.-I., Park, K.-S., Lim, J.-T.: Exponential stability of singularly perturbed systems with
time delay and uncertainties. Int. J. Syst. Sci. 46(1), 170–178 (2015)
7. Khalil, H.K.: Nonlinear Systems, 3rd edn. Prentice-Hall, Upper Saddle River (2002)
8. Kokotovic, P.V., Khalil, H.K., O’Reilly, J.: Singular Perturbation Methods in Control: Analysis
and Design. Academic, London (1986)
9. Liberzon, D.: Switching in Systems and Control. Birkhäuser, Boston (2003)
10. Liberzon, D., Morse, A.S.: Basic Problems is Stability and Design of Switched Systems. IEEE
Control Syst. Mag. 19(5), 59–70 (1999)
11. Liu, X., Shen, X., Zhang, Y.: Exponential stability of singularly perturbed systems with time
delay. Appl. Anal. 82(2), 117–130 (2003)
12. Morse, S.: Supervisory control of families of linear set-point controllers, part 1: exact matching.
IEEE Trans. Autom. Control 41(10), 1413–1431 (1996)
13. Seo, C.J., Kim, B.K.: Robust and reliable H1 control for linear systems with parameter
uncertainty and actuator failure. Automatica 32(3), 465–467 (1996)
14. Shorten, R., Wirth, F., Mason, O., Wulff, K., King, C.: Stability criteria for switched and hybrid
systems. SIAM Rev. 49(4), 545–592 (2007)
15. Veillette, R.J.: Reliable state feedback and reliable observers. In: Proceedings of the 31st
Conference on Decision and Control, Tucson, pp. 2898–2903 (1992)
16. Veillette, R.J., Medanic, J.V., Perkins, W.R.: Design of reliable control systems. IEEE Trans.
Autom. Control 37(3), 290–304 (1992)
17. Zhai, G., Hu, B., Yasuda, K., Michel, A.N.: Stability analysis of switched systems with stable
and unstable subsystems: an average dwell time approach. Proc. Am. Control Conf. 1(6), 200–
204 (2000)
Application of an Optimized SLW Model in CFD
Simulation of a Furnace
Masoud Darbandi, Bagher Abrar, and Gerry E. Schneider
Abstract Radiation is the most important part of heat transfer in combustion

processes in applications such as furnaces. A proper radiation model for CFD
simulations is the one, which provides the required accuracy with minimum
computational cost. In this work, we focus on radiation modeling in CFD simulation
of a laboratory scaled furnace. We use an optimized spectral line-based weighted-
sum-of-gray-gases (SLW) model, which only needs four radiation transfer equations
(RTE) solution for accurate prediction of radiative heat transfer in non-gray combus-
tion fields. This is while the classic non-optimized SLW model needs at least 10–20
RTEs solution for the same case. We apply both the optimized and non-optimized
SLW model to CFD simulation of the furnace. To evaluate the achieved results,
we compare them with available measured data. We further compare the results of
optimized SLW model with those of non-optimized SLW model. The comparisons
demonstrate that the optimized SLW model provides the same accuracy of non-
optimized SLW mode, while it requires less than 80% computational time.
1 Introduction
The fully coupled computational fluid dynamic (CFD) simulation is widely used for
design and optimization proposes in different industrial furnaces, boilers, and fire-
heaters. Radiation is the most important part of heat transfer in such combustion
devices. The process of turbulent combustion and heat transfer strongly depend on
each other and the small influence of thermal radiation may be magnified by the
non-linear processes of turbulent combustion. Moreover, radiation can affect flame
temperature and species concentration calculations. The inclusion of radiative heat
transfer reduces the size of flame region, where the maximum temperatures occur.
M. Darbandi () • B. Abrar

Department of Aerospace Engineering, Center of Excellence in Aerospace Systems,
Sharif University of Technology, Tehran, P.O. Box 11365-8639, Iran
G.E. Schneider
Department of Mechanical and Mechatronics Engineering, University of Waterloo,
Waterloo, ON, N2L 3GI, Canada

Therefore, the effects of thermal radiation should take into account even in case of
non-luminous flames [1–3]. This requires the solution of radiative transfer equation
(RTE), which determines the variation of radiation intensity in spatial and angular
space. However, accurate solution of RTE in combustion fields is a very complicated
task. It is mainly because of non-gray behavior of combustion gases. The absorption
coefficient of non-gray gases, which appears in RTE, varies very rapidly and
strongly with the wave number. Cumber and Fairweather [4] recommended that
non-gray radiation models should be used for accurate simulations of combustion
processes.
Among different non-gray radiation models, the spectral line-based weighted-
sum-of-gray-gases (SLW) model has been the subject of many researches in the last
two decades [5]. The SLW is categorized in the modern class of global models. The
global models can be considered as improved versions of the weighted-sum-of-gray-
gases (WSGG) model, which benefit form definition of absorption-line-black-body-
distribution-function (ALBDF) to obtain the radiation parameters. Goutiere, et al.
[6] evaluated different radiation models to calculate the radiative heat transfer in two
dimensional enclosures. They showed that the SLW model can accurately predict the
radiation heat transfer in their test cases. Similar investigations were also conducted
by Coelho [7] in three-dimensional enclosures. He also confirmed the accuracy of
SLW model. Johansson, et al. [8] evaluated WSGG and SLW models in several 1D
tests mimic the conditions in oxy-fired boilers. Their results showed that the SLW
model usually yields more accurate predictions than the WSGG model. Recently,
Modest [9] reviewed historical development of nongray models. He proposed to
include global models in simulations of complicated combustion systems. However,
one should consider the computational cost of global models, such as the SLW
model, before applying it in CFD simulations. The classic SLW model approximates
the non-gray gases by sum-up of 10–20 gray gases. Therefore, it requires 10–20
RTEs solution for these gray gases, which could be very time consuming in real
scale problems. It is well known that reducing the number of SLW’s gray gases and
therefore reducing the computational cost of RTEs solution is not readily possible,
because it could introduce some errors in the predicted results [5].
In this work, we conduct a CFD simulation for turbulent reacting flow in a
laboratory scaled gas fired furnace. We apply an optimized SLW model for efficient
calculation of radiation heat transfer. The optimized SLW model approximates
non-gray combustion gas mixture by sum-up of only 3 gray gases plus 1 clear
gas. However, the accuracy of the optimized SLW model is preserved through the
use of an optimization procedure. This would make the optimized SLW model
computationally more efficient than the classic non-optimized SLW model.
Application of an Optimized SLW Model in CFD Simulation of a Furnace 391
2 Governing Equations and Physical Models
2.1 Turbulent Flow Field and Energy Transport Equations
The mass and momentum governing equations for a turbulent flow can be expressed
as
r .V/ D 0 (1)
r .VV/ D pI C r N (2)
in which, is the density, V is the velocity vector, p is the pressure, and N is the
stress tensor. Here, we use the -" model along with the standard wall functions
[10], which is commonly accepted for turbulence modeling in industrial applications
such as furnaces and boilers [11, 12].
The governing equation for turbulent energy transport is given by
keff
r .VH/ D r . r H/ C SH (3)
cp
in which, H is enthalpy. The viscous heating terms are neglected here. The effective
thermal conductivity is given by keff D k C .cp t /=Prt , with Prt D 0:85. k, cp ,
t are thermal conductivity coefficient, specific heat at constant pressure, and fluid
viscosity coefficient, respectively. SH also represents source term due to radiation
heat transfer.
2.2 Combustion Model
Here, we use the mixture fraction theory for combustion modeling. The mixture
fraction is defined as f D .Yi Yi;o /=.Yi;f Yi;o /, in which Yi is the mass fraction of
species i. The subscripts f and o also refer to fuel and oxidizer at inlet streams.
In this theory, a transport equation is solved for each of mean mixture fraction
and mixture fraction variance. One advantage of using mixture fraction theory
for turbulent non-premixed combustion modeling is that, under the assumption of
chemical equilibrium, instantaneous values of species concentration, density, and
temperature can uniquely relate to mixture fraction via 'i D 'i . fi ; Hi /. However,
only averaged values of these scalars, i.e. ', f , and H are predicted in turbulent
flow modeling. The relationship between the averaged values and the instantaneous
values can be achieved through the use of presumed ˇ-shaped probability density
function (PDF)[13].
Here, we perform the chemical equilibrium calculations by means of Gibbs’ free
energy minimization. We assume there are 20 species and radicals in the equilibrium
mixture including CH4 , C2 H6 , C3 H8 , C4 H10 , CO2 , N2 , O2 , H2 O, CO, H2 , OH, O,
O3 , HOCO, HCO, CHO, H2 O2 , HONO, HO2 , and H. More details on governing

equations and combustion modeling in turbulent reacting flow can be found in
Ref. [14].
2.3 Radiation Model
Radiation effects are taken into account in the CFD simulation of combustion
processes via addition of radiation source term in the right hand side of energy
transport equation, i.e., Eq. (3). In the SLW model, this source term is given by
X
J Z
SH D j .4aj Ib Ij d˝/ (4)
jD1 4
The calculation of above source term requires the solution of j RTE s as follows:
dIj
D j .aj Ib Ij / (5)
ds
in which, I is the radiation intensity, j is the absorption coefficient, and aj is
the emissivity weighting factor. j is related to the absorption cross section as
described by j D NCj . To achieve Cj , the entire range of absorption cross
section between Cmax and Cmin is divided into J logarithmically spaced intervals.
The interval borders named as supplemental absorption cross section q are obtained
.j=J/
using CQ j D Cmin .Cmax =Cmin / , where Cj is defined as Cj D CQ j CQ jC1 . The
emissivity weighting factor aj is also calculated by employing the ALBDF definition
as follows:
Z
1
F.C ; Tg ; Tb ; Y/ D Ib .Tb ; /d (6)
Ib .Tb / fWC .Tg ;Y/Cg
Using the above definition, aj is written as aj D F.CQ jC1 / F.CQ j /. To obtain the
values of ALBDFs, here we use the tabulated values, which are presented by Pearson
and Webb [15] for H2 O, CO2 , and CO gases based on the recent HITEMP2010
spectroscopic database [16].
2.4 The Optimized SLW Model
As outlined in the previous sections, the SLW model is involved with J C 1 RTEs
solutions selections. About J D 10 20 gray gases are usually required to achieve
the required accuracy in the SLW model. However, 10–20 RTEs solutions may need
much computational efforts in real scale combustion application problems. This is
a disadvantage, which makes the use of classic SLW model prohibitive in CFD
simulations. On the other hand, choosing a value of J bellow 10 may deteriorate the
accuracy of SLW model [5]. Here, we propose the use of optimized SLW model
to reduce the number of gray gases. Instead of logarithmic discretization in the
classic non-optimized SLW model, we benefit from the optimization procedure to
obtain gray gas parameters aj and j in the optimized SLW model. The optimization
procedure minimizes the difference between the calculated values of total emissivity
based on only three gray gases parameters and the true values of total emissivity
over the path lengths Li . Mathematically, it is required to minimize the following
objective function:
X
error D Œ"optimized SLW .Li / "true .Li /2 (7)
i
where the values of total emissivity, based on the three gray gases parameters, are
calculated from
X
JD3
"optimized SLW .Li / D aj .1 ej Li / (8)
jD1
and the true values of total emissivity "true can be obtained using the detail spectro-
scopic data in the HITEMP2010 database and taking the following integration:
Z
1
"true .Li / D Ib .1 e Li /d (9)
Ib
The path lengths used in Eqs. (7), (8) and (9) should cover a range of typical lengths
consistent with the problem in hand. Here, we use a typical range from 0.01 to 10
times of the characteristic length of the problem including about 10 path lengths in
each decade. The minimization of the objective function in Eq. (7) is in fact a non-
linear curve-fit problem, in which the values of aj and j parameters are obtained for
each of the three gray gases as of the results. We use the trust region algorithm for
solving this problem. In this way, we can accurately approximate the non-gray gases
mixture by sum-up of only 3 gray and 1 clear gases. In the optimization procedure,
the true values of " are first calculated over a range of distances L typical of problem
under investigation. Then, the optimized parameters aj and j for each of 3 gray
gases are directly achieved using the least square curve fit to these " data.
3 Numerical Methods and Computational Procedure
We use the finite-volume method to treat the flow governing equations and to
derive the sets of linear algebraic equations. In this method, the solution domain is
discretized to a number of control volumes. The governing equations are integrated
over the entire faces of each control volume. We also use the finite-volume method
to derive a conservative statement for the RTE. For this purpose, each octant of the
angular space 4 is discretized into N N' non-overlapping control angles ˝l . In
this work, we use the SIMPLE algorithm to solve the flow governing equations.
Details of the employed finite volume method for the flow and RTE governing
equations and the SIMPLE algorithm are given in our previous works [17, 18].
4 Problem Description
We consider a laboratory scale furnace with available measured data in this study
[19]. The furnace is a circumferentially symmetric 300 kW BERL combustor, with
properly insulated octagonal cross section, conical hood and cylindrical exhaust
duct. It is equipped with a vertically fired burner in the bottom. The burner has 24
radial fuel injection holes and a bluff center body. Natural gas is radially introduced
through these holes. Air is introduced through an annular inlet equipped with a
proper swirler. Therefore, a non-premixed turbulent swirl stabilized flame forms in
the furnace. We model this problem as axisymmetric by appropriate adjustment with
the real 3D furnace. Figure 1 presents a schematic of the furnace and a close up of
burner’s head along with main dimensions.
The natural gas is assumed to be composed of 96.5 % CH4 , 1.7 % C2 H6 , 0.1 %
C3 H8 , 0.1 % C4 H10 , 0.3 % CO2 , and 1.3 % N2 . Fuel jet is considered to have a
mean radial velocity of v D 157:77 m/s at fuel inlet boundary. Air stream is
also considered to have a mean axial velocity of u D 31:35 m/s and a mean
swirl velocity of w D 20:97 m/s. The fuel and air inlet temperatures are 312 and
Fig. 1 Geometry of the furnace and close-up of the burner, Sayre et al. [19]
308 K, respectively. The temperature of the bottom wall, side wall, conical wall,
and exhaust duct wall are set to be constant at 1100, 1220, 1305, and 1370 K,
respectively. All walls are treated to be diffuse with an internal emissivity of 0.5.
The gauge pressure is also set to be zero at the furnace outlet.
Here we solve the problem using both the optimized and non-optimized SLW
models. In the non-optimized SLW model, j D 20 gray gases are used. In this
section, we first provide the results, which are achieved using the optimized SLW
model and compare them with the available measured data of furnace. Then we
further compare the achieved radiative quantities of the optimized SLW model with
those of the non-optimized SLW model. We obtain our results using a non-uniform
structured grid of about 10,000 control volumes and N N' D 4 4 control angles.
Figure 2 gives qualitative representations of the achieved flame and using the
optimized SLW model. Figure 2a displays the temperature contour, Fig. 2b and c
displays the H2 O and CO2 mole fraction contours, and Fig. 2d displays the CO mole
fraction contour. All the presented contours in Fig. 2 have important role in radiation
calculations.
Figure 3 presents radial temperature profiles in different axial locations of the
furnace. We compare the results of optimized SLW model with the available
measured data of Sayre, et al. [19]. It seems that our results present a thinner flame
than the measured data. This may be partially because of the axisymmetric modeling
of a three-dimensional furnace. Also the sharp gradients of temperature variations
in our results can be partially due to ignoring the finite rate of reactions in our
equilibrium chemistry calculations. Moving downstream in the furnace, these are
better agreement between our predictions and the measured data.
Figure 4 compares the results of the optimized and non-optimized SLW models
with each other. Figure 4a presents the contours of radiation source term. For a better
comparison, we limit the range of depicted source term. Comparison in this figure
reveals that the results of optimized SLW model are in complete agreement with
those of the non-optimized SLW model. Figure 4b displays the variation of incident
radiation heat flux along the furnace walls and the variation of temperature along
the furnace centerline. Again, there are complete agreement between the results
of optimized and non-optimized SLW models. These comparisons reveal that the
optimized SLW model can provide the same accuracy as it is provided by the non-
optimized SLW model.
Table 1 compares the number of equations and the required CPU times for
1 iteration of coupled CFD solution using different SLW models. The optimized
SLW model only needs 4 RTEs to be solved, while the non-optimized SLW model
needs 21 RTE solutions. This would indicate a great computational advantage for
the optimized SLW model. Here, all the computations are conducted on a laptop
equipped with an Intel CORE i5 CPU and 4 GB RAM. As is seen in Table 1,
0.6 Temperature (K): 400 600 800 1000 1200 1400 1600 1800
r (m) 0.4
0.2
0
0 0.5 1 x (m) 1.5 2 2.5 3
(a) Temperature Contour
0.8
mole fraction H2O (%): 1 3 5 7 9 11 13 15 17
0.6
r (m)
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3
x (m)
(b) H2O mole fraction Contour
0.8
0.6 mole fraction CO2 (%): 0.5 1.5 2.5 3.5 4.5 5.5 6.5 7.5 8.5
r (m)
0.4
0.2
0
0 0.5 1 1.5 2 2.5 3
x (m)
(c) CO2 mole fraction Contour
0.6
mole fraction CO (%): 0.2 0.6 1 1.4 1.8 2.2 2.6
r (m)
0.4
0.2
0
0 0.5 1 x (m) 1.5 2 2.5 3

(d) CO mole fraction Contour
Fig. 2 Contours of a: temperature and b: H2 O, c: CO2 , and d: CO mole fractions in the furnace
each solution iteration only lasts 0.8 s with the optimized SLW model. However, the
required time for each solution iteration is 4.1 s with the non-optimized SLW model.
In other words, using the optimized SLW model makes over 80 % reductions in
the required computational time in comparison with using the non-optimized SLW
model. This is where both models provide the same level of accuracy, see Fig. 4.
0.5
0.4
x = 27 mm
x = 109 mm
x = 177 mm
x = 343 mm
x = 432 mm
0.3
r (m)
0.2
0.1
0
1000
2000
1000
2000
1000
2000
1000
2000
1000
2000
T (K)
Fig. 3 Radial temperature profiles at different axial distances in furnace and comparison with the
measured data of Sayre, et al. [19]
–∇.q (kW/m2): –200 –150–100 –50 0 50 100

0.6
0.4 optimized SLW

0.2
r(m)
–0.2
non-optimized SLW
–0.4
–0.6
0 0.5 1 1.5 2 2.5

x(m)
(a) contour of radiation source term
2100 optimized SLW 300
non-optimized SLW
qr, inc (kW/m )
2
T (K)
1800
200
1500
100
0 0.5 1.5 1 2 2.5
x (m)
(b) variations of incident radiation heat flux along
the furnace wall and temperature along the furnace
centerline
Fig. 4 Comparison between the results of optimized and non-optimized SLW models, a: contours
of radiation source term, b: variations of incident radiation heat flux along the furnace wall and
temperature along the furnace centerline
Table 1 Number of equations and required CPU times for each coupled CFD solution iteration
using different SLW models
Model Number of RTEs CPU time (s)
Optimized SLW 4 0.8
Non-optimized SLW 21 4.1
6 Conclusion
In this work we introduced the optimized SLW model, which used only 3 gray gases
plus 1 clear gas to approximate the radiation in non-gray gas mixtures. Therefore, it
reduces the number of radiative transfer equatous to only 4 RTEs solution. This is
while the classic non-optimized SLW model requires about 10–20 RTEs solution to
support the same accuracy. We applied both the optimized and non-optimized SLW
models in the CFD simulation of turbulent reacting flow in a furnace. The achieved
results were in good agreements with available measured data in furnace. We further
compared the results of optimized SLW model with those of non-optimized SLW
model. The comparisons demonstrated that the accuracy of the optimized SLW
model was as good as the non-optimized SLW model. This is while the use of
optimized SLW model required 80 % less computational time. This improvement is
of great value for intensive CFD simulations. Therefore, we suggest the optimized
SLW model for radiation heat transfer calculation in CFD simulation of combustion
processes.
Acknowledgements The authors would like to thank the research deputy of Sharif University of
Technology for the financial support received during this research work.
References
1. Keramida, E.P., Liakos, H.H., Founti, M.A., Boudouvis, A.G., Markatos, N.C.: Radiative heat
transfer in natural gas-fired furnaces. IJHMT 43(10), 1801–1809 (2000)
2. Xu, X., Chen, Y., Wang, H.: Detailed numerical simulation of thermal radiation influence in
Sandia flame D. IJHMT 49(13), 2347–2355 (2006)
3. Kontogeorgos, D.A., Keramida, E.P., Founti, M.A.: Assessment of simplified thermal radiation
models for engineering calculations in natural gas-fired furnace. IJHMT 50(25), 5260–5268
(2007)
4. Cumber, P.S., Fairweather, M.: Evaluation of flame emission models combined with the
discrete transfer method for combustion system simulation. IJHMT 48(25), 5221–5239 (2005)
5. Denison, M.K., Webb, B.W.: A spectral line-based weighted-sum-of-gray-gases model for
arbitrary RTE solvers. J. Heat Transf. 115(4), 1004–1012 (1993)
6. Goutiere, V., Liu, F., Charrette, A.: An assessment of real-gas modelling in 2D enclosures.
JQSRT 64(3), 299–326 (2000)
7. Coelho, P.J.: Numerical simulation of radiative heat transfer from non-gray gases in three-
dimensional enclosures. JQSRT 74(3), 307–328 (2002)
8. Johansson, R., Andersson, K., Leckner, B., Thunman, H.: Models for gaseous radiative heat
transfer applied to oxy-fuel conditions in boilers. IJHMT 53(1), 220–230 (2010)
9. Modest, M.F.: The treatment of nongray properties in radiative heat transfer: from past to
present. J. Heat Transf. 135(6), 061801 (2013)
10. Launder, B.E., Spalding, D.B.: The numerical computation of turbulent flows. Comput.
Methods Appl. Mech. Eng. 3(2), 269–289 (1974)
11. Belosevic, S., Sijercic, M., Oka, S., Tucakovic, D.: Three-dimensional modeling of utility
boiler pulverized coal tangentially fired furnace. IJHMT 49(19), 3371–3378 (2006)
12. Crnomarkovic, N., Sijercic, M., Belosevic, S., Tucakovic, D., Zivanovic, T.: Numerical
investigation of processes in the lignite-fired furnace when simple gray gas and weighted sum
of gray gases models are used. IJHMT 56(1), 197–205 (2013)
13. Cumber, P.S., Onokpe, O.: Turbulent radiation interaction in jet flames: sensitivity to the PDF.
IJHMT 57(1), 250–264 (2013)
14. Kuo, K.K.: Principles of Combustion. Wiley, Hoboken (2005)
15. Pearson, J.T., Webb, B.W., Solovjov, V.P., Ma, J.: Updated correlation of the absorption line
blackbody distribution function for H2O based on the HITEMP2010 database. JQSRT 128:10–
17 (2013)
16. Rothman, L.S., Gordon, I.E., Barber, R.J., Dothe, H., Gamache, R.R., Goldman, A., Tennyson,
J.: HITEMP, the high-temperature molecular spectroscopic database. JQSRT 111(15), 2139–
2150 (2010)
17. Darbandi, M., Abrar, B., Schneider, G.E.: Solving combined natural convection-radiation
in participating media considering the compressibility effects. In: Proceeding of the 52nd
Aerospace Sciences Meeting, AIAA, Maryland (2014)
18. Darbandi, M., Abrar, B.: A compressible approach to solve combined natural convection-
radiation heat transfer in participating media. Numer. Heat Transf. B 66(5), 446–469 (2014)
19. Sayre, A., Lallemant, N.D.J., Weber, R.: Scaling Characteristics of Aerodynamics and Low-
NOx Properties of Industrial Natural Gas Burners, The SCALING 400 Study, Part IV: The 300
kW BERL Test Results. International Flame Research Foundation (1994)
Numerical Investigation on Periodic Simulation
of Flow Through Ducted Axial Fan
Seyedali Sabzpoushan, Masoud Darbandi, Mohsen Mohammadi,

and Gerry E. Schneider
Abstract In this paper, the flow through an axial fan is suitably simulated
considering relatively low mesh sizes and benefiting from the periodic boundary
condition. The current periodic boundary condition implementation has major
differences with the past classical ones, which were routinely used in literature.
In this regard, we first discuss the ambiguities behind the current periodic geometry
and the proposed mesh generation and possible boundary condition choices. Then,
proper remedies are proposed to resolve them. One remedy returns to the proper
choice of pitch ratio magnitude. After practicing various pitch ratio magnitudes, we
eventually arrive to an optimum one, which provides suitable numerical accuracy
despite using sufficiently low number of mesh elements. In order to validate the
achieved numerical solutions, they are compared with the experiments and other
available numerical results. The proposed approach can be also used in simulations
of other turbomachinary cases, where there are serious computer memory and
computational time limitations. For instance, using the current periodic boundary
condition approach, one can readily simulate very huge wind tunnels considering a
full 3D model of its fans, i.e., including its rotor and stator instead of considering
simple fan pressure jump or actuator disc models.
S. Sabzpoushan • M. Mohammadi
Department of Aerospace Engineering, Sharif University of Technology, Tehran, P.O. Box
11365-8639, Iran
M. Darbandi ()
Department of Aerospace Engineering, Center of Excellence in Aerospace Systems,
Sharif University of Technology, Tehran, P.O. Box 11365-8639, Iran
G.E. Schneider
Department of Mechanical and Mechatronics Engineering, University of Waterloo, Waterloo,
ON, N2L 3G1, Canada

402 S. Sabzpoushan et al.
1 Introduction
Fan has various applications in different industrial equipment, e.g., heat exchangers,
air conditioning systems, and wind tunnels. In case of applications with large
volume of air, the industries may have to consider using parallel fans. However,
this may lead to a difficulty that the flow at upstream and/or downstream of fans
may not perform axisymmetric pattern. In such cases, one may not simply apply
the periodic boundary condition at the boundaries of a slice of flow passage. It is
because if one wishes to simulate such a flow field, this would cause pitch ratio at
the interfaces between the moving rotor domain and the stationary ducts domains,
which in turn will cause critical regions at such interfaces. We should be careful on
two key points at these interfaces. First, one should know that the face mesh at these
interfaces must be as similar as possible to avoid missing data from one zone to the
next one [1]. Second, there is an important constraint indicating that the pitch ratio
cannot be further up than a certain limit [2]. As a rule of thumb, the pitch ratio must
not exceed the upper limit of about 10 if one expects an almost reliable simulation
[3]. Literature shows that there have been several efforts to simulate axial fans
assuming both full rotor disc consideration and periodic blade boundary condition
implementation. Fidalgo [1] simulated a three-dimensional unsteady flow through a
full-annulus rotor disc with distortion in its inlet total pressure. Additionally, Denton
[2] raised some limitations in turbomachinary CFD indicating that if the engineers
do not care those limitations, they may not get solutions with sufficient accuracies.
He specially focused on specific geometries such as the tip clearance and leading
edge shapes. Cevik [4] performed a simulation on the full geometry of an axial fan
with a diameter of 3.13 m placed in a duct with 12.5 m length. Using the k-epsilon
turbulence model, the total number of mesh elements were about 2,500,000. There
are some other publications, which use the true periodic boundary condition. For
example, Le Roax [5] simulated the rotor disc and its upstream/downstream ducts
completely and all together using the periodic boundary condition.
At this stage, we need to clarify the “pitch ratio” importance in axial fan
modelling by introducing a simple example (see Fig. 1). Prior to any explanation,
we should mention that this simple example is so helpful to examine the validity
and reliability of current periodic simulation at next stage. This can be done for
more complicated cases by comparing its solution with the fan experimental data
and full blade rotor simulation results. As it is shown in Fig. 2(a), the target fan
has 12 blades located in a long duct. For this geometry, the pitch ratio is 360ı /360ı
D 1 at the interface of rotational and stationary zones. Now, suppose a periodic
domain in which, we consider only two blades of the fan with (2/12)360ı D 60ı
annulus slice cross section. This slice would interact with the full 360ı domains
at its downstream and upstream zones. For this geometry, the pitch ratio becomes
360ı/60ı D 6. In this article, we focus on mass flow rate and flow uniformity rather
than other parameters. In case of using parallel set of fans, interactions between the
inflow and outflow streams would also become important.
Numerical Investigation on Periodic Simulation of : : : 403
Fig. 1 Two types of rotor disc geometry. (a) Full rotor disc. (b) Sliced rotor disc
(a) (b)
Fig. 2 The fan and its rotor blade geometries. (a) The assembled fan. (b) The result of rotor blade
CMM
2 Geometry and the Solution Domain
The length of rotor blade shown in Fig. 2 is 0.5 m from root to tip and the diameter
of its casing is 2 m. The maximum allowable rotational speed is about 900 rpm. In
order to eliminate the swirl of flow through the rotor, a stator has been embedded
downstream of the rotor, if this swirl can have considerable effects. Fortunately, in
case of using periodic rotor disc, the flow passage in sliced rotor domain does not
need to be coincident with the streamlines. It is because any amount of air exiting
from one of the two periodic sides, would enter from the other side.
Before applying the current approach in complicated geometries with complex
conditions, it is better to examine it in a simple geometry test case. Figure 3 presents
this simple test case, which is constructed based on a complex wind tunnel air driver
geometry. According to the standards related to the methods of fan testing [4, 6, 7],
the minimum lengths for the upstream and downstream ducts of an axial fan are
assumed to be 10D and 5D respectively, where D is fan casing diameter. We should
satisfy these recommended minimum standard values to let the flow become fully
developed before passing through the rotor and also at the end of downstream duct.
Fig. 3 The fan between forward and aft ducts in a simple test geometry
3 Grid Distribution and Boundary Conditions
Figure 4 shows rotor and stator blade sections accompanied with the generated
mesh. The boundary layer mesh and mesh concentration have been used wherever
needed. The boundary layer mesh is controlled suitably to achieve a maximum
thickness of about 10 cm.
Evidently, it is unavoidable facing with some limitations in mesh generation. The
sources can be due to the geometric constrains. For example, we are faced with very
tiny gap at the blade tip clearance, which has considerable effects on the achieved
fan performance; especially the fan noise generation [8, 9]. This tiny gap makes it so
difficult to connect the boundary layers generated on the upper and lower surfaces
to the blade tip. However, there are two possible approaches to resolve this problem.
One is to reduce the number of layers at the blade’s tip and casing. The other one is
to ignore the layer’s growth rate and to compress them. This sometimes requires to
change the first layer height. The first approach needs a sudden detraction of layers,
which is harmful for the numerical solution, especially for capturing the separation
Fig. 4 Two sample sections of mesh distributions around the rotor blade and the stator vane.
(a) Rotor blade leading edge. (b) Stator vane
Fig. 5 The mesh cluster between the rotor blade tip and its casing
and tip vortices. Hence, the second approach is chosen, see Fig. 5. Controlling y+ in
the range of 1–100, we can employ k-! SST turbulence model. As another concern,
it is important for the surface mesh on the blade to preserve the real shape of leading
edge [5, 10]. Finally, Fig. 6 enforces this point that the surface mesh on both sides
of an interface with any pitch ratio must be as similar as possible.
If we choose a slice of rotor disc (and that of the stator disc if it exists) instead
of the entire rotor and stator discs, the number of elements (and consequently the
required computer memory) will decrease considerably. This needs to apply suitable
periodic conditions. After performing some mesh independency efforts, we achieved
to a grid resolution of approximately 600,000 elements. This is where the entire
rotor disc solution domain needed about 1,700,000 elements to result in mesh-
independent solution. Such mesh studies were conducted based on monitoring both
Fig. 6 Similar mesh generations at the stationary and rotational domains’ interface
0.510
0.500
0.490
0.480
Y (m)
0.470 Coarse (400,000)

medium (600,000)
0.460 fine (1,000,000)
0.450
0.440
0.430
0.0 20.0 40.0 60.0 80.0
X-velocity (m/s)
Fig. 7 The mesh refinement study via monitoring the boundary layer velocity profile in three
different grid resolutions
the mass flow rate magnitude and the boundary layer velocity profile. Figure 7 shows
the results of mesh refinement study for the sliced rotor considering the boundary
layer velocity profile inside the duct wall. This figure also verifies our estimation on
the boundary layer thickness which is about 6 cm, whereas the boundary layer mesh
had been generated with a total thickness of 10 cm.
Here, we would like to mention some points on the chosen boundary conditions.
The boundary conditions at the inlet and outlet sections of the duct are not fixed by
neither mass flow rate, nor velocity magnitude and direction. So, the solution can be
gradually converged during the numerical procedure. We determine the turbulence
intensity of 1 % at the inlet section and the static temperature of 298 K at this
boundary. Determining the total pressure at the inlet and the average static pressure
at the outlet could be other possible choices as the boundary conditions. These
two alternative choices require more flow parameters and constraints to be defined
at these boundaries and hence lower the problem generality [11]. The interface
connections are chosen in a manner to enforce equal area-weighted average of
normal velocities at both sides of the interface. To achieve more similarity with the
real conditions, the rotor casing (shroud) is forced to have a counter-rotating RPM
relative to the rotational domain in order to be motionless in stationary reference
frame [10, 12, 13]. The ambient pressure is equal to the value reported in standard
atmosphere table for Tehran altitude, which is about 87,000 Pa.
4 The Results and Discussion
According to the experimental tests performed at the atmospheric pressure of

87,000 Pa under the standard conditions, the air mass flow rate for a single fan at
the maximum rotational speed is about 53~54 kg/s. As seen in Fig. 8, a high pitch
ratio would lead to invalid solution with about 60 % deviation from the exact value.
However, it promptly drops below 10 % after this point and approaches rapidly to
the exact solution. Thus, this can not only help us to find the proper pitch ratio
magnitude, but also can be used as a type of validation for the proposed simulation
case.
Figure 9 shows the tangential velocity contours just after the exit of rotor (before
entering the stator) and right after the stator. To elaborate the swirl and the effect
of stator on its decay, Fig. 10 presents the streamlines passing the fan and compares
the flow swirl in the absence and presence of stator. The symmetry in these contours
Numerical simulation Experimental test

90
85
80
75
Air mass flow rate (kg/s)
70
65
60
55
50
45 Full rotor disc simulation
40
0 0.2 0.4 0.6 0.8 1
Pitch ratio (rotating-to-stationary)
Fig. 8 Effect of pitch ratio on the calculated mass flow rate through the axial fan
(a) (b)
Fig. 9 Tangential velocity contours, indicating the swirl strength. (a) Just at the downstream of
rotor. (b) right after the stator
and non-swirling flow at the upstream of the rotor (see streamlines in Fig. 10) also
imply that the current periodic simulation performed reliably.
To provide more qualitative validations, one can check the pressure change over
the fan. The current calculations showed that the area-weighted average of static
pressure (gauge) over two sections somewhere at middle lengths of upstream and
downstream ducts would be about 156 and 23 Pa, respectively. This value is about
180 Pa at inlet and 2 Pa at outlet. They indicate that the static head of fan would be
approximately 180 Pa. Figure 11 shows the static pressure contours at two sections
just before and after the rotor disc. According to available fan performance curves,
this is exactly what we expect for the static head of such a fan with a relatively low
incidence angle (about 17ı ) of rotor blades.
Figures 12 and 13 show two important detected phenomena, which support
the current periodic boundary condition approach in axial fan simulation. The
first one is small vortices, which can potentially cause flow to separate from the
blade surface. The second one is the flow leakage from the blade pressure side
to its suction side through the gap at tip clearance. It is also known as downwash
phenomenon.
5 Conclusion
We introduced a new periodic boundary condition implementation, which was

relatively different from the past classical ones. This new method can be readily
used to simulate a single fan or a set of several parallel fans despite facing with
asymmetric flow conditions. The major obstacle is the pitch ratio, which occurs at
the interfaces between the rotor domain and the stationary upstream and downstream
ducts’ domains. After practicing various pitch ratios, we eventually found an
optimum magnitude, which would provide suitable mesh sizes with guaranteed
numerical accuracy. Our study showed that the number of mesh elements required to
(a)
(b)
Fig. 10 The streamlines inside the duct at a rotor speed of 900 rpm (a) without stator (b) with
stator
simulate a full-geometry fan in such applications is more than 11 times the number
of elements required using the proposed periodic approach. Indeed, this reduction
in the number of elements is so essential in many real industrial applications. For
example, it is important when one wishes to simulate the set of several combined
fans, e.g., a set of four parallel fans, utilized in a huge air supply ventilation system.
Z
a Pressure
Contour 4
–266.91
b Pressure
Contour 5
363.20
Z
–275.37 326.86
–283.84 290.52
–292.30
–300.77 X Y 254.18 X Y
–309.24 217.84
–317.70 181.50
–326.17 145.16
–334.63 108.82
–343.10
–351.56 72.48
–360.03 36.14
–368.49 0.20
–376.96 36.54
–385.43 72.88
–393.89 109.22
–402.36
–410.82 145.56
–419.29 181.89
[Pa] [Pa]
Fig. 11 Contours of gauge static pressure. (a) Just before passing the rotor. (b) Just after passing
the rotor
Fig. 12 Demonstration of small vortices appearing close to the rotor blade face
Velocity in Stn Frame

Vector 1 [m s^–1]
7.
27
47
66
86
31
.1
.0
.9
.8
8
Rotor blade tip
Flow direction
Fig. 13 Demonstration of flow leakage from the high pressure side to the low pressure side
through the tip clearance gap
Acknowledgements The authors would like to greatly thank the financial supports received from
the Deputy of Research and Technology of Sharif University of Technology (SUT).
References
1. Fidalgo, J.: A Study of Fan-Distortion Interaction with the NASA Rotor-67 Transonic Stage.
In: ASME Turbo Expo, June 2010, Glasgow (2010)
2. Denton, J.D.: Some Limitations of Turbomachinery CFD. In: ASME Turbo Expo, June 2010,
Glasgow (2010)
3. Bhasker, C.: Simulation of three dimensional flows in industrial components using CFD
techniques. In: Minin, I. (ed.) Computational Fluid Dynamics Technologies and Applications.
InTech (2011). ISBN:978-953-307-169-5, doi:10.5772/19909
4. Cevik, F.: Design of an axial flow fan for a vertical wind tunnel for paratroopers. Master thesis,
Mechanical Engineering Department, Middle East Technical University (2010)
5. le Roux, F.N.: The CFD simulation of an axial flow fan. Master thesis, Department of
Mechanical and Mechatronic Engineering, University of Stellenbosch (2010)
6. AMCA 210: Laboratory Methods of Testing Fans for Aerodynamic Performance Rating. Air
Movement and Control Association International, Inc., Arlington Heights, IL (1999)
7. Cory, W.B.: Fans and Ventilation: A Practical Guide. Elsevier-Roles Ltd, Amsterdam (2005)
8. Srinivas, G., Srinivasa Rao., P.: Numerical simulation of axial flow fan using Gambit and
Fluent. IJRET 3(3), 586–590 (2014)
9. Raj, S.A., Pandian, P.P.: Effect of tip injection on an axial flow fan under distorted inflow.
IJASER 3(1), 302–309 (2014)
10. Dwivedi, D., Dandotiya, D.S.: CFD analysis of axial flow fans with skewed blades. IJETAE
3(10), 741–752 (2013)
11. Augustyn, O.P.H.: Experimental and numerical analysis of axial flow fans. Master thesis,
Stellenbosch University (2013)
12. Raj, A.S., Pandian, P.P.: Numerical simulation of static inflow distortion on an axial flow fan.
IJMERR 3(2), 20–25 (2014)
13. Sarmiento, A.L.E., Gamboa, Y.F.Q., Oliveira, W., Camacho, R.G.R.: Performance analysis
through computational fluid dynamics of axial rotor with symmetric blades used in tunnel
ventilation. HIDRO and HYDRO, PCH NOTICIAS and SHP NEWS 60(1), 22–25 (2014)
Numerical Analysis of Turbulent Convective
Heat Transfer in a Rotor-Stator Configuration
D.-D. Dang and X.-T. Pham
Abstract This paper presents the numerical analysis of convective heat transfer
of a rotor-stator configuration, which is typically found in hydro-generators. The
Reynolds Averaged Navier Stokes (RANS) turbulence models based on the eddy-
viscosity approximation were employed. Different steady and unsteady multiple
frames of reference models were used to deal with the flow interaction in the
rotor-stator system. The fluid flow and heat transfer analysis were performed using
conjugate heat transfer methodology, in which the governing equations for the fluid
dynamics, heat conduction with additional constraints on the fluid-solid interface
were simultaneously solved. The computed convective heat transfer coefficient was
compared against available experimental data to assess the suitability of turbulence
models.
1 Introduction
During the operation of hydro-generators, the heat generated by the electromagnetic

and electrical loses causes the internal heating in the solid components. The solid
temperatures that are too high might reduce the lifetime or at worst result in the
breakdown of the machine.
In order to maintain the temperature of the solid parts within a safe margin,
the heat generated is removed by a circulation of cooling air. The most important
parameter governing the thermal performance of the hydro-generators is therefore
the convective heat transfer coefficient (CHTC) on the boundary between the
cooling fluid and the solid surface. Traditionally, CHTCs were approximated by
the convective empirical correlations which are available in literature and widely
D.-D. Dang () • X.-T. Pham

Ecole de technologie superieure, Montreal, QC H3C 1K3, Canada
e-mail: [email protected],[email protected]

414 D.-D. Dang and X.-T. Pham
used for thermal analysis. However, those correlations were originally established
for simple configurations, and might not be applicable or insufficiently accurate for
the case of complex geometries.
Prediction of convective heat transfer coefficient in hydro-generators is crucially
important but not a trivial task due to the complexity of flow dynamics in the
machine. The flow field in the rotor-stator system exhibits a number of phenomena
which are challenging for numerical modelling such as rotation, turbulence and
unsteady nature.
With respect to methodology, Lumped-Parameter Thermal Network (LPTN),
Finite Element Analysis(FEA) and Computational Fluid Dynamics(CFD) are the
most common approaches for thermal analysis of electrical machines [2]. Although
the analytical LPTN and FEA have advantage of being very fast and efficient, these
approaches require the knowledge of CHTCs in prior to the computation, which are
challenging to obtain and therefore not always available. These methods are beyond
the scope of the current research that focuses mainly on CFD application.
Pickering et al. [11] are considered as the first authors who applied CFD to
predict the temperature in the solid of electrical machines. The numerical analysis
of an air-cooled, 4-pole generator case concluded that CFD showed to be a potential
tool for thermal analysis of electrical machines. However, because of limited
computing resources at the time, only simulations with a coarse mesh near the
pole surface could be performed. Depraz et al. [3] studied the cooling system of a
large hydro-generator using the three-dimensional flow simulation. The comparison
between the numerical results by CFD and traditional two-dimensional network
based flow calculation showed a good agreement. Li [7] employed the conjugate
heat transfer (CHT) methodology to predict the temperature distribution in the solid
pole of the large hydro-generators. The CHT calculations used a coarse mesh on
the wall surface for wall-functions utilization. However, the comparison with the
experimental data was not carried out in his work. Recently Toriano et al. [12]
proposed a hybrid method combining the numerical and experimental techniques to
evaluate the heat transfer coefficient on the pole-face of a scaled rotating prototype.
Although numerous factors might affect the heat transfer prediction by conjugate
heat transfer approach, the aim of the current study is limited to investigate the
sensitivity of implemented numerical parameters on the prediction of fluid flow
and heat transfer in the rotor-stator system. In particular the influence of turbulence
models and steady-state Multiple Frames of Reference models are analyzed.
2 Computational Domain and Meshes
Because of the limited access on the real generator for experimental measurement,
a scaled prototype was built. Since it is not realizable from the computational
perspective to perform simulations of the entire scaled rotating model with an
Numerical Analysis of Turbulent Convective Heat Transfer in a Rotor-Stator. . . 415
Fig. 1 Computational domain of scaled rotating model
adequately fine grid; calculations were carried out in two steps. First, the simulation
with the full scaled model (Fig. 1) has been computed with a coarse 86 M hexahedral
cells grid using the standard k turbulence model. This calculation was then used
to generate the conditions at the inlet boundary for the following second step of
the calculation. The second configuration was compromised to only one section of
10 degree single-pole in (r; ) plane which allowed to perform with a finer mesh
up to 1.6 M hexahedra cells (Fig. 2). The extensive parametric analysis will be
performed on this simplifed 2-D configuration using various turbulence and multiple
frame of reference models.
Fig. 2 Computational domain of two-dimensional simplified configuration
3 Mathematical Model and Numerical Method
3.1 Governing Equations
The governing equations for incompressible steady state flow [4] are expressed in
the Cartesian tensor notation for a rotating coordinate system as follows.
Continuity equations:
@
.Ui / D 0I (1)
@xi
Momentum transport equation:

@ @P @ @Ui @Uj
.Ui Uj / D C . C / ui uj
@xj @xi @xj @xj @xi

2"ijp ˝p Uj ˝j Xj ˝i ˝j Xi ˝j (2)
„ ƒ‚ … „ ƒ‚ …
centrifugalforce Coriolisforce
The two last terms in momentum equations were added to account for the Cori-
olis and centrifugal forces in the rotating frame of reference. These contributions
however are eliminated in the stationary frame of reference.
Energy equation:

@ @ @T @
Uj htot D f uj h C Ui ij ui uj C SE ; (3)
@xj @xj @xj @xj
where the specific total enthalpy for ideal gas is given by:
P U2 1
htotal D Cp T C C i C u2i (4)
2 2
The heat conduction equation in the solid has the same form as the energy
equation for fluid Eq. 3, except that the velocity components are set to be zero, and
the solid thermal conductivity s was substituted.
On the fluid-solid interface, additional constraints need to be added to ensure the
equilibrium of heat flux and the temperature continuity through the two mediums.
Tsolid D Tfluid
@T @T
s jsolid D f jfluid (5)
@n @n
3.2 Turbulence Models and Boundary Layer Modelling
The presence of Reynolds stresses ui uj in Eqs. 2–3 means that the derived
equations are not closed, which necessitates additional equations for closure. The
turbulent stresses and turbulent heat flux are obtained using the effective viscosity
approximation:

2 @Ui @Uj
ui uj D kıij t C (6)
3 @xj @xi
t @h
uj h D (7)
Prt @xj
The turbulent eddy-viscosity t in Eqs. (6) and (7) is calculated by the turbulence
models. Several turbulence model bases on the eddy viscosity approximation are
implemented to calculate the flow dynamics and heat transfer. Standard k
(SKE) [6] model was developed for high-Re flow and has been widely used in
industrial simulation because it compromises between the robustness and accuracy.
However, SKE was consistently found inappropriate for the low-Re flows or flows
with separation and reattachment of boundary layer. To improve the prediction, the
Shear Stress Transport (SST) model [8] was also implemented. Menter [9] presented
several test cases to show that the SST k ! predicts flows with separation in the
presence of adverse pressure gradients more accurately than the SKE model.
Boundary layer modelling Modelling of the near-wall region is of major impor-
tance for CFD simulations, especially for the cases in which the prediction of the
flow quantities on the wall (skin friction, convective heat transfer) is important.
In principle, low-Reynolds number (LRN) and wall-functions are two common
approaches for the near-wall modelling. The LRN model refers to an approach that
resolves the entire boundary layer using a very fine mesh in the near wall regions.
The grids used for the low-Re number model require typically a dimensionless wall
distance of the wall-adjacent cell about unity, i.e. yC < 1. The grid for the low-Re
number is shown in Fig. 3 (right), in which the wall-space of the first cell was set
as 6.103 mm. Because of high computational cost associated with LRN model, the
wall-functions are often used instead. The idea behind the wall-functions approach
is to place the first computational node into the logarithmic layer and employ a
semi-empirical dimensionless profile to obtain the wall shear stress. The advantage
of this approach is that the boundary layer can be resolved with an adequately small
number of grid points. The wall functions approach requires the yC value of the first
node between 30 and 300, i.e., 30 < yC < 300. In the present CFD code, the LRN
and wall-functions models are employed for the !base and based turbulence
models, respectively.
Rotor-stator interaction The fluid flows in a rotor-stator system having non-
axisymmetric components are always unsteady, which is a challenge for CFD
simulations [10]. In an effort to make the rotor-stator analysis practical with limited
computing resource, several steady-state Multiple Frames of Reference (MFR)
models have been developed. In these models, an interface is defined between
rotating and stationary components in a manner so that the steady-state calculations
in each frame are supported while maintaining as much interaction between the
components as possible [5]. There are two steady-state interface techniques are
Fig. 3 Grids for different

boundary layer models:
wall-functions (left) and
low-Re number model (right)
available in the present CFD code. The first model is called Mixing Plane, where the
upstream flow velocity profile is first averages circumferentially before transferring
to the downstream region. In this context, any non-uniformity in the circumferential
direction will not be preserved in the next region. The second type of steady-state
interface is called Frozen Rotor, where the flow profile variation is now preserved
across the interface, however, the relative position between the rotor and stator is
fixed in time and space. Also, the transient sliding interface is also available to model
the unsteady flow due to the relative motion between the rotating and stationary
components. In this model, the flow field variation is fully taken into account.
3.3 Boundary Conditions and Convergence Criteria
The boundary conditions for the configuration illustrated in Fig. 2 are defined
as follows. The fluid velocity at the inlet boundary was set to be uniform, and
normal to the inlet surface. The average value of the inlet profile was derived from
the 3D simulation on the coarse grid with SKE turbulence model. For the case
with rotational speed 50 rpm, Uinlet D 2:10 m/s. The inlet temperature was set to
298 K (25 ı C). For all wall boundaries, no-slip conditions with zero velocity were
assumed. The circumferential periodicity feature of the geometry was specified in
the numerical model by defining rotational periodic interfaces. On the rotor-stator
interface, the general grid interface (GGI) was defined, which allows performing
calculations of non-conforming meshes in a conservative manner. The governing
equations presented in Sect. 2 were solved by CFD code ANSYS CFX-15.0.
Considering the recommendation of ASME V&V numerical accuracy guideline [1],
all equations were solved using second-order accuracy schemes.
Convergence was judged by examining the residual levels as well as by moni-
toring the relevant variables at critical locations. The used criterion requires that the
maximum residual of all equations dropped at least by the order of 104 and the
variables of interest at considered locations keeps constant at the steady-state value.
4 Results and Discussions
The flow and turbulence structure in the inter-pole duct calculated using mixing-
plane model with the SST k ! is illustrated in Fig. 4. The flow structure observed
are typical for the flow in the duct with forward facing-step, in which flow with
uniform velocity at the inlet develops in upstream region until it impinges on the
step; a circulation bubble is formed in the step corner due to the adverse pressure
Fig. 4 Velocity contour (left) and turbulent kinetic energy contour (right) predicted by mixing-
plane
Fig. 5 Normalized radial velocity in the inter-pole duct predicted by different MFR models
gradient caused by step blockage. The boundary layer separates over the step and
reattaches to the pole surface thereafter. Due to the system rotation the flow structure
is asymmetric in the inter-pole duct; the Coriolis force have the effect of shrinking
the recirculation size on the leading edge and enlarging the bubble size on the
trailing edge.
The steady state models frozen rotor and mixing-plane was employed for an
inherently unsteady flow in rotor-stator interaction, which is necessary to verify. The
calculated result using these models are compared with the time-averaged transient
simulation as reference result. Figure 5 shows the normalized radial velocity in the
Fig. 6 Average heat transfer coefficient on the pole-face computed by different turbulence models
duct at r D 1.12 m for different MFR models. It is observed that the mixing-plane
model shows a better agreement with the time-average transient model meanwhile
the result computed by FR model significantly varies depending the relative position
of rotor and stator. This results implies that the circumferential flow variation that
each passages rotates during a full revolution is small at the rotor-stator interface.
Since the simulation using mixing-plane model requires less computational effort
than transient simulations, this model will be employed for the further parametric
analysis.
Figure 6 reports the comparision of average CHTC on the pole face predicted
by different turbulence models closure, including the standard k (SKE), the Re-
Normalisation Group k (RNG), the standard k ! (SKO) and SST k ! (SST)
models. The averaged value is calculated by Eq. 8.
R
qw dA
hD R A (8)
A Tw Tref dA
where the Tref is the reference temperature, which is fluid temeprature at 5 mm away
from the pole face for both numerical and experimental approach. The experimental
data is extracted from Toriano et al. [12]. It can be seen that all turbulence models
predict higher CHTC in comparison with experimental data. The results calculated
by SST k ! shows the best agreement with the experiments, the relative error is
calculated as 11 %.
In Fig. 7, the dimensionless temperature T profiles are compared for different
positions at the pole-face along lines normal to the surface, as functions of y ,
Fig. 7 Profiles of dimensionless temperature T as functions of y along lines normal to the pole
face
defined by:
1=4
C k1=2 .Tw T/ Cp
T D (9)
qw
1=4
C k1=2 y
y D (10)

According to Eq. 9, the higher T values results in a lower wall heat flux in
case of the same level of predicted turbulence kinetic energy(TKE). However, the
TKE predicted by wall-functions approach are found much higher than the low-
Re number model. The example of the TKE profiles at two positions 2ı and 3ı
are illustrated in Fig. 8. Overall, this results in the significant over-estimation of
the CHTC by the wall function, as showed in Fig. 6. The maximum discrepancy
between the CHTC predicted by wall-functions and experimental data is up to 50 %.
Fig. 8 Normalized turbulence kinetic energy in the air-gap predicted by different turbulence
models
5 Conclusion
High resolution steady-state RANS CFD simulations of conjugate heat transfer in a

rotor-stator configuration was performed. The focus was the sensitivity analysis of
numerical parameters on the fluid flow and convective heat transfer prediction on
the pole-face. The results showed that the mixing-plane model demonstrated a good
agreement with the transient model than the frozen rotor, whose results significantly
varies depending on the relative position of the rotor and stator. The heat transfer
coefficient prediction by different turbulence models shows that the SST with low-
Re model for the near-wall region is more appropriate than the wall-functions, which
over-estimates up to 50 % this coefficient on the pole face on average.
Acknowledgements The support of the Natural Sciences and Engineering Research Council
of Canada(NSERC) and Fonds de recherche du Quebec Nature et technologies(FRQNT) are
gratefully acknowledged. We are also grateful for the funding and computing infrastructure
provided by Institut de recherche d’Hydro-Quebec(IREQ).
References
1. ASME Standard for verification and validation in computational fluid dynamics and heat
transfer.: ASME V&V 20-2009, American Society of Mechanical Engineers (2009)
2. Boglietti, A., Cavagnino, A., et al.: Evolution and modern approaches for thermal analysis of
electrical machines. IEEE Trans. Ind. Electron. 56(3), 872–882 (2009)
3. Depraz, R., Zickermann, R., Schwery, A., Avellan, F.: CFD validation and air cooling design
methodology for large hydro generator. In: Proceedings of 17th International Conference on
Electrical Machines ICEM, Crete Island, Greece (2006)
4. Ferziger, J.H., Peric, M.: Computational Methods for Fluid Dynamics. Springer, Berlin/Hei-
delberg (2001)
5. Galpin, P., Broberg, R., Hutchinson, B.: Three-dimensional Navier-Stokes predictions of steady
state rotor-stator interaction with pitch change. In: Proceedings of 3rd Annual Conference of
the CFD Society of Canada, Banff, Alberta (1995)
6. Launder, B.E., Spalding, D.B.: The numerical computation of turbulent flows. Comput.
Methods Appl. Mech. Eng. 3(2), 269–289 (1974)
7. Li W, Guan C, Chen Y (2014) Influence of rotation on rotor fluid and temperature distribution
in a large air-cooled hydrogenerator. IEEE Trans. Energy Convers. 28, 117–24
8. Menter, F.R.: Improved Two-Equation k-omega Turbulence Models for Aerodynamic Flows.
NASA TM-103975 (1992)
9. Menter, F.R., Kuntz, M., Langtry, R.: Ten years of industrial experience with the SST
turbulence model. Turbul. Heat Mass Transf. 4, 625–32 (2003)
10. Moradnia, P., Chernoray, V., Nilsson, H.: Experimental assessment of a fully predictive CFD
approach, for flow of cooling air in an electric generator. Appl. Energy 124, 223–30 (2014)
11. Pickering, S.J., Lampard, D., Shanel, M.: Modelling ventilation and cooling of the rotors
of salient pole machines. In: IEEE International Electric Machines and Drives Conference,
Cambridge, MA, pp. 806–808 (2001)
12. Torriano, F., Lancial, N. et al.: Heat transfer coefficient distribution on the pole face of a
hydrogenerator scale model. Appl. Therm. Eng. 70(1), 153–162 (2014)
Determining Sparse Jacobian Matrices Using
Two-Sided Compression: An Algorithm
and Lower Bounds
Daya R. Gaur, Shahadat Hossain, and Anik Saha
Abstract We study the determination of large and sparse derivative matrices using
row and column compression. This sparse matrix determination problem has rich
combinatorial structure which must be exploited to effectively solve any reasonably
sized problem. We present a new algorithm for computing a two-sided compression
of a sparse matrix. We give new lower bounds on the number of matrix-vector
products needed to determine the matrix. The effectiveness of our algorithm is
demonstrated by numerical testing on a set of practical test instances drawn from
the literature.
1 Introduction
The determination of the Jacobian matrix F 0 .x/ of a mapping F W <n ! <m is a

key computation in Newton’s method for solving nonlinear optimization problems.
Using forward difference (FD), the product of the Jacobian matrix with a vector s
may be approximated as
ˇ
@F.x C ts/ ˇˇ 1
ˇ D F 0 .x/s As ŒF.x C s/ F.x/ b; (1)
@t tD0
with one extra evaluation of F at .x C s/, where > 0 is a small increment,

assuming F.x/ has already been evaluated. Algorithmic (or Automatic) Differenti-
ation (AD) [7] gives b D F 0 .x/s and c D w> F 0 .x/ in forward mode and reverse
mode, respectively. The numerical values are accurate up to the machine round-
off and the cost of the matrix-vector product is a small multiple of the cost of
one function evaluation. The key observation is that the product of the Jacobian
matrix of function F, evaluated at a given point x, with a given vector s can be
obtained as As or A> s at a cost that is a small multiple of the cost of evaluating
the function F at x [7]. Therefore, in what follows, we express the computational
D.R. Gaur • S. Hossain () • A. Saha

Department of Mathematics and Computer Science, University of Lethbridge, Lethbridge,
AB, Canada

426 D.R. Gaur et al.
Fig. 1 Algorithm for FD 1 for j 1 to n

estimation of the Jacobian 2 do
1
matrix of F at x 3 B(:, j) [F(x + I(:, j)) − F(x)]
Fig. 2 A row compression a11 0 0 0 0 a16 1 0 0 0

matrix S a21 0 a23 a24 a25 0 1 0 0 0
a31 0 a33 0 0 0 0 1 0 0
B=
0 a42 0 a44 0 0 0 0 1 0
a51 0 0 0 a55 0 0 0 0 1
0 a62 0 0 0 a66 0 1 0 0
A S
a11 a16 0 0
a21 a23 a24 a25
a31 a33 0 0
=
a42 0 a44 0
a51 0 0 a55
a62 a66 0 0
cost of determining a Jacobian matrix in terms of the number of matrix-vector

products (MVPs) of the kind As or A> s. The complete Jacobian can be determined,
for example, by differencing along n coordinate directions as shown in Fig. 1 where
I denotes the identity matrix. We use colon notation of [5] to specify submatrices
of a matrix. In AD reverse mode, the Jacobian matrix can be determined row-
by-row as vector-matrix products s> A. Many large-scale practical optimization
problems often contain useful structure. For example, the component functions Fi
nonlinearly depend only on a small subset of the n independent variables giving
rise to partially separable structure [6]. It is usually the case that zero-nonzero
pattern of the Jacobian is a priori known or can be computed easily. Consider the
matrix A in Fig. 2. Define matrix S where S.W; 1/ D e1 C e2 ; S.W; 2/ D e3 C e6 ;
S.W; 3/ D e4 ; S.W; 4/ D e5 and compute the product B D AS using, for example,
the algorithm in Fig. 1. It is clear from Fig. 2 that the FD estimation of the unknown
elements of matrix A can simply be read-off, i.e., without incurring any floating-
point operations, from the compressed matrix B at a cost of only 4 (instead of
6) additional evaluations of function F. Alternatively, we may define a column
compression matrix W where W.W; 1/ D e2 C e6 ; W.W; 2/ D e4 C e5 ; W.W; 3/ D
e1 ; W.W; 4/ D e3 and compute the compressed matrix C> D W > A in reverse mode
AD. In either case, 4 matrix-vector products are required to completely determine
the matrix. We say that matrix A is directly determined if all the nonzero unknowns
can be read-off from the compressed matrices B or C as defined above. The Jacobian
matrix determination problem (JMDP) based on matrix-vector product calculation
can be stated as below (See [10]).
Determining Sparse Jacobian: Algorithm and Lower Bound 427
Given the specification of the sparsity pattern for matrix A, obtain matrices S 2 f0; 1gnp
and W 2 f0; 1gmq such that matrix A is directly determined and p C q is minimum.
A group of columns satisfying the property that no two of them contain nonzero
entries in the same row position is called structurally orthogonal. The observation
that a group of structurally orthogonal columns can be approximated with only
one extra function evaluation was reported first in a seminal paper by Curtis,
Powell, and Reid [3]. Their row compression algorithm1 was further analyzed by
Coleman and Moré [1] who gave the first graph coloring interpretation of the
compression problem. The classical arrow-head matrix example demonstrates that
one-sided compression (i.e., either column or row) may not yield full exploitation of
sparsity [9]. Hossain and Steihaug [9], and Coleman and Verma [2] independently
proposed techniques for two-sided compression (i.e., combined row and column
compression) for the sparse Jacobian determination problem. Unfortunately, sparse
Jacobian determination using compressions (one-sided or two-sided) is NP-hard [2]
implying that exact methods are impractical except for small problem instances.
Greedy heuristics complete direct cover (CDC) by Hossain and Steihaug [9] and
minimum nonzero count ordering (MNCO) by Coleman and Verma [2] produce
matrices S and W by considering the matrix rows and columns in specified orders.
Recently, Juedes and Jones [11] proposed an approximation algorithm for minimum
2
star bi-coloring (ASBC) with an approximation guarantee of O.n 3 / of the optimal.
An easy to compute “good” lower bound on the number of groups allows one to
measure the effectiveness of such algorithms. For row compression (of A or A> ) a
satisfactory lower bound on the number of matrix-vector products (or AD passes)
is given by max .A/ D maxi i where i denotes the number of nonzero entries in
nnz.A/
row i. For two-sided compressions, [8] proposed the expression d maxfm;ng e where
nnz.A/ denotes the number of nonzero entries in matrix A. The bound is derived
using a consistency argument in solving the system of linear equations defined by
the products B D JS and C D W > J. Juedes and Jones [11] derive the same bound
using a graph theoretic argument. In this paper we propose a new and more effective
lower bound which generalizes the above lower bound. We propose a new heuristic
algorithm for two-sided compression of sparse Jacobian matrices. We demonstrate
the effectiveness of our lower bound and algorithm by extensive computational
experiments on a standard set of test instances. The remainder of the paper is
organized as follows. In Sect. 2, we derive the new lower bound on the number
of matrix-vector products which is followed by the description of the new two-sided
compression algorithm. Section 3 contains results of numerical experiments from
the lower bound computation and the two-sided compression algorithm on a set
of test matrices. Test results indicate that the new lower bound is superior to the
existing one. On many of the test matrices, our algorithm produced better results.
The paper is concluded in Sect. 4 with comments on future research.
1
A row compression is a one-sided compression where only seed matrix S is defined.
2 Two-Sided Compression
As discussed in Sect. 1 the example in Fig. 2 can be determined using 4 MVPs using
either row or column compression with one-sided compression. This is also the
minimum possible since the maximum number of nonzero entries in any row of A
or A> , max .A/ D max .A> /, is 4. However, a closer examination of Fig. 2 reveals
that the last two columns of the compressed matrix B are quite sparse – only two
of the entries are nonzero. In other words, there are available sparsity which are not
exploited by one-sided compressions. On the other hand, with the following two-
sided compression
0 1 0 1
1 0 1
B1 0C B1C
B C B C
B C B C
B0 1C B0C
SDB C; WDB C
B0 1C B0C
B C B C
@0 1A @0A
0 1 0
we can compute the MVPs

0 1
a11 a16
Ba a C a C a C
B 21 23 24 25 C
B C
B a31 a33 C
AS D B C; W > A D a11 C a21 0 a23 a24 a25 a16
B a42 a44 C
B C
@ a51 a55 A
a62 a66
such that all the nonzero entries are determined uniquely. A lower bound on the
number of MVPs for one-sided compressions is given by the easily evaluated
expression minfmax .A/; max .A> /g which has been found to be a good approxima-
tion to the chromatic number of graphs associated with the matrices on an extensive
set of test instances [1]. On the other hand, only a handful of published literature
consider two-sided compressions. The Ph.D. dissertation [9] derived the bound
nnz.A/
LB2S D d maxfm;ng e on the number of MVPs in a two-sided compression which
is also independently derived and analyzed experimentally in [11]. In this paper we
generalize this lower bound and show that the new lower bound is always at least as
good and in many test instances it yields better result.
2.1 A New Lower Bound
Following [9] the bound LB2S can be derived intuitively in the following way.
Consider a two-sided compression S 2 f0; 1gnp ; W 2 f0; 1gmq . With an MVP
of type As; s 2 f0; 1gn , at most m nonzero entries can be uniquely determined.
Similarly, with an MVP of type w> A; w 2 f0; 1gm, at most n nonzero entries can
be uniquely determined. If there are nnz.A/ nonzero entries in matrix A, then its
follows that

nnz.A/
nnz.A/ pm C qn maxfm; ng. p C q/ implying . p C q/:
maxfm; ng
In one-sided compression a row with the maximum number of nonzero determines a

lower bound for the number of MVPs. In graph-theoretic terminology this represents
a subgraph where each pair of vertices are connected by an edge. Our proposal is to
extend this key observation to the two-sided scenario. In other words, we claim that
a dense submatrix of a given matrix A yields a lower bound on the number of MVPs
to uniquely determine matrix A. Clearly, a bound of the type LB2S on a submatrix
0 0
of A is also a bound on the whole matrix A. Now consider a submatrix A0 2 <m n
sufficiently dense such that neither the rows nor the columns of the submatrix are
structurally orthogonal. A trivial example is a completely dense submatrix. Then, we
need at least minfm0 ; n0 g MVPs to uniquely determine matrix A. In graph-theoretic
terminology, this completely dense submatrix corresponds to a complete bipartite
subgraph Km0 ;n0 of the bipartite graph associated with matrix A. We want to find the
largest dense submatrix of the kind described above. Unfortunately, the complexity
of deciding whether or not a bipartite graph contains a complete bipartite subgraph
of size k is NP-complete [4]. Instead of trying to find the largest dense submatrix
of matrix A we therefore settle for a reasonably dense submatrix which is easy
to compute. The above discussion can be succinctly stated as: find permutation
matrices P and Q such that the largest dense submatrix of A is included in some
leading submatrix of the permuted matrix PAQ. The k-th leading submatrix of
matrix A is denoted by A.1 W k; 1 W k/. To move dense rows and columns toward
the top left corner of matrix A we use the following simple procedure.
Order the rows in non increasing order of the number of nonzero entries followed by the
same ordering procedure applied to the columns. Matrices P and Q represent the row and
column ordering, respectively.
We then apply the formula LB2S on each leading submatrix of permuted matrix
PAQ. The largest LB2S value is our lower bound LB2SX on the number of MVPs to
determine matrix A. Thus,
LB2SX D maxfLB2Si g
i
and

nnzi .PAQ/
LB2Si D ;
i
where nnzi .:/; i D 1; : : : ; minfm; ng denotes the number of nonzero entries in the
i-th leading submatrix of the argument.
The above discussion can be summarized in the following lemma.
Lemma 1 A lower bound on the number of matrix-vector products to directly
determine A 2 <mn is given by
LB2SX D maxfLB2Si g
i
with

nnzi .PAQ/
LB2Si D ;
i
where nnzi .:/; i D 1; : : : ; minfm; ng denotes the number of nonzero entries in the
i-th leading submatrix of the argument and P and Q are permutation matrices. If
m D n, then we have

nnz.A/
LB2SX LB2S D :
maxfm; ng
2.2 A Heuristic for Two-Sided Compression
The heuristic algorithm that we outline here is inspired by the Recursive Largest
First (RLF) coloring heuristic due to [12]. The main idea of our compression
algorithm is to group columns and rows such that for each nonzero aij ,
1. there is a column group k containing column j such that there is no column j0 ¤ j
also in group k for which aij0 ¤ 0, or
2. there is a row group l containing row i such that there is no row i0 ¤ i also in
group l for which ai0 j ¤ 0
This condition is termed as direct cover condition in [8] and the nonzero aij meeting
this is said to be directly covered or simply covered, for brevity. The complete
direct cover (CDC) algorithm of [9] groups the columns and rows while respecting
the cover condition until all the nonzero entries are covered. The resulting column
groups and row groups define the matrices S and W such that the Jacobian matrix is
directly determined. The columns and rows are assigned to groups according to the
number of nonzero entries that are yet to be covered: higher the number of entries
earlier they are grouped. In our algorithm 2SIDEDCOMPRESSION, as depicted in
Fig. 3, we too take into account the number of entries that are yet to be covered in
forming the groups. First, we introduce terminology that will enable us to describe
the algorithm. Given a sparse matrix A let I D f1; : : : ; mg and J D f1; : : : ; ng denote
the set of row indices and the set of column indices, respectively. For row index i
and I I define,
˚
fbdnI .i/ D i0 ji0 2 I and there is an index j 2 J for which aij ¤ 0;

ai0 j ¤ 0 are not covered :
Analogously, for column index j and J J define,

˚
fbdnJ . j/ D j0 jj0 2 J and there is an
index i 2 I for which aij ¤ 0;
aij0 ¤ 0 are not covered :
ˇ ˇ
Also, let degI .i/ D jfbdnI .i/j and degJ . j/ D ˇfbdnJ . j/ˇ denote the cardinality
of the respective index sets.
Thus, degI .i/ denotes the number of rows corresponding to set I that cannot
be grouped with row i. Similarly, degJ . j/ denotes the number of columns
corresponding to set J that cannot be grouped with column j.
On return from 2 SIDEDCOMPRESSION each column is given a label from the set
f0; 1; : : : pg indicating the column group it is included in and each row is given label
from the set f0; 1; : : : qg indicating the row group it is included in. A column (or row)
that is labeled with group number 0 indicates that its nonzero entries are directly
covered by other group(s). Variable ncovered represents the number of nonzero
entries that have been directly covered. Initially, each column/row is labeled with
group number 0. The while -loop in line 5 is iterated until all the nonzero entries of
matrix A are directly covered. Variables Iu and Ju contain the indices of rows and
columns, respectively, that are yet to be included in a group other than group 0. The
pseudocode in lines 7–16 and lines 17–25 compute the next row group and column
group, respectively. The if -construct in lines 26–34 chooses the group which covers
the most nonzero entries (Fig. 3).
In this section we provide numerical evidence to demonstrate the effectiveness of the

methods developed in Sect. 2. For numerical experiments we choose the set of test
instances used in ([11] Table 2). Table 1 compares the lower bound on the number
of MVPs given by LB2SX and LB2S. In Tables 1 and 2, max .:/ represents the
maximum number of nonzero entries in any row of the matrix given in the argument.
Table 1 includes the test matrices of ([11] Table 2) where the two bounds differ.
Out of 32 instances, LB2SX is strictly better on 14 of them; on the remaining 16
instances LB2SX and LB2S yield the same value. On problem eris1176 algorithm
Fig. 3 An algorithm for two-sided compression of a sparse Jacobian matrix
Table 1 Lower bound on the number of MVPs in two-sided compression of sparse Jacobian
Matrix m n max .A/ max .A> / LB2S LB2SX
abb313 313 176 6 26 5 6
arc130 130 130 2 12 10 16
bp0 822 822 266 20 4 7
bp200 822 822 283 21 5 7
bp400 822 822 295 21 5 7
bp800 822 822 304 21 6 8
bp1000 822 822 308 21 6 8
eris1176 1176 1176 99 99 16 80
lund_a 147 147 21 21 17 18
lund_b 147 147 21 21 17 18
str_0 363 363 34 34 7 17
str_200 363 363 30 26 9 17
str_400 363 363 33 34 9 18
will_57 57 57 11 11 5 6
Table 2 Number of MVPs in two-sided compression heuristics

Matrix m n max .A/ max .A> / ASBC 2SIDED COMPRESSION
abb313 313 176 6 26 17 10
ash219 219 85 2 9 8 4
ash292 292 292 14 14 19 14
ash331 331 104 2 12 10 6
ash608 608 188 2 12 11 6
ash958 958 292 2 13 12 6
bp0 822 822 266 20 16 17
bp200 822 822 283 21 17 19
bp400 822 822 295 21 19 20
curtis54 54 54 12 16 12 11
eris1176 1176 1176 99 99 92 80
fs_541_1 541 541 11 541 18 13
fs_541_2 541 541 11 541 18 13
ibm32 32 32 8 7 9 8
lund_a 147 147 21 21 26 24
lund_b 147 147 21 21 27 23
shl_0 663 663 422 4 7 4
shl_200 663 663 440 4 7 4
shl_400 663 663 426 4 6 4
str_0 363 363 34 34 25 27
str_200 363 363 30 26 31 32
will199 199 199 6 9 9 7
2SIDEDCOMPRESSION is optimal (see Table 2) since the number of MVPs needed

to directly determine the matrix is same as the lower bound given by LB2SX. The
matrix is pattern symmetric and will require at least 99 MVPs to determine it by any
one-sided compression method. With a two-sided compression, however, at least
80 MVPs will be needed to determine the matrix. The LB2S gives a lower bound of
16 MVPs which is much lower than the optimum number of MVPs (D 80) required
to directly determine the matrix. Table 2 compares the number of MVPs needed by
algorithm ASBC and our heuristic algorithm 2SIDEDCOMPRESSION . The instances
are chosen from ([11] Table 2) where the number of MVPs required by the two
methods are different. Out of 22 instances, algorithm 2SIDEDCOMPRESSION yields
better result than ASBC on 17 (marked in bold face) of them. For the instances
where ASBC is better, the difference is at most 2.
4 Conclusion
In this paper we have considered the problem of sparse Jacobian matrix determina-
tion with two-sided compressions. The new lower bound on the number of MVPs
generalizes the lower bound of [9, 11]. The results from numerical experiments on
a standard set of test instances provide strong evidence that the proposed lower
bound is superior to that of [9, 11]. The new lower bound is easy to compute,
never smaller than the bound proposed in [9, 11], and strictly larger on about
50 % of the test instances. The Heuristic algorithm 2 SIDEDCOMPRESSION saves,
on average, 13 % MVPs compared with the ASBC algorithm and, on average it is
within approximately 1:5 times the optimal.
There are a number of possible extensions to this research that we plan to pursue
in future. In our lower bound calculation we use the simple strategy of reordering of
rows and columns according to the nonzero density. More sophisticated ordering
techniques can be considered here. For example, lining up the nonzero entries
together in rows and columns may reveal the dense submatrices which can be
identified more easily. This will yield the added benefit of utilizing data locality
to improve cache memory performance in a computer implementation. Also, the
current computer implementation of the compression algorithm can be improved by
employing more advanced data structures for manipulating sparse matrices.
Acknowledgements This research is supported in part by Natural Sciences and Engineering

Research Council of Canada (NSERC) Discovery Grant (Individual).
References
1. Coleman, T.F., Moré, J.J.: Estimation of sparse Jacobian matrices and graph coloring problems.
SIAM J. Numer. Anal. 20(1), 187–209 (1983)
2. Coleman, T.F., Verma, A.: The efficient computation of sparse Jacobian matrices using
automatic differentiation. SIAM J. Sci. Comput. 19(4), 1210–1233 (1998)
3. Curtis, A.R., Powell, M.J.D., Reid, J.K.: On the estimation of sparse Jacobian matrices. IMA
J. Appl. Math. 13(1), 117–119 (1974)
4. Garey, M.R., Johnson, D.S.: Computers and Intractability: A Guide to the Theory of NP-
Completeness. Freeman, San Francisco (1979)
5. Golub, G.H., Van Loan, C.F.: Matrix Computations, 3rd edn. Johns Hopkins University Press,
Baltimore (1996)
6. Griewank, A., Toint, Ph.L.: On the unconstrained optimization of partially separable objective
functions. In: Powell, M.J.D. (ed.), Nonlinear Optimization, pp. 301–312. Academic, London
(1982)
7. Griewank, A., Walther, A.: Evaluating Derivatives: Principles and Techniques of Algorithmic
Differentiation, 2nd edn. Society for Industrial and Applied Mathematics, Philadelphia (2008)
8. Hossain, A.K.M.S.: On the computation of sparse Jacobian matrices and newton steps. Ph.D.
Dissertation, Department of Informatics, University of Bergen (1998)
9. Hossain, A.K.M.S., Steihaug, T.: Computing a sparse Jacobian matrix by rows and columns.
Optim. Methods Softw. 10, 33–48 (1998)
10. Hossain, S., Steihaug, T.: Graph models and their efficient implementation for sparse Jacobian
matrix determination. Discret. Appl. Math. 161(2), 1747–1754 (2013)
11. Juedes, D., Jones, J.: Coloring Jacobians revisited: a new algorithm for star and acyclic
bicoloring. Optim. Methods Softw. 27, 295–309 (2012)
12. Leighton, F.T.: A graph coloring algorithm for large scheduling problems. J. Res. Natl. Bur.
Stand. 84, 489–505 (1979)
An h-Adaptive Implementation
of the Discontinuous Galerkin Method
for Nonlinear Hyperbolic Conservation Laws
on Unstructured Meshes for Graphics
Processing Units
Andrew Giuliani and Lilia Krivodonova
Abstract For computationally difficult problems, mesh adaptivity becomes a

necessity in order to efficiently use computing resources and resolve fine solution
features. The discontinuous Galerkin (DG) method for hyperbolic conservation laws
is a numerical method adapted to execution on graphics processing units (GPUs).
In this work, we give the framework of an efficient h-adaptive implementation of
the modal DG method on NVIDIA GPUs, outlining implementation considerations
in the context of GPU computing. Finally, we demonstrate the effectiveness of our
implementation with a computed example.
1 Introduction
A relatively popular approach to numerical simulation is mesh adaptivity. Notably,

in general it is not possible to know a priori how to optimally distribute on a
domain the numerical degrees of freedom (DOFs) e.g. the number of elements and
local order of approximation, for finite element methods. This is especially true
for time-dependent simulations. For many difficult problems in computational fluid
dynamics (CFD), computing resources will quickly be expended if DOFs are not
added satisfactorily. Fortunately, mesh adaptation strategies exist that automatically
add or remove DOFs solely in areas of the solution where they are needed [12].
One approach is spatial refinement, or h-adaptivity, which attempts to selectively
subdivide or merge mesh cells thereby locally increasing or decreasing solution
resolution. H-adaptivity can be implemented in varying fashions; we discuss three
possible ways: component grids, and block or cell-based refinement.
Berger considered component grids in detail on structured regular meshes [1, 2].
This method achieves h-adaptivity through a hierarchy of independent meshes that
A. Giuliani () • L. Krivodonova

University of Waterloo, 200 University Avenue West, Waterloo, ON N2L 3G1, Canada
e-mail: [email protected],[email protected]

436 A. Giuliani and L. Krivodonova
are overlay one another. An error indicator first flags elements that do not satisfy
some criterion. These elements are then clustered into groups based on spatial
proximity and other criteria. Then, additional rotated rectangular grids with finer
mesh spacing are created and overlay these groups. This is done multiple times
until an error tolerance is reached, creating a series of embedded grids After an
appropriate number of timesteps, the mesh hierarchy is redetermined through a
process called ‘regridding’. Since Berger [1, 2], a simplified approach has been
applied to GPUs in [3, 10] for the shallow water equations. Other approaches include
block and cell-based refinement strategies.
In block-based refinement, cells are grouped into blocks of a predefined size. If
any elements in a group are flagged for refinement, then all elements in that group
are refined and replaced by a number of refined blocks. This allows connectivity
information to be stored with respect to blocks in a quadtree data structure for 2D
and octree for 3D [9, 11]. Block-based implementations have been discussed in [14].
In cell-based refinement, elements are refined individually rather than in groups
which maintains high-resolution locality. This strategy can be implemented both on
structured and unstructured meshes, though many cell-based approaches to date deal
with Cartesian grids [7]. A trade off of refinement locality is that more connectivity
information is required than in block-based refinement. This is because data must
be stored on the level of elements, rather than blocks. A quadtree and octree data
structure can be used again to represent parent-child relationships for determining
mesh connectivity [5, 13]. The advantage of this strategy is that irregular boundaries
can be dealt with very easily through cut-cells, higher order elements or curved
boundary conditions.
GPUs are useful tools in scientific computing applications. The compute capacity
of NVIDIA GPUs is quite impressive; single and double precision arithmetic
throughput can exceed 1 TFLOP/s depending on the model. When this is considered
in conjunction with the price of such devices, GPUs are an attractive scientific
computing platform.
We now focus on the main subject of this work: an efficient cell-based h-
adaptive DG-GPU algorithm in NVIDIA’s CUDA C on unstructured triangular
meshes for nonlinear hyperbolic conservation laws. We begin with the derivation
of a DG method for these partial differential equations (PDEs). We then outline
the implementation details and present a computed example, examining the solver’s
performance characteristics.
2 The Discontinuous Galerkin Method
In two dimensions, hyperbolic conservation laws are PDEs of the form

d
u C rxy F.u/ D 0; (1)
dt
with the solution u.x; t/ D .u1 ; u2 ; : : : ; uM /| defined on ˝ Œ0; T such that x D
.x; y/, x 2 ˝ R2 , T is a final time and F.u/ D .F1 .u/; F2 .u// is the flux function.
h-Adaptivity for the DG Method on Unstructured Meshes for GPUs 437
Additionally, the initial condition
u.x; 0/ D u0 .x/;
along with appropriate boundary conditions are applied. The DG method is a high-
order method without an extensive stencil that can successfully capture shocks
and discontinuities in the numerical solution of conservation laws. As opposed to
the standard finite element method, continuity at the interface between adjacent
elements is not imposed.
This method can be formulated by first dividing
S the domain ˝ into an unstruc-
tured mesh of N triangles such that ˝ D NiD1 ˝i . We note that nonconforming
tessellations are permitted. In our implementation, the difference in refinement level
between adjacent cells must not exceed 1, see Fig. 1.
We define Sp .˝i / to be the space of polynomials of degree at most p on ˝i .
Additionally, fevj gjD1::Np is a set of orthonormal basis functions for Sp .˝i /, where Np
represents the number of basis functions.
The weak form of the conservation law is obtained by multiplying Eq. (1) by a
test function evj 2 Sp .˝i / and integrating on an element ˝i . After integrating by
parts, we obtain
Z Z Z
ut e
vj d˝i F.u/rxy e
vj d˝i C vj F.u/ ndl D 0;
e (2)
˝i ˝i @˝i
where n is the outward facing normal on @˝i .

Each element ˝i is mapped to the canonical triangle ˝c , having vertices at
.0; 0/; .1; 0/; .0; 1/. We use the transformation
0 1 0 10 1
x xi;1 xi;2 xi;3 1rs
@yA D @yi;1 yi;2 yi;3 A @ r A; (3)
1 1 1 1 s
e7 e7
e13 e3
Ω2 Ω4 Ω2
e1
e10 e8 e2 e2
Ω5 e4
Ω1
e6 e12 e11
Ω0 Ω3
e5 e0 e9
(x2,1 , y2,1 )
Fig. 1 Examples of admissible meshes: (left) two element mesh (right) ˝5 is refined yielding a
five element mesh
where .xi;z ; yi;z /zD1;2;3 are the three original vertices of element ˝i in physical space
and .r; s/ are the transformed coordinates in the computational space. The Jacobian
of the transformation is denoted by Ji and we require that det Ji be positive, i.e.
that the vertices be ordered counter-clockwise. Vertex ordering also defines the
side numbers q D 1; 2 : : : Nisides of each cell, where Nisides is the number of sides
an element has; for example si;q refers to the qth side of ˝i beginning clockwise
from .xi;1 ; yi;1 /. In the refined mesh of Fig. 1, s2;4 is edge e11 with the indicated
.x2;1 ; y2;1 /.
Additionally, we map each edge of the mesh ek , i.e. si;q , to the canonical interval
Ic D Œ1; 1 with the transformation
1
x x x .1 /
D k;1 k;2 2
1 ; (4)
y yk;1 yk;2 2 .1 C /
where is the transformed coordinate in the computational space and .xk;z ; yk;z /zD1;2
are the two original vertices of ek . The determinant of the Jacobian of (4) is
q
1
li;q D .xk;1 xk;2 /2 C .yk;1 yk;2 /2 :
2
As for other Galerkin methods, the solution on element ˝i is approximated by
PNp
a linear combination of the basis functions e vj , i.e. Ui D jD1 ci;j v
ej with ci;j D
1 2 M |
Œci;j ; ci;j ; : : : ; ci;j ; : : : ; ci;j , the modal degrees of freedom. Finally, we separate the
m
second term in (2) into line integrals along the edges of the element. It follows that
equation (2) becomes
Z
d 1
ci;j D F.Ui / .Ji1 rrs vj / det Ji d˝c
dt det Ji ˝c
XZ
Nisides
vj;q F.Ui / ni;q li;q d ; (5)
qD1 I0
where vj and vj;q are values of the basis functions mapped from the physical
element ˝i to the canonical element ˝c and from side si;q to the canonical side
I0 , respectively. Finally, ni;q is the outward facing normal of si;q . The right-hand
side of equation (5) is composed of two terms: a volume and surface integral.
As continuity between elements is not imposed, the solution Ui is multivalued in
the surface integral. We therefore introduce a numerical flux F.Ui ; Uk / to allow
information to be exchanged between adjacent cells ˝i and ˝k . We can evaluate
these terms exactly for a linear flux and approximately for a nonlinear flux with
numerical quadrature. The equation now becomes
Z
d 1
ci;j D F.Ui / .Ji1 rrs vj / det Ji d˝c
dt det Ji ˝c
XZ
Nisides
vj;q F.Ui ; Ukq / ni;q li;q d : (6)
qD1 Ic
Because equation (6) is a system of ordinary differential equations (ODEs) for the
degrees of freedom ci;j , it can be solved in time with a standard ODE solver, e.g.
Runge-Kutta (RK) method.
As the volume and surface integral contributions can be computed independently
of one another, the method is predisposed to applications on highly parallel
GPU architectures [6]. Developing algorithms for this computing platform is an
active area of research. An efficient h-adaptive, block-based structured mesh DG-
GPU implementation has been presented in [3, 10]. Cell-based h-adaptivity on
unstructured meshes in the context of GPUs can be more difficult due to the
inevitable irregularity of memory accesses and operations both in time stepping and
mesh adaptivity operations. We will now describe the implementation details of our
efficient h-adaptive DG-GPU algorithm in NVIDIA’s CUDA C; this implementation
consists of a time stepping and mesh adaptation module.
3 Time Stepping Details
With the time stepping module, we calculate the right-hand side of (6), without mesh
modification. Every n regular timesteps, the adaptive module is executed which
refines and coarsens select elements. The time stepping aspects of the solver are
outlined in more detail in [6]. This module consists of three compute kernels that
efficiently evaluate the right-hand side of (6) for a standard RK ODE time integrator.
The first kernel, eval_volume, launches one thread per element ˝i ; each thread
evaluates the volume integral terms for its allocated element. Likewise, the second
kernel, eval_surface, launches one thread per edge ek ; each thread evaluates
the surface integral terms for its allocated edge. Finally, eval_rhs launches one
thread per element and appropriately sums the volume and surface terms over the
mesh. An essential aspect of these final two kernels is the use of connectivity data
on the GPU.
DOF storage In order fully utilize the compute capacity of the GPU, the modal
DOFs must be stored in a specific fashion. See [6] for further information.
Connectivity We store connectivity information in GPU DRAM memory. We store
simple pointers in linear arrays as integers, which encode the element and edge IDs
˝i and ek , respectively. The bidirectional pointers are displayed in Fig. 2. Elements
e0
Ω0 e1
e7
e2
e13 e3 Ω1 e3
Ω4 Ω2 e4
e1
e2 Ω2 e6 Boundary
e7
Ω1
e4 e6 e12 e11 Ω3 e9
Ω0 Ω3 e11
e0 e9
Ω4 e12
e13
Fig. 2 (Left) nonconforming mesh, (right) connectivity data
point to their edges and edges point to their two elements; note that ghost cells
enforce boundary conditions. Due to the nonconforming nature of our meshes, each
element requires Nisides pointers to its respective edges, see ˝2 in Fig. 2 (N2sides D
4 ¤ 3).
4 Mesh Adaptation
The mesh adaptation module of our implementation works to optimally distribute

the spatial DOFs on the computational domain. We note that work is completed
directly in GPU video memory with minimal memory transfer between host and
device.
Tree structures A tree structure for both elements and edges is implemented,
keeping track of parent-child relationships. These data structures, along with other
information, allow us to redetermine the mesh connectivity described in Sect. 3 after
elements and edges are refined and coarsened. In the tree, each element and edge
point to their children; likewise, each element and edge also point to their parents.
In Fig. 1, element ˝5 is refined into four new elements. We denote ˝5 as the
parent of the four child elements ˝3 , ˝0 , ˝4 and ˝1 . Similar terminology is used
for the parent edges: e5 , e8 and e10 and their children. The bidirectional pointers
that would be stored in GPU DRAM for this example are presented in Fig. 3. We
highlight the elements and edges in the trees without children, corresponding to
those components which are active in the current mesh.
Error indicator In order to determine which elements need to be refined and
which need to be coarsened, an error indicator is used. In the literature, there
are a substantial number of indicators available. The number of times an element
e0
e1
Ω0
e2 e3
Ω1 e5 e4
Ω5
e7 e9
Ω3
e8 e11
Ω2
Ω4 e10 e12
e13
Fig. 3 Tree data stored for the two sequence of meshes in Fig. 1. Elements and edges highlighted
in blue are active in the current mesh
1 Ωi 0
0
-1
-1 -1
-1
Fig. 4 (Left) initial configuration around ˝i along with the values of aflag for each element,
(right) mesh after refinement and smoothing; dotted line indicates the elements added
is refined and coarsened as determined from the error indicator is stored as an

integer in an array aflag. A positive integer indicates refinement, whereas a
negative integer indicates coarsening. Our implementation allows a maximum level
of refinement of lmax D 5.
Flag smoothing Flag smoothing is an operation which enforces that the flags in
aflag do not create elements that have an interelement difference in refinement
level greater than the maximum allowed, i.e. jli lj j 1 where li and lj are
the refinement levels of the adjacent elements ˝i and ˝j , respectively. We must
therefore modify aflag so that after the refinement and coarsening operation, an
incompatible mesh is not created.
For example, in Fig. 4 on the left we have the mesh configuration around an
arbitrary element ˝i with value of aflag on each element. The two elements being
refined once will induce an interelement jump in refinement level of 2 with ˝i . The
smoothing operation will supress this jump by refining element ˝i once as shown
in the mesh on the right.
Refinement and coarsening After aflag is smoothed, the elements and edges
may now be refined and coarsened. We implement refinement and coarsening
kernels which modify the element and edge trees. Solution coefficients are rede-
termined using an L2 projection and geometric data is adapted accordingly. The
final step in the mesh adaptivity algorithm is redetermining the mesh connectivity
detailed in Sect. 3. This is done in part with the element and edge trees presented in
Sect. 4.
Memory management In refining and coarsening the mesh, we will naturally have
to add or remove data from lists, e.g. the modal degrees of freedom C. A naive
solution would be to simply replace the undesired element by a null placeholder.
Unfortunately, this would lead to ‘holes’ in our arrays which is not an efficient use
of memory and further would limit the effectiveness of coalesced memory accesses.
We must therefore not only remove elements from arrays but also shift those that
remain so that the reduced array is dense.
On GPUs, this problem does not have a simple solution. Indeed, parallelizing
this operation, called a ‘reduction’, while avoiding race conditions must be done
with great care. There are application programming interfaces (APIs) that have this
task optimized; we use CUDPP in our implementation [4]. The details of a reduction
are not presented here for brevity, an involved examination of this issue can be found
in [8].
5 Computed Example
We now present an example demonstrating the effectiveness of this implementation.

Consider the two dimensional advection equation
d
u C rxy Œ2yu; 2xu D 0; (7)
dt
on the domain ˝ D Œ1; 12 , subject to the initial condition

.x 0:2/2 y2
u0 .x; y/ D 5 exp : (8)
2 0:152
This is called the rotating hill test problem, whereby a Gaussian pulse is advected
around the origin; the exact solution is expressed as
u.x; y; t/ D u0 .x cos.2t/ C y sin.2t/; x sin.2t/ C y cos.2t//: (9)
This problem would benefit from h-adaptivity as much of the variation in the
solution is spatially concentrated in a small, moving region of the domain.
We apply our adaptive code to this problem and advance the solution until a final
time of T D 1, i.e. one full rotation around the origin. The DG spatial discretization
uses the upwind numerical flux and is integrated in time with an RK time integrator
of appropriate order. The adaptive routines are executed every 100 regular timesteps,
allowing an element to be refined a maximum of 3 times. The initial mesh was
composed of 1080 elements; if uniformly refined three times, it would comprise
69,120 elements. The numerical solution and final adaptive mesh are plotted in
Fig. 5 for p D 1. In Table 1, we present the computed L2 error of the final solution
for both the adaptive and uniformly refined meshes. The advantage of h-adaptivity is
apparent as only a fraction of the computational work is required to achieve an error
comparable to that of a uniformly refined mesh. In all simulations, the refinement
and coarsening subroutines comprise at most
2 % of the total runtime.
Fig. 5 Final refined mesh at T D 1 with the solution isolines for p D 1

Table 1 L2 error for p D 1; 2; 3 of the h-adaptive solution, and uniformly refined solution
p Average number of elements Adaptive mesh L2 error Uniformly refined mesh L2 error
1 18,195 4.684E-4 4.453E-4
2 24,906 4.373E-6 4.226E-6
3 34,785 4.514E-8 4.274E-8
6 Conclusion
In conclusion, we have presented the framework of an efficient h-adaptive imple-

mentation of the DG method for hyperbolic conservation laws on GPUs. We
plan on implementing a combined hp-adaptive solver, and extending support to 3
dimensional unstructured tetrahedral meshes with local time stepping.
Acknowledgements This research was supported in part by the Natural Sciences and Engineering
Research Council of Canada (NSERC) grant 341373-07 and the NSERC CGS-M grant.
References
1. Berger, M.: Adaptive mesh refinement for hyperbolic partial differential equations. Ph.D.
thesis, Stanford University (1982)
2. Berger, M.J., Oliger, J.: Adaptive mesh refinement for hyperbolic partial differential equations.
J. Comput. Phys. 53(3), 484–512 (1984)
3. Brodtkorb, A.R., Sætra, M.L., Altinakar, M.: Efficient shallow water simulations on GPUs:
implementation, visualization, verification, and validation. Comput. Fluids 55, 1–12 (2012)
4. CUDA Data Parallel Primitives Library. http://cudpp.github.io/. CUDPP 2.2
5. Flaherty, J.E., Loy, R.M., Shephard, M.S., Szymanski, B.K., Teresco, J.D., Ziantz, L.H.:
Adaptive local refinement with octree load balancing for the parallel solution of three-
dimensional conservation laws. J. Parallel Distrib. Comput. 47(2), 139–152 (1997)
6. Fuhry, M., Giuliani, A., Krivodonova, L.: Discontinuous Galerkin methods on graphics
processing units for nonlinear hyperbolic conservation laws. Int. J. Numer. Methods Fluids
76(12), 982–1003 (2014)
7. Gao, X.: A parallel solution-adaptive method for turbulent non-premixed combusting flows.
Ph.D. thesis, University of Toronto (2008)
8. Harris, M., Sengupta, S., Owens, J.D.: Parallel prefix sum (scan) with CUDA. In: Nguyen, H.,
NVIDIA Corporation (eds.) GPU Gems 3. Addison-Wesley, Upper Saddle River (2008)
9. Ivan, L., Sterck, H.D., Northrup, S.A., and Groth, C.: Multi-dimensional finite-volume scheme
for hyperbolic conservation laws on three-dimensional solution-adaptive cubed-sphere grids.
J. Comput. Phys. 255(0), 205–227 (2013)
10. Sætra, M.L., Brodtkorb, A.R., Lie, K.-A.: Efficient GPU-implementation of adaptive
mesh refinement for the shallow-water equations. J. Sci. Comput. 63(1), 23–48 (2014).
doi:10.1007/s10915-014-9883-4. ISSN:1573-7691
11. Susanto, A.: High-order finite-volume schemes for magnetohydrodynamics. Ph.D. thesis,
University of Waterloo (2014)
12. Wang, Z.: Adaptive High-order Methods in Computational Fluid Dynamics. Advances in
Computational Fluid Dynamics. World Scientific, Singapore/Hackensack (2011)
13. Zeeuw, D.L.D.: A quadtree-based adaptively-refined cartesian-grid algorithm for solution of
the Euler equations. Ph.D. thesis, The University of Michigan, Ann Arbor (1993)
14. Zhang, J.Z.: Parallel anisotropic block-based adaptive mesh refinement finite-volume scheme.
Ph.D. thesis, University of Toronto (2011)
Extending BACOLI to Solve the Monodomain
Model
Elham Mirshekari and Raymond J. Spiteri
Abstract BACOLI is a numerical software package that solves systems of

parabolic partial differential equations (PDEs) in one spatial dimension. It is based
on high-order B-spline collocation and features adaptivity in space and time. The
monodomain model of cardiac electrophysiology is a multi-scale model that couples
electrical activity in myocardial tissue at the tissue scale with that at the cellular
scale. This leads to a (parabolic) reaction-diffusion PDE coupled with a set of
nonlinear (non-parabolic) PDEs that do not involve spatial derivatives. In this paper,
we extend BACOLI to solve this more general class of problem, of which the
monodomain model is one example. We demonstrate that the extended BACOLI
software package outperforms the Chaste software package, which is a powerful,
widely used, and well-respected software package for heart simulation, in terms
of execution time on the monodomain equation with two cell models of varying
stiffness.
1 Introduction
BACOLI is a numerical software package that solves systems of parabolic partial

differential equations (PDEs) of the form
ut .x; t/ D f .t; x; u.x; t/; ux .x; t/; uxx .x; t//; a x b; t0 t tf ; (1a)
where u.x; t/ 2 RMu is the unknown vector function, subject to separated boundary
conditions of the form
Ba .t; u.a; t/; ux .a; t// D 0; Bb .t; u.b; t/; ux .b; t// D 0; t0 t tf ; (1b)
E. Mirshekari
Department of Mathematics and Statistics, University of Saskatchewan, Saskatoon, SK, Canada
R.J. Spiteri ()
Department of Computer Science, University of Saskatchewan, Saskatoon, SK, Canada

448 E. Mirshekari and R.J. Spiteri
and the initial condition
u.x; t0 / D u0 .x/; a x b: (1c)
BACOLI is one of only a few software packages to feature adaptive error control
in both space and time [18]. The software package is designed to produce an
approximate solution such that the error is estimated to be less than a user-supplied
tolerance [9]. Adaptive error control increases both the efficiency as well as the
reliability of the simulation.
The monodomain model is a multi-scale mathematical model for the evolution
of the electrical potential in myocardial tissue that relates the propagation of ionic
currents at the tissue scale with their generation at the cellular scale. Because of
their global nature, the mathematical models of the propagation at the tissue level
depend on spatial derivatives. On the other hand, because of their local nature, the
models of a single cell do not.
The monodomain model can be written as
@v
Cm C Iion .s; v; t/ D r .i rv/;
@t 1C
(2a)
@s
D f .s; v; t/;
@t
with boundary condition
O i rv/ D 0;
n:. (2b)
where v is the transmembrane potential, s is a (model-dependent) vector of cellular

states such as gating variables and ionic concentrations, Iion is the ionic current
across the membrane per unit cell membrane area, is the area of cell membrane
per unit volume, i is the conductivity, Cm is the capacitance of the cell membrane
per unit area, and nO is the unit normal vector to the boundary [14]. The system (2)
is supplemented by appropriate initial conditions.
The cell models considered in this paper are those of Luo–Rudy [10] (considered
to be mildly stiff) and the epicardial variant of ten Tusscher and Panfilov [15]
(considered to be stiff). The cell models do not have any dependency on the
spatial derivatives and thus cannot be solved directly using BACOLI. We extend
BACOLI so that it can solve PDEs of the form (2a).
The remainder of this paper unfolds as follows. A review of the BACOLI
software package is provided in Sect. 2. Some basic information on the of the heart
is provided in Sect. 3. The performance in terms of execution time of the extended
version of BACOLI is demonstrated on the monodomain model with two different
cell models in Sect. 4.
Extending BACOLI to Solve the Monodomain Model 449
2 Review of BACOLI
This section describes some basic concepts behind the BACOLI software package,
including its structure and the modifications to extend it to solve multi-scale systems
like the monodomain model.
2.1 Basic Concepts
BACOLI uses the following discretization strategy. Consider the interval Œa; b
divided into Nx subintervals Œxi1 ; xi , where i D 1; 2; : : : ; Nx ; with a partition
of the form
W a D x0 < x1 < : : : < xNx D b:
On subinterval i, i D 1; 2; : : : ; Nx , the unknown function is approximated by a

local polynomial. The spline s.x/ obtained from joining the local polynomials is
a piecewise polynomial with certain properties; e.g., it can have several continuous
derivatives [6].
The standard B-representation of a spline is as a linear combination of B-splines.
A B-spline is piecewise polynomial that has minimal support with respect to its
degree, continuity, and knot sequence. For the knot sequence X1 X2 : : :
XDCMC1 , we have
X
D
s.x/ D Bj;M .x/aj ;
jD1
where Bj;M .x/ is B-spline j of degree M, D is the number of the B-spline basis
functions, and the coefficients aj are the unknown coefficients that are defined
according to the choice of method, e.g., collocation . There are different types of
knots mentioned in [7]. The type that is considered in BACOLI is open uniform,
defined as
8
ˆ
<Xk D X1 ;
ˆ if k M C 1;
X Xk D constant; M C 1 k < D C 1;
ˆ kC1
:̂X D X ; k D C 1;
k DCMC1
where k D 1; 2; : : : ; D C M C 1; i.e., at each boundary, the multiplicity of the knots

is .M C 1/.
With the choice of open uniform knots, the associated B-splines have two
properties at the left boundary:
B1;M .x1 / D 1; and B01;M .x1 / D B02;M .x1 /:
Similar properties hold at the right boundary.
2.2 Structure of the BACOLI Software Package
The exact solution of (1), u.x; t/, is approximated by a linear combination of these
B-spline functions as
X
D
U.x; t/ D Bj;M .x/yj .t/;
jD1
where yj .t/ is the vector of unknown coefficients of the B-spline j. In order to

determine yj .t/, the BACOLI software package requires U.x; t/ to satisfy (1a) at
each of the M 1 Gaussian collocation points of each subinterval [18].
The collocation equations defining fyk .t/gD
kD1 are of the form
d
U.l ; t/ D f .t; l ; U.l ; t/; Ux .l ; t/; U xx .l ; t//; (3)
dt
where l D 1 C .i 1/.M 1/ C j; i D 1; 2; : : : ; Nx ; and j D 1; 2; : : : ; M 1.

