Mathematics

Advances in Applied Mathematics

FOR
DYNAMICAL SYSTEMS
BIOLOGICAL MODELING
DYNAMICAL SYSTEMS
FOR BIOLOGICAL
MODELING
Dynamical Systems for Biological Modeling: An Introduction pro-
vides both biology and mathematics students with the understanding
and techniques necessary to undertake basic modeling of biological
systems. It achieves this through the development and analysis of
dynamical systems. AN INTRODUCTION
The approach emphasizes qualitative ideas rather than explicit com-
putations. Some technical details are necessary, but a qualitative ap-
proach emphasizing ideas is essential for understanding. The model-
ing approach helps students focus on essentials rather than extensive
mathematical details, which is helpful for students whose primary in-
terests are in sciences other than mathematics.
The book discusses a variety of biological modeling topics, including
population biology, epidemiology, immunology, intraspecies competi-
tion, harvesting, predator–prey systems, structured populations, and
more.
The authors also include examples of problems with solutions and
some exercises that follow the examples quite closely. In addition,
problems are included that go beyond the examples, both in math- Fred Brauer
Brauer • Kribs
ematical analysis and in the development of mathematical models for
biological problems, in order to encourage deeper understanding and
an eagerness to use mathematics in learning about biology. Christopher Kribs

C664X

w w w. c rc p r e s s . c o m

C664X_cover.indd 1 11/9/15 3:23 PM

DYNAMICAL SYSTEMS
FOR BIOLOGICAL
MODELING
AN INTRODUCTION
Advances in Applied Mathematics

Series Editor: Daniel Zwillinger

Published Titles
Green’s Functions with Applications, Second Edition Dean G. Duffy
Introduction to Financial Mathematics Kevin J. Hastings
Linear and Integer Optimization: Theory and Practice, Third Edition
Gerard Sierksma and Yori Zwols
Markov Processes James R. Kirkwood
Pocket Book of Integrals and Mathematical Formulas, 5th Edition
Ronald J. Tallarida
Stochastic Partial Differential Equations, Second Edition Pao-Liu Chow
Dynamical Systems for Biological Modeling: An Introduction
Fred Brauer and Christopher Kribs
 Advances in Applied Mathematics

DYNAMICAL SYSTEMS
FOR BIOLOGICAL
MODELING
AN INTRODUCTION

Fred Brauer
University of British Columbia
Vancouver, British Columbia, Canada

Christopher Kribs
University of Texas at Arlington
Arlington, Texas, USA
Back cover image credit: (Photo of healthcare workers) Cleopatra Adedeji, CDC PHIL public domain
collection.

CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2016 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business

No claim to original U.S. Government works

Version Date: 20151221

International Standard Book Number-13: 978-1-4987-7404-8 (eBook - PDF)

This book contains information obtained from authentic and highly regarded sources. Reasonable
efforts have been made to publish reliable data and information, but the author and publisher cannot
assume responsibility for the validity of all materials or the consequences of their use. The authors and
publishers have attempted to trace the copyright holders of all material reproduced in this publication
and apologize to copyright holders if permission to publish in this form has not been obtained. If any
copyright material has not been acknowledged please write and let us know so we may rectify in any
future reprint.

Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced,
transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or
hereafter invented, including photocopying, microfilming, and recording, or in any information stor-
age or retrieval system, without written permission from the publishers.

For permission to photocopy or use material electronically from this work, please access www.copy-
right.com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222
Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that pro-
vides licenses and registration for a variety of users. For organizations that have been granted a photo-
copy license by the CCC, a separate system of payment has been arranged.

Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are
used only for identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
to our children and grandchildren
This page intentionally left blank
Contents

Preface xi

Acknowledgments xiii

I Elementary Topics 1
1 Introduction to Biological Modeling 3
1.1 The nature and purposes of biological modeling . . . . . . . 3
1.2 The modeling process . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Types of mathematical models . . . . . . . . . . . . . . . . . 11
1.4 Assumptions, simplifications, and compromises . . . . . . . . 14
1.5 Scale, and choosing units . . . . . . . . . . . . . . . . . . . . 17

2 Difference Equations (Discrete Dynamical Systems) 23

2.1 Introduction to discrete dynamical systems . . . . . . . . . . 23
2.1.1 Linear difference equations . . . . . . . . . . . . . . . 24
2.1.2 Solution of linear difference equations . . . . . . . . . 26
2.1.3 Nonlinear difference equations . . . . . . . . . . . . . . 28
2.2 Graphical analysis . . . . . . . . . . . . . . . . . . . . . . . . 33
2.3 Qualitative analysis and population genetics . . . . . . . . . 37
2.3.1 Linearization and local stability . . . . . . . . . . . . . 37
2.3.2 A problem in population genetics . . . . . . . . . . . . 42
2.4 Intraspecies competition . . . . . . . . . . . . . . . . . . . . 50
2.4.1 Two metered fish models . . . . . . . . . . . . . . . . 54
2.4.2 Between contest and scramble . . . . . . . . . . . . . . 56
2.5 Harvesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
2.5.1 Fishery harvesting and graphical equilibrium analysis 60
2.6 Period doubling and chaos . . . . . . . . . . . . . . . . . . . 69
2.6.1 Dispersal . . . . . . . . . . . . . . . . . . . . . . . . . 75
2.6.2 Dynamical diseases and physiological control systems 78
2.7 Structured populations . . . . . . . . . . . . . . . . . . . . . 82
2.7.1 Spatial dispersal . . . . . . . . . . . . . . . . . . . . . 86
2.7.2 Two-stage populations . . . . . . . . . . . . . . . . . . 90
2.8 Predator-prey systems . . . . . . . . . . . . . . . . . . . . . . 93
2.8.1 A plant-herbivore model . . . . . . . . . . . . . . . . . 93
2.8.2 A host-parasitoid model . . . . . . . . . . . . . . . . . 94
Miscellaneous exercises . . . . . . . . . . . . . . . . . . . . . . . . 97

vii
viii Contents

3 First-Order Differential Equations (Continuous Dynamical

Systems) 99
3.1 Continuous-time models and exponential growth . . . . . . . 99
3.1.1 Exponential growth . . . . . . . . . . . . . . . . . . . 100
3.1.2 Radioactive decay . . . . . . . . . . . . . . . . . . . . 105
3.2 Logistic population models . . . . . . . . . . . . . . . . . . . 108
3.2.1 All creatures great and small? . . . . . . . . . . . . . . 110
3.2.2 Competition among plants . . . . . . . . . . . . . . . 114
3.2.3 The spread of infectious diseases . . . . . . . . . . . . 117
3.3 Graphical analysis . . . . . . . . . . . . . . . . . . . . . . . . 123
3.3.1 Solutions . . . . . . . . . . . . . . . . . . . . . . . . . 123
3.3.2 Direction fields . . . . . . . . . . . . . . . . . . . . . . 126
3.4 Equations and models with variables separable . . . . . . . . 130
3.4.1 A linear model for the cardiac pacemaker . . . . . . . 131
3.4.2 General procedure . . . . . . . . . . . . . . . . . . . . 135
3.4.3 Solution of logistic equations . . . . . . . . . . . . . . 139
3.4.4 Discrete-time metered population models . . . . . . . 141
3.4.5 Allometry . . . . . . . . . . . . . . . . . . . . . . . . . 144
3.5 Mixing processes and linear models . . . . . . . . . . . . . . 150
3.5.1 Chemostats . . . . . . . . . . . . . . . . . . . . . . . . 152
3.5.2 Drug dosage . . . . . . . . . . . . . . . . . . . . . . . . 155
3.5.3 Newton’s law of cooling . . . . . . . . . . . . . . . . . 158
3.5.4 Migration . . . . . . . . . . . . . . . . . . . . . . . . . 160
3.6 First-order models with time dependence . . . . . . . . . . . 165
3.6.1 Superposition . . . . . . . . . . . . . . . . . . . . . . . 165
3.6.2 Integrating factors . . . . . . . . . . . . . . . . . . . . 167
3.6.3 Substitution and integration . . . . . . . . . . . . . . . 169
3.6.4 Mixing processes with variable coefficients . . . . . . . 171
3.6.5 Bernouilli equation . . . . . . . . . . . . . . . . . . . . 174
Miscellaneous exercises . . . . . . . . . . . . . . . . . . . . . . . . 182

4 Nonlinear Differential Equations 185

4.1 Qualitative analysis tools . . . . . . . . . . . . . . . . . . . . 185
4.1.1 Possible end behaviors . . . . . . . . . . . . . . . . . . 187
4.1.2 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . 191
4.1.3 Phase portraits . . . . . . . . . . . . . . . . . . . . . . 196
4.1.4 Foraging ants and phase transitions . . . . . . . . . . 199
4.2 Harvesting . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
4.2.1 Constant-yield harvesting . . . . . . . . . . . . . . . . 203
4.2.2 Constant-effort harvesting . . . . . . . . . . . . . . . . 214
4.2.3 Migration and dispersal as harvesting . . . . . . . . . 222
4.2.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . 224
4.3 Mass-action models . . . . . . . . . . . . . . . . . . . . . . . 228
4.3.1 A simple chemical reaction . . . . . . . . . . . . . . . 228
4.3.2 The spread of infectious diseases, revisited . . . . . . . 231
 Contents ix

4.3.3 Contact rate saturation and the “Pay It Forward” model 237
4.4 Parameter changes, thresholds, and bifurcations . . . . . . . 242
4.4.1 Hysteresis . . . . . . . . . . . . . . . . . . . . . . . . . 249
4.4.2 The spruce budworm . . . . . . . . . . . . . . . . . . . 251
4.5 Numerical analysis of differential equations . . . . . . . . . . 258
4.5.1 Approximation error . . . . . . . . . . . . . . . . . . . 259
4.5.2 Euler’s method . . . . . . . . . . . . . . . . . . . . . . 260
4.5.3 Other numerical methods . . . . . . . . . . . . . . . . 264
4.5.4 Eutrophication . . . . . . . . . . . . . . . . . . . . . . 270
Miscellaneous exercises . . . . . . . . . . . . . . . . . . . . . . . . 277

II More Advanced Topics 281

5 Systems of Differential Equations 283
5.1 Graphical analysis: The phase plane . . . . . . . . . . . . . . 283
5.2 Linearization of a system at an equilibrium . . . . . . . . . . 291
5.3 Linear systems with constant coefficients . . . . . . . . . . . 297
5.3.1 A liver chemistry example . . . . . . . . . . . . . . . . 305
5.4 Qualitative analysis of systems . . . . . . . . . . . . . . . . . 311
Miscellaneous exercises . . . . . . . . . . . . . . . . . . . . . . . . 319

6 Topics in Modeling Systems of Populations 321

6.1 Epidemiology: Compartmental models . . . . . . . . . . . . . 321
6.1.1 An epidemic model . . . . . . . . . . . . . . . . . . . . 321
6.1.2 A model for endemic situations . . . . . . . . . . . . . 327
6.2 Population biology: Interacting species . . . . . . . . . . . . 331
6.2.1 Species in competition . . . . . . . . . . . . . . . . . . 331
6.2.2 Predator-prey systems . . . . . . . . . . . . . . . . . . 337
6.2.3 Symbiosis . . . . . . . . . . . . . . . . . . . . . . . . . 345
6.3 Numerical approximation to solutions of systems . . . . . . . 351
6.3.1 Example: A two-sex model . . . . . . . . . . . . . . . 352

7 Systems with Sustained Oscillations and Singularities 359

7.1 Oscillations in neural activity . . . . . . . . . . . . . . . . . . 359
7.1.1 The Fitzhugh-Nagumo equations . . . . . . . . . . . . 360
7.1.2 A model for cat neurons . . . . . . . . . . . . . . . . . 363
7.2 Singular perturbations and enzyme kinetics . . . . . . . . . . 366
7.2.1 Bursting . . . . . . . . . . . . . . . . . . . . . . . . . . 372
7.2.2 An example from enzyme kinetics . . . . . . . . . . . 374
7.3 HIV: An example from immunology . . . . . . . . . . . . . . 379
7.3.1 A basic model . . . . . . . . . . . . . . . . . . . . . . 381
7.3.2 Including infected cells . . . . . . . . . . . . . . . . . . 387
7.4 Slow selection in population genetics . . . . . . . . . . . . . . 392
7.4.1 Equally fit genotypes . . . . . . . . . . . . . . . . . . . 393
7.4.2 Slow genetic selection . . . . . . . . . . . . . . . . . . 396
x Contents

7.5 Second-order differential equations: Acceleration . . . . . . . 402

7.5.1 The harmonic oscillator . . . . . . . . . . . . . . . . . 402
7.5.2 The van der Pol oscillator . . . . . . . . . . . . . . . . 405
7.5.3 A model of oxygen diffusion in muscle fibers . . . . . . 408

III Appendices 413

A An Introduction to the Use of MapleTM 415
A.1 Plotting graphs of functions . . . . . . . . . . . . . . . . . . 416
A.2 Graphical solution of first-order differential equations . . . . 417
A.3 Graphical solution of systems of differential equations . . . . 419
A.4 The cobwebbing method for graphical solution of first-order
difference equations . . . . . . . . . . . . . . . . . . . . . . . 420
A.5 Solution of difference equations and systems of difference equa-
tions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 422
A.6 A bifurcation program . . . . . . . . . . . . . . . . . . . . . . 424

B Taylor’s Theorem and Linearization 425

C Location of Roots of Polynomial Equations 427

D Stability of Equilibrium of Difference Equations 429

Answers to Selected Exercises 433

Bibliography 459

Index 467
Preface

Some understanding of difference equations and differential equations is be-

coming essential for students in the biological sciences. There are several recent
texts providing an introduction to mathematical biology through dynamical
systems, but most of these are accessible primarily to students of mathematics
with some interest in applications to biology rather than to biology students
who have not (yet) developed a disposition to describe real systems in mathe-
matical terms. Many of these texts present fully developed models of biological
systems ready for mathematical analysis. Such models provide students with
exercises in applying mathematical techniques but little experience in the
(often iterative) translation of biological concepts into mathematical terms
and vice versa, which is at the heart of modeling. It is our intention to try
to address both of these issues, to prepare students of biology as well as of
mathematics with the understanding and techniques necessary to undertake
basic modeling of biological systems through the development and analysis of
dynamical systems.
We propose to present an introduction to dynamical systems, covering both
discrete (difference equation) and continuous (differential equation) types, for
students who have had an introduction to calculus. While we assume that
students have learned basic material including differentiation, integration, and
exponential and logarithmic functions, we will recall topics which may not
have made sufficient impression to have been absorbed completely and will
describe some topics from a slightly different perspective. Our motivation will
come from biological topics, including but not confined to population biology
and epidemiology. Our approach will emphasize qualitative ideas rather than
explicit computations; we feel that this approach is both easier for students
whose primary interests are not in mathematics and more useful in many
applications. This material is not, however, intended to serve as a differential
equations course for students in the physical sciences as it does omit many
techniques and topics that are essential for such students.
In presenting mathematical topics, we will attempt to tell the truth and
nothing but the truth, but not necessarily the whole truth. We will try to
emphasize the basic truths but not to overwhelm the student with precise
technical detail. Some results will be stated without proof, while others will
be accompanied by outlines of the reasons why they are true. We will normally
avoid detailed, rigorous proofs.
Mathematics is not a spectator sport, and can be learned only by solving

xi
xii Preface

problems. We include examples of problems with solutions and some exercises

which follow the examples quite closely. However, a library of solved examples
used as templates will not be sufficient to meet the needs of a developing sci-
entist. For this reason, we also include problems that go beyond the examples,
both in mathematical analysis and in the development of mathematical mod-
els for biological problems, in order to encourage deeper understanding and
an eagerness to use mathematics in learning about biology. We also include
some problems, marked with an asterisk (*), which are more challenging.
We recommend the introduction to modeling in Chapter 1 as a beginning
of any course on biological modeling using this text. The contents of Chapters
2 through 4 overlap considerably with the material which would be covered in
a unified course (probably two semesters in length) which covers calculus and
some difference equations and elementary differential equations. For students
who have already taken such a course, a suitable course could review these
topics as needed (after starting with Chapter 1) and then continue with Chap-
ters 5, 6, and possibly 7. For students who have had a semester of calculus
without difference or differential equations, a suitable course could consist of
the first four chapters.
We provide some support for helping readers use software packages to
generate numerical solutions and graphs (as we ourselves have done to make
some of the figures in this book), primarily in the sections dealing with nu-
merical analysis and in the appendices. These packages and the companies
that produce them are listed below.

Maple is a registered trademark of Waterloo Maple, Inc., www.maplesoft.com.

Mathematica is a registered trademark of Wolfram Research, Inc., at

www.wolfram.com.

MATLAB is a registered trademark of The Mathworks, Inc. For prod-

uct information please contact The Mathworks, Inc., 3 Apple Hill Drive,
Natick, MA 01760-2098 USA, tel. 508-647-7000, fax 508-647-7001, e-mail
[email protected], www.mathworks.com.

Trademarked names may be used in this book without the inclusion of a

trademark symbol. These names are used in an editorial context only; no
infringement of trademark is intended.
Acknowledgments

We thank Bob Stern at CRC Press for planting the initial idea for this book,
and both him and Bob Ross for their support and understanding in the face
of long delays during the preparation of the manuscript. (Although we do not
include delay equations in this book, there certainly were delays.) We also
thank everyone at Taylor & Francis who helped with the production of this
book, especially Shashi Kumar and Marcus Fontaine for critically useful help
involving LaTeX, Karen Simon, and Kevin Craig for graphic design assistance.
The figures and photos accompanying discussion of the various biologi-
cal systems studied and discussed play a crucial role in bringing the models
(and their motivations) to life for the reader, and many appear in this work
through the kind permission of others. We therefore acknowledge here all those
who generously permitted us to print their photos, or helped us obtain permis-
sion: Donna Anstey, Francine Bérubé, Daniel Bowen, John Calambokidis, Tom
Chrzanowski, Patricia Ernst, Carla Flores, Tim Gerrodette, Stefano Guerrieri,
Ray Hamblett, Alan M. Hughes, Duncan Jackson, Russell S. Karow, Carolyn
Kribs, Joel Michaelsen, Bernard E. Picton, Dave Powell, Francis Ratnieks,
Helen Sarakinos, Howard Swatland, Michael Tildesley, and S. Bradleigh Vin-
son.
We also thank all those photographers who contributed indirectly, in-
cluding photos in the public domain: Cleopatra Adedeji, Lennert B., David
Burdick, Janice Haney Carr, Mark Conlin, Karen Couch, Jan Derk, Gary
Fellers, Ryan Hagerty, William Chapman Hewitson, Steve Hillebrand, Tom
Hodge, John & Karen Hollingsworth, Al Mare, Maureen Metcalfe, Benjamin
Mills, A.J. Nicholson, Sergey Nivens, Zev Ross, Jeff Schmaltz, Greg Webster,
Gary Zahm, and the following organizations: the Centers for Disease Con-
trol, Florida Keys National Marine Sanctuary, Ken Gray Image Collection
at Oregon State University, MODIS/NASA, National Diabetes Information
Clearinghouse, NOAA, the Pennsylvania Dept. of Conservation and Natural
Resources, River Alliance of Wisconsin, the Sickle Cell Foundation of Georgia,
the Southwest Fisheries Science Center of the NOAA Fisheries Service, U.S.
Fish and Wildlife Service, U.S. Geological Survey, U.S. National Park Service,
and the Wisconsin Dept. of Natural Resources.

xiii
This page intentionally left blank
 Part I

Elementary Topics

1
This page intentionally left blank
Chapter 1
Introduction to Biological Modeling

1.1 The nature and purposes of biological modeling

Mathematical modeling has been used to explain biological systems for
centuries, going back to work such as that of Thomas Malthus (1798) on the
growth of populations, Daniel Bernouilli (1760) on smallpox vaccination and
arguably Fibonacci (Leonardo of Pisa, 1202) on the well-known rabbit prob-
lem. Modern mathematical biology has been an ongoing scientific endeavor
since the end of the nineteenth century, growing out of the work of individuals
such as P.D. En’ko in epidemiology, D’Arcy Wentworth Thompson in change
in biological organisms, and G. Stokes in fluid flows, and the technical develop-
ments of the late twentieth century have placed it in the forefront of scientific
research. The use of theoretical models to describe and predict the behavior of
biological systems is especially useful today in examining large-scale questions
where controlled experiments are either impossible or unethical to carry out
— the effects of large-scale alterations in environmental conditions on local
ecology, or of slight decreases in efficiency of certain disease control measures.
Even in situations where some experimental data is available, mathematical
models may validate existing theories or provide new insights by describing
the mechanisms through which biological processes occur, taking in a whole
spectrum of possible measurements in a single step. More specifically, one
might use modeling
• to predict unknown behavior (such as intervention effects) based on
given factors,
• to generate virtual experiments that inform general biological theory,
• to account for observed behavior as simply as possible,
• to compare competing hypotheses — which better accounts for observed
results?
• to evaluate or rule out hypotheses,
• to suggest conjectures and new hypotheses, or

3
4 Dynamical Systems for Biological Modeling: An Introduction

• to guide empirical inquiry — does this process make a key difference in

the behavior of the system?1
In short, modeling helps us consider the question, “What if ...?” The language
of mathematics, meanwhile, provides powerfully compact descriptions of com-
plicated ideas, and is capable, at its best, of offering equally simple insights,
such as the idea of a single threshold quantity which determines the outcome
of a complex biological process and brings together all the relevant factors in
a way that clarifies the relative importance of each (cf. Section 4.4).
Mathematical models, however, are far from perfect: as intricate as they
may appear at times, they are mere sketches or caricatures of the living world.
Biological systems are complex, with many underlying forces and interactions,
heterogeneity, and random variations and fluctuations. In modeling such a
system, we can only hope to develop a good enough approximation to reality
to accomplish our purposes. The engineering statistician George Box wrote,
“All models are wrong, but some are useful.”2 We use models to give rough
descriptions of reality in cases where rough is still accurate enough to give
useful insights.
By its nature, modeling involves both iteration and compromise. As model-
ing is imperfect, the results of a given model may not agree with observations,
or may have limitations that keep the model from being useful. In these cases,
the modeler must make changes to the model, or even replace it altogether,
and then analyze the new model to see if it describes the biology better (or
well enough). One has to be willing to tinker and try again. Iteration thereby
makes the modeling process cyclical rather than linear. The following section
sketches this iterative modeling process.
Section 1.3 goes into more detail about the types of mathematical models
one might use. The selection of a particular kind of mathematical structure as
a model follows from one’s research questions. A laboratory or field biologist,
interested in the specific biology of the organisms in the system, may be most
interested in quantitative or numerical results, and models which fit observed
data closely. Theoretical biologists, on the other hand, may be more interested
in qualitative, structural insights, which can suggest trends across models of a
number of similar systems, such as the role of genetic diversity in the survival
of populations. In each case, the researcher’s perspective and interests will
shape the research question and, through it, the type(s) of model and analysis
most likely to provide fruitful answers. The types of models we shall explore
in this text are called dynamical systems, which mean that they change as
functions of some independent variable(s), usually time.
Facing up to the imperfection of any model means acknowledging its lim-
itations. Modeling does not mean ignoring or brushing aside the assumptions
1 cf. E. Smith, S. Haarer, J. Confrey (1997). Seeking diversity in mathematics educa-

tion: mathematical modeling in the practice of biologists and mathematicians, Science and
Education 6(5): 441–472.
2 G.E.P. Box (1976). Science and statistics, Journal of the American Statistical Associ-

ation 71: 791–799.

 Introduction to Biological Modeling 5

and simplifications we make in constructing and interpreting models; rather,

it is an active exercise in compromise: including as much detail as necessary
while keeping the model manageable in terms of analysis and (as applica-
ble) data collection. A saying attributed to Albert Einstein runs, “Everything
should be made as simple as possible, but not simpler.” Section 1.4 discusses
some of the most common compromises faced in mathematical modeling, as
well as their consequences.
Finally, the modeling process should also include an explicit choice of scale
and unit(s) of measurement for the quantities or populations being modeled,
and Section 1.5 introduces two related issues which are discussed further as
they recur later in the text. In each section, we shall speak of biological systems
in general, as the focus here is on modeling, but the reader can substitute
specific systems, such as the human circulatory system, the current generation
of Monarch butterflies, or the population of Toronto at risk for SARS in March
2003, for the word “system” where it occurs.

1.2 The modeling process

Although mathematical modeling may look like a great variety of different
things, there are certain commonalities in the process, in the “big picture,”
and as this text is aimed at introducing students to biological modeling, it is
just as important to become familiar with the higher-order tools and elements
of modeling as with the mathematical techniques specific to each type of
model. As mentioned in the previous section, modeling is an iterative process.
Figure 1.1 sketches the flow of this process.
In practice, we begin by developing our first try at a model (which process
includes defining the problem at hand), translating the biological system into
a mathematical system. We then apply mathematical tools to analyze the
model. The mathematical results must then be interpreted back into biological
terms, and finally we must evaluate the answers our model has given us, to
decide whether they are satisfactory — biologically reasonable, and sufficient
to meet our objectives. If they are not (which is the case often enough), we
must adjust the model to try to remedy the observed shortcomings, and repeat
the process until the results are satisfactory.

modeling analysis interpretation

BIOLOGICAL MATHEMATICAL MATHEMATICAL BIOLOGICAL
PROBLEM PROBLEM RESULTS RESULTS

evaluation

FIGURE 1.1: Mathematical modeling as an iterative process.

6 Dynamical Systems for Biological Modeling: An Introduction

Mathematical biologists therefore need to be able to:

1. take biological features and hypotheses and turn them into mathematical
structures;
2. apply relevant mathematical analysis tools;
3. interpret mathematical results in biological terms; and
4. evaluate biological results critically in order to evaluate the underlying
mathematical model.

Let us consider each of these activities one by one.

(1) Modeling What does model construction look like? The details can
vary widely, but the philosophy is to describe the essential characteristics
of your system in mathematical terms. For us, this will typically mean a
population or group of interacting populations, although we will apply the
term population to collections of anything from molecules to households. One
perspective on modeling is that there are two directions in which to go about
it:
• forward — begin with observed characteristics, factors, rates, make your
model reflect these, and see if the results match observations; or
• backward — use observations to develop known (desired) model behav-
ior, and craft a model known to exhibit this behavior.
In other words, one might begin by assembling model elements which cor-
respond to the biological features of the system and consider the results (a
predictive model), or one might instead begin with a model of a type known
to behave as desired, and modify the structure to fine-tune (or fit data) (a
descriptive model). If we consider the three purposes suggested earlier for bi-
ological modeling, we see that a forward modeling approach suits the goal
of predicting unknown behavior based on known forces, while a backward (or
back-end) approach might best suit the goal of accounting for observed behav-
ior. Evaluating competing hypotheses may require a combination of the two
approaches. It is important, however, to acknowledge the distinction between
first principles models derived directly from known laws (common in physics
and chemistry, e.g., using Newton’s Laws, but less so in biology) and ad hoc
models which use terms, expressions or functions commonly used to represent
a particular feature or process of the system based on past observations of
their fit to data (e.g., using a logistic model to describe population growth
with limited resources (see Sections 2.4 and 3.2), or a mass-action term to
describe the interaction of two populations (see Section 4.3). Likewise there is
a distinction between broadly descriptive models, which attempt to capture as
much detail as possible in a biological system, and focused theoretical models
 Introduction to Biological Modeling 7

which incorporate only those features relevant to a specific research question.3

Of course, this perspective for developing models is by no means exclusive,
and should be considered only as long as it is helpful in providing direction.
The elements of model development include identifying the problem or
research question as specifically as possible (a complex task we shall not do
more than illustrate), writing careful definitions of relevant terms (variables
and parameters), specifying all the assumptions being made (see Section 1.4
for a fuller discussion), drawing diagrams (sometimes), choosing an appropri-
ate mathematical structure (see Section 1.3) and scale (see Section 1.5), and
assembling all these elements into a coherent mathematical description (for
us, an equation or set of equations). It is also important to articulate one’s
objectives — what kind of answers one expects — as they will influence the
types of model and analysis that one uses.
(2) Robustness In the physical sciences, models often are based on spe-
cific assumptions about the terms included in the model, and these assump-
tions are based on mechanistic derivations. In the biological sciences, terms
that are included in models do not necessarily have mechanistic derivations,
but are included to give tractable problems. Thus, for example, the logistic
equation often used to model the growth of a single population assumes a
specific form for the growth rate of the population. Many of the results that
can be deduced from a logistic equation model require only that the growth
rate be zero when the population size is zero and when the population size has
a positive value K called the carrying capacity, and that the growth rate be
positive when the population is between zero and the carrying capacity. Ro-
bustness of a model is the property that conclusions drawn from the model are
valid under a less specific set of assumptions. Often one does not know the de-
tails of a biological problem, and one has greater confidence in the conclusions
drawn if changes in detail do not affect the conclusions drawn.
(3) Analysis The mathematical analysis tools and techniques required
vary by model type; for example, stochastic models, which incorporate random
variations, typically require a numerical approach, in which one runs a large
battery of computer simulations where one or more parameters (model inputs)
vary according to a specified probability distribution. For the deterministic
(non-probabilistic) dynamical systems which will form the basis for models
considered in this text, we have three possible approaches: exact, qualitative,
and quantitative. Exact solutions, in which we obtain an explicit formula for a
biological quantity (usually as a function of time), can be obtained only for the
simplest possible types of dynamical systems. We will find them useful mostly
for linearizations, simplifications of complicated models that sketch model be-
havior near special points. Qualitative analysis provides general information
on how a given system behaves, and is typically used to identify what affects
the system in the long term. Quantitative methods use computers to approxi-
3 Castillo-Chávez, as quoted in E. Smith, S. Haarer, J. Confrey, Seeking diversity in math-

ematics education: mathematical modeling in the practice of biologists and mathematicians,

Science and Education, 6: 441–472 (1997).
8 Dynamical Systems for Biological Modeling: An Introduction

mate solutions numerically, and are useful for illustration purposes, or when a
model is too complicated to complete a qualitative analysis, or when we have
good estimates of the parameters involved. In this text we shall show how to
apply all three types of analysis, although qualitative analysis will receive the
most emphasis, as a way to prove how complicated models behave, regard-
less of the exact parameter values. Although the mathematical nature of such
models can generate convincingly real data, it is often important to remember
that, as gross oversimplifications of reality, mathematical models are just as
much rough sketches as the nonmathematical theories with which biologists
are familiar, and qualitative descriptions of behavior befit this notion.
(4) Interpretation Interpreting mathematical results in biological
terms requires a return to the original context of the problem, and is easier
with a well-articulated research question. That is, rather than simply asking,
“What happens if we model the following biological system?” one ought to
ask a specific question that can be addressed with a model — for example,
“What effects do seasonal temperature variations have on the size of this pop-
ulation?” or “How does behavior change based on observed infection levels
affect the eradicability of this disease?” (A general statistical analysis of an
experimental data set, however, may show up trends that shape subsequent
research directions.) Interpretation is also facilitated by a close correspon-
dence between each element of the mathematical model and the elements of
the biological system. The conclusion “this population will go extinct if beta
x is less than 1” is only useful if one can put “beta x” in biological terms —
what do beta and x mean? can we measure them (or at least their product)?
Which might we be able to control, and which are inherent properties of the
biological system?
The basis for an interpretation may be numerical, qualitative, or graphical.
When most parameters (fixed quantities) in a model can be estimated fairly
well, numerical information can provide ranges within which some varying
or controllable quantity causes the biological system to behave in a partic-
ular way, or critical values which cause that behavior to change. Identifying
thresholds between survival and extinction, or between settling down to an
equilibrium level and oscillating forever, or comparing the behavior of two
related but slightly different models, can answer qualitative questions such
as, “Which explanation for this phenomenon better accounts for our obser-
vations?” or “Does the system behave differently if we take into account this
additional factor?” The ready availability of numerous computational tools
also makes it possible to represent results graphically, and a graph can be
an especially compelling, concise statement of a complex result. Graphs can
illustrate both general trends and striking thresholds.
However, just as with algebraic expressions, graphical depictions of math-
ematical information can be so concise that the job of unpacking them into
nonmathematical terms can be significant. The ability of mathematics in all its
forms to represent complicated ideas compactly makes it both powerful and
challenging, and if one considers all the time invested in becoming familiar
 Introduction to Biological Modeling 9

with and manipulating algebraic expressions, it should not be surprising that

unpacking the information contained in graphs requires practice and thought.
For example, Roth4 found that even as apparently simple a graph as that of
population growth rates shown in Figure 1.2 created multiple interpretation
challenges. While a definitive primer in graph interpretation is not the aim of
this text, we shall discuss interpretation of figures we present in later chap-
ters, and also here consider some particular questions regarding the graph in
Figure 1.2 (cf. Figures 2.19 and 2.27 in Section 2.5).

Rate
births deaths

Population size
FIGURE 1.2: A graph comparing birth and death rates vs. population size,
after Roth (2001).

• Note first that the horizontal axis in the graph is not time but population
size. That is, the graph shows how birth and death rates vary depending
on the population, rather than over time. Since many continuous graphs
depict how some quantity changes over time, one must readjust one’s
reading of the graph to reflect what it is that is varying.
• Note second that the vertical axis in the graph gives birth and death rates
(measured in individuals per day, week, or year), and not total births
and deaths. That is, the graph shows how fast births and deaths occur,
as population size grows (varies). The bending downward of the birth
rate curve, for example, as population size increases does not address
what has happened to previous births in the past of any population that
reaches that size. The distinction (and relationship) between population
4 W.-M. Roth (1998), Unspecified things, signs, and “natural objects”: towards a phe-

nomenological hermeneutic of graphing. In S. B. Berenson, K. R. Dawson, M. Blanton,

W. N. Coulombe, J. Kolb, K. Norwood, and L. Stiff (Eds.), Proceedings of the Twentieth
Annual Meeting of the North American Chapter of the International Group for the Psy-
chology of Mathematics Education (Vol. I, pp. 291–297), ERIC Clearinghouse for Science,
Mathematics, and Environmental Education, Columbus, OH, 1998.
10 Dynamical Systems for Biological Modeling: An Introduction

sizes and their rates of change (mathematically, their derivatives with

respect to time) is a fundamental prerequisite of the models we will
develop in the chapters ahead.
• The simpler of the two curves to interpret is the death rate, which ex-
hibits a linear rise, corresponding to a constant per capita death rate (the
total death rate divided by the population size), visible in the graph as
the slope of this line. Since constant per capita death rates due to natural
causes are commonly assumed (see Section 2.1), one can interpret this
line as indicating that there are no density-dependent causes of death in
this system.
• The rise and fall of the birth rate curve, however, suggests that there are
density-dependent forces influencing the birth rate (otherwise this graph
would be a line, too). The decrease in birth rate for larger populations
suggests a factor such as limited resources and/or competition which
reduces reproduction in large populations. (The graph also suggests that
for large enough populations there will be no births at all, which is not
realistic, but see Section 2.4.)
• Ultimately, interpretation of this graph requires the reader to see the
“big picture” depicted here, what one might call the emergent properties
of the graph: what does it communicate? In this case, the superimposi-
tion of the two curves allows us to compare them, to see for what sizes
the population grows and for what sizes it shrinks. Where the birth rate
is greater, the population is growing; where the death rate is greater, it
is shrinking. Our attention should therefore focus on the point(s) where
the two curves intersect, in particular the point where the death rate
catches up with the birth rate. Beyond all the individual details de-
scribed above, the main conclusion this graph was drawn to suggest is
this population level, where (for reasons we shall explore in the chapters
ahead) the population will eventually settle.

Finally, it is also important to acknowledge explicitly the limitations on the

results given: the range within which one’s model is a good enough description
of biological reality, and the extent to which the interpreted results may be
expected to hold true. For example, one of the most frequent simplifications
we shall undertake in our mathematical analyses in this text involves making
nonlinear (complicated) models linear near certain key points. Almost none of
our models will themselves be linear, however. Taylor’s theorem (see Appendix
B) guarantees that linear approximations are reasonable within a very small
operating range, but to replace, say, the graph of the birth rate curve in
Figure 1.2 above with the tangent line at any one of its points would give a
very misleading idea indeed of the relationship between population size and
birth rate. (Recall that the tangent line is the linear approximation to a curve
at that point.)
 Introduction to Biological Modeling 11

(5) Evaluation Evaluating the results given by a particular model re-

quires asking questions such as the following: Are results reasonable? Do they
agree with observations to the extent possible? Do the assumptions made im-
pose too many limitations on the model’s ability to explain or describe the
biological system? The answers to these questions determine whether it is
necessary to change or replace the model. Discussions in the text where evalu-
ation receives special emphasis include those in Sections 2.4, 3.2, 5.1, 5.2, 6.1,
and 7.1.
Note that many of the decisions which need to be made in modeling — like
what features are relevant, what assumptions are reasonable — really become
easier to make only with experience, and the examples in the chapters ahead
provide a beginning. Keep at it!

1.3 Types of mathematical models

Once a research question has been defined and model development begins,
one must decide what type of mathematical model is most appropriate for the
given situation. In this text we shall discuss dynamic models, called in math-
ematics dynamical systems, which track the change of one or more variables
(for us, usually populations) over time (and occasionally over other variables
as well). This discussion therefore leaves out static models, such as statisti-
cal analyses of datasets, which are nevertheless useful for other purposes. We
shall organize our discussion by considering, in turn, the various choices to be
made.5
One recurring issue in modeling is the question, discrete or continuous?
We shall consider this question in three different regards. The terms discrete
dynamical system and continuous dynamical system generally refer to models
that are discrete or continuous in time. Discrete-time models measure the
size of a population at regular, fixed intervals, saying nothing about its size
at moments in between those snapshots. This approach is appropriate for
systems which change or reproduce in distinct generations (e.g., annually, like
many species of fish), or for which data are available at regular intervals,
and results in a model composed of difference equations, as we shall see in
Chapter 2. Continuous-time models depict population growth continuously
over time and are appropriate for systems with ongoing and/or rapid change,
and manifest as differential equations, as we shall see in Chapters 3 and 4.6 For
5 Another accessible and thorough discussion of types and purposes of mathematical

models applied specifically to the epidemiology of sexually transmitted diseases is given

in G.P. Garnett (2002), An introduction to mathematical models in sexually transmitted
disease epidemiology, Sexually Transmitted Infections 78: 7–12.
6 A detailed comparison between a discrete-time model and the continuous-time model

that corresponds to it is outlined in the review exercises at the end of Chapter 3.

12 Dynamical Systems for Biological Modeling: An Introduction

example, suppose we are studying how a certain bird population is affected

by fluctuations in its food source. If the birds are the swallows of Mission San
Juan Capistrano, in California, around which recent development has seriously
depleted nearby insect populations, we might consider a discrete-time model
spanning several years and focus on the number of swallows which return to
the mission on March 19, or the number still at the mission the day before
they migrate south on October 23. However, if we are studying the effects of
seasonal variation in insect availability on a non-migratory bird such as the
Carolina wren, we might instead consider a continuous-time model.
The discrete/continuous question may also arise in two different ways in
defining variables: heterogeneity and population size. If some property, such as
age, spatial location or risk level, varies significantly enough within a popu-
lation to warrant modeling that variation, we must decide whether to incor-
porate the heterogeneity discretely with a stratification or continuously with
a distribution. Stratifications divide a population into groups, each of which
follows similar rules or behaves similarly, and requires defining one state vari-
able for each group. The simplest possible stratification is to define two groups,
such as juveniles and adults, habitat 1 and habitat 2, or high-risk and low-
risk. Continuous distributions introduce an additional independent variable
like age or risk level, and typically lead to partial differential equations, which
are mathematically more complicated. Many human censuses gather data by
age ranges, such as 0–4 years, 5–9 years, 10–15 years of age, etc., and models
built to make use of such data will usually stratify accordingly by age. With
the injunction to simplify in mind, many models stratify into two or a few
groups rather than introduce a second independent variable. For example, a
landmark study on the transmission of gonorrhea in the United States7 fo-
cused on variation in sexual activity levels. Although in practice there is a
wide variation in such levels, the study was able to provide important insights
by stratifying the population into two groups: a low-activity group called the
non-core which includes most of the population, and a small, high-activity
group called the core. However, many applications require detailed model-
ing of spatial or age information, and continuous distributions are important
then. For instance, studies of pattern formation on growing animals (fish,
cats, crocodiles) require keeping careful track of location on the animal in a
continuous way that permits one to see the growth of waves in the patterns.
The state variables which keep track of the sizes of populations and other
dynamic quantities in the model can also be discrete or continuous. At some
level, real populations are always composed of a whole number of individuals
(organisms, molecules, etc.) although of course measures of size such as mass
and volume vary with each individual. Models which keep track of the status
of each individual in a population must by nature have a discrete population
size (even if that size changes with births, deaths, arrivals and departures), but
7 H.W. Hethcote and J.A. Yorke (1984). Gonorrhea transmission dynamics and control.

Lecture Notes in Biomathematics 56. New York: Springer-Verlag.

 Introduction to Biological Modeling 13

models which keep track only of the overall sizes of otherwise homogeneous
groups may allow the variables representing them to vary continuously, espe-
cially if the unit (see Section 1.5) makes this meaningful: biomass measured
in kilograms, or individuals measured in thousands or millions (where each
individual is such a small quantity that the difference between discrete and
continuous is minimal). In general, however, the choice between discrete and
continuous state variables is linked to a more fundamental decision: whether
to make the model deterministic or stochastic.
A deterministic model describes a system in terms of averages: every possi-
ble outcome — dying or not dying, reproducing or not reproducing, recovering
from a disease at any moment from just after infection to years afterward —
occurs, in the proportions corresponding to the relative probabilities of each
event happening. If the average lifetime for a population under study is 30
years, then the corresponding average (per capita) death rate is its reciprocal,
1 1
30 /yr ≈ 0.03/yr, and in a deterministic model precisely 1/30th, or 3 3 %, of the
population will die each year. In some sense, it is more accurate to say that
in a deterministic model, 1/30th of each individual dies per year. This type
of model is appropriate for large populations, where we call on the so-called
Law of Large Numbers, which says that as the number of trials of a random
event (such as whether or not a given individual dies) increases, the actual
experimental probability (here the proportion of the population that under-
goes a certain change) approaches the theoretical probability of the event.
Deterministic models also commonly use continuous state variables.
On the other hand, a stochastic, or probabilistic, model describes all events
in terms of probabilities, and essentially flips a coin or tosses a die for each
individual member of the population, to see whether that member dies or
not, reproduces or not, etc. Consequently, these models use discrete state
variables. A stochastic approach generally requires a little more mathematical
machinery — in particular, familiarity with probability distributions — but is
appropriate in cases where populations are small enough that the Law of Large
Numbers does not apply, or where for some other reason it is important to
couch the model in terms of probabilities: for example, to determine a possible
probability distribution for the outcome of a certain event (like how many
cases an outbreak of disease will cause). Recently, studies have been made of
the structure of social contact networks,8 and in some cases detailed data has
been recorded for contact networks (e.g., SARS outbreaks, and sexual contact
networks in small closed populations); studying how these networks affect
the spread of contact-driven phenomena such as infectious diseases requires
a stochastic approach. Likewise, studies of disease eradication must focus on
the last few infected individuals, where probabilistic effects are important. For
the large-scale study of gonorrhea mentioned earlier, however, a deterministic
8 See, for instance, S.H. Strogatz (2001). Exploring complex networks, Nature 410: 268–

276, and M.E.J. Newman (2003). The structure and function of complex networks, SIAM
Review 45: 167–256.
14 Dynamical Systems for Biological Modeling: An Introduction

approach is not only simpler but has provided significant enough insights to
inform disease control strategies.
In addition, statistical analyses can be especially important for some mod-
els. Two particular calculations of relevance here are sensitivity analysis and
uncertainty analysis. In a sensitivity analysis, a probability distribution of pa-
rameter values is chosen around the parameters’ estimated averages, in order
to determine how strongly the model results depend on each parameter; the
outcome is a ranking of parameters to which some particular model output —
say, the minimum influx of new individuals per year needed in order to sustain
a population which reproduces slowly — is sensitive. For example, one might
find this hypothetical minimum influx to be more sensitive to the existing
population’s death rate than to its reproductive rate. An uncertainty analysis
studies the impact of measurement errors (how far off the estimated param-
eter values are); the outcome is a probability distribution for some model
output, say, the spread of possible values for the required minimum influx
given estimated probability distributions for the birth and death rates.
No single text can hope to make a reader conversant with all of these
types of mathematical models. In this book we shall focus on deterministic
models, with continuous state variables, stratified discretely when necessary
to incorporate heterogeneity, the alternative being continuous variation us-
ing partial differential equations. Chapter 2 will consider models of biological
systems appropriate for discrete dynamical systems (recall that by discrete
we mean discrete in time), while the remaining chapters will consider mod-
els using continuous dynamical systems. In each case, we will begin our study
with simple (linear) single equations, proceed to more complicated (nonlinear)
single equations, and finally consider systems of interacting populations.

1.4 Assumptions, simplifications, and compromises

Modeling, as noted earlier, is an exercise in compromise: we include as
much detail as we can (and must) without losing our ability to analyze the
model. This means that, in developing a model and evaluating the results, we
must identify which simplifying assumptions we make, why we make them, and
how they influence (or limit) the conclusions we draw. It is worth mentioning
some of the trade-offs we make most frequently (in general practice, as well
as in the chapters ahead).
That we simplify at all means that to some extent we must be willing
to give up biological accuracy in exchange for tractability or “analyzability.”
For example, we may assume that the total size of a given population is
constant over time, in order not to have to keep track of it as a separate
variable. Almost always in this text, we shall assume that the amount of
time it takes for any biological process to happen is distributed exponentially,
 Introduction to Biological Modeling 15

the way radioactive decay works (as opposed to, say, a fixed period of time);
although some biological processes do work this way, the real reason we make
this assumption is to make our models be differential equations rather than
something more complicated, like delay or integral equations. Second, at the
same time our willingness to consider reasonably complex models means giving
up exact solutions to the equations in favor of qualitative information about
the solutions: do they eventually taper off toward some equilibrium, or do
they oscillate forever, or do they crash to zero in finite time, or do they do
something stranger? In the case of a quantitative model, we give up exact
solutions and broad, analytical guarantees of possible behaviors in exchange
for quantitative approximations given specific parameter values.9
Third, typically we also make some kind of assumption of homogeneity
of individuals — that, within a given group, they all behave the same way
with regard to any events described in the model. Recall also the discussion of
handling heterogeneity in the previous section; stratification typically includes
this assumption of homogeneity within each class, in exchange for a simpler
type of model. This assumption, as well as the first one above, arises from the
motivation to minimize the number of state variables used (the dimension of
the model).
Fourth, we sometimes give up immediate biological interpretability in re-
turn for tractability — that is, after initially defining a model in biological
terms, we may further simplify it by redefining variables and parameters, in a
way that makes the model mathematically simpler (usually reducing the num-
ber of different variables and/or parameters) but may complicate our ability
to interpret the new quantities in biological terms. For example, one model
for foraging ants10 studied in Section 4.1 begins with five parameters:

N — the total number of ants;

α — the (per capita) rate at which ants find a food site;
β — the (per capita) rate at which ants find a pheromone trail left
by other ants;
σ — the maximum rate at which ants retire from bringing food back
from the site;
K — the number of ants at which the retirement rate reaches 12 σ.

It is determined, however, that the model can be simplified to one involving

only three parameters:

αN K(βN − α) βK 2
a= , b= , c= .
σ σ σ
9 Computational schemes used in numerical analysis, however, are typically shown to

provide approximations that converge to the exact solution as the time-steps involved get
smaller. That is, one can make the approximate solutions as close as one likes to the exact
solution, if one is willing to wait longer for the computer to run its calculations.
10 D.J.T. Sumpter and S.C. Pratt (2003). A modelling framework for understanding social

insect foraging, Behav. Ecol. Sociobiol. 53: 131–144. DOI 10.1007/s00265-002-0549-0.

16 Dynamical Systems for Biological Modeling: An Introduction

The resulting model is simpler to analyze, but the new parameters are more
difficult to interpret in biological terms because the biological information
has been packed mathematically, and must be unpacked following analysis in
order to be interpreted more easily. This process is called rescaling or non-
dimensionalization and is discussed further in Section 1.5 below.
In articulating the simplifying assumptions involved in forming a model,
one should consider the consequences of introducing these “errors.” For ex-
ample, what do we lose from the results of a model if we assume homogeneous
mixing of sediments in a lake, or of a human population? The answers here,
of course, may be very different: in the first case, we lose the ability to dis-
tinguish the effects of sediment build-up as a function of depth, or distance
from the shore, including the resulting distribution of creatures in the lake
that are affected by those sediments (small animals may hide from predators
in the sediment, small animals and plants may even eat it; others, like the
predators, may be hindered by it). In the second case we lose our ability to
determine to what extent some individuals, who come into contact with more
people than most, such as a supermarket checker or a delivery person, may
be instrumental in the transmission of some contact-based process, such as an
infectious disease or the spreading of news and rumors by word of mouth.
There is, finally, another type of simplification which is sometimes made
without any intention of biological justification. In order to make inroads on
a tough problem, we may deliberately oversimplify our model — for instance,
by setting one or more quantities to zero — when the simpler model is easier
to analyze, with the purpose of bringing those results back to the original
problem to inform our intuition and expectations of how the full analysis will
turn out. For example, if there are two processes taking place and one proceeds
much more slowly than the other, we may assume temporarily that the “slow”
variables remain constant. We mention this practice (discussed in Chapter 6)
simply because it can be a useful tool for analyzing models at the edge of
our ability to handle, and because it might not occur to modelers focused
on justifying biologically all mathematical simplifications (the difference here
being that they are only temporary simplifications).
In any case, since models are imperfect descriptions of reality, unless our
main focus really is close fitting of data, our ultimate goal for conclusions
drawn from modeling should be a notion called robustness. This means that the
biological insights obtained from a mathematical model should be independent
of the particular mathematical structure and functions used, to the extent that
we are uncertain what the exact structure or function should be. For example,
as a population grows to fill its habitat, it runs up against the limitations
on the habitat’s resources, which begins to limit the population’s ability to
reproduce. This is true on scales anywhere from a Petri dish to the planet
Earth. This limitation on the birth rate can be described mathematically by
any function which tapers off eventually toward zero after an initial increase
(see, e.g., Figure 1.2). The most common function used for this purpose is
the quadratic that leads to the logistic equation, analysis of which shows that
 Introduction to Biological Modeling 17

the population will eventually level off at a size determined by the resources.
However, other functions of similar shape (that described above) and similar
properties can be used instead, with the same results. Likewise, we might
hope that stratifying a population into three or four age or risk categories will
provide results qualitatively similar to those of a model which stratifies the
population into ten or twenty classes. (We certainly would not like to analyze
the latter model!) Another example is that of threshold quantities, mentioned
above and discussed in Section 4.4. In this way, we justify to a limited extent
the imperfection of our model.
Sometimes models make predictions that are surprising or even alarming.
Such predictions from simple models imply a need for further careful study
of the phenomenon, in order to see if more detailed models make similar
predictions. Predictions which are robust across multiple models are more
likely to be accurate (but no more accurate than the models) and should be
taken seriously.

1.5 Scale, and choosing units

Last in our discussion of what biological modeling is like is a discussion
of scale and units. When one decides to model a certain problem, such as the
variable infectivity observed with HIV, the human immunodeficiency virus, it
is necessary to choose the scale on which to study the problem. Accordingly,
the state variables involved should have appropriate units. At one extreme,
we might choose to focus on the molecular level, in units of nanometers and
seconds — the biochemistry of the immune system’s initial response to the
viral invasion. Moving up a level, we might instead focus on the cellular level,
viewing the interactions of individual virus cells with helper T-cells and in-
fected cells in units of microns, minutes and days. Moving up further, we
might develop a model in the space of a single organ or individual, in units
of centimeters and days to years. Beyond this, we might focus on the inter-
actions of a small collective of individuals, in units of meters and weeks or
months. Finally, at the upper extreme, we might make a model at the level
of an entire population, in units of kilometers and weeks to years. Note that
each jump in scale also effectively uses the lower level as the building block
for the higher level: we model (and count) molecules interacting within a cell,
cells interacting within an organ or individual, individuals interacting within
a collective, and collectives (or individuals) interacting within a population.
This spectrum is illustrated in Figure 1.3.
Arguably the least common, and to some extent newest, of the levels in
this proposed spectrum is that of the collective — a small group within a
population which functions effectively as a unit with regard to the entire pop-
ulation, and within which individuals interact much more with each other
18 Dynamical Systems for Biological Modeling: An Introduction

MOLECULAR CELLULAR ORGAN[ISM] COLLECTIVE POPULATION

FIGURE 1.3: A diagram showing the spectrum of scale levels on which

modeling can take place.

than with others outside the collective. Recent research in mathematical bi-
ology has identified numerous situations in which collectives are important:
for example, individual households in studying infectious diseases. Within a
household, infections tend to spread faster among household members than
between a household member and an individual outside the household. This
is true of human respiratory diseases, as many families know, but also of pests
such as mice, which carry hantaviruses, and insects, which act as vectors for
diseases such as malaria (mosquitoes) and Chagas’ disease (triatomines or
“kissing bugs”). At the population level, it then also becomes reasonable to
make households the unit of measurement, and count infected houses rather
than infected individuals, with recovery coming only when no individuals in
a given household remain infected, or when treatment has eliminated all the
pests in the house.
Another issue related to scale and units is the rescaling procedure men-
tioned in the previous section. Sometimes we find it possible to reduce the
mathematical complexity of a model by reducing the number of variables
and/or parameters. Although the rescaled model can be more difficult to inter-
pret immediately, the resulting simplifications in our calculations are typically
worth it, especially when we can always undo the transformation afterward to
facilitate interpretation. Probably the most common type of rescaling is to re-
define variables relative to benchmark parameters. For example, in the model
of foraging ants mentioned in the previous section (and explored more fully
in Section 4.1), Sumpter and Pratt also rescaled both the state variable E(t),
representing the number of ants exploiting the food site, and the independent
variable t for time, as follows:
E t
x= , τ= ,
K K/σ

so that both x and τ are dimensionless. That is, x gives the number of ex-
ploring ants as a multiple of the benchmark value K, and τ gives the time
as a multiple of the benchmark value K/σ (which is how much time it takes,
on average, for K ants to retire). This is why the process is also called non-
dimensionalization. (The transformation to x in the article also involves elimi-
nation of another state variable for the number of ants not exploiting the food
source by assuming the total ant population constant, another simplifying
technique mentioned in the previous section.)
 Introduction to Biological Modeling 19

Exercises
Consider the following biological systems from a modeling perspective. For
each, determine which type of mathematical model and analysis you believe
would be most appropriate with regard to the criteria developed in this chap-
ter:
• discrete or continuous independent variable(s) (time, age, space),
• discrete or continuous dependent variable (population or quantity),
• type of stratification or continuous distribution of traits, if any,
• deterministic or stochastic,
• quantitative (numerical) or qualitative analysis,
• scale and units.
Defend or justify each of your choices with a sentence. (This exercise is most
useful if discussed in a group setting after each person has written down
his/her choices.)
1. the number of salmon that return to a particular pool to spawn each
year
2. a population of pea aphids whose genetic resistance to infection by fungal
spores varies inversely with their resistance to attack by parasitic wasps
(studied by students in the paper by Smith, Haarer and Confrey)
3. the frequency of a particular allele within each generation of a population
4. the likelihood of extinction of a rare allele (for reasons other than ge-
netic) over many generations of a population
5. the geographic dispersal (movement) of house sparrows within the
United States, starting with the original eight pairs released in the spring
of 1851 in Brooklyn, New York
6. the spread of tall grasses around the shore of a pond from an isolated
cluster
7. seasonal competition for space among different species of trees in a rain-
forest
8. weekly harvesting (removal) of fish in a fish hatchery
9. the growth of a yeast culture in a Petri dish
10. the mixing and growth of bacteria and nutrients in a chemostat
11. the growth of a deer population subject to an annual hunting season
20 Dynamical Systems for Biological Modeling: An Introduction

12. the spread of an epidemic in a large city

13. the role of “super-spreaders” in an epidemic’s growth
14. the growth of a population of blue whales, which become sexually mature
at age 10 years
15. the amount of a given drug remaining in the body over time
16. the kinetics of glucose, insulin and beta cells in the blood of diabetics
17. electrocardiac regulation
18. fluctuations in shark and fish populations in the Mediterranean (the
sharks eat the fish)
19. population density control as a means of eradicating fox rabies
20. comparing two alternative explanations for the mechanisms underlying
cell contraction
21. the concentrations of activator and inhibitor chemicals in pattern for-
mation processes in the hydra, a multicellular water-based creature
22. uptake mechanisms for the diagnostic dye bromosulfophthalein into liver
cells from the blood
23. competition between two species of bird for the same food source and
nesting sites
24. the gradual degradation of the human immune system by HIV
The following three questions provide further practice in graph interpretation.
25. Answer the following questions about Figure 1.2.
(a) What is the significance of the point where the birth rate curve
crosses the x-axis?
(b) What is the significance of the point where the birth rate curve has
a horizontal tangent?
(c) If the population begins at a level greater than that where the two
curves cross, what will happen to its size?
(d) If the population begins at a level less than that where the two
curves cross, what will happen to its size?

26. In what important ways does Figure 1.4 below differ from Figure 1.2?
Consider what will happen to the size of the population if it begins in
each of the three regions marked (a), (b) and (c). (This graph exhibits
something called an Allee effect, which is explored in Section 2.2.)
 Introduction to Biological Modeling 21

Rate
deaths
births

Population size
(a) (b) (c) Graph courtesy M. Tildesley

FIGURE 1.4: Graph for Exer- FIGURE 1.5: Graph for Exer-
cise 26. cise 27.

27. Figure 1.5 shows the distribution of farms in the U.K. during the 2001
outbreak of foot-and-mouth disease (FMD), in terms of the numbers of
cows and sheep on each farm. Color intensity indicates the number of
farms with the given numbers of animals.
(a) Where, on this graph, are the farms with large numbers of cows
but no sheep? Where are the farms with many sheep but no cows?
(b) What does the graph indicate about the distinct types of farms
that exist?
(c) Since there is no treatment for FMD, the epidemic was brought
under control by culling — slaughtering all susceptible animals on
farms within a certain radius of any animal found to be infected.
This wiped out numerous small farms altogether. How might the
small farmers use this graph to argue for a change in control policy?
28. Smith, Haarer and Confrey (1997) studied a group of graduate students
working together in a course on mathematical biology (see footnote 1).
They found that the goals of students from different fields, as well as the
course instructors, who had different backgrounds, differed at times. If
we consider three perspectives from which students using mathematical
models to represent biological systems might come, we might say that:
Those who came with the perspective of an experimental biologist wanted
models to incorporate all their data, and to fit and explain them.
Those who came with the perspective of a theoretical biologist wanted
to use the data obtained beforehand to suggest trends to explore in the
model, to use the model as a means of exploring the effects of particular
phenomena (like genetic variation) across particular systems, i.e., for
many different species and habitats. They saw the model as a virtual
22 Dynamical Systems for Biological Modeling: An Introduction

experiment, whose results become a single datum in the identification

of how general biological systems behave.
Those who came with the perspective of a mathematical biologist wanted
their models to be analyzable, focused only on research-related features
of the system rather than broadly descriptive, and were interested by
mathematical results and their biological significance.
Compare these goals and purposes with each other. To what extent are
they compatible? To what extent would the models produced by each
differ (that is, the models themselves, as opposed to the uses to which
they are put)?
29. Suppose you are investigating the effects of competing predators on the
life cycle of the alewife, a freshwater fish that serves as prey for other
fish such as salmon, trout and bass. Make a list of the five factors you
expect would be most important to include in a mathematical model
for this biological system, and another list of five factors which, while
relevant to the life cycle of the alewife, you would exclude from this
model. Justify both lists biologically.
30. Part of the homogeneity assumption discussed in this chapter is an as-
sumption that when some particular type of contact occurs between
members of two different groups — say, a species of predator and its
prey, or an infectious individual and a susceptible individual — it is
equally likely for the contact to involve any member of each group. In
reality, individuals don’t move about randomly. List five factors which
affect which member of the first group is likely to be the one which in-
teracts with a member of the second group, and indicate some way a
model might take each factor into account.
31. One simplification that most models of biological systems make (and
certainly those in this book) is to assume that events take place in-
stantaneously. That is, infectious contacts between two individuals last
only a moment, and predators and parasites locate and catch their prey
immediately. (The alternative is to have the model remove the two in-
dividuals involved from their respective classes for the time necessary
to make the contact, and after the given delay return the survivor(s)
to those classes, which complicates the model.) What biological justi-
fication can be made for this simplification? In what cases might this
assumption introduce important inaccuracies in the model and its be-
havior?
Chapter 2
Difference Equations
(Discrete Dynamical Systems)

2.1 Introduction to discrete dynamical systems

Many organisms have births, deaths, and other demographic processes that
occur at distinct, regular intervals (see Figure 2.1). Perhaps best-known among
these are semelparous animals that reproduce only once before dying, such as
the Pacific salmon (five species of the genus Oncorhynchus), which battles its
way back upstream from the sea to spawn and die in the river where it was
born. Likewise, the cicada family Cicadidae, which includes some 1500 species
worldwide, includes the genus Magicicada, native to North America, whose
members emerge from their underground hibernation every 13 or 17 years to
mate and die. Analogues also exist in the plant kingdom, such as western North
America’s monument plant (also known as green gentian) Frasera speciosa and
century plant (also known as Parry’s agave) Agavi americana, both of which
live for over twenty years before growing a tall, thin blooming spike which
produces flowers and seeds but saps the plant’s resources to the point that it
dies afterward. Such plants are called monocarpic. Some plants, such as the
carrot Daucus carota, also have fixed life cycles. The carrot and other similar
plants are biennial: at the end of the first year, they store nutrients in the root
that support flowering in the second year, at the end of which they die. On
a smaller scale, there are processes such as ovulation and even neuron firing
which occur discretely at certain set frequencies, rather than continuously.
Finally, human interactions with some populations can also cause discrete cy-
cles, such as harvesting or wildlife population management programs. This
includes situations in which empirical studies necessitate a discrete perspec-
tive on time because the only data available have been gathered at regular
intervals, making everything that happens in between observations equivalent
to one demographic fell swoop.
In cases such as these where it is appropriate to model some biological
quantity or process in terms of the number or amount present at discrete times,
we shall use models which are discrete dynamical systems, that is, systems
which evolve over time through equations (formulas) which tell how to figure
the size of the next generation, based on the size of the present generation.

23
24 Dynamical Systems for Biological Modeling: An Introduction

Photo courtesy U.S. Fish and Wildlife Service, dls.fws.gov Photo courtesy Joel Michaelsen

FIGURE 2.1: The semelparous salmon (left) and the monocarpic century
plant (right) reproduce in distinct generations. Note the different stages of the
blooming spikes on the three century plants.

These equations are called difference equations, since they give the difference in
size between one generation and the next, and first-order difference equations1
have the form yk+1 = f (yk ), which can be read as saying that the size of the
next generation (generation number k + 1 of population y) is a function f of
the present generation (generation number k of population y). In the rest of
this section, we will discuss the simplest type of difference equation and what
its solutions look like. Following that, the next sections in this chapter develop
some tools, both analytical and computational, for studying more complicated
difference equations. Later sections will look at classes of biological systems
that can be studied using discrete dynamical systems, including population
genetics, competition, harvesting, and finally systems involving more than one
quantity or population.

2.1.1 Linear difference equations

The rate of change of some quantity is often proportional to the amount of
the quantity present. This may be true, for example, of the size of a population
with unrestricted growth (say, lab bacteria in a petri dish). Let y(t) represent
the number of members of a population of simple organisms at time t. If we
assume that these organisms reproduce by splitting, and that on average a
fraction a of the members split into two members in unit time, then for a
small period h of time

y(t + h) − y(t) ≈ a h y(t), (2.1)

where the symbol ≈ signifying approximate equality means that the error in
this approximation is small for small h, in the sense that this error divided by
1 First-order difference equations will be our primary focus in this chapter.
 Difference Equations (Discrete Dynamical Systems) 25

h approaches zero as h → 0, i.e., that

y(t + h) − y(t) − a h y(t)
lim = 0.
h→0 h
For now, we neglect the small error and model the population size by the
equation
y(t + h) = y(t) + a h y(t). (2.2)
In this way we can predict y(t + h) given y(t) and a.
If a fraction b of the members reproduce by splitting and a fraction d
of the members die in unit time, then (2.1) would be true with a = b − d.
The constant a may be either positive or negative, depending on whether
b > d (more births than deaths) or b < d (more deaths than births). We will
therefore allow a to designate the constant of proportionality in (2.1) or (2.2),
whether it is positive or negative.
We may rewrite (2.2) in the form

y(t + h) = (1 + ah)y(t), (2.3)

which expresses y(t + h) as a function of y(t). In this chapter we shall think

of h as a fixed time interval, so that if we start at time t = 0 the function
y is defined not for all t but only for t = kh (k = 0, 1, ...). In other words,
we are observing the value of y only at regular intervals, such as every hour,
or every twenty-four hours. If we define tk = kh, the kth observation time,
and let yk = y(tk ), the value observed at that time, then the function y is
described by the sequence of values {yk }. Since tk + h = tk+1 , (2.3) becomes
the difference equation

yk+1 = (1 + ah)yk (k = 0, 1, 2, . . .). (2.4)

In order to simplify the notation, we let r = 1 + ah, so that (2.4) becomes

yk+1 = ryk (k = 0, 1, 2, . . .). (2.5)

By a solution of the difference equation (2.5) we mean an algebraic expres-

sion which gives us values for all the yk (k = 0, 1, . . .). If we know the initial
population size y0 then we can calculate first y1 = ry0 , then y2 = ry1 = r2 y0 ,
etc. Thus we may solve the difference equation (2.5) recursively. The specifi-
cation of the value y0 is called an initial condition. The graph of the solution
is the discrete set of points {(tk , yk ), k = 0, 1, 2, . . .}, but it is customary to
connect these points with line segments to give a continuous graph.
If we return to the bacteria multiplying in the petri dish, and measure time
in multiples of the bacteria’s doubling time, and measure y in multiples of the
original colony size, then Example 1 gives a model for this simple doubling
process. Example 2 gives a variation on this theme, with time steps during
which the colony size increases by 10%, beginning with a colony five times the
size of that in Example 1.
26 Dynamical Systems for Biological Modeling: An Introduction

Example 1.
Solve the difference equation yk+1 = 2yk , with y0 = 1.
Solution: From y1 = 2y0 (the difference equation with k = 0) we see that
y1 = 2y0 = 2. Then y2 = 2y1 = 4, y3 = 2y2 = 8, . . . We may then guess (and
prove by induction) that yk = 2k . This is the solution of the given problem.
Graphing will show the familiar exponential curve. 

Example 2.
Verify that yk = 5(1.1)k is the solution of the difference equation yk+1 =
(1.1)yk , y0 = 5.
Solution: First verify the initial condition: The expression yk = 5(1.1)k with
k = 0 gives y0 = 5(1.1)0 = 5, as desired. Now substitute
 the proposed solution
into the difference equation: yk+1 = (1.1) 5(1.1)k = 5(1.1)k+1 . Thus the
given yk satisfies the given difference equation and initial condition. 

Equation (2.5) fits the general form yk+1 = f (yk ) of a first-order difference
equation; it also belongs to a more specific class of equations of the form

yk+1 = ryk + b, (2.6)

called linear difference equations because the right-hand side f (yk ) is a linear
function of yk . The special case (2.5) where b = 0 is known as the homogeneous
case. These simplest types of difference equation have special names for two
reasons: first, as we shall see below, they are relatively easy to solve outright
(in fact, they are the only kind of difference equation we shall attempt to solve
explicitly); and, second, as we shall see in a later section, they turn out to be
key to understanding the behavior of more complicated difference equations.

2.1.2 Solution of linear difference equations

It is easy to find an explicit formula for the solution of the linear homoge-
neous difference equation with constant coefficients (2.5) for any constant r.
We merely observe that

y1 = r y0 , y2 = r y1 = r 2 y0 , y3 = r y2 = r 3 y0 , . . .

Then it is natural to guess that in general

yk = r k y0 , (2.7)

and it is easy to verify (by induction) that this is correct.

Example 3.
Find the solution of the difference equation yk+1 = −yk , y0 = 1.
Solution: From the formula (2.7) with r = −1 we have yk = (−1)k ; notice
that yk oscillates between positive and negative values. 
 Difference Equations (Discrete Dynamical Systems) 27

Example 4.
Find the solution of the non-homogeneous linear difference equation

yk+1 = −yk + 1, y0 = 1.

Solution: We begin by calculating

y1 = −y0 + 1 = 0, y2 = −y1 + 1 = 1, y3 = −y2 + 1 = 0,

and then conjecture that yk alternates between 0 and 1. To verify the correct-
ness of this conjecture, we need only note that if yk = 1, then yk+1 = 0 and
if yk = 0, then yk+1 = 1. Thus yk does alternate between 0 and 1, which we
may express explicitly as yk = 12 [1 + (−1)k ]. 

It is also not difficult to solve the general linear non-homogeneous difference

equation (2.6) with constant coefficients r and b. We calculate

y1 =r y0 + b
y2 =r y1 + b = r(r y0 + b) + b = r2 y0 + r b + b
y3 =r y2 + b = r(r2 y0 + r b + b) = r3 y0 + r2 b + r b + b, etc.

The formula for the general term is

yk = rk y0 + b 1 + r + r2 + . . . + rk−1 .


(2.8)

By using the formula

1 − rk
1 + r + r2 + . . . + rk−1 =
1−r
for the sum of a geometric series (with r 6= 1), we may write this solution in
the form
1 − rk
 
b b
yk = r k y0 + b = y0 − rk + . (2.9)
1−r 1−r 1−r
We may also consider what happens to solutions of linear difference equa-
tions over long periods of time – in mathematical terms, as k → ∞. Consider-
ing the solution (2.9), we see that if r is large in size, i.e., r > 1 or r < −1, then
rk grows unbounded as k → ∞, and thus yk grows unbounded too. The only
b b
exception occurs when y0 = 1−r , in which case yk is a constant 1−r for all
b
k (cf. (2.9)) and thus approaches 1−r as k → ∞. Small differences, however,
from this equilibrium solution will be magnified for large r (cf. again (2.9)).
If r is instead small in size (between −1 and 1), then rk approaches zero
b
as k → ∞, and in view of (2.9) the solution yk approaches the limit 1−r as
k → ∞ regardless of the initial value y0 .
If r = −1, rk alternates between −1 and 1, and yk does not have a limit
28 Dynamical Systems for Biological Modeling: An Introduction
b
(unless y0 = 1−r ). If r = 1, the formula (2.9) is meaningless, but the formula
(2.8) becomes yk = y0 + rk; thus yk becomes unbounded as k → ∞.
The way the solution of the linear homogeneous difference equation (2.5)
depends on the value of the survival-and-growth term r will be important
when we study qualitative behavior of solutions of difference equations in
Section 2.3. From the formula (2.7) we see that yk → 0 as k → ∞ if |r| < 1
and yk grows unbounded as k → ∞ if |r| > 1. More precisely, if 0 ≤ r < 1,
yk decreases monotonically to zero and if −1 < r < 0, yk oscillates between
positive and negative values in approaching zero. If r > 1, yk increases to +∞
and if r < −1, yk oscillates unboundedly. The “boundary” cases are r = −1,
in which yk oscillates between ±y0 and does not approach a limit, and r = 1,
in which yk is the constant y0 . The essential property we shall need is the
following result.

The solution of the difference equation yk+1 = r yk approaches zero as

k → ∞ if and only if −1 < r < 1.

Example 5.
Find for which values of a every solution of the difference equation yk+1 =
(1 + a)yk approaches zero as k → ∞.
Solution: Every solution approaches zero if and only if |1 + a| < 1, or −1 <
1 + a < 1, or −2 < a < 0. 

2.1.3 Nonlinear difference equations

Note that model (2.5) assumes a constant growth rate independent of pop-
ulation size. This assumption is unlikely to be reasonable for real populations,
except possibly while the population is small enough in size not to be subject
to the effects of overcrowding. Various nonlinear difference equation models
have been proposed as more realistic in the general case, where resource lim-
itations constrain growth. For example, the difference equations
ryk
yk+1 = (2.10)
yk + A

due to Verhulst,2 and the second-order Hill function

ryk2
yk+1 = (2.11)
yk2 + A2

have been suggested as descriptions for populations whose growth rates satu-
rate for large population sizes. Here A is the population size at 50% saturation
2 P. F. Verhulst, Récherches mathématiques sur la loi d’accroissement de la population,

Mem. Acad. Roy. Brussels 18 (1845), 1–38.

 Difference Equations (Discrete Dynamical Systems) 29

(plug in A for yk in (2.10) to see why). For small yk (relative to A), the denom-
inator of the right-hand side of (2.10) and (2.11) is essentially A, so that (2.10)
has yk ≈ Ar yk , similar to the unrestricted growth of the linear equation (2.5).
However, for large yk , A is relatively insignificant, making the right-hand sides
of (2.10) and (2.11) (and thus yk+1 ) approximately equal to r.
Another much-studied example is the logistic difference equation
 yk 
yk+1 = ryk 1 − , (2.12)
K
also introduced by Verhulst, with a growth rate which decreases to zero as yk
approaches the carrying capacity K and which becomes negative for yk > K.
The logistic difference equation should not be taken seriously as a model for
large population sizes as yk+1 is negative if yk > K. Other difference equations
which have been used as models to fit field data are

yk+1 = ryk (1 + αyk )−β (2.13)

and (
ryk1−β , for yk > ǫ,
yk+1 = (2.14)
ryk , for yk < ǫ.
None of the difference equations (2.10), (2.11), (2.12), (2.13), (2.14) are
derived from actual population growth laws. Rather, they are attempts to
give quantitative expression to rough qualitative ideas about the biological
laws governing population growth. For this reason, we should be skeptical
of the biological significance of any deduction from a specific model which
depends on the precise formula for the solution of that model. Our goal should
be to formulate principles which are robust, that is, which are valid for a large
class of models embodying some set of qualitative hypotheses. Therefore, we
shall be more concerned with qualitative properties of solutions of difference
equations than with formulae for solutions. In fact, although we have seen
some examples of linear difference equations for which a solution formula is
available, the solution of nonlinear difference equations like (2.10), (2.11),
(2.12), (2.13), and (2.14) usually cannot be found in general as functions of
the equation parameters. In Section 2.3 we will develop tools for analyzing
the behavior of these more complicated equations.
The most that we can do quantitatively with many nonlinear difference
equations is to calculate solutions numerically by iteration for particular
choices of parameter values, as illustrated by the following two examples.

Example 6
Find the first four terms of the solution of the difference equation yk+1 =
yk (1 − yk ) with y0 = 12 .
Solution: We have y1 = 12 (1 − 12 ) = 14 = 0.25, y2 = 14 (1 − 41 ) = 16
3
= 0.1875,
3 3 39 39 39
y3 = 16 (1 − 16 ) = 256 = 0.152344, y4 = 256 (1 − 256 ) = 0.129135. 
30 Dynamical Systems for Biological Modeling: An Introduction
0.5 1

0.4 0.8

0.3 0.6

0.2 0.4

0.1 0.2

0 1 2 3 4 0 5 10 15 20 25

FIGURE 2.2: Solution to Exam- FIGURE 2.3: A solution to yk+1 =

ple 6. 3.62yk (1 − yk ).

Example 7.
1
Verify that the constant sequence yk = 2 (k = 0, 1, 2, . . .) is a solution of the
difference equation yk+1 = 2yk (1 − yk ).
Solution: If yk = 12 , then 2yk (1 − yk ) = 2( 12 )(1 − 21 ) = 12 . Thus yk = 1
2 satisfies
yk+1 = 2yk (1 − yk ). 

Solutions of difference equations may be calculated and graphed eas-

ily using a computer algebra system such as MapleTM , Mathematica R , or
MATLAB R . Again, although a graph of a solution is technically a set of
discrete points (k, yk ), we shall follow the customary procedure of connecting
the points and presenting the graph as a continuous sequence of line seg-
ments. Figure 2.2, for example, shows the solution of the difference equation
of Example 6 graphed with the aid of Maple; a program for this is given in
Appendix A. Figure 2.3 gives an example which would be more difficult to
calculate by hand, namely 25 terms of the solution of the difference equation
yk+1 = 3.62yk (1 − yk ) with y0 = 0.52. In the next section we shall present a
graphical approach to analyzing difference equations.

Exercises
1. The bacteria E. coli doubles in a little over 20 minutes in good lab
conditions. (For purposes of this exercise, we will assume it takes exactly
20 minutes.)
(a) Write a difference equation that models the growth of an E. coli
 Difference Equations (Discrete Dynamical Systems) 31

colony in a large nutrient dish, with an initial population of y0 =

10000 bacteria, and time measured in increments of 20 minutes.
What is the solution to this equation and initial condition?
(b) How would the equation change if the time between measurements
was one hour instead of 20 minutes? (Consider what happens to
the size of the colony in one hour.)
(c) In light of the discussion in the previous chapter, why would these
models be inappropriate for a colony beginning with a single bac-
terium, or only a few?
(d) Why would these models be inappropriate for tracking the growth
of the colony over a period of months?
(e) Suggest an alternate (nonlinear) model that might be more appro-
priate in tracking the growth of the colony from its inception until
well after it fills the petri dish. Explain in biological terms any
numbers or parameters.
2. Gressel and Segel3 derived the following annual model for the density of
a certain weed species (in weeds per square meter):

wn+1 = F wn + µ,

where wn is the density of the weed in year n, µ accounts for growth

due to mutation from another, closely related species, and F is the ratio
of the fitness of the weed species relative to that of its main competitor
(so, e.g., F > 1 means the weed is more fit than its competitor).
(a) Solve this difference equation to obtain an expression for wn in
terms of µ, F , and the initial condition w0 .
(b) In light of the discussion in this section on difference equations
of this form, what eventually (after many years) happens to the
density of this particular weed if F < 1? Interpret this conclusion
biologically.
(c) If instead F > 1, what happens to the weed density after many
years? Interpret this conclusion biologically.
(d) Which of these conclusions are realistic consequences of the origi-
nal assumptions, and which aren’t? Evaluate this model, and give
a range within which it gives a useful description of the weed’s
growth.
3. Find the first three terms of the solution to yk+1 = yk (1 − yk ), y0 = 34 .
4. Find the first three terms of the solution to yk+1 = yk (1 − yk ), y0 = 41 .
3 J. Gressel and L.A. Segel, The paucity of plants evolving genetic resistance to herbicides:

possible reasons and implications, Journal of Theoretical Biology 75: 349–371, 1978.
32 Dynamical Systems for Biological Modeling: An Introduction
yk
5. Find the first three terms of the solution to yk+1 = yk +1 , y0 = 1.
2
yk
6. Find the first three terms of the solution to yk+1 = 2 +1 ,
yk
y0 = 1.

7. Verify that yk = 1 is a constant solution of the difference equation

yk+1 = yk e1−yk .
8. Verify that yk = 1 is a constant solution of the difference equation
yk+1 = y2y k
k +1
.
∗
9. Verify that if r > 2 the sequence given by
√
(r + 2) + r2 − 4
yk = (k even)
2r
√
(r + 2) − r2 − 4
yk = (k odd)
2r
is a solution of the difference equation yk+1 = yk + ryk (1 − yk ).
∗
10. Show that a constant solution yk = ŷ (k = 0, 1, 2, ...) of a difference
equation yk+1 = g(yk ) must satisfy the relation g(ŷ) = ŷ.
In the following exercises, find the solution of the given difference equa-
tion.

11. yk+1 = (1.1)yk , y0 = 1 15. yk+1 = (1.1)yk − 0.1, y0 = 1

12. yk+1 = −(1.1)yk , y0 = 1 16. yk+1 = −(1.1)yk + 0.1, y0 = 0
13. yk+1 = 12 yk , y0 = 1 17. yk+1 = 12 yk − 12 , y0 = 0
14. yk+1 = 31 yk , y0 = 1 18. yk+1 = 13 yk + 1, y0 = 1

In the following exercises, determine which difference equations have

solutions that approach a limit as k → ∞, and then find the limits in
those cases.

19. yk+1 = −0.2yk 21. yk+1 = 0.2yk + 1

20. yk+1 = − 21 yk 22. yk+1 = − 12 yk − 1
 Difference Equations (Discrete Dynamical Systems) 33

2.2 Graphical analysis

There is a simple graphical method for solving difference equations, called
the cobwebbing method. We shall illustrate the method first by applying it to
the linear homogeneous difference equation

yk+1 = r yk (2.15)

which we have already solved analytically in Section 2.1. The method is also
applicable to difference equations which cannot be solved analytically. It may
be carried out on a computer with the aid of a computer algebra system such
as Maple, Mathematica, or MATLAB. Some of the figures in this section were
produced using Maple, with a program given in Appendix A.
This method begins by drawing in the y − z plane the reproduction curve,
which is the graph of the function on the right-hand side of our difference
equation, and the line z = y, which we shall use for reflection purposes. For
equation (2.15), the reproduction curve is the line z = r y (taken from r yk ).
Next, we mark y0 on the y-axis, and go vertically to the reproduction curve
(meeting it at height ry0 ). The next step is to go horizontally to the line
z = y, meeting it at the point (y1 , y1 ). (Recall y1 = ry0 from (2.15).) Then
we repeat the process, going vertically to the reproduction curve (at height
ry1 ), then horizontally to the line z = y at the point (y2 , y2 ), where y2 = ry1 .
Continuing in the same manner, we reach successively the values y3 , y4 , y5 ,
. . .. The graphic portrayal will have four different cases — r > 1 (Figure 2.4),
0 ≤ r < 1 (Figure 2.5), −1 < r < 0 (Figure 2.6), r < −1 (Figure 2.7) —
corresponding to different relative positions of the reproduction curve and the
line z = y. In each case, the graphical solution illustrates the behavior already

z z=ry z=y z z=y

z=ry

y =r y
1 0
y =r y
2 1
y =r y
2 1

y =r y
1 0

y y
y y y y
0 1 1 0

FIGURE 2.4: Cobwebbing method FIGURE 2.5: Cobwebbing method

for equation (2.15), r > 1. for equation (2.15), 0 < r < 1.
34 Dynamical Systems for Biological Modeling: An Introduction

z z=y z z=y

y0 y0
y y

z=ry
z=ry

FIGURE 2.6: Cobwebbing method FIGURE 2.7: Cobwebbing method

for equation (2.15), −1 < r < 0. for equation (2.15), r < −1.

obtained analytically in Section 2.1. Note that we exclude the case r = 1,

which would make (2.15) yk+1 = yk .
The cobwebbing method may be applied to any difference equation of the
form
yk+1 = g(yk ) (2.16)
using the reproduction curve z = g(y) and the line z = y. It gives informa-
tion about the behavior of solutions and is particularly useful for difference
equations whose analytic solution is complicated.

Example 1.
Apply the cobwebbing method to describe the solutions of the Verhulst equa-
tion
r yk
yk+1 = .
yk + A
ry
Solution: The reproduction curve is z = y+A and its slope is

dz rA
= .
dy (y + A)2

At y = 0 the slope is r/A. If r < A, this slope is less than 1, so the reproduction
curve lies below the line z = y, while if r > A this slope is greater than 1, so the
reproduction curve begins above the line z = y but comes down to intersect
it at y = r − A. If r > A, every solution, regardless of the initial value y0 ,
approaches the limit y∞ = r − A (Figure 2.8), and if r < A, every solution
approaches zero (Figure 2.9). In either case, the limit is an intersection of the
reproduction curve and the line z = y. 
 Difference Equations (Discrete Dynamical Systems) 35
z z

z=y z=y

ry
z=
y+A

ry
z=
y+A

y
y <r−A r−A y0>r−A y y
0 0

FIGURE 2.8: r > A in Example 1. FIGURE 2.9: r < A in Example 1.

Example 2.
Apply the cobwebbing method to describe the solutions of the difference equa-
tion
r y2
yk+1 = 2 k .
yk + A

Solution: The reproduction curve is

ry 2
z= ,
y2 + A
which intersects the line z = y when y = 0 and
√
r ± r2 − 4A
y= .
2
√ √
Thus if r > 2 A there are three real √ intersections and if r < 2 A the only
real intersection is √at y = 0. If r < 2 A, every solution approaches zero (Fig-
ure 2.10). If r > 2 A, every solution with y0 less than the smaller positive
intersection approaches zero (Figure 2.11), and every solution with y0 √ greater
2
than the smaller positive intersection approaches the limit y∞ = r+ r2 −4A
(Figure 2.12). This model has been used to describe a population which col-
lapses if its initial size is too small but survives if√ the initial population size
2
exceeds a threshold, which in this example is r− r2 −4A . Such behavior in a
population is called an Allee effect. As in Example 1, limits of solutions are
intersections of the reproduction curve and the line z = y. 

In Section 2.5 we shall see how we can identify limits of solutions (defined
in the next section as equilibria) graphically for different parameter values, by
this same superposition of two curves.
36 Dynamical Systems for Biological Modeling: An Introduction
z
z z

z=y
z=y z=y

RC

RC
RC

y y y y
y y
0 0 0

FIGURE
√ 2.10: r < FIGURE
√ 2.11: r > FIGURE
√ 2.12: r >
2 A in Example 2. 2 A, y0 < y− in Exam- 2 A, y0 > y− in Exam-
ple 2. ple 2.

Exercises
In Exercises 1-6, apply the cobwebbing method to describe the solutions
of the given difference equation.

1. yk+1 = 2yk + 3 4. yk+1 = yk + 2yk (1 − yk )

2. yk+1 = −3yk + 1 5. yk+1 = yk e1−yk
2yk
3. yk+1 = 12 yk (1 − yk ) 6. yk+1 = 4yk +1

7. Breaking a problem down into distinct cases can be a useful problem

solving tool.
(a) Why are there four different cases to consider in cobwebbing the
generic homogeneous linear difference equation (2.15)?
(b) At what two points in Example 2 does the solution split into two
cases or subcases?
8. (a) What happens mathematically to solutions of (2.15) when r < 0
(cf. Figures 2.6 and 2.7)?
(b) What does this mean in biological terms?
(c) Why do Examples 1 and 2 not consider the case r < 0?
9. Consider the nonlinear difference equation

yk+1 = 8yk3 (1 − yk ),

which provides another way to model growth limitations for large pop-
ulations (here large means closer to 1 than to 0, so consider the units
of y to be the maximum population size at which reproduction can
occur). Draw cobweb diagrams for three different initial conditions:
 Difference Equations (Discrete Dynamical Systems) 37

(a) y0 = 0.32, (b) y0 = 0.64, (c) y0 = 0.96. Explain in biological terms

what the model predicts will happen in each case. To what extent is this
prediction realistic?
10. For most complicated models, we will typically apply qualitative anal-
ysis tools as developed in Section 2.3, to obtain information about the
eventual behavior of solutions. However, there are cases in which even
these methods are difficult to apply; in such cases, numerical explo-
ration is again useful. Apply cobwebbing to the following two models of
intraspecies competition (see Section 2.4) and try to describe qualita-
tively the behavior of solutions in each case.
(a) The logistic model yk+1 = ryk (1 − yk ) with r = 3.55
(b) Ricker’s model yk+1 = ryk e−yk with r = 3.55

2.3 Qualitative analysis and population genetics

As remarked in the previous section, nearly all models of biological systems
are nonlinear, and typically cannot be solved outright to obtain the population
size (or other quantity) as a function of time (here, a discrete time index). We
therefore need to develop tools that will allow us to describe the behavior of
solutions qualitatively. In this section we introduce a technique for qualitative
analysis based on a process called linearization, and apply it to some simple
problems in population genetics.

2.3.1 Linearization and local stability

The motivation for linearization is fairly simple. Take any spot on a rea-
sonably smooth curve (say, a curve with continuous first derivative) and zoom
in on that spot on the graph. No matter how many zigs and zags the curve
contains, if it is smooth then as you zoom in on the given point, restricting
your attention to a smaller and smaller neighborhood of the chosen point, the
graph will begin to straighten out, until it resembles — locally — a straight
line. This suggests that, very close to any particular point, any smooth func-
tion behaves like a line (in particular, the line tangent to the curve at the
given point). Figure 2.13 illustrates this concept for the graph of a particular
curve. Consequently, within a very limited range, sufficiently smooth nonlinear
dynamical systems behave like linear dynamical systems — and we already
know how those behave. The utility of this approach turns out to depend on
choosing the right points around which to linearize: a special kind of point
called an equilibrium.
38 Dynamical Systems for Biological Modeling: An Introduction
15 4
60
10
40
2
20 5

-4 -2 2 4 -0.4 -0.2 0.2 0.4 -0.04 -0.02 0.02 0.04

-20 -5
-2
-40
-10
-60
-15 -4

FIGURE 2.13: Three views of the function y = x3 − 10x + 10 sin 10x, each
centered at the origin. As we restrict our view to smaller x-intervals around
0, the graph appears increasingly flat.

The constant solutions of a difference equation

yk+1 = g(yk ), (2.17)

if there are any, are particularly easy to find. They are simply the solutions
of the equation g(y) = y. We define an equilibrium of the difference equation
(2.17) to be a solution y∞ of the equation g(y) = y. This is also referred to
as a fixed point of the map g. Geometrically, an equilibrium is an intersection
of the reproduction curve z = g(y) and the line z = y. In the nonlinear
difference equations used as examples in the previous section, we have observed
a tendency for solutions to approach a limit as k → ∞, and these limits are
equilibria. Go back and look at the cobwebbing diagrams in the previous
section to see this (note that in most cases, cobwebbing leads us to a point
where the reproduction curve meets the line z = y). It will turn out that
we can obtain a great deal of information about the solutions of a difference
equation by studying the nature of its equilibria.
An essential fact in studying the behavior of solutions of a difference equa-
tion near an equilibrium is that a difference equation may be approximated
by a linear difference equation near an equilibrium. If y∞ is an equilibrium
of the difference equation (2.17), so that g(y∞ ) = y∞ , we make the change of
variable
u k = yk − y∞
so that uk represents the deviation from the equilibrium. Substitution into
(2.17) gives
y∞ + uk+1 = g(y∞ + uk ).
By Taylor’s Theorem (see Appendix B), we may write

g ′′ (c) 2
g(y∞ + uk ) = g(y∞ ) + g ′ (y∞ )uk + u
2! k
for some value c between y∞ and y∞ + uk . Since g(y∞ ) = y∞ , the difference
equation (2.17) is equivalent to the difference equation

g ′′ (c) 2
uk+1 = g ′ (y∞ )uk + u (2.18)
2! k
 Difference Equations (Discrete Dynamical Systems) 39
′′
for u. The term g 2!(c) u2k is not really a quadratic function of uk because the
intermediate value c may depend on uk , but it is small compared to uk in the
sense that this term divided by uk approaches zero as uk → 0. The lineariza-
tion of (2.17) at the equilibrium y∞ is defined to be the homogeneous linear
equation
uk+1 = g ′ (y∞ )uk (2.19)
obtained from (2.18) by neglecting this “higher order” term. Note that the
constant term drops out only because we are linearizing about an equilibrium,
where g(y∞ ) = y∞ .

Example 1.
Find the linearization of the logistic difference equation
 yk 
yk+1 = ryk 1 −
K
at each of its equilibria.
Solution: The equilibria are the solutions of the equation
 y
y = ry 1 − ,
K
namely y = 0 and y = K(1 − 1r ) (the latter positive for r > 1). Here the
function g(y) is given by  y
g(y) = ry 1 − ,
K
and
2ry
g ′ (y) = r − .
K
Thus g ′ (0)
 = r, and the linearization at the equilibrium y = 0 is uk+1 = r uk .
Also, g ′ K(1 − 1r ) = r − 2r(1 − r1 ) = 2 − r, and so the linearization at
y = K(1 − 1r ) is uk+1 = (2 − r)uk . 