There are at most .M C1/ B-spline basis functions that do not vanish at each internal
jf
collocation point [5]. Generally, on subinterval i, only the B-splines fBj;M gjDj0
do not
vanish at the internal l , where j0 D .i 1/.M 1/ C 1 and jf D i.M 1/ C 2.
Taking advantage of this property,
X
jf
U.l ; t/ D Bj;M .l /yj .t/: (4)
jDj0
Inserting (4) into (3) results in the system of ODEs
X
jf
Bj;M .m /y0j .t/ D f .t; m ; U.m ; t/; Ux .m ; t/; U xx .m ; t//; (5)
jDj0
where m D 1; 2; : : : ; M 1. Once the system of ODEs is constructed, the discretized

boundary conditions are added to this system of ODEs to form a system of initial-
value differential algebraic equations (DAEs). In other words, BACOLI treats the
boundary conditions directly. For simplicity, on the spatial interval Œ0; 1, at the left
boundary,
U.0; t/ D B1;M .0/y1 .t/; Ux .0; t/ D B01;M .0/y1 .t/ C B02;M .0/y2 .t/:
Similarly, at the right boundary,
U.1; t/ D BD;M .1/yD .t/; Ux .1; t/ D B0D;M .1/yD .t/ C B0D1;M .0/yD1 .t/:
Coupling the ODEs in (5) with the discretized boundary conditions gives an index-1
initial-value system of DAEs of the form
Ba .t; U.0; t/; Ux .0; t// D 0;
X
jf
Bj;M .l /y0j .t/ D f .t; l ; U.l ; t/; Ux .l ; t/; Uxx .l ; t//; (6)
jDj0
Bb .t; U.1; t/; Ux .1; t// D 0;
where l , l D 1; 2; : : : ; M 1, is collocation point l in subinterval i, i D 1; 2; : : : ; Nx .

This is a system of DMu equations in DMu unknowns. The system (6) is also
supplemented by appropriate initial conditions.
This system of DAEs is integrated with respect to the independent variable t in
order to determine the unknown coefficients yj .t/. The variable-stepsize, variable-
order DAE solver DASSL [13] is applied to integrate the resulting initial-value
DAEs with respect to time. After solving (6) with DASSL, an approximate solution
is available. The BACOLI software package has two options for the estimation of the
spatial error. One is through the use of a super-convergent interpolant (SCI) [1], and
the other is a low-order interpolant (LOI) [2]. The SCI option generates a numerical
solution in the space of polynomials of degree M and an error estimate for that
solution. On the other hand, the LOI option generates a numerical solution in the
space of polynomials of degree M C 1 and an error estimate for that solution.
Because one can consider the SCI option to be a standard error control mode, we
use SCI in this paper.
More specifically, in the SCI option, at any time step, based on the solution
computed through the solver DASSL, U.x; t/, which is of degree M, an interpolant,
Q t/, of degree .M C 1/ is constructed using some super-convergent values. The
U.x;
difference between U.x; t/ and U.x; Q t/ gives an asymptotic estimate of the error
for the approximate solution U.x; t/. Finally, BACOLI computes a spatial error
estimate, and if necessary, performs a spatial adaptation in the form of remeshing.
2.3 The Extended BACOLI Software Package
The BACOLI software package has been extended to solve system of equations
given by
ut .x; t/ D f .t; x; v.x; t/; u.x; t/; ux .x; t/; uxx .x; t//;
vt .x; t/ D g.t; x; v.x; t/; u.x; t//;
a x b; t t0 ;
subject to the separated boundary conditions
Ba .t; v.a; t/; u.a; t/; ux .a; t// D 0; Bb .t; v.b; t/; u.b; t/; ux .b; t// D 0;
and the initial conditions
u.x; t0 / D u0 .x/; v.x; t0 / D v0 .x/;
where u.x; t/ and v.x; t/ are the unknown scalar and vector functions of sizes 1 and
Mv , respectively. Full details on the modifications and a detailed list and description
of the modified subroutines may be found in [12].
3 Electrophysiology Background
The heart can be modelled as an electric dipole with separated positive and negative
charges that create an electrical field. The body behaves like a conductor when it
is exposed to such an electrical field. In each heart cycle, an electrical current runs
through the body. This electrical current causes an electrical potential at each point
of the body, giving rise to a time-varying potential difference throughout the body.
The potential variations during a heart cycle represent the electrical activation of
the heart muscle cells. The electrocardiogram (ECG) is a time-dependent reading
of specific potential differences with respect to a reference potential [14]. Figure 1
represents recorded variations of the transmembrane potential as a function of time
for the cell model of Bondarenko et al. [3, 4]. A control volume is a fixed region
considered in space in order to study the energies or fluids that pass its boundary [8].
Continuum modelling of a tissue is a method of modelling based on the concept of
the control volume, in which specific sample points are considered that relate to
a quantity that is the average of its neighbouring cells. In continuum modelling,
instead of considering individual cells, only those sample points are considered.
The monodomain model is a continuum-based, multi-scale mathematical model
for the electrical activity in heart tissue [14]. The standard formulation of this
model is (2). In this paper, the constants , i , and Cm are taken to be 1400 cm1 ,
20
Transmembrane potential (mV)
0
–20
–40
–60
–80
–100
0 10 20 30 40 50 60 70
Time (ms)
Fig. 1 Transmembrane potential as a function of time steps for the cell model of Bondarenko
et al. [3, 4]
1:75 mS/cm, and 1 F/cm2 , respectively, as per the default values from the Chaste
software package (see below).
4 Numerical Results
This section presents the numerical results of solving the monodomain equation
coupled with the Luo–Rudy I cell model [10] and the epicardial variant of the 2006
cell model of ten Tusscher and Panfilov [15]. BACOLI accepts two tolerances, a
relative and an absolute tolerance.
The boundary conditions for both cases are insulating boundary conditions on a
domain chosen as Œ0; 1
vx .0; t/ D 0; vx .1; t/ D 0:
In order to measure the errors of the numerical solution obtained with the
extended BACOLI software package, a reference solution is generated with
Chaste, a powerful, widely used, and well-respected software package for heart
simulation. This software applies a semi-implicit method [19] to the monodomain
model (2a), and the cell models (which reduce to ordinary differential equations at
each discrete point in space) are solved with Heun’s method. The reference solution
is generated by comparing increasingly accurate solutions for which the time step is
halved and the number of spatial mesh points is doubled. The process of generating
these increasingly accurate solutions continues until a desired number of matching
digits are obtained. In our numerical experiments, the comparison is made at 21
equally spaced time steps in the temporal domain and 101 equally spaced spatial
points.
In order to measure the accuracy and efficiency of any numerical method, one
can compute an average of the error at Nt points in the temporal domain t 2 Œt0 ; tf
and at Nx points in the spatial domain, i.e., the average is over N D Nt Nx space-time
points. After generating the reference solution for all N points, the mixed root mean
square (MRMS) [11] error is computed as
v
u N
u1 X vO i vi 2
eMRMS Dt ;
N iD1 1 C jvO i j
where vO i and vi , respectively, denote the reference solution and the numerical
solution for the voltage at space-time point i.
4.1 Monodomain Equation with Luo–Rudy I Cell Model
The Luo–Rudy I cell model consists of 8 variables and simulates the action potential
of a guinea pig ventricular cell [10]. The initial values are taken from [16]. Table 1
shows some accuracy results. The first column contains the spatial points for the
output. The second column represents the reference solution generated by Chaste.
This reference solution has 4 matching digits and MRMS error 4:86 104 %
between two consecutive phases. The third column represents the solution generated
with the extended BACOLI at tolerance 1 108 and SCI spatial error estimate. The
results are reported at the time tf D 5 ms. The matching digits are 7 at x D 0:0 and
6 elsewhere.
For the extended BACOLI software package, the absolute and relative tolerances
are 1:25 102 , and the number of collocation points is 3. These tolerances are
roughly the largest that gives a numerical solution with MRMS error of about
5 %. Because it is possible to achieve MRMS error of about 5 % with relatively
large tolerances, it is most efficient for the extended BACOLI to use a relatively
small number of collocation points. The initial mesh is set to have 4 uniform
subintervals in order to minimize any bias in the extended BACOLI software
package determining a suitable mesh. The resulting numerical solution has MRMS
error 4:15 %, and that for Chaste is almost the same. The results show that with
Table 1 Solution of the x Reference solution Extended BACOLI solution

Luo–Rudy I for the voltage
(mV) after 5 ms, obtained 0.0 19.72929206746164 19.7292917671303982
with Chaste and the 0.2 18.86929313940373 18.8692734805216737
extended BACOLI software 0.4 18.85464593756643 18.8546280990272841
package (SCI scheme and 0.6 19.29529722142660 19.2952427151648926
tolerance 1 108 ) 0.8 16.61051308307339 16.6104957169274741
1.0 20.52934860573765 20.5293347833165818
the extended BACOLI software package it takes 0:18 s and with Chaste it takes
0:50 s. Therefore, in this case there is a speed-up of a factor of around three in
the extended BACOLI software package. Such a saving becomes significant when
such simulations are run many times sequentially as may be required for parameter
or other types of optimization; for example, running the Luo–Rudy I model a
million times consecutively with the extended BACOLI software package would
theoretically save about 320; 000 s or 3:7 days on a six-day run.
4.2 Monodomain Equation with Ten Tusscher Cell Model
The epicardial variant of the model of ten Tusscher and Panfilov [15] consists of
19 variables and simulates the human ventricle [15]. The initial values are taken
from [17]. Table 2 shows some accuracy results. The first column contains the
spatial points for the output. The second column represents the reference solution
generated by Chaste. This reference solution has 3 matching digits and MRMS
error 3:92 103 % between two consecutive phases. The third column represents
the solution generated with the extended BACOLI at tolerance 1 108 and SCI
spatial error estimate. The results are reported at the time tf D 5 ms. The matching
digits vary between 3 and 6.
For the extended BACOLI software package, the absolute and relative tolerances
are 1 102 , and the number of collocation points is 3. These tolerances are roughly
the largest that gives a numerical solution with MRMS error of about 5 %. For
consistency with the previous experiment, the initial mesh size is taken to be uniform
with 4 subintervals. The resulting numerical solution has MRMS error 4:53 % and
that for Chaste is 5:60 %. The results show that with the extended BACOLI
software package it takes 1:16 s and with Chaste it takes 2:58 s. Therefore, in this
case there is a speed-up of a factor of around two in the extended BACOLI software
package.
Table 2 Solution of the x Reference solution Extended BACOLI solution

epicardial variant of the
model of ten Tusscher and 0.0 15.73671511358962 15.7366969196549213
Panfilov [15] for the voltage 0.2 15.88153751011188 15.8815217931186137
(mV) after 5 ms, obtained 0.4 16.07890988375955 16.0788979751286440
with Chaste and the 0.6 16.40115130209244 16.4010866279204457
extended BACOLI software 0.8 17.36927102074231 17.3668857822675768
package (SCI scheme and
tolerance 1 108 ) 1.0 20.73937330189026 20.7486443492494281
5 Conclusions
Modifications to the BACOLI software package were made to extend it to solve

a more general class of problems. The extended BACOLI software package can
solve problems in one dimension that consist of a scalar parabolic PDE coupled with
any number of PDEs that have no dependency on spatial derivatives. The practical
application of such problems described in this paper is the mathematical model of
the electrical activity of the heart known as the monodomain model. The examples
considered in this paper are the monodomain equation coupled with the Luo–Rudy I
cell model and the monodomain equation coupled with the epicardial variant of the
cell model of ten Tusscher and Panfilov [15]. For these examples, we demonstrated
a speed-up of factors of 2–3 in execution time of the extended BACOLI software
package in comparison with the Chaste software package for accuracies typically
required in practice.
References
1. Arsenault, T., Muir, P.H., Smith, T.: Superconvergent interpolants for efficient spatial error
estimation in 1D PDE collocation solvers. Can. Appl. Math. Q. 17(3), 409–431 (2009)
2. Arsenault, T., Smith, T., Muir, P.H., Pew, J.: Asymptotically correct interpolation-based spatial
error estimation for 1D PDE solvers. Can. Appl. Math. Q. 20(3), 307–328 (2012)
3. Auckland Bioengineering Institute.: The CellML project. http://www.cellml.org/
4. Bondarenko, V.E., Szigeti, G.P., Bett, G.C.L., Kim, S.J., Rasmusson, R.L.: Computer model
of action potential of mouse ventricular myocytes. Am. J. Physiol.-Heart C. 287(3), H1378–
H1403 (2004)
5. de Boor, C.: A Practical Guide to Splines. Springer, New York (1978)
6. de Boor, C., The MathWorks Inc.: Spline Toolbox User’s Guide. The MathWorks, Inc., Natick
(1999)
7. Dodgson, N.: B-splines. http://www.cl.cam.ac.uk/teaching/2000/AGraphHCI/SMEG/node4.
html (Aug 2013)
8. Engineers Edge LLC.: Control volume – fluid flow. http://www.engineersedge.com/fluid_flow/
control_volume.htm (Feb 2014)
9. Holder, D., Huo, L., Martin, C.F.: The Control of Error in Numerical Methods. Springer, New
York (2007)
10. Luo, C., Rudy, Y.: A model of ventricular cardiac action potential. Circ. Res. 68(6), 1501–1526
(1991)
11. Marsh, M.E., Torabi Ziaratgahi, S., Spiteri, R.J.: The secrets to the success of the Rush – Larsen
method and its generalizations. IEEE Trans. Bio-Med. Eng. 59(9), 2506–2515 (2012)
12. Mirshekari, E.: Extending BACOLI to solve multi-scale problems. Master’s thesis, Department
of Mathematics and Statistics, University of Saskatchewan (2014)
13. Petzold, L.R.: A description of DASSL: a differential/algebraic system solver. Technical report,
Sandia National Labs., Livermore (1982)
14. Sundnes, J., Lines, G.T., Cai, X., Nielsen, B.F., Mardal, K.A., Tveito, A.: Computing the
Electrical Activity in the Heart. Springer, Berlin (2006)
15. ten Tusscher, K.H.W.J., Panfilov, A.V.: Alternans and spiral breakup in a human ventricular
tissue model. Am. J. Physiol. Heart Circ. Physiol. 291(3), 1088–1100 (2006)
16. The CellML Project.: A model of the ventricular cardiac action potential. http://models.cellml.
org/exposure/2d2ce7737b42a4f72d6bf8b67f6eb5a2/luo_rudy_1991.cellml/@@cellml_
codegen/F77 (Feb 2014)
17. The CellML Project: Alternans and spiral breakup in a human ventricular tissue model
(epicardial model). http://models.cellml.org/exposure/a7179d94365ff0c9c0e6eb7c6a787d3d/
ten_tusscher_model_2006_IK1Ko_epi_units.cellml/@@cellml_codegen/F77 (Apr 2014)
18. Wang, R., Keast, P., Muir, P.H.: A high-order global spatially adaptive collocation method for
1-D parabolic PDEs. Appl. Numer. Math. 50(2), 239–260 (2004)
19. Whiteley, J.P.: An efficient numerical technique for the solution of the monodomain and
bidomain equations. IEEE Trans. Bio-Med. Eng. 53(11), 2139–2147 (2006)
An Analysis of the Reliability of Error Control
B-Spline Gaussian Collocation PDE Software
Paul Muir and Jack Pew
Abstract B-spline Gaussian collocation software for the numerical solution of

PDEs has been widely used for several decades. BACOL and BACOLI are recently
developed packages of this class that provide control of estimates of the temporal
and spatial errors of the numerical solution through the use of adaptive time-
stepping/adaptive time integration method order selection and adaptive spatial mesh
refinement. Previous studies have investigated the performance of the BACOL
and BACOLI packages, primarily with respect to work-accuracy measures. In
this paper, we investigate the reliability of the BACOL and BACOLI packages,
focusing on the relationship between the requested tolerance and the accuracy
achieved. In particular, we consider the effect, on the reliability of the software,
of (i) the degree of the piecewise polynomials employed in the representation of the
spatial dependence of the approximate solution, (ii) the type of spatial error control
employed, and (iii) the type of spatial error estimate computed.
1 Introduction/Background
B-spline Gaussian collocation software for the numerical solution of PDEs has been
widely used for several decades. Although there has been some development of
software of this type for 2D PDEs, see, e.g., [9] and references within, software
for the 1D case is at a more advanced state, and it is this case that we focus on in
this paper. (However, since the approach for the 2D case considered in [9] builds
on the numerical algorithms used in the 1D case, the questions we consider in this
paper will be relevant for the 2D case.) A characterizing feature of such software
is the representation of the numerical solution as a linear combination of known
B-spline [4] basis functions (piecewise polynomials of a given degree p) in space
within unknown time-dependent coefficients. The B-spline basis is chosen to have
P. Muir () • J. Pew

Saint Mary’s University, Halifax, NS, B3H 3C3, Canada

460 P. Muir and J. Pew
C1 -continuity and depends on a spatial mesh, fxi g; i D 0; : : : ; NINT, that partitions

the spatial domain. The approximate solution has the form
X
NC
U.x; t/ D yi .t/Bi .x/; (1)
iD1
where yi .t/ is the time-dependent coefficient of the i-th B-spline basis function,
Bi .x/, and NC D NINT.p 1/ C 2. The B-spline coefficients are determined by
solving collocation conditions, obtained by requiring the approximate solution to
exactly satisfy the PDEs at the images of the Gauss points of a given order on
each subinterval of the spatial mesh, and additional equations that depend on the
boundary conditions. Assuming a system of PDEs of the form

ut .x; t/ D f t; x; u.x; t/; ux .x; t/; uxx .x; t/ ; a x b; t t0 ; (2)
with initial conditions
u.x; t0 / D u0 .x/; a x b; (3)
and boundary conditions

bL t; u.a; t/; ux .a; t/ D 0; bR t; u.b; t/; ux .b; t/ D 0; t t0 ; (4)
the collocation conditions are a system of ODEs the form

U t .l ; t/ D f t; l ; U.l ; t/; U x .l ; t/; U xx .l ; t/ ; (5)
for l D 2; : : : ; NC 1. The collocation points are l D xi1 C hi j ; l D 1 C

p1
.i 1/.p 1/ C j; i D 1; : : : ; NINT; j D 1; : : : ; p 1; fi giD1 are the images of
the Gauss points on Œ0; 1, and hi D xi xi1 . (Thus, kcol, the number of collocation
points per subinterval, is p 1.)
About 35 years ago, the B-spline Gaussian collocation package, PDECOL [10]
was developed. The B-spline basis was implemented using the de Boor B-spline
package [4] and the ODE solver, GEARIB [6], was used to solve the collocation
conditions (5) together with the time differentiated boundary conditions. Temporal
error control was provided by the ODE solver. PDECOL used a fixed spatial mesh
which meant that no control of the spatial error was provided.
About a decade later, a modification of PDECOL, called EPDCOL [8] was
released; EPDCOL replaced the band matrix solver employed by PDECOL with
an almost block diagonal (ABD) solver, COLROW [5], more appropriate for the
structure of the linear systems arising from the B-spline collocation process. Savings
in execution time of about 50 % were obtained.
The next update to this family of PDE solvers, BACOL [18], released a little
over a decade later, was a new implementation of the B-spline Gaussian collocation
Reliability of Error Control B-Spline Gaussian Collocation Software 461
algorithm, with several additional features. BACOL employed a Differential-

Algebraic Equation (DAE) solver, DASSL [3], so that the boundary conditions (4)
could be treated directly together with the ODEs (5) arising from the collocation
process. Temporal error control was provided by DASSL. The most significant new
feature was the implementation of spatial error estimation and control. The spatial
error estimates were obtained by computing a second (higher order) collocation
N t/, based on B-splines of degree p C 1, which was then used together
solution, U.x;
with U.x; t/ to compute a set of spatial error estimates of the form
s
Z 2
b
Uj .x; t/ U
N j .x; t/
Ej .t/ D dx; j D 1; : : : ; NPDE; (6)
a ATOLj C RTOLj jUj .x; t/j
where Uj .x; t/ is the jth component of U.x; t/, U N j .x; t/ is the jth component of
N
U.x; t/, ATOLj and RTOLj are absolute and relative tolerances corresponding to the
jth component of the solution, and the number of PDEs is NPDE. At the end of each
time step, t, the collocation solution, U.x; t/ was accepted if
max Ej .t/ 1: (7)

1 jNPDE
Otherwise, the time step was rejected and BACOL computed a second set of spatial
error estimates of the form,
v
uNPDE Z 2
uX xi
Uj .x; t/ U
N j .x; t/
OEi .t/ D t dx; i D 1; : : : ; NINT;
jD1 xi1 ATOLj C RTOLj jUj .x; t/j
(8)
which were then used within an adaptive mesh refinement (AMR) process for
which the goal was to determine a spatial mesh for the current time such that
the estimated spatial errors were (i) approximately equidistributed over the mesh
subintervals and (ii) less than the user tolerance [20]. BACOL was shown to have
superior performance compared to a number of comparable packages, especially for
problems with sharp moving layers and higher accuracy requirements [19].
In BACOL, the spatial error estimate is used to control the error in U.x; t/, the
approximate solution returned by the code. We refer to this as Standard (ST) error
control. A minor modification of BACOL would allow it to return U.x; N t/ and in
this case, the computation of the returned solution would be controlled based on an
error estimate for a collocation solution that is of one lower order. We refer to this as
Local Extrapolation (LE) error control; see, e.g., [7], in the context of Runge-Kutta
formula pairs for the numerical integration of ODEs.
The most recent update to this family of solvers is a modification of BACOL
known as BACOLI [12]. The key feature of BACOLI is the avoidance of the com-
N t/, which represented a substantial cost within BACOL. BACOLI
putation of U.x;
instead computes only one collocation solution and one low cost interpolant that is
used in the computation of the spatial error estimate. There are two options available
for the interpolant. One involves a superconvergent interpolant (SCI) [1] which is
based on the presence of certain points within the spatial domain where U.x; t/ is
of at least one order of accuracy higher. The SCI is of the same order of accuracy
N t/ and leads to a spatial error estimate for U.x; t/, similar to the situation in
as U.x;
BACOL. This option therefore provides ST error control. The other option involves
a lower order interpolant (LOI) [2], of one order of accuracy lower than U.x; t/. The
LOI is based on interpolation of U.x; t/ at certain points such that the interpolation
error of the LOI agrees asymptotically with the error of a collocation solution of one
order lower than U.x; t/. This option therefore provides LE error control. BACOLI
has been shown to be about twice as fast as BACOL [12].
When considering the effectiveness of a numerical software package, the most
common analysis involves work-precision (i.e., work-accuracy) diagrams that show
the relationship between the CPU time and the accuracy. A number of such studies
involving BACOL and BACOLI have been performed; see, e.g., [12, 19]. However,
an equally important measure of the quality of a software package is its reliability ,
and a key component of this measure involves assessing the relationship between the
tolerance requested and the accuracy achieved. This depends on the quality of the
error estimates and the algorithms that adapt the computation to attempt to control
the error estimates with respect to the user-specified tolerances.
The purpose of this paper is to present an experimental analysis of the reliability
of the error control PDE solvers BACOL and BACOLI on a standard test problem
chosen from the literature. (Additional results of this type for several other test
problems are reported in [11] and some preliminary results of this type for BACOL
were reported in [17].) In particular, we are interested in how the choice of the
degree, p, of the B-spline basis and the choice of error control mode (ST or LE)
impact on the reliability of the solvers. We also examine the effectiveness of the
interpolation-based spatial error estimates provided by the SCI and LOI schemes.
2 Numerical Results
In this section we present numerical results for four code combinations identified in
the previous section:
• BACOL with ST error control : BAC/ST
• BACOL with LE error control : BAC/LE
• BACOLI with the SCI scheme/ST error control : SCI/ST
• BACOLI with the LOI scheme/LE error control : LOI/LE
We consider the One Layer Burgers Equation (OLBE) with D 104 :
ut D uxx uux ; (9)

with boundary conditions at x D 0 and x D 1 (t > 0) and an initial condition at

t D 0 .0 x 1/ such that the exact solution is
!
1
1 1 x 2t 4
u.x; t/ D tanh :
2 2 4
The solution has a sharp layer region around x D 0:25 when t D 0. As t goes from
0 to 1, the layer moves to the right and is located around x D 0:75 when t D 1. A
plot of the solution for D 104 is given in [12].
We apply each of the four code combinations identified above to the OLBE
with D 104 , for a range of kcol values (recall that the degree of the piecewise
polynomials in space is p D kcol C 1) and for 91 tolerance values (with ATOL1 D
RTOL1 ) uniformly distributed from 101 to 1010 . For each experiment, we report
the L2 -norm of the error at the final time, Tout D 1. Additional results, including
results for other test problems, are given in [11].
2.1 Reliability vs. kcol
In this subsection, we consider, for the LOI/LE code, the relationship between the
error and the tolerance, for kcol D 3; 5; 7; 9. We provide plots of the tolerance vs.
the error and plots of the tolerance vs. the error tolerance ratio (ETR), i.e., the error
divided by the tolerance. See Figs. 1, 2, 3, and 4. In Fig. 1a, we see that there is good
correlation between the requested tolerance and the achieved accuracy; the points
are clustered around the reference line corresponding to equal values of the tolerance
and error. In Fig. 1b, considering the horizontal reference line corresponding to an
ETR value of 1, we see that the ETR values vary over the range of 91 tolerance
values, with some errors greater than the tolerance and some less than the tolerance.
The other horizontal line represents the average of the ETR values. This figure also
Fig. 1 (a) Tolerance vs. Error, LOI/LE, kcol D 3, (b) Tolerance vs. ETR, LOI/LE, kcol D 3
shows that maximum ETR is 5.99, the minimum is 0.08, and that on average the
error is 1.86 times the tolerance. Figures 2, 3, and 4, give similar results for the
kcol D 5; 7; 9 cases, and show maximum, minimum, and average ETR values of
{3.35, 0.15, 1.13}, {3.78, 0.01, 1.03}, and {5.72, 0.01, 0.90}, respectively. Similar
results are reported in [11] for the other code combinations and test problems.
2.2 Reliability vs. Error Control Mode
In this subsection we consider the BAC/ST and BAC/LE codes for the kcol D 5
case. From the results for these codes we can assess the effect that the error control
mode has on reliability since this is the only difference between these codes. From
Figs. 5 and 6 we can see that the two codes appear to have quite similar performance.
For the BAC/ST code, the maximum, minimum, and average ETR values are 4.49,
0.28, and 1.79, while for the BAC/LE code, the corresponding values are 4.49,
Fig. 5 (a) Tolerance vs. Error, BAC/ST, kcol D 5, (b) Tolerance vs. ETR, BAC/ST, kcol D 5
Fig. 6 (a) Tolerance vs. Error, BAC/LE, kcol D 5, (b) Tolerance vs. ETR, BAC/LE, kcol D 5
Fig. 7 (a) Tolerance vs. Error, SCI/ST, kcol D 5, (b) Tolerance vs. ETR, SCI/ST, kcol D 5
0.18, and 2.01. Similar results are reported in [11] for the other kcol values and
test problems.
2.3 Reliability vs. Error Estimation Scheme
In this section we compare BAC/ST with SCI/ST and BAC/LE with LOI/LE, for
kcol D 5. This will allow us to assess the effectiveness of the interpolation-
based error estimation schemes. The plots for the SCI/ST code are given in Fig. 7.
A comparison of Figs. 5 and 7 show that the BAC/ST and SCI/ST codes have
comparable reliability, while a comparison of Figs. 6 and 2 show that the BAC/LE
and LOI/LE codes also have comparable reliability. From Fig. 7b, we see that for
the SCI/ST code the maximum ETR is 6.85, the minimum ETR is 0.10, and the
average is 1.63. Generally similar results for the other kcol values and test problems
are reported in [11]. (The only exception, as documented in [12] and [11], is for the
SCI/ST, kcol D 3 case, where issues associated with the quality of its error estimates
have been observed.)
3 Discussion/Conclusions/Future Work
Several conclusions can be drawn from the results presented in the previous
section:
• The codes considered in this paper appear to be reliable; the error is generally
well correlated with the requested tolerance. We see that in almost all cases
the error is within the correct order of magnitude of the requested tolerance.
Furthermore the error is, on average, a small multiple of the requested tolerance.
This suggests that the error estimates computed by the codes are generally of
good quality and that the new SCI and LOI error estimation schemes are of
comparable quality to the original error estimation scheme employed by BACOL.
• The relationship between the error and tolerance for each of the two versions of
BACOL that employ different error control modes (ST vs. LE) is quite similar.
This suggests that the choice of error control mode does not have a strong impact
on the reliability of the solver.
• The degree p of the B-spline basis as specified by the choice of kcol (p D kcolC1)
does appear to have an impact on the reliability. A comparison of the results for
the kcol D 3; 5; 7, and 9 cases shows that as kcol is increased, the reliability
of the codes tends to increase; that is, the error is, on average, almost equal to
the tolerance for higher kcol values. For the kcol D 9 case, the error is often
significantly less than the tolerance. (Of course, while it is desirable for the error
to be about the same size as or even slightly less than the tolerance, if the error is
too much less than the tolerance this can lead to inefficiency in the computation.)
While the entire family of software packages discussed in this paper leave kcol
as a parameter that must be chosen by the user, it is in fact not clear how to
make an appropriate choice of this parameter for a given problem. The correlation
between higher kcol values and improved reliability that can be seen from the
results presented in this paper, as well as earlier results investigating the relationship
between the choice of kcol, the tolerance, and the efficiency of the solvers [12], will
provide a basis for further analysis that will allow us to develop a new release of
BACOLI in which the code itself will choose kcol appropriately to improve the
efficiency and reliability of the computation.
Since BACOL and BACOLI employ DASSL to compute the time-dependent B-
spline coefficients, the overall reliability of the computations performed by these
codes is dependent on the reliability of DASSL. An example of the relationship
between the tolerance and the error for DASSL can be seen in Figure 1 of [16];
it shows generally good correlation between the error and the tolerance, but with
variations in the ETR of approximately one order of magnitude in either direction,
similar to those reported here for BACOL and BACOLI.
Furthermore, for DASSL and the codes considered in this paper, the relationship
between the tolerance and the error is not particularly smooth. That is, while the
correlation between a given tolerance and the corresponding error is generally good
(i.e., the error is, on average, within a small multiple of the tolerance), a small
change in the tolerance does not necessarily lead to a similar small change in the
error. This is related to what is referred to as the “conditioning” or “stability” of
the software [16]. The papers [14, 15], and [13] describe how control theoretic
techniques can be employed within ODE or DAE software to provide an improved
condition number for the software. We plan to introduce these type of techniques
into the BACOL and BACOLI packages.
Finally, as mentioned earlier, since the approach for the 2D case considered in
[9] builds on the 1D case, we expect the results of this paper to be relevant for the
numerical treatment of 2D PDEs.
References
1. Arsenault, T., Smith, T., Muir, P.H.: Superconvergent interpolants for efficient spatial error
estimation in 1D PDE collocation solvers. Can. Appl. Math. Q. 17, 409–431 (2009)
2. Arsenault, T., Smith, T., Muir, P.H., Pew, J.: Asymptotically correct interpolation-based spatial
error estimation for 1D PDE solvers. Can. Appl. Math. Q. 20, 307–328 (2012)
3. Brenan, K.E., Campbell, S.L., Petzold, L.R.: Numerical Solution of Initial-Value Problems
in Differential-Algebraic Equations. Society for Industrial and Applied Mathematics (SIAM),
Philadelphia (1989)
4. de Boor, C.: A Practical Guide to Splines. Volume 27 of Applied Mathematical Sciences.
Springer, New York (1978)
5. Díaz, J.C., Fairweather, G., Keast, P.: Algorithm 603. COLROW and ARCECO: FORTRAN
packages for solving certain almost block diagonal linear systems by modified alternate row
and column elimination. ACM Trans. Math. Softw. 9(3), 376–380 (1983)
6. Gear, C.W.: Numerical Initial Value Problems in Ordinary Differential Equations. Prentice-
Hall, Englewood Cliffs (1971)
7. Hairer, E., Nørsett, S.P., Wanner, G.: Solving Ordinary Differential Equations. I. Volume 8 of
Springer Series in Computational Mathematics, 2nd edn. Springer, Berlin (1993)
8. Keast, P., Muir, P.H.: Algorithm 688: EPDCOL: a more efficient PDECOL code. ACM Trans.
Math. Softw. 17(2), 153–166 (1991)
9. Li, Z., Muir, P.H.: B-spline Gaussian collocation software for two-dimensional parabolic PDEs.
Adv. Appl. Math. Mech. 5, 528–547 (2013)
10. Madsen, N.K., Sincovec, R.F.: Algorithm 540: PDECOL, general collocation software for
partial differential equations. ACM Trans. Math. Softw. 5(3), 326–351 (1979)
11. Muir, P.H., Pew, J.: Tolerance vs. error results for a class of error control B-spline Gaussian
collocation PDE solvers. Saint Mary’s University, Department of Mathematics and Computing
Science Technical Report Series. http://cs.smu.ca/tech_reports (2015)
12. Pew, J., Li, Z., Muir, P.H.: A computational study of the efficiency of collocation software for
1D parabolic PDEs with interpolation-based spatial error estimation. Saint Mary’s University,
Dept. of Mathematics and Computing Science Technical Report Series. http://cs.smu.ca/tech_
reports (2013)
13. Pulverer, G., Söderlind, G., Weinmüller, E.: Automatic grid control in adaptive BVP solvers.
Numer. Algorithms 56(1), 61–92 (2011)
14. Söderlind, G.: Digital filters in adaptive time-stepping. ACM Trans. Math. Softw. 29(1), 1–26
(2003)
15. Söderlind, G., Wang, L.: Adaptive time-stepping and computational stability. J. Comput. Appl.
Math. 185(2), 225–243 (2006)
16. Söderlind, G., Wang, L.: Evaluating numerical ODE/DAE methods, algorithms and software.
J. Comput. Appl. Math. 185(2), 244–260 (2006)
17. Wang, R.: High order adaptive collocation software for 1-D parabolic PDEs. Ph.D. thesis,
Dalhousie University (2002)
18. Wang, R., Keast, P., Muir, P.H.: BACOL: B-spline Adaptive COLlocation software for 1D
parabolic PDEs. ACM Trans. Math. Softw. 30(4), 454–470 (2004)
19. Wang, R., Keast, P., Muir, P.H.: A comparison of adaptive software for 1D parabolic PDEs. J.
Comput. Appl. Math. 169(1), 127–150 (2004)
20. Wang, R., Keast, P., Muir, P.H.: A high-order global spatially adaptive collocation method for
1-D parabolic PDEs. Appl. Numer. Math. 50(2), 239–260 (2004)
On the Simulation of Porous Media Flow Using
a New Meshless Lattice Boltzmann Method
S. Hossein Musavi and Mahmud Ashrafizaadeh
Abstract We propose a new meshless lattice Boltzmann method (MLLBM) for

the simulation of nearly incompressible flows in porous media. As for the standard
lattice Boltzmann method, the collision and the streaming operators are split. While
the collision equation remains unaltered, the streaming equation is discretized using
the Lax-Wendroff scheme in time, and the meshless local Petrov-Galerkin scheme
in space. The Poiseuille flow is solved to validate the numerical scheme. The method
is then used to simulate a flow in a porous medium. Although in general, meshless
methods suffer from high computational costs, the present method shows promising
accuracy and performance at a lower memory and computational costs for complex
geometries, when compared with standard lattice Boltzmann methods.
1 Introduction
Since the beginning of the development of the lattice Boltzmann method (LBM),
a great number of studies have been conducted to extend the capabilities of the
standard lattice Boltzmann method to handle nonuniform and unstructured grids
[1–6]. A major trend in these studies is to apply some standard numerical tech-
niques, such as the finite difference (FD), the finite volume (FV), and the finite
element (FE) methods, for the discretization of the Boltzmann equation. However,
for geometrically complex problems, such as flows in porous media, where the time
and computational requirements of creating good quality meshes are significant, the
idea of developing a meshless lattice Boltzmann solver naturally arises.
In this study, we implement our newly proposed meshless lattice Boltzmann
method [7] for the simulation of nearly incompressible flows in porous media.
In our formulation, we split the collision and the advection equations following
the standard lattice Boltzmann method [8]. The collision equation is the same as
for the standard LBM, however, the advection equation is discretized using the
S.H. Musavi • M. Ashrafizaadeh ()

Department of Mechanical Engineering, Isfahan University of Technology,
Isfahan 8415683111, Iran

470 S.H. Musavi and M. Ashrafizaadeh
Lax-Wendroff scheme in time and the meshless local Petrov-Galerkin (MLPG)

method in space.
The meshless feature of the proposed method makes it a more flexible lattice
Boltzmann solver, especially for cases where using the usual mesh generation
techniques introduces significant numerical errors into the solution, or where
improving the mesh quality is a complex and time consuming process. These
features of the present method are well illustrated in a number of test cases described
in the next sections.
2 Formulation
The discrete Boltzmann equation (DBE) with the BGK collision approximation is
@fi @fi 1 eq
C ci;˛ D . fi fi /; i D 1; : : : ; nQ ; (1)
@t @x˛
where fi is the particle distribution function, is the relaxation time towards

equilibrium, nQ is the number of discrete microscopic velocities and ci is the
discrete microscopic velocity. For the D2Q9 lattice we have
c0 D 0 ;

ci D cos.i 1/ ex C sin.i 1/ ey ; i D 1; 3; 5; 7 ;
4 4
p
ci D 2Œcos.i 1/ ex C sin.i 1/ ey ; i D 2; 4; 6; 8 ;
4 4
(2)
eq
and fi is the equilibrium distribution, that is
!
eq ci;˛ u˛ .ci;˛ u˛ /2 u2˛
fi D ti 1C C : (3)
c2s 2c4s 2c2s
As usual, we break Eq. (1) into two steps, namely the collision step:
1 eq
fQi D fi . fi fi /; (4)

and the advection step:
@fi @fi
C ci;˛ D 0: (5)
@t @x˛
On the Simulation of Porous Media Flow Using a New Meshless Lattice. . . 471
We discreatize Eq. (5) in time using the Lax-Wendroff scheme, which reads
@fin ıt2 @2 fin

finC1 D fin ıtci;˛ C ci;˛ ci;ˇ : (6)
@x˛ 2 @x˛ @xˇ
where ıt is the time step size and the superscript n is the time step number.
In order to apply the meshless local Petrov-Galerkin scheme to discretize Eq. (6)
in space, first, the local weak form of Eq. (6) on the control volume ˝I of point I is
derived by taking its inner product with a local test function WI over ˝I , and using
integration by parts, so that we obtain
Z Z Z
@fin ıt2 @WI @fin
WI finC1 d˝ D WI fin d˝ ıtWI ci;˛ C ci;˛ ci;ˇ d˝
˝I ˝I ˝I @x˛ 2 @xˇ @x˛
2 Z
ıt @fin
C WI ci;˛ ci;ˇ nˇ d
; (7)
2
I @x˛
where
I is the boundary of the control volume ˝I and nˇ is the unit outward normal
vector of
I . Equation (7) is the local weak form of Eq. (6).
In the next step, the field variable fi is to be expressed in terms of nodal values
fi;J by a local interpolation scheme, that is
X
Ns
fi .x; t/ D J .x/fi;J .t/ D ˚ T .x/fs .t/; (8)
JD1
where Ns is the number of nodal points in a local interpolation domain of point x

called the support domain, ˚ T .x/ D f1 .x/ 2 .x/ : : : Ns .x/g is the transpose of
the vector of shape functions, and fs .t/ D ffi;1 .t/ fi;2 .t/ : : : fi;Ns .t/gT is the vector of
the nodal values of fi in the support domain. In this study, we make use of the local
radial point interpolation method (LRPIM) [9] which uses the local radial basis
functions (RBF) augmented with polynomials as the basis function; thus ˚ T .x/
in the interpolation equation (8) is the transpose of the vector of LRPIM shape
functions given as [9]
Q̊ T .x/ D fRT .x/ pT .x/gG1 ; (9)
where RT .x/ D fR1 .x/ R2 .x/ : : : RNs .x/g is the transpose of the vector of radial
basis functions (RBF), pT .x/ D f1 x y : : : pm .x/g is the transpose of the vector of
T
monomial basis functions, m is the number of monomial basis functions, Q̊ .x/ D
f˚ .x/ Ns C1 .x/ : : : Ns Cm .x/g is the transpose of the extended vector of the shape
T
functions, and G is a symmetric matrix defined in Ref. [9]. Substituting Eq. (8) in
Eq. (7) we write,
NI h Z
X i NI h Z
X
WI J d˝ fi;J
nC1
D WI J d˝
JD1 ˝I JD1 ˝I
Z
ıt2 @WI @J
ıtWI C ci;ˇ ci;˛ d˝
˝I 2 @xˇ @x˛
Z i
ıt2 @J
fi;Jn
; (10)
2
I @x˛
where NI is the number of nodal points involved in the interpolation of the field
variable on the inner and the boundary points of the control volume ˝I . By
introducing the mass matrix as
Z
MIJ D WI J d˝; (11)
˝I
and the stiffness matrix as

Z
ıt2 @WI @J
Ki;IJ D ıtWI C ci;ˇ ci;˛ d˝
˝I 2 @xˇ @x˛
Z
ıt2 @J
; (12)
2
I @x˛
we rewrite Eq. (10) as follows,
X
NI X
NI
nC1
MIJ fi;J D ŒMIJ C Ki;IJ fi;J
n
: (13)
JD1 JD1
To complete the discretization process, the integrals of equations (11) and (12) are
to be evaluated numerically. The Gauss quadrature scheme is employed for this
purpose. We have
X
NG
MIJ D k WI .xk /J .xk /jJ˝I j; (14)
kD1
and
X
ıt2 @WI ˇˇ
@J ˇˇ ˝I
NG
Ki;IJ D k ıtWI .xk / C ci;ˇ ˇ ci;˛ ˇ jJ j
kD1
2 @xˇ xk @x˛ xk

@ ˇ
b
ıt2 X
NG
Jˇ
C k WI .xk / ci;˛ ˇ .ci;ˇ nˇ /jJ
I j; (15)
2 kD1 @x˛ xk
where k is the Gauss weighting factor for the Gauss quadrature point xk , J˝I and
J
I are the mapping Jacobian matrices for the domain and the boundary integrations,
respectively, and NG and NGb are the number of Gauss points used for the domain
and the boundary integrations, respectively.
Now, Eq. (13) becomes the fully discretized equation for the nodal point I.
Writing this equation for all of the nodal points in the computational domain
(I D 1; : : : ; N), and assembling the resulting equations in a global system of
equations, we can write
MfnC1
i D ŒM C Ki fni ; i D 1; : : : ; nQ ; (16)
where M, K, and fi are the global mass matrix, stiffness matrix, and particle
distribution vector, respectively. Equation (16) is a system of N equations with N
unknowns which should be solved separately for each direction i after imposing the
boundary conditions.
The advection equation of the particle distributions is a hyperbolic equation,
which requires boundary conditions for the incoming particles at the boundary
(ci;ˇ nˇ < 0). In this study, we impose boundary conditions using the bounce-back
scheme of non-equilibrium distributions, i.e.
eq eq
fi fi D fi fi ; (17)
where fi is the outgoing particle distribution along the opposite direction of the
incoming distribution fi . Substituting the equilibrium distribution of Eq. (3) in the
above equation, we obtain
fi D fi C 2b ti .ci;˛ ub;˛ /=c2s ; (18)
where b and ub;˛ are the macroscopic density and velocity at the boundary. If fi
in Eq. (18) is considered to be the post-collision (pre-streaming) distribution, then
Eq. (18) becomes an explicit essential boundary condition for the discretized system
of Eq. (16).
The coefficient matrix in Eq. (16) is a sparse matrix which can be efficiently
solved using sparse iterative solvers such as BiCGStab. However, the explicit nature
of the standard lattice Boltzmann method, and the diagonally dominant character
of the mass matrix, motivated us to find rational ways of diagonalizing the mass
matrix, and thus save much of the computational time. In this study, we use row-sum
lumping, in which the sum of the elements of each row of the mass matrix is used as
the diagonal element. As a result, our meshless lattice Boltzmann method becomes
an explicit solver for the fluid flow problems. The maximum value for the time
step leading to a stable solution is determined using the Courant number, CFL D
max fjci jg ıt=ıxmin , where ıxmin is the minimum point spacing in the domain. For a
stable solution the CFL number should be smaller than 1.0.
3 Results
3.1 Poiseuille Flow
The first test case considered in this study is the pressure driven fully developed flow
between two parallel plates. In solving this test case using the proposed method,
we consider a square computational domain in the xy plane and discretize it using
9 9, 17 17, 23 23, and 65 65 uniform point distributions. The constant
pressure boundary condition is imposed in the inlet and outlet and the no-slip and
the impermeability boundary conditions are imposed at the walls. For each point
distribution, time iterations continue until a steady state is reached. The result of our
method for the velocity distribution is depicted and compared with the analytical
solution in Fig. 1.
0.8
0.6
U
0.4 Presentstudy
Analyticalsolution
0.2
0
0 0.2 0.4 0.6 0.8 1
Y
Fig. 1 The velocity distributions for the Poiseuille flow; line: analytical solution, symbols: present
study
0
10
10-1
E 10-2
10-3
-4
10
10 -2 10 -1 10 0
h
Fig. 2 Numerical convergence of the meshless Lattice Boltzmann method in L2 error norm with
respect to the point spacing for the Poiseuille flow
In order to determine the order of accuracy of the numerical scheme, we employ

the following relative L2 error norm,
PNe !1=2
ID1 .UaI UnI /2
ED PNe 2
; (19)
ID1 UaI
where UaI and UnI are the analytical and numerical solutions of the velocity at
point I, respectively, and Ne is the fixed number of points used for the error
analysis. For this test case, a 10 10 uniform point distribution is used for the error
analysis. The variation of the above error norm with respect to the point spacing
h is sketched in the logarithmic diagram of Fig. 2. The rate of the convergence of
our numerical method, computed by the linear regression of the data in Fig. 2, is
R D 2:15.
3.2 Flow in a Porous Medium
One of the most common applications of the standard lattice Boltzmann method is
the simulation of the flow in porous media. In order to illustrate the ability of the
present meshless lattice Boltzmann method to deal with complex geometries, we
simulate a pressure-driven fluid flow in a typical two-dimensional porous medium

shown in Fig. 3 and compute the permeability of the medium. One of the common
definitions of the Reynolds number for the porous media flow is
vDp
Re D ; (20)

where and are the density and viscosity of the fluid, v is the superficial flow
velocity, and Dp is a representative grain size for the porous media, which is usually
chosen to be the average grain diameter. Experimental studies have illustrated that
the flow regimes with low Reynolds numbers (typically less than 10) are Darcian,
meaning that the superficial velocity-pressure gradient relation obeys the Darcy’s
law
k
v D rp; (21)

where k is the permeability of the medium measured in m2 or Darcy, where

1 Darcy D 9:869233 1013 m2 .
In our simulation, the no-slip boundary condition is imposed on the top and
the bottom walls and on the grain surfaces using the bounce-back scheme and the
constant pressure boundary condition is imposed on the left and right boundaries.
The pressure differences between inlet and outlet boundaries in our simulations are
Fig. 3 Domain and a typical point distribution for the flow in a porous medium
set so that the Reynolds number of the flow is low and the flow remains in the
Darcian regime.
The domain is discretized using arbitrary sets of nodal points as shown in Fig. 3.
First, to investigate the convergence of the numerical method with respect to the
domain discretization, we discretize the domain with 2601, 7364, 25284, and 90043
arbitrary distributed nodal points and compute the superficial velocity for each case.
The results are depicted in Fig. 4. As shown in this figure, the difference between
the cases with 25284 and 90043 points is less than 1 % and therefore we use the case
with 25284 points in all the following simulations. However, as observed from the
figure, the standard LBM requires 640000 grid points for the accurate simulation of
the flow. It means that the required number of points in our method is about 1/25th
of that of the standard LBM.
Next, to obtain the permeability of the medium, we compute the superficial
velocity for different pressure gradients and perform the linear interpolation of the
results as is depicted in Fig. 5. From this figure, the permeability is obtained as
k D 1:53 105 m2 D 1:55 107 Darcy.
The pressure and velocity magnitude distributions of the porous medium flow
are depicted in Figs. 6 and 7, respectively, for the dimensionless pressure gradient
of 0.1.
0.0014
Standard LBM
Meshless LBM
0.0012
Superficial velocity
0.001 Standard LBM

640000 grid points
0.0008
Meshless LBM
25284 nodal points
0.0006
0.0004 3 4 5 6
10 10 10 10
Number of points
Fig. 4 The superficial velocity versus the number of nodal points used in the meshless LBM
0.002
0.0016
Superficial velocity
0.0012
0.0008
0.0004
0
0 0.02 0.04 0.06 0.08 0.1 0.12
Pressure gradient
Fig. 5 The superficial velocity with respect to the pressure gradient for the flow in the porous
medium
0.8 0.09
0.08
0.07
0.06
0.05
0.04
0.6
0.03
0.02
y
0.01
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
x
Fig. 6 Pressure distribution for the flow in the porous medium

vel
0.8 0.02
0.018
0.016
0.014
0.012
0.01
0.6
0.008
0.006
y
0.004
0.002
0
0.4
0.2
0
0 0.2 0.4 0.6 0.8 1
x
Fig. 7 Velocity magnitude distribution for the flow in the porous medium
4 Conclusions
A new meshless lattice Boltzmann method (MLLBM) has been developed for the
simulation of the nearly incompressible fluid flows. The main advantage of our
method with respect to the previous extensions of the lattice Boltzmann method
is to eliminate the need for any meshes. A feature that shows its superiority over the
standard lattice Boltzmann method in complex geometries, such as porous media.
Two test cases have been considered in this study. First, the Poiseuille flow has
been solved and compared with the analytical solution, showing the second order of
accuracy of our method. Next, the flow in a porous medium has been simulated to
illustrate the capability of the method in dealing with domains with a very complex
geometry. Although the results presented in this study are for two-dimensional
cases, the extension for three-dimensional problems is straightforward and remains
for our future works. However, our preliminary 3D developments show that the time
and memory saving for the meshless LBM method should be even higher than that
of the 2D model.
References
1. Bardow, A., Karlin, I.V., Gusev, A.A.: General characteristic-based algorithm for off-lattice
Boltzmann simulations. EPL (Europhys. Lett.) 75(3), 434 (2006)
2. Patil, D.V., Lakshmisha, K.N.: Finite volume TVD formulation of lattice Boltzmann simulation
on unstructured mesh. J. Comput. Phys. 228(14), 5262–5279 (2009)
3. Li, Y., LeBoeuf, E.J., Basu, P.K.: Least-squares finite-element scheme for the lattice Boltzmann
method on an unstructured mesh. Phys. Rev. E 72(4), 046711 (2005)
4. Shi, X., Lin, J., Yu, Z.: Discontinuous Galerkin spectral element lattice Boltzmann method on
triangular element. Int. J. Numer. Methods Fluids 42(11), 1249–1261 (2003)
5. Düster, A., Demkowicz, L., Rank, E.: High-order finite elements applied to the discrete
Boltzmann equation. Int. J. Numer. Methods Eng. 67(8), 1094–1121 (2006)
6. Min, M., Lee, T.: A spectral-element discontinuous Galerkin lattice Boltzmann method for
nearly incompressible flows. J. Comput. Phys. 230(1), 245–259 (2011)
7. Musavi, S.H., Ashrafizaadeh, M.: Meshless lattice Boltzmann method for the simulation of fluid
flows. Phys. Rev. E 91(2), 023310 (2015)
8. Wolf-Gladrow, D.A.: Lattice-Gas Cellular Automata and Lattice Boltzmann Models: An
Introduction, vol. 1725. Springer, Berlin (2000)
9. Liu, G.R., Gu, Y.T.: An Introduction to Meshfree Methods and Their Programming. Springer
Science and Business Media, Dordrecht/New York (2005)
A Comparison Between Two
and Three-Dimensional Simulations
of Finite Amplitude Sound Waves in a Trumpet
Janelle Resch, Lilia Krivodonova, and John Vanderkooy
Abstract Simplifying a three-dimensional problem of simulating sound propaga-

tion in musical instruments is frequently done by exploiting axial symmetry and
reducing the problem to one or two dimensions. We examine if such dimension
reduction is valid. We numerically solve the equations of motion of compress-
ible gases using the discontinuous Galerkin method to model nonlinear sound
propagation inside a trumpet. The numerical results in two and three dimensions
are then compared with experimental data. Experiments were carried out on a
trumpet in which the sound pressure waves of the Bb3 and Bb4 notes played at forte
were recorded. We found that it is crucial to consider the problem with all three
spatial dimensions to ensure reflections in the bell region are properly modelled.
Additionally, the shape of flare must be carefully approximated to compute the
propagating waves accurately.
1 Introduction
In order to accurately model sound propagation in brass musical instruments, higher

amplitude propagating waves must be considered. For musical instruments such
as the trombone or trumpet, pressure variations inside the instrument can be a
significant fraction of atmospheric pressure if loud, high frequency notes are played.
Assuming the pressure disturbance entering the narrow tubing of the bore is large
enough, nonlinear behaviour can cause waveform distortion. In particular, the crest
of the wave will travel faster than the trough causing the wave to steepen. Wave
steepening will excite higher harmonic components of the sound pressure waves
which influences the timbre of the sound by giving it a more ‘brassy’ effect [3]. The
J. Resch () • L. Krivodonova

Department of Applied Mathematics, University of Waterloo, 200 University Avenue W.,
Waterloo, ON N2L 3G1, Canada
J. Vanderkooy
Department of Physics and Astronomy, University of Waterloo, 200 University Avenue W.,
Waterloo, ON N2L 3G1, Canada

482 J. Resch et al.
linear acoustic equations used to model sound propagation outside an instrument

are not suitable to model the sound propagation inside. Instead, wave propagation
inside of the instrument should take compressibility into account which can be
modelled using the Euler equations. We solve these equations numerically using the
discontinuous Galerkin method (DGM). This method is particularly useful because
it can handle unstructured meshes and has excellent dissipative and dispersive
properties. In certain regimes (not considered here) shock waves can be produced
inside an instrument and this method can handle such nonlinear effects [2].
Previously, in [5], we modelled nonlinear wave propagation in a 2D trumpet. In
addition, we considered the consequences of neglecting the third spatial dimension
since the spreading of the waves in 2D and 3D differs. We calculated that the
2D – 3D dimensionality difference (which we call the dimension factor) in our
results would be approximately 14 dB. We arrived at this value by assuming that
the axial pressure in both 2D and 3D is a good measure of the total energy leaving
the bell. A full derivation of the dimension factor value can be found in [5]. After
taking this amplitude difference into account, we obtained a good match between
the experimental and numerical data, for the frequency components that propagate
out of the trumpet bell.
Here we present some of our preliminary results for extending the model to 3D.
In particular, we will be able to verify and further investigate the influence that
spatial dimensions have on such models. This is an important aspect to consider
since many mathematical descriptions of sound wave propagation in the literature
are reduced to 2D or even 1D.
2 Experimental Data and Numerical Setup
The experimental data that we collected from a Bb Barcelona BTR-200LQ trumpet

shown in Fig. 1 is discussed in detail in [5]. We took one period of the sound
pressure measurements obtained from a quarter-inch microphone carefully mounted
Fig. 1 Placement of microphones on the Barcelona BTR-200LQ trumpet

A Comparison Between 2D and 3D Simulations of Finite Amplitude Waves in. . . 483
on the shank of the trumpet mouthpiece, and applied Fourier analysis to the data
to reconstruct the waveform as a sum of 30 cosine waves. This experimental data
obtained at the mouthpiece was then used as a boundary condition on the pressure
for our simulations. The musical notes recorded and then simulated were the Bb3 and
Bb4 played at forte. A half-inch microphone was also placed outside the trumpet on
the central axis about 17 cm away from the bell. We compared our simulation results
with the pressure measurements obtained from this microphone outside the bell.
For all simulations, we assumed that the initial flow was at rest and related
pressure and velocity through the 1D expression derived from linear acoustic theory.
Since there is little return from the reflections at the bell, at least for the higher
frequencies, the plane wave velocity expression is a reasonable approximation. For
the boundary conditions, we prescribed reflecting boundary conditions on the inner
and outer walls of the trumpet mesh (excluding the mouthpiece). Since we modelled
a trumpet in a box, we used pass-through boundary conditions on the computational
domain.
3 The Influence of the Geometry and Dimension of the Bell
Several numerical simulations have been carried out to determine if the model
considered is adequate. Some such computational experiments are discussed in
[5]. We have discovered that considering the proper geometry of brass musical
instruments is critical to simulate wave propagation with any accuracy. We have
also investigated the importance of the bell geometry. One of the most significant
properties of the bell is how it influences the reflections of the sound pressure waves.
The location at which the harmonic waves reflect in the bell is dependent on their
frequency. For higher frequency waves, a larger portion of the energy will be lost or
be completely transmitted from the bell [1].
In the literature, it is common for the trumpet flare to be approximated by several
different functions or combination of functions, e.g., exponential functions, hyper-
bolic functions, etc. However, we were uncertain if such geometric descriptions
would be sufficient to accurately model wave propagation through the trumpet bell.
If the propagating sound pressure waves are particularly sensitive to the bell’s
curvature, simulations may produce exaggerated discrepancies if the bell shape
is poorly approximated. Simulations performed in 2D should be relevant for the
frequency components that are transmitted from the bell. Therefore, to reduce run
time, we first examined our problem by neglecting the third spatial dimension and
focused on the flare shape.
The first mesh, which will be called mesh 1, represented a 2D trumpet where
the bell was approximated by an elliptic function. To improve this approximation,
bell measurements were taken at several locations along the flare, which was then
interpolated by the obtained set of points and cubic splines. The measurements were
484 J. Resch et al.
done at the university’s machine shop by sampling the diameter of the bore near
expansion, 1.02 m from the mouthpiece. This flare shape was used to construct a
2D trumpet mesh called mesh 2. To obtain more precise points to approximate the
trumpet shape, a photo of the trumpet flare was taken. The grabit software from
Math Works Inc. was used to accurately trace out the shape of the trumpet bell. This
third bell shape was used to create a 2D mesh which we will call mesh 3. Since this
representation of the bell was most accurate, it was further used to extend mesh 3
into a full 3D mesh. This 3D mesh will be referred to as mesh 4. The total number
of cells is 8190, 8467, 8038 and 14,362 for mesh 1, mesh 2, mesh 3 and mesh 4,
respectively. The meshes consist of triangular or tetrahedral elements with adaptive
sizes to accurately resolve the geometric features of the trumpet. More details of
the mesh construction can be found in [5]. The computational domain and a sample
mesh are shown in Fig. 2.
Once these flare shapes were obtained, simulations were carried out on the
meshes of a trumpet that is 1.48 m in total length. The diameter of the bore is
constant except near the end of the instrument where it slowly increases and then
flares rapidly to give the bell shape. This expansion for all meshes begins 1.02 m
from the mouthpiece end. In [5], we justified that the bends of the instrument do not
Fig. 2 Mesh 3: Two-dimensional computational domain and trumpet mesh constructed by tracing
out the shape of the bell in Matlab
Fig. 3 Initial and reflected wave of simulated pressure pulse through mesh 1, mesh 2, mesh 3 and
mesh 4
greatly influence the wave propagation in the instrument, especially in comparison

to the bell. Therefore, for all the simulations discussed, the bends will not be
modelled, i.e., the bends and coils of the trumpet will be unwrapped and the tubing
will be straightened out.
Before simulating the musical notes however, we considered a simpler problem.
We sent a pressure pulse down the trumpet meshes from the mouthpiece to examine
their acoustic behaviour. As expected, the variations in the bell geometry greatly
influenced wave propagation. The pulse generated at the mouthpiece was given by
2
1:0 C .0:01 0:01 cos.1500t//; if t < 1500
pD
1:0; else
which corresponds to a unipolar pressure pulse with an amplitude of approx-

imately 2000 Pa. In Fig. 3, we show the pressure simulated at a point located at
the mid-length of the cylindrical bore. The first peak located at approximately
t D 0:0025 corresponds to the initial pulse moving from the mouthpiece towards
the bell. The second, inverted peak seen at about t D 0:007 corresponds to the
signal travelling back to the mouthpiece after it has been reflected by the bell.
In Fig. 4, we also plotted the reflected transfer data which is calculated from the
frequency content of the reflected pulse divided by that of the incident pulse. This
curve represents the power reflected by the bell meshes.
486 J. Resch et al.
Fig. 4 Corresponding reflected transfer function of pressure pulses depicted in Fig. 3
Table 1 The number of cells, and the frequency components that are mostly confined to the
instrument or transmitted from the bell are summarized for each mesh
Mesh name Cell count Reflected freq. (Hz) Transmitted freq. (Hz)
Mesh 1 8190 600 800
Mesh 2 8467 600 1000
Mesh 3 8038 400 1000
Mesh 4 14;362 900 1200
We observe in Figs. 3 and 4 a rather significant difference in the shape of the

returning pressure pulse for each mesh considered, especially when comparing
the 2D and 3D results. We found that different frequency components were
confined to the instrument and transmitted from the bell for each mesh. This was
accomplished by examining the relationship between the reflected transfer function
and its corresponding compliment, the transmission transfer function. Frequency
components less than approximately 550 Hz for mesh 1, 600 Hz for mesh 2, 400 Hz
for mesh 3, and 900 Hz for mesh 4, are mostly reflected before or within the bell [5].
For frequency components greater than 800 Hz for mesh 1, 1000 Hz for mesh 2 and
mesh 3, and 1200 Hz for mesh 4, the waves are being mostly transmitted from the
bell. The properties and number of cells for each mesh is summarized in Table 1.
Fig. 5 Two-dimensional numerical and experimental results for the Bb3
Further simulations were then carried out using the measured Bb3 and Bb4 pressure
waveforms. In particular, we solved the full 2D and 3D sets of compressible Euler
equations with the initial and boundary conditions mentioned above on the meshes.
Each note was initialized with 30 harmonics.
The frequency spectra of the 2D simulation results for the Bb3 and Bb4 notes
outside of the trumpet bell can be seen in Figs. 5 and 6, respectively. Figures 5
and 6 also depict the frequency spectra of the experimental data. A comparison
of the numerical and experimental data shows that the numerical amplitude is
approximately 19C dB off, where 14 dB are expected since the third spatial
dimension is neglected (see analysis in [5]). However, we are not certain why we
obtained an additional decibel difference of roughly 4C dB. As we will discuss in
the next section, a possible explanation is that the subtleties of the bore geometry
near the mouthpiece were not considered. Energy losses were also neglected and
some conjecture that this could make a difference of several decibels (up to 6 dB)
[4]. Nonetheless, to easily compare the harmonic distribution of the experimental
and 2D numerical data presented in Figs. 5 and 6, we shifted the experimental data
by the decibel difference stated in the plots.
488 J. Resch et al.
Fig. 6 Two-dimensional numerical and experimental results for the Bb4
We can see in Figs. 5 and 6 that the performed simulations on mesh 1, i.e., the
2D trumpet mesh where the bell is approximated by an elliptic function, gives the
poorest results in shape and amplitude. The 2D simulation results carried out on
mesh 2 and mesh 3 however show improvement in both aspects, i.e., in amplitude
and harmonic distribution. In particular, the spectral data aligns relatively well
with the shifted experimental curve, with the exception of the lowest frequency
components. The first three harmonics for the Bb3 and the fundamental frequency
for the Bb4 deviate most from the experimental data. However, we postulate that the
lowest frequencies are mostly confined within the instrument and consequentially,
are greatly influenced by the reflections that take place near the bell. We also
predicted that we would observe some amplitude differences and harmonic distri-
bution discrepancies from examining the pulse results. We further hypothesized that
the experimental and numerical data will better match (specifically for the lowest
harmonic components of the notes) if the equivalent simulations were carried out in
3D, i.e., on mesh 4.
Fig. 7 Three-dimensional numerical and experimental results for the Bb3
The frequency spectra of the measured Bb3 and Bb4 pressure waveforms outside
the trumpet bell are plotted in Figs. 7 and 8, respectively. The result of solving
the full 3D system on mesh 4 is also plotted for the Bb3 and Bb4 notes in Figs. 7
and 8, respectively. Overall, we see that the 3D simulations show greater similarity
to experimental data. There are however still some discrepancies, particularly in
the amplitude of the 3D solution curves. The Bb3 and Bb4 notes are overestimated by
approximately 10 dB. For comparative purposes, we again shifted the experimental
curves by the amplitude difference. When this shift is considered, the shape of the
experimental data and numerical solutions are in better agreement compared to the
results in Figs. 5 and 6. More specifically, the resulting lower frequency components
for both simulated notes demonstrate that the bell reflections are more accurately
modelled when the third spatial dimension is considered. In the next section, we
conjecture on possible sources for the 10 dB difference obtained in our 3D numerical
simulations.
490 J. Resch et al.
Fig. 8 Three-dimensional numerical and experimental results for the Bb4
4 Discussion Regarding the Geometry of the Bore Near

the Mouthpiece
Neglecting subtleties of the geometry near the mouthpiece of the trumpet may be
an explanation for the observed amplitude discrepancies in our simulation results.
In reality, the tube of the trumpet near the mouthpiece does not maintain constant
radius. The radius of the trumpet bore from the mouthpiece slowly increases for
approximately 22 cm. The tubing then remains roughly at a constant radius until the
bore begins to widen at 102 cm from the mouthpiece. Furthermore, the geometry
of the trumpet mouthpiece itself is also quite complex and varies in shape and
dynamics for each unique mouthpiece model. We tried to avoid some of these
potential effects by measuring the sound pressure waveforms at the shank of the
mouthpiece. However, the mouthpiece throat’s radius is approximately 0.4 the
shank’s radius; and the radius of the trumpet bore at 20 cm from the shank is
approximately 1.42 the shank’s radius. It is thus possible that neglecting these
geometric attributes of the trumpet may cause an overestimation in the simulated
amplitude.
We are currently investigating how the geometry of the bore between the
mouthpiece and the first bend of the trumpet influences the wave propagation being
modelled. We speculate that in addition to the geometric shape of the tube near
the mouthpiece, the rate at which the radius increases may be important to consider.
Just as the rapid flare of the trumpet influences the wave propagation, specifically the
harmonic reflections, we hypothesize that the increase in radius at the mouthpiece
may have a similar effect. We are also currently studying the influence of such
reflections near the mouthpiece.
5 Conclusion
We have presented our 2D and 3D simulation results of modelling sound production

in a trumpet. For comparative purposes, a plot of the experimental data and all
simulation results presented here for the Bb3 and Bb4 notes can be found in Fig. 9.
Although neglecting the third spatial dimension is frequently done, it is evident
that such an assumption will not accurately model the spreading of the waves in
all spatial directions. Furthermore, in Figs. 3, 4, 5, and 6, we can see that wave
propagation is very sensitive to the geometry of the flare, specifically where the
reflections at the bell are sensitive even to minor changes to its shape. We have
shown that it is not sufficient to merely approximate the bell geometry by a smooth
curve, it must be constructed with maximal agreement to physical geometry to
obtain good results.
Our 3D simulation results presented in Figs. 7 and 8 are significantly more
accurate than our 2D simulations. In addition, these findings have initiated an
investigation to determine the importance of the tube profile near the mouthpiece
of the trumpet on accuracy of the model. While the tube appears to be straight from
the outside, it has a complex geometry inside that cannot be inferred. We would need
to cut open an instrument to measure the true geometry and are currently looking
for a less invasive solution. Preliminary results indicate that better match of the
experimental and numerical data can be achieved.
Fig. 9 Numerical and experimental results for the Bb3 and Bb4
492 J. Resch et al.
References
1. Benade, A.H.: Fundamentals of Musical Acoustics. Dover Publications, New York (1900)
2. Flaherty, J.E., Krivodonova, L., Remacle, J.F., Shephard, M.S.: Some aspects of discontinuous
Galerkin methods for hyperbolic conservation laws. J. Finite Elem. Anal. Des. 38, 889–908
(2002)
3. Hirschberg, A.J., Gilbert, J., Msallam, R., Wijnands, A.P.J.: Shock waves in trombones. J.
Acoust. Soc. Am. 99, 1754–1758 (1995)
4. Kausel, W., Moore, T.: Influence of wall vibrations on the sound of brass wind instruments. J.
Acoust. Soc. Am. 128, 3161–3174 (2010)
5. Resch, J., Krivodonova, L., Vanderkooy, J.: A two-dimensional study of finite amplitude sound
waves in a trumpet using the discontinuous Galerkin method. J. Comput. Acous. (2012).
doi:10.1142/S0218396X14500076
A Dual-Rotor Horizontal Axis Wind Turbine
In-House Code (DR_HAWT)
K. Lee Slew, M. Miller, A. Fereidooni, P. Tawagi, G. El-Hage, M. Hou,

and E. Matida
Abstract This paper describes the DR_HAWT (Dual-Rotor Horizontal Axis Wind
Turbine) in-house code developed by the present authors, the National Renew-
able Energy Laboratory (NREL) Phase VI validation test case, and a dual-rotor
configuration simulation. DR_HAWT uses a combination of the Blade Element
Momentum theory (BEM) and the vortex filament method to accurately predict
the aerodynamic performance of horizontal axis wind turbines for a single or dual-
rotor configuration. Using two-dimensional (2D) airfoil coefficients obtained from
XFOIL, the aerodynamic power prediction produced by DR_HAWT resulted in an
average error of less than 20 % when compared to the NREL Phase VI experiment.
Upon correcting the 2D wind tunnel aerodynamic coefficient data for 3D flow
effects, good agreement was obtained with the exception of the deep-stall region.
In addition, a dual-rotor test case consisting of a combination of a half-scaled and
full-sized NREL Phase VI rotor, is presented. The dual-rotor configuration resulted
in a 16 % average increase in power generation as compared to the full-sized rotor
alone.
1 Introduction
With recent advancements in wind energy technology, and the increase in oil prices,
wind energy is now considered to be a favourable green-energy alternative [1]. In an
attempt to maximize energy extraction while minimizing costs, a Contra-Rotating
Dual-Rotor Horizontal Axis Wind Turbine configuration has been proposed. It is
hypothesized that by having both an upwind rotor and a downwind rotor, more
power can be produced on a single wind turbine tower.
K. Lee Slew () • M. Miller • E. Matida

Carleton University, 1125 Colonel By Drive, Ottawa, ON K1S 5B6, Canada
A. Fereidooni
National Research Council Canada, 2320 Lester Road, Ottawa, ON K1V 1S2, Canada
P. Tawagi • G. El-Hage • M. Hou
ZEC Wind Power, 1800 Woodward Drive, Ottawa, ON K2C 0P7, Canada

494 K. Lee Slew et al.
Due to significant improvements in computational technology and knowledge of

Horizontal Axis Wind Turbine (HAWT) aerodynamics, computational simulation
tools have become a preferred method in the aerodynamic performance prediction of
wind turbines. Commercially available CFD packages offer accurate aerodynamic
predictions whilst capturing the complex flow phenomena that occur with HAWTs,
however, these programs are typically computationally expensive. For this reason,
Blade Element Momentum theory (BEM) is often used by the wind energy industry
as a preliminary aerodynamic performance prediction tool.
BEM is computationally inexpensive and yields accurate results provided accept-
able airfoil aerodynamic data is available, however, BEM cannot capture all the flow
phenomena which occurs in the wake of a HAWT. To reduce runtimes compared
to CFD, while providing a more accurate representation of the wake, an in-house
code, utilizing a grid-less technique, for Dual-Rotor Horizontal Axis Wind Turbines
(DR_HAWT) was created by the present authors based on a combination of BEM
[1] and the vortex filament method [2, 3]. DR_HAWT possesses the ability to predict
the aerodynamic performance of a single or dual-rotor configuration as well as
parallel processing capabilities.
To verify and validate DR_HAWT, the National Renewable Energy Laboratory
(NREL) Phase VI wind turbine was simulated. The experiment was performed in
the NASA-Ames 24:4 m 36:6 m wind tunnel using a two-bladed, 10-m-diameter
HAWT. The tapered and twisted blades utilize the S809 airfoil; further geometric
details can be found in Ref. [4]. All comparisons in this paper, unless stated
otherwise, are to the aerodynamic power output of the Sequence S experiment [4].
This paper will briefly describe the BEM and vortex filament methods imple-
mented in DR_HAWT including the 3D correction factors used to account for the
effect of rotation. The NREL validation test case will be presented as well as a
dual-rotor HAWT configuration.
2 Methodology
2.1 Blade Element Momentum Theory
The well documented BEM theory [1] is a combination of Blade Element theory and
the Momentum theory. The wind turbine blade is discretized into elements along the
blade span, where geometric data are known. At a specified Reynolds number (Re),
velocity and angle of attack (AOA), the coefficient of lift and drag (CL and CD ,
respectively) can be obtained from look-up tables for each specific airfoil. Using
the aerodynamic coefficients, the tangential and normal forces acting on the blade
elements can be determined. These forces can then be integrated along the span of
the blade to determine the rotor torque, thrust and power.
The wake of the wind turbine is commonly accounted for in the BEM method
by induction factors. Due to the extraction of energy from the wind flow, these
Dual-Rotor Horizontal Axis Wind Turbine Code (DR_HAWT) 495
induction factors decrease the velocity of the flow in the axial and tangential
direction by means of a semi-empirical model. The BEM method, combined
with various other correction factors such as Prandtl’s tip loss factor, provides
an adequate aerodynamic prediction of single rotor configuration of HAWT [1].
However, for a dual-rotor configuration, where the influence of the upwind and
downwind rotor wake is an intrinsic aspect of the aerodynamic performance
prediction, a better means of representing the wake is necessary.
2.2 Vortex Filament Method
In order to determine the influence of the wake on the aerodynamic performance

of a HAWT, particularly in a dual-rotor configuration, a vortex filament method has
been adapted from Strickland et al. [2, 3]. In this three-dimensional (3D), grid-less
free vortex model, vortex filaments are used in place of the induction factors in the
BEM method. These vortex filaments are free to convect in the fluid domain thereby
providing a suitable spatial and temporal representation of the wake.
Similar to the BEM method, the blade is discretized into elements where the
geometric characteristics are known. Each element is defined by unit vectors in the
spanwise (s), chordwise (c), and normal (n) directions, located at the aerodynamic
center of each element. In addition, the position of the aerodynamic center of each
element and the shed vortices are defined in a global coordinate system (i, j, and
k axes) located at the center of the upwind hub. Figures 1 and 2 illustrate the
coordinate systems used in DR_HAWT.
With respect to these coordinate systems, the relative velocity, Vrel , seen by a
blade element as shown in Fig. 2, can be evaluated by vector sum of the blade
rotational velocity, VT , free stream velocity, Vo , and the induced velocity caused by
the vortex filaments, Vind . The local flow conditions (Re, AOA) are calculated and,
Fig. 1 Global coordinate system used in DR_HAWT

Fig. 2 Velocity components acting on a blade element
the lift and drag coefficients are linearly interpolated from look-up tables consisting
of experimental or numerical airfoil aerodynamic data.
Based on the Lifting Line Theory, each blade element is associated with a bound
vortex filament of strength,
B . This strength can be expressed using the Kutta-
Joukowski theorem as follows:
1

B D CL cjVrel j (1)
2
The bound vortex strength is a function of the CL , the chord length (c) and the
magnitude of the relative velocity. Since the relative velocity takes into account the
induced velocity, a predictor-corrector method is employed for which the bound
vortex strength is predicted based on the previous time step induced velocities.
The predicted strength is then used to correct the induced velocities which in turn
corrects the bound vortex strength. This process is repeated until the differences
between the predicted and corrected values are below 1 103 . Once the bound
vorticity and the induced velocities have been corrected, the strength of the vortices
that constitute the wake can be computed.
Figure 3 illustrates the representation of blade elements with the wake which is
made up of vortex filaments. For a time step NT, the bound vortex strength of an
element i is written as
NT i
. As the blade rotates, vortices are shed in the form of
spanwise and trailing vortex filaments parallel to the span and chord of the blade
respectively. The strength of these vortex filaments are calculated based on Kelvin’s
Theorem for which any change in the bound vortex strength over time must be
accompanied by an equal and opposite vortex strength in the wake. This forms the
spanwise vortices (bold lines) whereas the trailing vortices (thin lines) result from
the spanwise variation in bound vortex strength. The tip vortex is assumed to have
the same magnitude and opposite direction as the bound vortex closest to the tip.
Each vortex filament in the wake convects with a local velocity. As the vortex
filament first leaves the blade, its initial velocity is assumed to be equal to the sum
of the induced velocities at the blade element end and the oncoming freestream
Fig. 3 Spanwise and trailing vortices being shed from a blade
velocity. In order to calculate the distance travelled by any vortex filament in the
wake for a given time step, a second order Adams-Bashforth explicit integration
formula is used. Knowing the strength and position of the shed vortices, the induced
velocity at any point can be calculated using the Biot-Savart Law given by:
Z

r dl
Vind D (2)
4 l jrj3
where r is the position vector from the point of interest in the fluid domain to the
incremental length (dl) along the vortex filament. The total perturbation velocity
induced by all the vortex filaments represents the effect of the wake on the blades.
In order to calculate the torque, thrust and power generated by the blades, the normal
and tangential forces are integrated along the span of the blade by assuming linear
variation along each element [1]. Finally, the aerodynamic power produced by the
wind turbine is the product of the rotational velocity and the total torque of the rotor.
2.3 Aerodynamic Coefficients
This study uses the 2D (non-rotating) airfoil coefficient data (CL and CD ) obtained
either by using experimental wind tunnel data, acquired from Ref. [4] or XFOIL
which numerically calculates the lift and drag coefficients for a given airfoil profile
based on pressure distributions obtained using a panel-method [5]. For comparison
purposes, the same numeric values of Reynolds numbers are used in both cases
for the aerodynamic coefficient input files. In addition, the experimental data was
corrected to obtain a better understanding of 3D flow effects due to rotation.
In an attempt to model these 3D effects, several numerical correction factors
have been formulated by the likes of Du and Selig [6], Snel et al. [7], Bak et al. [8],
Lindenburg [9], and more. These various models all simply correct the 2D CL and
CD wind tunnel data using the following equations.
CL;3D D CL;2D C f .c=r; : : :/ CL (3)

CD;3D D CD;2D C f .c=r; : : :/ CD (4)
where f is the correction factor which in all cases is a function of at least the chord
to radial (c=r) location ratio. CL and CD is the difference between the 2D CL and
CD wind tunnel data and the 2D CL and CD if hypothetically the flow remained
attached at all AOA. Some 3D correction formulations, such as that of Du and
Selig, have empirical correction factors which are built into f . The fine tuning of
the 3D corrections is only viable when experimental 3D data, like the data obtained
from the NREL experiment, is available. Obtaining a suitable universal empirical
correction factor for the 3D corrections should be performed by simulating several
different experiments.
The 3D correction factors of Snel et al. and Du and Selig were implemented in
DR_HAWT for AOA up to 30ı , at which point the corrected values linearly decay
to the 2D value at 55ı as per Ref. [10]. Figure 4 illustrates an example of the Snel
et al. 3D corrections applied to the CL of the S809 airfoil at a Reynolds number
of 6:5 105 . It should be noted that the Snel et al. correction only corrects the CL
whereas Du and Selig corrects both the CL and CD .
Fig. 4 CL curves obtained by XFOIL, 2D wind tunnel data and 3D corrections (Snel et al.) for a
given Reynolds number
2.4 Convergence Analyses
Prior to simulating the Sequence S of the NREL Phase VI experiment, a sensitivity

analysis was performed to determine the effect of changing the number of elements,
time increment, and number of time steps in DR_HAWT for the performance
prediction of the NREL Phase VI wind turbine. Default values of 26 discretized
blade elements used to define the geometry of each blade in Ref. [4], a time
increment of 0.03145 s and 200 time steps were both halved and doubled. The
outcome of this study revealed that the default values were adequate for this
simulation.
Typically, converged values of power, torque and thrust for a given oncoming
wind speed are obtained after the wake has travelled 3–4 diameters downstream.
Correspondingly, a single-rotor configuration for 200 time steps with a time
increment of 0.03145 s, has a simulation runtime of approximately 60 min using
an i7 3.60 GHz processor with 16 GB of RAM; even faster runtimes can be obtained
if the code is run with parallel processing. This highlights DR_HAWT’s ability to
run parametric studies of a single or dual-rotor configuration in a timely manner
compared to existing commercially available CFD packages.
Figure 5 illustrates the output power predicted by DR_HAWT using aerodynamic

coefficients from XFOIL, wind tunnel (2D) experiments, wind tunnel data corrected
for 3D flow effects using Snel et al. and Du and Selig, as well as the UAE
experimental data. Based on XFOIL data, DR_HAWT resulted in an average
Fig. 5 Aerodynamic power prediction produced by DR_HAWT

error below 20 % when compared with the experimental data, which is within
the measurement error bars [11]. With the 2D wind tunnel data, however, large
differences in excess of 100 % were obtained in the stall region (10–25 m/s).
The large variation in aerodynamic power curves indicate that the input aero-
dynamic coefficient is an influencing factor in the output results produced by
DR_HAWT. As seen in Fig. 5, this mainly affects the outcome in the stall region
(10–25 m/s) where separation and 3D flow effects are present on the blade.
Due to the rotational nature of the wind turbine, the flow over the wind turbine
blade is not 2D but it is in fact 3D which has been found to delay stall. The pressure
gradient due to rotation of the blade invokes flow in the radial direction which alters
the boundary layer, hence resulting in higher lift and lower drag values. For this
reason, wind turbines generally will obtain higher power values than predicted when
using 2D wind tunnel airfoil data alone [10]. Although closer to the experimentally
obtained aerodynamic power, particularly in the pre-stall (below 7 m/s) and stall
regimes (7–15 m/s), the 3D corrections still do not produce favourable results when
deep-stall is prominent (above 15 m/s). According to Vermeer et al., the use of 3D
corrections at high wind velocity is questionable [12].
No matter if XFOIL, 2D wind tunnel or 3D corrected aerodynamic coefficient
data is used for the simulations, an over prediction in power up to 7 m/s wind
speeds can be observed. The consistency in power values obtained by the various
simulations is due to the fact that the majority of the blade is in the linear, pre-
stall, region of the lift and drag coefficient curves. As illustrated in Fig. 4, this
linear region does not vary substantially between the different methods of obtaining
the aerodynamic airfoil data since the flow is attached. DR_HAWT was found to
consistently over predict the tangential force compared to experimental values [11].
Under the assumption that the CL and CD are not contributing factors in the pre-
stall region, this discrepancy must result from an inaccuracy in the method used
to determine the AOA. The challenge of predicting the AOA was also identified
by Lindenburg [10], Sant et al. [13] and Jonkman [14] suggesting that further
investigation into the methods used by DR_HAWT is required.
4 Dual-Rotor Configuration
In order to examine the effects of adding a second rotor to the HAWT, a half-
sized, geometrically scaled NREL Phase VI rotor and an unmodified NREL UAE
Phase VI rotor were simulated in the upwind and downwind positions, respectively.
The upwind rotor had a rotational velocity twice that of the unmodified rotor in
order to maintain similar tip speeds. The rotors were separated coaxially by half the
diameter of the unmodified rotor. The contra-rotating dual-rotor configuration was
simulated using XFOIL 2D data for wind speeds ranging from 5 to 25 m/s. As can
be seen in Fig. 6, the addition of the upwind rotor resulted in an average increase
in power of 16 % over the single rotor configuration. The power generated by the
dual-rotor configuration is lower than that of the combined individual single rotors
by approximately 6 %. This is as expected since the presence of the wake from one
rotor onto another is typically detrimental to their performance. A visualization of
the dual-rotor configuration post-processed in ParaView [15] can be seen in Fig. 7.
Fig. 6 Aerodynamic power comparison of a single and dual-rotor HAWT configuration
Fig. 7 Wake visualization of a dual-rotor configuration where the dots represent the end points of
a vortex filament
5 Conclusion
An in-house code, named DR_HAWT was created by the present authors for the
aerodynamic performance prediction of a single or dual-rotor HAWT configuration.
Converged values of aerodynamic loads are obtained within approximately 60 min.
DR_HAWT was validated against the NREL UAE Phase VI experiment with
relatively good agreement. Power predicted by DR_HAWT using numerically
obtained data from XFOIL, fell within the experimental error bars and had an overall
error of less than 20 %. This adequate aerodynamic performance prediction and fast
computational run times facilitate parametric studies of a single or dual-rotor HAWT
configuration.
Further analyses revealed that the input lift and drag coefficient data have a large
influence on the predicted aerodynamic forces. 3D flow effects were found to have
a profound significance particularly in the stall region. 3D corrections formulated
by Snel et al. and Du and Selig were applied to 2D coefficient data and found to
show improvement with the exception of the deep stall region (above 15 m/s). A
contra-rotating dual-rotor HAWT configuration based on the NREL Phase VI rotor,
was found to produce 16 % more power than the single NREL Phase VI rotor alone,
demonstrating the potential benefits of a dual-rotor wind turbine configuration.
Acknowledgements The authors would like to thank NSERC (National Sciences and Engineering
Research Council of Canada) and OCE (Ontario Centres of Excellence) for the financial support
as well as Scott Schreck of the National Renewable Energy Laboratory for providing experimental
data for validation purposes.
References
1. Hansen, M.: Aerodynamics of Wind Turbines. EarthScan, London (2008)

2. Strickland, J., Webster, B., Nguyen, T.: A Vortex Model of the Darrieus Turbine: An Analytical
and Experimental Study. Sandia National Laboratories, Lubbock (1980)
3. Fereidooni, A.: Numerical Study of Aeroelastic Behaviour of a Troposkien Shape Vertical Axis
Wind Turbine – MASc Thesis. Carleton University, Ottawa (2013)
4. Hand, M., Simms, D., Fingersh, L., Jager, D., Cotrell, J., Schreck, S., Larwood, S.: Unsteady
Aerodynamics Experiment Phase VI: Wind Tunnel Test Configurations and Available Data
Campaigns. National Renewable Energy Laboratory, Golden (2001)
5. Drela, M.: XFOIL: An Analysis and Design System for Low Reynolds Number Airfoils. Dept.
of Aeronautics and Astronautics, MIT, Cambridge (1989)
6. Du, Z., Selig, M.S.: A 3-D Stall-Delay Model for Horizontal Axis Wind Turbine Performance
Prediction. Am. Inst. Aeronaut. Astronaut. 21, 9–19 (1998)
7. Snel, H., Houwink, R., Bosschers, J.: Sectional Predicition of Lift Coefficients on Rotating
Wind Turbine Blades in Stall. Energy Research Centre of the Netherlands, Petten (1994)
8. Bak, C., Johansen, J., Anderson, P.B.: Three-Dimensional Corrections of Airfoil Character-
istics Based on Pressure Distributions. In: European Wind Energy Conference & Exhibition,
Athens (2006)
9. Lindenburg, C.: Modelling of Rotational Augmentation Based on Engineering Considerations
and Measurements. In: European Wind Energy Conference, London (2004)
10. Lindenburg, C.: Investigation into Rotor Blade Aerodynamics: Analysis of the Stationary
Measurements on the UAE Phase-VI Rotor in the NASA-Ames Wind Tunnel. Energy research
Centre of the Netherlands, Petten (2003)
11. Schreck, S.: Sequence S Data. National Renewable Energy Laboratory via Private Communi-
cations, Golden (2015)
12. Vermeer, L., Sørensen, J., Crespo, A.: Wind Turbine Wake Aerodynamics. Prog. Aerosp. Sci.
39, 467–510 (2003)
13. Sant, T., van Kuik, G., van Bussel, G.: Estimating the Unsteady Angle of Attack from Blade
Pressure Measurements on the NREL Phase VI Rotor in Yaw using a Free Wake Vortex Model.
In: 44th AIAA Aerospace Sciences Meeting and Exhibition, Reno (2006)
14. Jonkman, J.: Modeling of the UAE Wind Turbine for Refinement of FAST_AD. National
Renewable Energy Laboratory, Golden (2003)
15. Ayachit, U.: The ParaView Guide: A Parallel Visualization Application. Kitware, Inc., Clifton
Park, NY (2015)
Numerical Study of the Installed Controlled
Diffusion Airfoil at Transitional Reynolds
Number
Hao Wu, Paul Laffay, Alexandre Idier, Prateek Jaiswal, Marlène Sanjosé,
and Stéphane Moreau
Abstract Reynolds-Averaged Navier-Stokes simulations have been carried out for

the self-noise study of a Controlled Diffusion airfoil at chord Reynolds number
of 1:5 105 . Two numerical setups of the anechoic open-jet facilities where
experimental data on aerodynamics have been collected using hot-wire and particle
image velocimetry measurements are investigated to capture installation effects and
to compare different turbulent models. The installation effects are observed between
different wind tunnel nozzle dimensions. Different turbulent models in simulations
have been compared for such flow case. The k!SST model shows better agreement
with the experimental data at a reasonable computational cost, providing a reliable
initialisation field and boundary conditions for future direct numerical simulation
studies.
1 Introduction
In the design process of a rotating machine such as an automotive engine cooling

fan, a wind turbine or an air-conditioning unit, one major evaluation index is the
noise level for a given loading. Where other noise sources can be reduced or
avoided by a careful design, the trailing-edge (TE) noise, or airfoil self-noise, is the
only remaining noise source when an airfoil encounters a homogeneous stationary
flow. Such a noise is generated by the interaction between the TE with pressure
fluctuations convecting in a boundary layer. If the boundary layer is laminar, tonal
noise will be introduced; if the boundary layer is turbulent, pressure fluctuations are
present over a wide range of frequencies, which leads to broadband noise radiation.
The study of TE noise received much attention mainly in the 1970s and early
1980s. Most of such studies during that time involved experimental measurements
of wall-pressure fluctuations and far field acoustic on various airfoils in open-jet
anechoic wind tunnels [1, 2]. In recent years, Computational Fluid Dynamics (CFD)
H. Wu () • P. Laffay • A. Idier • P. Jaiswal • M. Sanjosé • S. Moreau

Department of Mechanical Engineering, Université de Sherbrooke, Sherbrooke, QC J1K2R1,
Canada

506 H. Wu et al.
methods have become an important complementary part aside from the experimental
methods. In practice however, simplified flow configurations are often employed
in simulations whereas most trailing-edge aeroacoustics experiments have been
conducted in open-jet wind-tunnel facilities, where the airfoil is immersed in a jet
downstream of the nozzle exit. It has been seen that the proximity of the airfoil
to the jet nozzle exit and the limited jet width relative to the airfoil thickness
and chord length can cause the airfoil loading and flow characteristics to deviate
significantly from those measured in free air and hence, alter the radiated noise
field [3]. Installation effects thus take place for different jet configurations and
require simulations to model it.
As aerodynamics has direct influence on aeroacoustics, the present work aims at
presenting and quantifying the installation effects as well as studying the influence
of different turbulent models on such flow case simulation. Experimental and
numerical setups are firstly introduced with associated technical methods. Results
on a systematic CFD study, based on Reynolds-Averaged Navier-Stokes (RANS)
are then presented and are compared with flows data over the Controlled Diffusion
(CD) airfoil (a cambered airfoil originally developed at Valeo Motors and Actuators)
installed in the anechoic wind tunnels of Ecole Centrale Lyon (ECL) and of
Université de Sherbrooke (UdS). The evaluation of different turbulence models and
installation effects observed in simulations will be discussed. The results give a
synthesis of the previous RANS study and provide guidance for the appropriate
initial field needed in future direct numerical simulation (DNS) studies [4, 5].
2 Experimental Setup
All the measurements were done in an open-jet anechoic wind tunnel shown in
Fig. 1a. The airfoil mock-up (Fig. 1b) is held between two side plates in a 30
30 cm2 test section, to keep the flow two-dimensional. This setup is very similar to
that used in [3, 6] at ECL, of which, the jet nozzle dimension is 50 25 cm2 . The
self-noise is the noise radiated by the turbulent eddies coming from the turbulent
boundary layer. The setup should be free of any additional noise source that would
make the self-noise. In particular, very low turbulence intensity is achieved by a
very high convergent ratio of about 1:25. The measured turbulence intensity is 0.3–
0.4 %. The wind tunnel is also acoustically treated to achieve low background noise.
According to the above mentioned criteria, it becomes clear that open-jet anechoic
wind tunnel is best suited for the study of airfoil self-noise [7]. The hot wire and
PIV measurements have been performed in the wake and pressure sensor probes on
the airfoil to detect airfoil loading, as described below.
Numerical Study of CD Airfoil at Transitional Re Number 507
Fig. 1 (a) Front view of the CD airfoil in the anechoic wind tunnel and (b) Close side-view
2.1 Wall Pressure Measurements
The mean wall pressure on the CD airfoil was measured using a Baratron capaci-
tance manometer which is connected to a pin hole on the surface of the airfoil using
a capillary tube. There are in total 21 probes, 18 of them are placed in streamwise
direction and rest 3 in spanwise direction. All the streamwise probes are placed at
the mockup mid-span as shown in Fig. 1b.
2.2 Velocity Measurements
The streamwise velocity in the wake was obtained using a TSI 1210-T1.5 single
hot-wire probe The length of the probe is 1.27 mm and a diameter of 3:8 m. The
displacement of the hot wire in the wake was realized using a Superior Electric
M062-FD03 Slo-Syn Stepping Motor. The minimum displacement of the hot wire
(the highest spatial resolution) in the near wake was 0.05 mm. The hot-wire probe
is connected to a Constant Temperature Anemometer System IFA 300. The data
acquisition was achieved using a National Instrument BNC 2090 system controlled
with Labview at the sampling frequency of 20 kHz. The relative measurement error
was calculated taking into account parameters listed in [8]. It was found to equal to
2.215 %.
Planar PIV (Particle Image Velocimetry) measurements have been performed
using a single LaVision sCMOS 5.5 megapixel camera with a pixel pitch of 6.5 m
and a Evergreen 70 mj ND:YAG laser [9]. The laser-sheet thickness was measured
to be about 2 mm. 700 images in double frame were recorded using sCMOS
camera fitted with Nikon 50 mm lens at an acquisition rate of 15 Hz. The image
508 H. Wu et al.
magnification was about 0.07. The particle image diameter and depth of field were
adjusted using lens aperture [9], chosen to be 11 for the current experiments.
This is done so that particle image diameter is greater than 2 pixels (calculated
by neglecting any lens aberrations). At this image diameter under sampling of
particle image which can cause peak locking can be avoided [10]. The depth of field
calculated was about 60 mm, which is much larger than the measured laser sheet
thickness so all the particles can be assumed in focus. Time between frames was
selected based on many considerations like minimization of loss-of-pair due out of
plane motion, truncation error due to constant velocity assumption of the particle
between the two frames and minimization of relative error on the displacement
estimate[11]. Finally it must be remembered that the dynamic velocity range of
PIV increases with time separation between the two frames [12]. Taking all these
considerations into effect, time between two frames was chosen to be about 45 s.
The free stream displacement particle between the two frames was measured to
be roughly about 7.5 pixels in the object plane. Taking the smallest resolvable
displacement fluctuation to be 0.1 pixel, the dynamic velocity range was calculated
to be about 75 [9]. The seeding density during the experiments was kept at about
0.055 ppp to ensure a more that 10 particles were present for the final interrogation
domain for a better signal-to-noise ratio in cross correlation [13]. The parameters
used for PIV measurements are summarized below in Table 1. All the results were
processed using DAVIS 8 software. To improve the accuracy in peak fitting and
sub-pixel accuracy, normalized cross correlation option in DAVIS 8 was selected
for the calculation of cross correlation. The final interrogation window size was
kept at 16 16 pixel with a 50 % overlap. An adaptive window shape was selected
for image cross correlation to take into account the effect of shear in the near
wake. The error analysis in PIV is a topic of active research and depends upon
many factors. In absence of detailed error analysis the value of error was chosen
to be equal to 0.1 pixel which corresponds to a typical error in measurement of
displacement in PIV [9]. On the other hand typical error on pulse separation has
an order of magnitude in nanoseconds while time between frames is in the order
Table 1 Parameters used in Parameters Value

the PIV measurement
Depth of focus
59 mm
Seeding density > 0:055 ppp
Mode Double frame
Frequency 50 Hz
Number of images 700
Window of interrogation 16 16
Focal length 57.41 mm
FOV 240 mm
Particle image diameter 2.33 pixels
Magnification 0.07
of microseconds hence relative error is very small, compared to relative error in

displacement.
3 Numerical Setup
3.1 Flow Conditions and Physical Models
The flow conditions for CD airfoil are a freestream velocity U0 of 16 m/s measured
at the wind tunnel nozzle exit away from the airfoil. The Mach number is 0.05 and
the Reynolds number is 1:5 105 based on the airfoil chord length c D 0:1356 m.
The flow is therefore turbulent or transitional. To model the turbulence in RANS
simulations, 4 turbulence models are tested: k (standard and with low Reynolds
number correction), k! SST, and tr k kl !. As the RANS results will
serve as the initialization field for future DNS study, the transitional model tr
k kl w is chosen for its capacity to capture transition process in the simulation.
Air is supposed to be an incompressible perfect gas. The geometric angle of attack
with respect to the wind tunnel axis is 8ı . To evaluate the airfoil loading and the
installation effects, the pressure coefficient Cp and the friction coefficient Cf are
introduced as shown respectively in Eq. (1),
Ps P0 n
Cp D 1 2
and Cf D 1 2
(1)
2 0 U0
2 0 U0
where Ps stands for the static pressure, n for the local wall-shear stress and all
the parameters with the sub-index 0 correspond to the reference values in this flow
case. The values of the results shown afterwards in the section Results are non-
dimensionalized based on c and U0 .
3.2 Numerical Parameters
To introduce the installation effects, the simulation domain includes the complete
nozzle geometry to assess the effect of the interaction between the jet shear layer
and the airfoil (Fig. 2). A 2D domain which represents the mid-section of the
Experimental Setup is generated by Centaur. Hybrid grids with a total 69,000 cells
are employed to get a balance between simulation accuracy and computational cost
according to what was reported in previous simulations [3] Quadrilaterals are used
to refine the grid close to the walls in order to capture the boundary layer around
the airfoil and the shear layer effect of the wind tunnel exit as shown in Fig. 2b,
c. The dimensionless wall-normal grid spacing in wall units yC is smaller than
1 over most of the chord length on both pressure and suction sides. The mesh for
510 H. Wu et al.
Fig. 2 Hybrid mesh (a); Zoom view of airfoil trainling-edge and nozzle exit (b) (c)
ECL installation is generated using the same methodology as in Fig. 2. The RANS
simulations are performed using ANSYS FLUENT-15. The pressure-based solver
provides two types of pressure-velocity coupling algorithms: either in a segregated
manner (SIMPLE) or in a coupled manner (Coupled). The results demonstrate
no difference in the wake velocity profiles or loading of the airfoil. The Coupled
scheme takes 2
3 times longer than SIMPLE. In the present paper, calculations
presented are performed using the SIMPLE scheme only.
4 Results
4.1 Effects of Turbulence Models