The importance of the linearization lies in the fact that the behavior of
solutions to the linearization of a difference equation at an equilibrium de-
scribes the behavior of solutions of the original difference equation near the
equilibrium. In particular, what we want to know is, if a solution (the popula-
tion size) is close to a particular equilibrium, will it approach the equilibrium,
or will it go away from the equilibrium? This attractive-repulsive quality is
called stability, and because it is based on linearizations, which are only good
approximations of the original when we are very close to an equilibrium, the
stability information we get from them is called local stability, as it does not
(yet) tell us what happens when we start far away from an equilibrium.
An equilibrium of the difference equation (2.17) is said to be asymptot-
ically stable if every solution whose initial value is sufficiently close to the
equilibrium remains close to the equilibrium and approaches it as k → ∞.
An equilibrium is said to be unstable if there are solutions with initial values
40 Dynamical Systems for Biological Modeling: An Introduction

arbitrarily close to the equilibrium which fail to remain near the equilibrium.
For an example, consider the generic homogeneous linear difference equation
(2.15), yk+1 = r yk , which has the unique equilibrium y = 0. As seen in the
previous section, when −1 < r < 1 (the second and third cases, Figures 2.5
and 2.6), solutions approach the equilibrium at 0; in these cases, the equilib-
rium is asymptotically stable. However, when r > 1 or r < −1 (the first and
fourth cases, Figures 2.4 and 2.7), solutions grow unbounded away from 0; in
these cases, the equilibrium is unstable. This example actually provides the
basis for the equilibrium stability theorem below, as all linearizations have
this form. The two examples in the previous section can also be analyzed in
this way (see, for instance, Example 3 below).
In applications to biological or other sciences, a difference equation model
is only an approximation to reality. There are inevitably errors or approxima-
tions in the model, and there are experimental errors in the measurement of
data to determine the parameters of the model. If an equilibrium is asymp-
totically stable, these errors are unimportant in the long-time behavior of the
system, because a small movement away from the equilibrium will be wiped
out over time. On the other hand, an unstable equilibrium is not experimen-
tally observable because any small error will cause the solution to move away
from it. Thus asymptotic stability of an equilibrium is an essential property for
applications. (Sometimes errors or changes in parameter values can actually
affect the stability of an equilibrium; these changes in stability, which occur
for continuous-time systems also, are called bifurcations and are discussed in
Section 4.4.)
It is now possible to establish the following important result, which we shall
use often (we state it here without proof). It is valid not only for the first-
order difference equations we have been studying here but also for systems of
difference equations, and indeed for many other kinds of dynamical systems,
including differential equations, which we shall meet in Chapter 3, and systems
of differential equations, which we shall meet in Chapter 5.

LINEARIZATION THEOREM: If y∞ is an equilibrium of the dif-

ference equation yk+1 = g(yk ), and if all solutions of the linearization
uk+1 = g ′ (y∞ )uk at this equilibrium approach zero, then the equilibrium
y∞ is locally asymptotically stable. If the linearization has unbounded so-
lutions, then the equilibrium y∞ is unstable.

Example 2.
Show that if 1 < r < 3, the equilibrium y = 0 of the logistic difference equation
 yk 
yk+1 = ryk 1 −
K
is unstable and the equilibrium y = K(1 − 1r ) is asymptotically stable. Show
also that if r > 3 there is no asymptotically stable equilibrium.
 Difference Equations (Discrete Dynamical Systems) 41

Solution: As we have seen in Example 1, the linearization at the equilibrium

y = 0 is uk+1 = r uk . From Section 2.1 (see the boxed statement on page 28),
we know that all solutions of this linearization approach zero if and only if
−1 < r < 1. Thus if r > 1 the linearization has unbounded solutions, and by
the linearization theorem the equilibrium y = 0 is unstable.
The linearization at the equilibrium y = K(1 − 1r ) is uk+1 = (2 − r)uk ,
and as we have seen in Section 2.1 all solutions approach zero if and only
if |2 − r| < 1, or −1 < 2 − r < 1, or 1 < r < 3. Thus the equilibrium
y = K(1 − 1r ) is asymptotically stable if 1 < r < 3 and unstable if r > 3.
Since both equilibria y = 0 and y = K(1 − r1 ) are unstable if r > 3, there is
no asymptotically stable equilibrium if r > 3. 

According to the linearization theorem, we can decide the asymptotic

stability or instability of an equilibrium by examining the behavior of solu-
tions of the linearization at the equilibrium. In Section 2.1 we analyzed com-
pletely the behavior of solutions of the linear homogeneous difference equation
uk+1 = r uk (which every linearization resembles): if |r| < 1, every solution
approaches zero, and if |r| > 1, there are unbounded solutions. This fact
together with the linearization theorem gives the following result.

EQUILIBRIUM STABILITY THEOREM: Let y∞ be an equilibrium

of the difference equation yk+1 = g(yk ). If |g ′ (y∞ )| < 1, the equilibrium is
asymptotically stable, and if |g ′ (y∞ )| > 1, the equilibrium is unstable.

We have used the linearization theorem and our knowledge of how linear
difference equations behave to develop the stability theorem, because this
approach extends naturally to systems of difference equations. For the first-
order difference equations we are studying here, however, there is also a simple
direct approach given in Appendix D which uses the Mean Value Theorem.
It is also possible to refine this theorem and show that the approach to
an asymptotically stable equilibrium y∞ is monotone if 0 < g ′ (y∞ ) < 1 and
oscillatory if −1 < g ′ (y∞ ) < 0. The truth of this is suggested by the cobweb-
bing method (cf. Figures 2.5 and 2.6). For instance, we saw in Example 2 that
the equilibrium y∞ = K(1 − 1r ) of the logistic difference equation is asymp-
totically stable if 1 < r < 3. As g ′ (y∞ ) = 2 − r, it follows that solutions of
the logistic difference equation approach y∞ monotonically if 1 < r < 2 and
with oscillations if 2 < r < 3. It is natural to conjecture that if r > 3, so
that there is no asymptotically stable equilibrium, all solutions oscillate but
without approaching an equilibrium.
This is a different sort of question than asking what happens near a partic-
ular equilibrium; it changes the scope of our study from local to global. Once
the local stability of each equilibrium of a model has been determined, one
can put the results together to form a picture of the overall model behavior.
42 Dynamical Systems for Biological Modeling: An Introduction

In some cases, there is only one stable equilibrium, and within the state space
(the set of possible values for our variable(s)) that equilibrium may then be
not only locally stable but globally stable, that is, all solutions approach it no
matter what the initial condition. For example, when 0 < r < 1, the logistic
difference equation has only one equilibrium, at 0, and it is locally stable. We
can also see directly by inspection that yk+1 < yk when r < 1, and in fact in
this case the equilibrium y = 0 is globally stable. In biological terms, the pop-
ulation’s maximum reproductive ratio r is so low that the population cannot
sustain itself. Then, when 1 <
r < 3, the equilibrium y = 0 is unstable, but
the equilibrium y = K 1 − 1r exists and is locally stable. In this case, it can
be shown that the latter equilibrium is globally stable.
There are other, more complicated possible scenarios for a model’s global
behavior. One possibility, which we shall explore further at the end of this
section, is multiple locally stable equilibria. Another, which corresponds to
the logistic difference equation with r > 3, is no stable equilibria at all. We
shall return to this question in Section 2.6. For now, however, let us consider
one more abstract example of local stability analysis before applying this
technique to a biological problem.

Example 3.
Determine the asymptotic stability of each equilibrium of the Verhulst differ-
ence equation
r yk
yk+1 = .
yk + A
ry rA
Solution: We have g(y) = y+A , g ′ (y) = (y+A) 2 . For the equilibrium y = 0,

since g ′ (0) = r/A, the equilibrium is locally asymptotically stable if r < A

and unstable if r > A. For the equilibrium y∞ = r − A, which is biologically
significant only if r > A, the equilibrium is locally asymptotically stable if
r > A since g ′ (y∞ ) = A/r. Thus there is always exactly one (globally) asymp-
totically stable equilibrium, y = 0 if r < A and y = r − A if r > A. The
reader should compare this approach with the cobwebbing analysis given in
Example 1 of Section 2.2, and Figures 2.8 and 2.9. 

2.3.2 A problem in population genetics

We will now use a difference equation to model how the distribution of two
different alleles for the same gene changes from one generation to the next in a
diploid organism (which has two copies of each gene). This discussion follows
that in Segel,4 Chapter 3, and Maynard Smith.5
We must first make some assumptions in order to define the problem more
4 Lee A. Segel, Modeling dynamic phenomena in molecular and cellular biology, Cam-

bridge University Press, Cambridge, 1984.

5 J. Maynard Smith, Mathematical ideas in biology, Cambridge University Press, Cam-

bridge, 1968.
 Difference Equations (Discrete Dynamical Systems) 43

precisely. There are several factors which might influence the distribution of
alleles from one generation to the next, including: the number of alleles for
the gene under study; the effect of genotype (which alleles a single individual
has) on viability, the proportion6 of individuals who survive to reproduce; the
effect of genotype on mating preferences; the effect of genotype on fertility, the
number of offspring an average individual produces; the (potentially genotype-
dependent) mutation rate from one allele to another; and possible gender
asymmetries. Let us examine here the role of genotype-dependent viability
in shaping the distribution of alleles over many generations. To simplify our
discussion, we will then assume in what follows that
(1) the gene under study has two distinct alleles, A and a.
There are therefore three possible genotypes a given individual might have:
homozygous AA, heterozygous Aa, and homozygous aa. We further assume
that
(2) genotype affects viability (some genotypes make individuals
more likely to survive and reproduce than others), in the same
manner for each generation.
In order to examine only the effects of this particular factor, we will also
assume that the other potential influences do not play a role here; that is,
that
(3) fertility is genotype-independent;
(4) mating preferences are genotype-independent;
(5) the mutation rate for all alleles is zero;
(6) all effects are gender-independent.
Note also that we are implicitly assuming distinct generations, with repro-
duction occurring only within a given generation (i.e., no mating between
generations); this is appropriate only for some populations, although the gen-
eral conclusions we draw can be applied more broadly.
Following assumption (2) we will incorporate differential viability by defin-
ing three parameters, φAA , φAa , φaa , each of which measures the proportion
of individuals of the given genotype who survive to reproduce. φAA , φAa , φaa
will take on values between 0 and 1 and will be assumed not to vary from
one generation to the next. In order to measure the frequency of each allele,
we define our variable pn to be the proportion of alleles in the nth generation
which are A, and qn = 1−pn to be the complementary proportion, the relative
frequency of a alleles.
Now, in order to calculate the relative frequencies pn+1 and qn+1 in the
next generation, we observe that each new individual receives one allele from
6 Although we might more intuitively speak of the likelihood or probability of surviving

to reproduce, this is a deterministic model, so we must instead speak of proportions.

44 Dynamical Systems for Biological Modeling: An Introduction

each parent. From assumption (4) alleles are mixed at random, that is, in
proportion to the relative frequencies of the gametes (eggs and sperm) pro-
duced (which are, in turn, the same as the relative frequencies of the alleles,
by assumption (3)). Therefore a proportion pn pn = p2n of the new individuals
receive A alleles from both parents, a proportion pn qn receive an A allele from
the mother and an a allele from the father, a proportion qn pn receive an a
allele from the mother and an A allele from the mother, and a proportion
qn qn = qn2 receive a alleles from both parents. (The reader can verify that
these four terms sum to (pn + qn )2 = 1, thereby accounting for the entire
next generation.) Note that the relative frequency of heterozygotes Aa in the
next generation is 2pn qn , since the result (the zygote) is the same regardless
of which allele came from which parent.
The genotype-dependent viability of assumption (2) means that of those
AA-homozygotes born into generation n + 1, only a proportion φAA of them
survive to reproduce, that is, a proportion φAA p2n of the entire generation
of newborns. Likewise, a proportion φAa 2pn qn become mature heterozygotes,
and a proportion φaa qn2 become mature aa-homozygotes. Since each AA indi-
vidual has two A alleles, and each Aa individual has one, the total frequency
of A alleles is now
fA = 2 φAA p2n + (φAa 2pn qn ) .

Likewise, the total frequency of a alleles in generation n + 1 is

fa = 2 φaa qn2 + (φAa 2pn qn ) .

Therefore, finally, the relative frequency of the A allele in generation n + 1 is

2 φAA p2n + (φAa 2pn qn )

fA
pn+1 = =
fA + fa 2 (φAA p2n ) + 2 (φAa 2pn qn ) + 2 (φaa qn2 )
φAA p2n + φAa pn (1 − pn )
= , (2.20)
φAA pn + 2φAa pn (1 − pn ) + φaa (1 − pn )2
2

and the relative frequency of the a allele in generation n+ 1 is qn+1 = 1 − pn+1

(since p + q = 1 in each generation, we really only need keep track of one of
the two).
We can simplify equation (2.20) by reducing the number of parameters to
two. We do so by defining relative viabilities
φAA φaa
F = , G= .
φAa φAa
F gives a ratio of AA viability to Aa viability; 0 < F < 1 means a AA zygote
has less viability than an Aa (φAA < φAa ); F > 1 means an AA zygote has
higher viability than an Aa. Likewise G gives the viability of an aa zygote
relative to an Aa. Effectively, the Aa viability has been made the yardstick
by which others are measured. (The choice of Aa was arbitrary — we could
instead choose one of the other two — but this choice preserves a certain
 Difference Equations (Discrete Dynamical Systems) 45

symmetry.) If we now divide the numerator and denominator in (2.20) by

φAa , we obtain
F p2n + pn (1 − pn )
pn+1 = . (2.21)
F p2n + 2pn (1 − pn ) + G(1 − pn )2
We have now formulated a model that tracks the relative frequency of the
two alleles across generations. As it is highly nonlinear, however, we cannot
solve (2.21) for a non-recursive formula to determine pn as a function of F , G,
and the initial condition p0 . We will therefore resort to a qualitative analysis
as developed earlier in this section.
We begin by finding equilibria. These are solutions p to the equation
F p2 + p(1 − p)
p= . (2.22)
F p2 + 2p(1 − p) + G(1 − p)2
Some algebra shows that the three equilibria are p = 0 (everyone is AA),
p = 1 (everyone is aa), and p = F G−1
+G−2 (some of both). The last equilibrium,
of course, only makes sense as long as it is in the interval [0,1]. Further algebra
(Exercise 8) shows that the last equilibrium has a value between 0 and 1 only
when F and G are either both greater than 1 or both less than 1.
To determine the local stability of each equilibrium, we linearize (2.21)
about each one in turn. To apply the equilibrium stability theorem stated
earlier in this section, we need the derivative g ′ (p) of the right-hand side of
the difference equation pn+1 = g(pn ). It is
(F + G − 2F G)p2 + 2(F − 1)Gp + G
g ′ (p) = 2 .
[F p2 + 2p(1 − p) + G(1 − p)2 ]
 
G−1 F +G−2F G
This gives g ′ (0) = 1/G, g ′ (1) = 1/F , and g ′ F +G−2 = 1−F G . There-
1
fore, the linearization about the equilibrium p = 0 is un+1 = G un , and by our
equilibrium stability theorem the equilibrium p = 0 is locally asymptotically
stable if 1/G < 1, that is, if G > 1. Likewise, the equilibrium p = 1 is locally
asymptotically stable if F > 1. These are both biologically reasonable, as we
shall explain shortly.
Determining the stability condition for the third equilibrium, p∞ =
G−1
F +G−2 , is a little more complicated, but is simplified if we recall that this
equilibrium only makes sense when either F < 1, G < 1 or F > 1, G > 1. The
criterion for stability here is
 
′ G−1 F + G − 2F G
g = < 1. (2.23)
F +G−2 1 − FG
If F < 1, G < 1, then both the numerator and the denominator of g ′ (p∞ )
are positive,7 and (2.23) becomes F + G − 2F G < 1 − F G, which simplifies
7 F G < 1, so the denominator is positive, and the numerator is F + G − 2F G = (F −

F G) + (G − F G) = F (1 − G) + G(1 − F ), also positive.

46 Dynamical Systems for Biological Modeling: An Introduction

to G < 1. Thus the third equilibrium is always locally asymptotically stable

in this region. The other possibility is that F > 1, G > 1. In this case, both
the numerator and the denominator of g ′ (p∞ ) are negative, so (2.23) becomes
2F G − F − G < F G − 1, which also simplifies to G < 1. Thus the third
equilibrium is always unstable in this region.

G
6
p = 0 stable
p = 0 stable p∞ unstable
p = 1 unstable p = 1 stable
p → 0, a wins
IC selects winner
1
p = 0 unstable p = 0 unstable
p∞ stable p = 1 stable
p = 1 unstable p → 1, A wins
p → p∞ , coexist
-F
0 1

FIGURE 2.14: Stability of equilibria for equation (2.21), in terms of the

parameters F and G.

We can now put these results together to form a complete picture of the
model’s behavior. There are four cases or regions in terms of F and G, as
depicted in Figure 2.14. When F < 1 and G > 1 (which we can also write as
φAA < φAa < φaa , so that AAs are less viable than Aas, which are in turn
less viable than aas), the equilibrium at 0 is unstable, and the equilibrium at
1 is (globally) stable, so that pn → 0 as n → ∞, that is, the proportion of
A alleles dwindles to zero from one generation to the next, until eventually
almost everyone is aa homozygous. Similarly, when F > 1 and G < 1, so
that φaa < φAa < φAA and the A allele contributes more to viability than
the a allele, then the equilibrium at 0 is unstable, the equilibrium at 1 is
(globally) stable, and pn → 1 as n → ∞. In this case, eventually nearly
everyone is AA homozygous. Note that in both of these two cases, one allele
clearly contributes more to viability than the other, so it is not surprising
that it comes to dominance. (Also note that in both cases, the coexistence
equilibrium p∞ is not present.)
When F < 1 and G < 1, both AA and aa genotypes are less viable than
the heterozygous Aa genotype. This is the case, for example, in Africa with
regard to the allele for sickle cell anemia. This allele provides some immunity
against malaria, but homozygotes with two copies of the allele develop sickle
cell anemia and often do not survive to maturity. Thus it is beneficial to have
precisely one copy of the allele. In this scenario, the equilibria at 0 and 1
are both unstable, but the coexistence equilibrium is (globally) stable. Here
 Difference Equations (Discrete Dynamical Systems) 47

all three genotypes persist, although both homozygotic genotypes are at a

disadvantage.
Finally, when F > 1 and G > 1, both homozygotic genotypes are more
viable than the heterozygote. Here the equilibria at 0 and 1 are both locally
stable, but the coexistence equilibrium in between them is unstable: if p0 <
p∞ , then p1 will be even smaller, while if p0 > p∞ , then p1 will be even
larger. The unstable coexistence equilibrium acts as a boundary to separate
those initial conditions (ICs) for which pn → 0 as n → ∞ from those for
which pn → 1 as n → ∞. That is, if p0 < p∞ , then the a allele wins out,
while if p0 > p∞ , then the A allele wins out. In either case, eventually nearly
everyone is homozygous with two copies of the “winning” allele. This is our
first instance of competition, which we shall study more in later sections. As
Segel observes (p. 48), if we have multiple isolated habitats in which these
same criteria hold (F, G > 1, so heterozygotes are at a disadvantage) but with
different initial conditions, then we may expect to find the A allele “winning”
in some of them, and the a allele dominating in others, so that this model also
provides a possible explanation for genetic variability across similar isolated
habitats.
Our global stability analysis here is not entirely rigorous, in that we have
not proven conclusively that the simple assembly of local stability results pro-
vides a complete picture of the model’s global behavior. Quantitative analysis,
by having a computer solve the equation for particular values of the parame-
ters F and G and particular initial conditions p0 , can further boost our intu-
itive understanding and confidence that this global picture is correct. We will
forego further discussion of other possible global behaviors until Section 2.6
(but see Exercise 9).
Note, finally, how our decision to investigate the selective effects of
genotype-dependent viability informed and shaped the assumptions that we
made and the model we analyzed. This focused research question and the
subsequent assumptions help us to frame and delimit the extent to which our
conclusions (namely, that simple differential viability can account for either
exclusive selection of one genotype or an equilibrium distribution of coexis-
tence) hold. To complete the modeling cycle discussed in the previous chapter,
we must evaluate the extent to which our interpreted conclusions provide a
sufficient answer to our original question. Several of the exercises below inves-
tigate other factors for selection of alleles.

Exercises
In Exercises 1–6, find all positive equilibria of the given difference equation,
and for each equilibrium determine whether it is asymptotically stable or
unstable. If the difference equation involves one or more parameters, your
answer may depend on the range of values of the parameters.
48 Dynamical Systems for Biological Modeling: An Introduction

1. yk+1 = 2yk − yk2 + 2 4. yk+1 = ryk (1 − yk )

2. yk+1 = 12 yk + yk2 − 7
144 5. yk+1 = ryk e1−yk
1
∗
3. yk+1 = αyk + β 6. yk+1 = ryk (1 + yk )− 2

r y2
7. Find all positive equilibria of the difference equation yk+1 = y2 +Ak
from
k
Example 2 of Section 2.2, and for each equilibrium determine when (in
terms of r and A) it is locally asymptotically stable or unstable. Compare
your results to those given for the cobwebbing approach.
8. Solve the equilibrium condition (2.22) and show that the coexistence
G−1
equilibrium p∞ = F +G−2 can only exist when F and G are either both
greater than 1 or both less than 1.
9. Suppose that, in the genetics discussion in this section, survival is
genotype-independent, i.e., φAA = φAa = φaa . Show that the fre-
quency of each allele is then constant across generations, establishing
the Hardy-Weinberg law of genetics (so named for its independent dis-
coveries by G.H. Hardy and Wilhelm Weinberg in 1908 [33, 104]), as
follows: (a) Find the resulting values of the relative viabilities F and G.
Where does this place the scenario relative to Figure 2.14? (b) Simplify
equation (2.21).
10. Suppose that, in the genetics discussion in this section, we are studying a
gene for which aa homozygotes never reproduce (they are either infertile
or do not survive to maturity). The a allele may nevertheless survive
among heterozygotes. In this case φaa = 0, and thus G = 0. (Note also
in this case we must have p0 > 0, or else the entire population consists
of aa homozygotes, and there is no next generation.)
(a) How do equation (2.21) and its behavior change when G = 0?
(b) Interpret biologically the criterion developed in (a) for survival of
the a allele when G = 0.
(c) Show that the G = 0 model in (a) guarantees pn > 1/2 for n ≥ 1.
Explain biologically why this must be so.
(d) Sketch (by hand or by computer) a graph of the remaining equi-
libria p as a function of the remaining parameter F . This type of
graph is called a bifurcation diagram, and we shall discuss bifurca-
tion diagrams further in Section 2.6, and in later chapters.
∗
(e) Have a computer generate a three-dimensional bifurcation dia-
gram for the model (2.21) in terms of both parameters F and G.
(Segel8 , p. 50) If we reverse assumption (2) and suppose that all
8 Lee A. Segel, Modeling dynamic phenomena in molecular and cellular biology, Cam-

bridge University Press, Cambridge, 1984.

 Difference Equations (Discrete Dynamical Systems) 49

three genotypes are equally viable, what happens to the model

(2.21)? Explain this result in biological terms.
11. Suppose that, in the discussion on population genetics in this section,
we relax assumption (3) and allow fertility to be genotype-dependent.
Again following Segel (pp. 49–50), define the average number of gametes
produced by an adult of each genotype to be 2mAA , 2mAa , and 2maa ,
evenly divided among eggs and sperm, and in each case half of which
correspond to each of the parent’s two alleles.
(a) Develop a difference equation for pn+1 in terms of pn analogous to
equation (2.21) which models the distribution of alleles under an
assumption of genotype-dependent fertility.
(b) By comparing this new model with equation (2.21), show that the
behavior of the new model is qualitatively the same as that of
(2.21), the only difference being different expressions for the fitness
ratios K and L.
12. Segel also considers (p. 47) the possibility of mutation, rather than dif-
ferential viability, driving the distribution of genotypes. In terms of the
discussion in this section, assumptions (2) and (5) are reversed (and all
the φ’s are the same).
(a) Segel suggests imagining mutations only in one direction, say from
A to a. If a proportion µ of the A gametes in any generation mutate
to a, then what proportion of the A gametes remain? If we assume
all genotypes have the same viability, what is the corresponding
equation for pn+1 in terms of pn ? (It should be linear.)
(b) Solve the new model outright, and use the result to estimate how
many generations it will take for this mutation to reduce the fre-
quency of the A allele from 0.5 to 0.05, if µ has a value of 10−6 per
generation.
(c) Now suppose that back-mutations also occur, from a to A, with an
average proportion µ′ of the a gametes in any generation mutating
to A. Write a difference equation that incorporates both mutations.
Give a qualitative analysis of its behavior, assuming that µ > µ′ .
(d) Choose values of F and G which lead to the disappearance of the A
allele (i.e., F < 1 < G), and have a computer calculate solutions of
(2.21) using an initial condition of p0 = 0.5. How many generations
are necessary to reach a frequency of 0.05? Compare this number
to that obtained in (b) above for mutation alone to accomplish the
same result. Which selection mechanism appears to operate more
quickly? How does this compare with currently held theory?
∗
(e) Now write a difference equation that incorporates both differential
viability and A → a mutation. To get you started, first reconsider
50 Dynamical Systems for Biological Modeling: An Introduction

the question, what proportion of new individuals receive an A allele

from one parent? (It is now less than pn .)
∗
13. If we reverse assumption (4) and instead assume extreme like-with-
like mating preferences, so that individuals only mate with others of
the same genotype, then we must keep track of the relative frequencies
of each genotype, rather than those of alleles. If we define αn and βn ,
respectively, as the relative frequencies of AA homozygotes and aa ho-
mozygotes in generation n, then the relative frequency of Aa heterozy-
gotes must be (1 − αn − βn ).

(a) Assuming random mixing of alleles within each genotype, write

difference equations for αn+1 and βn+1 , each in terms of αn , βn ,
F , and G. (We will not, however, discuss the qualitative analysis
of systems of two or more difference equations until Section 2.7.)
(b) Show that the relative frequency of heterozygotes (1−αn −βn ) ≤ 21
for all n ≥ 1, regardless of initial conditions. Explain biologically
why this must be so.

2.4 Intraspecies competition

In discussing the linear homogeneous difference equation yk+1 = ryk as
a model for population growth in Section 2.1, we observed that the model
becomes inadequate as soon as the population grows enough to begin encoun-
tering the resource limitations of its habitat, whether on a cellular scale or
a continental scale. Various examples were given of functions that have been
proposed to address the effect of resource limitations on population growth for
different populations. All these models must address, even if only implicitly,
the way in which members of the population compete with each other when
resources are scarce. This type of competition is called intraspecies competi-
tion, and is distinct from the interspecies competitionwe shall study when we
introduce systems of interacting populations in Section 2.7 and in Chapter 6.
There are two general assumptions one might make about the way re-
sources are distributed when there are not enough for everyone. Scramble com-
petition, in which resources are divided evenly among all members, reflects an
assumption that all individuals are equally capable of capturing resources for
themselves, with the result that everyone suffers equally when resources are
inadequate. In this case, reproduction may be severely curtailed, or even cease
altogether, if no one is getting enough food to reproduce reliably. An example
of a population observed to exhibit scramble competition is that of Nicholson’s
blowflies. In 1954 A.J. Nicholson conducted experiments on Lucilia cuprina,
the Australian sheep blowfly, in which he observed large oscillations in the
 Difference Equations (Discrete Dynamical Systems) 51

Reproduced by permission of CSIRO Publishing

from http://www.publish.csiro.au/nid/90/paper/ZO9540009.htm

FIGURE 2.15: Sustained fluctuation in a blowfly (Lucilia cuprina) popula-

tion with constant food supply, as observed by Nicholson (1954, Figure 3).

population size over time. Nicholson discovered9 that the blowfly population,
which emerged in distinct (or nearly distinct) generations, was controlled en-
tirely by the rate at which food was provided, so that (when this rate was held
constant) a small generation of blowflies, encountering no resource limitations,
produced a large next generation. The next generation, however, was so large
that each individual had a minuscule share of the food, and consequently
very few were able to develop eggs, resulting in a very small next generation.
These oscillations were sustained over a period of months (see Figure 2.15),
and others have since observed similar fluctuations in fly populations.
Contest competition, on the other hand, assumes that some individuals —
the most fit — are able to secure enough resources for themselves, while the less
fit may end up with nothing. Here we may imagine a ranking in which resources
are distributed to individuals in decreasing order of fitness, until resources are
exhausted. Under contest competition, some number of individuals (as many
as resources can support) will always be able to reproduce. For example, if we
consider the number of trees of a given species in a certain patch of woods,
often the tall trees spread their branches to catch as much sunlight as needed
for growth through photosynthesis, crowding out smaller trees nearby (which
then stop growing or may even die) until each tall tree’s foliage meets that
of its neighbors, allowing little sunlight to reach the ground. New trees then
grow only when a large tree falls, creating an opening.
The differences between these two assumptions will lead naturally to dif-
ferences in the mathematical models we develop to describe such populations.
In particular, scramble competition is modeled with reproduction functions
which rise to a maximum and then fall off, as excessive size reduces repro-
9 A.J. Nicholson, An outline of the dynamics of animal populations, Australian J. Zoology

3: 9–65, 1954.
52 Dynamical Systems for Biological Modeling: An Introduction

scramble repr. fn. contest repr. fn.

pop. size pop. size

FIGURE 2.16: Scramble (left) and contest (right) competition reproduction

functions.

duction. In some cases reproduction may reach zero for some finite carrying
capacity, as with the logistic model (2.12), while in others it may taper off
gradually, as with Ricker’s model (first seen in the exercises at the end of
Section 2.2 and discussed in more detail below). Contest competition, mean-
while, is modeled with reproduction functions which rise monotonically and
flatten out as they reach the ceiling created by the resource limitations. The
Verhulst model (2.10) incorporates contest competition. Figure 2.16 sketches
the general form of these two types of functions.
We can consider the consequences of each type of intraspecies competition
via the properties described above of the corresponding reproduction func-
tions. In either case we will also require that the reproduction function g(y)
in the difference equation yk+1 = g(yk ) have the following properties:
• For small populations y, reproduction should be roughly linear in y,
corresponding to a constant (maximum) per capita ratio r called the
intrinsic growth rate, i.e.,

g(y)
lim = r > 0.
y→0 y

• For large populations, reproduction should be constrained, i.e., g(y) is

bounded.
(The reader can verify that all models described in this section have these two
properties.) Since, as we saw in Section 2.2, equilibria correspond graphically
to intersections of the reproduction curve z = g(y) (which is bounded by
assumption) with the line z = y (which grows unbounded), it follows that
there must be a largest equilibrium y∞ , at which the line z = y crosses the
curve z = g(y) for the last time, from below to above. At this point, the slope
of z = y (which is 1) must therefore be greater than the slope (derivative) of
z = g(y). That is, at this largest equilibrium y∞ , we have g ′ (y∞ ) < 1.
 Difference Equations (Discrete Dynamical Systems) 53

If, in addition, we assume contest competition, with monotone increasing

reproduction, then we also have that g ′ (y) ≥ 0 for all y, and more specifically
g ′ (y∞ ) ≥ 0. Therefore 0 ≤ g ′ (y∞ ) < 1, satisfying the stability condition
|g ′ (y∞ )| < 1, so that under contest competition there must always be at
least one stable positive equilibrium (in a discrete-time model). In fact, if we
consider that for any additional equilibria between 0 and y∞ , the crossing of
the reproduction curve by the line z = y alternates between below-to-above
and above-to-below (as y increases), then the stability argument above can
also be applied to every second equilibrium to the left of y∞ .
On the other hand, if we assume scramble competition, so that reproduc-
tion rises and then falls as y increases, then the last crossing of z = g(y) by
z = y may occur on either the increasing or the decreasing part of the curve,
depending on the value of the intrinsic reproduction ratio r. If r is small, then
z = y will outstrip z = g(y) more quickly, and the last equilibrium y∞ will
occur on the increasing part of the curve, leading to an argument like that for
contest competition in which 0 ≤ g ′ (y∞ ) < 1 and the equilibrium is stable.
However, for large r the last equilibrium y∞ will occur on the decreasing part
of the curve z = g(y), so that g ′ (y∞ ) < 0. In this case the previous conclusion
that g ′ (y∞ ) < 1 no longer helps bound the derivative, and it is quite possible
that g ′ (y∞ ) < −1, which implies instability for y∞ .
The logistic equation yk+1 = ryk (1 − yk ), introduced in Section 2.1, is
an example of a population model with scramble competition. (Another form
sometimes used for the logistic equation is yk+1 = yk + r yk (1 − yk ), but we
shall use that given above throughout this chapter.) In Section 2.6 we will
describe the behavior of a difference equation such as the logistic equation
with r > 3 (cf. Section 2.3, Example 2) modeling scramble competition and
having no asymptotically stable equilibrium.
We can therefore conclude that populations with contest competition al-
ways reach an asymptotically stable [nonzero] equilibrium, but populations
with scramble competition have an asymptotically stable [nonzero] equilib-
rium only if the intrinsic growth rate is sufficiently small (but still large
enough for the equilibrium to exist in the first place, cf. the linear equations
of Section 2.1). We can interpret this idea intuitively as saying that in contest
competition, those who obtain enough resources to reproduce will ensure a
subsequent generation which is as large as possible with the given resources,
while in scramble competition too high an intrinsic reproduction rate would
result in the division of available resources so many ways that reproduction
would be severely curtailed, thereby setting the stage for sustained oscilla-
tions. This principle is robust: it does not depend on the specific form of the
reproduction function but only on the qualitative assumptions on the nature
of the competition.
We will now consider two models that have been proposed for different fish
populations, one representing each type of intraspecies competition.
54 Dynamical Systems for Biological Modeling: An Introduction

2.4.1 Two metered fish models

Many species of fish have an annual cycle in which all new fish are born in
a short time span. It is reasonable to assume that the number of newborn fish
is a function of the adult stock which has survived to the birth season and the
newly matured fish, and this total will depend on the adult population from
the previous year. Thus a difference equation model is appropriate. However,
the survivorship from one year to the next is a continuous process, and the
form of the model will depend on this process. This type of model, in which
continuous short-term dynamics shape the long-term dynamics of discrete
generations, is called a metered model. We shall now describe two frequently
used metered models for different species of fish, although we will not be able
to establish the appropriate survivorship until Section 3.4.
Some fish, such as haddock and North Atlantic plaice, have very high
fertility rates but low survivorship to adulthood. For such fish, it is appropriate
to assume a linear (rather than constant) per capita mortality rate between
spawning seasons, and this assumption leads to the Beverton-Holt model10
ayk
yk+1 = (2.24)
1 + byk
where a and b are positive constants. This equation, which corresponds to
contest competition, is the same as the Verhulst equation introduced in Sec-
tion 2.1 with a replacing Ar and b replacing A 1
. From our study of the Verhulst
equation (Section 2.3, Example 3), we now see that there is one asymptot-
ically stable equilibrium, namely y = 0 if a < 1 and y = a−1 b if a > 1. In
this case, the rise in per capita mortality with population density creates con-
test competition, with the result that no matter how many fish are born in a
given year, only a certain number survive to the next year. (In addition, the
reproductive ratio a = g ′ (0) must be high enough — a > 1 — to keep the
population from dying out altogether. The parameter b, which gives a mea-
sure of the size of the density-dependent per capita mortality rate relative to
the density-independent per capita mortality rate, only affects the size of the
eventual [positive] equilibrium.)
Some species of fish, notably salmon, habitually cannibalize their eggs
and young. Now the appropriate assumption is instead that the birth rate is
proportional to the adult population size and that there is a per capita death
rate proportional to the initial size of the young population. This assumption
leads to the Ricker model11
yk+1 = ayk e−byk (2.25)
with a and b again positive constants. Here a is the constant of proportion-
ality in the birth rate, while b measures the cannibalism rate due to adults.
10 R.J.H. Beverton and S.J. Holt, The theory of fishing, Sea fisheries: their investigation

in the United Kingdom (M. Graham, ed.), Edward Arnold, London, 1956, pp. 372–441.
11 W.E. Ricker, Stock and recruitment, J. Fisheries Research Board Canada 11: 559–623,

1954.
 Difference Equations (Discrete Dynamical Systems) 55

Photo by Bernard E. Picton and Christine C. Morrow Ken Gray Image Courtesy of Oregon State University

FIGURE 2.17: The bottom-feeding plaice (left), hard to see against the
sea floor except when they move, have high fertility and low survivorship to
maturity, leading to contest competition. The Mediterranean flour moth or
mill moth (right), seen in its characteristic resting pose, undergoes different
types of intraspecies competition during its life cycle as a result of parasitism.

This model, corresponding to scramble competition (all young have a reduced

chance of survival), exhibits behavior similar to that of the logistic model.
Equilibria of the Ricker model are given by y = 0 and ae−by = 1. Thus there
is an equilibrium y = 0 describing extinction of the population and possibly
an equilibrium describing survival given by eby = a, or y∞ = logb a (by log
we mean the natural logarithm). This possible survival equilibrium is positive
only if a > 1. For the Ricker model, g(y) = aye−by , g ′ (y) = ae−by (1 − by).
Thus g ′ (0) = a, and the equilibrium y = 0 therefore is asymptotically stable
if and only if a < 1. Also, if a > 1, so that the survival equilibrium y∞ makes
biological sense, then the condition ae−by∞ = 1 makes

g ′ (y∞ ) = 1 − by∞ = 1 − log a.

Thus the equilibrium y∞ is asymptotically stable if and only if −1 < 1−log a <
1. This condition reduces to log a < 2, or a < e2 . The result suggests that fish
which cannibalize their young and have a high birth rate (a > e2 ) may have
an unstable equilibrium. We shall see in Section 2.6 that this may indicate
oscillations in population size. Oscillations in population size can lead to times
at which the population size is very small, so that a small perturbation could
lead to wiping out the population. Fish populations which satisfy a Ricker
model may be quite vulnerable to external influences.
In this and later sections, we will use the Ricker form g(y) = aye−by as a
ry
prototypical scramble recruitment function and the Verhulst form g(y) = y+A
as a prototypical contest competition recruitment function.
56 Dynamical Systems for Biological Modeling: An Introduction

2.4.2 Between contest and scramble

The equation
b
yk+1 = ayk (1 + yk )−n (2.26)
n
with a, b > 0 and n > 1, introduced by Hassell12 to model intraspecific larval
competition in insects, is intended to describe a situation between the extremes
of contest competition — n = 1, which gives the Beverton-Holt model —
and scramble competition — n → ∞, which gives the Ricker model since
limn→∞ (1+ nb yk )−n = e−by . Exercises 9–12 below develop some of the analysis
of this model. Maynard Smith and Slatkin13 proposed a similar model to fill
in the spectrum between contest and scramble competition:
ayk
yk+1 = (2.27)
1 + (byk )n

with a, b > 0 as before. Here, as in the Hassell model, n = 0 corresponds

to no competition (yk+1 = a2 yk ), n = 1 corresponds to contest competition
(Beverton-Holt), and n → ∞ corresponds to scramble competition. Some
researchers have used statistical techniques such as least-squares regression
to fit observed data to such models and take the best-fit value for n as a
measure of the amount and type of intraspecies competition exhibited by the
population under study.
In fact, some species of insects have been found to exhibit different types of
competition in different parts of their life cycles, or depending on interactions
with other species or the environment. Lane and Mills found14 that parasitism
can actually change the nature of intraspecific competition from scramble to
contest in some insects — in particular, the Mediterranean flour moth Ephesia
kuehniella shifts from scramble competition prior to parasitization to contest
competition following it, since parasitized larvae are not only poor competi-
tors but also more vulnerable to cannibalism than unparasitized larvae. Reeve,
Rhodes and Turchin, meanwhile, found15 that the southern pine beetle Den-
droctonus frontalis shifts from contest competition early in its life cycle (when
adults attack a tree to lay eggs in galleries) to scramble competition later
on (brood survivorship), with the latter dominating. Flexible models such as
those of Hassell (1975) and Maynard Smith and Slatkin (1973) can therefore
be important tools for modeling this variability.
12 M.P. Hassell, Density dependence in single species populations, J. Animal Ecology 44:

283–295, 1975.
13 J. Maynard Smith and M. Slatkin, The stability of predator-prey systems, Ecology 54:

384–391, 1973.
14 S.D. Lane and N.J. Mills, Intraspecific competition and density dependence in an Eph-

esia kuehniella–Venturia canescens laboratory system, OIKOS 101: 578–590, 2003.

15 J.D. Reeve, D.J. Rhodes and P. Turchin, Scramble competition in the southern pine

beetle, Dendroctonus frontalis, Ecological Entomology 23(4): 433–443, 1998.

 Difference Equations (Discrete Dynamical Systems) 57

Exercises
In Exercises 1–4 determine the values of r for which there is an asymptotically
stable equilibrium.
2
ryk
1. yk+1 = 2 +A
yk

2. yk+1 = ry(1 + αyk )−β , (0 < β < 1)

yk
3. yk+1 = ryk e1− K
4. yk+1 = ryk (e1−yk − d)
In Exercises 5–8 determine the range of values of H for which there is
an asymptotically stable equilibrium. [The range of values may depend
on r.]
ryk
5. yk+1 = yk +A −H
yk
6. yk+1 = yk + ryk (1 − K) −H
7. yk+1 = (0.9)yk + 0.1 − H
8. yk+1 = αyk + β with 0 < α < 1.
The following exercises deal with the Hassell model (2.26).
9. ∗
Show that limn→∞ (1 + nb yk )−n = e−by , so that the limiting case of the
Hassell model as n → ∞ is the Ricker model. [HINT: Use the relation
limn→∞ (1 + nx )n = ex .]
10. Show that (2.26) has an equilibrium y = 0 which is asymptotically stable
if a < 1.
n 1
11. Show that (2.26) has a positive equilibrium y = b (a
n − 1) if a > 1.
∗
12. Show that the positive equilibrium of (2.26) is asymptotically stable
if n(1 − a−1/n ) < 2, and in particular for all a if n < 2.
13. Take the example given in this section of crowding of trees in a forest and
develop a difference equation which models the contest competition that
regulates the density of trees in a given patch. Explain the parameters
in your model, and say how you could measure them.
14. Find the expression for the limiting (scramble competition) case of the
Maynard Smith and Slatkin model (2.27) where n → ∞, as a function
of yk . (You will need to consider three cases, depending on the value of
byk .)
15. The model in Exercise 3 is an alternate form of the Ricker model (2.25).
Show that they are equivalent.
58 Dynamical Systems for Biological Modeling: An Introduction

2.5 Harvesting
Another mechanism besides competition which regulates population sizes
is the removal of individuals by another species, usually for food. Fish eat
plankton, and aquarium owners often buy algae-eating fish to keep their tanks
clean. Wolves and coyotes hunt deer and smaller mammals in the woods of
North America; lions hunt gazelles on the plains of Africa. Humans also cut
down trees to make paper and wood for construction. In areas where humans
have removed natural predators, the prey populations often grow unchecked
and can cause damage to the environment or become a nuisance. Deer popu-
lations, for instance, in areas where wolves have been shot by farmers, rise to
the point that they run onto roads, causing traffic accidents, and wander into
residential neighborhoods and eat flowers, bushes and gardens bare. Hunting
seasons have been developed and managed as a response to this overpopula-
tion, as a means of population control. Human colonization of wild areas has
also led to other, more accidental removals of natural predator or competi-
tor species, necessitating further population management programs based on
careful study. All these types of removal and control are known as harvest-
ing, and harvesting has become a major concern in wildlife and environmental
management. Rempel and Kaufman, for instance, studied16 the effects of both
spatial and temporal tree harvesting patterns in the Nakina Forest of north-
western Ontario, Canada, on the preservation of wildlife habitat for animals
such as caribou, moose, and martens, balancing commercial timber production
objectives with a concern not to fragment forest habitats.
We shall consider here only one type of harvesting, that is constant-rate
or constant-yield harvesting, by which we mean removing a fixed number of
members just before the annual birth period. Thus a population described
in the absence of harvesting by a difference equation yk+1 = g(yk ) would
be described under harvesting by a difference equation yk+1 = g(yk ) − H,
where H is the number of individuals harvested per year. The questions of
interest to investigate involve the relationship between the harvesting rate and
the equilibrium population size. What is the maximum harvest rate that will
allow the population to persist? What is the minimum harvest rate necessary
to prevent excessive population growth? To what extent does harvesting make
a population more vulnerable to extinction from a catastrophic event such as
drought or disease? Can harvesting make a population less vulnerable to such
events? We begin by considering a simple example.

Example 1.
Consider a population governed by a logistic difference equation from which
16 R.S. Rempel and C.K. Kaufman, Spatial modeling of harvest constraints on wood sup-

ply versus wildlife habitat objectives, Environmental Management 32(5): 646–659, 2003.
doi: 10.1007/s00267-003-0056-8.
 Difference Equations (Discrete Dynamical Systems) 59

H members are removed in each generation. Determine the maximum value of

H for which the population will persist (i.e., will have a positive equilibrium),
and show that if H > 0 there can be an asymptotically stable equilibrium for
values of r greater than 3.
Solution: The population is modeled by
 yk 
yk+1 = ryk 1 − − H. (2.28)
K
y


An equilibrium of (2.28) is a solution of y = ry 1 − K − H, i.e.,
 
1 HK
y2 − K 1 − y+ = 0. (2.29)
r r

The quadratic equation (2.29) has two positive real roots

 
  s  2
1 1 1 4HK 
y = K 1 − ± K2 1 − − (2.30)
2 r r r

if and only if
 2
1 4HK
K2 1 − − ≥ 0,
r r
which we can rewrite as
(r − 1)2
H ≤ Hmax = K.
4r
This gives the range of values of H (for a given value of r) for which the
population will persist.
To determine the (local) stability of the two equilibria found in (2.30), we
recall the condition |g ′ (y∞ )| < 1 for the stability of an equilibrium y∞ of the
difference equation yk+1 = g(yk ). Here we find that for (2.28)
 
 y 2y
g(y) = ry 1 − − H, g ′ (y) = r 1 − ,
K K

and substituting (2.30) for y∞ we obtain

r
4rH
g ′ (y∞ ) = 1 ∓ (r − 1)2 − .
K
Since g ′ (y∞ ) > 1 for the smaller of the two equilibria, only the larger equilib-
rium in (2.30) can be asymptotically stable, and the condition −1 < g ′ (y∞ ) <
1 for its asymptotic stability simplifies to g ′ (y∞ ) > −1, or
r
4rH
(r − 1)2 − < 2,
K
60 Dynamical Systems for Biological Modeling: An Introduction

which reduces to
4rH
(r − 1)2 − < 4.
K
We can rewrite this in terms of H as
(r − 1)2 − 4 (r − 3)(r + 1)
H > Hmin = K= K,
4r 4r
which is satisfied for all H ≥ 0 when r < 3. Therefore, when r < 3, all
H < Hmax allow the population to persist at a (stable) positive equilibrium,
provided that the initial population is above the unstable (smaller) equilib-
rium. (Too small an initial population will go extinct from even a little har-
vesting.)
The perhaps surprising result comes from addressing the second half of
the original question. Recall (Section 2.3, Example 2) that for r > 3 the
logistic equation without harvesting has no stable equilibria. When r > 3,
however, (2.28) has a stable positive equilibrium for H in the interval Hmin <
H < Hmax . For H > Hmax , harvesting is too great, and the population is
wiped out in a finite number of generations (since there is no equilibrium). For
H < Hmin , harvesting is not strong enough to temper the high reproductive
rate, the equilibria remain unstable, and the population oscillates, as it does in
the absence of harvesting. Therefore, we see that for populations that exhibit
scramble competition (as in the logistic model) and a high fertility rate (r >
3, measured in appropriate units), it is actually possible to use harvesting
as a means of stabilizing the population, which leaves it less vulnerable to
fluctuations caused by other external or environmental factors. 

2.5.1 Fishery harvesting and graphical equilibrium analysis

Perhaps the greatest commercial application of harvesting to wildlife man-
agement is in fishery management. Fisheries around the world have often been
wiped out or reduced to extremely low levels by overfishing (over-harvesting)
in the past. An understanding of the effect of harvesting on a fishery popu-
lation model may be helpful in avoiding such problems. Here again, the most
important question is for what harvesting rates H there will be a positive
asymptotically stable equilibrium — in other words, how much harvesting
can the population sustain without dying out? In the remainder of this sec-
tion, we shall apply constant-yield harvesting to the Beverton-Holt and Ricker
models developed in the previous section. As the models become more com-
plex, we will use a graphical approach to equilibrium analysis which resembles
the setup for the cobwebbing approach developed in Section 2.2.
The harvested Beverton-Holt model is
ayk
yk+1 = − H, (2.31)
1 + byk
and its equilibria are the solutions of
ay
= y + H, (2.32)
1 + by
 Difference Equations (Discrete Dynamical Systems) 61

Photos courtesy U.S. Fish and Wildlife Service, dls.fws.gov

FIGURE 2.18: At left, fishery biologists pumping eggs from the Tenor River,
Alaska, 1966. At right, a fishery biologist returns fish to their habitat. Fishery
biologists play a key role in restoring such important species as Atlantic and
Pacific salmon, American shad, striped bass, and Great Lakes trout.

given by
 
  s 2
1 a−1 a−1 H
y∞ = −H ± −H −4 . (2.33)
2 b b b

With some algebra (see Exercise 4) √ one can show that these equilibria exist
and are positive for H ≤ Hc = ( a − 1)2 /b, a > 1. (For H = 0 these are
the equilibria 0 and a−1
b found in the previous section for the Beverton-Holt
model without harvesting.) Hc , which is the greatest harvesting yield that
allows the population to persist, is called the critical harvesting yield. Here
ay a
g(y) = 1+by − H and g ′ (y) = (1+by) ′
2 , so the criterion |g (y∞ )| < 1 for stability
√
becomes y∞ > a−1 b . With, again, some algebra, one can show (see Exercise 5)
√
a−1
that b lies between the two equilibria (2.33) when H ≤ Hc , a > 1, so
that only the larger equilibrium is stable. Thus, as with the logistic model
(Example 2.5), any amount of harvesting (H > 0) will drive a population
with low fertility (a < 1) to extinction in finite time (since there are no
equilibria). Populations with high fertility (a > 1) can survive harvesting
as long as the initial population is large enough (greater than the smaller,
unstable equilibrium) and there is no over-harvesting (i.e., as long as H < Hc ).
If we compare these results with those for the Beverton-Holt model without
harvesting derived in the previous section, we see that the effects of harvesting
on species exhibiting contest competition are threefold:
62 Dynamical Systems for Biological Modeling: An Introduction
1
z
ry
z=
z=y 1+by
z=y
0.8

0.6

z
ry
z= 0.4

1+by
0.2

y 0 0.2 0.4
y
0.6 0.8 1

FIGURE 2.19: Beverton-Holt model, FIGURE 2.20: Beverton-

a < 1, H = 0. Holt model, a > 1, H = 0.

• harvesting small populations causes extinction;

• over-harvesting causes extinction;
• extinctions occur in finite time rather than asymptotically.
Exercises 6 and 7 suggest further reflections on these issues.

Alternatively, we can use graphical methods to analyze this model. As

observed in Section 2.3, equilibria of the difference equation yk+1 = g(yk ) can
be identified graphically as intersections of the graphs of the reproduction
function z = g(y) and the line z = y. These graphs will give us a visual
approach to the analysis of this model. For (constant-yield) harvesting models,
which take the form yk+1 = g(yk ) − H, equilibria obey the equation y + H =
g(y), so we can instead view them as intersections of the graphs of z = g(y)
and z = y + H.
Equilibria of the harvested Beverton-Holt model (2.31) are intersections
ay ay
of the graphs of z = 1+by and z = y + H. The curve z = 1+by starts at the
origin with slope a and is monotone increasing, approaching the limit ab as
y → ∞. If a ≤ 1, the line z = y + H lies entirely above this curve, except
for an intersection with the curve at y = 0 if H = 0 (Figure 2.19). Thus if
a < 1 there is no equilibrium under harvesting and the population dies out.
Since this is also what happens for H = 0, it is hardly surprising that the
population cannot sustain any harvesting.
The case a > 1, in which the unharvested population has an asymptotically
stable positive equilibrium, is more interesting. In this case, the line z = y
starts below the curve at y = 0 and intersects it again at a positive value of y,
y = a−1
b (Figure 2.20). Increasing H from 0 means moving the line z = y + H
up parallel to itself (its slope remains 1, but its y-intercept H increases). For
 Difference Equations (Discrete Dynamical Systems) 63
1 1 1
z=y+H
stable
ry
z=
0.8 0.8 1+by 0.8

ry
z=
1+by
0.6 z=y+H 0.6 0.6

z z z

0.4 0.4 0.4

0.2 0.2 0.2

unstable

0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
y y y

FIGURE 2.21: FIGURE 2.22: FIGURE 2.23:

Beverton-Holt model, Beverton-Holt model, Beverton-Holt model,
a > 1, 0 < H < Hc ; two a > 1, H = Hc ; single a > 1, H > Hc ; no
equilibria. equilibrium. equilibrium.

sufficiently small H, there are two intersections, with one starting at 0 and
increasing and the other starting at a−1b and decreasing (Figure 2.21). At the
smaller intersection, the slope of the curve is greater than the slope of the line
(which is 1) and thus by the stability theorem of Section 2.4 this equilibrium
is unstable. At the larger intersection, the slope of the curve is positive but
less than 1 and so by the stability theorem this equilibrium is asymptotically
stable. This means that the limiting population size is the second equilibrium,
as long as the initial population size is greater than the first equilibrium
(otherwise the population will die out).
However, as H increases, it will reach a value where the two equilibria come
together and curve and line are tangent (Figure 2.22). If H increases beyond
this value, there will be no equilibrium (Figure 2.23). Thus the qualitative
behavior of the system will be that there is an asymptotically stable positive
equilibrium until H reaches this critical value, which we denote by Hc , and
then the equilibrium jumps to zero. A jump in equilibrium value is called a
catastrophe, because biologically it corresponds to the disaster of wiping out
the population.
√
It is also possible to use graphs to show that the critical harvest rate is
( a−1)2
and that the equilibrium population size at this critical harvest rate
√b
a−1 ay
is b . This may be shown by recognizing that, since the graphs of z = 1+by
and z = y + H are tangent when H = Hc (Figure 2.23), the equilibrium
population size at the critical harvest rate is the point on the reproduction
curve at which the slope is 1. Substituting this value of y into equation (2.32)
and then solving for H will give us the critical harvest rate Hc (see Exercise 3
below).
We now consider a particular application.
64 Dynamical Systems for Biological Modeling: An Introduction

Example 2.
As reported in the previous section, Reeve, Rhodes and Turchin (1998) found
that the southern pine beetle Dendroctonus frontalis experiences contest com-
petition early in its life cycle, during oviposition. Their study measured the
numbers of eggs deposited in galleries in each attack site on a tree, and, match-
ing their data to the flexible competition model (2.27) proposed by Maynard
Smith and Slatkin (1973), gave an estimate of the exponent n which (as also
described in the previous section) measures the type of competition exhibited.
The estimate n = 1.5 corresponds roughly to the Beverton-Holt model. They
also gave parameter estimates of a = 54 and b = 0.226/egg.
Use these estimates to find the expected average (equilibrium) number of
eggs per attack site, and the critical harvest rate Hc , the maximum number
of eggs which could be removed or parasitized per attack site without driving
the population to extinction.
Solution: In the absence of harvesting, the equilibrium value is (a − 1)/b as
long as a > 1 (which it is here). This gives (54 − 1)/0.226 ≈ 235 eggs per
attack site, with a critical harvest yield
√ √
( a − 1)2 ( 54 − 1)2
Hc = = ≈ 178 eggs. 
b 0.226
The harvested Ricker model is

yk+1 = ayk e−byk − H, (2.34)

and its equilibria are the solutions of

aye−by = y + H, (2.35)

the intersections of the graphs of the two expressions. Since this equation is
transcendental and cannot be solved explicitly, we will proceed with a graph-
ical analysis. The curve z = aye−by starts at the origin with slope a, increases
a
to a maximum value of be at y = 1b and then decreases, approaching a limit
of zero as y → ∞. If a < 1, the line z = y + H lies above this curve, except
for an intersection with the curve at y = 0 if H = 0, as was the case for
the Beverton-Holt model. Thus if a < 1 there is again no equilibrium under
harvesting, and the population dies out.
If a > 1, the analysis proceeds very similarly to the harvested Beverton-
Holt model. For H = 0 (no harvesting), the line z = y starts below the curve
at y = 0 and intersects it again at a positive value of y (Figure 2.24). Again,
increasing H from zero means moving the line z = y + H up, parallel to itself.
For sufficiently small H there are two intersections, with one starting at 0 and
increasing, and the other starting at logb a and decreasing (Figure 2.25). At the
smaller intersection, the slope of the curve is greater than the slope of the line
(which is 1) and thus by the stability theorem of Section 2.4 this equilibrium
is unstable. At the larger intersection, stability depends on the slope of the
 Difference Equations (Discrete Dynamical Systems) 65
1

z z
0.8

z=y+H z=y+H
0.6
stable z=rye−by
z −by
z=rye
0.4

0.2
unstable
unstable
0 0.2 0.4 0.6 0.8 1 y y
y 1/b 1/b

FIGURE 2.24: FIGURE 2.25: Ricker model, r > 1, 0 < H <

Ricker model, r > 1, Hc ; second equilibrium has (left) y < 1/b, (right)
H = 0. y > 1/b.

curve. If the curve and line intersect to the left of y = 1/b (Figure 2.25a),
the slope of the curve is positive but less than 1 and by the stability theorem
this equilibrium is asymptotically stable. To the right of y = 1/b the curve’s
slope is negative (Figure 2.25b). If this slope is between 0 and −1 (which is
guaranteed if a < e2 ), the equilibrium will be asymptotically stable, but it is
possible for the curve’s slope to be less than −1 here, in which case neither
equilibrium is stable.
As H increases, it will reach a value Hc where the two equilibria come
together and curve and line are tangent (Figure 2.26). If H increases beyond
this value, there will be no equilibrium (Figure 2.27). Thus the qualitative
behavior of the system will be that for H near Hc there is an asymptoti-
cally stable positive equilibrium until H reaches this critical value, past which
the population size falls to zero. Again, we have a catastrophe, because this
situation corresponds biologically to the disaster of wiping out the population.
By matching the derivatives of the two graphs at the point of tangency
when H = Hc (Figure 2.26), we see that the equilibrium population size at
this point is the solution y∞ of the transcendental equation

ae−by∞ (1 − by∞ ) = 1

(the left-hand side is the derivative of the reproduction curve; the slope of the
line is 1). From the equilibrium condition (2.35) we also see that the critical
harvest rate is aye−by − y with this value of y∞ . However, it is not possible to
solve the equation (2.35) explicitly for y. In practice, one would use a numerical
approximation method to solve (2.35) and then one could calculate the critical
harvest rate. It would also be necessary to use a numerical approximation
method to find the equilibrium population size for a given harvest rate.
The stability analysis of the harvested Ricker equation is rather more com-
plicated, and we shall not attempt to carry it out. In fact, as with the logistic
66 Dynamical Systems for Biological Modeling: An Introduction
1 1

0.8 0.8

0.6 0.6

z z

0.4 0.4

0.2 0.2

0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

y y

FIGURE 2.26: Ricker FIGURE 2.27: Ricker

model, r > 1, H = Hc ; sin- model, r > 1, H > Hc ; no
gle equilibrium. equilibrium.

equation (Example 1), it is possible to choose the parameters of the model

so that the survival equilibrium with no harvesting is unstable but harvesting
tends to stabilize this equilibrium. We can summarize the effects of harvest-
ing on the Ricker model, our representative scramble competition model, by
observing that while over-harvesting can of course drive a population to extinc-
tion, there is an interval leading up to the critical harvesting rate within which
the population tends toward a stable equilibrium, even if without harvesting
there would be no stable equilibrium. This shows how careful harvesting can
be a tool for stabilizing a population that exhibits scramble competition. In
the following section we shall investigate in more detail what happens to a
population that reproduces in distinct generations when there is no stable
equilibrium.

Exercises
1. Reeve, Rhodes and Turchin (1998) also fit the number of pine beetle
oviposition (egg-laying) galleries per attack site on a tree to the Maynard
Smith and Slatkin model (2.27) and found that it too exhibits contest
competition, with an estimated n = 0.828 for naturally occurring attack
sites and n = 0.914 for field-experiment sites (recall n = 1 corresponds
to the Beverton-Holt model). For naturally occurring sites, the other
parameter estimates were a = 27.596 and b = 0.114/gallery; for the
field-experiment trees, they estimated a = 25.639 and b = 0.133/gallery.
Assume a Beverton-Holt model in both cases.
(a) Find the equilibrium gallery densities for both types of sites.
(b) Calculate the critical harvest rate for both types of sites.
 Difference Equations (Discrete Dynamical Systems) 67

(c) Find the limiting numbers of galleries per attack if a certain num-
ber H of galleries are “harvested” (removed or prevented by ex-
perimenters) for H = 75 and 150. Which of the two types of site
appears more sensitive to harvesting?
2. Suppose we now apply the parameter estimates from the previous prob-
lem (a = 27.596 and b = 0.114/gallery for naturally occurring attack
sites) to a Ricker model with harvesting.
(a) Estimate the critical harvest rate numerically.
(b)∗ Estimate numerically the minimum harvest rate required to sta-
bilize the population (i.e., the positive equilibrium). [Hint: Plot
g ′ (y∞ (H)) as a function of H and see where it falls inside the in-
terval (−1, 1).]
ay
3. (a) Show that the point on the curve z = 1+by which has slope 1 is
√
a−1
y= b .
√
ay a−1
(b) Solve the equation 1+by = y + H with y = b for H.

4. Show that the equilibria (2.33) of the harvested

√ Beverton-Holt model
exist and are positive precisely when H ≤ ( a − 1)2 /b, a > 1. [Hint:
First find conditions on H under which y∞ is real; then find conditions
on H under which y∞ is positive; finally, show that when a > 1 the
positivity threshold lies between the two thresholds for y∞ being real.]
5. Complete the stability
√
analysis for the harvested Beverton-Holt model
a−1
by showing
√ that b lies between the two equilibria (2.33) if and only
if H ≤ ( a − 1)2 /b, a > 1. [Hint: Begin with the desired inequality and
work backward.]
6. In this section we saw that over-harvesting (harvesting beyond the criti-
cal rate) can create a situation in which no nonnegative equilibria exist.
Choose one of the models studied in this section, and select parame-
ter values that cause over-harvesting. Select an initial condition y0 and
calculate the next ten generations, y1 , ..., y10 . What happens to the pre-
dicted population size? What do the results tell you about the valid
operating range of the model?
7. In this section we saw cases in which harvesting below the critical rate
can create a situation in which the smallest positive equilibrium is un-
stable and there is no zero equilibrium. In such a case, what happens
to a population which falls below the level of the smallest equilibrium?
(Try some calculations if necessary, to see what happens.) What is a
reasonable way to interpret this result biologically?
8. If yk < H for some k, then it is not possible to harvest H individuals from
every generation. To make the harvesting yield more realistic (and the
68 Dynamical Systems for Biological Modeling: An Introduction

model well-posed), suggest a way to make yield a function of population

size for small populations.
9. Another type of harvesting model is constant-effort harvesting, in which
a constant “effort” E is applied, and the yield Ey is proportional to
the population size — the more fish there are, for example, in a lake,
the easier it should be to catch one. The Beverton-Holt model with
constant-effort harvesting is
ayk
yk+1 = − Eyk .
1 + byk

(a) Show that this model has a zero equilibrium which is stable when
E − 1 < a < E + 1.
 
a
(b) Show that this model has a positive equilibrium E+1 − 1 /b
(E+1)2
which is stable when E < 1 or a < E−1 .
2
(c) What happens when a > (E+1)
E−1 ? when a < E − 1? It may help to
try some numerical examples or graph these criteria in the E − a
plane.
(d) Based on this analysis, summarize the effects of constant-effort har-
vesting on a population undergoing contest competition.
10∗ . Write the difference equation for the logistic model with constant-effort
harvesting, and derive the stability criterion for its zero equilibrium.
 Difference Equations (Discrete Dynamical Systems) 69

2.6 Period doubling and chaos

We now return to the question first raised in Section 2.3 of what behavior
is possible for a system modeled by a difference equation which has no stable
equilibria. We exclude the possibility of unrestricted growth first seen in the
simple linear difference equation yk+1 = ryk where r > 1 since, as we have
discussed, every biological system exists within an environment which has
limited (even if very great) resources. We will begin by also returning to the
context in which this question first arose.
A population undergoing scramble competition in discrete generations may
be described by the logistic equation
 yk 
yk+1 = rg(y) = ryk 1 − , (2.36)
K
which we have seen in Section 2.3 has an asymptotically stable equilibrium
at zero for 0 ≤ r < 1 (i.e., the population dies out if the intrinsic growth
rate is too low) and an asymptotically stable positive equilibrium K(1 − r1 )
for 1 < r < 3 (i.e., the population approaches a steady level for intermediate
growth rates). Figures 2.28 and 2.29 show graphical analyses for the latter
case, using the initial value y0 = K/2. As the growth rate parameter r passes
through the value 3, there must be a fundamental change in the behavior of
solutions of (2.36). While the equilibrium y = K(1 − 1r ) exists for all values of
r greater than 1, no solution other than the constant solution y = K(1 − 1r )
approaches it if r > 3.
What could happen to a population whose intrinsic growth rate is too high
from one generation to the next? The first generation, each of whose members

1
0.6

0.5
0.8

0.4
0.6

y
0.3

0.4
0.2

0.2
0.1

0 2 4 6 8 10 12 14 16 18 20
k
0 0.2 0.4 0.6 0.8 1

FIGURE 2.28: Solution to logistic FIGURE 2.29: Cobweb diagram for

equation (2.36), 1 < r < 3. logistic equation (2.36), 1 < r < 3.
70 Dynamical Systems for Biological Modeling: An Introduction

may have plenty of resources, will produce a second generation which is too
great for the local resources to support. If the population exhibits scramble
competition as in the logistic model, then the second generation will then be
nearly unable to reproduce — that is, the third generation will be quite small,
as very few members of the second generation will have had children. The third
generation, however, being few in number, will have all the resources needed
to reproduce at their intrinsic capacity, and the fourth generation will again
be quite large. We observe an oscillation or alternation here between small and
large generations, based on three factors: discrete generations (reproduction
occurs simultaneously, rather than continually until environmental resources
are all in use), scramble competition, and a very high intrinsic reproduction
rate. In fact, we see this kind of oscillation in the logistic model when 2 <
r < 3 — for instance, in Figures 2.28 and 2.29, where r = 2.5 — but the
oscillation dies down, and the population level eventually approaches a steady
intermediate state. We might therefore conjecture that populations with an
extremely high intrinsic reproduction rate may cause such wild fluctuations
that the population size never settles down. Indeed, as we shall see, this is
precisely what happens in the logistic model with r > 3.
When r increases beyond 3, a solution of period 2 appears in the lo-
gistic model. By this we mean that there is a pair of values y + and y −
such that rg(y + ) = y − and rg(y − ) = y + . Thus the alternating sequence
{y + , y − , y + , y − , . . .} is a solution of the difference equation (2.36). Since
this solution repeats after two terms, we call it a solution of period 2, or a
2-cycle. We can verify this behavior and find the values of y + and y − either
numerically (choose values for the parameters, say K = 1 and r just beyond 3,
and then iterate the solution many times) or analytically. In order to find the
values analytically, we must solve the equation yk+2 = yk , i.e., rg[rg(y)] = y.
As the function g(y) is a quadratic polynomial, this is a fourth degree polyno-
mial equation. We already know two of the roots of this polynomial equation,
namely the equilibria y = 0 and y = K(1 − 1r ). By dividing out the factors
corresponding to these two roots, we may reduce the fourth-degree equation
to the quadratic equation
   
2 r+1 2 r+1
y −K y+K =0
r r2

whose roots are

Kh p i Kh p i
y+ = r + 1 + r2 − 2r − 3 , y− = r + 1 − r2 − 2r − 3 .
2r 2r
These roots are real, and therefore meaningful, provided r2 − 2r − 3 = (r −
3)(r + 1) ≥ 0, that is, if r > 3 (since we know r > 0). Thus, if r > 3
the logistic difference equation (2.36) does indeed have a periodic solution of
period 2. Figure 2.30 shows the graph of the solution and Figure 2.31 shows
the cobweb diagram, again using y0 = K/2.
 Difference Equations (Discrete Dynamical Systems) 71
1
0.8

0.8

0.6

0.6

y
0.4

0.4

0.2
0.2

0 2 4 6 8 10 12 14 16 18 20
k
0 0.2 0.4 0.6 0.8 1

FIGURE 2.30: Solution to√logistic FIGURE 2.31: Cobweb diagram for

equation (2.36), 3 < r < 1 + 6. logistic equation (2.36), 3 < r < 1 +
√
6.

This periodic solution can be considered as an equilibrium of the second-

order difference equation
yk+2 = rg[rg(yk )], (2.37)
and more generally the study of 2-cycles of first-order difference equations can
be treated in terms of equilibria of second-order difference equations

yk+2 = g2 (yk ). (2.38)

We may find equilibria of such equations by looking for constant solutions; for
(2.38) the condition is y∞ = g2 (y∞ ). The linearization about each equilibrium
is obtained by replacing any nonlinear functions by the linear part of their
Taylor expansion and forming the corresponding linear difference equation; for
the second-order equation (2.38) we replace g2 (y) by g2′ (y∞ )u and obtain the
linearization uk+2 = g2′ (y∞ )uk . Finally, the stability criterion is then obtained
by recalling from Section 2.1 that solutions of linear difference equations have
the form uk = u0 λk for some number λ, and substituting u0 λk+2 for uk+2
and u0 λk for uk ; for (2.38) this gives λ2 = g2′ (y∞ ). The criterion |λ| < 1 for
stability then becomes |g2′ (y∞ )| < 1.
y
For the logistic function g(y) = y(1 − K ), these computations are com-
plicated. Exercises 3, 4 and 5 at the end of this section suggest one way to
∗ ∗
manage these calculations. It is possible to show that the equilibria y+ and y−
found above for (2.37) are asymptotically√stable (and hence the corresponding
2-cycle of (2.36) is stable) if 3 < r < 1 + 6 ≈ 3.4495. That is, for r between 3
and 3.4495, every solution of the logistic equation (2.36) approaches a solution
of period 2.
For r > 3.4495, the solution of period 2 is unstable, but it is possible to
show (in the same way) that a solution of period 4 appears, and that this
72 Dynamical Systems for Biological Modeling: An Introduction
1

0.8

0.8

0.6

0.6

0.4

0.4

0.2
0.2

0 2 4 6 8 10 12 14 16 18 20
k
0 0.2 0.4 0.6 0.8 1

FIGURE 2.32: Solution to logistic FIGURE 2.33: Cobweb diagram for

equation (2.36), 3.4495 < r < 3.544. logistic equation (2.36), 3.4495 < r <
3.544.

solution is asymptotically stable if 3.4495 < r < 3.544. Figure 2.32 shows the
graph of the solution, and Figure 2.33 shows the cobweb diagram. When the
solution of period 4 becomes unstable, a solution of period 8 appears, and this
solution is asymptotically stable for 3.544 < r < 3.564 (see Figures 2.34 and
2.35).
This period doubling phenomenon continues until r = 3.570, when periodic
solutions whose periods are not powers of 2 begin to appear. Figure 2.36
shows what appears to be close to a solution of period 5 for r = 3.75, and
Figure 2.37 shows the corresponding cobweb diagram. If we iterate more terms,
the cobweb diagram looks less like a periodic solution. (Figures 2.38 and 2.39
show a solution for r = 3.9.) For any positive integer p there is some value of
r > 3.828 for which the logistic equation has a periodic solution with period p,
but different initial values give solutions whose behaviors are quite different.
Such behavior is called chaotic.
These facts, whose proofs are difficult and require a close examination of
the properties of continuous functions and of the fixed points of iterates of
continuous functions, are not restricted to the logistic difference equation. It
is a remarkably robust fact that for every difference equation of the form
yk+1 = g(yk ), where g(y) is a function which increases from zero to a unique
maximum and then decreases and approaches zero as y → ∞ (possibly hitting
zero for finite y), the period doubling phenomenon and the onset of chaos
occur. The Ricker model introduced in Section 2.4 is another example of such
a g(y). (The Beverton-Holt model is not, because its g(y) never decreases.)
In fact, the period doubling begins to happen at the same rate: the intervals
of r values during which a given solution of period 2n is stable begin to decrease
in size by the same factor, as n increases. More specifically, if rn is the value
of r for which the asymptotically stable solution of period 2n appears, then
 Difference Equations (Discrete Dynamical Systems) 73
1

0.8

0.8

0.6

0.6

0.4

0.4

0.2
0.2

0 2 4 6 8 10 12 14 16 18 20
k
0 0.2 0.4 0.6 0.8 1

FIGURE 2.34: Solution to logistic FIGURE 2.35: Cobweb diagram for

equation (2.36), 3.544 < r < 3.564. logistic equation (2.36), 3.544 < r <
3.564.

for the logistic equation, r1 = 3, r2 ≈ 3.4495, r3 ≈ 3.544, and r4 ≈ 3.564.

The period 2 solution is stable for an interval of length r2 − r1 ≈ 0.4495, the
period 4 solution for an interval of length r3 − r2 ≈ 0.0945, and the period 8
solution for r4 − r3 ≈ 0.020. Comparing these periods,
r2 − r1 r3 − r2
≈ 4.73, ≈ 4.725.
r3 − r2 r4 − r3
The periods appear to be shortening by a factor of about 4.7 each time. For
any model of the form described above (logistic or otherwise), it has been
shown that this factor approaches a limit,
rn+1 − rn
lim = 4.66920176 . . . ,
n→∞ rn+2 − rn+1

a number which is called the Feigenbaum constant. Usually, the limiting value
is approached very rapidly. This means that the period doubling values of r
occur closer and closer together.
One way to illustrate period doubling and chaotic behavior is by means of
a bifurcation diagram, which we shall discuss in greater detail in Section 4.4. A
bifurcation diagram in general plots equilibria as a function of the parameter
r. If there is an orbit of period 2 for a given value of r, then the bifurcation
diagram will have two points for that value of r, and similarly an orbit of any
period will show in the bifurcation diagram as a number of points equal to the
period of the orbit for that value of r. To draw a bifurcation diagram for, say,
the logistic difference equation, we iterate the logistic function for a sequence
of values of r and plot the results obtained starting after enough iterations
for the orbit to have approached the periodic orbit. This may be done using
74 Dynamical Systems for Biological Modeling: An Introduction
1 1

0.8 0.8

0.6 0.6

0.4
0.4

0.2
0.2

0 2 4 6 8 10 12 14 16 18 20
k
0 0.2 0.4 0.6 0.8 1

FIGURE 2.36: Solution to logistic FIGURE 2.37: Cobweb diagram for

equation (2.36), r = 3.75. logistic equation (2.36), r = 3.75.

a computer algebra system such as Maple (for which a program is given in

Appendix A), and the result is displayed in Figure 2.40.
From a biological point of view, these results are also remarkable. Chaotic
behavior in a population model is a very disturbing phenomenon, as it chal-
lenges our notion that if we know the initial state of a system and the law
governing its development we should be able to predict the behavior. In chaotic
behavior, two solutions whose initial values are very close together — so close
that their difference is smaller than we can measure experimentally — may be-
have very differently. This means that knowing initial values to within as low
a tolerance as we like is still not enough to allow us to predict the long-term
behavior of the system.
One interpretation of the presence of chaos in the logistic equation is that
even very simple models can give rise to apparently unpredictable behavior.
This suggests that complicated behavior may arise even under simple gov-
erning laws. There is experimental data on insect populations which appears
to support the possibility of chaotic behavior. Chaotic behavior would mean
that experimental results may not be repeatable. In the range of values of
the parameter r for which there is an asymptotically stable solution, we may
predict behavior. In the chaotic range we will have to be satisfied with less
information; perhaps upper and lower bounds can be obtained for solutions
even if it is not possible to describe behavior more precisely.
It should also be noted, however, that (as will be seen in the following chap-
ters) chaotic behavior occurs for single-population models only in discrete-time
(difference) equations, where the built-in time lags due to having discrete time
steps allow overshooting an equilibrium. Biologically this suggests that species
which reproduce in distinct generations may be more subject to complicated
fluctuations from one generation to the next. The foregoing discussion also
highlights intraspecies scramble competition as a biological mechanism which
 Difference Equations (Discrete Dynamical Systems) 75
1 1

0.8 0.8

0.6 0.6

0.4
0.4

0.2
0.2

0 10 20 30 40
k
0 0.2 0.4 0.6 0.8 1

FIGURE 2.38: Solution to logistic FIGURE 2.39: Cobweb diagram for

equation (2.36), r = 3.9. logistic equation (2.36), r = 3.9.

opens the door to chaos for such populations. We may wonder what other
biological mechanisms may produce chaotic behavior in populations with dis-
tinct generations. Although we cannot provide an exhaustive answer to this
question, in the remainder of this section we shall consider two other examples
of biological phenomena which have been associated with chaotic behavior.

2.6.1 Dispersal
The term dispersal refers to spatial redistribution of some or all of a popu-
lation, often in response to population pressures (intraspecies competition as
population density builds in a single area). While models for biological dis-
persal typically keep track of population sizes in several different habitats or
patches (as we shall see in the following section), some simple studies have
been made on the effects of dispersal in a single patch.
Since dispersal from a single source tends to reduce the population den-
sity in that location, we might suspect that it will also counter the chaotic
tendencies of populations with high reproduction rates. We can see this very
simply if we consider the usual form of dispersal, in which a given proportion
d of individuals leave the patch (habitat) under study following reproduction.
In this case, the model yk+1 = g(yk ) becomes

yk+1 = (1 − d)g(yk ),

and the effective growth ratio r as in the logistic equation is reduced to (1 −

d)r < r. For the logistic model, the criterion for avoiding period doubling and
chaos then becomes r < 3/(1 − d), which is higher (less restrictive) than 3 for
0 < d < 1.
We might also consider migration into the patch, that is, dispersal from
76 Dynamical Systems for Biological Modeling: An Introduction
1

0.8

Photo courtesy Lennert B.

0.6
FIGURE 2.41: An
eosinophil, a type
of granulocyte, a
specialized white
0.4 blood cell, surrounded
here by erythrocytes
(red blood cells).
Granulocytes fight
infection by phago-
0.2
cytosis; eosinophils
are identifiable by
their multi-lobed nu-
clei. Their production
exhibits regular oscil-
0 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4 lation patterns but can
be affected by diseases
FIGURE 2.40: Bifurcation diagram for the logis- such as leukemia.
tic difference equation, with the parameter r on the
horizontal axis.

other nearby patches into this one. Since we are not modeling the population
densities in the other patches, we incorporate immigration as a constant c,
making the basic form
yk+1 = g(yk ) + c. (2.39)
The result of this immigration is to increase the equilibrium value y∞ above
what it would be without immigration, but the effect upon |g ′ (y∞ )| is difficult
to predict, and so in terms of simplicity of dynamics, constant migration may
either stabilize or destabilize a population’s dynamics. By the same token,
then, a constant dispersal (i.e., c < 0, as opposed to the proportional disper-
sal d discussed above) can also have either effect, in addition to reducing the
value of the positive equilibrium. Note that constant dispersal is equivalent to
constant-yield harvesting, so a high enough dispersal rate will drive a popula-
 Difference Equations (Discrete Dynamical Systems) 77

tion to extinction in finite time (as seen in Section 2.5). Doebeli discusses17 a
way to use variable dispersal to stabilize a population’s dynamics by making
the magnitude of the dispersal dependent on population density.
We shall now look briefly at examples where immigration and dispersal
can either stabilize or destabilize a population. We shall follow Doebeli’s lead
in using the flexible model (2.27) proposed by Maynard Smith and Slatkin
(1973) to describe a population’s natural growth. Constant migration applied
to this model gives us the equation
ayk
yk+1 = + c. (2.40)
1 + (byk )n

Example 1.
Describe the behavior of the Maynard Smith and Slatkin model with constant
immigration (2.40) and parameter values a = 32 , b = 1/pop. unit, and n = 5,
for migration values of c = 0, c = 12 , and c = 1 pop. units.
Solution: The parameter values given correspond to scramble competition
(n = 5) with an intermediate intrinsic growth ratio(a = 3/2). We may then
expect the positive equilibrium to be stable, and it is. We calculate
 1/5
(a − 1)1/n 1 1 − (a − 1)(n − 1) 2
y∞ = = ≈ 0.870551, g ′ (y∞ ) = =− ,
b 2 a 3

so without immigration (c = 0) the population density will stabilize around

0.87. For immigration of c = 1/2 per generation, we calculate y∞ ≈ 1.1146,
g ′ (y∞ ) ≈ −1.19213; the immigration has destabilized the equilibrium, caus-
ing periodic behavior. For immigration of c = 1 per generation, however, we
find y∞ ≈ 1.36049, g ′ (y∞ ) ≈ −0.825852, signaling damped oscillations and a
return to stability. 

What happens here is that a little immigration increases the first genera-
tion just enough to make the second generation considerably smaller — that
is, the decrease in natural population size outweighs the boost given by immi-
gration — but a lot of immigration will end up dominating the population’s
natural growth, so that each new generation will consist almost entirely of
newly arrived immigrants, and very few descendants of the previous genera-
tion (which was so numerous it was practically unable to reproduce).

Example 2.
Describe the behavior of the Maynard Smith and Slatkin model with constant
dispersal (2.40) and parameter values a = 10, b = 1/pop. unit, and n = 2, for
dispersal values of c = 0 and c = −1 pop. units.
Solution: The parameter values correspond to intermediate or very weak
17 M. Doebeli, Dispersal and dynamics, Theoretical Population Biology 47: 82–106, 1995.
78 Dynamical Systems for Biological Modeling: An Introduction

scramble competition (n = 2) with an extremely high intrinsic growth ratio

(a = 10). Here the competition is so weak that despite the high growth ratio,
the positive equilibrium is stable in the absence of dispersal. We compute

(a − 1)1/n 1 − (a − 1)(n − 1) 4
y∞ = = 3, g ′ (y∞ ) = =− ,
b a 5
verifying stability for c = 0. With a dispersal of 1 pop. unit per generation,
however, we compute y∞ ≈ 2.47419, g ′ (y∞ ) ≈ −1.00983, which is just enough
to destabilize the equilibrium. As the dispersal amount increases, things get
worse; for c < −4.085, there are no positive equilibria at all, and the pop-
ulation must “crash” to zero in finite time. Here, in some sense, dispersal
magnifies the effect of the competition, so that instability occurs. 

Of course, a full study of the effects of dispersal should include multi-

patch models in order to see how dispersal between two or more neighboring
habitats affects population densities in each location, and how the single-patch
dynamics interact. We shall not attempt a complete discussion of dispersal in
this text, but will return to the idea in the next section, when we take up
systems of interacting populations with discrete generations.

2.6.2 Dynamical diseases and physiological control systems

Many physiological systems normally display predictable patterns, either
remaining almost constant or with regular oscillations. However, there are
diseases, called dynamical diseases, which exhibit changes in these patterns
such as changes in the nature of the oscillations. Generally, these diseases
occur in physiological control systems. One area in which dynamical diseases
have been observed is the blood system. Some forms of anemia and leukemia
have been identified as dynamical diseases.
Except for lymphocytes, a type of white blood cell produced in lymphatic
tissues, blood cells are formed from primitive stem cells in the bone marrow.
The development process takes about six days. There is also a steady elimina-
tion process whose rate depends on the type of cell. For granulocytes (a kind
of white blood cell, see Figure 2.41), the cell destruction rate is about 10% per
day. While the cell production process is not well understood, it appears that
the production rate should be small for small cell density levels, increasing to
a maximum, and then decreasing to zero as cell density increases (a form simi-
lar to scramble competition reproduction functions). One form which Mackey
and Glass suggested and used successfully in fitting data18 is a production
18 M.C. Mackey and L. Glass, Oscillation and chaos in physiological control systems,

Science 197: 287–289, 1977; L. Glass and M.C. Mackey, Pathological conditions resulting
from instabilities in physiological control systems, Ann. N.Y. Acad. Science 316: 214–235,
1979.
 Difference Equations (Discrete Dynamical Systems) 79

rate in terms of the cell density y, of the form

bθn y
p(y) =
θn + y n
with positive constants b, θ, n. We assume that the elimination rate has the
form cy, and this suggests that if the production and elimination processes
occurred at discrete intervals, we might model the blood cell production by a
difference equation
bθn yk bθn yk
yk+1 = yk + n − cyk = n + (1 − c)yk . (2.41)
n
θ + yk θ + ykn
However, the actual process is continuous but with a time lag correspond-
ing to the length of the production process. If there were no time lag, we
would model the process by a differential equation using the balance relation
that the rate of change of the blood cell density is equal to the production
rate minus the elimination rate. This would give a continuous model
bθn y
y′ = − cy.
θn + y n
We will see in Chapter 4 how to analyze such models, but in overlooking the
time lag in the production process we actually lose an essential feature of the
system. To incorporate a time lag τ in the production process, we would have
to use as model a differential-difference equation
bθn y(t − τ )
y ′ (t) = − cy(t). (2.42)
θn+ [y(t − τ )]n
The analysis of differential-difference equations is complicated, well beyond
the scope of this book. Instead, we shall use a difference equation model as
a way of simulating the time lag. We think of the production time lag as the
unit of time in a discrete model, and then we obtain (2.41) as a very simplified
model.
Some parameter values which have been estimated for the model (2.42)
are b = 0.2/day, c = 0.1/day, θ = 1, n = 10. In order to use these for (2.41),
we must change the time scale. In (2.41) we take 6 days (the time lag) as
the unit of time, and this means we should use b = 6(0.2) = 1.2. Over 6
days, an elimination of 10% per day (c = 0.1/day) leads to a remainder of
(1 − c)6 = (0.9)6 = 0.53. Thus we should use c = 0.47 in (2.41).
n
Equilibria of (2.41) are given by y = 0 and c = θnbθ+yn . Thus there is

1/n
a positive equilibrium y = θ b−c c provided b > c. With the function
bθ n y
g(y) = θn +yn + (1 − c)y, we may calculate g ′ (0) = b + 1 − c, and from this the
stability condition −1 < g ′ (0) < 1 reduces to 0 < c−b < 2. The determination
of the stability condition for the positive equilibrium is more complicated, but
it can be calculated that this equilibrium is asymptotically stable if
 
nc − (n − 1)b
0<c 1− < 2.
b2
80 Dynamical Systems for Biological Modeling: An Introduction

Neither of these conditions is satisfied with the parameter values c = 0.47,

b = 1.2 obtained above. Simulations with these parameter values indicate
oscillations with a period of about 20 days (3.33 time units). In chronic myel-
ogenous leukemia, it is thought that the production time of blood cells in-
creases. An increase in the time units would cause changes in the parameters
b and c in the model (2.41) and would lead to more irregular behavior. This
is not incompatible with observations, although the state of knowledge of the
mechanisms and of experimental data precludes real comparisons.

Exercises
1. Find the value of r for which a solution of period 2 appears for the
difference equation
yk+1 = ryk e1−yk .

2. Consider the Hassell model yk+1 = ayk (1 + nb yk )−n with a, b > 0, n > 1.
Show that this model does not exhibit period doubling if n < 2 and find
the value of a for which a solution of period 2 appears if n > 2. [Hint:
See Exercise 12, Section 2.4.]
3. Let {y+ , y− } be a solution of period 2 of the difference equation yk+1 =
g(yk ). Show that both y+ and y− are equilibria of the second-order
difference equation yk+2 = g[g(yk )].
4. ∗
Define a new index m = k2 for k even and define the iterated function
g2 (y) = g[g(y)]. Show that the solutions y+ , y− of period 2 from Exer-
cise 3 are equilibria of the first-order difference equation ym+1 = g2 (ym ).
[Remark: Exercise 4 and the stability theorem of Section 2.3 show that
an equilibrium y∗ of this second-order difference equation is asymptot-
ically stable if |g2′ (y∗)| < 1. Exercise 5 below gives another criterion
for the asymptotic stability of a solution of period 2 of the difference
equation yk+1 = g(yk ).]
∗
5. (a) Let {y+ , y− } be a solution of period 2 of the difference equation
yk+1 = g(yk ). Use the chain rule of calculus to show that if g2 (y) =
g[g(y)], then
 
d
g2′ (y+ ) = {g[g(y)]} = g ′ (y− )g ′ (y+ ).
dy y=y+

(b) Deduce from part (a) that the solution of period 2 is asymptotically
stable if |g ′ (y− )g ′ (y+ )| < 1.
∗
6. Doebeli (1995) writes regarding constant dispersal (or immigration)
applied to the Maynard Smith and Slatkin (1973) model (2.27), “one can
choose the parameters in [this equation] so that ... dispersal destabilizes
the system. For example, [the equation] can have a stable equilibrium
 Difference Equations (Discrete Dynamical Systems) 81

while dispersal leads to a 2-cycle.” Show that the possibility of dispersal

or immigration destabilizing the positive equilibrium of (2.27) depends
on a and n but not on b. More specifically, show that for the equation
(2.40)
(a) there is an interval of negative values for c (i.e., dispersal) which
cause a stable equilibrium to become unstable if and only if 1 <
4n
n < 3 and the value of a lies between the curves a = (n−1) 2 and
n
a = n−2 (for 2 < n < 3);
(b) there is an interval of positive values for c (i.e., immigration) which
cause a stable equilibrium to become unstable if and only if n > 3
and  
4n n
max 1, <a< .
(n − 1)2 n−2
[Hint: Note that in both cases we need |g ′ (y∞ )| < 1 and |g ′ (y∗ )| > 1,
where y∞ is the equilibrium of the original model (c = 0) and y∗ is
the point where g ′ (y) reaches a minimum. Since the equilibrium of the
model (2.39) with dispersal or immigration increases with c, a third and
final condition is to place y∗ on whichever side of y∞ will make the
equilibrium approach y∗ as c is moved from 0.]
7. Use a graphical argument similar to those made in Section 2.5 to show
that a constant immigration term as in (2.39) increases the value of the
fixed point y∞ .
8. The proportional dispersal model yk+1 = (1 − d)g(yk ) discussed in this
section assumes that individuals disperse after reproduction, so that in
order to calculate the size of the next generation, we first apply the re-
production function, and then dispersal, to the previous generation yk .
Suppose instead that dispersal occurs before reproduction, or, equiva-
lently, that we measure the size of each generation immediately after
reproduction occurs.
(a) What is the general form of the resulting model?
(b) How does the effective intrinsic growth rate compare to that of the
original proportional dispersal model if g(y) is logistic?

9. Identify and discuss at least three modeling issues raised in the discus-
sion of the blood cell model in this section.
10. What practical difference is there between constant-yield harvesting and
constant-rate dispersal?
82 Dynamical Systems for Biological Modeling: An Introduction

2.7 Structured populations

The simplifying assumption of homogeneity within a population is some-
times unrealistic because of important differences among its members. In many
such cases, one can capture the essential variations among members by break-
ing the population into subgroups, called compartments, with movement often
possible from one compartment to another. Two of the most common ways to
structure a population are by age and by spatial distribution . In this section
we shall consider simple examples of both of these kinds of structure.
The modeling of structured population systems requires a system of two
or more equations. For simplicity, we shall consider here only systems of two
equations, which may also be referred to as two-dimensional systems, but
the methods apply equally to systems of higher dimension than two. First,
however, we must consider how the mathematical tools we developed for single
difference equations extend to systems of difference equations.
A general system of two first-order difference equations has the form

yk+1 = f (yk , zk ),
(2.43)
zk+1 = g(yk , zk ),

where the growth of each new generation depends on the sizes of both popula-
tions in the previous generation. Algebraically, an equilibrium of this system
is a solution of the pair of equations

y = f (y, z),
(2.44)
z = g(y, z).