4.1.1 Velocity and Turbulence Kinetic Energy Field
All models are able to establish a converged wake zone except for the tr k kl !
model as shown in Fig. 3. Even from the established field obtained with the k!SST
model that can be seen as a reference from previous studies[3], the scaled residuals
of the tr k kl ! model oscillate around much higher levels, two orders of
magnitude higher than other models. With regards to turbulence, the k over-
predicts the turbulence kinetic energy (TKE) (Fig. 3). By using damping functions,
the k model with low Reynolds-number correction has been implemented in
ANSYS Fluent and shows better behaviour than the standard k model. Yet it
Fig. 3 Velocity (left) and turbulence kinetic energy (right) iso-contours
still over-predicts the turbulence over the whole profile. The k! SST model only
sees the transition process to turbulence close to the trailing edge. The tr k
kl ! has unphysical boundary-layer development at the TE which is caused by
high numerical instabilities. To make a fair comparison, the simulations for tr
k kl ! have then been conducted in a URANS (Unsteady Reynolds Averaged
Navier-Stokes) mode and then averaged to get a steady solution.
4.1.2 Airfoil Loading
The tr k kl ! model averaged data is closer to the mean wall-pressure

measurement in the TE region and has similar prediction capacity as the k! SST
model at other positions along the airfoil (Fig. 4). Compared with the k! SST
model or other k! based models, the tr k kl ! model includes three transport
equations to solve the turbulent kinetic energy (kt ), the laminar kinetic energy
(kl ) and the scale-determining variable ! where ! D =kt , the ratio of isotropic
dissipation and the turbulence kinetic energy respectively. In the ! equation, a
transitional production term is introduced to produce a reduction in turbulence
length scale during the transition breakdown process, which essentially helps
predict the magnitude of low-frequency velocity fluctuations in the pre-transitional
512 H. Wu et al.
-Cp 0
k-ε
k-ε Low-Re
kω-SST
tr-k-kl-ω (RANS)
tr-k-kl-ω (URANS)
exp
-1
-1 -0.8 -0.6 -0.4 -0.2 0
x/c
Fig. 4 Comparison of mean wall-pressure coefficient
boundary layer that have been identified as the precursors to transition [14]. As the
numerical transitional process at UdS appears around TE, this model has shown its
advantage on capturing such fluctuations compared with other tested models. The
unsteadiness appearing in the tr k kl ! RANS simulation at the TE can be
potentially verified in the future by measurements of wall-pressure fluctuations.
4.1.3 Velocity Profile in the Wake Region
The wake velocity profile is a consequence of the transition process on the airfoil.
Here the velocity magnitude from different models are compared with both hot-wire
and PIV measurements in Fig. 5. Simulation results tend to over-predict the velocity
deficit at all measuring positions but are consistent with their respective production
of turbulence. Some of the discrepancies can be attributed to the isotropic hypothesis
on calculating the turbulence kinetic energy made in the RANS simulation. The k
with low Reynolds number correction model improves the result from the standard
k model by nearly 10 %. Yet, the k!SST model shows better overall comparison
with the experimental data.
In summary, for RANS simulations of such a flow case, the tr k kl !
model must be run in an unsteady mode and its mean solution is similar to that
of the k! SST model with an airfoil loading even slightly better compared
with experiment. The k! SST model provides the best overall prediction among
the tested models considering its reasonable computational time. It is thus used
as a reference model for future comparisons. By using damping functions, the
(a) k-ε
(b) (c)
k-ε Low-Re
kω-SST
0.1 tr-k-kl-ω (RANS) 0
tr-k-kl-ω (URANS)
HW
PIV
0
y/c y/c y/c

0 -0.1
-0.1
-0.1 -0.2
0.5 1 0.5 1 0.5 1
Ut/U0 Ut/U0 Ut/U0
0.1 0
y/c y/c y/c

0 -0.1
-0.1
-0.1 -0.2
0 0.02 0.04 0.06 0.08 0 0.01 0.02 0.03 0.04 0.05 0 0.005 0.01 0.015 0.02 0.025 0.03
2
2 2
k/U0 k/U0 k/U0
Fig. 5 Wake velocity and TKE profiles. (a) x=c D 0:07. (b) x=c D 0:22. (c) x=c D 0:44
two-equation k model with low Reynolds number correction shows better

behaviour in wake velocity profiles than the the standard k model.
4.2 Installation Effects
As firstly mentioned by Moreau et al. [3], installation effects are significant between
an isolated airfoil case and one that is installed in an open-jet wind tunnel. In this
paper, the effects of nozzle jet width is specifically investigated by comparing results
in two different setups at ECL and UdS respectively.
The loading of the airfoil is changed as can be seen from the wall pressure
coefficient in Fig. 7. This is mostly caused by a different transition process. At
ECL, a separation bubble turns the laminar boundary layer into a turbulent one
at the leading edge, which is hardly the case in the UdS set-up for which, the
transition process is less severe. Indeed Fig. 6 shows that the friction coefficient
Cf is barely negative in the UdeS case and the laminar recirculation bubble is hardly
formed. Because of the more limited nozzle jet width at UdS, the boundary layer
514 H. Wu et al.
0.2
RANS-UdS
RANS-ECL
Leading-edge area
0.15 0.03
0.02
0.1
Cf 0.01
0
0.05
-0.01
-1 -0.95 -0.9
-1 -0.8 -0.6 -0.4 -0.2 0

x/c
Fig. 6 Mean friction coefficient: global and zoom in the leading-edge area
2
RANS-UdS
RANS-ECL
Exp-UdS
Exp-ECL
1
-Cp
-1
-1 -0.8 -0.6 -0.4 -0.2 0
x/c
Fig. 7 Mean wall pressure coefficient
at the leading edge develops in a different way as a consequence of the larger flow
confinement. Besides, there is a difference of 15 % on Cp at the TE in Fig. 7 between
the experimental data and the simulation results at UdS as no flow separation is
captured in this simulation. Consequently, the RANS simulations have a hard time
predicting the transition process correctly, and the flow separation either at the
leading edge (laminar recirculation bubble) or at the TE (driven by the adverse
pressure gradient).
5 Conclusion
RANS simulations have yielded the transitional flow around a CD airfoil installed
in anechoic wind tunnels for future TE noise study. The results have been system-
atically compared with experimental data. Different turbulence models have been
tested to compare their capacity to capture transitional process. The k! SST
model has a better global behaviour considering the reasonable computational cost
and is thus considered as the reference model for future RANS simulations. The
transitional model trkkl! slightly improves the airfoil mean loading according
to the averaged URANS data. Installation effects for the two setups with different jet
nozzle size have been studied for flow with same inlet condition. In the larger nozzle
case, a laminar recirculation bubble that triggers the boundary layer development
on the suction side is observed while in the other case not. RANS simulations show
limitations in predicting the laminar transition process and flow separations under
adverse pressure gradient.
References
1. Brooks, T.F., Hodgson, T.H.: Trailing edge noise prediction from measured surface pressures.
J. Sound Vib. 78(1), 69–117 (1981)
2. Fink, M.R.: Experimental evaluation of theories for trailing edge and incidence fluctuation
noise. AIAA J. 13(11), 1472–1477 (1975)
3. Moreau, S., Henner, M., Iaccarino, G., Wang, M., Roger, M.: Analysis of flow conditions in
freejet experiments for studying airfoil self-noise. AIAA J. 41(10), 1895–1905 (2003)
4. Wang, M., Moreau, S., Iaccarino, G., Roger, M.: LES prediction of wall-pressure fluctuations
and noise of a low-speed airfoil. Int. J. Aeroacou. 8(3), 177–197 (2009)
5. Winkler, J., Sandberg, R.D., Moreau, S.: Direct numerical simulation of the self-noise radiated
by an airfoil in a narrow stream. In: 18th CEAS/AIAA Aeroacoustics Conference, Colorado
Springs, pp. 2012–2059 (2012)
6. Neal, D.R.: The effects of rotation on the flow field over a controlled-diffusion airfoil. Ph.D.
thesis, Michigan State University (2010)
7. Vathylakis, A., Kim, J.H., Chong, T.P.: Design of a low-noise aeroacoustic wind tunnel facility
at Brunel University. In: 20th AIAA/CEAS Aeroacoustic Conference and Exhibit, Atlanta
(2014)
8. Jorgenson, F.: How to Measure Turbulence with Hot Wire Anemometers. Dantec Dynamics,
Skovlunde (2004)
9. Raffel, M., Willert, C.E., Kompenhans, J.: Particle Image Velocimetry: A Practical Guide.
Springer, Heidelberg/New York (2013)
10. Westerweel, J.: Fundamentals of digital particle image velocimetry. Meas. Sci. Technol. 8(12),
1379 (1997)
11. Adrian, R.J., Westerweel, J.: Particle Image Velocimetry, Number 30. Cambridge University
Press, Cambridge/New York (2011)
12. Adrian, R.J.: Dynamic ranges of velocity and spatial resolution of particle image velocimetry.
Meas. Sci. Technol. 8(12), 1393 (1997)
13. Keane, R.D., Adrian, R.J.: Optimization of particle image velocimeters. I. Double pulsed
systems. Meas. Sci. Technol. 1(11), 1202 (1990)
14. Walters, D.K., Cokljat, D.: A three-equation eddy-viscosity model for Reynolds-averaged
Navier–Stokes simulations of transitional flow. J. Fluids Eng. 130(12), 121401 (2008)
Part IV
Mathematics and Computation in Finance,
Economics, and Social Sciences
Financial Markets in the Context of the General
Theory of Optional Processes
M.N. Abdelghani and A.V. Melnikov
Abstract A probability space is considered “unusual” if the information flow is

not right or left continuous or is not complete. On these probability spaces lives
certain type of stochastic processes known as optional processes including optional
semimartingales. Optional semimartingales have right and left limits but are not
necessarily right or left continuous. Here, we present a short summary of the
calculus of optional processes and define stochastic logarithms and present some
of its properties. Moreover, we develop a financial market model based on optional
semimartingales and methods for finding local martingale deflators for this market.
Finally, we present some financial examples.
1 Introduction

The assumption that the stochastic basis ˝; F ; F D .F t /t0 ; P satisfy the usual
conditions, where F is complete and right-continuous, is a foundation concept in
the theory of stochastic processes. Stochastic processes that are adapted to this basis
form a large class of processes known as semimartingales whose paths are right-
continuous with left limits (RCLL). This theory has been instrumental in generating
many important results in mathematical finance and in the theory of stochastic
processes. It is difficult to conceive of a theory of stochastic processes without the
usual conditions and RCLL processes.
However, it turns out that it is not difficult to give some examples of stochastic
basis without the usual conditions. In 1975 Dellacherie [4] started to study the theory
of stochastic processes without the assumption of the usual conditions, termed
“unusual conditions”.
Further developments of this theory were carried on by Lepingle [12], Horowitz
[8], Lenglart [11], and mostly by Galtchouk [6, 7].
We believe that the theory optional processes will offer a natural foundation
and a versatile set of tools for modeling financial markets. We can mention here
few research problems that were not treated with the methods of the calculus of
M.N. Abdelghani () • A.V. Melnikov

University of Alberta, Edmonton, AB, Canada

520 M.N. Abdelghani and A.V. Melnikov
optional processes but possibly should be. The first one, is a recent development
in mathematical finance specially in pricing of derivative contracts and hedging
under transaction costs (see [2] for details) that hints to the needed application of the
calculus of optional processes to price derivatives and hedge under transaction costs.
Furthermore, in models [1] with stochastic dividends paid at random times, there is
an opportunity to treat these problems in the context of optional semimartingale
theory in a natural way. Duffie [5] presented a new approach to modeling term
structures of bonds and contingent claims that are subject to default risk. Perhaps
Duffie’s method could be studied with the methods of optional calculus. The aim of
this paper is to present the theory of optional processes on unusual stochastic basis,
develop new results and bring its methods to mathematical finance. The paper is
organized as follows. Section 2 presents foundation material on optional processes.
Section 3 introduces stochastic exponentials and logarithms. Section 4 describes
optional semimartingale model of a financial market and two methods for finding
local martingale deflators for these markets. Section 5 presents examples of optional
semimartingale markets. Finally, we give some concluding remarks.
2 Foundation

Suppose we are given ˝; F ; F D .F t /t0 ; P , t 2 Œ0; 1/, where F t 2 F; F s
F t ; s t, a complete probability space. It is complete because F contains
all P null sets. But this space is unusual because the family F is not assumed
to be complete, right or left continuous. On this space, we introduce O.F/ and
P.F/ the -algebras of optional and predictable processes, respectively (see [7]).
A random process X D .Xt /, t 2 Œ0; 1/, is said to be optional if it is O.F/-
measurable. In general, optional processes have right and left limits but are not
necessarily continuous. For an optional process we can define the following: (a)
X D .Xt /t0 , a left continuous version of the process X and XC D .XtC /t0 ,
the right continuous version of X; (b) The jump processes 4X D .4Xt /t0 and
4Xt D Xt Xt and (c) 4C X D .4C Xt /t0 , 4C Xt D XtC Xt . A random
process .Xt /; t 2 Œ0; 1/, is predictable if X 2 P.F/ and strongly predictable if
X 2 P.F/ and XC 2 O.F/. We denote by P s .F/ the set of strongly predictable
processes. An optional semimartingale X D .Xt /t0 is an optional process that
can be decomposed to an optional local martingale M 2 M loc and an optional
finite variation processes A 2 V , i.e. X D X0 C M C A, [7]. A semimartingale
X is called special if the above decomposition exists but with A being a strongly
predictable process (A 2 A loc the set of locally integrable finite variation processes
[7]). Let S denote the set of optional semimartingales and S p the set of special
optional semimartingales. If X 2 S p then the semimartingale decomposition
is unique. Using the decomposition optional martingales and of finite variation
processes we can decompose a semimartingale further to X D X0 C X r C X g D
X0 C .Ar C M r / C .Ag C M g / where Ar and Ag are right and left continuous finite
Financial Markets in the context of The General Theory of Optional Processes 521
variation processes, respectively. M r 2 M rloc is a right-continuous local martingale

g
and M g 2 M loc is a left-continuous local martingale.
The stochastic integral with respect to optional semimartingale X is defined as a
bilinear form .'; / ı Xt where
Z t Z t
g g
Yt D .'; / ı Xt D Y0 C ' Xtr C ˇ Xt D 's dXsr C s dXsC :
0C 0
where Yt is again an optional semimartingale ' 2 P.F/; and 2 O.F/, such that
2 1=2 1=2 Rt
' ŒX r ; X r 2 A loc and 2 ˇ ŒX g ; X g 2 A loc . The integral 0C 's dXsr
is our usual
R t stochastic integral with respect to RCLL semimartingale however the
g
integral 0 s dXsC is Galchuk stochastic integral [7] with respect to left-continuous
semimartingale.
3 Stochastic Exponentials and Logarithms
Stochastic exponentials and logarithms are indispensable tools of financial mathe-

matics (see, Melnikov et al. [13]). They describe relative returns, link hedging with
the calculation of minimal entropy and therefore utility indifference. Moreover, they
determine the structure of the Girsanov transformation.
For optional semimartingales the stochastic exponential was defined by Galchuk
[7]. If X 2 S then there exists a unique semimartingale Z 2 S such that
Z t Z t
g
Zt D Z0 C Z ı Xt D Z0 C Zs dXsr C Zs dXsC D Z0 E .X/t ; (1)
0C 0
Two useful properties of stochastic exponentials are the inverse and product
formulas. The inverse of stochastic exponential is E 1 .h/ D E .h /, such that,
X .4hs /2 X .4C hs /2
ht D ht hhc ; hc it :
0<st
1 C 4hs 0s<t 1 C 4C hs
And, the product of stochastic Pexponentials is E .U/E

P .V/ D E .U C V C ŒU; V/
where ŒU; Vt D hU c ; V c it C 0<st 4Us 4Vs C 0s<t 4C Us 4C Vs :
The stochastic logarithm is defined by the following theorem.
Theorem 1 Let Y be a real valued optional semimartingale such that the processes
Y and Y do not vanish then the process
Z Z
1 t
1 t
1 g
Xt D ı Yt D dY r C dY ; X0 D 0; (2)
Y 0C Ys s 0 Ys sC
also denoted by X D L ogY is called the stochastic logarithm of Y. The process

X is a unique semimartingale such that Y D Y0 E .X/. Moreover, if X ¤ 1 and
C X ¤ 1 we also have
ˇ ˇ X ˇˇ ˇ
ˇ Yt ˇ 1 Ys ˇˇ Ys
L ogYt D log ˇˇ ˇˇ C 2 ı hY c ; Y c it log ˇˇ1 C (3)
Y0 2Y 0<st
Ys ˇ Ys
ˇ ˇ
X ˇ C Ys ˇˇ C Ys

log ˇˇ1 C :
0s<t
Ys ˇ Ys
It is important to note that the process Y need not be positive for L og.Y/ to exist,
in accordance with the fact that the stochastic exponential E .X/ may take negative
values.
Proof
The assumptions that Y and Y don’t vanish implies that: Sn D
inf t W jYt j 1n " 1, hence 1=Y is locally bounded; likewise, Tn D

inf t W jYt j 1n " 1, hence 1=Y is also locally bounded. Therefore, the stochastic
integral in (2) makes sense. Let YQ D Y=Y0 then YQ 0
D 1. By equation

(2) we have
1 1
that X D .1=Y/ ı Y. Therefore, 1 C Y ı X D 1 C Y ı YQ ı Y D 1 C Y YQ ı YQ D Y,
Q Q Q Q Q Q Q
i.e. YQ D E .X/. Furthermore, X D Y=Y ¤ 1 and C X D C Y=Y ¤ 1. To
obtain uniqueness let XQ be any other semimartingale satisfying Y D E .X/. Q Since
QY D Y=Y0 then YQ D E .X/. Q Therefore YQ D 1 C YQ ı XQ and Y D Y0 C YQ ı X.Q But since
Q
X0 D 0 we have XQ D YYQ ı XQ D Y1Q ı .Y Y0 / D Y1Q ı Y D X and we obtain uniqueness.
To deduce equation (3) we apply Gal’chuk-Ito’s lemma (see [7]) for the optional
semimartingale log jYj. But the log function explodes at 0. To circumvent this
problem consider for each n the C2 functions fn .x/ D log jxj on R such that
jxj 1=n. Consequently for all n, t < Tn and t < Sn we get log jYt j D log jY0 j C Y1 ı
P
P

C Ys
Yt 2Y1 2 ı hY c ; Y c it C 0<st log jYs j Y
Ys
s
C 0s<t C
log jYs j Ys
.
This result together with equation (2) yields equation (3) for t < Tn and t < Sn .
Since, Tn " 1 and Sn " 1 we obtain equation (3) everywhere.
Now, we present some of the properties of stochastic logarithms.
Lemma 1 (a) If X is a semimartingale satisfying X ¤ 1 and C X ¤ 1 then
L og.E .X// D X X0 . (b) If Y is a semimartingale such that Y and Y do not
vanish, then E .L og.Y// D Y=Y0 . (c) For any two optional semimartingales X and
Z we get L og .XZ/ D L ogX C L ogZ C ŒL ogX; L ogZ;
the following identities:
L og X1 D 1 L og .X/ X; X1 :
Proof (a) L og.E .X// D E .X/1 ı E .X/ D EE .X/.X/

ı X D X X0 . (b) Let Z D
E .L og.Y// then by (a) we find that L og .Z/ D L og .E .L og.Y/// D L og.Y/
L og.Y0 / D L og.Y=Y0 /. Therefore, Z D Y=Y0 . (c) By the integral definition of
1
the stochastic logarithm we find L og .XZ/ D XZ ı .X ı Z C Z ı X C ŒX; Z/ D
L ogX C L ogZ C ŒL ogX; L ogZ : Using integration by parts and the integral
1 1 1

definition the stochastic logarithm, L og DXı D 1 ı X X; X1 D
X X X
1 L og .X/ X; X1 :
4 Markets of Optional Processes

Let ˝; F ; F D .F t /t0 ; P , t 2 Œ0; 1/, be the unusual stochastic basis and
that the financial market stays on this space. The market consists of two types of
securities x and X. A portfolio D .; /, is composed of the optional processes
and . is the volume of the reference asset x while is the volume of the security
X. Suppose xt > 0 and Xt 0 for all t 0 and write the ratio process Rt D Xt =xt .
Then, the value of the portfolio is
Yt D t C t Rt : (4)
We restrict the portfolio to be self-financing that is we must have,
Yt D Y0 C ı Rt . (5)
Reconciling equations (4) and (5) we obtain Ct D t C R ı t C Œ; Rt D Y0 D C0

where Ct is the consumption process with its initial value C0 . Since the ratio process
R is optional semimartingale then , evolves in the space P.F/ O.F/ with the
predictable part determining the volume of Rr and the optional part determining the
volume of Rg . Also, belongs to the space O.F/. Furthermore,
R1 for the integral in
equation (5) to be well defined must be R-integrable, 0 s2 dŒR; Rs 2 A loc .
One can interpret the trading strategy in several different ways. First, even-
though gives us the option to trade different parts of the ratio-process, Rr
predictably and Rg optionally, we can still trade both parts in the same way,
predictably. Alternatively, in a portfolio of a sum of left-continuous and right-
continuous assets (see example Sect. 5) we can trade each asset independently.
The left-continuous asset will be traded with an optional strategy while the right-
continuous asset will be traded by a predictable one. This certainly a possible
portfolio in the current structure of financial markets. Moreover, with a theory
of optimal portfolio with consumption and endowment streams we can interpret
the predictable trades of Rr as the optimal portfolio while the optional trades of
Rg as an optimal consumption or endowment streams. Yet another way to think
about optional portfolios in financial markets is in terms of market-orders. Market-
orders can either be predictable or optional. Optional in the sense that trades can be
executed based on a condition, for example the stock price passing some boundary.
Conditional trades are optional processes that act on the left-continuous part of the
ratio-process Rg .
Finally, note that the ratio process is an optional semimartingale and must be
transformed to a local optional martingale for any pricing and hedging theory to
be viable. In the next section we show how to find local martingale transforms that
change the ratio process to some optional local martingales. A special subset of these
martingale deflators can transform RLL optional semimartingales to RCLL local
martingale. This special subset of transforms can be useful in cases where markets
only allow for predictable trading strategies. We will illustrate this procedure in
example Sect. 5.
As in Melnikov et al. [13] for RCLL We suppose that the dynamics of securities
in our market follows the stochastic exponential, Xt D X0 E t .H/ and xt D x0 E t .h/
where x0 and X0 are F 0 -measurable random variables. h D .ht /t0 and H D .Ht /t0
are optional semimartingales admitting the representations, ht D h0 C at C mt and
Ht D H0 C At C Mt with respect to (w.r.t) P. a D .at /t0 and A D .At /t0 are
locally bounded variation processes and predictable. m D .mt /t0 and M D .Mt /t0
are optional local martingales. A local martingale deflator is a strictly positive
supermartingale multiplier used in mathematical finance to transform the value
process of a portfolio to a supermartingale (i.e. a local martingale). Here we will
develop methods for finding local martingale deflators. We can write R as,
Xt
Rt D D R0 E .H/t E 1 .h/t D R0 E .Ht /E .ht / D R0 E . .ht ; Ht //; (6)
xt
t D .ht ; Ht / D Ht ht ŒH; h t D Ht ht C hhc ; hc H c it C J d C J g ;

X 4hs.4hs 4Hs / X 4 C h s 4 C h s 4 C Hs
d
J D ; J Dg
:
0<st
1 C 4hs 0s<t
1 C 4C hs
if is a local optional martingale then R is a local optional martingale and we are

done. Otherwise, we have to find a strictly positive transformation Z 2 M loc that
will render ZR 2 M loc . Z is known as the local martingale deflator. For a strictly
positive Z, we can define N 2 M loc with N D L og.Z/ D Z 1 ı Z or Z D E .N/.
To find N we have the following theorem;
Theorem 2 Given R D R0 E . .h; H// where .h; H/ as in equation (6) and Z D
E .N/ where Z; N 2 M loc .P; F/ and Z > 0 then ZR 2 M loc is a local optional
martingale if and only if .A a/ C hmc N c ; mc M c i C KQ d C KQ g D 0, where KQ d
and KQ g are the compensators of the processes

X .4hs 4Ns / .4hs 4Hs / X 4C hs 4C Ns 4C hs 4C Hs
Kd D ; Kg D :
0<st
1 C 4hs 0s<t
1 C 4C hs
Proof Suppose Zt D E .N/t 2 M loc , Zt > 0 for all t such that ZR 2 M loc thenZR D
R0 E .N/E . .h; H// D R0 E . .h; H; N//, where .h; H; N/ D Nt C Ht ht C
g
hhc ; hc H c it C Jtd C Jt C ŒN; H ŒN; h C ŒN; J d C ŒN; J g , hence,
g
.h; H; N/ D Nt C Ht ht C hhc ; hc H c it C Jtd C Jt
X X
C hN c ; H c it C 4Ns 4Hs C 4C Ns 4C Hs
0<st 0s<t
X X
hN c ; hc it 4Ns 4hs 4C Ns 4C hs
0<st 0s<t
X 4hs .4hs 4Hs /
C 4Ns
0<st
1 C 4hs

X 4 C h s 4 C h s 4 C Hs
C
C 4 Ns :
0s<t
1 C 4C hs
therefore, .h; H; N/ D Nt C Ht ht C hhc N c ; hc H c it C K d C K g . So,

if .h; H; N/ 2 M loc then ZR 2 M loc . And if 4C .h; H; N/ ¤ 1 and
4 .h; H; N/ ¤ 1 then .h; H; N/ 2 M loc , ZR 2 M loc . Now, let us
take into consideration the decomposition of H and h and write .h; H; N/ D
.A a/ C .M m C N/ C h.m N/c ; .m M/c i C K d C K g : So, .h; H; N/ is
a local optional martingale under P if .A a/ C hmc N c ; mc M c i C KQ d C KQ g D 0
where KQ d and KQ g are the compensators of K d and K g , respectively.
By finding all N 2 M loc such that the above equation (4) is valid and E .N/ > 0 we
find the set of all appropriate local optional martingale transforms Z such that ZR is
a local optional martingale. Note that if Z is a local martingales transform such that
ZR is a local martingale then it is true for all self financing strategies .
Theorem 3 If Z is a local martingale transform of R, that is ZR is a local optional
martingale, and is a self financing portfolio which is R-integrable then ZYt is a
local optional martingale.
Proof Z is a local martingale transform of R therefore Z > 0. D .; / is self

and R-integrable, then Yt D Y0 C ı Rt and Zt Yt can be written as,
financing
d Zt Yt D t ŒZt dRt C dZt Rt C d ŒZt ; Rt C dZt t D t d .Zt Rt / C dZt t : This leads
us to the following result Zt Yt D ıZt Rt CıZt : ıZt and ıZt Rt are local optional
martingales therefore their sum Zt Yt is a well defined local optional martingale.
Note, that we have implicitly used the factRthat is bounded, i.e. comes from the
1
fact that is a self financing and also that, 0 t2 d ŒZRt 2 A loc .
On the other hand, if we know that there exist a Z such that ZY is a local optional
martingale then what can we say about the portfolio and the product ZR? It is
reasonable to suppose that Z D E .N/ > 0, -self-financing, is R-integrable and
is bounded. In this case, ı Zt Rt D Zt Yt ı Zt is a sum of two local optional
martingales and therefore a local optional martingale it self, for any optional process
, in particular for D 1; therefore ZR is a local optional martingale.
An alternative approach can be developed using stochastic logarithms. What is
interesting about this approach is that we don’t have to define the process R as
a stochastic exponential of an underlying process . All that is required is that the
ratio process R and its predictable version R don’t vanish, except on sets of measure
zero. This approach is technically based on the following lemmas and results.
Lemma 2 Suppose X D X0 C A C M, x D x0 C a C m, R D X=x and R ¤

0 and R ¤ 0 a.s. P, then L og.R/ is a local optional martingale if and only if
1 1 1 1
X ı A x ı a C x2 Œm; m xX ı ŒM; m D 1.
Lemma 3 Suppose that R and R don’t vanish then L og.R/ 2 M loc , R 2

M loc .
If R is not a local martingale then there exists a local martingale Z > 0 such that ZR
is a local martingale. The following lemma helps us with finding Z.
Lemma 4 Let X D X0 C A C M, x D x0 C a C m, R D X=x and suppose that
R and R don’t vanish then L og.ZR/ is a local optional martingale if and only if
1C X1 ıA 1x ıaC x12 ıŒm; m xX
1 1
ıŒM; mC ZX 1
ıŒZ; M xZ ıŒZ; m D 1 furthermore
if Z D E .N/ > 0 then X ıA x ıaC x2 ıŒm; m xX ıŒM; mC X1 ıŒN; M 1x ıŒN; m D
1 1 1 1
1:
Corollary 1 Suppose that R and R don’t vanish then L og.ZR/ 2 M loc , ZR 2
M loc .
5 Illustrative Examples
Consider a market composed of a bond x and an asset X evolving according to

g
xt D x0 E .h/t and Xt D X0 E .H/t where ht D rt C bLt ; h0 D 0, Ht D t C
g
Wt C aLt ; H0 D 0. Lt D Lt t, Lt D LN t C t, and r, , , a, and b are
d d
constants. W is diffusion term and L and LN are Poisson with constant intensity and
respectively. Let F t be the natural filtration that is neither right or left continuous.
Here the bond is modeled by a left continuous process for which we have assumed
that its jumps don’t necessarily avoid the jumps of the asset (see also [5] for bonds
that can experience defaults). We believe our model gives a better description of a
portfolio of stocks and bonds than models that assume RCLL processes on usual
probability space.
Given x and X the ratio process is Rt D Xx00 E .Ht ht ŒH; h t /. We want to find
Z D E .N/ such that ZR is a local martingale. In Sect. 4 we showed that associated
with the product ZR D Xx00 E . .h; H; N// is the process .h; H; N/. To compute a
g
reasonable form for .h; H; N/ we suppose that Nt D &Wt C cLdt C Lt an optional
local martingale for which Z an optional local martingale deflator, hence
g g
.h; H; N/ D & Wt C cLdt C Lt C t C Wt C aLdt rt C bLt
˝ g c g c g c c ˛
C rt C bLt & Wt C cLdt C Lt ; rt C bLt t C Wt C aLdt

X 2 X b .b / 4C Lgs 2
C ac 4Lds C g
0<st 0s<t
1 C b4C Ls
g
D . r C & / t C .& C / Wt C .c C a/ Ldt C . b/ Lt C acLt
g g g
C .b / Lt C b . b/ Lt ; Lt
D . r C & / t C .& C / Wt C .c C a/ Ldt C acLt C b . b/ LN t :
because Wt and Ldt are martingales, .h; H; N/ is a martingale if and only if

r C & C ca C b b2 D 0. The solution of this equation leads to infinitely
many solutions which means the market is incomplete. Here are some solutions. Let
D 0 which leads to right continuous local martingale deflator. Another solution is
a one which will eliminate jumps on drift that is by letting D 1=b
the effects of
and c D 1=a, then & D r C b2 =.
This is a simple but important example that we have alluded to in Sect. 4. Suppose
that the market participants can’t trade the left-continuous part of the market ratio
process R with an optional trading strategy. A way around this problem is to
transform RLL semimartingles to RCLL local optional martingales. Consider a
market ratio process R D E .M/ > 0, Mt D rt C aWt C bDt C cGt where W is a
diffusion process, D a right-continuous compensated Poisson with intensity and G
a left continuous compensated Poisson with intensity . Let N D ˛W C ˇD C G.
We like to find Z D E .N/ > 0 such that ZR is RCLL. ZR D E .N/E .M/ D
E .N C M C ŒN; M/. Therefore, ZR D E ..a C ˛/W C .b C ˇ/D C .c C /G C .r C
a˛ C bˇ C c /t/. If we choose D c we get rid of the left-continuous part of
ZR. And, if we choose ˛ and ˇ such that r C a˛ C bˇ c2 D 0 we obtain the
RCLL local optional martingale ZR we are searching for. However, we still need to
make sure that our choice Z D E .˛W C ˇD cG/ > 0. This is true, if and only if
.1 C ˇ 4 Dt /.1 c 4C Gt / > 0 or .1 C ˇ/.1 c/ > 0 for all time t.
6 Conclusion
Optional processes including left continuous ones are a natural occurrence in

financial markets. For example, a defaultable bond is modeled by a left continuous
process in [5], stochastic dividends in [16] and transaction costs in [2, 3]. Also
optional processes appears for optimal consumption from investment [15], for
consumption from investment with random endowment [9, 14, 17]. Furthermore,
optional processes naturally arise in the context of super-replication in incomplete
markets as a result of optional decomposition theorem [10]. Therefore, to study
financial markets with optional processes, left-continuous processes and a mixture
of right and left continuous processes the calculus of optional processes is going
to be an indispensable tool in the future study of mathematical finance. We have
introduced the notions of the unusual basis and optional semimartingales, presented
a summary of the calculus of optional processes and introduced new results of
stochastic logarithms. We have also described the optional semimartingale model
of financial market and described a procedure of finding local martingale deflators
for this market. Finally we presented illustrative examples. The first example is of a
portfolio of a left continuous bond and right continuous stock where we showed how
to find local martingale deflators for this market. The second examples demostrates
a method for finding a subset of local martingale deflators that transforms a ladlag
semimartingale to cadlag local optional martingale.
Acknowledgements The research is supported by the NSERC discovery grant #5901.
References
1. Albrecher, H., Bauerle, N., Thonhauser, S.: Optimal dividend-payout in random discrete time.
Stat. Risk Model. Appl. Financ. Insur. 28(3), 251–276 (2011)
2. Czichowsky, C., Schachermayer, W.: Duality theory for portfolio optimisation under transac-
tion costs. arXiv:1408.5989 [q-fin.MF] (2014)
3. Davis, M.H., Panas, V.G., Zariphopoulou, T.: European option pricing with transaction costs.
SIAM J. Control Optim. 31(2), 470–493 (1993)
4. Dellacherie, C.: Deux remarques sur la separabilite optionelle. Sem. Probabilites XI Univ.
Strasbourg, Lecture Notes in Mathematics, vol. 581, pp. 47–50. Springer, Berlin (1977)
5. Duffie, D., Singleton, K.J.: Modeling term structures of defaultable bonds. Rev. Financ. stud.
12(4), 687–720 (1999)
6. Gal’chuk, L.I.: Optional martingales. Matem. Sb. 4(8), 483–521 (1980)
7. Gal’chuk, L.I.: Stochastic integrals with respect to optional semimartingales and random
measures. Theory Probab. Appl. XXIX(1), 93–108 (1985)
8. Horowitz, J.: Optional supermartingales and the Andersen-Jessen theorem. Z. Wahrschein-
lichkeitstheorie und Verw Gebiete. 43(3), 263–272 (1978)
9. Karatzas, I., Zitkovic, G.: Optimal consumption from investment and random endowment in
incomplete semimartingale markets. Ann. Probab. 31(4), 1821–1858 (2003)
10. Kramkov, D.O.: Optional decomposition of supermartingales and hedging contingent claims
in incomplete security markets. Probab. Theory Relat. Fields 105, 459–479 (1994)
11. Lenglart, E.: Tribus de Meyer et théorie des processus. Lecture Notes in Mathematics, vol. 784,
pp. 500–546. Springer, Berlin/New York (1980)
12. Lepingle, D.: Sur la représentation des sa uts des martingales. Lectures Notes on Mathematics,
vol. 581, pp. 418–434. Springer, Berlin/New York (1977)
13. Melnikov, A.V., Volkov, S.N., Nechaev M.L.: Mathematics of Financial Obligations. Trans-
lations of Mathematical Monographs, vol. 212. American Mathematical Society, Providence
(2002)
14. Mostovyi, O.: Optimal investment with intermediate consumption and random endowment.
Math. Financ. (2014). Published on-line
15. Mostovyi, O.: Necessary and sufficient conditions in the problem of optimal investment with
intermediate consumption. Financ. Stoch. 19(1), 135–159 (2015)
16. Wenshen, X., Ling, Z., Zhin, W.: An option pricing problem with the underlying stock paying
dividend. Appl. Math. JCU 12(B), 447–454 (1997)
17. Zitkovic, G.: Utility maximization with a stochastic clock and an unbounded random endow-
ment. Ann. Appl. Probab. 15(1B), 748–777 (2005)
A Sufficient Condition for Continuous-Time
Finite Skip-Free Markov Chains to Have Real
Eigenvalues
Michael C.H. Choi and Pierre Patie
Abstract We provide a sufficient condition for the negative of the infinitesimal

generator of a continuous-time finite skip-free Markov chain to have only real
and non-negative eigenvalues. The condition includes stochastic monotonicity and
certain requirements on the transition rates of the chain. We also give a sample path
illustration of Markov chains that satisfy the conditions and its Siegmund dual. We
illustrate our result by detailing an example which also reveals that our conditions
are not necessary.
1 Introduction
A Markov chain on a denumerable state space is said to be upward skip-free if the

only upward transition is of unit size, yet it can have downward jump of any arbitrary
magnitude. In this paper, we consider a continuous time upward skip-free Markov
chain X D .Xt /t0 on a finite state space E WD f0; 1; : : : ; ag with the infinitesimal
generator G WD .G.i; j//i;j2E and transition semigroup P D .Pt /t0 , where
0
Pt D Pt G ; t 0:
Since X is upward skip-free, G.i; j/ D 0 for j > i C 1; i; j 2 E. The generator G

necessarily satisfies 0 G.i; j/ < 1 for i ¤ j and for all j 2 E,
X
G. j; j/ G. j; k/ : (1)
k¤j
In what follows, we assume that X is conservative (i.e. the equality in (1) holds) and
G. j; j/ < 1 for j 2 E. We refer the interested readers to Asmussen [1], Norris [8]
for a general reference in continuous-time Markov chains.
M.C.H. Choi () • P. Patie

School of Operations Research and Information Engineering, Cornell University, Ithaca, NY,
USA

530 M.C.H. Choi and P. Patie
The spectrum of a skip-free Markov chain plays an important role in the study of
upward hitting time as well as the fastest strong stationary time, see e.g. Diaconis
and Fill [3] and Fill [4]. It is shown in Fill [4] that if the eigenvalues of G are all
real and non-negative, then the law of the upward hitting time can be interpreted
as a convolution of exponential random variables with parameters given by the
aforementioned eigenvalues. However, it is not known under which condition(s)
do a skip-free Markov chain admit real and non-negative spectrum. We aim to fill
in this gap in the literature by providing a sufficient condition. The main idea is to
develop a similarity transformation of G with its so-called Siegmund dual GO which
is the generator of a birth-and-death process.
The rest of the paper is organized as follow. We review Siegmund duality and
the concept of stochastic monotonicity in Sect. 2. They serve as important tools for
our main result, which is presented in Sect. 3. Finally, we provide a sample path
illustration in Sect. 4.
2 Siegmund Duality and Stochastic Monotonicity
Siegmund duality of Markov processes was first introduced by Siegmund [9], which
was further developed and applied by Diaconis and Fill [3], Huillet and Martinez [5],
Jansen and Kurt [6].
Definition 1 (Siegmund kernel) The Siegmund kernel HS D .HS .i; j//i;j2E is
defined to be
HS .i; j/ WD 1fijg ;
where 1 is the indicator function. Its inverse HS1 is
HS1 .i; j/ D 1fiDjg 1fiDj1g :
Definition 2 (Siegmund dual) We say that GO is the Siegmund dual (or HS -dual) of
G if GO is an infinitesimal generator such that
O j/ WD .HS1 GHS /T .i; j/ 0 ;

G.i; i ¤ j:
That is, apart from the diagonal entries .G.

O j; j//j2E , .G.i;
O j//i¤j is elementwise non-
negative.
Next, we recall the definition of stochastic monotonicity for Markov chains.
Definition 3 (Stochastic monotone) G is said to be stochastically monotone if
X X
G.h; k/ G.i; k/ ; 0 h i < j; (2)
kj kj
Real Spectrum for Skip-Free Markov Chains 531
X X
G.h; k/ G.i; k/ ; 0 l < h i: (3)
kl kl
It is known that the existence of Siegmund dual is equivalent to stochastic

monotonicity, which is shown in the following lemma. The proof of a more general
case can be found in Asmussen [1, Proposition 4.1].
Lemma 1 Suppose that .Xt /t0 is a continuous-time Markov chain on E with
infinitesimal generator G. The Siegmund dual GO exists if and only if G is stochasti-
cally monotone.
P
Proof For any i;P j 2 E, since HS GO T .i; j/ D GHS .i; j/ ; HS GO T .i; j/ D ki G.
O j; k/
and GHS .i; j/ D kj G.i; k/ , we have
X X
O j; i/ D
G. O j; k/
G. O j; k/
G.
ki k>i
X
D .G.i; k/ G.i C 1; k// (4)
kj
X
D .G.i C 1; k/ G.i; k// (since G is conservative) : (5)
k>j
O j/ 0 for i ¤ j if and only if G is

In view of (3) and (4), and (2) and (5), G.i;
stochastically monotone. u
t
Finally, we recall a lemma shown in Huillet and Martinez [5] that will be used in
the proof of our main result Proposition 1.
Lemma 2 Suppose that .Xt /t0 is a continuous-time Markov chain on E with
infinitesimal generator G. If G is irreducible on E (resp. on Enf0g, where 0 is an
absorbing state) and is stochastically monotone, then
P
1. GO is conservative except at 0 (resp. on E). That is, G.0;
O 0/ > k¤0 G.0; O k/,
and (1) holds (with G replaced by G)O for j ¤ 0.
O
2. a is the unique absorbing state in G.
3 Main Result
As X is upward skip-free, the infinitesimal generator G is non-symmetric and non-

reversible (with respect to the stationary distribution of the chain), making the
analysis of the spectrum of G difficult. One way to overcome this is to develop
conditions such that its Siegmund dual GO corresponds to a birth-and-death process,
which leads us to the following result.
Proposition 1 Suppose that .Xt /t0 is an upward skip-free Markov chain on E D

f0; 1; : : : ; ag with infinitesimal generator G (and a 3), and assume that .Xt /t0 is
either irreducible on E or irreducible on Enf0g with 0 being an absorbing boundary.
If
1. G is stochastically monotone, and for every i 2 Œ0; a 2,
X X
G.i C 1; k/ > G.i C 2; k/ ; and
ki ki
2. for every i 2 Œ0; a 3,

X X
G. j; k/ D G. j C 1; k/ 8j 2 Œi C 2; a 1 ;
ki ki
then all the eigenvalues of G are real, distinct and non-negative.

Proof Since G is stochastically monotone, its Siegmund dual GO exists by Lemma 1.
For i 2 Œ2; a 1, G.i;
O i 1/ D G.i; i C 1/ > 0 since G is upward skip-free,
O j/ D 0 8j i 2. We also have
conservative and irreducible on Enf0g, and G.i; P
G.1; 0/ D G.1; 2/ > 0. For i 2 Œ0; a2, condition 1 gives G.i;
O O iC1/ D ki G.iC
P
1; k/ ki G.i C 2; k/ > 0. Condition 2 guarantees that G.i; O j/ D 0 for each
i 2 Œ0; a 3 and j 2 Œi C 2; a 1. That is, G is an irreducible birth-and-death
O
process when restricted to the state space f0; : : : ; a 1g. Moreover, by Lemma 2, a
O and as a result the row corresponding to state a is
is the unique absorbing state of G,
zero. Denote by G the restriction of GO to f0; : : : ; a 1g. By breaking off the last
O BD
row and last column of G,O we can write

GO BD h
GO D D .HS1 GHS /T ; (6)
0T 0
O a/ for
where 0 is a column vector of zero, and h is a column vector storing G.i;
i 2 Œ0; a 1. Since GO is a similarity transformation of G, they share the same
eigenvalues. In addition, we observe from (6) that the eigenvalues of G are 0 and
that of GO BD . Now, for i 2 Œ0; a 1, let
BD .0/ D c > 0 ;
GO BD .0; 1/ : : : GO BD .i 1; i/
BD .i/ D BD .0/ ;
GO BD .i; i 1/ : : : GO BD .1; 0/
then BD .i/GO BD .i; j/ D BD . j/GO BD . j; i/ on Enfag, where both sides are non-zero
only when ji jj 1. Define on Enfag
A.i; j/ WD .i/1=2 . j/1=2 GO BD .i; j/ :

1=2 1=2
If D is a diagonal matrix with D BD .i; i/ D BD .i/, then A D D BD GO BD D BD .
Since GO BD is reversible, A is symmetric, and the spectral theorem for symmetric
matrices tells us that A has real eigenvalues. As a result, GO BD has real eigenvalues,
and hence G has real eigenvalues.
We will now show that G has distinct and non-positive eigenvalues. Since GO is
the infinitesimal generator of a finite Markov chain, GO is uniformizable. Let b WD
O i/ < 1. Then PO WD I C 1 GO is a stochastic matrix of a discrete time
maxi2E G.i; b
upward skip-free Markov chain with a being the unique absorbing state. By the
result of Kent and Longford [7, Sec. 4 Lemma 1], PO has distinct eigenvalues, which
implies GO has distinct eigenvalues as well. If G has a positive eigenvalue , then
1 C b > 1 is an eigenvalue for P, O which is not possible since a stochastic matrix
has eigenvalues less than or equal to 1. Therefore, all the eigenvalues of G are non-
positive. t
u
Condition 1 and 2 are by no means necessary for G to admit a real spectrum. This
can be illustrated with the following example:
Example 1 (Condition 1 and 2 are not necessary)
0 1
0:3 0:3 0 0
B 0:4 0:7 0:3 0 C
GDB
@ 0:5
C
0:3 0:85 0:05 A
0:1 0:2 0:4 0:7
has eigenvalues 0; 0:63; 0:89; 1:04, and does not satisfy condition 1 (since
O 1/ D G.1; 0/ G.2; 0/ D 0:1 < 0) and 2 (since G.0;
G.0; O 2/ D G.2; 0/
G.3; 0/ D 0:4 > 0). Therefore, condition 1 and 2 are not necessary for real, distinct
and non-positive eigenvalues. t
u
4 Pathwise Illustration of the Main Result
We recall the notion of strong pathwise duality for Markov processes. Relying
on the fact that Siegmund duality is a strong pathwise duality, we demonstrate
Proposition 1 from a sample path perspective in Fig. 2.
Definition 4 (strong pathwise duality) Let .Xt /t0 and .Yt /t0 be two continuous-
time Markov chains on finite state spaces E and F respectively, and H W E F ! R
be a measurable and bounded function. Suppose that for every T > 0 there are
families of processes f.Xsx /s2Œ0;T gx2E and f.Ysy /s2Œ0;T gy2F , defined on a common
probability space .˝; F ; P/, such that the following holds:
1. For all x 2 E and y 2 F, the finite dimensional distributions of .Xsx /s2Œ0;T under
P (resp. .Ysy /s2Œ0;T under P) agree with that of .Xt /t2Œ0;T under Px (resp. .Yt /t2Œ0;T
under Py ).
2. For all s 2 Œ0; T and all x 2 E, y 2 F,

y y
H.x; YT / D H.Xsx ; YTs / D H.XTx ; y/ ; P a:s::
Then .Xt /t0 and .Yt /t0 are said to be strongly pathwise dual with respect to H.
Theorem 1 (Siegmund duality is strong pathwise duality (Clifford and Sudbury
[2])) Let .Xt / and .Yt / be two continuous-time Markov chains on a finite state space
E, which are dual with respect to the Siegmund kernel HS .i; j/ D 1fijg . Then they
are strongly pathwise dual.
We now discuss a procedure to simulate .Xtx / with generator G and its Siegmund
y
dual .Yt / with generator GO over the time frame Œ0; T. We summarize the procedure
in Clifford and Sudbury [2] as follows: since the state space E is finite, both
y
.Xtx / and .Yt / are uniformizable. Suppose that X0x D x. Let us write c D
maxfmaxi G.i; i/; maxi G.i; O i/g > 0. Then the inter-arrival time (or holding time)
.ti /niD1 with t0 D 0 of the transition of .Xtx / is independent and exponentially
distributed with mean c1 , and the transition follows that of the discrete time
embedded Markov chain with transition matrix P D I C Gc , which is driven by
n i.i.d. .Ui /niD1 uniformly distributed numbers on .0; 1/.
y
To simulate .Yt /, we reverse in both jump directions and jump time of that of
y y
.Xtx /. Precisely, suppose that Y0 D y. The holding time of .Yt / is .T tnk /nkD0 ,
O
and its transition is governed by the transition matrix PO D I C Gc driven by .Vi /niD1 ,
y y
where Vi D 1 UnC1i . Let YQ t WD YTt . The above procedure is summarized into
Algorithm 1:
To illustrate Algorithm 1, we use it to simulate a continuous time birth and death
processes .Xt3 / on E D f0; : : : ; 7g with birth rate and death rate both equal to 0:5
for all states apart from 0 and 7, where we have an absorbing boundary at 0,and a
reflecting boundary at 7 with G.7; 6/ D 0:5. The sample paths are plotted in Fig. 1.
Algorithm 1 Simulate a stochastically monotone CTMC and its Siegmund dual

Require: x; y; T, G, GO
c maxfmaxi G.i; i/; maxi G.i; O i/g
G GO
P I C ; PO IC
c c y
i 1; t0 0; X0x
x; Y0 y
repeat
Sample ti
Exponential(mean D c1 )
i PiC1
until i ti > T
n length.ti / 1
Sample .Ui /niD1
Uniform.0; 1/; Set Vi 1 UnC1i for i D 1; : : : ; n
Sample n steps of P that starts at x and is driven by .Ui /niD1 , say .Xix /niD1
Sample n steps of PO that starts at y and is driven by .Vi /niD1 , say .Yi /niD1
y
y
return .ti /niD0 , .Xix /niD0 , .Yi /niD0
0
0 5 10 15 20
time t
Fig. 1 Sample paths of birth-death .Xt3 / with an absorbing boundary at 0, G.i; i C 1/ D G.i; i
1/ D 0:5, .YQ4t / and .YQ5t / with T D 20, a D 7
This figure should be compared with Jansen and Kurt [6, Fig. 1 Sec. 4.1]. We can
3
observe that strong pathwise duality (cf. Definition 4) holds, that is, X20 y if and
y
only if YQ 0 3.
Next, to demonstrate our main result Proposition 1, we apply Algorithm 1 to
simulate an upward skip-free process .Xt4 / with generator G in (7). It can be checked
that G fulfills condition .1/ and .2/ in Proposition 1. Strong pathwise duality holds
4
y if and only if YQ0 4. In essence, while .Xt4 / is upward
y
in Fig. 2, that is, X20
skip-free, its Siegmund dual can be considered as a birth-and-death process before
it gets absorbed to state 7.
0 1
0 0 0 0 0 0 0 0
B 0:3 0:5 0:2 0 0 0 0 0 C
B C
B 0:1 0:25 0:65 0:3 0 0 0 0 C
B C
B C
B 0:1 0:2 0:4 1:1 0:4 0 0 0 C
GDB C (7)
B 0:1 0:2 0:3 0:35 1:45 0:5 0 0 C
B C
B 0:1 0:2 0:3 0:3 0:6 2:1 0:6 0 C
B C
@ 0:1 0:2 0:3 0:3 0:5 1:0 3:1 0:7 A
0:1 0:2 0:3 0:3 0:5 0:7 1:0 3:1
0
0 5 10 15 20
time t
Fig. 2 Sample paths of .Xt4 / with G given by (7), .YQ1t / and .YQ2t / with T D 20, a D 7
Acknowledgements The authors would like to thank an anonymous referee for insightful
suggestions which improved an earlier version of the paper.
References
1. Asmussen, S.: Applied Probability and Queues. Volume 51 of Applications of Mathematics

(New York), 2nd edn. Springer, New York (2003). ISBN: 0-387-00211-1. Stochastic Modelling
and Applied Probability
2. Clifford, P., Sudbury, A.: A sample path proof of the duality for stochastically monotone Markov
processes. Ann. Probab. 13(2), 558–565 (1985). ISSN: 0091-1798
3. Diaconis, P., Fill, J.A.: Strong stationary times via a new form of duality. Ann. Probab. 18(4),
1483–1522 (1990). ISSN: 0091-1798
4. Fill, J.A.: On hitting times and fastest strong stationary times for skip-free and more general
chains. J. Theor. Probab. 22(3), 587–600 (2009). ISSN: 0894-9840, doi:10.1007/s10959-009-
0233-7
5. Huillet, T., Martinez, S.: Duality and intertwining for discrete Markov kernels: rela-
tions and examples. Adv. Appl. Probab. 43(2), 437–460 (2011). ISSN: 0001-8678,
doi:10.1239/aap/1308662487
6. Jansen, S., Kurt, N.: On the notion(s) of duality for Markov processes. Probab. Surv. 11, 59–120
(2014). ISSN: 1549-5787, doi:10.1214/12-PS206
7. Kent, J.T., Longford, N.T.: An eigenvalue decomposition for first hitting times in random walks.
Z. Wahrsch. Verw. Gebiete 63(1), 71–84 (1983). ISSN: 0044-3719, doi:10.1007/BF00534178
8. Norris, J.R.: Markov Chains. Volume 2 of Cambridge Series in Statistical and Probabilistic
Mathematics. Cambridge University Press, Cambridge (1998). ISBN: 0-521-48181-3. Reprint
of 1997 original
9. Siegmund, D.: The equivalence of absorbing and reflecting barrier problems for stochastically
monotone Markov processes. Ann. Probab. 4(6), 914–924 (1976)
Bifurcations in the Solution Structure of Market
Equilibrium Problems
F. Etbaigha and M. Cojocaru
Abstract In this study, the well-known market disequilibrium model with excess
supply and demand is investigated to determine if it exhibits changes in the structure
and the number of equilibrium states for specific choices of parameter values. We
propose to examine the effects of changing separately each price functions and
unit transaction cost functions. We study the bifurcation problem (i.e., qualitative
change in equilibrium states) as a parametrized variational inequality problem (VI).
We conduct our analysis based on modeling the markets via a projected dynamical
system (PDS), which is a type of constraint ordinary differential equations whose
critical points are the market equilibrium states of the economic model. Numerical
simulation for two examples is carried out to see if and when the behavior of these
market steady states exhibits any qualitative change.
1 Introduction
The market equilibrium models considered in this work are spatial price equilibrium
and market disequilibrium models [11, 14]. The extensive work on the spatial
price equilibrium induced various improvements in both the formulation and the
computation of this problem. In the formulation, it evolved from the formulation of
linear complementarity [1], to the formulation via variational inequality theory as
in [8]; then to PDS models as in [12]. In the computation, numerical methods and
approximation algorithms are utilized to exploit the solutions at different settings
and configurations [1, 12]. It also extended to a market disequilibrium case where the
prices have been regulated [16]. The initial market disequilibrium problem appeared
in [16]. Nagurney et al. [14] have introduced a new market equilibrium model with
excess supply and demand, which extended from the spatial price equilibrium model
presented in [11]. In this paper, we examine the possible structural changes of the
behavior of the market equilibrium and disequilibrium problems under a parameter
variation. Two frameworks are combined: the theory of PDS, and the theory of
evolutionary variational inequality problems (EVI) (see for example [2, 5, 7] and
F. Etbaigha () • M. Cojocaru

University of Guelph, Guelph, ON, Canada

538 F. Etbaigha and M. Cojocaru
the references therein). The rest of the paper is organized as follows: in Sect. 2, we
give a brief review of both market equilibrium and disequilibrium models; moreover,
their variational inequalities forms and corresponding adjustment dynamics are also
presented. Section 3 provides a study of bifurcation questions using the concept
of EVI problem. Numerical simulations for two market equilibrium examples are
presented in Sect. 4. The paper is concluded with a discussion and some suggestions
for future work.
We assume the reader is familiar with the definitions of a convex cone, and those
of the tangent and normal cone to a closed convex nonempty set (see for instance
[3]).
2 Brief Review of Market Equilibrium Models
2.1 Market Equilibrium Model with Excess Supply

and Demand (MESD)
The disequilibrium market model is presented in [14] and summarized below. It

presented m supply markets and n demand markets which involved the production
of a homogeneous commodity under perfect competition. The supply at market i
is expressed as si and demand at market j as dj . Further, the supply price at each
supply market i is denoted by i , and j denotes the demand price at demand market
j. All the supplies are grouped into a column vector s 2 Rm , and the supply prices
are grouped in a row vector 2 Rm . In the same context, the demands and demand
prices are grouped respectively into a column vector d 2 Rn and a row vector 2
Rn . Moreover, the nonnegative commodity shipment attached with a supply and
demand market pair .i; j/ is denoted by Qij . Each trade between the pair .i; j/ is
associated with a unit transaction cost cij . The unit transaction costs which include
the transportation cost and the tax are grouped into a row vector c 2 Rmn , and the
commodity shipments are grouped in a column vector Q 2 Rmn . It is assumed that
there are excess supply and excess demand at each market, denoted as ui and vj . All
excess supplies and excess demands are grouped respectively into a column vector
u 2 Rm and a row vector v 2 Rn . Therefore, a feasible model requires the following
constraints for all supplies, demands and shipments:
X X
si D Qij C ui ; i D 1; : : : m and dj D Qij C vj ; j D 1; : : : :n: (1)
j i
Note that in [11] the equilibrium market is given as above with ui D 0 and vj D 0.
Furthermore in [14], it is assumed that each supply price at the supply market is
regulated by a fixed minimum supply price i , called the price floor at supply market
i. The fixed maximum demand price at demand market j is denoted by Nj , called the
price ceiling at demand market j. The supply price floors and the demand price
Bifurcations in the Solution Structure of Market Equilibrium Problems 539
ceilings are grouped into vectors 2 Rm and N 2 Rn , respectively. The vector,

defined as Q 2 Rmn , contains m vectors of fQ i g, and each of these vectors consists of
n components fi g. The vector Q 2 Rmn consists of m vectors fQj g and each of these
vectors has n components f1 ; 2 ; : : : :; n g. In addition to the above restrictions, the
supply, demand, the commodity shipment, excess supply and excess demand, which
constitute the disequilibrium pattern must also satisfy the following conditions at all
supply and demand markets:
(
D j ; if Qij > 0;
i C cij (2)
j ; if Qij D 0;
( (
D i ; if ui > 0 D Nj ; if vj > 0;
i ; j (3)
i ; if ui D 0 Nj ; if vj D 0:
The conditions in (2) are known as equilibrium conditions [11]. The conditions (3)
are presented in detail in [14]. It is also considered that the supply price at any supply
market i depends on the supplies at all the markets, that is D .s/. Similarly, the
demand price at any demand market j depends on the demands at all the markets,
that is D .d/. In the same context, the unit transaction cost associated with the
pair .i; j/ depends on the commodity shipments from i to j, that is c D c.Q/, where
all ; and c are smooth functions. The vectors O D 2 Rm and O D 2 Rn
are also provided, where O D .Q; O u/ and O D .Q;O v/: The vector QO 2 Rmn
Q
consists of m vectors fO i g and each these vectors contains n components fO i g and
the vector QO 2 Rmn consists of m vectors fQOj g, where each vector has n components
fO1 ; O2 : : : :; On g.
2.2 Variational Inequality Formulation and Adjustment

Dynamics of MESD
The pattern .Q ; u ; v / satisfies the conditions (2) and (3) governing the disequi-
librium market problem if and only if it satisfies the VI problem
QO ; u / C c.Q / .Q
..Q QO ; v //:.Q Q / C ..Q
O ; u / /:.u u /C
O ; v //:.v v / 0 ; 8.Q; u; v/ 2 K and K D Rmn
.N .Q C RC RC :
m n
(4)
Remark: note that in [11] the market equilibrium model without the excess supply
and demand is formulated as a VI as follows:
hF.Q /; Q Q i 0 8 Q 2 K 1 ; K 1 D Rmn
C: (5)
Thus, according to [12], it is known that the adjustment to spatial price equilibrium
states in a VI problem can be obtained by studying the nonsmooth dynamical
system:
dQ.t/
D PTKQ.t/ .F.Q.t///; Q.0/ 2 K 1 ; K 1 D Rmn
C; (6)
dt
where F W K 1 ! Rmn and Fij .Q/ D i .s/ C cij .Q/ j .d/.

In the same manner, we can write the adjustment process of the MESD [7] as
d.Q.t/; u.t/; v.t//

D PTK.Q.t/;u.t/;v.t// .F.Q.t/; u.t/; v.t///; K D Rmn
C RC RC : (7)
m n
dt
For simplicity, we will write .Q; u; v/ instead of .Q.t/; u.t/; v.t//. F W K 7! Rmn
Rm Rn is defined by F.Q; u; v/ D .A.Q; u; v/; G.Q; u/; D.Q; v//; and A W K 7!
Rmn , G W Rmn m m mn n
C RC 7! R and D W RC RC 7! R are defined by: Aij D
n
QO i .Q; u/ C cij .Q/ QOj .Q; v/; Gi D O i .Q; u/ i ; Dj D Nj Oj .Q; v/. Let the vector
x D .Q; u; v/ 2 K, and F.x/ D F.Q; u; v/ then (7) can be written in the form,
xP D PTKx.t/ .F.x.t///: (8)
3 Bifurcations in MESD
3.1 Formulation of the Bifurcation Problem for Constrained

Systems
Bifurcation theory for nonsmooth dynamical systems has not been studied as
extensively as smooth dynamical systems (see for instance [15] and the references
therein). However, the works in [9] and [10] give results for the existence of bifurca-
tions in discontinuous Filippov systems and VI. The nonconventional bifurcations
of fixed points and periodic solutions in Filippov systems have been addressed in
[10]. In [9], it was shown that if V is the Hilbert space, B W V R ! V and K is
closed convex subset of V then the VI of the form
hu B.u; /; v ui 0 8v 2 K (9)
subject to B.0; / D 0 8 2 R; B.u; / D Bu C R.u/; R.u/ D o.kuk/ as u ! 0

has a bifurcation point .0; 0 / for 0 2 R provided there are a sequences .un ; n / of
solutions of (9) such that kun k ¤ 0 and as n ! 1,n ! 0 and un ! 0, for
all n. The above constrain can not be satisfied in our VI formulation, therefore (9)
can not be applied here. Since PDS lies outside of the field of classical dynamical
systems [13], one way to formulate and study the bifurcation problem is by using
the concept of a parametrized VI problem, sometimes called an EVI.
Let us assume that an MESD is formulated as in Sect. 2, such that its solution(s) are
given by equilibrium points of the dynamical system (7). Let 2 R and consider
the system:
d.Q; u; v/
D PTK.Q;u;v/ .F.Q; u; v; //; K D Rmn
C RC RC ;
m n
(10)
dt
which is equivalent to xP D PTKx.t/ .F.x.t/; //; where x D .Q; u; v/ 2 K so that

whenever D 0 we have the initial system (8).
We want to study the problem of finding points x 2 K so that [6]
PTK.x / .F.x ; // D 0 , F.x ; / 2 NK .x /; (11)
for some interval of values of 2 Œa; b. From (11), using the definition of a normal
cone, we have to:
find points x 2 K s.t. hF.x ; /; y x i 0; 8y 2 K: (12)
In (12), F depends on x and and hence, we can reformulate it as a parametrized

VI problem, or as an EVI problem (see for example [4, 5] for the current use of EVI
models) where the bifurcation parameter replaces the “time” parameter t:
hF.x./; /; y./ x./i 0; 8y./ 2 K./: (13)
According to the EVI theory, all the functions will be in Hilbert space
L2 .Œa; b; RnCmCnm /, then F W Œa; b K ! L2 .Œa; b; RnCmCnm / and the feasible set
K can be written as
K D fx 2 L2 .Œa; b; RnCmCnm / j0 x./ M; a:e in Œa; bg; with x; y 2 K: (14)
Note that: K./ D fw 2 RnCmCnm j 0 w Mg; where M is a large.

Essentially, the bifurcation problem becomes an EVI problem where the bounds
of the constraint set K./ are fixed, and do not depend on the parameter . In order
to solve such an EVI problem, one can use existing methods in the literature [5].
Cojocaru et al. in [5] showed that the linkage between the projected dynamical
system and EVI time dependent problem where the critical points of the PDS are
the same as the solutions of EVI problem. Using this approach, the PDS associated
to dependent EVI can be formulated as
dx.; /
D PTKx.; / .F.x.; /; //; (15)
d
where the time represents the time evolution to the equilibrium x.; :/. This
formulation is employed to solve the problem in (13).
3.2 Bifurcations for MESD
The equilibrium problems are analyzed using different parameterization strategies.

Initially, by introducing a parameter that represents the supply price floors, for which
the system behavior did not change and the equilibrium is still unique. Secondly we
impact the supply prices by introducing a parameter to represent the variation of
the supplies on the supply price functions, the equilibrium is still unique and minor
changes are observed only in the trajectories convergence for different values of .
The same experiments are repeated for the demand prices and the same results are
observed. These results suggest that the bifurcations are not likely to occur with
such parametrization strategies. Parametrizing the unit transaction cost functions
by introducing to represent the variation of the shipments will likely impact the
market steady states behavior significantly such that c D c.Q; /. Another direction
is to introduce into both the supply and demand prices functions such that D
.s; /, and D .d; /. It is expected that this change will impact the market
equilibria since changing both the supply and the demand prices at the same time
will dramatically impact the equilibrium and disequilibrium conditions (2) and (3).
The next section shows two examples of market equilibrium problems using this
parametrization strategy.
It is known that whenever F is strictly monotone or beyond (strongly monotone,
etc.) [11] as a function of x, the solution of the MESD problem is unique [13]. Thus,
there is only one equilibrium point of (8). From the EVI theory [5, 7], whenever
F.x; / with 2 Œa; b remains strictly monotone, the parametrized MESD (13)
still has a unique solution. It is therefore logical to seek values of which will
lead to a non-strictmonotonic F.x; /. In these ranges of values we may find distinct
behaviour of the equilibrium.
4 Examples
4.1 Example 1
We consider the example of a market equilibrium problem with two supply and two
demand markets which has a unique equilibrium as shown in [11]. In order to detect
if the change on the cost functions impacts the equilibrium, we modify the example
and introduce in the cost function. At D 0 the equilibrium is still unique
.0:8; 2:96; 0; 0/ since the vector field F is still strictly monotone. Additionally, the
variables are adjusted to be dependent where 2 Œa; b.
The supply price functions are defined by W R2 ! L2 .Œa; b; R2 /,
1 .s.// D 5s1 ./ C s2 ./ C 2, 2 .s.// D s1 ./ C 2s2 ./ C 3;
the demand price functions are defined by W R2 ! L2 .Œa; b; R2 /,
1 .d.// D 2d1 ./ d2 ./ C 28:75, 2 .d.// D d1 ./ 4d2 ./ C 41;
and the unit transaction costs are defined by c W Œa; b K ! L2 .Œa; b; R4 /,
c11 .Q; / D Q11 ./ C .0:5 C /Q12 ./ C 1, c12 .Q; / D .2 C /Q12 ./ C
Q22 ./ C 1:5, c21 .Q; / D .3 /Q21 ./ C 2Q11 ./ C 25, c22 .Q; / D .2
/Q22 ./ C Q12 ./ C 30. The feasible set is K D fQ 2 L2 .Œa; b; R4 /j0 Q./
5; a:e in Œa; bg.
We consider a vector field F defined as F W Œa; b K ! L2 .Œa; b; R4 / where
F11 .Q./; / D 1 .s.// C c11 .Q./; / 1 .d.//
D 8Q11 ./ C .6:5 C /Q12 ./ C 3Q21 ./ C 2Q22 ./ 25:75.
F12 .Q./; / D 1 .s.// C c12 .Q./; / 2 .d.//
D 6Q11 ./ C .11 C /Q12 ./ C 2Q21 C 6Q22 ./ 37:5.
F21 .Q./; / D 2 .s.// C c21 .Q./; / 1 .d.//
D 5Q11 ./ C 2Q12 ./ C .7 /Q21 ./ C 3Q22 ./ 0:75.
F22 .Q./; / D 2 .s.// C c22 .Q./; / 2 .d.//
D 2Q11 ./ C 6Q12 ./ C 3Q21 ./ C .8 /Q22 ./ 8.
Our EVI problem is given by hF.Q ./; /; Q./ Q ./i 0; 8Q./ 2 K; and
the associated PDS can be written as in (15).
We examine the effect of changing in the cost functions considering 2 Œ0; 8
to ensure that the costs are always nonnegative. We obtain the equilibrium at each
value of by the method in [5] implemented in Matlab. For 0 < 8, for any
arbitrary initial condition, the system shows just one equilibrium corresponding to
each value of . At D 8, the shipments converge to three boundary equilibria,
which means that F is non-strictmonotonic. The numerical simulations of the system
at D 8 are presented in the Figs. 1 and 2. Shipments at supply market 1 are
4.5
3.5
2.5
1.5
0.5
Q11 Q12 Q21 Q22
Fig. 1 Values of .Q11 ; Q12 ; Q21 ; Q22 / at equilibrium are presented as a heatmap. The heatmap
shows three equilibrium patterns, the equilibria are color-coded such that each repeated pattern is
represented by the same color
a 5
4.5
3.5
3
22
2.5
Q
1.5
0.5
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Q21
b 5
4.5
3.5
3
Q12
2.5
1.5
0.5
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Q
11
Fig. 2 Phase portraits for D 8, both figures show trajectories starting from 40 initial conditions
(equilibria shown by square markers). There are three distinct equilibrium points
balanced at: .Q11 ; Q12 / 2 f.0; 1:44/; .1:97; 0/; .0; 1:97/g, and shipments at supply
market 2 are balanced at: .Q21 ; Q22 / 2 f.5; 0/; .0; 5/; .0; 0/g as shown in Fig. 2.
As a result of the change in the number of equilibria, D 8 is a bifurcation
value.
4.2 Example 2
Here we modify the above example to paramterize the supply and demand price
functions. The example is extended to the case of excess supply and excess demand.
Furthermore is introduced into both price functions. Without introducing
the system shows one equilibrium .1:6; 2:8; 0; 0; 0; 0; 0; 0/ since F is still strictly
monotone. The supply price functions are defined by
1 .s./; / D .5C/s1 ./Cs2 ./C12, 2 .s./; / D .1C/s1 ./C2s2 ./C30.
The demand price functions are defined by
1 .d./; / D 2d1 ./ .1 /d2 ./ C 45,
2 .d./; / D .1 /d1 ./ 4d2 ./ C 55:
The unit transaction cost functions are as in the Example 1 with D 0.
The feasible set is K D fx 2 L2 .Œa; b; R8 /j0 x./ 5; a:e in Œa; bg, where
x./ D .Q11 ./; Q12 ./; Q21 ./; Q22 ./; u1 ./; u2 ./; v1 ./; v2 .//. We consider
the supply price floors at the markets are 1 D 10 and 2 D 15 and the demand
price ceilings are N1 D 45 and N2 D 55. Then the vector field F is defined as
F W Œa; b K ! L2 .Œa; b; R8 /; for simplicity we assume Q./ D .x1 ; x2 ; x3 ; x4 /.
Then F11 .Q./; u./; v./; / D .8 C /x1 C .6:5 C /x2 C .5 C /u1 C 3x3 C .2
/x4 C 2v1 C u2 C .1 /v2 32.
F12 .Q./; u./; v./; / D .6 C /x1 C .11 C /x2 C .5 C /u1 C 6x4 C 4v2 C
u2 C .1 /v1 C .2 /x3 41:5.
F21 .Q./; u./; v./; / D .5 C /x1 C .1 C /u1 C .4 C /x2 C .3 /x4 C 2v1 C
.1 /v2 C 7x3 C 2u2 .
F22 .Q./; u./; v./; / D 2x1 C .5 C /x2 C .3 /x3 C .1 /v1 C 4v2 C 6x4 C
.1 C /u1 C 2u2 15,
.1 1 ; 2 2 / D ..5 C /x1 C .5 C /x2 C .5 C /u1 C x3 C x4 C u2 C 2; .1 C
/x1 C .1 C /x2 C .1 C /u1 C 2x2 C 2x4 C 2u2 C 15/
.N1 1 ; N2 2 / D .2x1 C 2x3 C 2v1 C .1 /x2 C .1 /x4 C .1 /v2 ; .1
/x1 C.1 /x3 C.1 /v1 C4x2 C4x4 Cv2 /. Then the EVI problem can be written
as hF.Q ./; u ./; v ./; /; .Q./; u./; v.// .Q ./; u ./; v .//i 0.
We set 2 Œ7; 8 to ensure the prices are nonnegative. The experiments is
carried out as discussed in Example 1. For each value of there exist only one
equilibrium except for the value of D 6 at which two equilibria occur. Then at
D 6, F is non-strictmonotonic, and D 6 is a bifurcation value. Figure 3
shows the two equilibria. Shipments at supply market 1 .Q11 ; Q12 / 2 f.0; 5/; .5; 0/g
and shipments at supply market 2 .Q21 ; Q22 / 2 f.0; 0:83/; .0; 2:4/g are shown in
4.5
3.5
2.5
1.5
0.5
Q11 Q12 Q21 Q22 u1 u2 v1 v2
Fig. 3 The heatmap shows the equilibrium pattern .Q11 ; Q12 ; Q21 ; Q22 ; u1 ; u2 ; v1 ; v2 /. Each value
is represented by singular color, the number of equilibria in this case two which is indicated by
switch in color
(a) 5 (b) 4
4.5 3.5
4
3
3.5
2.5
Q 22
3
Q12
2.5 2
2 1.5
1.5
1
1
0.5 0.5
0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5
Q11 Q21
Fig. 4 Phase portraits for D 6 show the convergence of the shipments at each market. (a)
.Q11 ; Q12 / convergence. (b) .Q21 ; Q22 / convergence
Fig. 4. The excess supply and demand at market 1 .u1 ; v1 / 2 f.0; 0/; .4:9; 0/g and
the excess supply and demand at market 2 .u2 ; v2 / 2 f.4:18; 0/; .5; 0/g are presented
in Fig. 5.
(a) 5 (b) 5
4.5 4.5
4 4
3.5 3.5
3 3
v1
v2
2.5 2.5
2 2
1.5 1.5
1 1
0.5 0.5
0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
u1 u2
Fig. 5 Phase portraits for D 6 show the convergence of the excess supply and excess demand
at each market. (a) .u1 ; v1 / convergence. (b) .u2 ; v2 / convergence
5 Conclusion
In this study we considered the bifurcation problem for the market equilibrium
model as an EVI problem. Furthermore, we studied the impact of changing the
supply price, demand price and the cost functions on the market equilibrium states.
With both the cost functions and price functions, the effect of the variations of
the parameter is seen on the number of equilibria occurring at specific values.
The equilibrium states were obtained using trajectories of the associated projected
dynamics. The empirical results on two examples showed that bifurcations occur
in such systems. Further investigation is required to extend the applicability of the
proposed method to other equilibrium models. Moreover, additional future work
is to study the usage of nonlinear parameter dependencies to explore existence of
bifurcations in the solution structure of market equilibrium problems.
References
1. Asmuth, R., Eaves, B.C., Peterson, E.L.: Computing economic equilibria on affine networks
with lemke’s algorithm. Math. Oper. Res. 4(3), 209–214 (1979)
2. Barbagallo, A., Mauro, P.: Evolutionary variational formulation for oligopolistic market
equilibrium problems with production excesses. J. Optim. Theory Appl. 155(1), 288–314
(2012)
3. Cojocaru, M.-G.: Monotonicity and existence of periodic orbits for projected dynamical
systems on hilbert spaces. Proc. Am. Math. Soc. 134(3), 793–804 (2006)
4. Cojocaru, M.-G.: Double-layer dynamics theory and human migration after catastrophic
events. In: Cojocaru, M.-G. (ed.) Nonlinear Analysis with Applications in Economics, Energy
and Transportation, pp. 65–86. Bergamo University Press, Bergamo (2007)
5. Cojocaru, M.-G., Daniele, P., Nagurney, A.: Projected dynamical systems and evolutionary
variational inequalities via hilbert spaces with applications1. J. Optim. Theory Appl. 127(3),
549–563 (2005)
6. Cojocaru, M.-G., Jonker, L.: Existence of solutions to projected differential equations in hilbert
spaces. Proc. Am. Math. Soc. 132(1), 183–193 (2004)
7. Daniele, P.: Dynamic Networks and Evolutionary Variational Inequalities. Edward Elgar
Publishing, Cheltenham/Northampton (2006)
8. Florian, M., Los, M.: A new look at static spatial price equilibrium models. Reg. Sci. Urban
Econ. 12(4), 579–597 (1982)
9. Le, V.K., Schmitt, K.: Global Bifurcation in Variational Inequalities: Applications to Obstacle
and Unilateral Problems, vol. 123. Springer Science & Business Media, Berlin (1997)
10. Leine, R., Van Campen, D., Van de Vrande, B.: Bifurcations in nonlinear discontinuous
systems. Nonlinear Dyn. 23(2), 105–164 (2000)
11. Nagurney, A.: Network Economics: A Variational Inequality Approach, vol. 10. Springer
Science & Business Media, Berlin (1998)
12. Nagurney, A., Takayama, T., Zhang, D.: Massively parallel computation of spatial price
equilibrium problems as dynamical systems. J. Econ. Dyn. Control 19(1), 3–37 (1995)
13. Nagurney, A., Zhang, D.: Projected Dynamical Systems and Variational Inequalities with
Applications, vol. 2. Springer Science & Business Media, Berlin (2012)
14. Nagurney, A., Zhao, L.: Disequilibrium and variational inequalities. J. Comput. Appl. Math.
33(2), 181–198 (1990)
15. Perko, L.: Differential Equations and Dynamical Systems, vol. 7. Springer Science & Business
Media, Berlin (2013)
16. Thore, S.: Spatial disequilibrium. J. Reg. Sci. 26(4), 661–675 (1986)
Pricing Options with Hybrid Stochastic
Volatility Models
Glynis Jones and Roman Makarov
Abstract We introduce a hybrid stochastic volatility model where the asset price
process follows the Heston model and interest rates are governed by a two-factor
stochastic model. Two cases are considered. First, it is assumed that interest rates
and asset prices are uncorrelated. The characteristic function method is used to
derive semi-analytical pricing formulae for plain vanilla options. In the second case
we introduce a correlation between the asset price process and the short rate process
and use Monte Carlo simulations for pricing options. To reduce the stochastic error,
we implement the control variate method where an estimator of the option value
for the uncorrelated case is used as a control variate. The options are priced with a
varying correlation coefficient. We observe that the control variate method allows
us to speed up Monte Carlo computations by a factor with the magnitude of several
hundreds. The efficiency of the method is higher for smaller values of the correlation
coefficient. We then study the impact a correlation between the two processes has on
option prices. It has been noticed that the call option price is an increasing function
of the correlation coefficient.
1 Introduction
In 1973, Fischer Black, Myron Scholes and Robert Merton [2] introduced a
technique for finding a closed-form solution for the price of a plain European option.
The price, V D V.t; S/, of a European-style option is a function of time t and asset
price S (at time t) that satisfies the partial differential equation
@V 1 @2 V @V
C 2 S2 2 C rS rV D 0 (1)
@t 2 @S @S
subject to the terminal condition V.T; S/ D f .S/, where r is the risk-free interest
rate, is the volatility of the stock, and f .S/ is the payoff at maturity T. Here we
G. Jones • R. Makarov ()

Department of Mathematics, Wilfrid Laurier University, 75 University Avenue West, Waterloo,
ON, Canada

550 G. Jones and R. Makarov
assume no dividend on the stock. The solution to (1) is given by the discounted
expectation of the terminal payoff:
V.t; S/ D er.Tt/ EQ Œ f .ST / j St D S ; (2)
where Q is the risk-neutral probability measure with the bank account Bt D ert taken
as a numeraire.
The Black–Scholes–Merton formula (2) assumes that the price of an underlying
asset such as a stock, fSt gt0 , follows geometric Brownian motion (under the risk-
neutral measure Q):
dSt D rSt dt C St dWt :
The main drawback of this model is that the volatility of the underlying risky
asset is assumed to be constant. This assumption does not fit many factors observed
on financial markets such as volatility smiles and skews. Option pricing theory has
evolved since 1973. Other models that take into account many market phenomena
have been developed. Examples of such models include models with stochastic
volatility and stochastic interest rates. In this paper, we consider a hybrid model
that combines a stochastic volatility model for asset prices with a stochastic short
rate process (see [1, 8, 14] and references wherein). In particular, we develop a
modification of the Heston model [9] with stochastic interest rates that are governed
by a two-factor model (see [4]). We consider a two-factor modification of the
Vasicek model [15] but other models such the Hull–White model [10] and the Cox–
Ingersoll–Ross (CIR) model [5] can also be used.
To specify the hybrid model, we combine two systems of stochastic differential
equations. The first one defines the price of a risky asset and its volatility. The second
system defines the short rate process. As a result, we obtain a four-factor stochastic
volatility model named here as the Heston–Vasicek model. This hybrid model fits
in the class of affine diffusion processes (see [6]) for which a closed-form solution
of the characteristic function exists. However, the derivation of the characteristic
function for a hybrid multi-factor model with a full matrix of correlations is a
challenging problem. In this paper we study a simpler approach. First, we assume
that the stock price process and the interest rate process are uncorrelated. Fourier
techniques are used to obtain no-arbitrage prices of European-style options. Second,
we introduce a correlation between asset prices and interest rates and use the Monte
Carlo method to simulate the coupled system of SDEs and to price options. To
reduce the stochastic error, we use the control variate method (see, e.g., [7]), where
an estimator of the option value for the uncorrelated case is used as a control variate.
The rest of this paper is organised as follows. In Sect. 2, we formulate the hybrid
model and obtain characteristic functions for the uncorrelated case. A correlation
between the asset price and interest rate processes is then introduced and options are
priced using the Monte Carlo method. In Sect. 3, we discuss and analyse numerical
results, and our conclusions are drawn in Sect. 4.
Pricing Options with Hybrid Stochastic Volatility Models 551
2 The Hybrid Model
In this section, we construct a four-factor hybrid stochastic volatility model where a

stochastic interest rate is introduced into a stochastic volatility model. We define our
model such that the log-price process, fXt D ln St gt0 , follows the Heston model
and the short rate process, frt gt0 , is governed by the two-factor Vasicek model.
Therefore, it is called the Heston–Vasicek model. The hybrid model consists of four
coupled SDEs:
t p
dXt D .rt /dt C t dWts ; (3)
2
p
dt D . t /dt C t dWt ; (4)
drt D . C ut aN rt /dt C 1 dWtr ; (5)
N t dt C 2 dWtu ;
dut D bu (6)
where Wts , Wt , Wtr , Wtu are standard Brownian motions such that dWts dWt D s dt
and dWtr dWtu D ru dt. The two-factor Vasicek model (5)–(6) can be written as a
two-factor Gaussian model (the equivalence is proved in [4]):
dxt D axt dt C dWtx ; (7)

y
dyt D byt dt C dWt ; (8)
rt D xt C yt C .t/ ; (9)
where xt and yt are Gaussian processes, .t/ is a deterministic function given by

.t/ D r0 eaT C a .eat 1/, and W x and W y are correlated standard Brownian
motions with correlation coefficient xy . The parameters of (5)–(6) and (7)–(8) are
related as follows:
aN D a; bN D b;
p
1 D 2 C 2 C 2xy ; 2 D .a b/;
xy C
ru D p :
2 C2 C2xy
First, we assume that W x and W y are not correlated with W s and W . Thus, the
correlation matrix for the four Brownian motions is given by
2 3
1 s 0 0
6s 1 0 07
6 7: (10)
40 0 1 xy 5
0 0 xy 1
The option pricing formula for a standard European call takes the following form:
h RT i
Q
C.t; S/ D Et;r;S e t ru du .ST K/C : (11)
Q
Here, the expectation Et;r;S Œ computed under the risk-neutral measure Q is
conditional on frt D r; St D Sg. The discounting factor cannot be pulled out since
it is stochastic, so the mathematical expectation is computed as follows:
h RT i
Q
C.t; S/ D Et;r;S e t ru du .ST K/C
h RT i
Q
D Et;r;S e t ru du .ST K/ fST >Kg (12)
h RT i h RT i
Q
D Et;r;S e t ru du ST fST >Kg Et;r;S
Q
e t ru du K fST >Kg :
To calculate the conditional expectations in the r.h.s. of (12), we apply the change
of numeraire technique. The first expectation can be simplified by using the asset
price as a numeraire and the second expectation can be simplified by using the zero-
coupon bond as a numeraire. So, we have

Q1 BT =Bt RtT ru du
C.t; S/ D Et;r;S e ST fST >Kg
ST =St
RT

BT =Bt
Q2
Et;r;S e t ru du K fST >Kg
Z.T; T/=Z.t; T/
(13)
Q1 Q2
D SEt;S fST >Kg KZ.t; T/Et;S fST >Kg
D SQ1 .ST > K j St D S/ KZ.t; T/Q2 .ST > K j St D S/

D SF1 KZ.t; T/F2 ;
Rt h RT i
Q
where Bt D e 0 ru du is the bank account, Z.t; T/ D Et;r e t ru du is the time-t price
of a unit zero-coupon bond with maturity T, Q1 is the EMM (equivalent martingale
measure) with the stock St used as a numeraire asset, and Q2 is the T-forward EMM
with the zero-coupon bond Z.t; T/ used as a numeraire.
As in [9], the characteristic functions need to be found for computing the
probabilities F1 D Q1 .ST > K j St D S/ and F2 D Q2 .ST > K j St D S/. Since the
interest rate in this model is no longer constant, the characteristic functions for the
Heston model have to be modified to include the stochastic interest rate component.
To determine the form of the new characteristic functions we start with the log-price
process for the SDE (3) re-writing it as follows [11]:
t p
dXt D rt dt C . dt C t dWts / D rt dt C dXtH ; (14)
2
where fXtH gt0 is the log-price process for the Heston model with interest r D 0.
Integrating both sides gives us the following:
Z T Z T Z T
dXs D rs ds C dXsH H) XT D RT C XTH ;
t t t
RT
where RT D t ru du. Then, we have the characteristic functions fHVj ./, j D 1; 2 of
the hybrid model as follows:
Œei.RT CXT / D Et;rj ŒeiRT Et;Xj ŒeiXT ;

Q Q H Q Q H
fHVj ./ D Et;r;X
j
ŒeiXT D Et;r;X
j
(15)
where we used the fact that RT and XTH are independent. Therefore, the new
characteristic functions admit the following form:
fHVj ./ D fVj ./fj ./; j D 1; 2 ;
where fVj is the characteristic function for the Vasicek model and fj is the character-
istic function for the Heston model (under Qj ).
The functions fj , j D 1; 2 are already known [9, 13, 16] so we only need to
determine the characteristic functions of the integrated short rate process, RT , under
EMMs Q1 and Q2 , respectively,
fV1 ./ D Et;r Œe

Q1 iRT
; fV2 ./ D Et;r Œe
Q2 iRT
: (16)
If the short rate rt is a Gaussian process, then the integrated process Rt is Gaussian
2 z2
as well. Hence, we can use the identity EŒeiZ D eiz 2 for a normal random
variable Z with mean z and variance z2 . The probability distribution of the short
rate process does not change under the EMM Q1 . Therefore, Eqs. (7) and (8) can be
used to find the mean and the variance of RT under Q1 :
r0 aT r0 eaT T
R D EŒRT D e C 2C 2
C ;
a a a a a

2 2 1 3
VR D Var.RT / D T C eaT e2aT
a a 2a 2a

2 2 1 3
C T C ebT e2bT
b b 2b 2b

2xy eaT 1 ebT 1 e.aCb/T 1
C TC C :
ab a b aCb
Under the T-forward EMM Q2 the short rate process, rt D xt C yt C .t/, is

defined by the following SDEs [4]:
h 2
i
dxt D axt 2 .1 ea.Tt/ / xy b .1 eb.Tt/ / dt C dWtx ; (17)
h 2
i
dyt D byt 2 .1 eb.Tt/ / xy a .1 ea.Tt/ / dt C dWt :
y
(18)
Therefore, the mean and variance of RT under the EMM Q2 are, respectively,
EŒRT D xint C yint C int ;

2 2 aT 1 2aT 3
Var.RT / D TC e e
a a 2a 2a
2
2 bT 1 2bT 3
C TC e e
b b 2b 2b

2xy eaT 1 ebT 1 e.aCb/T 1
C TC C ;
ab a b aCb
where
Z
T
2 T2 s T 2
xint D E x.s/ds D .1 eaT / .1 ebT / ;
0 a 2 b 2
Z T
2 T 2 s T 2
yint DE y.s/ds D .1 ebT / .1 eaT / ;
0 b 2 a 2
Z T
r0 r0 eaT T
int D .s/ds D eaT C 2 C 2
C :
0 a a a a a
Substituting these into Eq. (16) gives us the new characteristic function fV2 . The
price of the call option under the new model is as follows:
C.t; S/ D SF1 KZ.t; T/F2 ; where (19)

Z
1 1 1 exp.i ln K/
Fj D C Re fHVj ./ d; j D 1; 2 ; (20)
2 0 i
and Z.t; T/ is the price of a zero-coupon bond. The integral in (20) can be evaluated
numerically using a quadrature rule or the fast Fourier transform (FFT) method
[13, 16]. Note that the price of the put option can be calculated using the put-call
parity. Clearly, we are able to derived the pricing formula for European options in
closed form (19)–(20) due to the assumption that the asset price process and the
short rate process are uncorrelated.
We then move on to the correlated case. Let us assume that W x is correlated

with W s , i.e. dWtx dWts D dt
O with O 2 .1; 1/. Alternatively, we can introduce the
correlation between W y and W s or between W x and W . We would like to see how
O affects the option prices as it changes. The correlation matrix for the system of
SDEs (3), (4), (7), (8) is as follows:
2 3
1 s O
O xy
6 s 1 O xy s 7
O s
6 7: (21)
4 O
O s 1 xy 5

O xy
O xy s xy 1
One can easily show that this matrix is positive definite for all s ; xy ; O 2 .1; 1/.
The Cholesky decomposition method can then be applied to generate correlated
normal random variables. These correlated random variables are used in Monte
Carlo simulations of the hybrid model.
As is seen from (21), the correlation coefficients Corr.W s ; W y /, Corr.W ; W x /,
Corr.W ; W y / cannot approach 1 or 1 as jj
O ! 1 since xy ; s 2 .1; 1/. So, the
selection of two processes between which the correlation is introduced affects the
range for the correlation coefficient between another pair.
3 Computational Results and Analysis
We would like to calibrate the model using historical data and then use the
parameters obtained for option pricing. First, the two-factor Gaussian model is
calibrated using yield rates obtained from the US treasury website (see [3, 4]). The
parameters obtained for the short rate model are as follows:
a D 0:2071; b D 2:6414; D 0:0100; D 0:0117; xy D 0:8317; D 0:0113:
Using these parameters, the Heston model has been calibrated using the least square
method [13]. The historical prices of options on SPDR S&P 500 ETF (SPY) for
June 10, 2010 were used with different strike prices and different maturities (option
prices were obtained from the BloombergTM database). The initial asset value is
S0 D 108:88. The least squares method is used to minimize the difference between
market prices and model prices of options. This method minimizes the sum of the
squared errors (the market price minus the model price) made in every iteration. To
achieve the best possible results, the same routine has been run multiple times with
different starting points to get an optimal solution. The parameters that provide the
best fit to market prices are
0 D 0:0648; D 0:0166; D 0:3534; D 0:1223; s D 0:8380:

Monte Carlo estimate of the option price Confidence interval for the option price
10.7 11.5
MC estimate
10.6 Exact value 11
Option price
Option price
10.5 10.5
10.4 10
10.3 9.5
0 1 2 3 4 5 6 7 8 9 10 0 1 2 3 4 5 6 7 8 9 10
M (the number of simulations) x 105 M (the number of simulations) x 105
Fig. 1 Monte Carlo simulations for the at-the-money call option with strike price K D 108:88
Substituting these values into the option pricing formula (19), we can calculate no-
arbitrage prices of call options for the uncorrelated case.
To verify the pricing formula (19), the exact option prices were compared
with prices obtained using Monte Carlo simulations. The Monte Carlo prices
are approaching the respective exact price in all our tests. For example, Fig. 1
demonstrates the convergence of Monte Carlo prices for the at-the-money call
option (with K D S0 ).
To construct Monte Carlo estimates for option prices, the Euler method (see,
e.g., [12]) is used. Discrete-time approximations of sample paths, .XO i ; O i ; xO i ; yO i /
.Xi t ; i t ; xi t ; yi t / with i D 0; 1; : : : ; n 1 and t D Tn , are generated using the
following scheme:
p p
XO iC1 XO i C t .Ori 0:5maxf0; O i g/ C maxf0; O i g tzs ;
p p
O iC1 O i C t.v O i / C maxf0; O i g tz ;
p
xO iC1 xO i aOxi t C tzx ;
p
yO iC1 yO i bOyi t C tzy ;
a.iC1/ t
rOiC1 xO iC1 C yO iC1 C r0 ea.iC1/ t e :
a
The initial conditions are xO 0 D 0, yO 0 D 0, rO0 D r0 , and XO 0 D ln S0 . For every

time step, four correlated normal random variables, .zs ; z ; zx ; zy /, are required; they
are obtained using the Cholesky decomposition of the correlation matrix given by
RT C
(10) or (21). A sample value of the discounted payoff, e 0 ru du eXT K , is then
computed. The trapezoidal rule is used to approximated the discounting factor.
Monte Carlo simulations usually have a large stochastic error associated with
them. To reduce the variance of our Monte Carlo estimator of the option value we
use the control variate method. The uncorrelated case (for which the option pricing
formula is available in closed form) is used to construct the control variate. The
same i.i.d. normal samples are used to calculate the estimates for both correlated
and uncorrelated cases.
In Table 1 we examine how the option prices vary with . O As O approaches 0,
the option price gets closer to the analytical price calculated for O D 0. We find
that the option price increases as O increases for at-the-money (K D 108:88), in-
the-money (K 2 f100; 106g), and out-of-the-money (K 2 f110; 115g) options. The
reasoning behind this is that as interest rates go up, the return on stocks increases
thereby increasing the value of a call option. We can say the call option price
and the interest rates are positively correlated. Therefore, increasing the correlation
coefficient between the interest rates and log-price processes increases the price of
a call option as is seen from our simulations.
The ratio of the standard deviations is calculated (the standard deviation of the
crude estimator over the standard deviation of the controlled estimator). This gives
us the factor by which the control variate method improves the crude Monte Carlo
method. This ratio ranges from 238 to 3846 depending on the values of O and K.
As is seen from Table 1 and Fig. 2, the control variate method gives a significant
improvement over the standard estimator when the correlation coefficient is close to
zero.
4 Conclusion
In this paper we have successfully derived a semi-analytical formula for no-arbitrage

pricing a standard European option under the hybrid four-factor Heston–Vasicek
model, where the interest rate process and the log-asset price process are uncorre-
lated. We then studied the impact a correlation between the two processes had on
option pricing using Monte Carlo simulations. It has been noticed that the price of
a European call is an increasing function of the correlation coefficient. The use of
a variance reduction method greatly improved our results for the correlated case as
we can see from the ratios of standard deviations presented in Table 1. For future
work, it will be interesting to study a hybrid model that combines the Heston model
and the two-factor Hull–White model and compare its performance with the model
studied in this paper.
558
Table 1 The impact of O on option prices

O K D 100 Std ratio K D 106 Std ratio K D 108:88 Std ratio K D 110 Std ratio K D 115 Std ratio
0.9 13.0576 314.4654 9.3820 284.3563 7.8718 273.9726 7.3296 268.0965 5.2128 240.9639
0.8 13.0597 389.1051 9.3843 357.1429 7.8742 340.1361 7.3321 332.2259 5.2153 298.5075
0.7 13.0616 476.1905 9.3866 436.6812 7.8765 414.9378 7.3344 406.5041 5.2176 363.6364
0.6 13.0636 477.0095 9.3889 537.6344 7.8789 510.2041 7.3368 500.0000 5.2200 448.4305
0.5 13.0656 729.9270 9.3912 671.1409 7.8813 641.0256 7.3392 625.0000 5.2224 558.6592
0.4 13.0676 934.5794 9.3935 862.0690 7.8837 826.4463 7.3416 813.0081 5.2248 719.4245
0.3 13.0695 1282.0513 9.3958 1176.4706 7.8861 1136.3636 7.3440 1098.9011 5.2272 980.3922
0.2 13.0715 1960.7843 9.3981 1785.7143 7.8885 1724.1379 7.3464 1666.6677 5.2296 1492.5373
0.1 13.0735 3846.1538 9.4004 3571.4286 7.8909 3448.2759 7.3489 3448.2759 5.2320 2941.1765
0 13.0755 – 9.4027 – 7.8932 – 7.3513 – 5.2344 –
0.1 13.0775 3846.1538 9.4050 3571.4286 7.8956 3448.2759 7.3537 3448.2759 5.2368 2941.1765
0.2 13.0794 1923.0769 9.4073 1818.1818 7.8980 1724.1379 7.3560 1694.9153 5.2392 1492.5373
0.3 13.0814 1282.0513 9.4096 1176.4706 7.9004 1123.5955 7.3585 1098.9011 5.2416 980.3922
0.4 13.0834 934.5794 9.4119 862.0690 7.9028 819.6721 7.3608 806.4516 5.2440 735.2941
0.5 13.0854 724.6377 9.4142 666.6667 7.9051 636.9427 7.3632 625.0000 5.2464 555.5556
0.6 13.0873 584.7953 9.4165 534.7594 7.9075 510.2041 7.3656 500.0000 5.2487 444.4444
0.7 13.0893 476.1905 9.4187 434.7826 7.9099 414.9378 7.3680 404.8583 5.2511 361.0108
0.8 13.0913 389.1051 9.4210 355.8719 7.9123 340.1361 7.3704 332.2259 5.2535 296.7359
0.9 13.0933 312.5000 9.4233 286.5330 7.9146 272.4796 7.3728 266.6667 5.2559 238.6635
Calculations were done using the maturity time T D 0:5, the time step t D 0:001, and the number of sample paths M D 1;000;000
G. Jones and R. Makarov
Fig. 2 The impact of the correlation coefficient O on the performance of the control variate
methods
Acknowledgements R. Makarov wishes to acknowledge the generous support of the NSERC

Discovery Grant program.
References
1. Ahlip, R., Rutkowski, M.: Pricing of foreign exchange options under the Heston stochastic
volatility model and CIR interest rates. Quant. Financ. 13(6), 955–966 (2013)
2. Black, F., Scholes, M.: The pricing of options and corporate liabilities. J. Pol. Econ. 81(3),
637–654 (1973)
3. Blanchard, A.: The two-factor Hull-White model: pricing and calibration of interest rates
derivatives. Accessed 28 Mar 2014
4. Brigo, D., Mercurio, F.: Interest Rate Models-Theory and Practice: With Smile, Inflation and
Credit. Springer, Berlin/New York (2007)
5. Cox, J.C., Ingersoll, J.E., Ross, S.A.: A theory of the term structure of interest rates.
Econometrica 53(2), 385–407 (1985)
6. Duffie, D., Pan, J.: Transform analysis and asset pricing for affine jump-diffusions. Economet-
rica 68(6), 1343–1376 (2000)
7. Glynn, P.W., Szechtman, R.: Some new perspectives on the method of control variates. In:
Monte Carlo and Quasi-Monte Carlo methods, 2000 (Hong Kong), pp. 27–49. Springer, Berlin
(2002)
8. Grzelak, L.A., Oosterlee, C.W.: On the Heston model with stochastic interest rates. SIAM J.
Financ. Math. 2(1), 255–286 (2011)
9. Heston, S.L.: A closed-form solution for options with stochastic volatility with applications to
bond and currency options. Rev. Financ. Stud. 6(2), 327–343 (1993)
10. Hull, J., White, A.: Pricing interest-rate-derivative securities. Rev. Financ. Stud. 3(4), 573–592
(1990)
11. Kienitz, J., Kammeyer, H.: An implementation of the Hybrid-Heston-Hull-White model.
Available at SSRN 1399389 (2009)
12. Kloeden, P.E., Platen, E.: Numerical Solution of Stochastic Differential Equations. Volume 23
of Applications of Mathematics (New York). Springer, Berlin (1992)
13. Rouah, F.D.: The Heston Model and Its Extensions in Matlab and C. Wiley, Hoboken (2013)
14. Van Haastrecht, A., Lord, R., Pelsser, A., Schrager, D.: Pricing long-maturity equity and FX
derivatives with stochastic interest rates and stochastic volatility. Insur.: Math. Econ. 45(3),
436–448 (2009)
15. Vasicek, O.: An equilibrium characterization of the term structure [reprint of J. Financ.
Econ. 5(2), 177–188 (1977)]. In: Financial Risk Measurement and Management. Volume
267 of International Library of Critical Writings in Economics, pp. 724–735. Edward Elgar,
Cheltenham (2012)
16. Zhu, J.: Applications of Fourier Transform to Smile Modeling, 2nd edn. Springer Finance.
Springer, Berlin (2010) Theory and implementation
Delay Stochastic Models in Finance
Anatoliy Swishchuk
Abstract The paper is devoted to an overview of delay stochastic models in finance

and their applications to modeling and pricing of swaps. The volatility process
is an important concept in financial modeling. This process can be stochastic
or deterministic. In quantitative finance, we consider the volatility process to be
stochastic as it allows to fit the observed market prices under consideration, as
well as to model the risk linked with the future evolution of the volatility, which
deterministic model cannot. Most stochastic dynamical systems (SDS) in finance
describe the state of security (asset or volatility) as a value at time t;-instantaneous
or current value at time t: We would like to take into account not only the current
value at time t; but also values over some time interval Œt ; t; where is a
positive constant and is called the delay. In this way, we incorporate path-dependent
history of the volatility under consideration. In this paper we will mainly focus
on newly developed so-called delayed Heston model that significantly improve
classical Heston model with respect to the market volatility surface fitting by 44 %.
Review of some other delay stochastic models in finance will be given as well.
1 Introduction: Variance and Volatility Swaps
A security is a tradable financial asset of any kind, e.g., stock. A stock is a portion
of ownership in a corporation. If S.t/ is a value of stock at time t; then S.t /
is a delayed stock price. Volatility is a measure for variation of price of a financial
instrument (stock, etc.) over time. If V.t/ is the volatility at time t; then V.t / is
a delayed volatility.
One of the main problems in finance is pricing of volatility and variance (square
of volatility) swaps. A forward contract or simply a forward is a contract between
two parties to buy or sell an asset at a specified future time at a price agreed upon
today. Volatility swaps are defined as forward contracts on future realized stock
volatility. Variance swaps are similar contracts on variance, the square of the future
volatility.
A. Swishchuk ()
University of Calgary, 2500 University Drive NW, Calgary, AB, T2N 1N4, Canada

562 A. Swishchuk
Volatility swaps allow investors to profit from the risks of an increase or decrease
in future volatility of an index of securities or to hedge against these risks. If you
think current volatility is low, for the right price you might want to take a position
that profits if volatility increases.
Payoff of variance swap is defined as
N.VR Kvar /;
where VR is the realized stock variance (quoted in annual terms) over the life of the
contract, Kvar is the strike price, N is the notional amount,
Z
1 T
VR WD Vs ds:
T 0
Price of variance swap is then:
Pvar D EQ ferT .VR Kvar /g;
where r is the risk-free discount rate corresponding to the expiration date T; and EQ
denotes the risk-neutral expectation.
Payoff of volatility swap is defined as
p
N. VR Kvol /;
p
where VR is the realized stock volatility (quoted in annual terms) over the life of
the contract, and
s
p Z
1 T
VR WD Vs ds:
T 0
Price of volatility swap is then:

p
Pvol .x/ D EQ ferT . VR Kvol /g:
If dS.t/ is a change of stock price at time t; then usually
dS.t/ D a.t; S.t//dt C V.t; S.t//dW.t/ C b.t; S.t//dL.t/

(1)
D “drift” C “noise” C “random jumps”;
where a.t; S.t// is a drift coefficient, V.t; S.t// is a diffusion coefficient, b.t; S.t// is
a jump size coefficient, and W.t/ and L.t/ is a Brownian motion and a jump process
(e.g., Lévy process, see [12]) (independent or correlated).
Delay Stochastic Models in Finance 563
There are many papers (see, e.g., [1, 4, 5, 7, 8, 10, 11, 13]) in finance that consider
a a.t; S.t //; V V.t; S.t // or both parameters t and S depending on delay
: In this paper, we will focus on volatility or variance models with delay in finance.
The volatility process V.t; S.t// in (1) is an important concept in financial
modeling. This process can be stochastic or deterministic. In quantitative finance,
we consider the volatility process V.t; S.t// to be stochastic as it allows to fit the
observed market prices under consideration, as well as to model the risk linked with
the future evolution of the volatility, which deterministic model cannot.
If V.t; S.t// is the volatility at time t; then we consider some models for the
change of V.t/ at time t; dV.t/:
dV.t; S.t// D c.t; V.t; S.t///dt C d.t; V.t; S.t///dW.t/ C e.t; V.t; S.t///dL.t/
D “drift” C “noise” C “random jumps”: (2)
In delayed stochastic volatility models the coefficients c; d; e depend on delay

or contain path-dependent history of volatility.
2 Some Delayed Stochastic Models in Finance: An Overview
Significance of path-dependency (a.k.a. delay) may be seen from the Fig. 1 below:
the price of variance swap crucially depends on delay, where we used S&P60
Canada Index as our real data (vertical line stands for delivery price, and two plane
lines stand for jump intensity (left) and for delay (right)).
1. Continuous-time GARCH model for Stochastic Volatility with Delay (see [9]).
We assume in this paper the following model for volatility V.t; S.t//:
Z p 2
dV.t; S.t// ˛ t
D C V.s; S.s//dW.s/ .˛ C /V.t; S.t//:
dt t
Here, all the parameters ˛; ; ; are positive constants and 0 < ˛ C < 1:
All the parameters of this model were inherited from its discrete-time
analogue:
˛ 2
Vn D C ln .Sn1 =Sn1l / C .1 ˛ /Vn1 ; lD ;
l
which, in the special case l D 1, is a well-known GARCH(1,1) model for
stochastic volatility without conditional mean of log-return (see [2, 3]). Here,
Sn is a stock price at time n:
We used this model to find the prices of variance and volatility swaps (see
[14]).
2. Multi-Factor Stochastic Volatility with Delay (see [15]).
564 A. Swishchuk
Dependence of Delivery Price on Delay and Jump Intensity
−4
x 10
5
Delivery Price
1 1
20
15
0.5 10
5
0
Jump Intensity Delay
Fig. 1 Dependence of delivery price on delay and jump intensity (S&P60 Canada Index)
The main idea is to consider continuous-time GARCH model for stochastic

volatility with coefficient above that itself is stochastic and depends on drift
and noise .t; chance/:
In [15], CAMQ, we considered three two-factor and one three-factor stochas-
tic volatility models with delay and we calculated the prices of variance swap for
each model.
One of the examples is:
8 hR p i2
ˆ
< dV.t/
D t C ˛ t
V.s/dW.s/ .˛ C /V.t/:
dt t
:̂
dt =t D dt C ˇdW1 .t/;
where S.t/ is defined from the equation

p
dS.t/ D S.t/dt C V.t/St dW.t/;
St WD S.t / and Wiener processes W.t/ and W1 .t/ may be correlated.