Geometrically, an equilibrium is an intersection of the two curves y = f (y, z)

and z = g(y, z) in the y − z plane. If (y∞ , z∞ ) is an equilibrium of (2.43), then
the system has a constant solution yk = y∞ , zk = z∞ [k = 0, 1, 2, . . .].
The description of the behavior of solutions of systems near an equilibrium
parallels the description given in Section 2.3 for a single first-order difference
equation. If (y∞ , z∞ ) is an equilibrium of (2.43), we make the change of vari-
ables uk = yk − y∞ , vk = zk − z∞ [k = 0, 1, 2, . . .], so that (uk , vk ) represents
deviation from the equilibrium. We then have the system

uk+1 = f (y∞ + uk , z∞ + vk ) − y∞ = f (y∞ + uk , z∞ + vk ) − f (y∞ , z∞ ),

vk+1 = g(y∞ + uk , z∞ + vk ) − z∞ = g(y∞ + uk , z∞ + vk ) − g(y∞ , z∞ ).
(2.45)
If we use Taylor’s theorem to approximate f (y∞ + uk , z∞ + vk ) and g(y∞ +
uk , z∞ + vk ) by their linear parts and neglect the remainder terms,

f (y∞ + uk , z∞ + vk ) ≈f (y∞ , z∞ ) + fy (y∞ , z∞ )uk + fz (y∞ , z∞ )vk ,

g(y∞ + uk , z∞ + vk ) ≈g(y∞ , z∞ ) + gy (y∞ , z∞ )uk + gz (y∞ , z∞ )vk ,
 Difference Equations (Discrete Dynamical Systems) 83

where fy , fz , gy , gz denote the partial derivatives of the functions f and g

with respect to the variables y and z, respectively, then we obtain a linear
system
uk+1 = fy (y∞ , z∞ )uk + fz (y∞ , z∞ )vk ,
(2.46)
vk+1 = gy (y∞ , z∞ )uk + gz (y∞ , z∞ )vk
called the linearization of the system (2.43) at the equilibrium (y∞ , z∞ ), which
approximates the true system (2.43) near the equilibrium (y∞ , z∞ ). The ana-
logue of the linearization theorem of Section 2.3 is true for systems: If all
solutions of the linearization at an equilibrium approach zero, then the equi-
librium is asymptotically stable.
The condition that all solutions of the linearization approach zero is more
complicated for systems than it was for a single equation, however. The idea
behind solving this problem begins by supposing that solutions of the lin-
earization have the form uk = u0 λk , vk = v0 λk for some number λ, since this
is the form that solutions to single linearized difference equations take (cf.
equation (2.7) of Section 2.1). We then find the conditions that make |λ| < 1.
If |λ| < 1, then limk→∞ λk = 0, so solutions uk = u0 λk , vk = v0 λk of the
linearization approach zero (that is, deviations from the equilibrium approach
zero, so the equilibrium must be locally asymptotically stable).
In order to find solutions to (2.46) of the form given above, it must be
possible to have nontrivial solutions (u0 , v0 ) to the system of linear equations

u1 =fy (y∞ , z∞ )u0 + fz (y∞ , z∞ )v0 = λu0 ,

v1 =gy (y∞ , z∞ )u0 + gz (y∞ , z∞ )v0 = λv0 .

If we rewrite these equations in matrix form,

       
u1 fy (y∞ , z∞ ) fz (y∞ , z∞ ) u0 u0
= =λ ,
v1 gy (y∞ , z∞ ) gz (y∞ , z∞ ) v0 v0

and denote the 2 × 2 matrix above as A(y∞ , z∞ ), and the vector [u0 v0 ]T as
~u0 , then we can write the equation in the form A~u0 = λ~u0 , or

(A − λI)~u0 = 0, (2.47)

where I is the 2 × 2 identity matrix. From linear algebra, we know that the
only way for (2.47) to have nonzero solutions ~u0 is for the matrix (A − λI) to
be singular. That is, this system has nontrivial solutions if λ satisfies

det[A(y∞ , z∞ ) − λI] = 0. (2.48)

(Students of linear algebra will recognize λ as the eigenvalue(s) of A.) Equa-

tion (2.48) is called the characteristic equation of this system, and for two-
dimensional systems can be written in scalar form as

λ2 − tr A(y∞ , z∞ )λ + det A(y∞ , z∞ ) = 0, (2.49)

84 Dynamical Systems for Biological Modeling: An Introduction

where tr A and det A denote the trace and determinant of A, namely

tr A(y∞ , z∞ ) =fy (y∞ , z∞ ) + gz (y∞ , z∞ ),

det A(y∞ , z∞ ) =fy (y∞ , z∞ ) gz (y∞ , z∞ ) − fz (y∞ , z∞ ) gy (y∞ , z∞ ).

Our result, therefore, is that all solutions of the linearization (2.46) approach
zero if all roots λ of the characteristic equation (2.48) or (2.49) satisfy |λ| < 1.
In this matrix form, the stability theorem of Section 2.3 generalizes to two-
dimensional systems, and indeed to systems of any dimension. (Note that for
single difference equations, A and I are simply scalars, and the characteristic
equation for yk+1 = g(yk ) at y∞ has the unique solution λ = A = g ′ (y∞ ).)
We shall state this result first for a system of any dimension, and then more
specifically for two-dimensional systems.

EQUILIBRIUM STABILITY THEOREM (I): Let

~y∞ = (y1∞ , y2 , . . . , yn∞ )

be an equilibrium of the system of difference equations

y1k+1 = f1 (y1k , y2k , . . . , ynk ),

y2k+1 = f2 (y1k , y2k , . . . , ynk ),
...
ynk+1 = fn (y1k , y2k , . . . , ynk ),

and let A(~y∞ ) be the matrix of the linearization of the system at this
equilibrium. If all roots of the characteristic equation (2.48) satisfy |λ| <
1, then the equilibrium is asymptotically stable, and if the characteristic
equation has a root with |λ| > 1, the equilibrium is unstable.

There is a set of conditions, known as the Jury criterion, which gives nec-
essary and sufficient conditions that all roots of a polynomial equation satisfy
|λ| < 1. (In general, for an equation of degree n, the Jury criterion consists
of n different inequalities.) The characteristic equation at an equilibrium of a
two-dimensional system of difference equations is a quadratic equation. The
Jury criterion for two-dimensional systems is as follows: Both roots of the
quadratic equation
λ2 + a1 λ + a2 = 0
satisfy |λ| < 1 if and only if

|a1 | < a2 + 1 < 2.

Appendix C offers a proof of the Jury criterion for two-dimensional systems.

 Difference Equations (Discrete Dynamical Systems) 85

For the characteristic equation (2.48), since a1 = −tr A(y∞ , z∞ ), a2 =

det A(y∞ , z∞ ), we obtain the following stability criterion:

EQUILIBRIUM STABILITY THEOREM (II): Let (y∞ , z∞ ) be an

equilibrium of the two-dimensional system of difference equations (2.43). If
A(y∞ , z∞ ) is the matrix of the linearization of the system at the equilib-
rium, and if

0 < |tr A(y∞ , z∞ )| < det A(y∞ , z∞ ) + 1 < 2, (2.50)

then the equilibrium is asymptotically stable.

Example 1.
For each equilibrium of the system yk+1 = 2zk , zk+1 = yk2 , determine whether
it is asymptotically stable.
Solution: Equilibria are solutions of y = 2z, z = y 2 . Substituting the first of
these equations into the second, we obtain z = 4z 2 , which has two solutions,
z = 0 and z = 14 . If z = 0, then y = 0, and if z = 14 , then y = 12 . Thus there
are two equilibria: (0,0) and ( 12 , 14 ). The matrix of the linearization at (y, z) is
 
0 2
.
2y 0

If y = 0, this matrix has trace 0 and determinant 0 and the conditions (2.50)
are satisfied. Thus the equilibrium (0,0) is asymptotically stable. If y = 12 , the
matrix has trace 0 and determinant −2, and the condition |tr A| < det A + 1
is violated. The equilibrium ( 21 , 14 ) is therefore unstable. 

In our study of period doubling in Section 2.6, we were obliged to study

a second-order difference equation, of the form yk+2 = g2 (yk ). An alternative
method for studying higher-order difference equations is to convert them to
systems of first-order difference equations. For example, we can rewrite the
second-order difference equation yk+2 = g2 (yk ) by defining zk = yk+1 , so that
yk+2 = zk+1 = g2 (yk ). This gives the system yk+1 = zk , zk+1 = g2 (yk ).

Example 2.
Determine the condition for asymptotic stability of an equilibrium y∞ of the
second-order difference equation yk+2 = g(yk ) by rewriting it as a first-order
system.
Solution: We rewrite the equation as the system yk+1 = zk , zk+1 = g(yk ). An
equilibrium is a solution of y = z = g(y), and the matrix of the linearization
at an equilibrium y is  
0 1
.
g’(y) 0
86 Dynamical Systems for Biological Modeling: An Introduction
p
The characteristic equation is λ2 − g ′ (y) = 0, with roots ± g ′ (y). Thus
the condition for asymptotic stability is |g ′ (y)| < 1. 

It is, of course, not necessary to transform a second-order difference equa-

tion to a system in order to study equilibria and their stability. This approach
merely gives another proof for the stability of solutions of period 2.
We now consider some common biological systems which must be modeled
by systems of [two] first-order difference equations.

2.7.1 Spatial dispersal

In Section 2.6 we considered spatial dispersal of a species through the lens
of a single patch (habitat), with constant dispersal either in or out of the
patch. Since in reality dispersal is a function of population density, it is more
realistic to consider this phenomenon as an interaction among all the patches
between which dispersal occurs. We will limit ourselves here to two patches
for simplicity.
In studying the spread of disease in an amphibian population, Stock and
Emmert considered19 a population of amphibians which disperses between
an aquatic habitat and a land-based habitat (see Figure 2.42). In the 1990s,
unexpectedly high mortality rates were observed in many amphibian species.
In many cases, the deaths were found to be linked to pesticides and other
contaminants found in the environment, even in relatively remote areas; in
other cases the deaths were caused by viral or fungal infections. These die-offs
are of great concern because amphibians are good barometers of significant
environmental changes that may initially go undetected by humans.
We will suppose that, of the adults of a given generation in either patch, a
proportion p of them survive to the next generation and remain in the same
patch, a proportion q of them survive to the next generation and disperse to
the other patch, and the remaining µ = 1 − p − q of them do not survive
to the next generation. In addition, all reproduction occurs in the aquatic
patch, with a density-dependent per capita reproduction ratio B(y) (which
we assume to be differentiable and monotone decreasing, and hence invertible
on the interval (0, B(0))).
If we denote the amphibian population in the aquatic (playa) habitat by
yk and the population on land by zk , and let the proportions p and q vary by
patch, we can describe this system with the equations

yk+1 = py yk + qz zk + yk B(yk ),
(2.51)
zk+1 = qy yk + pz zk .

The system (2.51) has two equilibria: extinction (0,0), which always exists,
19 E.M. Stock, Deterministic discrete-time epidemic models with applications to amphib-

ians, master’s thesis, Tarleton State University, 2006.

 Difference Equations (Discrete Dynamical Systems) 87

Photos by Gary Fellers, courtesy U.S. Geological Survey, www.usgs.gov

FIGURE 2.42: Left: a Pacific tree frog. Right: a high sierran lake with
mountain yellow-legged frogs. Amphibians such as these frogs move between
lakes and the surrounding land; the lakes, which serve as breeding sites, also
act as focal points for viral and fungal infections and pesticide runoffs.

 
qy
and survival B −1 (m), 1−p z
B −1
(m) , which exists when B(0) > m, where m
is an average effective mortality ratio for the playa patch,
qy qz µz
m = (1 − py ) − = µy + qy
1 − pz qz + µz
(in every generation of reproductive adults yk , a proportion qy of them leave
for land; a proportion qzµ+µ
z
z
of those die before returning to the water).
The asymptotic stability of the extinction equilibrium can be analyzed via
the matrix for the linearization of (2.51) about (0,0),
 
py + B(0) qz
A= .
qy pz
Applying the Jury criterion, we see that the extinction equilibrium is stable
if (and only if)

0 < py + pz + B(0) < 1 + (py + B(0))pz − qy qz < 2.

Rewriting these inequalities in terms of B(0), the stability conditions become

1 − py pz + qy qz
B(0) < m and B(0) < k = .
pz
A little algebra shows that m < k (mpz (1 − pz ) < kpz (1 − pz ) simplifies to
0 < (1 − pz )2 + qy qz ), so the extinction equilibrium is stable if B(0) < m, that
is, if the effective mortality ratio exceeds the maximum reproductive ratio.
The asymptotic stability of the survival equilibrium depends on the form
of B(y). The matrix A for the linearization of (2.51) about the survival equi-
librium is
py + m + B −1 (m)B ′ (B −1 (m)) qz
 
.
qy pz
88 Dynamical Systems for Biological Modeling: An Introduction

The second half of the Jury criterion, det A < 1, becomes

py + m + B −1 (m)B ′ (B −1 (m)) pz − qy qz < 1,

which can be rewritten as m + B −1 (m)B ′ (B −1 (m)) < k. Since by assumption

B is monotone decreasing, B ′ (B −1 (m)) < 0, so m + B −1 (m)B ′ (B −1 (m)) <
m < k, and the condition is met. The first half of the Jury criterion, |tr A| <
1 + det A, becomes

|py + pz + m + B −1 (m)B ′ (B −1 (m))|

< 1 + py pz − qy qz + pz m + B −1 (m)B ′ (B −1 (m)) . (2.52)

If tr A > 0, then we drop the absolute value bars and rewrite the condition
(with some algebra20) as m+ B −1 (m)B ′ (B −1 (m)) < m, which is again always
true because B is decreasing. However, if tr A < 0, the condition can be
rewritten as
qy qz
m + B −1 (m)B ′ (B −1 (m)) > − (1 + py ), (2.53)
1 + pz

which does not hold if B ′ (B −1 (m)) (which is negative) is large enough.

If we assume that the reproduction ratio is determined by contest com-
petition, then the total reproduction is an increasing function of population
d
density, i.e., dy (yB(y)) = B(y) + yB ′ (y) > 0 for all values of y including the
equilibria, so that tr A = py + pz + B(y∞ ) + y∞ B ′ (y∞ ) > 0, and the survival
equilibrium is stable whenever it exists (i.e., for B(0) > m). If, on the other
hand, the reproduction ratio is determined by scramble competition, then it
is possible for B to decline so fast that (2.53) does not hold.

Example 3.
Show that the survival equilibrium of (2.51) is always stable when it exists, if
the fertility is described by the Beverton-Holt model.
ay a
Solution: We have yB(y) = 1+by , so that B(y) = 1+by . Thus B −1 (m) =
1 a


b m − 1 , and the matrix for the linearization of (2.51) about the survival
equilibrium is  2 
py + ma qz
A= .
qy pz
m2
Here tr A = py + pz + a > 0, so the Jury criterion holds as noted above. 

Example 4.
Determine the stability condition for the survival equilibrium of (2.51), if the
fertility is described by the Ricker model.
20 Move p + p to the right-hand side, move the last term on the right to the left, and
y z
divide through by 1 − pz . The right-hand side becomes m, and the given inequality is
obtained.
 Difference Equations (Discrete Dynamical Systems) 89
1 a
Solution: We have yB(y) = aye−by , so that B(y) = ae−by . Now y∞ = b log m
exists for a > m, and the case tr A < 0 becomes
 
py + pz
a > m exp 1 + .
m

Also, in this case (2.53) becomes

 a qy qz
m 1 − log > − (1 + py ),
m 1 + pz
which, after some algebra, can be rewritten
 
2 pz
a < m exp 2 + P , where P = py + qy qz < py + qy < 1. (2.54)
m 1 − p2z
Since
py + pz 2
0<1+ <2+ P
m m
(the second inequality requires some algebra), the threshold value for a in
(2.54) gives a > m and tr A < 0, so (2.54) (together with a > m) is the
stability criterion for the survival equilibrium. 

The general behavior of the models in the two examples above is similar to
that of the simpler Beverton-Holt and Ricker models analyzed in Section 2.4.
So in what way do these dispersal models give a better understanding of the
effects of dispersal on the stability of the underlying populations?
In order to answer this question, we should compare the behaviors of the
two two-patch dispersal models with the behaviors of the one-patch models
that most closely correspond to them: models with the given reproductive
functions, mortality ratios µ, and no dispersal (i.e., q = 0, so that p + µ = 1
and we consider the entire population as a whole, without separating those in
the lake or playa from those on land). For Beverton-Holt fertility, this is
ayk
yk+1 = + (1 − µ)yk ;
1 + byk
for Ricker fertility, it is

yk+1 = ayk e−byk + (1 − µ)yk .

The analyses of these models are left as an exercise for the reader. 
The one-

1 a
patch model with Beverton-Holt fertility has a survival equilibrium b µ − 1 ,
which exists and is stable when a/µ > 1. The corresponding two-patch model
replaces µ with m. The one-patch model with Ricker fertility has a survival
equilibrium 1b log µa which is stable for 1 < a/µ < e2/µ . The two-patch Ricker
model has a survival equilibrium with y∞ = 1b log m a
which is stable when
a > m and (2.54) holds. Therefore the effects of water-land dispersal for
90 Dynamical Systems for Biological Modeling: An Introduction

amphibians are twofold: in increasing the effective mortality ratio from µx

to m due to deaths of nonreproducing adults on land, and, in the case of
scramble (Ricker) competition, in making the survival equilibrium less likely
to be stable, by reducing the upper bound for stability (it takes some algebra
to show that µ2y > 2 + m 2
P ). The model also gives a way to estimate the
number of adults on land relative to the number in the water, although it is
important to note that the model implicitly assumes that adults remain on
land for an average period of time longer than that between breeding times
(which is used as the time step).21

2.7.2 Two-stage populations

Insect populations normally pass through at least two developmental
stages, larval and adult. Often there is a pupal stage as well. We shall construct
a simple two-stage model covering a larval stage and an adult stage to give an
idea how one might go about devising and analyzing more accurate models.
We let yk denote the larval population size in the kth generation and zk the
adult population size in the kth generation. We assume that the birth rate of
larvae depends only on the adult population size and is rze−bz when the adult
population size is z (scramble competition). We assume that a fraction p of
the larvae survive to become adults in the next generation and that a fraction
µ of the adults die from one generation to the next. These assumptions lead
to the model
yk+1 = rzk e−bzk ,
(2.55)
zk+1 = pyk + (1 − µ)zk .
(Note that this model also implicitly assumes that the time required for larvae
to mature is approximately the same as the time between reproductive cycles.)
Equilibria of the model (2.55) are solutions of the equations y = rze−bz
and z = py + (1 − µ)z or µz = py. We substitute y = µp z into the second
equation, obtaining µz = rpze−bz . Thus either z = 0, which implies y = 0, or
rpe−bz = µ, which implies
1 rp µ rp
z= log , y = log . (2.56)
b µ pb µ
This equilibrium has positive values of y and z, corresponding to survival of
the species, only if rp
µ > 1, or rp > µ. The matrix of the linearization of the
system (2.55) at an equilibrium (y, z) is

0 re−bz (1 − bz)
 
.
p 1−µ

At the equilibrium (0,0), this is

21 It can be shown that the average number of generations that an individual spends in

the land patch before dying or returning to water is 1/(1 − pz ) in this model.
 Difference Equations (Discrete Dynamical Systems) 91

 
0 r
,
p 1−µ
with trace 1 − µ and determinant −rp. The equilibrium is asymptotically
stable if 1 − µ < 1 − rp < 2. Thus the asymptotic stability condition is rp < µ,
which is just the condition that (0,0) is the only equilibrium. At the survival
equilibrium, the matrix is

0 µp (1 − bz)
 
,
p 1−µ

with trace 1−µ and determinant µ(bz −1). The asymptotic stability condition
is
1 − µ < µ(bz − 1) + 1 < 2.
The first inequality is satisfied automatically and the second condition is
µ(bz − 1) < 1. Using the equilibrium value of z, we reduce this to µ log rp
µ < 1.
rp 1/µ 1/µ
Taking exponentials, we have µ < e , or rp < µe . Since the existence
of this equilibrium requires rp > µ, we have an asymptotically stable survival
equilibrium if and only if
1
µ < rp < µe µ .
This result suggests that populations described by this model will tend to
reach a (more or less) constant level when the effective reproductive ratio rp
(the reproductive ratio multiplied by the proportion of larvae that survive to
reproductive age) outpaces natural mortality µ but not by too much. Popula-
tions undergoing intraspecific scramble competition which have an excessively
high growth ratio will tend to suffer fluctuations, just as observed for the
simple logistic model in Section 2.6.
In evaluating the utility of this model, we should consider what additional
detail it gives over the single-stage model with Ricker-function fertility and
proportional mortality described in the subsection on dispersal above. The
general behavior is the same; in the simpler model, the population tends to
the same equilibrium value (with a = rp) as long as the effective reproductive
ratio outpaces natural mortality µ, but not by too much. In the simpler model,
2
the ratio a/µ should fall in the interval (1, e µ ), while for the two-stage model
1
it should fall within the interval (1, e µ ) in order for the population density to
stabilize. This suggests that the added delay incurred by newborns requiring
a full generation of time to reach reproductive age makes it more difficult to
maintain stability; this conclusion is consistent with our previous observation
that delays (as represented by the discrete time steps implicit in difference
equations) make fluctuations more likely. The other difference, of course, is
that the two-stage model allows us to estimate the size of the larval population
relative to the adult population at any time.
Emmert and Allen used a similar model (with an additional term qyk in
the first equation denoting those juveniles who survive without having reached
92 Dynamical Systems for Biological Modeling: An Introduction

maturation) as a basis for studying the fungal infection chytridiomycosis in

amphibian juveniles and adults;22 they first studied the model without disease
in order to better understand the behavior of the model when the disease is
introduced. Many of the other models we have studied in this chapter can like-
wise serve as baseline models to understand the behavior of a given species in
the absence of some phenomenon (like disease) which can alter that behavior.

Exercises
In each of the following exercises, find all equilibria and determine which of
them are asymptotically stable.

1. yk+1 = 3yk − 2zk , 4. yk+1 = ryk e−byk ,

zk+1 = 4yk + zk zk+1 = ayk (1 − e−byk )
2. yk+1 = zk + zk2 , zk+1 = yk + 1 5. yk+2 + 4yk+1 + 4yk = 0
3. yk+1 = 2yk , zk+1 = 4e−yk 6. yk+2 = e−yk

7. For the model with Beverton-Holt fertility and proportional mortality µ

ayk
yk+1 = + (1 − µ)yk ,
1 + byk
find all equilibria and analyze their asymptotic stability.
8. For the model with Ricker fertility and proportional mortality µ
yk+1 = ayk e−byk + (1 − µ)yk ,
find all equilibria and analyze their asymptotic stability.
9. (a) Rewrite the third-order difference equation yk+3 = g3 (yk ) as a first-
order system of dimension three.
∗
(b ) Linearize the resulting system and obtain the criterion for stability
of an equilibrium y∞ from the characteristic equation.
10. Write a model describing the dispersal of a species among three patches
in which the second patch is connected to the first and third, but the
first and third are not connected directly to each other. This situation is
appropriate for three patches arranged in a more or less linear fashion,
or for an amphibian population where the end patches are bodies of
water and the center patch is the patch of land between them.
11. Write a model describing the dispersal of a species among three patches
in which dispersal occurs directly among all three patches.

22 K.E. Emmert and L.J.S. Allen, Population persistence and extinction in a discrete-

time, stage-structured epidemic model, Journal of Difference Equations and Applications

10(13–15): 1177–1199, 2004.
 Difference Equations (Discrete Dynamical Systems) 93

2.8 Predator-prey systems

In the previous section we studied single populations structured by age or
spatial distribution. Another type of biological system which requires multiple
equations to model is one in which two or more species interact. One of the
most common is a predator-prey system, in which one population benefits at
the expense of the other. For instance, in modeling a population with discrete
generations, it may be important to include in the model the resources on
which the species being studied feeds, as well as the species of interest. This
leads to a system of two equations, one each for the species and the resource.
Another kind of species interaction is a host-parasitoid situation, most com-
mon among insects, in which one species lays its eggs in a host individual.
Both of these systems are special cases of predator-prey relationships.

2.8.1 A plant-herbivore model

The cinnabar moth Tyria jacobaeae lives for one year, feeding on the peren-
nial plant ragwort (see Figure 2.43). At the end of the year it lays eggs that
hatch the following spring, and then dies. We let yk denote the total biomass
of ragwort and zk the number of insect eggs at the start of the kth year. We
assume that the number of eggs laid is proportional to the amount of rag-
wort at the start of the previous year, so that zk+1 = ayk . We assume that,
without the moth feeding, the ragwort would have a total biomass of B, but
consumption reduces the stock of ragwort so that its biomass falls exponen-
tially with the ratio of insect eggs to plant biomass the previous year, so that
yk+1 = Be−bzk /yk . This gives the model23
bzk
−
yk+1 = Be yk
, zk+1 = ayk . (2.57)

Equilibria of the system (2.57) are solutions of the pair of equations

bz
z = ay, y = Be− y ,

and we may eliminate z to give y = Be−ab , z = aBe−ab . The matrix of the

linearization at an equilibrium (y, z) is
"     #
B ybz2 e−bz/y −B yb e−bz/y
.
a 0
23 E. van der Meijden, M.J. Crawley, and R.M. Nisbet, The dynamics of a herbivore-plant

interaction, Insect Populations: in Theory and Practice (J.P. Dempster and I.F.G. McLean,
eds.), Chapman and Hall, London, 1998.
94 Dynamical Systems for Biological Modeling: An Introduction

Photo courtesy Ray Hamblett Photo courtesy S. Bradleigh Vinson

FIGURE 2.43: A cinnabar moth FIGURE 2.44: A parasitoid wasp at-

caterpillar eating ragwort. This tall, tacking a host larva.
yellow flowering weed can be fa-
tal if eaten by horses and cattle,
and cinnabar caterpillars are used
in some places as biocontrol. In Le-
icestershire, England, the caterpil-
lars have reduced the ragwort pop-
ulation by up to 80%.

bz ab b b
Because of the equilibrium conditions e−bz/y = e−ab , y2 = Be−ab , y = Be−ab ,
this matrix reduces to  
ab −b
.
a 0
The trace of this matrix is ab and the determinant is also ab; thus the stability
condition is ab < ab + 1 < 2, or ab < 1. Thus the equilibrium is asymptotically
stable if and only if ab < 1 (note that ab is a dimensionless measure of the
amount by which the moth egg density reduces the ragwort biomass). Obser-
vations indicate both asymptotically stable equilibria and unstable equilibria
with oscillations are possible.

2.8.2 A host-parasitoid model

Insect parasitoids lay eggs in the larvae of a host species, causing the
death of the host larvae and the survival to adulthood of the parasitoid (see
Figure 2.44). We let yk denote the host population size in generation k and
zk denote the parasitoid population size in generation k. We suppose that
each host lays enough eggs to produce r larvae. The assumptions that (i)
the number of encounters between host larvae and parasitoids is proportional
to host population size and parasitoid population size, and that (ii) these
encounters are distributed randomly among the available hosts, lead to the
conclusion that the fraction of host larvae surviving to adulthood is e−bzk for
some b. Both hosts and parasitoids live for only one generation of time. Then
 Difference Equations (Discrete Dynamical Systems) 95

the host equation is yk+1 = ryk e−bzk . Since each parasitized host larva leads
to one adult parasitoid, the parasitoid equation is zk+1 = ryk (1 − e−bzk ). This
gives the Nicholson-Bailey model for host-parasitoid dynamics,24

yk+1 = ryk e−bzk ,

(2.58)
zk+1 = ryk (1 − e−bzk ).

Equilibria of (2.58) are solutions of the pair of equations y = rye−bz ,

z = ry(1 − e−bz ). One equilibrium is y = 0, z = 0, corresponding to extinction
of both populations. A second equilibrium is found by solving re−bz = 1,
which gives z = 1b log r, and substituting into z = ry(1 − e−bz ), which then
gives 1b log r = ry(1 − 1r ) = (r − 1)y, or y = b(r−1)
1
log r provided r > 1. Thus
there is a survival equilibrium
1 1
y∞ = log r, z∞ = log r
b(r − 1) b
if r > 1. The matrix of the linearization of (2.58) at an equilibrium (y, z) is

re−bz −rbye−bz
 
.
r(1 − e−bz ) rbye−bz

At the extinction equilibrium (0,0), this matrix is

 
r 0
,
0 0

with trace r and determinant zero. The stability conditions are r < 1 < 2,
or r < 1. Thus if r < 1, the extinction equilibrium is the only equilibrium
and it is asymptotically stable. At the survival equilibrium, the matrix of the
linearization is  
1 −by∞
,
r − 1 by∞
with trace 1 + by∞ and determinant rby∞ . The stability conditions are 1 +
by∞ < 1 + rby∞ < 2 or by∞ < rby∞ < 1. The first of these conditions is
r
satisfied if and only if r > 1, and the second is satisfied if rby∞ = r−1 log r < 1.
However, this condition is not satisfied for any value of r > 1. Thus the survival
equilibrium of the Nicholson-Bailey model can not be asymptotically stable.
This means that either both host and parasitoid populations will be wiped
out (if r < 1) or both populations will oscillate. These oscillations may have
large amplitude and may bring one of the populations to a very low level from
which a small perturbation could wipe it out.
Thus this rather simple model supplies an explanation for outbreaks and
subsequent crashes of an insect population with a parasitoid. However, host-
parasitoid populations which persist for many generations have been observed,
24 A.J. Nicholson and V.A. Bailey, The balance of animal populations, Proc. Zoological

Soc. London 3: 551–598, 1935.

96 Dynamical Systems for Biological Modeling: An Introduction

and this can not be explained with a Nicholson-Bailey model. In order to

allow the possibility of an asymptotically stable survival equilibrium we must
make some different assumptions. One assumption which would lead to this
conclusion is that adult hosts may live longer than one generation. If we
assume that a fraction p of hosts survive to the next generation, the model
(2.58) would be replaced by

yk+1 =ryk e−bzk + pyk ,

zk+1 =ryk (1 − e−bzk ).

This model does allow for the possibility of an asymptotically stable survival
equilibrium, but the computations to show this are somewhat complicated
(see Exercise 5). The result is that most of the adult hosts must survive from
one generation to the next in order to overcome a high reproductive ratio r
and stabilize the equilibrium.

The biological systems described in this section share the property that the
equilibrium is unstable if the natural growth rate is too large. As discussed in
previous sections, many discrete models have this property. In the next chap-
ter, we shall consider continuous models, or ordinary differential equations,
which do not share this property. Discrete models have a built-in time-lag,
and this is what makes instability possible.

Exercises
In each of the following exercises, find all equilibria and determine which of
them are asymptotically stable.
1. Determine which of the systems in Exercises 1–6 of Section 2.7 represent
predator-prey systems, by considering yk and zk as two separate species
in each case, and examining the effect that each population has on the
other (via the sign of the associated coefficients).
2. In developing the model (2.57), van der Meijden, Crawley, and Nisbet
assumed that the number of new eggs laid is proportional only to the
biomass of ragwort, and not to the number of adult moths. What implicit
assumption is being made about the relevant factors? That is, why might
this be a reasonable assumption?
3. Suppose that the assumption discussed in the previous problem is not
valid, and that in fact the number of new eggs laid is proportional to
both the biomass of ragwort and the number of adult moths, i.e., zk+1 =
ayk zk . Find the equilibria of this system, and derive criteria for their
asymptotic stability.
4. Compare the behavior of the plant-herbivore model in the main text with
that of the modified system proposed in the previous problem. Explain
 Difference Equations (Discrete Dynamical Systems) 97

in biological terms what difference it makes whether the number of new

eggs laid depends upon the number of adult moths.
5. (a) Show that the revised Nicholson-Bailey host-parasitoid model

yk+1 =ryk e−bzk + pyk ,

zk+1 =ryk (1 − e−bzk ).

has an extinction equilibrium (0,0) which always exists, and a sur-

vival equilibrium
 
1 r 1 r
log , log
b(r + p − 1) 1−p b 1−p

which exists if and only if r + p > 1.

(b) Show that the extinction equilibrium is stable if and only if r + p <
1, and that the survival equilibrium is stable if and only if

(r + p)(1 − p) r
log < 1.
(r + p − 1) 1−p

(c∗ ) Use a computer to sketch a graph of the region in the p − r param-

eter plane for which the survival equilibrium is stable.

Miscellaneous exercises
For each of the difference equations in Exercises 1–4, verify that the given
function is a solution of the given difference equation.

1. yk+1 = yk (1 − yk2 ), yk = −1
2. yk+1 = 1/yk , yk = c, (k odd), yk = 1/c (k even)
3. yk+1 = yk er(1−yk ) , yk = 1
4. yk+1 = yk /(1 + yk ), yk = 0
Solve each of the difference equations in Exercises 5–8 with initial value
y0 = c.
5. yk+1 = (k + 1)yk 7. yk+1 = yk /(k + 1)
k
6. yk+1 = 2 yk 8. yk+1 = yk + 1
For each of the difference equations in Exercises 9–12, find the equilibria, and
determine which equilibria are asymptotically stable.
98 Dynamical Systems for Biological Modeling: An Introduction

9. yk+1 = 2(yk − yk2 ) 11. yk+1 = yk eryk

10. yk+1 = yk2 + 1/4 12. yk+1 = 4(yk − yk3 )/3
For each of the difference equations in Exercises 13–16, determine for which
values of H (possibly depending on r) there is a positive equilibrium which is
asymptotically stable.
13. yk+1 = ryk − H 15. yk+1 = r y1k − H
14. yk+1 = ryk2 − H 16. yk+1 = − 21 yk2 − yk + 1
−H
2
yk
17. Show that the difference equation yk+1 = has no non-zero con-
1 + yk2
stant solution.
R∞
18. Define the function Γ(x) = 0 tx−1 e−t dt for x ≥ 0.
(a) Show that Γ(1) = 1 and Γ(n + 1) = nΓ(n) if n is a positive integer.
(b) Deduce that Γ(n + 1) = (n + 1)!.
1 a
19. Consider the difference equation yk+1 = (yk + ).
2 yk
(a) Find all equilibria.
√
(b) For which choices of y0 does yk approach a?
1 1
20. Consider the difference equation yk+1 = − yk2 − yk + .
2 2
(a) Show that there is no asymptotically stable equilibrium.
(b) Show that the initial value y0 = 1 leads to a solution of period 3.
(c) Use the cobwebbing method to draw several solutions correspond-
ing to different initial values.
Solve each of the following systems.
21. yn+1 = yn + zn , zn+1 = −2yn + 4zn , y0 = 1, z0 = 1
22. yn+1 = −yn + zn , zn+1 = yn , y0 = 1, z0 = 2
23. yn+1 = −yn + 2zn , zn+1 = 3yn , y0 = 1, z0 = 0
24. yn+1 = 2yn − zn , zn+1 = yn + 3zn , y0 = 0, z0 = 1
In Exercises 25–28 find all equilibria of the given system and determine which
are asymptotically stable.
25. yn+1 = (1 − yn ) − zn , zn+1 = −yn
ayn bzn −yn
26. yn+1 = 1+zn , zn+1 = 1+yn

27. yn+1 = zn , zn+1 = −yn + 1

28. yn+1 = yn /2 − zn2 + wn , zn+1 = yn − zn + wn , wn+1 = yn − zn + wn /2
Chapter 3
First-Order Differential Equations
(Continuous Dynamical Systems)

3.1 Continuous-time models and exponential growth

Most biological systems can be described on some scale in terms of contin-
uous change over time, whether it is the gradual accumulation of algae on the
edges of a stagnant pond or the rapid ebb and flow of species such as the mayfly
Dolania americana, the females of which generally live less than five minutes
as adults (Figure 3.1). For populations which do not reproduce in distinct,
synchronized generations, or for which frequent data is available, considering
changes to occur continuously in time may allow a model to capture impor-
tant features of growth. Although populations of discrete individuals should
increase or decrease by discrete numbers, rather than continuous amounts,
for large populations the inaccuracies incurred by treating the population size
as a continuous quantity are small. In cases where changes are continuous
in time, we can describe those changes by describing the rate of change in
terms of the time-derivative of the quantity, typically a population, rather
than in terms of the difference between population sizes in two consecutive
generations, as is done for discrete-time models. Doing so results in models
composed of differential equations, rather than difference equations.
The models we shall consider in this chapter involve a particular sort of
differential equation, and it is worth delineating the territory we shall explore.
First, in this chapter we shall consider population sizes to vary as a function
of time only, and not other continuous variables such as age, spatial location,
temperature, etc. Differential equations which involve only functions of one
independent variable are called ordinary differential equations, or sometimes
simply ODEs, in order to distinguish them from partial differential equations,
or PDEs, which involve two or more independent variables. Equations which
involve only the first derivative of the population size (or other quantity under
study) are again called first-order equations, to distinguish them from higher-
order equations (e.g., second-order equations, which involve second deriva-
tives, the most common physical application of which is acceleration as the
second derivative of location). In this chapter and the next, we shall restrict

99
100 Dynamical Systems for Biological Modeling: An Introduction

Photo and illustration courtesy U.S. Fish and Wildlife Service, dls.fws.gov

FIGURE 3.1: Left, algae accumulation, Kesterson National Wildlife Refuge,

Alaska. Photo by Gary Zahm, USFWS. Right, the mayfly, some species of
which live only minutes as adults after a longer nymph stage spent underwater.
Illustration by Karen Couch, USFWS.

our study to biological systems that can be modeled with first-order ordinary
differential equations.
We saw in the previous chapter that the simplest sort of change other
than a constant-valued derivative is one in which the rate of change is a
constant multiple of the quantity under study. This type of equation is called
linear, and, although many linear models are far too simple to capture the
complicated behavior of real biological systems, we also saw that the behavior
of more complex, nonlinear equations can be analyzed in terms of the behavior
of related linear equations. We thus begin our study of biological systems that
change continuously in time with a look at the simplest sort of first-order
ODE: the linear equation. Later sections will develop the tools necessary to
examine nonlinear models, and consider some of the types of models and
systems that have contributed most to our understanding of the biological
world, from chemical and neural functions within a single organism to the
growth and management of entire populations.

3.1.1 Exponential growth

The rate of change of some quantity is often proportional to the amount of
the quantity present. This may be true, for example, of the size of a population
with enough resources that its growth is unrestricted, and depends only on an
inherent per capita reproductive rate. The idea can also apply to a decaying
population — for example, the mass of a piece of a radioactive substance. In
such a case, if y(t) is the quantity at time t, then y(t) satisfies the first-order
homogeneous linear differential equation
dy
= ay, (3.1)
dt
 First-Order Differential Equations (Continuous Dynamical Systems) 101

where a is a constant which represents the proportional growth or decay rate.

a is positive if the quantity is increasing and negative if the quantity is decreas-
ing. The idea that populations grow in this way is old, used most famously
by Thomas Robert Malthus to hypothesize geometric (exponential) growth of
human populations.1
We can develop simple linear models like this from basic principles. For
instance, let y(t) represent the number of members of a population of simple
organisms at time t. If we assume that these organisms reproduce by splitting,
and that a fraction a of the members split into two members in unit time, then

y(t + h) − y(t) ≈ ah y(t), (3.2)

with the approximate equality (≈) signifying that the difference between the
two sides of the relation is small compared to h, in the sense that this difference
divided by h approaches zero as h → 0, i.e., that
y(t + h) − y(t) − ahy(t)
lim = 0. (3.3)
h→0 h
So far, this derivation is exactly the same as the one which began Section 2.1.
Now, however, instead of thinking of h as a fixed time interval, we allow h to
approach zero. It follows from (3.3) that

y(t + h) − y(t)
y ′ (t) = lim = ay(t).
h→0 h
Thus the population size y(t) satisfies the differential equation (3.1). If a frac-
tion b of the members reproduce by splitting and a fraction d of the members
die in unit time, then (3.2) would be true with a = b − d. The constant a may
be either positive or negative, depending on whether b > d or b < d. We will
use a to designate the constant of proportionality in (3.1); this constant may
be either positive or negative.
Although linear models may appear too simple to provide useful pre-
dictions, the hypotheses of constant per capita birth rates in the absence
of resource limitations, and constant per capita death rates, do have some
experimental support, even for relatively complex organisms. For instance,
Hutchinson cited2 the growth of the collared dove Streptopelia decaocto in
Great Britain, which it invaded in 1954 after spreading across much of west-
ern Europe. In the absence of initial competition among the first pioneer birds
to settle throughout Great Britain, the data (reproduced in Figure 3.2) show
a remarkably close fit to exponential growth (the figure is on a logarithmic
scale, making exponential curves appear linear) for several years, until pop-
ulation density slowed the growth. Hutchinson also observed a constant per
1 T.R. Malthus, An essay on the principle of population, Harmondsworth, Middlesex

(1798) [republished by Penguin, New York (1970)].

2 G. Evelyn Hutchinson, An introduction to population ecology, Yale University Press,

New Haven, 1978.

102 Dynamical Systems for Biological Modeling: An Introduction

Figure reprinted from G. Evelyn Hutchinson, Figure reprinted from G. Evelyn Hutchinson,
An introduction to population ecology, Yale An introduction to population ecology, Yale
University Press, 1978. Image copyright 1978 University Press, 1978. Image copyright 1978
Yale University Press, used with permission. Yale University Press, used with permission.

FIGURE 3.2: Population estimates FIGURE 3.3: Composite age-

for the collared turtledove Strep- specific survivorship curve for the
topelia decaocto in Great Britain from lapwing Vanellus vanellus, based on
1955 to 1964, on a logarithmic scale. ringed birds found dead in Europe.
Data show an initial exponential in- Data are presented from the end of
crease. the first year of life onward.

capita mortality rate reflected in data gathered by several researchers on some

species of birds. Figure 3.3, also plotted on a logarithmic scale, shows an expo-
nential decay describing mortality for the lapwing Vanellus vanellus following
an initial high juvenile mortality period (omitted).
Functions (such as population sizes) which obey the linear differential
equation (3.1) are exponential in form. To see, first, that exponential functions
obey (3.1), we need only verify that y = c eat is a solution of the differential
equation (3.1) for every choice of the constant c. By this we mean that if we
substitute the function y = c eat into the differential equation (3.1) it becomes
an identity. If y = ceat , then y ′ = aceat = ay, and this is the necessary verifi-
cation. Thus the differential equation (3.1) has an infinite family of solutions
 First-Order Differential Equations (Continuous Dynamical Systems) 103

(one for every choice of the constant c, including c = 0),

y = c eat . (3.4)
In order to show that all solutions to the differential equation (3.1) have the
form (3.4), suppose that y(t) is a solution of the differential equation y ′ = ay,
that is, that y ′ (t) = ay(t) for every value of t. If y(t) 6= 0, division of this
equation by y(t) gives
y ′ (t) d
= log |y(t)| = a. (3.5)
y(t) dt
We will always use log to denote the natural logarithm, rather than ln, which
may have been used in your calculus course. Integration of both sides of (3.5)
gives log |y(t)| = at + k for some constant of integration k. Then
|y(t)| = eat+k = ek eat .
Because eat and ek are positive for every value of t, |y(t)| cannot be zero, and
thus y(t) cannot change sign. We may remove the absolute value and conclude
that y(t) is a constant multiple of eat , y = ceat . We note also that if y(t) is
different from zero for one value of t then y(t) is different from zero for every
value of t. Thus the division by y(t) at the beginning of the proof is legitimate
unless the solution y(t) is identically zero. The identically zero function is a
solution of the differential equation, as is easily verified by substitution, and
it is contained in the family of solutions y = ceat with c = 0.
The absolute value which appears in the integration produces some com-
plications which may be avoided if we know that the solution must be non-
negative, so that |y(t)| = y(t). This is the case in many applications. If we
know that a solution y(t) of the differential equation y ′ = ay is positive for all
t, we could replace (3.5) by
d
log y(t) = a
dt
and then integrate to obtain log y(t) = at + k, y(t) = eat+k = eat ek = ceat .
The logical argument in the above proof is that if the differential equation
has a solution, then that solution must have a certain form. However, it also
derives the form and thus serves as a method of determining the solution.
In order for a mathematical problem to be a plausible description of a
scientific situation, the mathematical problem must have only one solution;
if there were multiple solutions we would not know which solution represents
the situation. This suggests that the differential equation (3.1) by itself is not
enough to specify a description of a physical situation. We must also specify
the value of the function y for some initial time when we may measure the
quantity y and then allow the system to start running. For example, suppose
we impose the additional requirement, called an initial condition, that
y(0) = y0 . (3.6)
104 Dynamical Systems for Biological Modeling: An Introduction

A problem consisting of a differential equation together with an initial con-

dition is called an initial value problem. We may determine the value of c
for which the solution (3.4) of the differential equation (3.1) also satisfies the
initial condition (3.6) by substituting t = 0, y = y0 into the form (3.4). This
gives the equation
y0 = c e0 = c
and thus c = y0 . We now use this value of c to give the solution of the dif-
ferential equation (3.1) which also satisfies the initial condition (3.6), namely
y = y0 eat . This procedure may be followed in any situation described by an
initial value problem, including population growth and radioactive decay.

Example 1.
Suppose that a given population of protozoa develops according to a simple
growth law with a growth rate of 0.7944 per member per day, that there are
no deaths, and that on day zero the population consists of two members. Find
the population size after 6 days.

Solution: The population size satisfies the differential equation (3.1) with a =
0.7944, and is therefore given by y(t) = ce0.7944t . Since y(0) = 2, we substitute
t = 0, y = 2, and we obtain 2 = c. Thus the solution satisfying the given initial
condition is y(t) = 2 e0.7944t , and the population size after 6 days is y(6) =
2 e(0.7944)(6) = 235 (rounding the population size to the nearest integer). 

If we know that a population grows exponentially according to a simple

growth law but do not know the rate of growth we view the solution y =
ceat as containing two parameters (c and a) which must be determined. This
requires knowledge of the population size at two different times to provide two
equations which may be solved for these two parameters.

Example 2. Suppose that a population which follows a simple growth law

has 100 members at a starting time and 150 members at the end of 100 days.
Find the population at the end of 150 days.

Solution: The population size at time t satisfies y(t) = ceat and y(0) = 100,
y(100) = 150. Thus y(0) = 100 = ce0 , y(100) = ce100a = 150. It follows
that c = 100 and 150 = 100e100a . We obtain e100a = 1.5, a = log 1.5
100 =
−3
4.05465 × 10 . Finally, we obtain
−3
y(150) = 100 e150(4.05465×10 )
= 183.71.

Rounding off to the nearest integer, we obtain the population size 184 after
150 days. 
 First-Order Differential Equations (Continuous Dynamical Systems) 105

3.1.2 Radioactive decay

Radioactive materials decay because a fraction of their atoms decompose
into other substances. If y(t) represents the mass of a sample of a radioactive
substance at time t, and a fraction k of its atoms decompose in unit time,
then y(t + h) − y(t) is approximately −ky(t), and we are led to the differential
equation (3.1) with a replaced by −k. If it is clear from the nature of the
problem that the constant of proportionality must be negative, we will use
−k for the constant of proportionality, giving a differential equation

y ′ = −ky (3.7)

with k > 0.

Example 3.
The radioactive element strontium 90 has a decay constant 2.48 × 10−2
years−1. How long will it take for a quantity of strontium 90 to decrease
to half of its original mass?
Solution: The mass y(t) of strontium 90 at time t satisfies the differential
equation (3.7) with k = 2.48 × 10−2 . If we denote the mass at time t = 0 by
−2
y0 , then y(t) = y0 e−(2.48×10 )t . The value of t for which y(t) = y0 /2 is the
solution of
y0 −2
= y0 e−(2.48×10 )t .
2
If we divide both sides of this equation by y0 and then take natural logarithms,
we have
1
−(2.48 × 10−2 )t = log = − log 2
2
so that t = (log 2)/(2.48 × 10−2 ) = 27.9 years. 

The time required for the mass of a radioactive substance to decrease to

half of its starting value is called the half-life of the substance. The half-life T
is related to the decay constant k by the equation
log 2
T =
k
because if y(t) = y0 e−kt and (by definition) y(T ) = y20 , then e−kT = 12 , so
that −kT = log 12 = − log 2. For radioactive substances it is common to give
the half-life rather than the decay constant.

Example 4.
Radium 226 is known to have a half-life of 1620 years. Find the length of
time required for a sample of radium 226 to be reduced to three fourths of its
original size.
log 2
Solution: The decay constant for radium 226 is k = 1620 = 4.28×10−4 years−1 .
106 Dynamical Systems for Biological Modeling: An Introduction

In terms of k, the mass of a sample at time t is y0 e−kt if the starting mass is

y0 . The time τ at which the mass is 3y40 is obtained by solving the equation

3y0
= y0 e−kτ
4
3
or 4 = e−kτ . Taking natural logarithms we obtain −kτ = log 34 , which gives

log 43 1620(log 34 )
τ =− = = 672 years. 
k log 2

The radioactive element carbon 14 decays to ordinary carbon (carbon 12)

with a decay constant 1.244 × 10−4 years−1, and thus the half-life of carbon
14 is 5570 years. This has an important application, called carbon dating, for
determining the approximate age of fossil materials. The carbon in living mat-
ter contains a small proportion of carbon 14 absorbed from the atmosphere.
When a plant or animal dies, it no longer absorbs carbon 14 and the propor-
tion of carbon 14 decreases because of radioactive decay. By comparing the
proportion of carbon 14 in a fossil with the proportion assumed to have been
present before death, it is possible to calculate the time since absorption of
carbon 14 ceased.

Example 5.
Living tissue contains approximately 6 × 1010 atoms of carbon 14 per gram of
carbon. A wooden beam in an ancient Egyptian tomb from the First Dynasty
contained approximately 3.33 × 1010 atoms of carbon 14 per gram of carbon.
How old is the tomb?
Solution: The number of atoms of carbon 14 per gram of carbon, y(t), is given
by y(t) = y0 e−kt , with y0 = 6 × 1010 , k = 1.244 × 10−4, and y(t) = 3.33 × 1010
for this particular t value. Thus the age of the tomb is given by the solution
of the equation
−4 3.33 × 1010 3.33
e−(1.244×10 )t = = ,
6 × 1010 6
and if we take natural logarithms this reduces to
log 3.33 − log 6
t=− = 4733 years. 
1.244 × 10−4

Exercises
In Exercises 1–6, assume the population size satisfies a simple growth law.

1. Suppose that the birth rate of a given population is 0.36 per member
per day with no deaths. If the population size on day zero is 50, what
is the population size 10 days later?
 First-Order Differential Equations (Continuous Dynamical Systems) 107

2. Suppose that the birth rate of a given population of protozoa is 0.2 per
member per day with no deaths. If the population size on day zero is
10, find the population size 20 days later.
3. Suppose a population has 173 members at t = 0 and 262 members at
t = 10. Estimate the population size at t = 5.
4. Suppose that a population has 87 members at t = 0 and 125 members
at t = 4. Estimate the population size at t = 6.
5. Suppose that a population has 12 members at t = 3 and 5 members at
t = 10. What was the population size at t = 0?
6. Suppose that a population has 13 members at t = 4 and 20 members at
t = 6. What was the population size at t = 0?
7. If the half-life of a radioactive substance is 30 days, how long would it
take until 99 % of the substance decays?
8. How long does it take for a piece of carbon 14 to decrease to 20 % of its
original size?
9. In a sample of uranium 238, it is found that 0.0000154 % of the mass
disintegrates in 1000 years. Find the half-life of uranium 238.
10. How old is a fossil in which 85 % of the carbon 14 has disintegrated?
11∗ . Show that the solution of the initial value problem y ′ = ay, y(t0 ) = c is
y(t) = cea(t−t0 ) .
12∗ . Use the graph in Figure 3.2 to estimate (a) the per capita birth rate
of the collared turtledove, in units of per individual per year, and (b)
how long it would take for a population of this bird to double in an
environment with enough resources to prevent competition. Hint: What
form does log y have if dy/dt = ay? How can you read a from the graph?
13. Explain why the graphs of solutions to the linear differential equation
dy/dt = ry appear linear on a logarithmic scale. Calculate the slope of
such a line (supposing an initial condition y(0) = y0 ) and interpret it.
14. Suppose that a given population exhibits exponential growth when it is
small. How and why would you expect that to change when the popu-
lation is large? How would the graph of the population over time differ
from the graph of a purely exponential function?
108 Dynamical Systems for Biological Modeling: An Introduction

3.2 Logistic population models

As has been observed earlier in this text, linear models are useful either
for very simple situations such as those described in the previous section (e.g.,
unrestricted population growth) or as indicators of local behavior for more
complicated systems near equilibria. However, most of the biological systems
we may wish to model have essentially nonlinear features which are crucial in
determining how the system behaves. In particular, any expanding population
will eventually begin to feel growth restrictions as it runs up against the
resource limitations of its environment.3 At such a point in time, the per capita
growth rate must become density-dependent, that is, a function (typically
decreasing) of the population density in the local area. We have already seen
in Chapter 2 how different assumptions on the nature of this dependence can
produce important differences in the growth of populations which reproduce
in discrete generations. In this section we will look at the continuous-time
version of the simplest, most common model proposed to incorporate density-
dependent growth limitations: the logistic model.
The discrete-time logistic equation was introduced in Chapter 2 as a means
of accounting for restricted growth. The continuous-time logistic equation
 y
y ′ = ry 1 − (3.8)
K
can be derived as follows. Let y(t) be the size of a population at time t. As we
wish to consider y(t) as a differentiable function of t, it is not quite appropriate
to consider y to be the number of members of a population, although for
large populations it may be a reasonable approximation. We will think of y
as representing the biomass, that is, the total mass of the members of the
population, and we think of this biomass as increasing due to the conversion
of nutrients to increased mass of the members of the population. We let x(t)
be the amount of nutrient available at time t and we assume that consumption
of one unit of nutrient leads to an increase of a units of population biomass,
so that
y ′ (t) = −ax′ (t). (3.9)
y ′ (t)
We assume also that the per capita rate of population growth y(t) is propor-
tional to the amount of nutrient available, so that
y ′ (t)
= bx(t) (3.10)
y(t)
for some constant b. Integration of (3.9) gives
y(t) = −ax(t) + c (3.11)
3 For a more extensive discussion of the limitations of exponential growth models, see

Mark Kot, Elements of mathematical ecology, Cambridge University Press, Cambridge,

2001, pp. 5–6.
 First-Order Differential Equations (Continuous Dynamical Systems) 109

with c a constant of integration; in fact, c = y(0) + ax(0). Now, substitution

of (3.11) into (3.10) gives
 
c − y(t) bc  y(t) 
y ′ (t) = bx(t)y(t) = b y(t) = y(t) 1 − .
a a c

We now let r = bc/a, K = c, and we obtain the logistic differential equation

(3.8) for the total biomass of the population. We again consider r to be the
maximum intrinsic per capita growth rate (in units of per time) and K to be
the carrying capacity determined by the resources of the environment and the
needs of the population.
This equation was originally proposed by Verhulst4 to model the assump-
tion that the per capita growth rate should decrease as population size in-
creases. Verhulst chose a decreasing linear function as the per capita growth
rate for simplicity. Solutions of the logistic equation have been fitted to many
different population models with considerable predictive success; although the
notion of biomass may make one think of simple organisms, Verhulst’s original
application was to humans.
In Section 3.4, we will show how to solve the logistic differential equation
(3.8) explicitly, using a technique called separation of variables. For now, we
may verify that for every constant c the function
K
y= (3.12)
1 + ce−rt
is a solution of this differential equation.5 To see this, note that for the given
function y,
Kcr e−rt
y′ =
(1 + ce−rt )2
and
y K −y ce−rt
1− = = .
K K (1 + ce−rt )
Thus  y Krc e−rt
ry 1 − = = y′,
K (1 + ce−rt )2
and the given function satisfies the logistic differential equation for every
choice of c.
4 P.F. Verhulst, Notice sur la loi que la population suit dans son accroissement, Corr.

Math. et Phys. 10(1838), 113–121.

5 We observe that the family of solutions (3.12) of the logistic differential equation (3.8)

includes the constant solution y = K (with c = 0) but not the constant solution y = 0. The
existence and uniqueness theorem of Section 3.3 shows that, since we have now obtained a
solution corresponding to each possible initial condition, we have obtained all solutions of
the logistic differential equation.
110 Dynamical Systems for Biological Modeling: An Introduction

To find the solution which obeys the initial condition y(0) = y0 , we sub-
K
stitute t = 0, y = y0 into the form (3.12), obtaining 1+c = y0 which implies
K−y0
c = y0 as long as y0 6= 0 and gives the solution

K Ky0
y= = (3.13)
y0 + (K − y0 ) e−rt
 
K−y0
1+ y0 e−rt

to the initial value problem with y0 6= 0. Note that the denominator begins at
K (for t = 0) and moves toward y0 as t → ∞.
Now suppose that K represents some physical quantity such that K > 0.
One can see from the form (3.13) that if y0 > 0, then the solution y(t) exists
for all t > 0, and limt→∞ y(t) = K. If y0 < 0, then this solution does not
exist for all t > 0, because y(t) → −∞ wherethe denominator changes sign:
as y0 + (K − y0 )e−rt → 0, or t → log y0y−K 0
. If y0 = 0, the solution of the
initial value problem is not given by (3.13), but is the identically zero function
y = 0. If, instead, K < 0, as will occur in some examples presented in this
chapter, then the solution exists for all time if y0 < 0, but approaches positive
infinity in finite time if y0 > 0, for the same reason as above.
Note also that (for K > 0, y0 > 0) if r > 0 then y → K, while if r < 0 (as
also occurs for some examples in this chapter) then y → 0. In either case a
given solution approaches a limit monotonically, regardless of the magnitude
of r, in contrast to the behavior of the logistic difference equation of Chap-
ter 2, where more complicated behavior occurs for large values of the growth
constant r.

3.2.1 All creatures great and small?

Although the simple reproduction process involved in the original deriva-
tion (and that above) for the logistic model would appear to provide biological
justifications only for simple organisms which grow via cell division, in fact
the logistic model has been found to fit data for the growth of many more
complex organisms remarkably well. We shall now consider several such ex-
amples in order to get a feel for how this equation can be used to describe
population growth.
One notable application of the logistic equation to simple biological popu-
lations is an experiment on the protozoa Paramecium by the Soviet biologist
G.F. Gause.6 Gause fit his experimental data to the solution (3.12) using the
values of the constants given in Example 1, Section 3.1. Table 3.1 lists the
observations, the solution of the logistic model with these constants, and the
solution of the simple exponential population model with the same intrinsic
growth rate. The calculated values are rounded off to the nearest integer. The
reader will note that for the first 4 days the exponential model yields results
6 G.F. Gause, The Struggle for Existence, Williams and Wilkins, Baltimore (1934).
 First-Order Differential Equations (Continuous Dynamical Systems) 111

TABLE 3.1: Gause’s data for protozoan (Paramecium) growth,

compared to logistic and exponential models’ predictions.
t (days) 0 1 2 3 4 5 6 7
y (observed) 2 3 22 16 39 52 54 47
y (logistic) 2 4 9 17 28 40 51 57
y (simple model) 2 4 10 22 48 106 ...

t (days) 8 9 10 11 12 13 14 15 16
y (observed) 50 76 69 51 57 70 53 59 57
y (logistic) 61 62 63 64 64 64 64 64 64

TABLE 3.2: Yeast population y vs. time t in hours, from Carlson (1913).
t 0 1 2 3 4 5 6 7 8 9
y(t) 9.6 18.3 29.0 47.2 71.1 119.1 174.6 257.3 350.7 441.0

t 10 11 12 13 14 15 16 17 18
y(t) 513.3 559.7 594.8 629.4 640.8 651.1 655.9 659.6 661.8

comparable to those of the more sophisticated logistic model. However, for

t ≥ 5 the simple model is hopelessly inaccurate while the logistic model fits
the observations reasonably well.
Another example of the ability of the logistic model to predict the growth of
simple organisms is an experiment of Carlson7 on yeast cultures, later analyzed
by Pearl8 and Renshaw.9 Carlson measured the amount of yeast in a particular
culture every hour (see Table 3.2). Pearl and Renshaw (separately) fit this data
to a logistic model by rewriting the culture size y in terms of the logarithm
of relative growth
K − y(t)
z(t) = log ,
y(t)
which we can rewrite (solving (3.12) for (K − y)/y and substituting) as

K − y0
z = log − rt,
y0

the graph of which (versus time) is a line with slope −r and y-intercept
log(K − y0 )/y0 . Thus a modified linear regression10 can provide estimates
7 T. Carlson, Über Geschwindigkeit und Grösse der Hefevermehrung in Würze, Bio-

chemische Zeitschrift 57: 313–334, 1913.

8 R. Pearl, Introduction of medical biometry and statistics, Saunders, Philadelphia, 1930.
9 Eric Renshaw, Modelling biological populations in space and time, Cambridge Studies

in Mathematical Biology 11, Cambridge University Press, Cambridge, 1995, pp. 53–55.
10 In this case, the usual linear regression calculation must be modified slightly since the

data zn are functions of the parameter K (otherwise we must first fix a value for K and
then find the best r for that K). In searching for a minimum (with respect to K and r) of
112 Dynamical Systems for Biological Modeling: An Introduction
y[t] z[t]

4
600

500 2

400
t
2.5 5 7.5 10 12.5 15 17.5
300

200 -2

100
-4
t
2.5 5 7.5 10 12.5 15 17.5

FIGURE 3.4: Yeast culture growth FIGURE 3.5: Logarithmic (z) data
data from Carlson (1913) and its from Carlson (1913) and its best-fit
best-fit logistic curve. line.

for the parameters r and K (given y0 ). The best-fit values for the data in Ta-
ble 3.2 are K ≈ 664.5, r ≈ 0.539/hr. Graphs of y(t) and z(t) comparing data
and best-fit solution (similar to those in Renshaw, but using the parameter
values given here) are shown in Figures 3.4 and 3.5. The fit is clearly quite
good.
The logistic model has also been used with some success to fit data on
growth of larger, more complex organisms, as the following example illustrates.

Example 1.
Census figures (in millions) for the United States, P (t) in the table below, fit
the solution
265
y(t) =
1 + 69 e−0.03t
of the logistic differential equation, with t = 0 corresponding to the year 1790,
reasonably well. Predict the population in the years 1990 and 2000, and the
limiting population size.

Year P (t) y(t) Year P (t) y(t) Year P (t) y(t)

1790 4 3.8 1860 31 28.0 1930 123 130.2
1800 5 5.1 1870 39 36.5 1940 132 150.0
1810 7 6.8 1880 50 47.0 1950 152 169.0
1820 10 9.1 1890 63 59.7 1960 180 186.5
1830 13 12.2 1900 76 74.8 1970 204 202.0
1840 17 16.2 1910 92 91.8 1980 226 215.3
1850 23 21.4 1920 106 110.6

the sum of squared errors

X» „ «–2
K − y0
zn − log − rtn ,
n
y0

substituting zn = log[(K − yn )/yn ] transforms the problem into one in which a computer
can search simultaneously for the best-fit K and r.
 First-Order Differential Equations (Continuous Dynamical Systems) 113

Solution: The limiting population size as seen from the expression for y(t) is
265. The population size in 1990 (t = 200) would be predicted from P (t) as
265
= 226.3,
1 + 69 e−(0.03)(200)
and the prediction for 2000 (t = 210) would be
265
= 235.2.
1 + 69 e−(0.03)(210)
The actual population found in the 1990 census was approximately 250 million.
The actual population given in the 2000 census was about 281.4 million. 

Another notable application of the continuous-time logistic model to the

growth of more complex organisms provides a helpful perspective from which
to evaluate this equation’s overall utility in predicting population growth.
Davidson made use of the extensive records of the sheep populations imported
to South Australia and Tasmania in the 19th century — over a century of an-
nual records in each case — to study their growth in habitats (albeit artificial
— sheep cannot persist unaided in Australia) with a carrying capacity de-
termined principally by the amount of land available for pasture. Davidson
found11 that the data for both populations matched logistic curves closely
during the initial “sigmoid” (S-shaped) period of growth. Here again, the lo-
gistic parameters were fit via the transformation z = log[(K − y)/y] which
makes the curve linear, here first choosing a value for K and then using linear
regression to find r. The parameters c and r for (3.12) for the two popula-
tions are, in fact, remarkably close, differing only in the carrying capacities K.
However, as the populations reached their carrying capacities, they exhibited
ongoing, significant fluctuations, gaining or losing over 25% in a handful of
years (see Figures 3.6 and 3.7).
Davidson’s work is not the only instance where a logistic model fits data
well only during the initial rapid growth period. Renshaw discusses (e.g., p.
55) several other situations in which lab cultures and other populations were
observed to grow more or less logistically toward carrying capacity, but there-
after exhibited either sustained fluctuations or a gradual decline. The likely
explanation for this phenomenon is that when a population has not yet run
up against resource limitations, its reproductive capacity may often be the
strongest force affecting its growth, whereas other factors such as stochastic
effects, spatial heterogeneity, or changes in the environment (the fluctuations
observed in the Australian sheep populations in the graphs were attributed to
fluctuations in carrying capacity while pastures recovered from overgrazing)
dominate when the logistic growth is small. In any case, the conclusion to draw
11 J. Davidson, On the ecology of the growth of the sheep population in South Australia,

Transactions of the Royal Society of South Australia 62: 141–148, 1938.

J. Davidson, On the growth of the sheep population in Tasmania, Transactions of the
Royal Society of South Australia 62: 342–346, 1938.
114 Dynamical Systems for Biological Modeling: An Introduction
kilosh eep kilosh eep

8000

2000

6000
1500

4000
1000

2000
500

year year
1860 1880 1900 1920 1940 1840 1860 1880 1900 1920

FIGURE 3.6: South Australian FIGURE 3.7: Tasmanian sheep

sheep population (in thousands) from population (in thousands) from 1818
1838 to 1936 with the logistic curve to 1936 with the logistic curve y(t) =
y(t) = 7115/[1 + exp(249.11 − 1670/[1 + exp(240.81 − 0.13125 t)] su-
0.13369 t)] superimposed, after Ren- perimposed, after Renshaw (1995).
shaw (1995).

seems to be that the continuous logistic model often does a remarkably good
job of describing the rapid growth of even complex populations approaching
carrying capacity, but is of at best limited use in describing the growth of
populations at or near that capacity.
We finish this section with two extensions of the logistic model to multiple
populations, which provide a glimpse of the modeling of multi-population
biological systems undertaken in Chapters 5, 6, and 7.

3.2.2 Competition among plants

Bampfylde reviews12 a set of simple deterministic models for hierarchical
competition among plants in a habitat composed of multiple patches or sites,
developed by Tilman13 to account for the observed trade-offs between colo-
nization and competitive ability (as measured by longevity) as plants allocate
their energies between developing root systems and reproduction. One exam-
ple cited by Tilman involves a prairie habitat in Minnesota where primary
competition is for nitrogen in the soil; superior competitors such as bluestem
grass develop extensive root systems often running two meters or more be-
low the surface but spread slowly, whereas inferior competitors such as rough
bentgrass proliferate quickly but have relatively shallow root systems which
prevent it from invading areas where bluestem is established (see Figure 3.8
for photos).
If we take the habitat sites to be of sufficient size for one adult to survive,
we can describe the proportion of sites occupied by a given species in the
12 Caroline Bampfylde, Modelling rainforests, M.Sc. thesis, Oxford University, 1999, Sec-

tion 2.2, pp. 6–9.

13 David Tilman, Competition and biodiversity in spatially structured habitats, Ecology

75: 2–16, 1994.

 First-Order Differential Equations (Continuous Dynamical Systems) 115

Photo courtesy U.S. National Park Service, www.nps.gov Photo by Dave Powell, USDA Forestry Service

FIGURE 3.8: Left, big bluestem (Andropogon gerardii), a competitive prairie

grass with a deep root system. Right, rough bentgrass (Agrostis scabra), a
perennial clump grass whose wide inflorescence breaks off and floats like a
tumbleweed, enabling it to colonize quickly despite a relatively shallow root
system.

rainforest as a function of time (i.e., without being spatially explicit). Suppose

now that the proportion p1 of sites occupied by species 1 is determined by (a)
the per site rate c1 at which seeds or other propagules disperse and proliferate,
and (b) the per-site mortality rate m1 . Let us also assume that these rates are
unaffected by the presence of other species, i.e., that species 1 dominates any
competition with other species, and that colonization sites are determined at
random. Then the rate of production of new sites is given by the product of
the rate of seedling production by the occupied sites c1 p1 and the proportion
of sites available for colonization (1 − p1 ), while the rate at which sites are
vacated is given by m1 p1 . This gives the model
dp1
= c1 p1 (1 − p1 ) − m1 p1 (3.14)
dt
originated by Levins,14 which can be rewritten in logistic form as
!
dp1 p1
= (c1 − m1 )p1 1 − . (3.15)
dt 1− m 1
c1

We observe that r = c1 − m1 and K = 1 − m c1 are positive if and only if c > m.

1

From our analysis earlier in this section, we know that (3.15) has the solution
m1
1− c1
p1 = ,
1+ Ce−(c1 −m1 )t
14 R. Levins, Some demographic and genetic consequences of environmental heterogeneity

for biological control, Bulletin of the Entomological Society of America 15: 237–240, 1969.
116 Dynamical Systems for Biological Modeling: An Introduction
 
where C = 1 − m c1 − p1 (0) /p1 (0), as long as p1 (0) 6= 0. Therefore, if c1 >
1

m1 then p1 → 1 − m c1 , whereas if c1 < m1 then p1 → 0. This result agrees

1

with intuition, in that the criterion for the species to survive is that it should
reproduce faster than it dies out.
Now suppose that there is another species, species 2, which competes with
species 1 for resources (and space) in the rainforest. Suppose that species 1
dominates species 2 in this competition, so that species 2 can only grow in
those sites not occupied by species 1. Then this inferior competitor colonizes
the proportion (1−p1 −p2 ) of sites occupied by neither the superior competitor
nor itself at a base rate of c2 p2 , and vacates sites due to natural mortality at
a rate m2 p2 , and due to displacement by the superior competitor at a rate
c1 p1 p2 (species 1 colonizes at a rate c1 p1 ; a proportion p2 of those sites are
occupied by species 2, causing displacement). These assumptions, together
with those on species 1, give us a system of equations
dp1
=c1 p1 (1 − p1 ) − m1 p1 ,
dt
dp2
=c2 p2 (1 − p1 − p2 ) − m2 p2 − c1 p1 p2 . (3.16)
dt
As species 1 remains unaffected by species 2, this system is said to decouple,
that is, we can analyze the growth of the two species separately. Species 1
will approach its carrying capacity p∗1 as discussed in the previous paragraph,
p1 → p∗1 , at which point (3.16) becomes

dp2
= c2 p2 (1 − p∗1 − p2 ) − m2 p2 − c1 p∗1 p2 ,
dt
which we can rewrite in explicit logistic form as
 
dp2 p2
= r2 p2 1 − ,
dt r2 /c2

where
c2 m

c1 − m2 − (c1 − m1 ), c1 > m1 ;
1
r2 = c2 − m2 − (c1 + c2 )p∗1 =
c2 − m 2 , c1 < m 1 .

The behavior of p2 thus again depends on the sign of the coefficient r2 , ap-
proaching r2 /c2 if r2 > 0 and 0 if r2 < 0. Under the (reasonable) assumption
that for each species i found in the rainforest, ci > mi , that is, that in the
absence of competition the species would persist, we take the corresponding
form for r2 and write the persistence condition for species 2 as
c1
c2 > (m2 + c1 − m1 ).
m1
This system can be extended to a complete hierarchy of n species, ranked
 First-Order Differential Equations (Continuous Dynamical Systems) 117

by dominance in competition (with species i dominating and displacing

species j whenever i < j), so that, for instance,
dp3
= c3 p3 (1 − p1 − p2 − p3 ) − m3 p3 − (c1 p1 + c2 p2 )p3 ,
dt
and more generally
 
i i−1
dpi X X
= ci pi (1 − p j ) −  mi + cj p j  p i . (3.17)
dt j=1 j=1

Since the competition is completely ordered, each equation in the system can
be analyzed independently (beginning with species 1), rewritten as a logistic
equation, and shown to approach a unique equilibrium value, which is the
greater of 0 and
i−1   
mi X ∗ cj
p∗i = 1 − − pj 1 + . (3.18)
ci j=1
ci

Note that the last term in the expression for p∗i represents the reduction in
distribution of species i due to displacement and exclusion by its superior
competitors. Tilman observed that depending on parameter values (i.e., a
species’s ability to colonize and survive), an inferior competitor may actually
occupy a greater proportion of sites than a superior competitor. For instance,
suppose species 1 has a high mortality rate, so that p∗1 is small. Then if species 2
has a low mortality rate,
 
m2 c1 m2
p∗2 = 1 − − p∗1 1 + ≈1− > p∗1 .
c2 c2 c2
This model can be extended to accommodate, for example, habitat de-
struction and fragmentation (see Exercises 8 and 9 below). Bampfylde, who
was interested in modeling competition in rainforests, observed that displace-
ment does not occur among rainforest species, so competition is purely for
sites opened through natural mortality. In this case, the displacement term in
(3.17) is omitted, and the colonization term is adjusted to exclude displace-
ment. ThusPthe colonization rate ci pi is reduced by the proportion of open
n
sites, (1 − j=1 pj ), rather than merely by the proportion of sites not occu-
Pi
pied by superior competitors, (1 − j=1 pj ). The resulting model does not
decouple, however, so we shall reserve its further study for Chapter 6.

3.2.3 The spread of infectious diseases

Another biological science to which mathematics has made important con-
tributions is epidemiology. Mathematical models for the spread of infectious
diseases typically divide populations into distinct classes based on epidemi-
ological or demographic factors relevant to the transmission of the disease:
118 Dynamical Systems for Biological Modeling: An Introduction

susceptible and infected individuals, high-risk and low-risk individuals, highly

infectious and less infectious individuals, etc. For this reason they are often
referred to as compartmental models. Most compartmental models involve
keeping track of several populations, an activity which requires as many state
variables (and equations) as compartments. The simplest possible epidemic
model, however, can be rewritten as a single logistic equation, and so we in-
clude it here.
Consider a population of constant size N in which an infectious disease
is introduced. We divide the population into two classes: susceptibles (not
infected) and infectives (infected and contagious). Let S(t) denote the num-
ber of susceptibles at time t and let I(t) denote the number of infectives, so
that S(t) + I(t) = N . We assume that the disease is spread from infectives to
susceptibles through contact. Suppose that an “average” infective makes po-
tentially infective contacts with a constant number βN of individuals in unit
time. Then, presumably βI(t) of these individuals are already infected, and
the number of new infections caused by an “average” infective in unit time is
βS(t). Thus the total number of new infections in unit time is βS(t)I(t). We
assume also that the disease is never fatal, and that in unit time a fraction
γ of the infectives recover, so that the number of recoveries in unit time is
γI(t). This assumption is equivalent to the assumption that for each s ≥ 0
the fraction of the infectives who remain infective for a time interval s is e−γs ,
and that the average length of the infective period is 1/γ. Finally, we assume
that on recovery infectives have no immunity against re-infection and return
to the susceptible class.
We may formulate a model to describe S(t) and I(t) by thinking of a flow
rate βSI from the susceptible class to the infective class and a flow rate γI
from the infective class to the susceptible class. As noted with some previous
population models, S and I should, properly speaking, take on only integer
values, to count individuals. However, we will allow them to be real-valued
in this model, and consider the model either as an approximation valid for
large populations N , or (if we set N = 1) as a representation of proportions
of the population susceptible and infective. (Models for small populations are
usually written as discrete processes, and stochasticity also plays an important
role on this scale.)
We now have a pair of differential equations
S ′ = −βSI + γ I, I ′ = βSI − γ I,
a model first described by W. O. Kermack and A. G. McKendrick.15 Since
S(t)+ I(t) = N for all t, we may replace S by N − I and describe the situation
by a single differential equation:
" #
′ I
I = βI(N − I) − γ I = (βN − γ)I 1 − βN −γ . (3.19)
β

15 W.O. Kermack and A.G. McKendrick, Contributions to the mathematical theory of

epidemics, Part II, Proc. Royal Soc. London 138 (1932), 55–83.
 First-Order Differential Equations (Continuous Dynamical Systems) 119

The differential equation (3.19) has the same form as the logistic equation
(3.8), with r replaced by βN − γ) and K replaced by βNβ−γ . It is important
to note that, as was true for the plant competition model studied earlier in
this section, in (3.19) it is possible for r = βN − γ to be either positive or
negative. Recalling the solution (3.12) of (3.8), we again observe that if r > 0
(and K > 0), then every solution y(t) with y(0) > 0 has limt→∞ y(t) = K,
while if r < 0, then every solution y(t) > 0 has limt→∞ y(t) = 0. If we translate
this result to (3.19), we see that if βN/γ < 1, then limt→∞ I(t) = 0, while if
βN/γ > 1, then limt→∞ I(t) = N − βγ > 0.
This result is the famous threshold theorem of Kermack and McKendrick:
If the quantity βN/γ, called the basic reproductive number, is less than 1,
then the number of infectives approaches zero, and the number of susceptibles
approaches N . In epidemiological terms this means that the infection dies
out. On the other hand, if the basic reproductive number βN/γ exceeds 1,
then the number of infectives remains positive and tends to a positive limit.
In epidemiological terms, this means that the infection remains endemic. The
quantity βN/γ represents the number of secondary infections caused by each
infective over the duration of the infection. Since βN is a number of contacts
per infective in unit time, the dimensions of βN are time−1 . Thus the basic
reproductive number βN/γ is dimensionless, and does not depend on the units
used in describing the model.
One infectious disease which has been studied mathematically by models
such as (3.19), as well as by considerably more detailed models, is gonorrhea.
For gonorrhea the average infective period (1/γ) is known to be about one
month. However, the transmission coefficient β must be estimated by some
indirect means. One way to estimate β is to use the fact that for small values
of I a solution of (3.19) grows exponentially with exponent r = (βN − γ).
Thus if an infection in some region begins with a small number of infectives,
the time t̂ for the number of infectives to double (ert̂ = 2) is approximately

log 2 log 2
= .
r βN − γ
In one reported gonorrhea outbreak this doubling time was approximately
1.7 months, which leads to the estimate βN/γ = 1.4, which indicates a basic
reproductive number of 1.4.

Example 2.
A new strain of influenza is introduced into a town with 1200 inhabitants by
two visitors. Assume that the average infective is in contact with 0.4 inhabi-
tants per day and that the average duration of the infective period is 6 days.
Will the infection die out or will the flu persist?
Solution: The number of potentially infective contacts per infective per day
is βN = 0.4 days−1 . Thus the basic reproductive number is βN/γ =
(0.4 days−1 )(6 days) = 2.4; since this exceeds 1, the flu will persist. 
120 Dynamical Systems for Biological Modeling: An Introduction

There are other problems such as the spread of a rumor and rates of
chemical reactions which may also lead to a logistic differential equation model
(see Exercises 17 and 18 below, and Exercise 1 in Section 4.3).

Exercises
1∗ . Suppose that a population satisfies a logistic model (3.8). Show that its
growth rate y ′ is at a maximum when y = K 2 and that its maximum
rK ′
growth rate is 4 . Hint: Maximize y by setting

d y 
(y ′ )′ = ry(1 − ) = 0.
dt K

2. The population of the United States in 1970 was 202.0 (millions) and the
per capita rate of growth was approximately 0.7% per year. Use these
data and an exponential growth model to predict the population in the
years 1990, 2000, and 2100.

3. Suppose that the population of the United States satisfies a logistic

model with carrying capacity 300 (millions). (a) Estimate the value of r
in the logistic model by using the observed population size and per capita
growth rate in 1970 as given in Exercise 2. (b) Predict the population
in the years 1990, 2000, and 2100.
4. (a) Compare the predictions of the exponential and logistic models in
Exercises 2 and 3 for the years 1990 and 2000 with each other and with
actual data. (b) Compare their predictions for the year 2100. Which
seems more realistic? Explain.
5∗ . Evaluate the logistic model as a predictor of the growth of the U.S.
population (Example 1, this section).
(a) The actual population (P (t) in the table) is generally slightly ahead
of the prediction y(t) through 1910. Beginning in 1920, however, it
is significantly behind the prediction. Why?
(b) At what point, if any, does the model stop being an effective pre-
dictor? Speculate on possible causes for this change, and how they
might be incorporated into the model.
6. Suppose a population satisfies the logistic model (3.8) with r = 0.4,
K = 100, y(0) = 5. Find the population size for t = 10.
7. The Pacific halibut fishery has been modeled by the logistic equation
(3.8) with parameters estimated as r = 0.71 yr−1 , K = 80.5 × 106 kilo-
grams, where y(t) is the total biomass at time t measured in kilograms.
If y(0) = K4 , find the biomass 1 year later. Also, find how long it will
take for the total biomass to increase to K
2.
 First-Order Differential Equations (Continuous Dynamical Systems) 121

8. Tilman et al.16 generalized Tilman’s multi-species competition model to

include habitat destruction. If a proportion D of the sites are destroyed,
only the remaining proportion 1 − D can be occupied. Rewrite (3.17)
and the corresponding equilibrium values (3.18) to accommodate this
assumption.
9. For the one-species version of the habitat destruction model of Tilman
et al. discussed in Exercise 8, find the persistence criterion for the species
(which previously was simply c1 > m1 ) given D.
10. Find parameter values ci , mi for a three-species version of Tilman’s
multi-species competition model (3.17) such that 0 = p∗2 < p∗1 < p∗3 , that
is, the intermediate competitor is driven to extinction by competition
with species 1, while the inferior competitor occupies more sites than
the superior competitor by virtue of longevity. (Recall the assumption
that ci > mi .)
11. Calculate the net maximum growth rate ri for species i in Tilman’s
multi-species competitionmodel (3.17)
 by rewriting the equation in lo-
pi
gistic form, dpi /dt = ri pi 1 − p∗ , with p∗i as given in (3.18).
i

12. Show that Bampfylde’s displacement-free rainforest competition model

as described in this section does not decouple, by writing out the equa-
tions for the two-species version: alter (3.17) by omitting the final term
and replacing (1 − ij=1 pj ) with (1 − nj=1 pj ) in the colonization term.
P P

13. A disease begins to spread in a population of 800. The infective period

has an average duration of 14 days and the average infective is in contact
with 0.1 person per day. What is the basic reproductive number? To
what level must the average rate of contact be reduced so that the
disease will die out?
14∗ European fox rabies is estimated to have a transmission coefficient β of
80 km2 years−1 and an average infective period of 5 days. There is a
critical carrying capacity Kc measured in foxes per km2 , such that in
regions with fox density less than Kc rabies tends to die out, while in
regions with fox density greater than Kc rabies tends to be endemic.
Estimate Kc .
[Remark: It has been suggested in Great Britain that hunting to reduce
the density of foxes below the critical carrying capacity would be a way
to control the spread of rabies.]
15∗ . A communicable disease from which infectives do not recover may be
modeled by the pair of differential equations
S ′ = −βSI, I ′ = βSI.
16 David Tilman, Robert M. May, Clarence M. Lehman, and Martin A. Nowak, Habitat

destruction and the extinction debt, Nature 371: 65–66, 1994.

122 Dynamical Systems for Biological Modeling: An Introduction

(a) Show that in a population of fixed size K, such a disease will even-
tually spread to the entire population.
(b) Given the meaning in practical terms of “eventually,” what mod-
eling assumption made implicitly above accounts for the fact that
there always remain some individuals who never get infected, even
if the population remains closed and of fixed size, and infected in-
dividuals really never recover.
16∗ . Consider a disease spread by carriers who transmit the disease without
exhibiting symptoms themselves. Let C(t) be the number of carriers
and suppose that carriers are identified and isolated from contact with
others at a constant per capita rate α, so that C ′ = −αC. The rate
at which susceptibles become infected is proportional to the number of
carriers and to the number of susceptibles, so that S ′ = −βSC. Let C0
and S0 be the number of carriers and susceptibles, respectively, at the
time t = 0.

(a) Determine the number of carriers at time t from the first equation.
(b) Substitute the solution to part (a) into the second equation and
determine the number of susceptibles at time t.
(c) Find the number of members of the population who escape the
disease, limt→∞ S(t).
17∗ . Consider a population of fixed size K in which a rumor is being spread
by word of mouth. Let y(t) be the number of people who have heard
the rumor at time t and assume that everyone who has heard the rumor
passes it on to r others in unit time. Thus from time t to time t + h the
rumor is passed on rh y(t) times, but a fraction y(t)
K of the people
 who
hear it have already heard it and thus there are only rh y(t) K−y(t)
K
people who hear the rumor for the first time. Use these assumptions to
obtain an expression for y(t + h) − y(t), divide by h and take the limit
as t → ∞ to obtain a differential equation satisfied by y(t).
18. At 9 AM, 1 person in a village of 100 inhabitants has heard a rumor.
Suppose that everyone who has heard the rumor tells one other person
per hour. Using the model of Exercise 17, determine how long it will
take until half the village has heard the rumor.
 First-Order Differential Equations (Continuous Dynamical Systems) 123

3.3 Graphical analysis

In Chapter 2 we explored cobwebbing, a method for graphical analysis of
difference equations which yields both exact numerical solutions for discrete-
time initial value problems (difference equations with initial conditions) and
qualitative information about the nature of solutions to the given difference
equation. In this section we will explore the use of direction fields, a method
for graphical analysis of differential equations which yields similar information
for continuous-time initial value problems (differential equations with initial
conditions). First, however, we revisit the notions of solutions and families
of solutions introduced in Section 3.1 in order to present an important result
about when solutions are certain to exist, which will inform our use of direction
fields.

3.3.1 Solutions
Recall that by a differential equation we mean simply a relation between
an unknown function and its derivatives. The general form of a first-order
differential equation is
dy
y′ = = f (t, y), (3.20)
dt
with f a given function of the two variables t and y. By a solution of the
differential equation (3.20) we mean a differentiable function y of t on some
t-interval I such that, for every t in the interval I,

y ′ (t) = f {t, y(t)}.

In other words, differentiating the function y(t) results in the function f (t, y).
For example, as we saw in Example 1 of Section 3.1, the function y = 2eat is a
solution of the differential equation y ′ = ay on every t-interval. We see this by
differentiating 2eat to get a 2eat , which we can rewrite as ay. To verify whether
a given function is a solution of a given differential equation, we need only
substitute into the differential equation and check whether it then reduces to
an identity.

Example 1.
1
Show that the function y = t+1 is a solution of the differential equation
y ′ = −y 2 .
Solution: For the given function,
dy 1
=− = −y 2
dt (t + 1)2
and this shows that it is indeed a solution. 
124 Dynamical Systems for Biological Modeling: An Introduction

In the same way we can verify that a family of functions satisfies a given
differential equation. By a family of functions we mean a function which in-
cludes an arbitrary constant, so that each value of the constant defines a
distinct function. √ The family ce5t , for instance, includes the functions e5t ,
5t 5t
−4e , 12e , and 3 e5t , among others. When we say that a family of func-
tions satisfies a differential equation, we mean that substitution of the family
(i.e., the general form) into the differential equation gives an identity satisfied
for every choice of the constant.

Example 2.
1
Show that for every c the function y = t+c is a solution of the differential
′ 2
equation y = −y . (This equation models a decaying population whose per
capita mortality rate is linear in population size, rather than constant.)
Solution: For the given function,
dy 1
=− = −y 2 ,
dt (t + c)2

and this shows that each member of the given family of functions is a solution.


Example 3.
Show that the family of functions

1 + cet
y=
1 − cet
is a solution of the differential equation

y2 − 1
y′ =
2
for every value of the constant c.
Solution: For the given function y,

cet (1 − cet ) − (−cet )(1 + cet )

y′ = ,
(1 − cet )2

y2 − 1 1 (1 + cet )2 − (1 − cet )2
 
= ,
2 2 (1 − cet )2
and algebraic simplification shows that these two expressions are equal. 

In applications we are usually interested in finding not a family of solutions

of a differential equation but a single solution which satisfies some additional
requirement, usually an initial or boundary condition. Geometrically, an ini-
tial condition picks out the solution from a family of solutions which passes
 First-Order Differential Equations (Continuous Dynamical Systems) 125

through the point (t0 , y0 ) in the t-y plane. Physically, this corresponds to mea-
suring the state of a system at the time t0 and using the solution of the initial
value problem to predict the future behavior of the system.

Example 4.
1
Find the solution of the differential equation y ′ = −y 2 of the form y = t+c
which satisfies the initial condition y(0) = 1.
1
Solution: We substitute the values t = 0, y = 1 into the equation y = t+c ,
1
and we obtain a condition on c, namely 1 = c , whose solution is c = 1.
The required solution is the function in the given family with c = 1, namely
1
y = t+1 .

Example 5.
Find the solution of the differential equation y ′ = −y 2 which satisfies the
general initial condition y(0) = y0 , where y0 is arbitrary.
1
Solution: We substitute the values t = 0, y = y0 into the equation y = t+c and
1 1
solve the resulting equation y0 = c for c, obtaining c = y0 provided y0 6= 0.
Thus the solution of the initial value problem is
1
y=
(t + y10 )2

except if y0 = 0. If y0 = 0, there is no solution of the initial value problem of

the given form; in this case the identically zero function, y = 0, is a solution.
We have now obtained a solution of the initial value problem with arbitrary
initial value at y = 0 for the differential equation y ′ = −y 2 . 

A family of solutions may arise if we are considering a differential equa-

tion with no initial condition imposed, and we will then also be concerned
with the question of whether the given family contains all solutions of the
differential equation. To answer this question, we will need to make use of a
theorem which guarantees that each initial value problem for the given differ-
ential equation has exactly one solution. More specifically, if an initial value
problem is to be a usable mathematical description of a scientific problem, it
must have a solution, for otherwise it would be of no use in predicting be-
havior. Furthermore, it should have only one solution, for otherwise we would
not know which solution describes the system. Thus for applications it is vital
that there be a mathematical theory telling us that an initial value problem
has exactly one solution. This was not an issue in Chapter 2; for a difference
equation we can always iterate (in theory) to form a solution. For differential
equations, however, solutions are not obtained by a direct construction, and
it is necessary to have some result telling us that a problem really does have a
solution. Fortunately, there is a very general theorem which tells us that this
is true for the initial value problem (3.20), y(0) = y0 provided the function
126 Dynamical Systems for Biological Modeling: An Introduction

f is reasonably smooth. We will state this result and ask the reader to ac-
cept it without proof because the proof requires more advanced mathematical
knowledge than we have at present.

EXISTENCE AND UNIQUENESS THEOREM: If the function

f (t, y) is continuous and differentiable with respect to y in some region of
the plane which contains the point (t0 , y0 ), then the initial value problem
consisting of the differential equation y ′ = f (t, y) and the initial condi-
tion y(t0 ) = y0 has a unique solution which is defined on some t-interval
containing t0 in its interior.

Even though the function f (t, y) may be well-behaved in the whole t-y
plane, there is no assurance that a solution will be defined for all t. As we
1
have seen in Example 1, the solution y = t+1 of y ′ = −y 2 , y(0) = 1 exists only
for −1 < t < ∞. As we have seen in Example 2, each solution of the family
1
of solutions y = t+c has a different interval of existence. In Example 5 we
have shown how to rewrite a family of solutions for a differential equation in
terms of an arbitrary initial condition — that is, as a solution of an initial value
problem for that differential equation. We have also seen how to identify those
initial conditions which cannot be satisfied by a member of the given family.
Often there are constant functions which are not members of the given family
but which are solutions and satisfy initial conditions that cannot be satisfied
by a member of the family. The existence and uniqueness theorem tells us that
if we can find a family of solutions, possibly supplemented by some additional
solutions, so that we can find this collection contains a solution corresponding
to each possible initial condition, then we have found the set of all solutions
of the differential equation.

3.3.2 Direction fields

The geometric interpretation of a solution y(t) to a differential equation
(3.20) is that the curve y = y(t) has slope f (t, y) at each point (t, y) along its
length. Thus we might think of approximating the solution curve by piecing
together short line segments whose slope at each point (t, y) is f (t, y). To
realize this idea, we construct at each point (t, y) in some region of the plane
a short line segment with slope f (t, y). The collection of line segments is called
the direction field of the differential equation (3.20). The direction field can
help us to visualize solutions of the differential equation since at each point
on its graph a solution curve is tangent to the line segment at that point. We
may sketch the solutions of a differential equation by connecting these line
segments by smooth curves.
Drawing direction fields by hand is a difficult and time-consuming task.
 First-Order Differential Equations (Continuous Dynamical Systems) 127

There are computer programs, both self-contained and portions of more elab-
orate computational systems such as Mathematica, MATLAB and Maple,
which can generate direction fields for a differential equation and can also
sketch solution curves corresponding to these direction fields. We give some
examples here which have been produced by Maple (see Appendix A); the
reader with access to a facility which is capable of drawing direction fields is
urged to reproduce these examples before trying to produce other direction
fields.

Example 6.
Draw a direction field and some solutions of the differential equation y ′ = y.
(This is the prototypical linear homogeneous differential equation, with the
time unit rescaled so that the growth rate a is equal to 1; see Section 3.1 for
applications.)
Solution: See Figure 3.9. The direction field suggests exponential solutions,
which we know from Section 3.1 to be correct. 
4 1

0.8
3

0.6

y(t) 2 y(t)

0.4

1
0.2

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.2 0.4 0.6 0.8 1
t t

–0.2
–1

–0.4

–2

FIGURE 3.9: Direction field and FIGURE 3.10: Direction field and
solutions for y ′ = y. solutions for y ′ = −y 2 .

Example 7.
Draw a direction field and some solutions of the differential equation y ′ = −y 2 .
Solution: See Figure 3.10. The direction field indicates that solutions below
the t-axis become unbounded below while solutions above the t-axis tend to
zero as t → ∞, as could be seen from Example 2. 
128 Dynamical Systems for Biological Modeling: An Introduction

Example 8.
Draw a direction field and some solutions of the differential equation y ′ =
y(1 − y). This is the prototypical logistic differential equation, in which both
time and population units have been rescaled so that r and K are both equal
to 1.
Solution: See Figure 3.11. The direction field indicates that solutions below
the t-axis are unbounded below while solutions above the t-axis tend to 1 as
t → ∞, consistent with what we have established for this logistic differential
equation. 
1.5 1.5

1 1

y(t) y(t)

0.5 0.5

0 0.2 0.4 0.6 0.8 1 –2 –1 1 2

t t

–0.5 –0.5

–1 –1

FIGURE 3.11: Direction field and FIGURE 3.12: Direction field and
solutions for y ′ = y(1 − y). solutions for y ′ = 1 − t2 − y 2 .

Example 9.
The differential equation y ′ = 1 − y 2 is a slight variation on the logistic
equation y ′ = y − y 2 ; the important difference between the two for purposes
of modeling biological systems is that when y = 0 the new model has y ′ = 1,
rather than y ′ = 0 as the logistic model has. This allows the population to
grow to its asymptotic value of 1 even when measurement errors in the initial
population size, which may be very small (given the rescaling of units so that
K = 1), yield a value of 0 or even a small negative number. Although the new
model is not quite biologically sound at y = 0, it may be used with the given
caveat.
Suppose now that we wish to incorporate an additional feature into the
model: namely, an eventual decline in population once it has reached its carry-
ing capacity (as mentioned in Section 3.2, such declines have been observed fol-
 First-Order Differential Equations (Continuous Dynamical Systems) 129

lowing the initial growth period). A simple modification might involve adding
a negative term dependent on t, say −t2 , to y ′ . Draw a direction field and
some solutions of the differential equation y ′ = 1 − y 2 − t2 , to evaluate the
effects of this modification.
Solution: See Figure 3.12. It can be seen that regardless of initial conditions
(i.e., above or below the carrying capacity of 1), solutions eventually grow
negative and unbounded. The −t2 term does produce an eventual decline
from y = 1, but the decline is too strong; the term needs to be modified to
restrict its maximum size, so that solutions remain positive. 

As the reader can see, direction fields provide a useful visual indication
of system behavior for a range of initial conditions. Although the solutions
obtained by connecting the segments are approximate rather than exact, and
approximate numerical solutions for specific initial conditions are found more
commonly via numerical analysis, direction fields are useful tools in suggesting
qualitative behavior that varies by initial condition in cases when it may be
difficult to produce a complete analysis using the usual qualitative tools (which
will be developed in Section 3.5).
The geometric view of differential equations presented by the direction field
will appear again when we examine some qualitative properties in Section 4.1,
and is also the basis of the numerical approximation of solutions in Section 4.5.

Exercises
1
1. Show that y = (1 − t2 )− 2 is a solution of the differential equation y ′ =
ty 3 .
2. Show that y = 3t is a solution of the differential equation y ′ = y/t.
1
3. Show that y = (c−t2 )− 2 is a solution of the differential equation y ′ = ty 3
for every choice of the constant c.
4. Show that y = ct is a solution of the differential equation y ′ = y/t for
every choice of the constant c.
1
5. Among the family of solutions y = (c − t2 )− 2 of y ′ = ty 3 , find the
solution such that y(0) = 1.
6. Find the solution of y ′ = ty 3 such that y(0) = 0.
7. Show that the solution of the initial value problem y ′ = ty 3 , y(0) = −1
1
is y = −(1 − t2 )− 2 .
t
1+ce ′ 1 2
8. Among the solutions y = 1−ce t of the differential equation y = 2 (y −1),

find the ones which satisfy the initial conditions y(0) = 1, y(0) = 0, and
y(2) = 0.
130 Dynamical Systems for Biological Modeling: An Introduction

9. Show that if ŷ is a constant such that f (t, ŷ) = 0 for all t, then y = ŷ is
a constant solution of y ′ = f (t, y).
10∗ . * If y(t) is a solution of y ′ = f (t, y) which has a local maximum at the
point (t̂, ŷ), show that f (t̂, ŷ) = 0.
11. Draw a direction field and some solutions of the differential equation
y ′ = ty.
12. Draw a direction field and some solutions of the differential equation
y ′ = t2 + y 2 .
13. Draw a direction field and some solutions of the differential equation
y′ = 1 + y2.
14. Draw a direction field and some solutions of the differential equation
y ′ = y − t.
p
15∗ . For the differential equation y ′ = −t + 2y + t2 ,
(a) Show that y = ct+ 12 c2 is a solution for every choice of the constant
c.
2
(b) Show that y = − t2 is a solution.
2 2
(c) Find the tangent line to the curve y = − t2 at the point (−c, − c2 ).
16∗ . Improve the modification to the model y ′ = 1 − y 2 presented in Example
9 so that it does not cause all solutions to grow unbounded and negative.

3.4 Equations and models with variables separable

In this section we shall learn a method for finding solutions of a class
of differential equations. Although exact solutions are not necessary in order
to use models to answer questions about biological systems, they provide
additional information that can lead to helpful insights for many of the most
common model equations.
The method of separation of variables is based on the approach we used in
Section 3.1 to solve the differential equation of exponential growth or decay,
and is applicable to differential equations with variables separable. A differen-
tial equation is called separable, or is said to have variables separable, if the
function f which describes the rate of change of the population y over time
can be expressed in the form

f (t, y) = p(t)q(y); (3.21)

that is, if the changes in population size due to population density y can be
 First-Order Differential Equations (Continuous Dynamical Systems) 131

factored out from those due to the time index t (such as seasonal factors). For
y
example, y ′ = 1+t ′
2 and y = y
2
are both separable, whereas y ′ = sin t − 2ty is
not separable. The examples discussed in Section 3.1 are also separable, and
the method of solution described for equation (3.1) of Section 3.1 is a special
case of the general method of solution to be developed for separable equations.
The reason for the name separable is that the differential equation can then
be written as
y′
= p(t),
q(y)
with all the dependence on y on the left and all the dependence on t on the
right-hand side of the equation, provided that q(y) 6= 0 (indeed, this is the
first step in the method). An important special case, which includes most of
the examples in this chapter, is when f is completely independent of t and is
a function of y only (i.e., p(t) = 1); in this case, the equation is said to be
autonomous. The method of separation of variables applies to many relatively
complex differential equations, and is worth learning for this reason.
Although the models and systems explored in this section are thus united
by a mathematical, rather than biological, feature, two classes of models which
allow separation of variables are the linear and logistic differential equations,
which form the basis for many models of biological systems. We shall begin
our discussion with some applications of these models, and continue with some
other, more specialized separable models.

3.4.1 A linear model for the cardiac pacemaker

The firing of neurons and the regulation of cardiac activity have long been
the subjects of intense study, including numerous attempts to capture via
modeling the nuances of the cardiac pacemaker, the sinoatrial node in the
heart whose cells provide the impulses that regulate heartbeats. Most of these
complex models are well beyond the scope of this chapter, but we shall use
this context to adapt a simpler, linear model that has been proposed for the
periodic firing of neurons in general, and of the cardiac pacemaker wave in
particular.17
The sinoatrial node consists of a cluster of cells which depolarize, fire, and
then reset periodically to send a wave of electrical signals that cause the atria
of the heart to contract. After reaching the atrioventricular node, the pulses,
distributed through the His-Purkinje system, then cause the ventricles to con-
tract a fraction of a second later, pumping blood from the heart to the rest of
the body (see Figure 3.13). The mechanism that produces these impulses in
the sinoatrial node involves differences in concentrations of sodium, calcium,
and potassium ions inside and outside the cells. Small channels in the cell
17 See, e.g., B.W. Knight, Dynamics of encoding in a population of neurons, J. General

Physiology 59: 734–766, 1972, and D.S. Jones and B.D. Sleeman, Differential equations and
mathematical biology, Boca Raton, FL: CRC Press, 2003.
132 Dynamical Systems for Biological Modeling: An Introduction

membranes open and close to change the electrical potential via an exchange
of particular ions.
Single autonomous differential equations typically do not lead to sustained
oscillations like a heartbeat, but we can obtain periodic behavior by intro-
ducing a control variable u that is sensitive to the cellular voltage levels,
and switches on and off when the voltage crosses set thresholds. In this way
we may think of the control as a regulating force like that of an artificial
pacemaker (although such devices’ controls are typically calibrated to times,
rather than voltages). We shall assume, then, that we know four voltage lev-
els: the control values uL and uH and the threshold values yL and yH , with
uL < yL < yH < uH . We shall also assume that the control signal u works as
follows: when the cellular voltage level is low, y(t) ≤ yL , the control switches
on, u = uH ; when the voltage level rises above the upper threshold, y(t) ≥ yH ,
the neuron fires and the control switches off, u = uL . While the voltage level is
rising or falling between the two thresholds, yL < y < yH , the control remains
at its present level. Finally, we shall assume that, apart from the influence
of the control signal, the pacemaker signal y(t) naturally decays at a rate
proportional to its present level.
These assumptions lead to a linear heterogeneous model for the pacemaker
signal voltage,
y ′ = au − ay = a(u − y), (3.22)
where a > 0 is a rate constant which determines how quickly the signal decays
from its previous values (such constants are sometimes said to measure the
“forgetfulness” of the system). Although the control signal u may change from
time to time as the voltage y crosses the upper or lower threshold, it is said to
be piecewise constant—that is, constant on short periods of time—and so we
will study how this system behaves on these short periods, later connecting
the solutions during each period to form a long-term picture.
Equation (3.22) is separable, and we shall solve it using the method of sep-
aration of variables, which we will explain more generally later in this section.
(Another method for solving linear differential equations will be presented
in Section 3.5.) To separate variables, we first divide both sides of (3.22) by
a(u − y), which is permissible as long as y 6= u (and remember that the model,
if successful, will constrain y within [yL , yH ], so that it never reaches uL or
uH ). This yields
1 dy
= 1.
a(u − y) dt
Integrating both sides of this equation with respect to t, considering y as a
function y(t) of t,
1 dy
Z Z
dt = 1 dt,
a(u − y) dt
we simplify to
dy
Z Z
= dt. (3.23)
a(u − y)
 First-Order Differential Equations (Continuous Dynamical Systems) 133

Alternatively, and more colloquially, we can treat the derivative dy/dt as a

ratio of differentials dy and dt, and multiply equation (3.22) by dt/a(u − y),
to obtain
dy
= dt.
a(u − y)
This form of the equation is meaningless without an interpretation of differ-
entials, but integration then yields the meaningful form (3.23) above.
Integration of (3.23) gives
1
− log |y − u| = t + c,
a
with c an arbitrary constant of integration. Algebraic solution now gives

log |y − u| = −a(t + c)

|y − u| = e−a(t+c) = e−ac e−at

y − u = ±e−ac e−at
y = u ± e−ac e−at .
We now rename the arbitrary constant ±e−ac as a new arbitrary constant, C,
which measures the initial distance of the cellular voltage from the control.
We thus obtain the family of solutions

y = u + Ce−at . (3.24)

Note that although our separation of variables explicitly excluded the constant
solution y = u, this solution is contained in the family (3.24).
In order to find the solution of (3.22) which satisfies the initial condition

y(0) = y0 (3.25)

we substitute t = 0, y = y0 into (3.24), obtaining y0 = u + C or C = y0 − u.

This value of C gives the solution

y = u + (y0 − u)e−at = u(1 − e−at ) + y0 e−at (3.26)

of the initial value problem (3.24), (3.25).

In applications, the specific form of the solution of a differential equation
is often less important than the behavior of the solution for large values of
t. Because e−at → 0 as t → ∞ when a > 0, we see from (3.24) that every
solution of (3.22) tends to the limit u as t → ∞ if a > 0. This is an example
of qualitative information about the behavior of solutions which will be useful
in applications. In Section 4.1 we shall examine other qualitative questions—
information about the behavior of solutions of a differential equation which
may be obtained indirectly rather than by explicit solution.

If we now set time to zero at the moment when the control switches on,
134 Dynamical Systems for Biological Modeling: An Introduction
sinoatrial
(SA) node y(t)
uH
yH

yL
uL
bundle t
of His 0 T1 T1+T
2

u(t)
uH

uL
Purkinje
fibers t
atrioventricular
(AV) node 0 T1 T1+T
2
Illustration by Carolyn Kribs

FIGURE 3.13: The cardiac electri- FIGURE 3.14: Pacemaker and con-
cal system. trol voltage over one period.

we have y0 = yL , and u = uH for t on an interval [0, T1 ], where T1 is the time

required for the voltage to reach the upper threshold and cause the neuron to
fire. During this interval, we have, from (3.26),

y(t) = uH + (yL − uH )e−at .

Thus the voltage rises from yL toward uH , until it crosses the upper threshold
yH , at which point the neuron fires and the control resets to uL . At this
moment we have y(T1 ) = yH , which we use as a new initial condition in (3.26)
to obtain a description of the voltage as it falls back toward uL ,

y(t) = uL + (yH − uL )e−a(t−T1 )

(t − T1 is time since T1 ). This description of y holds until the voltage falls

below the lower threshold yL to begin the cycle over again T2 time units later
(i.e., for T1 < t ≤ T1 + T2 ). Figure 3.14 connects the graphs of y on these two
intervals to show a complete period, alongside the corresponding graph of the
control variable u. The graphs repeat every T1 + T2 time units.
In practice, we might use such a model to predict the behavior of the car-
diac rhythm under different conditions (such as an alteration of the threshold
voltages, or of the control signal—see Exercise 18 below), or to calculate infor-
mation that is not directly observable, based on known data. For instance, if
we know three of the voltage parameters and the two time periods T1 and T2 ,
 First-Order Differential Equations (Continuous Dynamical Systems) 135

we can determine the remaining voltage parameter and the rate constant a.
Conversely, if we know the four voltage parameters and a, we can determine
T1 and T2 . These relationships come from the continuity of the voltage y at
the points where the control signal changes:
y(T1 ) = uH + (yL − uH )e−aT1 = yH (3.27)
and
y(T1 + T2 ) = uL + (yH − uL )e−a(T2 ) = yL . (3.28)
Experimental observations have shown that the voltage level ranges from
yL = −70 mV to yH = +30 mV during one period, which lasts about one
second. If observations show T1 ≈ 600 msec and T2 ≈ 400 msec, and we
suppose a lower control voltage of uL = −100 mV, then we can calculate
yH − u H yL − u L
e−aT1 = , e−aT2 = ,
yL − u H yH − u L
so that
 
yL − u H
aT1 = log , (3.29)
yH − u H
 
yH − u L
aT2 = log . (3.30)
yL − u L
To find the rate constant a, we substitute into (3.30):
 
30 mV − (−100 mV) 130
a(400 msec) = log = log ,
−70 mV − (−100 mV) 30
and find that a = 0.003666 msec−1 ≈ 0.004 msec−1 , or 4 sec−1 . To find uH , we
can then substitute into (3.29) and solve, to find uH = +42.5 mV ≈ +40 mV.
While this model uses an external control variable rather than incorporat-
ing details of the ion channel biochemistry to generate periodic behavior, the
exercise above does illustrate the utility of even simple models, as well as the
procedure of separation by variables. Before considering further applications,
we shall next give a general description of this procedure.

3.4.2 General procedure

Suppose that we wish to solve the differential equation
y ′ = p(t)q(y) (3.31)
where p is continuous on some interval a < t < b and q is continuous on
some interval c < y < d. We begin by separating variables as done for equa-
tion (3.22), dividing by q(y) and integrating to give
dy
Z Z
= p(t)dt + c (3.32)
q(y)
136 Dynamical Systems for Biological Modeling: An Introduction

where c is a constant of integration. Evaluating the integrals on each side of

this equation requires knowing the integrands’ antiderivatives. Let us name
these antiderivatives
dy
Z Z
Q(y) = , P (t) = p(t)dt,
q(y)
by which we mean that Q(y) is a function of y whose derivative with respect
1
to y is q(y) and P (t) is a function of t whose derivative with respect to t is
p(t). Then (3.32) becomes
Q(y) = P (t) + c, (3.33)
and this equation describes a family of implicit solutions y of the differential
equation (3.31). (The solutions can be made explicit if Q is invertible: y(t) =
Q−1 (P (t) + c).) To verify that each function y(t) defined by (3.33) is indeed
a solution of the differential equation (3.31), we differentiate (3.33) implicitly
with respect to t, obtaining dQ dy dP
dy dt = dt , that is,

1 dy
= p(t),
q(y(t)) dt
on any interval on which q(y(t)) 6= 0, and this shows that y(t) satisfies the
differential equation (3.31). There may be constant solutions of (3.31) which
are not included in the family (3.33). A constant solution to the original
equation (3.31) corresponds to a solution of the equation q(y) = 0.
If we wish to solve an initial value problem and find the particular solution
of equation (3.31) which satisfies the initial condition y(t0 ) = y0 , then instead
we should make the indefinite integrals in equation (3.32) definite, integrating
forward from the initial condition (t0 , y0 ):
Z y Z t
du
= p(s) ds. (3.34)
y0 q(u) t0

(We use u and s as integration variables in place of y and t so as not to confuse

them with the bounds of integration.) Evaluation of this integral defines y
implicitly in terms of t.
Alternatively, to solve the initial value problem we can take the general
solution (3.33) and (since the constant of integration c is arbitrary) specify
1
Q(y) to be the indefinite integral of q(y) such that Q(y0 ) = 0, and P (t) to
be the indefinite integral of p(t) such that P (t0 ) = 0. Then, because of the
fundamental theorem of calculus, we have
Z y Z t
du
Q(y) = , P (t) = p(s) ds,
y0 q(u) t0

and we may write the solution (3.33) in the form

Z y Z t
du
= p(s) ds + c.
y0 q(u) t0
 First-Order Differential Equations (Continuous Dynamical Systems) 137

Now substitution of the initial conditions t = t0 , y = y0 gives c = 0, and the

solution of the initial value problem is again given implicitly by (3.34).
We have now solved the initial value problem in the case q(y0 ) 6= 0. Since
y(t0 ) = y0 and the function q is continuous at y0 , q(y(t)) is continuous, and
therefore q(y(t)) 6= 0 on some interval containing t0 (possibly smaller than the
original interval a < t < b). On this interval, y(t) is the unique solution of the
initial value problem.
If q(y0 ) = 0, the situation is more complicated. We now have the constant
function y = y0 as a solution of the initial value problem. However, the condi-
tion q(y0 ) = 0 also makes the integral on the left-hand side of equation (3.34)
improper. If it diverges, then equation (3.34) is meaningless. If it converges,
then (3.34) gives another solution in addition to the constant solution y = y0 .
It is possible to show that if q(y) is differentiable, so that the existence and
uniqueness theorem stated in Section 3.3 applies, then this improper integral
must diverge, and the initial value problem has only the constant solution
(which exists regardless of the convergence of the integral). However, if q(y)
is not differentiable at y0 , there may be more than one solution: a constant
solution and another solution found by separation of variables (see Exercise 29
below).
We shall see in Section 4.1 that the constant solutions of a separable dif-
ferentiable equation (3.31) play an important role in describing the behavior
of all solutions as t → ∞.

Example 1.
Find all solutions of the differential equation y ′ = −y 2 .
Solution: We divide the equation by y 2 , permissible if y 6= 0, to give
1 dy
= −1.
y 2 dt
We then integrate both sides of this equation with respect to t. In order to
integrate the left side of the equation we must make the substitution y = y(t),
where y(t) is the (as yet unknown) solution. This substitution gives
1 dy dy 1
Z Z
2
dt = 2
=− +c
y dt y y
and we obtain
1
− = −t − c (3.35)
y
1
or (since Q(y) = −1/y is easily inverted here) y = t+c , with c a constant of
integration. Observe that no matter what value of c is chosen this solution
is never equal to zero. We began by dividing the equation by y 2 , which was
legitimate provided y 6= 0. If y = 0, we cannot divide, but the constant
function y = 0 is a solution. We now have all the solutions of the differential
equation, namely the family (3.35) together with the zero solution. 
138 Dynamical Systems for Biological Modeling: An Introduction

In practice, it may be more convenient to use the more colloquial form of

separation of variables mentioned in analysis of the pacemaker model, rather
than the formal procedure of Example 1. We can write the differential equation
y ′ = −y 2 in the form dy 2
dt = −y and separate variables by dividing the equation
2
through by y and multiplying by dt to give
dy
− = dt.
y2
Although, as noted earlier, this form of the equation is meaningless without
an interpretation of differentials, the integrated form
dy
Z Z
− = dt
y2
is meaningful. Carrying out this integration yields equation (3.35) as before.

Example 2.
Solve the initial value problem

y ′ = −y 2 , y(0) = 1.

1
Solution: The results of Example 1 give the family of solutions y = t+c for
the differential equation y ′ = −y 2 . When we substitute t = 0 and y = 1, we
1
get 1 = 0+c = 1c , which implies that c = 1. Thus the solution to the initial
1
value problem is y = t+1 , defined for all t ≥ 0.
Alternatively, separating variables and integrating as before yields
dy
Z Z
= −1 dt + c,
y2
or, in its definite form using the initial condition (t, y) = (0, 1),
Z y Z t
du
2
= −1 ds.
1 u 0

Evaluating both integrals yields

1 y t
− = −s ,
u 1 0

or − y1 − (− 11 ) = (−t) − (−0), which simplifies to y = 1

t+1 . 

Example 3.
Solve the initial value problem

y ′ = −y 2 , y(0) = 0.

Solution: The procedure used in Example 1 leads to the family of solutions

 First-Order Differential Equations (Continuous Dynamical Systems) 139
1
y = t+c for the differential equation y ′ = −y 2 . When we substitute t = 0,
y = 0 and attempt to solve for the constant c, we find that there is no solution.
When we divided the differential equation by y 2 we had to assume y 6= 0, but
the constant function y ≡ 0 is also a solution of the differential equation, as
may easily be verified. Since this function also satisfies the initial condition,
it is the solution of the given initial value problem. 

Example 4.
Solve the initial value problem

y ′ = ty 3 , y(0) = 1.

Solution: Division by y 3 (assuming y 6= 0) and integration with respect to t

gives
dy
Z Z
= t dt.
y3
Carrying out the integration, we obtain

1 t2
− 2
= + c.
2y 2

The initial condition y(0) = 1 gives c = − 12 , which gives the solution implicitly
by
1
− 2 = t2 − 1.
y
In this example it is easy to find the solution explicitly. We have y 2 = (1 −
t2 )−1 . Since y(0) = 1, we must take the positive square root and thus we
obtain the solution y = (1 − t2 )−1/2 defined on the interval −1 < t < 1. 

3.4.3 Solution of logistic equations

We now return to the logistic differential equation
 y
y ′ = ry 1 − (3.36)
K
whose solutions were described in Section 3.2. Recall that this equation de-
scribes the growth of a population y whose natural per capita birth rate r
is reduced by resource limitations in an environment capable of supporting
a population of K members. In Section 3.2, we verified that the solutions
were given by equation (3.12) but did not show how to obtain these solutions.
The differential equation (3.36) is separable, and separation of variables and
integration gives
K
Z Z
dy = r dt
y(K − y)
140 Dynamical Systems for Biological Modeling: An Introduction

provided y 6= 0, y 6= K. In order to evaluate the integral on the left-hand side,

we use the algebraic relation (which may be obtained by partial fractions)
K 1 1
= +
y(K − y) y K −y
and then rewrite the left-hand side as
K dy dy dy y
Z Z Z
= + = log |y| − log |K − y| = log .
y(K − y) y K −y K −y
The right-hand side simplifies to rt+ c, so we can now exponentiate both sides
of the equation to obtain
y
= ert+c = ert ec .
K −y
If we remove the absolute value bars from the left-hand side, we can define a
new constant C on the right-hand side, equal to e−c if 0 < y < K, or to −e−c
otherwise, thus rewriting the right-hand side as C1 ert . Finally, we solve for y,
and obtain the family of solutions
K
y= (3.37)
1 + Ce−rt
as in Section 3.2. This family contains all solutions of (3.36) except for the
constant solution y = 0. A qualitative observation is that if r > 0 every
positive solution approaches the limit K as t → ∞. (The only other non-
negative solution is the constant solution y ≡ 0.) The derivation of the family
of solutions is the same if r < 0, and in this case every solution which begins
less than K approaches the limit 0 as t → ∞. (For r < 0 all solutions which
begin above the constant solution y ≡ K increase without bound.)

Example 5.
In Example 1 of Section 3.2 we saw that the population of the United States
can be described reasonably well by a logistic growth curve with intrinsic
(maximum) growth rate r = −0.03/yr and carrying capacity K = 265 (million
people). However, in the year 2000 the U.S. population was estimated at 281.4
million. Adapt the model to account for the observed population exceeding
the apparent carrying capacity.
Solution: One possible explanation is an increased carrying capacity due to
improved agricultural techniques (or the converting of forest land into farm-
land). In practice, such changes are continuous and ongoing, but making K a
continuously varying function of time keeps the model from being separable.
A simple way to incorporate an increase in K is to choose a single year in
which to consider the carrying capacity to have risen. Since, from the data
given in Example 1 of Section 3.2, 1970 was the first since 1910 in which the
real population surpassed the model prediction, let us choose 1970 as our new
 First-Order Differential Equations (Continuous Dynamical Systems) 141

starting point. This will give us an initial condition of y0 = 202. To determine

the increased carrying capacity, we may use estimates that have been made
of 470 million acres of arable cultivated land in the U.S. and the average need
per person of 1.2 acres to calculate K = 470/1.2 ≈ 392 (million people). We
substitute into the solution (3.37) to obtain
392
y(t) = ;
1 + Ce−0.03t
if we let the year 1970 correspond to t = 0, we get 202 = 392/(1 + C), which
gives a value of C = 0.94. Using these values, we predict a population in the
year 2000 of
392
y(30) = = 283.6 (million people),
1 + 0.94e−0.03(30)
which is within less than 1% of the actual value. 

3.4.4 Discrete-time metered population models

In Section 2.4, we described some models for fish species which have an
annual reproductive cycle. The number of new adults recruited each year is the
number of fish born the preceding year (a function of the adult population size
the preceding year) multiplied by the fraction of newborns who survive until
the next year. Thus the model formulated is discrete, but has a recruitment
function which is the result of a continuous process. We are now ready to derive
these recruitment functions by solving the separable differential equations that
describe the survivorship of newborn fish.
Species of fish which suffer cannibalism in the egg, larval and/or early
juvenile stages have a per capita mortality rate which is proportional to the
size of the adult population (which serves as predators). We shall also assume
that the initial number of newborn fish (larvae, or eggs that hatch) is directly
proportional to the size of the adult population. If we denote the adult pop-
ulation at the beginning (t = 0) of the kth year by yk , and the number of
surviving newborn/juvenile fish at time t in the kth year by u(t), then we can
write these two assumptions mathematically as
u′ (t) = −(b̂yk ) u(t), u(0) = ryk , (3.38)
where b̂ and r are constants which measure, respectively, the per-adult con-
sumption rate of young and the average number of young produced per adult.
This system can be solved using separation of variables (in fact, it is linear in
u):
du
Z Z
= − (b̂yk ) dt
u
log u(t) − log u(0) = −(b̂yk )t
u(t) = u(0)e−(b̂yk )t .
142 Dynamical Systems for Biological Modeling: An Introduction

Substituting the initial condition and a final time value of T corresponding to

one year, we have
u(T ) = ryk e−(b̂yk )T .
Defining a new constant b = b̂T , the average number of young consumed
per adult in one year, and renaming r as a, we have, finally the recruitment
function
g(yk ) = ayk e−byk (3.39)
which leads to the Ricker model obtained in Section 2.4. (Beverton and Holt
also showed that if instead we assume that mortality depends upon size of the
newborn fish, and that the time required to grow to the critical size beyond
which predation drops off is proportional to the adult population, we also get
a Ricker model. See Exercise 21 below for details.)
The other common stock-recruitment model assumes instead that newborn
mortality depends not upon the density of the adult population, but upon the
density of newborns themselves, due to such factors as competition for food
among the juveniles (often referred to as “nursery competition”). In this case,
one assumes a per capita death rate of newborn fish which is a linear function
of the current newborn population size. This leads to a more complicated but
still separable differential equation with the same initial condition as for the
Ricker model:
u′ = −(µ0 + µ1 u)u, u(0) = ryk .
Separating variables and integrating, we have
du
Z Z
= − dt.
u(µ0 + µ1 u)
To continue, we need the result, obtained by the method of partial fractions,
that !
1 1 1 1
= − .
u(µ0 + µ1 u) µ0 u u + µµ10
Substituting this expression and making the integrals definite, we have
Z u(T ) ! Z T
1 1 1
− dv = − dt,
u(0) µ0 v v + µµ10 0
" ! !#
1 u(T ) u(0)
log − log = −T.
µ0 u(T ) + µµ10 u(0) + µµ01

Solving for u(T ), we obtain (after some work)

e−µ0 T u(0)
u(T ) = µ1 .
µ0 (1 − e−µ0 T )u(0) + 1

We now substitute the initial condition u(0) = ryk and define new constants
 First-Order Differential Equations (Continuous Dynamical Systems) 143

a = re−µ0 T and b = r µµ10 (1 − e−µ0 T ), so that u(T ) gives the Beverton-Holt

recruitment function
ayk
g(yk ) = (3.40)
1 + byk
obtained in Section 2.4.
In practice, models are sometimes used to determine basic demographic
rates by fitting observed data to the given functions, and using estimates for
model parameters (a and b in the models above) to calculate the underlying
rates (such as the mortality parameters in the models above). For example,
Schmidt et al.18 studied the coral reef damselfish Dascyllus trimaculatus, the
three-spot dascyllus, in lagoons of Moorea, French Polynesia, and found its
stock-recruitment data fit the Beverton-Holt function fairly well (see Exer-
cise 22 below for a caveat), with estimated parameter values of a = 0.696 and
b = 0.0711 per (sub-adult per 0.1m2 anemone). They measured survivorship
over 6 months rather than a year, so we take t = 1/2 yr. If we assume an
average number of fertilized eggs per adult of r = 10, 000 (observed values
range from a few thousand to 50,000 or more), then we can back-calculate the
corresponding mortality rates as follows:
log(a/r) log(0.696/10, 000)
µ0 = − =− = 19.1 yr−1 ,
T 1/2
bµ0
µ1 = = 0.000136 yr−1 pop−1
r(1 − e−µ0 T )
where pop is the population unit given above.
As another example, we consider the study by Dumas and Prouzet19
of Atlantic salmon in the Nivelle River in southwestern France. They fit
their data to a Ricker model and estimated values of a = 8.13510−2 and
b = 3.8410−6 per adult for survivorship of parr (salmon which have not yet
left freshwater) over a season (let us again assume T = 1/2 yr). Here the
estimation is easier: the reproductive ratio r = a and the mortality rate is
b̂ = b/T = 7.6810−6 per adult-yr.
Finally, we observe that some studies adapt these common stock-
recruitment models to incorporate additional variables which they find im-
portant in explaining variations in survivorship. Jung and Houde20 studied
the stock-recruitment relationship in bay anchovies in Chesapeake Bay and
found the data fit best to a modified Ricker model
g(yk ) = ayk e−(b1 yk +b2 ∆L) ,
18 Russell J. Schmidt, Sally J. Holbrook, and Craig W. Osenberg, Quantifying the effects

of multiple processes on local abundance: a cohort approach for open populations, Ecology
Letters 2: 294–303, 1999.
19 J. Dumas and P. Prouzet, Variability of demographic parameters and population dy-

namics of Atlantic salmon Salmo salar L. in a south-west French river, ICES Journal of
Marine Science 60(2): 356–370, April 2003.
20 Sukgeun Jung and Edward D. Houde, Recruitment and spawning-stock biomass dis-

tribution of bay anchovy (Anchoa mitchilli) in Chesapeake Bay, Fishery Bulletin 102(1):
63–77, 2004.
144 Dynamical Systems for Biological Modeling: An Introduction

where ∆L is the mean latitude difference (weighted by biomass) from the

mouth of the bay (37◦ N), a variable which accounted for differences in the
amount of dissolved oxygen in the waters of the bay, a factor in juvenile
anchovy development. This model corresponds to a per capita mortality rate
of the form −(b̂1 yk + b̂2 ∆L), where b̂2 describes the differential mortality rate
per degree of latitude.

3.4.5 Allometry
Let x(t) and y(t) be the sizes of two different organs or parts of an indi-
vidual at time t. There is considerable empirical evidence to suggest that the
dy
relative growth rates dx
dt /x and dt /y are proportional. This means that there
is a constant k, depending on the nature of the organs, such that
1 dy 1 dx
=k . (3.41)
y dt x dt
The relation (3.41) is called the allometric law, and the identification of such
differential growth ratios is called allometry (the scientist Julian Huxley re-
ferred to it as heterogony). The single equation (3.41) does not provide enough
information to determine x and y as functions of t, but we can eliminate t and
obtain a relation between x and y. If we consider y as a function of x, then
according to the chain rule of calculus,
dy dy . dx
= . (3.42)
dx dt dt
Combining (3.41) and (3.42) we have
dy y
=k ,
dx x
R R

which can be solved by separation of variables dy/y = k dx/x to give

log y = k log x + a, (3.43)

and finally
y = ek log x ea = xk ea = cxk . (3.44)
In order to determine the constant c, we need the values of x and y at some
starting time.
In experiments, one might measure both x and y at various times and
then use these measurements to plot log y against log x, that is, to plot the
experimental data on logarithmically scaled graph paper. According to the
relation (3.43), the graph should be a straight line with slope k and y-intercept
a = log c. Because of experimental error, the points may not line up perfectly,
but it should be possible to draw a line fitting the data well. One can then
measure the slope and y-intercept of the line to determine the constants k and
c in the relation (3.44).
 First-Order Differential Equations (Continuous Dynamical Systems) 145

Example 6.
Thompson’s classic work On Growth and Form21 reports the following data
from Huxley for the weights of the claw and body of the fiddler crab Uca
pugnax:

Mass of claw (g) 5 9 14 25 38 53 59 78 105 135 165 196

Mass of body 58 80 109 156 200 238 270 300 355 420 470 536
without claw (g)

Use these data to determine the constants c and k in the allometric law
y = cxk describing the relation between claw mass x and body mass y.
Solution: Transforming the data, we have:

log x 1.61 2.20 2.64 3.22 3.64 3.97 4.08 4.36 4.65 4.91 5.11 5.28
log y 4.06 4.38 4.69 5.05 5.30 5.47 5.60 5.70 5.87 6.04 6.15 6.28

Linear least-squares regression for log y = k log x + a gives k = 0.602,

a = 3.09, and c = ea = 22.0. Thus we have the relation y = 22x0.602 which
allows prediction of body mass given claw mass, or vice versa. Figures 3.15
and 3.16 show the fit of this curve to Thompson’s data. 
7
500 y
6

5 400

4 300

3
200
2
100
1
x
1 2 3 4 5 6 50 100 150 200

FIGURE 3.15: Log-log plot of Hux- FIGURE 3.16: Linear plot of Hux-
ley’s allometric data, and the line of ley’s allometric data, and the corre-
best fit. sponding curve.

A caveat: Thompson warned that allometric laws hold only during periods
of rapid growth, and claimed that simpler relationships are likely to hold at
other times (see Exercise 24 for Thompson’s discussion of this with regard
to the fiddler crab). It is certainly true that all models have limitations; this
one is similar to those we observed in earlier sections for linear and logistic
models.
21 D’Arcy Wentworth Thompson, On growth and form, Vol. I, 2nd ed., Cambridge: Cam-

bridge University Press, 1942, pp. 205–212.

146 Dynamical Systems for Biological Modeling: An Introduction

Exercises
In Exercises 1–8 below, find all solutions of the given differential equation. Do
not overlook constant solutions.

1. y ′ = t2 y 5. y ′ = −2ty

2. y ′ = 7y 3 cos t y2
6. y ′ = 1−t
t
3. y ′ = y 7. y ′ = t2 y − 4t2
y+1
4. y ′ = ty 2 8. y ′ = t

In Exercises 9–16 below, find the solution of the given differential equation
which satisfies the given initial condition.
9. y ′ = t2 y, y(5) = 1 13. y ′ = ty 2 , y(0) = −2
2
10. y ′ = 7y 3 cos t, y(1) = 0 y
14. y ′ = 1−t , y(−1) = log1 2

11. y ′ = −2ty, y(0) = y0 15. y ′ = t2 y − 4t2 , y(0) = 0

y+1
12. y ′ = yt , y(0) = 1 16. y ′ = t , y(4) = 1
4

17. We can rewrite the pacemaker model (3.22) to emphasize how the pace-
maker signal y follows the control u: y ′ = −a(y−u). Since we are suppos-
ing the control signal u to be piecewise constant, then (y − u)′ = y ′ , and
we can make a substitution of variables, as is sometimes done in evalu-
ating integrals in calculus courses, to rewrite this equation in terms of
the difference z = y − u: (y − u)′ = −a(y − u) becomes

z ′ = −az. (3.45)

(a) Interpret equation (3.45) biologically.

(b) Solve the differential equation z ′ = −az for z(t), and then back-
substitute to return the solution to terms of y and u. Verify ex-
plicitly that the solution obtained in this way is equivalent to that
obtained in this section.

18. In this problem we consider the effects of irregularities in the threshold

voltages for the pacemaker model (3.22) studied in this section.
(a) Suppose that for some reason the control signal becomes less sen-
sitive to the pacemaker voltage, in such a way that yH begins to
rise toward uH . For example, with uH = +40mV, suppose yH rises
from +30 mV to +35 mV. What important effect does this have
on y(t)?
 First-Order Differential Equations (Continuous Dynamical Systems) 147

(b) Given the nature of the solution y(t), what will happen as yH ap-
proaches uH ? (Hint: Consider equation (3.29).)
(c) If the control signal becomes sufficiently desensitized, it may be-
come necessary to introduce an external pacemaker which drives
the control based explicitly on timing rather than on values of y.
That is, if the neuron does not fire within a certain amount of
time after last resetting, the artificial pacemaker forces the signal
to switch. This has the effect of controlling T1 and T2 directly.
If an external pacemaker of this type is applied to a heart with the
parameter values given in this section, and the maximum period
T1 allowed by the external pacemaker before causing the control
to switch off is 800 msec, then how high can the threshold yH rise
before the external pacemaker activates?

19. In the pacemaker model of this section, what happens to the firing times
T1 and T2 when uH − yH = yL − uL ?
20. In the logistic model, suppose that we have a population whose repro-
ductive rate undergoes seasonal fluctuations—many species of animal
reproduce only at certain times of year. This has the effect of making
the growth rate r a function of time. If we let t = 0 correspond to Jan-
uary and suppose that most reproduction occurs in June or July, then
we might model the growth rate as

r(t) = r0 t + r1 sin 2πt,

with r0 ≥ r1 > 0. Substitute this expression into the logistic equation

(3.36) and apply the method of separation of variables to solve it. What
does this model predict about the long-term behavior of this population?
(You may want to consider a graph.)
21. Beverton and Holt showed22 another way to derive the Ricker recruit-
ment function. If we assume that newborn fish suffer higher mortality
while very small, and that the time tc required to grow to the critical
size below which mortality drops is directly proportional to the adult
population (since a greater adult population means more food resources
are unavailable for the growing young), then we can model newborn sur-
vivorship with a per capita mortality rate of µ1 for 0 ≤ t < tc and µ2
for tc ≤ t ≤ T (where T represents one year).
(a) Write the linear differential equation describing newborn survivor-
ship during the period 0 ≤ t < tc . Solve the equation, using the
initial condition u(0) = ryk and following the method in the text.
Find the value of u(tc ).
22 Raymond J.H. Beverton and Sidney J. Holt, On the dynamics of exploited fish popula-

tions. Fishery Investigations Series 2 (19). London: Great Britain Ministry of Agriculture.
148 Dynamical Systems for Biological Modeling: An Introduction

(b) Write the linear differential equation describing newborn survivor-

ship during the period tc ≤ t ≤ T . Solve the equation, using the
initial condition for u(tc ) given in part (a) (and tc as a lower bound
for integration). Find the value of u(T ).
(c) Define new constants a and b so that a = re−µ2 T and (µ1 − µ2 )tc =
byk , where the latter equation reflects the assumption that tc is
directly proportional to yk . Show that u(T ) gives the Ricker re-
cruitment function g(yk ) as in (3.39).
22. In their study of the damselfish, Schmidt et al. (1999) initially used the
Maynard Smith and Slotkin model (equation 1.27, p. 27), a generaliza-
tion of the Beverton-Holt model, and estimated the exponent n which
measures the nature of the intraspecies competition at 1.14 (n = 1 cor-
responds to strictly contest competition and the Beverton-Holt model).
Using the parameter estimates a = 0.696, b = 0.0711 per (sub-adult
per 0.1m2 anemone) and r = 10, 000 eggs per adult given in the text,
determine the relative (percentage) error incurred in calculating the per
capita mortality parameters µ0 and µ1 .
23. Jung and Houde (2004) estimated parameter values for their modified
Ricker model of a = 365, b1 = 0.19 per metric ton, and b2 = 1.35
per degree. They measured survivorship over an approximately 6-month
span.
(a) Use these values to calculate the differential mortality rates b̂1 and
b̂2 .
(b) Solve the corresponding differential equation using the per capita
mortality rate given in the text. Substitute the parameter values
obtained in (a) to estimate survivorship over a full year.
24. Thompson’s response to the heterogony identified by Huxley and others
through what we called in this section the allometric law was to cau-
tion that such differential growth ratios hold only during unconstrained
growth. In particular, he claimed that the fiddler crab’s growth ratio
was linear rather than geometric past a certain point. Below are the rest
of Huxley’s data for the fiddler crab.

Mass of claw (g) 243 319 418 461 537 594

Mass of body 618 743 872 983 1080 1166
without claw (g)

Mass of claw (g) 617 670 699 773 1009 1380

Mass of body 1212 1299 1363 1449 1808 2233
without claw (g)
 First-Order Differential Equations (Continuous Dynamical Systems) 149

(a) Determine the constants in the allometric law that best fit these
additional data.
(b) Now try instead a simple linear fit, of the form y = mx + b.
(c) Which of the two models above (allometric or linear) appears to be
a better fit to these data?
(d) Repeat the analysis in parts (a)–(c) above, using the entire data set
(24 points). Which model is the better fit for the entire data set?
How would a model defined piecewise (allometric for part of the
data set, linear for the rest) compare to the two one-piece models?

25. As in Exercise 24, determine, evaluate and compare the allometric and
linear models for the following data on the stag beetle Lucanus cervus
from Huxley, republished by Thompson (p. 208).

Length of mandible (mm) 6.0 7.8 9.0 10.0 11.2 11.9 12.8 14.4
Total length (mm) 31.0 38.8 40.5 42.6 45.0 46.9 49.2 53.6

26. As in Exercise 24, determine, evaluate and compare the allometric and
linear models for the following data on the reindeer beetle Cyclommatus
tarandus from Huxley, republished by Thompson (p. 209).

Length of mandible (mm) 3.9 10.7 14.1 19.9 24.0 30.7 34.5
Total length (mm) 20.4 33.1 38.4 47.3 54.2 66.1 74.0

27. Suppose a population satisfies a logistic model with r = 0.4, K = 100,

y(0) = 5. Find the population size when t = 10.
28. Suppose a population satisfies a logistic model with r = 0.3, K = 50,
y(0) = 2. Find the population size when t = 25.
29. Find the limit as t → ∞ of the solution of each of the following initial
value problems:
(a) y ′ = −y + 1, y(0) = 0
(b) y ′ = −y + 1, y(0) = 100
(c) y ′ = y 2 , y(0) = −1
(d) y ′ = −y 2 , y(0) = 1.
30∗ . Find two solutions of the initial value problem y ′ = 3y 2/3 , y(0) = 0.
Rt
31∗ . Find all differentiable functions f (t) such that [f (t)]2 = 0 f (s) ds for
all t ≥ 0.
32∗ . Find all continuous (not necessarily differentiable) functions f (t) such
Rt
that [f (t)]2 = 0 f (s) ds for all t ≥ 0.
150 Dynamical Systems for Biological Modeling: An Introduction

3.5 Mixing processes, and first-order linear models with

constant coefficients
The autonomous first-order linear differential equation
dy
= −ay + b, (3.46)
dt
where a and b are given constants, serves as the basis for many models of
biological systems in two senses: first, it can serve as a simplified model for
many processes that occur in biological systems; and, second, the solution of
linear equations is the foundation for the qualitative analysis of nonlinear sys-
tems, as will be seen in Section 4.1.In this section we shall consider (3.46) as
a model for mixing processes in which chemicals, cells, organisms, and even
heat distribute themselves evenly within a system. The overall quantity y of a
given substance (or population) present within a biological system changes as
the substance enters the system at a rate b (here taken to be constant; it is,
at least, independent of y), and leaves the system at a rate ay proportional to
the quantity of the substance already present.23 Here we assume that the con-
centration of the substance within the system is relatively uniform, or at least
that the concentration leaving the system is close to the average concentration
throughout the system. See Figure 3.17 for an illustration. This assumption
of a uniform, or homogeneous, distribution within the given system has given
the name “mixing processes” to such phenomena. The assumption is often
fairly accurate in describing the mixing of simple, nonreactive chemicals in a
vat, but may be a gross oversimplification in describing the movement of more
complex populations. It is nevertheless useful as a first step in developing and
understanding more complex models.

SYSTEM
inflow outflow
uniform mixing
-  -
........
rate b  rate ay
population y(t)

FIGURE 3.17: A schematic representation of a simple mixing process.

Before we consider some common applications of this model to biological

systems, we will first show how to solve equation (3.46) and related initial value
problems. There are several techniques for solving first-order linear differential
23 Here we assume that a > 0, although all our calculations will be equally valid for a ≤ 0.
 First-Order Differential Equations (Continuous Dynamical Systems) 151

equations24 ; here we will introduce one which will also be of use when we
consider the more general first-order equation in Section 3.6, in which the
coefficients a and b may be functions of the independent variable t (that is,
they may change over time).
The technique of an integrating factor involves using the product rule for
derivatives in reverse, much like integration by parts. We begin by writing
(3.46) as
dy
+ ay = b
dt
and then look for a way to rewrite the left-hand side as an exact deriva-
tive: that is, the two-term derivative of a product of functions. If we multiply
through by a function g(t), then the left-hand side becomes gy ′ + agy. The
derivative of the product (gy) is gy ′ + g ′ y, Rso we have a match if g ′ = ag. Our
integrating factor g should therefore be e a dt (making g ′ = ag), which for
constant a reduces to eat . Now multiplying the original equation by g yields
dy
eat + aeat y = beat ,
dt
which by the above line of reasoning we can rewrite as
d at

e y = beat . (3.47)
dt
From this point we can apply either an indefinite or definite integral. In the
first case, we get Z
eat y = beat dt;

b at
for a and b constant, the right-hand side simplifies to ae plus an arbitrary
constant of integration C, which gives us the solution
b
y= + Ce−at . (3.48)
a
If we have the initial condition y(0) = y0 , then the definite integral of (3.47)
becomes Z t Z t
d aτ
(e y(τ )) dτ = beaτ dτ
0 dτ 0

(using τ as the variable of integration), which simplifies to

b at
eat y(t) − e0 y(0) = e − e0 ;


a
that is,
b
y(t) = y0 e−at + 1 − e−at ,


(3.49)
a
24 Including separation of variables, cf. the cardiac pacemaker model of section 3.4.
152 Dynamical Systems for Biological Modeling: An Introduction

a time-weighted average of the initial value y0 and the final (asymptotic) value
b/a. (In terms of (3.48), then, C = y0 − ab .)
In the discussions of related models that follow, we shall take as given the
solution (3.48) to the differential equation (3.46), and the solution (3.49) to the
corresponding initial value problem. While we wish to point out the breadth
and variety of examples, our main goal is to encourage some understanding
of the modeling process in areas of interest to the reader.

3.5.1 Chemostats
The chemostat is a laboratory device proposed in 1950 by Novick and
Szilard25 and Monod26 to reproduce classic continuous linear diffusion in cul-
tures of bacteria, and has since become a regular fixture in many biology
laboratories. It consists of a large container, the chemostat reactor, to which
are connected inflow and outflow tubes (connected, respectively, to a nutrient
supply tank and an outflow collector), as well as other instruments which help
maintain constant conditions inside the reactor. The underlying assumption
governing chemostat populations is that growth is limited by one key nutri-
ent (the rest being present in abundant supply). The chemostat is a physical
model of a natural biological system in much the same way as a differential
equation such as (3.46) is a mathematical model of a biological system. In ad-
dition to its use as a constant supply of bacterial cultures in the lab, it is used
to approximate and study the workings of systems from populations of cells
to entire ecosystems, especially where field and in vivo studies are impossible.
Figure 3.18 shows a typical setup.
Many chemostat models are complex, incorporating multiple populations
or substances within the reactor (for example, nutrient supply and the or-
ganisms that consume them), multiple spatial layers within the reactor, etc.,
but here we shall consider the simplest possible model, in which we track a
single substance or population within a system. For example, Carman and
Woodburn27 used a chemostat to study the impact of low levels of antibiotics
used in food-producing animals on human intestinal flora: in particular, the
extent to which low levels of the common antibiotic ciprofloxacin may induce
resistance in intestinal bacteria. Such drugs are commonly used in food ani-
mals, but residues may remain in meat ingested by consumers in levels low
enough to produce resistance (to the same or related drugs) in gut bacteria.
Carman and Woodburn used a chemostat containing 500 mL of culture, with
a nutrient inflow of 35 mL/hr. If we assume a constant balancing outflow of
25 Aaron Novick and Leo Szilard, Description of the chemostat, Science 112: 715–716,

1950.
26 Jacques Monod, La technique de culture continue: théorie et applications, Ann. Inst.

Pasteur 79: 390–410, 1950.

27 Robert J. Carman and Mary Alice Woodburn, Effects of low levels of ciprofloxacin on a

chemostat model of the human colonic microflora, Regulatory Toxicology and Pharmacology
33: 276–284, 2001.
 First-Order Differential Equations (Continuous Dynamical Systems) 153

Photo courtesy Thomas Chrzanowski Photo courtesy Alan M. Hughes

FIGURE 3.18: Left: A chemostat, perhaps the most classical physical model
of a linear mixing process. In the two chemostats pictured, the containers
in front are the chemostat reactors, the large bottles above are the nutrient
reservoirs, and the metal boxes circulate water around the reactors to maintain
temperature. The outflow collectors are not in view. Right: Loch Linnhe in
western Scotland, the fjordic lake studied by Ross et al. (2001), in which the
nutrient dynamics behave like a chemostat.

35 mL/hr and an initial concentration of 0 g of ciprofloxacin in the chemo-

stat, how long will it take for the concentration of cipro in the chemostat to
reach 0.1 µg/mL for each of the inflow concentrations used by Carman and
Woodhouse: 0.43, 4.3, and 43 µg/mL? What will the concentration be in each
case after an hour?
In order to answer these questions, we let y(t) be the total amount of cipro
(in µg) in the chemostat t hours after the inflow is turned on. The inflow b is
then 0.43 µg/mL × 35 mL/hr = 15.05 µg/hr in the first case (multiply by 10
or 100 for the larger inflows), and the outflow a, whose units are per-time, is
35 mL/hr÷500 mL = 0.07/hr, i.e., 7% of the chemostat’s volume is exchanged
each hour. From (3.49) with y0 = 0, we then have y(t) = 215 1 − e−0.07t .

After one hour, the amount of cipro in the chemostat should be

y(1) = 215 1 − e−0.07 ≈ 14.5 µg,

making the concentration about 14.5 µg/500 mL = 0.029, µg/mL (multiply

this by 10 or 100 for the higher inflow rates). In order for the concentration
to reach 0.1 µg/mL, we need a total amount of 0.1 µg/mL × 500 mL = 50 µg,
so we set y(t) to 50 and solve for t:

50/215 = 1 − e−0.07t , so e−0.07t ≈ 0.767 and t ≈ − log(0.767)/0.07 ≈ 3.78 hr.

For the higher inflow rates, we use 2150 or 21500 in place of 215 and obtain
time values of 0.336 hr and 0.0333 hr, respectively (about 20 min and 2 min).
We may observe that the primary adjustment necessary in applying (3.49) here
is the conversion between concentration and total amount; this is a common
issue with simple nonbiological mixing problems as well.
154 Dynamical Systems for Biological Modeling: An Introduction

Jannasch wrote28 that in practice, the number of biological systems that

can be reasonably modeled by a chemostat is extremely limited, because of
the assumptions of time-independence and constancy of conditions that are
rarely true even in controlled continuous cultures. Nevertheless, the chemo-
stat model continues to be used, in mathematics and in laboratories, as a
starting point for understanding the dynamics of well-mixed systems with
known inflows and outflows. Perhaps the largest recent modeling application
of this metaphor involves an entire ecosystem: a fjordic lake, whose high lev-
els of freshwater runoff (inflow) and tidal exchange (inflow and outflow) give
it characteristics not unlike the chemostat. (Most lakes turn over so little of
their volumes per day that they are more usually modeled as closed systems.)
Ross et al.29 studied the food chain in Loch Linnhe, on the western coast of
Scotland (see Figure 3.18), which was found to have an average of about 40%
of its surface-layer volume exchanged daily via the tides, and about 3.5% of its
surface-layer volume per day through freshwater runoff. Their model not only
counted separately the carnivores, zooplankton, phytoplankton, and nutrients
in the lake, but also differentiated levels of the lake by depth (fjordic lakes are
typically elongated, deep for their width, and open to the ocean). The surface
layer (about 10 m deep), which includes just under a quarter of the lake’s
total volume, accounts for most of the daily exchange. Ross et al. found that
the inorganic nutrient dynamics did indeed resemble those of a chemostat,
but that the metaphor was less accurate for the more complex dynamics of
plankton and carnivores in the lake. The main difference between sea loch and
chemostat dynamics lay in the occasional irregularities in the former due to
sporadic mixing among the different layers in the lake, which can cause brief
net outward nutrient flows when nutrients settled in the bottom are stirred
into the levels of higher exchange nearer the surface.
If we interpret simplistically the exchange rates given above for Loch
Linnhe (and assume a constant total volume for the lake), the coefficient
a = −0.435/day accounts for outflow in the surface layer (compare with
0.07/hr or 1.68/day in the Carman and Woodburn chemostat). This coef-
ficient can be used in tracking any of various populations in the lake’s surface
layer, such as nutrients like carbon or nitrogen, or chemicals collected in runoff.
For instance, suppose a given chemical accumulates from runoff, entering the
loch at a rate of b = 1 kg/day. The limiting value this model predicts is then
b/a ≈ 2.3 kg of the substance in the loch. Although seasonal disturbances will
prevent the true accumulation from approaching this steady state monotoni-
cally, a quick calculation shows that within a week, there will be

y(7) = 2.3 1 − e−0.435×7 ≈ 2.2

28 Holger W. Jannasch, Steady state and the chemostat in ecology, Limnology and

Oceanography 19(4): 716–720, 1974.

29 Alex H. Ross, William S. C. Gurney, Michael R. Heath, Steven J. Hay, and Eric W.

Henderson, A strategic simulation model of a fjord ecosystem, Limnology and Oceanography

38(1): 128–153, 1993.
 First-Order Differential Equations (Continuous Dynamical Systems) 155

kg of it in the lake, which is within 5% of its steady-state value.

3.5.2 Drug dosage

When a dosage of a drug is administered to a patient, the concentration
of the drug in the blood increases. Over time, the concentration decreases as
the drug is eliminated from the body. The question we would like to model
is how the drug concentration changes with repeated doses of the drug. This
area of study is called pharmacokinetics.
Experimental evidence indicates that the rate of elimination is proportional
to the concentration of drug in the bloodstream. Thus we assume that, if C(t)
is a function representing the concentration of drug in the blood at time t, its
derivative is given by
C ′ (t) = −qC(t).
Here q is a positive constant, called the elimination constant of the drug.
We assume also that when a drug is administered, it is absorbed completely
immediately. This is certainly at best an approximation to the truth, probably
a better approximation for a drug injected directly into the bloodstream than
for a drug taken by mouth. Let us assume that the drug is administered
regularly at fixed time intervals of length T with a dose capable of raising the
concentration in the blood by A. Then if we begin with an initial dose A at
time t = 0, C(t) satisfies the initial value problem C ′ (t) = −qC(t), C(0) = A
until the second dose. Thus for 0 ≤ t ≤ T , we have C(t) = Ae−qt , and just
before the second dose at time T , C(T − ) = Ae−qT .
We let Ci be the concentration at the beginning of the ith interval and Ri
the residual concentration at the end of the ith interval; we have shown that
C1 = A, R1 = Ae−qT . After the second dose, at time T , the concentration
jumps from R1 to C2 = R1 + A = A(1 + e−qT ). Now we have a new initial
value problem, C ′ (t) = −qC(t), C(T ) = C2 , for T < t ≤ 2T , whose solution
we can show to be C(t) = C2 e−q(t−T ) . Just before the third dose, then, at
time 2T , we have R2 = C(2T − ) = C2 e−qT .
More generally, the residual concentration Ri at the end of each interval
between doses is the concentration Ci at the beginning of the interval multi-
plied by e−qT , and the level Ci+1 at the beginning of the next interval (just
after the next dose is given) is this residual concentration plus A. Thus for
the second interval we have
C2 = A(1 + e−qT ), R2 = A(1 + e−qT )e−qT = A(e−qT + e−2qT ).
Next we see that for the third interval
C3 = A+A(e−qT +e−2qT ) = A(1+e−qT +e−2qT ), R3 = A(e−qT +e−2qT +e−3qT ).
We may give a simpler expression for each of these by using the formula for
the sum of a finite geometric series,
1 − rn+1
1 + r2 + r3 + . . . + rn = ,
1−r
156 Dynamical Systems for Biological Modeling: An Introduction

to obtain
1 − e−3qT 1 − e−3qT
C3 = A −qT
, R3 = Ae−qT .
1−e 1 − e−qT
From this we may conjecture (and prove by induction) that for every positive
integer n,
1 − e−nqT −qT 1 − e
−nqT
Cn = A , Rn = Ae .
1 − e−qT 1 − e−qT
As n → ∞, e−nqT → 0 because e−qT < 1. Thus Cn and Rn approach limits
C and R, respectively, as n → ∞, where
A Ae−qT
C= −qT
, R= . (3.50)
1−e 1 − e−qT
It is easy to see that both the initial and residual concentration increase
with each dose but never exceed the limit values. For example, if a drug with an
elimination constant of 0.1 hours−1 is administered every 8 hours in a dosage
e−0.8
of 1 mg/L, the limiting residual concentration is 1−e −0.8 = 0.816 mg/L.

If the time interval T between doses is short, then the ratio of residual
concentration to dose,
R e−qT 1
= = qT ,
A 1 − e−qT e −1
is large (for small T , eqT − 1 is near 0), as each new dose builds upon the
previous one before it has had time to dissipate. On the other hand, if the
time interval between doses is long, each Rn is close to zero, and each Cn is
close to A; each individual dose dissipates almost completely before the next
dose is given.

Example 1.
Harms et al.30 studied the pharmacokinetics of oxytetracycline, an antibiotic
also used as a marker to label bone for age studies, in the loggerhead sea turtle
Caretta caretta (see Figure 3.19). They administered a single dose of the drug
and monitored its gradual elimination from the turtles’ blood plasma. They
reported results in terms of the elimination half-life, which is the amount of
time required to eliminate half of the drug from the blood. For a single dose
of 25 mg/kg administered intramuscularly to turtles with an average mass of
8 kg, the mean elimination half-life was 61.9 hours.
What is the minimum [intramuscular] dosage required in order to keep the
level of oxytetracycline in an 8 kg loggerhead’s blood plasma above 100 mg,
if doses are given once every day?31 Once every three days? Once a week?

30 Craig A. Harms, Mark G. Papich, M. Andrew Stamper, Patricia M. Ross, Mauricio

X. Rodriguez, and Aleta A. Holm, Pharmacokinetics of oxytetracycline in loggerhead sea

turtles (Caretta caretta) after single intravenous and intramuscular injections, Journal of
Zoo and Wildlife Medicine 35(4): 477-488, 2004.
31 Concentrations are typically reported in units of µg/mL; here we simplify by considering

the total amount.

 First-Order Differential Equations (Continuous Dynamical Systems) 157

Photo courtesy U.S. Fish and Wildlife Service, dls.fws.gov Image courtesy Benjamin Mills

FIGURE 3.19: Left: A loggerhead sea turtle fitted with a tracking device.
Right: The molecular structure of oxytetracycline, so named for the oxygen
(dark atoms along top and bottom) bonded to four (tetra) hydrocarbon rings.

140 concentration[mg] 700 concentration[mg]

120 600

100 500

80 400

60 300

40 200

20 100
time[days] time[days]
5 10 15 20 5 10 15 20

FIGURE 3.20: Left: Projected drug concentration with daily doses of 31 mg.
Right: Projected drug concentration with weekly doses of 556 mg.

1
Solution: The equation = e−qt1/2 for the elimination half-life t1/2 can be
2
rewritten for the elimination constant q:
 
1
q = − log /t1/2 .
2

A half-life of 61.9 hours therefore corresponds to a value q = 0.0112/hr. We are

interested in the eventual residual concentration R. From (3.50), for T = 24 hr
we have R = 3.24A, for T = 72 hr, R = 0.807A, and for T = 168 hr R =
0.180A. This means that daily doses A must be at least 100/3.24 ≈ 30.9 mg,
while doses given once every three days must be at least 124 mg, and weekly
doses must be at least 556 mg. (For an 8 kg turtle, these figures correspond
to concentrations of 3.86, 15.5, and 69.5 mg/kg.) Observe that when dose
frequency is reduced by a factor of 7 (from once per day to once per week),
the dosage size needed increases by a factor of 18 (see Figure 3.20). 
158 Dynamical Systems for Biological Modeling: An Introduction

3.5.3 Newton’s law of cooling

When a body at some temperature is placed in surroundings at a differ-
ent temperature, heat flows from the body to the surroundings if the body
is hotter than the surroundings, and from the surroundings to the body if
the surroundings are hotter than the body. Newton’s law of cooling, which
is based on balance equations for the transfer of heat, states that the rate at
which body temperature changes is proportional to the temperature difference
between the body and its surroundings. Let us assume that the surroundings
are at a temperature T ∗ , called the ambient temperature, and that the body
is small enough compared to the surroundings that the heat exchange has a
negligible effect on the temperature of the surroundings. We assume also that
all of the surroundings are at the same temperature; that is, that heat is cir-
culated throughout the surroundings. According to Newton’s law of cooling,
the temperature T of the body at time t satisfies a differential equation of the
form
T ′ = −k(T − T ∗ ) = −kT + kT ∗ .
Here, k is a constant of proportionality which must be positive in order to
make T ′ > 0 when T < T ∗ and T ′ < 0 when T > T ∗ . Then T satisfies an
initial value problem

T ′ = −kT + kT ∗ , T (0) = T0 , (3.51)

where T0 is the temperature of the body when it is placed into the surround-
ings. Since (3.51) is the same as (3.46) with a and b replaced by k and kT ∗
respectively, its solution (from (3.49)) is

T = T0 e−kt + T ∗ (1 − e−kt ), or T (t) = T ∗ + (T0 − T ∗ )e−kt . (3.52)

From this we see that the temperature of the body will approach the ambient
temperature as t → ∞. In practical applications, the constant of proportion-
ality k cannot be measured directly but must be calculated from temperature
readings at more than one time.

Example 2.
In an experiment published in 1937, James Hardy tried to determine a con-
stant of proportionality for heat loss from the human body.32 Immediately
there are many complicating factors which must be acknowledged, includ-
ing wind, body position (which affects the amount of surface area exposed),
and substances such as water or fabric (clothes) which affect the transfer of
heat from the skin to the surrounding air. The body’s own defenses, such as
the chill or shivering reflex, which attempts to create heat through expend-
ing [kinetic] energy, also interrupt the heat flow described by Newton. Hardy
therefore limited his primary investigations to the transfer of heat from dry
32 James D. Hardy, The physical laws of heat loss from the human body, PNAS 23: 631–

637, 1937.
 First-Order Differential Equations (Continuous Dynamical Systems) 159

skin directly to still air, in the limited range of 23◦ C to 28◦ C, with his human
subjects lying perfectly still and straight. Hardy also wrote the heat exchange
rate coefficient as being directly proportional to surface area, k = KA, where
K is a constant rate per unit area and A is the surface area across which heat
is being exchanged.
Hardy used two human subjects, as well as an elliptical iron cylinder for
comparison purposes (the cylinder had dimensions roughly those of an adult
human torso, and afforded more detailed measurements). He reported heat
exchange rate coefficients of 5.2 cal/◦ C hr m2 for the human subjects and
6.75 cal/◦ C hr m2 for the cylinder. Hardy gave this latter figure as being
equivalent to 1.64 cal/◦ C hr overall (from which we may infer an effective
area of about 0.24 m2 , though Hardy’s figure is twice this), and gave the
cylinder’s mass as 24.98 kg. Using the common mass and surface area estimates
of 68 kg and 1.8m2 for an adult human male, determine the two coefficients
k for Newton’s law of cooling with the human subjects and the cylinder, and
the amount of time required for each to cool from 28◦ C to 23◦ C in a 20◦ C
environment. The specific heats of iron and the human body are 829 cal/kg◦ C
and 112 cal/kg◦C, respectively.
Solution: Hardy wrote in terms of heat exchange; we wish to write in terms
of cooling, or temperature change, so we need to use the specific heat of each
object to convert Hardy’s data (in calories) into purely temperature-based
terms. For the iron cylinder, we have
cal cal
k = 1.64 ◦ ÷ 112 ◦ ÷ 24.98kg = 0.000586/hr.
C hr kg C
For the human subjects, we multiply 5.2 cal/◦ C hr m2 by the estimated surface
area of 1.8m2 and then proceed similarly:
cal cal
k = 5.2 ◦ × 1.8m2 ÷ 829 ◦ ÷ 68kg = 0.000166/hr.
C hr m2 kg C

Substituting into the second form of (3.52), we have 23 = 20 + (28 − 20)e−kt ,

or t = − log(3/8)/k, which gives us 1674 hours and 5909 hours, respectively,
for the iron cylinder and the human body. 

Example 3.
Newton’s law of cooling may also be applied forensically, to estimate the time
of death based on body temperature measurements after death. This appli-
cation is especially common at the beginning and end of hunting seasons, to
determine whether game animals such as deer were killed within the legal
season. Kienzler et al.33 conducted a study on the cooling rate for white-
tailed deer.34 They found a linear cooling rate constant of k = 0.087656/hr,
33 J.M. Kienzler, P.F. Dahm, W.A. Fuller, A.F. Ritter, Temperature-based estimation for

the time of death in white-tailed deer, Biometrics 40(3): 849–854, 1984.

34 Their model incorporated the effects of weight and ambient temperature as well as
160 Dynamical Systems for Biological Modeling: An Introduction

with a standard deviation of 0.004912/hr. The normal body temperature for

white-tailed deer is 38.8◦ C.
If a deer carcass found in 0◦ C weather has a thigh temperature of 15◦ C,
how long is it likely to have been dead? Give a mean estimate as well as a
95% confidence interval (2 standard deviations above and below the mean),
assuming variation in k only.
Solution: The second form of (3.52) gives us 15◦ C = 0◦ C + (38.8◦ C −
0◦ C)e−0.087656t , from which t = log(15/38.8)/(−0.087656)hr ≈ 10.8 hr. Using
the endpoints of the 95% CI [0.077832,0.09748] for k yields t values of 12.2
and 9.75 hours, respectively. 

Example 4.
We can also sometimes use data to generate predictions directly without cal-
culating k explicitly. If a loaf of freshly baked bread cools from 100◦ C to 60◦ C
in the first 10 minutes it is on a cooling rack, when the surroundings are at
20◦ C, what will the loaf’s temperature be 30 minutes after it was placed on
the cooling rack?
Solution: From the second form of (3.52) with T0 = 100◦ C, T ∗ = 20◦ C, we
have T (10) = 20 + (100 − 20)e−10k = 60, and T (30) = 20 + (100 − 20)e−30k .
From the first of these, e−10k = 40/80 = 1/2, so that e−30k = (e−10k )3 = 1/8.
Thus T (30) = 20 + 80(1/8) = 30◦ C. 

3.5.4 Migration
Another process which can be described in terms of mixing is the migra-
tion of populations, between or among groups in general, and into and out of
a single group in the context of one-equation models. From cells to organisms
and households, migration adds an important dimension to the dynamics of
populations that are not entirely closed. Cell migration is a necessary mecha-
nism for such processes as cell differentiation and the healing of wounds, and
the migration of tumor cells is metastasis, the defining characteristic of ma-
lignant cancers. Many animals such as insects and birds undertake periodic
seasonal migrations, and transnational immigration of humans has become a
major political issue in the early twenty-first century.
Although some behavioral and social scientists would distinguish between
seasonal or temporary movements and those movements of populations in-
tended to be permanent, we will include in our definition of migration any
movement which can be captured on the time scale of interest to us. If we
are dealing with migration over the course of years, seasonal, month-to-month
movements into and out of the zone under study will only manifest in our
model to the extent that they accumulate. On a time scale of hours, however,
every daily activity has the potential to affect the model. When a population

time since death; see Exercise 15. A linear regression of their data assuming proportional
dependence on time only would yield a different value for k than the one used here.
 First-Order Differential Equations (Continuous Dynamical Systems) 161

TABLE 3.3: United States population and per capita demographic rates for
selected years.
Year Population Births Deaths Immigrations Emigrations
(millions) (/yr) (/yr) (/yr) (/yr)
1950 152 0.0241 0.0096 0.000681 0.000185
1960 180 0.0237 0.0095 0.001397 0.000236
1970 204 0.0184 0.0095 0.001628 0.000441
1980 226 0.0159 0.0088 0.001988 0.000520
1990 250 0.0167 0.0086 0.002935 0.000640
2000 281.4 0.0144 0.0085 0.003232 0.000831

gains or loses a significant proportion of its size through connections with

other populations, it is important to incorporate them into a model. These
movements can be classified as either immigration (movement into the pop-
ulation from outside) or emigration (departure from within the population).
Many demographic studies tally only net migration, but as immigration and
emigration may have different causes, they must often be modeled separately.
As we are limiting ourselves in this chapter and the next to the study of
single populations, we will give here only a single application of this idea,
and consider the exchange of members among multiple populations in later
chapters. For this example we return to the modeling of the United States
population first undertaken in Example 1 of Section 3.2. Table 3.3 lists per
capita birth, death, immigration, and emigration rates for the United States
for every tenth year beginning in 1950, as well as the corresponding population
sizes. (To get the total births, etc. for the given year, we would multiply
the per capita rates by the population size.) A population model designed
to incorporate migration will include a term for each of these. The simplest
approach might be to write an equation containing a linear term for each of
these processes:
dx
= bx − dx + ιx − ǫx, (3.53)
dt
where x is the population size and b, d, ι, and ǫ are the four per capita rates
that drive it.
The assumption underlying (3.53), however, is that each of these processes
occurs at a rate proportional to the size of the population. This may be reason-
able for processes that involve the population itself, including births, deaths,
and emigrations (ignoring for the moment the logistic refinement suggested
in Section 3.2). Immigration, however, may be driven by forces outside the
population under study. For example, the number of people who immigrate
to the United States in a given year may depend more on the populations of
other countries, or on political factors, than on the present population size of
the United States. If so, then the term describing immigration should not be a
function of x. One option would be to introduce a second variable to describe
the factor which drives the immigration rate, but as we wish to keep our model
162 Dynamical Systems for Biological Modeling: An Introduction

TABLE 3.4: United States population predictions (in millions) and percent
errors for models (3.54) and (3.53), alongside the actual values.
Updated parameters 1950 parameters
Year Const. % Prop’l % Const. % Prop’l % True
immig. error immig. error immig. error immig. error value
1950 152 — 152 — 152 — 152 — 152
1960 176.5 −1.94% 176.6 −1.89% 176.5 −1.94% 176.6 −1.89% 180
1970 209.7 2.78% 209.9 2.89% 204.8 0.38% 205.2 0.57% 204
1980 225.5 −0.23% 225.7 −0.15% 237.4 5.05% 238.4 5.47% 226
1990 246.0 −1.59% 246.2 −1.51% 275.1 10.03% 276.9 10.77% 250
2000 277.0 −1.57% 277.4 −1.43% 318.5 13.19% 321.7 14.33% 281.4

to a single variable, the only alternative is to make this rate a constant (or,
conceivably, a function of time alone: see Exercise 16 below for an exploration
of this possibility). This assumption yields an alternative model,
dx
= bx − dx + I − ǫx, (3.54)
dt
where I is the total immigration rate (in, say, people per year).
One way to decide which assumption about the factors driving immigration
gives a better description of population growth is to compare the predictions
of both models with actual data. Both (3.53) and (3.54) are of the form (3.46),
so we may give their solutions, from (3.49), as
I  
x(t) = x(0)e(b−d+ι−ǫ)t and x(t) = x(0)e(b−d−ǫ)t + e(b−d−ǫ)t − 1 ,
b−d−ǫ
respectively. Since both models make the simplifying assumption that the as-
sociated per capita rates are constant over time, as well as the assumption
that the underlying forces are not constrained by resource limitations (hence
the linear terms, rather than, say, logistic), the model should only be used
for limited ranges of time. Table 3.4 compares the predictions made by these
two models in two ways: first using the data for each given year to estimate
only the next decade’s population, and then using the true values for the
next decade to generate the next prediction; and then using the initial (1950)
data to project as far as 2000. One can see that the proportional immigration
model (3.53) always makes higher predictions, which is not surprising since
the immigration rate in (3.53) increases continuously with the population,
rather than remaining fixed as in (3.54). Thus the proportional immigration
model is slightly more accurate than the constant migration model when they
undershoot, and less accurate when they overshoot. In general, however, the
difference in predictions is less than one percent, which implies that the dif-
ference between these two assumptions on immigration rates is less important
for predicting population size than other factors.
As can be seen in Table 3.3, the per capita birth, death, and migration rates
are not really constant over time. In the following section we shall consider
 First-Order Differential Equations (Continuous Dynamical Systems) 163

the more general linear first-order differential equation, in which coefficients

may be functions of time.

Exercises
1. At time t = 0, a tank contains 3 lb. of salt dissolved in 100 gal. of water.
A mixture containing 12 lb. of salt per gallon is pumped in at a rate
of 5 gal./min., and the mixture is pumped out at the same rate. Find
the weight of salt in the tank at time t. How much salt is there after
one hour? Does the weight of salt in the tank at time t have a limit as
t → ∞? [Note Exercises 1–3 involve nonzero initial populations.]
2. A tank contains 100 liters of water and 10 kg. of salt, thoroughly mixed.
Pure water is added at a rate of 5 liters/minute and the mixture is
poured off at the same rate. How much salt is left in the tank after 1
hour, assuming instantaneous and complete mixing?
3. A tank contains 30 lb. of a pollutant dissolved in 200 gallons of water.
Fresh water is poured into the tank at a rate of 10 gal./min. and the
solution is pumped out at the same rate, with immediate and complete
mixing. Find the concentration of pollutant as a function of time, and
the time required for the concentration to drop to 1 lb. per 100 gallons.
4. Water containing 1 pound of salt per gallon is poured into a 50 gallon
tank of water at a rate of 2 gallons per minute and the mixture runs out
at the same rate. After 20 minutes, this process is stopped, and fresh
water is pumped into the tank at a rate of 2 gallons per minute, with
the mixture running out at the same rate. Find the amount of salt in
the tank after 10 more minutes.
5. At time t = 0 cigarette smoke containing 4% carbon monoxide is intro-
duced into a previously carbon monoxide-free room with volume 1800
cubic feet, at a rate of 0.6 cubic feet per minute. The mixture leaves
the room at the same rate. After how long will the carbon monoxide in
the room reach a concentration of 10−4 (which is considered harmful)?
[Make the simplifying assumption that the gas has constant density.]
6. Use either the technique of separation of variables or an integrating
factor to show that the solution to the initial value problem C ′ (t) =
−qC(t), C(T ) = C2 , for T < t ≤ 2T (between the first and second drug
doses), is C(t) = C2 e−q(t−T ) .
7. Suppose that a dose of d milligrams of a drug is injected into the blood-
stream. Assume that the drug is eliminated from the blood at a rate
proportional to the amount of the drug present in the blood. Assume
in addition that half of the drug dose has been eliminated after 1 hour.
Find the time at which the amount of drug in the bloodstream is 20%
of the original dose.
164 Dynamical Systems for Biological Modeling: An Introduction

8. A drug with an elimination constant of 0.1 hours−1 is administered every

8 hours in a dosage of 1 mg/L. Find the residual concentration at the
end of each of the first five intervals.
9. The same drug as in Exercise 8 is administered and it is known that the
maximum safe concentration is 0.5 mg/L (the concentration must never
be allowed to exceed 0.5 mg/L).
Find the shortest time interval T between doses which will be safe.
10. Solve the differential equation (3.51) by using the substitution u = T −
T ∗ to eliminate the constant term, and then back-substituting in the
resulting purely exponential solution.
11. If a body cools from 100◦ C to 60◦ C in surroundings at 20◦ C in 10
minutes, how long will it take for the body to cool to 25◦ C?
12. Water is heated to the boiling point (100◦ C) and is then removed from
the heat and placed in a room with a constant temperature of 20◦ C.
After 3 minutes the water temperature is 90◦ C.
(a) Find the water temperature after 6 minutes.
(b) How long will it take for the water temperature to reach 75◦ C?
13. A container of ice cream with temperature 0◦ F is put in a room with
temperature 75◦ F. After 15 minutes it is observed that the temperature
has risen to 10◦ F. How long will it take for the ice cream to reach a
temperature of 32◦ F?
14. An iron bar at 400◦F is moved to a room with temperature 80◦ F. After 5
minutes the temperature of the bar is 120◦ F. When will the temperature
of the bar reach 90◦ F?
15. The cooling model used by Kienzler et al. (1984) for white-tailed deer
was actually more complex than (3.52): their exponent was more than
a simple linear function of time. In place of −kt, they used β1 t + β2 t2 +
β3 W t + β4 T ∗ t + β5 Z(t), where W is weight and Z(t) = min((t − 3)2 −
9, −9). They estimated β1 = −0.087656/hr, β2 = −0.001174/hr2, β3 =
0.000515/hr kg, β4 = −0.001971/hr ◦ C, and β5 = −0.006168/hr2. Use
their model to repeat the calculation of Example 3 for a 68 kg deer.
What difference does it make in the estimate?
16. Suppose that the immigration rate for the United States is not constant,
as in (3.54), but a linear function of time, I(t) = ι0 + ι1 t. Solve the
resulting differential equation.
17∗ . Using the data in Table 3.3, compute the total immigration per year,
and use linear regression to determine the best-fit coefficients ι0 and ι1
for the model in Exercise 16. Compare the resulting predictions against
those given in Table 3.4 and evaluate the results.
 First-Order Differential Equations (Continuous Dynamical Systems) 165

3.6 First-order models with time dependence

The models presented so far for biological systems have involved param-
eters which represent various aspects of the environment within which the
system operates: both internal characteristics such as natural birth and death
rates and external factors such as resource limitations, inflow and outflow. For
simplicity we have assumed these parameters, which appear as coefficients of
various terms in the models, to be constant. However, there are many situ-
ations in which either this assumption makes model predictions wildly inac-
curate, or the variations in these coefficients are precisely what we want to
study. In mixing processes such as chemostat systems or migrating popula-
tions, inflow and outflow may not be balanced, or may vary periodically. An
environment’s resource capacity may increase or decrease over time. In such
cases, we must adapt our model by relaxing this simplifying assumption, and
writing one or more coefficients as functions of time (or, in some cases, of
population size).
In this section we will look at some ways to solve first-order differential
equations with variable coefficients. Sometimes we can use one of the tech-
niques we have already seen, with no more than minor adaptations, but at
other times it will be more convenient to use an extension of these methods
that shows us more directly the structure associated with the general first-
order linear differential equation. We will begin by developing and illustrating
these two kinds of solution methods, and then look at some applications to
chemostat, migration, fishery management, and population growth models
with variable coefficients.

3.6.1 Superposition
In order to understand the solution methods to be presented in this section,
it is helpful to understand how solutions to a linear ODE can be written in
terms of the solutions to a special, simplified version of the equation. The idea
is that each solution is the sum of two functions, one of which is a solution to
this simpler equation. This idea is called the superposition principle. Although
it applies, properly speaking, only to linear differential equations, for reasons
shown below, we will also use some helpful substitutions to allow us to apply
it to some nonlinear models such as the logistic equation.
The first-order linear differential equation seen in the previous section can
be written in its most general form as
dy
= −a(t)y + b(t), (3.55)
dt
where the coefficients a(t) and b(t) are functions of t. (Recall that a linear
differential equation y ′ = f (t, y) is one in which the function f (t, y) is linear in
166 Dynamical Systems for Biological Modeling: An Introduction

y.) In the special case that a and b are constants, equation (3.55) is separable,
and we found in the previous section that its solution has the form
b
y= + ce−at (3.56)
a
for some constant c. This solution is a sum of two terms, the first a specific
solution y = ab of y ′ + ay = b and the second a family of solutions y = ce−at
of the corresponding homogeneous differential equation y ′ + ay = 0. It turns
out that an analogous result holds for the more general linear ODE (3.55) —
that is, it turns out that solutions to (3.55) can be written as a sum of two
terms, as done in equation (3.56), even if a and b are not constants. First let
us state this result more precisely; then we will see why it is true.
We will call a differential equation such as (3.55) homogeneous if the func-
tion b(t) on the right side is the identically zero function, so that (3.55) takes
the form
dy
= −a(t) y. (3.57)
dt
We will say that the differential equation (3.55) is non-homogeneous (or hetero-
geneous) if b(t) is not the identically zero function. For a non-homogeneous dif-
ferential equation (3.55), the differential equation (3.57) is said to be the corre-
sponding homogeneous differential equation. A homogeneous equation (3.57)
is separable, and by separation of variables has a family of solutions of the
form R
y = c yH (t) = c e− a(t) dt . (3.58)
Here, yH (t) is a solution of the homogeneous differential equation (3.57), and
every solution y of (3.57) is a constant multiple of yH (t) (because this is a
first-order equation35 ). The relation between the solutions of the general equa-
tion (3.55) and the homogeneous equation (3.57) is called the superposition
principle.

THEOREM (SUPERPOSITION PRINCIPLE): Every solution of

the non-homogeneous differential equation (3.55) is the sum of a particular
solution of (3.55) and some solution of the corresponding homogeneous
differential equation (3.57).

This means that once we find a single solution of the non-homogeneous

differential equation (3.55) (perhaps a constant solution, if there is one), we
can obtain all the other solutions of (3.55) by adding the set of all solutions
of the homogeneous equation (3.57), as given by (3.58).
35 As we will see in later chapters, the solution space of an nth-order linear ODE is n-

dimensional; that is, every solution is a combination of constant multiples of n different

functions.
 First-Order Differential Equations (Continuous Dynamical Systems) 167

To establish the superposition principle, we can show that any two solu-
tions y1 (t) and y2 (t) of (3.57) differ by some constant multiple of yH . We first
observe that

y1′ (t) = −a(t)y1 (t) + b(t), y2′ (t) = −a(t)y2 (t) + b(t),

and then calculate

d
(y2 (t) − y1 (t)) = y2′ (t) − y1′ (t) = [−a(t)y2 (t) + b(t)] − [−a(t)y1 (t) + b(t)]
dt
= −a(t)[y2 (t) − y1 (t)].

Thus y2 − y1 is a solution of the homogeneous equation (3.57), for any two

solutions y1 and y2 of the general equation (3.55). Once we know a solution y1 ,
then, we can add the collected solutions to (3.57) to get the collected solutions
to (3.55). None will be left out. For this reason, finding solutions yH to the
homogeneous equation is often a first step in solving the general equation.

3.6.2 Integrating factors

The solution yH to the homogeneous equation (3.57) can now be seen
as the motivation behind the technique of integrating factors introduced in
the previous section for the case of constant coefficients. In Section 3.5 we
suggested looking at the term y ′ + ay as the derivative of a product of two
′
functions (one of them y). From equation (3.58) we can see that yH = −ayH ,
′
so that a = −yH /yH , which transforms (3.55) into
′
yH
y′ − y = b(t).
yH
This suggests the integrating factor 1/yH , by which we multiply to get
  ′
1 1 ′ 1 b(t)
y′ + y − 2 yH y = .
yH yH yH yH
Integration will then allow us to solve for y.

Example 1.
Find all solutions of the differential equation

y ′ + 2y = 2 et .

Solution: We begin with the corresponding homogeneous differential equation

y ′ + 2y = 0, for which we find (by separation of variables, for instance) the
family of solutions yH = ke−2t (we will arbitrarily choose k = 1 in what
follows). This makes our integrating factor 1/yH = e2t , which makes the
equation
′
y ′ e2t + y 2e2t = ye2t = 2 e3t .
168 Dynamical Systems for Biological Modeling: An Introduction

Integrating, we have
2 3t
ye2t = e +c
3
for arbitrary constant of integration c. Thus we have the family of solutions
2 t
y= e + c e−2t . 
3

Example 2.
Find the solution of the initial value problem
y ′ + 2y = 2et , y(0) = 4.

Solution: In Example 1 we found all solutions to the given differential equation,

so now we need only determine for which value of c the initial condition is
satisfied. Substituting t = 0, y = 4 into the solution above, we have 4 =
2 10
3 + c, from which c = 3 . Thus the solution of the initial value problem is
2 t 10 −2t
y = 3e + 3 e . 

Example 3.
Find all solutions of the differential equation
1
y′ + y = 1.
1+t

Solution: The corresponding homogeneous differential equation is

1
y′ = − y,
1+t
and separation of variables gives the solution
dy dt
Z Z
=− ,
y 1+t
log |y| = − log |1 + t| + k.
Exponentiation leads to the family of solutions
c
yH = c e− log |1+t| = .
|1 + t|
If we take t > −1 we can drop the absolute value sign. This yH suggests the
integrating factor (t + 1), which yields
′
(t + 1)y ′ + y = ((t + 1)y) = (t + 1).
Integrating, we obtain
1 t+1 c
(t + 1)y = (t + 1)2 + c, y = + .
2 2 t+1
(Integrating t + 1 to t2 /2 + t + c merely yields a different value for c.) 
 First-Order Differential Equations (Continuous Dynamical Systems) 169

3.6.3 Substitution and integration

We see that in applying an integrating factor, we end up first solving for the
quantity (y/yH ). Giving this quantity a name (typically u) and making the
substitution formal produces another approach to solving nonhomogeneous
equations. In this approach, we define u(t) by
y(t) = yH (t) u(t) (3.59)
and (once we know yH ) use this relation to transform the equation for y into
an equation for u which may be easier to solve outright. This kind of substi-
tution based on guessing a partial form for the solution was one of the chief
techniques developed in the days before computers were widely available to
solve differential equations numerically (or symbolically), when outright solu-
tion of a differential equation might be the only convenient way to interpret
the predictions of the corresponding model.
The change of variable defined in (3.59) actually reduces the solution of
the differential equation to an integration. Differentiating (3.59), we find
y ′ (t) = yH (t)u′ (t) + yH
′
(t)u(t) = yH (t)u′ (t) + [−a(t)yH (t)]u(t)
= yH (t)u′ (t) − a(t)y(t),
so that (comparing to (3.55)) yH (t)u′ (t) = b(t), or
b(t)
u′ (t) = . (3.60)
yH (t)
We can now find u(t) by integration, and then find y(t) = yH (t) u(t).
It is possible to give an explicit formula for y(t),
b(t)
Z
y(t) = yH (t) dt,
yH (t)
and an even more complicated formula may be obtained by using the formula
(3.58) for yH (t). However, we advise against attempting to memorize such a
formula, and recommend instead using the substitution (3.59) and obtaining
the solution by integration.
Note that, from (3.58) and by continuity of the exponential function, the
homogeneous solution yH (t) is continuous and different from zero on any in-
terval on which the function a(t) is continuous. The solution given above,
and hence every solution of (3.55), is defined on every interval on which the
coefficients a(t) and b(t) are continuous, in contrast to separable differential
equations, for which there may be values of t where the functions in the dif-
ferential equation are smooth but solutions do not exist.

Example 4.
Solve the initial value problem
ty ′ + 2y = t3 , y(1) = 0.
170 Dynamical Systems for Biological Modeling: An Introduction

Solution: We first put the differential equation in standard form (where the
coefficient of y ′ is 1) by dividing both sides of the differential equation by t to
give the initial value problem
2
y ′ + y = t2 , y(1) = 0.
t
Next, we need to solve the corresponding homogeneous differential equation
2
y ′ + y = 0.
t
Separation of variables gives the solution
dy dt
Z Z
= −2 ,
y t
log |y| = −2 log |t| + k,
c
y = ±e−2 log t ek = 2 .
t
With yH = 1/t2 , we can make the substitution y = u/t2 , from which

u′ 2u
y′ = 2
− 3.
t t
Substituting for y ′ and y in the original equation gives us
 ′
u′

u 2u 2u 2
− + = t , or = t2 ,
t2 t3 t t2 t2

from which u′ = t4 and thus u(t) = t5 /5 + c. (Alternatively, we can substitute

directly into (3.60) to find u′ .) Now we substitute yH and u into (3.59) to find
the general solution

1 t5 t3
 
c
y(t) = 2 +c = + 2.
t 5 5 t

Substituting t = 1, y = 0 gives that 15 + c = 0, so c = − 51 , and the solution to

3
the initial value problem is y = t5 − 5t12 . 