3. Stochastic Volatility with Delay and Jumps (see [16]).
In this paper, the volatility satisfies the following equation:

Z Z
dV.t; S.t// ˛h t p t i2
D C V.s; S.s//dW.s/C ys dN.s/ .˛ C /V.t; S.t//;
dt t t
where W.t/ is a Brownian motion, N.t/ is a Poisson process with intensity and
yt is the jump size at time t. We assume that EŒ yt D A.t/, EŒ ys yt D C.s; t/; s < t
and EŒ y2t D B.t/ D C.t; t/, where A.t/; B.t/; C.s; t/ are all deterministic
functions. We calculated the price of variance swap in this case.
4. Lévy-based Stochastic Volatility with Delay (see [17]). The stochastic volatility
in this paper satisfies the following Lévy-driven equation:
hR p i2
dV.t;S.t// ˛ t
dt
D C t V.u; S.u/dL.u/ .˛ C /V.t; S.t//
where L.t/ is a Lévy process independent of W.t/: Here, > 0 is a mean-

reverting level (or long-term equilibrium of V.t; St /), ˛; > 0; and ˛ C < 1:
We also calculated the price of variance swap in this case.
3 Delayed Heston Model
Classical Heston model (see [6]) is one of the most popular stochastic volatility
models in the industry as semi-closed formulas for vanilla option prices are
available, few (five) parameters need to be calibrated, and it accounts for the mean-
reverting feature of the volatility. In this section we will focus on newly developed
so-called delayed Heston model that significantly improve classical Heston model
with respect to the market volatility surface fitting by 44 %. In this model, we take
into account not only current state of volatility at time t but also its past history over
some interval Œt ; t; where > 0 is a constant and is called the delay. In this way,
our model incorporates path-dependent history for volatility. We will show how to
model and price variance and volatility swaps (forward contracts on variance and
volatility) for the delayed Heston model and how to hedge volatility swaps using
variance swaps (see [18]).
The main points of this section are motivation, advantage and goal:
• Motivation: to include past history (a.k.a. delay) of the variance (over some
delayed time interval Œt ; t);
• Advantage: Improvement of the Volatility Surface Fitting (44 % reduction of the
calibration error) compare with Classical Heston model;
• Goal: to price and hedge volatility swaps.
We consider in a first approach adjusting the Heston drift by a deterministic
function of time so that the expected value of the variance under our new delayed
Heston model is equal to the one under ‘delayed vol’. Our approach can therefore be
566 A. Swishchuk
seen as a variance 1st moment correction of the Heston model, in order to account
for the delay. It is important to note that our model is a generalization of the classical
Heston model (the latter corresponding to the zero delay case D 0 of our model).
We performed numerical tests to validate our approach. With recent market data
(Sept. 30th 2011, underlying EURUSD), we performed the model calibration on the
whole market vanilla option price surface (14 maturities from 1M to 10Y, 5 strikes
ATM, 25 Delta Call/Put, 10 Delta Call/Put). The results show a significant (44 %)
reduction of the average absolute calibration error compared to the Heston model
(i.e. average of the absolute differences between market and model prices).
Further, we consider variance and volatility swaps hedging and pricing in our
delayed Heston framework. These contracts are widely used in the financial industry
and therefore it is relevant to know their price processes (how much they worth at
each time t) and how we can hedge a position on them, i.e. theoretically cancel the
risk inherent to holding one unit of them.
Using the fact that every continuous local martingale can be represented as a
time-changed Brownian motion, as well as the Brockhaus and Long approximation
(that allows to approximate the expected value of the square-root of an almost
surely non negative random variable using a 2nd order Taylor expansion approach),
we were able to derive closed formulas for variance and volatility swaps price
processes. In addition, as variance swaps are relatively liquid instruments in the
market (i.e. they can be easily bought and sold), we considered the question of
hedging a position on a volatility swap using variance swaps in our framework.
We are able to derive a closed formula for the dynamic hedge ratio, i.e. the
number of units of variance swaps to hold at each time in order to hedge a position
on a volatility swap.
In the following subsections we give a list of main steps in pricing of variance
and volatility swaps and hedging of volatility swaps for delayed Heston model.
3.1 Delayed Heston Model for Variance
As we mentioned before, our main motivation is to take into account past history of
the varinace in its diffusion (over some delayed time interval Œt ; t).
The Heston model is one of the most popular stochastic volatility models in the
industry, as semi-closed formulas for vanilla option prices are available, few (five)
parameters need to be calibrated, and it accounts for the mean-reverting feature of
the volatility:
p
dSt D rSt dt C Vt St dWtQ
p
dVt D Œ. 2 Vt /dt C ı Vt dWtQ ;
where St is a stock price, Vt is the stochastic variance (just square of volatility),

WtQ is the Wiener process with respect to the risk-neutral probability Q and r is the
interest rate.
We would like to take into account not only its current state (as it is the case in
the Heston model) but also its past history over some interval Œt ; t; where is
a positive constant and is called the delay. Namely, at each time t; the immediate
future volatility at time t C will not only depend on its value at time t but also on
all its history over Œt ; t: Namely, at each time t, the immediate future volatility
at time t C will not only depend on its value at time t but also on all its history
over Œt ; t: We would like to mention that the non-Markovian continuous-time
GARCH model (see [14])
Z t p
dVt 1
D . 2 Vt / C ˛ . Vs dZsQ . r/ /2 Vt
dt t
is not suitable in this case, as there is no closed-form solution available for this
model. Therefore, we proposed the following Markovian delayed Heston model for
variance (see [18]):
2
p Q
dVt D Œ. Vt / 2C 1.t/dt
R t CQ ı Vt dWt Q (3)
.t/ WD ˛ . r/ C t E .Vs /ds E .Vt / :
We can notice that the classical Heston model and ‘delayed variance’ in (3) are
very similar in the sense that the expected values of the variances are the same –
when we make the delay tends to 0 in ‘delayed variance’. As mentioned before, the
Heston framework is very convenient for practitioners, and therefore it is naturally
tempting to adjust the Heston dynamics in order to incorporate – in some way – the
delay introduced in ‘delayed variance’. We note, that lim !0 supt2RC j .t/j D 0:
3.2 Variance & Volatility Swaps Pricing Techniques
We show in this section how to calculate the variance and volatility swaps.
RT
Consider the realized variance: VR WD T1 0 Vs ds; where Vt is defined in (3), and
suppose that
p
Kvar D EQ ŒVR ; Kvol D EQ Œ VR :
To calculate Kvar we need only EQ ŒVR ; but for calculating Kvol we need more,
due to Brockhaus and Long approximation:
p p VarŒZ
EŒ Z EŒZ ; (4)
8EŒZ3=2
namely, VarŒVR :
568 A. Swishchuk
Using time-changed Brownian motion representation for continuous local mar-

tingales:
if xt WD .V0 2 /e t C e t .Vt 2 /; and
R
dxt D f .t; xt /dWtQ ; xt D W O Tt ; Tt D< x >t D t f 2 .s; xs /ds; where
0
˛ .r/2 ˛
2 WD 2 C ; WD ˛ C C .1 e /; then
Vt D 2 C .V0 2 /e t C e t W

O Tt D EQ ŒVt C e t W
O Tt ; (5)
we get closed formula for Variance Swap and Volatility Swap fair strikes. Formula
(5) for the variance is the time-change representation for the variance in the delayed
Heston model and our main object in finding the variance and volatility swaps.
The parameter 2 in (5) can be interpreted as the delayed-adjusted long-range
variance. We note, that 2 ! 2 as ! 0:
The parameter in (5) can be interpreted as the delayed-adjusted mean-reverting
speed. We note, that ! as ! 0:
3.3 Volatility Swap Hedging
Consider the Price Processes: p

Variance Swap: Xt .T/ WD EtQ ŒVR ; and Volatility Swap: Yt .T/ WD EtQ Œ VR ;
where RT
VR WD T1 0 Vs ds:
Portfolio containing 1 Volatility Swap and ˇt Variance Swaps has the following
form:
˘t D er.Tt/ ŒYt .T/ Kvol C ˇt .Xt .T/ Kvar /

Rt
If It WD 0 Vs ds is the accumulated variance at time t; then:
1
RT
Xt .T/ D EtQ Œq
It
T
C T t Vs ds WD g.t; It ; Vt /
1
RT
Yt .T/ D EtQ Œ It
T C T t Vs ds WD h.t; It ; Vt /
We compute the infinitesimal variations of these processes (using the fact that
Xt .T/ and Yt .T/ are martingales):
@g p
dXt .T/ D @V ı Vt dWtQ ;
@h p
t
dYt .T/ D @Vt ı Vt dWtQ ; (6)
@h @g p
d˘t D r˘t dt C er.Tt/ Œ @V t
C ˇt @V t
ı Vt dWtQ ;
0.19
0.18
Naive Kvol
Kvol
0.17
0.16
0.15
0.14
0.13
0 1 2 3 4 5 6 7 8 9 10
T (year)
Fig. 2 Naive Volatility Swap vs. adjusted Volatility Swap strikes
hence, to keep our portfolio ˘t risk-neutral, we have to eliminate the coefficient at

the random part in (6): )
@h @Yt .T/
@Vt @V
ˇt D @g
t
D @X .T/ (7)
t
@Vt @Vt
-hedge ratio.
We take the parameters that have been calibrated in above-mentioned section
(vanilla options on September 30th 2011 for underlying EURUSD, maturities from
1M to 10Y, strikes ATM, 25D Put/Call, 10D Put/Call), namely, they are:
.v0 ; ; 2 ; ı; c; ˛; / D .0:0343; 3:9037; 108; 0:808;
p 0:5057; 71:35; 0:7821/:
We plot below the naive Volatility Swap strike Kvar and the adjusted Volatility
p
Swap strike Kvar Var .V R/
Q
3 along the maturity dimension, see Fig. 2, as well as the
2
8Kvar
VarQ .VR /
convexity adjustment 3 ; see Fig. 3, respectively. Also, we plot initial hedge
2
8Kvar
ratio ˇ0 .T/ with respect to the formula (7) with t D 0; see Fig. 4.
4 Discussion and Some Open Problems
In this paper is we gave an overview of delay stochastic models in finance and

their applications to modeling and pricing of swaps. Most stochastic dynamical
systems in finance describe the state of security (asset or volatility) as a value at
time t;-instantaneous or current value at time t: We took into account not only the
570 A. Swishchuk
0.025
Convexity Adjustment
0.02
0.015
0.01
0.005
0
0 1 2 3 4 5 6 7 8 9 10
T (years)
Fig. 3 Convexity adjustment (see (4))
–2.4
–2.6
Initial Hedge Ratio
–2.8
–3
–3.2
–3.4
–3.6
–3.8
0 1 2 3 4 5 6 7 8 9 10
T (years)
Fig. 4 Initial Hedge ratio ˇ0 .T/
current value at time t; but also values over some time interval Œt ; t; where
is a positive constant and is called the delay. In this way, we incorporated path-
dependent history of the security (asset or volatility) under consideration. In this
paper we mainly focused on newly developed so-called delayed Heston model that
significantly improve classical Heston model with respect to the market volatility
surface fitting by 44 %. Review of some other delay stochastic models in finance
has been given as well.
There are some very interesting open problems arising from the delay stochastic
models in finance. One of them is modeling and pricing of covariance and
correlation swaps in the case of two underlying assets with delayed Heston model
for variances. Another problem is a comparison of volatility swap or other volatility
derivatives pricing for delayed Heston model and Levy-based stochastic volatility
model Levy-based stochastic volatility model. Also, there are some financial models
that contain fractional Brownian motion as an uncertainty. It would be good to
compare our delay stochastic models with fractional models. We shall leave all these
problems for our future research papers.
References
1. Arriojas, M., Hu, Y., Mohammed, S-E.A., Pap, G.: A delayed Black and Scholes formula.
Stoch. Anal. Appl. 25, 471–492 (2007)
2. Bollerslev, T.: Generalized autoregressive conditional heteroskedasicity. J. Econom. 31, 307–
327 (1986)
3. Bollerslev, T., Chou, R., Kroner, K.: ARCH modeling in finance: a review of the theory and
empirical evidence. J. Econom. 52, 5–59 (1992)
4. Dokuchaev, N.: Modelling possibility of short-term forecasting of market parameters. Ann.
Econ. Finance 16(1), 43–161 (2015)
5. Dibeh, G., Harmanani, H.M.: A stochastic chartist fundamentalist model with time delays.
Comput. Econ. 40(2), 105–113 (2012)
6. Heston, S.: A closed-form solution for options with stochastic volatility with applications to
bond and currency options. Rev. Financ. Stud. 6(2), (1993), 327–343
7. Ivanov, A., Kazmerchuk, Yu., Swishchuk, A.: Theory, stochastic stability and applications of
stochastic differential delay equations. A survey of recent results. Differ. Equ. Dyn. Syst. 11
(1–2), 55–115 (2003)
8. Kazmerchuk, Y.I., Swishchuk, A.V., Wu, J.H.: Black-Scholes Formula Revisited: Security
Markets with Delayed Response, Bachelier Finance Society 2nd World Congress, Crete (2002)
9. Kazmerchuk, Y., Swishchuk, A., Wu, J.-H.: A continuous-time GARCH model for stochastic
volatility with delay. Can. Appl. Math. Q. 13(2), 123–148 (2005)
10. Kazmerchuk, Yu., Swishchuk, A., Wu, J.-H.: The pricing of options for security markets with
delayed response. Math. Comput. Simul. 75, 69–79 (2006)
11. Li, J.-C., Mei, D.-C.: The influences of delay time on the stability of a market model with
stochastic volatility. Physica A: Stat. Mech. Appl. 392(4), 763–772 (2013)
12. Sato, K.-I.: Lévy Processes and Infinitely Divisible Distributions. Cambridge University Press,
Cambridge (1999)
13. Stoica, G.: A stochastic delay financial model. Proc. Am. Math. Soc. 133(6) 1837–1841 (2004)
14. Swishchuk, A.: Modeling and pricing of variance swaps for stochastic volatilities with delay.
WILMOTT Mag. (19), 63–73 (Sept. 2005)
15. Swishchuk, A.: Modeling and pricing of variance swaps for multi-factor stochastic volatilities
with delay. Can. Appl. Math. Q. 14(4), 439–467 (2006)
16. Swishchuk, A., Li, X.: Pricing of variance and volatility swaps for stochastic volatilities with
delay and jumps. Int. J. Stoch. Anal. 27 (2011). http://dx.doi.org/10.1155/2011/435145
17. Swishchuk, A., Malenfant, K.: Variance swaps for local Levy based stochastic volatility with
delay. Int. Rev. Appl. Financ. Issues Econ. 3(2), 432–441 (2011)
18. Swishchuk, A., Vadori, N.: Smiling for the delayed volatility swaps. Wilmott Mag (November
2014)
Semi-parametric Time Series Modelling
with Autocopulas
Antony Ware and Ilnaz Asadzadeh
Abstract In this paper we present an application of the use of autocopulas for

modelling financial time series showing serial dependencies that are not necessarily
linear. The approach presented here is semi-parametric in that it is characterized
by a non-parametric autocopula and parametric marginals. One advantage of using
autocopulas is that they provide a general representation of the auto-dependency
of the time series, in particular making it possible to study the interdependence of
values of the series at different extremes separately. The specific time series that is
studied here comes from daily cash flows involving the product of daily natural gas
price and daily temperature deviations from normal levels. Seasonality is captured
by using a time dependent normal inverse Gaussian (NIG) distribution fitted to the
raw values.
1 Introduction
In this study, autocopulas are used to characterise the joint distribution between
successive observations of a scalar Markov chain. A copula joins a multivariate
distribution to its marginals, and its existence is guaranteed by Sklar’s theorem
[8]. In particular, a Markov chain of first order with any given univariate margin
can be constructed from a bivariate copula. A theoretical framework for the use
of copulas for simulating time series was given by [3], who presented necessary
and sufficient conditions for a copula-based time series to be a Markov process,
but not necessarily a stationary one. They presented theorems specifying when
time series generated using time varying marginal distributions and copulas are
Markov processes. Joe [4] proposed a class of parametric stationary Markov
models based on parametric copulas and parametric marginal distributions. Chen
and Fan [2] studied the estimation of semiparametric stationary Markov models,
This work was supported by MITACS, Direct Energy and by an NSERC Discovery Grant.
A. Ware () • I. Asadzadeh
University of Calgary, 2500 University Drive NW, Calgary, AB T2N 1N4, Canada

574 A. Ware and I. Asadzadeh
using non-parametric marginal distributions with parametric copulas to generate

stationary Markov processes.
The term ‘autocopula’ was first used to describe the unit lag self dependence
structure of a univariate time series in [7], and we adopt the terminology here.
We make use of the framework presented in [3] to produce Markov processes
such that the marginal distribution changes over time. The main benefit of using
autocopulas for univariate time series modelling is that the researcher is able
to specify the unconditional (marginal) distribution of Xt separately from the
time series dependence of Xt [6]. We apply a semiparametic method which is
characterized by an empirical autocopula and a parametric time varying marginal
distribution. This allows us to capture seasonal variations in a natural way. This is
an important feature of our model, motivated by the fact that many financial time
series exhibit seasonality, particularly those arising from energy and commodity
markets.
The remainder of this paper is organized as follows. In Sect. 2, we introduce the
data; Sect. 3 describes the model, including a review of copulas, and of the Normal
Inverse Gaussian distribution. This section also includes details of the calibration
and simulation procedures, and the final section presents some results.
2 The Data
The motivation for this project came from the desire to develop a parsimonious
model that could help to capture so-called load-following (or swing) risk. This is
one of the main sources of financial uncertainty for an energy retailer, and arises
from the combination of retail customer consumption (volume) uncertainty and
price uncertainty. Both volume (V) and price (P) are driven to a large extent by
weather. In particular, average daily temperature is one of the main drivers of daily
natural gas consumption in various North American markets: this in turn drives
market prices through a supply and demand process.
Some of the load-following risk exposure can be hedged using gas forwards and
temperature derivatives. A significant part of the risk exposure that cannot be easily
hedged and is linked to the daily product between the weather deviation from normal
(žF) and the daily price deviation from the expected value of the ex-ante forward
price. To make this more specific, let P denote the last-traded forward monthly
index price, and V the expected monthly average volume. Cash flows for the retailer
depend on the product PV, and the uncertainty in this quantity can be written
.P C P/.V C V/ PV D P V C V P C P V:
We can capture the first-order dependence between volume and weather deviations
by writing V D ˇ W C , where represents higher-order dependence on W
as well as exogenous risk factors; modelling these other contributions was beyond
the scope of this study. Forward instruments in weather and natural gas markets
Semi-parametric Time Series Modelling with Autocopulas 575
Fig. 1 Product of weather and gas price deviations ( P W) in Algonquin over 2003–2014.
Spikes correspond to combinations of high weather deviation from normal and high spot price
deviation from next forward month
can then be used to hedge risks corresponding to the terms P V and V P. It can
be seen that the term P W is an important component of unhedged risk in these
cashflows. One approach to modelling this component would be to develop separate
models for weather and natural gas prices (both daily and forward prices). However,
because of the desire for parsimony, we instead seek a model that allows us to study
the time-series Xt D . P W/t , in order to estimate the range and probabilities of
possible outcomes at the level of a complex portfolio of retail load obligations.
Here we focus on Algonquin Citygate gas prices and the Boston Logan station
for weather data. The data cover the period 1 January 2003–31 June 2014 on a
daily basis, and are shown in Fig. 1. The most dramatic feature of the graph is the
presence of intermittent clusters of spikes, during which the gas prices rise from
their approximate average daily value and at the same time temperature rises or
falls drastically. These mostly occur during winter, although large deviations also
occur at other times of the year. It is also clear that the marginal densities of these
observations will not be well-represented by normal distributions.
3 The Model
Here we introduce the simulation model in more detail, providing a brief review
of copulas, and the normal inverse Gaussian distribution, which we use for the
marginal densities.
3.1 Copulas and Autocopulas
A copula1 is a multivariate distribution function defined on a unit cube Œ0; 1n ,

with uniformly distributed marginals. In the following, we use copulas for the
interdependence structure of time series and, for simplicity and the fact that we
are interested in the first order lag interdependence, we focus on the bivariate case,
although the approach can be used to capture dependence on higher order lags.
Let F12 .x; y/ be the joint distribution function of random variables X and Y
whose marginal distribution functions, denoted as F1 and F2 respectively, are
continuous. Sklar’s theorem specifies that there exists a unique copula function
C.u; v/ D F12 .F11 .u/; F21 .v// that connects F12 .x; y/ to F1 .x/ and F2 .y/ via
F12 .x; y/ D C.F1 .x/; F2 .y//. The information in the joint distribution F12 .x; y/ is
decomposed into that in the marginal distributions and that in the copula function,
where the copula captures the dependence structure between X and Y. Various
families of parametric copulas are widely used (Gaussian, Clayton, Joe, Gumbel
copulas, for example).
In a time series setting, we use a copula (or autocopula) to capture the
dependence structure between successive observations. More generally, we have the
following definition [7].
Definition 1 (Autocopula) Given a time series Xt and L D fli 2 ZC ; i D
1; : : : ; dg a set of lags, the autocopula CX;L is defined as the copula of the d C 1
dimensional random vector .Xt ; Xtl1 ; : : : ; Xtld /.
If a times series Xt is modelled with an autocopula model with unit lag, with
autocopula function C.u; v/ D CX;1 .u; v/, and (time-dependent) marginal CDF
Ft .x/, then, for each t, the CDF of the conditional density of Xt given Xt1 can
be expressed
@C
FXt jXt1 .x/ D Ft1 .Xt1 /; Ft .x/ : (1)
@u
We will discuss issues related to calibration and simulation below.
Autocopula models include many familiar time series as special cases. For
example, it is straightforward to show that an AR(1) process, yt D ˛yt1 CˇC.t/,
can be modelled using theautocopula framework using the marginal distribution
yˇ=.1˛/
F1 .y/ D ˚ p (where ˚ denotes the standard normal CDF) and a
2 =.1˛ 2 /

2 1˛
Gaussian copula with mean D ˇ=.1 ˛/ and covariance 1˛2 .
˛1
Part of the motivation for the use of autocopulas in time series modelling is
that, while correlation coefficients measure the general strength of dependence, they
provide no information about how the strength of dependence may change across
1
For more discussion on the theory of copulas and specific examples, see [5].
Fig. 2 Estimated values of

the quantities C.u; u/=.u/ and
C.u; u/=.1 u/ showing
lower and upper tail
dependence in the observed
values of P W
the distribution. For instance, in the dataset we consider here there is evidence of
tail dependence, whereby correlation is higher near the tails of the distribution. We
can quantify this using the following definition ([4], Section 2.1.10).
Definition 2 (Upper and Lower Tail Dependence) If a bivariate copula C is
such that limu!1 C.u; u/=.1 u/ D U exists, where C.u; u/ D 1 C.1; u/
C.u; 1/ C C.u; u/, then C has upper tail dependence if U 2 .0; 1 and no upper tail
dependence if U D 0. Similarly, if limu!0 C.u; u/=.u/ D L exists, C has lower
tail dependence if L 2 .0; 1 and no lower tail dependence if L D 0
In Fig. 2 we show estimates of the quantities C.u; u/=.u/ and C.u; u/=.1 u/,
where here we use the order statistics of the time series Xt D . P W/t to generate
a preliminary empirical proxy for the copula function C. It is clear from the figure
that neither set of values tends towards zero in the limit u ! 0 or u ! 1, and we
conclude that the data exhibit nonzero tail dependence.
3.2 Time Varying Marginal Distribution
As noted above, the marginal densities for our time series will not be normal.
We found that the normal inverse Gaussian (NIG) distribution provided a more
satisfactory fit. More information about this distribution and its applications can
be found in [1]. Here we review its definition and properties.
3.2.1 Definition and Properties of the NIG Distribution
A non-negative random variable Y has an inverse Gaussian distribution with

parameters ˛ > 0 and ˇ > 0 if its density function is of the form
˛
.˛ ˇy/2
fIG .yI ˛; ˇ/ D p y3=2 exp ; for y > 0:
2ˇ 2ˇy
A random variable X has an NIG distribution with parameters ˛, ˇ, and ı if
XjY D y
N. C ˇy; y/ and Y
IG.ı; 2 /;
p
with WD ˛ 2 ˇ 2 , 0 jˇj < ˛ and ı > 0. We then write X
NIG.˛; ˇ; ; ı/.
Denoting by K1 the modified Bessel function of the second kind, the density is given
by

ı˛ exp ı C ˇ.x / p
fNIG .xI ˛; ˇ; ; ı/ D p K1 ˛ ı 2 C .x /2 :
ı 2 C .x /2
There is a one-to-one map between the parameters of the NIG distribution and the
mean, variance, skewness and kurtosis of the data. We first use moment matching to
determine initial estimates for the parameters, which we use as starting values for
a MLE search method. The corresponding fit to the data is shown in Fig. 3, where
the best fitting normal density is also shown. It can be seen that the NIG fit is quite
good. However, it is evident from Fig. 1 that the time series is strongly seasonal.
We seek to capture this seasonality through the marginal densities by making one of
the parameters of the NIG model time-dependent. This was achieved by assuming
the parameter to be constant in each month, and maximizing the resulting joint
likelihood across the entire data set. Of the four possible choices, ı gave the greatest
improvement to the AIC and BIC, as shown in Table 1. The estimated values of ı
are shown in Fig. 4, and the seasonal pattern that is evident in the original data is
evident again here.
Fig. 3 Histogram of observed data ( P W), with fitted normal distribution and NIG distribution.
The estimated NIG parameters are: ˛ D 0:0980, ˇ D 0:0131, D 0:0122 and ı D 2:3799
Table 1 Calibrated values of Model AIC BIC Log likelihood

AIC, BIC, and Log likelihood
˛ be time varying 34;622:85 35;793:11 17;133:42
ˇ be time varying 36;280:04 37;450:30 17;962:02
be time varying 35;850:14 37;020:40 17;747:07
ı be time varying 33;495:75 34;666:02 16;569:87
Fig. 4 Calibrated monthly values of ı from the combined NIG likelihood, ˛ D 0:0293, ˇ D
0:0205, and D 0:0323
Fig. 5 Calibrated monthly values of ı, together with an example of a simulated path, as well as
a colour contour plot of the quantiles from a large number of simulated paths. Seasonal mean
and variance in (2) are linear combination of a constant and sin 2nt and cos 2nt for increasing
n. The estimated parameters are as follows: a D 0:31, b D .0:93; 0:30; 0:09; 0:22; 0:12,
0:032; 0:01; 0:03; 0:01/, and D .0:15; 0:09; 0:01/
As can be seen in Fig. 4, the value of ı tends to be higher in winter and lower in
summer. The time series of values appears to be mean reverting with seasonal mean
and variance.
p We model the time series using a seasonal mean reverting process for
t D ı t :
tC1 D at C b.t/ C .t/zt ; (2)
where the zt are independent standard normal samples. The mean and variance are
estimated using periodic functions with periods from 1 year down to 3 months.
Simulated and estimated values of ıt are shown in Fig. 5. Twenty thousand paths
were simulated using (2), and for each month the set of values was used to determine
quantiles, which were then used to create the coloured patches shown in the figure.
The darker patches correspond to quantiles nearer to the centre of the distribution,
and the lighter patches to quantiles nearer the extremes.
Once we have values of ıt , we can obtain the time varying cumulative distribution
function and time varying density function. The NIG cumulative distribution
function does not have a closed form solution, so we can compute the CDF using
Gaussian quadrature to evaluate the following integral.
Z xt
F.xt I ˛; ˇ; ; ı/ D fNIG .Xt I ˛; ˇ; ; ıt /dXt (3)
1
In next section we explain the procedure to calculate the empirical autocopulas and
simulate cash flows.
3.3 Estimating the Empirical Autocopula
Having estimated the time-dependent NIG densities, we use these to produce a time
series of values Vt D Ft .Xt / 2 Œ0; 1. If the marginal densities were exact, these
would be uniformly distributed on Œ0; 1. In practice, they will only be approximately
uniform, and we generate an additional empirical marginal density and an empirical
(auto)copula to capture the joint density of .Vt ; Vt1 /.
The empirical autocopula C is estimated by first estimating an empirical joint
density for .Vt ; Vt1 / in the form of a strictly increasing continuous function ˚.; /
that is piecewise bilinear. The domain Œ0; 12 is partitioned into rectangles containing
approximately similar numbers of samples .Vt1 ; Vt /, and taking ˚ to be the
cumulative integral of the sum of indicator functions for these rectangles, scaled by
the number of samples in each rectangle. ˚ is then used to create strictly increasing
piecewise linear marginal densities ˚1 and ˚2 . The inverses of these densities
are therefore also piecewise linear, and when composed with ˚ they generate a
piecewise bilinear copula function C.u1 ; u2 / D ˚ ˚11 .u1 /; ˚21 .u2 / .
This process is illustrated in Fig. 6. In Fig. 6a we plot the pairs of transformed
values ˚1 .Vt1 /; ˚2 .Vt / , together with the outlines of rectangles used to generate
the piecewise bilinear function C. As mentioned, these rectangles contain roughly
equal numbers of points; constructing the empirical autocopula in this way ensures
that it is strictly increasing, and well-suited to enable the computations involved
in time series simulation (see below) to be carried out efficiently. The resulting
empirical autocopula C is shown in Fig. 6b. This function is binlinear on each of the
rectangles shown in Fig. 6a, but is less regular than it looks. The corresponding joint
2C
density, @u@1 @u2
.u1 ; u2 /, is shown in Fig. 6c. It can be seen that the density is higher
near .0; 0/ and near .1; 1/, which is consistent with the tail dependency observed
earlier.
Fig. 6 Generation of the empirical autocopula. (a) Scatter plot of ˚1 .Vt1 / against ˚2 .Vt /. Each
rectangle contains about 100 points (note that rectangles with around 25 points were used in the
simulation). (b) Empirical autocopula C.u1 ; u2 / defined to be bilinear on each of the rectangles
shown in (a). (c) The empirical density @2 C=@u1 @u2 , which is constant on each of the rectangles
shown in (a)
3.4 Simulation of Time Series Using Autocopula
Armed with the time-dependent NIG densities Ft ./, the empirical marginal densi-
ties FV;i ./ and the empirical autocopula C.; /, we can generate simulated values xt
as follows.
1. Given an initial value x0 , generate v0 D F0 .x0 /.
2. For t D 0; 1; : : : , given vt , generate vtC1 :
a. Set u1 D ˚1 .vt /.
b. Given u1 , create the piecewise linear function C.u/ WD C.u1 ; u/=u1 .
c. Set u2 D C1 .U/, where U is an independent uniform random draw.
d. Set vtC1 D ˚21 .u2 /.
3. For each t > 0, set xt D Ft1 .vt /.
Here we have used the fact (already alluded to in (1)) that, if U1 and U2 are uniform
random variables whose joint distribution is the copula C.u1 ; u2 /, then, for u1 > 0,
the cumulative density function for U2 , conditional on U1 D u1 , is
@C C.u1 ; u2 /
PŒU2 < u2 jU1 D u1 D .u1 ; u2 / D :
@u2 u1
The proof of this can be found in, for example, [3]. The fact that C is a piecewise
bilinear function means that C will be piecewise linear. Moreover, the construction
of the empirical copula as described in Sect. 3.3 ensures that it is an increasing
function with a limited number of corners. Its inverse can then be constructed
readily, and will also be an increasing piecewise linear function with a limited
number of corners, and so can be evaluated with little computational effort. Indeed,
in practice the computation of the final step in the above algorithm, the inversion
of the time-dependent NIG densities, took more time than the copula-related
computations.
4 Results
In Fig. 7 we show a 12-year sample time series for P W computed as described

in Sect. 3.4. In addition, we simulated around 700 independent time series and
computed, for each month, the 99th percentile of values produced in that month
across all simulations. What can be seen in the sample path is the same mixture
of quiescent periods and periods with large deviations from zero. There is some
evidence of ‘clumps’ of large deviations occurring in winter months, although this
is less clear than in the original data (see Fig. 1). There is, nevertheless, an increased
occurrence of large deviations in winter months, as can be seen from the plot of the
99th percentiles superimposed on the sample simulation, shown in Fig. 7.
Figure 8 reproduces the data from Fig. 2, with error bars corresponding to the 5th
and 95th percentiles of the values obtained from the simulations, showing that the
simulations have reproduced the observed tail dependence.
Fig. 7 One example of simulated daily values of P W, together with the 99th percentiles of
collected monthly values from around 700 simulations
Fig. 8 Estimated values of the quantities C.u; u/=.u/ and C.u; u/=.1 u/ for the original
observations of P W. Also shown are error bars corresponding to the 5th and 95th percentiles
of the values obtained from around 700 simulations
References
1. Barndorff-Nielsen, O.E., Mikosch, T., Resnick, S.I.: Lévy Processes: Theory and Applications.
Springer Science & Business Media, New York (2001)
2. Chen, X., Fan, Y.: Estimation of copula-based semiparametric time series models. J. Econom.
130(2), 307–335 (2006)
3. Darsow, W.F., Nguyen, B., Olsen, E.T., et al.: Copulas and Markov processes. Ill. J. Math. 36(4),
600–642 (1992)
4. Joe, H.: Multivariate Models and Multivariate Dependence Concepts. CRC Press/Taylor &
Francis Group, Boca Raton/London, New York (1997)
5. Nelsen, R.B.: An Introduction to Copulas, 2nd edn. Springer Science & Business Media,
New York (2007)
6. Patton, A.J.: Copula–based models for financial time series. In: Handbook of Financial Time
Series, pp. 767–785. Springer, Berlin/Heidelberg (2009)
7. Rakonczai, P., Márkus, L., Zempléni, A.: Autocopulas: investigating the interdependence struc-
ture of stationary time series. Methodol. Comput. Appl. Probab. 14(1), 149–167 (2012)
8. Sklar, M.: Fonctions de répartition à n dimensions et leurs marges. Université Paris 8, (1959)
Optimal Robust Designs of Step-Stress
Accelerated Life Testing Experiments
for Proportional Hazards Models
Xaiojian Xu and Wan Yi Huang
Abstract Accelerated life testing (ALT) is broadly used to obtain reliability

information of a product in a timely manner. The Cox’s proportional hazards (PH)
model is often utilized for reliability analysis. In this paper, we focus on designing
ALT experiments for hazard rate prediction when a PH model is adopted. Due to
the nature of prediction made from ALT experimental data, attained under the stress
levels higher than the normal condition, extrapolation is encountered. In such case,
the assumed model can not be tested. For possible imprecision in an assumed PH
model, the method of construction for robust designs is explored. As an example,
we investigate the robust designs for the situation where baseline hazard function
in the fitted PH model is simple linear but the true baseline hazard function is
actually quadratic. The optimal stress-changing times are determined by minimizing
the asymptotic variance and asymptotic squared bias.
1 Introduction
Failure data are needed in order to quantify the life characteristics of a product.
However, under normal design conditions such failure life data are very difficult
to obtain for a product with high reliability within a reasonable time period. To
overcome this problem, accelerate life testing (ALT) has been well-developed to
shorten the lifetime of a product and quickly obtain failures.
In general, test units in accelerated life testing experiments are subjected to
higher than normal design level of stresses. The use of such accelerating stresses for
a particular material is established by engineering practice. In an ALT experiment,
both censoring and time-dependent loading plan are often used. In this paper,
we consider time-censoring and time-dependent step-stress loading plan. Time-
censoring occurs when the experiment stops at a predefined censoring time.
In time-dependent step-stress ALT, all test units are subjected to stress levels
increasing by steps: first, all test units are subjected to a lower stress level for a
X. Xu () • W.Y. Huang

Department of Mathematics and Statistics, Brock University, St. Catharines, ON L2S 3A1,
Canada

586 X. Xu and W.Y. Huang
specified length of time, then the stress level increases and the test continues for all
unfailed test units, and so on.
To avoid invalid estimation, the range of stress levels should be restricted
in a stress loading plan. The failure data obtained at accelerated conditions are
extrapolated to estimate the characteristic of life distribution at normal design
conditions.
In the literature of designing a step-stress ALT experiment, most work has been
done for the cases when a simple step-stress plan is adopted, when the model is
fully parametric, or when model assumed is exactly correct. To mention a few:
Miller and Nelson [9] have first presented Q-optimal designs for simple step-stress
tests with complete data from an exponential distribution. Their designs minimize
the asymptotic variance (AVAR) of the maximum likelihood estimator (MLE) of
the mean life at a normal design stress. Bai et al. [2] have extended the theory of
Miller and Nelson [9] to censored data. They have constructed the optimal simple
step-stress ALT with time-censoring. Bai and Kim [1] have presented an optimal
simple step-stress ALT for a Weibull distribution under time-censoring. The optimal
low stress level and stress change time are obtained by minimizing the AVAR of
the MLE of a specified percentile at normal design stress. Fard and Li [5] have
also presented a step-stress ALT for a Weibull distribution under time-censoring;
however, their optimal stress-changing time is obtained by minimizing the AVAR of
the MLE for reliability instead. Recent work on optimal designs for step-stress ALT
has been reviewed by Hunt and Xu [6]. They have adopted a generalized Khamis-
Higgins model for the effect of changing stress levels and also assumed that the
lifetime of a test unit follows a Weibull distribution; however, their designs are also
optimal when both the shape and scale parameters are considered being functions of
the stress levels. The resulting optimal design chooses the stress-changing time in
order to minimize AVAR of the MLE of reliability at the normal design stress level
and at a pre-specified time.
Ma and Meeker [8] have extended the results of Bai and Kim [1] to provide
a general method for multiple step-stress ALTs for a log location-scale family
of distributions. They have presented an approach to calculate the large-sample
approximate variance of the MLE, computed from step-stress ALT data, for a
quantile of the failure time distribution at normal design conditions. Jiao [7] has
first developed the simple step-stress ALT plan when a PH model is utilized for
reliability prediction. Their optimal stress level is obtained by minimizing the
variance of the MLE of hazard rate at the normal design stress level and over a
pre-specified time period. Elsayed and Zhang [3] have presented an optimal simple
step-stress ALT plan based on a PH model to obtain the most accurate reliability
function estimates at normal design stress. They have also formulated a nonlinear
programming problem to minimize AVAR of the hazard rate estimator over a
prespecified time period at normal design stress.
The general theory for the asymptotic distribution of MLEs under model
misspecification has been derived by White [11]. He has examined the consequences
and detection of model misspecification when the maximum likelihood technique
is used for estimation. The proposed quasi-MLE converges to a well-defined
Optimal Robust Design of ALT 587
limit. In addition, the properties of their quasi-MLE and the information matrix
have been exploited to yield several useful tests when model misspecification is
suspected. Pascual [10] has presented the methodology for deriving the asymptotic
distribution of MLEs of model parameters with constant stress ALT when the stress-
life relationship is misspecified. When possible departures from an assumed ALT
model are suspected, his ALT plans can provide protection against potential bias
without much loss in efficiency. However, the methodology developed there can be
used to derive robust designs only for constant stress ALTs.
In a complete general setting, robust designs for one-point extrapolation which
is the case for ALT have been discussed in Wiens and Xu [12], for least squares
estimation of a mean response. As Fang and Wiens [4] pointed out, “Extrapolation
to regions outside of that in which observations are taken is, of course, an inherently
risky procedure and is made even more so by an over-reliance on stringent model
assumptions.” The classical optimal designs minimize the variance alone. However,
when the fitted models are incorrect, the estimation is biased. A robust design should
be obtained in an optimal way so that even when the fitted model was not exactly
correct, the designs can still be relatively efficient with a small bias.
The present paper extends previous work to construct optimal robust ALT
designs by taking into account multiple step-stress, semi-parametric model assump-
tion, and possible imprecision in the assumed model. Due to the nature of the
prediction made from ALT experimental data, attained under the stress levels
higher than the normal design condition, extrapolation is encountered. For possible
imprecision in an assumed PH model, the method of construction for robust designs
becomes significantly important. Therefore, we consider the situation where a PH
model with a simple linear baseline hazard function is fitted; however, the true
baseline hazard function is a quadratic function. Optimal robust designs will be
obtained in order to protect against possible departure from the assumed model with
a minimum loss of efficiency. We propose a two-stage design procedure, in Sect. 4,
where the optimal stress-changing times are derived by minimizing the asymptotic
squared bias (ABIAS2 ) at the first stage and then minimizing the AVAR at the second
stage.
2 Model Misspecification and Asymptotic Distribution

of MLE Attained
The Cox’s proportional hazards (PH) model is one of the most important means
for predicting the lifetime of a product. This model provides a flexible method for
identifying the effects of the covariates on failure rate. A PH model with a covariate
(applied or transformed stress) independent of time is generally expressed as
.tI s/ D 0 .t/exp.s/; (1)

where 0 .t/ is the baseline hazard rate function at the time of t, s is a stress level
used in an ALT test, and is a unknown parameter.
It is assumed that the true model is Model (1) with a quadratic baseline hazard
function, denoted by MT , i.e., Model (1) with 0 .t/ D 0 C 1 t C 2 t2 and D
˛. However, the fitting model is Model (1) with a linear baseline hazard function,
denoted by MF , i.e., Model (1) with 0 .t/ D 0 C 1 t and D ˇ.
The maximum likelihood method has been used for estimating the model
parameters. Define D Œ0 ; 1 ; 2 ; ˛T and D Œ0 ; 1 ; ˇT . Let ` .I / be the
log-likelihood function under MT , and ` .I / be the log-likelihood function under
MF , both with the same design, . The expected log-likelihood ratio under MT over
MF is
I. W / D EMT Œ`.I / `.I /: (2)
In this paper, we use a 3-step-stress design to demonstrate our design method;

however, the method developed can be applied to any m-step-stress ALT .m 2/.
We let fF .tI s/ and GF .tI s/ be the probability density function and the cumulative
distribution function at failure time t and stress level s under MF , respectively. We
also let c be the censoring time, and Fi .t/ D GF .tI si /, i D1, 2, 3. In addition,
we define a such that the cumulative distribution function under stress level s2 ,
evaluated at a is the same as that under stress level s1 , evaluated at 1 ; and we define
b such that the cumulative distribution function under stress level s3 , evaluated at b
is the same as that under stress level s2 , evaluated at a C 2 1 . For a 3-step-stress
ALT design, the likelihood function from observed lifetimes ti , i D 1; : : : ; n, is
P
n 3iD1 ni
Y
n1 Y
n2 Y
n3 Y
L .I / D fF .ti I s1 / fF .xi I s2 / fF .yi I s3 / .1 GF .zI s3 // ; (3)
iD1 iD1 iD1 iD1
where a D F21 .F1 . 1 //, b D F31 .F2 .a C 2 1 //, x D a Ct 1 , y D b Ct 2 ,

and z D b C c 2 . Given Model (1) with 0 .t/ D 0 C 1 t and D ˇ, we have
h
1 i
fF .tI s/ D .0 C 1 t/ exp.ˇs/ exp 0 t C t2 exp.ˇs/ ;
2
h
1 2 i
GF .tI s/ D 1 exp 0 t C t exp.ˇs/ :
2
The negative log-likelihood function is
X
n1
1
` .I / D n1 ˇs1 ln .0 C 1 ti / 0 ti C ti2 exp .ˇs1 /
iD1
2
n2 h
X 1 i
n2 ˇs2 ln .0 C 1 xi / 0 xi C x2i exp .ˇs2 /
iD1
2
n3 h
X 1 i
n3 ˇs3 ln .0 C 1 yi / 0 yi C y2i exp .ˇs3 /
iD1
2
!
X3 h
1 i
C n ni 0 z C z2 exp .ˇs3 / : (4)
iD1
2
Taking the derivative of (4) with respect to 0 ; 1 and ˇ, we have
Xn1
@` .I / 1
D ti exp .ˇs1 /
@0 iD1
.0 C 1 ti /
n2
X
1
xi exp .ˇs2 /
iD1
.0 C 1 xi /
n3
X
1
yi exp .ˇs3 /
iD1
. 0 C 1 yi /
3
!
X
C n ni Œz exp .ˇs3 / ; (5)
iD1
Xn1
@` .I / ti ti2
D exp .ˇs1 /
@1 iD1
.0 C 1 ti / 2
n2
X
xi x2
i exp .ˇs2 /
iD1
.0 C 1 xi / 2
n3
X
yi y2
i exp .ˇs3 /
iD1
.0 C 1 yi / 2
3
! 2
X z
C n ni exp .ˇs3 / ; (6)
iD1
2
and
1 h
X 1 i
n
@` .I /
D n1 s1 C s1 0 ti C ti2 exp .ˇs1 /
@ˇ iD1
2
n2 h
X 1 i
n2 s2 C s2 0 xi C x2i exp .ˇs2 /
iD1
2
n3 h
X 1 i
n3 s3 C s3 0 yi C y2i exp .ˇs3 /
iD1
2
!
X3 h
1 i
C n ni s3 0 z C z2 exp .ˇs3 / : (7)
iD1
2
T
Let D 0 ; 1 ; ˇ be the value of that minimizes (2), which is the
solutions of 0 ; 1 ; ˇ by setting all equations in (5), (6), and (7) equal to 0: Suppose
that the data are collected under design, , with sample size n: The experimenter
fits Model MF to the data by using MLE method. We adopt quasi-MLE approach
by taking ni ’s to be independent although they are p actually not. We let b denote
the quasi-MLE of . By Theorem 3.2 of White [11], n .b / is asymptotically
normal with mean 0 and covariance matrix C .I / which is defined as
C .I / D ŒA .I /1 B .I / ŒA .I /1 ; (8)
where
2
@ ` .; /
A .I / D EMT (9)
@@T
and

@` .; / @` .; /
B .I / D EMT : (10)
@ @T
3 Optimality Criteria
We consider to estimate the hazard rate over a given time period at the normal design
stress level. We determine the optimal stress-changing times 1 and 2 in order to
minimize the ABIAS2 and AVAR of the MLE of a hazard function average over
a specific period of time, .0; T, under normal design stress level sD . For a given
design, the MLE estimator of the hazard rate at sD can be obtained by:
b

0 C b1 t/ exp b̌sD :
MF .tI sD / D .b (11)
The ABIAS2 of (11) average over .0; T is

Z
ABIAS 2
T
b
MF .tI sD / jMT dt
0
Z 2
D
T
h
b i
EMT MF .tI sD / MT .tI sD / dt
0
Z T

2
0 C 1 t exp2 .ˇ sD / dt ; (12)
0 0 C 1 t C 2 t exp .˛sD /
and the AVAR of (11) average over .0; T is

Z T h
b i
AVAR MF .tI sD / jMT dt
0
1
Z T h iˇ h @b iT ˇˇ
@b
@b @b ˇ @b @b ˇ
D ˇ C WD ˇ dt. (13)
n @b 1 @b̌
0 @b b @b 1 @b̌
0 @b
0 D b D
4 Optimal Robust Designs
In ALT practice, a certain number of failures under each test stress level is often
required. Such requirement is made to avoid that the stress changing times occur
too soon to provide a reasonable step-stress ALT having the same number of stress
levels as planned. Please see Elsayed and Zhang [3] for an example of simple step-
stress ALT. We also take such practical requirement into our design consideration.
These requirements on the minimum number of failures at each stress level are as
follows:
(a) The expected number of failures at stress level s1 has a minimum value MNF1 :
n Pr Œt 1 js1 MNF1 I (14)
(b) The expected number of failures at stress level s2 has a minimum, called MNF2 :
.n n1 / Pr Œa C t 1 2 js2 MNF2 I (15)
(c) The expected number of failures at stress level s3 has a minimum, named MNF3 :
.n n1 n2 / Pr Œb C t 2 c js3 MNF3 ; (16)
where ni is the number of failures under si , i D 1; 2; 3:

Under the constraints of minimum number of failures at each stress level, we

propose a two-stage optimal robust design procedure. At the first stage, we construct
a benchmark design so that the estimation bias can be reduced to a minimal possible.
At the second stage, we update the benchmark design to an optimal robust design in
order to minimize the AVAR.
We illustrate our method of designing a 3-step-stress ALT experiment by
revisiting an example discussed in Elsayed and Zhang [3] with initial values:
0 D 0:0001, 1 D 0:5, 2 D 0:0015, ˛ D 3800, and T D 300: We take s2
be 197.5, the average of low (s1 D 145) and high (s3 D 250) stress levels. All other
values of the parameters in the example remain the same.
We first determine the parameters needed for (12). The value of 0 ; 1 can be
determined by minimizing the absolute distance between quadratic baseline hazard
function and linear baseline hazard function over the whole testing period, namely
Z
T
0 ; 1 D arg min 0 C 1 t C 2 t2 .0 C 1 t/ dt: (17)
0
We have considered five scenarios of the constraints which are listed in Table 1.
At the first stage, the optimal 1 and 2 for our benchmark design can be
determined in order to minimize (12) under each constraint. We denote B Ck as
the optimal design obtained in benchmark under the given constraint Ck , where
k D 1; ; 5. At the second stage, we adopt the benchmark design parameters.
We set the derivatives in (5), (6), and (7) to be zero, and solve for 0 ; 1 ; ˇ. This
is a constrained nonlinear problem. We solve it in an iterative way. First, we fix ˇ
at 3800, and search for f f
0 and 1 in positive ranges of 0 and 1 within their
95 % confidence intervals in order to maximize the log-likelihood function. Then,
B
Ck can be obtained for B D Œf f
0 ; 1 ; 3800 by minimizing (12). Second, with
0 and 1 obtained from the first step, we update our estimate of ˇ to ˇ . By

T
iterating these two steps, we have D 0 ; 1 ; ˇ . Then, the robust design can
be obtained for . We iterate these two steps, until either 0 ; 1 or ˇ remains
unchanged or the difference between its values obtained from the current step and
that from the previous step is sufficiently small within a pre-specified range. Finally,
we can obtain the optimal 1 and 2 based on the most updated estimates for .
Table 1 Constraint cases for 3-step-stress ALT

Names of the Expected number
constraints Constraint parameters of total failures
C1 MNF1 D 40; MNF2 D 20; MNF3 D 10: 70
C2 MNF1 D 40; MNF2 D 15; MNF3 D 15: 70
C3 MNF1 D 30; MNF2 D 30; MNF3 D 10: 70
C4 MNF1 D 30; MNF2 D 20; MNF3 D 20: 70
C5 MNF1 D 40; MNF2 D 10; MNF3 D 10: 60
Table 2 The optimal First stage Second stage Efficiency

stress-changing times for
Design 1 2 Design 1 2 effk .R; B/
both stages
B
C1 160 201 R
C1 160 201 2:0758
B
C2 161 189 R
C2 91 119 2:9840
B
C3 148 198 R
C3 78 128 3:6900
B
C4 147 177 R
C4 77 107 12:4564
B
C5 177 203 R
C5 177 203 1:6514
We denote R Ck as the robust design obtained for three-step-stress ALT under the
given constraint Ck , where k D 1; ; 5. We define the efficiency of R Ck relative to
B
Ck in terms of AVAR as

AVAR B Ck ; B
effk .R; B/ D : (18)
AVAR .R Ck ; /
Our resulting robust designs, R Ck , for both stages and their relative efficiencies are
presented in Table 2.
5 Discussion and Future Research
The example above indicates that our two-stage design procedure has provided
significant efficiency gains. The average efficiency of the optimal robust designs is
as 4.57 times as that of their corresponding benchmarks. The minimum gain is 65 %
for Scenario C5 . At a maximum, our resulting two-stage design for Scenario C4 can
be as 12.46 times efficient as its benchmark (Please see the value of effk .R; B/ in
the second last row of Table 2). Compared C1 to C5 , there only 10 required average
failures less, but in result, the efficiency gain has increased from 65 % to 107 %.
Among the five scenarios considered in the example, the optimal robust designs
obtained at the second stage for both C1 and C5 remains the same as their
corresponding benchmarks attained at the first stage. We note that when the
MNF3 is much lower than MNF1 , the constraint has limited the design space for
further minimizing the estimation variance at the second stage. For other scenarios,
the optimal robust designs obtained after two stages are far different from their
benchmarks. Both stress-changing times are much shorter for the optimal robust
designs than its benchmarks. Consequently, the testing stress levels are increased
more quickly and more failures are observed under higher stress levels, so that the
variability in resulting estimation shall be reduced.
In general, we would recommend that for designing a multiple step-stress ALT,
the experimenters should consider as less restriction on the total number of failures
as possible, at the same time, keep the number of expected failures at each stress
level as even as they can so that the estimation quality can be further improved.
In this paper, we focus on designing a multiple step-stress ALT experiment for

hazard rate estimation; however, the proposed design procedure can be applied
for estimation or prediction of any other quantity of interest, such as reliability
of a product at a given time under the normal design conditions. Furthermore,
the proposed method can also be extended with possible modification to an ALT
experiment involved multiple stress factors. For this situation, a more general
version of (1) can be considered as

.tI s/ D 0 .t/ exp T s ; (19)
where s is a column vector of multiple (applied or transformed) stresses used in

ALT, and is a column vector of unknown parameters. Optimal and robust designs
for step-stress ALT, when (19) is adopted, will be investigated in our future research.
References
1. Bai, D.S., Kim, M.S.: Optimum simple step-stress accelerated life tests for Weibull distribution
and type-I censoring. Nav. Res. Logist. 40, 193–210 (1993)
2. Bai, D.S., Kim, M.S., Lee, S.H.: Optimum simple step-stress accelerated life tests with
censoring. IEEE Trans. Reliab. 38, 528–532 (1989)
3. Elsayed, E.A., Zhang, H.: Design of optimum simple step-stress accelerated life testing plans.
In: Proceedings of the 2nd International IE Conference, Riyadh, Kingdom of Soudi Arabia
(2005)
4. Fang, Z., Wiens, D.P.: Robust extrapolation designs and weights for biased regression models
with heteroscedastic errors. Can. J. Stat. 27, 751–770 (1999)
5. Fard, N., Li, C.: Optimal simple step stress accelerated life test design for reliability prediction.
J. Stat. Plan. Inference 139, 1799–1808 (2009)
6. Hunt, S., Xu, X.: Optimal design for accelerated life testing with simple step-stress plans. Int.
J. Performability Eng. 8, 575–579 (2012)
7. Jiao, L.: Optimal Allocations of Stress Levels and Test Units in Accelerated Life Tests. Rutgers
University, New Brunswick (2001)
8. Ma, H., Meeker, W.Q.: Optimum step-stress accelerated life test plans for log-location-scale
distributions. Nav. Res. Logist. 55, 551–562 (2008)
9. Miller, R., Nelson, W.: Optimum simple step-stress plans for accelerated life testing. IEEE
Trans. Reliab. 32, 59–65 (1983)
10. Pascual, F.G.: Accelerated life test plans robust to misspecification of the stress-life relation-
ship. Technometrics 48, 11–25 (2006)
11. White, H.: Maximum likelihood estimation of misspecified models. Econometrica 50, 1–25
(1982)
12. Wiens, D.P., Xu, X.: Robust designs for one-point extrapolation. J. Stat. Plan. Inference 138,
1339–1357 (2008)
Detecting Coalition Frauds
in Online-Advertising
Qinglei Zhang and Wenying Feng
Abstract Online advertising becomes to play a major role in the global advertising
industry. Meanwhile, since publishers have strong incentives to maximize the
number of views, clicks, and conversions on advertisements, the publisher fraud
is a severer problem for advertisers and worth the endeavor to detect and prevent
them. By reviewing the literature of the frauds in online advertising, the frauds
can be categorized as non-coalition attacks and coalition attacks in general. In this
paper, we attempt to mitigate the problem of coalition frauds by proposing a new
hybrid detecting approach that identifies the coalition frauds from both economic
and traffic perspectives. Moreover, we propose an algorithm to detect the coalition
frauds efficiently with an inductive style and greedy strategy.
1 Introduction
With accordance to a recent report [21], online advertisements in 2014 account for
nearly one-quarter of the global advertising spend. Figure 1 depicts the general
process of the online advertising. The Internet publisher, the advertiser, and the
ad intermediaries (i.e. ad exchangers) are the three key participants in the online
advertising setting. The publishers make money through hosting websites with
advertisements, and advertisers pay for having their ads displayed on publishers’
websites. Ad-exchangers such as Google’s DoubleClick [4] and Yahoo!’s Right-
Media [18] are involved as brokers who pair the publishers’ ad requests with the
most profitable advertiser bid for the request. Since publishers earn revenue based
on the number of views, clicks, and actions that are generated by the advertisement,
publishers have strong incentives to maximize those numbers. While publishers
can apply legitimate methods to attract more traffic to their websites, dishonest
publishers attempt to generate invalid traffic to make more money. In particular,
invalid (fraudulent) traffic is identified as impressions, clicks, or even actions that are
not the result of genuine user interest [1]. The publisher fraud is a severer problem
for advertiser and worth the endeavor to detect and prevent them.
Q. Zhang • W. Feng ()

Trent University, Peterborough, ON, K9J 7B8, Canada

596 Q. Zhang and W. Feng
Fig. 1 The process of online Users

Brands
advertising
Advertisers Publishers
Intermediaries
One of the challenges for detecting the online advertising frauds is to determine
characteristics that signal the potential of fraudulent traffics. Currently, neither
practical systems nor academic research offer a standard mechanism for the fraud
detection in online advertising. Moreover, although using data analysis techniques is
an evident option, existing solutions are far from mature and there are many research
gaps need to be tackled. Among various attacks in online advertising, coalition
frauds are a type of attacks that become more and more prevalent nowadays. Since
such attacks can be launched with lower cost while still difficult to be identified. In
this paper, we attempt to mitigate the problem of coalition frauds by proposing a
new detecting algorithm [7].
The paper is structured as follows. Section 2 reviews the related works of
frauds in online advertising. In Sect. 3, we propose a hybrid detection approach for
coalition frauds. We also discuss the effectiveness and efficiency of the proposed
approach. We conclude and discuss the future work in Sect. 4.
2 Literature Review: Frauds in Online Advertising
With an extensive investigation in both academic and industrial context, we identify

a collection of known attacks in online advertising. In particular, fraudulent publish-
ers use two types of resources, machines, and web pages, to deploy an attack. The
attacks can be performed either by the machines, or by the web pages, or by both.
From the machine side, the attacks can be performed either by real human or by
software. The fraudsters hire malicious users to constantly reload a page and click on
ads. In order to make such attacks harder to trace and detect, some fraudsters abuse
legitimate services to hide identifying information like cookies, IP address, etc.,
from HTTP requests. Fraudsters also resort to bot to simulate impressions, clicks, or
conversions. Initially, those fraudulent bots sit on static machines and issue HTTP
requests to certain URLs, which makes the identification of them relatively easy.
Later on, more sophisticated techniques use anonymous proxies and malware-type
bots to achieve the IP diversity of the traffic. On the side of web pages, the fraudulent
Detecting Coalition Frauds in Online-Advertising 597
publishers can use various of different mechanisms to commit frauds. For example,
impression stuffing and keywords stuffing attacks inflate the number of impressions.
Fraudulent publishers put excessive numbers of banners on their pages to get a large
number of impressions for each page view. In particular, half a dozen or even more
of different ad banners could present on a single page [5]. Sometimes, those banners
are even stacked in front of each other, which make those back banners invisible.
Moreover, to drive higher traffic to the page than it deserves, fraudsters invisibly
use some high-value advertising keywords (e.g. in hidden HTML tags, or by text
with the same color as the backgrounds). Furthermore, the publisher can also inflate
the number of clicks with coercion attacks and force browser attacks. With coercion
attacks, the content of ads are made invisible or modified to something more relevant
to the user, and the users are instruct to perform clicks unknowingly on ads. The
fraudulent publisher inserts additional HTML code that force the browser to click
on the ads. With such client-side script, the users generate request implicitly and
click on the ads stealthily when they loading the webpages.
2.1 The Classification of the Online Advertising Frauds
The classical way to detect publisher frauds is achieved by monitoring the perfor-
mance of advertisements on sites. In other words, fraudulent traffic are identified
as low quality traffic. However, such approach can be too aggressive that even
legitimate traffic can be discarded. In addition, the malicious intents of fraudsters
are not identified directly. Since the fraudsters can control the performance of
advertisements loaded on their sites, they can easily fool the detection tools by
mimicking validity metrics to avoid being detected.
To complement the classical traffic monitoring approach, more advanced tech-
niques are proposed and applied to identify the pattern of fraudulent behaviors.
More specifically, a classification of the fraudulent behaviours help to recognize
a more generic pattern of frauds. A related work with a detailed discussion on the
classification of online advertising frauds can be found in [14]. As aforementioned,
fraudulent publishers use machines and sites to deploy an attack. Regardless
of different attacking mechanisms (e.g. coercion, robots, etc.), the premise for
data analysis approaches to detect publisher frauds is to identify the correlations
between the attacking machines and the corresponding sites. There are two types of
relationships between fraudulent machines and sites. In an one-to-one relationship,
fraudsters only control their own machines and sites. On the other hand, in a many-
to-many relationship, a group of fraudsters share their resources to perform the
attacks. Consequently, the online advertising frauds can be categorized as two types:
the former type of attacks is called the non-coalition attack, while the latter type is
called the coalition attack.
With respect to the above classification, the key to identify fraudulent traffic
is to identify the association between fraudulent websites and machines. More
sophisticated attacks blur the strong correlation between the fraudulent sites and
machines, which makes them difficult to be detected. In the case of non-coalition

attacks, fraudsters only control their own machines and sites. An elementary attack
with repeated views, clicks, or actions from one machine on one site can be easily
detected and blocked. Various data analysis techniques have been proposed to
detect such attacks, such as finding duplicates in data streams [12, 22], identifying
abnormal source IPs that are shared by large number of users [19], and etc. To
compromise those data analysis detection techniques, fraudsters need to develop
sophisticated techniques (e.g., simulate fake visitors, or frequently change the
identification of users) to avoid being detected. In other terms, to increase the
difficulty of being detected, the cost of launching non-coalition attacks will also
increase significantly. On the other hand, since the resources are shared by several
fraudsters, coalition frauds are difficult to be identified while the cost of launching
them is still relatively low.
2.2 Countermeasures for Frauds in Online Advertising
In the literature, various countermeasures have been explored to mitigate the frauds
problem in online advertising. In [3], an alternative pricing model is proposed to
remove the incentive of fraud. However, the approach modifies current pricing
model, which requires a global effort and is not likely to happen in the near future.
Preventing and detection are two general measurements to mitigate fraudulent
problems. In the context of online advertising, it is important to terminate the
identified publishers from re-join the ad-networks. A typical preventing approach
is to create a block list which specifies the signals that indicate fraudulent activities.
By looking at different characteristics of HTTP requests (e.g. IP address, publisher
ID, Customers ID), a potential fraudulent traffic might be filtered based on the
maintained block list. One key challenge of the preventing approach is to determine
characteristics that signal the potential of fraudulent traffics. The signal is consid-
ered as confidential information for the companies that run the ad network.
Detection techniques found in both practice and academic research can be
generally categorized as positive and active detection. The active detection means
that additional interactions are conducted with the web and/or users to verify the
validity of the traffic. Y. Peng et al. [16] proposed an active detection technique,
which combats malicious scripts clickbots by creating and validating an impression-
click identifier. Another example of active detection is using bluff ads [6]. The
strategy of bluff ads is to serve some ads that are purposely uninviting, and test the
legitimacy of the individual click. However, the active detection is usually limited to
address a specific type of fraud, which leads to an arms race between the fraudulent
publishers and the detectors. On the other hand, passive detection mainly depends
on applying data analysis techniques to the aggregate number of traffic logs.
Data mining techniques can be used for both signature-based detection (e.g. [9]),
and anomaly-base detection (e.g. [19]). While signature-based detection works well
for known malicious patterns, anomaly-based detection is more effective to identify
Table 1 Summary of the fraud detection techniques in the literature

Signature-based Anomaly-based
Active [6, 16] Need interactions with
the web and/or users
Positive [9] [8, 10, 12, 13, 15, 17, 19] Use data analysis
and our new techniques to the traffic
hybrid-approach logs
Work well for More effective to identify
known malicious new fraudulent pattern
patterns
fraudsters with updated types of attacks. Related research about passive fraud
detection mainly focuses on developing different algorithms to recognize a specific
behaviour pattern of attacks. For example, the pattern of duplicate clicks within a
short period of time from the same visitor is recognized as one elementary attack.
Algorithms about finding duplicates in data streams such as [12, 19] are applied to
detect such an attack and its variation. Moreover, different features are proposed to
distinguish the fraudulent behaviour patterns from normal behaviours. In [19], the
number of users sharing the source IP is used as a metric to train the normal pattern
of behaviors. Attacks are detected as an anomalous deviation from the expected
publishers’ IP size distribution. Recent efforts [9, 10, 17] have also been done by
using classification algorithms based on multiple features to classify the valid and
invalid traffic. For the coalition attacks, few published work [8, 13, 15, 20] have
proposed algorithms to combat them. The techniques in [13, 15, 20] mainly based
on the similarity among sites to identify the coalitions,while [8] aims to identify the
coalitions from the economic incentive of attackers. We articulate the similarity-
based and the gain-based metrics in detail later and adopt both of them in our
detection technique. A summary of existing detection techniques in the literature
is given in Table 1.
3 Detecting Coalition Frauds in Online Advertising
Unlike non-coalition attacks, the main challenge of detecting coalition attacks is to

identify the correlations of multiple publishers instead of merely focusing on the
pattern of a single publisher.
3.1 Identifying the Coalition Frauds
In order to identify the coalition groups, we can consider the problem from the
economic perspective. In particular, there are two observations for the coalition
groups: (1) the coalition groups inherently have higher ROI (return of investment)
than normal, (2) the expect gain to each member increases while more fraudsters
are involved.
The ratio of gain and cost can be used as an estimator of ROI. Therefore, the
metric relates to the first observation can be expressed by
W.g/
GPR.g/ D GPR ; (1)
R.g/
where W.g/ is the gain derived by the members of the coalition g, R.g/ denotes the
amount of the resources of the coalition, and GPR is a predefined threshold of the
minimum GPR (Gain per Resource) for coalition groups.
A coalition has a high value of GPR, but not vice versa. In other words, a coalition
group with some normal publishers or visitors may still have a high value of GPR.
To capture the property of the second observation, we use the definition of GPR
group. Formally, we define g is a GPR-GPOUP iff 8.g j g0 g W GPR.g0 / <
GPR.g/ /, where g0 denotes any subgroup of g. Therefore, a coalition group can be
defined in terms of GPR as follows:
Definition 1 (gain-based) A group of publishers and visitors, denoted as g, is
called a coalition iff GPR.g/ GPR and g is GPR group. Moreover, a coalition
is a maximal coalition iff : 9.g00 j g g00 W g00 is a coalition /.
As another key character of coalition attacks, the traffic to an attacker’s sites are
from a relatively larger set of attacking machines that shared by many attackers. To
capture the correlations between publishers, we can measure the traffic similarity
among them. Several general similarity measurements can be applied in our context
to capture the similarity metric.
Since the number of repeat visitors is an important indicator for detecting frauds
in online advertising, we use a bag model to denote the visitors to a publisher. Unlike
sets, elements can be duplicated in bags. Let Bp1 and Bp2 denote the bags of visitors
to the publishers p1 and p2 , respectively. The pairs of sites p1 and p2 have similar
traffic iff
jBp1 u Bp2 j
Similarity. p1 ; p1 / D SIM ; (2)
jBp1 t Bp2 j
where u and t denote the intersection and union of bags. We use a pair .n; e/ to
represent an element in a bag, where n indicates how many times the e repeats in
the bag. The definitions of u and t are given as follows:
8..n; e/ j .n; e/ 2 A u B W 9.m1 ; m2 j .m1 ; e/ 2 A ^ .m2 ; e/ 2 B W n D min.m1 ; m2 / / /

8..n; e/ j .n; e/ 2 A t B W ..n; e/ 2 A ^ : 9.k j k 2 N W .k; e/ 2 B //
_..n; e/ 2 B ^ : 9.k j k 2 N W .k; e/ 2 A //
_. 9.m1; m2 j .m1 ; e/ 2 A ^ .m2 ; e/ 2 B W n D max.m1; m2 / / /
The lhs of the Equation (2) is a variant of Jaccard coefficient [2] capturing
the traffic similarity of Bp1 and Bp2 , and SIM on the rhs denotes the minimum
threshold of similarity for legitimate pairs of sites. Since two legitimate sites only
have negligible similarity, it is unlikely that every pair of sites in a random group is
similar. In other terms, the coalition attackers form a group that is pair-wise similar.
Formally, the definition of coalition groups with regard to the traffic similarity is
given as follows:
Definition 2 (similarity-based) A group of publishers g is a coalition group iff
8.p1 ; p2 j .p1 2 g/ ^ .p2 2 g/ ^ .p1 ¤ p2 / W Similarity.p1 ; p2 / SIM /
The aforementioned two approaches capture the key characters of coalition attacks
from different perspectives. The gain-based approach calculates the economic gain
of each group (see Definition 1), while the similarity based approach considers the
traffic similarity within each group (see Definition 2). The reader can refer to [8]
and [15] respectively for more details of these two approaches. However, a main
challenge for both approaches is to define the proper threshold (see Equations 1
and 2). For example, some legitimate popular sites may have equal or even greater
GPR than the coalition groups, and legitimate sites with similar traffic may be falsely
identified as fraudsters. Therefore, to increase the confidence for detecting coalition
attacks, it would be beneficial to consider the characters from both perspectives
and combine both measurements. While the GPR is a measurement for a group of
entities, the similarity is a measurement only for a pair of entities. In order to get
much more confidence on the signal of coalitions, we further give the definition of
coalitions as follows:
Definition 3 (hybrid) Let g denote a subset of all publishers, we say g is a coalition
group iff
8.g0 j g0 < g W GPR.g0 / < GPR.g/ / ^ GPR.g/ GPR

^ 8.p1 ; p2 j .p1 2 g/ ^ .p2 2 g/ ^ .p1 ¤ p2 / W similarity.p1 ; p2 / SIM /
3.2 Detecting the Coalition Frauds
With accordance to the above definition of coalition frauds, we have three strategies
to detect such coalitions. The first strategy is to detect the GPR groups that satisfy
the GPR properties and their GPR value are greater than GPR . Then we verify the
pair-wise similarity property for each identified group. The second strategy is to
calculate the similarity for all pairs and identify all pair-wised similar groups. For
each identified group, we further verify the GPR properties of the group and its
GPR value. The third strategy is to dynamically check the two properties while
identifying the coalition groups. Such strategy will be the most efficient way, which
is critical when dealing with big data. Common advanced analytics disciplines
such as statistical analysis, data mining, and predictive analytics all can be used
to deal with big data. Machine learning algorithms (e.g. neural network, decision
tree) is especially useful for the analysis of big data, since it can explore the
hidden characters with less reliance on human direction. On the other hand, the key
neck-bottle of using machine learning algorithms in our context is the demand of
first extracting related features for all potential groups, which is a computationally
expensive process. However, the adaption of the third strategy is not straightforward.
We should validate that the derived detection algorithm can correctly identify the
frauds corresponding to the definition given in the previous section.
Given a group of publishers and visitors g, let Gclick .g/ and Gsimilarity .g/ denote
the click graph and the similarity graph associated with g, respectively. Before
giving the detection algorithm, we first define two key properties used in the
algorithm as follows:
• GPR-Core property: g is a GPR group H)
8.g0 j Gclick .g0 / is a connected subgraph of Gclick .g/ W g’ is a GPR group /
• similar-clique property: g is a similar clique H)
8.g0 j Gsimilarity .g0 / is a subgraph of Gsimilarity .g/ W g’ is a similar clique /
With accordance to the definition of coalition frauds given in Definition 3, we
have the following proposition:
Proposition 1 A group g is a coalition group iff GPR.g/ GPR and g satisfies
both the CPR-Core property and the similar-clique property.
Due to the limit of space, we omit the proof of the proposition. Algorithm 1
depicts the pseudocode to detect coalition frauds with the hybrid approach. To
combine the aforementioned two criteria dynamically, we apply an inductive style
and greedy strategy in the algorithm, which can dramatically prune the search space
and improve the efficiency of the detection technique. It can be proved that the
GPR-Core and the similar-clique properties are anti-monotone. In other words, the
GPR-Core and the similar-clique properties (used in Line 5 and 6, respectively) hold
with the induction. Consequently, Proposition 1 indicates that the given algorithm
can correctly detect the coalition frauds in our context.
Algorithm 1 Detecting coalition frauds

1: G calculate the base case value for gpr measurement
2: E calculate the pair similarity for the similarity measurement
3: for size in range(2, len(V)) do
4: potential INDUCTIVE_ KEY _ GENERATOR (previous,V)
5: current1 check the potential group with the GPR-Core property
6: current2 check the potential group with the similarity-clique property
7: current current1 \ current2
8: coalition[size] append if it satisfies the GPR group threshold
9: previous coalitions[size]
10: end for
11: max_coalitions MAXIUM.coaltions/
Another key challenge for both gain-based and similarity-based approaches is the
complexity of the algorithm. Given the set of all publishers as P and visitors as V, the
GPR approach require calculating the GPR for all potential subgroups of G, which
complexity is exponential to jPj C jVj. On the other hand, the similarity approach
requires to identify the pairwise similar group which is equivalent to discovering the
maximal cliques in the sites’ similarity graph. In general, finding maximal cliques
in a graph also has exponential complexity. The problem of detecting coalitions is
NP-hard in general. In practice, the problem of detection frauds in online advertising
usually involves processing large sets of data. To further improve the efficiency of
the detection algorithm, we adopt a MapReduce [11] computing model which can
facilitate the parallel, distributed computing on multiple clusters. The lines of 1, 2, 4,
5, 6 and 11 in the pseudocode of Algorithm 1 are implemented with the MapReduce
paradigm. However, the implementation of the algorithm is out the scope of this
paper.
3.3 Preliminary Result
To evaluate the effectiveness of the proposed detecting techniques, we apply our

detection technique to various traffic patterns in this section. As we mentioned
earlier, the value of the threshold is critical when deciding whether or not a group
is a coalition. To avoid the negative false alarms, both the parameters SIM and GPR
should set to low values. However, the positive false alarms could increase when the
values are too low. Existing techniques for detecting coalition approach (e.g. [8, 15])
usually determine the value of the threshold by making a trade-off between the
positive and negative false alarms. Unlike those techniques with a singe metric, the
hybrid approach adapted in this paper could minimize the rate of negative false
alarms by setting the threshold to relatively low values. On the other hand, the rate
of positive false alarms is reduced by combing the decisions from two different
perspectives, Fig. 2 illustrates click graphs of three simplified traffic patterns. The
nodes of p1 , p2 , and p3 denote the publishers, while the nodes of v1 , v2 , v3 , and
a b c
V1 P1 2
P1 P1 V1 V1
V2 P2 V2 P2 2
P2 V4 V2
P3 V3 2
P3 V3 P3 V3
Fig. 2 Different traffic patterns with/without coalition Attacks

Table 2 The results of the three different detection techniques

Similarity-based [15] GPR-based [8] Hybrid (our approach)
Case (a) Fraud (✔) Fraud (✔) Fraud (✔)
Case (b) Frauds (✖) Normal (✔) Normal (✔)
Case (c) Normal (✔) Frauds (✖) Normal (✔)
v4 denote the visitors. The case (a) is an example of coalition attacks that the three
publishers share the resources. The case (b) and case (c) are two examples that the
resources are not really shared by those publishers. However, the two latter cases
are easy to be mis-identified since they show similar traffic metrics as coalition
attacks. Table 2 shows the results of checking the three cases with different detection
techniques. To identify the coalition attack of case (a), the parameter GPR should
be at least 1, and the parameter SIM should be equal to or less than 13 . However,
the case (b) is falsely identified as coalition frauds with SIM D 13 , while the case
(c) is falsely identified with GPR D 1. But with the proposed hybrid approach,
we can avoid such positive false alarms while remaining the precision for detecting
coalition frauds.
4 Conclusions and Future Work
In this paper, we adapt a hybrid approach to identify the coalition frauds from two
distinguish perspectives. To detect the coalition groups, we proposed an algorithm
with inductive styles and greedy strategy, and discuss the correctness and efficiency
of the algorithm. Finally, we evaluate the effectiveness of the proposed detection
technique by applying it to several trial cases.
As further work, we can apply the proposed technique to different data sets to
further evaluate the algorithm. Since there are no benchmark data sets for coalition
frauds in online advertising, we can generate synthetic data sets by simulating the
traffic of both normal and fraudulent traffics. We also can apply the algorithm to
industrial data sets and validate the performance of the algorithm by using third
party services that can label the traffic logs. Moreover, we can improve the precision
of our hybrid detection technique by exploring the coalition frauds problems from
more other perspectives and integrating them to our detection algorithm.
Acknowledgements The authors thank the referee for some helpful comments. The project was
supported by the Mathematics of Information Technology and Complex Systems (MITACS) of
Canada and the EQ Advertising Group Ltd.
References
1. Ad traffic quality resource center. http://www.google.com/intl/en_ALL/ads/adtrafficquality/

index.html. Last accessed on Dec 2015
2. Charika, M.: Similarity estimation techniques from rounding algorithms. In: Proceedings of
the 34th ACM STOC Symposium on Theory of Computing, Montreal, 9pp., 19–21 May 2002
3. Daswani, N., Mysen, C., Rao, V., Weis, S., Gharachorloo, K., Ghosemajumder, S., and the
Google Ad Traffic Quality Team: Online Advertising Fraud. The paper is available at https://
books.google.ca/books?id=U8bh6a9Jvz0C
4. Doubleclick by google. https://www.doubleclickbygoogle.com/. Last accessed on Dec 2015
5. Edelman, B.G.: Securing online advertising: rustlers and sheriffs in the new wild west. In:
Viega, J. (ed.) Beautiful Security, pp. 89–105. O’Reilly Media, Inc., Farnham (2009). Available
at SSRN: http://ssrn.com/abstract=1267311 or http://dx.doi.org/10.2139/ssrn.1267311
6. Haddadi, H.: Fighting online click-fraud using bluff Ads. ACM SIGCOMM Comput. Com-
mun. Rev. 40(2), 21–25 (2010)
7. Jakobsson, M., Ramzan, Z. (ed.): Crimeware: Understanding New Attacks and Defenses
(Symantec Press), pp. 325–354. Addison-Wesley Professional, Boston (2008)
8. Kim, C., Miao, H., Shim, K.: Catch: a detecting algorithm for coalition attacks of hit inflation
in internet advertising. Inf. Syst. 36(8), 1105–1123 (2011)
9. Kitts, B., Zhang, J.Y., Roux, A., Mills, R.: Click fraud detection with bot signatures. In: ISI
2013, Seattle, pp. 146–150 (2013)
10. Landergren, M.H.T.: Implementing best practices for fraud detection on an online advertising
platform. Master’s thesis, Chalmers University of Technology (2010)
11. MapReduce. http://hadoop.apache.org/docs/r1.2.1/mapred_tutorial.html. Last accessed on Dec
2015
12. Metwally, A., Agrawal, D., EI Abbadi, A.: Duplicate detection in click streams. In: Inter-
national World Wide Web Conference, WWW 2005, Chiba, pp. 12–21, 10–14 May. ACM
1-59593-046-9/05/0005 (2005)
13. Metwally, A., Agrawal, D., EI Abbadi, A.: Using association rules for fraud detection in web
advertising networks. In: Proceedings of the 31st VLDB Conference, Trondheim, pp. 169–180
(2005)
14. Metwally, A., Agrawal, D., El Abbadi, A., Zheng, Q.: On hit inflation techniques and
detection in streams of web advertising networks. In: Proceedings of the 27th Interna-
tional Conference on Distributed Computing Systems, Toronto, p. 52, 25–27 June 2007.
doi:10.1109/ICDCS.2007.124
15. Metwally, A., Agrawal, D., EI Abbadi, A.: Detectives: detecting coalition hit inflation attacks
in advertising networks streams. In: Proceedings of the 16th International Conference on World
Wide Web, Banff, pp. 241–250, 8–12 May, 2007. doi:10.1145/1242572.1242606
16. Peng, Y., Zhang, L., Chang, J.M., Guan, Y.: An Effective Method for Combating Malicious
Scripts Clickbots. Lecture Notes in Computer Science, vol. 5789, pp. 523–538. Springer,
Berlin/Heidelberg (2009)
17. Perera, B.K.: A class imbalance learning approach to fraud detection in online advertising.
Master’s thesis, Masdar Institute of Science and Technology (2013)
18. Rightmedia from yahoo! https://advertising.yahoo.com/index.htm. Last accessed on Dec 2015
19. Soldo, F., Metwally, A.: Traffic anomaly detection based on the IP size distribution. In:
Proceedings of the INFOCOM International Conference on Computer Communications, Joint
Conference of the IEEE Computer and Communications Societies, Orlando, pp. 2005–2013,
25–30 Mar 2012
20. Springborn, K., Barford, P.: Impression fraud in online advertising via pay-per-view networks.
In: SEC’13 Proceedings of the 22nd USENIX Conference on Security, Berkeley, pp. 211–226
(2013)
21. Zenithoptimedia. http://techcrunch.com/2014/04/07/. Last accessed on Dec 2015
22. Zhang, L., Guan, Y.: Detecting click fraud in pay-per-click streams of online advertising
networks. In: Proceedings of the 28th International Conference on Distributed Computing
Systems, Beijing, pp. 77–84, 17–20 June 2008. doi:10.1109/ICDCS.2008.98
Part V
New Challenges in Mathematical Modeling
for Scientific and Engineering Applications
Circle Inversion Fractals
B. Boreland and H. Kunze
Abstract In 2000, Frame and Cogevina (Comput Graph 24(5):797–804, 2000)

introduced a method for constructing fractals using circle inversion maps. The
literature (Clancy and Frame, Fractals 3:689–699, 1995; Frame et al., A fractal
representation of grammatical complexity in time series. Manuscript in Preparation,
1996; Frame and Cogevina, Comput Graph 24(5):797–804, 2000; Leys, Comput
Graph 29(3):463–466, 2005) focuses on the graphical aspect of such fractals, with-
out presenting a careful development of the underlying mathematical framework.
In this paper, we present such a framework, making a strong connection to iterated
function systems (IFS) theory. Our final result establishes that the set valued system
of circle inversion maps induced by a collection of non-touching circles in the plane
has a unique set attractor. We present a graphical example using the chaos game.
1 Introduction
Circle inversion was introduced in Apollonius of Pregas’ book, Plane Loci, and has
drawn interest in geometry. Mandelbrot introduced the early concepts of fractals in
the 1970s and, in The Fractal Geometry of Nature, Mandelbrot discusses successive
inversion with respect to a family of M circles. The literature applies the chaos game
to circle inversion maps [2, 4] to explore the graphical aspect, while casually stating
that there are contraction maps involved. As we will see in Sect. 2, a circle inversion
map is only contractive on part of its domain, making clear the need for care. In this
paper, we build the needed rigorous mathematical framework. In Sect. 2, we present
some background concepts on circle inversion maps. In Sect. 3, we establish some
contractivity results for a system of two circle inversion maps, which we extend in
Sect. 4 to a general system of non-touching circles. In Sect. 4, we also establish the
existence of a unique set attractor to the system of set-valued circle inversion maps.
Approximations of this fractal set can be drawn using the chaos game; we provide
an example in Sect. 5. We can use the already established content to now build a
B. Boreland () • H. Kunze


610 B. Boreland and H. Kunze
framework to show why this random inversion algorithm arrives at a fractal, or in

other words, converges to a set attractor.
2 Inversion in a Circle
Let C 2 R2 be a solid circle with centre o and boundary @C; then the following
property is true. e
Property 1 @C divides R2 into three pieces

• the interior of the circle, Cint
• the boundary of the circle, @C
• the exterior of the circle, Cext
where C D Cint [ @C
Definition 1 Given C in R2 with fixed centre o and radius r we define the inversion
map T on R2 n fog by e
e
jjT.x/ ojj jjx ojj D jj˘.x/ ojj2 D r2 ; 8x 2 R2 n fog (1)
e e e e e e e e
where jj jj is the Euclidean norm and ˘.x/ is the projection of x onto @C along the
e write the inversion e
radial ray from o to x. Alternatively, we can map as
e e
d.T.x/; o/d.x; o/ D d 2 .˘.x/; o/ D r2 ; 8x 2 R2 n fog (2)
e e ee e e e e
where d.x; y/ D jjx yjj.
ee e e
Lemma 1 Given a circle C with inversion map T we have the following:
i. T W Cint n fog ! Cext
ii. T W Cext ! eC
int
iii. T W @C ! @C
Proof Follows from the definition of T and Property 1.