Example 5.
Solve the initial value problem

y ′ + 2ty = 2t, y(0) = y0 .

Solution: The corresponding homogeneous equation y ′ + 2ty = 0 has solutions

given by
dy
Z Z
= −2 t dt
y
 First-Order Differential Equations (Continuous Dynamical Systems) 171
2 2
so that log |y| = −t2 + k, and y = c e−t . Thus we may use yH = e−t and
2 2
(3.59) or (3.60) to derive that u′ = 2tet , from which u = et + c. Then
2
 2  2
y = yH u = e−t et + c = 1 + ce−t .

Imposition of the initial condition y(0) = y0 gives y0 = c + 1, or c = y0 − 1,

and this gives
2
y = 1 + (y0 − 1)e−t . 
Example 6.
Solve the initial value problem

y ′ + 2ty = sin t, y(0) = y0 .

Solution: The corresponding homogeneous equation is the same as in Exam-

2 2
ple 5, so we can use yH = e−t and (3.60) to derive that u′ = et sin t. Thus
Z
2
u(t) = et sin t dt + c,

and, writing the solution as a definite integral,

Z t
−t2 2 2
y = yH u = e es sin s ds + ce−t .
0

Imposition of the initial condition shows c = y0 and gives, finally,

Z t
2 2 2
y = e−t es sin s ds + y0 e−t . (3.61)
0

Although the integral in (3.61) cannot be evaluated in terms of elementary

functions, it gives a function whose value for t = 0 is zero and whose derivative
2
is et sin t. Using this information and the product rule for differentiation we
may verify that the function defined by (3.61) is actually a solution of the
initial value problem. (See Exercise 15 below.) 

3.6.4 Mixing processes with variable coefficients

We have already looked at some mixing problems which led to linear dif-
ferential equations whose coefficient functions a(t) and b(t) were constants.
These differential equations were not only linear but also separable, and we
treated them in Section 3.5. In these problems, the total volume of fluid re-
mained constant because the rates of flow in and out of the system were the
same. If the rates of flow into and out of the system are not constant, we will
obtain linear differential equations which are not separable. Following are two
examples of mixing processes discussed in the previous section in which the
simplifying assumption of constant inflow and outflow has been removed.
172 Dynamical Systems for Biological Modeling: An Introduction

Example 7.
A culture tank contains 10 liters of water. A solution containing 1 kg/liter of
nutrient is added at a rate of 0.5 liters/hour, and the mixture is drained off at
a rate of 0.3 liters/hour. Find the concentration of the nutrient solution after
24 hours.
Solution: Let s(t) be the weight of nutrient in the tank at time t. Since the
volume of solution in the tank at time t is (10 + 0.2t) liters, the concentration
of nutrient at time t is s(t)/(10 + 0.2t). Thus the weight of nutrient being
removed at time t is 0.3s(t)/(10 + 0.2t) kg/hr. The weight of nutrient being
added is 0.5 kg/hr. Thus the rate of change of the nutrient weight is given by

ds 0.3s
= 0.5 − ,
dt 10 + 0.2t
which we write in standard form as
3
s′ + s = 0.5.
100 + 2t
To find the appropriate integrating factor, we solve the corresponding ho-
mogeneous differential equation by separation of variables,
ds 3 dt
Z Z
=− ,
s 100 + 2t

which leads to log s = − 23 log(100 + 2t) + k1 . Exponentiating, we have

3 3
s = e− 2 log(100+2t) ek1 = (100 + 2t)− 2 k2 .

Since we need only one solution to obtain an integrating factor, we may omit
3
the constant of integration and write sH = (100 + 2t)− 2 . Then the integrating
3
factor is (100 + 2t) 2 , multiplication by which gives
3 1 3 3
(100 + 2t) 2 s′ + (100 + 2t) 2 s = [(100 + 2t) 2 s]′ = 0.5(100 + 2t) 2 .

Integration gives
Z
3 3 5
(100 + 2t) 2 s = 0.5(100 + 2t) 2 dt = 0.1(100 + 2t) 2 + c.

5
Substitution of the initial condition s(0) = 0 gives 0 = 0.1(100) 2 + c, c =
5
−0.1(100) 2 = −104 , leading to the solution
3 5
(100 + 2t) 2 s = 0.1(100 + 2t) 2 − 104 ,
3
s = 0.1(100 + 2t) − 104 (100 + 2t)− 2 .
3
When t = 24, s = 0.1(148) − 104(148)− 2 ≈ 9.25 kg. Also, when t = 24, the
 First-Order Differential Equations (Continuous Dynamical Systems) 173

volume of water in the tank is 14.8 liters, and the concentration of nutrient is
9.25
14.8 ≈ 0.625 kg/liter. The concentration at time t, when the volume of water
is (10 + 0.2t) liters, is
3 − 52
10 + 0.2t − 104 (100 + 2t)− 2

10
= 1− kg/liter. 
10 + 0.2t 10 + 0.2t

Example 8.
If we examine the per capita demographic rates for the United States given in
Table 3.3, we can see clear trends of change in each rate: decreasing birth and
death rates and increasing immigration and emigration rates. Graphs of these
rates show that the changes can be approximately described as linear in time.
Assume rates of this form and the hypothesis discussed in Section 3.5 that
immigration is independent of the U.S. population size, to derive the linear
model x′ = r(t)x + I(t), where r(t) = b − d − ǫ. Compare the resulting popu-
lation predictions with those of the constant-coefficient model of Section 3.5.
Does this refinement affect the model’s predictions significantly?
Solution: Let us define t as time in years since 1950, as was done in the
previous section. To make the demographic coefficients linear in time, we write
r(t) = r0 + r1 t and I(t) = I0 + I1 t, and substitute into the original equation
to obtain
x′ (t) = (r0 + r1 t)x + (I0 + I1 t), (3.62)
for which the corresponding homogeneous equation x′ (t) = (r0 + r1 t)x can be
solved by separation of variables:
dx
Z Z
= (r0 + r1 t)dt
x
t2
log |x| =r0 t + r1 + k
2
2
xH =cer0 t+r1 t /2

and we again take ek = c = 1 for simplicity. Then from (3.60) we have

u′ (t) = (I0 + I1 t)e−(r0 t+r1 t /2)

2
,

from which Z t
(I0 + I1 s)e−(r0 s+r1 s /2)
2
u(t) = x0 + ds,
0

since u(0) = x(0)/xH (0) = x0 /1. We can simplify part of the integral via the
substitution w = r0 s + r1 s2 /2: then
   
r0 I1 r0 I1 dw
I0 + I1 s = I0 − I1 + (r0 + r1 s) = I0 − I1 + ,
r1 r1 r1 r1 ds
174 Dynamical Systems for Biological Modeling: An Introduction

and
I1 I1 I1
Z Z
−(r0 s+r1 s2 /2)
(r0 + r1 s)e ds = e−w dw = − e−w + c
r1 r1 r1
I1 −(r0 s+r1 s2 /2)
=− e + c.
r1
Thus
 Z t
r0 I1  −(r0 t+r1 t2 /2) 
e−(r0 s+r1 s /2) ds −
2
u(t) = x0 + I0 − I1 e −1 ,
r1 0 r1
and
 
I1 2 I1
x(t) = xH u = x0 + er0 t+r1 t /2 −
r1 r1
  Z t
r0
e−(r0 s+r1 s /2) ds.
2
r0 t+r1 t2 /2
+ I0 − I1 e
r1 0

As in Example 6, the unresolved integral cannot be evaluated in terms of ele-

mentary functions, but values can be obtained through numerical integration
for any given value of t, once we have values for the demographic coefficients.
To determine coefficient values for the demographic rates, we use simple
linear regression on the values in Table 3.3. This yields b0 = 0.0240095/yr,
b1 = −0.000205714/yr2, d0 = 0.00971905/yr, d1 = −0.0000254286/yr2,
ǫ0 = 0.000152571/yr, ǫ1 = 0.0000129171/yr2 (so that r0 = 0.0141379/yr,
r1 = −0.000193203/yr2), I0 = 63728.6 people/yr, I1 = 15982.9 people/yr2 .
We compare the resulting solution with that obtained in Section 3.5 for fixed
coefficients, taking for the coefficient values here the averages of those in Ta-
ble 3.3, which yield r = 0.0093/yr and I = 463, 300 people/yr. Figure 3.21
compares the two solutions with the actual population values; one can see
that the model with coefficients linear in time makes predictions with about
half the error of the model predictions using constant coefficients for most of
the period given. Only toward the end of the period does a sudden jump in
census population throw the variable-coefficient model’s estimate off. Overall,
the two estimates are within 3% of each other during this period. 

3.6.5 Bernouilli equation

Earlier in this section we observed the use of a substitution (change of
variables) to transform a differential equation into another equation which may
be easier to solve. The methods outlined in this section apply only to linear
differential equations, but there are classes of nonlinear ODEs which can be
transformed into linear ODEs — and hence solved — via a substitution. One
such class of equation is the Bernouilli equation (named after Jacob Bernouilli
(1654–1705)), of the form
y ′ + a(t)y = b(t)y n , (3.63)
 First-Order Differential Equations (Continuous Dynamical Systems) 175

FIGURE 3.21: Migration model predictions for the U.S. population using
fixed coefficients (short dashes) and coefficients linear in time (long dashes),
compared to census values (solid line).

where n is a constant which is neither 0 nor 1. (If n = 0 or n = 1 the differential

equation (3.63) is linear.) The change of dependent variable z = y 1−n in (3.63)
gives
1
y = z 1/(1−n) , y ′ = z n/(1−n) z ′ ,
1−n
so that
1
z n/(1−n) z ′ + a(t)z 1/(1−n) = b(t)z n/(1−n) ,
1−n
z ′ + (1 − n)a(t)z = (1 − n)b(t). (3.64)
Since (3.64) is linear it may be solved by finding the appropriate integrating
factor and integrating. Note, however, that the identically zero solution of
(3.63) may not be obtained by this method, because we have divided through
by z n/(1−n) .

Example 9.
Solve the initial value problem
1
y ′ − y = −ty 2 , y(0) = − .
2

Solution: The given differential equation is a Bernouilli equation with n = 2.

′
Therefore we let z = y −1 , or y = z1 , y ′ = − zz2 , and the differential equation
becomes
1 1 t
− 2 z′ − = − 2 ,
z z z
z ′ + z = t.
176 Dynamical Systems for Biological Modeling: An Introduction

The appropriate integrating factor is et , and multiplying by it yields

z ′ et + zet = (zet )′ = tet ,

Z
zet = tet dt = (t − 1)et + c.

Then we have
1 1
y= = .
z (t − 1) + ce−t
Substitution of the initial condition y(0) = 21 gives 12 = c−1
1
, or c = 3. Thus
the solution of the given initial value problem is y = [(t − 1) + 3e−t ]−1 . 

Bernouilli equations may arise in many different areas, as an artifact of

the mathematics involved. For example, Benchekroun and Van Long stud-
ied a problem in international fishing by casting it in terms of game theory;
the resulting equation for the optimal reactive strategy took the form of a
Bernouilli equation36 with exponent n = 2. In the situation they modeled,
migratory fish that travel along a coastline belonging to two nations, such as
the Canadian salmon which travel along the Alaskan coastline on their way
back to their Canadian breeding grounds, are subject to catching by groups
from both countries, with the first player in an interceptor role and the sec-
ond in a reactive role. Overfishing will drive the population to extinction. The
optimal reaction of the second player in cases where the first player deviates
from his optimal catch can be described by a differential equation which the
authors transform into a Bernouilli equation.
However, by far the most common Bernouilli equation to arise in the study
of biological systems is the logistic equation, in which again n = 2. We have
already solved the logistic equation with constant coefficients using separation
of variables, but when the coefficients — the intrinsic growth rate r and the
carrying capacity K — vary over time the equation is no longer separable.
In practice, these coefficients do vary with time: on a short time scale, peri-
odic (daily or annual) fluctuations affect system resources and reproduction,
while on a long time scale gradual changes in both the environment and the
organisms which live in it can accumulate and become appreciable.
We have seen in Example 9 that the change of variables y = 1z transforms
a Bernouilli equation with n = 2 into a first-order linear differential equation.
The result of this substitution is the same whether r and K are constants
or functions of t. Thus we can use this substitution to transform the logistic
differential equation with variable coefficients
 
′ y
y = r(t)y 1 − (3.65)
K(t)
36 Hassan Benchekroun and Ngo Van Long, Transboundary fishery: a differential game

model, Economica 69: 207–221, 2002. See their Appendix B for details.
 First-Order Differential Equations (Continuous Dynamical Systems) 177

into the linear first-order differential equation

r(t)
z ′ = −r(t)z + . (3.66)
K(t)
In order to find an integrating factor, we find a solution ofRthe correspond-
ing homogeneous equation z ′ = −r(t)z, namelyR zH (t) = e− r(t)dt , and this
r(t)dt
indicates an integrating factor 1/zH (t) = e . Multiplication of (3.66) by
this integrating factor gives
R
′
R
r(t)dt
R
r(t)dt
h R
r(t)dt
i′ r(t) e r(t)dt
z e + zr(t)e = ze = ,
K(t)
and integration gives
R
r(t) e r(t)dt
R
Z
r(t)dt
ze = dt + c, (3.67)
K(t)
from which we can obtain an explicit formula for z and then an explicit formula
for the original unknown function y = 1/z.
If we assume the carrying capacity K to be constant in order to study the
effects of a variable intrinsic growth rate, (3.67) gives
R
1 e r(t)dt
R
Z R
r(t)dt r(t)dt
ze = r(t) e dt + c = + c,
K K
and we obtain
1 R
z= + c e − r(t)dt ,
K
1 K
y= 1
R = R . (3.68)
K + c e− r(t)dt 1 + cKe− r(t)dt
(Compare with the solution (3.37)R for the case whereR r is constant, too.)
∞ ∞
We see that if the infinite integral r(t)dt diverges, r(t) dt → ∞, then
limt→∞ y(t) = K. This will be the case whenever the growth rate r(t) does
not die down to zero fast enough: if, for example, r is constant, or periodic
(seasonal), say r(t)R = r0 + a cos ωt with r0 > 0. When r(t) is integrable as
∞
t → ∞, however, r(t) dt → L, then limt→∞ y(t) < K. Thus (3.68) gives
the quantitative behavior of the population when the growth rate varies, and if
in addition r(t) dies down to zero fast enough (for whatever biological reason)
the population will settle at a lower level.
If instead r is constant but the carrying capacity K is a function of t,
(3.67) gives
rert
Z
rt
ze = dt + c.
K(t)
In terms of definite integrals we may write
Z t
−rt r ers
z=e ds + z(0) e−rt . (3.69)
0 K(s)
178 Dynamical Systems for Biological Modeling: An Introduction

In order to find lim z(t), we may use L’Hôpital’s rule for the indeterminate
t→∞
form R t r ers r ert
0 K(s) ds K(t) 1
lim rt
= lim rt
= lim .
t→∞ e t→∞ re t→∞ K(t)
1
Thus if K(t) approaches a limit as t → ∞, then z(t) → limt→∞ K(t) , and thus
y(t) and K(t) have the same limit. If the carrying capacity does not approach
a limit (for instance, if it is seasonal), then the population will not do so, either
(since its “target level” is moving). However, the solution can be shown (see
Exercise 15 below) to be asymptotically bounded within the range of values
taken on by K(t).

Example 10.
In Section 3.2 we discussed the use of the logistic model with constant coeffi-
cients in accounting for Australian sheep populations, noting that the model
described the initial period of rapid growth well but failed to account for
the sustained significant (10% or more) fluctuations seen after the population
reached its carrying capacity. One hypothesis has been that the pasture land
required some time to recover from the overgrazing it experienced. Assume
that the pastures undergo cycles of overgrazing followed by recovery. To what
extent can this assumption account for the variations in the Tasmanian sheep
population seen in Figure 3.7?
Solution: From the figure, if the period 1846–1856 represents the period of
initial overgrazing (causing growth greater than expected for the amount of
system resources), then 4 cycles of fluctuation can be observed during the ap-
proximately 68 years which follow, with amplitude approximately 170 (thou-
sand sheep). This can be modeled most simply by adding a sine wave of
amplitude 170 and period 17(= 68 ÷ 4) to the estimated carrying capacity of
1670 (thousand sheep) given in the figure caption, making
2π
K(t) = 1670 + 170 sin (t − 1856),
17
where t is the year. If we assume the sheep’s intrinsic growth rate still constant
at r = 0.13125/yr, then we have the second special case described above, which
leads to (3.69). Taking our starting point as t0 = 1857, for which the original
solution y(t) = 1670/[1 + exp(240.81 − 0.13125 t)] has the value 1585, so that
z0 = 1/1585, we can then rewrite (3.69) as
Z t rs
re
z(t) = z0 e−r(t−t0 ) + e−rt ds.
t0 K(s)

Finally, recall y(t) = 1/z(t).

Although the integral in the expression for z cannot be evaluated in terms
of elementary functions, it can be evaluated numerically for any given value of
t. We can therefore plot a graph of y and compare it with the given data; see
 First-Order Differential Equations (Continuous Dynamical Systems) 179

FIGURE 3.22: A logistic model FIGURE 3.23: A population

with sinusoidally varying carrying ca- (dashed curve) governed by a logistic
pacity approximates observed fluctu- model with periodic carrying capac-
ations in the Tasmanian sheep popu- ity (system resources) oscillates out
lation. of phase with the carrying capacity
(solid curve). Also shown are simple
logistic bounds (see Exercise 15).

Figure 3.22. We see that the population size does not approach the variable
carrying capacity asymptotically; rather, it increases for that part of the cycle
when y(t) < K(t), and decreases as long as y(t) > K(t) (see Figure 3.23). We
also see that the fluctuations appear to have a negative effect overall on the
population, in that the population predicted using a variable carrying capacity
is less than the average carrying capacity more than half of the time (that is,
it is below the line more often than above it). In general, we can see that this
model is more consistent with observed population values than the original,
though we should note that this is still only a rough approximation at best. 

Exercises
In Exercises 1–12, solve the given initial value problem.

1
1. y ′ + 2y = 2, y(0) = 2 7. ty ′ + 2y = t2 , y(−2) = 5

2. y ′ + 2y = et , y(0) = 1 8. t2 y ′ + 2ty = 1, y(1) = −1

3. y ′ + 3y = e−2t , y(0) = 5 9. ty ′ + y = et , y(1) = 1
4. y ′ + y = t, y(1) = −2 10. ty ′ − y = t2 , y(1) = 1
5. y ′ + 1t y = 1, y(1) = 1 11. y ′ = −2t(y − t2 ), y(0) = 0
6. y ′ − 1t y = 1, y(1) = 1 12. (t + 1)(y ′ + y) = e−t , y(0) = 0

y y3
13. Find the solution of y ′ + t = t2 passing through the point (6,3).
14. Find the solution of y ′ − y = ty 2 passing through the point (1, −1).
180 Dynamical Systems for Biological Modeling: An Introduction

15. Find the solution of the differential equation ty ′ = 2t2 y + y log y through
the point (1,1). [Hint: Make the change of dependent variable v = log y.]
16∗ . Discuss the behavior of solutions of the initial value problem y ′ = λy,
y(0) = y0 as t → ∞ for each of the cases λ > 0, λ = 0, and λ < 0. (We
will see in Section 4.1 that this is a key issue in the qualitative analysis
of differential equations.)
17∗ . Solve and then discuss the behavior of solutions as t → ∞ for each of
the following differential equations satisfying the generic initial condition
y(t0 ) = y0 , where t0 and y0 are given constants: (a) y ′ = −2y + e−t (b)
1
y ′ = −2y +et (c) y ′ = −2y +1 (d) y ′ = −2y + 1+t 2 [Hint: Use L’Hôpital’s

rule to evaluate the limit.]

18∗ . Find the solution of the initial value problem y ′ + y = g(t), y(0) = 0,
where g(t) = 1 if 0 ≤ t < 2 and g(t) = 0 if t ≥ 2. This equation models
a system with an external influence (i.e., independent of the population
y) which shuts off after a given time.
Rt
19∗ . Solve the integral equation y(t) = 0 y(s) ds + t + a. [Hint: Differentiate
the equation and note that y(0) = a.]
20∗ . (a) Show (by constructing it) that the differential equation y ′ + y = f (t)
has a unique solution which is bounded for −∞ < t < ∞ if f (t) is
continuous and |f (t)| ≤ M for −∞ < t < ∞. [Hint: Consider the
solution yA with yA (−A) = 0, where A > 0, and let A → ∞, y(t) =
limA→∞ yA (t).] (b) If the function f (t) is periodic with period 2π (i.e.,
f (t + 2π) = f (t) for −∞ < t < ∞), show that the solution obtained in
part (a) is periodic with period 2π. [Hint: Show by a suitable change of
the variable of integration that the solution obtained in part (a) satisfies
y(t + 2π) = y(t) for every t.]
21∗ . Let a be a positive constant and let limt→0+ f (t) = b. By choosing an
appropriate initial condition, show that the differential equation

ty ′ + ay = f (t)

has a unique solution that is bounded as t → 0+, and find the limit of
this solution as t → 0+.
22. Verify that the function defined by equation (3.61) is the solution of
Example 6.
23. A tank contains 10 liters of water to which is added a nutrient solution
containing 0.3 kg of nutrient per liter. This nutrient solution is poured
in at the rate of 3 liters/min, is thoroughly mixed, and then the mixture
is drained off at the rate of 2 liters/min. How much nutrient is in the
tank after 5 minutes?
 First-Order Differential Equations (Continuous Dynamical Systems) 181

24. A 500 gallon runoff containment tank contains 200 gallons of a solution
containing 15 lb. of pollutant, when heavy rains flood the system. Fresh
water pours into the tank at a rate of 12 gallons per minute, and the
mixture is pumped out for processing at a rate of 6 gallons per minute.
Find the amount of pollutant in the tank when the tank overflows.
25. A catfish pond has 1000 L of water containing 2 kg of dissolved salt.
Flooding causes seawater to enter the pond at 60 L/min. (Seawater has
an average salt content of 36 g/L.) At the same time, a drainage system
turns on and pumps the water out at a rate of 100 L/min. (a) Find the
amount of salt in the tank at any time t. (b) Find the concentration of
salt (in grams per liter) as a function of time. Chervinski found37 that
young catfish can survive salinity levels up to 1/4 that of seawater. How
long does the pond’s caretaker have to rescue the catfish? (c) Sketch the
graphs of the amount of salt and the concentration of salt against time,
and determine the absolute maximum of each quantity.
26. A 400 liter tank contains a mixture of water and chlorine with 0.05 grams
of chlorine per liter. In order to reduce the concentration of chlorine in
the tank, fresh water is pumped into the tank at a rate of 4 L/sec. After
thorough stirring, the mixture is pumped out at a rate of 10 L/sec. Find
the amount of chlorine in the tank as a function of t.
27. A very crude model of a population with migration is given by
y ′ + ay = M (t), y(0) = y0 ,
where a is the proportional rate of decrease without migration and M (t)
is the rate at which members are added to the population. Here, positive
values of M (t) correspond to immigration, and negative values of M (t)
correspond to emigration. Find the population size as a function of t if
M (t) is a positive constant M0 .
28. Find the population size y modeled by y ′ +ay = M0 +A sin ωt, y(0) = y0
where the sine term models fluctuations in migration. [Observe that if
|M0 | < |A|, then the migration rate takes on both positive and negative
values, indicating that there are periods of net immigration and periods
of net emigration.]
29. Solve the logistic equation with constant coefficients y ′ = ry(1 − y/K)
via the Bernouilli substitution.
30. Suppose that a population governed by the logistic model (3.65) has
a carrying capacity bounded by constants Kmin ≤ K(t) ≤ Kmax for
all time (and a constant intrinsic growth rate r). Show that the result-
ing population size y(t) is bounded by the two solutions of the form
37 J. Chervinski, Salinity tolerance of young catfish, Clarias lazera (Burchell), Journal of

Fish Biology 25(2): 147, August 1984.

182 Dynamical Systems for Biological Modeling: An Introduction

(3.13) to the logistic equation with constant carrying capacities Kmin

and Kmax , respectively (as illustrated in Figure 3.23), and thus asymp-
totically bounded by the constants Kmin and Kmax . [Hint: Use the above
inequality to derive bounds on the integrand in (3.69), and use the re-
sulting bounds on z to derive bounds for y = 1/z.]

Miscellaneous exercises
In Exercises 1–4, verify that the given family of functions is a solution of the
given differential equation for every value of the constant c.
sin t − 2 cos t ′
1. y = ce2t + , y − 2y = cos t
5
2
2. y = cet /2
, y ′ − ty = 0
e−1/t 2 ′
3. y = c , t y = y(1 − t)
t
y
4. y = ct − t log t, y ′ = −1 + t

In Exercises 5–8, draw a direction field for the given differential equation on
the given rectangle.
5. y ′ − 2y = cos t; 0 ≤ t ≤ 5, −2 ≤ y ≤ 2
6. y ′ − ty = 0; −2 ≤ t ≤ 2, −2 ≤ y ≤ 2
1
7. t2 y ′ = y(1 − t); 2 ≤ t ≤ 4, −1 ≤ y ≤ 1
8. y ′ = −1 + yt ; 1
2 ≤ t ≤ 4, 0 ≤ y ≤ 4
For each of the differential equations in Exercises 9–12, construct the direction
field and use it to sketch the solution passing through the origin.
9. y ′ = t + y
t
10. y ′ =
y−t
11. y ′ = (y + 1)(t + 1)
12. y ′ = y 3 − 1
In Exercises 13–26, find the solution of the given initial value problem.
13. y ′ = t2 y 2 , y(1) = 1
14. y ′ = ty log y, y(0) = e
2
15. y ′ = tey et , y(1) = 0
First-Order Differential Equations (Continuous Dynamical Systems) 183

16. y ′ = (1 + y 2 )(1 + t2 ), y(0) = 0

17. yy ′ = t, y(1) = 0
18. y ′ = tan y, y( π4 ) = 1
y
19. y ′ = t + t2 , y(1) = 1
20. y ′ + ty = t2 , y(0) = 0
21. ty ′ − y = ty, y(1) = 4
22. y ′ − y = et , y(0) = 1
1
23. y ′ = ty + y 2 , y(−1) = 1
24. ty ′ + y = t2 y 2 , y(0) = −1
1
25. (1 + t2 )y ′ + 2ty = , y(0) = 0
1 + t2
26. (1 + t2 )y ′ = 1 + ty, y(1) = 0
27. A population of organisms governed by the law of simple population
growth has a birth rate of 0.35 per member per week. How long does it
take for the population size to triple?
28. A population governed by the law of simple population growth has a
birth rate of 0.20 per member per week. If the initial population size is
100 members, what is the population size after 6 weeks?
29. The half-life of radium is 1620 years. How long does it take for one
quarter of a given amount of radium to disintegrate?
30. The half-life of plutonium is 50 years. How long does it take for one
tenth of a given amount of plutonium to disintegrate?
31. A population grows according to the logistic law with an initial popula-
tion size of 100 members and a limiting population size of 500 members.
The population size reaches 250 in 1 week. What is the population size
after 2 weeks?
32. A population grows according to the logistic law with an initial popu-
lation size of 100 members, a population size of 200 members after 1
week, and a limiting population size of 500 members. Find the rate at
which the population size is growing after 1 week.
33. A tank contains 1 gallon of a salt solution consisting of 1 lb. salt per
gallon. A solution containing 2 lb. salt per gallon runs into the tank at
a rate of 5 gallons per minute and the mixture runs out of the tank at
the same rate. Find the weight of salt in the tank as a function of time.
184 Dynamical Systems for Biological Modeling: An Introduction

34. To a tank containing 100 lb. salt dissolved in 100 gallons of water is
added a solution containing 2 lb. salt per gallon at a rate of 2 gallons
per minute. The mixture runs out at a rate of 4 gallons per minute. Find
the weight of salt and the concentration of the mixture after 10 minutes.
35. A body with initial temperature 100◦ C cools in air at 20◦ C, taking 5
minutes to cool to a temperature of 80◦ C. How long does it take to reach
a temperature of 60◦ C?
36. A hot body is cooled in air at 20◦ C, reaching temperatures of 45◦ C 10
minutes after it starts to cool and 40◦ C 20 minutes after it starts to
cool. What was its original temperature when cooling began?
p
37∗ . Explain why the initial value problem y ′ = 1 − y 2 , y(0) = 2 has no
solution.
38∗ . Show that the function y(t) defined as y = 0 for 0 ≤ t < 1 and as y = t3
for 1 ≤ t ≤ 2 is differentiable for 0 ≤ t ≤ 2 and solves the initial value
problem y ′ = 3y 2/3 , y(0) = 0.
ry
39. The differential equation y ′ = y+A − y may be thought of as a limiting
case of the Verhulst difference equation yk+1 = ykry+A
k
(think of yk+1 − yk
′
as an approximation to the derivative y ). Compare the behaviors of the
solution of the difference equation and the solution of the differential
equation, with A = 1 and various values of r. For what values of r do
the two equations behave similarly?
40. (a) Formulate a differential equation which is a limiting case of the
Beverton-Holt difference equation
ayk
yk+1 = .
1 + byk

(b) For what values of the parameters a and b do the solutions of the
difference equation and the differential equation derived in part (a)
behave similarly?
41∗ . As has been mentioned in the text, use of a continuous variable to rep-
resent a discrete population is an approximation, the error involved in
which we are usually willing to overlook if it is small relative to the
population size; that is, if a model predicts a population of 761.23 birds
at a given time, we can use the general size rounded to the nearest inte-
ger as an indication that the population will include something close to
761 birds. However, if we keep in mind the fact that deterministic mod-
els such as differential equations represent an average over all possible
outcomes, we can interpret non-integer (or indeed all) results in these
terms. Write probabilistic interpretations of the results in the first two
simple examples of Section 3.1.
Chapter 4
Nonlinear Differential Equations

In the previous chapter, we developed techniques for solving first-order differ-

ential equations which are either linear or separable. However, in general the
models we develop to describe biological systems will not be linear—indeed,
we shall shortly see several types of biological systems which are driven by
distinctly nonlinear interactions. Most nonlinear differential equations cannot
be solved outright; we can, however, develop tools to analyze the long-term
behavior of these models, as was done for difference equations in Chapter 2,
and often the long-term behavior of the biological system is what really inter-
ests us. (If we are interested in short-term behavior of the system, we typically
use computers to approximate the solution numerically.1 )
In this chapter, we shall first develop these qualitative analysis tools for
single first-order differential equations, taking care to use notation which will
help us extend them to systems of differential equations in the next chap-
ter. Following this, we shall study particular types of nonlinear interactions
such as harvesting (last seen in Section 2.5) and contact processes. One of
the fundamental results in biological modeling is the description of threshold
quantities which control sharp changes in the nature of a system—for exam-
ple, the difference between survival and extinction—and we shall also take
time to examine common biological thresholds under the mathematical lens
of bifurcations.

4.1 Qualitative analysis tools

As an example of a simple nonlinear model for a biological system, let
us consider the model developed by Beekman, Sumpter, and Ratnieks2 to
describe the foraging of Pharaoh ants, Monomorium pharaonis. Like many
1 Before computers became readily available, researchers often expended great effort to
solve differential equations outright; this fact accounts for the many solution techniques
that have historically been taught in courses on ordinary differential equations.
2 Madeleine Beekman, David J. T. Sumpter, and Francis L. W. Ratnieks, Phase transition

between disordered and ordered foraging in Pharaoh’s ants, PNAS 98(17): 9703–9706, 14
August 2001; and David J. T. Sumpter and Madeleine Beekman, From nonlinearity to
optimality: pheromone foraging by ants, Animal Behaviour 66: 273–280, 2003.

185
186 Dynamical Systems for Biological Modeling: An Introduction

ants, these foragers lay pheromone trails connecting any food sources they find
back to their nests, in order to help nestmates find the food (see Figure 4.1).
These pheromone trails are volatile, lasting only about 10 minutes unless ants
still on the trail lay down more. Beekman, Sumpter, and Ratnieks studied
the effects of colony size, distance from food source to nest, and strength
of the pheromone trail on the number of ants using the trail. Their model
hypothesized, first, that ants begin to forage at a given food site either by
finding it independently, at a rate α inversely proportional to the food site’s
distance from the nest, or by being drawn to the pheromone trail, at a rate
dependent both on the number of ants laying the trail down and on the quality
of the food source, measured by a parameter β. If we let x(t) represent the
number of ants foraging on the trail at time t, and denote the colony size by
n, then the rate at which an ant begins using the trail is given by α + βx, the
number of ants not using the trail is n−x, and thus the total rate at which ants
join the trail is (α+βx)(n−x). Their model also hypothesized that ants lose or
stop using the pheromone trail at a rate dependent, in part, on the strength
of the trail. More specifically, they hypothesized that the rate at which an
individual ant loses the trail is inversely proportional to the number of ants x
renewing the trail, in such a way that the total rate at which ants leave the
trail never passes a maximum rate s. (Note the distinction here between the
rate for a single ant, or per capita rate, and the total rate.) This hypothesis
sx
gives a trail loss rate of x+K , where K is the number of ants on the trail at
which the total trail loss rate reaches s/2. Since the total rate never surpasses
s, this function is said to saturate; we shall see other examples of biological
processes that saturate in the sections ahead.
The model that results from the above hypotheses is
dx sx
= (α + βx)(n − x) − . (4.1)
dt x+K
Note that although equation (4.1) is separable, the technique of separation
of variables does not lead to a closed-form expression for x(t) because of the
algebraic complexity of the terms. We can nevertheless learn a lot about the
behavior this set of rules produces, results which Beekman, Sumpter, and
Ratnieks confirmed experimentally and interpreted to reveal a remarkable
shift in the efficiency of such behavior as colony size grows. In order to do so,
we now develop the tools we will need.

Photo courtesy Francis Ratnieks

FIGURE 4.1: Foraging Pharaoh ants following a pheromone trail.

 Nonlinear Differential Equations 187

4.1.1 Possible end behaviors

We might begin our consideration of long-term behavior by asking what
can happen to a biological system in the end. A quick review of the continuous-
time systems we studied in Chapter 3 reveals two types of behavior observed so
far: the quantity or population in question either (1) settles down to a constant
level, called an equilibrium, or (2) in the case of our simplest models, grows
unbounded. Recalling the period doubling and chaos exhibited in Chapter 2
by nonlinear discrete-time models describing a population under scramble
competition (such as the discrete logistic equation), we might ask whether it
is possible for continuous-time models dy dt = f (y) of a single population to
predict sustained oscillations and other complicated behaviors. It turns out
that (under some very simple hypotheses, such as f (y) being a continuous
function) they cannot: they can only do one of the two things listed above.
We can see why this is true if we think for a moment about the nature of the
model. Since the rate of change dy dt of y depends only on y, then for a given
y-value, say y1 , every time the solution reaches that value, its rate of change
will be the same, f (y1 ). More specifically, if f (y1 ) > 0, then the solution will
always increase from y1 ; if f (y1 ) < 0, then the solution will always decrease
from y1 . This means that a solution can never return to a point it has already
passed: if f (y1 ) > 0, then once the solution has passed above y1 , it can never
pass below it, because any solution that gets close enough to y1 is forced
upward. In other words, every solution is monotone: either always increasing
or always decreasing (or constant).
Note that this argument holds only for one-dimensional, continuous-time
models: solutions to discrete-time models make discrete jumps, enabling a
solution to pass from above y1 to below y1 without being affected by f (y1 ),
and solutions to multi-dimensional systems can pass below (or even return to)
a given point without first passing through it.
Before continuing, we should state this result a little more precisely. We
restrict our study here to autonomous differential equations, which do not
depend explicitly on the independent variable t, of the general form
dy
= f (y). (4.2)
dt
(Autonomous differential equations are always separable, R but solution by sep-
aration of variables may be impractical if the integral fdy (y) is difficult or im-
possible to evaluate, as with the ant model (4.1) above.) We will also assume
that the function f (y) is smooth enough that the existence and uniqueness
theorem of Section 3.3 holds, and there is a unique solution to (4.2) for each
given initial condition. For simplicity, we shall consider only non-negative val-
ues of t, and we will think of solutions as determined by their initial values
for t = 0. Under these constraints, we can make use of certain properties of
autonomous differential equations, beginning with the role of fixed points of
the system.
188 Dynamical Systems for Biological Modeling: An Introduction

Property 1: Equilibria. The constant solutions of the differential equa-

tion y ′ = f (y) are precisely the zeroes of the function f (y). That is, for
a given number y ∗ , y = y ∗ is a constant solution of (4.2) if and only if
f (y ∗ ) = 0.

This is true because the derivative of a constant function is the zero func-
tion, so if y = y ∗ , then dy
dt = 0, and this function obeys the differential equation
dy ∗
dt = f (y) if and only if f (y ) = 0. Such a solution is called an equilibrium
or fixed point of the differential equation. Qualitative analysis of differential
equations begins with the study of the equations’ equilibria.
Because any solution which includes an equilibrium remains at that equi-
librium for all time, by virtue of the informal argument raised earlier, the
equilibria of a single differential equation separate the state space (the set of
all possible values of y(t)) into disjoint intervals: no solution can cross an equi-
librium (or else two solutions pass through that point, violating uniqueness),
so any solution must stay within a single interval. Furthermore, since f (y) is
continuous, and is zero only at equilibria, f (y) cannot change sign within an
interval (between equilibria): it will either be uniformly positive, in which case
all solutions in that interval will increase monotonically, or uniformly nega-
tive, in which case all solutions in that interval will decrease monotonically.
If we consider the graph of y(t) over time, these intervals become horizontal
bands on the graph.

Property 2: Monotonicity. The graph of every solution curve of the

differential equation y ′ = f (y) remains in the same band and is either
monotone increasing (y ′ (t) > 0) or monotone decreasing (y ′ (t) < 0) for all
t ≥ 0, depending on whether f (y) > 0 or f (y) < 0, respectively, in the
band.

It may be easiest to see these properties at work via an example already

familiar to us.

Example 1.
Describe the bands for the logistic differential equation
dy  y
= ry 1 − (4.3)
dt K
with r, K > 0, and find which solutions are increasing, and which decreasing.
Solution: The

bands are bounded by equilibria, which are the roots of f (y) =
y
ry 1 − K , namely y = 0 and y = K. There are therefore three bands to
 Nonlinear Differential Equations 189
y


consider (Figure 4.2). If y > K (Band 1), then 1 − K < 0, and the function
y


f (y) = ry 1 − K is negative. Therefore, solutions y(t) of (4.3) with y(0) > K
are decreasing for all t. If 0 < y < K (Band 2), the function f (y) is positive,
and therefore solutions y(t) with 0 < y(0) < K are increasing for all t. If y < 0
(Band 3), then f (y) is negative, and therefore solutions y(t) with y(0) < 0 are
decreasing for all t. (Compare Figure 4.2 with the direction field in Figure 3.11,
where r = K = 1.) 

3
y
Band 1 2.5

y(t) decreasing
2
K

Band 2 y(t)1.5

y(t) increasing
1

0 t

Band 3 0.5

y(t) decreasing
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
t

FIGURE 4.2: The t − y plane di- FIGURE 4.3: Some solutions of the
vided into bands by the equilibria of logistic equation (4.3) with r = 2,
the logistic equation. K = 2.

Note that if y(0) is above the largest equilibrium of (4.2), then the interval
or band containing the solution is unbounded, and if f (y) > 0 in this band,
then the solution y(t) may be (positively) unbounded and does not neces-
sarily exist for all t ≥ 0. Likewise, if y(0) is below the smallest equilibrium
of (4.2) and f (y) < 0 in the band containing the solution, then the solution
may be unbounded (negatively) and may fail to exist for all t ≥ 0. In the
example above of the logistic equation, solutions with y(0) < 0 (Band 3) grow
negatively without bound, but nevertheless exist for all t ≥ 0.
In practice, it is often easy to check for unbounded growth in solutions of a
single differential equation, by seeing whether the growth rate f (y) is bounded
away from zero for large y. A solution may become unbounded positively
(i.e., y(t) → +∞) if it is monotone increasing when y is large. Models of
most biological systems account for limitations on system resources, so that
190 Dynamical Systems for Biological Modeling: An Introduction

f (y) < 0 for very large y, and in such a case every solution remains bounded,
because solutions which become large and positive must be decreasing. If, as is
also frequently the case in applications, y = 0 is an equilibrium and only non-
negative solutions are of interest, then solutions cannot cross the line y = 0
or become negatively unbounded (i.e., y(t) → −∞). In many applications, y
stands for a quantity such as the number of members of a population, the
mass of a radioactive substance, the quantity of money in an account, or the
height of a particle above ground which cannot become negative. In such a
situation, only non-negative solutions are significant, and if y = 0 is not an
equilibrium but a solution reaches the value zero for some finite t, we will
consider the population system to have collapsed and the population to be
zero for all larger t. If this is the case, we need not be concerned with the
possibility of solutions becoming negatively unbounded even if y = 0 is not
an equilibrium.
In the above case of the logistic equation (4.3), we can see that every
non-negative solution remains bounded, by observing that f (y) < 0 (and thus
dy/dt < 0) when y > K (Band 1).
We can now combine the properties of boundedness and monotonicity to
argue that any solution of a differential equation (4.2) which does not grow
unbounded must instead approach a limit. From calculus, any function which
is either bounded above and monotone increasing, or bounded below and
monotone decreasing, must approach a limit. (Consider the analogy of a car
stuck in forward gear approaching a wall along a narrow lane. At some point—
the wall, if not sooner—it must approach a stopping point.) However, it can
be shown that if a differentiable, monotone function—a solution of (4.2)—
tends to a limit as t → ∞, then its derivative meanwhile must tend to zero.
That is, if y(t) → L, then dy d
dt → dt L = 0. At the same time, by continuity
dy
of f , f (y(t)) → f (L); since dt = f (y(t)), this suggests that f (L) = 0: that
is, that L is a root of f (y). By Property 1, this means that L must be an
equilibrium of the system. Thus every bounded solution of (4.2) approaches
an equilibrium.

Property 3: Bounded solutions. Every solution of the differential

equation dy
dt = f (y) which remains bounded for 0 ≤ t < ∞ approaches
an equilibrium of the equation as t → ∞.

In the case of the logistic equation (4.3), this means that since all non-
negative solutions are bounded, all non-negative solutions must approach an
equilibrium, either y = 0 or y = K. Note, however, that different solutions
to the same equation may approach different equilibria. Here it follows from
Property 3 and the arguments in Example 1 that (as seen in Section 3.2)
all non-negative solutions approach y = K, except for the constant solution
 Nonlinear Differential Equations 191

y(t) = 0, which is already at the zero equilibrium. Figure 4.3 illustrates the
behavior with r = 2, K = 2.
Note also that these last two properties are carefully worded to apply
only to single differential equations, and not to systems of several differential
equations. We shall see in Chapter 5 that solutions to systems of two or more
differential equations need not be monotone, and consequently their bounded
solutions need not approach equilibria.
Before continuing on, we invite the reader to consider how these properties
apply to the other equation familiar to us from Chapter 3: the first-order linear
differential equation y ′ = −ay + b. We will discuss the results in an example
later in this section.

4.1.2 Stability
The question of whether or not solutions approach a given equilibrium
motivates us to study more formally the concept of stability, which has to
do with whether solutions approach (and remain near) a given equilibrium.
In applications, the initial condition usually comes from observations and is
subject to experimental error. For a model to be a plausible predictor of what
will actually occur, it is important that a small change in the initial value not
produce a large change in the solution. For example, the solution of the logistic
differential equation with initial value zero remains zero for all values of t, but
every solution with positive initial value, no matter how small, approaches
the limit K as t → ∞. Because of its extreme sensitivity to changes in the
initial value, we do not ascribe practical significance to the equilibrium zero
as a limiting value for solutions of the logistic equation.
An equilibrium y ∗ of (4.2) is said to be stable if every solution with initial
value close enough to y ∗ remains close to it for all time. The equilibrium
is said to be locally asymptotically stable when every solution with initial
value close enough to y ∗ actually approaches it as t → ∞. The equilibrium
is said to be globally asymptotically stable when every solution approaches it,
regardless of initial condition. If there are solutions which start arbitrarily
close to an equilibrium but move away from it, then the equilibrium is said to
be unstable. For the logistic differential equation (4.3) the equilibrium y = 0
is unstable, and we can see from examining Bands 1 and 2 in Example 1 that
the equilibrium y = K is locally asymptotically stable; in fact, it is globally
asymptotically stable in the positive (y > 0) state space. (In asserting the
global nature of the stability of y = K, we discount the initial condition
y = 0: in applications, unstable equilibria have no significance, because they
can be observed only if the initial condition is “just right.”)
Note that the process for determining global asymptotic stability, which
requires knowledge of other equilibria, the state space, and boundedness of so-
lutions, is distinct from that for determining local asymptotic stability, which
requires only information about what happens very close to one equilibrium
point. Therefore, the standard qualitative analysis process we shall now de-
192 Dynamical Systems for Biological Modeling: An Introduction

velop involves three steps: (1) identifying all equilibria, (2) performing a local
stability analysis on each equilibrium, and (3) assembling the pictures of local
behavior into a single (global) portrait of how solutions may behave in gen-
eral. From Property 1 we already have a methodical way to find equilibria of
a differential equation (4.2): they are the zeroes of the function f (y).
There is also a methodical way to see whether or not a given equilibrium is
locally asymptotically stable or not, beyond testing the sign of f (y) just above
and below the equilibrium (this latter method also does not extend gracefully
to systems of more than one equation, where there are many directions in
which one can be close to an equilibrium). As was the case with discrete-time
models, this way involves the derivative (slope) of the growth function f (y)
at the equilibrium.
Let us look at the piece of the graph of f (y) near an equilibrium y ∗ of
(4.2). By Property 1, f (y ∗ ) = 0; normally this means that the graph of f (y)
crosses the y-axis from one side to the other at y ∗ (in a few unusual cases, f
might turn around and go back the way it came, as in Figure 4.4(c)). We can
determine the direction of the crossing from the sign of the derivative at the
equilibrium, f ′ (y ∗ ). If f ′ (y ∗ ) < 0, then f (y) has a negative slope there, so that,
close to y ∗ , f (y) is positive if y < y ∗ and negative if y > y ∗ (Figure 4.4(a)).
In this case solutions above the equilibrium decrease toward the equilibrium,
while solutions below the equilibrium increase toward the equilibrium. This
shows that an equilibrium y ∗ with f ′ (y ∗ ) < 0 is locally asymptotically stable.
By a similar argument, if f ′ (y ∗ ) > 0, solutions above the equilibrium in-
crease and solutions below the equilibrium decrease, with both moving away
from the equilibrium (Figure 4.4(b)). In this case we see that the equilibrium
is unstable. The only case in which this approach fails to determine an equi-
librium’s stability is when f ′ (y ∗ ) = 0, one example of which is illustrated
in Figure 4.4(c). In such a case (which occurs relatively rarely in practice),
both stability and instability are possible, and we must investigate the sign of
f (y) above and below the equilibrium. If f (y) < 0 above the equilibrium and
f (y) > 0 below the equilibrium, then the equilibrium is locally asymptotically
stable.

y’=f(y) y’=f(y) y’=f(y)

(a) (b) (c)

y y y
y* y* y*
FIGURE 4.4: Equilibria y ∗ with (a) f ′ (y ∗ ) < 0, (b) f ′ (y ∗ ) > 0, (c) f ′ (y ∗ ) = 0.

We can state this test more formally as follows:

 Nonlinear Differential Equations 193

Property 4: Stability. An equilibrium y ∗ of y ′ = f (y) with f ′ (y ∗ ) < 0

is (locally) asymptotically stable; an equilibrium y ∗ with f ′ (y ∗ ) > 0 is
unstable.

We now apply this test on a few examples.

Example 2.
For each equilibrium of the logistic differential equation (4.3), determine
whether it is locally asymptotically stable or unstable.
y
, f (y) = r − 2ry 2y

′

Solution: We have f (y) = ry 1 − K K = r 1 − K . Since
f ′ (K) = −r < 0, the equilibrium y = K is asymptotically stable, and since
f ′ (0) = r > 0, the equilibrium y = 0 is unstable. 

Example 3.
Show that the equilibrium y = 0 of the differential equation y ′ = −y 3 is locally
asymptotically stable.
Solution: f ′ (0) = −3(0)2 = 0, so we cannot apply Property 4 here. However,
for y small and negative we see that −y 3 > 0, while for y small and positive
−y 3 < 0. Thus all solutions approach the limiting value zero, and the equilib-
rium y = 0 is locally asymptotically stable. The behavior of solutions is shown
in Figure 4.5. 
1 1

0.8

0.6
y(t)0.5
y(t)
0.4

0.2 t

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
t
–0.2

–0.4 –0.5

–0.6

–0.8
–1

–1

FIGURE 4.5: Some solutions of the FIGURE 4.6: Some solutions of the
equation y ′ = −y 3 near 0. equation y ′ = −y 2 near 0.
194 Dynamical Systems for Biological Modeling: An Introduction

Example 4.
Show that the equilibrium y = 0 of the differential equation y ′ = −y 2 is
unstable.
Solution: The function −y 2 is negative for all y except y = 0. Thus solutions
of the differential equation y ′ = −y 2 are decreasing. Since solutions with
negative initial value move away from y = 0, the equilibrium y = 0 is unstable.
The behavior of solutions is shown in Figure 4.6. Note that for each of the
differential equations in Examples 3 and 4 we have f (0) = 0, f ′ (0) = 0;
no conclusion about stability can be drawn from f ′ (0) = 0 without further
examination. 

Property 4 is the principal result of this section. Before stating it formally

as a theorem, we shall extend it to consider the special case when the per
capita growth rate is written explicitly in the differential equation.
In many applications, the population’s overall growth rate is written so
as to make the per capita growth rate explicit: that is, the function f (y)
has the form f (y) = yg(y), which has the effect of guaranteeing that y = 0
is an equilibrium. Also, any nonzero equilibrium y ∗ must have g(y ∗ ) = 0.
We calculate f ′ (y) = g(y) + yg ′ (y) and note that at the equilibrium y = 0,
f ′ (0) = g(0), while at any nonzero equilibrium y ∗ , f ′ (y ∗ ) = y ∗ g ′ (y ∗ ) (since
g(y ∗ ) = 0). Thus the equilibrium y = 0 is locally asymptotically stable if
g(0) < 0 and unstable if g(0) > 0; a nonzero equilibrium y ∗ > 0 is locally
asymptotically stable if g ′ (y ∗ ) < 0 and unstable if g ′ (y ∗ ) > 0.
With the logistic equation (4.3), for instance, the per capita growth rate is
g(y) = r(1 − y/K). Thus g(0) = r, so that the equilibrium y = 0 is unstable,
while g ′ (y) = −r/K for all y, so that the equilibrium y = K is asymptotically
stable.

EQUILIBRIUM STABILITY THEOREM: An equilibrium y ∗ of

y ′ = f (y) with f ′ (y ∗ ) < 0 is asymptotically stable; an equilibrium y ∗ with
f ′ (y ∗ ) > 0 is unstable. The equilibrium y = 0 of y ′ = yg(y) is asymptoti-
cally stable if g(0) < 0 and unstable if g(0) > 0, while a nonzero equilibrium
y ∗ is asymptotically stable if y ∗ g ′ (y ∗ ) < 0 and unstable if y ∗ g ′ (y ∗ ) > 0.

We note that this theorem does not cover the case f ′ (y ∗ ) = 0. As we have
seen in Examples 3 and 4 above, a situation with f (y ∗ ) = f ′ (y ∗ ) = 0 must be
examined individually.
Let (y ∗ ) be an equilibrium of a differential equation (4.2), that is, a solution
of the equation f (y) = 0. We assume that the equilibrium is isolated, that is,
that there is an interval around y ∗ that does not contain any other equilibrium.
We shift the origin to the equilibrium by letting y = y ∗ + u, and then use
Taylor’s theorem to approximate f (y ∗ + u). Our approximation is
f (y ∗ + u) = f (y ∗ ) + f ′ (y ∗ )u + h1 = f ′ (y ∗ )u + h (4.4)
 Nonlinear Differential Equations 195

where h is a function that is “quadratic” in u in the sense that it is negli-

gible relative to the linear term in (4.4) when u is small (i.e., close to the
equilibrium).
The linearization of the system (4.2) at the equilibrium y ∗ is defined to be
the linear equation with constant coefficients obtained by omitting the higher
order term h in (4.2).
u′ = f ′ (x∗ )u. (4.5)
If f ′ (x∗ ) < 0, all solutions of (4.5) approach zero as t → ∞, while if
f (x∗ ) > 0, all solutions of (4.5) grow unbounded as t → ∞. There is a
′

theorem stating that if f ′ (x∗ ) 6= 0 all solutions of (4.2) starting sufficiently

close to x∗ behave like the corresponding solution of (4.5). In other words,
if f ′ (x∗ ) < 0, the equilibrium x∗ is asymptotically stable and if f ′ (x∗ ) > 0
the equilibrium x∗ is unstable. We have been able to obtain this result (in
fact a somewhat stronger result) by the arguments described in this section.
However, the arguments used here do not extend to systems of differential
equations and in Chapter 5 we will see how to use the idea of the linearization
to obtain a criterion for determining the asymptotic stability of an equilibrium
of a system.

The final step in the qualitative analysis of a differential equation involves

moving from a local to a global perspective, assembling all the information
about behaviors near each equilibrium to form a picture of how all possible
solutions of the equation behave. As noted earlier in this section, the equi-
libria of an autonomous differential equation (4.2) separate the state space
into distinct intervals, within each of which solutions are either all increasing
or all decreasing. After determining the local stability of each equilibrium,
one must determine whether the model allows any unbounded solutions. If
f (y) is negative for values of y above the largest equilibrium, then no solu-
tions become positively unbounded. If f (y) is positive for values of y above
the largest equilibrium, then this equilibrium is unstable, and solutions with
initial value above this equilibrium become unbounded. If f (y) is negative
for values of y below the smallest equilibrium, then this equilibrium is like-
wise unstable, and solutions with initial value below this equilibrium become
negatively unbounded.
If, within the state space (the set of y values of interest), all solutions
approach the same locally asymptotically stable equilibrium, then that equi-
librium is said to be globally asymptotically stable. If, instead, some solutions
grow unbounded, or there are multiple locally stable equilibria, then there is
no globally asymptotically stable equilibrium, and the end behavior of the
model depends upon initial conditions. In the latter case, accurate measure-
ment of initial conditions is important to the extent that it places solutions on
the correct side of the threshold values which separate one end behavior from
another, and these threshold values are key to understanding the biological
system being studied.
196 Dynamical Systems for Biological Modeling: An Introduction

4.1.3 Phase portraits

As is the case with discrete-time models, it sometimes helps to have a
graphical approach to complement our algebraic analysis tools. Continuous-
time models’ behavior can be described pictorially by graphing the state space
and marking important points and boundaries, including stability and direc-
tion of “flow” of solutions, in what are called phase portraits. Since the state
space of a single differential equation is one-dimensional, we refer to such a
diagram as the equation’s phase line; however, in later chapters we will also
consider phase planes for two-dimensional systems. (Phase portraits for sys-
tems with more than two dimensions are difficult to produce in their entirety,
for obvious reasons.)
A one-dimensional phase portrait consists of the phase line (which may
or may not be infinite in both directions, depending on how the state space
is defined) with the equilibria marked on it, and an arrow on each interval
between equilibria, indicating the direction of change on that interval. To see
the relationship between the phase line and the time-derivative dy/dt, it may
help to begin by graphing the function f (y) and examining the y-axis. An
example is given in Figure 4.7. Note first that the graph of f (y) crosses the
axis at each equilibrium. In between equilibria, the graph of f (y) is either
always above the axis, corresponding to an increasing y(t), or always below
the axis, corresponding to a decreasing y(t). We therefore draw arrows on the
y-axis, to the right where the graph is above the axis, and to the left where
the graph is below the axis. From the directions of the arrows to either side of
an equilibrium we can tell whether the equilibrium is locally asymptotically
stable or unstable; we mark stable equilibria with a solid dot and unstable
equilibria with a hollow dot. After enough experience with this approach, we
no longer need the graph of f (y) superimposed, and can simply draw the
phase line itself. (Note that we can also construct a phase line portrait by
identifying the equilibria analytically, and then substituting a representative
point from each interval to determine the sign of f (y) on that interval.)
2
(a) f(y) (b)
1

-3 -2 -1 1 2
y −2 0 1 y
-1

-2

-3

FIGURE 4.7: A phase line (a) superimposed on the corresponding graph of

f (y), and (b) by itself.

To read a phase line portrait, we think of the phase line as the state space
within which the solution curve y(t) moves, thinking of t as a parameter.
 Nonlinear Differential Equations 197

Thus the solution is described by motion along the line, in the direction given
by the arrows. The graph of Figure 4.7 describes a situation in which there
are asymptotically stable equilibria at y = −1 and y = 2, and an unstable
equilibrium at y = 0.
We illustrate this approach with two examples, the first the familiar linear
differential equation studied in Chapter 3.

Example 5.
dy
Describe the asymptotic behavior of solutions of the differential equation dt =
−ay + b, analytically and with a phase line.
Solution: Here f (y) = −ay + b, f ′ (y) = −a for all y, and the only equilibrium
is y = ab . If a > 0, then any equilibrium is asymptotically stable (since f ′ (y) =
−a < 0), and there are no unbounded solutions (since f (y) < 0 if y is large and
positive, f (y) > 0 if y is large and negative). This means that every solution
is bounded and approaches the limit ab .
If a < 0, however, the equilibrium is unstable, as now f ′ (y) = −a > 0 for
all y. Further, since f (y) > 0 above the equilibrium and f (y) < 0 below the
equilibrium, every solution grows unbounded, either positively or negatively.
Figure 4.8 shows the two corresponding phase lines. 

(a) (b)

b/a b/a
dy
FIGURE 4.8: Phase lines for dt = −ay + b with (a) a > 0, (b) a < 0.

For our second example we consider a model which has been proposed as
an alternative to the logistic model. Recall that the logistic model was origi-
nally conceived to account for the effect of resource limitations on population
growth, by making the per capita growth rate a decreasing function of pop-
ulation size, positive for small y but negative for large y. Another way to do
so is with a per capita birth rate re−y and a constant per capita death rate
d, which yields the differential equation
dy
= y re−y − d .


(4.6)
dt
Since the precise form of the logistic model was derived empirically, and we
want model predictions that are robust in the sense of being tied more to
the qualitative assumptions than to the specific function(s) used, we can now
apply the analysis tools derived in this section to see whether this alternative
model predicts the same kind of population growth as the logistic model.
198 Dynamical Systems for Biological Modeling: An Introduction

Example 6.
Describe the asymptotic behavior of solutions with y(0) ≥ 0 of the differential
equation (4.6), with r, d > 0.
Solution: The equilibria of (4.6) are the solutions of y(re−y − d) = 0. Thus
there are two equilibria, namely y = 0 and the solution y ∗ of re−y = d, which is
y ∗ = log dr . If r < d, then y ∗ < 0, and only the equilibrium y = 0 is of interest.
In this case, f (y) < 0 for y > 0, and solutions with y(0) > 0 decrease to 0
(Figure 4.9). That is, the zero equilibrium is globally asymptotically stable in
the non-negative state space.
If instead r > d, we define K = log dr > 0, so that the positive equilibrium
is y = K, and e−K = d/r.
We may now rewrite the differential equation (4.6)
as y ′ = ry e−y − e−K −y −K


. For the function f (y) = ry e − e , we have
f ′ (y) = r e−y − e−K − rye−y and f ′ (0) = r(1 − e−K ) > 0, implying that the

equilibrium y = 0 is unstable. The equilibrium y = K is asymptotically stable

since f ′ (K) = −rKe−K < 0. All positive solutions are bounded, because
f (y) < 0 for y > K. Thus every solution with y(0) > 0 tends to the limit K
(which is thus globally asymptotically stable in the non-negative state space),
while the solution with initial value zero is the zero function and has limit
zero (Figure 4.10).
Alternatively, using the per capita growth rate form of the equilibrium
stability theorem, we have a per capita growth rate of g(y) = re−y − d, so
that g ′ (y) = −re−y . Then g(0) = r − d, making the equilibrium y = 0 stable
(g(0) < 0) precisely when r < d, while g ′ (K) = −re−K = −d < 0, making the
equilibrium y = K stable (Kg ′ (K) < 0) precisely when r > d (K > 0). 

fHyL

fHyL

y y

FIGURE 4.10: A phase line and

graph of f (y) for (4.6), with r > d.
FIGURE 4.9: A phase line and
graph of f (y) for (4.6), with r < d.

Example 6 shows that solutions of the differential equation (4.6) behave

 Nonlinear Differential Equations 199

qualitatively in the same manner as solutions of the logistic equation. Some

other differential equations which have been proposed as population models
and which exhibit the same behavior are suggested in Exercises 9–12 at the
end of this section. It turns out that one can show that every population model
of the form (4.2) for which the per capita growth rate is a decreasing function
of the population size y, and which is positive for 0 < y < K and negative
for y > K, has the property that every solution with y(0) > 0 approaches the
limit K as t → ∞ (see Exercise 13 below).

4.1.4 Foraging ants and phase transitions

We now return to the model of ant foraging developed by Beekman et
al., who wished to investigate the notion that in order for insect collectives
to function effectively, the population must be larger than some critical size.
In the case of Pharaoh ants, the question was whether the effective use of
pheromone trails to guide fellow ants from the same colony to a given food
source is dependent upon colony size. Their model, given in equation (4.1),
provides a means to determine whether the underlying assumptions imply a
dependence on colony size in terms of the number of ants eventually following
the pheromone trail.
To apply the qualitative analysis techniques developed in this section, we
first write the equilibrium condition dx/dt = 0, or
sx
f (x) = (α + βx)(n − x) − = 0.
x+K
We see immediately that x = 0 is not an equilibrium of this system (since
f (0) = αn > 0: if no ants are initially on a trail to the given food source, some
will nonetheless find the food source and (re)establish the trail). Likewise
the number of ants on the trail cannot increase without bound; as we would
expect, it is in particular bounded by the colony size n (and note f (n) < 0,
and f (x) < 0 for x > n). We can multiply the equation through by x + K to
make it polynomial, obtaining the cubic equation

βx3 + (α + β(K − n))x2 + (s + α(K − n) − βKn)x + (−αKn) = 0 (4.7)

(cf. equation 1 in Beekman et al.). Since f (0) > 0 and f (n) < 0, by continuity
of f there must be at least one equilibrium x∗ in between; we see from the
cubic equilibrium condition that there is in fact either one or three. There is
a cubic formula which can be used to find any equilibria, but the resulting
expressions have many terms and may be difficult to interpret.
Since the expressions for the equilibria are complicated, the first derivative
test given in the equilibrium stability theorem may be difficult to apply in
general. Instead, we shall use the phase line, together with the fact that the
one or three equilibria are simple zeroes of f , to determine their stability. If
there is only one equilibrium, then the phase line is split into two intervals, on
the lower of which f is increasing, and on the upper of which f is decreasing
200 Dynamical Systems for Biological Modeling: An Introduction

(Figure 4.11(a)). The unique equilibrium is therefore locally asymptotically

stable. Since there are no unbounded solutions, the equilibrium is also globally
asymptotically stable.
If, however, there are three equilibria, the phase line is comprised of four
intervals, with the sign of f alternating from one to the next (+, −, +, −;
Figure 4.11(b)). In this case, the first and third equilibria (E1 and E3 ) are both
locally asymptotically stable, while the second (E2 ) is unstable. Here there is
no globally asymptotically stable end state: all solutions with initial value
x(0) < E2 therefore approach E1 , while solutions with initial value above E2
approach E3 . The unstable equilibrium E2 acts as a critical threshold number
of ants using the trail, above which a high proportion of ants uses the trail,
and below which the ants’ foraging behavior remains disorganized.

(a) (b)

0 E n 0 E1 E2 E3 n

FIGURE 4.11: Phase lines for (4.1) with (a) one equilibrium, (b) three
equilibria.

Since the model’s behavior differs significantly depending on whether there

is one or three equilibria, we can use a quantity called the cubic discriminant
∆ to determine which is the case (this can be found in references on solving
cubic equations). If ∆ < 0, there are three equilibria; if ∆ > 0, there is one.
For (4.7) the discriminant is

∆ =4β[s + αK − (α + βK)n]3 + 27α2 β 2 K 2 n2

− [(α + βK) − βn]2 [s + αK − (α + βK)n]2 − 4αKn[(α + βK) − βn]3
+ 18αβKn[(α + βK) − βn][s + αK − (α + βK)n].

Although this quantity is complicated, we can investigate two specific as-

sertions made in Beekman et al.: first, that the more complicated behavior
(with three equilibria) occurs when food sources are difficult for ants to locate
without a pheromone trail (i.e., when α is small), and second, that pheromone-
based foraging is more efficient in larger colonies. If we set αp= 0, then
∆ = 4β[s − βKn]2 [s − β(K + n)2 /4], so that ∆ < 0 whenever n > 4s/β − K.
Beekman et al. measured the actual value of α as 0.0052/min. and (in Sumpter
and Beekman) used values for the other parameters of β=0.00125/ant-min.
(an average of two values
p used), s = p 1 ant/min., and K = 10 ants. This gives
a threshold of about 4s/β − K = 4/0.00125 − 10 = 57 ants when α = 0,
and a threshold of about 628 ants for the measured value of α. The thresh-
old colony size of 57 ants for α = 0 is so small that we can verify that the
more complicated behavior occurs when food sources are sufficiently difficult
to find, and the threshold colony size of 628 ants for the measured value of
α agrees with Beekman et al.’s experimental results, which found a marked
 Nonlinear Differential Equations 201

improvement in foraging organization for colony sizes above 600 ants. We can
also observe that if we expand the expression for ∆ as a polynomial in n,
the lead term is −β 2 (α − βK)2 n4 , implying that for sufficiently large colonies
∆ < 0 and the more complicated behavior occurs.
We can also address the hypothesis of more efficient foraging (i.e., a higher
proportion of ants using the trail) in larger colonies by dividing the equilib-
rium condition (4.7) by n3 in order to rewrite it in terms of the equilibrium
proportion y = x/n of ants using the trail:
   
α + βK s + αK α + βK αK
βy 3 + − β y2 + 2
− y − 2 = 0.
n n n n

If we let n → ∞, so that 1/n → 0, then this equation simplifies to βy 3 −

βy 2 = 0, the solutions of which are y = 0 (twice) and y = 1. Thus as colony
size increases, the upper equilibrium approaches 1 (all ants are on the trail),
while the middle (unstable threshold) equilibrium approaches 0, making the
upper equilibrium nearly globally asymptotically stable (since very few initial
conditions will be less than E2 ).
Therefore in general this model predicts the existence of a critical threshold
(E2 ) in pheromone trail usage when food sources are difficult enough to find,
and when a colony is large enough (roughly 700 ants or more); in addition,
as colony size increases, the initial proportion of the colony’s ants required
to establish the trail (in order for it to become well used) decreases, and the
proportion of the colony’s ants who will then use the trail increases toward 1.
The transition from one to three equilibria marks a major change in the
global behavior of this model, and we will return to this equation later in this
chapter, when we shall study the topic of bifurcations more generally, both
mathematically and in terms of their biological consequences.

Exercises
For each of the differential equations in Exercises 1–8, draw the phase line,
find all equilibria and describe the behavior of solutions as t → ∞.

1. y ′ = y 5. y ′ = y(1 − y)(2 − y)
2. y ′ = −y 6. y ′ = −y(1 − y)(2 − y)
3. y ′ = y 3 7. y ′ = y 2 (y + 1)
4. y ′ = sin y 8. y ′ = −c(y − 1)2

For each of the differential equations in Exercises 9–12, describe the behavior
of solutions with y(0) > 0.

9. y ′ = ry log K
y
202 Dynamical Systems for Biological Modeling: An Introduction
 
10. y ′ = ry −1 + Ky

ry(K−y)
11. y ′ = K+Ay
h i
y θ
12. y ′ = ry 1 −


K

∗
13. Consider a differential equation y ′ = g(y) with g(y) = yh(y), where
h(y) is a decreasing function of y, so that h′ (y) < 0, and h(y) > 0 for
0 < y < K, but h(y) < 0 for y > K. Show that the only equilibria are
y = 0 and y = K, and that the equilibrium y = 0 is unstable while the
equilibrium y = K is asymptotically stable.
∗
14. (a) Show that if y(t) is a solution of an autonomous differential equa-
tion y ′ = g(y), then y ′′ (t) = g ′ {y(t)} g{y(t)}. (b) Deduce from part (a)
that if the solution y(t) has an inflection point, this inflection point must
occur at a point (t, y) with g ′ (y) = 0.
15. In this section, we analyzed the long-term behavior of Beekman et al.’s
model for ants following a pheromone trail to a single food source. In
what sense is this behavior long-term, i.e., what is the time scale implicit
in this discussion? How does this time scale relate to the assumption that
the colony size is constant?
16. How would the ant foraging model (4.1) change if
(a) the ants did not use pheromone trails?
(b) ants using the trail never left it?
(c) ants using the trail left it at a constant per capita rate σ?
In each case, indicate how the resulting model’s behavior differs from
that of Beekman et al.
17. How would the interpretation of the rate at which ants leave the trail
be affected if it were modeled by
 p
x
s ,
x+K

for some number p > 1?

18. In practice, why is it advantageous for some ants using the trail to stop
using it (before the food source is exhausted)?
 Nonlinear Differential Equations 203

4.2 Harvesting
In Section 2.5 we studied harvesting, or population management, in
discrete-time models, where both reproduction and harvesting occur at spe-
cific discrete times. If the harvesting events are frequent enough, they can be
considered to be nearly continuous and ongoing, in the same way that we con-
sider ongoing reproduction and mortality for many biological systems. In this
section we will look at the management of populations from a continuous-time
perspective, including not only such activities as hunting, fishing and logging,
but also production of bacteria in laboratory settings and even the dispersal
of human populations.
In our previous study of population harvesting we considered primarily
constant-yield harvesting, a practice in which members of a population are
removed at a constant rate. Constant-yield harvesting models3 are appropriate
for systems in which there is a set quota or deliberate, controlled management
of the population. In other situations, however, such as the notorious bycatch
of dolphins in the tuna industry in the late 20th century that eventually led to
changes in fishing practices, harvesting may be either an incidental byproduct
of an activity not directly designed to control that population, or a process
with a duration or intensity set by factors other than the population size: for
instance, occasional fishing in a small lake or pond, where fishermen go a given
number of times, and the number of fish caught depends upon the availability
(population density) of fish in the pond. These situations are better modeled
by constant-effort harvesting, in which members of the population are removed
at a rate proportional to the population size.4 In this section we shall consider
biological systems of both types.

4.2.1 Constant-yield harvesting

Population harvesting affects populations of every size, from microscopic
organisms to blue whales, the largest animals on Earth. Constant-yield har-
vesting occurs when a fixed quota is set for removing members of the popula-
tion, as often occurs in hunting. In general, if a population’s natural growth
without harvesting is described by the model y ′ = f (y), then under constant-
yield harvesting it is modeled by the differential equation

y ′ = f (y) − H, (4.8)
3A primary reference is Fred Brauer and David A. Sánchez, Constant rate population
harvesting: equilibrium and stability, Theor. Pop. Biol. 8(1): 12–30, August 1975.
4 A primary reference is M.B. Schaefer, Some aspects of the dynamics of populations

important to the management of commercial marine fisheries, Bull. Inter-Amer. Trop. Tuna
Comm. I: 25–56, 1954.
204 Dynamical Systems for Biological Modeling: An Introduction

where H > 0 is the constant harvest rate (yield per time) at which members are
removed from the population. (Note that the units of H here are population
per time, whereas in discrete-time harvesting models it had simply population
units.) We assume that the rate of removal is uniform (not seasonal as in many
types of hunting).
Although with a few reasonable assumptions we can prove results describ-
ing the behavior of any constant-yield harvesting model (4.8), for illustrative
purposes we will assume that the populations under study in this section obey
a logistic law, so that f (y) = ry(1 − y/K). Then the population size satisfies
a differential equation of the form
 y
y ′ = ry 1 − − H, (4.9)
K
where r, K, and H are positive constants. While this equation can be solved
explicitly by separation of variables, the integration must be handled by ex-
amining three different cases depending on the values of the constants (see
Exercise 1). It is simpler (and probably more informative) to carry out a
qualitative analysis.
r 2
The equilibria of (4.9) are the solutions of ry − K y − H = 0, or

HK
y 2 − Ky + = 0. (4.10)
r
These are given by " r #
1 4HK
y= K ± K2 − . (4.11)
2 r
We must distinguish three cases:
4HK rK
(i) K 2 − r > 0, i.e., H < 4 ,
4HK rK
(ii) K 2 − r = 0, i.e., H = 4 ,
4HK rK
(iii) K 2 − r < 0, i.e., H > 4 .

The number and nature of the equilibria are different in the three cases.
In case (i) there are two distinct equilibria given by (4.11), and it is clear
that one, which we shall call y1 , is smaller than K 2 while the other, which
we shall call y2 , is larger than K 2 . If we let f (y) = ry − Kr 2
y − H, then
2r 2r K K
f (y) = r − K y = K 2 − y . Thus f (y) > 0 if y < 2 , and f ′ (y) < 0
′ ′

if y > K 2 . We now see, using the equilibrium stability theorem, that any
equilibrium y ∗ of (4.9) with y ∗ > K/2 is asymptotically stable, while any
equilibrium y ∗ with y ∗ < K/2 is unstable. More specifically, in case (i), we
see that y1 is unstable while y2 is asymptotically stable.
To consider the global picture for case (i), we note that since f (y) < 0 for
0 < y < y1 , any solution y(t) with 0 < y(0) < y1 is monotone decreasing and
reaches zero in finite time. When this happens, we consider the population to
 Nonlinear Differential Equations 205

have been wiped out and the model to have collapsed. If instead y(0) > y1 ,
the solution y(t) approaches the limit y2 as t → ∞. This is illustrated in
Figure 4.12 with r = 2, K = 2, H = 95 .
In case (ii) there is a single equilibrium K
2 which is a double root of (4.10),
′ K
and f ( 2 ) = 0. The equilibrium stability theorem does not apply here, but if
we rewrite f (y) as
   2
r 2 rK r 2 rK r K
f (y) = ry − y − =− y − ry + =− y− ,
K 4 K 4 K 2

we see that f (y) < 0 if y 6= K2 . Thus every solution is monotone decreasing

(except the equilibrium solution), and solutions starting above K 2 approach
the equilibrium K2 , while solutions starting below K
2 reach zero in finite time.
This behavior is illustrated in Figure 4.13 with r = 2, K = 2, H = 1.
In case (iii) the roots of (4.10) are complex, so (4.9) has no equilibria.
Since f (y) has no real zeroes and f (0) = −H < 0, f (y) < 0 for all y. Thus
every solution is monotone decreasing and reaches zero in finite time; the
population is wiped out no matter what the initial population size is, as shown
in Figure 4.14 with r = 2, K = 2, H = 1.5.
Phase portraits for all three cases are shown in Figure 4.15.
2 2 2

1.8 1.8 1.8

1.6 1.6 1.6

1.4 1.4 1.4

1.2 1.2 1.2

y(t) 1 y(t) 1 y(t) 1

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0.5 1 1.5 2 2.5 3 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 0 0.5 1 1.5 2 2.5 3
t t t

FIGURE 4.12: FIGURE 4.13: FIGURE 4.14:

Solutions to (4.9), Solutions to (4.9), Solutions to (4.9),
with H < rK/4. with H = rK/4. with H > rK/4.

If we now compare the three cases, we can understand the overall effect
of constant-yield harvesting on a population with logistic growth. Recall that
with no harvesting (H = 0) the population approaches the carrying capacity
K regardless of its initial size. As H increases from 0 to rK 4 , the limiting
population size y2 decreases monotonically from K to K/2, while at the same
time the minimum population size y1 required for survival increases from 0
toward K/2. That is, low harvest rates affect the population in two ways:
206 Dynamical Systems for Biological Modeling: An Introduction
(i) (ii) (iii)

0 y y K 0 K/2 K 0 K
1 2

FIGURE 4.15: Phase lines for (4.9) in the three cases given in the text. Note
the “semi-stable” equilibrium in case (ii).

by lowering its long-term size and by requiring a minimum size for survival.
As the harvest rate increases, the long-term size and minimum size approach
each other, narrowing the window for survival. When the harvest rate reaches
the critical value Hc = rK 4 , these two values meet at K/2, half the original
carrying capacity.
However, for harvest rates beyond the critical value (H > rK 4 ), the pop-
ulation is wiped out in a finite amount of time. Thus the limiting population
size decreases continuously (from K to K 2 ) as H increases from 0 to rK/4,
and then drops to zero as H passes through the critical value rK/4. Such a
discontinuity in the limiting behavior of a system is called a (mathematical)
catastrophe. The corresponding biological catastrophe is the extinction of the
population.
The implications of this model for the management of populations under
a quota system are entailed in the critical harvest rate rK/4. In cases such as
fishing and logging where harvesting is tied to industry, the economic goal of
maximizing the harvest rate must be tempered by an awareness of the fragility
created by harvesting. In this case, the harvest rate must be held sufficiently
far below the critical rate to ensure that occasional or seasonal fluctuations
in reproduction or mortality do not cause the population to drop below the
minimum survival size.
As suggested at the beginning of this section, the logistic growth model
is representative of a wide range of population growth functions. It turns out
that for any general population model of the form y ′ = f (y) whose per capita
growth rate is a monotone decreasing function of population size, and positive
for small y but negative for large y,5 the response to constant-yield harvesting
is similar. (The per capita growth rate in the logistic model, r(1 − y/K), fits
this description.) For such models, with initial conditions close to the natural
carrying capacity, we can show (see Exercise 3d for a special case) that the
limiting population size is positive when H is less than a critical harvest rate
Hc , but drops to zero as H passes Hc . We observed similar behavior for discrete
models in Section 2.5, but for discrete models the situation is complicated by
the possibility that all equilibria may be unstable.
Note that for the logistic model, the critical harvesting rate Hc = rK/4
is the maximum value of the function f (y) = ry(1 − y/K). This is not a
5 Note that these conditions are similar to our definition of scramble competition for

discrete-time models in Chapter 2.

 Nonlinear Differential Equations 207

Photo courtesy Wisconsin Dept. of Natural Resources Graph courtesy Wisconsin Dept. of Natural Resources

FIGURE 4.16: A white-tailed buck. FIGURE 4.17: Wisconsin white-

tailed deer population.

coincidence; in general Hc = maxy f (y), as can be seen graphically (see again

Exercise 3d).
We will now consider some case studies to which our constant-yield har-
vesting model can be applied.

Example 1.
Wisconsin is one of many states in the U.S. where human expansion has
removed most of the predators of the deer native to the area, with the result
that the deer population grows unchecked, until deer frequently appear on
roads and in yards, causing accidents and damage to gardens and shrubs.
As a result, the Wisconsin Department of Natural Resources (WDNR) issues
hunting permits for a fixed number of deer to be removed each hunting season,
in order to manage the burgeoning population.
The Wisconsin Department of Natural Resources (WDNR) first estab-
lished population goals for white-tailed deer (Figure 4.16) in 1962. As deer
range expanded and hunting interest increased, the goal has grown, to the
point that in the early 21st century it stands at about 700,000. As can be seen
in Figure 4.17, the deer population has continued to grow steadily nonethe-
less; beginning in the mid-1990s, winter populations have been about 50% over
goal. In the 2006–2007 season, just over 500,000 deer were hunted, but by the
fall of 2007, the population had reached about 1.7 million deer, in keeping
with the reproductive rate of about 0.5/yr which can be estimated from the
graph.
(a) What is the minimum equilibrium population value we can infer from
the continued growth shown in the graph together with these values for
H and r?
(b) If hunting half a million deer per year eventually keeps the population
around 3 million deer, how many deer would there be with no hunting?
(This question requires many unrealistic simplifying assumptions, such
as keeping the deer within a fixed habitat.)
208 Dynamical Systems for Biological Modeling: An Introduction

(c) Using this value of K, what are the critical harvesting rate and resulting
equilibrium level?
(d) Using this value of K, what harvesting rate H would meet the WDNR
goal of 700,000 deer?

Solution:
(a) From the graph, we can see that the population keeps growing even with
1/2 million deer harvested per year. Therefore, harvesting is presently
below the critical rate, that is, H < rK/4. Solving for K, we have K >
4H/r = 4(0.5)/0.5 = 4 million deer as a minimal carrying capacity. From
(4.11), the stable equilibrium must be greater than K/2 = 2 million deer.
(b) Solving the equilibrium condition (4.11) for K, we get (after some alge-
bra)
y ∗2
K= ∗ H.
y − r
Substituting r = 0.5/yr, H = 0.5M deer/yr, y ∗ = 3M deer yields a value
of K = 4.5M deer.
(c) Hc = rK/4 = 0.5625M deer/yr, and the equilibrium there is K/2 =
2.25M deer.
(d) As mentioned in the answer to (a), the constant-yield harvesting model
(4.9) cannot give a stable equilibrium value lower than K/2. Of course,
our estimates for r and y ∗ are very rough, but in general the true carry-
ing capacity must be less than twice the goal, in order for the goal to be
reachable through constant-yield harvesting. Since present pre-hunt pop-
ulations are more than twice the present population goal, the goal can
only be reached with variable yield harvesting: that is, harvesting that
depends upon the population size. WDNR continues to adjust their har-
vest quotas annually; we will consider one type of population-dependent
harvesting later in this section. 

The great whales once numbered in the millions, but whaling emerged as
a major industry beginning in the eighteenth century, hunting whales of sev-
eral types for their oil and other products, and by the late twentieth century
many species had been reduced to just a few thousand members each (see Ta-
ble 4.16 ). As a result, in 1986 the International Whaling Commission (IWC)
implemented a moratorium on commercial whale hunting, and began a careful
study of the different species’ ability to recover, while setting guidelines for
6 Data from G. A. Knox, The key role of krill in the ecosystem of the southern ocean with

special reference to the convention on the conservation of Antarctic marine living resources,
Ocean Management 9: 113–156, 1984.
 Nonlinear Differential Equations 209

TABLE 4.1: Population data (in thousands) for five species of whales prior
to the 1986 IWC moratorium, from Knox, 1984.
Species \ Year 1900 1910 1920 1930 1940 1950 1960 1970 1980
Fin 400 380 360 320 275 220 140 90 80
Blue 200 185 150 95 30 20 10 7 3
Humpback 100 95 70 50 35 20 10 5 3
Sei 95 90 87 83 80 76 70 50 20
Minke 25 25 25 30 34 38 65 96 100

future whale hunting quotas. At present, no commercial whaling is allowed

(aboriginal groups such as those in Alaska continue their traditional subsis-
tence whale hunting, but under fixed quotas); this ban is to remain in effect
until species have recovered to at least 54% of their pre-whaling populations,
which will probably take at least 50–100 years. Norway and Japan continue to
harvest whales, Norway by declaring its own quotas for commercial whaling
in exception of the IWC moratorium, and Japan by using a loophole in the
regulation which allows limited whale hunting for scientific research purposes.
Both countries hunt minke whales, now one of the most numerous species,
Norway harvesting an average of 577 per year from 1997 to 2006 and Japan
about 300 per year.
Biologists worldwide are hoping the damage done can be reversed. Com-
plex chain reactions are already occurring as a result of the overharvesting
of whales: Over-harvesting of great whales (1946–1979), a major food source
for killer whales, forced killer whales to hunt smaller marine mammals: har-
bor seals (1970s), fur seals (mid-70s to mid-80s), sea lions (1980s–1990s), and
finally sea otters (1990s–present). Sea urchins, a major food source for sea
otters, then grew unchecked, decimating the forests of kelp eaten by the sea
urchins.

Example 2.
The International Whaling Commission decided in 1965 to protect blue
whales, whose population decreased to fewer than 2000 before the moratorium
was enacted, in spite of an estimated carrying capacity of 200,000 whales and
intrinsic growth rate of 0.05/yr, as a result of annual catches averaging 13,000
whales per year over the period 1926–1940. (a) Assuming that blue whales
obey a logistic law, estimate the maximum number of whales that can be
caught per year without wiping out the population. (b) The IWC estimated a
population of 2,300 blue whales (excluding pygmy blues) in the southern hemi-
sphere in 1997–1998, and a growth rate of 0.082/yr. If we make the simplistic
assumption that the southern hemisphere’s carrying capacity for blue whales
is half that for the entire planet, how long will it take before the population
reaches the IWC’s goal of 54% of carrying capacity using their figures?
Solution: (a) The critical harvest rate is rK/4 = (0.05)200, 000/4 = 2500
whales/yr, less than one fifth the historical catch rate cited.
210 Dynamical Systems for Biological Modeling: An Introduction

(b) With no harvesting, we assume the population obeys a simple logistic

law, for which equation (2.13) gives the solution as a function of time. Solving
this equation for time yields
  
1 K − y0 K
tf = ln −1 .
r y0 y(tf )

Substituting r = 0.082/yr, K = 100, 000 whales, y0 = 2300 whales, and

y(tf ) = 0.54K gives tf ≈ 48 years. Under this prediction, southern hemisphere
blue whales could reach IWC goal levels by 2046. 

Example 3.
Another species of whale hunted to near-extinction in the early twentieth cen-
tury is the right whale,whose post-moratorium fate in the two hemispheres
has been quite different. Southern right whales off the coast of Argentina
numbered about 7500 at the end of the twentieth century, according to the
IWC, and reproduction rates of approximately 0.07/yr have been cited. They
numbered about 100,000 a century earlier. However, the North Atlantic right
whale is now extinct in the eastern Atlantic, and numbered about 300 in the
western Atlantic as of the end of the twentieth century. In the first fifteen years
of the IWC moratorium, their mortality rate increased more than fivefold, al-
though they too numbered about 100,000 a century before. This population
has been observed to be declining at a rate of 0.024/yr.7 (a) If some southern
right whales were to be transported somehow to bolster the northern popu-
lation, what removal rate could the present southern population sustain and
still make progress toward carrying capacity? (b) If we suppose the northern
right whales to be in exponential decline at the given rate (a logistic law makes
no sense in this case) and add whales from the southern hemisphere at the
rate identified in (a), what will the expected equilibrium value be?
Solution: (a) In order for removal not to threaten the southern population,
the present population level must be above the lower equilibrium y1 in (4.11).
If we set y1 = 7500 whales, r = 0.07/yr, and K = 100, 000 whales and solve
for H, we get H = 485 whales/yr.
(b) Exponential decline with importation at a constant rate yields the
mixing equation y ′ = −ay + b studied in Section 3.5, the equilibrium value for
which is b/a. Taking b = 485 whales/yr and a = 0.024/yr yields b/a = 20, 200
whales, a large enough population that one might hope its decline could be
reversed. 

The plan suggested in the above example is overly simplistic for many rea-
sons but illustrates the strategy of importing members of related populations
to save endangered populations that has been used successfully with some
7 Hal Caswell, Masami Fujiwara, and Solange Brault, Declining survival probability

threatens the north Atlantic right whale, Proceedings of the National Academy of Sciences
USA 96(6): 3308–3313, March 16, 1999.
 Nonlinear Differential Equations 211

Photo courtesy John Calambokidis, Cascadia Research Photo courtesy NOAA

FIGURE 4.18: An aerial view of FIGURE 4.19: A right whale.

two blue whales.

other species, especially birds. The failure of some species, like the northern
right whale, to recover following the cessation of decades of overharvesting is
not yet understood, and remains the subject of intense study. One possibility
is a lack of sufficient diversity (genetic and otherwise) in small populations,
but there are many other, behavioral aspects of these populations’ lives that
have been completely disrupted and are not well-known. Some dolphin species
are suffering a similar problem, as we shall see later in this section. A few other
whale populations are considered in the exercises at the end of this section.

Logging is another major industry, producing both wood and paper for uses
around the world. This type of harvesting has great impact on the world’s
forests. For example, clear-cutting—a technique in which all the trees in a
given area are cut down, leaving bare ground (see Figure 4.20)—has destroyed
many primary (old-growth) forest habitats. In population terms, clear-cutting
has the effect of reducing the area of wooded land and thus the carrying
capacity of the system. However, in some places lumber and paper companies
replant trees on the lands where they cut, so that, with many such areas
in a given region, each in a different stage of regrowth, the total wooded
area remains constant, even though the harvesting is ongoing. In the United
States, conservation efforts are holding the amount of forested land roughly
constant. About one third of the U.S., or 303 million hectares,8 is forested
(including Alaska). Of this, just over one third (104.2 million hectares) is
classified as primary forest, the most biodiverse form of forest. In general, the
world’s temperate forests (including those of the U.S. and Canada) are being
roughly conserved. However, the world’s tropical forests, almost all of them
in developing countries, are disappearing at a rate of 8% per decade, while
global wood-consumption is growing at 26% per decade.
Logging yields depend on the intensity of forest management. Intensive
management (used only in Europe as of the end of the twentieth century)
8A hectare is 10,000 m2 .
212 Dynamical Systems for Biological Modeling: An Introduction
MSY
6
intensive
5.5 production
5
4.5
4
3.5 consumption semi-intensive
production
t
2 4 6 8 10 12 14
Photo by Steve Hillebrand, courtesy U.S. FWS
FIGURE 4.21: Graphs of estimated
FIGURE 4.20: This photo of a global wood production vs. consump-
clear-cut hill in Oregon was taken in a tion, in 109 m3 /yr vs. years since 1984,
study of habitat loss for the northern in Example 4.
spotted owl. .

TABLE 4.2: Estimated maximum sustainable stemwood yield for the world’s
forests as of 1984, assuming intensive management and, for tropical forests, a
75-year rotation cycle.
Area Productivity MSY
Forest type (106 km2 ) (m3 /km2 /yr) (109 m3 /yr)
Cold temperate closed 10.56 180 1.91
Cold temperate open 3.09 27 0.09
Moderate temperate closed 3.88 280 1.09
Moderate temperate open 1.73 42 0.08
Tropical closed 11.9 207 2.46
Tropical open 7.3 31 0.23
Total 38.46 152 5.86

involves gradual complete replacement of all trees in a given region on a fixed

cycle under high-quality supervision, and implies the eventual replacement
of nearly all old growth. Non-intensive management may preserve some old
growth but generally gives sustainable yields about half that of intensive man-
agement. Yields per area also depend on the density of trees: closed forests
have dense woods, with trees as close together as natural growth allows, while
open forests are sparser. Optimal harvest-cycles are 3–4 decades for wood pulp
and firewood, and 9–12 decades for sawtimber. Shorter harvest cycles reduce
sustainable wood-yield. Table 4.2 gives one published estimate for maximum
sustainable yields of the world’s forests as of 1984. The maximum sustainable
yield (MSY) is the maximum harvest rate which can be sustained without driv-
ing the population in question to extinction. In the table, temperate forests
are divided into two classes, as those in harsher climates (Canada and the
former USSR) have historically had lower yields.
 Nonlinear Differential Equations 213

Example 4.
If we make the wildly unrealistic assumption that all the world’s forests began
intensive management in 1984, and take the conservation rates and consump-
tion growth rates given in the text, along with an estimated 1984 global wood
consumption rate of 3.67 × 109 m3 /yr, in what year would simple exponen-
tial models predict consumption to outpace production? What if we assume
intensive management of only temperate forests?
Solution: An exponential model for production under the intensive maximum
sustainable yield levels in Table 4.2 would give

M SY (t) = 3.17 + 2.69(0.92)t

where t is measured in years since 1984 and MSY in 109 m3 /yr; the first (con-
stant) component comes from temperate woods and the second (exponential)
from tropical woods, shrinking at the rate of 8% per year. Consumption, mean-
while, can be calculated in the same units as 3.67(1.26)t/10 (growth at the rate
of 26% per decade). Although we cannot set these two expressions equal and
solve for t in closed form, by graphing both quantities as functions of time we
can easily find the point at which they intersect, which occurs for t = 8.63,
that is, in 1992 or 1993.
If we instead assume intensive management of only temperate forests, pro-
duction is M SY (t) = 3.17 + 1.35(0.92)t. In this case, consumption outpaces
production at t = 4.65, i.e., in 1988 or 1989. Both production curves are
graphed vs. consumption in Figure 4.21. Of course, without intensive man-
agement consumption is already beyond production at t = 0. 

Example 5.
Brazil is home to one of the largest tropical forests in the world. 57%—or about
477.7 million hectares (Mha)—of Brazil was forested as of 2005. Between 1990
and 2005, Brazil lost 8.1% of its forest cover, or around 42.33 million hectares.
This land was largely clear-cut for agricultural or other non-logging reasons, so
that the habitat is effectively lost. Suppose, however, that in 2005 developed
countries began providing subsidies to protect and restore the remaining forest
land. If individuals, meanwhile, continued clear-cutting at the same rate (by
1980, over 72% of Brazil’s forest clearing was due to ranching), what effective
inherent growth rate r would be necessary in order for replanting to save the
forests?
Solution: The average harvesting rate is H = 42.33Mha/15yr= 2.822 Mha/yr.
In order for the constant-yield harvesting model to have an equilibrium, we
must have H < rK/4, from which r > 4H/K = 4(2.822)/477.7 = 0.02363/yr.
At this rate, however, the equilibrium forest area is only half its 2005 size.
This solution also ignores issues of old (primary) vs. new growth. 
214 Dynamical Systems for Biological Modeling: An Introduction

Example 6.
A less controversial example of harvesting is the use of laboratory bacteria
cultures to provide a source of bacteria for use in experiments. If a given
culture of bacteria doubles in 12 hours in a nutrient dish capable of sustaining
106 bacteria, what is the maximum sustained rate at which bacteria can be
removed from the dish? How would the answer change if we assume simple
exponential growth rather than logistic?
Solution: The inherent reproduction rate r can be written either as
(ln 2)/12hr = 0.058/hr or as (ln 2)/0.5day = 1.39/day, using the methods
of Section 3.1. The MSY under the logistic constant-yield harvesting model is
rK/4 = (1.39)106 /4 = 347, 000 bacteria/day. 

An exponential constant-yield harvesting model would have

y ′ = ry − H,

whose only equilibrium is H/r. However, stability analysis of this

equilibrium—with f (y) = ry − H, f ′ (y) = r > 0—shows that it is unsta-
ble. This means that, for a given harvest rate H, if the initial condition is
greater than H/r, the population will grow unbounded (just as in Malthus’s
unharvested exponential growth model), while if the initial condition is less
than H/r, the harvesting will cause the population to crash (reach zero in fi-
nite time). In either case, the model becomes unrealistic at some point in time,
just as the Malthus model does. Also, there is no maximum harvesting rate,
if one is willing to wait long enough before beginning harvesting. If we define
harvesting to begin at time zero, then we must have H < ry(0). In an envi-
ronment with limited resources, of course, the initial condition is limited by
the environment’s carrying capacity, but the exponential model cannot reflect
this limitation. This criterion is identical to that for the logistic harvesting
model if we take y(0) = K/4, at which point (in the logistic model) growth
has already been reduced to 43 r. These ambiguities are why we do not use the
above exponential harvesting model.

4.2.2 Constant-effort harvesting

In constant-effort harvesting, members are removed from a population by
a process dependent upon both effort E (a rate or frequency with units of
1/time) and the population size y, under the assumption that the larger the
population, the more often members of it will be removed through this process.
For example, if a fisherman goes fishing on a private pond a certain number of
times per year, and each trip lasts the same length of time, the number of fish
he catches each year may be taken as dependent upon the pond’s population,
rather than upon his effort, which we assume to be fixed. Alternatively, if a
certain type of animal eats a farmer’s crop, the farmer will attempt to kill
or remove the animal when he sees it in his fields; he does not set out with
 Nonlinear Differential Equations 215

the idea of finding a certain number of animals, but the more of them there
are in his fields, the more often he will find them. The key distinction from
constant-yield harvesting is that here the process is not tied to a harvesting
quota; instead, the frequency of encounters is assumed to be proportional to
the population size (an idea central to the notion of contact processes, which we
shall investigate in the next section). For a population governed by the model
y ′ = f (y), constant-effort harvesting produces the equation y ′ = f (y) − Ey.
Once again we shall consider the specific case of a population governed by
a logistic law. In this case the model becomes
 y
y ′ = ry 1 − − Ey. (4.12)
K
Clearly y = 0 is one equilibrium of this equation; the other we calculate to be
 
∗ E
yH = K 1 − . (4.13)
r
Note that this second equilibrium is only biologically meaningful (i.e., positive)
if E < r, that is, if harvesting events occur less frequently than reproduction.
If instead E > r, then the zero equilibrium is the only equilibrium.
To determine local asymptotic stability of the equilibria, we compute the
derivative of the right-hand side of (4.12) and use the criterion
2y ∗
 
r 1− − E < 0,
K
which simplifies to  
1 E 1 ∗
y∗ > K 1− = yH .
2 r 2
If 0 ≤ E < r, then this criterion is clearly true for the positive equilibrium
∗ ∗
yH and false for the zero equilibrium. If, however, E > r, then yH < 0, so
that the stability criterion is true for the zero equilibrium (and false for the
∗
negative equilibrium yH ). In both cases, the local asymptotic stability extends
to global asymptotic stability in the non-negative state space, since there are
no unbounded solutions (f (y) is negative for large y). Figure 4.22 gives phase
line diagrams for both cases.

E<r E>r

0 y* K 0 K
H

FIGURE 4.22: Phase lines for (4.12) in the two cases given in the text.

Constant-effort harvesting therefore reduces the size of a logistic popula-

tion by a proportion E/r of the carrying capacity, unless E ≥ r, in which case
216 Dynamical Systems for Biological Modeling: An Introduction

it gradually drives the population extinct. The extinction in the latter case
is gradual rather than sudden (as with constant-yield harvesting) because as
the population becomes scarcer, its members become harder to find with the
same level of effort. (One might argue that when the population becomes very
small, its members may become too difficult to find to guarantee a constant
yield. Exercise 15a at the end of this section suggests a way to address this
scenario.)
We may wish to compare the maximum sustainable yield (MSY) for
constant-effort harvesting with that obtained for constant-yield harvesting
(rK/4 for a population with logistic growth). The maximum sustainable ef-
fort is a little below r, but recall that the actual harvest rate is Ey rather
∗
than E. We therefore find the maximum of EyH over all possible efforts:

E2
   
∗ E
M SY = max EyH = max E K 1 − = max K E − . (4.14)
E E r E r

To find the maximum of this last expression with respect to E, we set its
derivative to zero:  
2E r
K 1− =0 ⇔ E= .
r 2
Since the MSY is a downward-facing quadratic in E (to see this, just graph
it vs. E), we know that this unique critical point is indeed a maximum. If we
substitute E = r/2 into (4.14), we obtain

(r/2)2
 
r r r  rK
M SY = K − =K − = .
2 r 2 4 4

The reader who is surprised to find the MSY for constant-effort harvesting to
be the same as that for constant-yield harvesting should note that, given a
constant effort E, the population y will eventually settle down to the corre-
∗ ∗
sponding equilibrium yH , so that the product EyH which represents the yield
is asymptotically constant as well. The difference in practice is that constant-
yield harvesting is often managed to obtain the MSY, while constant-effort
harvesting typically describes scenarios in which the effort is fixed by other
considerations than the MSY.
We now consider two applications of constant-effort harvesting to ecology.

A striking example of incidental harvesting that became notorious in the

late twentieth century was the bycatch of dolphins in the commercial tuna
fishing industry. In the late 1950s, the method of purse-seine fishing was de-
veloped to catch yellowfin tuna (Thunnus albacares) associated with schools
of dolphin in the eastern tropical Pacific Ocean. Although this method was
intended to allow dolphins to escape unharmed while trapping the tuna, over
the next four decades this approach resulted in the deaths of over six million
dolphins, the highest known for any marine harvest. In comparison, the total
 Nonlinear Differential Equations 217

Photo courtesy NOAA Photo courtesy NOAA

FIGURE 4.23: Northeastern offshore FIGURE 4.24: Eastern spinner dol-

spotted dolphins, Stenella attenuata. phins, Stenella longirostris orientalis.

number of whales of all species killed commercially during the entire twenti-
eth century is only two million. The two primary species of dolphins affected
are the offshore pantropical spotted dolphin, Stenella attenuata (Figure 4.23),
and the eastern spinner dolphin, Stenella longirostris orientalis (Figure 4.24).
An informative and very readable reference on this issue is the report by Tim
Gerrodette,9 from which Figures 4.25–4.28 are taken.
In purse-seine fishing, a net is set around schools of tuna, often (about
half the time) associated with schools of dolphins (the reasons the two species
swim together are not well understood).10 Once the net is set around the dol-
phins and tuna (Figure 4.25), the net is closed, and a “backdown procedure”
initiated, in which one end of the net is held slightly underwater, so that the
dolphins may escape close to the surface (Figure 4.26). Most of the time, the
dolphins do escape (meaning that any given dolphin may be chased, netted
and freed many times during its lifetime), but sometimes things go wrong,
and the scale and frequency of these operations led over years to the millions
of dolphin deaths cited above, reducing the northeastern spotted and eastern
spinner dolphin stocks to 20% and 30%, respectively, of their population sizes
when the fishery began, as illustrated by the graphs in Figures 4.27 and 4.28.
Early 21st-century population estimates for these spotted and spinner dolphin
stocks are 640,000 and 450,000, respectively.
What is still not understood about this situation is the failure of these
two populations to recover following the halting of this massive harvesting. In
9 Tim Gerrodette, The tuna-dolphin issue, in W.F. Perrin, B. Würsig and

J.G.M. Thewissen, eds. Encyclopedia of Marine Mammals. Academic Press, San

Diego, 2002. pp. 1269–1273. Updated 2005 at http://swfsc.noaa.gov/textblock.aspx?
Division=PRD&ParentMenuId=228&id=1408.
10 One study suggested that the association may be related to reducing predation risk in

zones where low oxygen content in the water compresses the tuna habitat into a shallow
[oxygen-rich] region near the surface: M.D. Scott, S.J. Chivers, R.J. Olson, P.C. Fiedler,
K. Holland, Pelagic predator associations: tuna and dolphins in the eastern tropical Pacific
Ocean, Marine Ecology Progress Series 458: 283–302, 2012.
218 Dynamical Systems for Biological Modeling: An Introduction

Photo courtesy NOAA SWFSC Photo courtesy NOAA SWFSC

FIGURE 4.25: Purse seine be- FIGURE 4.26: Backdown procedure in

ing set on tuna and dolphins. The progress. As the tuna vessel moves back-
net is not yet closed, and four ward, the net is drawn into a long chan-
speedboats are driving in tight nel. The corkline at the far end is pulled
circles near the opening to pre- under water slightly, and the dolphins
vent the dolphins (and tuna) from escape. Speedboats along the corkline
escaping. From Gerrodette, 2002. help keep the net open. From Gerrodette,
2002.

the 1970s the United States passed laws that required American fishing ships
to reduce their dolphin bycatch to “insignificant levels approaching zero.” As
the fleets scaled down, however, tuna fleets from Latin American countries
grew, and so it was only through an international effort, including pressure
from U.S. consumers, who continued to buy most of the world’s commercially
processed tuna, that by the end of the century the annual dolphin mortality
had been reduced from onetime highs in the hundreds of thousands to under
3000. As can be seen in Figure 4.27, the two affected dolphin populations
stopped their steep decline in the mid-1970s, but even into the 21st century,
with relatively low mortality rates, the two populations persist at the reduced
levels to which purse-seine harvesting brought them. Median recovery rates of
1.7%/yr and 1.4%/yr have been calculated for northeastern offshore spotted
and eastern spinner dolphins, respectively.11
In studying the present populations, Gerrodette and colleagues have sug-
11 Two studies with preliminary data suggesting a stronger recovery in the early twenty-
first century are: Tim Gerrodette, George Watters, Wayne Perryman, Lisa Ballance, Esti-
mates of 2006 dolphin abundance in the Eastern Tropical Pacific, with revised estimated
from 1986-2003, NOAA Technical Memorandum NMFS-SWFSC-422, April 2008, available
at https://swfsc.noaa.gov/publications/TM/SWFSC/NOAA-TM-NMFS-SWFSC-422.pdf;
and International Dolphin Conservation Program Scientific Advisory Board,
7th meeting, Updated estimates of NM IN and stock mortality limits, Doc-
ument SAB-07-05, La Jolla, California, 30 October 2009, available at
http://www.iattc.org/PDFFiles2/SAB-07-05-Nmin-and-Stock-Mortality-Limits.pdf.
 Nonlinear Differential Equations 219

Reprinted from Encyclopedia of Marine Mammals, 2nd edition, Tim Gerrodette, The tuna-dolphin issue,
pp. 1192–1195, Copyright 2009, with permission from Elsevier

FIGURE 4.27: Estimated population trajectories of northeastern offshore

spotted dolphins and eastern spinner dolphins in the eastern tropical Pacific
Ocean. The populations declined due to high numbers of dolphins killed in
the tuna fishery from 1960–75.

gested several hypotheses to explain the observed failure to recover to earlier

levels:12
• Dolphin bycatch (including orphaned calves) may be higher than re-
ported.
• The catch-and-release experience, which a given dolphin may experience
2 to 50 times per year, causes stress, increased predation, and separation
of mothers and calves. Behavior changes and reduced reproduction have
been observed.
• Dolphin habitat is not constant (although no major long-term changes
have been measured).
• Many other unmeasured factors may be responsible, such as competitive
displacement by other species, disruptions to the population structure,
and the ongoing large-scale removal of tuna associated with the dolphins.
The first two hypotheses may be grouped as fishery effects, and the last two
as ecosystem effects. Those causes which directly affect individual dolphins
may be considered incidental harvesting (whether literal or figurative: that is,
whether the individual dies or merely isolates himself from the population),
while those which affect the population as a whole through environmental or
ecosystemic changes may be considered changes to the baseline parameters
(for a logistic model, r and K).
12 T. Gerrodette and J. Forcada, Non-recovery of two spotted and spinner dolphin pop-

ulations in the eastern tropical Pacific Ocean, Marine Ecology Progress Series 291: 1–21,
2005.
P.R. Wade, G.W. Watters, T. Gerrodette, and S.R. Reilly, Depletion of spotted and spin-
ner dolphins in the eastern tropical Pacific: modeling hypotheses for their lack of recovery,
Marine Ecology Progress Series 343: 1–14, 2007.
220 Dynamical Systems for Biological Modeling: An Introduction

Example 7.
Assume the two dolphin species are effectively at equilibrium. (a) If the ob-
served depression of the two dolphin species is to be explained by de facto
constant-effort harvesting, what is the effective frequency of these harvest-
ing events? (b) If instead harvesting “effort” is limited to the 3000 dolphins
per year cited in the discussion above, and the population depressions are
to be explained by changes to the baseline parameter r, what reproductive
rate would allow such limited harvesting to hold the population so far below
carrying capacity? (c) If harvesting effort is set as in (b) and the population
equilibrium is to be explained instead by changes to K, what is the resulting
carrying capacity from which the present populations are being depressed?
Solution: For the northeastern offshore spotted dolphins we have an estimated
present population of 640,000, original (pre-fishing) population of 3.2 million,
and r = 0.017/yr. For the eastern spinner dolphins we have an estimated
present population of 450,000, original (pre-fishing) population of 1.5 million,
and r = 0.014/yr.
(a) If we assume constant-effort harvesting, then the original population is
∗
the carrying capacity K and the present population is the equilibrium yH =
∗
K(1 − E/r), from which E = r(1 − yH /K). For spotted dolphins this gives
E = 0.017(1 − 0.64/3.2) = 0.0136/yr. For spinner dolphins this gives E =
0.014(1 − 0.45/1.5) = 0.0098/yr.
(b) We calculate effort from the given figures for harvesting, Ey ∗ , and the
present equilibrium, y ∗ . First, we arbitrarily divide the 3000 harvesting deaths
between the two species in proportion to their population sizes:
640, 000
3000 × = 1761
640, 000 + 450, 000
deaths/yr for the spotted dolphins, and the remaining 1239 deaths/yr for
the spinner dolphins. Then we divide: E = 1761/640, 000 = 0.00275/yr for
the spotted dolphins, and for the spinner dolphins E = 1239/450, 000 =
0.00275/yr. We now solve the equilibrium condition (4.13) for r:
E 0.00275
r= y∗ = = 0.00344/yr
1− K 1 − 0.2
for the spotted dolphins, and similarly we find r = 0.00393/yr for the spinner
dolphins. Both of these values are exceptionally low: that is, the reproductive
rates would have to be depressed by an order of magnitude from the usual
range of cetacean reproductive rates (a common estimate is 0.04/yr) in order
for changes in r to explain the depression. This would include both a reduced
birth rate and high calf mortality.
(c) We use the E = 0.00275/yr value calculated in (b) and the r and
equilibrium (y ∗ ) values from the literature to calculate K: for the spotted
dolphins
y∗ 640, 000
K= = ≈ 760, 000,
1 − Er 1 − 0.00275
0.017
 Nonlinear Differential Equations 221

Reprinted from Encyclopedia of Marine Mammals, 2nd Photo by John and Karen Hollingsworth, courtesy
edition, Tim Gerrodette, The tuna-dolphin issue, U.S. Fish and Wildlife Service.
pp. 1192–1195, c 2009, with permission from Elsevier.

FIGURE 4.29: Mississippi sand-

FIGURE 4.28: Estimated annual
hill cranes in the National Wildlife
number of dolphins killed in the east-
Refuge created specifically to pre-
ern tropical Pacific purse-seine tuna
serve them. Once reduced to a pop-
fishery, total for all dolphins and sep-
ulation of only 30 birds, this sub-
arately for the two dolphin stocks
species has now more than doubled
with the highest number killed.
that amount.

and for the spinner dolphins we likewise compute K = 560, 000 as the envi-
ronmentally reduced carrying capacity. 

The sandhill crane (Grus canadensis, Figure 4.29) is the most populous of
the world’s cranes, with over 500,000 distributed across North America, from
the tip of Siberia to Cuba. There are six recognized subspecies, three of them
migratory and three nonmigratory subspecies restricted to very small habitats
(and populations) in Mississippi, Florida and Cuba. Most migratory sandhill
cranes pass through the Platte River in the northern U.S. where they form
breeding pairs that are typically lifelong.
Sandhill cranes have lost much of their original wetlands habitats to human
development, and this continues to be the main threat to their existence (espe-
cially loss of roosting grounds on the Platte). However, overhunting also poses
a major threat to some populations in the central U.S. and Canada, along with
collisions with vehicles, utility lines, and fences. Some farmers hunt the cranes
because of the damage they do to crops, in areas where their original habitat
has been drained for agricultural purposes. The rate at which these sandhill
cranes encounter farmers (and their crops), and other manmade artifacts, is
proportional to the population size, rather than occurring at a fixed rate as
with quota-managed populations. For this reason we can model these encoun-
ters as constant-effort harvesting (although see Exercise 11 for an application
of constant-yield harvesting).
222 Dynamical Systems for Biological Modeling: An Introduction

Example 8.
It has been estimated13 that of the 420,000 mid-continental sandhill cranes,
between 25,000 and 32,000 cranes are lost each year to hunting or collisions.
If we take the current population as an equilibrium level (this population is in
fact believed to be stable, although the present population size may not be a
true equilibrium), what is the average length of time before a crane has such
a fatal encounter, and what is the true carrying capacity K if we assume the
cranes’ intrinsic growth rate to be 0.098/yr?
Solution: If the current population is at equilibrium (y ∗ = 420000), then the
harvest rate is Ey ∗ . If we take an average of the two extremes given, this gives
Ey ∗ = 28500 cranes/yr, from which E = 28500/420000 = 0.0679/yr. The
average time before an encounter happens is the reciprocal of this figure, or
14.74 years. To find the carrying capacity, we solve the equilibrium condition
(4.13) for K, to get
∗
yH 420000
K= = = 1.37 million cranes. 
1 − Er 1 − 0.0679
0.098

4.2.3 Migration and dispersal as harvesting

Although we do not typically think of harvesting or population manage-
ment as pertaining to humans, the processes of human migration and dispersal,
typically driven by resource considerations, affect local patterns of growth in
the same way as harvesting. Here harvesting corresponds to emigration, and
we extend our notion to include immigration as negative harvesting. In Sec-
tion 3.5 we considered migration through the lens of mixing processes; here
we look at migration and dispersal through the lens of harvesting. There we
considered migration terms that were either constant or linear (in the popu-
lation size); these correspond to the two harvesting models discussed in this
section. In Section 2.6 we observed for discrete-time models that constant
dispersal is equivalent to constant-yield harvesting, and that it may stabilize
or destabilize survival (positive) equilibria. For continuous-time models we
do not observe the same destabilizing effect but the correspondence of con-
stant dispersal or migration to constant-yield harvesting persists, as does the
correspondence of linear dispersal or migration to constant-effort harvesting.
Following are two illustrations of how we may turn harvesting terms positive
to reflect immigration.
13 Curt D. Meine and George W. Archibald (Eds). 1996. The cranes: Status survey

and conservation action plan. IUCN, Gland, Switzerland, and Cambridge, U.K. 294pp.
Northern Prairie Wildlife Research Center Online. http://www.npwrc.usgs.gov/resource/
birds/cranes/index.htm (Version 02MAR98).
 Nonlinear Differential Equations 223

Example 9.
In Example 1, Section 3.2 we fitted the population of the United States to the
curve
265
y= , (4.15)
1 + 69 e−0.03t
with population size measured in millions and time measured in years from
1790. Suppose the population is governed by a logistic differential equation
for which this is the solution but that in addition there was an immigration
of one million people per year beginning in 1790. What would be the limiting
population size under these conditions?
Solution: For the solution y(t) given in (4.15), limt→∞ y(t) = 265, and thus
in the logistic model K = 265. From the form of the solution of the logistic
equation obtained in Section 3.3, comparison with (4.15) shows that r =
0.03/yr. Constant
immigration of M per year would change the logistic model
y r 2
to y ′ = ry 1 − K + M , with equilibria given by ry − K y + M = 0, whose
only positive solution is
" r # " r #
1 2
4KM K 4M
y= K+ K + = 1+ 1+ .
2 r 2 rK

With r = 0.03, K = 265, M = 1, this gives the equilibrium (limiting) popu-

lation 295 million. 

Example 10.
An island with resources to support 1000 people becomes so attractive that
other people begin to hear about it from the inhabitants and come. On average,
each inhabitant’s messages convince inspire someone to immigrate there once
every fifteen years. If, as in the previous example, the normal reproduction
rate is r = 0.03/yr, what will the equilibrium size of the population be?
Solution: This situation corresponds to constant-effort harvesting, where the
harvesting (immigration) is proportional to the size of the population. Here the
immigration rate is Iy, where I = 1/(15yr) = 0.067/yr, and the equilibrium
size, adapting (4.13), is
   
I 0.067
y∗ = K 1 + = 1000 1 + ≈ 3222.
r 0.03

Once every fifteen years doesn’t sound like much, but in the end, despite the
impossibility of further natural growth (because of resource limitations), peo-
ple are arriving so fast (Iy ∗ ≈ 215 people/yr) that the immigration dominates
the system dynamics. 
224 Dynamical Systems for Biological Modeling: An Introduction

4.2.4 Conclusions
In Section 2.5 we found that for discrete-time constant-yield harvesting of
logistic (and other scramble-competition) populations there are both maxi-
2
mum (Hmax = (r−1) 4r K) and minimum (Hmin =
(r−3)(r+1)
4r K) harvest rates
that allow stability since scramble populations can be naturally unstable. In
continuous-time harvesting we have only a maximum (Hmax = rK/4, quite
close to its discrete counterpart) since continuous-time scramble populations
do not have that inherent instability. A similar comparison can be made for
constant-effort harvesting: for example, discrete-time models have zero equi-
libria stable for r − 1 < E < r + 1 (see, e.g., Exercises 9 and 10 in Section 2.5),
while the continuous-time model (4.12) has a zero equilibrium stable for E > r.
We can also distinguish the results of constant-yield or quota harvesting,
where overharvesting leads to catastrophe, from those of constant-effort or
proportional harvesting, where overharvesting leads to a gradual decline. Re-
call, however, that the maximum sustainable harvesting rate turns out to be
the same for both models.

Exercises
1. Solve the logistic constant-yield harvesting equation (4.9) using separa-
tion of variables and the technique of partial fraction expansion. Note
that the quadratic expression in y involved in dy/dt factors differently
(or not at all) in each of the three cases discussed in the text.
∗
2. Solve the logistic constant-effort harvesting equation (4.12) using sep-
aration of variables and the technique of partial fraction expansion.
∗
3. Consider a general population model y ′ = f (y) for which the popula-
tion growth rate f (y) is differentiable, has f (0) = 0, rises to a unique
maximum fmax , and decreases thereafter, eventually becoming negative.
Under constant-yield harvesting, this population size is governed by a
differential equation y ′ = f (y) − H.
(a) Show graphically that if 0 ≤ H < fmax , there are two intersections
of the curve z = f (y) and the horizontal line z = H, corresponding
to equilibria of the differential equation.
(b) Show that the larger of the two equilibria is asymptotically stable
and the smaller of the two equilibria is unstable.
(c) Show that if H = fmax there is only one equilibrium.
(d) Show that if H > fmax there are no equilibria and every solution
reaches zero in finite time.
4. Over 40,000 deer-vehicle collisions were reported in 2007, when the deer
population rose as high as 1.7 million deer. From this data (and the
 Nonlinear Differential Equations 225

value r = 0.5/yr obtained in this section) we can estimate a per capita

encounter rate E = 0.04M/1.7M = 0.023/yr (for a harvest rate Ey).
(a) If we now consider a deer population at an initial equilibrium of
about 3 million, and disregard other harvesting, by how much does
this “constant-effort harvesting” reduce the deer population?
(b) Write a model that incorporates both constant-yield (hunting) har-
vesting and constant-effort (deer-vehicle collisions) harvesting.
(c) Using the same parameters given above, with H = 0.5 million
deer/yr and K = 4.5 million deer as in Example 1 to generate an
equilibrium population of 3 million deer under hunting alone, find
the impact of deer-vehicle collisions. How does it compare with the
simpler estimate obtained in part (a)?
5. It has been estimated that 240,000 humpback whales lived in the North
Atlantic prior to commercial whaling, but according to Knox there were
only 3,000 by 1980 (see Table 4.1). One international study estimated
the 1986 population at 5,500 and the 1993 population at 10,600. Repro-
duction rate estimates for humpback whales have varied widely across
different populations: The IWC estimated an annual rate of population
increase of 3.1% in the Gulf of Maine for the period 1979–1993 (note
the 1986 moratorium marks the halfway point of this interval) but re-
ported rates of increase of about 7% for the eastern North Pacific during
the period 1990–2002, and rates of increase of 12.4% in East Australia
(1981–1996), and of 10.9% in West Australia (1977–1991). Use the pop-
ulation estimates for 1986 and 1993, and the given carrying capacity,
to estimate r for the North Atlantic population, as well as the critical
harvest rate.
6. It has been estimated that 265,000 minke whales lived in the North
Atlantic prior to commercial whaling. The IWC estimated 174,000 minke
whales in the North Atlantic (Central and Northeastern stocks) from
1996 to 2001. Norway harvested an average of 577 minke whales per year
commercially from these stocks between 1997 and 2006. (a) If we take
the population estimate as an equilibrium under whaling (an admittedly
unrealistic assumption), what is the corresponding reproduction rate r?
(b) With the given harvesting rate, what is the expected equilibrium
value if the true value of r is 0.03/yr? 0.06/yr? 0.09/yr?
7. The IWC estimated that there were about 10,500 bowhead whales in the
Bering-Chukchi-Beaufort Seas stock in 2001, with an observed annual
net growth rate of 3.2% during the period 1978-2001. It also estimated
that there were 1,230 bowhead whales in a separate stock off western
Greenland in 2006. A century earlier, the world bowhead whale popu-
lation is estimated to have been 120,000. Compare the maximum sus-
tainable yield (harvest rate) of this population with the average annual
drop in population during the twentieth century.
226 Dynamical Systems for Biological Modeling: An Introduction

8. By the early twentieth century, the population of gray whales in the

North Pacific had been reduced to about 2,000, but by the end of the
century it had rebounded to about 22,000 (most of them in the Eastern
North Pacific, off the coasts of Canada and the U.S.—only 121 were
estimated to remain in the Western North Pacific by 2007). Assuming
a carrying capacity of 24,000 and the reported (1967–1996) growth rate
of 2.5%/yr, estimate the maximum sustainable yield, and compare it
with the average annual catch of 174 whales/yr reported during the last
20 years before the IWC moratorium (during which the IWC reported
a net growth rate of 3%/yr in this population). What effect does the
model predict for a harvest rate of 174 whales/yr?
9. A 2007 study used DNA analysis to estimate a pre-whaling gray whale
population of between 76,000 and 118,000. What effect would an annual
catch of 174 whales (see previous exercise) have on this population if
the carrying capacity were 100,000? (Again use r = 0.025/yr.)
∗
10. Assume that blue whales obey a logistic law in the absence of whaling,
with carrying capacity of 200,000 whales and intrinsic growth rate of
0.05/yr, and that the average annual catch of 13,000 whales (1926–1940)
held in fact from 1925 to 1965, when the population had fallen to about
2000. Use either numerical integration on a computer, or the case (iii)
solution to the logistic constant-yield harvesting equation obtained in
Exercise 1, to estimate the blue whale population in 1925.
11. The sandhill crane (Grus canadensis) is hunted by farmers because of the
damage it causes to crops. Miller and Botkin found14 that the sandhill
crane population that winters in New Mexico, if not harvested, would
follow a logistic growth law with carrying capacity 194,600 and intrinsic
growth rate 0.098/yr. Find the critical harvest rate which would wipe
out the sandhill crane population, and the limiting population size if
3000 sandhill cranes per year are shot by irate farmers.
12. Suppose that a given population of sandhill cranes follows a logistic
growth law with carrying capacity 194,600 and intrinsic growth rate
0.098/yr. If a single crane, on average, dies from encounters with man’s
world roughly 0.04 times per year, by how much is the equilibrium pop-
ulation reduced?
13. Each breeding pair of sandhill cranes needs its own territory of 20–
100 acres of marshland. Of the 19,300 acres of land in the Mississippi
Sandhill Crane National Wildlife Refuge, about 12,500 acres can be used
by cranes. (a) What is the carrying capacity here, if about half of a crane
population consists of breeding pairs? (b) As of September 1994, there
14 R.S. Miller and D.B. Botkin, Endangered species: models and predictions, American

Scientist 62: 172–181, 1974.

 Nonlinear Differential Equations 227

were 120 cranes in the refuge. About how fast should the population be
growing at that point?
14. Suppose a small breeding population of 10 chipmunks is displaced into
a neighborhood capable of feeding 100 of them, but with an unfriendly
cat whom they encounter only very occasionally, about 5 times a year.
Assume the chipmunks’ reproduction satisfies a logistic model with
r = 0.3/yr, and that the feline encounters can be considered constant-
effort harvesting. How many chipmunks will there be after ten years? At
equilibrium? (Note: This is, of course, a facetious example, as at such
small sizes, stochastic effects dominate.)
15. In reality, the harvesting rate may depend both upon population size
and upon external (e.g., economic) considerations. Consider a scenario
in which harvesting activities are driven by a desired (constant-yield)
harvest rate H, but when harvesting begins to drive the population size
down, harvesting becomes more difficult, and is instead dependent upon
the (constant) effort expended to harvest them. In this case the effective
harvest rate becomes

Ey, y ≤ H/E;
R(y) =
H, y > H/E.

(a) Explain the system’s behavior if the population’s natural carrying

capacity K is less than the threshold value H/E.
(b)∗ . Explain the system’s behavior if K > H/E.
∗
16. Suppose a population of a species that follows a logistic growth pattern
is being used as a breeding reserve to produce individuals that can be
released back into the animal’s native habitat. Given r and K, what
is the maximum rate at which individuals in this population can be
released back into the wild, if every winter a proportion p of those in
the reserve die from cold or hunger?
17. Show that the positive equilibrium is stable, and draw the phase line,
for: (a) the constant-yield model with positive immigration M (as in
Example 9), (b) the constant-effort model with positive immigration Iy
(as in Example 10).
228 Dynamical Systems for Biological Modeling: An Introduction

4.3 Mass-action models

One of the key influences on the growth of any population—whether of
cells or complete individuals–is encounters with other populations. So-called
contact processes are at the heart of the dynamics of many population models.
Contact processes appear in models as nonlinear terms that describe the rates
at which one population interacts in a certain way (for good or ill) with
another population. The simplest way to describe such a rate is to assume
that it is directly proportional to the size of each population—that is, the
more members either population has, the more frequently encounters between
the two groups will happen. Under such an assumption, the contact rate has
the form α x(t) y(t), where x(t) and y(t) are the two populations in question,
and α is some constant of proportionality that gives the baseline rate at which
an encounter happens. This form is called bilinear, because it is linear in each
of the two variables.
The use of a bilinear contact rate has its origin in the study of chemical
reactions. In elementary chemical reactions with two reactants, call them A
and B, and one product C, the rate at which the reactants combine to form
the product has been determined to have the form k+ AB, where k+ is a
rate constant and A(t) and B(t) give the quantities (or concentrations) of the
respective reactants at time t. This description of the reactants “encountering”
each other is known as the law of mass action, first identified in 1864 by
Guldberg and Waage. For this reason, similar bilinear contact rates are said
to have mass-action incidence, even in contexts outside chemistry (such as
population biology).

4.3.1 A simple chemical reaction

For an understanding of what the law of mass action produces in terms
of describing this simple reaction, let us complete the model by supposing
that the reaction is reversible: that is, that the product C spontaneously
decomposes back into the reactants A and B at a rate k− C proportional to
the amount of C. This yields the chemical reaction
k+
A + B ⇋ C,
k−

and the differential equations

dA
= −k+ AB + k− C,
dt
dB
= −k+ AB + k− C,
dt
dC
= k+ AB − k− C. (4.16)
dt
 Nonlinear Differential Equations 229

Since dA dC
dt + dt = 0, we can deduce that A + C must be a constant, call it KA ,
so that A(t) + C(t) = A(0) + C(0) = KA , and we can write A in terms of C:

A(t) = KA − C(t).

Likewise, we have that B(t) + C(t) = B(0) + C(0) = KB . Therefore, we can

just keep track of C, rewriting (4.16) as
dC
= k+ (KA − C)(KB − C) − k− C. (4.17)
dt
We can now find the eventual (equilibrium) level of each chemical predicted
by the model. The equation (4.17) has two equilibria, which we can write

∗ 1n p o
C± = (KA + KB + Kd ) ± (KA + KB + Kd )2 − 4KA KB , (4.18)
2
where the ratio Kd = kk− +
of the reverse reaction rate to the forward reaction
rate is called the dissociation constant. (Note that the units for Kd are in
moles (amount), the same as KA and KB , because the units for k− and k+
are different.) A careful inspection of the expression inside the radical will show
that it is always positive, but smaller than (KA + KB + Kd ). Therefore both
equilibria are always positive. To determine their local asymptotic stability,
we calculate the derivative of dC/dt = f (C):

f ′ (C) = k+ (2C − KA − KB − Kd ).

We find that f ′ (C) < 0 if and only if C < 12 (KA + KB + Kd ). Since this is true
∗ ∗
for C− and false for C+ , the former must be locally asymptotically stable, and
the latter unstable (see the phase portrait in Figure 4.30).

0 *
C_ C*+ C

FIGURE 4.30: A phase portrait for equation (4.17).

This appears to suggest that there is a region in phase space for which the
amount of chemical C grows without bound; however, this is illusory, as the
∗
equilibria C± are defined in terms of the (presumably known) initial condition
∗
C(0), and we can show that C(0) is well below the threshold level C+ :
∗ ∗
C− + C+ 1 1
= (KA + KB + Kd ) = C(0) + (A(0) + B(0) + Kd ) > C(0),
2 2 2
∗ ∗ ∗
that is, C(0) is below the midway point between C− and C+ . Therefore C−
is globally asymptotically stable, and the level of chemical C will approach
∗
C− (from which we can also calculate the equilibrium levels of A and B; cf.
Exercise 6).
230 Dynamical Systems for Biological Modeling: An Introduction

Example 1.
The dissociation reaction for an acid with chemical formula HA (where A
represents the ion that bonds with hydrogen to form the acid) is typically
written as
k+
H + + A− ⇋ HA,
k−

even though in practice this dissociation occurs in water, so that it is more

properly
k+
H3 O+ + A− ⇋ HA + H2 O.
k−

The water is commonly omitted since it does not affect the reaction (except
to bond with the dissociated hydronium ions H + ). In this context the disso-
ciation constant for acids is often denoted Ka rather than Kd .
Strong acids (distinguished by having Ka > 1), such as hydrochloric acid
(HCl) and sulfuric acid (H2 SO4 ), dissociate almost completely in water, leav-
ing little of the original acid left, while weak acids (Ka < 1) dissociate very
little. If the dissociation constant for sulfuric acid is 1000, find how much
sulfuric acid remains of 0.50 M (moles) when it is left to dissociate in water.
Solution: Here we begin with no reactants, A(0) = B(0) = 0, and an initial
condition of C(0) = C0 = 0.50 M. From (4.18) we have
 
1
q
∗ 2 2
C− = (2C0 + Ka ) − (2C0 + Ka ) − 4C0
2
1h p i
= 1001 − 10012 − 1 ≈ 0.00025M. 
2

Example 2.
Acetic acid, CH3 COOH, is a weak acid, with a dissociation constant of 1.7 ×
10−5 M. How much acetic acid would remain if 0.50 M were left to dissociate
in water?
Solution: With the same initial conditions as in the previous example, we have
 
1
q
∗
C− = (2C0 + Ka ) − (2C0 + Ka )2 − 4C02
2
1h p i
= 1.000017 − 1.0000172 − 1 = 0.4999915M ≈ 0.50M.
2
In this case one can easily observe that almost none of the acid dissociates,
whereas in the previous example almost all of it did. 

The law of mass action has formed the basis for describing contact pro-
cesses in many contexts within population biology, perhaps most notably
predator-prey relationships and the spread of infectious diseases. We shall
delay our own study of predator-prey systems until Chapter 6, in which we
 Nonlinear Differential Equations 231

Photo courtesy of Hinochika / Shutterstock.com

FIGURE 4.31: On a street near Sannomiya JR Station in Kobe, Japan,

people wear face masks during the 2009 outbreak of swine flu. Disease con-
trol strategies center around reducing the rate of infectious contacts between
susceptible and infected individuals.

examine systems of interacting populations. However, as we saw in Section 3.2,

a simple two-compartment epidemic model can be represented with a single
equation in the case when the total population remains constant. We shall now
take advantage of this fact to use such a model to consider how mass action
incidence and its alternatives affect predictions about whether an outbreak of
infectious disease can persist.

4.3.2 The spread of infectious diseases, revisited

The driving process for the spread of any infectious disease is contacts
between infectious individuals and susceptible individuals (see Figure 4.31).
From its inception, mathematical epidemiology has used some version of the
law of mass action to describe the rate at which these contacts occur. Com-
partmental models such as those we shall consider assume a certain amount of
homogeneity within each compartment, so that we can describe contact rates
in terms of averages for each compartment and ignore individual differences.
In Section 3.2 we presented the differential equation
dI 1 1
= βSI − I = β(N − I)I − I, (4.19)
dt τ τ
where the total population N (assumed constant) is composed of infectives I
and susceptibles S, as a simple model for the spread of an infectious disease.
The fact that S = N − I, so that as the number of infectives increases, the
232 Dynamical Systems for Biological Modeling: An Introduction

number of susceptibles available to be infected decreases, introduces into the

mass-action term βSI the self-limiting term that makes (4.19) fit the logistic
form. In this way we were able to use the solution to the logistic equation in
order to derive the threshold quantity that determines whether an outbreak
will persist or die out. We will now reconsider this model from a slightly more
general framework of contact processes that will help us understand precisely
what the law of mass action means as a description of contact rates between
two populations.
Let us suppose, as before, that we have a population of N individuals
classified as either infectives I or susceptibles S. Also as before, infected in-
dividuals recover (and become susceptible again) after an average τ units of
time. Now let us assume that the rate at which individuals contact other
members of the population is a function c(N ) of the population size, with the
property that c(N ) is a nondecreasing function of N . That is, the larger the
population, the more (or possibly the same) potentially infectious contacts
one makes each day. (In other words, having more people around should not
mean fewer contacts.) We will say more about the specific form of this function
shortly.
If we now also assume that everyone in this population mixes homoge-
neously, so that a person’s infection status (infected or not) has no effect on
the [infection] classes of people one contacts, then the proportion of one’s
contacts that are made with susceptible individuals is just S/N , and the pro-
portion of contacts made with infected individuals is I/N . (This assumption
is known as proportional mixing or random mixing.) Thus, each day an aver-
age infected individual contacts c(N ) individuals, c(N ) S/N of which are with
susceptibles. Then a total of (c(N ) S/N )I infectious contacts are made each
day (or other unit of time).15
We can now write this more general model as
dI SI 1 (N − I)I 1
= c(N ) − I = c(N ) − I. (4.20)
dt N τ N τ
Here the important distinction is that the per capita contact rate is not a con-
stant β, but a function c(N ) of the population size. In fact, for populations of
small or moderate size, this per capita contact rate increases roughly linearly
with population size, say c(N ) = βN , in which case equation (4.20) simplifies
to the mass-action model (4.19) which we have seen before. However, for very
large populations the contact rate saturates, approaching a constant maxi-
mum, say c(N ) = α, at which point it is simply not possible to come into
[potentially infectious] contact with more people in a single day. This assump-
tion, which generates a rate of new infections of αSI/N , is called standard
incidence. Standard incidence is not appropriate for small populations, or pop-
ulations in danger of being driven to extinction, because it suggests that even
15 An alternative, equivalent way to derive the new infections term is to start with a single

susceptible’s daily contacts c(N ), multiply by the proportion of infectious contacts I/N and
the number of susceptibles S.
 Nonlinear Differential Equations 233

as N → 0 the contact rate remains steady. In reality, the contact rate should
be described as increasing more or less linearly with population size at first,
and then leveling off, possibly but not necessarily smoothly, as the population
size increases.
We can now take a qualitative approach to analyzing this more general
model. By inspection, one equilibrium of (4.20) is I = 0, called the disease-
free equilibrium. Factoring I out of the equilibrium condition, we are left with
N −I 1
c(N ) − = 0,
N τ
from which we find the endemic equilibrium
 
∗ 1
I =N 1− .
c(N )τ

To determine each equilibrium’s local asymptotic stability, we differentiate the

function f (I) = dI/dt, calculating:

I∗
 
′ 1
f (I) = c(N ) 1 − 2 − .
N τ

Thus (after some algebra)

   
1 1
f ′ (0) = c(N ) 1 − , f ′ (I ∗ ) = −c(N ) 1 − .
β(N )τ β(N )τ

Regardless of the form of c(N ), therefore, the persistence of the infection is

determined (as before) by its basic reproductive number, usually denoted R0 ,
and given in this case by R0 = c(N )τ . (Note this quantity is dimensionless).
If R0 < 1, then f ′ (0) < 0, making the zero equilibrium locally asymptotically
stable, while the endemic equilibrium I ∗ is not only unstable but outside the
state space altogether, for its value in this case is negative. On the other
hand, if R0 > 1, then f ′ (0) > 0, so that the zero equilibrium is unstable,
while I ∗ > 0 and f ′ (I ∗ ) < 0, so that the endemic equilibrium is locally
asymptotically stable, and the infection will persist in the population. Note,
finally, that in either case, the unique locally asymptotically stable equilibrium
is in fact globally asymptotically stable.
We can interpret the expression for R0 biologically by recalling from the
original discussion in Section 3.2 that an “average” infective makes c(N ) po-
tentially infective contacts in unit time, and remains infective an average of
τ units of time. Therefore R0 = c(N )τ represents the average number of sec-
ondary infections caused per infected individual before recovery (hence the
term basic reproductive number). If each infection is able to replace itself and
more (R0 > 1), then we should expect the disease to persist, while if each infec-
tion cannot, on average, replace itself before the individual recovers (R0 < 1),
then we would say the disease is doing a poor job of reproducing itself and
234 Dynamical Systems for Biological Modeling: An Introduction

Left: photo by C.M. Kribs. Right: photo by Daniel Bowen/Public Transport Users Assoc., Melbourne, Australia

FIGURE 4.32: Left, Main Street in Hope, Arkansas, USA (population

10,000); right, a commuter train in Melbourne, Victoria, Australia (popula-
tion 3.74 million). The number of [potentially infectious] contacts one makes
in a day depends very heavily upon population density, but beyond a certain
point it saturates, and can increase no further.

should die out. This kind of insight illustrates the reason why such mathemat-
ical models are useful. For this model, one might be able to make this same
argument without going through all the details of a qualitative analysis, but
not so for the more complicated models which arise in studying the problems
of interest to us. Mathematical results give rise to biological insights, and that
is our motivation.
So what difference does the form of the contact rate c(N ) make? In this
case, we have assumed the population size N to be constant, in order to be
able to study the system using a single differential equation. However, in many
more complicated systems, the population size may change during an outbreak
(especially for endemic situations). For populations whose size is on the order
of hundreds, or a few thousands, the mass-action assumption that the contact
rate is directly proportional to population size is reasonable: If you live in
a small town and the population doubles overnight, then you probably will
encounter double the usual number of people throughout the course of your
day: busier sidewalks and stores, fuller buses, etc.
However, at some point one’s ability to contact others in a way that might
transmit infection begins to saturate as the population grows. If you live in a
major metropolitan area like New York or Tokyo and the population doubles
overnight, you will probably not come into contact with twice as many people
on your daily routine, because the sidewalks, subways, and stores are already
all filled to capacity (compare the photos in Figure 4.32). In such a situation,
where the contact rate has saturated, we might instead assign c(N ) = a,
a constant value independent of changes in population size (as long as the
population remains at saturation levels). As discussed above, this assumption
produces the standard incidence model
dI SI 1
=α − I,
dt N τ
 Nonlinear Differential Equations 235

with a resulting constant R0 = ατ , as opposed to that generated by mass-

action incidence as in Section 3.2: R0 = βN τ , which is proportional to the
population size (and in particular may cross the critical value of 1 as the
population size changes during an outbreak). The conclusion to draw from this
discussion is that for low population densities the basic reproductive number
may change if the population changes, whereas for high population densities
it is not sensitive to fluctuations in population size.
In a scenario where the population size may change drastically during the
course of an outbreak—for example, if the mortality rate from infection is
quite high, as with the Ebola virus—it may be necessary to incorporate both
the unsaturated, mass-action phase for low population densities and the sat-
urated, standard incidence phase for high population densities. In this case,
the function c(N ) would rise almost linearly near the origin and then taper
off, either gradually or with a ramp-like cutoff. Figure 4.33 illustrates two
different possible contact rate functions, both of which saturate at a rate α
for large populations: a smooth, gradual saturation csm and a ramp-like func-
tion csw that switches sharply between linear and constant phases at some
threshold density A. In both cases, the overall form R0 = c(N )τ of the basic
reproductive number remains the same, but its dependence on population size
changes, again in ways that may prevent a disease from driving a population
extinct, because as the population size drops, so does R0 . (We should note
that although the expression we derived above for R0 came from a model in
which the total population is constant, the overall dependence of R0 on c(N )
remains true for models in which the population size may vary, as we shall see
in later chapters.)

Example 3.
Suppose that the infectious contact rate for the common cold saturates sharply
(like csw in Figure 4.33) at a level of α = 1.6/day for population sizes above
1 million, and increases linearly for smaller populations. (a) If the average
infectious period for the cold is 1 day, what is the critical population size
above which an outbreak of the cold will persist? (b) For large populations,
at what endemic level will the infection establish itself? (c) Despite model
predictions, outbreaks are observed in small populations, and outbreaks come
and go in large populations, rather than remain in a constant endemic state.
What limitations on the model are involved in these apparent contradictions?
Solution: (a) For large populations, we have the saturated R0 = ατ =
(1.6/day)(1day) = 1.6 > 1, so that the model predicts persistence of the
infection in populations of over 1 million. For smaller populations, we have
c(N ) = (1.6/day)(N/106) and R0 = (1.6/day)(N/106)(1day) = 1.6N/106, so
that R0 > 1 when N > 106 /1.6 = 625, 000. That is, the model predicts that
the infection will persist in populations of 625,000 people or more.
(b) The endemic equilibrium for populations of over 1 million has
I∗
   
1 1
= 1− = 1− = 0.375.
N R0 1.6
236 Dynamical Systems for Biological Modeling: An Introduction

That is, the model predicts that, eventually, 37.5% of the population will be
infected at any given time.
(c) The occurrence of outbreaks in small populations does not contradict
the model’s predictions, which simply state that when R0 < 1, eventually the
outbreak will die out. It is possible, however, that an outbreak might grow
within a small subset of the population which makes [potentially infectious]
contacts much more often than the general population—for instance, children
enrolled in school; this is a limitation of the model, which assumes that ev-
eryone in the population has roughly the same average contact rate. A more
detailed model could subdivide the population by contact rate, into high-risk
and low-risk groups.
The periodic disappearance and reappearance of outbreaks of the common
cold in large populations is most often attributed to seasonal factors which
are also not considered in this model, the most evident being weather. 

As we have seen before, re-examining the assumptions underlying a given

model may help us understand better why the model makes the predictions
that it does. We will here consider an alternative to one of the assumptions un-
derlying this simple epidemic model, and reconsider the others in the exercises
at the end of this section.

Example 4.
In formulating the epidemic model (4.20), we assumed that the outbreak is
occurring on such a short time scale that demographic changes such as births
and deaths are negligible. How can the model be adapted to account for such
changes without contradicting the assumption of a constant total population
size?
Solution: As discussed in our initial attempts to model population growth,
established populations run up against the carrying capacity of their environ-
mental resources. For an established population, it is therefore reasonable to
assume that, on a time scale of a few years or less, the total birth rate roughly
evens out the total death rate, so that the population size remains constant
even though it is constantly gaining and losing members. We can incorporate
this assumption via a constant per capita mortality rate µ, which can be es-
timated by noting that 1/µ is then the average lifetime in the system. If we
assume that the disease is not transmitted vertically, so that all newborns (or
new recruits) enter the susceptible class, then our model becomes
1
S ′ (t) = µN + I − β(N )SI/N − µS,
τ
1
I ′ (t) = c(N )SI/N − I − µI.
τ
Adding the two equations yields dN/dt = dS/dt + dI/dt = µN − µS − µI = 0,
so N remains constant, and we can again write simply the equation for I,
substituting S = N − I.
 Nonlinear Differential Equations 237

We leave the analysis of this model as an exercise but note that the basic
qualitative results remain unchanged from the original model (4.20), and the
quantitative results change only to incorporate µ into R0 . 

4.3.3 Contact rate saturation and the “Pay It Forward”

model
We shall now apply the above discussion of how contact rates saturate as
populations grow large to a new context. Many other phenomena besides the
transmission of infectious diseases are driven by human interactions, and can
be modeled by descriptions of contact processes. Studies of numerous sociolog-
ical phenomena driven by peer-pressure influence have extended the metaphor
of mass-action contact rates by applying epidemic modeling techniques to eat-
ing disorders, crime and gang activity, drug and substance abuse, grassroots
political movements, and even learning environments. Although these behav-
iors are far removed from the chemical reactions for which the law of mass
action was developed, such descriptions of the rate at which individuals change
their behavior can lead to powerful insights about why phenomena like these
can be so robust, and difficult to eradicate once established.
One study16 considered the effects of contact rate saturation on the growth
of a grassroots movement which depends upon individuals’ actions in order
to motivate others to participate. In this case, the phenomenon in question
was the philosophy of proactive charitable acts described in Catherine Ryan
Hyde’s novel Pay It Forward.17 The growth of this movement occurs when
participating individuals “pay it forward” (do a significant favor for one per-
son in return for one done to them by someone else) to non-participating
individuals, inspiring those who benefit to “pay it forward” themselves. Po-
tential recruits’ inclination to join the movement upon receiving such a favor
may be conditioned on their familiarity with the movement from secondary
sources like the news and anecdotes from friends. In this way, the effective
contact rate becomes an increasing function of the size of the movement (the
larger the movement, the more likely a person is to have heard of it, and thus
be predisposed to join upon receiving a favor).
If we let N (t) denote the number of people in the movement at time t and
P be the (constant) size of the overall population, then we can describe the
growth of the phenomenon as follows:

N ′ = f (N ) = c(N )N (P − N )/P − µN, (4.21)

where c(N ) is the per capita contact rate (making c(N )(P − N )/P the per
capita recruitment rate) and µ is the per capita rate at which individuals die or
drop out of the movement. We shall consider two ways to model the saturation
16 C.M. Kribs-Zaleta, To switch or taper off: the dynamics of saturation, Math. Biosci.

192(2): 137–152, Dec. 2004.

17 Catherine Ryan Hyde, Pay it forward, Simon and Schuster, New York, 1999.
238 Dynamical Systems for Biological Modeling: An Introduction
c(N )
6

α csw
.
...
............. csm
. . ....
.... (a)
.. . .
..... 0 N
....
...
...... -N
0 A (b)
0 *
N_ N*+ N

FIGURE 4.33: Two different ways FIGURE 4.34: Phase portraits for
to model contact rate saturation for the “Pay It Forward” model (4.21).
large populations.

that occurs as the movement reaches a critical size (at which point the move-
ment is so well known that further growth does not make non-participants any
more familiar with it): smooth and sharp saturation corresponding to those
graphed in Figure 4.33. If we keep α as the maximum per capita contact rate
and A as the critical population level, then we can define the two contact rates
as
csw = α min(N/A, 1), csm = αN/(N + A).
A qualitative analysis of the model (4.21) using smooth saturation csm
reveals that, in addition to the zero equilibrium, a pair of positive equilibria
∗
N± 1 p 
= (1 − m) + (1 − m)2 − 4ma
P 2
p
exists if and only if m ≤ 1 + 2a − 2 a(a + 1), where m = µ/α and a = A/P .
Stability analysis using the derivative

α (N + A)(2P N − 3N 2 ) − N 2 (P − N )
f ′ (N ) = −µ
P (N + A)2

shows that the zero equilibrium is locally asymptotically stable (f ′ (0) = −µ <
0), and when the positive equilibria exist, some algebra (which we omit here)
∗ ∗
can show that the lower one N− is always unstable and the upper one N+ is
always stable. This leaves us with a situation where, if individuals retire from
the movement faster than new ones can ever be recruited (m > 1), or if the
critical size at which the movement is well known to the public is large enough
(a > (1 − m)2 /4m), the movement is sure to die out (the zero equilibrium is
globally asymptotically stable, Figure 4.34(a)). If, on the other hand, the
maximum recruitment rate exceeds the retirement rate and that maximum
occurs at a small enough critical size, then survival of the movement depends
 Nonlinear Differential Equations 239
∗
upon having a large enough initial core (N (0) > N− , see Figure 4.34(b)). In
this latter case, we have two locally asymptotically stable equilibria, 0 and
∗
N+ ; for this reason the equation is said to be bistable.
The analysis of this same model using the switching function csw to de-
scribe the contact rate is a little more complicated, but not more difficult.
It yields qualitatively similar results to those above (an indication that these
results are robust), and we leave the details as an exercise (see Exercise 12).
We may wonder how the phase portrait in Figure 4.34(a) can become the
one in Figure 4.34(b) simply by shifts in some parameters, or more generally
what consequences this change from a globally stable extinction equilibrium
to a bistable state implies for the future of the movement. These questions
can best be answered through the lens of bifurcations, first mentioned in
Section 2.6 as changes in the qualitative behavior of a dynamical system. In
the next section, we will develop techniques to study bifurcations, and study
or revisit several situations (including this one) where bifurcations identify
the important changes in a system’s dynamics.

Exercises
1. In a chemical reaction a substance S1 with initial concentration K is
transformed into a substance S2 . Let y(t) be the concentration of S2
at time t, so that K − y(t) is the concentration of S1 at time t. If the
reaction is autocatalytic (meaning that the reaction is stimulated by
S2 ), then dy
dt is proportional to y and K − y. Thus

dy
= αy(K − y)
dt
for some constant α. If the reaction is started at time t = 0 by intro-
ducing an initial concentration A of S2 , find the concentration of S2 as
a function of t.
2. In a second-order chemical reaction, a molecule of a substance S1 and a
molecule of a substance S2 interact to produce a molecule of a new sub-
stance S3 . Suppose the substances S1 and S2 have initial concentrations
a and b, respectively, and let y(t) be the concentration of S3 at time t.
Then the concentrations of S1 and S2 at time t are a − y(t) and b − y(t),
respectively. The rate at which the reaction occurs is described by the
differential equation
dy
= α(a − y)(b − y),
dt
where α is a positive constant (cf. (4.17)). Find the concentration y as
a function of t if y(0) = 0.
3. Find the concentration y(t) in Exercise 2 if a = b.
240 Dynamical Systems for Biological Modeling: An Introduction

4. Find the amount of each of the following acids that remains at equilib-
rium after 0.50 M of it is left to dissociate in water:
(a) boric acid, B(OH)3 (Ka = 5.8 × 10−10 M)
(b) iodic acid, HIO3 (Ka = 0.17 M)
(c) trichloroacetic acid, Cl3 COO2 H (Ka = 0.22 M)
5. Give the units for the reaction rate constants k+ and k− in (4.17).
∗
6. Show that the expression for C− in (4.18) is equivalent to the formula
used by chemists,
A∗ B ∗
Kd = .
C∗

7. Sketch a complete phase portrait for the epidemic model (4.20).

8. Complete the substitution suggested in Example 4 and write the differ-
ential equation for I(t) describing the epidemic model with demographic
renewal. Then complete the model’s qualitative analysis.
9. In formulating the epidemic model (4.20), we assumed that the infectious
contact rate was a function of population size, β(N ). This is a reason-
able assumption in most cases, but outbreaks of disease can themselves
change the nature of the daily contact structure, either because of the
severity of their symptoms (people may be scared into wearing breath-
ing masks, as in the 2003 SARS epidemic, or staying home altogether)
or because of the size of the outbreak (healthcare facilities may become
overburdened and unable to cope, as in the 1918 Spanish flu pandemic).
How would the model and its analysis change if the contact rate were
instead dependent on I(t), and what form would you suggest for the
function β(I)?
10. In formulating the epidemic model (4.20), we assumed that no deaths
occurred due to the disease. If instead we assume that infected individ-
uals die at a rate of δ/day, how does the model change? Hint: Consider
this assumption’s impact on the system overall.
11. Some researchers18 have suggested a contact rate with exponents to
describe the incidence of new infections, something of the form βS m I n ,
where at least one of m, n > 1. In each case this unusual form was
motivated by a particular characteristic of the scenario being modeled.
(a) Formulate and analyze such a model for the case m = 1, n = 2.
How does the behavior of this model differ from (4.20) with the usual
mass-action incidence?
18 e.g., Pauline van den Driessche and James Watmough, A simple SIS epidemic model

with a backward bifurcation, J. Math. Biol. 40(6): 525–540, 2000, Example 4.1.
 Nonlinear Differential Equations 241

(b) Formulate and analyze such a model for the case m = 2, n = 1.

How does the behavior of this model differ from (4.20) with the usual
mass-action incidence, and from the model in part (a)?
∗
12. The analysis of switching models, such as the “Pay It Forward” model
(4.21) using rsw (N ), has three parts: first, analyze the part of the model
that falls below the switching point; second, analyze the part above the
switch; and third, address the continuity of these results at the switch
point.
(a) Complete a qualitative analysis of (4.21) with r(N ) = β0 N/A, and
note under which conditions the equilibrium values N ∗ have N ∗ < A.
(b) Complete a qualitative analysis of (4.21) with r(N ) = β0 , and note
under which conditions the equilibrium values N ∗ have N ∗ > A.
(c) Make a composite phase portrait for (4.21) with rsw (N ) by pasting
together the part of the portrait for part (a) below the switch point A
and the part of the portrait for part (b) above A. Show that the switching
model is bistable when a ≤ 1 − m or 1/2 < a < 1/(4m) (where a and m
are as defined in this section).
13. Predator-prey systems, like the other systems studied in this section,
are driven by a contact process: encounters between predator and prey.
This process is also commonly modeled with a mass-action term kxy,
where k is an encounter rate, x is the number of predators, and y is the
number of prey.
(a) If the prey exhibit logistic growth in the absence of predators (with,
say, growth rate r and carrying capacity K), what differential equation
describes their growth dy/dt in the presence of predators?
(b) Suppose the prey are the predators’ only food source, and that the
predator birth rate is a mass-action term (with a different rate constant,
say c, reflecting the efficiency with which predation is “converted” into
reproduction), while the predators’ per capita natural mortality rate is
a constant µ. Write the differential equation describing the predators’
growth rate dx/dt.
(c) Now suppose that in one predator-prey system, the prey are so plen-
tiful that we can consider their population size a constant Y , no matter
how great the predation (for instance, because any depletion is quickly
compensated for by migration from nearby habitats). Analyze the be-
havior of the predator model in part (b) under this assumption.
(d) Suppose that in a different predator-prey system, the predators have
many other food sources in addition to the prey, so that their population
is a constant X. Analyze the behavior of the prey model in part (a) under
this assumption.
242 Dynamical Systems for Biological Modeling: An Introduction

4.4 Parameter changes, thresholds, and bifurcations

We have now seen several examples of models for biological systems in
which the nature of a population’s long-term behavior depends upon the values
of one or more parameters in the model: beginning with the discrete logistic
[difference] equation, which we saw in Section 2.6 exhibits period doubling for
some values of the intrinsic growth rate r, and earlier in the present chapter
with models of foraging ants, harvesting, infectious diseases, and collective
behaviors. In all of these models, small changes in a parameter may cause
a system to cross a threshold beyond which the model’s behavior changes
qualitatively and sometimes abruptly. Even the simplest of all differential
equations we have studied, that attributed to Malthus, has solutions which
behave differently (grow unbounded or dwindle away to zero) depending upon
the sign of the coefficient r in dx/dt = rx. We will use the term bifurcation to
refer to such qualitative changes, which involve the appearance, disappearance,
or change in stability of the equation’s equilibria.
Although a bifurcation gives a mathematical description of a change in a
model’s behavior, it usually also has important biological consequences, most
commonly the survival or extinction of the population under study. There
are two principal reasons why we should be interested in knowing the be-
havior a given model predicts for a wide range of parameter values. In some
cases, values of model parameters may change over time, as the correspond-
ing biological system reacts to external factors; in others, model parameters
may remain constant but be difficult or impossible to measure accurately and
precisely. Bifurcation diagram (graphs) and threshold quantities (the combi-
nations of model parameters that mark changes in a system’s behavior) are
mathematical tools that help us interpret bifurcations in the context of the
systems in which they occur. In this section we shall develop some familiarity
with these tools by revisiting the models mentioned above to look at the un-
derlying bifurcations and get an idea for the kinds of systemic changes they
herald. We shall not try to give an exhaustive list of the possibilities, but the
examples we already have in hand will illustrate the most common types of
bifurcation that can occur in a single differential equation.
We begin by recalling the simple epidemic model presented in Sections 3.2
(equation (2.19)) and 4.3 (equation (4.19)). We have previously rewritten it
to fit the form of a logistic equation, but now we will instead use rescaling to
reduce the number of parameters to 1:
 
dI 1 βN τ − 1 2
= β(N − I)I − I = β I −I . (4.22)
dt τ βτ

The technique of rescaling, first mentioned in Chapter 1, involves defining

new variables, and often new corresponding parameters, relative to certain
benchmark quantities: populations relative to a given size, time relative to a
 Nonlinear Differential Equations 243

given period or rate. It also involves using the Chain Rule from calculus to
dz dz dy
write a new differential equation: dx = dy dx . In this case, we need rescale only
the time variable, to eliminate the β; we define s = βt, rename y = I, and
define the bifurcation parameter p = βNβτ τ −1
= Rβτ
0 −1
; then, using the Chain
Rule,
dy dI dt 1
= = β[py − y 2 ] ,
ds dt ds β
so that the epidemic model (4.22) becomes

y ′ = py − y 2 . (4.23)

In this simplified form, we can better see the effects of the bifurcation param-
eter p.
Applying qualitative analysis techniques, we find two equilibria, y ∗ = 0
and y ∗ = p. We have f (y) = py − y 2 , f ′ (y) = p − 2y, so that f ′ (0) = p and
f ′ (p) = −p. Thus if p < 0, the equilibrium y = 0 is asymptotically stable and
the equilibrium y = p is unstable, while if p > 0, the equilibrium y = p is
asymptotically stable and the equilibrium y = 0 is unstable. (It is left as an
exercise to draw the two corresponding phase portraits and verify that there
are no non-negative unbounded solutions, so that local stability extends to
global.)
We can represent this information graphically in a bifurcation diagram by
graphing the equilibrium values y ∗ as functions of the bifurcation parameter
p. We draw solid lines or curves to represent stable equilibria and broken
or dashed lines to represent unstable equilibria. The bifurcation diagram for
equation (4.23) is shown in Figure 4.35. We see that as p increases through
zero, the two equilibria exchange stability; this type of exchange is called a
transcritical bifurcation. We can also now identify the bifurcation point via the
coordinates p = 0 and y = 0, since the two equilibria coincide at y = 0 when
p = 0.
We can also return to the original equation (4.22) and describe the bifur-
cation in terms of I ∗ , the equilibrium number of infectives, and R0 , the infec-
tion’s basic reproductive number (which has more biological meaning than p).
The endemic equilibrium y ∗ = p becomes I ∗ = (R0 − 1)/βτ , and its stability
criterion p > 0 becomes R0 > 1. Since we consider only a nonnegative state
space I ≥ 0 and by definition our epidemiological parameter R0 = βN τ ≥ 0,
we include in the bifurcation diagram (Figure 4.36) only the positive quad-
rant. In general, this bifurcation at R0 = 1, I ∗ = 0, which is so central to the
analysis of even the most complex epidemic models, is normally transcritical.

The “Pay It Forward” model studied in Section 4.3 exhibits a different

type of bifurcation. Equation (4.21) with smooth saturation is

dN αN P −N
= N − µN. (4.24)
dt N +A P
244 Dynamical Systems for Biological Modeling: An Introduction
y* I*

R0
1

FIGURE 4.35: Bifurcation diagram FIGURE 4.36: Bifurcation diagram

for the transcritical bifurcation in for the transcritical bifurcation in
(4.23). (4.22).

One equilibrium is N ∗ = 0; if we define a = A/P and b = α/µ (in terms of

the analysis in Section 4.3, b = 1/m), there are two additional equilibria
∗
N± bh p i
= (b − 1) ± (b − 1)2 − 4ab
P 2
p
if and only if b > bc = 1 + 2a + 2 a(a + 1). If we fix the parameter a
and use b as our bifurcation parameter, we can use the equilibrium stability
information determined in Section 4.3 to sketch the bifurcation diagram shown
in Figure 4.37. Note that here, rather than an exchange of stability between
two existing equilibria, we have a bifurcation point b = bc , N ∗ /P = b(b − 1)/2
on either side of which two equilibria (one stable and one unstable) appear
and disappear (respectively). This type of bifurcation is called a saddle-node
bifurcation, because the two equilibria thus created are commonly referred to
in higher dimensions as a saddle point (the unstable one) and a stable node.
(We shall see the reasons for these names when we study systems of equations
beginning in the next chapter).
Formal bifurcation analysis often involves shifting a bifurcation point to
the origin to simplify the equation. In its simplified form, the saddle-node
bifurcation is given by the equation

y′ = p − y2. (4.25)
√
This has no equilibria if p < 0, and the two equilibria y = ± p if p > 0.
√ √
Since f (y) = p − y 2 , f ′ (y) = −2y, and we have f ′ (± p) = ∓2 p. Thus
the positive equilibrium is asymptotically stable and the negative equilibrium
is unstable. Figure 4.38 shows the bifurcation diagram for (4.25). The term
bifurcation literally means “splitting,” and it is from situations such as this
that the phenomenon takes its name.
We may also observe that bifurcations are the reason why in some cases
we have needed to draw more than one phase portrait for a given differen-
tial equation. In fact, phase portraits can be seen as vertical “slices” of a
 Nonlinear Differential Equations 245
N* y*

b
bc

FIGURE 4.37: Bifurcation dia- FIGURE 4.38: Bifurcation dia-

gram for the saddle-node bifurca- gram for the saddle-node bifurca-
tion in (4.24). tion in (4.25).

bifurcation diagram, on which the bifurcation parameter is constant. Drawing

vertical lines to either side of the bifurcation in Figure 4.37, for example, and
marking the equilibria on those lines with circles (solid for stable and hollow
for unstable) produces the two phase portraits given for the “Pay It Forward”
model in Figure 4.34 in Section 4.3.

A third common type of bifurcation, which combines elements of both of

the previous two types, is illustrated by the following simplified model used by
Bernard et al.19 in their study of the dynamics of white blood cell production.
They identified three processes that affect the population of hematopoietic
stem cells (see Figure 4.40) in the resting phase: differentiation into mature
white blood cells, assumed to occur at a per-cell rate F (here assumed con-
stant); re-entry into the proliferative phase, at a per-cell rate driven by neg-
ative feedback to limit the number of proliferating cells; and entry from the
proliferative phase by the newly divided cells. The negative feedback to avoid
over-proliferation was modeled by Bernard et al. using a Hill function, of the
form
kθs
K(S) = s , (4.26)
θ + Ss
where k is the maximum rate of re-entry into the proliferative phase (when
S ≈ 0) and θ is the half-saturation constant [cell density]. Here the exponent
s measures the sharpness of the saturation; for s = 1 we have a Verhulst
function like the one used in the smooth-saturation “Pay It Forward” model.
Finally, the proliferative phase is assumed to last precisely τ units of time,
during which proliferating cells may experience apoptosis (cell death), at a
per-cell rate of γ. Therefore, if S(t − τ )K(S(t − τ )) cells re-enter the prolifer-
ative phase at time t − τ , then at time t (τ units later) they will return to the
quiescent (resting) phase, multiplied in number by two (because of prolifera-
tion) and reduced by a proportion e−γτ due to apoptosis. (To understand this
19 Samuel Bernard, Jacques Bélair, and Michael C. Mackey, Bifurcations in a white-blood-

cell production model, Comptes Rendus Biologies 327: 201–210, 2004.

246 Dynamical Systems for Biological Modeling: An Introduction

last factor, consider a fixed initial population P of cells in the proliferative

phase undergoing apoptosis: we have P ′ = −γP , so that P (t) = P (0)e−γt .)
All together, these assumptions yield the model
dS
= −F S(t) − S(t) K(S(t)) + 2e−γτ S(t − τ ) K(S(t − τ )). (4.27)
dt
Because of the fixed time delay involved, equation (4.27) is a delay dif-
ferential equation, rather than an ordinary differential equation. As such, its
analysis is beyond the scope of this course, but we can nevertheless identify its
equilibria, since at an equilibrium S(t) and S(t − τ ) are the same. We can see
immediately that S ∗ = 0 is one equilibrium; factoring this out of the equilib-
rium condition, we are left with the equation −F − K(S ∗ ) + 2e−γτ K(S ∗ ) = 0,
which (using (4.26)) we can solve to get
1/s
k(2e−γτ − 1)

∗
S =θ −1 .
F
If the exponent s is odd, then this expression gives a single (possibly negative)
value of S ∗ regardless of parameter values. In this case, were we to choose
a single parameter, say the maximum reactivation rate k, as a bifurcation
parameter, we would find a transcritical bifurcation at k = F/(2e−γτ − 1)
where the nonzero equilibrium crosses the zero equilibrium and exchanges
stability with it. (We leave sketching the bifurcation diagram as an exercise
for the reader.) However, if the exponent s is even, then the expression for S ∗
gives either 2 equilibria (the positive and negative roots), if k > F/(2e−γτ −1),
or none if k < F/(2e−γτ − 1) (since negative numbers have no real sth roots
for s even). Bernard et al. show that the zero equilibrium is stable when
k < F/(2e−γτ − 1), and that near the bifurcation point the nonzero equilibria
are stable, leading to the bifurcation diagram shown in Figure 4.39. In reality
we assume that k must exceed this critical value, since the stem cell density
remains positive, constantly producing new blood cells.
As the figure suggests, this type of behavior is called a pitchfork bifurcation.
In its simplest form, a pitchfork bifurcation has the equation
y ′ = py − y 3 . (4.28)
√
This equation has equilibria y ∗ = 0 and y ∗ = ± p (if p > 0). To determine
′ 2 ′
stability, we calculate f (y) = p − 3y . Since f (0) = p, the zero equilibrium
√
is asymptotically stable if p < 0 and unstable if p > 0. Since f ′ (± p) = −2p,
the two nonzero equilibria which exist if p > 0 are both locally asymptotically
stable. As p increases through p = 0, a single, globally asymptotically sta-
ble equilibrium loses its stability and gives rise to two locally asymptotically
stable equilibria. The corresponding bifurcation diagram showing the three
equilibria as functions of p is given in Figure 4.41. Here we have both the ap-
pearance/disappearance of equilibria associated with saddle-node bifurcations
and the stability exchange seen in transcritical bifurcations.
 Nonlinear Differential Equations 247
S*

Photo courtesy Patricia Ernst

FIGURE 4.39: Bifurcation diagram
for the pitchfork bifurcation in (4.27) FIGURE 4.40: Hematopoietic stem
with s = 2, using F as bifurcation cells (HSCs) develop into many dif-
parameter. ferent types of blood cells. These
HSCs may proliferate (subdivide)
many times before differentiating into
mature white blood cells.

Having now seen the three most common types of bifurcation that a single
differential equation may exhibit, we can now revisit some other models that
involve bifurcations, to see what type of bifurcation each model involves.

Example 1.
Classify any bifurcations exhibited by the constant-yield harvesting model
(4.9) of Section 4.2.
Solution: Since H is an externally controlled parameter (as opposed to r and
K, which are innate), it is the most sensible choice for bifurcation parame-
ter. As discussed in Section 4.2, the model’s only equilibria, given by equation
(4.11), exist if and only if H < rK/4. At H = rK/4 the equilibria coalesce into
a single point, and for H > rK/4 there are no equilibria. From this descrip-
tion, from their stabilities as determined earlier, and from the corresponding
bifurcation diagram graphing the equilibria as functions of H (see Exercise
8), we can identify the bifurcation at H = rK/4, y ∗ = K/2 as a saddle-node.


Example 2.
Classify any bifurcations exhibited by the constant-effort harvesting model
(4.12) of Section 4.2.
Solution: Here the effort E is the most reasonable choice of bifurcation param-
eter. As discussed in Section 4.2, the model’s two equilibria cross at E = r,
y ∗ = 0, with each one stable on a different side of the bifurcation point. From
this description and the bifurcation diagram (see Exercise 9), the bifurcation
can be identified as transcritical. 
248 Dynamical Systems for Biological Modeling: An Introduction
y* y*

1-
---
v----
2

p r
1
--
4
1
----
-v----
2

FIGURE 4.41: Bifurcation diagram FIGURE 4.42: Bifurcation diagram

for the pitchfork bifurcation in (4.28). for Example 3.

Example 3.
Describe the behavior of solutions of the differential equation y ′ = ry +y 3 −y 5 .
Solution: Equilibria are y = 0 and the solutions of r + y 2 − y 4 = 0, which is
quadratic in y 2 ; solutions to the latter equation are given by
√
1 ± 1 + 4r
y2 =
2
if r ≥ − 14 . To determine stability, we substitute into f ′ (y) = r + 3y 2 − 5y 4
and find that f ′ (0) = r, so that the zero equilibrium is locally asymptotically
stable if and only if r < 0, while
√ √ √
 
1
f′


(1 ± 1 + 4r) = − 1 + 4r 1 + 4r ± 1 .
2
∗2
√
Thus for the two equilibria given√ by y√ = (1 + 1 + 4r)/2 which exist for
r ≥ −1/4 we have f ′ (y ∗ ) = − 1√+ 4r( 1 + 4r + 1), implying stability, while
for the two given√by y ∗2 = √ (1 − 1 + 4r)/2 which exist for −1/4 ≤ r ≤ 0 we
have f ′ (y ∗ ) = − 1 + 4r( 1 + 4r − 1) < 0, implying instability. We plot the
bifurcation curve (Figure 4.42), which appears to show a pitchfork bifurcation
√
at r = 0, y ∗ = 0 and two saddle-node bifurcations at r = − 41 , y ∗ = ±1/ 2. 

If we consider what would happen to a solution of the differential equa-

tion in Example 3 as the parameter r changes through values in the range
[−1/4, 0], we will see that the bifurcations cause sudden jumps from one equi-
librium value to another. If r < −1/4 initially, for instance, the solution y will
approach the zero equilibrium, but if r increases past 0, the zero equilibrium
becomes unstable, and the two outer equilibria become stable. If we suppose
y to represent a population, so that y(t) ≥ 0, then the solution y(t) will jump
toward the upper equilibrium, approaching it asymptotically. If later r de-
creases again, the solution will remain close to the upper equilibrium until the
latter vanishes at r = −1/4, at which point the solution will jump back down
toward zero. This discontinuous dependence of equilibrium population size on
 Nonlinear Differential Equations 249

a parameter is known as hysteresis, a phenomenon with important enough

biological consequences to warrant a brief discussion of its own.

4.4.1 Hysteresis
Hysteresis can be described mathematically by having two (or more) bi-
furcations at each of which a stable equilibrium either vanishes or becomes
unstable, without a direct “hand-off” of stability to an intersecting equilib-
rium. These multiple discontinuities in the limiting value for solutions give the
system what is thought of as “memory”: in the interval between the discon-
tinuities, there are multiple locally asymptotically stable equilibria, and the
one which a given solution approaches depends upon which of the two regions
beyond this interval the bifurcation parameter has most recently occupied.
In the case of Example 3 above, if −1/4 < r < 0, then the solution y(t) ap-
proaches the upper equilibrium if, prior to entering the interval [−1/4, 0], r
was most recently greater than 0, and approaches the zero equilibrium if r
was most recently less than −1/4. In this way a directed loop is formed.
Hysteresis loops have significant biological consequences as well. First, a
small change in environmental parameters may cause a sudden and very large
change in population size. Second, once the change occurs, it may be difficult
to undo: returning to the previous equilibrium requires a large change in the
environmental parameter(s), far beyond the value(s) held before the change.
For instance, in Example 3, if r increases past 0, in order to return the popu-
lation to the zero equilibrium one must get r below −1/4. These consequences
will be easier to understand in context, and so in the remainder of this sec-
tion we will look at the effects of hysteresis through the lens of two biological
systems which exhibit it.

The first such system is the ant foraging model (4.1) with which we began
this chapter,
dx sx
= (α + βx)(n − x) − .
dt x+K
In Section 4.1 we saw that this equation may have either one or three equilib-
ria. The appearance or disappearance of two equilibria signals a bifurcation.
Although the equilibrium condition (4.7) is cubic in the number x of ants
foraging at a given source, it is linear in the colony size n, and we can easily
solve for n in terms of x:
 
s
n= 1+ x,
(α + βx)(x + K)
which allows us to plot a bifurcation diagram of x∗ versus n. More generally,
when studying the relationship between a bifurcation parameter and some
property of a model (such as equilibria), we may often derive expressions which
are complicated in the state variable (here, x) but simple, even linear, in the
bifurcation parameter. Therefore, although we would really like to write an
250 Dynamical Systems for Biological Modeling: An Introduction
x* /K
5

n
35 40 45 50 55 60
x* /K
4

3 Photo courtesy Duncan Jackson

2
FIGURE 4.44: Foraging Pharaoh ants,
1 Monomorium pharaonis, are considered a
n
major nuisance in many places such as
10 20 30 40 50
hospitals. Their ability to send out new
colonies makes them difficult to eradi-
FIGURE 4.43: Bifurcation dia- cate.
grams for the ant foraging model
(4.1) with and without hysteresis.

expression for the state variable as a function of the parameter, we may instead
do the reverse, writing the parameter as a function of the state variable, in
order to graph it. This approach involves the notion of inverse functions, and
rather than graphing the independent variable on the horizontal axis and the
function on the vertical, we graph the variable (x, which we wish to interpret
as a function of the bifurcation parameter) on the vertical, and the inverse
function (which gives the parameter value) on the horizontal, thus giving us
the graph we wanted in the beginning. This approach may also be referred to as
defining the equilibrium value implicitly in terms of the bifurcation parameter.
Using the parameter estimates given by Beekman et al. and used in Sec-
tion 4.1, we see that the behavior varies with α, the rate at which ants find
the food source independently. The two possibilities are given in Figure 4.43:
either there are no bifurcations, and the stable equilibrium is always unique,
or else there are two saddle-node bifurcations, and a hysteresis loop exists in
the region where three equilibria exist. In the latter case, a growth in colony
size may result in a sudden leap in Pharaoh ants’ (Figure 4.44) collective for-
aging ability at a single site, whereas a drop in colony size below the lower
bifurcation point will suddenly wipe out the ants’ ability to make use of their
fellows’ pheromone trails. Either of these sudden changes will be difficult to
reverse, in that a much larger change in population size is necessary to undo
the change than was necessary to cause it in the first place.
A caveat is in order here: as can be seen from the two cases in Figure 4.43, n
is not the only bifurcation parameter. In fact, through rescaling we can show
(see Exercise 12 below) that our model has three fundamental bifurcation
 Nonlinear Differential Equations 251

parameters. It is important to be aware of how many model parameters really

affect a system’s bifurcation behavior. In all the other equations analyzed in
this section, we have derived explicit conditions for bifurcations; for instance,
each of the harvesting models from Section 4.2 only has one real bifurcation
parameter, since for any given parameter for which we sketch a bifurcation
diagram for those models, the graph remains qualitatively the same for all
allowable values of the other model parameters. This is not true, however, for
Bernard et al.’s white blood cell production model, nor is it true for the ant
foraging model.
Note that we have now classified all the bifurcations observed in mod-
els presented in earlier sections and chapters, save one: the discrete logistic
model. Single differential equations cannot undergo period doubling as single
difference equations can, because state variables change continuously rather
than making discrete jumps, and therefore cannot jump past an equilibrium
value. We will, however, see an equivalent behavior in systems of differential
equations in later chapters.

4.4.2 The spruce budworm

The spruce budworm Choristoneura fumiferana (see Figure 4.45) is an in-
sect which attacks spruce and fir trees in the forests of eastern North America.
Most years its population level is relatively low, but in some years (historically
roughly every forty years) it exhibits outbreaks in which the population may
increase by a factor of 1000. During an outbreak, budworms may eat enough
new needles in an evergreen forest to kill 80% of the trees in the forest and
effectively destroy the forest (see Figure 4.46). The destruction of the forest
also eliminates the budworms’ food supply, causing a collapse of the budworm
population, after which the forest can begin a slow recovery. A full model of
this behavior would include both the spruce budworm and the forest, and
these operate on very different time scales (weeks vs. decades). Here, we shall
describe a model for the budworm population (the fast variable) but rather
than also modelling the forest (the slow variable) we shall allow some of the
parameters of the model to vary to describe the change in the forest in re-
sponse to the budworm dynamics. This technique is often used to simplify
the analysis of models for systems with processes or components that act on
vastly different time scales.
This insect has been extensively studied, notably by Ludwig, Jones and
Holling.20 Consistent with this prior work, we let y denote the budworm pop-
ulation density (in larvae/acre) and assume that its natural growth is logistic,
with capacity K a function of the forest’s size (in particular, the average
leaf surface area per tree), and maximum growth rate r intrinsic to budworm
biology.
20 D. Ludwig, D.D. Jones and C.S. Holling, Qualitative analysis of insect outbreak systems:

the spruce budworm and forest, Journal of Animal Ecology 47(1): 315–332, February 1978.
252 Dynamical Systems for Biological Modeling: An Introduction

Reproduced with permission from Natural Resources Reproduced with permission from Natural Resources
Canada, Canadian Forest Service, 2015. Canada, Canadian Forest Service, 2015.

FIGURE 4.45: Mature spruce bud- FIGURE 4.46: Spruce budworm in-
worm larva. duced whole tree mortality in a nat-
ural conifer stand.

In addition, the spruce budworm is subject to predation by birds and par-

asites which saturates at a level H for high budworm densities. We assume
that the predators are opportunistic, choosing the most abundant prey, and
therefore preying little upon spruce budworms until they exceed a threshold
density A. This assumption leads us to use what is called a Holling type III
function to describe the predation: in particular, a second-order Hill function,
like that used in the blood cell production model presented earlier in this
section. The reason for this choice is that Hill functions stay close to zero
before the population passes the threshold A, and rise quickly to the maxi-
mum thereafter (see Figure 4.47). Ludwig et al.’s research suggested that the
saturation threshold A is, like K, proportional to the average leaf surface area
per tree.

pred./H
other prey
0.8 plentiful,
budworms
0.6 rare
budworms plentiful,
0.4
predation high
0.2

y/A
1 2 3 4

FIGURE 4.47: A second-order Hill function illustrates Holling type III pre-
dation, with opportunistic predators preying only on the most populous prey.
 Nonlinear Differential Equations 253

These assumptions lead to the following model for spruce budworm growth:
 y y2
y ′ = ry 1 − −H 2 . (4.29)
K y + A2
The model involves the four parameters r, K, H, and A. As discussed in Chap-
ter 1, it is often convenient to reduce the number of parameters by rescaling
the model: rescaling both population and time relative to certain benchmark
values will reduce the number of parameters by two (and with only two left, we
can more easily graph the different cases that arise). There are several options
for the rescaling benchmarks. Ludwig et al. used the half-saturation constant
A and the per-budworm-density predation rate H/A (see Exercise 13); here
we shall instead use the carrying capacity K as our population benchmark and
[the reciprocal of] the budworms’ natural growth rate r as our time bench-
mark, as did Robert May in his review of the work of Ludwig et al.21 The
qualitative results will be the same, and we encourage the interested reader
to compare our analysis below with that of Ludwig et al.
We therefore define the rescaled variables x = y/K (budworm larva den-
sity relative to the forest’s carrying capacity for them) and, if we call the
original time variable τ , the new time scale t = rτ , measured in budworm
“generations” (t = τ / 1r , where r1 is the average time for a budworm to re-
produce). We will also find it convenient to define the rescaled parameters
a = A/K, which gives the threshold density for predation relative to the for-
est’s carrying capacity, and h = H/Hc = H/ rK 4 , the predation rate expressed
as a proportion of the critical harvesting rate Hc (identified in Section 4.2).
Invoking the Chain Rule twice,
dx dy dx dτ
= ,
dt dτ dy dt
with dx/dy = 1/K, dτ /dt = 1/r, we obtain the rescaled model
dx h x2
= x(1 − x) − . (4.30)
dt 4 x2 + a2
∗
Looking for equilibria, we find that either x∗ = 0 or (1−x∗ )− h4 x∗2x+a2 = 0.
In the latter case we can multiply through by x∗2 + a2 to make the equation
polynomial:
h
G(x∗ ) ≡ (1 − x∗ )(x∗2 + a2 ) − x∗ = 0.
4
The function G is cubic with G(0) = a > 0 and G(1) = − h4 < 0, so it
2

has either 1 or 3 roots in (0,1). To determine local asymptotic stability, we

differentiate f (x), the right-hand side of equation (4.30), and find that
h 2a2 x
f ′ (x) = 1 − 2x − .
4 (x∗2 + a2 )2
21 Robert M. May, Thresholds and breakpoints in ecosystems with a multiplicity of stable

states, Nature 269: 471–477, 06 October 1977.

254 Dynamical Systems for Biological Modeling: An Introduction
G[x] G[x] G[x]
0.01 0.03
0.02
0.025
x
0.2 0.4 0.6 0.8 1 0.01 0.02
-0.01 0.015
x
0.2 0.4 0.6 0.8 1 0.01
-0.02
-0.01 0.005
-0.03 x
0.2 0.4 0.6 0.8 1

FIGURE 4.48: As model parameters a and h change, the polynomial G(x)

goes from having three roots in (0,1) to having only one; the change occurs
at a bifurcation point where G has a double root (center graph).