When a point y D T.x/ satisfies (2), y is called the inverse of x where T.y/ D x
e e e
and T.b/ D b, foreb 2 @C. e e
e e e
Definition 2 Given a collection of non touching circles Ci , i D 1; : : : ; N, with
centres oi and radii ri we define the inversion map Ti on R2 n foi g by
e e
jjTi .x/ oi jj jjx oi jj D jj˘i .x/ oi jj2 D ri2 ; 8x 2 R2 ; x ¤ oi (3)
e e e e e e e e
where ˘i .x/ is the projection of x onto @Ci along the radial ray from oi to x.
e e e e
Circle Inversion Fractals 611
We can build (3) when x and Ti .x/ satisfy the expression

e e
jj˘i .x/ oi jj2
Ti .x/ D oi C e e .x oi /
e e jjx oi jj2 e e
e e
Any point x 2 R2 can be represented in terms of the parametrization for the circle
Ci : e
x D ari w.t/ C oi for some a 0; t 2 Œ0; 2; (4)

e e e
where w.t/ D .cos.t/; sin.t//. If a < 1 then x 2 Ci;int , if a D 1 then x 2 @Ci and if
e x2C .
a > 1 then e e
i;ext
e 2
For x 2 R n foi g, we have a > 0, and we find that with some easy simplifications
e
we obtain e
jjri w.t/jj2 1
Ti .x/ D Ti .ari w.t/ C oi / D oi C e 2
.ari w.t// D oi C ri w.t/ (5)
e e e e jjari w.t/jj e e a e
e
Thus, we see clearly why Ti is called a circle inversion map: under the action of Ti ,
the radial scaling factor ai > 0 becomes a radial scaling factor of a1i .
3 Mathematical Framework for Two Non-touching Circles
We now develop a mathematical framework for a system of two non-touching

circles. We begin by observing properties along the radial rays. We define some
notation for use in upcoming results. Given a circle C1 in R2 with centre o1 , let
e
Ro1 .o1 ; b1 / D semi infinite ray from o1 through b1 2 @C1 ; with endpoint o1
e e e e e e
Ro1 ;1 .o1 ; b1 / D infinite line through o1 and b1
e e e e e
Ro1 ;int .o1 ; b1 / D Ro1 .o1 ; b1 / \ C1
e e e e e e
Ro1 ;ext .o1 ; b1 / D Ro1 .o1 ; b1 / n fRo1 ;int .o1 ; b1 /; b1 g
e e e e e e e e e e
Lemma 2 For C1 2 R2 with centre o1 , b 2 @C1 , b1 2 @C1 , and T1 the inversion
map for C1 , we have the following: e e e
i. T1 W Ro1 ;int .o1 ; b1 / n fo1 g ! Ro1 ;ext .o1 ; b1 /

T1 W Re e e e e e
ii. o1 ;ext .o1 ; b1 / ! Ro1 ;int .o1 ; b1 /
iii. e
T1 .b/ D b e e e e e
2e e
iv. T1 W Ro1 ;int .o1 ; b1 / n fo1 g ! Ro1 ;int .o1 ; b1 /
e e e e e e
v. T12 W Ro1 ;ext .o1 ; b1 / ! Ro1 ;ext .o1 ; b1 /
e e e e
vi. T12 .a/eD a e
e e
Proof (i), (ii) and (iii) follow from Lemma 1. (iv), (v) and (vi) follow by iterating
Lemma 1 a second time.
Now, given two non-touching circles C1 and C2 in R2 with centres o1 and o2 ,
respectively, define .Ci /int , .Ci /ext , @Ci , Ti and projection maps ˘i . Let e e
Ro1 o2 .x; y/ D portion of the open line segment between o1 and o2 from x to y
ee e e e e e e
Ro1 o2 .oi ; bi / D Ro1 o2 .o1 ; o2 / \ Ci
ee e e ee e e
Ro1 o2 .b1 ; b2 / D Ro1 o2 .o1 ; o2 /nfRo1 o2 .oi ; bi /; i D 1; 2g
ee e e ee e e ee e e
Lemma 3 For i; j 2 1; 2 with i ¤ j, and the preceding set up, we have
i. Ti W Roi oj .oj ; bi / ! Roi oj .oi ; bi /
ii. Ti W Reee e ee e e
oi oj .oi ; bi / ! Roi oj .oj ; bi /
ee e e ee e e
Proof Both (i) and (ii) are easily proved using Lemmas 2(ii) and 2(i), respectively.
It is worth noting that in Lemma 3, the centres of the circles need to special
consideration.
Theorem 1 Let C1 and C2 be two circles in R2 with centres o1 , o2 , respectively.
Then there exists a c 2 Œ0; 1/ such that, e e
d.T1 .x1 /; T1 .x2 // c d.x1 ; x2 /; 8x1 ; x2 2 Ro1 o2 .o2 ; b2 /

e e e e e e ee e e
Proof Introduce the x-axis so that Ro1 o2 .o2 ; b2 / lies along it. This means that x1 , b1
and o1 correspond to numbers alongethe e e we think of T as mapping numbers
e x-axis; e e
1
e
to numbers on the axis. Since x1 > o1 ” T1 .x1 / > o1 , the circle inversion map
along our x-axis is
.b1 o1 /2
T1 .x1 / D o1 C
.x1 o1 /
Now,
0 .b1 o1 /2 0 .b1 o1 /2
T1 .x1 / D 2
H) T1 .x1 /
.x1 o1 / .b2 o1 /2
By the Mean Value Theorem we get

0
jT1 .x1 / T1 .x2 /j D jT1 ./jjx1 x2 j, 2 R
.b1 o1 /2
jx1 x2 j
.b2 o1 /2
D c jx1 x2 j, c < 1
Returning to the vector notation, we have proved there exists a c 2 Œ0; 1/ such that,
d.T1 .x1 /; T1 .x2 // c d.x1 ; x2 /; 8x1 ; x2 2 Ro1 o2 .o2 ; b2 /

e e e e e e ee e e
Theorem 1 leads to the following significant corollary, which follows from two
applications of Theorem 1.
Corollary 1 There exists a c 2 Œ0; 1/ such that,
d..T1 ı T2 /.x1 /; .T1 ı T2 /.x2 // c d.x1 ; x2 /; 8x1 ; x2 2 Ro1 o2 .o1 ; b1 /

e e e e e e ee e e
In Corollary 1, .Ro1 o2 .o1 ; b1 /; j j/ is a complete metric space and T1 ı T2 is a

contraction map fromee .Roe e
1 o2 .o1 ; b1 /; j j/ to itself. Banach’s Fixed Point Theorem
applies: there exists a uniqueexN on
e e e the radial ray in C such that .T ı T /.Nx/ D xN .
1 1 2
e
A similar statement can be made for T2 ı T1 and the radial ray in C2 . Beginning e e
at any point on the ray from o1 to o2 , repeated alternating application of T1 and T2
e e of these two fixed points.
approaches the two cycle consisting
4 Collections of Non-touching Circles
We present the following result in R2 with the understanding that this may be a two
dimensional subspace of a higher dimensional setting. In this result we consider a
collection of non-touching circles in the plane.
Lemma 4 Given a circle C1 with centre o1 , radii r1 and inversion map T1 ,
e
1
d.T1 .x1 /; T1 .x2 // D d.x1 ; x2 /; 8x1 ; x2 2 R2 n fo1 g;
e e a1 a2 e e e e e
where, for j D 1; 2, xj D aj r1 w1 .tj / C o1 for some tj 2 Œ0; 2, aj > 0.

e e e
Proof We consider two cases.
Case 1: x1 ; x2 and o1 lie on the same radial ray. Without loss of generality, we
e
caneplace thee three points on the x-axis at locations x D .a r ; 0/, x D
1 1 1 2
.a2 r1 ; 0/ and o1 D .0; 0/. Then, e e
e
d.T1 .x1 /; T1 .x2 // D jjT1 .x1 / T1 .x2 /jj
e e e e
jj˘1 .x1 /jj2 jj˘1 .x2 /jj2
D e
kx k2 e1 x e x
kx2 k2 e
2
1
2e e
r1 r 2

D 2 2 1 1.a r ; 0/ 1
.a r ; 0/
r1 a1 2 2 2 1
r1 a2
ˇ ˇ
ˇ 1 1 ˇ 1
ˇ
Dˇ r1 ˇˇ D r1 ja2 a1 j
a1 a2 a1 a2
1
D d.x1 ; x2 /
a1 a2 e e
Case 2: x1 ; x2 and o1 are not on the same radial ray. Once again the boundary @C1
canebe parametrized
e e as in (4). There exists t1 ; t2 2 Œ0; 2, t2 > t1 , without
loss of generality, such that
˘1 .xj / D r1 w1 .tj / C o1 and xj D aj r1 .cos.tj /; sin.tj // C o1 ; j D 1; 2:

e e e e e
Note that this case includes the choice t2 D t1 C , two points on the
same diameter but on opposite sides of the centre. After straightforward
calculations, we arrive at
d2 .x1 ; x2 / D a21 r12 C a22 r12 2a1 a2 r12 cos.t2 t1 / (6)

e e
By (5),
1
T1 .xj / D o1 C r1 w.tj /
e e aj e
Thus by using (5), (6) and some simplifying we have,
1 1 2 2
d2 .T1 .x1 /; T1 .x2 // D 2 r12 C 2 r12 r cos.t2 t1 /
e e a1 a2 a1 a2 1
1
D d2 .x1 ; x2 /;
a21 a22 e e
which upon taking the square root gives the desired result.
We will use Lemma 4 and the following definition when establishing the
subsequent contractivity result.
Definition 3 Given a collection of non-touching circles Ci , i D 1; : : : ; N, with
centres oi and radii ri , for x D ai ri wi .t/ C oi 2 R2 , ai > 0, t 2 Œ0; 2, we define
e e e e
ai;min D min fai g > 1
x2Cj ;j¤i
e
corresponding to the radial scaling of the closest point to oi in all other circles Cj ,
j ¤ i. e
Theorem 2 Ti : Cj ! Ci , i ¤ j, is contractive. That is, there exists a c 2 Œ0; 1/

such that,
d.Ti .x1 /; Ti .x2 // c d.x1 ; x2 /; 8x1 ; x2 2 Cj ; i ¤ j

e e e e e e
Proof By Lemma 4 we have that
1 1
d.Ti .x1 /; Ti .x2 // D d.x1 ; x2 / 2 d.x1 ; x2 / D c d.x1 ; x2 /
e e a1 a2 e e ai;min e e e e
Since Ti : Ci;int ! Ci;ext is expansive, we redefine the map as follows.

Definition 4 Given a collection of non-touching circles Ci , i D 1; : : : ; N, with
centres oi , define
e
[
N
XD Ci
iD1
and, relative to each Ci , define
Ti W X ! X
by
(
Ti .x/ if x 2 Cj ; j ¤ i
T i .x/ D e e
e x if x 2 Ci
e e
Theorem 3 T i W X ! X satisfies
.i/ d.T i .x1 /; T i .x2 // c d.x1 ; x2 / for some c 2 Œ0; 1/; 8x1 2 Cj ; x2 2 Ck ; j; k ¤ i
e e e e e e
.ii/ d.T i .x1 /; T i .x2 // D d.x1 ; x2 / 8x1 ; x2 2 Ci
e e e e e e
.iii/ d.T i .x1 /; T i .x2 // c d.x1 ; x2 / for some c 2 Œ0; 1/; 8x1 2 Ci ; 8x2 2 Cj ; j ¤ i
e e e e e e
Proof (i) follows from Theorem 2, (ii) follows immediately since T i is the identity
map in this case. For (iii), with x1 2 Ci and x2 2 Cj , we have a2 > 1 > a1 0, and,
for values t1 ; t2 2 Œ0; 2, e e

d2 .x1 ; x2 / D ri2 a21 C a22 2a1 a2 cos.t2 t1 / ; (7)
e e
using (6), and with some expanding and factoring we are left with
r2
d2 .T i .x1 /; T i .x2 // D i2 a22 a21 C 1 2a1 a2 cos.t2 t1 / (8)
e e a2
The bracketed expression in (8) is less than or equal to the bracketed expression
in (7) provided that
a22 a21 C 1 a21 C a22 ; where a2 > 1 > a1 0: (9)
Hence, letting y D a21 and z D a22 , we consider the function
f .y; z/ D y yz C z:
If we show that f .y; z/ 1 for y 2 Œ0; 1/ and z > 1, then (9) holds. We write
f .y; z/ D y yz C z D y.z 1/ C .z 1/ C 1 D .z 1/.1 y/ C 1
and we see that f .y; z/ > 1 for y < 1 and z > 1, proving the result. This means that
the statement in case (iii) of our theorem holds with c D a12 .
Now, using Theorem 3 we can prove the following contractivity result involving
compositions of (non-touching) circle inversion maps.
Theorem 4 For i ¤ j, T i ı T j : X ! X is contractive:
d..T i ı T j /.x1 /; .T i ı T j /.x2 // c d.x1 ; x2 / for some c 2 Œ0; 1/; 8x1 ; x2 2 X

e e e e e e
Proof If x1 ; x2 2 Ck ; k ¤ j, then the result follows by applying Theorem 3(i) twice.

If x1 ; x2 e 2 Cej , then the result follows by applying Theorem 3(ii) and 3(i). If x1 2
e
Cj ; x2 2 Ck ; k ¤ j, then the result follows by applying Theorem 3(iii) and 3(i). e
e
e
By Theorem 4 and Banach’s Fixed Point Theorem, for each i we can say that
there exists a unique xN j in Ci such that .Ti ı Tj /.Nxj / D xN j for each choice of j. In fact,
by Corollary 1 we know e lieseon the appropriate radial ray,
e that each such fixed point
Roi oj .oi ; bi /.
ee Wee now e formulate this multi-circle framework in the context of sets. For a set
2
A R , define
TO i .A/ D fT i .x/; 8x 2 Ag
e e
Define the distance from a set A to a set B by
d.A; B/ D sup d.x; B/ D sup inf d.x; y/

x2A x2A y2B
The Hausdorff distance between the non-empty compact sets A and B is defined
by
dH .A; B/ D maxfd.A; B/; d.B; A/g
Let H .X/ denote the set of all non-empty compact subsets of X. It is well known
that .H .X/; dH / is complete.
Theorem 5 For i ¤ j,
1. TO i ı TO j W H .X/ ! H .X/
2. TO i ı TO j is contractive on .H .X/; dH /
Proof The first claim follows because T i ı T j W H .X/ ! H .X/. The proof relies
on the continuity of the map ([1], Lemma 2, page 80). By Theorem 4, we can denote
by cij the contractivity factor with respect to d of T i ı T j . For A; B 2 H .X/,
dH ..TO i ı TO j /.A/; .TO i ı TO j /.B//

8 9
< =
D max sup inf d.x1 ; x2 /; sup inf d.x1 ; x2 /
: O O e e x 2.TO ıTO /.B/ x2 2.TOi ıTOj /.A/ e e ;
x1 2.TOi ıTOj /.A/ x2 2.T i ıT j /.B/ 1
e e e
i j
e
(
D max sup inf d..T i ı T j /.x1 /; .T i ı T j /.x2 //;
x1 2A x2 2B e e
e e )
sup inf d..T i ı T j /.x1 /; .T i ı T j /.x2 //
x1 2B x2 2A e e
e( e )
cij max sup inf d.x1 ; x2 /; sup inf d.x1 ; x2 /
x1 2A x2 2B e e x1 2B x2 2A e e
e e e e
D cij dH .A; B/
We reach our final theorem, which establishes the contractivity of the union of
the maps in Theorem 5.
O S
N
Theorem 6 For A 2 H .X/, define T.A/ D .TOi ı TO j /.A/. Then
i;jD1
i¤j
1. TO W H .X/ ! H .X/
2. TO is contractive on .H .X/; dH /
Proof The first claim follows from Theorem 5 because the union is finite. For the
second claim we have
0 1
B[N [
N
C
O
dH .T.A/; O
T.B// D dH B
@ . O ı TO /.A/;
T i j .TO i ı TO j /.B/C
A
i;jD1 i;jD1
i¤j i¤j
max dH ..TO i ı TOj /.A/; .TO i ı TO j /.B//; property of dH

1i;jN
max cij dH .A; B/; by Theorem 5

1i;jN
D c dH .A; B/
Finally, by Theorem 6 and Banach’s Fixed Point Theorem we can conclude that
O satisfying
there exists a unique fixed point A 2 H .X/, the set attractor of T,
O /DA
T.A
5 The Chaos Game
Following Sect. 3, we can use the Chaos Game to plot approximations of the set
attractor A of T. O Choose an initial point x 2 X. With equal probability select one
0
e
of the composition maps T i ı T j , i ¤ j. Apply the map to x0 to produce x1 2 X (in Ci ,
of course). Continue randomly selecting composition maps e to produceethe sequence
1
fxn gnD0 . If we start with a point on A , all points are on A . In practice, we stop after
e large number of iterations, dependent upon the resolution of the plot we are
some
producing.
An example of a fractal generated by a five-circle system is given in Fig. 1.
Fig. 1 Five circles used to generate a set attractor with 10,000 iterations of the chaos game
References
1. Barnsley, M.F.: Fractals Everywhere. Academic, London (1988)

2. Clancy, C., Frame, M.: Fractal geometry of restricted sets of circle inversion. Fractals 3, 689–699
(1995)
3. Frame, M., Clancy, C., Peak, D.: A fractal representation of grammatical complexity in time
series. Manuscript in Preparation (1996)
4. Frame, M., Cogevina, T.: An infinite circle inversion limit set fractal. Comput. Graph. 24(5),
797–804 (2000)
5. Leys, J.: Sphere inversion fractals. Comput. Graph. 29(3), 463–466 (2005)
Computation of Galois Groups in magma
Andreas-Stephan Elsenhans
Abstract We give an outline of a degree independent algorithm for the computation

of Galois groups, as it is implemented in magma (van der Waerden BM, Algebra 1,
Springer, Berlin 1960) (version 2.21). Further, we summarize several performance
tests and list hard examples.
1 Introduction
Given a polynomial f with rational coefficients, one can ask for the Galois group
of the splitting field (viewed as an extension of the rationals). This is a classical
problem in algorithmic algebra. The first algorithm to solve it was described by van
der Waerden in his famous book on algebra [12]. Later, more practical algorithms
were given [2, Sec. 6.3].
A crucial point on the above question is the way one wants to present the
result. One could give the Galois group as an abstract group. But this would be
an incomplete answer, as this does not give the action on the splitting field. On the
other hand, describing the action on the splitting field would require to construct it.
And this would be impractically large in many cases.
A good compromise is to give complex or p-adic root approximations and the
action of the Galois group on them. This is the starting point of the Stauduhar
method [11]. The approach generalizes to any global field.
1.1 Previous Implementations
Several people implemented Galois group algorithms that used precomputed tables
with all the combinatorial data necessary for the computation. Most notable is the
implementation of K. Geißler that covered irreducible polynomials up to degree
23 [7].
A.-S. Elsenhans ()

Mathematisches Institut, Universität Paderborn, Warburger Str. 100, 33098 Paderborn, Germany

622 A.-S. Elsenhans
The first degree independent implementation that could also handle reducible
polynomials was given by C. Fieker and J. Klüners [6]. However, this implementa-
tion ran out of memory or needed hours of CPU time with several polynomials of
degree 30.
Recently, the author worked on bottlenecks of this package. The aim of this
article is to give a summary of what is now possible.
2 Stauduhar’s Step
Definition 1 Let U G Sn be permutation groups. They act on ŒX1 ; : : : ; Xn

by permutation of the variables.
A polynomial I 2 ŒX1 ; : : : ; Xn is called a U-invariant if I D I for all 2 U.
It is called a relative invariant for U G if its stabilizer in G is equal to U.
Lemma 1 Relative invariants exist of all pairs of subgroups U G Sn .
Proof Choose a monomial
P with trivial stabilizer in Sn . E.g., m WD X1 X22 Xn1
n1
.
Then the orbit sum 2U m is a relative invariant. t
u
Definition 2 Let U G be two groups. We denote by G==U a set of coset
representative of G=U.
With this preparation, we can describe the Stauduhar step [11].
Theorem 1 Let f 2 ŒT be a degree n polynomial with distinct roots r1 ; : : : ; rn
in an arbitrary splitting field. Further, let G Sn be a permutation group that
contains the Galois group of f , U be any subgroup of G and I 2 ŒX1 ; : : : ; Xn a
relative invariant for U G. Finally, we assume that the values .I/.r1 ; : : : ; rn / for
2 G==U are pairwise distinct.
Then the Galois group of f is contained in U if and only if .I/.r1 ; : : : ; rn / is
rational.
The above theorem leads to the following algorithm for the computation of
Galois groups.
1. Choose a starting group (e.g., G D Sn ).
2. Compute the conjugacy classes of the maximal subgroups of G.
3. For each subgroup class representative U G do the following:
a. Compute a relative invariant I.
b. Determine .I/.r1 ; : : : ; rn / for 2 G==U.
c. Use the theorem above to find all the conjugates of U that contain the Galois
group.
4. If no subgroup of G is found that contains the Galois group, return G as the
Galois group.
Galois Groups 623
5. Intersect all the subgroups found that contain the Galois group.
6. Redo all the steps above with the intersection as starting group.
Remark 1 There are various problems that have to be solved to make the above
approach practical.
1. First, we need a good strategy to pick an initial group that is as small as possible.
One approach for this is the use of subfields [7, Chap. 5.1]. In case of a reducible
polynomial, one can compute the Galois group of each factor and take the direct
product as a starting group [6].
2. In magma, the computation of maximal subgroups is done using the Cannon-
Holt-algorithm [1].
3. How can we prove the rationality of .I/.r1 ; : : : ; rn / when we only work with
approximations of the roots? One way to do this is described in [7, Chap. 3.3].
This requires to work with p-adic precisions that are proportional to the index of
the subgroup. This is practical only when the index is small.
However, it is possible to work with a moderate p-adic precision and derive a
heuristic result that has to be proven later.
3 The Use of the Frobenius
3.1 Using Cycle Types
Let a monic polynomial f 2 ŒT be given and a prime p that does not divide its
discriminant. It is well known that the local Galois group of the p-adic splitting field
of f is a subgroup of the Galois group of the global splitting field.
The local Galois group is generated by the Frobenius element. When we take the
degrees of the irreducible factors of the reduction of f modulo p, we get the cycle
type of the Frobenius element.
Doing this for several primes, we derive some information about the Galois group
that can be used as follows:
1. If one gets sufficiently many different cycle types then one can prove that the
Galois group is the full symmetric group Sn . More precisely, in the case that
f 2 ŒT is irreducible, n is at least 8 and one has a cycle of prime length l with
n
2
< l < n 2, the alternating group is contained in the Galois group. If, in
addition, an odd permutation is detected then the Galois group is Sn .
2. Before we apply Stauduhar’s step to a subgroup, we can check that it contains all
the cycle types found.
624 A.-S. Elsenhans
3.2 Short Cosets
A more sophisticated way to use the local Galois group is provided by the so called
short cosets [7, Chap. 5.2]. The idea behind them is that, when working with p-adic
root approximations, the Frobenius Frobp is known as an explicit permutation of the
roots.
Thus, a necessary condition for U to contain the Galois group is that it contains
the Frobenius permutation. We call the remaining coset representatives
G==Frobp U WD f 2 G==U j Frobp 2 U g
the short cosets. They can be computed without listing all the representatives G==U
([7, Algorithmus 5.12], [4]).
In many cases, the number of short cosets is very small. A naive explanation for
this is the following:
1 /
Frobp 2 U ” Frob.
p 2U
1
Assuming the conjugates Frob. p
/
to be equidistributed in G, the probability to hit
1
U is ŒGWU . Thus, for large index subgroups this number is very small. Of course, this
is just a very coarse heuristic.
4 The Invariants
To make the Stauduhar method run, we need a relative invariant for each pair of
groups, the algorithm may run into. As there are 25,000 transitive permutation
groups in degree 24, it is not practical to store them in a database. We have to
compute them at run time. A more detailed description of the construction of the
invariants if given in [3, 5–7]. The main tool for this is the usage of block systems.
Definition 3 A partition B1 ; : : : ; Bk of f1; : : : ; ng is called a system of blocks for a
transitive subgroup G Sn if
Bi 2 fB1 ; : : : ; Bk g
for all 2 G.
Theorem 2 Let U be maximal in G Sn and B D fB1 ; : : : ; Bk g be a block system
for U with B1 D f1; : : : ; jg.
P P
1. If B if not a block system for G then B2B . i2B Xi /2 is a relative invariant.
2. If B is a block system of G, denote by 'B the action of G on the blocks.
Galois Groups 625
P P
• If 'B .U/ 6D 'B .G/ then I0 . i2B1 Xi ; : : : ; i2Bk Xi / is a relative invariant.
Here, I0 is a relative invariant for 'B .U/ P'B .G/.
• If StabU .B1 /jB1 ¤ StabG .B1 /jB1 then 2U== StabU .B1 / I0 .X1 ; ::; Xj / is a
relative invariant. Here, I0 is a relative invariant for StabU .B1 /jB1
StabG .B1 /jB1 .
A proof of this is given in [7, Satz 6.14, 6.16].
If these constructions do not apply then we have to do a deeper inspection of the
structure of the permutation groups. Here are a few examples:
Example 1
1. Consider the groups
H D f.1 ; : : : ; 10 / 2 A10 5367
3 j 1 10 D idg Ì S10 D T30
5407
A3 o S10 D T30 D G:
A relative invariant is given by
10
Y
I WD X3i2 C 3 X3i1 C 32 X3i :
iD1
5396 5421
2. The above example has the extension T30 T30 by adding the generator
D .1 2/.4 5/ .28 29/ to both groups.
A relative invariant is given by I C I . In invariant theory, this construction of
an invariant out of an invariant of a subgroup is known as the Reynolds operator.
3. Consider the groups
4831
G D T30 D .=2/15 Ì GL4 .2 / D S2 o GL4 .2 / S2 o A15
and H D T30 3819

D N Ì GL4 .2 / G. Here, N is the Hamming code in 15
2 .
Recall the combinatorial structure of P3 .2 / and the Hamming code:
• P3 .2 / has 15 points and 15 planes.
• The stabilizer of one plane is a subgroup U0 GL4 .2 / of index 15.
• The subgroup U0 decomposes the points into two orbits of size 7 and 8.
• The Hamming code N is given as the subspace of 15 2 . To construct it, we label
the coordinates with the points of P3 .2 /. The linear equations that define N
are
X
Ci D 0
i2P3 . 2 /nE
for all the planes E P3 .2 /.

• The automorphism group of the Hamming code is GL4 .2 /.
626 A.-S. Elsenhans
• We have #G D 660;602;880; ŒG W H D 16.

• The important property of the subgroup U0 is that it stabilizes one of the linear
relations that define the Hamming code N.
The interested reader may consult [10, Sec. 3.3] for a more general introduction
to the Hamming code.
To construct an invariant, we start with the preimage U of U0 in G. In a first
step, we construct an H \ U-invariant that is not a U-invariant by putting
Y
I0 WD .X2 X1 / :
2U==StabU f1;2g
This means, we turn the linear relation of N that is stabilized by U0 into

an invariant. Now, a direct application of the Reynolds operator gives us the
invariant
X
I WD I0
2H==H\U
as a sum of 15 products.
Experiment 1 To get an overview of the complexity of the invariants that may be
used by our code, we use the database of transitive groups up to degree 30 [8] and
apply our invariant constructions to all the maximal subgroups. We use the notation
Tkn for the k-th group of degree n in the database. The construction of all these
invariants took about 1 h.
We use the cost function
.Number of multiplications in one evaluation/ deg.IG;U / min.45; ŒG W U/
to identify hard cases. They are listed in Table 1.

Example 2 When we compute the Galois group of a polynomial with Galois group
T1153
30 , we run into the hardest case listed above.
To get the group of a test polynomial took about 10 s. About 1 s was used in the
Stauduhar step that involves the hard invariant.
The computation was done on one core of an Intel i7-3770 processor running at
3.40 GHz.
Table 1 Pairs of groups with costly invariants

Groups Index deg.I/ Evaluation costs
P
L2 . 8 / D T932 A9 120 6 387 multiplications
M24 D T24;680
24 A24 1,267,136,462,592,000 6 759 powers
T1153
30 T30
4863
20,160 8 5221 multiplications
Galois Groups 627
Experiment 2 Using the database of polynomials [9], we can pick 1954 polynomi-
als in degree 16 and 1117 polynomials in degree 20. One for each transitive group
of degree 16 (resp. 20).
The total running time to get all the Galois groups for the degree 16 polynomials
is 1094 s. The time for all the degree 20 polynomials is 907 s. The slowest examples
are polynomials with large coefficients. They took 5 (resp. 13) s. This is explained
by the larger p-adic precision that has to be used.
A slow example with small coefficients is P455 WD x16 11x8 C 9. The Galois
group of this polynomial is T45516 . It takes 2.2 s to compute it. This is explained by
the fact that our method has to inspect 219 conjugacy classes of subgroups.
The computation was done on one core of an Intel i7-3770 processor running at
3.40 GHz.
5 Proving the Galois Group
The uses of heuristic p-adic precision and the short cosets result in an unproven
Galois group. The first proof algorithm was described in [7, Chap. 5.3]. Our new
proof algorithm is a slight variation.
The idea to confirm the results is that the Galois group predicts the factorizations
of resolvent polynomials. We can compute these resolvents and check that the
prediction is in fact correct. This proves that the Galois group is contained in the
stabilizer of the factorization pattern. These set stabilizers can be computed in
magma. This results in a subgroup of Sn that is proven to contain the Galois group.
We repeat this with several resolvent polynomials until we run into a contradiction
or confirm the heuristic result.
6 Conclusion
The Galois group package of magma (version 2.21) provides a degree independent
implementation of the Stauduhar method. It can compute the Galois group of a
rational degree 30 polynomial with moderate coefficient size in a few seconds on a
modern standard PC. The initial result is heuristic. But, a proof algorithm is available
as well.
References
1. Cannon, J., Holt, D.: Computing maximal subgroups of finite groups. J. Symb. Comput. 37,
589–609 (2004)
2. Cohen, H.: A Course in Computational Algebraic Number Theory. Springer, Berlin/New York
(1993)
628 A.-S. Elsenhans
3. Elsenhans, A.-S.: Invariants for the computation of intransitive and transitive Galois groups. J.
Symb. Comput. 47, 315–326 (2012)
4. Elsenhans, A.-S.: A note on short cosets. Exp. Math. 23, 411–413 (2014)
5. Elsenhans, A.-S.: Improved methods for the construction of relative invariants for permutation
groups (preprint, 2015)
6. Fieker, C., Klüners J.: Computation of Galois groups of rational polynomials. LMS J. Comput.
Math. 17, 141–158 (2014)
7. Geißler, K.: Berechnung von Galoisgruppen über Zahl- und Funktionenkörpern. Dissertation,
Berlin (2003)
8. Hulpke, A.: Constructing transitive permutation groups. J. Symb. Comput. 39(1), 1–30 (2005)
9. Klüners J., Malle G.: A database for field extensions of the rationals. LMS J. Comput. Math.
4, 182–196 (2001)
10. van Lint, J.H.: Introduction to Coding Theory. Springer, Berlin/New York (1992)
11. Stauduhar, R.P.: The determination of Galois groups. Math. Comput. 27, 981–996 (1973)
12. van der Waerden, B.L.: Algebra I. Springer, Berlin (1960)
Global Dynamics and Periodic Solutions
in a Singular Differential Delay Equation
Anatoli F. Ivanov and Zari A. Dzalilov
Abstract Differential delay equation " Œx 0 .t/ C cx 0 .t 1/ C x.t/ D f .x.t 1//

is considered where " > 0 and c 2 R are parameters, and f W R ! R is piece-
wise continuous. For small values of the parameter " a connection is made to the
continuous time difference equation x.t/ D f .x.t 1//; which is further linked to
the one-dimensional dynamical system x 7! f .x/. Two cases of the nonlinearity
f are treated: when it is continuous and of the negative feedback with respect to
a unique equilibrium, and when it is of the so-called Farrey-type with a single
jump-discontinuity. Several properties are studied, such as continuous dependence
of solutions on the singular parameter " and the existence of periodic solutions.
Open problems and conjectures are stated for the case of genuinely neutral equation,
when c ¤ 0.
1 Introduction
Differential delay equations (DDEs) of the form

" x 0 .t/ C cx 0 .t 1/ C x.t/ D f .x.t 1//; t0 (1)
appear as mathematical models of various real world phenomena. They were first
mentioned, to the best of our knowledge, in paper [7] as exact reductions of
nonlinear boundary value problems for one-dimensional wave equations modelling
violin string oscillations. The very same idea of reduction was later used in paper
[4] describing complex oscillations in electrical circuits with tunnel diodes. Note
that in both papers the reduction is to the neutral type DDEs (1), when c ¤ 0.
This is a well known idea and approach that certain boundary value problems
for hyperbolic partial differential equations can be exactly reduced to differential
A.F. Ivanov ()

Department of Mathematics, Pennsylvania State University, Lehman, PA 18627, USA
Z.A. Dzalilov
Federation University Australia, Mt Helen Campus, Ballarat, VIC 3350, Australia

630 A.F. Ivanov and Z.A. Dzalilov
difference equations of various types [5]. The retarded type DDEs (1), when c D 0,
appear in numerous other applications, a partial list of which can be found e.g.
in [1, 3, 6]. This case is more studied compared with the neutral case of c ¤ 0.
Very few results are available in the latter case. This paper is an attempt to initiate
a comprehensive study of neutral differential delay equations (1) in one particular
approach as a singular perturbation problem.
When " D 0 Eq. (1) becomes a continuous time difference equation (DE) of the
form
x.t/ D f .x.t 1//; t 0: (2)
The dynamics of the latter as t ! 1 are largely determined by the interval map f ,
or the equivalent scalar difference equation
xnC1 D f .xn /; n 2 N0 WD N [ f0g: (3)
In this paper we attempt to relate several properties of solutions of DDEs (1) to

the well-studied and understood properties of solutions of difference equations (2)
and (3).
In Sect. 2 we list basic facts and recall necessary knowledge about differential
delay equations (1), difference equations with continuous time (2), and interval
maps associated with scalar difference equation (3). They all can be readily found
in e.g. monographs [1, 2, 5, 6].
Section 3 contains our main results and is made of two subsections. Section 3.1
deals with the continuous dependence on the singular parameter " > 0 between the
solutions of DDE (1) and its limiting case " D 0, the continuous time difference
equation (2). Since the solutions to Eq. (2) are typically discontinuous at tk D k 2
N0 the closeness between the solutions of equations (1) and (2) can be proved
everywhere on finite intervals Œ0; T except a small vicinity of the integer points
tk . In the case when the solutions to (2) are continuous the closeness is uniform
on the entire interval Œ0; T. The precise statements are given by Theorem 1 and its
Corollary. Section 3.2 deals with the existence of periodic solutions to DDE (1) and
their asymptotic shape as " ! 0C. The case of continuous negative feedback and
one-sided bounded f is considered in Sect. 3.2.1. The other case of this paper, when
f is a Farey-type nonlinearity, is treated in Sect. 3.2.2. We show the existence of
the so-called square-wave periodic solutions for small " > 0 in both cases which
correspond to cycles of the map f . The proof is similar in both cases, and it is
outlined for the case of retarded DDEs (1), when c D 0.
Section 4 concerns some open questions and conjectures about the general
neutral DDEs (1) when c ¤ 0. It particular, it is a natural expectation that the
periodicity results of Sects. 3.2.1 and 3.2.2 can be extended to this case. This
conjecture is confirmed by preliminary numerical simulations. At present we do
not have a rigorous proof of the periodicity in the neutral case.
This paper is a result of authors’ presentation at the AMMCS-CAIMS Congress
2015 held at the WLU during 7–12 June 2015. It contains some recently obtained
Global Dynamics and Periodic Solutions in a Singular Differential Delay Equation 631
mathematical results which ideas and proofs are only outlined here, due to their
length and complexity as well as to the space limitation of the proceedings. A full
scale paper with all the mathematical details and proofs included is a forthcoming
work.
2 Preliminaries
For arbitrary '.t/ 2 C1 .Œ1; 0; R/ WD X1 there exists unique solution x"' .t/
to Eq. (1) defined and continuous for all t 1. The solution is differentiable
everywhere for t 1 except at the integer values tk D k 2 N0 ; where it is
typically continuous only. The solution is formally obtained for t > 0 by successive
integration.
Let X0 WD C0 .Œ1; 0/; R/ be a set of continuous initial functions such that the
limit limt!0 .t/ exists. X0 can be viewed as a restriction of the standard space
of initial functions C.Œ1; 0; R/ when its elements are considered on the reduced
domain, the half open interval Œ1; 0/. Thus
X0 WD f 2 C0 .Œ1; 0/; R/ j lim .s/ D .0/ existsg:

s!0
For arbitrary 2 X0 there exists unique solution x .t/ to Eq. (2) defined for all
t 1. The solution is formally obtained for t 0 by successive iterations. The
solution is typically discontinuous at the integer values tk D k 2 N0 with a jump
discontinuity. In the case when the contiguity condition
lim .s/ D f . .1// .cc/

s!0
holds, and f is continuous, the solution x .t/ is also continuous for all t 1.
We consider two distinct cases of the nonlinearity f in DDEs (1):
(H1) f is continuous and satisfies the negative feedback assumption
xf .x/ < 0 8x 2 Rnf0g: .nf /
Also, f is one-sided bounded on R: f .x/ M or f .x/ M for all x 2 R

and some M > 0;
(H2) f is a Farey-type type nonlinearity defined by: f .x/ D mx C A if x 0; and
f .x/ D mx B if x > 0, for some 0 < m < 1 and A; B > 0.
The first case is motivated by a standard set of assumptions normally imposed
for the retarded type DDEs (1) when c D 0; see e.g. [3] and further references
therein. The second case is an exact reduction of a specific nonlinear boundary value
problem for a scalar wave equation [5]. Though many aspects of the dynamics of
equations (1) are studied in the case c D 0, virtually no comprehensive research is
done and very little is known in the neutral case, when c ¤ 0.
3 Main Results
3.1 Continuous Dependence on Singular Parameter "
In this subsection we establish certain continuous dependence results on the singular

parameter " for the differential delay equation (1). We show that its solutions and
solutions of the difference equation with continuous argument (2) are close on finite
intervals in a properly chosen metric for all sufficiently small ". Since the solutions
to Eq. (1) are continuous for all t 0, while the solutions to Eq. (2) are typically
discontinuous at integer values tk D k 2 N0 , we need a proper means of comparison
of those solutions.
We shall use the uniform metric to measure the closeness between two functions
˛.t/ and ˇ.t/ defined on a set S R:
jj˛ ˇjjS WD supfj˛.t/ ˇ.t/j; t 2 Sg:
For the differential delay equation (1) initial functions '1 ; '2 2 C1 .Œ1; 0; R/ D
X1 are compared in the standard uniform metric:
jj'1 '2 jjŒ1;0 D supfj'1 .t/ '2 .t/j C j'10 .t/ '20 .t/j; t 2 Œ1; 0g:
For the difference equation (2) initial functions 1; 2 2 C0 .Œ1; 0/; R/ D X0

are compared in the following uniform metric
jj 1 2 jjŒ1;0/ D supfj 1 .t/ 2 .t/j; t 2 Œ1; 0/g:
Given ' 2 X1 and 2 X0 we shall compare them on the initial interval Œ1; 0 as
follows:
jj' jjŒ1;0 D supfj'.t/ .t/j; t 2 Œ1; 0; where .0/ WD lim .t/g:

t!0
Given two initial functions, ' 2 X1 and 2 X0 , which are close on the initial
interval Œ1; 0, we would like to estimate the closeness between their respective
solutions to Eqs. (1) and (2) on any finite interval Œ0; T. Since x"' .t/ is continuous
for all t 0, while x .t/ is typically discontinuous at tk D k 2 N0 , the closeness
in the uniform metric on the interval Œ0; T cannot happen. However, the closeness
takes place everywhere else, if the integer values tk D k are excluded on the interval
Œ0; T. For the purpose of such comparison we introduce the following notation
ŒT
[
JT WD Œ0; T n U .k/;
kD0
where U .k/ D .k ; k C / is the -neighborhood of the point tk D k, and Œ is

the integer value function.
Theorem 1 Suppose f 2 C.R; R/ satisfies the Lipschitz condition, j f .u/ f .v/j
Kju vjj, and ' 2 X1 is fixed. For arbitrary T > 0; > 0, and > 0 there
exist ı > 0 and "0 > 0 such that if 2 X0 is such that jj 'jjŒ1;0 < ı then
jjx x"' jjJT < for all 0 < " < "0 .
The proof of the theorem follows from the following three lemmas.
Lemma 1 Suppose that 1 ; 2 2 X0 are given, and let x 1 .t/; x 2 .t/ be the
corresponding solutions to difference equation (2). Then
jjx 1 x 2 jjŒ0;1 Kjj 1 2 jjŒ1;0 :
The proof follows from the Lipschitz continuity of function f .

Lemma 2 Suppose that '1 ; '2 2 X1 are given, and let x"'1 .t/; x"'2 .t/ be the
corresponding solutions to differential delay equation (1). Then there exists "0 > 0
such that for all 0 < " < "0 one has
jjx"'1 x"'2 jjŒ0;1 M2 jj'1 '2 jjŒ1;0
for some constant M2 > 0.

Proof The proof is similar to that of Lemma 2 of paper [3], page 185. Solutions to
Eq. (1) also solve the following integral equation
Z
t 1 t st
x.t/ D x.0/ expf g C f .x.s 1// expf g ds (4)
" " 0 "
Z t
st
c x 0 .s 1/ expf g ds:
0 "
The reasoning and major steps of the proof of Lemma 2 should be repeated and
applied now to the integral equation (4). We leave details to the reader. t
u
Lemma 3 Suppose that ' 2 X1 is given, and let x"' .t/ and x' .t/ be the correspond-
ing solutions to equations (1) and (2), respectively. Then for arbitrary > 0 and
> 0 there exists "0 > 0 such that for all 0 < " < "0 one has
jjx"' x' jjŒ;1 :
Proof The proof follows ideas and repeats major steps of the proof of Lemma 3 of
paper [3] when one estimates the difference jx"' .t/ f .x.t 1//j for t 2 Œ0; 1. The
integral equation (4) is used in the new calculations.
By integration by parts the last term of Eq. (4) one arrives at the following integral
equation
t
x.t/ D Œx.0/ C cx.1/ expf g (5)
"
Z
1 t st
C Œ f .x.s 1// C cx.s 1/ expf g ds cx.t 1/:
" 0 "
Using the identity

Z
1 t
st t
F.t/ D F.t/ expf g ds C expf gF.t/
" 0 " "
the difference x"' .t/ x' .t/ D x"' .t/ f .'.t 1//; t 2 Œ0; 1, can be represented as
t
x"' .t/ f .'.t 1// D Œ'.0/ C c'.1/ f .'.t 1// c'.t 1/ expf g
"
Z
1 t st
C Œ f .'.s 1// f .'.t 1// expf g ds
" 0 "
Z
c t st
C Œ'.s 1/ '.t 1/ expf g ds:
" 0 "
The rest of the estimates can be done very similar to those in Lemma 3 of [3]. We
leave details to the reader. t
u
The proof of Theorem 1 now follows by induction and the triangle inequality
jjx x"' jjJT jjx x' jjJT C jjx' x"' jjJT :
Corollary 1 When the contiguity condition (cc) is satisfied Theorem 1 is valid in

the uniform metric on the whole interval Œ0; T.
This essentially follows from the last chain of estimates in Lemma 3 for the
difference x"' .t/ f .'.t 1//; t 2 Œ0; 1, where one can choose D 0.
3.2 Periodicity
This subsection deals with the existence of periodic solutions to differential delay
equation (1). We consider the two cases of the nonlinearity f as described above by
the hypotheses (H1) and (H2).
We first introduce a notion of periodic solutions to difference equation (2)
associated with cycles of the corresponding map f .
Definition 1 (Square wave periodic solutions to DE (2)) Let fa1 ; a2 ; : : : ; aN g be

an N-cycle of map f , f .ak / D akC1 ; k D 1; : : : ; N 1; f .aN / D a1 : Define the
piece-wise constant function by
Sqwv.t; a1 ; : : : ; aN / WD ak ; for t 2 Œk 1; k/; k D 1; 2; : : : ; N;
and extend it periodically to all t 2 R. The function Sqwv.t; a1 ; : : : ; aN /; t 2 R; is

called the square wave periodic solution of difference equation (2) corresponding
to the cycle fa1 ; a2 ; : : : ; aN g.
We define next the convergence of a parameter dependent family of functions to
the discontinuous function Sqwv.t; a1 ; : : : ; aN / on the interval Œ0; N. Let a family
p" .t/ 2 C.Œ0; N; R/ with the parameter " be given.
Definition 2 We say that p" .t/ converges to Sqwv.t; a1 ; : : : ; aN / on the interval
Œ0; N as " ! 0C if for arbitrary > 0 and > 0 there exists "0 > 0 such
that for all 0 < " < "0 one has jjSqwv.; a1 ; : : : ; aN / p" ./jjJN < .
3.2.1 Negative Feedback Nonlinearity
Theorem 2 Assume that (H1) is satisfied and f 0 .0/ < 1. There exists "0 > 0
such that for all 0 < " < "0 DDE (1) has a periodic solution x"p .t/ with the period
T D 2 C O."/ where O."/ ! 0C as " ! 0C . The periodic solution x"p .t/ converges
as " ! 0 to one of the square wave periodic solutions Sqwv.t; a1 ; a2 / of DE (2),
where fa1 ; a2 g is a cycle of period two of the map f .
Proof We outline main steps of the proof for the case c D 0, omitting the
intermediate details due to the space limitation.
Consider the set of initial functions
X D f' 2 C0 .Œ1; 0; R/ j .i/; .ii/; .iii/ are satisfied g;
where the assumptions (i), (ii), and (iii) are specified as follows:
(i) ' is non-negative with '.1/ D 0; '.t/ > 0 8t 2 .1; 0, and ' is bounded
from above, '.t/ M 8t 2 Œ0; 1 and some M > 0;
(ii) ' has an exponential growth in a right neighborhood of t D 1 given by:

.tC1/˛

'.t/ l .t/ WD K 1 e " 8t 2 Œ1; 1 C ˇ;
for some K > 1; 0 < ˛ < 1, and ˇ D O."/;

(iii) ' is bounded away from zero, '.t/ ı 8t 2 Œ1 C ˇ; 0; for some
sufficiently small and fixed ı > 0.
Loosely speaking, the set X consists of non-negative bounded above initial
functions such that all of them start at 0 when t D 1 (assumption (i)), they quickly
grow with the exponential rate in a small right ˇ-neighborhood of t D 1 beyond

the value of ı > 0 (assumption (ii)), and they remain bounded away from 0 by the
value ı for the remainder Œ1 C ˇ; 0 of the initial interval Œ1; 0 (assumption (iii)).
Consider also a negatively symmetric to X set defined by
X WD f 2 C.Œ1; 0; R/j 2 X g:
When c D 0 solutions to Eq. (1) satisfy the following integral equation

Z
t 1 t
st
x.t/ D x.0/ expf g C f .x.s 1// expf g ds: (6)
" " 0 "
It is obtained from the integral equation (4) by setting c D 0.

By using (6) one can make precise forward estimations on the solution x"' .t/ for
' 2 X . One can show that there exists "10 > 0 such that for every ' 2 X and all
0 < " < "10 there exists time t1 D t1 .'/ > 0 such that the corresponding solution
x' .t/ to DDE (1) has the properties:
(a) the segment x' .t1 C 1 C s/; s 2 Œ1; 0; of the solution belongs to X ;
(b) 0 < t1 P1 " for some P1 > 0.
Define now a mapping F 1 on X with the values in X by the formula
F 1 .'/ WD x' .t1 C 1 C s/; s 2 Œ1; 0:
Likewise, there exists "20 > 0 such that for every 2 X and all 0 < " < "20 there
exists time t2 D t2 . / > 0 such that the corresponding solution x .t/ to DDE (1)
has the properties:
(c) the segment x .t2 C 1 C s/; s 2 Œ1; 0; of the solution belongs to X ;
(d) 0 < t2 P2 " for some P2 > 0.
Define a mapping F 2 on X with the values in X by the formula
F 2 . / WD x .t2 C 1 C s/; s 2 Œ1; 0:
For arbitrary ' 2 X and all 0 < " < minf"1 ; "2 g WD "0 define now the mapping
F as the composition of the above mappings F 1 and F 2 , F WD F 2 ı F 1 . F
maps the convex set X into itself. It is a compact map, for when c D 0 the DDE (1)
is a retarded type equation. Also the set F .X / is bounded due to the one-sided
boundedness of the function f . Therefore, by the Schauder fixed point theorem there
exists '0 2 X such that .F /.'0 / D '0 . Clearly, that the corresponding solution
x'0 .t/ is periodic. Its period is T D 2 C t1 C t2 2 C .P1 C P2 /". The convergence
x'0 .t/ ! Sqwv.t; a1 ; a2 /, where fa1 ; a2 g is a 2-cycle of the map f , can be proved
by using the continuous dependence on " (Theorem 1) and the fact that its period
T D 2 C O."/ is close to 2 for small " > 0. Details are omitted due to their length.
t
u
3.2.2 Farey-Type Nonlinearity
Theorem 3 Assume that (H2) is satisfied, and let fa1 ; : : : ; aN g be the unique
globally attracting cycle of the map f . There exists "0 > 0 such that for all
0 < " < "0 DDE (1) has a periodic solution x"p .t/ with the period T D N C O."/,
where O."/ ! 0 as " ! 0. The periodic solution x"p .t/ converges as " ! 0 to the
square wave periodic solution Sqwv.t; a1 ; : : : ; aN / of DE (2).
Proof The idea of the proof is similar to that in the proof of Theorem 2. Associated
with the unique cycle of the map f we construct a sequence of convex bounded sets
Xk C.Œ1; 0; R/; k D 1; : : : ; N; and a sequence of compact maps F k such that
F k W Xk ! XkC1
.N C 1 WD 1/. The composition map F WD F N ı ı F 1 then
maps X1 into itself. By the Schauder fixed point theorem it has a fixed point '0 2 X1

there. The corresponding solution x"'0 .t/ to DDE (1) is periodic. Its properties and
the convergence to Sqwv.t; a1 ; : : : ; aN / as " ! 0 follow from the construction of
the sets Xk outlined below.
Let fa1 ; : : : ; ak ; akC1 ; : : : ; aN g be the globally attracting cycle of the map f with
f .ak / D akC1 ; k D 1; : : : ; N 1 and f .aN / D a1 . In order to be specific in the
definition of Xk we assume that akC1 > ak . Define now Xk as follows:
Xk D f' 2 C0 .Œ1; 0; R/ j .i/; .ii/ are satisfied g;
where the assumptions (i) and (ii) are specified as follows:

(i) ' starts in a neighborhood of ak with a fast exponential transition to akC1
defined as: '.1/ D ak C ı, and l .t/ '.t/ u .t/ 8t 2 Œ1; 1 C ˇ;
where
.tC1/˛ .tC1/
l .t/ D akC1 C .ak C ı akC1 / e " ; u .t/ D M C .ak C ı M/ e "
with ˇ D O."/; 0 < ˛ < 1 and ı > 0 is small and fixed;

(ii) ' stays in the ı-neighborhood of akC1 for the rest of the time on the initial
interval Œ1; 0:
akC1 ı '.t/ akC1 C ı 8t 2 Œ1 C ˇ; 0:
In the opposite case of akC1 < ak the inequalities in part (i) of the definition of Xk
will have to be reversed, as well as the starting point for ' to be moved to ak ı,
'.1/ D ak ı.
By using the integral equation (6) one can show that there exists "k0 > 0 such that
for all 0 < " < "k0 and every ' 2 Xk there exists time tk > 0 for its solution x"' .t/
such that the following holds:

(a) the segment x' .tk C 1 C s/; s 2 Œ1; 0; of the solution belongs to XkC1 ;
(b) 0 < tk Pk " for some Pk > 0.
Therefore, the map F k W Xk ! XkC1

can be defined by
.F k '/.t/ WD x' .tk C 1 C t/; t 2 Œ1; 0:
The composite map F D F N ı ı F 1 has then a fixed point '0 2 X1 which
generates the periodic solution x"'0 .t/ to DDE (1) with the stated properties. t
u
4 Open Problems and Conjectures
The existence of periodic solutions in Theorem 2 is a well known fact due to Hadeler
and Tomiuk [1, 2, 6]. The existence itself only requires the instability of the trivial
solution x 0 and the one-sided boundedness of f . In the case of globally attracting
cycle of period two for the map f , fa1 ; a2 g, the square wave limiting shape of the
periodic solutions was derived by Mallet-Paret and Nussbaum [2, 3]. We prove the
asymptotic shape in the general case of the negative feedback assumption on f . The
questions of uniqueness and asymptotic stability of the periodic solutions remain
largely open problems. In fact, multiple periodic solutions and chaotic behaviors
are shown to exist in the case of globally attracting two-cycle [3]. Therefore, to
establish at least some sufficient conditions for the existence and stability of periodic
solutions for the neutral DDE (1) appears to be quite challenging mathematical
problem.
The theoretical results on periodic solutions of this paper can be proved for
the retarded DDEs (1) at this time. The principal reasons being that the Schauder
Fixed Point Theorem is used, with the essential requirements of the convexity and
boundedness of the sets X ; X ; Xk and of the compactness of the shift operators
F along the solutions of DDE (1) when c D 0. Though the estimates in the proofs
of the Theorems 2 and 3 seem to remain valid also in the case c ¤ 0 we are not
aware of the existence of appropriate fixed point theorems that can be applied in the
neutral case.
Conjecture 1 Theorems 2 and 3 remain valid for the case of neutral DDE (1), when
c ¤ 0.
Acknowledgements The first author would like to express his gratitude and appreciation for the
support and hospitality extended to him during his visit and stay at the CIAO of the Federation
University Australia, Ballarat, in December 2014–January 2015. This paper is a result of the
collaborative research work initiated during the visit.
References
1. Erneux, T.: Applied Delay Differential Equations. Ser.: Surveys and Tutorials in the Applied
Mathematical Sciences, vol. 3, 204 pp. Springer, New York (2009)
2. Hale, J.K., Lunel, S.M.V.: Introduction to Functional Differential Equations. Springer Applied
Mathematical Sciences, vol. 99, 447 pp. Springer-Verlag, New York (1993)
3. Ivanov, A.F., Sharkovsky, A.N.: Oscillations in singularly perturbed delay equations. In: Jones,
C.K.R.T., Kirchgraber, U., Walther, H.O. (eds.) Dynamics Reported (New Series), vol. 1,
pp. 165–224. Springer, New York (1991)
4. Nagumo, J., Shimura, M.: Self-oscillation in a transmission line with a tunnel diode. Proc. IRE
49(8), 1281–1291 (1961)
5. Sharkovsy, A.N., Maistrenko, Yu.L., Romanenko, E.Yu.: Difference Equations and Their
Perturbations, vol. 250, 358 pp. Kluwer Academic Publishers, Dordrecht/Boston/London (1993)
6. Smith, H.: An Introduction to Delay Differential Equations with Applications to the Life
Sciences. Ser.: Texts in Applied Mathematics, vol. 57, 172 pp. Springer, New York/Dor-
drecht/Heidelberg/London (2011)
7. Witt, A.A.: On the theory of the violin string. Zhurn. Tech. Fiz. 6, 1459–1470 (1936, in Russian)
Localized Spot Patterns on the Sphere for
Reaction-Diffusion Systems: Theory and Open
Problems
Alastair Jamieson-Lane, Philippe H. Trinh, and Michael J. Ward
Abstract A new class of point-interaction problem characterizing the time evolu-

tion of spatially localized spots for reaction-diffusion (RD) systems on the surface of
the sphere is introduced and studied. This problem consists of a differential algebraic
system (DAE) of ODEs for the locations of a collection of spots on the sphere,
and is derived from an asymptotic analysis in the large diffusivity ratio limit of
certain singularly perturbed two-component RD systems. In Trinh and Ward (The
dynamics of localized spot patterns for reaction-diffusion systems on the sphere.
Nonlinearity Nonlinearity 29(3), 766–806 (2016)), this DAE system was derived for
the Brusselator and Schnakenberg RD systems, and herein we extend this previous
analysis to the Gray-Scott RD model. Results and open problems pertaining to the
determination of equilibria of this DAE system, and its relation to elliptic Fekete
point sets, are highlighted. The potential of deriving similar DAE systems for more
complicated modeling scenarios is discussed.
1 Introduction
Spatially localized patterns can occur for two-component reaction-diffusion (RD)

systems in the singularly perturbed limit corresponding to a large diffusivity ratio
between the two components in the system. In particular, on the surface of the unit
sphere, the stability and dynamics of localized spot patterns, whereby the solution
concentrates at discrete points on the sphere, have been analyzed recently for the
singularly perturbed Brusselator and Schnakenberg RD systems (cf. [24, 27]). It was
also shown that, in certain parameter regimes, these localized patterns can exhibit
various instabilities, including either spot-shape instabilities, leading to spot self-
replication events, or competition instabilities, leading to the annihilation of spots
A. Jamieson-Lane • M.J. Ward ()

Department of Mathematics, University of British Columbia, Vancouver, BC, Canada
P.H. Trinh
OCIAM, Mathematical Institute, University of Oxford, Oxford, UK

642 A. Jamieson-Lane et al.
in the pattern. This analysis of [24] and [27] extends the previous studies of localized
spot patterns on 2-D planar domains for related RD systems (cf. [6, 11, 29]).
By using formal asymptotic methods on the RD system in the singularly
perturbed limit, it is possible to derive a differential algebraic ODE system char-
acterizing the dynamics of spot patterns. This new class of dynamically interacting
particle system has some common features with the well-known ODE system
characterizing the dynamics of Eulerian point vortices in fluid mechanics. In this
latter context, there has been an intense study of the dynamics and equilibria of
point vortices on the sphere over the past three decades (cf. [2, 3, 10, 18, 19, 23]).
Similar to the original asymptotic derivation of the limiting point vortex problem
in [3] starting with the Euler equations of fluid mechanics, the main result for spot
dynamics in [27] for the Brusselator and Schnakenberg model, and herein for the
Gray-Scott RD model, provides a reduced dynamical system for the time evolution
of the centres of the localized spots on the sphere.
Motivated by specific questions related to the development of biological patterns
on both stationary and time-evolving surfaces (cf. [5, 12, 17, 20, 21]), there have
been many numerical studies of RD patterns on the sphere and other compact
manifolds (cf. [1, 13, 14, 28] see also the references therein). Prior analytical
studies of RD pattern formation on the surface of the sphere, have focused on
using weakly nonlinear and equivariant bifurcation theory to derive normal form
equations characterizing the development of small amplitude spatial patterns that
bifurcate from a spatially uniform steady-state (cf. [4, 7, 16]). However, as a result
of the typical high degree of degeneracy of the eigenspace associated with spherical
harmonics of large mode number, these amplitude equations typically consist of a
rather large coupled set of nonlinear ODEs. The latter are known to have an intricate
subcritical bifurcation structure (cf. [4, 7, 16]). As a result, the preferred spatial
pattern that emerges from an interaction of these weakly nonlinear modes is difficult
to predict theoretically. This intrinsic difficulty is accentuated for RD systems where
there is a large diffusivity ratio, which effectively yields a large aspect ratio system
where center manifold analysis is of more limited use [25]. For such large aspect
ratio RD systems, there is typically a rather wide band of unstable modes [24, 27],
and so the prediction of pattern development based on the conventional paradigm of
using both a Turing and a weakly nonlinear analysis is not generally possible.
However, it is in this singular limit of a large diffusivity ratio that localized
spot patterns robustly appear from a transient process starting from small random
perturbations of a spatially uniform state [24]. A discussion of results and open
problems relating to the study of such “far-from equilibrium patterns” is the topic
of this short article. In particular, in certain cases the equilibria of this DAE
system for spot dynamics have a close relationship to the classical problem in
approximation theory of determining a set of elliptic Fekete points, which are the
globally minimizer of the discrete logarithmic energy for N points on the sphere.
Dynamics and Equilibria of Localized Spots on the Sphere 643
The outline of this brief article is as follows. In Sect. 2, we briefly present the
DAE system for the dynamics of spots for the Brusselator models as derived in
[27]. In Sect. 3 we discuss some results and open questions related to determining
equilibria for these DAE systems. New results for the equilibria of patterns with
either 9 or 10 spots are presented. In Sect. 4 we give a new result for the DAE
dynamics for spot patterns for the well-known Gray-Scott model. Finally, in Sect. 5
we list a few open problems related to deriving similar DAE dynamics for spot
interactions for more complicated models.
2 Dynamics of Spots on the Sphere
Under the assumption of a large diffusivity ratio and a small “fuel” supply, the
Brusselator model of [22] posed on the surface of the sphere, can be scaled into
the following system for u D u.x; t/ and the inhibitor v D v.x; t/ (cf. [24, 27]):
@u @v
D "2 S u C "2 E u C fu2 v ; D S v C "2 u u2 v ; (1)
@t @t
for some O.1/ constants E > 0, > 0, and 0 < f < 1. We refer to E as the “fuel”
parameter. Here S is the Laplace-Beltrami operator on the sphere.
Spatial patterns for which u concentrates as " ! 0 at a discrete set of points
x1 ; : : : ; xN on the sphere are called spot patterns. For " ! 0, we have u D O.1/
in the core of the spot, where jx xj j D O."/, and u
"2 E away from the spot
centers where jx xj j D O.1/. Then for " ! 0, the effect of the localized spots on
the inhibitor field v in (1) is to introduce a sum of Dirac-delta “forces” where the
strength of the “force” induced by the spot at xj is proportional to Sj (see [24] for
details). As such, v can be represented as a superposition of the well-known source
neutral Green’s function for the sphere. In this way, in [24] a quasi-equilibrium
spot pattern with frozen locations x1 ; : : : ; xN was constructed using the method of
matched asymptotic expansions by formulating a nonlinear algebraic system for the
spot locations x1 ; : : : ; xN and the spot strengths S1 ; : : : ; SN . The linear stability of
this quasi-equilibrium spot pattern to O.1/ time-scale perturbations was analyzed
in [24].
Provided that the quasi-equilibrium spot pattern is linearly stable, the slow
dynamics of the spot pattern on the long time scale D "2 t was derived in [27].
The collective coordinates characterizing this slow dynamics are the spot locations
x1 ; : : : ; xN and their corresponding spot source strengths S1 ; : : : ; SN , that both evolve
slowly on the long time-scale D "2 t. The slow dynamics derived in [27] is a
differential algebraic system of ODEs as given by the following result:
Principal Result 1 (Slow spot dynamics (cf. [27])) Let " ! 0. Provided that
there are no O.1/ time-scale instabilities of the quasi-equilibrium spot pattern, the
time-dependent spot locations, xj for j D 1; : : : ; N, on the surface of the sphere vary
on the slow time-scale D "2 t, and satisfy the dynamics
dxj 2 X Si xiN
D I Qj ; Q j xj xTj ; j D 1; : : : ; N ;
d A j .Sj / iD1
jxi xj j2
i¤j
(2a)
coupled to the constraints for S1 ; : : : ; SN in terms of x1 ; : : : ; xN given by the roots of
a nonlinear algebraic system involving the Green’s matrix G
h i 2E
N .S/ I .I E 0 /G S C .I E 0 /.S/ e D 0: (2b)
N
Here I is N N identity matrix, .E 0 /ij D N1 , .S/i D Si , ..S//i D .Si /, .G /ij D

log jxi xj j for i ¤ j and .G /ii D 0, .e/i D 1, and D 1= log " 1.
From (2) the spot locations are coupled to the spot strengths by (2b), yielding a
DAE system for the slow spot evolution. Since the spot strengths can be calculated
in terms of the locations from (2b), the DAE system has index 1. One readily
established feature of the DAE system (2) is that it is invariant under an orthogonal
transformation, corresponding to a rotation of the spots on the sphere.
In this DAE system (2), there are two functions .Sj / and A j .Sj / < 0, which
depend only on Sj and the Brusselator parameter f , that must be determined
numerically in terms of the local profile of the spot xj (cf. [27]). These are shown
in Fig. 1 for a few values of f . In the limit " ! 0, and for E D O.1/, the only
possible O.1/ time-scale instability of the quasi-equilibrium spot pattern is a linear
instability of the local spot profile to a peanut-shape if Sj > ˙2 . f /. This linear
instability is found in [24] to lead to a nonlinear spot self-replication event. The
threshold values ˙2 . f / for spot self-replication for a few values of f are given in the
caption of Fig. 1.
0
25
20 −5
15 −10
10
Aj −15
χ(Sj) 5
0 −20
−5
−25
1 2 3 4 5 6 7 8 9 0 1 2 3 4 5 6 7 8 9
Sj Sj
Fig. 1 Left panel: the function .Sj I f / in (2) for f D 0:4 (heavy solid), f D 0:5 (solid), f D 0:6
(dotted), and f D 0:7 (widely spaced dots). The spot self-replication threshold ˙2 . f / for Sj is
shown by the thin vertical lines in this figure. If Sj > ˙2 . f / the local spot profile is linearly unstable
to a peanut-splitting instability (cf. [24]). The threshold values are ˙2 .0:4/ 8:21, ˙2 .0:5/
5:96, ˙2 .0:6/ 4:41, and ˙2 .0:7/ 3:23. Right panel: the function A j .Sj / < 0 in (2) with the
same labels as in the left panel
3 Equilibria of the DAE Dynamics and Open Questions
In this section we discuss some previous results obtained in [27] as well as some
new results for the equilibria of (2) that have large basins of attraction to initial
conditions. We consider patterns for small values of N which Sj D O.1/ as ! 0.
A few open problems are mentioned.
To determine the possible equilibrium spot configurations of (2) with large basins
of attraction for N 3 when E D O.1/, f , and are given, we performed numerical
simulations of (2) for both pre-specified and for randomly generated uniformly
distributed initial conditions for the spot locations on the surface of the sphere. To
generate N initial points that are uniformly distributed with respect to the surface
area we let h and h be uniformly distributed random variables in .0; 1/ and define
spherical coordinates D 2h and D cos1 .2h 1/. Newton’s method was
used to solve (2b) for the initial set of N points. If the Newton iterates failed to
converge, indicating that no quasi-equilibrium exists for the initial configuration
of spots, a new randomly generated initial configuration was generated. The DAE
dynamics (2a) was then implemented by using an adaptive time-step ODE solver
coupled to a Newton iteration scheme to compute the spot strengths.
In the simulations below we took f D 0:5 and " D 0:02. We remark that
whenever e D .1; : : : ; 1/T is an eigenvector of the Green’s matrix G , then the
constraint (2b) admits a solution where S D Sc e. For such an equal spot–source
strength pattern, the equilibrium spatial configuration of spots for (2) is independent
of E, f , and .
In our discussion below, we refer to a ring pattern as a collection of N equally-
spaced spots lying on an equator of the sphere. We refer to an .N 2/ C 2 pattern
as a spot pattern consisting of two antipodal spots with the remaining N 2 spots
equally-spaced on the equatorial mid-plane between the two polar spots.
The simulations of [27] of 50 randomly generated initial spot for N D 3; : : : ; 8
yielded the following results for equilibria of (2) with large basin of attractions:
• N D 3: three equally-spaced spots that lie on a plane through the center of the
sphere. (Common spot strength pattern.)
• N D 4: four spots centered at the vertices of a regular tetrahedron. (Common
spot strength pattern.)
• N D 5; 6; 7: an .N 2/ C 2 pattern consisting of a pair of antipodal spots, with
the remaining N 2 spots equally-spaced on the equatorial mid-plane between
the two polar spots. (Two different spot strengths for polar and mid-plane spots.)
• N D 8: a “twisted cuboidal” shape, consisting of two parallel rings of four
equally-spaced spots, with the rings symmetrically placed above and below an
equator. The spots are phase shifted by 45ı between each ring. The perpendicular
distance between the two planes is 1.12924 as compared to a minimum distance
of 1.1672 between neighboring spots on the same ring, so that the pattern does
not form a true cube. (Common spot strength pattern.)
We remark that for the case N D 2 it was shown in Lemma 5 of [27] that any two
initial spots on the sphere will become antipodal in the long-time limit ! 1. This
was done by deriving a simple ODE for the angle ./ between the spot centers x1
and x2 at time , as measured from the center of the sphere, i.e. xT2 x1 D cos , and
establishing from this ODE that ! as ! 1 for any .0/.
In [27] the linear stability of ring configuration of spots was studied numerically.
The following conjecture was formulated in [27] based on numerical experiments.
• A ring pattern of N D 3 is orbitally stable, but is unstable if N 4.
• For N D 4; 5; 6; 7, an .N 2/ C 2 pattern is orbitally stable, but such a pattern is
unstable if N 8.
More specifically for N D 3, the numerical computations of [27] suggest that a
ring solution is orbitally stable to small random perturbations in the spot locations
in the sense that as time increases the perturbed spot locations become colinear on a
nearby (tilted) ring. For N 4, a similar small, but otherwise arbitrary, perturbation
of the spot locations on the ring leads to a breakup of the ring pattern. Similar an
.N 2/ C 2 pattern for N D 8 breaks up and forms a twisted cuboidal shape.
Open Problem: Establish analytically these results for the equilibria of spot
patterns for N D 3; : : : ; 8 using group theory methods for ODEs. Analyze the linear
stability of ring patterns by using an approach similar to that done in [2] for the
corresponding problem of the linear stability of Eulerian point vortices on a ring.
For larger values of N it becomes increasingly difficult to visualize the symme-
tries of the final equilibrium pattern that emerges under the DAE system (2) from
initial data. More specifically, it becomes increasingly challenging to find a rotation
matrix to put the pattern in a standard reference configuration. We now discuss two
new results for N D 9 and N D 10 not obtained in [27]. For even larger values of N
point-matching algorithms from computer science may be useful for classifying the
symmetries of the final pattern.
For N D 9 the equilibrium state of (2) with a large basin of attraction for initial
conditions is the pattern shown in the lower right subfigure of Fig. 2. Our simulations
with 50 random initial configurations have shown that the limiting pattern consists
of 3 parallel planes of 3 spots each. The spots on the equatorial plane and the other
two planes are phase-shifted 60ı (see the caption of Fig. 2 for details.)
For N D 10 the equilibrium state of (2) with a large basin of attraction for
initial conditions is the pattern shown in the lower right subfigure of Fig. 3. The
equilibrium pattern consists of two polar spots together with two parallel planes
with four equally-spaced spots on each plane. The relative phase-shift of the spots
on the two planes is 45ı (see the caption of Fig. 3 for details).
A classical problem of point configurations on the sphere is theP problem
P of find-
ing the global minimizer of the discrete logarithmic energy V i¤j log jxi
xj j on the sphere where jxj j D 1, which also has applications to minimizing
the mean first passage time for Brownian motion on the sphere (cf. [8]). Such
optimizing configurations are called elliptic Fekete point sets. By comparing our
results for equilibrium spot configurations with the optimal energies V of elliptic
Fig. 2 The evolution of a 9-spot pattern at different time for the Brusselator (2) when f D 0:5,
E D 18, and D 0:02. (a) The initial state D 0. (b) D 1. (c) D 3. (d) The computed steady
state after a suitable rotation. The steady-state consists of 3 planes of 3 spots each. The spots on
the equatorial plane and the other two planes are phase-shifted 60ı . The distance d between the
equatorial plane and each of the other two planes is d 0:7014. The common value of the spot
strengths on the equatorial plane differs from that of the other 6 spots
Fekete point sets, as given in Table 1 of [26], we conclude that our equilibrium
spot configurations for N D 3; : : : ; 10 having a large basin of attraction are indeed
elliptic Fekete point sets. As a remark, if we were to set Sj D 1 in (2) and ignore
the constraint (2b), then it is readily seen upon introducing Lagrange multipliers
that local and global minima of the discrete energy V are stable equilibria of the
simplified DAE dynamics.
Open Problem: Explore computationally for N 10 whether there is a relation-
ship between elliptic Fekete point sets and equilibria of the full DAE dynamics (2)
that have large basins of attraction of initial conditions.
Fig. 3 The evolution of a 10-spot pattern at different times for the Brusselator (2) when f D 0:5,
E D 22, and D 0:02. (a) The initial state D 0. (b) D 1. (c) D 3. (d) The computed steady
state after a suitable rotation. The steady-state consists of two polar spots together with two parallel
planes of 4 equidistantly spaced spots. The relative phase-shift of the spots on the two planes is
45ı . The distance d between the equator and either of the two planes is d 0:4234
4 Dynamics of Spots on the Sphere: The Gray-Scott Model
In this section we give a new result for the slow dynamics of spots on the unit sphere
for the well-known Gray-Scott RD model (cf. [6]).
vt D "2 S v v C Buv 2 ; ut D D s u C .1 u/ uv 2 : (3)
Although the analysis of spot dynamics for this problem follows the methodology
done in [27] for the Brusselator model, this new analysis requires the reduced-wave
Green’s function G.xI / on the sphere satisfying
1
S G G D ı.x / ; (4a)
D
1
G.xI /
log jx j C R C o.1/ ; as x ! ; (4b)
2
for some R independent of . In terms of this Green’s function we can derive the
following result for the slow dynamics of a collection of spots for the GS model.
Principal Result 2 (Slow spot dynamics for the GS model) Let " ! 0, and
assume that B D O."=/, where D 1= log ". Then, provided that there are
no O.1/ time-scale instabilities of the quasi-equilibrium spot pattern, the time-
dependent spot locations, xj for j D 1; : : : ; N, on the surface of the sphere vary on
the slow time-scale D "2 t, and satisfy the dynamics for j D 1; : : : ; N,
dxj X N
D 2"2 j .Sj / I Q j Si rx G.xj I xi / ; Q j xj xTj ; (5a)
d iD1
i¤j
coupled to the nonlinear constraints for S1 ; : : : ; SN in terms of x1 ; : : : ; xN given by
X
N
B D Sj .1 C 2R/ C .Sj / C 2 Si G.xj I xi / ; j D 1; : : : ; k : (5b)

iD1
i¤j
p
where D 1= log " and B B"1 D D O.1/.
In this system .Sj / and .Sj / depend on the core problem near the spot and are
specific to the GS model. Since the reduced-wave Green’s function and its regular
part R is not available in simple explicit form, and can only be written in terms of
the Legendre function (see [24]), it is more challenging to investigate the dynamics
and equilibria of spot patterns on the sphere for the GS model. This topic requires
further investigation.
5 DAE Spot Dynamics in Complex Models: Open Problems
There are several possible extensions of the methodology for deriving and analyzing
localized spot patterns for other scenarios.
We remark that explicit DAE dynamics for spot patterns is only possible when
the source-neutral Green’s function , satisfying s G D j˝j1 ı.xx0 / is explicitly
available. Such a Green’s function is well-known for the sphere, and this fact is key
to deriving (2). However, recently in [9], this Green’s function has been provided
analytically for a particular class of surfaces of revolution. For this class, DAE
dynamics on a manifold of varying curvature, and the possibility of pinning of
localized spot, can be analyzed.
The second possible extension of the modeling framework is to allow for spatial
heterogeneity in the fuel supply, so that E D E.x/. The corresponding outer solution
v will involve a sum ofR Dirac
distributions,
one near each spot, together with a term
vp of the form vp D ˝ E./ EN G.xI / d, where EN denotes the spatial average
of E over the sphere. In this way, it can be shown that the corresponding DAE
dynamics will yield a nonlocal system for the evolution of the spots.
Finally, a more intricate method to introduce spatial heterogeneity in (1) is
to consider spot patterns on the sphere for a model that couples passive bulk-
diffusion in the interior of the sphere to the Brusselator PDE on the surface. In
this context, the fuel supply E represents an exchange, or flux, between the surface-
bound concentrations and their bulk counterparts. This coupling should lead to rich
behavior in the DAE dynamics for spots on the sphere. This paradigm of studying
coupled surface-bulk models is becoming more prominent in scientific computation
and in applications (cf. [14, 15]).
Acknowledgements PHT thanks Lincoln College, Oxford and the Zilkha Trust for generous
funding. MJW gratefully acknowledges grant support from NSERC.
References
1. Barreira, R., Elliott, C.A., Madzvamuse, A.: The surface finite element method for pattern
formation on evolving biological surfaces. J. Math. Biol. 63, 1095–1109 (2011)
2. Boatto, S., Cabral, H.E.: Nonlinear stability of a latitudinal ring of point vortices on a non-
rotating sphere. SIAM J. Appl. Math. 64(1), 216–230 (2003)
3. Bogomolov, V.A.: Dynamics of vorticity at a sphere. Fluid Dyn. (USSR) 6, 863–870 (1977)
4. Callahan, T.K.: Turing patterns with O.3/ symmetry. Physica D 188(1), 65–91 (2004)
5. Chaplain, M.A.J, Ganesh, M., Graham, I.G.: Spatio-temporal pattern formation on spherical
surfaces: numerical simulation and application to solid tumour growth. J. Math. Biol. 42(5),
387–423 (2001)
6. Chen, W., Ward, M.J.: The stability and dynamics of localized spot patterns in the two-
dimensional Gray-Scott model. SIAM J. Appl. Dyn. Syst. 10(2), 582–666 (2011)
7. Chossat, P., Lauterbach, R., Melbourne, I.: Steady-sate bifurcation with O.3/ symmetry. Arch.
Rat. Mech. Anal. 113, 313–376 (1990)
8. Coombs, D., Straube, R., Ward, M.J.: Diffusion on a sphere with localized traps: mean first
passage time, eigenvalue asymptotics, and Fekete points. SIAM J. Appl. Math. 70(1), 302–332
(2009)
9. Dritschel, D.G., Boatto, S.: The motion of point vortices on closed surfaces. Proc. R. Soc. A
471, 20140890 (2015)
10. Kidambi, R., Newton, P.K.: Motion of three vortices on the sphere. Physica D 116(1–2), 143–
175 (1998)
11. Kolokolnikov, T., Ward, M.J., Wei, J.: Spot self-replication and dynamics for the Schnakenberg
model in a two-dimensional domain. J. Nonlinear Sci. 19(1), 1–56 (2009)
12. Kondo, S., Asai, R.: A reaction–diffusion wave on the skin of the marine angelfish Pomacan-
thus. Nature 376, 765–768 (1995)
13. Landsberg, C., Voigt, V.: A multigrid finite element method for reaction-diffusion systems on
surfaces. Comput. Vis. Sci. 13, 177–185 (2010)
14. Macdonald, C.B., Merriman, B., Ruuth, S.J.: Simple computation of reaction-diffusion pro-
cesses on point clouds. Proc. Natl. Acad. Sci. U.S.A. 110(23), 9209–9214 (2013)
15. Madzvamuse, A., Chung, A.H.W., Venkataraman, V.: Stability analysis and simulations of
coupled bulk-surface reaction diffusion systems. Proc. R. Soc. A 471, 20140546 (2015)
16. Matthews, P.C.: Pattern formation on a sphere. Phys. Rev. E 67(3), 036206 (2003)
17. Nagata, W., Harrison, L.G., Wehner, S.: Reaction-diffusion models of growing plant tips:
Bifurcations on hemispheres. Bull. Math. Biol. 6(4), 571–607 (2003)
18. Newton, P.K.: The N-Vortex Problem: Analytical Techniques. Springer, New York (2001)
19. Newton, P.K., Sakajo, T.: Point vortex equilibria and optimal packings of circles on a sphere.
Proc. R. Soc. A 467, 1468–1490 (2011)
20. Painter, K.J.: Modelling of pigment patterns in fish. In: Maini, P.K., Othmer, H.G. (eds.)
Mathematical Models for Biological Pattern Formation. IMA Volumes in Mathematics and
Its Applications, vol. 121, pp. 58–82. Springer-Verlag, New York (2000)
21. Plaza, R.G., Sánchez-Garduño, F., Padilla, P., Barrio, R.A., Maini, P.K.: The effect of growth
and curvature on pattern formation. J. Dyn. Diff. Equ. 16(4), 1093–1121 (2004)
22. Prigogine, I., Lefever, R.: Symmetry breaking instabilities in dissipative systems II. J. Chem.
Phys. 48(4), 1695–1700 (1968)
23. Roberts, G.: Stability of relative equilibria in the planar n-vortex problem. SIAM J. Appl. Dyn.
Syst. 12(2), 1114–1134 (2013)
24. Rozada, I., Ruuth, S., Ward, M.J.: The stability of localized spot patterns for the Brusselator
on the sphere. SIAM J. Appl. Dyn. Syst. 13(1), 564–627 (2014),
25. Sewalt, L., Doelman, A., Meijer, H., Rottschafer, V., Zagarias, A.: Tracking pattern evolution
through extended center manifold reduction and singular perturbations. Physica D 298, 48–67
(2015)
26. Stortelder, W., de Swart, J., Pintér, J.: Finding elliptic Fekete point sets: two numerical solution
approaches. J. Comput. Appl. Math. 130(1–2), 205–216 (1998)
27. Trinh, P., Ward, M.J.: The dynamics of localized spot patterns for reaction-diffusion systems
on the sphere. Nonlinearity 29(3), 766–806 (2016)
28. Varea, C., Aragón, J.L., Barrio, R.A.: Turing patterns on a sphere. Phys. Rev. E. 60(4), 4588–
4592 (1999)
29. Wei, J., Winter, M.: Stationary multiple spots for reaction-diffusion systems. J. Math. Biol.
57(1), 53–89 (2008)
Continuous Dependence on Modeling in Banach
Space Using a Logarithmic Approximation
Matthew Fury, Beth Campbell Hetrick, and Walter Huddell
Abstract We prove regularization for certain ill-posed problems in Banach space

by considering small changes in the model. The problems are of the form du=dt D
Au; 0 t < T; u.0/ D where A is assumed to generate a uniformly bounded
holomorphic semigroup fezA W <z 0g. Continuous dependence on modeling is
established by considering an approximate problem where the operator A is replaced
by an operator fˇ .A/, ˇ > 0 which approximates A in some sense as the parameter ˇ
tends to zero. The particular logarithmic approximation we apply originates from a
modified quasi-reversibility method recently introduced by Boussetila and Rebbani.
1 Introduction
Ill-posed problems have been the focus of considerable attention over the last few
decades. In this paper we focus on the abstract Cauchy problem
du
D Au ; 0t<T ; (1)
dt
u.0/ D
where A is the infinitesimal generator of a uniformly bounded holomorphic

semigroup fS.z/ D ezA W <z 0g in a Banach space .X; k k/, 2 X, and u
is a function u W Œ0; T ! X. This problem is generally ill-posed in that a solution
may not exist, or if it does, it may not depend continuously on the data. To estimate
M. Fury ()
Penn State Abington, 1600 Woodland Road, Abington, PA 19001, USA
B. Campbell Hetrick
Gettysburg College, 300 North Washington Street, Gettysburg, PA 17325, USA
W. Huddell
Eastern University, 1300 Eagle Road, St. Davids, PA 19087, USA

654 M. Fury et al.
the solution, we introduce an approximate well-posed problem
dv
D fˇ .A/v ; 0t<T ; (2)
dt
v.0/ D
where ˇ > 0 and fˇ .A/ approximates A. We prove continuous dependence on

modeling; i.e., that if fˇ .A/ is a suitable approximation of A, then a solution of
the ill-posed problem (if it exists) is appropriately close to the solution of the
well-posed model problem. This approach, introduced by Lattes and Lions in
[11], has been used with several different approximations fˇ . Examples include
fˇ .A/ D A ˇA2 , used in [2, 3, 5, 11, 12, 14], and fˇ .A/ D A.I C ˇA/1 , as
used in [3, 5, 10, 13]. However, both of these cases induce an error of order eC=ˇ ,
which is problematic in extending results to the nonlinear case. In this paper we use
a different approximation introduced by Boussetila and Rebbani [4]:
1
fˇ .A/ D ln.ˇ C epTA / ; ˇ>0; p1: (3)
pT
Boussetila and Rebbani use this approximation in Hilbert space (see also [15] and
[8]); Huang [9] extends the results to Banach space. In both cases, the authors use
the quasi-reversibility method: they use the solution to a well-posed model problem
to approximate the initial value for the original problem, then solve the original
problem backward. They show that the final value of the solution obtained in this
manner depends continuously on the initial data. Our work differs in that we prove
directly continuous dependence on the model. Specifically, we show that
p
ku.t/ vˇ .t/k C. ˇ/1 T M T ;
t t
0t<T (4)
where u.t/ is an assumed solution of (1), vˇ .t/ is the solution of (2) using the
logarithmic function fˇ in (3), and C and M are constants independent of ˇ.
The problems above share the same initial data, but we note that in practice the
observed data may differ. Thus we are interested in the approximate problem (2)
with replaced by ı where ı > 0 yields a change in the initial data satisfying
k ı k ı. In this context, we directly obtain regularization results by applying
the estimate (4).
In Sect. 2, we introduce the functional calculus used for fˇ .A/ and prove several
auxiliary results. In Sect. 3, we use these results to prove continuous dependence
on modeling (Theorem 1). In Sect. 4, we prove regularization for the problem (1).
Finally, Sect. 5 demonstrates the theory of this paper applied to certain partial
differential equations. Below, B.X/ denotes the space of bounded linear operators on
X, and .A/ the resolvent set of A. Also, a classical solution u.t/ of (1) is a function
u W Œ0; T ! X such that u.t/ 2 Dom.A/ for 0 < t < T, u 2 CŒ0; T \ C1 .0; T/, and
u satisfies (1) in X.
Continuous Dependence Using a Logarithmic Approximation 655
2 Approximation
In order to make sense of the operator fˇ .A/, we employ a functional calculus by

deLaubenfels [6]. By the fact that A generates fS.z/ D ezA W <z 0g, define the
function
Z
1 1 1 cos.sr/ irA
G.s; A/ D e dr ; s 0 :
1 r2
It may be shown that G.s; A/ is a continuous

function in s mapping into B.X/. Also
while kG.s; A/k s sup<z0 kezA k , by switching within equivalent norms, we
may assume without loss of generality that kG.s; A/k s for all s 0 (see [9]).
From this emerges deLaubenfels’s functional calculus
h i Z 1
f .A/ D lim f .t/ I C f 00 .s/G.s; A/ds (5)
t!1 0
for f 2 ACr1 Œ0; 1/ WD fh ı g W h 2 AC1 Œ0; 1g where g.t/ D .1 C t/1 and

AC1 Œ0; 1 WD f f W f 0 exists and is absolutely continuous on Œ0; 1g. In [9], Huang
provides a concise layout of applying (5) in order to find a formula for fˇ .A/ defined
1
by (3). Inserting f .s/ D pT ln.ˇ C epTs / in (5) yields
Z 1
1 ˇpTepTs
fˇ .A/ D ln ˇ G.s; A/ds :
pT 0 .ˇ C epTs /2
Also, Huang shows that

3
p fˇ .A/ is a bounded operator on X satisfying kfˇ .A/k
pT ln ˇ for 0 < ˇ < . 5 1/=2.
Lemma 1 Let A be the infinitesimal generator of a uniformly bounded
p holomor-
phic semigroup fS.z/ D ezA W <z 0g on X and let 0 < ˇ < . 5 1/=2.
Then
8ˇ 1
k Ax C fˇ .A/xk kS .2pT/xk
pT
for all x 2 Dom.S1 .2pT// D Ran.S.2pT//.

Proof Note that in the statement of the lemma we have used implicitly the fact that
S.t/ is a bounded, injective operator for each t 0 (cf. [7, Lemma 3.1]). Now, define
for s 0,
1 1
h.s/ D .s ln.ˇ C epTs //e2pTs D ln.ˇepTs C 1/e2pTs :
pT pT
656 M. Fury et al.
Using the fact that ln.x C 1/ x for x 0, we have

ˇ 2 ˇ
ˇ 4ˇ pTe2pTs C 3ˇpTepTs ˇ 2pTs
jh00 .s/j D ˇˇ 4pT ln.ˇe pTs
C 1/ ˇe
ˇ
.ˇepTs C 1/2
2
4ˇ pTe2pTs C 3ˇpTepTs
C 4ˇpTe pTs
e2pTs 8ˇpTepTs :
.ˇepTs C 1/2
Then for x 2 Dom.S1 .2pT// D Ran.S.2pT// Dom.A/,
k Ax C fˇ .A/xk D k.A C fˇ .A//e2pTA S1 .2pT/xk

Z 1
Dk h00 .s/G.s; A/S1 .2pT/x dsk
0
Z 1
8ˇ 1
8ˇpTs epTs kS1 .2pT/xk ds D kS .2pT/xk :
0 pT
t
u
In light of Lemma 1, for x 2 Dom.A/, define the operator gˇ .A/ in X by
gˇ .A/x D Ax C fˇ .A/x :
Proposition 1 Let A be the infinitesimal generator of a uniformly p bounded

holomorphic semigroup fS.z/ D ezA W <z 0g on X and let 0 < ˇ < . 5 1/=2.
Then gˇ .A/ generates a C0 semigroup fetgˇ .A/ gt0 on X satisfying

tgˇ .A/ t pT
ke k2 ln C1 for 0<tT:
pT ˇt
Proof For s 0 and 0 < t T, define the function h.t; s/ D .ˇepTs C 1/t=pT .
1
Also, let sˇ D pT ln. pT
ˇt /. Then for 0 < t T and x 2 X,
Z 1
tgˇ .A/ @2
ke xk D k h.t; s/G.s; A/x dsk
0 @s2
Z 1 tpTˇe pTs
. pTt ˇepTs 1/
Dk G.s; A/x dsk
.ˇepTs C 1/ pT C2
t
0
Z sˇ tpTˇs epTs . pTt ˇepTs C 1/ Z 1 tpTˇs epTs . pTt ˇepTs 1/
t kxk ds C t kxk ds
0 .ˇepTs C 1/ pT C2 sˇ .ˇepTs C 1/ pT C2
!
2ˇepTsˇ .tsˇ C 1/ C 2 1
D t t kxk
.ˇepTsˇ C 1/ pT C1 .ˇ C 1/ pT
!

ln. pT
ˇt
/ 1 t pT
< 2kxk C 2kxk ln C1 :
ˇt
t t
. pTt C 1/ pT C1 . pTt C 1/ pT pT
Finally, for any y 2 X, it may be shown by a dominated convergence argument that

R1 2
limt!0C 0 @s@ 2 f.h.t; s/ 1/epTs gG.s; A/y ds D 0. In other words, ketgˇ .A/ epTA y
epTA yk ! 0 as t ! 0C for all y 2 X. The result
ketgˇ .A/ x xk ! 0 as t ! 0C for all x 2 X (6)
then follows from the density of Ran.epTA / in X together with the fact that
pT C
pT ln. ˇt / ! 0 as t ! 0 .
t
t
u
Lemma 2 For sufficiently small ˇ > 0,
r
2 t
ketgˇ .A/ k p C1 for 0tT :
ˇ pT
p
Proof By Proposition 1 and the fact that x ln. 1x / x for x > 0, we have
s ! r
tgˇ .A/ 1 ˇt 2 t p 2
r
t

ke k2 C1 D p C ˇ p C1
ˇ pT ˇ pT ˇ pT
for 0 < t T. This estimate also holds for sufficiently small ˇ when t D 0 by (6).
3 Continuous Dependence on Modeling
The results proved in Sect. 2 will be used to establish continuous dependence on

modeling. For this, we assume u.t/ is a solution of (1) and let vˇ .t/ be the unique
solution of (2). We extend these solutions into the complex plane and apply the
Three Lines Theorem. Following [7], for " > 0 define a family of bounded operators
Z
1 2
C" D e"w .w A/1 dw
2i

where
is a complex contour contained within .A/, running from 1ei to 1ei
with 0 < < 2 . It may be shown that fC" g">0 is a strongly continuous holomorphic
semigroup on X generated by A2 .
658 M. Fury et al.
Lemma 3 Let " > 0 and let u.t/ be a classical solution of (1). Then
C" etfˇ .A/ D etgˇ .A/ C" u.t/ for all t 2 Œ0; T :
Proof The result follows from uniqueness of solutions to well-posed problems since
each item is a classical solution of (2) with initial data replaced by C" . t
u
We also require the following based on work of Agmon and Nirenberg [1].
Lemma 4 ([1, p. 148]) Let .z/ be a complex function which is bounded and
continuous on S D fz D C i j 2 Œ0; T; 2 Rg. For ˛ D t C ir 2 S,
define
Z Z
1 1 1
˚.˛/ D .z/ C dd :
S z˛ zN C 1 C ˛
N
Then ˚.˛/ is absolutely convergent, @˚.˛/ D .˛/ where @N denotes the Cauchy-
Riemann operator, and there exist constants K > 0 and L > T such that
Z ˇ ˇ
ˇ 1
1
1 ˇ
ˇ C ˇ d K 1 C ln L if ¤t:
ˇ zN C 1 C ˛ ˇ
1 z ˛ j tj
Theorem 1 Let A be the infinitesimal generator of a uniformly bounded p holo-

morphic semigroup fS.z/ D ezA W <z 0g on X and let 0 < ˇ < . 5 1/=2.
Let u.t/ and vˇ .t/ respectively be classical solutions of (1) and (2) and assume that
u.t/ 2 Dom.S1 .2pT// D Ran.S.2pT// and kS1 .2pT/u.t/k M 0 for all t 2 Œ0; T.
Then there exist constants C and M, each independent of ˇ, such that
p t t
ku.t/ vˇ .t/k C. ˇ/1 T M T for 0t<T:
Proof Define S to be the complex strip S D ft C ir j t 2 Œ0; T; r 2 Rg. Since eirA

is a bounded operator on X for every r 2 R, we may define for ˛ D t C ir 2 S,
" .˛/ D eirA C" .u.t/ vˇ .t// D eirA C" .u.t/ etfˇ .A/ / :
Also, following work of Agmon and Nirenberg [1], for ˛ D t C ir 2 S define

Z Z
1 1 1
˚" .˛/ D @" .z/ C dd ;
S z˛ zN C 1 C ˛

where z D Ci 2 S and @N denotes the Cauchy-Riemann operator @N D 12 @t@ C i @r@ .
In view of Lemma 4, we will show that @ N " .˛/ is bounded and continuous on S.
Note, using the fact that Dom.AC" / D X (cf. [7, Propoisition 2.10]), we have

@ d d
" .˛/ D eirA C" u.t/ etfˇ .A/ D eirA C" .Au.t/ fˇ .A/etfˇ .A/ / ;
@t dt dt
@ @
" .˛/ D eirA C" .u.t/ etfˇ .A/ / D eirA .iA/C" .u.t/ etfˇ .A/ / :
@r @r
Therefore,
N " .˛/ D eirA C" .Au.t/ fˇ .A/etfˇ .A/ / eirA AC" .u.t/ etfˇ .A/ /
2 @
D eirA .A fˇ .A//C" etfˇ .A/ :
Recall, fC" g">0 is a holomorphic semigroup satisfying C" x ! x as " ! 0 for

every x 2 X. Hence for small ", we may set C D sup0<"<1 kC" k. Also, set J D
sup<z0 kezA k. Then by Lemma 3, Lemma 2, Lemma 1, and stabilizing condition,
keirA .A fˇ .A//C" etfˇ .A/ k D keirA etgˇ .A/ .A fˇ .A//C" u.t/k

r r
2 t 8ˇ 1 2 t 8ˇ
Jp C1 kS .2pT/C" u.t/k J p C1 CM 0
ˇ pT pT ˇ pT pT
since C" commutes with A. Hence we have shown

p
N " .˛/k C0 ˇ
k@ (7)
where C0 is a constant independent of ˇ, ", and ˛. Thus, @ N " .˛/ is bounded on S.

N " .˛/ is continuous on S. Hence, by Lemma 4, ˚" .˛/
It is also readily shown that @
is absolutely convergent, @˚N " .˛/ D @ N " .˛/ and there exist constants K > 0 and
L > T such that
Z 1ˇ ˇ
ˇ 1 1 ˇ
ˇ C ˇ d K 1 C ln L if ¤ t :
ˇ zN C 1 C ˛ ˇ
1 z ˛ j tj
Next, define w" W S ! C by
w" .˛/ D x " .˛/ x ˚" .˛/ ;
where x is in the dual space of X. We will show that the Three Lines Theorem may
be applied to w" . First for ˛ D t C ir 2 S, using standard properties of semigroups
660 M. Fury et al.
and our stabilizing assumptions, we have
k" .˛/k JkC" u.t/ C" etfˇ .A/ k D JkC" u.t/ etgˇ .A/ C" u.t/k
Z t
tgˇ .A/ d sgˇ .A/
D Jk.I e /C" u.t/k D Jk .e C" u.t//dsk
0 ds
Z t Z t r
2 s 8ˇ
D Jk esgˇ .A/ gˇ .A/C" u.t/dsk J p C1 CM 0 ds (8)
0 0 ˇ pT pT
showing that " .˛/ is bounded on S. Next, for z D C i 2 S, by Lemma 4 and (7),
Z Z ˇ ˇ
1 1 T 0 p ˇˇ 1 1 ˇ
ˇ dd
k˚" .˛/k C ˇˇ C
1 0 z˛ zN C 1 C ˛ ˇ
Z p Z T
1 T 0p L e L
C ˇ K 1 C ln d D K ˇ 1 C ln d (9)
0 j tj 0 j tj
where eK is a constant. It follows that w" is bounded on S. Furthermore, w" is

continuous on S as well, and by Lemma 4, for ˛ in the interior of S, @w N " .˛/ D
N N
x @" .˛/ x @˚" .˛/ D 0 showing that w" is analytic on the interior of S.
Therefore, by the Three Lines Theorem,
jw" .t/j M.0/1 T M.T/ T

t t
for 0 t T, where M.t/ D supr2R jw" .t C ir/j. Looking at the sides of the strip
S, we have from (9),
M.0/ D sup jw" .ir/j kx kk" .ir/k C kx kk˚" .ir/k
r2R
D kx kkeirA .u.0/ vˇ .0//k C kx kk˚" .ir/k D kx kk˚" .ir/k
p Z T
e L
K ˇkx k 1 C ln d :
0
Next by (8), we have k" .T C ir/k J 0 where J 0 is a constant independent of ˇ, ",

and r since ˇ < 1. Hence, together with (9) we have that
M.T/ D sup jw" .T C ir/j kx kk" .T C ir/k C kx kk˚" .T C ir/k
r2R
Z T
L
kx k J 0 C e
K 1 C ln d
0 j tj
where again we have used the fact that ˇ < 1. Therefore

p Z T 1 Tt Z T Tt
L L
jw" .t/j kx k e

K ˇ 1C ln d J Ce
0
K 1C ln d :
0 0 j tj
Taking the supremum over all x 2 X with kx k 1, we have

h p i1 Tt t p
k" .t/ ˚" .t/k e J0 C e
t t
K ˇM1 KM2 T D C. ˇ/1 T M T
where C and M are constants independent of both ˇ and ". Again using (9),
kC" .u.t/ vˇ .t//k D k" .t/k k" .t/ ˚" .t/k C k˚" .t/k
p p Z T L

C. ˇ/1 T M T C e
t t
K ˇ 1 C ln d
0 j tj
p
C. ˇ/1 T M T
t t
for a possibly different value of C still independent of ˇ. Here the bound on the right
is independent of ", so we let " ! 0 to obtain
p t t
ku.t/ vˇ .t/k C. ˇ/1 T M T :
4 Regularization for Problem (1)
We prove regularization for problem (1) using our estimate from Theorem 1.
Definition 1 [10, Definition 3.1] A family fRˇ .t/ j ˇ > 0; t 2 Œ0; Tg B.X/ is
called a family of regularizing operators for the problem (1) if for each solution u.t/
of (1) with initial data and for any ı > 0, there exists ˇ.ı/ > 0 such that
1. ˇ.ı/ ! 0 as ı ! 0 ,
2. ku.t/ Rˇ .t/ı k ! 0 as ı ! 0 for each t 2 Œ0; T whenever k ı k ı .
Theorem 2 Let A be the infinitesimal generator of a uniformly bounded holomor-
phic semigroup fS.z/ D ezA W p <z 0g on X and let fˇ .A/ be defined by (3). Then
tfˇ .A/
fRˇ .t/ WD e j 0 < ˇ < . 5 1/=2; t 2 Œ0; Tg is a family of regularizing
operators for problem (1).
Proof Let u.t/ be a classical solution of (1) satisfying kS1 .2pT/u.t/k M 0 for all
t 2 Œ0; T and let k ı k ı. Note that since fˇ .A/ 2 B.X/, etfˇ .A/ satisfies
3t 3t
ketfˇ .A/ k etkfˇ .A/k e pT ln ˇ D ˇ pT
p
for all t 2 Œ0; T, for 0 < ˇ < . 5 1/=2.
662 M. Fury et al.
p
First let t 2 Œ0; T/ and choose ˇ D ı 6 . Then ˇ ! 0 as ı ! 0 and from
Theorem 1,
ku.t/ Rˇ .t/ı k ku.t/ etfˇ .A/ k C ketfˇ .A/ etfˇ .A/ ı k

D ku.t/ vˇ .t/k C ketfˇ .A/ . ı /k
p 3t
C. ˇ/1 T M T C ˇ pT ı
t t
p t t t
D C.ı 12 /1 T M T C ı 1 2T ! 0 as ı ! 0 :
p
In the case t D T, from (8) and (9) the estimate ku.T/ vˇ .T/k N ˇ can
easily be
p
obtained where N is a constant independent of ˇ. Then as above, still with
ˇ D ı6,
p 3 p p
ku.T/ Rˇ .T/ı k N ˇ C ˇ p ı D Nı 12 C ı ! 0 as ı ! 0 :
5 Example
For an example, we use that of deLaubenfels in [6]. Let T be the unit circle in
the complex plane and for 1 p 1, let H p .T/ be the set of all functions in
Lp .T/ that can be extended to a holomorphic function on the unit disc. Define A D
H 1 .T/ \ C.T/ and let A D i d d d
. Then iA D d is the generator of the familiar
translation group on A
.eirA /.ei / D .ei. Cr/ / ; r2R
for 2 H 1 .T/. This group extends to a uniformly bounded holomorphic semigroup

fezA W <z 0g as shown in [6].
In this setting, (1) becomes the ill-posed partial differential equation
ut D iu ; 0t<T ;
u.e ; 0/ D .e / :
i i
By Theorem 2, Rˇ .t/ D etfˇ .A/ D .ˇ C epTA / pT . If 2 A , deLaubenfels in [6]

t
shows G.s; A/ in this case to be given by
Œs
X
G.s; A/ D .s k/.k/g
O k
kD0
1
R
where gk .ei / D eik and .k/
O D 2 .ei /gk .ei /d. Defining the function
t
pTs pT
h.s/ D .ˇCe / and employing the functional calculus (5), we find for 2 A
Z 1
pT
t
Rˇ .t/ D h.A/ D ˇ C h00 .s/G.s; A/ds
0
Z Œs
t
1
tpTepTs epTs .t C pT/ X
D ˇ pT C 1 .s k/.k/g
O k ds :
pT.ˇ C epTs / kD0
t
0 .ˇ C epTs / PT C1
Acknowledgements The authors would like to thank Rhonda J. Hughes for her guidance and also
the 2015 AMMCS-CAIMS Congress.
References
1. Agmon, S., Nirenberg, L.: Properties of solutions of ordinary differential equations in Banach
space. Commun. Pure Appl. Math. 16, 121–151 (1963)
2. Ames, K.A.: On the comparison of solutions of related properly and improperly posed Cauchy
problems for first order operator equations. SIAM J. Math. Anal. 13, 594–606 (1982)
3. Ames, K.A., Hughes, R.J.: Structural stability for ill-posed problems in Banach space.
Semigroup Forum 70, 127–145 (2005)
4. Boussetila, N., Rebbani, F.: A modified quasi-reversibility method for a class of ill-posed
Cauchy problems. Georgian Math. J. 14, 627–642 (2007)
5. Campbell Hetrick, B.M., Hughes, R.J.: Continuous dependence results for inhomogeneous ill-
posed problems in Banach space. J. Math. Anal. Appl. 331, 342–357 (2007)
6. deLaubenfels, R.: Functional calculus for generators of uniformly bounded holomorphic
semigroups. Semigroup Forum 38, 91–103 (1989)
7. deLaubenfels, R.: Entire solutions of the abstract Cauchy problem. Semigroup Forum 42, 83–
105 (1991)
8. Fury, M.: Modified quasi-reversibility method for nonautonomous semilinear problems, Ninth
Mississippi state conference on differential equations and computational simulations. Electron.
J. Differ. Equ. Conf. 20, 99–121 (2013)
9. Huang, Y.: Modified quasi-reversibility method for final value problems in Banach spaces. J.
Math. Anal. Appl. 340, 757–769 (2008)
10. Huang, Y., Zheng, Q.: Regularization for a class of ill-posed Cauchy problems. Proc. Am.
Math. Soc. 133, 3005–3012 (2005)
11. Lattes, R., Lions J.L.: The Method of Quasi-Reversibility, Applications to Partial Differential
Equations. American Elsevier, New York (1969)
12. Miller, K.: Stabilized quasi-reversibility and other nearly-best-possible methods for non-well-
posed problems. In: Symposium on Non-Well-Posed Problems and Logarithmic Convexity.
Lecture Notes in Mathematics, vol. 316, pp. 161–176. Springer, Berlin (1973)
13. Showalter, R.E.: The final value problem for evolution equations. J. Math. Anal. Appl. 47,
563–572 (1974)
14. Trong, D.D., Tuan, N.H.: Regularization and error estimates for non homogeneous backward
heat problems. Electron. J. Differ. Equ. 4, 1–10 (2006)
15. Trong, D.D., Tuan, N.H.: Stabilized quasi-reversibility method for a class of nonlinear ill-
posed problems. Electron. J. Differ. Equ. 84, 1–12 (2008)
Solving Differential-Algebraic Equations
by Selecting Universal Dummy Derivatives
Ross McKenzie and John D. Pryce
Abstract A common way of making a high index DAE amenable to numerical

solution is that of index reduction. A classical way of reducing a DAE’s index is the
dummy derivative method of Mattsson and Söderlind, however for many problems
this method only provides a local index 1 DAE. Using the Signature Matrix based
structural analysis of Pryce to inform the dummy derivative method we present a
way to make this reduction global, where instead of picking new dummy derivatives
at run time and thus changing the overall structure of the problem you instead have
to update a list of parameters.
1 Introduction
We consider DAEs of the form:
fi .t; the xj and derivatives of them/; i D 1; : : : ; n (1)
where xj .t/, j D 1; : : : ; n are state variables and functions of some independent vari-
able t, usually considered to be time. Usually to solve such a problem differentials
of some of the equations are added to the system, these equations are called the
hidden constraints. A DAE has an associated index, mainly we are concerned with
the differential index, which is the minimum number of differentiations required
to obtain an ODE. Generally speaking the higher the index the more difficult
the numerical solution—there exist good solvers for index 1 systems and as such
we would like to reduce any DAE to an equivalent (same solution set) index 1
formulation. We are interested in reducing the index of our DAE via the Dummy
Derivative (henceforth DD method) method introduced in [4] by using the structural
analysis of the Signature Matrix method introduced in [2] (henceforth SA) to
R. McKenzie () • J.D. Pryce

School of Mathematics, Cardiff University, Cardiff, CF24 4AG, Wales, UK

666 R. McKenzie and J.D. Pryce
identify the hidden constraints and otherwise inform the DD method more than is
possible than in the classical approach of DDs via the Pantelides algorithm [1]. The
DD method introduces a choice of derivatives of variables to be considered algebraic
so that our enlarged system containing the hidden constraints is made square. This
process is inherently local because it relies on the non-singularity of potentially
dynamic matrices. We show in Sect. 2 how to pick a global index 1 system for a
simple example, in Sect. 3 we demonstrate how this approach extends to any DAE
on which Pryce’s SA works and in Sect. 4 we present numerical results.
2 Avoiding Dummy Pivoting for a Simple Example
Example 1 Consider the infamous simple pendulum:
s x-
9 @@
f1 .t/ D x00 .t/ C .t/x.t/ D 0>
> @ y
=
00 @ ?
f2 .t/ D y .t/ C .t/y.t/ g D0 @~ (2)
>
>
2 2 2 ;
f3 .t/ D x .t/ C y .t/ L D0 Pendulum bob, mass=1
where g and L are gravitational acceleration and length of the rod respectively. The
problem’s signature matrix as specified by the Signature Matrix Method [2] is:
x y ci
f1 2 1 0ı ! 0
f 1 2ı 0 0
˙D 2
f3 0ı 0 1 2
dj 2 2 0
Note: there are two highest value transversals (henceforth HVTs) for the simple
pendulum, marked by and ı in the signature matrix above. The offsets c and d tell
us how many times to differentiate each equation and the associated derivative order
for the variables that will be found. We now proceed to carry out the DD algorithm
in [4] but note whilst the matrices we use are the same the numbering is changed
so as to line up with the SA notation. The DD algorithm proceeds in stages, using
matrices GŒ and H Œ for D 0; 1; : : : where GŒ0 is the n n system Jacobian
.c / .d /
J D @fi i =@xj j . Deleting appropriate rows of GŒ gives H Œ . Deleting appropriate
columns of H Œ to form a nonsingular matrix gives GŒC1 . This example illustrates
Solving DAEs by Selecting Universal Dummy Derivatives 667
the process; for space reasons we refer to [4] for details. We have an initial Jacobian
(with highest order equations and variables) and a secondary non-square Jacobian:
x00 y00 ci
f1 1 0 x! 0 00 00
Œ0 f2 0 1 y 0 Œ0 x y
G D 00 and H D f300 2x 2y 0 :
f3 2x 2y 0 2
dj 2 2 0
We have to select 1 column from H Œ0 to get a square non-singular matrix GŒ1 .

We cannot choose column 3 because it is structurally 0, as is always the case
with columns corresponding to non-differentiated variables. We thus either choose
column 1 or column 2. Consider instead that we want to eliminate the need to choose
between any variables and instead pick both, since any choice may become invalid
(due to a subsequent G matrix becoming locally ill conditioned, i.e. as x or y ! 0)
and it will be expensive to pivot between potential systems as this happens for a
general DAE. If we want to eliminate the choice of candidate dummy derivatives at
this stage we will need to add equations to the DAE so that we can choose a square
submatrix of H Œ0 that contains the columns corresponding to all candidate dummy
derivatives, i.e. x and y. This is achieved by adding an equation of form:
Z1 WD ˛x0 C ˇy0 z1 D 0 (3)
to the original DAE, where ˛ and ˇ are some parameters that will be chosen at
run time so the row using new equations (newly introduced equations of the form
Zi D 0) in GŒ1 is roughly orthogonal to the row using old equations (equations
part of the original DAE including hidden constraints) in GŒ1 and z1 is some new
variable to solve for. We would like the H Œ0 for our new DAE to be of the form:
00 00 0
x y z1
f3002x 2y 0 0
H Œ0 D
Z10˛ ˇ 0 1
so that it is now possible to pick x00 and y00 as DDs for all time and update ˛ and ˇ
along the solution to keep the resulting GŒ1 matrix well-conditioned. Note: checking
the condition number of the corresponding G matrix and using a Gram-Schmidt or
QR type procedure for ‘new’ rows if the matrix becomes ill conditioned would be a
reasonable way of updating ˛ and ˇ dynamically. Consider however the new DAEs
signature matrix and canonical (element-wise minimum) offsets:
x y z1 ci
f1 0 2 1 0ı 11 0
f2 B1 2ı 0 1C 0
˙ D f3 @ 0ı 0 1 1A 2
Z1 1 1 1 0 0
dj 2 2 0 0
where the block structure is highlighted, see [3]. Unfortunately we see that entries
in positions .4; 1/ and .4; 2/ are both structurally zero (meaning dj ci ¤ i;j ), so
won’t appear in H Œ0 , see [5]. There is a key inequality that comes from the SA for
finding the offsets c and d:
dj ci i;j with equality on a HVT. (4)
Noting this one can see that changing d4 D c4 (because entry .4; 4/ must be on
a HVT) to 1 will not affect the other offsets. Due to our choice of Z1 we see that
d4 D c4 D 1 makes the entries in positions .4; 1/ and .4; 2/ structurally non zero. If
we write down H Œ1 for this new system (i.e. the DAE using equations f1 ; f2 ; f3 and
Z1 ) we get:
0 0
Œ1 x y
H D f30 2x 2y I
again we add an equation to the system so that we can always choose the variables
that are DD candidates at this stage:
Z2 WD x C ıy z2 D 0: (5)
So that our new DAE’s signature matrix with canonical offsets is:
x y z1 z2 ci
f1 0 2 1 0ı 1 11 0
f2 B1 2ı 0 1 1C 0
B ı C
0 0
˙ D 3B
f 1 1 1C 2
@ A
Z1 1 1 1 0 1 0
Z2 0 0 1 1 0 0
dj 2 2 0 0 0
Now to have all necessary entries of H Œ0 and H Œ1 structurally non zero we need to
set c4 D d4 D 1 and c5 D d5 D 2. Note that adding these equations to the system
does not change the original solution of the DAE, because the equations we add
form a new block dependent on the original DAE’s (block 1 above) solution. If we
now consider the DD stages of our new system we get:
x00 y00 z01 z002 ci

f300 2x 2y 0 0 0 !2
Z10 ˛ ˇ 0 1 0 1
H Œ0 D
Z200 ı 0 0 1 2
dj 2 2 0 1 2
We see that a valid choice is x00 , y00 and z002 —what we’re doing is adding DDs for all
variables in an old block that doesn’t contain an equation introduced at this stage
in the original system and ‘borrowing’ DDs from the new blocks that do use an
equation introduced at this stage in the original system. We’re forcing ourselves in
to a scheme that is unobtainable by performing a block decomposition of the new
DAE and doing DDs on each block. When proceeding to find H Œ1 we see that the
equation introduced for stage 1 is removed and we select x0 and y0 as DDs. Now note,
that since the second equation we introduced is almost the antiderivative of the first
we can condense both in to only one additional equation and finish with a DAE that
has the following signature matrix and valid offsets and now only selects x00 , x0 , y00
and y0 as DDs. The reason this is possible to do is perhaps not clear at first glance
(since whilst such a reduction is clearly structurally valid it may fail numerically).
Note that due to Griewank’s Lemma [2] our new DAE’s DD Jacobians GŒ1 and GŒ2
are equal, so if a set of parameters works for stage 1 it will also work for stage 2.
x y z2 ci
f1 0 2 1 0ı 11 0
f2 B1 2ı 0 1C 0
˙ D f3 @ 0ı 0 1 1A 2
Z2 0 0 1 0 2
dj 2 2 0 2
3 Finding ‘Universal’ Dummy Derivatives in General
Before giving a general algorithm for adding such equations to the system we need
the following definition:
Definition 1 Let the m and n be the number of rows and columns in the DD
matrix H Œ1 respectively.
We have the following algorithm to find a static selection of states (Note: we only
suggest to carry out the algorithm below in practice if a static selection cannot
already be found.):
Algorithm 1 The ‘Universal’ Dummy Derivative Algorithm

1: When finding the matrix GŒ :
2: if m D n
3: Proceed as normal
4: else
5: Create a vector say, candidatelist, of variables occurring in H Œ1 ; : : :
6: with structurally non zero columns, to order dj
7: S Dsize(candidatelist)m
8: for j D 1; : : : ; S (using a new vector, parameterlist, in each equation)
9: Add an equation to the DAE of the form:
Psize.candidatelist/
10: iD1 candidatelist.i/ parameterlist.i/ vj D 0
11: Set each new equation’s offsets equal to
12: Add DDs to the system for the differentials of entries in candidatelist
13: if > 1
14: for j D 1; : : : ; S
.2/ .max ci /
15: Add DDs to the system for vj ; : : : ; vj
16: Form the now non-square GŒ comprising of all rows for variables in candidatelist
17: Proceed with DD algorithm, i.e. form H Œ and proceed to the next stage
18: Tidy up final system:
19: Check if any equation introduced is the antiderivative (barring new variables and parameters)
of one introduced at a later stage.
20: Remove all such equations and any corresponding new variables and DDs from the system.
Note: The final check at the end is not needed to keep a static selection of dummy
derivatives, but can reduce the size of the resulting index 1 problem, as seen in
Sect. 2. Note also line 15 where we introduce extra dummy derivatives for the new
variables—we need the new equations at some stage but we do not need them at
previous stages and will instead solve their block based DD system at those stages,
which by construction is always square. Let us consider the following example:
Example 2 Consider a problem with signature and initial H matrix:
x1 x2 x3 x4 x5 c i
f1 0 2 111 x000
1 x002 x003 x004 x005
f2 B 2 2 C0 f10 0 0 0 0 1
B C
f3 B 1 0 C2 f300 B 0 0 0C
˙D @ A and H Œ0 D @ A
f4 0 0 2 f400 0 0 0
f5 1 1 1 f50 0 0 0
dj 3 2 2 2 2
In H Œ0 is being used to denote a structurally non-zero entry. We have
candidatelist D .x001 ; x02 ; x03 ; x04 ; x05 / (6)

and S D 1. This makes sense: there are only two possible choices of GŒ1 . All
variables that appear in a GŒ1 are also in candidatelist, either we have
x001 x02 x03 x04 x001 x03 x04 x05

0
f1 0 0 01 f1 0 0 0 1
f0 0 0C f30 B 0 0C
D 30 B

GŒ1 @ A or GŒ1 D 0@ A:
f4 0 0 f4 0 0
f5 0 0 0 f5 0 0 0
We add an equation of the form:
Z1 WD ˛x001 C ˇx02 C x03 C ıx04 C x05 v1 D 0 (7)
to the DAE and choose x000 00 00 00 00

1 , x2 , x3 , x4 and x5 to be ‘Universal’ DDs, setting d6 D
c6 D 1.
We form the following non square Jacobian:
x001 x02 x03 x04 x05

0
f1 0 0 0 1 00 0 0 0 0
x1 x2 x3 x4 x5
f30 B 0 0 0C f30 0 0 0
GŒ1 D 0@ H Œ1 D
0A
and get
f4 0 0 f40 0 0 0
f5 0 0 0
again, this system is not square, we have
candidatelist D .x01 ; x3 ; x4 / (8)
and S D 1 again. We add an equation of the form:
Z2 WD x01 C x3 C x4 vv1 D 0 (9)
to the DAE and choose x001 , x03 , x04 and vv100 to be ‘Universal’ DDs, setting d7 D
c7 D 2.
Let us just confirm our choice of DDs is valid (given the potential need to update
our parameters to keep any structurally non singular G matrix numerically non
singular). Our enlarged system has signature matrix and initial H matrix:
x1 x2 x3 x4 x5 v1 vv1 ci
f1 0 2 1 11 x000
1 x002 x003 x004 x005 v10 vv100
f2 B 2 2 C0 f10 0 0 0 0 0 0 1
B C
f3 B 1 0 C2 f300 B 0 0 0 0 0 C
B C B C
0 0 f400 B 0 0 0 0 0 C
˙ D 4B C2
f
B C and H Œ0 D 0 B C:
f5 B 1 1 C1 f5 B 0 0 0 0 0 C
@ A @ A
Z1 2 1 1 1 1 0 1 Z10 ˛ ˇ ı 1 0
00
Z2 1 0 0 0 2 Z2 0 0 0 1
dj 3 2 2 2 2 1 2
Where as expected a valid GŒ1 is:
x001 x02 x03 x04 x05 vv10

f1 0 0 0 0 0 1
x001 x02 x03 x04 x05 vv10
f30 B 0 0 0 0 C
0 !

B C f30 0 0 0
f0 0 0 0 0 C
D 4B

GŒ1 B C yielding H Œ1 D f40 0 0 0 0
f5 B 0 0 0 0 C
Z20 0 0
Z1 @ ˛ 0 A
1
ˇ ı
Z20 0 0 1
again, as expected a valid choice for GŒ2 is:
x01 x3 x4
f3 0!
GŒ2 D f4 0
Z2
where the algorithm terminates.

One may think a way to tidy up the above system would be to increase c6 D
d6 D 2 and remove the final equation and variable. Structurally this will still work,
but numerically we may fall in to trouble as GŒ2 may be singular.
We now go about proving the above algorithm provides a valid DD scheme with
same solution as the original problem. First we need the following definition:
Definition 2 (Transversals) A transversal for an m n matrix, with m n is a set
of m tuples .i; j/ so that i 2 f1; : : : ; mg and j 2 f1; : : : ; ng where no i or j is repeated.
Theorem 1 If a variable’s derivative does not appear in any potential DD scheme
the corresponding entries in the H matrix are all 0.
This result is easily derived from the following:
Theorem 2 Given any rectangular matrix with more columns than rows, and of
structural full row rank every non empty column contains an element of some
transversal.
Proof Assume for the sake of contradiction there exists a non empty column that
does not contain any elements that belong to a transversal. Without loss of generality
rearrange so that a non 0 entry is in the bottom right corner, say position .m ; n /.
Since we have a valid DD scheme at the stage (i.e. the system is of full row rank) we
know there exists some transversal, T say. This transversal contains some element
.m ; J/, for some J 2 f1; : : : ; n 1g. We can therefore form a new transversal of
the form fT .m ; J/g [ .m ; n /.
Theorem 1 is why we consider all non empty columns in our candidatelist. We now
seek to justify our removal of equations in the last part of the algorithm above.
Theorem 3 If a scheme produced by the above algorithm yields equations that can
be removed the solution is the same as the scheme without removing those equations.
Proof Structurally it is clear that such removal still leaves us with the same potential
choice of DDs, since if such a removal is possible we will have (at least) structurally
identical equations at some stage , (at least one) of which is being treated as
its own block system until its new variable is needed to make a square G matrix.
Numerically if such a reduction is possible then the size of the candidatelist vector
must remain unchanged. Since we only remove equations in our DD stages we have
two possibilities, either the difference between the size of candidatelist and m has
stayed the same, or it has increased. If it has stayed the same then the proposed
removal of equations gives us the same structural matrix, and by Griewank’s Lemma
the same numerical matrix. If it has reduced we will be adding new equations
that are not the antiderivatives of equations previously introduced. In which case
we can still make our corresponding G matrix non-singular by varying only the
new parameters introduced at this stage, since we know the rows are linearly
independent.
Note: The above proof yields a necessary condition for introducing equations that
can be removed that could shorten our algorithm. If at some stage ci > 1 for
each i considered at that stage then at the next stage we must introduce equations
which are the ‘antiderivatives’ of the ones introduced at this stage. We could shorten
Algorithm 1 by introducing such a condition. Finally we have the following:
Theorem 4 The above algorithm provides an always static selection of DDs that
is always valid (provided one chooses suitable parameters throughout integration)
and has same solution as the original DAE.
Proof Theorems 1 and 3 give us that we have a valid DD scheme, hence all that
is left to prove is that such equation additions do not change the solution set. As
illustrated in above example the inclusion of additional equations in the described
manner is equivalent to adding more dependent blocks to the system, which will not
change the original block’s solution.
Note: ‘same solution’ in the above Theorem may be confusing on a first read since
the reformulated DAE is of a larger size. We mean that the value of any xi .t/ in the
original DAE is also the value of that same xi .t/ in the reformulated DAE.
4 Preliminary Numerical Results for Universal Dummy

Derivatives
We solve the simple pendulum index 1 universal DD reformulation given below

(where xd denotes the DD corresponding to selecting x0 as a DD in the algorithm
above, xdd denotes x00 and similarly for y):
9
f1 .t/ D xdd .t/ C .t/x.t/ D 0>
>
>
>
f2 .t/ D ydd .t/ C .t/y.t/ g D 0>
>
>
>
>
>
>
f3 .t/ D x2 .t/ C y2 .t/ L2 D 0>
>
>
>
>
>
Z2 .t/ D x.t/ C ıy.t/ z2 .t/ D 0=
(10)
f30 .t/ D 2x.t/xd .t/ C 2y.t/yd .t/ D 0>
>
>
>
>
>
>
Z20 .t/ D xd .t/ C ıyd .t/ z02 .t/ D 0>
>
>
>
>
f300 .t/ D 2x.t/xdd .t/ C 2x2d .t/ C 2y.t/ydd .t/ C 2y2d .t/ D 0>
>
>
>
>
;
Z200 .t/ D xdd .t/ C ıydd .t/ z002 .t/ D0
in MATLAB using ode45 (using variable step size with initial conditions x D 6,
y D 8, x0 D y0 D 0 and parameters L D 10 and G D 9:81) by reformulating
the problem as an ODE in z2 and z02 and switching parameters whenever the
angle between .x; y/ and .; ı/ becomes small. We compare the solution with one
produced by DAETS [6], an accurate order 30 Taylor series solution, see Fig. 1.
We briefly present some information on switching the system (changing and
ˇ), a full discussion is left to a future work. If we choose our switching condition the
angle between .; ı/ and .x; y/ being less that =4, i.e. trying to keep the G matrices
well conditioned then the change in energy from t D 0 to t D 100 grows to around
105 . If we instead switch after every time to step to a new .; ı/ orthogonal to .x; y/
this change is greatly reduced to around 1012 . One could also attempt to solve the
new ‘Universal’ DD system by coming up with a continuous choice of parameters
so that the G matrices are globally non-singular, however in general this will likely
be difficult to achieve so we do not present results for doing this with the simple
pendulum.
−6
10
−8
10
−10
10
−12
10
−14
10 x
y
−16
10
0 50 100 150 200 250 300 350 400 450 500
Time
Fig. 1 Difference between the solution of the ‘Universal’ Dummy Derivative index 1 formulation
of the simple pendulum solved in MATLAB by reformulating to an ODE and using ode45 and
the solution to the original index 3 formulation solved via an order 30 Taylor Series method using
DAETS
5 Conclusion
We have given and demonstrated by examples a way of finding dummy derivatives

such that one does not have to change the structure of the reduced index 1 problem
when solving high index DAEs. Instead we have changed the problem from one
of structural pivoting of variables to numerical switching of parameters, which we
believe to be much easier for a general DAE code to handle, since one can exploit
underlying structures of the index 1 formulation throughout a simulation. In a future
work we aim to more formally develop a good switching condition for an arbitrary
DAE or discover a way of choosing continuous parameters for an arbitrary DAE.
References
1. Pantelides, C.C.: The consistent initialization of differential-algebraic systems. SIAM J. Sci.

Stat. Comput. 9, 213–231 (1988)
2. Pryce, J.D.: A simple structural analysis method for DAEs. BIT Numer. Math. 41(2), 364–394
(2001)
3. Pryce, J.D., Nedialkov, N.S., Tan, G.: DAESA—a matlab tool for structural analysis of
differential-algebraic equations: theory. ACM TOMS 41(2), 1–20 (2015)
4. Mattsson, S.E., Söderlind, G.: Index reduction in differential-algebraic equations using dummy
derivatives. SIAM J. Sci. Comput. 14(3), 677–692 (1993)
5. McKenzie, R., Pryce, J.D.: Structural analysis and dummy derivatives: some relations. In: Cojo-
caru, M.G. (ed.) Interdisciplinary Topics in Applied Mathematics, Modeling and Computational
Science, pp. 293–299. Springer, Cham (2015)
6. Nedialkov, N.S., Pryce, J.D.: DAETS User Guide. 2008–2009
On a Topological Obstruction in the Reach
Control Problem
Melkior Ornik and Mireille E. Broucke
Abstract This paper explores aspects of the Reach Control Problem (RCP) to drive
the states of an affine control system to a facet of a simplex without first exiting from
other facets. In analogy with the problem of nonlinear feedback stabilization, we
investigate a topological obstruction that arises in solving the RCP by continuous
state feedback. The problem is fully solved in this paper for the case of two and
three dimensions.
1 Problem Statement
This paper studies a topological obstruction that arises in solving the Reach Control
Problem (RCP) using continuous state feedback. We consider a simplex S WD
cofv0 ; : : : ; vn g with vertices fv0 ; : : : ; vn g and facets fF0 ; : : : ; Fn g. Each facet is
indexed according to the vertex it does not contain. Facet F0 is called the exit facet.
Let hj be the normal vector to facet Fj pointing outside S. Define I D f1; : : : ; ng. Let
I.x/ f0; 1; : : : ; ng be the minimal set of indices such that x 2 cofvi ; j i 2 I.x/g.
That is, x is in the interior of cofvi j i 2 I.x/g.
We consider the affine control system on S
xP D Ax C Bu C a ; (1)
where x 2 Rn and u 2 Rm where 1 m < n. Let B WD Im.B/ and O WD fx 2

Rn j Ax C a 2 Bg. Let u .t; x0 / denote the trajectory of (1) starting at x0 under a
feedback u.x/. The problem is to find a state feedback u.x/ such that all trajectories
u .; x0 / starting in S exit S through F0 in finite time without first leaving S. The
RCP has been extensively studied [4–6]. The purpose of this paper is to announce
a topological obstruction in the RCP, paralleling the analogous problem arising in
the problem of continuous state feedback stabilization [3], and to give preliminary
results on the problem for low dimensional systems.
M. Ornik () • M.E. Broucke

Department of Electrical & Computer Engineering, University of Toronto, Toronto,
ON M5S 3G4, Canada

678 M. Ornik and M.E. Broucke
A parallel study of the same problem was made in [10] (some preliminary work
appeared in [8]). The contributions of this paper significantly differ from [10]. This
paper primarily uses retraction theory to study the case of dim.OS / D n 1.
This leads to a simple solution in low dimensions of n. Studies of low-dimensional
systems are prevalent [1]. On the other hand, [10] uses homotopy theory to study the
case of dim.B/ D 2. While the conclusions of [10] could also potentially lead to a
solution in low dimensions of n, this situation is not explored in [10]. Furthermore,
the elegant cone condition for the topological obstruction developed in this paper,
B \ cone.OS / D 0, does not make an appearance at all in [10]; its place is taken by
a more involved result based on null-homotopic maps on a circle.
A supplement to this paper is found in [11] where we present supporting results
and a study of the case of an obstruction using affine feedback. The results of this
paper for the case of n D 2; 3, Theorems 3 and 4, could be obtained from [11].
However, while it is tempting to forgo more advanced topological methods in favour
of the brute force linear algebra arguments in [11], we insist on the importance of
the topological approach in order to have a hope of generalizing the results to higher
dimensions.
For each x 2 S, we define the cone
C.x/ D fy 2 Rn j hj y 0; j 2 InI.x/g : (2)
In other words, C.x/ is the set of all vectors y which, when attached at x, point into
S or through the exit facet F0 . (We note that if x 2 Int.S/, C.x/ D Rn .) In order
for the trajectory u .t; x0 / to not leave S through any facet except the exit facet, we
require [6]:
du
D Ax C Bu.x/ C a 2 C.x/ ; x2S: (3)
dt
In addition to this necessary condition, if u.x/ solves the RCP then there are no
closed-loop equilibria in S. The equilibria of an affine system can only lie in the
affine space O, and for all x 2 S \ O and u 2 Rm , Ax C Bu C a 2 B. Defining the
closed-loop vector field f .x/ D Ax C Bu.x/ C a, the previous statements suggest that
a necessary condition to solve the RCP by continuous state feedback is: there exists
a non-vanishing continuous map f .x/ on the set S \ O such that f .x/ 2 B \ C.x/.
Motivated by Brockett’s work [3], we will say that if such a function does not exist,
the system contains a topological obstruction.
Define OS D S \ O. We define cone.OS / D \x2OS C.x/. For the remainder
of the paper we assume that OS ¤ ;. We will also assume v0 … OS , as well as
1 dim.OS / n 1. The cases of dim.OS / D 0; n and v0 2 OS are trivial to
analyze. For the sake of completeness, this analysis was formally done in Lemma 9,
Lemma 10 and Corollary 11 of [11].
On a Topological Obstruction in the Reach Control Problem 679
We study the following problem.

Problem 1 Let S, B, O, and OS be as above. Does there exist a continuous map
f W OS ! Bnf0g such that for every x 2 OS , f .x/ 2 C.x/?
2 Main Results
This section presents the main results on solving Problem 1. We show that if
dim.OS / D n 1, then it is possible to characterize the solution of Problem 1
in terms of a smaller polytope OS0 , and OS0 will be amenable to a complete analysis
of the problem in low dimensions. The main result is presented in Theorem 1. The
consequences of Theorem 1 to low dimensional systems are presented in Theorem 3.
Let us assume OS is .n 1/-dimensional. According to [7, 9], this means S is cut
by O into two parts: one part containing v0 and p 0 other vertices, and the other
containing the other n p 1 vertices of S. W.l.o.g. we assume fv0 ; v1 ; : : : ; vp g
are on one side of OS and fvpC1 ; : : : ; vn g are on the other side, where we assume
vertices of S on OS are in the set fvpC1 ; : : : ; vn g. The vertices of OS lie on those
edges of S connecting vi ’s which are on different sides of OS . Thus, we employ the
notation oij to denote a vertex of OS with I.oij / D fi; jg. If there are no vertices of S
on OS , then OS has .p C 1/.n p/ vertices [7], but if OS contains r vertices of S,
then OS has .p C 1/.n p/ pr vertices. At this point we introduce a mild abuse
of notation with the convention that if vj 2 OS , then oij D vj for all i D 0; : : : ; p.
Let us introduce the following notation. Let
fi1 ; i2 ; : : : ; ik jj1 ; j2 ; : : : ; jl g D cofoi˛ jˇ W 1 ˛ k; 1 ˇ lg :
We observe that since I.oij / D fi; jg, if x 2 fi1 ; i2 ; : : : ; ik jj1 ; j2 ; : : : ; jl g then I.x/
fi1 ; i2 ; : : : ; ik g [ fj1 ; j2 ; : : : ; jl g. Also observe that OS D f0; : : : ; pjp C 1; : : : ; ng.
Lemma 1 Let OS D f0; : : : ; pjpC1; : : : ; ng, A D fi1 ; : : : ; ik jj1 ; : : : ; jl g OS , and
A0 D fi01 ; : : : ; i0k0 jj01 ; : : : ; j0l0 g OS . Let L D fi1 ; : : : ; ik g \ fi01 ; : : : ; i0k0 g. Analogously,
let R D fj1 ; : : : ; jl g \ fj01 ; : : : ; j0l0 g. Then, A \ A0 D fLjRg.
Proof By definition every vertex of fLjRg is a vertex of A and of A0 , so fLjRg
A \ A0 . Conversely, suppose x 2 A \ A0 . Since x 2 A, I.x/ fi1 ; : : : ; ik ; j1 ; : : : ; jl g
and since x 2 A0 , I.x/ fi01 ; : : : ; i0k0 ; j01 ; : : : ; j0l0 g. Hence, I.x/ L [ R, where we
use the fact that fi1 ; : : : ; ik g \ fj01 ; : : : ; j0l0 g D ; and fi01 ; : : : ; i0k0 g \ fj1 ; : : : ; jl g D ;. It
follows x 2 fLjRg. t
u
Before getting to the crux of the problem, let us introduce the notions of a
homeomorphism and a retraction. Let X and XQ be topological spaces. X and XQ
are homeomorphic if there exists a continuous bijection h W X ! XQ which has
a continuous inverse. Furthermore, if A is a subspace of X , a continuous map
r W X ! A is a retraction if rjA id.
Example 1 Let us consider the case of n D 3, with OS a quadrilateral with vertices

o02 2 cofv0 ; v2 g, o03 2 cofv0 ; v3 g, o12 2 cofv1 ; v2 g, and o13 2 cofv1 ; v3 g. By
definition, C.o02 / C.o12 / and C.o03 / C.o13 /. In fact, one can easily show that
the cone of any point on the edge o12 o02 will be larger than C.o02 /, and analogously
for the edge o13 o03 . Thus, we reach the idea that if a continuous function satisfying
Problem 1 exists on the convex set cofo02 ; o03 g containing the most restrictive cones,
then that function can easily be extended to the entire OS .
Theorem 1 will serve to prove the above claim. The procedure outlined in the
proof of Theorem 1, adapted to this example, is as follows. If a function f satisfying
Problem 1 can be defined on the edge o02 o03 , we can also define it on the edge o12 o02
by f .x/ D f .o02 / and on the edge o13 o03 by f .x/ D f .o03 /. We note that such f is
non-zero and satisfies the cone condition f .x/ 2 C.x/ because C.o02 / C.o12 / and
C.o03 / C.o13 /. So far f has been defined on three edges of OS . Then f can be
defined on the remainder of OS , which consists of its interior as well as the relative
interior of the edge o12 o13 , by using a retraction r — a continuous map from OS to
the three edges of OS on which f is already defined, such that r is identity on those
three edges. More formally, it can be shown, as in Theorem 1, that the function
f .x/ D f .r.x// exists and solves Problem 1.
Theorem 1 (Dimension Reduction) Let dim OS D n 1, v0 62 OS , and p > 0.
0
Define VO S
D fo 2 VOS j .9j 2 f1; : : : ; ng/o 2 cofv0 ; vj gg, and let OS0 D co.VO
0
S
/.
Then the answer to Problem 1 is affirmative if and only if it is affirmative for OS0 .
0
Proof Since VO S
VOS it follows that OS0 OS . Thus, if there exists f W
OS ! Bnf0g solving Problem 1 then f jOS0 W OS0 ! Bnf0g also solves Problem 1.
Conversely, suppose there exists f 0 W OS0 ! Bnf0g solving Problem 1. From our
0
notational convention, OS D f0; 1; : : : ; p j p C 1; : : : ; ng, VO S
D fo0.pC1/; : : : ; o0n g,
0
and OS D f0 j p C 1; : : : ; ng.
We now proceed with the main topological argument. Informally, we build a
skeleton of OS , starting with OS0 , and adding in each step additional edges and faces
of OS until in the last step all of OS is added. We then use topological methods to
show that Problem 1 for the set obtained in each step can be reduced to the same
problem applied to the set from the previous step, thus going back from OS to OS0 .
We build a skeleton of OS as follows. Let OS1 D OS0 , and for all 2 k n, let
[
OSk D OSk1 [ f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g.
0<i1 <:::<i˛ p;
p<j1 <:::<jˇ n;
˛CˇDk,˛;ˇ1
Observe that each H D f0; i1 ; : : : ; i˛ j j1 ; : : : ; jˇ g is a closed, convex polytope of

some dimension d, so it is homeomorphic to the closed ball Bd , and its boundary is
homeomorphic to the .d 1/-dimensional sphere Sd1 [2]. We claim that
@f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g n OSk1 D Int.fi1 ; : : : ; i˛ jj1 ; : : : ; jˇ g/ : (4)

There are three points to the proof of the claim.

(i) fi1 ; : : : ; i˛ jj1 ; : : : ; jˇ g 2 @f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g. To show that, we note that if
x 2 fi1 ; : : : ; i˛ jj1 ; : : : ; jˇ g, then
I.x/ fi1 ; : : : ; i˛ ; j1 ; : : : ; jˇ g f0; i1 ; : : : ; i˛ ; j1 ; : : : ; jˇ g,
so x 2 f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g. Moreover, @f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g consists of

points with jI.x/j ˛ C ˇ.
(ii) Int.fi1 ; : : : ; i˛ jj1 ; : : : ; jˇ g/\OSk1 D ;. Assume x 2 Int.fi1 ; : : : ; i˛ jj1 ; : : : ; jˇ g/.
Then 0 62 I.x/ and I.x/ D ˛ C ˇ D k. However, if x 2 OSk1 , then either
0 2 I.x/ or jI.x/j k 1.
(iii) @f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ gn Int.fi1 ; : : : ; i˛ jj1 ; : : : ; jˇ g/ OSk1 . This follows
because if x 2 @f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ gnInt.fi1 ; : : : ; i˛ jj1 ; : : : ; jˇ g/, then either
0 2 I.x/ and jI.x/j k so x 2 OSk1 ; or 0 62 I.x/ and jI.x/j k 1, so again
x 2 OSk1 .
What we have shown so far is that f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g is homeomorphic to a
closed ball, and @f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g \ OSk1 is homeomorphic to its boundary
sphere Sd1 with an open connected set Int.fi1 ; : : : ; i˛ jj1 ; : : : ; jˇ g/ of dimension d1
cut out of it. Now we show there exists a retraction from Bd to the punctured sphere
Sd1 n P, where P is homeomorphic to an open ball of dimension d 1. Since P
has dimension d 1, Sd1 is split into two connected parts: P and Sd1 nP. The
retraction argument is standard in topology. We provide the proof since it is integral
to our results.
First, let us note that Bd is homeomorphic to the upper half-ball BCd D fx 2 Bd W
x1 0g. The precise homeomorphism is not difficult to find, but one can simply
imagine taking the ball and flattening its lower half. Now, our sphere Sd1 was
mapped by this to the boundary of BCd . Furthermore, without loss of generality,1
we can assume that the closed part Sd1 nP makes up the bottom of the half-ball:
fx 2 Bd W x1 D 0g, while the open part fx 2 Bd W x1 > 0g corresponds to P.
0
Let us define the function rH W BCd ! fx 2 Bd W x1 D 0g by rH 0
.x1 ; x2 ; : : : ; xn / D
.0; x1 ; x2 ; : : : ; xn /. Clearly, this is a valid retraction and thus, we have obtained a
retraction from BCd to fx 2 Bd W x1 D 0g. Now, using the fact that BCd is
homeomorphic to Bd , while the same homeomorphism takes fx 2 Bd W x1 D 0g
00
to Sd1 nP, we know there thus exists a retraction rH W Bd ! Sd1 nP. Finally,
reminding ourselves that there exists a homeomorphism between H and Bd which
00
takes Sd1 nP to @OS \ OSk1 , by “pushing” rH through that homeomorphism, we
obtain a retraction rH W f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g ! @f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g\OSk1 .
Now we glue these retractions to each other. In order to do that, we need
to know that for H’s with constant ˛ C ˇ D k, all the different retractions
rH W f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g ! @f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g \ OSk1 agree on the
1
Really, this is done through another homeomorphism: this time, imagine, before flattening the
ball, choosing the part that needs to be flattened to be Sd1 nP .
intersections of their domains. That is, if H ¤ H0 , then rH jH\H0 rH0 jH\H0 .

(The claim is obvious if H D H0 .) Let H D f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g. Analogously,
let H0 D f0; i01 ; : : : ; i0˛0 jj01 ; : : : ; j0ˇ0 g. We noted in Lemma 1 that H \ H0 D
f0; fi1 ; : : : ; i˛ g \ fi01 ; : : : ; i0˛0 gjfj1 ; : : : ; jˇ g \ fj01 ; : : : ; j0ˇ0 gg. Since we assumed that
H ¤ H0 , there needs to be an element in fi1 ; : : : ; i˛ ; j1 ; : : : ; jˇ g which is not an
element of the set fi01 ; : : : ; i0˛0 ; j01 ; : : : ; j0ˇ0 g and vice versa (note that both of those sets
have k elements, so one cannot be a subset of the other).
Thus, H \ H0 will not contain more than k 1 non-zero vertices of S in its
notation, and hence it will be in both OSk1 (by the definition of OSk1 ), and in @H
(as none of its elements can be in the interior of H: the expansion as a convex sum of
every element in the interior needs to contain every vertex mentioned in the notation
of H). Analogously, H \ H0 2 @H0 \ OSk1 , which is the image of the retraction
rH0 .
Hence, we know that rH0 jH\H0 is an identity map, and so is rH jH\H0 . Thus,
these two retractions can indeed be glued together. By iterating this procedure for
all H, we obtain a glued retraction rk W OSk ! OSk1 which takes each k-dimensional
edge in OSk to its boundary. Let us note what this retraction does. For every point
x 2 OSk , if x is also in OSk1 , it will not do anything. Thus, C.rk .x// D C.x/.
If x 62 OSk1 , then either x is in the interior of some H D f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g
that is being added to OSk , or it is in the interior of fi1 ; : : : ; i˛ jj1 ; : : : ; jˇ g. Now, if x is
in the interior of f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g, rk maps it to a point in the boundary of H.
In that case, it is easy to verify that C.rk .x// C.x/. This has formally been done in
Lemma 6 of [11].
If x is in the interior of fi1 ; : : : ; i˛ jj1 ; : : : ; jˇ g, then I.x/ D fi1 ; : : : ; i˛ ; j1 ; : : : ; jˇ g.
On the other hand, H D f0; i1 ; : : : ; i˛ jj1 ; : : : ; jˇ g, so for any point y 2 H, I.y/
fi1 ; : : : ; i˛ ; j1 ; : : : ; jˇ g. Thus, C.y/ C.x/. Thus, specifically C.rk .x// C.x/.
In all three cases, we deduce that C.rk .x// C.x/. By composing r D r2 ı
r ı ı rn , we obtain a retraction r W OS D OSn ! OS1 D OS0 . Define
3
f .x/ D f 0 .r.x//. We obtained a nowhere vanishing function f on OS such that f .x/ D

f 0 .r2 .r3 .: : : .rn .x// : : :///. Thus, f .x/ is contained in C.r2 .r3 .: : : .rn .x// : : :///
C.r2 .r3 .: : : .rn1 .x// : : :/// : : : C.x/. This function satisfies the conditions
of Problem 1. t
u
We now proceed to resolving Problem 1 for the case of n D 2 and n D 3. We
have previously assumed that dim OS ¤ 0 and dim OS ¤ n, as these cases are
simple to analyze. The case of n D 2 is thus reduced to dim OS D 1. As we have
also required 1 dim B < n, we conclude that dim B D 1. However, the case of
dim B D 1 is resolved by Theorem 1 in [12].
This resolves the case of n D 2, as well as dim B D 1. The only cases that remain
are when n D 3, dim B D 2, and dim OS is either 1 or 2. We will see that, when
dim OS D 1, an argument based on linear algebra applies. On the other hand, a
purely topological argument applies when dim OS D 2.
First we examine why a sufficiently high dimension for B resolves Problem 1.
Lemma 2 Suppose OS D cofo1 ; : : : ; oC1 g where the oi ’s are the vertices of OS .

If there exists a linearly independent set fbi 2 B \ C.oi / j i D 1; : : : ; C 1g, then
the answer to Problem 1 is affirmative.
P P
Proof Let f W OS ! B be defined by f . C1 iD1 ˛i oi / D ˛i bi , where ˛i D 1 and
˛i 0. Necessarily f .x/ ¤ 0 for x 2 OS for otherwise the bi ’s would be linearly
dependent. Also, by a standard convexity argument f .x/ 2 C.x/, x 2 OS . t
u
The following is the key result in the case of dim OS D 1.
Lemma 3 Let n D 3, dim B D 2, and let o1 and o2 be vertices of OS . Then there
exist linearly independent vectors fb1 ; b2 j bi 2 B \ C.oi /g. Moreover, if OS D
cofo1 ; o2 g, the answer to Problem 1 is affirmative.
Proof First we assume o1 2 Int.Fi / for some i 2 f0; 1; 2; 3g. By the definition of
C.o1 /, it is a closed half space or R3 , so there exist linearly independent vectors
b11 ; b12 2 B \ C.o1 /. We claim B \ C.o2 / ¤ 0. If o2 2 Int.Fi / for some i 2
f0; 1; 2; 3g then the argument above proves the claim. Instead, assume w.l.o.g. that
o2 2 F1 \ F2 . Then C.o2 / D fy 2 R3 jh1 y 0; h2 y 0g. Let B D Ker.M T / for
some M 2 R31 . Finding 0 ¤ y 2 B \ C.o2 / is equivalent to solving
2 3 2 3
h1 T s1
4 h2 T 5 y D 4 s2 5 (5)
MT 0
for some s1 ; s2 2 R 0 and y ¤ 0. Because fh1 ; h2 g are linearly independent,

rank.H/ 2, where H is the matrix appearing on the left hand side of equation 5.
If rank.H/ D 3, then let
2 3
1 0
Œy1 y2 D H 1 4 0 1 5 :
0 0
Since .1; 0; 0/ and .0; 1; 0/ are linearly independent, y1 and y2 are linearly
independent as well.
Next, assume rank.H/ D 2. In other words, M D c1 h1 C c2 h2 for some c1 ; c2 2
R. Then, by taking s1 D s2 D 0, equation (5) reduces to Œh1 h2 T y D 0. By the rank-
nullity theorem, there exists y ¤ 0 satisfying this equation. Moreover, if w.l.o.g.
v0 D 0, then y 2 F1 \ F2 D cofv0 ; v3 g, and we can take y D v3 . We have shown
there exist linearly independent b11 ; b12 2 C.o1 / and there exists 0 ¤ b2 2 C.o2 /.
We claim at least one of the pairs fb11 ; b2 g and fb12 ; b2 g is linearly independent. For
otherwise there exist c1 ; c2 2 R such that b2 D c1 b11 D c2 b12 , implying b11 and
b12 are linearly independent, a contradiction. We conclude there exists a linearly
independent set fb1 ; b2 j bi 2 C.oi /g.
Next we assume neither o1 nor o2 lies in the interior of a facet. W.l.o.g. suppose
o1 2 F1 \ F2 and o2 2 F1 \ F3 . If either C.o1 / or C.o2 / contains two linearly
v0
o03
v3
o02
o13
v1 v2
o12
0
Fig. 1 The set OS for Example 1. The edge from o02 to o03 forms OS
v0 v0 v0
v3 v3 v3
v1 v2 v1 v2 v1 v2
Fig. 2 Three of the four possible configurations of set OS for n D 3 and dim OS D 2, with the
fourth one given in Fig. 1. The leftmost configuration is addressed by Theorem 2, while the other
two can be reduced using Theorem 1
independent vectors, then by the previous argument, we are done. Otherwise, by

the previous argument again v3 2 C.o1 / and v2 2 C.o2 /. Since fv2 ; v3 g are linearly
independent, we are done. Finally, if OS D cofo1 ; o2 g, then by Lemma 2 the answer
to Problem 1 is affirmative. t
u
The remaining case to study is when n D 3, dim OS D 2, and dim B D 2.
Assuming v0 62 OS (which is a trivial case discussed in Lemma 10 of [11]), there
are four topologically distinct cases for OS , depending on the way O cuts S. These
are given in Figs. 1 and 2. In Fig. 1, OS is a quadrangle. In that case, p D 1; that
is, there are two vertices of S on each side of O. Then we can apply Theorem 1 to
reduce OS to OS0 , and according to the construction in the proof, OS0 has dimension
1, and we can apply Lemma 3. Similarly, in the cases given in middle and the
rightmost configuration of Fig. 2, we can apply Theorem 1 to reduce OS to OS0
with dim OS0 being either 0 or 1, respectively. Finally, in the situation given by the
leftmost configuration of Fig. 2, we draw upon a proof method already utilized in
[4], which is based on Sperner’s lemma. Here we employ a variant found in [13].
Lemma 4 Let P D cofw1 ; : : : ; wnC1 g be an n-dimensional simplex. Furthermore,
let fQ1 ; : : : ; QnC1 g be a collection of sets covering P such that
(P1) Vertex wi 2 Qi and wi 62 Qj for j ¤ i.
(P2) If w.l.o.g. x 2 cofw1 ; : : : ; wl g for some 1 l n C 1, then x 2 Q1 [ [ Ql .
T
Then nC1
1D1 Qi ¤ ;.
Theorem 2 Let n D 3 and suppose OS D cofo1 ; o2 ; o3 g with v0 62 OS and oi 2

.v0 ; vi , i D 1; 2; 3. The answer to Problem 1 is affirmative if and only if
B \ cone.OS / ¤ 0.
Proof Sufficiency is clear: if 0 ¤ b 2 B \ cone.OS /, a constant function f .x/ D b

satisfies Problem 1. For necessity, suppose there exists f W OS ! B n f0g such that
f .x/ 2 C.x/, x 2 OS . By way of contradiction suppose B \ cone.OS / D 0. Since
oi 2 .v0 ; vi , i D 1; 2; 3, we have
cone.OS / D fy 2 Rn j hj y 0 ; j D 1; 2; 3g :
Define the sets
Qi WD fx 2 OS j hi f .x/ > 0g ; i D 1; 2; 3 : (6)
Now we verify the conditions of Lemma 4.

Firstly, we claim that fQi g cover OS . For suppose not. Then there exists x 2
OS such that hj f .x/ 0, j D 1; 2; 3. Hence f .x/ 2 B \ cone.OS /, so f .x/ D
0, a contradiction to f being non-vanishing on OS . Secondly, we verify property
(P1). We claim that oi 2 Qi for i D 1; 2; 3. For suppose not. Then hi f .x/ 0.
Additionally, because f .oi / 2 C.oi /, hj f .x/ 0, j 2 f1; 2; 3g n fig. We conclude
f .oi / 2 B \ cone.OS /, so f .oi / D 0, a contradiction. Next we claim oi 62 Qj ,
j ¤ i. This is immediate since f .oi / 2 C.oi / implies hj f .oi / 0, j ¤ i. Thirdly,
we verify property (P2). Suppose w.l.o.g. (by reordering the indices f1; 2; 3g) x 2
cofo1 ; : : : ; or g for some 1 r 3. We claim x 2 Q1 [ [ Qr . For suppose
not. Then hj f .x/ 0, j D 1; : : : ; r. Also, it is easily verified that C.x/ D fy 2
Rn j hj y 0 ; j D r C 1; : : : ; 3g. Thus, hj f .x/ 0, j D r C 1; : : : ; 3. Hence,
f .x/ 2 B \ cone.OS /, so f .x/ D 0, a contradiction to f being non-vanishing on OS .
T3We have verified (P1)-(P2) of Lemma 4. Applying the lemma, there exists x 2
iD1 Qi ; that is, hj f .x/ 0, j D 1; 2; 3. We conclude that f .x/ 2 B \ cone.OS /,
so f .x/ D 0, a contradiction. t
u
The following result finally resolves Problem 1 in all cases of interest.
Theorem 3 Let S, B and OS be as above, and let n 2 f2; 3g. If n D 3, dim B D 2
and OS does not satisfy the conditions of Theorem 2, then the answer to Problem 1
is affirmative. Otherwise, the answer to Problem 1 is affirmative if and only if B \
cone.OS / ¤ 0.
Proof The discussion prior to Lemma 2, as well as Lemma 3 and Theorem 2,
covered all the cases except for the one where dim B D dim OS D 2 and OS
does not satisfy the conditions of Theorem 2. However, in that case, as described
prior to Lemma 4, Theorem 1 reduces OS to OS0 of dimension 0 or 1. Applying
Lemma 3, we now obtain that the answer to Problem 1 is affirmative. t
u
An interesting modification of Problem 1 requires that f be not only continuous,

but also affine. This corresponds to a classic problem of designing an affine state
feedback, applied to RCP. We note that in most of the configurations considered
above, the same claims that work for Problem 1 can also be used in the affine case.
The only significantly different case is dim OS D dim B D 2, i.e., the situation
covered by Theorem 1. In this case, the full analysis of the problem of affine
obstruction can be done using linear algebra. Such an analysis is not difficult, but
is computationally long. It is presented in full in Section 3.2.2 of [11]. With the
results contained in [11] (in fact, Theorem 16 in [11] is essentially the same as our
Theorem 3), we reach the following theorem:
Theorem 4 Let n 2 f2; 3g. The problem of affine obstruction is solvable if and only
if Problem 1 is solvable.
It is not known if such a result holds in general. Hence, we end this paper with
the following conjecture.
Conjecture 1 Let n 4. The problem of affine obstruction is solvable if and only
if Problem 1 is solvable.
3 Conclusion
This paper introduces a topological obstruction to solving the RCP via continuous
state feedback. The results show an interplay between linear algebra-based argu-
ments regarding the number of control inputs and purely topological arguments
regarding a cone condition on B. We show that for n D 2 and n D 3 these two
properties together fully characterize when a topological obstruction arises.
References
1. Bemporad, A., Morari, M., Dua, V., Pistikopulos, E.N.: The explicit linear quadratic regulator
for constrained systems. Automatica 38(1), 3–20 (2002)
2. Bredon, G.E.: Topology and Geometry. Springer, New York (1997)
3. Brockett, R.W.: Asymptotic stability and feedback stabilization. In: Brockett, R.W., Millman,
R.S., Sussmann, H.J. (eds.) Differential Geometric Control Theory, pp. 181–191. Birkhauser,
Boston (1983)
4. Broucke, M.E.: Reach control on simplices by continuous state feedback. SIAM J. Control
Optim. 48(5), 3482–3500 (2010)
5. Broucke, M.E., Ganness, M.: Reach control on simplices by piecewise affine feedback. SIAM
J. Control and Optim. 52(5), 3261–3286 (2014)
6. Habets, L.C.G.J.M., van Schuppen, J.H.:A control problem for affine dynamical systems on a
full-dimensional polytope. Automatica 40(1), 21–35 (2004)
7. Kettler, P.C.: Vertex and Path Attributes of the Generalized Hypercube with Simplex Cuts.
University of Oslo, Oslo (2008)
8. Mehtaa, K.: A topological obstruction in a control problem. Master’s thesis, University of

Toronto (2012)
9. Min, C.: Simplicial isosurfacing in arbitrary dimension and codimension. J. Comput. Phys.
190(1), 295–310 (2003)
10. Ornik, M., Broucke, M.E.: A topological obstruction to reach control by continuous state
feedback. In: This work has now been published in 54th IEEE Conference on Decision and
Control (CDC), 2015, pp. 2258-2263, doi: 10.1109/CDC.2015.7402543. For full citation, see
here: http://ieeexplore.ieee.org/xpls/abs_all.jsp?arnumber=7402543&tag=1
11. Ornik, M., Broucke, M.E.: Some results on an affine obstruction to reach control.
arXiv:1507.02957 [math.OC] (2015)
12. Semsar Kazerooni, E., Broucke, M.E.: Reach controllability of single input affine systems.
IEEE Trans. Autom. Control 59(3), 738–744 (2014)
13. Talman, D.: Intersection theorems on the unit simplex and the simplotope. In: Imperfections
and Behavior in Economic Organizations, pp. 257–278. Springer, New York (1994)
Continuous Approaches to the Unconstrained
Binary Quadratic Problems
Oksana Pichugina and Sergey Yakovlev
Abstract Two continuous approaches to discrete problems over sets inscribed

into a sphere are presented. They use an analytic description of the sets and a
convex extension of objective functions from the sets onto Euclidean space. The
first approach is the Branch and Bound Polyhedral-Spherical Method (B&BPSM)
for “two-layer sets”, which utilizes the sets representation as an intersection of a
polyhedron with a sphere and the global minimizers on a sphere. The concept of a
functional representation of a discrete set as an intersection of surfaces is introduced
and used, for constructing penalty functions, in the second approach – the Penalty
Method based on Functional Representations (FRPM). The methods are applied to
the Unconstrained Binary Quadratic Problem. For the purpose, the representation of
the binary set as a “touching set” of smooth surfaces is derived.
1 Introduction
The Unconstrained Binary Quadratic Problem (UBQP):
z D f .x/ ! min; (1)

f .x/ D x Ax C c x;
T T
(2)
x 2 Bn D f0n ; 1n g (3)
is a famous NP-hard problem known by numerous applications and a variety

of exact and heuristic methods (the solution methods alongside with different
applications on graphs, facility and resources allocation problems, clustering and
partitioning problems, variations of assignment problems, etc. see in details in [4]).
The methods are roughly divisible into combinatorial and continuous. The division
O. Pichugina ()
Department of Mathematics & Statistics, Brock University, St. Catharines, ON, Canada
S. Yakovlev
Department of IT & Protection of Information, National University of Internal Affairs, Kharkov,
Ukraine
690 O. Pichugina and S. Yakovlev
is not very precise since the continuous techniques are often used for solution of
subproblems in Branch & Bound, Branch & Cut and other combinatorial algorithms.
In [3] connection between continuous and combinatorial approaches to discrete
problems are discussed, and an overview of the continuous ones is given. The last
approaches are typically classified into continuous relaxation and reformulation
based. Among the relaxation techniques there are polyhedral and semi-definite
relaxations. For instance, the semi-definite programming is applicable to the UBQP
after reformulation of (1), (2) and (3) as a linear problem in Rnn over the cut
polytope. Since the analytic description of the polytope is unknown, its semi-definite
relaxation is used [2] and efforts of researchers are focused on strengthening this
basic relaxation [5].
The UBQP equivalent reformulations depend on ways how the discrete con-
straint (3) is rewritable as continuous one, e.g.:
– Way 1 [3]: xi C yi D 1; xi yi D 0; xi ; yi 0; i 2 Jn D f1; : : :; ng, transform the
UBQP into a constrained quadratic problem in R2n ;
– Way 2 [3, 4, 6]: replacement of (3) by (4) allows to consider the unconstrained
problem (5) instead of the UBQP.
The next two approaches utilizing (4) reduce the UBQP to concave or convex
problems and are based on the following property: if f .x/ satisfies (7), then
9min ; max ; min max such that ˚ 1;2 .x; / are concave 8 max and convex
8 min .
xi .1 xi / D 0; i 2 Jn : (4)
X
˚ l .x; / D f .x/ C .xi .1 xi //l ! min : (5)
i
0 x 1; (6)
f .x/ 2 C2 .Rn /; @2 f .x/=.@xi @xj / < 1; i; j: (7)
For example, concave methods [3] are applied to (1), (2) and (3) equivalent
reformulation (2), (5), (6) (l D 1) where is chosen to provide concavity of (5).
The smoothing methods [6] utilize (5), (6) (l D 2) in a penalty function; use the
boxing constraints (6) in forming barriers and providing convexity of the function
for a range of .
We continue study of the continuous approaches to binary optimization based
on nonlinear reformulations at the same space. For the purpose, we introduce the
concepts of an extension of a function from sets onto larger sets (see Sect. 2)
and a functional representation of discrete sets (see Sect. 2.3). The presented in
Sects. 2.2 and 2.3 methods operate with sets inscribed into a sphere and convex
extensions from the sets. Ways of constructing the extensions, in particular for (2),
are described in Sect. 2.1. The Branch and Bound Polyhedral-Spherical Method
(B&BPSM) for “two-layer sets”, combining branch and bound techniques with
polyhedral and spherical relaxations, is described and specified for the UBQP
Conltinuous Approaches to the Unconstrained Binary Quadratic Problems 691
in Sect. 2.2. The Penalty Method based on Functional Representations (FRPM),

presented in Sect. 2.3, is applicable to any sets representable as intersection of
surfaces. In particular, for the binary set (3) a new representation as a touching
set of a sphere and an order four surface is given and recommended to use in the
FRPM.
2 Main Part
In the section we consider the following generalization of the UBQP – the

problem (1) over a finite discrete set inscribed into a hypersphere:
x 2 E Rn ; (8)
jEj D N < 1; (9)
E Sr .a/; (10)
2 2
Sr .a/ D fx 2 R W .x a/ D r g:
n
(11)
There is a number of combinatorial sets inscribed into a sphere, such as the binary
set Bn [6, 12], the permutation and polypermutation sets, some classes of the partial
permutations and combinations [8]. So, the class (1), (8), (9), and (10) includes all
unconstrained problems over these sets, in particular, the UBQP.
The set (9), (10) is vertex located, i.e. E coincides with the vertex set of the
corresponding polyhedron:
E D vert P where P D conv E is the polyhedron: (12)
Definition 1 ([11]) Function F.x/ is an extension of f .x/ from M onto M 0 M if

it is defined over M 0 and coincides with f .x/ over M: F.x/ D f .x/; x 2 M.
Depending on a type of the function F.x/ the f .x/-extension can be continuous,
differentiable, smooth, convex or concave, quadratic, etc.
Theorem 1 ([11]) If f .x/ is defined on a vertex located set E, then there exists F.x/,
which is a convex differentiable extension (CDE) of f .x/ from E onto Rn .
By (12), Theorem 1 is applicable to the sets (9)–(10). Therefore, we can assume
that, instead of the original function (1), its CDE from E onto Rn is considered.
A CDE can be constructed in different way, e.g., by using the next result:
Theorem 2 ([10]) If f .x/ satisfies (7) then

9K W 8k > K F.x/ D f .x/ C k .x a/2 r2 (13)
is a CDE of f .x/ from Sr .a/ onto Rn .

Remark 1 In addition, in [10] the lower estimated bound of K is given.

This property of reducing the original problem (1), (8), (9), and (10) to optimization
of a convex function, F.x/ ! min, seems useful but the problem is still discrete.
E
The convexity can be really helpful if continuous optimization is applicable to it.
The methods presented in Sects. 2.2 and 2.3 are especially effective for the
functions and the combinatorial sets with the known solutions of the following
subproblems:
1. an analytic description of the polyhedron P: P D fx W Ax bg;
2. a linear problem over E: xLP D argmin cT x;
x2E
3. the global solution of (1) on a sphere:
zS D min f .x/; xS D argmin f .x/I (14)

x2Sr .a/ x2Sr .a/
4. the minimal circumscribed sphere around E, Smin .

Alongside with (10), these four cases can simplify the solution of the next problems
considered typically complicated:
– minimization of f .x/ over a polyhedron:
zP D minf .x/; xP D argmin f .x/: (15)

x2P x2P
Knowledge of the exact solution of the subproblem 2 allows to utilize the

modification of conditional gradient methods such as the Frank-Wolfe algorithm,
where linear direction-finding subproblems are solved explicitly [12];
– search of a projection of an arbitrary point x 2 Rn onto a discrete set E, y D PrE x.
This problem for sets satisfying (10) is reduced to the linear subproblem 2.
Note that for the UBQP the optimization approaches are applicable effectively.
Indeed, the explicit solution of the subproblems: #1 is the unit hypercube (6); #2 –
i D 0; if ci 0; otherwise 0 .i 2 Jn /; #3 – is presented in [1]; #4 – is:
is xLP
Smin D S pn .0.5/ : (16)

2
2.1 Convex Extensions of Quadratic Functions
We present two ways of constructing convex extensions of (2) from Bn onto Rn :

Way 1: Notice that ˚ 1 .x; / in (5) is rewritable as

˚ 1 .x; / D f .x/ .x 0.5/2 n=4 : (17)
By (16), (17) is an extension of any f .x/ from Smin and 8E Smin onto Rn .
Using Theorem 2 and taking D k in (13), we obtain that (17) is the CDE
of (7) 8 < min D K, where K can be estimated according to Remark 1.
In particular, (17) is a CDE of (2) from Bn onto Rn if < min , where the bound
min can be found exactly: min is the minimal eigenvalue of the matrix A.
Way 2 : The convex extensions can be formed specifically for a discrete set. For
instance, a convex extension of (2) from Bn onto Rn can be constructed as follows:
– the function (2) is represented as a sum of a convex linear form f1 .x/ and a
quadratic form f2 .x/:
X
n X
n X
n X
n X
n
f1 .x/ D ci xi ; f2 .x/ D aij xi xj D jaij j .˙xi xj /I
iD1 iD1 i j iD1 i j
– the convex extensions are formed for each nonconvex term jaii j x2i and ˙jaij j
xi xj using the property (4): 8i 2 Jn ; x2i D xi . For the both types of nonconvex
Bn
terms the convex extensions are formed as follows:
(a) if i D j and aii < 0, then the CDE of the component fii D x2i is Fii D xi ;
(b) if i ¤ j and aij
¤ 0, then the CDEFij of
the component fij D ˙xi xj is:
1
2 2 2 1
2
fij D ˙xi xj D 2 xi ˙ xj xi xj D 2 xi ˙ xj xi xj D Fij ;
Bn
– all these convex extensions are incorporated to the final CDE F.x/ of (2):
X
n X
n X X
F.x/ D f .x/ C D x2i D xi , where D D jaii j C jaij j:
iD1 iD1 aii <0 aij ¤0;i<j
2.2 The Polyhedral Spherical Method
The discrete problem (1), (8)–(10) is equivalent to the f .x/-optimization problem (1)
over intersection of a sphere (11) and a polyhedron (12), f .x/ ! min. It yields
P\Sr .a/
two types of continuous relaxations: Relaxation 1 – optimization over a polyhedron,
which is convex problem (15); Relaxation 2 – optimization over a sphere, which is
nonconvex problem (14).
The Polyhedral-Spherical Method (PSM) is the cutting-plane method where
cuttings of infeasible parts of a hypersphere are conducted by polyhedron facets.
In [9, 12] the approximating version of this method was presented and applied
to quadratic optimization over the permutation set and the binary set, respectively.
This version essentially uses: (a) the convex extensions of a quadratic function;
(b) simplicity of search of all stationary points of a convex quadratic function
on a sphere; (c) easiness of solving (15) over the hypercube (6). The search of
x D argminf .x/ is restricted to facets cutting the stationary points. Then the
E
problem is reduced to search of these points and solving the subproblems of
type (14), (15) on these facets, etc. As a result, a series of optimization problems of
different dimensions with convex objective functions over sets inscribed into spheres
is solved.
In [7] the method is generalized for any sets inscribed into a sphere and is named
the Polyhedral-Spherical Method.
In [12] the exact version for (1), (2), and (3) is also presented, which utilizes the
decomposition of Bn into two subsets located on parallel facets of (6).
In this paper, we modify the exact method and generalize it, in the
Branch&Bound PSM (B&BPSM), to any “two-layer” combinatorial set, i.e. it
allows the decomposition: 8a; b W ax D b is a P-facet dim conv.fx 2 E W ax ¤
bg/ < dim P.
The B&BPSM Algorithm for two-layer sets
Step 1. The unconstrained convex problem (1) is solved, z D min
n
f .x/ D f .x /,
R
and x is projected onto E, y D PrE x . The initial upper zu and lower zl
estimated bounds are: zl D z ; zu D z.y / D zy .
Step 2. Depending on location of x we choose: Scheme 1 – solving the Relax-
ation 2 problems at each step – if x 2 P (xP D x ); Scheme 2 – solution of
Relaxation 1–2 problems at each iteration – if x … P.
Step 3. Initialization H D f;g, BH D Bf;g is the root of the search tree. The
candidate solution set B D Bf;g .
Step 4. Take the branch BH 2 B with the least lower bound and jHj < n. If the
0
choice is not unique, then among BH 2 fBh ; Bh g the branch Bh is examined first,
h0
B – next.
If for BH the relaxation problems were solved, go to Step 4.2, otherwise to Step
4.1:
Step 4.1. For BH solve the relaxation and projection problems on RnjHj -facet
depending on the scheme: (a) for Scheme 1 – xS;H ; yS;H , the bounds are zl;H D
x ; z D min.z ; zy /; (b) for Scheme 2 – x
zS;H ; xP;H ; yS;H ; yP;H , the bounds
u u S;H S;H
are z D max.zx ; zx /; z D min.z ; zy ; zy /;

l;H S;H P;H u u S;H P;H
Step 4.2. Let:

– ˘.xS;H / D fx W ax D bg be a cutting plane for the point xS;H nearest to xP
such that fx W ax bg is a right cut of xS;H , EH D E \ ˘.xS;H /;
0
– ˘ 0 .xS;H / D fx W a0 x D b0 g be the second layer, EH D EnEH , facet.
0
The sets EH ; EH form two branches from the node BH – BfH;jHC1jg and
0 fH;jHC1jg fH;jHC1j0 g
BfH;jHC1j g - with the same lower bound zu;B
H
D zu;B D zu;B . They
0
are added to B instead of BH , B D BnBH [ fBfH;jHC1jg ; BfH;jHC1j g g.
Step 4 is repeated until the candidate solutions with lower bound less than zu
remain.
The B&BPSM for the UBQP. Step 4.2 becomes: ˘.xS;H / D fx W xi D bg. The rule
for cutting plane choice on Step 4.2 becomes: ˘.xS;H / D fx W xi D bg, ˘ 0 .xS;H / D
fx W xi D 1 bg where b 2 f0; 1g, i 2 Jn is chosen depending on which one of the
restrictions xi 0 or xi 1 is mostly violated at xS;H . Respectively, new branches
0
BfH;jHC1jg ; BfH;jHC1j g correspond to fixing the coordinate xi .
Example 1 Solve the following problem: f .x/ D 33:1 x21 C 23:7 x22 C 21:6 x23
11:4 x1 x2 C 38:8 x1 x3 15:6 x2 x3 52 x1 4 x2 39 x3 ! min; x 2 B3 :
Solution outline:
1. n D 3, f .x/ is convex, x D 12 A1 c D .0:51; 0:40; 0:59/, y D .1; 0; 1/. The
initial bounds are zl D z D 25:54; zu D zy D 2:5;
2. x is an interior point of the unit hypercube (xP D x , yP D y ). The results of
the relaxation Scheme 1: xS D .0:06; 0:51; 1:25/, yS D .0; 1; 1/. The bounds are
zl D zS D 21:30, zu D min.zu ; zSy / D min.2:50; 13:30/ D 13:30 (the record
occurs at xu D .0; 1; 1/).
3. H D f;g, the right cut for xS;H D xS is uniquely defined by x3 1. Respectively,
x3 is fixed and exploring of the search tree starts with the branch BfH;jHC1jg D
0 0
B1 W x3 D 1, the next one is BfH;jHC1j g D B1 W x3 D 0.
4. The reduction is conducted (H D f1g) and, at first, the Relaxation 2 problem on
the plane x3 D 1 is solved yielding xS1 D .0:09; 0:11; 1/, zl1 D zS1 x D 17:7.
yS1 D .0; 0; 1/, zS1 y D 17:4 > zu
; new record zu
D 17:4 occurs at xu D
.0; 0; 1/.
The right cut for xS1 is x1 0 yields two branches with H D fH; jH C 1jg D
0
f1; 2g, B12 W x1 D 0; x3 D 1 and B12 W x1 D 1; x3 D 1, with fixed x1 and the
0
lower bound zl12 D zl12 D zl1 D 17:7.
0 0 0 S10
5. The same is conducted for B1 : xS1 D .1:04; 0:04; 0:00/, zS1 x D 18:92; y D
0
.1; 0; 0/, zS1
y D 18:9. New record is zu
D 18:9 at x u
D .1; 0; 0/ and the
0 0 0
bound is zl1 D 18:92. The right xS1 -cut is x1 1 yields the branches B1 2 W
10 20 l10 2 l10 20
x1 D 1; x3 D 0 and B W x1 D 0; x3 D 0, with the lower bound z Dz D
0
zl1 D 18:9.
0 0 0
6. The least lower bound 18:92 corresponds to B1 2 and B1 2 , which are examined
0
consecutively. First, for H D f1 ; 2g Relaxation 2 is conducted. Since, jHj D n
0 0
1, B1 2 corresponds to 1-sphere, where, in addition to yS1 D .1; 0; 0/, it is enough
S10 0
to analyse another endpoint t D .1; 1; 0/ of the hypercube edge, zS1 t D 10:6.
l10 2 S10 S10 0 0
z D min.zy ; zt / D min.18:9; 10:6/ D 18:9. For H D f1 ; 2 g and B1 2 0 0
0 0 0 0 0 0 0 0
the result of Relaxation 2 is xS1 2 D .0; 0; 0/. Since zl1 2 D zS1 2 D 0 > zu , B1 2
is discarded.
7. No bud nodes with less lower bound than zu D 18:9 left. The examination of
the search tree is terminated. The solution is x D .1; 0; 0/, z D 18:9.
Illustration of the solution is shown in Figs. 1, 2, and 3.
Fig. 1 The touching the level

surface at xS . In Figs. 2 and 3
the reduced problems solving
at Step 4 are demonstrated,
namely, the ellipsoids
0
yielding xS1 ; xS1 alongside
0
with the projections yS1 ; yS1 ,
their cuts and the branches of
the next level
Fig. 2 B1 exploration 1.2
0.8
0.6
x2
B12’
0.4 B12
0.2
xS1
–0.2 0 0.2 0.4 0.6 0.8 1 1.2
x1
–0.2
yS1
2.3 The Penalty Method Based on Functional Representations
Definition 2 The following system of functions:
f.x/ D f fi .x/gi2Jm (18)
is called a functional representation of a set M Rn if
x0 2 M , fi .x0 / D 0 8fi .x/ 2 f.x/: (19)
Similar to the extensions of functions, depending on a type of functions in (18), the

sets’ representations can be continuous, differentiable, convex, smooth, etc.
0
Fig. 3 B1 exploration 1.2
1
tS1’2
0.8
x2
0.6
B1’2’
0.4 B1’2
0.2 S1’2’
X
xS1’
–0.2 0 0.2 0.4 0.6 0.8 1 1.2
x1 yS1’=XS1’2’
–0.2
Definition 3 The representation (18) is irreducible if it is impossible to extract

anything from (18) without violating (19): 8j 2 Jm M ¤ fx W fj .x/ D 0gfi .x/2f.x/;i¤j .
The underlying idea of the Penalty Method based on Functional Representations
(FRPM) is constructing and applying irreducible differential representations of the
initial discrete set E in penalty optimization.
Suppose, for E its differentiable representation is known. Then, by (19), the
discrete problem (1), (8) is transformed into the constrained continuous problem (1)
subject to f.x/ D 0. After placing these constraints in a penalty function, say
quadratic, the problem becomes unconstrained one of minimizing the differentiable
function:
X
m
2
˚.x; / D f .x/ C f i .x/ ! min: (20)
>0
iD1
The problem (20) is solved for an increasing sequence
2 D fk W k < kC1 ; 0 D 0gk2JK0 DJK [f0g : (21)
Its solutions gradually approach a local minimizer x of (1): xk ! x . Notice

k!1
that ˚.x; / in (20) is an extension of (1) onto Rn and a sum of a convex objective
function (the first term) and a nonconvex penalty function (the second term). Hence
9k0 W ˚.x; k / is convex for k k0 and the solution of (20) is unique, otherwise is
not and the rest of the search is held in the vicinity N.x / of x .
For the problem (1), (8), (9), and (10), by using Theorem 2, the region of ˚.x; /-
convexity can be increased and, respectively, the vicinity N.x / can be narrowed by
switching from (20) to the following extension of (1) from E onto Rn :
X
m
2
˚.x; ; / D f .x/ C ..x a/2 r2 / C f i .x/ .; > 0/: (22)
iD1
The problem (22) can be solved for two increasing sequences, (21) and fj W j <
jC1 gj , in the way described below:
The FRPM Algorithm for (1), (8) based on (22).
Step 0. Set j D 0, 0 D 0, " > 0, the upper and lower bounds zl D 1, zu D 1;
Step 1. Solve (22) for (21), D j , and obtain fxjk gk ; xj D xjK . zl D max.zl ; zj /,
zu D min.zu ; f .PrE .xjk //;
k
Step 2. For (22), by Remark 1, find the lower bound 0j : 8 > 0j ˚.x; K ; /
is convex. Solve (22) for D K , 2 Mj D fj0 > 0 W j0 <
0
j0 C1 ; J 0 > 0j gj0 2JJ0 , the solutions are fxjK gj ; xK D xJ K . zl D max.zl ; zK /,
zu D min.zu ; f .PrE .xjK //;
j
Step 3. Set f .x/ D ˚.x; K ; J 0 /;
Step 4. Repeat Steps 1–3. The process terminates if zu zl < ".
Assuming that n > 3, we classify functional representations of discrete sets
depending on a number of functions in (18) as follows: (a) intersecting functional
representation if m D n implying that E is formed as an intersection of n surfaces;
(b) touching functional representation if m D 2, hence, E is the set of touching
points of two surfaces; (c) otherwise mixed functional representation.
Deriving functional representations of combinatorial sets is an interesting on its
own problem, because it provides better understanding the sets topological structure.
Among this class of problems finding the touching representations is a problem of
particular interest since it investigates extremal properties of the sets. Indeed, let
assume that a touching representation of E is known and the functions in (18) are
enumerated such that the surface S2 D f f2 .x/ D 0g is inscribed into S1 D f f1 .x/ D
0g. Then E is the set of local minimizers of f2 .x/ on the surface S1 .
Touching sets representations, especially differentiable ones, are useful in opti-
mization, e.g., they allows usage in the FRPM, instead of (22), the penalty function:
2
X 2
X 2
˚.x; ;
/ D f .x/ C i f i .x/ C i f i .x/ ; (23)
iD1 iD1
and managing a few penalty parameters 1 ; 2 ; 1 ; 2 for obtaining the solutions

closer to exact one.
Fig. 4 Illustration of the

touching representations (24).
The Fig. 4 shows the unit
hypercube inscribed into the
order four surface S2 . In turn,
S2 is inscribed into the sphere
S1 D Smin . The binary set B3
is the “touching set” of S1 , S2
The Functional Representations of Bn . We show that the FRPM based on any

of (22), (23) can be used for (1), (3):
– Equation (22) is applicable since the irreducible intersecting representation
fi .x/ D xi x2i ; i 2 Jn , is derived directly from (4) and Bn is inscribed into
the sphere (16).
– Equation (23) can be used for the following irreducible touching differentiable
representation of Bn , which we derived (the geometric interpretation is shown in
Fig. 4):
X
n X
n
f1 .x/ D .xi 0:5/2 D n=4; f2 .x/ D .xi 0:5/4 D n=16: (24)
iD1 iD1
We recommend for the UBQP utilizing (24) in the penalty function (23) of (1),
(2), and (3).
3 Conclusion
Two approaches, the B&BPSM and FRPM, to discrete optimization over sets
inscribed into a sphere are presented. For the UBQP they use analytic descriptions
of Bn as an intersection of a hypercube with a sphere and as a touching set; convex
extensions of quadratic functions. The approaches can be applied to other sets with
known analytic description of a convex hull, a circumscribed surface, a functional
representation, and to any function with known the global minimizer on the surface.
References
1. Dahl, J.: Convex optimization in signal processing and communications. Ph.D. dissertation.
Department of Communication Technology, Aalborg University (2003)
2. Helmberg, C., Rendl, F.: Solving quadratic (0,1)-problems by semidefinite programs and
cutting planes. Math. Program. 82(3), 291–315 (1998)
3. Hillier, F.S., Appa, L., Pitsoulis, L., Williams, H.P., Pardalos, P.M., Prokopyev, O.A., Busygin,
S.: Continuous approaches for solving discrete optimization problems. In: Handbook on
Modelling for Discrete Optimization, pp. 1–39. Springer, New York (2006)
4. Kochenberger, G., Hao, J.-K., Glover, F., Lewis, M., Lu, Z., Wang, H., Wang, Y.: The
unconstrained binary quadratic programming problem: a survey. J. Comb. Optimiz. 1, 58–81
(2014)
5. Krislock, N., Malick, J., Roupin, F.: Improved semidefinite bounding procedure for solving
Max-Cut problems to optimality. Math. Program. 143(1–2), 61–86 (2014)
6. Murray, W., Ng, K.-M.: An algorithm for nonlinear optimization problems with binary
variables. Comput. Optimiz. Appl. 47(2), 257–288 (2008)
7. Pichugina, O., Yakovlev, S.: Polyhedral-spherical approach to solving some classes of com-
binatorial optimization problems. In: Proceedings of the 6th International School-Seminar on
Decision Theory, Uzhgorod, pp. 152–153 (2012)
8. Stoyan, Y.G., Yemets, O.: Theory and Methods of Euclidean Combinatorial Optimization.
ISSE, Kiev (1993)
9. Stoyan, Y.G., Yakovlev, S.V., Parshin, O.V.: Quadratic optimization on combinatorial sets in
Rn . Cybern. Syst. Anal. 27(4), 562–567 (1991)
10. Stoyan, Y., Yakovlev, S., Emets, O., Valuiskaya, O.: Construction of convex continuations for
functions defined on a hypersphere. Cybern. Syst. Anal. 34(2), 176–184 (1998)
11. Yakovlev, S.: The theory of convex continuations of functions at the vertices of convex
polygons. Comput. Math. Math. Phys. 34(7), 959–965 (1994)
12. Yakovlev, S., Grebennik, I.: Certain classes of optimization problems on set distributions and
their properties. Izvestiya Vysshikh Uchebnikh Zavedenii Mathematika 11, 74–86 (1991)
Fixed Point Techniques in Analog Systems
Diogo Poças and Jeffery Zucker
Abstract Analog computation is concerned with continuous rather than discrete

spaces. Most of the physical processes arising in nature are modeled by differential
equations, either ordinary (example: spring/mass/damper system) or partial (exam-
ple: heat diffusion). In analog computability, the existence of an effective way to
obtain solutions (either exact or approximate) of these systems is essential.
We develop a framework in which the solutions can be seen as fixed points
of certain operators on continuous data streams, using the framework of Fréchet
spaces. We apply a fixed point construction to retrieve these solutions and present
sufficient conditions on the operators and initial inputs to ensure existence and
uniqueness of these corresponding fixed points.
1 Introduction
Analog computation, as conceived by Kelvin [10], Bush [1], and Hartree [4], is
a form of experimental computation with physical systems called analog devices
or analog computers. Historically, data are represented by measurable physical
quantities, including lengths, shaft rotation, voltage, current, resistance, etc., and
the analog devices that process these representations are made from mechanical or
electromechanical or electronic components [5, 7, 9].
The main objects of our study are analog networks or analog systems, [6, 11–13],
whose main components are described as follows:
Analog network D data C time C channels C modules:
We will model data as elements from a topological vector space A that is

actually a Fréchet space. We will use the nonnegative real numbers as a continuous
model of time T D Œ0; 1/. Each channel carries a continuous stream, represented
as a function u W T ! A (this space is denoted by C.T; A /). Each module M is
D. Poças () • J. Zucker

Department of Mathematics and Statistics, McMaster University, Hamilton, ON, Canada

702 D. Poças and J. Zucker
Fig. 1 Analog network (with g

two modules and three
channels) for the time t
u Lu u
evolution problem L
0
specified by a stream function
FM W C.T; A /k A ` ! C.T; A /:
In this way, we can think of analog networks as directed graphs where modules
are nodes and channels are edges (see Fig. 1). We will use analog systems to study
the time evolution problem (also called the Cauchy problem).
Definition 1 (Time evolution problem, [3]) For a given initial condition g 2 A
and an operator L W A ! A , the time evolution problem is given by the system
du
dt
D Lu; t 2 TI
(1)
u.0/ D g:
We look for a solution u 2 C.T; A /.

To construct an analog system that represents the time evolution problem, we can
simply integrate the differential equation (1) to obtain
Z t
u.t/ D g C Lu.s/ds DW Fg .u/; (2)
0
where we use the right hand side to define an operator Fg W C.T; A / ! C.T; A /,
which can be computed using an analog network with two modules. Introducing a
feedback to implement the equality, we obtain the analog system of Fig. 1.
We can informally define a specification of the analog network as a tuple of
streams describing the data on all channels which satisfy the equations given by the
modules. We can then observe the equivalence between the notions of (a) solutions
to the time evolution problem of Definition 1; (b) specifications of the analog system
of Fig. 1; and (c) fixed points of the operator Fg of Equation (2). Henceforth we will
focus on the last notion. Our goal is to provide sufficient conditions on Fg that ensure
existence and uniqueness of fixed points, as well as the existence of a constructive
method to obtain fixed points when they exist.
In Sect. 2 we introduce Fréchet spaces, which form the framework for our
problem at hand. In Sect. 3 we assume analyticity of g to prove local existence and
convergence of fixed points (Theorem 2). In Sect. 4 we first extend our results to
global existence and convergence (Theorem 3) and then extend our constructive
Fixed Point Techniques in Analog Systems 703
method for different choices of initial input (Theorem 5). Finally, we turn to
uniqueness of fixed points and prove it for certain operators (Theorem 6).
2 Fréchet Spaces
For the remainder of this paper, we use the following notation:

Notation 1 A is the space of infinitely differentiable functions, A D C1 .R/.
Notation 2 The operator L W A ! A is given by Lu D ˛@x u, for some ˛ 2 R.

Thus, the operator Fg becomes
Z t
Fg .u/.t; x/ D g.x/ C ˛ @x u.s; x/ds: (3)
0
As we know, A D C1 .R/ is not complete under the supremum norm; however,

for each x0 2 R, X 2 RC and k 2 N, we can define a pseudonorm by
ˇ k ˇ
ˇ@ f ˇ
k f kX;x0 ;k ˇ
D sup ˇ k .x/ˇˇ :
jxx0 jX @x
It turns out that the notion of interest is that of Fréchet space, which we now
briefly review (a detailed exposition can be found in [8, Ch. V]).
Definition 2 (Fréchet space [8]) A Fréchet space is a topological vector space
X whose topology is induced by a countable family of pseudonorms fk k˛ g˛2A .
Moreover,
• the family fk k˛ g˛2A separates points, that is,
u D 0 if and only if kuk˛ D 0 for all ˛I
• X is complete with respect to fk k˛ g˛2A , that is, for every sequence .xn / which is
Cauchy with respect to each pseudonorm k k˛ , there exists x 2 X such that .xn /
converges to x with respect to each pseudonorm k k˛ .
Example 1 The space A D C1 .R/ of infinitely differentiable functions in R is a
Fréchet space with the countable family of pseudonorms given by
ˇ k ˇ
ˇ@ f ˇ
k f kN;k ˇ
D sup ˇ k .x/ˇˇ ; (4)
NxN @x
for N; k 2 N.
Example 2 The space C.T; A / of continuous streams is also a Fréchet space, with
the countable family of pseudonorms given by
ˇ k ˇ
ˇ@ u ˇ
kukM;N;k D sup sup ˇˇ k .t; x/ˇˇ ; (5)
0tM NxN @x
for M; N; k 2 N.
We can see that the family of pseudonorms in Example 2 is closely related
to the family in Example 1. In fact, this illustrates a useful property of Fréchet
spaces; in general, the space of continuous streams over a Fréchet space is itself a
Fréchet space. In other words, Fréchet spaces work well with the operation of taking
continuous streams.
Proposition 1 (New Fréchet spaces from old) If A is a Fréchet space with a
countable family of pseudonorms fk k˛ g˛2A , then so is C.T; A / with the countable
family of pseudonorms fk kM;˛ gM2N;˛2A , where
kukM;˛ D sup ku.t/k˛ :

0tM
Even though Fréchet spaces, as they stand, are not necessarily normed spaces, we
can define a metric from the pseudonorms, under which these spaces are complete.
Proposition 2 Given a Fréchet space, we can define a complete metric from the
family of pseudonorms which induces the same topology.
For a proof, see [8, Ch. V], in particular Theorem V.5.
The usefulness of complete metric spaces is evident due to the following.
Theorem 1 (Banach fixed point theorem) Given a complete metric space .X; d/,
suppose that T W X ! X is a contracting operator in the sense that there exists
0 < 1 with
d.T.x/; T.y// d.x; y/ for all x; y 2 X:
Then T has a unique fixed point x . Moreover, for all x0 2 X the sequence of
iterations xn WD T n .x0 / converges to x .
3 Local Convergence Theorem
Consider the space C.T; A /. Take any (arbitrary but fixed) x0 2 R, X 2 RC , T 2 T.

Then, for any k 2 N, we have a pseudonorm
ˇ k ˇ
ˇ@ u ˇ
kukT;X;x0 ;k ˇ
D sup ˇ k .t; x/ˇˇ : (6)
0tT @x
jxx0 jX
Fig. 2 Compact rectangles t
· T,X,x0 ,k
x
x0 − X x0 x0 + X
Observe that we are taking suprema on compact rectangles of the form Œ0; T
Œx0 X; x0 C X (see Fig. 2). The reason for taking suprema on compact rectangles
will be made clear shortly with Theorem 2 (local convergence theorem). We also
observe that, for each compact rectangle X0 D Œ0; TŒx0 X; x0 CX, we can define
the space of compact continuous streams C.Œ0; T; C1 .x0 X; x0 CX//. Clearly, any
function in C.T; A / can be mapped to a function in C.Œ0; T; C1 .x0 X; x0 C X//
via the restriction u 7! uX0 . Moreover, C.Œ0; T; C1 .x0 X; x0 C X// can be seen
to be a Fréchet space with the family of pseudonorms k kT;X;x0 ;k given by (6). Note
that x0 , X and T are fixed and the indexing is on k 2 N.
Finally, observe that the operator Fg W C.T; A / ! C.T; A / has a restriction
FgX0 to the space C.Œ0; T; C1 .x0 X; x0 C X//.
Our next step is to prove contraction inequalities, which play an important role
in fixed point techniques.
Lemma 1 (Contraction inequalities) Consider the Fréchet space C.T; A / with
pseudonorms kkT;X;x0 ;k given by (6). Let g 2 C1 .R/ and Fg W C.T; A / ! C.T; A /
be given by (3). Then, for any u; v 2 C.T; A /, any pseudonorm k kT;X;x0 ;k and any
m 2 N, we have the following bound:
.j˛jT/m
kFgm .u/ Fgm .v/kT;X;x0 ;k ku vkT;X;x0 ;kCm : (7)
mŠ
Proof By induction on m. t
u
Let us see how we can use these bounds in a proof.
Theorem 2 (Local Fréchet space convergence theorem) Consider the Fréchet
space C.T; A / with pseudonorms k kT;X;x0 ;k given by (6). Take an initial input
u0 2 C.T; A / and initial condition g 2 C1 .R/. Assume also that u0 D 0 and g is
analytic at x0 with some radius of convergence1 R. Let Fg W C.T; A / ! C.T; A /
be given by (3). Then, for any T; X 2 RC such that j˛jT C X < R, the sequence .um /
1
Or equivalently, that g has a holomorphic extension on a disk of the complex plane with center x0
and radius R; see Remark 1.
given by um D Fgm .u0 / converges in the rectangle X0 D Œ0; T Œx0 X; x0 C X to

a fixed point of FgX0 .
Proof To ease the exposition we introduce the pseudonorms on g given by
ˇ k ˇ
ˇ@ g ˇ
kgkX;x0 ;k ˇ
D sup ˇ k .x/ˇˇ ; for x0 2 R; X 2 RC ; k 2 N: (8)
jxx0 jX @x
Since g is analytic at x0 with radius of convergence R, there is a sequence of real

coefficients .aj / such that, for all x 2 .x0 R; x0 C R/,
1
X
g.x/ D aj .x x0 /j :
jD0
p 1
It also follows that lim supn!1 n jan j R
(see Footnote 1). Moreover, we have
the following bound, for any X < R:
ˇ ˇ
ˇ X1 ˇ X 1
ˇ . j C k/Š ˇ . j C k/Š
kgkX;x0 ;k D ˇˇ sup ajCk .x x0 /j ˇˇ jajCk jX j : (9)
ˇjxx0 j<X jD0 jŠ ˇ jD0 jŠ
Let T; X 2 RC such that j˛jT C X < R. We show that .um / is a Cauchy sequence
with respect to the pseudonorm k kT;X;x0 ;k . First observe that
1
X 1
X
kumC1 um kT;X;x0 ;k D kFgm .g/ Fgm .0/kT;X;x0 ;k
mD0 mD0
1
X
1 Tm
j˛jm kgkX;x0 ;kCm
mD0
mŠ
1 X
X 1
2 .k C m C j/Š
j˛jm T m X j jakCmCj j
mD0 jD0
mŠjŠ
1 X
X s
3 .k C s/Š
D j˛jm T m X sm jakCsj
sD0 mD0
mŠ.s m/Š
1
X
4 .k C s/Š
D .j˛jT C X/s jakCsj ;
sD0
sŠ
where (1) is justified by the Contraction Inequalities (Lemma 1), (2) by equation (9),
(3) by rearranging the sum and adding over diagonals s D m C j and (4) by taking
the binomial expansion of .j˛jT C X/s .
By the root test, the above series is convergent, since

r
s .k C s/Š p R
lim sup .j˛jT C X/s jakCs j D .j˛jT C X/ lim sup n jan j 1 < D 1:
s!1 sŠ n!1 R
Since the series is convergent, it follows that, for i < j,
X
j1 1
X
kuj ui kT;X;x0 ;k kumC1 um kT;X;x0 ;k kumC1 um kT;X;x0 ;k ! 0:
i!1
mDi mDi
Hence .um / is a Cauchy sequence with respect to the pseudonorm k kT;X;x0 ;k .

Since this holds for all k 2 N and C.Œ0; T; C1 .x0 X; x0 C X// is complete, it
follows that .um / has a limit in X0 . Now, using continuity of FgX0 , we conclude that
this limit must be a fixed point of FgX0 . t
u
Remark 1 The reader should distinguish between the following two concepts:
• the existence of a holomorphic function, defined in a disk of the complex plane
C, which coincides with g at the real axis fy D 0g;
• the convergence of the construction um D Fgm .0/ to a fixed point u, defined in a
rectangle of T R, which coincides with g at initial time ft D 0g.
As seen from Theorem 2, the existence of a holomorphic extension implies
convergence to a fixed point. Both these functions (the holomorphic extension
and the fixed point) can be depicted by planar diagrams, and both can be seen as
extensions of g (see Fig. 3). However, these functions, and the domains which they
inhabit, are substantially different.
y t
C T×R
R
T
g(x + iy)
u(t, x)
· · · x · · · x
x0 − R x0 x0 + R x0 − X x0 x0 + X
Fig. 3 On the left: a function g.x C iy/ of type C ! C, defined in a disk, that coincides with g
at fy D 0; x0 R < x < x0 C Rg. On the right: a fixed point u.t; x/ of type T R ! R, defined
in a rectangle, that coincides with g at ft D 0; x0 X < x < x0 C Xg. The rectangle and disk
dimensions follow the relation j˛jT C X < R
4 Global Convergence Theorems
As an immediate corollary of Theorem 2, we have:

Theorem 3 (First global Fréchet space convergence theorem) Consider the
Fréchet space C.T; A / with pseudonorms k kT;X;x0 ;k given by (6). Take an initial
input u0 2 C.T; A / and initial condition g 2 C1 .R/. Assume also that u0 D 0
and g is entire (i.e. has a holomorphic extension to the complex plane). Let
Fg W C.T; A / ! C.T; A / be given by (3). Then the sequence .um / given by
um D Fgm .u0 / converges to a fixed point of F.
Proof Since g is entire, it is analytic at any x0 with any radius of convergence R.
Thus, by Theorem 2, the sequence um converges to a fixed point on any compact
rectangle Œ0; T Œx0 X; x0 C X. Therefore, we have convergence of um for any
pseudonorm k kT;X;x0 ;k , so that we have convergence in C.T; A /. t
u
The next step is to generalize Theorem 3 to a larger class of initial functions u0
(other that u0 D 0). We do that proof in two steps: first assume g D 0 to establish
sufficient conditions on u0 ; then consider the more general case g 2 C1 .R/.
Definition
P1 3 We say that a function u 2 C.T; A / is uniformly entire if u.t; x/ D
jD0 a j .t/x j
for some sequence of functions .aj / 2 C.R/ such that
r
lim j sup jaj .t/j D 0 for all T 2 T:
0tT
The motivation for this terminology is that, for p such a function u, the section
x 7! u.t; x/ is entire for all t, and the convergence j jaj .t/j ! 0 is uniform in t.
Theorem 4 Consider the Fréchet space C.T; A / with the family of pseudonorms
k kT;X;x0 ;k given by (6). Let also u0 2 C.T; A / be an initial input, and g 2 C1 .R/
be an initial condition. We assume in addition that u0 is uniformly entire and g D 0.
Let F0 W C.T; A / ! C.T; A / be given by
Z t
F0 .u/.t; x/ D ˛ @x u.s; x/ds: (10)
0
Then the sequence .um / given by um D F0m .u0 / converges to 0.

Proof To ease the exposition we introduce the pseudonorms on aj given by
kaj kT D sup jaj .t/j, for T 2 T: (11)

0tT
X
We show that kum kT;X;0;k is a convergent series for any pseudonorm kkT;X;x0 ;k
m
with x0 D 0. We have that (see proof of Theorem 2)
1
X 1
X
kum kT;X;0;k D kF0m .u0 / F0m .0/kT;X;0;k
mD0 mD0
X1
j˛jm T m
ku0 kT;X;0;kCm
mD0
mŠ
ˇ ˇ
1 ˇX ˇ
X j˛jm T m ˇ 1 .j C k C m/Š ˇ
D sup ˇˇ ajCkCm .t/xj ˇˇ
mD0
mŠ 0tT ˇ jD0 jŠ ˇ
jxjX
1 X
X 1
.j C k C m/Š
j˛jm T m X j kajCkCm kT
mD0 jD0
mŠjŠ
1 X
X s
.k C s/Š
D j˛jm T m X sm kakCs kT
sD0 mD0
mŠ.m s/Š
X1
.k C s/Š
D .j˛jT C X/s kakCs kT :
sD0
sŠ
By the root test, the above series is convergent, since

r
s .k C s/Š
.j˛jT C X/s ! j˛jT C X
sŠ s!1
p P
and s
kakCs kT ! 0 by assumption. Therefore kum kT;X;0;k is convergent, so that
s!1
kum kT;X;0;k ! 0 and thus um converges to 0, as we wanted to prove. t
u
m!1
We now combine Theorems 3 and 4 to prove our most general result.