We calculate f ′ (0) = 1 > 0, so the equilibrium x∗ = 0 is unstable. This means

that the graph of f (x) crosses 0 from below at x = 0, so the next time it
crosses 0, it must do so from above, making f ′ (x∗ ) < 0 for the first positive
equilibrium, and alternating in this way thereafter. Thus the first positive
equilibrium is locally asymptotically stable, and, if there are three positive
equilibria, the second and third will be unstable and stable, respectively.
Returning to the question of bifurcations, we can identify the combinations
of the parameters a and h where bifurcations occur by noting that, at a
bifurcation point, there must be a double root of G(x): that is, G goes from
crossing the x-axis once on the interval (0,1) to crossing it thrice (or vice
versa) via a situation in which it just touches the x-axis tangentially at a
single point (see Figure 4.48). At such points, not only is G(x) = 0, but also
G′ (x) = 0. Although these equations are cubic and quadratic, respectively, in
x, they are linear in both a2 and h, so (again using an implicit approach) we
can solve them to obtain a and h parametrically, in terms of x:
h h
G′ (x) = 2x − 3x2 − a2 − = 0 ⇔ a2 = 2x − 3x2 − .
4 4
We substitute this expression for a2 into the equation G(x) = 0:
  
2 2 h h
(1 − x) x + 2x − 3x − − x = 0.
4 4

Solving for h, we get h = 8x(1 − x)2 , from which a2 = x2√ (1 − 2x). Since by
definition a > 0, we take the positive square root: a = x 1 − 2x. This last
expression gives us the necessary bounds on the parameter x: 0 ≤ x ≤ 1/2;
that is, different combinations of a and h produce bifurcations at x values any-
where from 0 to 1/2, and letting x vary from 0 to 1/2 will allow us to trace out
on the a–h plane precisely the combinations that produce those bifurcations.
(We can refer to this approach as reverse parametrization.) Modern scientific
computing software such as Mathematica, Maple and MATLAB can create
such parametric plots, and the result is given in Figure 4.49.
Thus for most values of a and h there are only two equilibria (the un-
stable zero equilibrium and a globally stable positive one), but for small a
 Nonlinear Differential Equations 255
h
1.2
x*
1

0.8 capacity
4 2 outstrips
0.6 predation

0.4

0.2 predation
dominates
a S
0.05 0.1 0.15 0.2 0.25 S1 S2

FIGURE 4.49:√ The paramet- FIGURE 4.50: A bifurcation diagram for

ric curve a = x 1 − 2x, h = (4.30) showing equilibrium budworm den-
8x(1 − x)2 (0 ≤ x ≤ 12 ) marks sity as a function of forest size S.
bifurcations dividing the a–h
plane into regions where (4.30)
has either two or four equilib-
ria.

and small enough h—that is, when predation saturates (“turns on”) for even
relatively low budworm densities, and is too weak to keep up with budworm
reproduction—there are four (the unstable zero equilibrium, and two stable
positive equilibria separated by an unstable one).
We now return to the context in which we developed the model: the bud-
worm population developing on a fast timescale, and the forest on a slow
timescale. In order to see what effect the forest’s growth and decay has on the
budworm population, we need to establish what relationship exists between
the two. Of the model’s four original parameters, the carrying capacity K and
predation saturation threshold A were both found to be directly proportional
to the state of the forest (in particular, to the leaf surface area density, mea-
sured in branches per acre). If we consider our rescaled parameters a = A/K
and h = H/ rK 4 , we find a to be independent of forest development as the two
linear factors—say A = αS and K = κS for foliage density S—cancel each
other out, while h = 4H/rκS varies inversely with foliage density. Since the
equilibrium condition G(x) = 0 is linear in h, it is a simple matter to solve it
4H
for h in terms of x, and then to substitute into the expression S = rκh , to get
∗
the equilibrium budworm density x implicitly as a function of S:
H x
S= .
rκ (1 − x)(x2 + a2 )

Figure 4.50 shows the resulting

√ bifurcation diagram for a representative value
of a in the interval (0, 1/3 3) where the bifurcations occur (see Exercise 14
for the identification of these limiting values).
256 Dynamical Systems for Biological Modeling: An Introduction

TABLE 4.3: Empirical parameter estimates for (4.30), from Ludwig et al.
(1978).
Parameter Value Units
r 1.52 /yr
α 1.11 larvae/branch
κ 355 larvae/branch
H 43 200 larvae/acre/yr

Looking at the bifurcation diagram, we see that in young forests (or those
largely degraded by infestation), with S small, the budworm density is kept
relatively low, first because the carrying capacity K = κS is low, and second
because the predation threshold A = αS is also low, so that predators choose
the budworms as primary prey. Here A is low because there are as yet so
few [healthy] branches that it does not take predators long to find budworms.
In this way, predation keeps the budworm population in check. As the for-
est grows, both the carrying capacity and the predation threshold increase,
leading to a gradual increase in budworm density. When S reaches the upper
bifurcation point, marked S2 in Figure 4.50, the food supply has become so
great that predation can no longer contain the budworm reproduction, and
the budworm density rises exponentially quickly toward the upper equilib-
rium, where budworm growth is limited only by the carrying capacity of the
food supply. At this upper equilibrium, the budworm population begins to de-
grade the forest, and S begins to decrease. If the deforestation caused by the
budworms causes it to decrease below the lower bifurcation point, marked S1
in Figure 4.50, then the food supply—and cover for the budworms—disappear,
leaving the budworms again dominated by predation. The upper equilibrium
disappears, and the population crashes back down to the lower equilibrium,
where they remain while the forest begins to rebuild, thus completing the
hysteresis loop. Since the forest growth takes decades, these cycles will also,
explaining the roughly 40-year budworm irruptions that were long observed.
The remaining question, given that hysteresis only occurs in our model
for a certain range of parameter values, is whether in fact observations yield
parameter estimates within this range. Ludwig et al. report two sets of pa-
rameter estimates, one based upon general principles and the other based
upon empirical observations. In most cases the two estimates are fairly close.
We reproduce in Table 4.3 the relevant parameters from their Table 1. From
them, we can calculate
√ a = α/κ = 1.11/355 ≈ 0.00313, which is well within
the range a < 1/3 3 ≈ 0.192 where hysteresis occurs.
The phenomenon of hysteresis has been observed in the outbreak and sub-
sequent crash of many different insect populations. There is, however, another
possible behavior for the system which has been observed in practice. This in-
volves human intervention to keep the insect population down by spraying,
in order to avoid the collapse of the forest (of course, this is not incorporated
in our model). In this case, the insect population remains at a high equilib-
 Nonlinear Differential Equations 257

rium level, held in check only by ongoing intervention. Such behavior is called
“perpetual outbreak.”

Exercises
In Exercises 1–7, describe the behavior of solutions of the given differential
equation as the parameter r is varied.

1. y ′ = r + y 2 5. y ′ = y + ry 3
2. y ′ = ry − yey y
6. y ′ = ry(1 − y) − y+1
3. y ′ = ry + y 2
y2
4. y ′ = r − y + y 2 +1 7. y ′ = ry + 1 + y 2

8. Draw a bifurcation diagram for the constant-yield harvesting model (4.9)

of Section 4.2 using H as a bifurcation parameter, and show how it
includes the three phase portraits given in Figure 4.15.
9. Draw a bifurcation diagram for the constant-effort harvesting model
(4.12) of Section 4.2 using E as a bifurcation parameter, and show how
it includes the two phase portraits given in Figure 4.22.
10. Use the bifurcation diagrams for transcritical, saddle-node, and pitchfork
bifurcations to sketch all [qualitatively distinct] phase line portraits for
each model, superimposed on the bifurcation diagrams.
11. Use Figure 4.43 to sketch all [qualitatively distinct] phase line portraits
for equation (4.1), superimposed on the bifurcation diagrams.
12. Rescale both the colony size x and the time variable t in the foraging
ant model (4.1) to reduce the number of model parameters to three.
13. (a) Rescale the spruce budworm model (4.29) using the benchmarks
Ludwig et al. originally used, defining the new variables u = y/A and
s = τ H/A and the new parameters Q = K/A and R = rA/H. Which of
these parameters is more reasonable to use as a bifurcation parameter?
(b) A technique that we used frequently in Chapter 2 to analyze equilib-
ria for discrete-time models was to consider equilibria as the intersections
of the graphs of two functions f (x) and g(x), in cases where the equi-
librium condition can be written as f (x) − g(x) = 0 or f (x) = g(x). Use
this technique to identify and analyze the stability of the equilibria for
the rescaled budworm model you derived in part (a).
(c) Interpret the results in (b) in terms of forest growth through the
bifurcation parameter chosen in (a).
258 Dynamical Systems for Biological Modeling: An Introduction

14. We can identify the range of values of the parameter a for which the
rescaled budworm model (4.30) undergoes two bifurcations √ as h (or S)
varies, by using the fact that the parametric curve a = x 1 − 2x, h =
8x(1 − x)2 (0 ≤ x ≤ 1/2) identified in the text “turns around” at the
cusp point. That is, as x increases from 0, the curve moves up and to
the right from the origin (that is, da/dx > 0 and dh/dx > 0). At the
cusp point, however, the curve begins to move down and to the left.
Therefore, at the cusp point da/dx = dh/dx = 0. Use this fact to find
the values of a and h at which this occurs.

4.5 Numerical analysis of differential equations

We began this chapter by observing that most differential equation models
for biological systems are nonlinear and cannot be solved outright. Section 4.1
presented some qualitative tools for analyzing the long-term behavior of such
models. In this section we shall conside briefly how numerical (quantitative)
analysis is used to study the short-term behavior of such models—what will
the spruce budworm density be in this forest a month from now? How long
will it take for an epidemic outbreak to drop below a given number of cases?
(Quantitative analysis can also be used to develop intuition about the long-
term behavior of systems in cases where we cannot complete a qualitative
analysis.) Numerical analysis, which uses computers to approximate solutions
of dynamical systems such as differential equations, has a long and detailed
history, and today many different methods (some quite specialized) are used
to analyze mathematical models numerically. Even a full introduction to nu-
merical analysis would fill an entire book, so in this section we shall simply try
to give the reader an understanding of the general approach behind numerical
methods, the relationship between the true solution to a differential equation
and the approximation to that solution produced by numerical methods, and
the trade-offs involved in quantitative modeling.
There is a great deal of computer software available to generate numerical
approximations of solutions to differential equations. Some of this software is
part of a larger computing system, such as Mathematica, MATLAB or Maple;
other programs are smaller and more narrowly aimed. Most useful programs
can generate both a table of values for a solution and a graphic portrayal of
solution curves. We encourage the reader to become familiar with some com-
puter software for approximating solutions of differential equations, and to
experiment with it, but not to use it to the exclusion of all other approaches.
In particular, numerical approximations are usually not a substitute for the in-
formation provided by the qualitative approach of Section 4.1. One reason for
this is that numerical approximations require as inputs specific values for all
model parameters, and consequently the results only give information about
 Nonlinear Differential Equations 259

the system’s behavior for those particular parameter values; the system’s be-
havior for other parameter values may be quite different, even if the differences
in values are slight. Another reason is that all approximations involve errors,
as we will discuss in more detail below.
The basic problem we would like to solve is to estimate the value of the
population y(t) at some particular moment in time T , given an initial popula-
tion size y0 and a description (a differential equation) of how that population
changes over time. In mathematical terms, we want to solve the initial value
problem
y ′ = f (t, y), y(t0 ) = y0 . (4.31)
Numerical analysis works by converting this continuous-time problem into a
discrete-time problem (i.e., a difference equation yk+1 = g(k, yk )), which can
be solved one time-step at a time, just as we did when we first encountered
nonlinear difference equations in Section 2.1. In order to accomplish this con-
version, we first subdivide the time interval of interest, t0 ≤ t ≤ T , into N
subintervals of length h, by defining t1 = t0 + h, t2 = t1 + h = t0 + 2h, . . .,
tN = T = t0 + N h. The number of subintervals N and the length of each
subinterval h are related by T − t0 = N h; that is, once the time interval is
known, choosing the time-step h is equivalent to choosing the number of time-
steps N = (T − t0 )/h. The fundamental idea is to proceed to y(T ) by first
approximating y(t1 ), then y(t2 ), then y(t3 ), etc. Each calculation is based on
an approximate value for y ′ (t) at that step, obtained using the value of the
solution at one or more points (but not involving any limiting process). In
other words, at each step we try to make the approximation yk as close as
possible to the real solution at that point, y(tk ), so that the final value yN is
as close as possible to the real solution y(tN ) = y(T ). A numerical method
must then have a way to use the function f (t, y) from the original differential
equation to derive a function g(k, y) that will produce faithful estimates.

4.5.1 Approximation error

Any numerical approximation method is subject to two sources of error:
(i) Discretization error, caused by the approximations used in for-
mulating the method (converting the differential equation into a
difference equation).
(ii) Round-off error, caused by the limitation of the computer or
calculator in the number of digits it can retain in its calculations.
The errors from previous steps may tend to propagate and build up. In gen-
eral, discretization error can be reduced by choosing time-steps small enough
that the discrete approximation cannot diverge far from the original system’s
solution. (More specifically, if we can make the approximation as close as we
like to the true solution by choosing a small enough time-step, then the nu-
merical method is said to converge, a prerequisite for any useful method.)
260 Dynamical Systems for Biological Modeling: An Introduction

However, a smaller time-step also means more calculations to reach the final
time T , which in turn increases the cumulative round-off error.
In practice, therefore, the choice of step size h is a challenge: If the step size
is too large, the discretization error at each stage may be large, and these errors
may accumulate catastrophically. On the other hand, if the step size is too
small the computation may require a long time, and the large number of steps
may produce large accumulated round-off errors. One common procedure is to
vary the step size, selecting an optimal h value at each stage to minimize the
total error of both kinds, but one may instead simply repeat the calculation
with successively smaller values of h until further reduction of h makes no
significant change in the approximate value of y(T ) for the chosen time T .
In practice, computer software often carries out the numerical approximation
automatically without asking the user to provide the step size and without
indicating what method is being used.
We should also be concerned that the qualitative behavior of the original
(continuous-time) system and its (discrete-time) approximation agree: for in-
stance, that any equilibrium of one is also an equilibrium of the other, and
that each equilibrium is stable for the same range of parameter values in both
models. Since we have seen that even with a simple logistic model can behave
very differently in continuous and discrete time, we must verify each of these
properties for any given numerical method.

4.5.2 Euler’s method

In this section we will describe three common numerical methods. In each
case we will only use a constant step size h, although in practice many sophis-
ticated methods adjust the step size during the approximation process. The
first, and simplest, of these methods is known as Euler’s method, originated
by the great mathematician Leonhard Euler (1707–1783).
The idea behind this method is to extrapolate at each step along a line
tangent to the true solution: in effect, to linearize from one step to the next. If
we imagine a direction field (see Section 3.3) imposed on a graph, we begin at
the initial condition and jump forward h time units in the direction signaled
by the direction field at that point. Then, from this new point, we again
jump ahead following the direction field. Although the direction field varies
continuously, the idea is that if we make small enough jumps we should not
go too far afield. While the Euler method is much less accurate than the
approximation methods used in standard computational software, it gives us a
great deal of insight into the underlying principles of numerical approximation
of solutions to differential equations.
Mathematically, Euler’s method is based on the difference quotient

y(t + h) − y(t)
y ′ (t) ≈
h
 Nonlinear Differential Equations 261

(true for small h). We can rewrite this approximation as

y(t + h) ≈ y(t) + h y ′ (t).

Euler’s method therefore replaces the differential equation (6.18) with the
difference equation
y(t + h) = y(t) + h f (t, y), (4.32)
or in discrete terms (tk = t0 + kh)

yk+1 = yk + h f (tk , yk ), k = 0, 1, 2, . . . , N. (4.33)

If we start with t = t0 , we can use the difference equation (4.32) to calculate

an approximation

y1 = y(t0 ) + hf (t0 , y(t0 )) = y0 + hf (t0 , y0 )

to y(t1 ). We now take t = t1 and use the approximation y1 for y(t1 ) in (4.32)
to obtain an approximation

y2 = y1 + hf (t1 , y1 )

to y(t2 ). We repeat this procedure, letting yk be the approximation to y(tk )

for k = 1, 2, . . . , N and calculating these approximations recursively using
(4.33) until we have our estimate yN for y(T ) = y(tN ). In general, a nu-
merical method is defined by the difference equation it uses to approximate
a differential equation; the Euler method is identified by (4.33). The Euler
method is usable on a microcomputer or programmable calculator, or even on
a simple non-programmable hand calculator if speed and convenience are not
vital.

Example 1.
Use Euler’s method with a step size of 1 to approximate the solution of the
logistic equation y ′ = ry(1 − y/K) on the interval 0 ≤ t ≤ 10 for r = 1,
K = 10, and y0 = 1.
Solution: Following (4.33), we compute

y1 = y0 + hry0 (1 − y0 /K)= 1 + 1(1)1(1 − 1/10) = 1.9,

y2 = y1 + hry1 (1 − y1 /K)= 1.9 + 1(1)1.9(1 − 1.9/10) = 3.439,
y3 = y2 + hry2 (1 − y2 /K)= 3.439 + 1(1)3.439(1 − 3.439/10) = 5.6953279,

and so forth, up to y10 ≈ 10. The graphs of the true solution (given by
equation (2.13)), the approximation u(t) based on the above values for yk ,
and a coarser approximation based on a step size of h = 1.5 are shown in
Figure 4.51. 

In the example above and in Figure 4.51, we use impractically large step
262 Dynamical Systems for Biological Modeling: An Introduction
y y
10
10
8
8
6
6
4 4
2 2
t t
2 4 6 8 10 2 4 6 8 10

FIGURE 4.51: Comparison of the FIGURE 4.52: Comparison of the

actual (solid curve) and two approx- true solution (solid curve) to the lo-
imate solutions (h = 1 long dashes, gistic equation of Example 1 and
h = 1.5 short dashes) to the logistic (4.32) with the approximations (h =
equation of Example 1. 1) generated by Euler’s method
(long dashes) and the modified Euler
method (short dashes).

sizes to illustrate some basic properties of approximations (in practice we

would use step sizes orders of magnitude smaller, perhaps one-millionth the
length of the entire interval). Note first of all that the places where the approx-
imations diverge most from the true solution are those where the derivative
y ′ (t) = f (t, y) changes rapidly; because Euler’s method essentially assumes
that dy/dt is constant on each subinterval, the discretization error is greatest
following a subinterval on which the true value of dy/dt changes significantly.
For the S-shaped logistic curve, this means underestimating the initial jump
in growth and then overshooting when the population levels off near the car-
rying capacity K. Second, note that the smaller step size (h = 1) produces a
closer approximation than the larger one (h = 1.5).
Properly speaking, the approximations produced by any numerical method
are discrete sequences rather than continuous functions, and should be
graphed discretely, with dots; however, in order to compare such approxi-
mations with the true solution, we commonly graph the approximations by
connecting the points yk to form continuous functions u(t) which can be used
to approximate y(t) at any point in time. In this way, the graph of the function
u(t) defined on each subinterval by

u(t) = yk + (t − tk )f (tk , yk ) [tk ≤ t ≤ tk+1 ]

is made up of straight line segments joining the points (tk , yk ) and (tk+1 , yk+1 ).
Each straight line segment has the same direction as the element of the direc-
tion field (Section 3.2) at the point (tk , yk ). A good approximation is one for
which the graph of the function u(t) is close to the graph of the true solution
y(t).
We can also use this simple example to look at errors in more detail. Since
we used large step sizes in Example 1, the approximation error is due primarily
 Nonlinear Differential Equations 263

to discretization. For example, we calculated y3 = 5.6953279, while the true

10
solution has y(3) = 1+9e −3 ≈ 6.90568, an error of about 17.5%. However, if we

had rounded the value of y2 = 3.439 to 3.4 in calculating y3 , we would have

obtained the estimate 3.4+1(1)3.4(1−3.4/10) = 5.644 for y3 , which introduces
an additional 0.0513279, or 0.74%, round-off error. Furthermore, if we then
round this estimate for y3 to 5.6 in order to calculate y4 , the round-off error
will accumulate (note the true value of y3 rounds to 5.7). Although rounding
to two digits is an extreme example, the true value of y10 in Example 1 requires
over 1000 decimal places to represent completely, which is more than even the
most precise software normally uses, so there is round-off error even at high
precision.
It is possible to show using Taylor approximations (see Appendix B) that,
in general, the discretization error of the Euler method in each step is no
greater than M h2 , where M is some constant whose value depends on the
function f (t, y) but not on h. Taking a smaller value of h will decrease the
discretization error at each step, but taking a smaller value of h means taking
more steps and may increase the round-off error. The best accuracy may
require an intermediate value of h. The accumulated discretization error of
the Euler method after N = (T − t0 )/h steps is then at most M h2 (T − t0 )/h =
M (T − t0 )h, that is, Kh for some constant K. The size of the round-off error
(relative to h) depends on the precision used in the calculations.
We can also compare the qualitative behaviors of solutions to the discrete-
time model used by a given numerical method with the solutions to the original
continuous-time model. For Euler’s method, we compare the difference equa-
tion (6.10) with the differential equation (4.31). Any equilibrium y ∗ of the
differential equation (4.31) must have f (t, y ∗ ) = 0 for all t; substituting this
into (6.10) gives yk+1 = yk , so y ∗ must also be a fixed point of the difference
equation (6.10), and vice versa. We should be concerned about the possibility
that the equilibrium y ∗ might be asymptotically stable for (4.31) but unstable
for (6.10). An equilibrium y ∗ of (4.31) is asymptotically stable if f ′ (y ∗ ) < 0,
while an equilibrium y∞ of (6.10) is asymptotically stable if |1+hf ′ (y∞ )| < 1,
or
−2 < hf ′ (y∞ ) < 0.
The upper inequality is equivalent to f ′ (y∞ ) < 0, the asymptotic stability
condition for (4.31). The other condition is f ′ (y∞ ) > − h2 , and this condition
can always be satisfied if h is sufficiently small. Thus there is no instability
problem using the Euler method if h is chosen sufficiently small.

Example 2.
Use the Euler method to estimate the value of e as y(1), where y(t) is the
solution of the initial value problem y ′ = y, y(0) = 1, and determine the
discretization error.
Solution: The Euler method gives

y0 = 1, y1 = 1 + hy0 = 1 + h, y2 = y1 + hy1 = y1 (1 + h) = (1 + h)2 , . . .

264 Dynamical Systems for Biological Modeling: An Introduction

If we take h = 0.1, so that N = 10, then the approximation to e is y10 =

(1 + h)10 = (1.1)10 ≈ 2.5937. If we take h = 0.01, we would have yk+1 = (1 +
h)yk = 1.01 yk , and this would approximate e by y100 = (1.01)100 ≈ 2.7048.
Similarly, h = 0.001 would lead to the approximation (1.001)1000 = 2.7169.
In general, the approximation to e given by Euler’s method is (1 + h)1/h , and
we know from calculus that the limit of this expression as h → 0 is e. Here
we can see that reducing h from 0.01 to 0.001 changes our approximation of
e by about 0.01, so we may be confident that our approximation is correct
to the nearest tenth, i.e., e ≈ 2.7. (The actual value of e is approximately
2.7183, which yields discretization errors of about 0.1246, 0.0135, and 0.0014,
or 4.6%, 0.5%, and 0.05%, for h = 0.1, 0.01, and 0.001, respectively.) 

4.5.3 Other numerical methods

Although the Euler method can provide accurate approximations for small
enough step sizes, in practice more sophisticated methods are used to approxi-
mate the solutions of differential equations. They all share the ideas of division
into subintervals and the use of difference equations, but make more compli-
cated calculations at each step, in order to generate less error.
As an example, the modified Euler method uses a slight change in method
to improve results. The Euler method is based on the approximation y ′ (t) ≈
[y(t+h)−y(t)]/h, which uses the derivative at the beginning of each subinterval
as the derivative over the entire interval. We might suppose that the derivative
in the middle of a subinterval might be a better approximation of the derivative
over the entire interval, and indeed it can be shown (although we leave the
proof to a text on numerical analysis) that such an approach generates a
smaller error in general. There are several numerical methods based upon this
idea; one is the modified Euler method, which uses the difference quotient
y(t + h) − y(t − h)
y ′ (t) ≈ .
2h
(For another such method, see the midpoint method in Exercise 9.) Here we are
actually using the derivative at t as the derivative over the double subinterval
[t − h, t + h]. To implement this method, we substitute for y ′ (t) in (4.31):

y(t + h) − y(t − h)
= f (t, y(t))
2h
or
y(t + h) = y(t − h) + 2hf (t, y(t)).
For a given time tk , we have tk + h = tk+1 and tk − h = tk−1 , so that we can
rewrite the difference equation as

yk+1 = yk−1 + 2hf (tk , yk ) [k = 1, 2, . . . , N − 1]. (4.34)

If the initial value y0 = y(t0 ) and an approximation y1 for y(t1 ) are known,
 Nonlinear Differential Equations 265

then the approximation y2 for y(t2 ) is given by

y2 = y0 + 2hf (t1 , y1 ).

In order to use this method, it is necessary to estimate y1 by some other means.

For example, we could use the Euler method (preferably with a smaller value
of h) to approximate y(t1 ).
It is possible to show that the discretization error in the modified Euler
method at each step is at most M̃ h3 , as compared to M h2 for the Euler
method. The higher power of h in this error bound indicates greater accuracy
for the modified Euler method (since h is small). However, the constants in the
two bounds are different, and the constant M̃ in the bound for the modified
Euler method may be larger than M . Thus the improvement may be less than
expected unless h is small enough.

Example 3.
Use the modified Euler method with a step size of 1 to approximate the
solution of the logistic equation y ′ = ry(1 − y/K) on the interval 0 ≤ t ≤ 10
for r = 1, K = 10, and y0 = 1. Compare the accuracy to that of Euler’s
method using the same step size.
Solution: As noted above, before we can apply the modified Euler method we
must first obtain an estimate y1 for y(t1 ). To keep things simple, let us use
the estimate y1 = 1.9 given by the original Euler method (with h = 1) in
Example 1 (a smaller h would give a better y1 ). Then, following (4.34), we
compute

y2 = y0 + 2h ry1 (1 − y1 /K) = 1 + 2(1) 1(1.9)(1 − 1.9/10) = 4.078,

y3 = y1 + 2h ry2 (1 − y2 /K) = 1.9 + 2(1) 1(4.078)(1 − 4.078/10) = 6.7299832,
y4 = y2 + 2h ry3 (1 − y3 /K) = 4.078 + 2(1) 1(6.7299832)(1 − 6.7299832/10)
= 8.479431625543553,

etc., up to y10 ≈ 10. If we compare the results with those given by the simpler
Euler method, we find significant differences immediately: for y(2) ≈ 4.50853,
the modified Euler method gives y2 = 4.078, an error of less than 10%, rather
than the Euler method’s y2 = 3.439 which has an error of 23.7%. At the next
step, for y(3) ≈ 6.90568, the modified Euler method has an error of 2.5%
compared to the Euler method’s 17.5% error. However, in this case the equi-
librium y∞ = K of the difference equation (4.34) corresponding to the logistic
differential equation is not stable, unlike both the true solution and the Euler
method approximation with h = 1. Figure 4.52 graphs the approximations
u(t) based on the Euler method and modified Euler method along with the
true solution y(t) for the given interval; one can see the instability evidenced
by the oscillation as the modified Euler method solution approaches K. There-
fore, although the modified Euler method gives a much better approximation
here during the initial period of growth (say 0 ≤ t ≤ 5) than the simple Euler
266 Dynamical Systems for Biological Modeling: An Introduction

method, it fails to converge to the carrying capacity K and hence provides a

markedly worse estimate for the settled period t > 5. 

The final numerical method we shall mention in this section is perhaps the
most well known for this purpose: the fourth-order Runge-Kutta method (also
called RK4), which uses the approximation

K1 + 2K2 + 2K3 + K4
yk+1 = yk + h ,
6
where  
h h
K1 = f (tk , yk ), K2 = f tk + , yk + K1 ,
2 2
 
h h
K3 = f tk + , yk + K2 , K4 = f (tk + h, yk + h K3 ).
2 2
An interpretation of these terms is that K1 is the same approximation to y ′ (tk )
given by the Euler method, K2 is a first approximation to y ′ (tk + h2 ) (the slope
halfway between tk and tk+1 ) based on K1 , K3 is a second approximation to
y ′ (tk + h2 ) based on K2 , and K4 is an approximation to y ′ (tk+1 ) (the slope at
the end of the interval) based on K3 . yk+1 then uses a weighted average of
these estimates of the slope y ′ (t) over the time interval [tk , tk+1 ].
It is possible to show that the discretization error at each step is no greater
than a constant multiple of h5 for this method, so that the accumulated dis-
cretization error is no greater than a constant multiple of h4 . This considerable
improvement is evident in the following example.

Example 4.
Use the fourth-order Runge-Kutta method with a step size of 1 to approximate
the solution of the logistic equation y ′ = ry(1−y/K) on the interval 0 ≤ t ≤ 10
for r = 1, K = 10, and y0 = 1. Compare the accuracy to that of Euler’s
method and the modified Euler method using the same step size.
Solution: Here we must first calculate the four Ki prior to finding each yk . We
therefore proceed as follows: To calculate y1 , we find
K1 = ry0 (1 − y0 /K) = (1)1(1 − 1/10) = 0.9,
h
K1 1 − y0 + h2 K1 /K
` ´ˆ ` ´ ˜
K2 = r y0 + 2
= (1)1.45(1 − 1.45/10) = 1.23975,
h
K2 1 − y0 + h2 K2 /K
` ´ˆ ` ´ ˜
K3 = r y0 + 2
≈ (1)1.62(1 − 1.62/10) ≈ 1.35748,
K4 = r(y0 + K3 )(1 − (y0 + K3 )/10) ≈ (1)1.68(1 − 1.68/10) ≈ 1.80171,

and then finally obtain

K1 + 2K2 + 2K3 + K4 0.9 + 2(1.24) + 2(1.36) + 1.80
y1 = y0 +h = 1+1 ≈ 2.31603.
6 6
If we compare the discretization errors involved in just this first time-step
across all three of the methods presented so far, we find errors [y1 − y(1)]/y(1)
 Nonlinear Differential Equations 267

of 18.09% for Euler’s method (and the modified Euler method) but only 0.16%
for RK4. This dramatic (by a factor of 100) improvement in accuracy continues
throughout the interval of interest, with the error in later time-steps never even
reaching 1/5 of 1% for RK4, whereas the average magnitude of error is 7%
for Euler’s method and almost 6% for the modified Euler method over the
interval [0,10]. Table 4.4 gives the true solution and the approximations for
all three methods at each step on this interval, rounded to 4 decimal places.


TABLE 4.4: Comparison of the true solution to the logistic equation of

Examples 1, 2, and 3 with the approximations (h = 1) generated by Euler’s
method, the modified Euler method, and RK4.
t True solution Euler method Modified Euler RK4
(k) y(t) yk method yk yk
0 1 1 1 1
1 2.3197 1.9 1.9 2.32
2 4.5085 3.44 4.0780 4.5007
3 6.9057 5.6953 6.7300 6.8976
4 8.5849 8.1470 8.4794 8.5759
5 9.4283 9.6566 9.3087 9.4187
6 9.7818 9.9882 9.7665 9.7747
7 9.9186 10.0000 9.7649 9.9144
8 9.9699 10.0000 10.2257 9.9678
9 9.9889 10.0000 9.3033 9.9879
10 9.9959 10.0000 11.5220 9.9955

In[1]:= r=1; K=10; y0=1;

In[2]:= truesol=NDSolve[{y’[t]==r*y[t]*(1-y[t]/K),y[0]==y0},
{t ,0,10}]
Out[2]= {{y[t] → InterpolatingFunction[{{0.,10.}},<>] [t]}}
In[3]:= truey[t ]=y[t]/.truesol[[1]]
Out[3]= InterpolatingFunction[{{0.,10.}},<>][t]
In[4]:= truey[2]
Out[4]= 4.50854

FIGURE 4.53: Sample syntax for solving a differential equation numerically

in Mathematica.

Common computational software normally uses even more sophisticated

methods with variable step sizes h from one sub-interval to the next and
even variable order (RK4 is fourth-order because it involves four levels of
estimation at each step). Figures 4.53, 4.54 and 4.55 show the syntax used to
obtain numerical solutions of differential equations in Mathematica, Maple,
268 Dynamical Systems for Biological Modeling: An Introduction

> r=1; K=10; y0=1;

> p:= dsolve({D(y)(x) = r*y(x)*(1-y(x)/K), y(0)=y0}, y(x),
type=numeric):
> p(2);
x = 2, y(x) = 4.50854...

FIGURE 4.54: Sample syntax for solving a differential equation numerically

in Maple.

in file MAIN.M:
in file LOGISTIC.M:
t0 = 0;
function yout = logistic(t,y)
tf = 10;
y0 = 1;
r = 1;
[t,y] = ode45(’logistic’,[t0
K = 10;
tf],y0);
yout = r*y*(1-y/K);
ans = interp1(t,y,2)

FIGURE 4.55: Sample syntax for solving a differential equation numerically

in MATLAB.

and MATLAB, respectively. Note that in each case the precise step size and
method are left implicit, although documentation accompanying each software
package describes the methods used. Here we solve the logistic differential
equation used in the foregoing examples.
We may now consider some examples for which the point is not the me-
chanics of the approximation method, but rather the insights that it produces.

Example 5.
In the budworm model (4.29) of Section 4.4, with the parameter values as
given in Table 4.3 (and A = αS, K = κS as discussed therein), suppose that
an initial budworm density is measured of approximately 30,000 larvae/acre
for a given forest.
(a) Find the range of forest densities [S1 , S2 ] (cf. Figure 4.50) for which a
hysteresis loop exists.
(b) If the forest density is 500 branches/acre, what will the budworm
density be in 1 year? 2 years? How long will it take for the budworm density
to be within 10% of its equilibrium value?
(c) If the forest density is 5000 branches/acre, what will the budworm
density be in 1 year? 2 years? How long will it take for the budworm density
to be within 10% of its equilibrium value?
(d) Interpret and explain the differences in the answers to (b) and (c).
Solution: (a) Using the parameter values, we can use a computer or graph-
ing calculator to graph S as a function of the equilibrium rescaled budworm
 Nonlinear Differential Equations 269

Log10 [y[t]]
6
5
4
3
2
1
t
2 4 6 8 10

FIGURE 4.56: Solutions to the differential equation (4.29) showing the log
budworm densities of Example 5 over time, for forest densities of 500 (solid
curve) and 5000 (dashed curve) branches/acre.

density x, as described in Section 4.4, and find the two places where the
graph has a horizontal tangent. These prove to be (x, S) = (0.0031366, 12842)
and (0.49998, 320.22), so that S1 = 320.22 branches/acre and S2 = 12842
branches/acre.
(b) Using a computer to solve the differential equation numerically, we find
that y(1) ≈ 20, 644 larvae/acre, y(2) ≈ 39 larvae/acre, and it will take about
3.36 years for the population density to come within 10% of its equilibrium
value of 10.84 larvae/acre.
(c) Numerical solution with S = 5000 branches/acre indicates that y(1) ≈
36, 620 larvae/acre, y(2) ≈ 63, 613 larvae/acre, and it will take about 6.12
years for the population density to come within 10% of its equilibrium value
of 1.75 million larvae/acre.
(d) In the smaller forest (S = 500 branches/acre), the budworm density
drops by a factor of 1,000 in just 2 or 3 years, while in the larger forest
(S = 5000 branches/acre), the same initial density leads to a slower growth,
by a factor of nearly 100. Both forest densities fall within the hysteresis range,
but in one case the initial density given falls below the separatrix (the unstable
equilibrium), causing the density to crash, while in the other that same initial
density lies above the separatrix, leading the density to skyrocket to a major
infestation. In ecological terms, the tenfold increase in forest density allows the
given budworms to reproduce beyond the predators’ ability to keep them in
check, whereas the lower forest density does not provide an abundant enough
food supply for reproduction to overcome predation. Figure 4.56 shows the
two populations over the first ten years, graphed on a logarithmic scale since
the equilibria are orders of magnitude apart. 
270 Dynamical Systems for Biological Modeling: An Introduction

Photo courtesy River Alliance of Wisconsin Photo by Jeff Schmaltz, MODIS/Courtesy NASA

FIGURE 4.57: The surface of this lake FIGURE 4.58: This satellite
at Mounds Dam, Wisconsin is covered photo from June 2008 shows phy-
with algae blooms caused by eutroph- toplankton blooms from exten-
ication. Excess nitrogen and phospho- sive eutrophication in the Sea of
rus runoff from fertilizers make microbe Azov (upper right) and where the
populations mushroom, removing oxygen Danube River empties into the
from the water and creating dead zones Black Sea (upper left). The pale
where neither plants nor animals can live. swirls in the Black Sea reflect sed-
imentation.

4.5.4 Eutrophication
Aquatic ecosystems are affected strongly by the concentrations of many
different substances in the water, which determine the ability of plant and an-
imal species to survive. Some of the substances found in bodies of water serve
to feed the organisms in the ecosystem, in particular oxygen and various nutri-
ents, while others may be toxic. The extent to which sunlight is blocked may
also affect plants that use photosynthesis. The level of nutrients in the water
is characterized on a spectrum between oligotrophic, a state with relatively
low levels of nutrients and plant production, and consequently clear water,
and eutrophic, a state with high levels of nutrients and plant production, and
consequently murky water. Eutrophic lakes actually suffer from their over-
abundance of nutrients, as phytoplankton (microscopic photosynthetic plants
that live near the water’s surface) and other microbes use the excess nutrients
to reproduce exponentially, creating so-called blooms which deplete the oxy-
gen in the water that larger plants and animals need in order to survive. Left
unchecked, the eutrophication process can ultimately wipe out an ecosystem.
Eutrophication of lakes and even larger bodies of water has become a
major problem in the past century because of the rise of chemical fertilizers
in agriculture and urban lands, which pump high concentrations of nitrogen
and phosphorus into the soil, much of which runs off into the local water
table, streams, rivers, and lakes. In the twenty-first century, eutrophication
has been observed not only in ponds and lakes (see Figure 4.57) but in the
world’s seas and oceans (see Figure 4.58), including the Sea of Azor and the
 Nonlinear Differential Equations 271

Gulf of Mexico. Because the effects of eutrophication can be devastating and in

some cases irreversible, it is important to understand how each of the various
contributing factors affects eutrophic dynamics.
Carpenter, Ludwig and Brock (1999)22 studied a model for the amount of
phosphorus in a lake over time, as a function of three factors: runoff, natu-
ral clearance, and recycling. The primary source of new phosphorus is from
runoff, and is assumed to enter the lake at a constant rate (L). The lake’s
ability to flush the excess (free) phosphorus out is proportional to the con-
centration of phosphorus in the water. The recycling of phosphorus, however,
is a major factor in sustaining eutrophic environments: Living organisms in
the water ingest phosphorus and other nutrients, taking them out of the wa-
ter and binding them into themselves, notably the benthic plants along the
lake bottoms. When eutrophication begins, the multiplication of microbes on
the surface kills these other organisms, and the dead matter that remains
then decomposes, releasing the phosphorus back into the ecosystem. At low
phosphorus levels, little of this recycling occurs, because the lake is in an
oligotrophic state favorable to larger organisms, but at some point there is a
threshold beyond which eutrophication causes recycling of phosphorus from
sediment at a high rate. In eutrophic systems, the rate at which phosphorus
is recycled in this way may even exceed the rate at which it accumulates due
to runoff. In such cases the eutrophication may become irreversible, as even
if the runoff is halted altogether, the recycling can sustain the phosphorus
concentration at eutrophic levels.
Carpenter et al. therefore used a Holling Type III, or sigmoid (S-shaped),
form to describe the recycling rate, as Ludwig et al. did to describe spruce
budworm predation as described in Section 4.4:
Pq
,
mq + Pq
for phosphorus level P , where m is the 50% saturation constant (analogous to
the constant A in the budworm model) and q is an exponent greater than 1.
This is again a Hill function (of order q), which remains very low for P < m
and quickly rises to a maximum (of 1) for P > m. We can think of m as the
threshold level of phosphorus above which microbial blooms kill the other,
larger organisms in the ecosystem. We can therefore write the model as
dP Pq
= L − sP + r q , (4.35)
dt m + Pq
where s is the phosphorus elimination rate and r is the maximum phospho-
rus recycling rate. The higher-order terms P q complicate the analysis of this
model, which is why we introduce it here, as an appropriate subject for nu-
merical analysis.
22 S.R. Carpenter, D. Ludwig, W.A. Brock, Management of eutrophication for lakes sub-

ject to potentially irreversible change, Ecological Applications 9(3): 751–771, Aug. 1999.
272 Dynamical Systems for Biological Modeling: An Introduction
fi [p]

L+r

L P
0

P
oligotrophic m eutrophic 0 P

FIGURE 4.59: A graphical equilib- FIGURE 4.60: The two possible

rium analysis for (4.35) reveals either phase portraits for (4.35).
one or three equilibria, depending on
the slope of f2 (P ) (short dashes) and
the vertical offset L (long dashes).

If we attempt a qualitative analysis, although we cannot in general write

expressions for any equilibria the model may have, we can fairly quickly de-
termine the possible numbers of equilibria and their stability, using graph-
ical methods. If we group the positive (i.e., runoff and recycling) terms
in (4.35) together, we can write dP/dt = f1 (P ) − f2 (P ), where f1 (P ) =
L + rP q /(mq + P q ) > 0 and f2 (P ) = sP > 0. Then equilibria of (4.35) are
intersections of the two functions, f1 (P ) = f2 (P ). We can consider the pos-
sible intersections graphically (Figure 4.59). The graph of f1 is the sigmoid
recycling function, shifted vertically by the constant L, while the graph of f2
is simply a straight line through the origin with slope s > 0. As the figure
shows, depending upon the slope s of the line and the vertical offset L, there
are three possibilities:
• for high slopes s (fast clearance rates), an oligotrophic equilibrium fea-
turing little or no recycling;
• for low slopes s (slow clearance rates), a eutrophic equilibrium with
recycling near its maximum;
• for intermediate slopes s and vertical offset L within a given range,
there may be three equilibria—one oligotrophic, one eutrophic, and one
in between.
We can use phase line analysis to determine the local asymptotic stability
of these equilibria, armed with the observations that (for dP/dt = f (P ))
f (0) = L > 0 but f (P ) < 0 for any P ≥ L+r s . Therefore (see Figure 4.60)
if there is only one equilibrium, it is globally asymptotically stable, while if
there are three, the first and last are locally asymptotically stable, and the
middle one is unstable.
Carpenter et al. classified lakes into one of three categories depending
upon the values of s and L. If the slope (phosphorus clearance rate) s is
 Nonlinear Differential Equations 273
fi [p] fi [p] fi [p]

Lmin Lmin Lmin

P P P

FIGURE 4.61: Graphical descriptions of Carpenter et al.’s three categories

of lakes: (a) reversible, (b) hysteretic, (c) irreversible. Hysteretic lakes may
also have only an oligotrophic equilibrium when L = Lmin .

higher than the maximum slope (derivative) of the recycling function,23 then
the line can only intersect the sigmoid curve once, and for low enough L the
only equilibrium will be oligotrophic (Figure 4.61(a)). Lakes such as these
are called reversible, since lowering the phosphorus runoff rate L will reverse
eutrophication. Reversible lakes tend to be deep and cool (a notable example
is Lake Washington, in the U.S. state of Washington).
If the lake’s phosphorus clearance rate s is not higher than the recycling
rate’s maximum slope, then the potential for reversing eutrophication depends
upon the extent to which the runoff rate L can be reduced. It is not possible
in practice to eliminate all phosphorus runoff, and if the minimum runoff
rate that can be achieved, Lmin , is too high, then there may never be an
oligotrophic equilibrium (Figure 4.61(a)). Such lakes are called irreversible,
since eutrophication is permanent even under the strongest restrictions on
runoff. These lakes are often shallow and experience high phosphorus runoff
rates (a well-studied example is Shagawa Lake in the U.S. state of Minnesota).
Finally, lakes whose clearance rate s falls below the critical level for re-
versibility, but whose minimum phosphorus runoff rate Lmin is low enough
to permit an oligotrophic equilibrium (Figure 4.61(b)), are called hysteretic,
as they undergo hysteresis as a function of runoff control. Such lakes, typ-
ically small and shallow with high recycling rates, can have eutrophication
permanently reversed through a temporary reduction in runoff rates.
It is possible to carry out further bifurcation analysis (see Exercise 14), but
not to find explicit expressions for the equilibrium values, nor conditions under
which three equilibria exist (hysteresis), so we turn to numerical analysis for
further insights.
Carpenter et al. studied the particular case of Lake Mendota, which is
immediately adjacent to the campus of the University of Wisconsin,24 in the
center of the city of Madison. They measured parameter values of Lmin =
3800 kg/yr, s = 0.817/yr, r = 731, 000 kg/yr, m = 116, 000 kg, q = 7.88,
and in a related study, Lathrop et al. (1998)25 published records of the annual
23 See Exercise 13 for the derivation of this maximum slope.
24 where both they and the authors of this book worked
25 Richard C. Lathrop, Stephen R. Carpenter, Craig A. Stow, Patricia A. Soranno, and
274 Dynamical Systems for Biological Modeling: An Introduction

phosphorus input into Lake Mendota during the period 1976–1996, which had
an average input rate of L = 34, 000 kg/yr, a range of 15,000–67,000 kg/yr,
and a mean phosphorus load of 57,600 kg.
If we use these parameter values in our model (4.35), we find Lake Men-
dota to be in a hysteretic state, with equilibria at 41,900 kg (oligotrophic),
77,500 kg (unstable), and 936,000 kg (eutrophic). However, during 14 of the
21 years observations were made, the estimated phosphorus load was in the
range 44,000–70,000 kg. Therefore, it is likely that the lake is closer to an
equilibrium value like the reported mean of 57,600 kg. Indeed, Carpenter et
al. estimated as much as a tenfold uncertainty in the value of the maximum
recycling rate r (that is, the true value of r might be up to ten times as
much as their estimate, or as little as one tenth), and a 100-fold uncertainty
in the value of the eutrophication threshold m. There is likewise some un-
certainty in L, since Lathrop et al. measured phosphorus runoff coming from
the principal streams feeding the lake, but could not measure that coming
directly from some urban areas next to the lake shore. Changes in the param-
eters r and m are not likely to affect the value of the oligotrophic equilibrium
much (cf. Figure 4.61(b)), but we can easily estimate the value of L which
would correspond to an equilibrium at the observed level: At equilibrium, we
q
have L = sP − r mqP+P q = 43, 873 kg/yr, which also generates equilibria at
69,900 kg (unstable) and 948,000 kg (eutrophic). This suggests that the lake
is currently in a largely oligotrophic state, although a one-time influx of as
little as 69, 900 − 57, 600 = 12, 300 kg would be enough to cause the lake to
experience eutrophication toward the upper equilibrium. Numerical analysis
reveals that this change would probably not be detected immediately simply
from data: in the first five years, the phosphorus level would only increase by
1000 kg, after seven years it would be up by about 5500 kg (still less than
10%), but after eleven years it would be within 10% of the 948,000 kg eu-
trophic equilibrium. (We invite the reader to duplicate these results using any
standard numerical software and plot the graph of P (t).)
In terms of management, we can also see that aggressive reduction of
phosphorus runoff (L = Lmin ) could reduce the oligotrophic equilibrium by
more than an order of magnitude, down to about 4650 kg. In such a situation,
beginning at the present level of 57,600 kg, we can use numerical analysis on
our model (4.35) to find that it would then take about 2.81 years to reduce
the phosphorus content in the lake to below 10,000 kg.
Numerical analysis allows us to quantify the rates of change in this study
in ways that help us recognize it in the data we collect. Carpenter et al. go
further to study the economics of managing the phosphorus content in the lake:
certain economic benefits that result from having an oligotrophic lake (such as
recreational activities and the health benefits associated with water quality)
compete with other economic benefits that generate much of the phosphorus

John C. Panuska, Phosphorus loading reductions needed to control blue-green algal blooms
in Lake Mendota, Canadian Journal of Fisheries and Aquatic Sciences 55: 1169–1178, 1998.
 Nonlinear Differential Equations 275

runoff (such as agricultural products, farming income, and weed-free lawns).

It is difficult to negotiate the tradeoffs between these costs and benefits, but
understanding the effects of our collective choices helps provide a framework
for making decisions.

Exercises
1. Estimate e by using the modified Euler method to approximate y(1)
where y(t) is the solution of the initial value problem y ′ = y, y(0) = 1.
Compare your results with those of Example 2 using h = 0.1.
In Exercises 2–7, use the technology available to you to approximate the so-
lution of the initial value problem for the given interval.
2. y ′ = −y, y(0) = 1; 0 ≤ t ≤ 5 5. y ′ + 2ty = sin t, y(0) = 1; 0 ≤ t ≤ 1
1
3. y ′ = sin y, y(0) = 1; 0 ≤ t ≤ 1 6. y ′ = |y| 2 , y(0) = 1; 0 ≤ t ≤ 1
1
4. y ′ = sin y, y(0) = 0; 0 ≤ t ≤ 1 7. y ′ = |y| 2 , y(0) = 0; 0 ≤ t ≤ 1

8. What is the maximum allowable step size h that would ensure that the
approximations generated by Euler’s method for solutions of the logistic
equation with parameters as given in Example 1 approach the (correct)
equilibrium?
9. Another relatively simple numerical method which, like the modified
Euler method, uses the idea of the derivative at a subinterval’s mid-
point being a good estimate for the derivative throughout the entire
subinterval, is the midpoint method, which approximates the differential
equation dy/dt = f (t, y) with the difference equation
 
h h
yk+1 = yk + hf tk + , yk + f (tk , yk .
2 2

Compare the accuracy (in terms of discretization error) of this method

with those of the three methods explored in this section, by
(a) using it to solve the logistic equation with r = 1, K = 10, y(0) = 1;
(b) using it to approximate the value of e as in Exercise 1.

10. Use numerical methods to find how long it would take an ant colony
of size 1000 ants governed by equation (4.1) to reach 90% engagement
in foraging, if initially only 25 ants find the source and establish the
trail. Use Beekman et al.’s parameter values of α = 0.0052/min., β =
0.00125/ant-min., s = 1 ant/min., and K = 10 ants.
11. Repeat the previous exercise with a colony size of 500 ants, and conjec-
ture a biological explanation for the difference in times.
276 Dynamical Systems for Biological Modeling: An Introduction

12. In the foraging ant model (4.1) using the parameter values suggested
in Section 4.1 carry out simulations for values of K close to the criti-
cal threshold value determined in Section 4.1 to exhibit the change in
behavior at the threshold.
13. For the eutrophication model of Carpenter et al., find the maximum
slope of the phosphorus inflow rate f1 (P ) by first finding the inflection
point P̄ at which f1′′ (P̄ ) = 0, and then calculating f1′ (P̄ ). (Your answer
should be r/m times a function of q.) Phosphorus clearance rates s
greater than this critical value indicate eutrophication-reversible lakes.
∗
14. For the eutrophication model of Carpenter et al., use rescaling and
the reverse parametrization method applied to the budworm model in
Section 4.4 to identify the parameter values for which there exist three
equilibria:
(a) Rescale both the state variable and the time variable to reduce
the number of model parameters to two. (Think about meaningful
benchmarks by which to rescale each variable.) Write the rescaled
differential equation using the new parameters a (replacing L) and
b (replacing s).
(b) Use the fact that at a bifurcation point not only dp/dt = f (p) = 0
but also f ′ (p) = 0, to write two equations which are linear in the
two new parameters a and b. Solve these equations to find expres-
sions for a and b as functions of p (this is the reverse parametriza-
tion).
(c) Show that (a, b) moves from (−1, 0) to (q, 0) for 0 ≤ p < ∞, with
a increasing in p while b hits a (positive) maximum followed by a
(negative) minimum.
(d) Using the known zeroes of b(p), show that the reverse-parametrized
curve (a(p), b(p)) passes through the positive quadrant if and only
if q > 2. What is the significance of this result?

15. In their study of eutrophication of Lake Mendota, Carpenter et al. es-

timated a tenfold uncertainty in their value for the maximum recycling
rate r, and a 100-fold uncertainty in their value for the eutrophication
threshold m.
(a) Evaluate the worst-case scenario: that is, imagine their values re-
ally are off by respective factors of 10 and 100. What are the cor-
responding equilibria? What kind of lake would this be?
(b) Now evaluate the best-case scenario: again, their values for r and
m are off by factors of 10 and 100, respectively, but in the other
direction. What are the equilibria and lake classification?
(c) Judging by the equilibrium value in each case, to which extreme is
the present situation closer?
 Nonlinear Differential Equations 277

Miscellaneous exercises
For each of the differential equations in Exercises 1–4, draw the phase line,
find the equilibria, and describe the asymptotic behavior of solutions.
2
1. y ′ = y 2 (y − 1)
y
2. y ′ = 1−y

3. y ′ = ye−y
4. y ′ = y sin y
5. (a) Consider the logistic differential equation
 y
y ′ = ry 1 − , y(0) = A
K
with K = 100, A = 25 and several different values of r including
r = 1.5, r = 2.5, r = 3.5, and r = 4.5. Describe the behavior of the
solution for each value of r qualitatively.
(b) Consider the discrete logistic equation
 yk 
yk+1 = ryk 1 − , y0 = A
K
with K = 100, A = 25 and several different values of r including
r = 1.5, r = 2.5, r = 3.5, and r = 4.5. Describe the behavior of the
solution for each value of r qualitatively.
(c) Consider the alternative discrete version
yk
yk+1 = ryk e(1− K ) , y0 = A

with K = 100, A = 25 and several different values of r including

r = 1.5, r = 2.5, r = 3.5, and r = 4.5. Describe the behavior of the
solution for each value of r qualitatively.
Approximate the solutions of the problems in parts (a), (b), (c) numer-
ically for each given value of r. Compare and contrast the behaviors of
the three solutions for each r.
∗
6. Carry out the method of separation of variables for equation (4.1), as
follows:
(a) Rewrite the equation dx
R dx R
dt = f (x) in the form f (x) = dt, and then
rewrite the left-hand integrand as a rational polynomial, i.e., write
1/f (x) as A(x)/B(x), where A(x) and B(x) are both polynomials.
278 Dynamical Systems for Biological Modeling: An Introduction

(b) Use the cubic formula to factor B(x); depending upon parameter
values, it will have either one or three real roots. Next, apply the
method of partial fraction expansion to A(x)/B(x). In the case
where B(x) has three real roots, you can rewrite A(x)/B(x) as the
sum of three terms of the form Ai /(x−ri ) (where i = 1, 2, 3). In the
case where B(x) has only one real root, you can rewrite A(x)/B(x)
A1
as the sum x−r + x2A+B
2 x+A3
1 x+B2
. (You may have to look up the details
for both of these techniques.)
(c) Carry out the integration.
(d) Exponentiate both sides of the equation, and use the initial condi-
tion x(0) = x0 to rewrite any constant of integration.
7. Explain why no two consecutive equilibria can both be (locally) asymp-
totically stable. (Consider the behavior of solutions on the interval be-
tween them.)
8. Suppose a population of whales grows according to an exponential
growth law but with constant-effort harvesting. Then the population
size satisfies a differential equation y ′ = ry − Ey. (a) If r = 0.04/yr,
what catch per year will maintain the population at a constant size of
8000 whales? (Note the catch is not E but Ey.) (b) If harvesting takes
place for ten years with E = 0.06, how long will it take after harvesting
ceases until the population returns to its original size of 8000 whales?
9. A forest of 10,000 acres suffers from forest fires which destroy an average
of 1% of its area annually. If the natural growth rate is r = 0.05/yr, find
the equilibrium forest size.
∗
10. (a) Show that the minimum survival threshold y1 in the constant-
yield harvesting model has y1 > H/r for all values of K. (Hint: Work
backward.)
(b) Use calculus to show that y1 → H/r as K → ∞: that is,
" r #
1 4HK H
lim K − K2 − = .
K→∞ 2 r r

11. In the mid-1990s, the Canadian harp seal population was estimated at
4.8 million seals. In 1997 an estimated 344,000 harp seals were killed.
The harp seal’s reproductive rate has been estimated at 7%/year. (a)
Under a constant-yield scenario, what is the minimum overall carrying
capacity, if the observed harvest is sustainable? (b) Use the result of
Exercise 10 to evaluate the sustainability of this hunt.
12. In formulating the epidemic model (4.20), we assumed that individuals
mix randomly in making contacts with others, so that a proportion I/N
 Nonlinear Differential Equations 279

of anyone’s contacts are made with infectives, and a proportion S/N

are made with susceptibles. If the symptoms of the disease are severe
enough, however, it may be more appropriate to assume that infected
individuals’ contact rates are reduced by some factor σ (0 < σ < 1) to σβ
instead of just β, either because they are so ill they must remain in bed,
resting, or because they deliberately avoid contact with others whom
they might infect. In this case, at what rate would new infections occur?
What is the resulting model? Hint: First determine the new effective
population size for contact rate purposes—it’s less than N .
∗
13. In the previous exercise with less active infectives, determine R0 and
the endemic equilibrium under the assumption of (a) mass-action inci-
dence, (b) standard incidence.
14. In formulating the epidemic model (4.20), we assumed that infected
individuals recover from the disease at a given rate (1/τ ). If, instead,
infected individuals never recover (as with HIV), (a) what happens to
the model? to the expression for R0 ? Hint: Consider what happens to τ
under this assumption. (b) What eventual outcome does the model then
predict? Why is this outcome inconsistent with the observed data for
unrecoverable infections such as HIV?
15. Describe the behavior of solutions of the differential equation y ′ = r − y 2
as the parameter r is varied.
16. Describe the behavior of solutions of the differential equation y ′ = ry−y 2
as the parameter r is varied.
17. Draw a bifurcation diagram for the two epidemic models in Exercise 11
of Section 4.3.
18. (a) Show graphically that all equations of the form y ′ = p − xn exhibit
either a saddle-node bifurcation at the origin, if n is even, or no bifur-
cation at all, if n is odd. (b) Show graphically that all equations of the
form y ′ = px − xn exhibit either a transcritical bifurcation at the origin,
if n is even, or a pitchfork bifurcation at the origin, if n is odd. (c) What
changes occur in (a) and (b) above when the − sign in the equation
becomes a +?
This page intentionally left blank
 Part II

More Advanced Topics

281
This page intentionally left blank
Chapter 5
Systems of Differential Equations

In Chapter 3 we saw how to model and analyze continuously changing quan-

tities using differential equations. In many applications of interest there may
be two or more interacting quantities — populations of two or more species,
for instance, or parts of a whole, which depend upon each other. When the
amount or size of one quantity depends in part on the amount of another, and
vice versa, they are said to be coupled, and it is not possible or appropriate
to model each one separately. In these cases we write models which consist of
systems of differential equations. In this chapter, we will find that the quan-
titative and qualitative approaches we used to analyze individual differential
equations in Chapters 3 and 4 extend in a more or less natural way to cover
systems of differential equations. Extending them will require some basic mul-
tivariable calculus, principally the use of partial derivatives; we provide some
support in the worked examples for the unfamiliar reader, but recommend
using a calculus text for more complete reference.
We begin below with an important example from population biology, a
predator-prey system, in which one species preys upon another for its food
supply.

5.1 Graphical analysis: The phase plane

One of the landmarks in the development of mathematical ecology is the
Lotka-Volterra model for the population sizes of two interacting species. About
1925 the famous Italian mathematician Vito Volterra was asked if it was possi-
ble to give a scientific explanation for the large fluctuations in fish populations
in the Adriatic Sea. These fluctuations were of great concern to fishermen,
both in periods of low population sizes when fish catches were small and there
was little income for the fishing industry, and in periods of high population
sizes when fish were abundant and the large supply made selling prices low.
Volterra constructed a simple model which has become known as the Lotka-
Volterra model (because A. J. Lotka constructed a similar model about the
same time in a different context), based primarily on the hypothesis that fish
and sharks were in a predator-prey relationship (see Figure 5.1). Even though
it turned out later that this model was unreasonably simplistic and not re-

283
284 Dynamical Systems for Biological Modeling: An Introduction

Photo by Mark Conlin, Southwest Fisheries Science Ctr, NOAA Fisheries Svc Photo by Stefano Guerrieri

FIGURE 5.1: The most common sharks and fish in the Adriatic Sea, the
subject of Volterra’s landmark study, are the blue shark, Prionace glauca (left),
and members of the goby family (pictured at right is one example, Didogobius
schlieweni).

ally consistent with observations, it gave great insight into the modeling of
predator-prey relationships and led to many important developments.
Let y(t) denote the number of fish and z(t) the number of sharks at time
t. We make the quite unreasonable simplifying assumption that all fish are of
the same kind and that all sharks are identical.1 We assume that the plankton
on which the fish feed is available in unlimited quantities, and thus that the
fish population would grow exponentially in the absence of sharks. Then, if
there were no sharks, the fish population would satisfy an exponential growth
equation
y ′ = λy.
The sharks, on the other hand, are assumed to depend on the fish as their food
supply. We assume that in the absence of fish the shark population would die
out exponentially, and thus that the shark population satisfies an exponential
decay equation
z ′ = −µz
if there are no fish. To describe the interaction between sharks and fish, we
assume that the presence of fish produces a linear increase in the per capita
shark growth rate, and the presence of sharks produces a linear decrease in
the per capita fish growth rate. That is, the rate of encounters between fish
y and sharks z is proportional to the size of each group, making it a multiple
of yz. Thus we assume per capita growth rates of λ − bz for fish and −µ + cy
for sharks, where b and c are constants which describe how sharply the shark
population affects the fish reproduction, and vice versa. This gives us a system
1 This assumption will allow us to write a reasonably simple model and investigate

whether the basic idea of interdependence can account for the observed fluctuations.
 Systems of Differential Equations 285

of two first-order differential equations

y ′ = y(λ − bz) (5.1)

z ′ = z(−µ + cy)

called the Lotka-Volterra equations.

We cannot solve the system (5.1) analytically, but we can use qualitative
tools to examine it. In addition to the techniques developed in the previous
chapter, which we will apply presently, we can now also eliminate t and reduce
the system to a single equation, by using the relation
dz dz . dy
=
dy dt dt

from calculus (an application of the Chain Rule) to obtain a first-order differ-
ential equation for z as a function of y,

dz z(−µ + cy)
= . (5.2)
dy y(λ − bz)

The differential equation (5.2) has variables separable, and may be solved by
the methods of Section 3.4 (which we will now do).
Separation of variables in (5.2) gives

−µ + cy λ − bz
Z Z
dy = dz,
y z
and integration gives

−µ log y + cy = λ log z − bz + h, (5.3)

where h is a constant of integration. We now have a relation between y and z

which defines z implicitly as a function of y for each choice of the constant of
integration h. (Substituting initial conditions for y and z into (5.3) will allow
us to find the appropriate value for h.) If we define a function V (y, z) of two
variables by
V (y, z) = −µ log y + cy − λ log z + bz,
the relation (5.3) takes the form

V (y, z) = h. (5.4)

That is, as y and z change (over time, but we have made that implicit here),
they will follow the curve given by (5.4), so that the value of V (y, z) remains
constant, at its initial value h.
If we return briefly to (5.1), by setting both equations equal to zero we
can find that the system has two equilibria: y = z = 0 and y = µ/c, z = λ/b.
V (y, z) is undefined at (0,0) — it approaches ∞ as we approach the origin
286 Dynamical Systems for Biological Modeling: An Introduction

or either axis — but defined everywhere in the interior of the first quadrant;
since we saw above that V (y, z) must remain constant, we can conclude that
(y, z) can only approach (0,0) if we begin on one of the axes: in particular,
the z-axis. That is, the fish and shark populations will die out entirely only
in the uninteresting case where there are no fish, and the sharks die out in
the absence of their food supply. (In the remainder of our discussion below we
will assume that both fish and sharks are present.)
At the other equilibrium, however, we find that
µ λ
V (y, z) = h0 ≡ −µ log + µ − λ log + λ.
c b
In fact, we can show that h0 is the minimum value V (y, z) can ever take on,
by examining the first and second derivatives of V , just as with a function of
one variable. In particular, if we take the partial derivatives of V with respect
to y and z, we find2 that

∂V µ ∂V λ ∂2V µ ∂2V λ
= − + c, = − + b, = > 0, = 2 > 0,
∂y y ∂z z ∂y 2 y2 ∂z 2 z

so the equilibrium y = µ/c, z = λ/b is the only place where both ∂V ∂y = 0 and
∂V
∂z = 0. Furthermore, we see from the second derivatives that the function
is concave up in the interior of the first quadrant, so this critical point must
be a minimum.3 (Here we begin to see our first hint of new and interesting
behavior: since the value of V (y, z) must remain constant, and V (µ/c, λ/b)
is different (less) than the values of V for all the points surrounding it, no
trajectories lead toward this equilibrium, either.) Therefore every solution of
the system (5.1) is described by the relation (5.4) with some choice of h ≥ h0 ,
determined by initial conditions.

Recall now that our purpose in writing this model was to see if the in-
terdependence between the fish and shark populations could account for the
fluctuations observed. In order to see that we do have periodic solutions here,
we will extend the notion of a phase line introduced in Section 4.1. Since we
now have a system of two dimensions rather than one, we will speak of the
phase plane, i.e., the y − z plane. For each value of h ≥ h0 , the relation (5.4)
defines implicitly a curve in this plane. More directly, it gives the relation
between the two population sizes from the differential equation (5.2). It does
2 Note to the reader not accustomed to partial derivatives: the process is just like normal

differentiation with respect to the variable in question, while treating all other independent
variables as constants. The notation ∂V /∂y is used above, for instance, instead of dV /dy, in
order to show that V is a function of more than one variable. As a short example, consider
A(y, z) = y 2 sin z + 4. Then ∂A/∂y = 2y sin z, while ∂A/∂z = y 2 cos z. In this case we have
three different second-order partial derivatives: ∂ 2 A/∂y 2 = 2 sin z, ∂ 2 A/∂z 2 = −y 2 sin z,
and ∂ 2 A/∂y∂z = ∂ 2 A/∂z∂y = 2y cos z.
2
3 We can also note that ∂ V = 0, but we are borrowing just enough from multivariable
∂y∂z
calculus here to make our point.
 Systems of Differential Equations 287

not, however, give the populations as functions of the time t. Such a curve is
called an orbit of the system (5.1).
To show that solutions to (5.1) are periodic, we must show that the curves
given by V (y, z) = h are closed. We can, of course, use computer programs
such as Maple, Mathematica or MATLAB to generate numerical solutions
and plot them, and Figure 5.2 below shows such a plot. An analytical way to
show that orbits of (5.1) are closed involves using a polynomial approximation
to the logarithms involved in the formula for V , similar to the linearization
process introduced in Section 4.1 to analyze stability of equilibria. Since it is
often important to be able to show results analytically — here, for example,
to distinguish genuine periodic orbits from spirals which grow or decay very
slowly — we will now consider this approach.
We begin by making the change of variables
µ λ
y= + u, z = + v
c b
in (5.4), to center our system (u, v) around the equilibrium ( µc , λb ). This gives
us
 
µ λ µ  µ 
V + u, + v = −µ log +u +c +u
c b c c
   
λ λ
−λ log +v +b + v = h. (5.5)
b b

Using rules of logarithms, we write

µ     
 µ cu µ cu
log + u = log 1+ = log + log 1 + .
c c µ c µ
2
x
We now use the quadratic Taylor approximation log(1 +  x) ≈ x −  2 (see
Appendix B) for values of x close to 0, to approximate log 1 + µc u ≈ µc u −
c2 2
2µ2 u for values of u close to 0 — that is, for (y, z) close to ( µc , λb ), or values
of V (y, z) = h close to h0 . This gives us
µ  µ c c2
log + u ≈ log + u − 2 u2 ,
c c µ 2µ
b2 2
and similarly we may approximate log λb + v by log λb + λb v − 2λ



2 v . (This

is the same approximation process we would use in linearizing about the equi-
librium, as introduced in Section 4.1, but here we are keeping the quadratic
as well as the linear terms.) When we substitute these approximations into
(5.5) we obtain the curve

c2 2 b2 2
µ  
λ
−µ log − cu + u + µ + cu − λ log − bv + v + λ + bv = h
c 2µ b 2λ
288 Dynamical Systems for Biological Modeling: An Introduction

or
c2 2 b 2 2 µ λ
u + v = h + µ log − µ + λ log − λ = h − h0
2µ 2λ c b
as an approximation to the curve (5.5). This curve, whose equation in the
y − z plane is
2
c2  µ 2 b2

λ
y− + z− = h − h0 , (5.6)
2µ c 2λ b
represents an ellipse with center at ( µc , λb ) as long as h > h0 .
This shows that the curve (5.6) is a closed curve around the equilibrium
( µc , λb ) if h − h0 is small and positive. Since the solution runs around a closed
curve, it must repeat itself and is therefore periodic. Thus the Lotka-Volterra
model predicts periodic fluctuations as had been observed in the shark and
fish populations. It is possible to prove that the√period of oscillation, or time
for a cycle to repeat itself, is approximately 2π/ λµ. The phase portrait (Fig-
ure 5.2), or sketch of orbits in the phase plane, shows that, because the orbit
is traversed in a counterclockwise direction (see Exercise 7 below), the maxi-
mum prey population (reached at the rightmost end of each orbit) occurs one
quarter of a cycle before the maximum predator population (reached at the
top of each orbit).
2

1.8

1.6

1.4

1.2

z 1

0.8

0.6

0.4

0.2

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

y

FIGURE 5.2: Phase portrait for the Lotka-Volterra model.

Now that we have seen an example, let us restate more generally the termi-
nology we shall use in this chapter. The Lotka-Volterra system is an example
 Systems of Differential Equations 289

of a two-dimensional autonomous system of first-order differential equations,

y ′ = F (y, z), (5.7)

′
z = G(y, z).

The phase plane for such a system is the y − z plane.

An equilibrium of the system (5.7) is a solution (y∞ ,z∞ ) of the pair of
equations
F (y, z) = 0, G(y, z) = 0.
Geometrically, an equilibrium is a point in the phase plane. In terms of the
system (5.7), an equilibrium gives a constant solution y = y∞ , z = z∞ of
the system. This definition is completely analogous to the definition of an
equilibrium given for a difference equation in Section 2.3, and for a first-order
differential equation in Section 4.1.
The orbit of a solution y = y(t), z = z(t) of the system (5.7) is the curve
in the y − z phase plane consisting of all points (y(t), z(t)) for 0 ≤ t < ∞. A
closed orbit corresponds to a periodic solution.
There is a geometric interpretation of orbits which is analogous to the
interpretation given for solutions of first-order differential equations in Sec-
tion 3.3. Just as the curve y = y(t) has slope y ′ = f (t, y) at each point (t, y)
along its length, an orbit of (5.7) (considering z as an implicit function of y)
has slope
dz z′ G(y, z)
= ′ =
dy y F (y, z)
at each point of the orbit. The phase plane direction field for a two-dimensional
autonomous system is a collection of line segments with this slope at each point
(y, z), and an orbit must be a curve which is tangent to the direction field at
each point of the curve. Computer algebra systems such as Maple, MATLAB
and Mathematica may be used to draw the direction field for a given system.
For example, Figure 5.3 gives a direction field drawn using Maple for the
Lotka-Volterra system (5.1) with λ = 1, µ = 1, b = 1, c = 1, with the orbits
from Figure 5.2 superimposed on it.

Example 1.
Describe the orbits of the system

y ′ = z, z ′ = −y.

Solution: If we consider z as a function of y, we have dz

dy = z ′ /y ′ = − yz .
Solution by separation of variables gives
Z Z
z dz = − y dy,

2 2
and integration gives z2 = − y2 + c. Thus every orbit is a circle y 2 + z 2 = 2c
with center at the origin, and every solution is periodic. 
290 Dynamical Systems for Biological Modeling: An Introduction

2 2

1.5 1.5

z z
1 1

0.5 0.5

0.5 1 1.5 2 0.5 1 1.5 2

y y

FIGURE 5.3: Direction field and FIGURE 5.4: Direction field and
phase portrait for the Lotka-Volterra nullclines for the Lotka-Volterra
model. model (the coordinate axes are also
nullclines).

To find equilibria of a system (5.7), it is helpful to draw the nullclines,

the curves F (y, z) = 0 on which y ′ = 0, and G(y, z) = 0 on which z ′ = 0.
An equilibrium is an intersection of these two curves. In order to distinguish
between the two nullclines, we will use a solid curve for the y-nullcline and a
dotted curve for the z-nullcline. Thus for the Lotka-Volterra system (5.1) we
may add the y-nullclines y = 0 and z = λb , and the z-nullclines z = 0 and
y = µc to the direction field of Figure 5.3 to obtain Figure 5.4.

Exercises
In each of Exercises 1–6, describe the orbits of the given system.

1. y ′ = yz 2 , z ′ = zy 2 3. y ′ = cos z, z ′ = y
2. y ′ = ez , z ′ = e−y 4. y ′ = yez , z ′ = yzey

5. y ′ = −βyz, z ′ = βyz (see Exercise 7, Section 3.4)

6. y ′ = −αy, z ′ = −βyz (see Exercise 8, Section 3.4)
7. * Show that the orbits of the Lotka-Volterra system (5.1) are traversed
in a counterclockwise direction as t increases. [Hint: For a point on the
orbit with z = λb and y > µc , y ′ = 0 and z ′ > 0; thus at this point y is a
maximum and z is increasing].
8. * Find in which direction the system in Example 1 traverses its orbits.
 Systems of Differential Equations 291

5.2 Linearization of a system at an equilibrium

Sometimes it is possible to find the orbits in the phase plane of a system of
differential equations, as we were able to do for the Lotka-Volterra system in
the preceding section, but it is rarely possible to solve a system of differential
equations analytically. For this reason, our study of systems will concentrate
on qualitative properties. The linearization of a system of differential equations
at an equilibrium is a linear system with constant coefficients, whose solutions
approximate the solutions of the original system near the equilibrium. In this
section we shall see how to find the linearization of a system. In the next
section we shall see how to solve linear systems with constant coefficients, and
this will enable us to understand much of the behavior of solutions of a system
near an equilibrium.
Let (y∞ , z∞ ) be an equilibrium of a system

y ′ = F (y, z), (5.8)

′
z = G(y, z),

that is, a point in the phase plane such that

F (y∞ , z∞ ) = 0, G(y∞ , z∞ ) = 0. (5.9)

We will assume that the equilibrium is isolated, that is, that there is a circle
centered around (y∞ , z∞ ) which does not contain any other equilibrium. We
shift the origin to the equilibrium by letting y = y∞ + u, z = z∞ + v, and
then use Taylor’s theorem for two variables (see Appendix B) to approximate
F (y∞ + u, z∞ + v) and G(y∞ + u, z∞ + v). The difference here between a
one-dimensional system and a two-dimensional one is that the approximation
via Taylor’s theorem in two or more dimensions uses partial derivatives. Our
approximations are

F (y∞ + u, z∞ + v) = F (y∞ , z∞ ) + Fy (y∞ , z∞ )u + Fz (y∞ , z∞ )v + h1 , (5.10)

G(y∞ + u, z∞ + v) = G(y∞ , z∞ ) + Gy (y∞ , z∞ )u + Gz (y∞ , z∞ )v + h2 ,

where h1 and h2 are functions which are “quadratic” in u and v, in the sense
that they are negligible relative to the linear terms in (5.10) when u and v are
small (i.e., close to the equilibrium).
The linearization of the system (5.8) at the equilibrium (y∞ , z∞ ) is defined
to be the linear system with constant coefficients

u′ = Fy (y∞ , z∞ )u + Fz (y∞ , z∞ )v, (5.11)

′
v = Gy (y∞ , z∞ )u + Gz (y∞ , z∞ )v.

To obtain it, we first note that y ′ = u′ , z ′ = v ′ , and then substitute (5.10)

292 Dynamical Systems for Biological Modeling: An Introduction

into (5.8). By (5.9) the constant terms are zero, and for the linearization we
neglect the higher-order terms h1 and h2 . The coefficient matrix of the linear
system (5.11) is the matrix of constants
 
Fy (y∞ , z∞ ) Fz (y∞ , z∞ )
.
Gy (y∞ , z∞ ) Gz (y∞ , z∞ )

In population models, this matrix is often called the community matrix of

the system at equilibrium. Its entries describe the effect of a change in each
variable on the growth rates of the two variables.

Example 1.
Find the linearization at each equilibrium of the Lotka-Volterra system

y ′ = y(λ − bz), z ′ = z(−µ + cy).

Solution: As noted in the previous section, the equilibria are the solutions of
y(λ − bz) = 0, z(−µ + cy) = 0. One solution is the origin, y = 0, z = 0,
representing extinction of both fish and sharks, and a second is y = µc , z = λb ,
representing coexistence. Because the partial derivatives of the functions on
the right side of the system are, respectively,
∂ ∂
[y(λ − bz)] = λ − bz, [y(λ − bz)] = −by,
∂y ∂z
∂ ∂
[z(−µ + cy)] = cz, [z(−µ + cy)] = −µ + cy,
∂y ∂z
the linearization at an equilibrium (y∞ ,z∞ ) is

u′ = (λ − bz∞ )u −by∞ v,
′
v = cz∞ u +(−µ + cy∞ )v.

Thus the linearization at (0,0) is

u′ = λu, v ′ = −µv,

and the linearization at ( µc , λb ) is

bµ cλ
u′ = − v, v ′ = u.
c b
(What these linearized systems tell us about the predator-prey system’s be-
havior is investigated later in this section, cf. Example 3). 

Example 2.
Find the linearization at each equilibrium of the system

y ′ = −βyz, z ′ = βyz − γz.

 Systems of Differential Equations 293

This corresponds to the classical Kermack-McKendrick SIR model for an epi-

demic, which we shall study in Section 6.1. Here y corresponds to the number
of susceptible, uninfected individuals, z to the number of infected, infective
individuals, and β and γ to the infection and recovery rates, respectively.
Solution: The equilibria are the solutions of −βyz = 0, βyz − γz = 0. To
satisfy both these equations, we must have z = 0, but there is no restriction
on y (so we have an entire line of equilibria (y, 0)). Since
∂ ∂
[−βyz] = −βz, [−βyz] = −βy,
∂y ∂z
∂ ∂
[βyz − γz] = βz, [βyz − γz] = βz − γ,
∂y ∂z
the linearization at an equilibrium (y∞ ,z∞ ) is

u′ = −βz∞ u −βy∞ v,
′
v = βz∞ u +(βy∞ − γ)v.

The equilibria are the points (y∞ ,0) with arbitrary y∞ , and the corresponding
linearization is
u′ = −βy∞ v, v ′ = (βy∞ − γ)v.
If we examine the second of these equations, we see that v(t) experiences
either exponential growth or exponential decay, depending on the sign of the
coefficient (βy∞ − γ). If y∞ > γ/β, this coefficient is positive and solutions
to the linearized system exhibit exponential growth in v, which corresponds
to growth in z, the number of infected individuals. Only if y∞ < γ/β does v
decay, corresponding to a drop in the number of infectives. As we will discover
in Section 6.1, this comes from the modeling assumption that infectious con-
tacts, like the predator-prey contacts in the Lotka-Volterra model, occur at
a rate proportional to the product of both populations, yz. Infection control
might therefore concentrate on reducing y, the number of susceptibles who
make potentially infectious contacts; two common ways to do so, illustrated
in Figure 5.5, are vaccination and protection from infectious contact. 

An equilibrium of the system (5.8) with the property that every orbit
with initial value sufficiently close to the equilibrium remains close to the
equilibrium for all t ≥ 0, and approaches the equilibrium as t → ∞, is said to
be asymptotically stable. An equilibrium of (5.8) with the property that some
solutions starting arbitrarily close to the equilibrium move away from it is said
to be unstable. These definitions are completely analogous to those given in
Section 2.4 for difference equations and Section 4.1 for first-order differential
equations.
The fundamental property of the linearization which we will use to study
stability of equilibria is the following result, which we state without proof. The
proof may be found in any text which covers the qualitative study of nonlinear
294 Dynamical Systems for Biological Modeling: An Introduction

Photo courtesy CDC PHIL Photo by Eneas De Troya

FIGURE 5.5: Two common strategies for reducing the number of individuals
at risk of infection in an epidemic are vaccination and protection from contact.
At left, a man vaccinates a child (in his mother’s arms) against smallpox in
the Republic of Chad during a worldwide vaccination program that began in
1967 and lasted over a decade. At right, riders on a Mexico City subway train
wear breathing masks during the 2009 H1N1 influenza epidemic.

differential equations. Here we suppose F and G to be twice differentiable —

that is, smooth enough for the linearization to give a correct picture.

LINEARIZATION THEOREM: If (y∞ , z∞ ) is an equilibrium of the

system
y ′ = F (y, z), z ′ = G(y, z),
and if every solution of the linearization at this equilibrium approaches
zero as t → ∞, then the equilibrium (y∞ ,z∞ ) is asymptotically stable. If
the linearization has unbounded solutions, then the equilibrium (y∞ ,z∞ )
is unstable.

For a first-order differential equation y ′ = g(y) at an equilibrium y∞ , the

linearization is the first-order linear differential equation u′ = g ′ (y∞ )u. We
may solve this differential equation by separation of variables and see that all
solutions approach zero if g ′ (y∞ ) < 0, and there are unbounded solutions if
g ′ (y∞ ) > 0. We have seen in Section 4.1, without recourse to the linearization,
that the equilibrium is asymptotically stable if g ′ (y∞ ) < 0 and unstable if
g ′ (y∞ ) > 0. The linearization theorem is valid for systems of any dimension
and is the approach needed for the study of stability of equilibria for systems
of dimension higher than 1.
Note that there is a case where the theorem above does not draw any
conclusions: the case where the linearization about the equilibrium is neither
asymptotically stable nor unstable. In this case, the equilibrium of the original
(nonlinear) system may be asymptotically stable, unstable, or neither. To see
 Systems of Differential Equations 295

that this case may indeed occur, recall the Lotka-Volterra model we studied
in the previous section, where periodic solutions arose.

Example 3.
Show that the equilibrium (0,0) of the Lotka-Volterra system

y ′ = y(λ − bz), z ′ = z(−µ + cy)

is unstable.
Solution: As we have seen in Example 1, the linearization of the system at the
equilibrium (0,0) is u′ = λu, v ′ = −µv. These two equations can be solved sep-
arately by separation of variables, and every solution of the linearization has
the form u = c1 eλt , v = c2 e−µt . As every solution with c1 6= 0 is unbounded,
the linearization theorem shows that the equilibrium (0,0) is unstable.

If we try to apply the theorem to the interior equilibrium µc , λb , we
must consider the linearization u′ = − bµ ′ cλ
c v, v = b u. Although we have
not yet discussed any way to analyze such linearizations in general, methods
we will discuss later in this chapter will enable us to show that all√solutions
√ linearization are periodic — namely, combinations of sin λµ t and
of this
cos λµ t. Since the linearization’s solutions neither approach zero nor be-
come unbounded, the linearization theorem above does not allow us to draw
any conclusions about the stability of this equilibrium; instead, we must turn
to other methods, such as the phase plane approach we took in the last section.


Example 4.
For each equilibrium of the system

y ′ = z, z ′ = −2(y 2 − 1)z − y,

determine whether the equilibrium is asymptotically stable or unstable.

Solution: The equilibria are the solutions of z = 0, −2(y 2 − 1)z − y = 0, and
∂ ∂
thus the only equilibrium is (0,0). Since ∂y [z] = 0, ∂z [z] = 1, and

∂ ∂
[−2(y 2 − 1)z − y] = −4yz − 1, [−2(y 2 − 1)z − y] = −2(y 2 − 1),
∂y ∂z
the linearization at (0,0) is u′ = 0u + 1v = v, v ′ = −u + 2v. We can actually
solve this system of equations outright by using a clever trick4 to reduce it
4 It has been said that the best way to solve a differential equation is by correctly guessing

the solution. It has also been said that a technique is a trick that can be used more than once.
Many techniques (and tricks) that have been developed to solve differential equations take
advantage of some peculiarity of the particular equations under consideration. The trick used
above certainly falls into that category. However, as our approach in this text concentrates
on qualitative methods, we shall not attempt to provide a list of such tricks, referring
the interested reader instead to a more general introductory text on solving differential
equations.
296 Dynamical Systems for Biological Modeling: An Introduction

to a single equation. We subtract the first equation from the second to give
(v − u)′ = (v − u). This is a first-order differential equation for (v − u),
whose solution is v − u = c1 et . This partial result is already enough to tell
us that the equilibrium is unstable, as since the difference between u and v
grows exponentially, at least one of them must therefore grow exponentially
as well. However, to give a more explicit conclusion we shall complete the
solution: Substitution of v = u + c1 et into u′ = v gives another first-order
linear equation u′ − u = c1 et . The solution of this equation, obtained by the
method of Section 3.5, is u = (c1 t + c2 )et , and therefore v = (c1 t + c1 + c2 )et .
As this has unbounded solutions, we (again) conclude that the equilibrium
(0,0) is unstable. 

Exercises
In Exercises 1–8, find the linearization of the given system at each equilibrium.

1. y ′ = y + z − 2, z ′ = z − y 6. y ′ = z, z ′ = sin y
′ ′
2. y = y + z − 1, z = y 7. y ′ = y(λ − ay − bz),
′ 2 ′
3. y = y + z , z = y + 1 z ′ = z(µ − cy − dz), a, b, c, d > 0
4. y ′ = z − 2, z ′ = y 2 − 8z 8. y ′ = y(λ − ay + bz),
z ′ = z(µ + cy − dz), a, b, c, d > 0
5. y ′ = ez , z ′ = e−y

In each of Exercises 9–12, for each equilibrium of the given system determine
whether the equilibrium is asymptotically stable or unstable.
9. y ′ = −2y, z ′ = −z 11. y ′ = −y, z ′ = y 2 − z
10. y ′ = y, z ′ = −z 12. y ′ = y + z, z ′ = z − 1

13. Show that the equilibrium (0,0) of the system y ′ = y(λ − ay − bz),
z ′ = z(µ − cy − dz) is unstable (see Exercise 7).
14. Show that the equilibrium (0,0) of the system y ′ = y(λ − ay + bz),
z ′ = z(µ + cy − dz) is unstable (see Exercise 8).
15. How would the epidemic model of Example 2 have to change if
(a) the whole population were partially protected from infection?
(b) part of the population were born [permanently] partially protected
from infection?
(c) part of the population were initially temporarily protected?
(d) the outbreak were extended enough to warrant including demo-
graphic renewal (births and deaths) in the model?
In each case write the new model.
 Systems of Differential Equations 297

5.3 Linear systems with constant coefficients

We have seen in the preceding section that the stability of an equilibrium
of a system of differential equations is determined by the behavior of solutions
of the system’s linearization at the equilibrium. This linearization is a linear
system with constant coefficients (recall that a linear system has right-hand
sides linear in the state variables, here y and z). Thus, in order to be able
to decide whether an equilibrium is asymptotically stable, we need to be able
to solve linear systems with constant coefficients. We were able to do this in
the examples of the preceding section because the linearizations took a simple
form with one of the equations of the system containing only a single variable.
In this section we shall develop a more general technique.
The problem we wish to solve is a general two-dimensional linear system
with constant coefficients,

y ′ = ay + bz, (5.12)
z ′ = cy + dz,

where a, b, c, and d are constants. We look for solutions of the form

y = Y eλt , z = Zeλt , (5.13)

where λ, Y , and Z are constants to be determined, with Y and Z not both zero.
When we substitute the form (5.13) into the system (5.12), using y ′ = λY eλt ,
z ′ = λZeλt , we obtain two conditions

λY eλt = aY eλt + bZeλt ,

λZeλt = cY eλt + dZeλt ,

which must be satisfied for all t. Because eλt 6= 0 for all t, we may divide these
equations by eλt to obtain a system of two equations which do not depend on
t, namely

λY = aY + bZ,
λZ = cY + dZ

or

(a − λ)Y + bZ = 0, (5.14)
cY + (d − λ)Z = 0.

The pair of equations (5.14) is a system of two homogeneous linear algebraic

equations for the unknowns Y and Z. For certain values of the parameter
λ this system will have a solution other than the obvious solution Y = 0,
Z = 0. In order that the system (5.14) have a non-trivial solution for Y and
298 Dynamical Systems for Biological Modeling: An Introduction

Z, it is necessary that the determinant of the coefficient matrix, which is

(a − λ)(d − λ) − bc, be equal to zero. This gives a quadratic equation, called
the characteristic equation, of the system (5.12) for λ. We may rewrite the
characteristic equation as

λ2 − (a + d)λ + (ad − bc) = 0. (5.15)

We will assume that ad − bc 6= 0, which implies that λ = 0 is not a root

of (5.15). Our reason for this assumption is the following: If λ = 0 is a root
(or, equivalently, ad − bc = 0), then equations (5.14) become the same as the
equilibrium conditions for (5.12), obtained by setting the right-hand sides of
(5.12) to zero. With ad − bc = 0, the equilibrium conditions will reduce to a
single equation: for instance, substituting d = bc/a into the second equation
and then multiplying by a/c will result in the first equation. Consequently,
there will be a line of non-isolated equilibria. We do not wish to explore this
problem (but see Exercise 19 below), in part because the treatment of non-
isolated equilibria is complicated, and in part because it is difficult to apply
the linearization theorem of the previous section when the linearization of a
system about an equilibrium is of this form.
With ad−bc 6= 0, the characteristic equation has two roots λ1 and λ2 , which
may be real and distinct, real and equal, or complex conjugates. (Students of
linear algebra will take note that these roots are in fact the eigenvalues of
the community matrix; the characteristic equation of the system is also the
characteristic equation of this matrix.) If λ1 and λ2 are the roots of (5.15),
then there are a solution (Y1 , Z1 ) of (5.14) corresponding to the root λ1 , and
a solution (Y2 , Z2 ) of (5.14) corresponding to the root λ2 . These, in turn, give
us two solutions

y = Y1 eλ1 t , z = Z1 eλ1 t and y = Y2 eλ2 t , z = Z2 eλ2 t

to the system (5.12). We note that if λ is a root of (5.15), then equations

(5.14) reduce to a single equation, so that we may always choose a value for
one of Y , Z, with the other then determined by (5.14). If (5.14) sets one of
Y , Z to zero, then we may choose a value for the other.
Because system (5.12) is linear, it is possible to show that if we have two
different solutions of the system (5.12), then every solution of (5.12) is a
constant multiple of the first solution plus a constant multiple of the second
solution. By “different” we mean that neither solution is a constant multiple of
the other. Here, if the roots λ1 and λ2 of the characteristic equation (5.15) are
distinct, we do have two different solutions of (5.12), and then every solution
of system (5.12) has the form

y = K1 Y1 eλ1 t + K2 Y2 eλ2 t , (5.16)

z = K1 Z1 eλ1 t + K2 Z2 eλ2 t

for some constants K1 and K2 . The form (5.16) with two arbitrary constants
 Systems of Differential Equations 299

K1 and K2 is called the general solution of the system (5.12). If initial values
y(0) and z(0) are specified, these two initial values may be used to determine
values for the constants K1 and K2 , and thus to obtain a particular solution
in the family (5.16).

Example 1.
Find the general solution of the system

y ′ = −y − 2z, z ′ = y − 4z

and also the solution such that y(0) = 3, z(0) = 1.

Solution: Here a = −1, b = −2, c = 1, d = −4, so that a+ d = −5, ad− bc = 6,
and the characteristic equation is λ2 +5λ+6 = 0, with roots λ1 = −2, λ2 = −3.
With λ = −2, both equations of the algebraic system (5.14) are Y − 2Z = 0,
and we may take Y = 2, Z = 1. The resulting solution of (5.12) is y = 2e−2t ,
z = e−2t . With λ = −3, both equations of the algebraic system (5.14) are
Y − Z = 0, and we may take Y = 1, Z = 1. The resulting solution of (5.12)
is y = e−3t , z = e−3t . Thus the general solution of the system is

y = 2K1 e−2t + K2 e−3t , z = K1 e−2t + K2 e−3t .

To satisfy the initial conditions, we substitute t = 0, y = 3, z = 1 into this

form, obtaining a pair of equations 2K1 + K2 = 3, K1 + K2 = 1. We may
subtract the second of these from the first to give K2 = 2, and then K1 = −1.
This gives, as the solution of the initial value problem,

y = 4e−2t − e−3t , z = 2e−2t − e−3t . 

If the characteristic equation (5.15) has a double root, this method gives
only one solution of the system (5.12), and we need to find a second solution
in order to form the general solution. The characteristic equation has a double
root if the discriminant (a+d)2 −4(ad−bc) = (a−d)2 +4bc = 0. It is possible to
show (and the reader can verify) that if λ is a double root of (5.15), λ = a+d
2 ,
then in addition to the solution y = Y1 eλt , z = Z1 eλt of (5.12) there is a
second solution, of the form

y = (Y2 + Y1 t)eλt , z = (Z2 + Z1 t)eλt ,

where Y1 , Z1 are as in (5.14) and Y2 , Z2 are given by

(a − λ)Y2 + bZ2 = Y1 ,
cY2 + (d − λ)Z2 = Z1 .

Thus the general solution of the system (5.12) in the case of equal roots is

y = (K1 Y1 + K2 Y2 )eλt + K2 Y1 teλt , (5.17)

z = (K1 Z1 + K2 Z2 )eλt + K2 Z1 teλt .
300 Dynamical Systems for Biological Modeling: An Introduction

Note that if (5.15) has a single root and b = 0, then λ = a = d, and we have
Y1 = 0, Z1 = cY2 , so that the general solution becomes

y = K3 eλt , z = K4 eλt + cK3 teλt ,

where K3 ≡ K2 Y2 and K4 ≡ cK1 Y2 + K2 Z2 are arbitrary constants. Likewise

if (5.15) has a single root and c = 0, then λ = a = d, Z1 = 0, Y1 = bZ2 , and
the general solution reduces to

y = K5 eλt + bK6 teλt , z = K6 eλt ,

with arbitrary constants K5 ≡ bK1 Z2 + K2 Y2 and K6 ≡ K2 Z2 . (We cannot

have b and c both zero here, or else the system becomes uncoupled: y ′ = ay,
z ′ = dz.)

Example 2.
Find the general solution of the system

y ′ = z, z ′ = −y + 2z,

and also the solution such that y(0) = 2, z(0) = 3.