Theorem 5 (Second global Fréchet space convergence theorem) Consider the
Fréchet space C.T; A / with pseudonorms k kT;X;x0 ;k given by (6). Take an initial
input u0 2 C.T; A / and initial condition g 2 C1 .R/. Assume also that u is
uniformly entire and that g is entire. Let Fg W C.T; A / ! C.T; A / be given by (3).
Then the sequence .um / given by um D F m .u0 / converges to a fixed point of F.
Proof Let Fg ; F0 W C.T; A / ! C.T; A / be given by (3), (10). We observe that, for
any u; v 2 C0;1 .X/ we have
Z t Z t Z t
Fg .u C v/ D g C ˛ .u C v/x ds D g C ˛ ux ds C ˛ vx ds D Fg .u/ C F0 .v/:
0 0 0
We can then infer that u1 D Fg .u0 / D Fg .0 C u0 / D Fg .0/ C F0 .u0 /. Also,

u2 D Fg .u1 / D Fg .Fg .0/ C F0 .u0 // D Fg2 .0/ C F02 .u0 /, and, in general,
um D Fgm .0/ C F0m .u0 /:
By Theorem 3, .Fgm .0// converges to a fixed point of Fg . By Theorem 4, .F0m .u0 //

converges to 0. Therefore, .um / and .Fgm .0// have the same limit. In particular, .um /
converges to a fixed point of Fg . t
u
A nice consequence of the proof is that it allows us to also establish uniqueness
in a certain class of functions.
Theorem 6 (Uniqueness of uniformly entire fixed points) Consider the Fréchet
space C.T; A / with pseudonorms k kT;X;x0 ;k given by (6). Take an initial condition
g 2 C1 .R/ and assume also that g is entire. Let Fg W C.T; A / ! C.T; A / be
given by (3). Then there is at most one uniformly entire fixed point of Fg .
Proof Let u be any uniformly entire fixed point of Fg . By the proof of Theorem 5,
we know that u D Fgm .u/ D Fgm .0/ C F0m .u/. Since .F0m .u0 // converges to 0, we
get that .Fgm .0// converges to u. Thus any uniformly entire fixed point of Fg must
coincide with the limit of .Fgm .0//. t
u
5 Conclusion and Further Research
In this paper we have seen how to study solutions to differential equations as outputs
of analog networks, and how to obtain them using fixed point techniques. The
example we have considered (L D ˛@x ) is a well-known problem whose solution
can be obtained analytically by taking a Fourier transform. However, the method
presented in this paper provides a different perspective which is suitable for analog
computability and the study of analog systems as in [6, 11–13], where Fréchet
spaces clearly provide a natural framework.
We intend to investigate this approach (fixed points in Fréchet spaces) in more
general settings. In fact, as a next step one can look at a more general operator
L W A ! A using higher-order derivatives, for example with bounds of the form
kLukT;X;x0 ;k Ck@`x ukT;X;x0 ;k CkukT;X;x0 ;kC` : (12)
We observe that analyticity of the initial condition g is not enough to ensure

existence of solutions. A counterexample is given by the heat equation ut D uxx
1
with initial condition g.x/ D 1x . Even though g is analytic near zero, the solution
fails to be analytic at a neighborhood of the origin (see [2]). Thus, the general case
may require different tools such as explicit bounds on the pseudonorms of g.
We also plan to investigate properties of the fixed points, such as computability,
continuity and stability, as functions of the parameters and initial conditions.
References
1. Bush, V.: The differential analyzer. A new machine for solving differential equations. J. Frankl.
Inst. 212(4), 447–488 (1931)
2. Egorov, Y.V., Shubin, M.A.: Foundations of the Classical Theory of Partial Differential
Equations. Springer-Verlag, Berlin (1998)
3. Hadamard, J.: Lectures on Cauchy’s Problem in Linear Partial Differential Equations. Dover,
New York (1952)
4. Hartree, D.R.: Calculating Instruments and Machines. Cambridge University Press, Cambridge
(1950)
5. Holst, P.A.: Svein Rosseland and the Oslo Analyzer. IEEE Ann. Hist. Comput. 18(4), 16–26
(1996)
6. James, N.D., Zucker, J.I.: A class of contracting stream operators. Comput. J. 56, 15–33 (2013)
7. Johansson, M.: Early analog computers in Sweden – with examples from Chalmers University
of Technology and the Swedish aerospace industry. IEEE Ann. Hist. Comput. 18(4), 27–33
(1996)
8. Reed, M., Simon, B.: Methods of Modern Mathematical Physics: Functional Analysis.
Academic Press Inc., San Diego (1980)
9. Small, J.S.: General-purpose electronic analog computing: 1945–1965. IEEE Ann. Hist.
Comput. 15(2), 8–18 (1993)
10. Thompson, W., Tait, P.G.: Treatise on Natural Philosophy, vol. 1, (Part I), 2nd edn. Cambridge
University Press, Cambridge (1880)
11. Tucker, J.V., Zucker, J.I.: Computability of analog networks. Theor. Comput. Sci. 371, 115–146
(2007)
12. Tucker, J.V., Zucker, J.I.: Continuity of operators on continuous and discrete time streams.
Theor. Comput. Sci. 412, 3378–3403 (2011)
13. Tucker, J.V., Zucker, J.I.: Computability of operators on continuous and discrete time streams.
Computability 3, 9–44 (2014)
A New Look at Dummy Derivatives
for Differential-Algebraic Equations
John D. Pryce and Ross McKenzie
Abstract We show the dummy derivatives index reduction method for DAEs,
introduced in 1993 by Mattsson & Söderlind, is a particular case of the Pryce
˙-method solution scheme. We give a pictorial display of the underlying block
triangular form.
This approach gives a simple general method to cast the reduced system in semi-
explicit index 1 form, combining order reduction and index reduction in one process.
It also shows each DD scheme for a given DAE is uniquely described by an
integer “DDspec” vector ı.
The method is illustrated by an example.
We give various reasons why, contrary to common belief, converting further from
semi-explicit index 1 form to an explicit ODE, can be a good idea for numerical
solution.
1 Introduction
We give a brief summary, wherein terms in slanted font are defined more fully later.
The dummy derivatives (DDs) method for a differential-algebraic equation
(DAE), described as a way to reduce it to a (locally) equivalent DAE of index 1,
was introduced by Mattsson and Söderlind [2] (here, MS denotes this paper or its
authors).
MS use the Pantelides method [4] for structural analysis (SA) of the DAE; this
can be replaced by the Pryce signature matrix method (˙-method) [5] which can do
the same analysis more simply, and is more powerful in that Pantelides only applies
to DAEs of first order. Hence the DAEs to which DDs, as described by MS, can
be applied are precisely the SA-friendly DAEs: those for which SA (Pantelides or
Pryce) succeeds , by giving a nonsingular System Jacobian at some point.
The batches of derivatives formed by DDs are exactly those in stages of the ˙-
method’s standard solution scheme , but in reverse order, as proved in [3]. This paper
follows up this connection. Adapting notation in [5], we give a concise description
J.D. Pryce () • R. McKenzie

School of Mathematics, Cardiff University, Cardiff CF24 4AG, Wales, UK

714 J.D. Pryce and R. McKenzie
of the block-triangular form (BTF) of dependencies between batches, plus a pictorial

view. Each DD scheme for a given DAE is uniquely described by an integer DD-
specification vector ı D .ı1 ; : : : ; ın /. The DAE resulting from DDs is generally
called “index 1” but has the somewhat stronger property of being an implicit ODE .
MS ignores the conceptually trivial but practically important issue of reduction
to order 1, needed when giving a DAE to a solver such as DASSL [1]. Our approach
unifies index and order reduction into one process.
2 Basic Theory
DAEs and index As in MS (but with a different notation), we consider DAEs of

the general form
fi . t; the xj and derivatives of them / D 0; i D 1; : : : ; n; (1)
where the xj .t/; j D 1; : : : ; n are state variables that are functions of an independent
variable t, usually regarded as time.
The classic differentiation index d is the largest number of times some equation
fi of the system must be differentiated w.r.t. t to produce an enlarged system of
equations, solvable as an algebraic system to give an equivalent ordinary differential
equation system (ODE) for the state variables xj .t/.
Example 1 To contrast the above with the DDs viewpoint below, consider the DAE
0 D A WD x u.t/; 0 D B WD xP y; (dot denotes d=dt); (2)
where u is a given forcing function. It is an edge-case (in computer science terms) in

having a unique solution—zero degrees of freedom—but the general d definition
still applies. One must differentiate B once to give an equation with yP in it; now xR
appears, so differentiate A twice to produce xR which can then be eliminated. Thus
d D 2, since combining xR yP D 0 and xR uR .t/ D 0 with the original set one obtains
an ODE
xP D y; yP D uR .t/: (3)
Then (3) has the same solution as (2) through any consistent point of (2). But (3) has
DOF D 2 so its general solution x D u.t/C˛ Cˇt; y D u P .t/Cˇ (with ˛; ˇ arbitrary
scalars) includes many beside the unique solution x D u.t/; y D uP .t/ of (2).
(DOF means the phrase “degrees of freedom”, DOF is the corresponding
number.)
Solving the ODE numerically as a proxy for the DAE results in progressive drift
from the DAE’s solution manifold, which in the above example, as a subset of the
ODE’s .t; x; y/ state space, is the set M D f .t; x; y/ 2 R3 j x D u.t/, y D uP .t/ g.
A New Look at Dummy Derivatives for Differential-Algebraic Equations 715
By contrast the DDs method gives a system genuinely equivalent to the original,
but with no extra DOF that cause numerical drift. It is not concerned with producing
an ODE in all the xj but, like the ˙-method, seeks merely the least number of
differentiations that allows us to find all the xj as functions of t. In Example 1 the
answer is clearly to differentiate A once and B not at all, and rearrange to get
x D u.t/; y D uP .t/; (4)
which unlike (3) has no degrees of freedom, i.e. shows the unique solution. Strictly
by DD rules one obtains (4) via a DAE with three equations and variables, namely
0 D A D x u.t/; 0 D B D x0 y; 0 D AP D x0 uP .t/;
where the “genuine” derivative xP has been replaced by the dummy derivative x0 .
˙ -method summary The ˙-method finds the n n signature matrix ˙ D .ij /,
where ij is the order of the highest derivative of xj in fi , if xj occurs in fi , and 1
if not. A transversal is a set T of n matrix positions .i; j/, with just one in each
row and each column, P and we seek a highest-value transversal (HVT), such that the
number Val.T/ D .i;j/2T ij is maximised. We find corresponding dual variables,
the offsets, integer n-vectors c D .c1 ; : : : ; cn / and d D .d1 ; : : : ; dn / satisfying
dj ci ij ; with equality on a HVT: (5)
By default we choose the canonical offsets, the unique elementwise smallest ones
having mini ci D 0. Then ci says how many times equation fi D 0 is to be
differentiated, and (5) fulfils the requirement stated in Sect. 2.2 of MS, that “the
differentiated problem is structurally nonsingular with respect to its highest-order
.d / .c /
derivatives”—i.e., xj j occurs in fi i for each .i; j/ in some transversal.
The ˙-method’s standard solution scheme (SSS) says: let kd D maxj dj and
kc D maxi ci , then
for k D kd ; kd C1; : : : do
Solve the mk equations: fi.kCci / D 0 for those i such that k C ci 0
.kCdj / (6)
for the nk unknowns: xj for those j such that k C dj 0,
using already known values from previous (with smaller k) stages.
The sizes mk ; nk satisfy mk nk , i.e., each set of equations is underdetermined or

square. Clearly, mk ; nk increase with k, and D n for k 0. The iteration starts at
k D kd because mk D 0 for k < kc , nk D 0 for k < kd , and kd kc .
When solving a DAE by Taylor series expansion, k goes to some positive
expansion order. In the DDs context we stop at k D 0. Here, a consistent point
.q/
of the DAE shall mean a solution of (6) up to stage k D 0, i.e. values of xj for
. p/
q D 0 W dj , j D 1 W n, that satisfy fi D 0 for p D 0 W ci ; i D 1 W n. A fuller definition
is in [5].
Henceforth we assume SA succeeds—or the DAE is SA-friendly—meaning the

n n System Jacobian

.c / .d /
J D @fi i =@xj j : (7)
i;jD1 W n
is nonsingular
P at some
P consistent point. Then, the value Val.T/ of a HVT, which
also equals j dj i ci , is the number of degrees of freedom DOF of the DAE; also

.kCc / .kCd /
the mk nk Jacobian Jk D @fi i =@xj j of (6) is of full row rank; (8)
and Jk is the sub-matrix (rows where k C ci 0, columns where k C dj 0) of J.

An implicit ODE is an SA-friendly DAE with c D 0, so one can solve to get
an ODE without differentiating any equations. For instance the DAE xP C yP D 1,
x y D 0 is index 1 but not an implicit ODE, while xP C y D 1, x y D 0 is an
implicit ODE.
Example 2 Figure 1 shows the DAE of the simple pendulum in Cartesian coordi-
nates, with signature matrix ˙ annotated with the canonical offsets, and its HVTs
marked ; ı . The Jacobian J is annotated in the style of MS, also showing the offsets.
For conciseness we use notation introduced in [5]. Write xjl to mean the lth
.l/ .l/
derivative xj of variable xj D xj .t/, similarly fil to mean the lth derivative fi of
function fi . Thus for an index-pair .i; l/ or . j; l/ unless said otherwise it is assumed
that i; j; l are integers with i; j 2 1 W n and l 0.
For a set S of index-pairs let xS be the vector comprising all the xjl for . j; l/ 2 S,
listed in some arbitrary but fixed order. It may mean a vector of symbolic variables,
or real numbers, or functions of t, depending on context. Similarly, let fS be the
vector comprising all the fil for .i; l/ 2 S, in some arbitrary but fixed order.
For a given DAE with offsets ci , dj and for any integer k define index-pair sets
Ik D f .i; l/ j l D k C ci 0 g; Jk D f . j; l/ j l D k C dj 0 g: (9)
Write jXj for the number of elements in a finite set X. Then the sizes mk ; nk in (6)
are given by jIk j D mk and jJk j D nk .
Further define Ik to be the union of Il for all l k, and similarly I<k , Jk , J<k .
Then the SSS, equation (6), taken up to k D 0, can be written
Fig. 1 Pendulum DAE with signature matrix and Jacobian

Fig. 2 Ik , Jk , fIk and xJk for the pendulum, using offsets c D .0; 0; 2/ and d D .2; 2; 0/. For
example, J1 is the set f.1; 1/; .2; 1/g so that xJ1 D .x11 ; x21 /, which is the same as .Px; yP /
for k D kd ; kd C1; : : : ; 0 do
(10)
solve fIk .t; xJk / D 0, or equivalently fIk .t; xJ<k I xJk / D 0, for xJk .
The second form separates the “already known” t and xJ<k from the “to be
found” xJk .
For example, Fig. 2 tabulates Ik , Jk , fIk and xJk for the pendulum, taking x1 ; x2 ; x3
to mean x; y; and f1 ; f2 ; f3 to mean A; B; C, in the order given.
3 DDs as a Particular Case of the SSS
In the “vanilla” SSS, all components in xJk are on an equal footing. For numerical
solution, in those stages k < 0 where an underdetermined system is to be solved
because mk < nk , typically a trial xJk vector is projected on the manifold defined by
fIk D 0, using a Gauss-Newton method or similar, to obtain the accepted solution.
Example 3 For the pendulum, stage k D 2 consists in projecting a trial value of
xJ2 D .x10 ; x20 / D .x; y/ on the zero set of fI2 D . f30 / D .h/, i.e. on the circle
x2 C y2 D L2 in the x; y plane. Stage k D 1 consists in projecting a trial value
of xJ1 D .x11 ; x21 / D .Px; yP / on the zero set of fI1 D . f31 / D .h/,
P i.e. on the line
xPx C yPy D 0 in the xP ; yP plane defined by the previously found values of x and y.
In DDs, the components in xJk are not on an equal footing. The method as
.d /
presented in MS starts with the high-order derivatives xj j and works down, at each
stage selecting some derivatives to be dummy, which essentially means they are
found, as functions of others. For the pendulum, cf. [2, Sect. 4, Example 3], when
y ¤ 0 one can choose yR ; yP to be dummy derivatives y00 ; y0 , giving a reduced DAE
where everything is a function of x and xP ; when x ¤ 0 one can do the reverse.
From the ˙-method angle it is natural to go the opposite way, starting with the
low-order derivatives. DDs then becomes a particular way to perform the SSS, as
we now describe. It is proved in [6] that for a given DAE, the set of DD schemes
constructible by this “forward” method is identical with the set constructible by the
classical “reverse” method.
For brevity let all the unknowns that are found by the SSS in stages k D kd W 0 be
.q/
called items. That is, an item is any xj with 0 q dj , j 2 1 W n. Equivalently, the
items are precisely the components of xJ0 .
The forward DDs method We assume henceforth that the functions fi are suffi-
ciently smooth for all needed uses of the Implicit Function Theorem (IFT). For each
stage k the mk nk system (10) is of full row rank by (8). Hence we can find subsets
Fk Jk and Sk D Jk n Fk such that, denoting nk mk by DOFk
jFk j D mk ; jSk j D DOFk ;
and the mk mk sub-matrix of Jk defined by
Gk D .columns of Jk corresponding to Fk /; (11)
is nonsingular at the current consistent point. (Gk is the same as GŒ in MS, where
D 1 k.) Then by the IFT the mk items forming the vector xFk can locally be
found as functions of the remaining DOF k items, which form xSk .
DOF k is the number of DOF introduced at solution stage k. E.g, for the pendulum,
nk mk D 2 1 D 1 for k D 2 and k D 1. That is, one DOF is introduced at
each of these stages: an arbitrary position on the circle, and an arbitrary velocity.
The subvector xFk of xJk comprises the items found at this stage as functions of
the items in xSk . We call the latter state items. We call the sequence of Fk , which
defines all the found items, a solving scheme.
The state vector is the vector of all state items, that is
S
xS ; where S D kd k1 Sk : (12)
Note we stop at k D 1 because S0 is empty. The components of xS are the state

variables of the implicit ODE, to which this solving scheme reduces the DAE. All
the remaining items are “found” and form the found vector
S
xF ; where F D kd k0 Fk : (13)
Since each vector xJk is the concatenation of xFk and xSk , the SSS in the form (10),
combined with the IFT used in a solving scheme, becomes
for k D kd ; kd C1; : : : ; 0 do
(14)
solve fIk .t; xS<k ; xF<k I xSk ; xFk / D 0 for xFk as a function of xSk and xF<k .
This set of equations has a block-triangular form (BTF) such that, by repeated use
of the IFT, and repeated substitution, one has
Theorem 1 Each found item can be expressed as a function of the state vector xS .
Proof We show this for the case kd D 2, from which the general proof is clear.
For kD2, xS<k and xF<k are empty vectors, so we find xF2 as a function of xS2 .
For k D 1, we find xF1 as a function of xS1 , xS2 and xF2 , which by substituting
xF2 becomes a function of xS1 . For k D 0, we find xF0 as a function of xS0 , xS1 ,
xS2 , xF1 and xF2 , which by substituting xF1 and xF2 becomes a function of
xS0 .
Since xS0 equals xS , of which xS1 and xS2 are subvectors, the result follows.
Numerically, this involves one root-finding at each k-stage, taking given input
values of xS and using the nonsingular matrices Gkd ; Gkd C1 ; : : : ; G0 D J in turn.
Not all solving schemes are useful in practice. As so far described, for the
pendulum one can choose state items y at stage 2 and xP at stage 1, which does
not lead to an implicit ODE. The extra rule needed to make a useful scheme is
Definition 1 A dummy derivative scheme (DD scheme) is a solving scheme for
which, if an item is a derivative of a found item, then it is also a found item. Equival-
ently, the projections FO k Df j j . j; l/ 2 Fk g, of Fk on the j component, increase
with k.
For instance, in the preceding paragraph, choosing y and xP as state items violates
this constraint, since x must be a found item while its derivative xP is not.
Some derivative of each xj is a found item, since stage k D 0 is an n n system
.d /
that solves for each leading derivative xj j . Hence there is a unique least integer
.ı /
ıj 2 0 W dj such that xj j is a found item. The vector ı D .ı1 ; : : : ; ın / is the DD-
specification vector, or DDspec, of a DD scheme. Then
.l/
– An xj with l < ıj is a state item.
.l/
– An item xj with l > ıj is necessarily a found item and the derivative of a found
item. It is a dummy derivative—this accords with the MS definition.
.ı /
– The remaining n items xj j are loop-closers if ıj > 0, as they create the relations
in the reduced DAE that make an ODE, and algebraic items if ıj D 0 (see below).
The total number of state items equals DOF, and there are ıj state items
P for each j.
Hence any DDspec vector satisfies 0 ı d (elementwise), and j ıj D DOF .
Not every ı with these properties defines a DD scheme. E.g., for the pendulum,
ı D .2; 0; 0/, .0; 2; 0/ and .1; 1; 0/ obey these constraints but only the first two
“work”.
Figure 3 shows the process pictorially. On the left is the general scheme for the
case kd D 2. Each set of equations fIk D 0 has its output (what is solved for) above
it, and downward lines from it to the previously computed data it uses as input. On
the right is the more detailed data for the particular case of Example 4 below. These
are created by MATLAB functions, in the second case using structural data produced
by DAESA’s function daeSA from the MATLAB code for the DAE in Example 4.
The diagram uses natural notation for derivatives, e.g. x02 (MATLAB graphics is
poor at dots) instead of the x21 notation used in Example 4.
Fig. 3 DDs process picture. Left: general scheme, kd D 2. Right: specific picture for Example 4
Scheme (14), whether done conceptually or numerically, splits the xj into

“differential variables” having ıj > 0, and “algebraic variables” having ıj D 0.
E.g., if there are variables u; v; w whose ı’s are 2; 1; 0, then u; v are differential and
w is algebraic. Two loop-closer equations create a 2nd-order ODE uR D U.u; uP ; v/,
vP D V.u; uP ; v/. After one solves the ODE to get u; uP ; v as functions of t, the relation
wDW.u; uP ; v/ gives w. As another example, the DAE in (2) has empty differential
part.
Depending on DAE structure, one might find higher derivatives along the way,
e.g. if doing DDs for the pendulum with state variables x; xP , one cannot find the
algebraic variables y and without also finding yP and yR at the same time.
4 Preparing for Numerical Solution
Converting to First Order Implicit ODE How to reduce the result of DDs to
first order seems not well explained in the literature. Online Modelica tutorials
commonly order-reduce before SA and DDs: clumsy and giving a larger final system
than necessary. In fact this task fits almost trivially into DDs, in the following steps.
In Step 1, write down the SSS equations (6) up to stage 0—this is independent of
any DD scheme. We use xjl notation to show the itemsP count as unrelated variables.
P
There are DOF more variables than equations, ND.nC j dj / versus MD.nC i ci /.
In Step 2, make the system “square” by adding DOF new equations
xP jl D xj;.lC1/ ; (15)
one for each state variable xjl in xS , to say what its derivative “really is”. If xj;.lC1/ is
a state item, this implements order-reduction; otherwise it is a loop-closer.
In Step 3, rearrange the variables with the state vector xS followed by the found
vector xF , see (12) and (13), to get the result of DDs in the form
xP S D E.t; xS ; xF /; 0 D F.t; xS ; xF /: (16)
Function E is very simple: its components are all the xj;.lC1/ ’s of (15); it does not
really depend on t. Function F is actually fI0 . Write this as the concatenation of fIk
listed k D kd ; kd C1; : : : ; 0, and xF as the concatenation of xFk in the same order.
Then @F=@xF is block lower triangular with the nonsingular matrices Gk on its
block diagonal, so it is nonsingular. Hence the two equations (16) form a semi-
explicit index-1 DAE, see [1, p. 34]. From this easily follows:
Theorem 2 System (16) is SA-friendly with offsets ci all 0, so it is an implicit ODE.
Example 4 MS use the following linear, constant coefficient DAE [2, Example 1]:
MSEXAMPLE1 x1 x2 x3 x4 ci
A 20 32 xR 1 xR 2 xR 3 xP 4
0ı
0 D A D x1 Cx2 Cu1 .t/ AR2 1 1 0 03
0 D B D x1 Cx2 Cx3 Cu2 .t/ B6 0 0 0ı 7 2 R 1
˙D C 4 0 JD B6 1 1 0 7:
; 1 0ı 5 1 4 5
0 D CD x1 CPx3 Cx4 Cu3 .t/ CP 0 0 1 1
D 2ı 2 2 1 0
0 D DD 2Rx1 CRx2 CRx3 CPx4 Cu4 .t/ D 2 1 1 1
dj 2 2 2 1
The ui .t/ are given forcing functions. The two HVTs of ˙ are shown and ı . We
have c D .2; 2; 1; 0/ and d D .2; 2; 2; 1/, so J D @.A;R B;
R C;P D/[email protected] ; xR 2 ; xR 3 ; xP 4 /. It
is nonsingular, so SA succeeds and there are 2 DOF.
Step 1, using the xil notation, writes down M D 9 equations in N D 11 variables,
grouped here by stages k D 2; 1; 0:

0 D A D x10 Cx20 Cu1 .t/
(17)
0 D B D x10 Cx20 Cx30 Cu2 .t/
9
0 D AP D x11 Cx21 CPu1 .t/ =
0 D BP D x11 Cx21 Cx31 CPu2 .t/ (18)
;
0 D C D x10 Cx31 Cx40 Cu3 .t/
9
0 D AR D x12 Cx22 CRu1 .t/ >>
=
0 D BR D x12 Cx22 Cx32 CRu2 .t/
(19)
0 D CP D x11 Cx32 Cx41 CPu3 .t/ >>
;
0 D D D 2x12 Cx22 Cx32 Cx41 Cu4 .t/
There are two possible DDspecs, ı D .2; 0; 0; 0/ and .0; 2; 0; 0/. Following MS,
we choose the latter, so the state variables are x20 and x21 . Step 2 now defines two
equations; the upper one does order reduction and the lower is a loop-closer:

xP x
xP S D 20 D 21 D E.t; xS ; xF /: (20)
xP 21 x22
Step 3 splits the variables into xS D .x20 I x21 /T and xF D .x10 ; x30 I x11 ; x31 ; x40 I
x12 ; x22 ; x32 ; x41 /T , where the semicolons group into stages k D 2; 1 for xS and
2; 1; 0 for xF . Then (17), (18), and (19) jointly define the equations
0 D F.t; xS ; xF /; (21)
and (20), (21) give the desired semi-explicit index 1 form.

The overall Jacobian is below, where a blank means h 1 an all-zero
i block of the
00
appropriate size, and G0 D J with sub-matrices G1 D 1 1 0 , G2 D 11 01 .
011
xS xF2 xF1 xF0

Œx20 ; x21 Œx10 ; x30 Œx11 ; x31 ; x40 Œx12 ; x22 ; x32 ; x41
" #2 " # " # 3
x21 01 0000
ED 6 7
"x22 #6
6 "0 0# 0100 7
7
A 66 10 7
7
fI2 D 6 G2 7
B 6 1 0 7
2 36 2 3 2 3 7
@.E; F/ AP 6 0 1 00 7
D 6 7
: (22)
@.xS ; xF / 6 P 76 6 7 6 7 7
fI1 D 4 B 56 40 15 40 05 G1 7
6 7
C 6 00 10 7
2 36 6 2 3 7
7
AR 6 000 7
6 76 6 7 7
6 BR 76 60 0 07 7
fI0 D 6 76 6 7 7
6 P 76 6 7 G0 7
4C 54 41 0 05 5
D 000
This formulation was confirmed by giving it to the MATLAB implicit solver ode15i,
for certain ui .t/ and initial values, and comparing with the analytic solution.
Converting to explicit ODE Conceptually it is obvious from the nonsingularity of
@F=@xF in the DAE (16) that near a consistent point we can convert it to an explicit
ODE yP D f.t; y/—solve the second equation for xF and substitute in the first. It may
seem strange to do this numerically, in view of the impressive track record of codes
such as DASSL [1], for solving DAEs of the form G.t; z; zP / D 0.
But there are reasons for doing so. Some applications generate models with n in
the thousands and only a few DOF, and for which stiffness is not a problem. Then
it makes sense to convert to a small explicit ODE and solve by, say, a Runge–Kutta
code, using far less working memory than would a DAE code. The ESI-CyDesign
Modelica system does numerical solution this way, having found it more efficient
for their typical models. The Numerical Algorithms Group (NAG) Ltd have recently
put a reverse communication RK code into their library, for just such uses.
Also, ways to compute the Jacobians Gk for an SA-friendly DAE are well known,
while the full Jacobian of the DASSL-style formulation—in the example, the below-
diagonal blocks of (21)—is messier to find; but see the discussion in Sect. 5.
Further (Andreas Rauh, Rostock, personal communication 2015), some calcula-

tions of robust design optimisation in control theory are far easier if the governing
DAE (which is SA-friendly) has been converted to an explicit ODE.
5 Discussion
We have shown the dummy derivatives method is a particular case of the ˙-

method’s standard solution scheme, and displayed the underlying BTF pictorially.
This gives a simple method to cast the DD’ed system as a semi-explicit index 1
DAE, hence an implicit ODE, combining order reduction and index reduction in
one process. We pointed out that there may be compelling reasons for the further
step of converting to an explicit ODE before numerical solution.
Step 3 of the method in Sect. 4 is optional—being an implicit ODE is independent
of how the equations and variables are arranged. Thus different DD schemes for
a DAE are distinguished purely by their sets of added equations (15). This set is
deducible immediately from the DDspec vector ı. All this, except the vector ı, is
implicit in Mattsson and Söderlind’s paper, but the simplicity of the algorithm—
hence of programming it—does not seem to be known in the literature.
A Jacobian such as (21) is used for Newton iterations inside a DASSL-style
solver. In this context it seems one can ignore its “messy to compute” below-
diagonal blocks, i.e. set them to 0, assuming a sufficiently good initial guess.
Namely, as G2 is correct it will make the xF2 component converge quadratically.
Once this has happened, as G1 is correct it will now make the xF1 component
converge quadratically; and so on. In the context of solving Example 4 by ode151
we have checked that this (surely well known) fact is true, but we have done no
experiments to validate it on harder, nonlinear, problems.
References
Differential-Algebraic Equations, 2nd edn. SIAM, Philadelphia (1996)
3. McKenzie, R.: Structural analysis based dummy derivative selection for differential-algebraic
equations. Technical report, Cardiff University (2015). Submitted to BIT Numerical Analysis
4. Pantelides, C.C.: The consistent initialization of differential-algebraic systems. SIAM. J. Sci.
Stat. Comput. 9, 213–231 (1988)
(2001)
6. Pryce, J.D.: A simple approach to Dummy Derivatives for DAEs. Technical report, Cardiff
University, July 2015. In preparation
New Master-Slave Synchronization
Criteria of Chaotic Lur’e Systems
with Time-Varying-Delay Feedback Control
Kaibo Shi, Xinzhi Liu, Hong Zhu, and Shouming Zhong
Abstract This study focuses on the issue of designing a time-delay output feedback
controller for master-slave synchronization of chaotic Lur’e systems (CLSs). The
time delay is assumed to be a time-varying continuous function which is bounded
below and above by positive constants. By constructing an appropriate Lyapunov-
Krasovskii functional (LKF), a novel delay-dependent synchronization condition is
obtained. Besides, by employing a new free-matrix-based inequality (FMBI), the
desired controller gain matrix can be achieved by solving a set of linear matrix
inequalities (LMIs). Finally, one numerical example of Chua’s circle is given to
illustrate the effectiveness and advantages of the proposed results.
1 Introduction
During the past few decades, the chaotic synchronization problem has attracted
increasing attention due to its extensive applications in many fields including
secure communication, physical, chemical and ecological systems, human heartbeat
regulation, and so on [1–4]. As is well-known that many nonlinear systems can be
modeled precisely in the form of Lur’e systems, such as Chua’s circuit, network
systems and hyper chaotic attractors, which include a feedback connection of
K. Shi () • H. Zhu

School of Automation Engineering, University of Electronic Science and Technology of China,
Chengdu, Sichuan 611731, P.R. China
X. Liu
S. Zhong
School of Mathematical Sciences, University of Electronic Science and Technology of China,
Chengdu, 611731, P.R. China

726 K. Shi et al.
a linear system and a nonlinear element satisfying the sector condition [5–7].
Therefore, master-slave synchronization for CLSs has been a focused research topic.
Recently, the synchronization problem of chaotic systems with time delay has
been intensively investigated due to the unavoidable signal propagation delay
frequently encountered in remote master-slave synchronization scheme [8–10].
Especially, many delay-independent and delay-dependent synchronization criteria
have been derived by constructing an appropriate LKF in [11]. Compared with the
results in [11], less conservative synchronization criteria were obtained and some
fairly simple algebraic conditions are derived for easier verification in [12]. Based
on Lyapunov method and LMIs approach, the authors in [13] further generalized
and improved the proposed results in [11, 12]. However, in order to obtain sufficient
conditions for master-slave synchronization, the authors in [11, 13] employed model
transformation, which leads to some conservatism for inducing additional terms.
Based on the free weighting matrix approach, a delay-dependent synchronization
condition is obtained in [14]. By using a delay-partition approach, several delay-
dependent synchronization criteria are derived in the form of LMIs in [15, 16]. By
constructing an appropriate LKF including the information of time-varying delay
range, new delay-range-dependent synchronization criteria for Lur’e systems are
established in [17]. Besides, it should be noted that only constant delay is considered
in [11–16, 18]. In practice, as everyone knows that the range of time-varying delay
non-zero lower bound is often encountered, and such systems are referred to as
interval delayed systems.
Motivated by the issues discussed above, the delay-dependent master-slave
synchronization problem of CLSs with time-varying-delay feedback control is
investigated in this paper. By taking full advantage of the information of time-
varying-delay range and nonlinear term of the error system, a less conservative
delay-dependent synchronization criterion is obtained. In addition, based on an
appropriate LKF combined with a new FMBI, an explicit expression of the desired
control law can be obtained in terms of LMIs. Finally, one numerical example of
Chua’s circle is presented to demonstrate the effectiveness and advantages of the
design methods.
Notation Notations used in this paper are fairly standard: Rn denotes the n-
dimensional Euclidean space, Rnm the set of all n m dimensional matrices;
I the identity matrix of appropriate dimensions, AT the matrix transposition of
the matrix A. By X > 0 (respectively X 0), for X 2 Rnn , we mean
that the matrix X is real symmetric positive definite (respectively, positive semi-
definite); diagfr1 ; r2 ; ; rn g denotes block diagonal matrix with diagonal elements
ri ; i D 1; ; n, the symbol represents the elements below the main diagonal of a
symmetric matrix, SymfMg is defined as SymfMg D 12 .M C MT /.
New Master-Slave Synchronization Criteria of Chaotic Lur’e Systems with. . . 727
2 Preliminaries
First, consider the following master-slave synchronization scheme of two identical

chaotic Lur’e systems with time-varying delay feedback control:
(
xP .t/ D Ax.t/ C B'.Dx.t//;
M W (1)
p.t/ D Hx.t/;
(
yP .t/ D Ay.t/ C B'.Dy.t// C u.t/;
S W (2)
q.t/ D Hy.t/;
C W u.t/ D K.p.t d.t// q.t d.t///; (3)
which consists of master system M , slave system S and controller C . M and S

with u.t/ D 0 are identical chaotic time-delay Lur’e systems with state vectors x.t/,
y.t/ 2 Rn , outputs of subsystems p.t/, q.t/ 2 Rl , respectively, u.t/ 2 Rn is the
slave system control input, and A 2 Rnn , B 2 Rnnd , D 2 Rnd n , H 2 Rln are
known real matrices, K 2 Rnl is the time-varying delay controller gain matrix to
be designed. It is assumed that './ is the nonlinear function in the feedback path.
Assumption A Time-varying delay d.t/ is differential function and satisfies the
following condition:
0 < dL d.t/ dU ; P < 1:

d.t/ (4)
where dL , dU and are three real constants.

Assumption B The nonlinear function 's .˛/ satisfies the following sector condi-
tion: .s D 1; ; nd /
's .˛/ 2 KŒk ;ksC D f's .˛/ j 's .0/ D 0; ks ˛ 2 ˛'s .˛/ ksC ˛ 2 ; ˛ ¤ 0g (5)
s
Next, given the synchronization schemes (1), (2) and (3), the synchronization
error is defined as r.t/ D x.t/ y.t/, and we can get the following synchronization
error system:
rP .t/ D Ar.t/ C B.Dr.t/; y.t// KHr.t d.t//; (6)
where .Dr.t/; y.t// D '.Dr.t/ C Dy.t// '.Dy.t//. Let D D Œd1 ; ; dnd T with
ds 2 Rn .s D 1; 2; ; nd /. Under Assumption B, it is easy to obtain that s .dTs r; y/
satisfies the following condition
s .dTs r; y/ 's .dTs .r C y// 's .dTs y/

ks D ksC ; 8r; y 2 Rn ; dTs r ¤ 0: (7)
dTs r dTs r
728 K. Shi et al.
Lemma 1 ([19]) For differentiable signal r.t/ in Œa; b ! Rn , for symmetric

matrices R 2 Rnn , and Z1 2 R3n3n , Z3 2 R3n3n , and any matrices Z2 2 R3n3n ,
2 3 2 3 2 3
N11 N21 Z1 Z2 N1
and N1 D 4 N12 5, N2 D 4 N22 5 2 R3nn satisfying R N D 4 Z3 N2 5 > 0, the
N13 N23 R
following inequality holds:
Z b
rP T .s/RPr.s/ds $ T Œ˝0 C SymfN1 ˙1 C N2 ˙2 g$;
a
where ˝0 D .b a/.Z1 C 13 Z3 /, ˙1 D ŒI; I; 0, ˙2 D ŒI; I; 2I,

1
Rb T
$ D ŒrT .a/; rT .b/; ba a r .s/ds .
T
Remark 1 It should be noted that the set of slack variables in the above inequality
can provide great freedom in deriving less conservative stability conditions, which it
is possible to obtain a much sharper bound. It is easy to prove that some well-known
integral inequalities are special cases of this one in [19]. For example, if we let N1 D
ŒYT ; 0T , N2 D 0, Z1 D diagfX; 0g, Z2 D 0 and Z3 D 0, this integral inequality
1
can reduce to the integral inequality in [20]. And if we let N1 D ba ŒR; R; 0T ,
3 T 1 T 1 T 1
N2 D ba ŒR; R; 2R , Z1 D N1 R N1 , Z2 D N1 R N2 , Z3 D N2 R N2 , this
T
integral inequality also becomes the celebrated Wirtinger integral inequality [21].
3 Main Results
Theorem 1 Under Assumptions A and B, for given scalars , dL , dU , the error

system (6) is globally asymptotically stable if there exist positive matrices P, Ri
.i D 1; ; 5/, any positive definite diagonal matrices G D diagfg1; ; gnd g,
L D diagfl1 ; ; lnd g, W D diagfw1; ; hnd g, symmetrical matrices X1 , Y1 ,
Z1 , X3 , Y3 and Z3 , and any matrices X2 , Y2 , Z2 , Ni .i D 1; 2; ; 6/, N and U
with the appropriate dimensions, such that the following symmetric linear matrix
inequalities hold:
2 3 2 3 2 3
X1 X2 N1 Y1 Y2 N3 Z1 Z2 N5
4 X3 N2 5 > 0; 4 Y3 N4 5 > 0 and 4 Z3 N6 5 > 0; (8)
R4 R5 R5
< 0; (9)
where
D e1 .R1 C R3 DK WKC DT /eT1 C e2 .dL R4 C dUL R5 /eT2 e3 .R1 R2 /eT3

e4 R2 eT4 .1 /e5 R3 eT5 2e6 WeT6 C Symfe1 Pe2 C e1 D.K C KC /eT6
C e2 DT .G L/eT6 C e1 DT ŒKC L K GDeT2 C Œe1 x C e2 yNŒeT2 C AeT1 C BeT6

1
KHeT4 g C Œe1 ; e3 ; e7 dL X1 C X3 C SymfM1 ˘1 C M2 ˘2 g Œe1 ; e3 ; e7 T
3

1
C Œe3 ; e4 ; e8 dUL Y1 C Y3 C SymfM3 ˘1 C M4 ˘2 g Œe3 ; e4 ; e8 T
3

1
C Œe4 ; e5 ; e9 dUL Z1 C Z3 C SymfM5 ˘1 C M6 ˘2 g Œe4 ; e5 ; e9 T ;
3
Moreover, the gain matrix of state estimator is given by K D N1 U.

Proof Consider the following a newly augmented LKF for the error system (6):
V.rt / D V1 .rt / C V2 .rt / C V3 .rt /; (10)
where
Z t
V1 .rt / D r .t/Pr.t/ C
T
rT .s/R1 r.s/ds
thL
Z tdL Z t
C r .s/R2 r.s/ds C
T
rT .s/R3 r.s/ds; (11)
tdU td.t/
Z 0 Z t Z dL Z t
V2 .rt / D rP T .s/R4 rP .s/dsd C rP T .s/R5 rP .s/dsd; (12)
dL tC dU tC
X
nd Z dsT r.t/ X
nd Z dsT r.t/
V3 .rt / D 2 gs Œs ./ ks d C2 ls ŒksC s ./d;
sD1 0 sD1 0
(13)
Taking the derivative of V.xt / along the trajectory of the error system (6) yields:
P 1 .rt / D 2rT .t/PPr.t/ C rT .t/.R1 C R3 /r.t/ rT .t dL /.R1 R2 /r.t dL /

V
rT .t dU /R2 r.t dU / .1 d.t//r
P T
.t d.t//R3 r.t d.t//
2rT .t/PPr.t/ C rT .t/.R1 C R3 /r.t/ rT .t dL /.R1 R2 /r.t dL /
rT .t dU /R2 r.t dU / .1 /rT .t d.t//R3 r.t d.t//; (14)
730 K. Shi et al.
P 2 .rt / DPrT .t/.dL R4 C dUL R5 /Pr.t/

V
Z t Z tdL
rP .s/R4 rP .s/ds
T
rP T .s/R5 rP .s/ds
tdL tdU
Z t
DPr .t/.dL R4 C dUL R5 /Pr.t/
T
rP T .s/R4 rP .s/ds
tdL
Z tdL Z td.t/
rP T .s/R5 rP .s/ds rP T .s/R5 rP .s/ds; (15)
td.t/ tdU
By using Lemma 1, we can have

Z t
rP T .s/R4 rP .s/ds
tdL

1
$1T .t/ dL X1 C X3 C SymfM1 ˘1 C M2 ˘2 g $1 .t/;
3
(16)
Z tdL
rP T .s/R5 rP .s/ds
td.t/

1
$2T .t/ dUL Y1 C Y3 C SymfM3 ˘1 C M4 ˘2 g $2 .t/;
3
(17)
Z td.t/
rP T .s/R5 rP .s/ds
tdU

1
$3T .t/ dUL Z1 C Z3 C SymfM5 ˘1 C M6 ˘2 g $3 .t/;
3
(18)
Rt
where dUL D dU dL , $1T .t/ D ŒrT .t/; rT .t dL /; d1L tdL rT .s/ds, $2T .t/ D
1
Rt
ŒrT .t dL /; rT .t d.t//; d.t/d tdL r .s/ds,˘1 D ŒI; I; 0 $3 .t/ D Œr .t
T T T
1
R t
L
d.t//; rT .t dU /; dU d.t/ tdL r .s/ds, ˘2 D ŒI; I; 2I.
T
X
nd X
nd
P 3 .rt / D2
V P s .dT r.t// C 2
.gs ls /dsT r.t/ P s kC gs k /dT r.t/
dsT r.t/.l
s s s s
sD1 sD1
D2Pr .t/D .G L/.Dr.t// C 2r .t/DT ŒKC L K GDPr.t/;

T T T
(19)
Next, for any scalars x and y, and arbitrary matrix N with appropriate dimensions,
we can obtain
0 D2ŒPrT .t/x C rT .t/yNŒPr.t/ C Ar.t/ C B.Dr.t/; y.t// KHr.t d.t//;

(20)
From Eq. (7), for any positive diagonal matrix W D diagfw1 ; ; wnd g, it yields
that
X
nd
2 Œs .dTs r.t/; y.t// ks dTs r.t/ws Œs .dTs r.t/; y.t// ksC dTs r.t/
sD1
D 2rT .t/DK WKC DC r.t/ C 2rT .t/D.K C KC /W.DT r.t/; y.t//

2 T .DT r.t/; y.t//W.DT r.t/; y.t//; (21)
Combining Eqs. (14), (15), (16), (17), (18), (19), (20), and (21) yields
V.rt / T .t/ .t/; (22)
where T .t/ D ŒrT .t/; rP T .t/; rT .t dL /; rT .t d.t//; rT .t dU /; T .DT r.t/; y.t//;

$1T .t/; $2T .t/; $3T .t/,
From (9), we can have V.t; rt / " k r.t/ k2 holds for any sufficiently small
" > 0.
Therefore, this implies the error system (6) is globally asymptotically stable. The
proof is completed. t
u
4 Numerical Example
In this section, a numerical simulation example is given to show the effectiveness

and correctness of the main results derived above.
Consider Chua’s circuit example and the system equation in [14, 16–18] is given
by
8
ˆ
ˆ xP 1 .t/ D˛.x2 .t/ h.x1 .t///;
<
xP 2 .t/ Dx1 .t/ x2 .t/ C x3 .t/; (23)
ˆ
:̂
xP 3 .t/ D ˇx2 .t/;
732 K. Shi et al.
with nonlinear characteristic
h.x1 .t// D m1 x1 .t/ C 0:5.m0 m1 /Œjx1 .t/ C cj jx1 .t/ cj;
and parameters m0 D 17 , m1 D 27 , ˛ D 9, ˇ D 14:28, and c D 1. The system can

be written in chaotic Lur’e system framework (1) with
23 2 3 2 3
˛m1 1 ˛ 0 ˛.m0 m1 / 1
AD4 1 2 1 5 ; B D 4 0 5 ; D D H D 4 05:
0 ˇ 1 0 0
Under the above conditions and D 0, the maximum upper bounds on the
allowable delays of h obtained from the above references [14, 16–18] and Theorem 1
are listed in Table 1. Thus, it is also to see that our result is more effective than the
recently reported ones.
In addition, set '.˛/ D 0:5.j˛C1jj˛1j/, h D 0:191, x.0/ D Œ0:2; 0:3; 0:2
and y.0/ D Œ0:5; 0:1; 0:6, the simulation results are shown in Figs. 1, 2, 3, 4, 5,
and 6 for the above gain matrix. The trajectories of the master-slave systems with
Table 1 Maximum allowed delays h and the best gain matrices K for D 0
Method [18] [14] [17] [16] Theorem 1
h 0.141 0.180 0.183 0.185 0.188
2 3 2 3 2 3 2 3 2 3
6:0229 3:9125 4:1455 4:0779 4:6512
6 7 6 7 6 7 6 7 6 7
K 4 1:3367 5 4 0:9545 5 4 0:9250 5 4 0:9087 5 4 0:5992 5
2:1264 3:8273 4:2596 4:3430 4:5177
0.4
0.2
x 2 (t)
−0.2
−0.4
4
2 3
2
0 1
0
x (t) −2 −1
3
−2 x (t)
−4 −3 1
Fig. 1 State trajectories of master system x.t/

0.4
0.2
y 2 (t)
−0.2
−0.4
4
2 3
2
0 1
y (t) 0
3 −2 −1
−2 y1(t)
−4 −3
Fig. 2 State trajectories of slave system y.t/ without u.t/
4
x (t) x (t) x (t) y (t) y (t) y (t)
1 2 3 1 2 3
1
Amplitude
−1
−2
−3
0 1 2 3 4 5 6 7 8 9 10
t
Fig. 3 State trajectories of master-slave systems x.t/ and y.t/ with u.t/
u.t/ D 0 are shown in Figs. 1 and 2. Under the above gain matrix K, the responses
of the state x.t/ and y.t/, the error signal r.t/, outputs of subsystems p.t/ and q.t/
are represented in Figs. 3, 4, 5, and 6, respectively. Therefore, it is clear to show that
the synchronization error is tending asymptotically to zero.
734 K. Shi et al.
r (t) r (t) r (t)

1.5 1 2 3
0.5
Amplitude
−0.5
−1
−1.5
−2
0 1 2 3 4 5 6 7 8 9 10
t
Fig. 4 State trajectories of synchronization error system r.t/
2
p(t)
1.5
0.5
Amplitude
−0.5
−1
−1.5
−2
−2.5
0 1 2 3 4 5 6 7 8 9 10
t
Fig. 5 State trajectories of output vector p.t/

2
q(t)
1.5
0.5
0
Amplitude
−0.5
−1
−1.5
−2
−2.5
0 1 2 3 4 5 6 7 8 9 10
t
Fig. 6 State trajectories of output vector q.t/
5 Conclusions
This paper consider the delay-dependent master-slave synchronization problem of

CLSs with time-varying-delay feedback control. First, by choosing an augmented
LKF, a novel delay-dependent synchronization criterion is derived. Second, by
employing a new FMBI, which has been proved to be lesser conservative than the
existing integral inequalities and showed to have a great potential efficient in the
actual operation, the desired controller gain matrix can be achieved by solving a set
of LMIs. Finally, one numerical example of Chua’s circle is given to illustrate the
effectiveness and advantages of the proposed results. The foregoing methods have
the potential to be useful for the further study of CLSs. Meanwhile, it is expected
that these approaches can be further used for other time-delay systems.
Acknowledgements This work was supported by National Basic Research Program of China
(2010CB732501), National Natural Science Foundation of China (61273015), The National
Defense Pre-Research Foundation of China (Grant No. 9140A27040213DZ02001), The Program
for New Century Excellent Talents in University (NCET-10-0097).
References
1. Liao, T.L., Tsai, S.H.: Adaptive synchronization of chaotic systems and its application to secure
communications. Chaos, Solitons Fractals 11, 1387–1396 (2000)
2. Lu, J.N., Wu, X.Q., LRu, J.H.: Synchronization of a unified chaotic system and the application
in secure communication. Phys. Lett. A 305, 365–370 (2002)
736 K. Shi et al.
3. Grzybowski, J.M.V., Rafikov, M., Balthazar, J.M.: Synchronization of the unified chaotic
system and application in secure communication. Commun. Nonlinear Sci. Numer. Simul. 14,
2793–2806 (2009)
4. Voss, H.: Anticipating chaotic synchronization. Phys. Rev. E 61, 5115–5119 (2000)
5. Mkaouar, H., Boubaker, O.: Chaos synchronization for master slave piecewise linear systems:
application to Chua’s circuit. Commun. Nonlinear Sci. Numer. Simul. 17, 1292–1302 (2012)
6. GKamez-GuzmKan, L., Cruz-HernKa ndez, C., LKopez-GutiKerrez, R.M., GarcKia-Guerrero, E.E.:
Synchronization of Chua’s circuits with multi-scroll attractors: application to communication.
Commun. Nonlinear Sci. Numer. Simul. 14, 2765–2775 (2009)
7. LRu, J.H., Murali, K., Sinha, S., Leung, H., Aziz-Alaoui, M.A.: Generating multi-scroll chaotic
attractors by thresholding. Phys. Lett. A 372, 3234–3239 (2008)
8. Stepp, N.: Anticipation in feedback-delayed manual tracking of a chaotic oscillator. Exp. Brain
Res. 198, 521–525 (2009)
9. Pyragas, K., Pyragas, T.: Extending anticipation horizon of chaos synchronization schemes
with time-delay coupling. Philos. Trans. R. Soc. A 368, 305–317 (2010)
10. Milton, J.G.: The delayed and noisy nervous system: implications for neural control. J. Neural
Eng. 8, 065005 (2011)
11. Yalcin, M.E., Suykens, J.A.K., Vandewalle, J.: Master-slave synchronization of Lur’e systems
with time-delay. Int. J. Bifurc. Chaos 11(6), 1707–1722 (2001)
12. Liao, X.X., Chen, G.R.: Chaos synchronization of general Lur’e systems via time-delay
feedback control. Int. J. Bifurc. Chaos 13(1), 207–213 (2003)
13. Cao, J.D., Li, H.X., Daniel, W.C.H.: Synchronization criteria of Lur’e systems with time-delay
feedback contro. Chaos, Solitons Fractals 23, 1285–1298 (2005)
14. He, Y., Wen, G.L., Wang, Q.G.: Delay-dependent synchronization criterion for Lur’e systems
with delay feedback control. Int. J. Bifurc. Chaos 16(10), 3087–3091 (2006)
15. Ding, K., Han, Q.L.: Master-slave synchronization criteria for horizontal platform systems
using time delay feedback control. J. Sound Vib. 330(11), 2419–2436 (2011)
16. Ge, C., Hua, C.C., Guan, X.P.: Master-slave synchronization criteria of Lur’e systems with
time-delay feedback control. Appl. Math. Comput. 244, 895–902 (2014)
17. Li, T., Yu, J.J., Wang, Z.: Delay-range-dependent synchronization criterion for Lur’e systems
with delay feedback control. Commun. Nonlinear Sci. Numer. Simul. 14(5), 1796–1803 (2009)
18. Han, Q.L.: New delay-dependent synchronization criteria for Lur’e systems using time delay
feedback control. Phys. Lett. A 360(4), 563–569 (2007)
19. Zeng, H.B., He, Y., Wu, M., She, J.H.: Stability of time-delay systems via Wirtinger-based
double integral inequality. IEEE Trans. Autom. Control. 60(10), 2768–2772 (2015)
20. Zhang, X.M., Wu, M., She, J.H., He, Y.: Delay-dependent stabilization of linear systems with
time-varying state and input delays. Automatica 41, 1405–1412 (2005)
21. Seuret, A., Gouaisbaut, F.: Wirtinger-based integral inequality: application to time-delay
systems. Automatica 49, 2860–2866 (2013)
Robust Synchronization of Distributed-Delay
Systems via Hybrid Control
Peter Stechlinski and Xinzhi Liu
Abstract Drive and response systems which exhibit time-delays and uncertainties
are synchronized using hybrid control. Classes of dwell-time satisfying switching
rules are identified under which synchronization can be achieved in a robust manner.
The theoretical results are established using multiple Lyapunov functions and
Halanay-like inequalities.
1 Introduction
Synchronization problems arise in secure communications [1], multi-agent net-

works [2], chaotic systems [3], neural networks [4], and many other important
applications. Many real-world phenomena exhibit time-delays (e.g., see [5, 6])
and uncertainties in parameter measurements and data input [7], which can have
significant impacts on control strategies for synchronization. The main focus of
the present article is to investigate the synchronization of uncertain systems with
distributed delays and nonlinear perturbations. The approach here is to use a
combination of switching and impulsive control. This type of hybrid control, which
can have a number of advantages over continuous control (see [8–10] for details),
leads to a switched system model for the error between the drive and response
systems.
Switched systems, which are a type of hybrid system, are governed by a mixture
of continuous/discrete dynamics and logic-based switching, and have applications
in a wide-range of problems in applied mathematics, engineering, and computer
science [10]. In the switched systems literature, finding classes of switching rules
which guarantee a performance goal (e.g., stability, synchronization, etc.) is a major
area of research (e.g., see [8, 10–14]), where concepts such as multiple Lyapunov
functions and dwell-time satisfying switching rules are used. This line of research
has been extended to switched systems with time-delays composed of stable and
unstable modes by way of Halanay-like inequalities (e.g., see [15–20]). Robust
control (e.g., see [21, 22]) is important to take into account since relatively small
P. Stechlinski () • X. Liu

University of Waterloo, Waterloo, ON, N2L 3G1, Canada

738 P. Stechlinski and X. Liu
uncertainties can lead to instability (or, desynchronization in the present context),

and has been studied in the time-delay switched systems literature (e.g., see [7, 23–
25]).
Motivated by the above discussion, the main objective of this article is to
extend the literature by studying the robust hybrid synchronization of systems with
distributed delays. Contributions include establishing a switched and impulsive
Halanay-like inequality using multiple Lyapunov functions to prove synchroniza-
tion that is robust to uncertainties, and developing verifiable synchronization
conditions in the face of time-delays. Classes of dwell-time, average dwell-time,
and periodic switching rules are identified under which synchronization can be
achieved via hybrid control. Methods are developed in the present article to deal
with switching desynchronization (where impulsive control is crucial), nonlinear
perturbations, as well as impulsive disturbances. The results are generalizable to
stability studies of switched systems composed of a mixture of stable and unstable
modes. The theoretical findings are augmented with numerical simulations.
2 Problem Formulation
Let RC denote the set of non-negative real numbers, let N denote the set of positive
integers, and let Rn denote the Euclidean space of n-dimensions (equipped with the
Euclidean norm k k). For a positive constant , denote PC D PC.Œ ; 0; Rn / to be
the space of piecewise continuous functions mapping Œ ; 0 to Rn , equipped with
the norm k k WD sup s0 k .s/k. Let max ./ and min ./ denote the maximum
and minimum eigenvalues of a symmetric matrix, respectively.
In the spirit of Guan et al. [14], consider the following drive system:
Z 0
xP .t/ D .A C A.t//x.t/
Q C .C C C.t//
Q x.t C s/ds C F.t; xt /; (1)

where > 0 is an upper bound on the distribution of time-delays; xt 2 PC is

defined by xt .s/ WD x.t C s/ for s 0; F W R PC ! Rn satisfies
F.t; 0/ 0 for all t t0 ; A; C 2 Rnn are known; A; Q CQ are uncertainties of
the form A.t/ WD
A %A .t/A , C.t/ WD
C %C .t/C , where
A ,
C , A , C are
Q Q
known constant matrices of appropriate dimensions and %A , %C are time-dependent
uncertainty matrices.
The response system is given by
Z 0
yP .t/ D .A C A.t//y.t/
Q C .C C C.t//
Q y.t C s/ds C F.t; yt / C u; (2)

Robust Synchronization of Distributed-Delay Systems via Hybrid Control 739
where u WD u1 C u2 is a controller designed as follows:

1
X Z 0
u1 .t; x; y/ WD .Hik C Hik .t//e.t/ C .Gik C Gik .t//
Q Q e.t C s/ds 1I k .t/;
kD1
1
X
u2 .t; x; y/ WD .Eik C EQ ik .t//e.t/ı.t tk /;
kD1
where e WD y x is the error between the drive and response systems; Hi , Gi , Ei ,

i D 1; : : : ; m (for some positive integer m) are known constant control matrices,
H Q i , EQ i , i D 1; : : : ; m, are time-dependent uncertainties of the form H
Q i, G Q i .t/ WD
H;i %H;i .t/H;i , G Q i .t/ WD

G;i %G;i .t/G;i , EQ i .t/ WD
E;i %E;i .t/E;i , where
H;i ,
G;i ,
E;i , H;i , G;i , E;i are known constant matrices of appropriate dimensions and
%H;i , %G;i , %E;i are time-dependent uncertainty matrices; ik 2 P WD f1; : : : ; mg for
each k 2 N, where m is a positive integer; I k WD Œtk1 ; tk / is called the switching or
hybrid interval, 1I k ./ is the indicator function, and ı./ is the generalized Dirac
delta function. Under this construction, u1 and u2 are switching and impulsive
controllers containing some uncertainties, respectively (see, e.g., [14] for a similar
hybrid control construction without time-delays or uncertainties present).
The error system can be written as the following hybrid system:
Z 0
eP .t/ D A .t/e.t/ C C .t/ e.t C s/ds C F.t; yt / F.t; xt /; t ¤ tk ;

e.t/ D .I C E .t //e.t /; t D tk ;

et0 D 0 ; k D 1; 2; : : : ;
(3)
where t0 2 R is the initial time; 0 WD yt0 xt0 2 PC is the initial error; Ai .t/ WD
Q C Hi C H
A C A.t/ Q i .t/, Ci .t/ WD C C C.t/
Q C Gi C G Q i .t/, and Ei .t/ WD Ei C EQ i .t/ for
each i D 1; : : : ; m. The switching rule W Œt0 ; C1/ ! P is a piecewise constant
function that maps each switching interval Œtk1 ; tk / to an integer in P, i.e., it selects
the active mode of (3). Denote the set of all such switching rules by S .
Definition 1 The uncertainties are said to be admissible if they are continuous
matrix functions satisfying k%A .t/k < 1, k%C .t/k < 1, k%H;i .t/k < 1, k%G;i .t/k <
1, and k%E;i .t/k < 1 for all t t0 , i D 1; : : : ; m. The drive system (1) and the
response system (2) are said to be robustly synchronized (RS) if limt!1 ke.t/k D 0
for all 0 2 PC and admissible uncertainties. The systems are said to be robustly
exponentially synchronized (RES) if there exist > 0, C > 0 such that, for t t0 ,
ke.t/k Ck0 k expŒ.t t0 / for all 0 2 PC and admissible uncertainties.
Remark 1 If 2 S inf-dwell WD f 2 S W 9 > 0 s.t. infk2N ftk tk1 g g,
Fi is composite-PC in the sense of Ballinger and Liu [26], continuous in its first
variable, and locally Lipschitz in its second variable for each i D 1; : : : ; m, then
there exists a unique solution of (3) by [26] and using the method of steps over the
hybrid moments [27].
Given the hybrid control matrices f.Hi ; HQ i ; Gi ; G
Q i ; Ei ; EQ i / W i 2 Pg, the goal of
this article is to design the hybrid time mode sequences (i.e., the set f.ik ; tk / W k 2
Ng) such that the drive and response systems achieve robust synchronization.
2.1 Main Results
Consider the following notions of activation time and total number of switches.
Definition 2 Let t1 ; t2 2 R such that t2 t1 and let A P. Denote the total
R t2
activation time of the modes in A by TA .t1 ; t2 / WD t1 1A ..t//dt. Denote the total
number of switches to a mode in A by ˚A .t1 ; t2 / WD jftk W .tk / 2 A; t1 tk < t2 gj.
Zhu established the following Halanay-like lemma.
Lemma 1 ([18]) Let a; b > 0. Assume that u W Œt0 ; 1/ ! RC satisfies uP .t/
bkut k au.t/ for t t0 . If b a 0 then u.t/ kut0 k expŒ.b a/.t t0 /
for t t0 , while if b a < 0 then there exists a positive constant satisfying
C b exp. / a < 0 such that u.t/ kut0 k expŒ.t t0 / for t t0 :
For use in the main theorem, the following Halanay-like result is proved.
Proposition 1 Let ai ; bi ; di ; hi 0 for i D 1; : : : ; m. Assume that u W Œt0
; C1/ ! RC satisfies
(
uP .t/ b kut k a u.t/; t ¤ tk ; t t0 ;
(4)
u.t/ d u.t / C h kut k ; t D tk ; k D 1; 2; : : : ;
for some 2 S which satisfies tk tk1 for all k 2 N. Then, for t t0 ,

0 1 8 9
˚P .t0 ;t/
Y <X X =
u.t/ kut0 k @ ıij A exp i Tfig .t0 ; t/ i TQ fig .t0 ; t/ ; (5)
: ;
jD1 i2P u i2P s
where TQ fig .t0 ; t/ WD Tfig .t0 ; t/ ˚fig .t0 ; t/ , i WD bi maxi2P f1=ıi ; 1g ai , ıi WD

di Chi e , WD maxi2P s fi g, i > 0 is chosen for i 2 P s so that i Cbi ei ai < 0,
P u WD fi 2 P W i 0g, P s WD fi 2 P W i < 0g, and i > 0 is chosen for
i 2 P s so that i C bi maxi2P f1=ıi ; 1g exp.i / ai < 0.
Proof Lemma 1 implies that, for t 2 Œt0 ; t1 /,
(
kut0 k expŒi1 .t t0 /; if i1 2 P u ;
u.t/
kut0 k expŒi1 .t t0 /; if i1 2 P s ;
where kut0 k D sup s0 u.t0 C s/. Let

0 1
˚P .t0 ;t/
Y
w.t/ WD kut0 k @ ıij A exp Œ .t0 ; t/
jD1
where
X X
.t0 ; t/ WD i Tfig .t0 ; t/ i TQ fig .t0 ; t/:
i2P u i2P s
Suppose the result holds for t 2 Œtk1 ; tk /, that is, u.t/ w.t/. We claim that u.t/
w.t/ for t 2 Œtk ; tkC1 /. If not, then there exists a time t 2 Œtk ; tkC1 / such that u.t / D
w.t /, u.t/ w.t/ for all t 2 Œtk ; t / and for any " > 0 there exists a time t" 2
.t ; t C "/ such that u.t" / > w.t" /. Suppose that ikC1 2 P s , then
uP .t / ˇikC1 sup u.t C s/ ˛ikC1 u.t /;

s0
0 1
Y
k1
ˇikC1 maxfıi ; 1gkut0 k @ ıij A expŒ .t0 ; tk / expŒikC1 .t tk /
i2P
jD1
0 1
Yk1
˛ikC1 ıik kut0 k @ ıij A expŒ .t0 ; tk / expŒikC1 .t tk /;
jD1
and so
0 1
Yk
1
uP .t / ˇikC1 max ; 1 exp.ikC1 / ˛ikC1 kut0 k @ ıij A exp. .t0 ; t //:
i2P ıi jD1
It follows that uP .t / ikC1 w.t / D w.t

P /, a contradiction to the definition of t" .
On the other hand, if ikC1 2 P u , then
uP .t / ˇikC1 sup u.t C s/ ˛ikC1 v.t /;

s0
0 1
Y
k1
ˇikC1 maxfıi ; 1gkut0 k @ ıij A expŒ .t0 ; tk / expŒikC1 .t tk /
i2P
jD1
0 1
Y
k1
˛ikC1 ık kut0 k @ ıij A expŒ .t0 ; tk / expŒikC1 .t tk /;
jD1
and therefore,
0 1
Yk
1

uP .t / ˇikC1 max ; 1 ˛ikC1 kut0 k @ ıij A expŒ .t0 ; t /:
i2P ıi jD1
Thus, uP .t / ikC1 w.t / D w.t

P /, a contradiction. The result holds by induction.
The nonlinear perturbations are assumed to satisfy the following condition.
Assumption 1 Assume there exist #1 ; #2 0 such that, for all .t; ; / 2 RC
R0
PC PC, kF.t; / F.t; /k #1 k .0/ .0/k C #2 k .s/ .s/kds.
The following lemmata are required for the main result.
Lemma 2 (Matrix Cauchy Inequality) For any symmetric positive definite matrix
W 2 Rnn and x; y 2 Rn , 2xT y xT Wx C yT W 1 y.
Lemma 3 ([22]) For any positive scalar ", matrices U; V; W with W T W I,
UWV C V T W T U T "UU T C "1 V T V.
Lemma 4 ([28]) For any positive definite symmetric matrix W 2 Rnn ,
R v T R v
nonnegative
Rv T scalar v, and w W Œ0; v ! Rn , 0 w.s/ds W 0 w.s/ds
v 0 w .s/Ww.s/ds:
Lemma 5 ([23]) Let P be a symmetric positive definite matrix and let M, U, V and
W.t/ be real matrices of appropriate dimensions, with W.t/ being a matrix function.
Then, for any " > 0 such that P1 "UU T is positive definite and W.t/T W.t/ I,
.M C UW.t/V/T P.M C UW.t/V/ M.P1 "UU T /1 M C "1 V T V for all t 2 R.
We are now in a position to present the main synchronization result.
Theorem 1 Suppose that Assumption 1 holds. For i D 1; : : : ; m, let Pi 2 Rnn be
positive definite symmetric matrices and let "A;i ; "C;i ; "G;i ; "H;i be positive constants
such that P1
i "E;i
E;i
E;i is positive definite. Let
T T
ˇi WD 2 max .P1 1
i .C C Gi / Pi .C C Gi // C 2 kPi k#2 max .Pi /
T
C 2 Œ"G;i max .P1 1 T

i
G;i
G;i / C "C;i max .Pi
C
C /;
T
˛i WD 1 2kPi k#1 max .P1 1 1

i / max .Pi Œ.A C Hi / Pi C Pi .A C Hi / C "A;i Pi A A Pi
T T
C "A;i
AT
A C "1 T T 1 T 1
H;i Pi H;i H;i Pi C "H;i
H;i
H;i C "C;i Pi C C Pi C "G;i Pi G;i G;i Pi /;
T

max .Pi / 1
i WD ; WD max i ; i WD ˇi max ; 1 ˛i ;
min .Pi / i2P i2P ıi
ıi WDmax .P1 T 1
i Œ.I C Ei / .Pi "E;i
E;i
T

E;i /1 .I C Ei / C "1
E;i E;i E;i /;
T
where i > 0, i 2 P s , satisfy i C ˇi maxi2P f1=ıi; 1g exp.i / ˛i < 0. Let

P u WD fi 2 P W i 0g, P s WD fi 2 P W i < 0g, C WD maxi2P u i ,
WD mini2P s i , T C .t0 ; t/ WD TP u .t0 ; t/, T .t0 ; t/ WD TP s .t0 ; t/ ˚P s .t0 ; t/,

ı WD maxi2P ıi . Then, the following results hold:
(i) If there exist > 0, > 0 such that tk tk1 for all k 2 N, ı < 1,
T C .t0 ; t/ T .t0 ; t/, and ln.ı/= C C < 0, then RES is achieved.
(ii) If there exist > 0, > 0 such that tk tk1 for all k 2 N, ı > 1,
T C .t0 ; t/ T .t0 ; t/, and ln.ı/= C C < 0, then RES is achieved.
(iii) If there exist > 1, k > 0 for k 2 N such that tk tk1 k , P s ;,
and ln. ıik / C ik k 0 for all k 2 N, then RS is achieved.
(iv) If there exist > 1, k > 0 for k 2 N such that k tk tk1 , P u ;,
and ln. ıik / ik .k / 0 for k 2 N, then RS is achieved.
(v) If there exist tavg > 0, N0 > 0 such that k 1 N0 C .t t0 /=tavg for any
t 2 Œtk1 ; tk /, ı > 1, and ln.ı/ tavg < 0, then RS is achieved.
(vi) If there exist i > 0 for i 2 P such that, for t 2 Œtk1 ; tk /, .t/ D k and
.t C / D .t/, where i WD ti ti1 , WD 1 C 2 C : : : C m , then RS is
achieved if
X
m X X
ln ıi C i i i .i / < 0: (6)
iD1 i2P u i2P s
R0
Proof Let Vi .e/ WD eT Pi e for i D 1; : : : ; m, and let eQ .t/ WD e.t C s/ds. The
time-derivative of Vi along the ith mode of (3) for t ¤ tk is given by
dVi
.e.t// D eT .t/Œ.A C A.t/ Q i .t//T Pi C Pi .A C A.t/
Q C Hi C H Q C Hi C H Q i .t//e.t/
dt
C 2eT .t/Pi .C C C.t/
Q C Gi C G Q i .t//Qe.t/C2eT .t/Pi ŒFi .t; yt / Fi .t; xt /:
By Lemma 2, 2eT .t/Pi .C C Gi /Qe.t/ eT .t/Pi e.t/ C eQ T .t/.C C Gi /T Pi .C C Gi /Qe.t/.

Furthermore, Lemma 3 gives that
eT .t/.AQ T .t/Pi C Pi A.t//e.t/

Q "1
A;i e .t/Pi A A Pi e.t/ C "A;i e .t/
A
A e.t/;
T T T T
eT .t/.H Q i .t//e.t/ "1

Q iT .t/Pi C Pi H H;i e .t/Pi H;i H;i Pi e.t/ C "H;i e .t/
H;i
H;i e.t/;
T T T T
2eT .t/Pi .C.t/

Q CG Q i .t//Qe.t/ "1
C;i e .t/Pi C C Pi e.t/ C "C;i e
T T
Q T .t/
CT
C eQ .t/
C "1
G;i e .t/Pi G;i G;i Pi e.t/ C "G;i e
T T
Q T .t/
G;i
T

G;i eQ .t/:
Lemma 4 implies that

Z 0
eQ T .t/.C C Gi /T Pi .C C Gi /Qe.t/ eT .t C s/.C C Gi /T Pi .C C Gi /e.t C s/ds;

and can be applied similarly to the terms eQ T .t/

C;i
T

C;i eQ .t/ and eQ T .t/
G;i
T

G;i eQ .t/.
By Assumption 1,
Z 0
2eT .t/Pi ŒF.t; yt / F.t; xt / 2ke.t/kkPi k #1 ke.t/k C #2 ke.t C s/kds :

At the switching and impulsive moment t D tk ,
VikC1 .e.tk // eT .tk /.I C Eik C EQ ik .tk //T Pik .I C Eik C EQ ik .tk //e.tk /
eT .tk /Œ.I C Eik /T .P1 1
i "E;ik
E;ik
E;ik / .I C Eik /e.tk /
T
C eT .tk /Œ"1

E;ik E;ik E;ik e.tk /;
T
R0
by Lemma 5. Then, using eT .tCs/Pi e.tCs/ds sup s0 eT .tCs/Pi e.tCs/
and the well-known fact that for any positive definite matrix P 2 Rnn , symmetric
matrix Q 2 Rnn , and x 2 Rn , min .P1 Q/xT Px xT Qx max .P1 Q/xT Px, it
follows that v.t/ WD V .e.t// satisfies the conditions of Proposition 1 with bi WD ˇi ,
ai WD ˛i , hi WD 0, di WD ıi .
To prove case (i), note that P .t0 ; t/ D k 1 for t 2 Œtk1 ; tk /, and tk tk1
implies that t t0 k for t 2 Œtk1 ; tk /. Equation (5) implies that

v.t/ kvt0 k exp .k 1/ ln ı C C T C .t0 ; t/ T .t0 ; t/ (7)

kvt0 k .1=ı/ exp .t t0 / ln.ı/= C C T C .t0 ; t/ T .t0 ; t/

kvt0 k .1=ı/ exp .t t0 / ln.ı/= C .C /T .t0 ; t/

kvt0 k .1=ı/ exp .ln.ı/= C C /.t t0 /=M
for some M > 0 since suptt0 f.t t0 /=T .t0 ; t/g is finite. Robust exponential
synchronization of the drive and response systems follows. Since tk tk1
implies that t t0 k for t 2 Œtk1 ; tk /, case (ii) is proved by similar arguments
used to show case (i).
To prove case (iii), RS is implied from Eq. (5) since, for t 2 Œtk1 ; tk /,
2 3
Xk1 X
v.t/ kvt0 k exp 4 ln ıij C i Tfig .t0 ; t/5
jD1 i2P u
kvt0 k exp Œln ıi1 C i1 C ln ıi2 C i2 C : : : C ln ıik1 C ik1 C ik :
That is, v.t/ c k where c WD kvt0 k = mini2P ıi . Case (iv) is proved similarly.
Beginning from Eq. (7), case (v) implies that, for t 2 Œtk1 ; tk /,
v.t/ kvt0 k .ı/N0 expŒ.t t0 / ln.ı/=tavg C .C /T .t0 ; t/

kvt0 k .ı/N0 expŒ.ln.ı/=tavg C C /.t t0 /=M;
implying RS. For case (vi), Eq. (5) implies that, for j 2 N,
v.t0 C j!/
2 3
X
jm
X X
kvt0 k exp 4 ln ıi C i Tfig .t0 ; t0 C j!/ i TQ fig .t0 ; t0 C j!/5
iD1 i2P u i2P s
2 3
X
m X X
D kvt0 k exp 4j ln ıi C j i Tfig .t0 ; t0 C !/ j i TQ fig .t0 ; t0 C !/5
iD1 i2P u i2P s
2 0 13
Xm X X
kvt0 k exp 4j @ ln ıi C i i i .i /A5 :
iD1 i2P u i2P s
Since v.t/ is also bounded on any compact interval, the result follows.
3 Example
Consider system (3) with n D 2, P D f1; 2; 3; 4g, t0 D 0, distributed delay D

0:1,

0 1 2 0:5
AD ; CD ;
1 0 0 1:4

3:1 0 1:8 0:1 7:1 1:3 15 2
H1 D ; H2 D ; H3 D ; H4 D ;
0 3:1 0:1 1:6 1:3 22:3 2 8:8

1:05 1:1 0:4 0:4 0:1 0:1 0:12 0:13
G1 D ; G2 D ; G3 D ; G4 D ;
1:11 1:04 0:3 0:1 0:1 0:1 0:3 0:1

0:1 0:1 0:4 0 0:2 0 0:3 0:1
E1 D ; E2 D ; E3 D ; E4 D :
0:2 0:1 0:01 0:4 0 0:2 0:1 0:3
For 2 PC, let F. / D .0; 0:1 ln.cosh. 1 .0////. Let A;i D C;i D H;i D
G;i D E;i D I,
A;i D
C;i D
H;i D
G;i D
E;i D 0:01I, %A;i .t/ D %C;i .t/ D
%H;i .t/ D %G;i .t/ D %E;i .t/ D cos.5t/I. Suppose the initial conditions are given by
xt0 .s/ D .14; 20/ and y0 .s/ D .22; 33/ for s 0, and suppose that the
switching rule is periodic with 1 D 0:4, 2 D 0:4, 3 D 0:2, and 4 D 1 (i.e.,
! D 2).
(a) (b)
400 100
x (t)
1 x1(t)
300
5 σ(t)
x2(t)
80 x2(t)
4
y1(t) (u=0)
y (t)
200 1 3
y2(t) (u=0)
y (t)
60 2 2
100 1
40 0
0 0 2 4
Time
−100 20
−200 0
−300
−20
−400
0 0.5 1 1.5 2 2.5 3 3.5 4 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
Time Time
Fig. 1 Simulations of drive system (1) and response system (2). (a) Trajectories without hybrid
control. (b) Trajectories with hybrid control
Choose

10 10 0:354 0:020 0:253 0:037
P1 D ; P2 D ; P3 D ; P4 D ;
01 01 0:020 0:204 0:037 0:374
and "A;i D "H;i D "C;i D "H;i D "E;i D 10 for i 2 P, #1 D 0:1, #2 D 0. Then,

ˇ1 D 0:021, ˇ2 D 0:006, ˇ3 D 0:060, ˇ4 D 0:059, ˛1 D 5:602, ˛2 D 6:812,
˛3 D 12:866, ˛4 D 11:867, D 1:768, ı1 D 1:070, ı2 D 0:825, ı3 D 2:087,
ı4 D 1:028. Hence 1 D 5:623, 2 D 6:818, 3 D 12:806, 4 D 11:808, so
that P s D f3; 4g and P u D f1; 2g. Choose 1 D 11, 2 D 12, to get
4
X X X
ln ıi C i i i .i / D 5:485;
iD1 i2P u i2P s
which implies robust synchronization of the drive and response systems by Theo-
rem 1. See Fig. 1 for an illustration.
4 Conclusions
The sets P s and P u in Theorem 1 represent the synchronizing and desynchronizing

modes, respectively, while the conditions ı < 1 and ı > 1 represent impulsive
control and possible impulsive perturbations, respectively. Each theorem condition
establishes threshold conditions involving the mode activation times, mode syn-
chronization/desynchronization effects, and impulsive effects, for various classes
of switching rules (i.e., dwell-time (i)–(iv), average dwell-time (v), and periodic
switching (vi)). The results found are robust to model uncertainties, nonlinear
perturbations, and impulsive perturbations. One possible future direction is to
consider the robust hybrid synchronization from an optimal control point of view.
Acknowledgements This research was financially supported by the Natural Sciences and Engi-
neering Research Council of Canada (NSERC).
References
1. Khadra, A., Liu, X., Shen, X.: Application of impulsive synchronization to communication
security. IEEE Trans. Circuits Syst. I: Fundam. Theory Appl. 50(3), 341 (2003)
2. Olfati-Saber, R., Murray, R.M.: Consensus problems in networks of agents with switching
topology and time-delays. IEEE Trans. Autom. Control 49(9), 1520 (2004)
3. Li, C., Liao, X., Yang, X., Huang, T.: Impulsive stabilization and synchronization of a class of
chaotic delay systems. J. Nonlinear Sci. 15(4), 043 (2005)
4. Li, P., Cao, J., Wang, Z.: Robust impulsive synchronization of coupled delayed neural networks
with uncertainties. Phys. A: Stat. Mech. Appl. 373, 261 (2007)
5. Lakshmikantham, V., Rao, M.R.M.: Theory of Integro-Differential Equations. Gordon and
Breach Science Publishers, Amsterdam (1995)
6. Burton, T.: Volterra Integral and Differential Equations. Elsevier B.V., Amsterdam (2005)
7. Liu, J., Liu, X., Xie, W.C.: Delay-dependent robust control for uncertain switched systems with
time-delay. Nonlinear Anal. Hybrid Syst. 2(1), 81 (2008)
8. Liberzon, D.: Switching in Systems and Control. Birkhauser, Boston (2003)
9. Davrazos, G., Koussoulas, N.T.: A review of stability results for switched and hybrid systems.
In: Proceedings of 9th Mediterranean Conference on Control and Automation, Dubrovnik
(2001)
10. Evans, R.J., Savkin, A.V.: Hybrid Dynamical Systems. Birkhauser, Boston (2002)
11. Shorten, R., Wirth, F., Mason, O., Wulff, K., King, C.: Stability criteria for switched and hybrid
systems. SIAM Rev. 49(4), 545 (2007)
12. Hespanha, J.P., Morse, A.S.: Stability of switched systems with average dwell-time. In:
Proceedings of the 38th IEEE Conference on Decision and Control, Phoenix, vol. 3, p. 2655
(1999)
13. Guan, Z.H., Hill, D., Shen, X.: On hybrid impulsive and switching systems and application to
nonlinear control. IEEE Trans. Autom. Control 50(7), 1058 (2005)
14. Guan, Z.H., Hill, D., Yao, J.: A hybrid impulsive and switching control strategy for synchro-
nization of nonlinear systems and application to Chua’s chaotic circuit. Int. J. Bifurc. Chaos
16(1), 229 (2006)
15. Alwan, M.S., Liu, X.: On stability of linear and weakly nonlinear switched systems with time
delay. Math. Comput. Model. 48(7–8), 1150(2008)
16. Zhang, Y., Liu, X., Shen, X.: Stability of switched systems with time delay. Nonlinear Anal.
Hybrid Syst. 1(1), 44 (2007)
17. Yang, C., Zhu, W.: Stability analysis of impulsive switched systems with time delays. Math.
Comput. Model. 50(7–8), 1188 (2009)
18. Zhu, W.: Stability analysis of switched impulsive systems with time delays. Nonlinear Anal.
Hybrid Syst. 4(3), 608 (2010)
19. Niamsup, P.: Stability of time-varying switched systems with time-varying delay. Nonlinear
Anal. Hybrid Syst. 3(4), 631 (2009)
20. Wang, X., Li, X.: A generalized Halanay inequality with impulse and delay. Adv. Dev. Technol.
Int. 1(2), 9 (2012)
21. Wang, Y., Xie, L., de Souza, C.E.: Robust control of a class of uncertain nonlinear systems.
Syst. Control Lett. 19, 139 (1992)
22. de Souza, C.E., Li, X.: Delay-dependent robust H-infinity control of uncertain linear state-
delayed systems. Automatica 35, 1313 (1999)
23. Chen, W.H., Zheng, W.X.: Robust stability and H-infinity control of uncertain impulsive
systems with time-delay. Automatica 45, 109 (2009)
24. Xiang, Z., Chen, Q.: Robust reliable control for uncertain switched nonlinear systems with
time delay under asynchronous switching. Appl. Math. Comput. 216(3), 800 (2010)
25. Liu, X., Zhong, S., Ding, X.: Robust exponential stability of impulsive switched systems with
switching delays: a Razumikhin approach. Commun. Nonlinear Sci. Numer. Simul. 17, 1805
(2012)
26. Ballinger, G., Liu, X.: Existence and uniqueness results for impulsive delay differential
equations. Dyn. Contin. Discret. Impuls. Syst. 5(3), 579 (1999)
27. Liu, J., Liu, X., Xie, W.C.: Input-to-state stability of impulsive and switching hybrid systems
with time-delay. Automatica 47, 899 (2011)
28. Hien, L.V., Phat, V.N.: Exponential stabilization for a class of hybrid systems with mixed
delays in state and control. Nonlinear Anal. Hybrid Syst. 3(3), 259 (2009)
Regularization and Numerical Integration
of DAEs Based on the Signature Method
Andreas Steinbrecher
Abstract Modeling and simulation of dynamical systems often leads to

differential-algebraic equations (DAEs). Since direct numerical integration of DAEs
in general leads to instabilities and possibly non-convergence of numerical methods,
a regularization is required. We present three approaches for the regularization of
DAEs that are based on the Signature method. Furthermore, we present a software
package suited for the proposed regularizations and illustrate its efficiency on two
examples.
1 Introduction
Modeling and simulation of dynamical systems is an important issue in many

industrial applications and often leads to systems of differential-algebraic equations
(DAEs). In this paper, we consider nonlinear DAEs of the form
F.t; z; zP/ D 0; (1)
with sufficiently smooth functions F W I Rn Rn ! Rm and z W I ! Rn , with a

compact time interval I R.
In general, DAEs are not suited for direct numerical integration due to the
so called hidden constraints, which restrict the solution but are not explicitly
stated as equations. Hidden constraints can be determined by a certain number
of differentiation of (parts of) the DAE. This number of differentiations gives a
classification of difficulties in the treatment of DAEs, see [2, 4, 5, 12].
To guarantee a stable and robust numerical integration, a regularization of DAEs
is required. The basic idea of an approach to obtain a regularization for DAEs is
to consider the original system together with a sufficient number of its derivatives.
Then a system with the same solution set as the original system can be derived that
now contains all (hidden) constraints explicitly as algebraic equations. Then it can
A. Steinbrecher ()
Institute of Mathematics MA 4-5, TU Berlin, Straße des 17. Juni 136, 10623 Berlin, Germany

750 A. Steinbrecher
be guaranteed that all (hidden) constraints are satisfied within the numerical solution
and instabilities or drift from the solution manifold are avoided.
In most modeling and simulation tools, the current state-of-the-art to regularize
DAEs is to use some kind of structural analysis based on the sparsity pattern of
the system, to obtain necessary information for a regularization. The advantage of
a structural analysis in comparison to classical algebraic regularizations is that fast
algorithms based on graph theory can be applied. Usually, the Pantelides algorithm
[9] in combination with the Dummy Derivative Method [6] is used. In this paper,
we present regularization approaches based on the Signature method (˙-method)
[10].
In Sect. 3, we discuss regularization techniques that are based on the information
provided by the ˙-method which is reviewed in Sect. 2. Two of the regular-
ization techniques yield overdetermined systems of DAEs. Then, in Sect. 4 we
shortly discuss the software package QUALIDAES for the numerical integration
of (overdetermined) quasi-linear DAEs. In Sect. 5 we illustrate the applicability of
QUALIDAES in combination with the three proposed regularized formulations.
2 Signature Method: Structural Analysis of DAEs
The ˙-method, see [7, 8, 10] and also [11], can be applied to regular nonlinear
DAEs of arbitrary high order p of the form F.t; z; zP; : : : ; z. p/ / D 0. We denote by Fi
the ith component of the vector-valued function F and by zj the jth component of
the vector z. Then, the ˙-method consists of the following steps:
1. Building the signature matrix ˙ D Œij i;jD1;:::;n with
(
highest order of derivative of zj in Fi ,
ij D
1 if zj does not occur in Fi .
2. Finding a highest value transversal (HVT) of ˙, i.e., a transversal T D

f.1; j1 /; .2; j2 /; : : : ; .n; jn /g,Pwhere . j1 ; : : : ; jn / is a permutation of .1; : : : ; n/, with
maximal value Val.T/ WD .i;j/2T ij .
3. Computing the offset vectors c D Œci iD1;:::;n and d D Œdj jD1;:::;n with ci ; dj 2 N0
such that
dj ci ij for all i; j D 1; : : : ; n; and dj ci D ij for all .i; j/ 2 T: (2)
4. Forming the ˙-Jacobian J D ŒJij i;jD1;:::;n with

8
< @Fi
.ij / if dj ci D ij ;
Jij WD @zj
:
0 otherwise.
Regularization and Numerical Integration of DAEs Based on the Signature Method 751
5. Building the reduced derivative array

2 3
2 3 Fi .t; Z /
F 1 .t; Z / 6 7
dt Fi .t; Z / 7
d
6 :: 7 6
F .t; Z / D 4 : 5D 0 with F i .t; Z / D 6 :: 7 for i D 1; : : :n;
4 : 5
F n .t; Z / dci
dtci
Fi .t; Z /
(3)
h i
with Z T D z1 zP1 : : : z1.d1 / : : : zn zPn : : : zn.dn / :
6. Success check: if F .t; Z / D 0,Pconsidered locally as an algebraic system, has
a solution .t ; Z / 2 I RnC iD1 di , and J is nonsingular at .t ; Z /, then
n
.t ; Z / is a consistent point and the method succeeds.

If the ˙-method succeeds, the DAE (1) is called structurally regular with the
structural index S defined as S WD maxi ci if all dj > 0 or S WD maxi ci C 1
if some dj D 0. Note that the success check of the ˙-method is performed locally at
a fixed point .t ; Z /, such that the result may hold only locally in a neighborhood.
The HVT as well as the offset vectors c and d are not uniquely defined. However,
there exists a unique element-wise smallest solution, the so-called canonical offsets.
These are uniquely determined and independent of the chosen HVT, see [10].
The reduced derivative array (3) contains all necessary equations to extract all
hidden constraints by use of analytical manipulations. This forms the basis for all in
the following proposed regularization approaches.
Example 1 Let us consider the simple pendulum (Fig. 1) of mass m > 0, length
` > 0 under gravity g, see also [10, 11]. The state of the pendulum is described
by the position x and y, the velocities v and w of the mass point, and the Lagrange
multiplier . The system equations are given by
2 3 2 3
F1 .z; zP/ xP v
6 F .z; zP/ 7 6 7
6 2 7 6 yP w 7
6 7 6 7
0 D F.z; zP/ D 6 F3 .z; zP/ 7 D 6 mvP C 2x 7 (4)
6 7 6 7
4 F4 .z; zP/ 5 4 mwP C 2y C mg 5
F5 .z; zP/ x2 C y2 `2 (5)
T
where z D x y v w . The signature matrix ˙ for (4) and the canonical offset
vectors c and d are given by
2 3 2 3 2 3
1 1 0 1 1 1 2
61 1 0 17 617 627
6 1 7 6 7 6 7
6 7 6 7 6 7
˙ D6 0 1 1 1 0 7; c D 6 0 7; d D 6 1 7; (6)
6 7 6 7 6 7
41 0 1 1 0 5 405 415
0 0 1 1 1 2 0
752 A. Steinbrecher
Fig. 1 Simple pendulum Y

g
X
m
(x(t),y(t))
where the two possible HVTs are marked by light and dark gray boxes. The
corresponding ˙-Jacobian J and the reduced derivative array (3) takes the form
2 3
xP v
2 3 6 xR vP 7
1 0 1 0 0 6 7
6 P
y w 7
60 1 0 1 0 7 6 7
6 7 6 R
y wP 7
6 7 6 7
JD6 0 0 m 0 2x7 ; 0 D F .t; Z / D 6 mvP C 2x 7 (7)
6 7 6 7
40 0 0 m 2y5 6 mwP C 2y C mg 7
6 7
2x 2y 0 0 0 6 x2 C y2 `2 7
4 5
2xPx C 2yPy
2xRx C 2Px2 C 2yRy C 2Py2

with Z T D x xP xR y yP yR v vP w wP . Since det.J/ D 4m.x2 C y2 / D 4m`2 ¤ 0,
the ˙-Jacobian J is nonsingular at every consistent point, and the ˙-method
succeeds with S D maxi ci C 1 D 3. G
3 Regularization Approaches for DAEs
In the following, we discuss three different regularization approaches for struc-

turally regular DAEs (1) with m D n. These regularization approaches mainly use
the reduced derivative array (3) provided by the ˙-method and differ in its degree
of analytical preprocessing and possible degree of automation of the approach.
3.1 Regularization via State Selection

P
The reduced derivative array (3) consists of M D i ci Cn equations in n unknowns
z1 ; : : : ; zn .P
Thus, to obtain the same number of equations and unknowns, we have to
introduce i ci new variables.
For that we choose an HVT T from the set of all existing HVTs. Then for each
j D 1; : : : ; n we have a unique i with .i; j/ 2 T , and a regularization can be obtained
. C1/ .d /
from the reduced derivative array (3) by replacing the derivatives zj ij ; : : :; zj i
C1
with the algebraic variables !j ij ; : : :; !jdi if di > ij . We obtain the structurally
extended (StE) formulation
S .t; z; zP; !/ D 0 (8)
h i
C1
with ! T D !1T !nT and !jT WD !j ij !jdi 2 Rci if dj > ij or !j WD Œ 2
R0 if dj D ij as new algebraic variables. Note that dj D ij C ci by (2).
The obtained structurally extended formulation (8) is of increased size, not
unique, since it depends on the chosen HVT T , and in general it is only valid
locally in a neighborhood of a consistent point. Furthermore, it may lead to an
inappropriate regularization, if an inappropriate HVT T is used (see Example 2).
To choose a suitable HVT, i.e., one that is valid in a preferably large neighborhood
of a consistent
Q point, we define the (local) weighting coefficient for each HVT T
as T WD .i;j/2T jJij j and choose one with largest value T , i.e., we choose T
as an HVT of ˙ with T D maxT .T /. In [11] it is shown that the structurally
extended system (8) is (locally) a regular system of structural index S 1 if T
is chosen as described above. This formulation has locally the same set of solutions
for the original unknowns z as the original DAE (1). Due to the local validity, the
determination of a new regularized formulation during the (numerical) integration
may be necessary. In this case, the integration has to be interrupted and to be
restarted with the newly regularized model equations. Unfortunately, this leads to
an increase in simulation time and influences the obtained precision negatively.
This approach completely can be automated using automatic or symbolic
differentiation tools. Therefore, an analytical preprocessing is not necessary.
In the proposed method, the selection of variables for which derivatives are
replaced is directly prescribed by the HVT T and the offset vectors. Once the ˙-
method is done, this selection of variables is easy to achieve and requires no further
numerical computations. In contrast, in the Dummy Derivative Method, see [6]. As
a result, the two approaches might result in different regularized systems.
Example 2 The system (4) is regularPand of structural index S D 3. The reduced
derivative array (7) consists of M D ci C n D 9 equations in n D 5 unknowns.
754 A. Steinbrecher
For the HVT T D T1 marked by the light gray boxes in the signature matrix (6)
(assuming that T1 D 4x2 T2 D 4y2 ), we have to introduce new algebraic
variables as in the following table:
For j with i replace derivatives with new algebraic variables

. C1/ .d / C1 d
(s.t. .i; j/ 2 T ) zj ij ; : : :; zj j !j ij ; : : :; !j j
.1/ .2/
1 5 z1 ; z1 !11 ; !12
.2/
2 2 z2 !22
.1/
3 1 z3 !31
4 4 ; ;
5 3 ; ;
Then we construct the structurally extended formulation as
2 3
!11 v
6 !12 !31 7
6 7 2 3
6 7 2 13
6 yP w 7 x !1
6 !22 wP 7 6 7
6 7 6y7 6 !2 7
6 7 6 7 6 7
S 1 .t; z; zP; !/ D 6 m!31 C 2x 7 D 0 with z D 6 v 7; ! D 6 12 7:
6 7 6 7 4 !2 5
6 mwP C 2y C mg 7 4w5
6 7 !31
6 x2 C y2 `2 7
6 7
4 2x!11 C 2yPy 5
2 1 2 2 2
2x!1 C2.!1 / C2y!2 C2Py (9)
For the HVT T D T2 marked by the dark gray boxes in the signature matrix (6)
(assuming that T2 D 4y2 T1 D 4x2 ), we have to introduce new algebraic
variables as in the following table:
For j with i replace derivatives with new algebraic variables

. C1/ .d / C1 d
(s.t. .i; j/ 2 T ) zj ij ; : : :; zj j !j ij ; : : :; !j j
.2/
1 1 z1 !12
.1/ .2/
2 5 z2 ; z2 !2 ; !22
1
3 3 ; ;
.1/
4 2 z4 !41
5 4 ; ;
Then we construct the structurally extended formulation as

2 3
xP v
6 !12 vP 7
6 7 2 3
6 !21 w 7 2 23
6 7 x !1
6 2
!2 !4 1 7 6 7
6 7 6y7 6 !1 7
6 7 6 7 6 27
S 2 .t; z; zP; !/ D 6 mvP C 2x 7 D 0 with z D 6 v 7; ! D 6 2 7:
6 7 6 7 4 !2 5
6 m!41 C 2y C mg 7 4w5
6 7 !41
6 x2 C y2 `2 7
6 7
4 2xPx C 2y!21 5
2 2 2 1 2
2x!1 C2Px C2y!2 C2.!2 /
(10)
Applying the ˙-method to (9) and (10), we obtain a structurally regular system
with structural index S D 1 as long as x ¤ 0 while (10) forms a structurally regular
system with structural index S D 1 as long as y ¤ 0. G
3.2 Regularization via Algebraic Derivative Arrays
As we have seen in the previous section, the structurally extended regularization is

only valid locally. Therefore, it may be necessary to switch to another structurally
extended formulation within a numerical integration (also known as dynamic state
selection). This influences the time integration negatively. As a remedy in this
section we propose an extended regularization.
As mentioned above, the reduced derivative array (3) contains all necessary
equations to obtain all hidden constraints. Now we replace all derivatives of all
. j/ j
variables by algebraic variables, i.e., we replace zi by !i for i D 1; : : :; n and
j D 1; : : :; di , in the reduced derivative array to obtain the algebraic derivative array

Œ!i1 ; : : :; !idi T if di > 1;
0 D F .t; z; !/ D F .t; z1 ; !1 ; : : : ; zn ; !n / with !i D
Œ if di D 0:
From this it is possible to extract the set of all hidden constraints in an algebraic way
which, in particular, can be done automatically within the numerical integration.
In combination with the original DAE (1) we obtain the algebraic derivative
array (ADA) formulation

F.t; z; zP/
0 D A .t; z; zP; !/ D (11)
F .t; z; !/
which forms an overdetermined DAE for z and ! and has the same set of solutions
for the original unknowns z as the original DAE (1). The obtained algebraic
756 A. Steinbrecher
derivative array formulation is of more increased size as the structurally extended

formulation but globally valid such that the regularized formulation has to be
determined only once before the numerical integration. This approach also can
be automated using automatic or symbolic differentiation tools. Therefore, an
analytical preprocessing is not necessary.
Example 3 Let us use the model equations of the simple pendulum to illustrate
the regularization based on the algebraic derivative array formulation. The reduced
derivative array is given in (7). Replacing all occurring derivatives with algebraic
variables, i.e., xP ! !11 , xR ! !12 , yP ! !21 , yR ! !22 , vP ! !31 , and wP ! !41 , leads to
the algebraic derivative array
2 3
!11 v
6 !12 !31 7
6 7
6 1
!2 w 7
6 7
6 2
!2 !4 1 7
6 7
6 7
0 D F .t; z; !/ D 6 m!31 C 2x 7: (12)
6 7
6 1
m!4 C 2y C mg 7
6 7
6 2
x Cy L 2 2 7
6 7
4 1
2.x!1 C y!2 / 1 5
2.x!12 C.!11 /2 Cy!22 C.!21 /2 /
Together with the original DAE (4), we get the algebraic derivative array formula-
tion (11) with F.t; z; zP/ given in (4) and F .t; z; !/ given in (12). This corresponds
to an overdetermined set of DAEs containing 14 equations for 11 unknowns. G
3.3 Regularization via Overdetermined Formulation
In the previous sections, we did obtain regularizations that increased in its size more
then necessary. Therefore, in the numerical integration more computational time is
needed than necessary.
As mentioned in Sect. 2 the reduced derivative array (3) allows the determination
of the set of (hidden) constraints
0 D H.t; z/
in an analytical preprocessing. In combination with the original DAE (1), we obtain

the set

F.t; z; zP/
0 D O.t; z; zP/ D ;
H.t; z/
which forms the regularized overdetermined (OVD) formulation for z and which
has the same set of solutions as the original DAE (1). The obtained overdetermined
formulation is of increased number of equations for the same unknown variables as
in the original DAE. For more details see [12].
Example 4 For the simple pendulum, see Example 1, we get the set of (hidden)
constraints as
2 3
x2 C y2 `2
0 D H.t; z/ D 4 2xv C 2yw 5 (13)
v 2 Cw2 m2 .x2 C y2 /yg
and consequently the regularization via overdetermined formulations is determined

by (1), (13) as a set containing eight equations for five unknowns. G
4 The Software Package QUALIDAES
The software package QUALIDAES (QUAsi LInear DAE Solver) is suited for the
numerical integration of (overdetermined) quasi-linear DAEs of the form
" # " #
E.z; t/ k.z; t/ differential part;
zP D
0 G.z; t/ algebraic constraints:
QUALIDAES is implemented in FORTRAN. The discretization scheme is the 3-

stage Runge-Kutta RADAU IIa of order 5 and uses adapted decomposition methods
w.r.t. the overdeterminedness to solve the (hidden) constraints (numerically) pre-
cisely, while the differential part is solved in an ‘approximative sense’.
Further features of QUALIDAES are variable step size control, continuous
output, and check and correction of the initial values with respect to its consistency.
The specification of the system DAEs can be done in FORTRAN source code or
alternatively in MODELICA using a simple MODELICA-parser. For MODELICA
a Matlab interface is provided, see [1]. For more information to QUALIDAES we
refer to [13].
5 Numerical Results
In the following, we illustrate the efficiency of QUALIDAES with the proposed

regularization techniques on two examples. The first one is the simple pendulum,
and the second one is a mass-spring-chain with path constraints.
758 A. Steinbrecher
Efficiency (absolute error vs. simulation time)
QUALIDAES(ADA)
10 2
absolut error ERR=||numsol-refsol|| 2
QUALIDAES(StE)
DASSL(DI1)
10 0
10 -2
QUALIDAES(OVD)
RADAU5(GGL)
10 -4
10 -1 10 0 10 1
simulation time
Fig. 2 Efficiency (simulation time vs. absolute error for different prescribed tolerances)
The numerical integrations are done on an AMD Phenom(tm) II X6 1090T,

3210 MHz, 16 GB RAM, openSuSE 13.1 (Linux 3.11.10), GNU Fortran compiler
gcc version 4.8.1, no compiler options.
Example 5 The model equations of the simple pendulum, see Example 1, are stated
in (4) and have structural index S D 3 and differentiation index d D 3. The mass
and length are given by m D 1 and ` D 1 while we use the gravitational acceleration
g D 13:7503716373294544. In this case the exact solution has a period of 2 s which
allows the comparison of the accuracy every period.
The simulation is done for the time domain I D Œ0 s; 2000 s with initial values
T
z.0/ D 1 0 0 0 0 . To show the efficiency (see Fig. 2) of the numerical integration
with QUALIDAES in comparison we use the widely used solvers RADAU5, with the
Gear-Gupta-Leimkuler (GGL) formulation [3] and DASSL with the d-index-1 (DI1)
formulation consisting the first four equations of (4) and the last equation of (13).
The numerical solution QUALIDAES(OVD) offers the best efficiency, i.e., less
computational effort and small absolute error of the numerical solution. This
numerical solution is slightly more efficient as the numerical result obtained with
RADAU5(GGL) followed by the numerical solutions of QUALIDAES(ADA) and
QUALIDAES(StE). The large time consumption for the integration of the StE-form
lies in its local validity as discussed in Example 2. Within one period of 2 s it has to
be switched between (9) and (10) four times. Furthermore, due to the increased
size of the ADA-form and the StE-form both are less efficient than the OVD-
Fig. 3 Mass-spring-chain
F Control
p m
3
Path
p m
2
p m
1
F Control
form. The numerical integration using DASSL is not successful at all due to the
stability properties of BDF methods, the occurring drift off effect by use of the DI1-
form, where the constraints on position level and on velocity level are lost, and the
large time domain I. Nevertheless, the maximally obtained precision is excellent for
QUALIDAES(OVD) and QUALIDAES(ADA). G
Example 6 In the following we are interested in a path following problem of a mass-

spring-chain as in Fig. 3. On the first and the last body the force F is applied while
the center body has to follow a prescribed path (14g). We have pi and vi as the
position and velocity of body i D 1; 2; 3 and F as unknown variables. The model
equations are given by
pP 1 D v1 ; (14a)
pP 2 D v2 ; (14b)
pP 3 D v3 ; (14c)
mvP 1 D c. p1 p2 / C F; (14d)
mvP 2 D c. p1 p2 / c. p2 p3 /; (14e)
mvP 3 D c. p2 p3 / C F; (14f)
0 D p2 sin.t/ (14g)
760 A. Steinbrecher
Efficiency (absolute error vs. simulation time)

2
10
DASSL(DI1)
absolut error ERR=||numsol-refsol|| 2 10 0
RADAU5(DI1)
10 -2
QUALIDAES(ADA)
10 -4
QUALIDAES(StE)
-6
10
QUALIDAES(OVD)
10 -1
simulation time
Fig. 4 Efficiency (simulation time vs. absolute error for different prescribed tolerances)
with structural index S D 5 and differentiation index d D 5. Therefore, we have

hidden constraints up to level 4, i.e., up to the 4th derivative of (several) model
equations are necessary to determine all hidden constraints. We use m D 1 and
c D 1=6 with the exact solution p2 .t/ D sin.t/, v2 .t/ D cos.t/, p1 .t/ D p3 .t/ D
2 sin.t/, v1 .t/ D v3 .t/ D 2 cos.t/, F.t/ D 3 sin.t/=2. The simulation is done for
the time do-main I D Œ0 s; 500 s.
In the same style as in the last example, the efficiency of the numerical solutions
is illustrated in Fig. 4.
Again the numerical solution QUALIDAES(OVD) offers the best efficiency
followed by the numerical solution QUALIDAES(StE). In contrast to the previous
example, the efficiency using the StE-form is better since the StE-form is valid for
the whole time domain and has not to be switched. Furthermore, the numerical solu-
tion QUALIDAES(ADA) offers a good efficiency. The efficiency of RADAU5(DI1)
is similar, but the obtained precision is not as good as for the previous results. The
reason is that we used the DI1-form, where the hidden constraints of level 4 are
included instead of (14g). Therefore, the (hidden) constraints up to level 3 are lost,
and the solution drifts away from these lost constraints. Nevertheless, the maximally
obtained precision is excellent for QUALIDAES with all regularizations proposed in
Sect. 3, i.e., for the StE-form, the ADA-form, and the OVD-form. G
6 Summary
The aim of this article was to discuss several approaches for the regularization of
differential-algebraic equations that benefit numerical integration. For this purpose,
we have proposed three different regularization approaches which end in the
structurally extended formulation, the algebraic derivative array formulation, and
the regularized overdetermined formulation. All these regularization approaches
are based on the signature method, which was reviewed in Sect. 2, and two of
them require numerical integration methods suited for overdetermined differential-
algebraic equations. We also briefly introduced the software package QUALIDAES
and illustrated its efficiency for two examples.
Acknowledgements This work has been supported by the European Research Council through
Advanced Grant MODSIMCONMP.
References
1. Altmeyer, R., Steinbrecher, A.: Regularization and Numerical Simulation of Dynamical

Systems Modeled with Modelica. Preprint 29-2013, Institut für Mathematik, Berlin (2013)
Differential Algebraic Equations. Classics in Applied Mathematics, vol. 14. SIAM, Philadel-
phia (1996)
3. Gear, C.W., Leimkuhler, B., Gupta, G.K.: Automatic integration of Euler-Lagrange equations
with constraints. J. Comput. Appl. Math. 12–13, 77–90 (1985)
4. Hairer, E., Wanner, G.: Solving Ordinary Differential Equations II – Stiff and Differential-
Algebraic Problems, 2nd edn. Springer, Berlin (1996)
5. Kunkel, P., Mehrmann, V.: Differential-Algebraic Equations. Analysis and Numerical Solution.
EMS Publishing House, Zürich, 2006.
variables. SIAM J. Sci. Comput. 14, 677–692 (1993)
7. Nedialkov, N., Pryce, J., Tan, G.: DAESA – a Matlab tool for structural analysis of DAEs: soft-
ware. Technical report CAS-12-01-NN, Department of Computing and Software, McMaster
University, Hamilton (2012)
8. Nedialkov, N.S., Pryce, J.D.: Solving differential-algebraic equations by Taylor series (i):
computing Taylor coefficients. BIT Numer. Math. 45(3), 561–591 (2005)
9. Pantelides, C.C.: The consistent initialization of differential-algebraic systems. SIAM J. Sci.
Stat. Comput. 9(2), 213–231 (1988)
10. Pryce, J.D.: A simple structural analysis method for DAEs. BIT 41(2), 364–394 (2001)
11. Scholz, L., Steinbrecher, A.: Regularization of DAEs based on the signature method. BIT
Numer. Math. 1–22 (2015). doi:10.1007/s10543-015-0565-x
12. Steinbrecher, A.: Numerical solution of quasi-linear differential-algebraic equations and
industrial simulation of multibody systems. Ph.D. thesis, Technische Universität Berlin (2006)
13. Steinbrecher, A.: QUALIDAES: a software package for the numerical integration of quasi-
linear differential-algebraic equations. Technical report, Institut für Mathematik, Technische
Universität Berlin, Berlin (in preparation)
Symbolic-Numeric Methods for Improving
Structural Analysis of Differential-Algebraic
Equation Systems
Guangning Tan, Nedialko S. Nedialkov, and John D. Pryce
Abstract Systems of differential-algebraic equations (DAEs) are generated rou-

tinely by simulation and modeling environments, such as MapleSim and those based
on the Modelica language. Before a simulation starts and a numerical method is
applied, some kind of structural analysis is performed to determine which equations
to be differentiated, and how many times. Both Pantelides’s algorithm and Pryce’s
˙-method are equivalent in the sense that, if one method succeeds in finding
the correct index and producing a nonsingular Jacobian for a numerical solution
procedure, then the other does also. Such a success occurs on many problems of
interest, but these structural analysis methods can fail on simple, solvable DAEs and
give incorrect structural information including the index. This article investigates ˙-
method’s failures and presents two symbolic-numeric conversion methods for fixing
them. Both methods convert a DAE on which the ˙-method fails to a DAE on which
this SA may succeed.
1 Introduction
We consider differential-algebraic equations (DAEs) of the general form
fi . t; the xj and derivatives of them / D 0; i D 1; : : : ; n; (1)
G. Tan ()
School of Computational Science and Engineering, McMaster University, 1280 Main Street West,
Hamilton, ON, L8S 4L8, Canada
N.S. Nedialkov
Department of Computing and Software, McMaster University, 1280 Main Street West, Hamilton,
ON, L8S 4L8, Canada
J.D. Pryce
School of Mathematics, Cardiff University, Senghennydd Road, Cardiff CF24 4AG, Wales, UK

764 G. Tan et al.
where the xj .t/; j D 1; : : : ; n are state variables that are functions of an independent
variable t, usually regarded as time.
Pryce’s structural analysis (SA), the ˙-method [10], determines for (1) its
structural index, number of degrees of freedom (DOF), variables and derivatives that
need initial values, and the constraints of the DAE. These SA results can help decide
how to apply an index reduction algorithm [3], perform a regularization process
[12], or design a solution scheme for a Taylor series method [5–7]. The ˙-method
is equivalent to Pantelides’s algorithm [9]: they both produce the same structural
index [10, Theorem 5.8], which is an upper bound for the differentiation index, and
often both indices are the same [10].
The ˙-method succeeds on many problems of practical interest, producing a
nonsingular System Jacobian. However, this SA can fail—hence Pantelides’s algo-
rithm can fail as well—on some simple, solvable DAEs, producing an identically
singular System Jacobian.
We investigate such SA’s failures and present two symbolic-numeric conversion
methods for fixing them. After each conversion, provided some conditions are
satisfied, the value of the signature matrix is guaranteed to decrease. We conjecture
that such a decrease should result in a better problem formulation of a DAE, so that
the SA may produce a nonsingular System Jacobian and hence succeed.
Section 2 summarizes the ˙-method. Section 3 describes SA’s failures. Section 4
presents our two conversion methods, each of which is illustrated with an example
therein. Section 5 presents concluding remarks.
This article is a succinct version of the technical report [13]; the reader is referred
to it for more detailed explanations and examples.
2 Summary of the ˙ -Method
This SA method [10] constructs for a DAE (1) an n n signature matrix ˙, whose
.i; j/ entry ij is either an integer 0, the order of the highest derivative to which
variable xj occurs in equation fi , or 1 if xj does not occur in fi .
A highest-value transversal (HVT) of ˙ is a set T of n positions .i; j/ with one
position in each row and each column, such that the sum of the corresponding entries
is maximized. This sum is the value of ˙, written Val.˙/. If Val.˙/ is finite, then
the DAE is structurally well posed (SWP); otherwise it is structurally ill posed.
We assume henceforth the SWP case. Using the HVT, we find n equation and n
variable offsets c1 ; : : : ; cn and d1 ; : : : ; dn , respectively, which are integers satisfying
ci 0 for all i; dj ci ij for all i; j with equality on a HVT . (2)
These offsets are valid but not unique. There exists an elementwise smallest solution
of (2) termed the canonical offsets [10].
Symbolic-Numeric Methods for Improving Structural Analysis of DAEs 765
We associate with given valid offsets an n n System Jacobian J, defined as

( .ij /
@fi =@xj if dj ci D ij ; and
Jij D (3)
0 otherwise .
If J is nonsingular at a consistent point as defined in [5], then we say the ˙-method

succeeds and there exists (locally) a unique solution through this point [10].
In the success case, the canonical offsets determine the structural index and the
number of DOF, defined as
(
1 if minj dj D 0
S D max ci C
i 0 otherwise
X X X
and DOF D Val.˙/ D ij D dj ci :
.i;j/2T j i
(DOF means “degrees of freedom”, while DOF is the corresponding number.)

Example 1 We illustrate1 the above concepts with the simple pendulum DAE of
differentiation index 3.
0 D f1 D x00 C x x y ci x y
f1 "2 0ı # 0 f1 " 1 x#
0 D f2 D y00 C y g ˙ D f2 2 0 0
ı J D f2 1 y
ı
f3 0 0 2 f3 2x 2y
0 D f3 D x2 C y2 L2
dj 2 2 0
The state variables are x; y, and ; G is gravity and L > 0 is the length of the
pendulum. There are two HVTs of ˙, marked with and ı, respectively. A blank
in ˙ denotes 1, and a blank in J denotes 0.
Since det.J/ D 2.x2 C y2 / D 2L2 ¤ 0, the System Jacobian is nonsingular,
and the SA succeeds. The structural index is S D mini ci C 1 D 2 C 1 D 3 (because
minj dj D d3 D 0),Pwhich equals
P the differentiation index. The number of DOF is
DOF D Val.˙/ D d
j j c
i i D 4 2 D 2.
1
When we present a DAE example, we show its signature matrix ˙, canonical offsets ci and dj ,
and the associated System Jacobian J.
766 G. Tan et al.
3 Structural Analysis’s Failure
We say that the ˙-method fails, if a DAE (1) has a finite Val.˙/ and an identically
singular J. In the failure case, the SA reports Val.˙/ as an “apparent” DOF but not
a meaningful one.
In this article, we focus on the case where such an identically singular J is
structurally nonsingular. That is, there exists a HVT T of ˙ such that Jij is
generically nonzero for all .i; j/ 2 T.
Example 2 We illustrate a failure case with the following DAE2 in [1, p. 23].
0 D f1 D x0 C ty0 g1 .t/ x y ci x y
f 1 1 0 f 1 t
˙D 1 JD 1
0 D f2 D x C ty g2 .t/ f2 0 0 1 f2 1 t
dj 1 1
The SA fails since det.J/ D 0. Now J is identically singular but structurally

nonsingular: on a HVT marked by , each of J11 D 1 and J22 D t is generically
nonzero.
One simple fix is to replace f1 by f 1 D f1 C f20 , which results in the system
below; cf. [3, Example 5].
0 D f 1 D y C g1 .t/ g02 .t/ x y ci x y

˙ D f1 0 0 J D f1 1
0 D f2 D x C ty g2 .t/ f2 0 0 0 f2 1 t
dj 0 0
Now that det.J/ D 1 ¤ 0, the SA succeeds. Notice Val.˙ / D 0 < 1 D Val.˙/.

Another simple fix is to introduce a new variable z D x C ty and eliminate x in f1
and f2 .
0 D f 1 D y C z0 g1 .t/ y z ci y z
f 0 1 0 f 1 1
˙D 1 JD 1
0 D f2 D z g2 .t/ f2 0 1 f2 1
dj 0 1
For this resulting DAE, det.J/ D 1 ¤ 0, and the SA succeeds. After solving for y
and z, we can obtain x D z ty. This fix also gives Val.˙ / D 0 < 1 D Val.˙/.
The above two manipulations illustrate the two conversion methods presented in
Sect. 4, respectively.
2
In the original formulation, the driving functions are f1 ; f2 . Here they are renamed g1 ; g2 .
4 Conversion Methods
We present two conversion methods for systematically fixing SA’s failures. The first
method is based on replacing an existing equation by a linear combination of some
equations and derivatives of them. We call this method the linear combination (LC)
method and describe it in Sect. 4.1. The second method is based on substituting
newly introduced variables for some expressions and enlarging the system. We call
this method the expression substitution (ES) method and describe it in Sect. 4.2.
We only present the main features of these methods; the reader is referred to [13]
for more details and especially the proofs of Theorems 1 and 2 below.
Given a DAE (1), we assume it has a finite Val.˙/ and a System Jacobian J that is
identically singular but structurally nonsingular. We also assume the equations in (1)
are sufficiently differentiable, so that our methods fit into the ˙-method theory.
After a conversion, we obtain a system with signature matrix ˙ and System
Jacobian J. If Val.˙/ is finite and J is still identically singular, then we can
perform another conversion using either of the methods, provided the corresponding
conditions are satisfied. Suppose a sequence of conversions produces a solvable
DAE with Val.˙ / 0 and a generically nonsingular J. Given the fact that each
conversion reduces some Val.˙/ by at least one, the total number of conversions
does not exceed the value of the original signature matrix.
If the resulting system is structurally ill posed after a conversion, that is,
Val.˙ / D 1, then we say the original DAE is ill posed [13].
4.1 Linear Combination Method
Let u be a nonzero n-vector function in the cokernel coker.J/ of J, that is, JT u is

identically zero. We consider J and u as functions of t, the xj ’s and derivatives of
them.
Let xj ; u denote the order of the highest derivative to which xj occurs in the
expressions defining the function u, or 1 if xj does not occur in u at all.
Denote
˚
I D f i j ui 6 0 g; c D min ci ; and L D i 2 I j ci D c : (4)
i2I
Here ui 6 0 means that ui is generically nonzero. The LC method is based on the

following theorem.
Theorem 1 If

xj ; u < d j c for all j D 1; : : : ; n ; (5)
768 G. Tan et al.
and we replace an fl with l 2 L by

X .ci c/
fl D ui fi ; (6)
i2I
then Val.˙/ < Val.˙/, where ˙ is the signature matrix of the resulting DAE.
We call (5) the condition for applying the LC method.
Example 3 We illustrate this method with the following problem:
0 D f1 D x01 C x3 0 D f3 Dx1 x2 C g1 .t/

0 D f2 D x02 C x4 0 D f4 Dx1 x4 C x2 x3 C x1 C x2 C g2 .t/ ;
where g1 and g2 are given driving functions.
x1 x2 x3 x4 c i x1 x2 x3 x4
f1 2 1 0 30 f1 2 1 1 3

1 0 70 17
˙D 26 JD 26
f f 1
f3 4 0 0 51 f3 4 x2 x1 5

f4 0 0 0 0 0 f4 x2 x1
dj 1 1 0 0
Here Val.˙/ D 1. A shaded entry ij in ˙ marks a position .i; j/ where dj ci >

ij 0 and hence Jij D 0 by the formula (3) for J. The SA fails here since J is
identically singular (but structurally nonsingular).
T ˚
We choose u ˚D x2 ; x1 ; 1; 1 2 coker.J/. Then (4) becomes I D 1; 2; 3; 4 ,
c D 0, and L D 1; 2; 4 . The condition (5) holds since
.x1 ; u/ D 0 < 1 0 D d1 c; .x2 ; u/ D 0 < 1 0 D d2 c ;

.x3 ; u/ D 1 < 0 0 D d3 c; .x4 ; u/ D 1 < 0 0 D d4 c :
Selecting l D 4 2 L for example, we replace f4 by

X .ci c/
f4 D ui fi D x2 f1 C x1 f2 C f30 f4 D x1 x2 C g01 .t/ g2 .t/ :
i2I
The resulting DAE is 0 D . f1 ; f2 ; f3 ; f 4 / with the SA results as follows.
x1 x2 x3 x4 c i x1 x2 x3 x4
f1 2 1 0 30 f1 2 1 1 3

1 0 70 17
˙D 26 JD 26
f f 1
f3 4 0 0 51 f3 4 x2 x1 5
f 4 0 0 1 f 4 1 1
dj 1 1 0 0
Now Val.˙ / D 0 < 1 D Val.˙/. The SA succeeds at all points where det.J/ D
x2 x1 ¤ 0.
From (4) and (6), we can recover the replaced equation fl by

P .c c/ ı
fl D f l i2Inflg ui fi i ul :
Provided ul ¤ 0 for all t in the interval of interest, it is not difficult to show that the
original DAE and the resulting one have the same solution (if there exists one); see
[13, § 5.3].
4.2 Expression Substitution Method
Let v be a nonzero n-vector function in the kernel of J; write v 2 ker.J/. Denote

˚
JD j j vj 6 0 ; s D jJj ;
˚ (7)
MD i j dj ci D ij for some j 2 J ; and c D max ci :
i2M
We choose an l 2 J, and introduce s 1 new variables
.dj c/ vj .dl c/ ˚

yj D xj x for all j 2 J n l : (8)
vl l
In each fi with i 2 M, we

vj .d c/ .cci /
substitute yj C xl l
vl (9)
. / ˚
for every xj ij with ij D dj ci and j 2 J n l :
Denote by f i the equations that result from these substitutions, and write f i D fi for
i … M. Using (8), we append to these f i ’s the equations
.dj c/ vj .dl c/ ˚

0 D gj D yj C xj x for all j 2 J n l (10)
vl l
that prescribe the substitutions. Hence the resulting enlarged system consists of
˚
equations 0 D f 1 ; : : : ; f n and 0 D gj for all j 2 J n l
˚
in variables x1 ; : : : ; xn and yj for all j 2 J n l :
Key to applying the ES method is the following theorem.

770 G. Tan et al.
Theorem 2 Let J, s, M, and c be as defined in (7). Assume

(
< dj c if j 2 J
xj ; v and dj c 0 for all j 2 J : (11)
dj c otherwise ,
For any l 2 J, if we
• introduce .s 1/ new variables as defined in (8),
• perform substitutions in fi for all i 2 M by (9), and
• append the equations gj in (10),
then Val.˙/ < Val.˙/, where ˙ is the signature matrix of the resulting DAE.
We call (11) the conditions for applying the ES method.
Example 4 We illustrate the ES method with the artificially constructed DAE below.
0 00
0 D f1 D x1 C ex1 x2 x2 C h1 .t/ x1 x2 ci x1 x2
f 1 2 0 f ˛ ˛x2
˙ D f1 0 1 1
J D f1 1 x2
0 D f2 D x1 C x2 x02 C x22 C h2 .t/ 2 2
dj 1 2
0 00
Here h1 and h2 are given driving functions, and ˛ D ex1 x2 x2 . Obviously det.J/ D
0 and the SA fails.
Suppose we choose v D .x2 ; 1/T 2 ker.J/. Then (7) becomes
˚ ˚
J D 1; 2 ; s D jJj D 2; M D 1; 2 ; and c D max ci D 1 :
i2M
We can apply the ES method as the conditions (11) hold:
.x1 ; v/ D1 1 D 1 1 1D d1 c 1; d1 c D 1 1 D 0 0 ;

.x2 ; v/ D 0 0 D 2 1 1D d2 c 1; d2 c D 2 1 D 1 0 :
˚ ˚
For example, we choose l D 2 2 J. Now J n l D 1 . Using (8) and (10), we
introduce a new variable
.d c/ v1 .d2 c/ .11/ x2 .21/
y1 D x1 1 x2 D x1 x D x1 C x2 x02 ;
v2 1 2
and append the equation 0 D g1 D y1 C x1 C x2 x02 . Then we substitute .y1 x2 x02 /0
for x01 in f1 to obtain f 1 , and substitute y1 x2 x02 for x1 in f2 to obtain f 2 . The resulting
DAE and its SA results are shown below.
0 02
0 D f 1 D x1 C ey1 Cx2 C h1 .t/ 2x1 x2 y1 3 ci
"
x1 x2 y1
#
f1 0 1 1 0 f1 1 2x02 ˇ ˇ
0 D f 2 D y1 C x22 C h2 .t/ ˙ D f2 4 0 0 5 1 J D f2 2x2 1
g1 0 1 0 0 g1 1 x2
0 D g1 D y1 C x1 C x2 x02 dj 0 1 1
0 02
Here ˇ D ey1 Cx2 . Now Val.˙/ D 1 < 2 D Val.˙/. The SA succeeds at all points
where det.J/ D 2ˇ.x2 C x02 / x2 ¤ 0.
From the steps of applying the ES method, we can “undo” the expression
substitutions to recover the original DAE. Similar to the LC method, the ES method
also guarantees that, provided vl ¤ 0 for all t in some interval of interest, the original
DAE and the resulting one have equivalent solutions (if any) in a natural sense; cf.
[13, § 6.3].
Our experience suggests that it is effective to attempt the LC method first, and
if its condition (5) is violated, then we try the ES method. Also, it is desirable to
choose an l 2 L [resp. l 2 J] in the LC [resp. ES] method, such that an ul [resp. vl ]
never becomes zero. For example, it can be a nonzero constant, x21 C 1, or 2 C cos x2 .
Such a choice of l guarantees that the resulting DAE is “equivalent” to the original
one—that is, they always have the same solution, if there exists one.
We show below that the LC method does not succeed on the DAE in Example 4
because the condition (5) is not satisfied.
T 0 00
Choose u D 1; ˛ 2 coker.J/, where ˛ D ex1 x2 x2 . Then (4) becomes
˚ ˚ ˚ ˚
ID i j ui 6 0 D 1; 2 ; c D min ci D 0; and L D i 2 I j ci D c D 1 :
i2I
Since x1 and x2 occur of orders 1 and 2 in u, respectively, we have
.x1 ; u/ D 1 6< 1 0 D d1 c and .x2 ; u/ D 2 6< 2 0 D d2 c :
Hence (5) is not satisfied. Choosing l D 1 2 I for example and replacing f1 by

f 1 D u1 f1 C u2 f20 D x1 C h1 .t/ C ˛ 1 C x01 C x2 x002 C .x02 /2 C 2x2 x02 C h02 .t/

results in the DAE 0 D f 1 ; f2 .
x1 x2 ci
x1 x2
˙ D f1 1 2 0
J D f1 ˛ ˛x2
f2 0 1 1 f2 1 x2
dj 1 2
Here D x01 C x2 x002 C .x02 /2 C 2x2 x02 C h02 .t/. The SA fails still, since J is identically
singular. Now Val.˙ / D Val.˙/ D 2.
5 Conclusions
We proposed two symbolic-numeric conversion methods for improving the ˙-

method. They convert a DAE with a finite Val.˙/ and an identically singular (but
structurally nonsingular) System Jacobian to a DAE that may have a generically
772 G. Tan et al.
nonsingular System Jacobian. A conversion guarantees that both DAEs have (at
least locally) the same solution. The conditions for applying these methods can be
checked automatically, and the main result of a conversion is Val.˙ / < Val.˙/,
where ˙ is the signature matrix of the resulting DAE.
An implementation of these methods requires
(1) computing a symbolic form of a System Jacobian J,
(2) finding a vector in coker.J/ [respectively ker.J/],
(3) checking the LC condition (5) [respectively ES conditions (11)], and
(4) generating the equations for the resulting DAE.
These symbolic computations may seem expensive. However, for DAEs whose J
can be permuted into a block-triangular form [8, 11], we can locate the diagonal
blocks that are singular and then apply our methods to these blocks only, instead of
working on the whole DAE. This approach is presently under development.
We combine MATLAB’s Symbolic Math Toolbox [14] with our structural analysis
software DAESA [8, 11], and have built a prototype code that applies our conversion
methods automatically. We aim to incorporate them in a future version of DAESA.
With our prototype code, we have applied our methods on numerous DAEs. They
are either arbitrarily constructed to be “SA-failure cases” for our investigations, or
borrowed from the existing literature: Campbell and Griepentrog’s robot arm [2], the
transistor amplifier and the ring modulator [4], and the linear constant coefficient
DAE in [12]. Our conversion methods succeed in fixing all these solvable DAEs;
see [13] and especially Appendix B thereof. We believe that our assumptions and
conditions are reasonable for practical problems, and that these methods can help
make the ˙-method more reliable.
Finally, we conjecture that reducing Val.˙/ tends to give a better formulation of
a DAE from the SA’s perspective and then a nonsingular System Jacobian.
Acknowledgements The authors acknowledge with thanks the financial support for this research:
GT is supported in part by the Ontario Research Fund, Canada, NSN is supported in part by the
Natural Sciences and Engineering Research Council of Canada, and JDP is supported in part by
the Leverhulme Trust, the UK.
References
1. Brenan, K., Campbell, S., Petzold, L.: Numerical Solution of Initial-Value Problems in
Differential-Algebraic Equations, 2nd edn. SIAM, Philadelphia (1996)
2. Campbell, S.L., Griepentrog, E.: Solvability of general differential-algebraic equations. SIAM
J. Sci. Comput. 16(2), 257–270 (1995)
4. Mazzia, F., Iavernaro, F.: Test set for initial value problem solvers. Technical report 40,
Department of Mathematics, University of Bari (2003). http://pitagora.dm.uniba.it/~testset/
5. Nedialkov, N.S., Pryce, J.D.: Solving differential-algebraic equations by Taylor series (I):
computing Taylor coefficients. BIT Numer. Math. 45(3), 561–591 (2005)
6. Nedialkov, N.S., Pryce, J.D.: Solving differential-algebraic equations by Taylor series (II):
computing the system Jacobian. BIT Numer. Math. 47(1), 121–135 (2007)
7. Nedialkov, N.S., Pryce, J.D.: Solving differential-algebraic equations by Taylor series (III): the
DAETS code. JNAIAM J. Numer. Anal. Ind. Appl. Math. 3, 61–80 (2008)
8. Nedialkov, N.S., Pryce, J.D., Tan, G.: Algorithm 948: DAESA—a Matlab tool for structural
analysis of differential-algebraic equations: software. ACM Trans. Math. Softw. 41(2), 12:1–
12:14 (2015). doi:10.1145/2700586
9. Pantelides, C.: The consistent initialization of differential-algebraic systems. SIAM. J. Sci.
Stat. Comput. 9, 213–231 (1988)
(2001)
11. Pryce, J.D., Nedialkov, N.S., Tan, G.: DAESA—a Matlab tool for structural analysis of
differential-algebraic equations: theory. ACM Trans. Math. Softw. 41(2), 9:1–9:20 (2015).
doi:10.1145/2689664
12. Scholz, L., Steinbrecher, A.: A combined structural-algebraic approach for the regularization of
coupled systems of DAEs. Technical report 30, Reihe des Instituts für Mathematik Technische
Universität Berlin, Berlin (2013)
13. Tan, G., Nedialkov, N., Pryce, J.: Symbolic-numeric methods for improving structural analysis
of differential-algebraic equation systems. Technical report, Department of Computing and
Software, McMaster University, 1280 Main St. W., Hamilton, ON, L8S4L8, Canada (2015).
84pp. Download link:http://www.cas.mcmaster.ca/cas/0reports/CAS-15-07-NN.pdf
14. The MathWorks, Inc.: Matlab symbolic math toolbox (2015). https://www.mathworks.com/
help/symbolic/index.html
Pinning Stabilization of Cellular Neural
Networks with Time-Delay Via Delayed
Impulses
Kexue Zhang, Xinzhi Liu, and Wei-Chau Xie
Abstract This paper studies the pinning stabilization problem of time-delay

cellular neural networks. A new pinning delayed-impulsive controller is proposed
to stabilize the networks. Based on a Razumikhin-type stability result, a global
exponential stability of time-delay cellular neural networks with delayed impulses is
constructed. It is shown that the global exponential stabilization of delayed cellular
neural networks can be effectively realized by controlling a small portion of units in
the networks via delayed impulses. Numerical simulations are provided to illustrate
the theoretical result.
1 Introduction
Cellular neural networks (CNNs), introduced by Chua and Yang in [1, 2], have
attracted the attentions of several researchers in recent years. This is mainly due
to their broad applications in many areas, including image processing and pattern
recognition (see, e.g., [2, 3] ), data fusion [4], odor classification [5], and solving
partial differential equations [6].
In real-world applications, it is inevitable for the existence of time delay in
the processing and transmission of signals among units of CNNs. Hence, it is
practical to investigate CNNs with time-delay (DCNNs) (see, e.g., [7–12]). Stability
of DCNNs, as a prerequisite for their applications, has been studied extensively in
the past decades, and various control methods have been introduced to stabilize
DCNNs, such as intermittent control [9], sliding mode control [10], impulsive
control [11], and sampled-data control [12]. Among these control algorithms, the
impulsive control method has been proved to be an effective approach to stabilize
DCNNs. The control mechanism of this method is to control the unit states of
K. Zhang () • X. Liu

W.-C. Xie
Department of Civil and Environmental Engineering, University of Waterloo, Waterloo, ON,
N2L 3G1, Canada

776 K. Zhang et al.
the CNN with small impulses, which are small samples of the state variables of
the CNN, at a sequence of discrete moments. Since the time delay in unavoidable
in sampling and transmission of the impulsive information is dynamical systems,
many control problems of dynamical systems have been investigated via delayed
impulses in recent years, such as stabilization of stochastic functional systems [13]
and synchronization of dynamical networks [14].
The regular impulsive control method to stabilize a CNN is to control each
unit of the network to tame the unit dynamics to approach a steady state (i.e.,
equilibrium point). However, a neural network is normally composed of a large
number of units, and it is expensive and infeasible to control all of them. Motivated
by this practical consideration, the idea of controlling a small portion of units,
named pinning control, was introduced in [15, 16], and many pinning impulsive
control algorithms have been reported for many control problems of dynamical
networks (see, e.g., [17–23]) It is worth noting that no time delay is considered
in these pinning impulsive controllers proposed in the above literatures. However,
it is natural and essential to consider the delay effects when processing the impulse
information in the controller.
Due to the cost effective advantage of impulsive control method and pinning
control strategy and the wide existence of time delay, it is practical to investigate
the pinning impulsive control approach that takes into account of delays. However,
to our best knowledge, no such result has been reported for stabilization of DCNNs.
Therefore, in this paper, we propose a novel pinning delayed-impulsive controller
for the DCNNs. The remainder of this paper is organized as follows. In Sect. 2,
we formulate the problem and introduce the pinning delayed-impulsive control
algorithm. In Sect. 3, a global exponential stability result of the impulsive DCNNs
is obtained. Then, in Sect. 4, a numerical example is considered to illustrate the
theoretical result. Finally, conclusions are stated in Sect. 5.
2 Preliminaries
Let N denote the set of positive integers, R the set of real numbers, RC the set of
nonnegative real numbers, and Rn the n-dimensional real space equipped with the
Euclidean norm. For a; b 2 R with a < b and S Rn , we define
n ˇ
ˇ
PC .Œa; b; S/ D W Œa; b ! Sˇ .t/ D .tC /; for any t 2 Œa; b/I .t /
exists in S; for any t 2 .a; bI .t / D .t/ for all but

o
at most a finite number of points t 2 .a; b ;
where .tC / and .t / denote the right and left limit of function at t, respec-
tively. For a given constant > 0, the linear space PC .Œ ; 0; Rn / is equipped
with the norm defined by jj jj D sups2Œ ;0 jj .s/jj, for 2 PC .Œ ; 0; Rn /.
Pinning Stabilization Via Delayed Impulses 777
Consider the following CNN with delays:
xP D Cx.t/ C Af .x.t// C Bf .x.t 1 //; (1)
where x D .x1 ; x2 ; : : :; xn /T 2 Rn , and xi is the state of the ith unit; n denotes

the number of units in CNN (1); f .x.t// D . f1 .x1 .t//; f2 .x2 .t//; : : :; fn .xn .t///T and
fj .xj .t// denotes the output of the jth unit at time t; A D Œaij nn and B D Œbij nn ;
constants aij and bij represent the strengths of connectivity between units i and j
at time t and t 1 , respectively; 1 corresponds to the transmission delay when
processing information from the jth unit; C D Diagfc1 ; c2 ; : : :; cn g and constant ci
denotes the rate at which the ith unit will reset its potential when disconnected with
other units of the network.
Throughout this paper, we make the following assumptions:
(A) f .0/ D 0 and there exists a constant L such that jj f .u/ f .v/jj Ljju vjj
for all u; v 2 Rn .
The Lipschtiz condition on the nonlinear activation function fi has been widely
considered due to its significance in the application of CNNs (see, e.g., [2, 24, 25]).
Based on the assumption, system (1) admits the trivial solution.
The objective of this paper is to design the following delay-dependent pinning
impulsive controller to exponentially stabilize CNN (1):
P1
kD1 Œ1 xi .t/ C 2 xi .t 2 /ı.t tk /; i 2 Dlk ;
Ui .t; xi / D (2)
0; i 62 Dlk ;
for i D 1; 2; : : :; n, where 1 and 2 are impulsive control gains to be determined;

2 > 0 denotes the time delay in controller Ui ; the impulsive instant sequence
ftk g satisfies ftk g R, 0 t0 < t1 < : : : < tk < : : :, and limk!1 tk D 1;
ı./ is the Dirac Delta function, and xi .tk / denotes the left limit of xi at time tk .
Let l denote the number of units to be pinned at each impulsive instant, and the
index set Dlk is defined as follows: reorder states x1 .tk /, x2 .tk /; : : :; xn .tk / such
that jxp1 .tk /j jxp2 .tk /j : : :jxpl .tk /j jxplC1 .tk /j : : : jxpn .tk /j, then
define Dlk D fp1; p2; : : :; plg. The pinning impulsive control mechanism can be
explained as follows: at each impulsive instant, we only control l units that have
larger deviations from the trivial state than the remaining n l units.
The definition of Dlk is borrowed from [26]. Since no delay has been considered
in the pinning impulsive controller in [26], the pinning algorithm introduced in [26]
can be treated as a special case of our pinning delayed-impulsive control strategy
(i.e., 2 D 0).
The impulsive controlled CNN (1) can be written in the form of an impulsive
system:
8
< xP .t/ D Cx.t/ C Af .x.t// C Bf .x.t 1 //; t 2 Œtk1 ; tk /;
x .t / D 1 xi .tk / C 2 xi .tk 2 /; i 2 Dlk ; k 2 N; (3)
: i k
xt0 D ;
778 K. Zhang et al.
where D .1 ; 2 ; : : :; n /T and i 2 PC .Œa; b; R/ is the initial function,

xi .tk / D xi .tkC / xi .tk /, and xt0 is defined by xt0 .s/ D x.t0 C s/ for s 2 Œ ; 0
and D maxf 1 ; 2 g. Then the pinning impulsive stabilization problem of CNN (1)
is transformed into the stability problem of impulsive system (3).
Definition 1 The trivial solution of system (3) is said to be globally exponentially
stable (GES), if there exist constants ˛ > 0 and M 1 such that, for any initial data
xt0 D , the following inequality holds,
jjx.t; t0 ; /jj Mjjjj e˛.tt0 / ; for all t t0 :
3 Stabilization
In this section, we will use a Razumikhin-type stability criterion to construct

verifiable conditions for the GES of impulsive CNN (3). First, a class of Lyapunov
functions for impulsive systems is introduced.
Definition 2 Function V W RC Rn ! RC is said to belong to the class 0 if the
following is true:
(1) V is continuous in each of the sets Œtk1 ; tk / Rn , and for each x 2 Rn , t 2
Œtk1 ; tk /, and k 2 N, lim.t;y/!.tk ;x/ V.t; y/ D V.tk ; x/ exists;
(2) V.t; x/ is locally Lipschiz in all x 2 Rn , and for all t t0 , V.t; 0/ 0.
To establish a stability result for system (3), the following lemma is required.
Lemma 1 Assume that there exist V 2 0 , and constants p; q; c; w1 ; w2 ; 1 > 0,
and 2 0, such that
(i) w1 jjxjjp V.t; x/ w2 jjxjjp , for t 2 RC and x 2 Rn ;
(ii) for t 2 Œtk1 ; tk /, 2 PC .Œ ; 0; Rn /, and k 2 N,
V 0 .t; .0// cV.t; .0//;
whenever V.t C s; .s// < qV.t; .0// for all s 2 Œ ; 0;

(iii) for k 2 N and 2 PC .Œ ; 0; Rn /,
V.tk ; .0/ C Ik .tk ; // 1 V.tk ; .0// C 2 sup fV.tk C s; .s//gI

s2Œ ;0
1
(iv) q > 1 C2
> ecd , where d D supk2N ftkC1 tk g.
Then the trivial solution of system (3) is GES.
Lemma 1 is a direct consequence of Theorem 3.1 in [27]. Next, verifiable
conditions will be constructed for the GES of system (3) by utilizing a quadratic
Lyapunov function.
Theorem 1 If the following inequality is satisfied
ecd < 1; (4)

q l 2 jjBjj
where D 1 nl C n
j1 C 1 j C j2 j , c D 2 mini fci g C 2L.jjAjj C p /

> 0,
and d D supk2N ftkC1 tk g, then the trivial solution of system (3) is GES.
Proof Consider the Lyapunov function V.x/ D xT x. Note that condition (i) of
Lemma 1 is satisfied with w1 D w2 D 1 and p D 2. Taking the time-derivative
along solutions of (3) yields
P
V.x/ D 2xT .t/Pxh.t/ i
D 2xT .t/ Cx.t/ C Af .x.t// C Bf .x.t 1 //

2cmin C 2jjAjjL V.x.t// C 2jjBjjLjjx.t/jjjjx.t 1 /jj

2cmin C 2jjAjjL C jjBjjL"1 V.x.t// C "jjBjjLV.x.t 1 //; (5)
where cmin D mini fci g and constant " > 0. It can be seen from (4) that there exists
a constant q > 0 such that
1
q> > ecN d ; (6)

p
where cN D 2 mini fci g C 2L.jjAjj C qjjBjj/. If V.x.t C s// < qV.x.t// for all
s 2 Œ ; 0, then we can obtain from (5) that

P
V.x/ 2cmin C 2jjAjjL C jjBjjL."1 C q"/ V.x.t//: (7)
Define h as a function of ": h."/ D 2cmin C 2jjAjjL C jjBjjL."1 C q"/. Then,

for " > 0, function h attains its minimum value cN at " D p1q (that is, h0 ."/ D 0 at
" D p1 ). Therefore, we can get from (7) that V.x/
q
P cN V.x/.
Given a constant > 0, letting N1 D 1 nl Œ1 .1 C /.1 C 1 /2 , then
X
.1 N1 / x2i .tk / .1 N1 /.n l/ min fx2i .tk /g
i2Dlk
i62Dlk
D lŒN1 .1 C /.1 C 1 /2 min fx2i .tk /g

i2Dlk
X
ŒN1 .1 C /.1 C 1 /2 x2i .tk /;
i2Dlk
780 K. Zhang et al.
i.e.,
X X X
n
2
.1 C /.1 C 1 / x2i .tk / C x2i .tk / N1 x2i .tk /:
i2Dlk i62Dlk iD1
Then, for t D tk , we have

X X
V.x.tk // D x2i .tk / C x2i .tk /
i2Dlk i62Dlk
X X
D Œ.1 C 1 /xi .tk / C 2 xi .tk 2 /2 C x2i .tk /
i2Dlk i62Dlk
X
Œ.1 C /.1 C 1 /2 x2i .tk / C .1 C 1 /22 x2i .tk 2 /
i2Dlk
X
C x2i .tk /
i62Dlk
X X
.1 C /.1 C 1 /2 x2i .tk / C x2i .tk /
i2Dlk i62Dlk
X
n
C .1 C 1 /22 x2i .tk 2 /
iD1
N1 V.x.tk // C .1 C 1 /22 V.x.tk 2 //

1 V.x.tk // C 2 sup fV.x.tk C s//g;
s2Œ ;0
where 1 D N1 , 2 D .1 C 1 /22 , and constant > 0 to be determined to minimize

the value of 1 C 2 .
2 p n
N
Let h./ D nl .1 C 1 /2 C 22 1 , then, for D j 1C1
j l , we have hN 0 ./ D 0,
N
i.e., h./ attains its minimum for > 0. Hence,
l
r l 2
D minf1 C 2 g D 1 C j1 C 1 j C j2 j :
>0 n n
Based on the above discussion, we can conclude that all the conditions of
Lemma 1 are satisfied. Thus, the trivial solution of system (3) is GES. t
u
Remark 1 Parameter is related to impulsive control gains 1 , 2 and the ratio l=n.
It can be seen from (4) that, the fewer units are controlled at impulsive instants, the
more frequently the impulsive controllers need to be added to the network.
4 Numerical Simulations
In this section, we will consider an example to demonstrate our theoretical result. In

order to observe the pinning control process clearly, we will investigate a CNN with
only two units in the following example.
Example 1 Consider CNN (1) with n D 2, c1 D c2 D 1, 1 D 1,

2 0:1 1:5 0:1
AD ; BD ;
5 3 0:2 2:5
and f .x/ D . f1 .x1 /; f2 .x2 //T with f1 ./ D f2 ./ D tanh./. The chaotic attractor of
CNN (1) is shown in Fig. 1.
We consider two types of impulsive controllers:
(1) l D 1, i.e., impulsive control one unit at each impulsive instant. Let tk
tk1 D 0:03, 2 D 1, 1 D 0:868, and 2 D 0:2, then (4) is satisfied.
Thus, Theorem 1 implies that the trivial solution of (3) is GES. See Fig. 2 for
numerical simulations.
(2) l D 2, i.e., impulsive control two units at each impulsive instant. Let tk tk1 D
0:08, and 2 , 1 , 2 are the same as those in the first scenario, then (4) is satisfied
and Theorem 1 implies that the trivial solution of (3) is GES. Numerical results
are shown in Fig. 3.
The initial data in Figs. 2 and 3 is chosen the same as that in Fig. 1, and the red
dot denotes the state x at initial time t D 0. The vertical (or horizontal) lines in
Fig. 2a represent the state jump of x2 (or x1 ) while the other state is unchanged.
Since both units are controlled in Fig. 3a, no vertical and horizontal lines can be
observed. It can be seen from Fig. 2 that different unit may be controlled at different
impulsive instants. This is consistent with our pinning algorithm of controlling the
unit which has the largest state deviation with the equilibrium. However, it is more
practical to control one specific unit at all impulsive instants. Next, we apply the
pinning impulsive controller to the first and second unit at all impulsive instants
respectively, and numerical results are shown in Fig. 4a, b. The impulsive control
782 K. Zhang et al.
1
x2
−1
−2
−3
−4
−5
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
x1
Fig. 1 Chaotic behavior of CNN (1) with the parameters given in Example 1. The initial data for
this simulation is .s/ D Œ1; 1T for s 2 Œ 1 ; 0, and the red dot denotes the state x at the initial
time t D 0
(a) 0.4 (b) 1

x
x12
0.2
0.5
0
−0.2 0
x2
−0.4
x
−0.6 −0.5
−0.8
−1
−1
−1.2 −1.5
−0.2 0 0.2 0.4 0.6 0.8 1 1.2 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x1 t
Fig. 2 Impulsive control one unit of CDN (1) at each impulsive instant. (a) Phase portrait.
(b) State trajectories
gains 1 , 2 , and the impulsive sequence ftk g are chosen the same as those in Fig. 2.
Figure 4 implies that stabilization cannot be realized via this type of pinning strategy
with the given parameters, and more strict conditions may be required to guarantee
the stability which will be investigated in our future research.
(a) 0 (b) 1
x
−0.2
x12
0.5
−0.4
−0.6 0
x2
x
−0.8 −0.5
−1
−1
−1.2
−1.4 −1.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
x t
1
Fig. 3 Impulsive control both units of CDN (1) at each impulsive instant. (a) Phase portrait.
(b) State trajectories
(a) (b)
2.5 x
1.5 x
1 1
2 x x
2 2
1.5 1
1 0.5
0.5
0 0
x
−0.5
−1 −0.5
−1.5 −1
−2
−2.5 −1.5
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
t t
Fig. 4 Impulsive control one specific unit of CDN (1) through all the impulsive instants.
(a) Control the first unit. (b) Control the second unit
5 Conclusion
We have studied the stabilization problem of cellular neural networks with time-
delay. An impulsive controller that takes into account both the pinning control
algorithm and time delays has been proposed. Sufficient conditions for the global
exponential stability of cellular neural networks with time-delay and delayed
impulses have been derived by using a Razumikhin-type stability result. Our result
has shown that the delayed cellular neural networks can be exponentially stabilized
by pinning controlling a small portion of units at each impulsive instant. Numerical
simulations have been provided to demonstrate the effectiveness of our theoretical
result.
784 K. Zhang et al.
References
1. Chua, L.O., Yang, L.: Cellular neural networks: theory. IEEE Trans. Circuits Syst. 35(10),
1257–1272 (1988)
2. Chua, L.O., Yang, L.: Cellular neural networks: applications. IEEE Trans. Circuits Syst.
35(10), 1273–1290 (1988)
3. Jamrozik, W.: Cellular neural networks for welding arc thermograms segmentation. Infrared
Phys. Technol. 66, 18–28 (2014)
4. Chatziagorakis, P., Sirakoulis, G.Ch., Lygouras, J.N.: Design automation of cellular neural
networks for data fusion applications. Microprocess. Microsyst. 36(1), 33–44 (2012)
5. Ayhan, T., Yalçin, M.E.: An application of small-world cellular neural networks on odor
classification. Int. J. Bifurc. Chaos 22(01), 1250013 (12pp.) (2012)
6. Hadada, K., Piroozmanda, A.: Application of cellular neural network (CNN) method to the
nuclear reactor dynamics equations. Ann. Nucl. Energy 34(5), 406–416 (2007)
7. Civalleri, P.P., Gilli, M., Pandolfi, L.: On stability of cellular neural networks with delay. IEEE
Trans. Circuits Syst. I: Fundam. Theory Appl. 40(3), 157–165 (1993)
8. Mohamad, S., Gopalsamy, K.: Exponential stability of continuous-time and discrete-time
cellular neural networks with delays. Appl. Math. Comput. 135(1), 17–38 (2003)
9. Yu, J., Hu, C., Jiang, H., Teng, Z.: Exponential lag synchronization for delayed fuzzy cellular
neural networks via periodically intermittent control. Math. Comput. Simul. 82(5), 895–908
(2012)
10. Gan, Q., Xu, R., Yang, P.: Synchronization of non-identical chaotic delayed fuzzy cellular
neural networks based on sliding mode control. Commun. Nonlinear Sci. Numer. Simul. 17(1),
433–443 (2012)
11. Wang, Q., Liu, X.: Impulsive stabilization of cellular neural networks with time delay via
Lyapunov functionals. J. Nonlinear Sci. Appl. 1(2), 72–86 (2008)
12. Kalpana, M., Balasubramaniam, P.: Stochastic asymptotical synchronization of chaotic Marko-
vian jumping fuzzy cellular neural networks with mixed delays and the Wiener process based
on sampled-data control. Chin. Phys. B 22(7), 078401 (10pp.) (2013)
13. Cheng, P., Deng, F., Yao, F.: Exponential stability analysis of impulsive functional differential
equations with delayed impulses. Commun. Nonlinear Sci. Numer. Simul. 19(6), 2104–2114
(2014)
14. Zhang, W., Tang, Y., Miao, Q., Fang, J.: Synchronization of stochastic dynamical networks
under impulsive control with time delays. IEEE Trans. Neural Netw. Learn. Syst. 25(10), 1758–
1768 (2014)
15. Wang, X.F., Chen, G.: Pinning control of scale-free dynamical networks. Phys. A: Stat. Mech.
Appl. 310(3–4), 521–531 (2002)
16. Li, X., Wang, X.F., Chen, G.: Pinning a complex dynamical netwrok to its equilibrium. IEEE
Trans. Circuits Syst. I: Regul. Pap. 51(10), 2074–2087 (2004)
17. Lu, J., Kurths, J., Cao, J., Mahdavi, N., Huang, C.: Synchronization control for nonlinear
stochastic dynamical networks: pinning impulsive strategy. IEEE Trans. Neural Netw. Learn.
Syst. 23(2), 285–292 (2012)
18. Mahdavi, N., Menhaj, M.B., Kurths, J.: Pinning impulsive synchronization of complex
dynamical networks. Int. J. Bifurc. Chaos 22(10), 1250239 (14pp.) (2012)
19. Hu, A., Xu, Z.: Pinning a complex dynamical network via impulsive control. Phys. Lett. Sect.
A: Gen. Atom. Solid State Phys. 374(2), 186–190 (2009)
20. Zhou, J., Wu, Q., Xiang, L.: Pinning complex delayed dynamical networks by a single
impulsive controller. IEEE Trans. Circuits Syst. I: Regul. Pap. 58(12), 2882–2893 (2011)
21. Lu, J., Wang, Z., Cao, J.: Pinning impulsive stabilization of nonlinear dynamical networks with
time-varying delay. Int. J. Bifurc. Chaos 22(7), 1250176 (12pp.) (2012)
22. Yang, X., Cao, J., Yang, Z.: Synchronization of coupled reaction-diffusion neural networks
with time-varying delays via pinning-impulsive controller. SIAM J. Control Optim. 51(5),
3486–3510 (2013)
23. Zhang, K., Liu, X., Xie, W.C.: Impulsive control and synchronization of spatiotemporal chaos
in the Gray Scott model. In: Interdisciplinary Topics in Applied Mathematics, Modeling
and Computational Science. Springer Proceedings in Mathematics & Statistics, pp. 549–555.
Springer, Cham (2015)
24. Trivedi, A.R., Datta, S., Mukhopadhyay, S.: Application of Silicon-Germanium source Tunnel-
FET to enable ultralow power cellular neural network-based associative memory. IEEE Trans.
Electron Devices 61(11), 3707–3715 (2014)
25. Yildiz, N., Cesur, E., Kayaer, K., Tavsanoglu, V., Alpay, M.: Architecture of a fully pipelined
real-time cellular neural network emulator. IEEE Trans. Circuits Syst. I: Regul. Pap. 62(1),
130–138 (2015)
26. Lu, J., Kurths, J., Cao, J., Mahadavi, N., Huang, C.: Sychronization control for nonlinear
stochastic dynamical networks: pinning impulsive strategy. IEEE Trans. Netw. Learn. Syst.
23(2), 285–292 (2012)
27. Cheng, P., Deng, F., Yao, F.: Exponential stability analysis of impulsive stochastic functional
differential systems with delayed impulses. Commun. Nonlinear Sci. Numer. Simul. 19(6),
2104–2114 (2014)
Convergence Analysis of the Spectral Expansion
of Stable Related Semigroups
Yixuan Zhao and Pierre Patie
Abstract The purpose of this note is to carry out a convergence analysis of the
spectral representation, derived recently in Patie and Savov (Spectral expansions
of non-self-adjoint generalized Laguerre semigroups, submitted, 2015), of some
non-self-adjoint Markovian semigroups related to spectrally negative ˛-stable Lévy
processes conditioned to stay positive. More specifically, we start by performing an
error analysis for the spectral type series expansions. Moreover, these semigroups
are closely related to a class of invariant semigroups, whose speed of convergence to
equilibrium has been studied in Patie and Savov (Spectral expansions of non-self-
adjoint generalized Laguerre semigroups, submitted, 2015). Our second aim is to
carry out a numerical analysis on the convergence rate which is illustrated with two
examples.
1 Introduction
The class of Lévy stable and related processes have attracted much attention
over the past decades. In particular, this class of processes have been shown to
be an appropriate statistical description for a large number of phenomena. Such
applications include, e.g., the study of turbulence in physics [4], seismic series
and earthquakes in geography [6], risk process in financial mathematics [10], and
primary sequences of protein-like copolymers in biomedical science [3]. Therefore,
being able to explicitly represent the transition kernel and the solution to the Cauchy
problem associated to stable related semigroups has become a critical issue in
many areas. In our context, we consider a spectrally negative ˛-stable process
Z D .Zt /t0 , with ˛ 2 .1; 2/ . It means that Z is a process with stationary and
independent increments, having no positive jumps, and its law is characterized by
Y. Zhao ()
Cornell University ORIE Department, 288 Rhodes Hall, Cornell University, Ithaca, NY, USA
P. Patie
Cornell University ORIE Department, 206 Rhodes Hall, Cornell University, Ithaca, NY, USA

788 Y. Zhao and P. Patie
its characteristic exponent which takes the form, for t > 0; <.z/ > 0,
Z 0
˛ jyj˛1
ln EŒe D z t D
zZt
.ezy 1 zy/ dy
1
.˛/
where throughout
stands for the Gamma function. Next, let K 0 D .Kt0 /t0 be the
semigroup of the process Z killed upon entering the negative half-line. That is, for
f 2 Bb .RC /, the set of bounded Borelian functions on RC D Œ0; 1/, and for any
t; x > 0, Kt0 is defined by
Kt0 f .x/ D Ex Œ f .Zt /Ift<T0 g ;
where T0 D infft > 0I Zt < 0g. By [2, Section 3.2], p˛1 .x/ D x˛1 ; x > 0; is
an excessive function for K 0 , i.e. p˛1 0 on RC and for all t > 0, Kt0 p˛1 .x/
"
p˛1 .x/. Hence one may construct a semigroup K " D .Kt /t0 through Doob’s h-
transform as follows,
Kt" f .x/ D p1 0

˛1 .x/Kt .p˛1 f /.x/; x > 0;
which turns out to be the Feller semigroup of the ˛-stable process conditioned to
stay positive, and will be referred to as ˛-SP+ throughout . It has been shown in [2]
that the infinitesimal generator of K " , denoted by L" , can be expressed in terms of
an Erdélyi-Kober type operator in the form, for a smooth function f ,
Z
d2 x
y˛1 f .y/ dy
L" f .x/ D x˛ D˛0C;1;1 f .x/ D x1˛ :
dx2 0 .x y/ ˛1
.2 ˛/
It should be noted that K " is also a ˛1 -self-similar semigroup as in the definition of

[5], i.e. Kt" f .cx/ D Kc"1=˛ t dc f .x/ for any c > 0, where dc f .x/ D f .cx/. Hence it is
easily seen that the semigroup K D .Kt /t0 defined by
1
Kt f .x/ D Kt" f ı p˛ .x ˛ /
is a 1-self-similar semigroup, as p˛ .x/ D x˛ is a homeomorphism of RC to RC .

For sake of presentation, it will be more convenient to work with the semigroup K.
According to a celebrated result by Lamperti [5], there exists a bijection between
the set of convolution semigroups on R and the one of positive self-similar Feller
semigroups. In particular, it has been shown, see [2, 8], that K is associated via
the Lamperti mapping to the semigroup of a spectrally negative Lévy process D
.t /t0 , with Laplace exponent given, for z 2 CC D fz 2 CI <.z/ > 0g, by

.˛z C ˛/
log EŒez1 D .z/ D : (1)

.˛z/
Convergence Analysis of the Spectral Expansion of Stable Related Semigroups 789
Another semigroup of interests is the so-called associated generalized Laguerre (gL)

semigroup denoted by P D .Pt /t0 and defined as follows
Pt f .x/ D Ket 1 f ı det .x/; x > 0: (2)
Moreover, P admits a stationary measure with absolutely continuous density .x/dx,

and by [9, Theorem 1.6], P can be extended uniquely to a strongly continuous
2
contraction semigroup, also denoted
2
R 1by2 P, on the weighted Hilbert space L ./
endowed with the norm jj f jj D 0 f .x/.x/dx. From Bertoin and Yor [1], the
R1 n Qn
kD1 .k/
integer moments of are given by 0 x .x/dx D nŠ
D
.˛nC˛/
.˛/nŠ
, which
yields that its Mellin transform is
Z 1

.˛s/
M .s/ D xs1 .x/dx D ; s 2 CC : (3)
0
.˛/
.s/
Hence standard Mellin inversion technique gives
X1
.1/nC1 sin. n
˛ /
. ˛ C 1/ ˛n
n
.x/ D x ; (4)
nD0

.˛/nŠ
which, by Stirling approximation, defines a function analytical on CC and by [9,

Theorem 2.1], is a positive probability density on RC . It can be easily shown that
the semigroup P is non-self-adjoint in L2 ./ and thus the spectral theorem can
not be applied. However, Patie and Savov [9] have recently developed an original
approach to provide a spectral representation of the class of gL semigroups, which
enables us to write Pt f and Kt f as infinite series expansions under some mild
conditions on f and t. In particular, we shall observe that when f is a power function,
the series expansion is reduced to finitely many terms and thus can be evaluated
precisely. Hence the first aim of this paper is to perform a numerical convergence
analysis for the series expansion. Moreover, in [9], an interesting study on the speed
of convergence of a gL semigroup to its equilibrium is presented. In particular,
depending on the subspace of L2 ./ one considers, one can observe either the
hypocoercivity phenomena, that is bounds of the form Cet or a non-classical bound
of the form C.e2t 1/1=2 for some C > 0. Therefore, the second objective of this
paper is to carry out a numerical Ranalysis on the speed of convergence for Pt f to f
1
in L2 ./ as t ! 1, where f D 0 f .x/.x/dx. Such numerical plots can serve as
illustrations for the theoretical results in [9].
In this paper, we also study another class of semigroups called the Gauss-
Laguerre semigroup, denoted by G D .Gt /t0 and defined as follows. For any
1
˛ 2 .1; 2/ and ˇ 2 Œ1 ˛1 ; 1/, we introduce the Gauss-Laguerre operator defined
for a smooth function f and x > 0, by
Z 1
sin..˛ 1//
L˛;ˇ f .x/ D .d˛;ˇ x/f 0 .x/ C x f 00 .xy/g˛;ˇ .y/dy; (5)
0

.˛ˇC˛ˇ/
where d˛;ˇ D
.˛ˇC1ˇ/ and

.˛ 1/ ˇC 1 C1 1
g˛;ˇ .y/ D 1
y ˛1 2 F1 .˛ˇ C˛ ˇ; ˛I ˛ˇ C˛ ˇ C1I y ˛1 /; (6)
ˇ C ˛1 C1
with 2 F1 the Gauss hypergeometric function. Then it is shown in [7, Theorem 1.1]
that L˛;ˇ is the generator of a non-self-adjoint contraction Markov semigroup G in
L2 .e˛;ˇ /, where
1 1
xˇC ˛1 1 ex ˛1
e˛;ˇ .x/ D ; x > 0; (7)

..˛ 1/ˇ C 1/
is its unique invariant measure. The motivation for studying this class of semigroups
is that while they share many similar properties with P, the expressions of their
spectral expansions are much simpler and easier to evaluate. In this paper, we will
also study the speed of convergence to equilibrium for G.
2 The ˛-SP+ Semigroup and Its Associated gL Semigroup
2.1 Spectral Expansions and Convergence to Equilibrium
We start by introducing a few notations. For x 2 RC , let

n
X
n

.˛/kŠ k
P n .x/ D .1/ k k
x; (8)
kD0

.˛k C ˛/
1
.xn .x//.n/ 1 X .1/kC1 sin. k
˛ /
.n C
k
˛ C 1/ k
W n .x/ D D x˛ ; (9)
nŠ nŠ kD0
.˛/kŠ
where, for a smooth function f , f .n/ .x/ D d dxf .x/ Wn .x/

n
n , and we let V n .x/ D .x/ , which,
2

L ./.Finally, we let ˛ to be the largest solution to the equation .1
by [9], is in
1 arcsin./
/ ˛1 cos ˛1 D 12 , and set

.˛ 1/ 1
T˛ D log sin and T˛ ;˛ D max T˛ ; 1 C :
2 ˛
Then we have the following theorem.

Theorem 1
(a) For any f 2 L2 ./, x > 0 and t > T˛ ;˛ , we have in L2 ./,
1
X
Pt f .x/ D ent h f ; V n i P n .x/; (10)
nD0
1
X ˝ ˛
Kt f .x/ D .t C 1/n f ı d.tC1/ ; V n P n .x/: (11)
nD0
Particularly, recalling that pm .x/ D xm , we have for any t > 0,
X
m

.˛m C ˛/
Kt pm .x/ D xm C xmn tn : (12)
nD1

.˛.m n/ C ˛/nŠ
(b) There exists C > 0 and an integer k 0 such that for any f 2 L2 ./ and
t > T˛ ;˛ ,
s
.k/
k 1
kPt f f k C 2 2 kf f k : (13)
e2.tT˛ ;˛ / 1
Moreover, when f D pm , we have, for all t > 0,

.˛m C ˛/ t
kPt pm pm k e kpm epm ke ; (14)

.˛/.mŠ/2
where e.x/ D ex ; x > 0.

Remark 1 Note that (12) indeed coincides with the expression of the entire
moments given by Bertoin and Yor [1, Proposition 1], where the form of as in
their theorem is given by (1).
0
Proof We first observe that limu!1 u˛ .u/ D limu!1 u˛

.˛uC˛/
.˛u/
D C˛ for
0
some C˛ > 0, therefore the spectral expansion (10) follows from [9, Theo-
rem 1.9(1)]. Next, by applying a simple deterministic time change on (2), we get
that for any f 2 L2 ./ and t > 0,
Kt f .x/ D Plog.tC1/ f ı d.tC1/ .x/;
which further yields (11) from (10). Moreover, by [9, Proposition 9.3], the Mellin
transform of W n is, for s 2 CC ,
.1/n
.s/ .1/n
.˛s/
M Wn .s/ D M .s/ D : (15)

.n C 1/
.s n/
.˛/
.n C 1/
.s n/
Hence taking f D pm , we have
.1/n .t C 1/m
.˛m C ˛/
hpm ı dtC1 ; V n i D .t C 1/m M Wn .m C 1/ D ;

.˛/
.n C 1/
.m C 1 n/
which, since
has a pole at any negative integers, vanishes when n m C 1.
Then (11) becomes a finite series and easy algebra yields the representation (12).
For the speed of convergence, (13) comes from [9, Theorem 1.9(3c)]. Moreover, by
[9, Theorem 2.9], P intertwins with Q D .Qt /t0 , the classical Laguerre semigroup,
which is self-adjoint on L2 .e/ with invariant measure e.dx/. More specifically, there
exists a multiplicative kernel such that Pt f D Qt f for any f 2 L2 .e/, and we
.˛/.mŠ/2
have for all m 2 N, pm .x/ D Qm mŠ .k/ pm .x/ D
.˛mC˛/ pm .x/. Hence we see that

kD1 k

.˛mC˛/
1 pm .x/ D p .x/

.˛/.mŠ/2 m
and by [9, Theorem 1.9(3a)], we have

.˛m C ˛/ t
kPt pm pm k k1 pm e1 pm ke D e kpm epm ke

.˛/.mŠ/2
which completes the proof. t

u
2.2 Numerical Illustrations
In this section, we first provide the plots of em .N/, the relative truncation error
of (12) up to N terms which is formally defined by
ˇP ˝ ˛ ˇ
ˇ m n
f ı d.tC1/ ; W n P n .x/ ˇˇ
ˇ nDNC1 .t C 1/
em .N/ D ˇ Pm ˝ ˛ ˇ;
ˇ nD0 .t C 1/
n f ı d
.tC1/ ; W n P n .x/ ˇ
for N running from 1 up to m 1 > 0. Figures 1, 2, and 3 depict the plots for
.m D 20; x D 1/; .m D 40; x D 1/; .m D 40; x D 5/, respectively. Each
plot consists of three experiments ˛ D 1:2; 1:5; 1:8., and in each figure, the left
panel plots log em .N/ with respect to N while the right panel plots em .N/. It can be
seen from the plots that em .N/ exhibits a super-exponential decay, and the relative
error becomes negligible within only a few terms. Comparing Fig. 1 with Fig. 2,
we observe that keeping the value of x constant, the convergence is slower for
larger m (for example, 6 terms are sufficient for a fairly good convergence when
m D 20, but not when m D 40), which is indeed expected since there are more
non-zero terms in the series summation (12) when m gets large. Comparing Fig. 2
with Fig. 3, we also observe that the rate of convergence among different values
of ˛’s depends on the value of x. When x D 1, ˛ D 1:2 converges the fastest
and ˛ D 1:8 converges the slowest. But when x D 5, ˛ D 1:5 converges the
fastest but ˛ D 1:2 becomes the slowest. Hence the relation of convergence rate
Log Truncation Error m = 20, x = 1 Truncation Error

10 3
α = 1.2 α = 1.2
α = 1.5 α = 1.5
α = 1.8 α = 1.8
0
2.5
−10
2
loge(N)
−20
e(N)
1.5
−30
1
−40
0.5
−50
−60 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10
Number of terms N Number of terms N
Fig. 1 Truncation error plots for m D 20; x D 1; ˛ D 1:2; 1:5; 1:8

20 40
α = 1.2 α = 1.2
α = 1.5 α = 1.5
α = 1.8 35 α = 1.8
0
30
−20
25
log e(N)
−40 e(N)
20
−60
15
−80
10
−100 5
−120 0
0 5 10 15 20 25 30 35 40 0 2 4 6 8 10

20 3.5
α = 1.2 α = 1.2
α = 1.5 α = 1.5
α = 1.8 α = 1.8
0 3
−20 2.5
log e(N)
−40 2
e(N)
−60 1.5
−80 1
−100 0.5
−120 0
0 5 10 15 20 25 30 35 40 0 2 4 6 8 10

Log Invariance Error for α-SP+, m = 20 x 10

−4
Invariance Error
−8 1.2
α = 1.2 α = 1.2
α = 1.5 α = 1.5
α = 1.8 α = 1.8
−10
1
−12
0.8
log (t)
−14
(t)
0.6
−16
0.4
−18
0.2
−20
−22 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6
t t
Fig. 4 Invariance error plots for m D 20; ˛ D 1:2; 1:5; 1:8
Log Invariance Error for α-SP+, m = 40 x 10

−8
Invariance Error
−18 1.2
α = 1.2 α = 1.2
α = 1.5 α = 1.5
−20 α = 1.8 α = 1.8
1
−22
0.8
−24
log (t)
(t)
−26 0.6
−28
0.4
−30
0.2
−32
−34 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6
t t
Fig. 5 Invariance error plots for m D 40; ˛ D 1:2; 1:5; 1:8
with respect to ˛ is inconclusive. Next, we provide some numerical plots on the

speed of convergence to equilibrium for the semigroup P. Figures 4 and 5 depict
the plots for .t/ D kPkpt pmmep
pm k
m ke
with m D 20 and m D 40, respectively. Each plot
consists of three experiments ˛ D 1:2; 1:5; 1:8, and each experiment runs for t D 1
to 10. In each figure, the left panel plots log .t/ while the right panel plots .t/.
From the figures, it is easy to observe exponential decrease for each case. Under the
same conditions, by comparing ˛ D 1:2; 1:5; 1:8, we can see that the convergence
becomes faster as ˛ increases. Moreover, by comparing m D 20 with m D 40, we
also observe that the convergence rate increases as m grows, which can be attributed
to the fact that mC1 is indeed the number of non-zero terms in the series summation.
In other words, the convergence is faster using more terms for evaluation.
3 Convergence to Equilibrium of the Gauss-Laguerre

Semigroup
The spectral expansion of G can be found in [7] and in this section, we only state
the speed of convergence to equilibrium as in the following theorem.
Theorem 2 Denoting T˛ D log.2˛1 1/, there exists a constant C˛ > 0 such
that for all f 2 L2 .e˛;ˇ / and t > T˛ ,
r
1
kGt f e˛;ˇ f ke˛;ˇ C˛ k f e˛;ˇ f ke˛;ˇ : (16)
e2.tT˛ / 1
In particular, when f D pm , we have for all t > 0,

..˛ 1/.m C ˇ/ C 1/ t
kGt pm e˛;ˇ pm ke˛;ˇ e kpm epm ke : (17)
mŠ
..˛ 1/ˇ C 1/
Proof By [9, Example 3.3], G corresponds to a spectrally negative Lévy process

˛;ˇ with Laplace exponent ˛;ˇ .u/ D u
..˛1/.uCˇ1/C1/
..˛1/.uCˇ/C1/
, hence (16) follows from
[9, Theorem 1.9 (3b)]. Moreover, as in the previous case, G also intertwins with the
classical Laguerre semigroup Q via an intertwining kernel denoted by ˛;ˇ , i.e. we
have Gt ˛;ˇ f D ˛;ˇ Qt f for all f 2 L2 .e/. Furthermore, by [7, Proposition 3.2] and
its proof, we have for all m 2 N,
mŠ
..˛ 1/ˇ C 1/
˛;ˇ pm .x/ D pm .x/:

..˛ 1/.m C ˇ/ C 1/

..˛1/.mCˇ/C1/
Hence 1
˛;ˇ pm .x/ D mŠ
..˛1/ˇC1/ pm .x/ and by [9, Theorem 1.9(3a)], we have

..˛ 1/.m C ˇ/ C 1/ t
kGt pm e˛;ˇ pm ke˛;ˇ et k1 1
˛;ˇ pm e˛;ˇ pm ke D e kpm epm ke :
mŠ
..˛ 1/ˇ C 1/
This completes the proof of this proposition. t

u
Now we proceed to give the numerical plots for the convergence of Gt pm to
kGt pm e˛;ˇ pm ke
˛;ˇ
e˛;ˇ pm . Figures 6 and 7 depict the plots .t/ D kpm epm ke
with m D 20
and m D 40 respectively. Each plot consists of three experiments .˛; ˇ/ D
.1:2; 4/; .1:5; 1/; .1:8; 0/ and each experiment runs for t D 1 to 10. In each
figure, the left panel plots log .t/ while the right panel plots .t/.
It is easy to observe that in all cases, we still have exponential decay, which
verifies our bound (17). Moreover, we still observe faster convergence when m gets
larger, which can be explained exactly as the previous ˛-SP+ example since for
larger value of m, there are more terms used in the series evaluation.
Log Invariance Error for Gauss-Lauguerre semigroup m = 20 x 10

14
Invariance Error
36 7
α = 1.2, β = −4 α = 1.2, β = −4
α = 1.5, β = −1 α = 1.5, β = −1
34 α = 1.8, β = 0 6 α = 1.8, β = 0
32 5
log (t)
30 4
(t)
28 3
26 2
24 1
22 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6
t t
Fig. 6 Invariance error plots for m D 20; .˛; ˇ/ D .1:2; 4/; .1:5; 1/; .1:8; 0/
Log Invariance Error for Gauss-Lauguerre semigroup m= 40 x 10

40
Invariance Error
94 3.5
α = 1.2, β = −4 α = 1.2, β = −4
α = 1.5, β = −1 α = 1.5, β = −1
92 α = 1.8, β = 0 α = 1.8, β = 0
3
90
2.5
88
log (t)
86 2
84
(t)
1.5
82
1
80
0.5
78
76 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6
t t
Fig. 7 Invariance error plots for m D 40; .˛; ˇ/ D .1:2; 4/; .1:5; 1/; .1:8; 0/
Acknowledgements The authors would like to thank an anonymous referee for insightful
suggestions which improved an earlier version of the paper.
References
1. Bertoin, J., Yor, M.: On the entire moments of self-similar Markov processes and exponential
functionals of Lévy processes. In: Annales de la faculté des sciences de Toulouse 6e série, tome
11, pp. 33–45 (2002)
2. Caballero, M.E., Chaumont, L.: Conditioned stable Lévy processes and the Lamperti represen-
tation. J. Appl. Probab. 43(4), 967–983 (2006)
3. Govorun, E.N., Ivanov, V.A., Khokhlov, A.R., Khalatur, P.G., Borovinsky, A.L., Grosberg,
A.Y.: Primary sequences of proteinlike copolymers: Lévy–flight-type long-range correlations.
Phys. Rev. E 64, 040903–040903-4 (2001)
4. Kolmogorov, A.N.: Dissipation of energy in isotropic turbulence. Dokl. Akad. Nauk SSSR 32,
19–21 (1941)
5. Lamperti, J.: Semi-stable Markov processes. I. In: z. Wahrscheinlichkeitstheorie verw. Geb. 22,
pp. 205–225. Springer (1972)
6. Posadas, A., Morales, J., Vidal, F., Sotolongo-Costa, O., Antoranz, J.C.: Continuous time
random walks and South Spain seismic series. J. Seismol. 6, 61–67 (2002)
7. Patie, P., Savov, M.: Cauchy problem of the non-self-adjoint Gauss-Laguerre semigroups and
uniform bounds of generalized Laguerre polynomials (2013, Submitted)
8. Patie, P., Simon, T.: Intertwining certain fractional operators. Potent. Anal. 36, 569–587 (2012)
9. Patie, P., Savov, M.: Spectral expansions of non-self-adjoint generalized Laguerre semigroups
(2015, Submitted)
10. Yang, H., Zhang, L.: Spectrally negative Lévy processes with applications in risk theory. Adv.
Appl. Prob. 33, 281–291 (2001)
Index
Absorbing boundaries, 22 Apparent permeability, 5

Abstract Cauchy problem, 653 Apparent velocity (convective flux), 7
Accelerated life testing experiments, 585 Aquaponic agriculture, 189
Acoustics, 481 Asymptotic stability, 121, 130
Acute Bee Paralysis Virus (ABPV), 299 Augmented body, 48
Adaptive error control, 448 Augmented gyrostat, 60
Adaptive mesh refinement, 461 Average absolute calibration error, 566
Adjusted volatility swap strike, 569 Average dwell time, 355, 381, 746
Adjusting the Heston drift, 565 Axial fan, 401–403, 407, 408
Advection, 469
Advection-diffusion PDE, 4
Aeroacoustics, 505 Backward time, 159
Aerodynamics, 493 BACOL, 460–462, 467
Aerosol, 89–91, 95, 97, 98 BACOLI, 447–451, 453, 461, 462, 467
Airfoil, 505 Banach fixed point theorem, 704
Air volume fraction, 284 Barotropic vorticity equation, 37
Algebraic derivative array Basic reproduction number, 255

formulation for differential-algebraic BEM, 493
equations, 755 Bessel function, 17
regularization of differential-algebraic Beta-plane, 37
equations via
s, 755 BGK collision approximation, 470
Algorithmic (or automatic) differentiation Bidomain models, 209
(AD), 425 Binary set, 691
Almost block diagonal, 460 Biofilm, 267
ALT with step-stress plan, 585 Biofilm detachment rate, 267
Analog system, 701 Biot-Savart Law, 493
Analysis of differential-algebraic equations Birth, 291, 293–295
signature method, 749 Blade Element Momentum Theory, 493
Analytic description of the polyhedron, 692 Bounce-back, 473
Analytic descriptions of the binary set, 699 Boundary conditions, 18, 447, 448, 450–453
Analytics, 233 Dirichlet, 18
Anomalous diffusion, 15 Neumann, 18
ANSYS Fluent, 280 Brain mechanics, 224
Anti-monotone, 602 Branch and Bound Polyhedral-Spherical
Apparent diffusivity, 7 Method, 690

Modern Science and Engineering, DOI 10.1007/978-3-319-30379-6
800 Index
Brass instruments, 481 Consistent point, 715

Brownian motion process, 369 Constraints of minimum number of failures,
Brusselator, 641–643, 647, 648 591
B-spline, 459–462, 467 Continuous approaches, 689
Bubble column, 280 Continuous dependence on modeling, 654, 657
Bubble injection, 277 Continuous dependence on parameter, 629
Bubble plumes, 278 Continuous time random walk, 15
Bubble velocity, 280, 283 Continuous-time GARCH model for stochastic
Burner, 89–91, 93–98 volatility with delay, 563
Contractivity, 617
Control Theory, 687
Cardiac electrophysiology, 209 Control variate method, 557
Cauchy-Riemann operator, 658 Conversion
Cell model, 447, 448, 452–456 to explicit ODE, 722
Cell-task method, 133–136, 138, 139 to implicit ODE, 720
Cellular neural networks, 775 Convex differentiable extension of function
Cerebrospinal fluid dynamics, 224 from a set, 691
CFD, 389, 390, 392, 393, 395, 398, 505 Convex extension, 693
CFD modelling, 277 Convexity adjustment, 569
Chain of gyrostats, 59 Coordinate change, 142
Chain of rigid bodies, 48 Coupled lattice Boltzmann, 209
Chandrasekhar, 161 Coupled nonlinear parabolic/elliptic PDEs, 209
Change of numeraire technique, 552 Covariance and correlation swaps, 571
Channel radius, 4 Covariance matrix, 141
Chaos game, 618 Critical pressure, 6
Chaotic Lur’e system, 725 Critical temperature, 6
Characteristic function, 553 C0 -semigroup, 121
Chaste, 447, 453–456 Cytokine, 331
Chemostat, 267
Cholera, 290
Circle inversion, 609 DAE, 450, 451, 665, 666, 714
Circle inversion map, 610 consistent point, 715
Circulation, 277, 285 semi-explicit index-1, 721
Circumscribed surface, 699 Damping matrix, 101
Civan’s transport model, 7 Darcy’s Law, 7
Classical Heston model, 565 DASSL, 461, 467, 722
Classical solution, 654 Data assimilation, 141
Click graph, 602 Davey-Stewartson system, 27
Closed formulas for variance and volatility DD scheme, 719
swaps price processes, 566 DD-specification vector, 714, 719
Co-infection, 290, 291, 293, 295–297 DDspec, 719
Coalition attacks, 597 Death, 291, 293–296
Collage Theorem, 115 Defaultable bond, 526
Collision, 469 Deformation, 141
Collocation, 447, 449–451, 454, 455 Degrees of freedom (DOF), 714, 716
COLROW, 460 Delay, 561, 565
Combinatorial algorithms, 690 Delayed Heston model, 561, 565
Combinatorial set, 694 Delayed Heston model for variance, 566
Combinatorial sets inscribed into a sphere, 691 Delayed stochastic volatility, 563
Combustion, 89–93, 96–98, 389–392, 398 Delayed-adjusted long-range variance, 568
Competition instabilities, 641 Delayed-adjusted mean-reverting speed, 568
Compound laser modes, 83 Delta function, 24
Compressibility factor, 4 Demand for pollution control, 341
Conformal invariance, 145 Demand markets, 538
Index 801
Derivative based global sensitivity measures, Error control, 460–462, 465, 467
194 Escape velocity, 163
Detailed chemical kinetic, 89, 93 Escherichia coli, 245
DGM, 481 ESI-CyDesign, 722
Differential algebraic system, 641, 643 Euler method, 556
Differential-algebraic equations (DAEs), 461, Euler-Bernouli, 101
763 Eulerian approach, 280
dummy derivatives, 750 Eutrophication, 277
dynamic state selection, 755 Evolutionary variational inequality problems,
hidden constraints, 749 541
numerical integration of
, 749 Exponential, 521
numerical integration of
with Exponential stability, 355, 368
QUALIDAES, 757 Extended BACOLI, 447, 448, 452–456
regularization of
, 749
structural index, 751
Differentiation index, 665 FDE, 153
Direct cover condition, 430 Feedback control, 725
Discontinuous Galerkin, 435 Fekete point sets, 641, 642, 646, 647
Discrete Boltzmann equation (DBE), 470 Finite difference scheme, 18
Discrete and continuous time difference explicit, 18
equations, 629 implicit, 18
Discrete logarithmic energy, 642, 646 Fishery, 342, 345, 347
Disease elimination, 256 Fitting model, 588
Disease intervention strategies, 256 Flame, 89–98
Dissipative, 125 Flow conditions, 5
Dual-Rotor, 493 Food-borne illness, 245
Dummy derivative, 665–667, 669, 719 Fortran, 493
scheme, 719 Forward contract, 561
universal, 667, 669, 674 Found vector, 718
Dummy derivatives Fourier spectrum, 40

for differential-algebraic equations, 750 Fox function, 17
regularization of differential-algebraic Fractals, 609
equations via
, 753 Fractional Brownian motion, 571
Dynamic feedback controller, 121 Fractional calculus, 15
Dynamic friction, 153, 161 Fréchet space, 703
Dynamic hedge ratio, 566 Free Vortex, 493
Dynamic models, 224 FRPM, 691
Dynamic state selection for differential- Fuel-injector, 89, 96, 97
algebraic equations, 755 Functional differential equation, 153
Functional representation of a discrete set, 689
Eigenvalues, 123
Electrophysiology, 447, 448 Galois group, 621
Elliptic PDEs, 113 Gas phase, 280
˙-method, 713 Gaussian collocation, 459–461
standard solution scheme, 715 Gaussian interest rate model, 551
system Jacobian, 713 Geography, 241
Entrainment, 153 Geometric configurations, 279
Environmental Monitoring, 233 Geometrically complex, 469
EPCA, 367 Geometry of similar motions of Hess tops, 54
EPDCOL, 460 Global relative sensitivity, 195
Epidemic model, 256 Globally exponentially stable, 778
Equilibrium, 291, 292 Global solution on a sphere, 692
Equivalent martingale measure, 552 GPR (Gain per Resource), 600
802 Index
GPR-Core property, 602 Index-pair, 716

GPU, 435 Infinitesimal generator, 653
Gravitation, 153 Injector, 89, 91, 95, 96, 98
Gravitational current, 158 Integration
Gravitational field, 154 numerical
of differential-algebraic
Gray-Scott, 641–643, 648 equations, 749
Green’s function, 643, 649 numerical
of differential-algebraic
Green’s matrix, 644, 645 equations with QUALIDAES, 757
Ground beef, 245 Interface, 402, 405–408
Grunwald-Letnikov derivative, 18 Interfacial forces, 280, 281, 283
Gyrostat, 59, 60 Interferon, 331
Intersecting functional representation, 698
Intracranial pressure, 223
H-adaptivity, 435 Intrinsic permeability, 5
H1 , 367 Invalid traffic, 595
Harmonic maps, 143 Inverse Laplace transform, 17
Hausdorff distance, 617 Inverse problem, 102, 113
Hazard rate estimator, 590 Ionic models, 209
Heat transfer, 389, 390 Irreducible representation, 697
Hedge ratio, 569 Isotropy subgroup, 83
Hedge volatility swap, 565 ISS, 367, 370
Helmholtz Theorem, 493 Itô integral, 369
Helminths, 289 Itô Lemma, 370
Hess solution, 51 Item, 718
Hess top, 51 algebraic, 719
Heston model, 549, 566 found, 718
Hidden constraints, 665, 667 loop-closer, 719
Hidden constraints of differential-algebraic state, 718
equations, 749
HIV, 290
AIDS, 290 Jacobian matrix, 425
Holomorphic semigroup, 653 Jordan triples, 103
Homoclinic orbit, 27
Homology, 235
Honeybee, 299 Kelvin-Voigt damping, 121
Hopf bifurcation, 224 Knudsen number, 4
Human-environment system, 341, 343, 349 Kozeny-Carman relation, 5
Humans, 290, 291, 293, 294
HVT, 666, 668, 715
Hybrid control, 739 Lévy process, 565
Hybrid parallelization, 139 Lévy-based stochastic volatility with delay,
Hybrid system, 367, 687, 739 565
Hyperbolic conservation laws, 435 Lagmuir isotherm, 6
Lagrange inversion, 162
Lang-Kobayashi rate equations, 81
Ill-posed problem, 653 Laplace transform, 16
Immunity, 290, 291, 293 Laplace-Beltrami, 643
Impingement, 89, 95–98 Lattice Boltzmann method (LBM), 469
Implicit ODE, 714, 716 Law of total probability, 15
Impulse extension equation, 199 Lax-Wendroff scheme, 470
Impulsive switched singular systems, 355 Least square method, 555
Index (differentiation), 714 Levy-based stochastic volatility model, 571
Index for differential-algebraic equations Liénard-Wiechert potentials, 153
structural
, 751 Lifting Line Theory (LLT), 493
Index 803
Linear matrix inequality, 725 Negative false alarms, 603

Linear problem over a discrete set, 692 Neutral differential delay equations, 629
Liquid phase, 280 Newton’s law, 156
Local radial point interpolation method Next-generation matrix, 256
(LRPIM), 471 Next-generation method, 291, 292
Logarithm, 521 Non-coalition attacks, 597
Loop-closer, 719 Non-gray, 389, 390, 393, 398
Lotka-Volterra, 342 Non-Markovian continuous-time GARCH
Lower bound, 427 model, 567
Lower order interpolant, 462 Nonlinear, 343, 345
Lyapunov function, 367 Nonlinear wave propagation, 481
Lyapunov method, 355 Normalization matrix, 102
Lyapunov-Krasovkii functional, 725 NREL Phase VI, 493
Numerical integration of differential-algebraic
equations, 749
Magma, 623
with QUALIDAES, 757
Malaria, 289–297
Markets of Optional Processes, 523
Martingale, 568 Offsets, 715
Mass matrix, 101 canonical, 666, 668, 715
Mathematical model, 189, 225 valid, 668, 672
Mathlab Toolbox, 194 Online advertising, 595
Matlab code, 19 Optimal robust designs, 587
Mean free path, 4 Optional martingale deflator, 524
Mellin transform, 17, 789 Optional processes, 520
inverse, 17 Optional semimartingale, 520
Meshless, 469 Option pricing, 549
Meshless lattice Boltzmann method Orbitally stable, 646
(MLLBM), 479 Ordinary differential equation model, 313
Meshless local Petrov-Galerkin (MLPG), 470 Ordinary differential equations (ODE), 189,
Metabolic heat generation rate, 168 450, 451, 665
Metapopulation model, 256 Orthogonal matrix, 103
Micro-jet, 89, 95–98 Overdetermined formulation
Minimal circumscribed sphere, 692 regularization of differential-algebraic
Mixed functional representation, 698 equations via
s, 756
Modelling, 290
SIR model, 290
Molecular-dynamics, 133 Parabolic, 447, 456
Momentum exchange, 280 Paraffin-oil, 89–91, 93–98
Monodomain model, 447–449, 452, 453, 456 Parallel speedup, 135
Monte Carlo method, 556 Parameter estimation, 113
Monte Carlo sampling methods, 194 Particle distribution function, 470
Mosquitoes, 290, 291, 293, 294, 296 Past history, 565, 567
Multi-factor stochastic volatility with delay, Path-dependent history, 561
563 Pathogen die-off, 309
Multi-scale, 447–449, 452 Payoff of variance swap, 562
Multiple Lyapunov functions, 355, 379, 380, Payoff of volatility swap, 562
387 PDE, 447, 448, 456
Multi-scale system, 209 PDECOL, 460
Penalty method based on functional
representations, 691
Naive volatility swap strike, 569 Pendulum, 751
Nano-particle, 89, 93 Pennes equation, 168
Nearly incompressible flows, 469 Perforated domain, 113
804 Index
Periodic boundary condition, 401, 402, 408 Regularization of differential-algebraic

Periodic solutions, 629 equations, 749
Permanent rotation of gyrostat, 61
via algebraic derivative arrays, 755
Permeability, 476
via dummy derivatives, 753
Persistent homology, 235
via overdetermined formulation, 756
Perturbation expansion, 162
via state selection, 753
Pinning impulsive control, 777 algebraic derivative array formulation, 755
Plot initial hedge ratio, 569 regularized overdetermined formulation,
Point vortices, 642 757
Poiseuille flow, 474 structurally extended formulation, 753
Polar decomposition, 102 Regularized overdetermined formulation for
Pollutant, 89, 90, 98 differential-algebraic equations, 757
Pollution epidemic, 341–344, 346–349 Relative efficiency, 593
Polyhedral and spherical relaxations, 690 Relative invariant, 622
Polyhedron, 689 Relativity, 153
Population dispersal, 255 Reliability, 462, 465–467
Population dynamics, 189 Reliable control, 380
Porosity, 4 Reliable Controllers, 379, 381, 383, 385, 387
Porosity-pressure correlation, 5 Reproduction number, 291, 292, 294, 295, 297
Porous media, 469 Retarded field, 153
Portfolio, 568 Retracts, 687
Positive false alarms, 603 Riemann-Liouville derivative, 18
Post-Newtonian, 153 Riemannian manifold, 142
Pressure field, 7 Rock density, 3
Price of variance swap, 562, 567 Rock permeability, 3
Price of volatility swap, 562, 567 Rotating electrical machine, conjugate heat
Projected dynamical system, 537 transfer, turbulence modelling,
Projection onto a discrete set, 692 CFD., 413
Proportional hazards models, 587 Rotational domain, 406, 407
Pseudo-spectral numerical method, 39 Rotor disc, 402, 403, 405, 408
PSM, 693 Row-sum lumping, 474
RTE, 389, 390, 392, 394, 395, 398
Quadratic function, 693

SA, 665, 668, 713
QUALIDAES
SA-friendly, 713, 716
numerical integration of differential-
Schistosomiasis, 289–297
algebraic equations, 757
Schnakenberg, 641, 642
Quasi-MLE method, 590
Schrödinger equation, 121
Quasi-reversibility method, 654
Security, 561
Semigroup
˛-SP+, 788
Radial basis functions (RBF), 471 Gauss-Laguerre, 789
Radiation, 389–392, 395, 397, 398 generalized Laguerre, 789
Rarefaction coefficient correlation, 5 self-similar, 788
Razumikhin, 367 series expansion, 791
Razumikhin-type stability criterion, 778 convergence, 792
Reach control problem, 687 stationary distribution, 789
Reaction-diffusion, 641 convergence, 791, 794, 795
Real gas deviation factor, 6 Sensitivity, 291, 293–295, 297
Recovery, 291, 293–295 Sensitivity analysis, 10, 189
Refinement, 435 SensSB, 194
Reflecting boundaries, 23 SEPCA, 367
Regularization, 654, 661 Shale gas, 4
Index 805
Shale rock, 4 Stochastic volatility with delay and jumps, 564

Shape functions, 471 Stock, 561
Short cosets, 624 Strong pathwise duality, 533
Siegmund duality, 530 Strongly predictable, 520
Signature matrix, 668, 670, 715 Structural analysis, 665, 668, 713
method, 665, 666, 713 Structural analysis of differential-algebraic
Signature method for differential-algebraic equations, 763
equations, 749 signature method, 749
Similar motions of rigid bodies’ chain, 50 Structural index for differential-algebraic
Similar-clique property, 602 equations, 751
Similarity graph, 602 Structurally extended formulation for
Simple pendulum, 751 differential-algebraic equations, 753
Singular impulsive system, 358 Succeed, 713, 716
Singular perturbations, 629 Superconvergent interpolant, 462
Singularly perturbed, 379, 380, 641 Supply markets, 538
Singularly perturbed systems, 379, 381, 383, Swarming, 299
385, 387 Switched, 379–381, 383, 385, 387
SIRS model, 255 Switched systems, 367, 737
Skew-symmetric, 108 Synchronization, 725, 740
SLW, 389, 390, 392, 395, 397, 398 System Jacobian, 666, 713
Snails, 290, 291, 293–295
Social learning, 342, 350
Social misperception, 342 Target reproduction number, 256
Social stigmatization, 342, 344 Test function, 471
Software, 447–456 Three Lines Theorem, 660
Soliton, 27 Tight gas, 3
Solving scheme, 718 Time delay, 775
Soot formation, 90, 91, 97 Time evolution problem, 702
Sound propagation, 481 Time-censoring, 585
Spatial error estimate, 461, 462 Time-changed Brownian motion, 568
Spectral analysis, 125 Time-fractional advection-diffusion equation,
Spectral mapping theorem, 108 16
Spherical harmonics, 642 Time-fractional diffusion equation, 16
Spot self-replication, 641 Time-scale tolerance, 203
Spot dynamics, 642, 643, 648, 649 Tip clearance, 402, 404, 408, 410
Spot patterns, 641–643, 649 Topological data analysis, 233
Spot self-replication, 644 Topological obstruction, 687
SSS, see Standard solution scheme Topology, 233
Stabilization, 121 Tortuosity, 5
Stable process, 787 Touching functional representation, 698
Standard solution scheme, 713, 715 Touching set of smooth surfaces, 689
Star bi-coloring, 427 Traffic similarity, 600
State selection Trailing-edge (TE) noise, 505
regularization of differential-algebraic Translation group, 662
equations via
, 753 Transmissibility, 293–295
State variables, 714 Transport model, 4
State vector, 718 Transversal, 666, 672, 715
Stator, 401, 403–405, 407 highest-value, 715
Stauduhar method, 621 Tropical cyclone, 35
Steady state pressure field, 8 True model, 588
Stiffness matrix, 101 Trumpet, 481
Stochastic integral, 521 Tuberculosis, 290
Stochastic monotonicity, 530 Turbulent, 89–93, 95–98
Stochastic volatility, 549 Turing, 642
806 Index
Twisted cuboidal, 645 Varroa destrutor, 299

Two-layer sets, 690 Vasicek model, 549
Two-sided compressions, 427, 431 Vertex located set, 691
Two-stage design, 587, 592 Volatility, 561, 563
Volatility or variance models with delay, 563
Volatility surface fitting, 565
UBQP, 689 Volatility swap, 561, 568
UGR - unconventional gas reservoirs, 7 Volatility swap fair strike, 568
Unconstrained binary quadratic problem, 689 Volatility swap hedging, 568
Unconventional reservoirs, 3, 4 Vortex, 35
Uniform exponential regulator, 202 Vortex filament, 493
Uniform motion, 157 Vortex Rossby wave, 36
Uniformly entire function, 708
Uniformly exponentially regulated in the mean,
202 Water flow streamlines, 282
Unit hypercube, 692 Water quality, 278
Unusual conditions, 520 Water velocity, 281–283
Upward skip-free, 529 Wave mean-flow interaction, 36
Weak form, 471
Well-posedness, 123
Vaccine, 289 Wiener process, 369
Vanilla option price surface, 566 Wind energy, 493
Vanilla options, 569 Wind turbine, 493
Variance and volatility swaps hedging, 566 Witness complexes, 237
Variance swap, 561, 568
Variance swap fair strike, 568
Variational inequality, 539 XFOIL, 493

Mathematical and Computational Approaches in Advancing Modern Science and Engineering

Uploaded by

Copyright:

Available Formats

Mathematical and Computational Approaches in Advancing Modern Science and Engineering

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Mathematical and Computational Approaches in Advancing Modern Science and Engineering

Uploaded by

Copyright:

Available Formats

Jacques

Herb Kunze Roman Makarov

Roderick Melnik Raymond J. Spiteri

ISBN 978-3-319-30377-2 ISBN 978-3-319-30379-6 (eBook)

Library of Congress Control Number: 2016943639

© Springer International Publishing Switzerland 2016

Printed on acid-free paper

This Springer imprint is published by Springer Nature

• Mathematical and Computational Methods in Life Sciences and Medicine

has a strong multidisciplinary focus, supported by fundamental theories, rigorous

Montreal, Canada Jacques Bélair

Part I Theory and Applications of Mathematical Models

Effect of Jet Impingement on Nano-aerosol Soot Formation

Part II Mathematical and Computational Methods in Life

Dynamics and Bifurcations in Low-Dimensional Models

Part III Computational Engineering and Mathematical

A Dual-Rotor Horizontal Axis Wind Turbine In-House Code

Part IV Mathematics and Computation in Finance,

Part V New Challenges in Mathematical Modeling

Global Dynamics and Periodic Solutions in a Singular

Convergence Analysis of the Spectral Expansion of Stable

Iftikhar Ali, Bilal Chanane, and Nadeem A. Malik

Abstract Transport models for gas flow in unconventional hydrocarbon reservoirs

I. Ali () • B. Chanane • N.A. Malik

© Springer International Publishing Switzerland 2016 3

and these must be known as functions of p and px . However, most applications to

2 Physical Properties of Shale Gas Reservoirs

Table 1 Classification of Knudsen number Flow regimes

The correlation between porosity and intrinsic permeability is given by the

where  < ˛KC 1, 0 ˇKC < 1 and

 D a exp.b pc /; (3)

where a , b and c are model constants. Tortuosity is related to porosity by,

where a is also a model constant.

where f .Kn / is the flow condition function given by

f .Kn / D .1 C Kn / .1 C .4 bSF /Kn / .1 bSF Kn /1 ; (6)

where is called the Rarefaction Coefficient Correlation [6] given by

Z D ap2r C bpr C c (10)

 D Sc exp.AB / (14)

The apparent diffusivity D (m2 /s) is given by,

and the apparent convective flux (velocity) U (m/s) is given by,

where the 1 , 2 and 3 appearing in D and U are given by

3 . p/ D Œ . p/ C Ka . p/  . p/ : (20)

where Ka D K C f which is obtained from Eq. (5).

4 Model Validation Under Steady State Conditions

with boundary conditions, p.0/ D pL and p.L/ D pR ; pL and pR assumed known.

where the summation is over the N D 30 data-points in [16]. Case 1 in Table 2

Table 2 List of models Cases K f  Error

D 0. Case 16 is the fully pressure-dependent case. An additional case, from

a Darcy’s Law b Civan (2011)

205 kPa 205 kPa

0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3

200 240 kPa

0 0.5 1 1.5 2 2.5 3 0 5 10 15

Table 3 Reservoir model Parameter Parameter

Table 4 The range of Cases ˛KC ˇKC

K D 1  1015 m2 . From the present calculations the porosity is pressure dependent

5 Sensitivity Analysis and Estimation of Model Parameters

275 kPa 275 kPa

275 kPa 275 kPa

where < ˛KC 1, 0 ˇKC < 1 and

D a exp.b pc /; (3)

where a , b and c are model constants. Tortuosity is related to porosity by,

D Sc exp.AB / (14)

3 . p/ D Œ . p/ C Ka . p/ . p/: (20)

Table 2 List of models Cases K f Error

K D 1 1015 m2 . From the present calculations the porosity is pressure dependent

where k is the wave number, ! is the frequency, m, , , A and B are material

iut C ıuxx C uyy D juj2 u C b.'1;x C '2;y /u;

so .1 /.m1 m2 / D n2 , m1 > m2 and < 1. It should be also noted that

i"t C "xx C "yy D jaj2 ." C " / b.'1" C '2" /;

".x; y; t/ D AŒei.n xCNn y/Cn t C BŒei.n xCNn y/Cn t ;

2jaj2 .m1 1/

G ia2 t 2.1 ı/ 2.1 ı/

2t 1 C 2b˙ cos. p1 x C p2 y/e˝˙ tC C b2˙ e2.˝˙ tC / sec2

˙ 4.ı 1/p21 2e˝˙ tC C cos. p1 x C p2 y/.1 C e2.˝˙ tC / sec2 /

˙ 4.ı 1/p22 2e˝˙ tC C cos. p1 x C p2 y/.1 C e2.˝˙ tC / sec2 /

where b˙ D cos.2/ sin.2/.

.r; ; t/ D N .r/ C " .r; ; t/; (4)

This equation is examined in the annular region r1 r < 1; 0 < 2 shown

and j.r; ; t/j is shown as a function of wavenumber at r D rc D 12 and t D 100.