Solution: Since a = 0, b = 1, c = −1, d = 2, we have a + d = 2, ad − bc = 1.
The characteristic equation is λ2 − 2λ + 1 = 0, with a double root λ = 1. With
λ = 1, both equations of the system (5.14) are Y − Z = 0, and we may take
Y = 1, Z = 1 to give the solution y = et , z = et of (5.12). Substituting these
values into the equations for Y2 , Z2 , we find that they reduce to the single
equation −Y2 + Z2 = 1, so we may take Y2 = 0, Z2 = 1. Equations (5.17) now
give the general solution y = K1 et + K2 tet , z = (K1 + K2 )et + K2 tet . To find
the solution with y(0) = 2, z(0) = 3, we substitute t = 0, y = 2, z = 3 into
this form, obtaining the pair of equations K1 = 2, K1 + K2 = 3. Then K2 = 1,
and the solution satisfying the initial conditions is y = 2et + tet , z = 3et + tet .


Another complication arises if the characteristic equation (5.15) has com-

plex roots. While the general solution of (5.12) is still given by (5.16), in this
case the constants λ1 and λ2 are complex, and the solution is in terms of com-
plex functions. Complex exponentials, however, can be defined with the aid of
trigonometric functions (eiθ ≡ cos θ + i sin θ for real θ), so it is still possible to
give the solution of (5.12) in terms of real exponential and trigonometric func-
tions. In this case, if the characteristic equation (5.15) has conjugate complex
roots λ = α ± iβ, where α and β are real and β > 0, equations (5.16) become

y = (K1 Y1 + K2 Y2 )eαt cos βt + i(K1 Y1 − K2 Y2 )eαt sin βt,

z = (K1 Z1 + K2 Z2 )eαt cos βt + i(K1 Z1 − K2 Z2 )eαt sin βt.

We can eliminate the imaginary coefficients by defining Q1 ≡ i(K1 − K2 ),

 Systems of Differential Equations 301

Q2 ≡ K1 + K2 , and taking (a − λi )Yi + bZi = 0 for i = 1, 2 from (5.16) to

arrive at the form

y= Q1 beαt sin βt + Q2 beαt cos βt,

z = −[Q1 (a − α) − Q2 β]eαt sin βt + [Q1 β − Q2 (a − α)]eαt cos βt.

Example 3.
Find the general solution of the system

y ′ = −2z, z ′ = y + 2z,

and also the solution with y(0) = −2, z(0) = 0.

Solution: We have a = 0, b = −2, c = 1, d = 2, and the characteristic equation
is λ2 − 2λ + 2 = 0, with roots λ = 1 ± i. The general solution is then

y = −2K1 et sin t − 2K2 et cos t, z = (K1 − K2 )et sin t + (K1 + K2 )et cos t.

To find the solution with y(0) = −2, z(0) = 0, we substitute t = 0, y = −2,

z = 0 into this form, obtaining −2K2 = −2, K1 + K2 = 0, whose solution is
K1 = −1, K2 = 1. This gives the particular solution

y = 2et sin t − 2et cos t, z = −2et sin t. 

In many applications, especially in analyzing stability of an equilibrium,

the precise form of the solution of a linear system is less important to us than
the qualitative behavior of solutions. It will turn out that often the crucial
question is whether all solutions of a linear system approach zero as t → ∞.
In the next section we will give an algebraic criterion to answer this question,
but we may also infer the answer from a look at the direction field near the
origin. It requires more effort, but a phase portrait consisting of several orbits
may be even more helpful.
For example, the direction field and phase portrait for the system y ′ = −y−
2z, z ′ = y − 4z, solved explicitly in Example 1 above, are shown in Figure 5.6.
The direction field suggests that every orbit approaches the origin with a
fixed limiting direction, and this is correct. The origin is an equilibrium for
the linear system, and an equilibrium with this behavior for orbits is called a
node. Nodes occur when the roots of the corresponding characteristic equation
are real and of the same sign (here both negative).
The direction field and phase portrait for the system y ′ = y + 2z, z ′ = −z
are shown in Figure 5.7. Orbits starting on the line z = −y approach the
origin, but all other orbits are repelled from the origin. Such an equilibrium is
called a saddle point. Saddle points occur when the roots of the corresponding
characteristic equation are real but have different signs.
The direction field and phase portrait for the system y ′ = z, z ′ = −2y − 2z
302 Dynamical Systems for Biological Modeling: An Introduction

FIGURE 5.6: Direction field and phase portrait showing a stable node.

FIGURE 5.7: Direction field and phase portrait showing a saddle point.

are shown in Figure 5.8. Every orbit approaches the origin, but the approach
is by an inward spiral. Such an equilibrium is called a spiral point. Here the
roots of the characteristic equation are complex.
The direction field and phase portrait for the system y ′ = z, z ′ = −y are
shown in Figure 5.9. Every orbit is a closed orbit around the origin, and such
an equilibrium is called a center. In this case the roots of the characteristic
equation are purely imaginary.
In the above examples, every orbit approaches the origin for the node
and spiral point, but we could also give examples in which the directions
 Systems of Differential Equations 303

FIGURE 5.8: Direction field and phase portrait showing a stable spiral
point.

FIGURE 5.9: Direction field and phase portrait showing a center.

are reversed and all orbits are unbounded. For a center, no orbits approach
the origin, and for a saddle point there are orbits which approach the origin
and also unbounded orbits. Every equilibrium point for a linear autonomous
system is of one of these four types.
As noted above, the nature of the origin as an equilibrium of the linear sys-
tem (5.12) depends on the roots of the characteristic equation (5.15). If both
roots of (5.15) are real and of the same sign, then the solutions of (5.12) are
combinations of either positive exponentials or negative exponentials. This
304 Dynamical Systems for Biological Modeling: An Introduction

Photo by Steve Hillebrand, US FWS

FIGURE 5.10: A system of interconnected ponds acts like a mixing system.

′
implies that the slope of an orbit, which is yz ′ (t)
(t)
, must approach a limit as
t → ∞, and thus that the origin is a node. If the roots of (5.15) are real and
of opposite sign, then there are solutions which are positive exponentials and
solutions which are negative exponentials. Thus there are solutions approach-
ing the origin and solutions moving away from the origin, and the origin is a
saddle point. If the roots of (5.15) are complex, the solutions contain trigono-
metric functions and the orbits oscillate. If the real parts of the roots are zero,
the orbits are periodic, and the origin is a center. If the real parts are different
from zero, the orbits spiral, and the origin is a spiral point.

In addition to its use in equilibrium stability analysis, the solution of lin-

ear systems with constant coefficients has some direct applications. We now
conclude this section with a direct application to mixing problems. In Sec-
tion 3.6 we described some mixing problems with a single compartment. Mix-
ing problems with two compartments, whether natural (such as a system of
interconnected ponds, see Figure 5.10) or artificial (connected tanks, such as
in Example 4 below), may be described by characterizing the rate of change
in each compartment to give a system of two differential equations. If the mix-
ture flows from one compartment to the other, the concentrations in the two
compartments are linked, and we would expect a coupled system of differential
equations.
 Systems of Differential Equations 305

Example 4.
A tank contains 100 liters of water and 10 kg. of salt, thoroughly mixed. Pure
water is added at the rate of 5 liters/minute, and the mixture is poured off at
the rate of 5 liters/minute into a second tank which initially contains 80 liters
of water. The mixture in the second tank is then poured out as waste, at the
same rate. Formulate and solve a model to describe the weight of salt in each
tank as a function of time.
Solution: Let y(t) denote the weight of salt in the first tank and z(t) the
weight of salt in the second tank at time t. Then y(0) = 10 and z(0) = 0. The
y
concentration of salt in the first tank is 100 kg./liter, and the weight of salt
y y
poured into the second tank per minute is 5 100 = 20 kg. The concentration
of salt in the second tank is z/80 kg./liter, and the weight of salt poured out
z z
is 5 80 = 16 kg. Thus

y ′ = −y/20, z ′ = y/20 − z/16.

In addition to this system of differential equations, we must impose the initial

conditions y(0) = 10, z(0) = 0.
1 1 1
To solve this problem, we take a = − 20 , b = 0, c = 20 , d = − 16 . This
1 1 1
gives the characteristic equation λ2 + 16 λ + 1600 = 0, with roots λ1 = − 20
1 1
and λ2 = − 80 . With λ = − 20 , the system (5.14) reduces to 4Y1 + 3Z1 = 0,
so we can take Y1 = 3, Z1 = −4, giving a solution y = 3e−t/20 , z = −4e−t/20
1
of the system. With λ = − 80 , the system reduces to Y2 = 0, so we can take
Z2 = 1, giving a solution y = 0, z = e−t/80 of the system. Thus the general
solution is y = 3K1 e−t/20 , z = −4K1 e−t/20 + K2 e−t/80 . To satisfy the initial
conditions y(0) = 10, z(0) = 0, we substitute t = 0, y = 10, z = 0, and
obtain the equations 10 = 3K1 , 0 = −4K1 + K2 , with solution K1 = 10 3 ,
K2 = 40 3 . Thus the solution of the initial value problem is y = 10e −t/20
,
40 −t/20 40 −t/80
z=−3e + 3e . We note that since the solution is a sum of negative
exponentials, it approaches zero as t → ∞; that is, the amount of salt in each
tank approaches zero asymptotically as time goes on. 

5.3.1 A liver chemistry example

The chemical bromosulfophthalene (BSP) is used to measure hepatic
metabolism. BSP is administered directly into the blood via an intravenous
(IV) injection, and then BSP levels in the blood are recorded periodically in
order to measure the liver’s ability to clear toxins, chemicals and other foreign
substances from the blood. When blood flows through the liver, there is an
exchange of these substances between the blood and the liver, at rates depen-
dent upon the concentration of the given substance in the system of origin
(blood or liver). At the same time, the liver eliminates excess amounts of the
substance via other pathways (such as the digestive system). This exchange
306 Dynamical Systems for Biological Modeling: An Introduction

is described by Jolivet5 in a simple linear two-compartment model, describing

the three exchanges from blood to liver, from liver to blood, and from liver to
elimination from the body. If we denote the blood with a subscript 1, the liver
with a subscript 2, and elimination with a subscript 0, then we can denote
the three exchange rates, respectively, as k21 , k12 , and k02 . Each rate can be
interpreted as the reciprocal of the average time required for each exchange
to occur.
Figure 5.11 provides a stylized flow chart for this exchange process. If we
define q1 (t) to be the quantity (or, alternatively, the concentration) of BSP
in the blood and q2 (t) the quantity (or concentration) in the liver, then we
can describe the rates of change of q1 and q2 via the following two differential
equations:
dq1
= −k21 q1 +k12 q2 (5.18)
dt
dq2
= k21 q1 −(k12 + k02 )q2
dt
where the first term in each equation represents flow of BSP from the blood to
the liver, the second term in each equation represents the flow in the reverse
direction, and the remaining term in the second equation represents elimi-
nation of BSP from the liver. To complete the statement of an initial value
problem, we add the initial conditions q1 (0) = Q, q2 (0) = 0 representing the
injection at time 0 of a quantity (or concentration) Q of BSP into the blood,
at which point the liver has no BSP in it.
Since the system (5.18) is a linear system with constant coefficients, we
assume that there are exponential solutions of the form q = aeλt , so that
q ′ = λq; then the equations (5.18) become

λq1 (t) = −k21 q1 (t) +k12 q2 (t),

λq2 (t) = k21 q1 (t) −(k12 + k02 )q2 (t).
 
−k21 k12
Solutions λ are eigenvalues of the coefficient matrix ,
k21 −(k12 + k02 )
as straightforward substitution leads to elimination of q1 and q2 , leaving the
characteristic equation

λ2 + (k21 + k12 + k02 )λ + (k21 k02 ) = 0.

The two roots are

1 √ 
λ1,2 = − k21 + k12 + k02 ± ∆ , where ∆ = (k21 + k12 + k02 )2 − 4k21 k02
2
(5.19)
is the discriminant of the quadratic characteristic equation. Some algebra
5 Emmanuel Jolivet, Introduction aux modèles mathématiques en biologie,
INRA/Masson, Paris, 1983. Chapter 3, pp. 37–52.
 Systems of Differential Equations 307

FIGURE 5.11: A sketch of a FIGURE 5.12: A comparison of

flow chart depicting the exchanges the liver chemistry data (dots), the
of the chemical BSP between the slow-decaying component of the so-
blood and the liver. lution (dashed curve), and the full
solution (solid curve).

shows that√∆ = (k21 − k02 )2 + k122

+ 2k12 (k02 + k21 ) > 0, so both roots are real
and (since ∆ < k21 +k12 +k02 ) negative, corresponding to exponential decay.
In particular, if we take λ1 > λ2 , then λ1 describes a slower rate of elimination
of BSP from the body, while λ2 describes a portion which is eliminated more
quickly. Overall, then, solutions to the system of differential equations have
the form

q1 (t) = a1 eλ1 t + a2 eλ2 t , (5.20)

q2 (t) = b1 eλ1 t + b2 eλ2 t , (5.21)

where the constants a1 , a2 , b1 and b2 are determined by initial conditions.

Knowing that the amounts (or concentrations) of BSP in the blood and
liver can be described by these functions, we can use measurements of BSP
concentration in the blood over time to calculate first the coefficients in the
above equations, and then the original exchange rates, which give an under-
standing of how quickly the liver is able to clear BSP from the body. As an
example of how this model might be used in practice, Jolivet provides the
data in Table 5.1 from Feldmann and Schneider,6 in which 0.8 mg of BSP was
introduced via IV injection.

TABLE 5.1: Data from Feldmann and Schneider [24] showing the concen-
tration q1 of BSP in the blood as a function of time t since introduction.
t (min) 3 6 9 12 15 20 30 40 60
q1 (t) (µg/L) 49 20 14 5 4 3 2 2 1

6 U. Feldmann and B. Schneider, A general approach to multicompartment analysis

and models for the pharmacodynamics, in J. Berger et al. (eds.), Mathematical models
in medicine: workshop, Mainz, March 1976 (Berlin/New York: Springer-Verlag), Lecture
Notes in Biomathematics 11: 243–277, 1976.
308 Dynamical Systems for Biological Modeling: An Introduction

As always, real data contains “noise,” so will not fit any theoretical form
(i.e., (5.20)) perfectly. There are many sophisticated numerical and statistical
methods for deriving the best possible estimate for the coefficients in the
equation, but for the sake of simplicity here we will use only some of these
data and derive simple estimates (which nevertheless compare well with those
Jolivet obtained, q.v.). First, we assume that the difference in decay rates
between λ1 and λ2 is great enough that after enough time, most of the fast-
decaying eλ2 t component has dwindled to a negligible level, leaving primarily
the slower-decaying eλ1 t component: q1 (t) ≈ a1 eλ1 t . If we assume that this is
true for the last half of the data, then we can use the concentrations after 20
and 60 minutes to calculate the coefficients a1 and λ1 :
q1 (20) = 3 ≈ a1 e20λ1 , q1 (60) = 1 ≈ a1 e60λ1 ,
log 3 ≈ log a1 + 20λ1 , log 1 = 0 ≈ log a1 + 60λ1 ,
from which log a1 ≈ −60λ1 , and, substituting, log 3 ≈ −40λ1 , making λ1 ≈
(log 3)/40min = −0.0275/min and thus a1 ≈ e−60λ1 = 5.2µg/L (note all
logarithms here are natural, not base ten).
To recover a2 and λ2 we return to the earliest data points (3 and 6 minutes
after injection), where we assume the fast-decaying component can still be
detected:
q1 (3) = 49 = a1 e3λ1 + a2 e3λ2 , q1 (6) = 20 = a1 e6λ1 + a2 e6λ2 ,
49 = 5.2e3(−0.0275) + a2 e3λ2 , 20 = 5.2e6(−0.0275) + a2 e6λ2 ,
49 = 4.788 + a2 e3λ2 , 20 = 4.41 + a2 e6λ2 ,
log 44.212 = 3.789 = log a2 + 3λ2 , log 15.59 = 2.747 = log a2 + 6λ2 ,
so log a2 = 2.747 − 6λ2, making 3.789 = 2.747 − 3λ2, and thus λ2 = −(3.789 −
2.747)/3min = −0.347/min. Then, finally, a2 = e2.747−6(−0.347) ≈ 125µg/L.
This gives the concentration of BSP in the blood of
q1 (t) = 5.2e−0.0275t + 125e−0.347t . (5.22)
From this we see that most of the BSP is removed fairly quickly from the
blood (after an average of 1/λ2 = 2.88min), while the rest takes longer to be
eliminated (an average of 1/λ1 = 36.4min). Figure 5.12 superimposes plots
of the original data, the solution, and the slow-decaying component of the
solution.
To recover the original exchange rates we use a somewhat ad hoc process,
first observing that q1 (0) = 5.2 + 125 ≈ 130µg/L, q2 (0) = 0 so that (from
(5.18)
dq1
(0) = −k21 q1 (0) + k12 q2 (0) = −(130µg/L)k21 ,
dt
while at the same time (from (5.22))
dq1
(0) = 5.2(−0.0275) + 125(−0.347) = −43.518(µg/L)/min,
dt
 Systems of Differential Equations 309

making k21 = −43.518/(−130)min = 0.334/min. Then, to recover the other

two rates, we observe that (from (5.19), and after a little algebra)

λ1 λ2 = k02 k21 , λ1 + λ2 = −(k21 + k12 + k02 ).

The first of these two equations becomes (−0.0275)(−0.347) = 0.334k02 , mak-

ing k02 = 0.0286/min, and then the latter equation becomes −0.0275−0.347 =
−(0.334 + k12 + 0.0286), so that k12 = 0.0119/min. The reciprocals of these
rates give the corresponding average passage times.

Exercises
In each of Exercises 1–14, find the general solution of the given system by
analytic solution, use a computer algebra system to examine the behavior
of solutions, and classify the origin as a node, saddle point, center, or spiral
point.

1. y ′ = y + 5z, z ′ = y − 3z
2. y ′ = y − z, z ′ = z
3. y ′ = 2y + z, z ′ = z
4. y ′ = −y, z ′ = y − z
5. y ′ = 4y, z ′ = 2y + 4z
6. y ′ = z, z ′ = −y + 2z
7. y ′ = z, z ′ = −2y − 3z
8. y ′ = y − z, z ′ = 4y − 3z
9. y ′ = y + 2z, z ′ = −3y + 6z
10. y ′ = 3y − 4z, z ′ = y − 2z
11. y ′ = 3y + 5z, z ′ = −5y + 3z
12. y ′ = z, z ′ = −y
13. y ′ = y + z, z ′ = z
14. y ′ = 5y + z, z ′ = 5z
15. A tank contains 1000 liters of water and a salt solution containing 10
kg./liter is pumped into it at a rate of 200 liters /minute. The mixture
is led to a second tank containing 1000 liters of water at a rate of 200
liters/minute, and the mixture is pumped out of the second tank at the
same rate. What is the concentration of salt in the second tank after
one hour?
310 Dynamical Systems for Biological Modeling: An Introduction

16. Two tanks begin with 10 kg. of salt dissolved in 100 liters of water.
Water is pumped into the first tank at a rate of 10 liters/minute, the
mixture is pumped from the first tank to the second tank at a rate of
10 liters/minute, and the mixture is pumped out of the second tank at
a rate of 10 liters/minute. What is the amount of salt in each tank after
one hour, and what is the amount of salt in each tank after a very long
time?
17. Obtain the general solution of the system y ′ = ay, z ′ = cy + dz if a 6= d
by solving y ′ = ay, substituting the result into z ′ = cy + dz and solving.

18. Obtain the general solution of the system y ′ = ay, z ′ = cy + az by

solving y ′ = ay, substituting the result into z ′ = cy + az and solving.
19. * Consider the system

y ′ = y − z,
z ′ = 3y − 3z,

for which the condition ad − bc 6= 0 is violated.

(a) Show that every point of the line z = y is an equilibrium.
(b) Use the fact that z ′ = 3y ′ to deduce that z = 3y + c1 for some
constant c1 .
(c) Use the result of part (b) to eliminate z from the system and obtain
a first-order linear differential equation for y.
(d) Solve for y and obtain the solution y = − c21 + c2 e−2t , z = − c21 +
3c2 e−2t .
(e) Show that as t → ∞ every orbit approaches the point (− c21 , − c21 )
on the line of equilibria, and that the slope of the line joining this point
to any point on the orbit is the constant 3.
(f) Use the information obtained to sketch the phase portrait of the
system.
 Systems of Differential Equations 311

5.4 Qualitative analysis of systems

In order to apply the linearization theorem of Section 5.2 to questions
of stability of an equilibrium, we must determine conditions under which all
solutions of a linear system with constant coefficients

y ′ = ay + bz, (5.23)
z ′ = cy + dz

approach zero as t → ∞. As we saw in Section 5.3, the nature of the solutions

of (5.23) is determined by the roots of the characteristic equation

λ2 − (a + d)λ + (ad − bc) = 0. (5.24)

If the roots λ1 and λ2 of (5.24) are real, then the solutions of (5.23) are made
up of terms eλ1 t and eλ2 t , or eλ1 t and teλ1 t if the roots are equal. In order that
all solutions of (5.23) approach zero, we require λ1 < 0 and λ2 < 0, so that
the terms will be negative exponentials. If the roots are complex conjugates,
λ = α ± iβ, then in order that all solutions of (5.23) approach zero, we require
α < 0. Thus if the roots of the characteristic equation have negative real part,
all solutions of the system (5.23) approach zero as t→ ∞. In a similar manner,
we may see that if a root of the characteristic equation has positive real part,
then (5.23) has unbounded solutions.
It turns out, however, that it is not necessary to solve the characteristic
equation in order to determine whether all solutions of (5.23) approach zero,
as there is a useful criterion in terms of the coefficients of the characteristic
equation. The basic result, whose proof may be found in Appendix C, is that
the roots of a quadratic equation λ2 + a1 λ + a2 = 0 have negative real part
if and only if a1 > 0 and a2 > 0. Applying this to the characteristic equation
(5.24) and the system (5.23), we obtain the following result for linear systems
with constant coefficients:

STABILITY THEOREM FOR LINEAR SYSTEMS: Every solu-

tion of the linear system with constant coefficients (5.23)

y ′ = ay + bz, z ′ = cy + dz

approaches zero as t → ∞ if and only if the trace a + d of the coeffi-

cient matrix of the system is negative and the determinant ad − bc of the
system’s coefficient matrix is positive. If either the trace is positive or the
determinant is negative, there is at least one unbounded solution.
312 Dynamical Systems for Biological Modeling: An Introduction

Example 1.
Determine whether all solutions tend to zero or whether there are unbounded
solutions for each of the following systems:

(i) u′ = −u − 2v, v ′ = u − 4v
(ii) u′ = v, v ′ = −u − 2v
(iii) u′ = −2v, v ′ = u + 2v

Solution: (i) The characteristic equation is λ2 + 5λ+ 6 = 0, with roots λ = −2,

λ = −3. Thus all solutions tend to zero. Alternatively, since the trace of the
coefficient matrix is −5 < 0 and the determinant is 6 > 0, the stability theorem
gives the same conclusion. For (ii), the characteristic equation is λ2 +2λ+1 = 0
with a double root λ = −1, and thus all solutions tend to zero. For (iii), the
characteristic equation is λ2 − 2λ + 2 = 0, and since the trace is positive, there
are unbounded solutions. As we indicated in the previous section, we could
also have drawn this conclusion from a phase portrait. 

If we apply the stability theorem for linear systems to the linearization

u′ = Fy (y∞ , z∞ )u + Fz (y∞ , z∞ )v, (5.25)

v′ = Gy (y∞ , z∞ )u + Gz (y∞ , z∞ )v

of a system
y ′ = F (y, z), z ′ = G(y, z) (5.26)
at an equilibrium (y∞ , z∞ ), we obtain the following result.

EQUILIBRIUM STABILITY THEOREM: Let (y∞ , z∞ ) be an equi-

librium of a system y ′ = F (y, z), z ′ = G(y, z), with F and G twice differ-
entiable. Then if
Fy (y∞ , z∞ ) + Gz (y∞ , z∞ ) < 0 (5.27)
and

Fy (y∞ , z∞ )Gz (y∞ , z∞ ) − Fz (y∞ , z∞ )Gy (y∞ , z∞ ) > 0, (5.28)

the equilibrium (y∞ , z∞ ) is asymptotically stable. If either

Fy (y∞ , z∞ ) + Gz (y∞ , z∞ ) > 0

or
Fy (y∞ , z∞ )Gz (y∞ , z∞ ) − Fz (y∞ , z∞ )Gy (y∞ , z∞ ) < 0
the equilibrium (y∞ , z∞ ) is unstable.
 Systems of Differential Equations 313

Example 2.
Determine whether each equilibrium of the system

y ′ = z, z ′ = 2(y 2 − 1)z − y

is asymptotically stable or unstable.

Solution: The equilibria are the solutions of z = 0, 2(y 2 − 1)z − y = 0, and
thus the only equilibrium is (0,0). Here F (y, z) = z, with partial derivatives
0 and 1, respectively, and G(y, z) = 2(y 2 − 1)z − y, with partial derivatives
4yz − 1 and 2(y 2 − 1) respectively. Therefore the community matrix at the
equilibrium is  
0 1
−1 −2
with trace −2 and determinant 1, as in Example 1(ii). Thus the equilibrium
(0,0) is asymptotically stable. 

Example 3.
Determine whether each equilibrium of the system

y ′ = y(1 − 2y − z)
z ′ = z(1 − y − 2z)

is asymptotically stable or unstable. This system could model a population of

two competing species, say two species of fish sharing the same habitat (see
Figure 5.13): each species lives according to a logistic rule in the absence of the
other, with a carrying capacity of 1/2 unit (measured in thousands or millions
of individuals), y ′ = y(1 − 2y) or z ′ = z(1 − 2z), and both species suffer
from the presence of each other since they compete for the same resources
(food, shelter, etc.), with the negative effect given by the same “bilinear”
term (yz) which we have seen used to model encounters in predator-prey and
epidemic models. We shall study populations in competition in more detail in
Section 6.2.1; this is a special case where the two competitors are essentially
evenly matched (same carrying capacity, same competitive disadvantage).
Solution: The equilibria are the solutions of y(1−2y −z) = 0, z(1−y −2z) = 0.
One solution is (0,0), representing extinction of both species; a second is the
solution of y = 0, 1 − y − 2z = 0, which is (0, 12 ) [the second species displaces
the first]; a third is the solution of z = 0, 1 − 2y − z = 0, which is ( 21 ,0) [the
first species displaces the second]; and a fourth is the solution of 1−2y −z = 0,
1−y−2z = 0, which is ( 13 , 13 ), representing coexistence. The community matrix
at an equilibrium (y∞ , z∞ ) is
 
1 − 4y∞ − z∞ −y∞
.
−z∞ 1 − y∞ − 4z∞

At (0,0), this matrix has trace 1 and determinant 1, and thus the equilibrium is
314 Dynamical Systems for Biological Modeling: An Introduction

Photo by David Burdick, courtesy NOAA

FIGURE 5.13: Yellow tang (Zebrasoma flavescens), black Achilles tang

(Acanthurus achilles), and chub swim near a coral reef off the coast of Maui,
Hawaii. Fish which share the same habitat may compete for food, shelter, and
other resources.

unstable. At (0, 21 ), this matrix has trace − 12 and determinant − 21 , and thus the
equilibrium is unstable. At ( 12 ,0), this matrix has trace − 21 and determinant
− 12 , and thus the equilibrium is unstable. At ( 13 , 13 ), this matrix has trace − 34
and determinant 13 , and thus this equilibrium is asymptotically stable. Thus
the model suggests that evenly matched competitors can coexist. 

The careful reader will have noticed that, like the equilibrium stability
theorem of Section 4.1, the equilibrium stability theorem given above has a
hole of sorts in its result, in that the theorem says nothing about the stabil-
ity of equilibria for which the trace and determinant lie on the boundary of
conditions (5.27) and (5.28) — in other words, for which the linearization has
solutions which do not approach zero as t → ∞ but stay bounded. The reason
for this “hole” is that in such cases, the linearization does not give enough
information to determine stability. The following example recalls such a case,
from the application with which we began this chapter.
 Systems of Differential Equations 315

Example 4.
Determine the asymptotic stability or instability of each equilibrium of the
Lotka-Volterra system

y ′ = y(λ − bz),
z ′ = z(−µ + cy).

Solution: We showed in Example 3, Section 5.2 that the equilibrium (0,0) is

unstable. Thus we need only examine the equilibrium (y∞ , z∞ ) with y∞ = µc ,
z∞ = λb . By the computation carried out in Example 1, Section 5.2, the
community matrix at this equilibrium is

0 − bµ
 
c .
cλ
b 0

This matrix has a positive determinant, but the trace is zero. In this case,
the stability theorem does not give any information. However, as we saw in
Section 5.1, the orbits of the system neither tend to the equilibrium nor move
away from the equilibrium. Thus the equilibrium is neither asymptotically
stable nor unstable, but behaves like a center (cf. Figures 5.2, 5.9). 

If all orbits beginning near an equilibrium remain near the equilibrium

for t ≥ 0, but some orbits do not approach the equilibrium as t → ∞, the
equilibrium is said to be stable, or sometimes neutrally stable. If the origin is
neutrally stable for the linearization at an equilibrium, then the equilibrium
may also be neutrally stable for the nonlinear system, as for a Lotka-Volterra
system. However, it is also possible for the origin to be neutrally stable for
the linearization at an equilibrium, while the equilibrium is asymptotically
stable or unstable. Thus neutral stability of the origin for a linearization at
an equilibrium gives no information about the stability of the equilibrium.
We have seen in Section 4.1 that a solution of an autonomous first-order
differential equation is either unbounded or approaches a limit as t → ∞. For
an autonomous system of two first-order differential equations, these same
two possibilities exist. In addition, however, there is the possibility of an orbit
which is a closed curve, corresponding to a periodic solution. Such an orbit is
called a periodic orbit because it is traversed repeatedly.
There is a remarkable result which says essentially that these are the only
possibilities.

POINCARÉ-BENDIXSON THEOREM: A bounded orbit of a sys-

tem of two first-order differential equations which does not approach an
equilibrium as t → ∞ either is a periodic orbit or approaches a periodic
orbit as t → ∞.
316 Dynamical Systems for Biological Modeling: An Introduction

Standard techniques from the study of complex variables can be used to

show that a periodic orbit must enclose an equilibrium point in its interior. In
many examples, there is an unstable equilibrium, and orbits beginning near
this equilibrium spiral out toward a periodic orbit. A periodic orbit which is
approached by other (non-periodic) orbits is called a limit cycle. One example
of a limit cycle involves the system

y ′ = y(1 − y 2 − z 2 ) − z,
z ′ = z(1 − y 2 − z 2 ) + y,

which has its only equilibrium at the origin. The equilibrium is unstable, and
Figure 5.14 illustrates the fact that all orbits not beginning at the origin spiral
counterclockwise [in or out] toward the unit circle y 2 + z 2 = 1.

-2 -1 1 2

-1

-2
FIGURE 5.14: Trajectories approaching a limit cycle.

In many applications the functions y(t) and z(t) are restricted by the
nature of the problem to non-negative values. For example, this is the case
if y(t) and z(t) are population sizes. In such a case, only the first quadrant
y ≥ 0, z ≥ 0 of the phase plane is of interest. For a system

y ′ = F (y, z), z ′ = G(y, z)

which has F (0, z) ≥ 0 for z ≥ 0 and G(y, 0) ≥ 0 for y ≥ 0, then since y ′ ≥ 0

along the positive z-axis (where y = 0) and z ′ ≥ 0 along the positive y-axis
(where z = 0), no orbit can leave the first quadrant by crossing one of the
axes. The Poincaré-Bendixson theorem may then be applied to orbits in the
first quadrant. In such a case, the first quadrant is called an invariant set, a
region with the property that orbits must remain in the region.
If instead F and G are identically zero along the respective [half-]axes,
 Systems of Differential Equations 317

then y ′ = 0 for {y = 0, z ≥ 0}, and z ′ = 0 for {y ≥ 0, z = 0}. In this case,

orbits which begin on an axis must remain on that axis, and orbits beginning
in the interior of the first quadrant (with y(0) > 0, z(0) > 0) must remain
in the interior of the first quadrant (i.e., y(t) > 0 and z(t) > 0 for t ≥ 0).
If there is no equilibrium in the first quadrant, there cannot be a periodic
orbit, because a periodic orbit must enclose an equilibrium. Thus, if there is
no equilibrium in the first quadrant every orbit must be unbounded.

Example 5.
Show that every orbit in the region y > 0, z > 0 of the system
yz
y′ = y(2 − y) − ,
y+1
yz
z′ = 4 −z
y+1
approaches a periodic orbit as t → ∞.
Solution: We have y ′ = 0 when y = 0, and z ′ = 0 when z = 0, so orbits
starting in the first quadrant remain in the first quadrant. Equilibria are the
z 4y
solutions of either y = 0 or 2 − y = y+1 , and either z = 0 or y+1 = 1. If y = 0,
z
we must also have z = 0. If 2 − y = y+1 , we could have z = 0, which implies
y = 2, or y = 31 , which implies z = 20 9 . Thus there are three equilibria, namely
(0,0), (2,0), and ( 13 , 20
9 ). By checking the values of the trace and determinant
of the community matrix, which is
y∞
" #
z∞
2 − 2y∞ − (1+y ∞)
2 − y∞ +1
4z∞ 4y∞ ,
(y∞ +1)2 y∞ +1 − 1

we may see that each of the three equilibria is unstable. In order to apply the
Poincaré-Bendixson theorem, we must show that every orbit starting in the
first quadrant of the phase plane is bounded.
To show this, we might like to show that y ′ and z ′ are negative when
y and/or z are sufficiently large, but a glance at the equations tells us this
isn’t necessarily so. Therefore we instead consider some positive combination
of y and z whose time derivative does become negative far enough from the
origin. In particular, consider the function V (y, z) = 4y + z. If an orbit is
unbounded, then along this orbit the function V (y, z) must also be unbounded.
The derivative of V (y, z) along an orbit is

d
V [y(t), z(t)] = 4y ′ (t) + z ′ (t) = 4y(2 − y) − z.
dt
This is negative except in the bounded region defined by the inequality z <
4y(2 − y). Therefore the function V (y, z) cannot become unbounded, because
it is decreasing (dV /dt < 0) whenever it becomes large (z > 4y(2−y), which is
true, for example, whenever V > 9). This proves that all orbits of the system
318 Dynamical Systems for Biological Modeling: An Introduction

are bounded. Now we may apply the Poincaré-Bendixson theorem to see that
every orbit approaches a limit cycle. 

For autonomous systems of more than two differential equations there is no

result analogous to the Poincaré-Bendixson theorem. Orbits of such systems
may behave in very strange ways, and it is not possible to give a description
of all the possibilities. One possibility is chaotic behavior similar to what
we saw in Section 2.6 for difference equations. However, the analogue of the
linearization theorem is valid, and it is still possible to decide whether an
equilibrium is asymptotically stable.

Exercises
In Exercises 1–6, for each equilibrium of the given system determine whether
the equilibrium is asymptotically stable or unstable.

1. y ′ = y + z − 2, z ′ = z − y (cf. Exercise 1, Section 5.2)

2. y ′ = y + z − 1, z ′ = y (cf. Exercise 2, Section 5.2)
3. y ′ = y + z 2 , z ′ = y + 1 (cf. Exercise 3, Section 5.2)
4. y ′ = z − 2, z ′ = y 2 − 8z (cf. Exercise 4, Section 5.2)
5. y ′ = ez , z ′ = e−y (cf. Exercise 5, Section 5.2)
6. y ′ = z, z ′ = sin y (cf. Exercise 6, Section 5.2)
7. * Determine the behavior of orbits of the system

y ′ = y(1 − y − 2z),
z ′ = z(1 − 2y − z).

8. * Determine the behavior of orbits of the system

yz
y ′ = y(2 − y) − y+1 ,
yz
z ′ = 2 y+1 − z.

9. * Determine the behavior of orbits of the system

y ′ = y(λ − ay − bz), z ′ = z(−µ + cy).

10. * Determine the behavior of orbits of the system

   
y z
y ′ = ry 1 − , z ′ = sz 1 − .
K + az M + by
 Systems of Differential Equations 319

11. * Consider the system

y ′ = z, z ′ = −y − z 3 .

(a) Show that (0,0) is the only equilibrium.

(b) Show that the linearization of the system at the origin has a center.
(c) Show that the non-negative function V (y, z) = y 2 + z 2 decreases
along every orbit of the system and tends to zero, which implies that
the origin must be a spiral point of the system.

12. * Consider the system in polar coordinates (r, θ)

r′ = r(1 − r), θ′ = 1.

(a) Show that r approaches 1 as t → ∞, and θ increases unboundedly.

(b) Deduce that the circle r = 1 is a limit cycle which every orbit except
the constant solution r = 0 approaches.
13. * The system in Example 5 has a predator-prey structure, with the term
common to both equations representing the interaction rate between
predators and prey.
(a) Which variable represents the predators, and which the prey?
(b) What is the significance of the 2 in the dy/dt equation?
(c) The interaction rate saturates as which of the two populations
grows?

Miscellaneous exercises
In each of Exercises 1–2, describe the orbits of the given system.

1. y ′ = y 2 z, z ′ = zy 2
2. y ′ = e−z , z ′ = ey
320 Dynamical Systems for Biological Modeling: An Introduction
In each of Exercises 3–6, find the linearization of the given system at each
equilibrium.
3. y ′ = y + z − 4, z ′ = y − z 5. y ′ = y 2 z, z ′ = zy 2
4. y ′ = z − 1, z ′ = y 6. y ′ = e−z , z ′ = ey

In each of Exercises 7–8, for each equilibrium of the given system determine
whether the equilibrium is asymptotically stable or unstable.
7. y ′ = y 2 z, z ′ = zy 2 8. y ′ = e−z , z ′ = ey

In each of Exercises 9–16, find the general solution of the given system by
analytic solution, use a computer algebra system to examine the behavior
of solutions, and classify the origin as a node, saddle point, center, or spiral
point.
9. y ′ = z − 3y, z ′ = 5y + z 13. y ′ = −z, z ′ = y
10. y ′ = y, z ′ = y + z 14. y ′ = y − z, z ′ = y + z
11. y ′ = 2y + z, z ′ = z 15. y ′ = y + z, z ′ = y + 2z
12. y ′ = −y, z ′ = y − z 16. y ′ = y + z, z ′ = y + z

17. A tank contains 1000 liters of water and a salt solution containing 20
kg./liter is pumped into it at a rate of 100 liters /minute. The mixture
is led to a second tank containing 1000 liters of water at a rate of 100
liters/minute, and the mixture is pumped out of the second tank at the
same rate. What is the concentration of salt in the second tank after
one hour?
18. Two tanks begin with 10 kg. of salt dissolved in 100 liters of water.
Water is pumped into the first tank at a rate of 20 liters/minute, the
mixture is pumped from the first tank to the second tank at a rate of
20 liters/minute, and the mixture is pumped out of the second tank at
a rate of 20 liters/minute. What is the amount of salt in each tank after
one hour, and what is the amount of salt in each tank after a very long
time?

In Exercises 19–22, for each equilibrium of the given system determine whether
the equilibrium is asymptotically stable or unstable.
19. y ′ = y + z − 4, z ′ = y − z 21. y ′ = y 2 z, z ′ = zy 2
20. y ′ = z − 1, z ′ = y 22. y ′ = e−z , z ′ = ey
Chapter 6
Topics in Modeling Systems of
Populations

There are many questions involving the interaction of two different popula-
tions. These different populations may be members of a single population but
distinguished by gender, age, or their infection status with respect to a dis-
ease present in the population, or they may be members of two quite different
species, cooperating, competing for a common resource, or in a predator-prey
relationship. The modeling of such questions leads naturally to systems of
differential equations, typically with each differential equation describing one
of the interacting populations. This chapter is devoted to some examples, de-
scribing applications of the general theory of systems of differential equations
developed in the previous chapter.

6.1 Epidemiology: Compartmental models

6.1.1 An epidemic model
In Sections 3.2.3 and 4.3.2 we considered models for the spread of an
infectious disease in which a population was divided into susceptibles and in-
fectives; the underlying assumptions were that there was a rate of contracting
the infection which depended on the number of susceptibles and the number
of infectives, that there was a rate of recovery depending on the number of
infectives, and that on recovery infectives returned to the susceptible class. In
other words, it was assumed that there was no immunity against re-infection
after recovery from the infection.
In this section, we shall consider some models for the spread of infectious
diseases which include a third class, of removed members. Many diseases,
especially diseases caused by viral agents, including smallpox, measles, and
rubella (German measles), provide immunity against re-infection.
We let S(t) denote the number of susceptibles, I(t) the number of infec-
tives, and R(t) the number of removed members. We assume that the popula-
tion has constant total size K, so that S(t) + I(t) + R(t) = K. We will derive
differential equations expressing the rate of change of the size of each of the

321
322 Dynamical Systems for Biological Modeling: An Introduction

three classes, but in each case one of the equations may be eliminated since
we can use the above relation to find S, I or R in terms of the other two. Thus
we will obtain a system of two differential equations to describe the spread of
diseases for which there is a removed class.
A model was proposed by W. O. Kermack and A. G. McKendrick1 to
explain the rapid rise and fall of cases frequently observed in epidemics, in-
cluding the Great Plague of 1665–66 in England, cholera in London in 1865,
and plague in Bombay in 1906. This model is

S ′ = −βSI,
I ′ = βSI − γI, (6.1)
R′ = γI.

The only difference from the model of Section 3.2.3 is that the term γI now
represents a rate of transition from the class I to the class R, instead of a rate
of return to the class S. The rate of recoveries in unit time is γI, and the rate
of transmission of infection from infectives to susceptibles is βSI. Note that
this model is only appropriate if the duration of an outbreak is short enough
that demographics (natural births and deaths) can be ignored.
We consider the model as a system of two equations, viewing R as deter-
mined by S and I, R = K − S − I, since the first two equations do not involve
R:

S ′ = −βSI, (6.2)
′
I = βSI − γI.

The equilibria of the system (6.2) (seen in slightly different form in Section 5.2,
Example 2) are the solutions of the pair of equations βSI = 0, βSI − γI = 0.
The first of these implies that either S = 0 or I = 0. If S = 0, the second
equation is satisfied only if I = 0, while if I = 0 the second equation is satisfied
for every S. Thus there is a line of equilibria (S∞ ,0) with S∞ arbitrary, 0 ≤
S∞ ≤ K. If we compute the linearization of the system (6.2) at an equilibrium
(S∞ ,0), we obtain

u′ = −βS∞ v,
′
v = (βS∞ − γ) v,

and the linearization theorem of Section 5.2 cannot be applied.

In order to obtain an understanding of the qualitative behavior of solutions
of the system (6.2), we observe from (6.2) that S ′ < 0 whenever S > 0, I > 0.
This means that the function S(t) decreases for all t. In addition, I ′ < 0
whenever I > 0, βS < γ, while I ′ < 0 if I < 0, βS > γ. Thus if S(0) < βγ ,
S(t) remains less than βγ for all t, and I(t) decreases to zero as t increases.

1 W.O. Kermack and A.G. McKendrick, A contribution to the mathematical theory of

epidemics, Proc. Roy. Soc. London 115 (1927), 700–721.

 Topics in Modeling Systems of Populations 323

However, if S(0) > βγ , then I(t) increases so long as βS > γ, and thus I(t)
increases initially before decreasing to zero. We think of introducing a small
number of infectives into a susceptible population so that I(0) = ǫ > 0,
S(0) = K − ǫ. Then if βK/γ < 1, I(t) decreases monotonically to zero and
the infection dies out. On the other hand, if βK/γ > 1, an epidemic occurs,
as S(0) > βγ (for ǫ small), so I(t) increases to a maximum and then decreases
to zero. This is another threshold theorem of Kermack and McKendrick with
the threshold quantity βK/γ. This threshold quantity distinguishes between
two possible behaviors just like the threshold quantity in Section 3.2.3, but
the possible behaviors are not the same as in Section 3.2.3.
One might suppose that the reason for the eventual disappearance of the
infection in the epidemic case is that all susceptibles become infected, but
observations of epidemics indicate that this is not the case. The model (6.2)
agrees with observation in that it implies that the limiting value S(∞) =
limt→∞ S(t) of every solution of the system (6.2) obeys S(∞) > 0. We may
see this by calculating

γ S ′ (t)
 
d γ
S(t) + I(t) − log S(t) = S ′ (t) + I ′ (t) −
dt β β S(t)
γ
= −βSI + [βSI − γI] − (−βI) = 0
β

(motivated by observing that S ′ + I ′ = −γI, and finding a function of S and I

whose time derivative is γI). Thus S(t) + I(t) − βγ log S(t) is a constant. Since
I(∞) = 0 and I(0) ≈ 0, this gives
γ γ
S(∞) − log S(∞) = S(0) − log S(0).
β β
It follows that
γ γ S(0)
S(0) − S(∞) = [log S(0) − log S(∞)] = log ,
β β S(∞)

and therefore h i
S(0)
log S(∞)
β/γ = . (6.3)
S(0) − S(∞)
The quantity β/γ is known as the contact number. Not only does (6.3) imply
S(∞) > 0 (because if S∞ were zero, the right side of (6.3) would be infinite but
the left side is finite), but it also gives a means of estimating the contact rate β,
which generally cannot be measured directly. By making a serological survey
(testing for immune responses in the blood) in the population before and after
an epidemic, one may estimate S(0) and S(∞), and then (6.3) gives β/γ. If
the mean infective period 1/γ is known as well, then β can be calculated. The
contact rate β depends on the disease as well as on other factors such as the
rate of mixing in the population.
324 Dynamical Systems for Biological Modeling: An Introduction

30

25

20

l(t)
15

10

0 50 100 150 200 250

S(t)
Photo by Smb1001

FIGURE 6.1: During the Great Plague of FIGURE 6.2: A phase por-
1665–66, the village of Eyam in England vol- trait for model (6.2).
untarily quarantined itself in hopes of pre-
venting the plague from spreading to neigh-
boring villages. Inhabitants of the neighbor-
ing populations left food and other supplies
for Eyam residents at the Boundary Stone
(pictured) just outside the village.

For example, the village of Eyam in England maintained isolation from

other villages during the Great Plague of 1665–66 (see Figure 6.1), and its
population decreased from 350 to 83 during the course of the epidemic. There
is reason to believe that there were actually two separate epidemics, the first of
which reduced the susceptible population to 254. By substituting S(0) = 254,
S(∞) = 83 into (6.3), we obtain

log 254
β/γ = 83
= 6.54 × 10−3 .
254 − 83
The infective period was 11 days, or 0.3667 months. Using a month as the
unit of time we obtain the estimate β = 0.0178. This data, with 7 initial
infectives, gives the phase portrait of Figure 6.2, traversed from right to left
as time progressed and the number of susceptibles decreased. Note that in this
case infected individuals were removed to the R class through death for the
most part, rather than recovery with immunity. Our simple model (6.2) still
describes this process, however different the interpretation may be, as the R
class of the model simply includes individuals no longer involved in the spread
of the disease.
The criterion βK/γ > 1 for the establishment of a disease can also be
expressed as the requirement that the susceptible population density exceeds
a certain critical value βγ . For fox rabies in Europe, observations indicate a
 Topics in Modeling Systems of Populations 325

critical population density of approximately 1 fox/km2 ; rabies dies out in re-

gions which are more sparsely populated. This data together with the average
life expectancy of 5 days for a rabid fox gives the estimate β ≈ 72 km2 /fox
year.
In order to avoid an epidemic, it is necessary to reduce below 1 the quan-
tity βK/γ, which is called the basic reproductive number and often denoted
by R0 . This may sometimes be achieved by immunization, which has the ef-
fect of transferring members from the susceptible class to the removed class
and thus reducing S(0). (This idea was first mentioned briefly in Section 5.2,
Example 2.) If we immunize a fraction p of the susceptible population, we
would replace K by (1 − p)K, and this would give a basic reproductive num-
ber βK(1 − p)/γ. In order to make this basic reproductive number less than 1,
γ γ
we require βK(1 − p)/γ < 1. This is equivalent to 1 − p < βK , or p > 1 − βK .
A population is said to have herd immunity if a sufficiently large fraction
has been immunized to reduce the basic reproductive number below 1 and
thus assure that the disease will not spread if an infective is introduced into
the population (the term suggests that one can be protected from infection
if enough of one’s neighbors are immune, see Figure 6.3). The only infectious
disease for which this has actually been achieved worldwide is smallpox. For
measles, epidemiological data in the U.S. indicates a basic reproductive num-
ber ranging from 5.4 to 6.3 in rural areas, requiring vaccination of 81.5% to
84.1% of the population to achieve herd immunity. In urban areas, the basic
reproductive number ranges from 8.3 to 13.0, requiring vaccination of 88.0%
to 92.3% of the population. As measles vaccination is only about 95% effec-
tive (for vaccination at age 15 months) and not all people are willing to allow
vaccination, it is impossible in practice to achieve herd immunity. For small-
pox, the basic reproductive number is about 5, requiring 80% vaccination to
achieve herd immunity. This is feasible because the consequences of smallpox
are dire enough to encourage immunization.

Example 1.
A survey of freshman students at Yale University2 found that 25% were sus-
ceptible to rubella at the beginning of the year and 9.65% were susceptible at
the end of the year. What fraction would have had to be immunized to avoid
the spread of rubella?
Solution: Using S(0) = 0.25, S(∞) = 0.0965 and substituting in (6.2), we
obtain
0.25
log 0.0965
β/γ = = 6.20.
0.25 − 0.0965
γ
In order to avoid the spread of rubella, the requirement is S(0) < β = 0.16.

2 A.S. Evans, Viral Infections of Humans, 2nd ed., Plenum Press, New York (1982),

reported by H. W. Hethcote, Three basic epidemiological models, Applied Mathematical

Ecology, S. A. Levin, T. G. Hallam, and L. J. Gross (eds), Biomathematics 18, Springer-
Verlag, New York-Heidelberg (1989), 119–144.
326 Dynamical Systems for Biological Modeling: An Introduction

Photo by AlMare

FIGURE 6.3: The notion of herd immunity comes from the fact that mem-
bers of a homogeneously mixing population (a “herd”) making potentially
infectious contacts with neighbors can be protected from infection if enough
of those neighbors are immune. Imagine here that the individuals in the car
in the background are in close contact with the neighboring members of the
herd surrounding them (here, a herd of goats on a road in Greece). If enough
of those neighbors are vaccinated against infection, then by the time the in-
dex individual comes into contact with an unvaccinated individual, the latter
may no longer be infected. With the infection rate reduced so drastically, the
infection dies out in the population.

This could be achieved by immunizing an additional 9% of the class, or

0.25−0.16
0.25 = 36% of the susceptible students. 

Example 2.
In Example 2, Section 3.2.3, a disease was described spreading in a population
1
of 1200 members. Suppose the disease, with β = 3000 , 1/γ = 6 days, had
conferred immunity on recovered infectives. How many members would have
had to have been immunized to avoid an epidemic?
1
Solution: The basic reproductive number is βK/γ = 3000 × 6 × 1200 = 2.4.
γ
Thus herd immunity, p > 1 − βK , would require immunization of a fraction
1
1 − 2.4 = 0.5833, or 700 members. 

It may be important to know the maximum number of infectives at any

given time, for example to be able to arrange enough facilities for isolation
and treatment. From the model (6.2), we know that the maximum of I occurs
when S = βγ , and also that the quantity S + I − βγ log S is constant. Thus if
 Topics in Modeling Systems of Populations 327
γ
t∗ is the time when S = β, we have

γ γ
S(t∗ ) + I(t∗ ) − log S(t∗ ) = S(0) + I(0) − log S(0)
β β
and
γ γ γ γ
+ I(t∗ ) − log = S(0) − log S(0).
β β β β
From this we conclude that I(t∗ ), the maximum number of infectives, is given
by

γ γ γ γ γ γ log βS(0)
I(t∗ ) = S(0) − log S(0) + log − = S(0) − − . (6.4)
β β β β β β γ
For the Great Plague in Eyam, this gives a maximum infective population of
30.4, confirmed by the phase portrait of Figure 6.2.

Example 3.
What is the maximum number of infectives in the rubella epidemic of Exam-
ple 1?
Solution: In Example 1 we were given S(0) = 0.25, and we calculated β/γ =
6.20. Then (6.4) gives
1 1
I(t∗ ) = 0.25 − − log(6.20)(0.25) = 0.018.
6.20 6.20
Thus at most 1.8% of the population is infected at any one time. 

6.1.2 A model for endemic situations

The model (6.2) studied earlier in this section is appropriate for describing
an epidemic, a single outbreak of a disease. Many diseases are always present,
especially in less-developed countries, and this is described by saying that they
are endemic. In order to model a long-term endemic situation, we must add
demographics—births and natural deaths—to the model (6.2). If we assume a
proportional death rate µ in each compartment and a birth rate µK (to keep
the total population size constant), we would have a model

S′ = µK − βSI − µS (6.5)
I′ = βSI − (γ + µ)I.

We analyze this model by finding equilibria and checking their asymptotic

stability. Equilibria (S, I) are solutions of the pair of equations

βSI = µ(K − S), βSI = (γ + µ)I.

The second of these equations factors into two possibilities, namely I = 0

328 Dynamical Systems for Biological Modeling: An Introduction

(disease-free equilibrium) and βS = γ + µ (endemic equilibrium). For a mean-

ingful endemic equilibrium, we must have S ≤ K, or
βK
R0 = ≥ 1.
γ+µ
This quantity R0 has an epidemiological meaning, because a single infective in
a wholly susceptible population would make βK contacts in unit time, all of
which would produce new infections, and would continue to transmit infection
for a mean infective period (corrected for natural deaths) of 1/(γ + µ). Thus
R0 is the total number of secondary infections caused by a single infective
inserted into a wholly susceptible population.
At the disease-free equilibrium, S = K, I = 0. At an endemic equilibrium,
if there is one (that is, if R0 > 1), it is easy to calculate that
 
γ +µ µ 1
S= , I= 1− K.
β γ+µ R0
The matrix of the linearization of (6.5) at an equilibrium (S, I) is
 
−(βI + µ) −βS
.
βI βS − (γ + µ)

At the disease-free equilibrium, this matrix is

 
−µ −βK
,
0 βK − (γ + µ)

and it is easy to see that this has negative eigenvalues if and only if R0 < 1.
Thus the disease-free equilibrium is asymptotically stable if and only if R0 < 1.
At the endemic equilibrium (S, I) the matrix is
 
−(βI + µ) −βS
.
βI 0

Since this matrix has negative trace and positive determinant, the endemic
equilibrium is always asymptotically stable if it exists. We may summarize the
analysis by saying that there is always a single asymptotically stable equilib-
rium, the disease-free equilibrium if R0 < 1 and the endemic equilibrium if
R0 > 1.
For most human diseases, the mean infective period 1/γ is less than one
month, much smaller than the mean life span 1/µ, which is onthe order of 70
years. Thus an endemic equilibrium, with
 
µ 1
I= 1− K,
γ +µ R0
the number of infectives is a tiny fraction of the carrying capacity. In a pop-
ulation of moderate size, say 1000, the number of infectives at an endemic
 Topics in Modeling Systems of Populations 329

equilibrium might be less than 1, and small random effects might be enough
to wipe out the infective population. A numerical simulation of an epidemic
model (6.2) and an endemic model (6.5) would appear indistinguishable. How-
ever, in a large population of, say 1, 000, 000, there might be 1, 000 infectives
at an endemic equilibrium, and this is a number of disease cases large enough
to be significant.
For animal diseases, where the life span may be much shorter and the dis-
ease infective period may be much longer, the differences between an epidemic
and an endemic situation may be much more readily noticed. An extreme
example would be a disease with no recovery, such as rinderpest (a cattle
disease of ancient origin which has quite recently become the second disease,
after smallpox, to be eliminated). Such a disease would be described by an SI
model, like (6.5) but with γ = 0,

S′ = µK − βSI − µS, (6.6)

I′ = βSI − µI,

having
βK
R0 = ,
µ
and an endemic equilibrium
 
µ 1
S= , I= 1− K.
β R0

Exercises
1. The same survey of Yale students described in Example 1 reported that
91.1% were susceptible to influenza at the beginning of the year, and
51.4% were susceptible at the end of the year. Estimate the contact
number β/γ and decide whether there was an epidemic.
2. An influenza epidemic was reported at an English boarding school in
1970 which spread to 512 of the 673 students. Estimate the contact
number β/γ.
3. What fraction of the Yale students of Exercise 1 would have had to be
immunized to prevent an epidemic?
4. What fraction of the boarding school students of Exercise 2 would have
had to be immunized to prevent an epidemic?
5. What was the maximum number of Yale students of Exercises 1 and 3
missing classes because of influenza at any given time?
6. What was the maximum number of boarding school students of Exercises
2 and 4 suffering from influenza at any given time?
330 Dynamical Systems for Biological Modeling: An Introduction

7. * If some members of a population susceptible to a disease which pro-

vides immunity against re-infection moves out of the region of the epi-
demic, at a per capita rate λ, the situation may be modeled by a system

S ′ = −βSI − λS, I ′ = βSI − γI.

Show that both S and I approach zero as t → ∞.

8. * Consider a model for a disease which confers only temporary immunity
after recovery, so that recovered individuals lose their immunity at a per
capita rate of c (per time unit).
(a) Reformulate model (6.1) to reflect the new structure for the disease
cycle.
(b) Reduce the model to a two-dimensional system as done for (6.1), by
eliminating R.
(c) Identify the equilibria.
(d) Describe the qualitative behavior of the system. Is there a threshold
condition for this SIRS model, as there was with the SIR model?
9. Perform numerical simulations for the models (6.2) and (6.5) with pa-
rameter values K = 1000, β = 1/2000, γ = 1/4, µ = 1/25, 000 and a
variety of initial values.
10. Perform numerical simulations for the SI model (6.6) with parameter
values A = 100, β = 1/2000, µ = 1/10 and a variety of initial values.
Calculate the basic reproduction number and equilibrium susceptible
and infective population sizes. Does the model approach an endemic
equilibrium directly or does it go through an epidemic-like behavior
first like the SIR model?
11. Suppose that a vaccine is developed and approved for an infection which
confers immunity upon recovery, and that susceptible individuals get
vaccinated at a per capita rate φ. How would the SIR model systems
with (6.5) and without (6.2) demographics change to incorporate this
vaccination program, if vaccinated individuals have the same immunity
(assumed lifelong) as recovered individuals?
12. * Identify the equilibria of the model in the previous exercise for the
system with demographics, and analyze their stability. From the results,
deduce the expression for the control reproductive number, analogous to
R0 , and interpret the difference from the expression given in the text for
R0 .
13. Rewrite the SI model (6.6) to incorporate vertical transmission, in which
a proportion p of offspring born to infected mothers are born infected.
State any additional assumptions.
 Topics in Modeling Systems of Populations 331

6.2 Population biology: Interacting species

The Lotka-Volterra system considered in Section 5.1 is an example of a
model for the sizes of two interacting populations. The formulation of models
for two interacting species depends on the nature of the interaction as well as
on the assumptions about the behavior of each population in the absence of
the other population. In this section, we shall examine some models for two
different kinds of interaction of two species, namely species in competition,
species in a predator-prey relation, and symbiotic relations. These are the
three possible relationships between two species: both species may be dam-
aged (competition), one species may profit at the expense of the other species
(predator-prey relationship), or each species may benefit from the presence of
the other species (symbiosis).

6.2.1 Species in competition

Let us consider two species whose population sizes at time t are y(t) and
z(t), respectively. Suppose that each species would grow according to a logistic
law if there were no interaction with the other species. Suppose also that the
two species are competing for resources, and that the effect of this competi-
tion is to decrease the per capita growth rate of each species by an amount
proportional to the population size of the other species. These assumptions
lead to a model of the form
y ′ = y(λ − ay − bz) (6.7)
z ′ = z(µ − cy − dz)
with λ, µ, a, b, c, and d positive constants. The carrying capacities of the two
species are, respectively, λa and µd . It is easy to see that (0,0), ( λa ,0), and (0, µd )
are equilibria of the system (6.7). In addition, there may be an equilibrium
which we will call (y∞ ,z∞ ) with y∞ > 0, z∞ > 0, if the lines ay + bz = λ and
cy + dz = µ intersect in the first quadrant. Their point of intersection,
dλ − bµ aµ − cλ
y∞ = , z∞ = , (6.8)
ad − bc ad − bc
which exists as long as ad 6= bc, lies in the first quadrant when the numerators
and denominator in (6.8) are either all positive or all negative. Some algebra
shows that this happens when the relative growth ratio λ/µ (of y to z) lies
between a/c and b/d.
Because F (y, z) = λy − ay 2 − byz and G(y, z) = µz − cyz − dz 2 , the
community matrix at an equilibrium (y, z) is
 
λ − 2ay − bz −by
.
−cz µ − cy − 2dz
332 Dynamical Systems for Biological Modeling: An Introduction

Thus at the equilibrium (0,0) this matrix has trace λ + µ > 0 and determinant
λµ > 0, and the equilibrium is unstable. At the equilibrium ( λa ,0), the matrix
has trace −λ+ aµ−cλ b and determinant −λ( aµ−cλ
a ), so the trace is negative and
the determinant positive if aµ − cλ < 0. A similar argument shows that the
equilibrium (0, µd ) is asymptotically stable if dλ − bµ < 0. For the equilibrium
(y∞ ,z∞ ), if there is one, the community matrix is
 
−ay∞ −by∞
.
−cz∞ −dz∞

Thus the trace is negative, and the determinant is positive if and only if
ad − bc > 0 (or a/c > b/d).
Therefore the equilibrium (0,0) is always unstable, and we may summarize
the results for the other three equilibria as follows:

I. If b/d < λ/µ < a/c, there is an equilibrium (y∞ ,z∞ ) which is asymptot-
ically stable, and the other three equilibria are unstable.
II. If a/c < λ/µ < b/d, there are an unstable equilibrium (y∞ ,z∞ ) and two
locally asymptotically stable equilibria ( λa ,0) and (0, µd ).

III. If λ/µ < a/c, b/d, the equilibrium ( λa ,0) is unstable and the equilibrium
(0, µd ) is asymptotically stable.

IV. If λ/µ > a/c, b/d, the equilibrium ( λa ,0) is asymptotically stable and the
equilibrium (0, µd ) is unstable.

Writing the possibilities in terms of the above ratios allows us to consider

the relative importances of the natural growth rates (λ/µ) and the limiting
effects of species y (a/c) and species z (b/d). Thus in case I the relative growth
and self-limiting ratios for each species dominate the relative effects of com-
petition — for species y relative to species z, this is b/d < λ/µ < a/c as
written above (competition for species y means the effects of species z), while
for species z relative to species y this is c/a < µ/λ < d/b, with c/a repre-
senting the relative effects of competition with species y. Here competition
is not severe enough to prevent coexistence of the two species. Mathemati-
cally speaking, in this case there is one asymptotically stable equilibrium with
both species surviving, and every orbit in the first quadrant approaches this
equilibrium.
In case II, the reverse of the above is true: the relative effects of competition
dominate both the relative natural growth and self-limiting ratios, a/c <
λ/µ < b/d as given above for species y relative to z, or d/b < µ/λ < c/a for
species z relative to y. Here the level of competition is so high that the two
species cannot coexist. Therefore there are two locally asymptotically stable
equilibria, each corresponding to survival of one species and extinction of the
other. Which species wins the competition depends on the initial values.
 Topics in Modeling Systems of Populations 333

FIGURE 6.4: Case I: Coexistence. FIGURE 6.5: Case II: Competitive

exclusion.
@

FIGURE 6.6: Case III: z survives. FIGURE 6.7: Case IV: y survives.

In case III, the relative growth ratio for species y (relative to z) is dom-
inated by both of the limiting ratios, λ/µ < a/c, b/d, while the relative
growth ratio for species z (relative to y) exceeds the two limiting ratios,
µ/λ > c/a, d/b. Therefore the more robust z-species survives, and the y-
species is wiped out.
In case IV, the tables are turned: the relative growth ratio for y (to z)
exceeds the two limiting ratios, λ/µ > a/c, b/d, while the relative growth
ratio for z (to y) is less than the two limiting ratios, µ/λ < c/a, d/b. Thus the
y-species survives, and the z-species is wiped out.
The four cases may also be distinguished by the locations of the isoclines,
which are all straight lines, as shown in Figures 6.4–6.7 for the four cases,
respectively. In these figures, each asymptotically stable equilibrium has been
marked @. By drawing the two nullclines for any competitive system and not-
ing their relative positions, we may identify which of the four cases describes
the system.
334 Dynamical Systems for Biological Modeling: An Introduction

Photo courtesy Pennsylvania Dept. of Conservation Photo by Greg Webster

and Natural Resources—Forestry Archive, Bugwood.org

FIGURE 6.8: Competition occurs in southeastern Canada and the north-

eastern United States between the gypsy moth, Lymantria dispar (left), and
the northern tiger swallowtail, Papilio canadensis (right).

Suppose now that we have two species of moth competing for the same
food supply, with respective initial per capita reproduction rates of 100 and
60 (in per capita per time units), and carrying capacities on a given patch
of habitat (such as a tree or field) of 25 and 30. (Figure 6.8 illustrates one
such competition, between the gypsy moth and the northern tiger swallowtail
butterfly.) In the next two examples we shall see that the extent to which
competition adversely affects each species determines whether the two can
coexist, or, if not, which one survives.

Example 1.
Determine the outcome of a competition modeled by the system

y ′ = y(100 − 4y − 4z), z ′ = z(60 − y − 2z).

Solution: A coexistence equilibrium is found by solving the system of algebraic

equations 4y + 4z = 100, y + 2z = 60. If we subtract double the second
equation from the first we obtain 2y = −20, and thus there is no coexistence
equilibrium. Alternatively, we note that λ/µ = 5/3 is not between a/c = 4
and b/d = 2.
The two equilibria with one species surviving are (25,0) and (0,60). We
know that every orbit must approach one of these two equilibria. In order to
′
decide which, we note that as (y, z) → (0, 60), yy = 100 − 4y − 4z → −240 < 0,
′
while as (y, z) → (25, 0), zz = 60 − y − 2z → 35 > 0. This shows that (y, z)
must approach the equilibrium (0,60), because in order to approach (25,0) it
would be necessary to have z ′ < 0. Alternatively, λ/µ < a/c, b/d, so we are in
case III above. We now see that the z-species wins the competition. 
 Topics in Modeling Systems of Populations 335

Example 2.
Now suppose instead that species y is only affected one quarter as much by
the presence of species z — that is, that each z individual reduces species y’s
per capita growth rate by one instead of four. Determine the outcome of the
competition modeled by the resultant system

y ′ = y(100 − 4y − z), z ′ = z(60 − y − 2z).

Solution: The system of equilibrium conditions 4y + z = 100, y + 2z = 60

can be solved by eliminating one of the variables to obtain the coexistence
equilibrium (20,20).To determine this equilibrium’s asymptotic stability, we
use the fact that the community matrix at the equilibrium (y∞ , z∞ ) is
   
−ay∞ −by∞ −80 −20
=
−cz∞ −dz∞ −20 −40

with negative trace and positive determinant. Thus the coexistence equilib-
rium is asymptotically stable, and every orbit approaches it. Alternatively, we
note that λ/µ now lies between the other two ratios, b/d = 1 < λ/µ = 5/3 <
a/c = 4, as in case I. In this case, the level of competition (coefficients b and
c) is small enough for both species that they can coexist. 

Finally, we revisit the model for plant competition first explored in Sec-
tion 3.2.2. Equation (3.17) gave the general equation for a ranked system
of competitors, in which the proportion pi of habitat occupied by species i
changes according to four processes: colonizing uninhabited patches, natural
mortality, displacing inferior competitors (with index j > i) and being dis-
placed by superior competitors (with index j < i). Analysis showed that each
species persisted at a positive equilibrium p∗i if and only if its colonization rate
ci exceeded its own mortality rate Pm i and the superior competitors left some
i−1
space available for colonization, j=1 pj < 1. However, Bampfylde observed
that displacement does not occur in rainforests;
Pn a competition model includ-
ing only colonization of open spaces (1 − j=1 pj for n species) and natural
mortality, but no displacement, becomes (cf. (2.18))
 
n
dpi X
= ci pi 1 − p j  − mi p i (6.9)
dt j=1

for i = 1, 2, ..., n.
We analyze system (6.9) through its equilibria,
Pn found by setting dpi /dt = 0.
For each i this yields either p∗i = 0 or 1 − j=1 p∗j = mi /ci . In general the
values of mi and ci are independent from one species to another, so it is
unlikely to have mi /ci = mj /cj for
Pnany i 6= j; this then makes it impossible for
the proportion of free sites (1 − j=1 pj ) to match simultaneously more than
one mi /ci as in the equilibrium conditions. Thus no more than one species
336 Dynamical Systems for Biological Modeling: An Introduction

can have a nonzero equilibrium value—that is, any equilibrium represents

a scenario where all species but one die out, an illustration of the so-called
principle of competitive exclusion in ecology, which holds that only one species
or population can occupy a given ecological niche (food, habitat, etc.) at a
given time.
An example may help to illustrate. Consider a system with three such
competing species. The equilibrium conditions are (dp1 /dt = 0) either p∗1 = 0
P3 P3
or 1 − j=1 p∗j = m1 /c1 , (dp2 /dt = 0) either p∗2 = 0 or 1 − j=1 p∗j = m2 /c2 ,
P3
and (dp3 /dt = 0) either p∗3 = 0 or 1 − j=1 p∗j = m3 /c3 . If we assume that
m1 /c1 , m2 /c2 , and m3 /c3 all have different values, then at most one of them
P3
can be equal to 1 − j=1 p∗j . Thus the equilibria are (0, 0, 0) [extinction of all
3 species], (1 − m m2
c1 , 0, 0) [only species 1 persists], (0, 1 − c2 , 0) [only species 2
1

m3
persists], and (0, 0, 1 − c3 ) [only species 3 persists]. We can determine when
each equilibrium is asymptotically stable using the community matrix, which
calculations show to be
 
c1 (g − p1 ) − m1 −c1 p1 −c1 p1
 −c2 p2 c2 (g − p2 ) − m2 −c2 p2 ,
−c3 p3 −c3 p3 c3 (g − p3 ) − m3

where g = 1 − p1 − p2 − p3 is the gap (the proportion of unoccupied sites). At

the first equilibrium this simplifies to
 
c1 − m 1 0 0
 0 c2 − m 2 0 ,
0 0 c3 − m 3

which readers familiar with linear algebra will observe has the eigenvalues
(solutions to the characteristic equation) c1 − m1 , c2 − m2 and c3 − m3 . Thus
the extinction equilibrium is asymptotically stable if and only if ci < mi for all
3 species (that is, each species’ mortality rate outstrips its ability to colonize
new territory), just as for the displacement model of Chapter 3.
At the second equilibrium, in which species 1 wins the competition, the
community matrix simplifies to

−c1 p∗1 −c1 p∗1

 
m 1 − c1
 0 c2 (m1 /c1 ) − m2 0 ;
0 0 c3 (m1 /c1 ) − m3

once again the eigenvalues are the diagonal entries, and this equilibrium is
asymptotically stable if they are all negative, i.e.,
 
c1 c2 c3
> max 1, , .
m1 m2 m3

We can extrapolate similar conditions for the asymptotic stability of the other
single-survivor equilibria, so that in the end the species with the highest ci /mi
 Topics in Modeling Systems of Populations 337

ratio wins the competition. (We omit here the discussion of issues related to
global stability which would be required for a rigorous proof.)
This result suggesting competitive exclusion may appear counter-intuitive
when the dominant competitor will not (at equilibrium) occupy the entire
rainforest: the model which allows displacement predicts coexistence as long
as there is enough habitat for all, and without displacement the inferior com-
petitors would appear to be at an advantage relative to the same scenario with
displacement. However, what constitutes a superior competitor is quite differ-
ent in this new model (the species with the greatest ci /mi ratio) than in the
one studied in Chapter 3 (where species are ranked entirely independently of
their ci /mi ratios), and a little algebra shows that any “inferior competitor” in
this new model (say c2 /m2 < c1 /m1 ) would also approach a zero equilibrium
value in the prior model.
In the boundary case, mentioned as unlikely above, of a “tie” for highest
ci /mi ratio, there are an infinite number of [non-isolated] equilibria such that
the “tied” species’ proportions sum to their common mi /ci ratio.

6.2.2 Predator-prey systems

The Lotka-Volterra system
y ′ = y(λ − bz) (6.10)
z ′ = z(−µ + cy)
was our first example of a model for a predator-prey system, formulated under
the assumptions that the presence of the z-species (predators) reduces the
growth rate of the y-species (prey), that presence of the prey increases the
growth rate of the predators, and that by themselves the prey species would
grow exponentially while the predator species would die out exponentially.
If we assume instead that the prey species obeys a logistic law, we replace
the Lotka-Volterra system (6.10) by
y ′ = y(λ − ay − bz), (6.11)
z ′ = z(−µ + cy).
The system (6.11) obviously has an equilibrium at (0,0). There is a second
equilibrium with z = 0 and ay = λ, or y = λa (prey but no predators).
A coexistence equilibrium (y∞ , z∞ ), in the interior of the first quadrant, is
found by solving the pair of equations
ay + bz = λ,
cz = µ,
obtaining y∞ = µc , z∞ = cλ−aµ
c . This equilibrium is relevant only if z∞ > 0,
or aµ − cλ < 0. In this case the community matrix at (y∞ , z∞ ) is
 
−ay∞ −by∞
,
cz∞ 0
338 Dynamical Systems for Biological Modeling: An Introduction

calculated much as in Example 4, Section 5.4. As the conditions for stability of

this equilibrium, namely −ay∞ < 0 and bcy∞ z∞ > 0, are satisfied automati-
cally, the equilibrium (y∞ , z∞ ) is asymptotically stable when it exists. We can
show similarly that if aµ − cλ < 0 the equilibrium ( λa ,0) is unstable, while if
aµ − cλ > 0 this equilibrium is asymptotically stable. The equilibrium (0,0) is
always unstable. Thus the model (6.11) always has exactly one asymptotically
stable equilibrium, which every orbit approaches. If aµ − cλ < 0 (λ/µ > a/c)
this equilibrium is ( λa ,0), corresponding to extinction of the predators. In nei-
ther case can there be a periodic orbit, and thus the model (6.11) is less
“realistic” than the original Lotka-Volterra model. The self-limiting property
of the prey dynamics prevents sustained oscillations, and as a increases can
even prevent damped ones.

Example 3.
Determine the behavior of a predator-prey system (for example, moths and
birds) modeled by the system
y ′ = y(180 − y − z), z ′ = z(−500 + 10y).

Solution: An interior equilibrium is a solution of the system

y + z = 180, 10y = 500.
As this system has the solution (50,130) in the first quadrant, there is an
equilibrium (50,130) which (from the above discussion) all orbits approach.
The two species co-exist, with the prey at less than 1/3 its natural carrying
capacity of 180. 

In the model (6.11) the term −byz in the equation for y ′ represents the
rate at which predators consume prey. Thus it is assumed that each predator’s
consumption is proportional to the prey population size. Biologically, it is
more plausible to assume that the rate of prey consumption per predator
increases with prey population size but is bounded as the prey population
becomes unbounded, that is, that there is a maximum rate of consumption per
predator no matter how plentiful the food supply. Beyond a certain point, the
prey population no longer limits the resources of the predators. For example,
we may assume that the rate of consumption of prey per predator has the
qy
form y+A where q and A are positive constants.3 Instead of the term cyz in
yz
the equation for z ′ , we incorporate a term proportional to z+A , representing
the conversion of food (prey) into predator biomass. This leads us to a model
of the form
 y ayz
y ′ = ry 1 − − , (6.12)
K y+A
 
y J
z ′ = sz − .
y+A J +A
3 See Section 2.1.3 for an interpretation of this Verhulst-type expression, including the

significance of A.
 Topics in Modeling Systems of Populations 339

Photo courtesy Florida Keys National Marine Sanctuary

FIGURE 6.9: A barracuda eats another fish off the Florida keys.

Here, we have also changed the names of some of the parameters in (6.11),
r sJ qyz
replacing λ by r, a by K , and µ by J+A . The term y+A (replacing byz)
in the first equation of (6.12) is called the predator functional response, and
syz
the term y+A (replacing cyz) in the second equation of (6.12) is called the
predator numerical response; the constant qs is the conversion efficiency of prey
into predators. The model (6.12) assumes that the prey population would obey
a logistic law in the absence of predators, and that the predator population
would die out exponentially in the absence of prey.
Here we have rewritten the natural decay rate of the predator population
in terms of J, which we can see from (6.12) is the minimum prey population
required to sustain the predator population (z ′ ≥ 0). As J decreases, so does
the rate at which the predator population would die out in the absence of the
prey. In the following two examples, we shall see that the parameter J (or,
equivalently, µ) is capable of changing the nature of the system’s behavior.

Example 4.
Determine the qualitative behavior of a predator-prey system (imagine this
time large and small fish in a pond, cf. Figure 6.9) modeled by the differential
equations
 y yz
y′ = y 1 − − ,
30 y + 10
 
y 3
z′ = z − .
y + 10 5
340 Dynamical Systems for Biological Modeling: An Introduction

Solution: Equilibria are solutions of the pair of equations

 
y z
y 1− − = 0, (6.13)
30 y + 10
 
y 3
z − = 0.
y + 10 5

One equilibrium is given by y = 0, z = 0. If z = 0, (6.13) implies

 y
y 1− = 0,
30
and thus another equilibrium is given by y = 30, z = 0. An equilibrium with
y and z both positive satisfies
y z y 3
1− − = 0, = .
30 y + 10 y + 10 5
The second of these equations is 5y = 3y + 30, or y = 15, and substitution into
z
the first equation gives 1 − 21 − 25 = 0, or z = 12.5. Thus a third equilibrium
is given by y = 15, z = 12.5.
We now use the equilibrium stability theorem of Section 5.4 with
 
 y yz y 3
F (y, z) = y 1 − − , G(y, z) = z − .
30 z + 10 y + 10 5
Then the partial derivatives are
y 10z y
Fy (y, z) = 1− − , Fz (y, z) = −
15 (y + 10)2 y + 10
10z y 3
Gy (y, z) = 2
, Gz (y, z) = − .
(y + 10) y + 10 5
At the equilibrium (0,0), the community matrix is
 
1 0
,
0 − 53

and thus the equilibrium is unstable. At the equilibrium (30,0), the community
matrix is
−1 − 43
 
3 ,
0 20
3
and since its determinant is − 20 < 0 this equilibrium is also unstable. At the
equilibrium (15,12.5), the community matrix is
 1
− 5 − 53

1 ,
5 0

and since this has trace − 51 < 0 and determinant 3

25 > 0, this equilibrium is
 Topics in Modeling Systems of Populations 341

asymptotically stable. It is possible to show that every orbit with y(0) > 0
and z(0) > 0 — not just those which start close to (15,12.5) — approaches
this equilibrium. Thus predator and prey coexist here, with prey at half their
natural carrying capacity. 

Example 5.
Suppose now that the environment of the pond in Example 4 improves in such
a way that the predator fish tend to live longer (or die off more slowly), so
that the natural per capita mortality rate drops from 53 (in per time units)
to 13 . Determine the qualitative behavior of this new predator-prey system,
modeled by the differential equations
 y yz
y′ = y 1 − − ,
30 y + 10
 
y 1
z′ = z − .
y + 10 3

Solution: Equilibria are solutions of the pair of equations

 
y z
y 1− − = 0,
30 y + 10
 
y 1
z − = 0.
y + 10 3
As in Example 4, there is an equilibrium at (0,0) and a second, predator-free
(z = 0) equilibrium with y = 30. An equilibrium with y and z both positive
satisfies
y z y 1
1− − = 0, = .
30 y + 10 y + 10 3
The second of these equations is 3y = y + 10, or y = 5, and substitution into
5 z
the first equation gives 1 − 30 − 15 = 0, or z = 12.5. Thus a third equilibrium,
representing coexistence, is given by y = 5, z = 12.5.
The community matrix is the same as in Example 4, except that 35 is
replaced by 13 in G(y, z). Thus the community matrix at (0,0) is
 
1 0
,
0 − 31
and this equilibrium is unstable. The community matrix at (30,0) is
−1 − 34
 
5 ,
0 12
5
and since this has determinant − 12 < 0, this equilibrium is unstable. The
community matrix at (5,12.5) is
1
− 31

9 .
5
9 0
342 Dynamical Systems for Biological Modeling: An Introduction

Since this matrix has positive trace, the equilibrium (5,12.5) is also unstable,
and the system has no asymptotically stable equilibrium. In order to show that
all orbits in the first quadrant are bounded, we apply the technique introduced
in Example 5, Section 5.4, adding the two equations of the model to obtain
 y z
(y + z)′ = y 1 − − .
30 3
Thus y + z is decreasing except in the bounded region defined by z3 <
y


y 1 − 30 . In order for an orbit to be unbounded, y + z must be unbounded,
and, as in Example 5, Section 5.4, this is impossible since y + z is decreasing
whenever y + z is large. Thus all orbits in the first quadrant are bounded,
and the Poincaré-Bendixson theorem may be applied to show that there must
be a limit cycle (with the equilibrium (5,12.5) in its interior) to which every
orbit tends. Thus the two species co-exist, but their population sizes fluctuate
periodically. Some orbits are shown in Figure 6.10. 
36
34
32
30
28
26
24
22
20
z18

16
14
12
10
8
6
4
2

5 10 15 20 25
y

FIGURE 6.10: Some orbits of the system in Example 5.

Examples 4 and 5 show that for a model of the form (6.12) there may be
periodic orbits, periodic orbit] or every orbit may approach an equilibrium.
Which behavior occurs depends on the values of the parameters in the model
rather than on the form of the model. In fact, there is a class of models consid-
erably more general than (6.12) exhibiting the same two possible behaviors.
We shall now consider models of the general form
y ′ = yf (y) − yzφ(y), (6.14)
z ′ = z[syφ(y) − c].
 Topics in Modeling Systems of Populations 343

In this model, as before, y(t) is the prey population size and z(t) is the predator
population size. Here the term yf (y) represents the prey population growth
rate in the absence of predators. We assume that the per capita growth rate
of the prey decreases as prey population size increases, and that there is a
prey carrying capacity K, so that

f ′ (y) < 0 [y ≥ 0], f (K) = 0. (6.15)

The term yzφ(y) represents the predator functional response, with consump-
tion of yφ(y) prey per predator in unit time. We assume that consumption of
prey is positive, and that the prey consumption per predator yφ(y) increases
with prey population size, but that the fraction φ(y) of prey population con-
sumed per predator decreases:

φ(y) > 0, [yφ(y)] > 0, φ′ (y) ≤ 0. (6.16)

The term syzφ(y) is the predator numerical response, and cz is predator mor-
tality. Because of (6.16), the predator per capita growth rate syφ(y) − c in-
creases with prey population size and is positive for prey population size above
some minimum J, sJφ(J) = c. Normally, it is assumed that the minimum prey
population size for predator survival J is less than the prey carrying capacity
K, as otherwise it is clear that the predator population cannot survive.
Let us try to analyze the behavior of solutions of a system (6.14) under
the assumptions (6.15) and (6.16). The equilibria of (6.14) are the solutions
of

y[f (y) − zφ(y)] = 0, (6.17)

syzφ(y) = cz. (6.18)

If y = 0, then (6.17) is satisfied. In order to satisfy (6.18) as well we must

take z = 0. Thus one equilibrium is (0,0). If z = 0, then (6.18) is satisfied
and (6.17) reduces to yf (y) = 0, which implies that either y = 0 or y = K.
Thus (K,0) is a second equilibrium, as we should expect. If y and z are both
positive, (6.17) becomes f (y) − zφ(y) = 0, and (6.18) becomes syφ(y) = c,
f (J)
which implies y = J. If y = J, (6.17) implies z = φ(J) . Since φ(J) > 0, then
z > 0 provided f (J) > 0, i.e., provided J < K. Thus if J < K, there is a
third equilibrium (y∞ ,z∞ ) with y∞ > 0, z∞ > 0 given by

f (J)
y∞ = J, z∞ = .
φ(J)

In order to calculate the community matrix at each equilibrium of (6.14),

we take F (y, z) = yf (y) − yzφ(y), G(y, z) = syzφ(y) − cz. Then

Fy (y, z) = yf ′ (y) + f (y) − z[φ(y) + yφ′ (y)], Fz (y, z) = −yφ(y)

Gy (y, z) = sz[φ(y) + yφ′ (y)], Gz (y, z) = syφ(y) − c.
344 Dynamical Systems for Biological Modeling: An Introduction

Thus the community matrix at the equilibrium (0,0) is

 
f (0) 0
.
0 −c

Since f (0) > 0 and −c < 0, this equilibrium is unstable (by applying the
equilibrium stability theorem of the previous section).
The community matrix at the equilibrium (K, 0) is
Kf ′ (K)
 
−Kφ(K)
.
0 sKφ(K) − c

Because f ′ (K) < 0 and sKφ(K)−c > 0 (if K > J), we see that the equilibrium
(K, 0) is unstable if K > J. If instead K < J, then sKφ(K) − c < 0, and in
this case the equilibrium (K, 0) is asymptotically stable.
If K > J, there is a third equilibrium (y∞ ,z∞ ) with community matrix
cc[yf (y)]′y∞ − z∞ [yφ(y)]′y∞ −y∞ φ(y∞ )
 
.
sz∞ [yφ(y)]′y∞ 0
The conditions for asymptotic stability are

[yf (y)]′y∞ − z∞ [yφ(y)]′y∞ < 0 (6.19)

and
y∞ φ(y∞ ) sz∞ [yφ(y)]′y∞ > 0. (6.20)
Because φ(y∞ ) > 0, and [yφ(y)]′y∞
> 0 from (6.16), the condition (6.20) is
satisfied, and the equilibrium (y∞ , z∞ ) is asymptotically stable if and only if
(6.19) is satisfied. The condition (6.19) is equivalent to

y∞ f ′ (y∞ ) + f (y∞ ) − y∞ z∞ φ′ (y∞ ) − z∞ φ(y∞ ) < 0,

f (y∞ )
and since z∞ = φ(y∞ ) , this is the same (after multiplying by φ(y∞ )/y∞ ) as

f ′ (y∞ )φ(y∞ ) − f (y∞ )φ′ (y∞ ) < 0. (6.21)

f (y)
The equation of the prey nullcline is z = φ(y) , and differentiation shows that
the slope of the prey nullcline is
dz φ(y)f ′ (y) − f (y)φ′ (y)
= . (6.22)
dy [φ(y)]2
Comparing (6.22) and (6.21), we obtain a simple geometric characterization
of stability of the equilibrium (y∞ , z∞ ):

The equilibrium (y∞ , z∞ ) of the system (6.14) with y∞ > 0, z∞ > 0 under
the hypotheses (6.15) and (6.16) is asymptotically stable if and only if the
prey nullcline has negative slope at the equilibrium.
 Topics in Modeling Systems of Populations 345

Because of the hypotheses (6.15) and (6.16), the slope of the prey nullcline
′
at its y-intercept (K, 0) is fφ(K)
(K)
< 0. The slope of the prey nullcline at y = 0
may be either positive or negative, depending on the functions f (y) and φ(y)
y q


and the parameter values. For example, if f (y) = r 1 − K and φ(y) = y+A ,
the prey nullcline has positive slope at y = 0 if A < K. Thus there are cases
in which the prey nullcline increases to a maximum and then decreases to zero
as y increases. In such cases the equilibrium (y∞ , z∞ ) is unstable if it is on
the increasing portion of the prey nullcline, and asymptotically stable if it is
on the decreasing portion of the prey nullcline.
Increasing the carrying capacity of the prey species amounts to increasing
the food supply for the predators. Geometrically, this would have the effect
of moving the prey nullcline upward and to the right. This could move the
equilibrium from the decreasing portion of the prey nullcline to the increasing
portion of the prey nullcline and thus change it from an asymptotically stable
equilibrium to an unstable equilibrium. This possibility, that increasing the
food supply for the predators could destabilize the population system, has
been called the “paradox of enrichment.”
If the equilibrium (y∞ , z∞ ) is unstable, the system (6.14) has no asymp-
totically stable equilibrium. It is possible to show that every solution of the
system (6.14) in the first quadrant is bounded as t → ∞ (see Exercise 13
below). Then by the Poincaré-Bendixson theorem (Section 5.4), every orbit
must approach a limit cycle, and the prey and predator populations oscillate
as in Example 5. It is possible that an orbit could come very close to one of
the axes, where a small perturbing force, say an environmental change, could
wipe out one of the populations and lead to the collapse of the population
system. Thus the oscillations caused by enrichment could turn out to be very
harmful to the predator species.

Example 6.
Describe the long-term behavior of a predator-prey system modeled by the
system
y ′ = y(10 − y) − yz, z ′ = z(y − 5).

Solution: This system is of the form (6.14) with f (y) = 10−y, so that K = 10,
and φ(y) = y, so that J = 5. The hypotheses (6.15) and (6.16) are satisfied.
The prey nullcline is the line z = 10 − y with negative slope, and thus the
coexistence equilibrium given by y = 5, z = 10 − y, or z = 5, is asymptotically
stable. Every orbit approaches this equilibrium, so the species will coexist. 

6.2.3 Symbiosis
There are situations in which the interaction of two species is mutually
beneficial, for example, plant-pollinator systems (Figure 6.11 gives another
example). Such an interaction is called mutualistic or symbiotic. The interac-
tion may be facultative, meaning that the two species could survive separately,
346 Dynamical Systems for Biological Modeling: An Introduction

Photo by Jan Derk

FIGURE 6.11: Two goby fish and a shrimp in a symbiotic relationship: the
shrimp maintains a burrow in the sand on the sea floor which also serves as a
safe refuge for the goby and their eggs, while the goby alert the shrimp (which
has poor eyesight) if danger approaches.

or obligatory, meaning that each species would become extinct without the as-
sistance of the other.
If we model a symbiotic system by a pair of differential equations with
linear per capita growth rates

y′ = y(λ − ay + bz), (6.23)

z′ = z(µ + cy − dz),

the mutualism of the interaction is modeled by the positive nature of the in-
teraction terms cy and bz. In a facultative interaction, the constants λ and µ
are positive, while in an obligatory relation the constants λ and µ are nega-
tive. In each type of interaction there are two possibilities, depending on the
relationship between the slope a/b of the y isocline and the slope c/d of the
z-isocline. If ad > bc, the mutualistic effects are smaller than the self-limiting
terms in the per capita growth rates, and the slope of the y-isocline is greater
than the slope of the z-isocline.
In both facultative and obligatory interactions, if ad < bc there is a region
of the phase plane in which solutions become unbounded, and this suggests
that either we must restrict models of this form by requiring ad > bc, or we
must consider models with nonlinear per capita growth rates.
For models with linear per capita growth rates and ad > bc, the only
asymptotically stable equilibrium in the facultative case is the intersection
(y∞ , z∞ ) of the lines ay − bz = λ, −cy + dz = µ with y∞ > 0, z∞ > 0, and
 Topics in Modeling Systems of Populations 347

every orbit tends to this equilibrium. To see this, we calculate the equilibrium
dλ + bµ cλ + aµ
y∞ = > 0, z∞ = > 0.
ad − bc ad − bc
The community matrix at (y∞ , z∞ ) is
   
λ − 2ay∞ + bz∞ by∞ −ay∞ by∞
=
cz∞ µ − 2dz∞ + cy∞ cz∞ −dz∞

and since this matrix has negative trace and positive determinant the equi-
librium (y∞ , z∞ ) is asymptotically stable. It is easy to verify that the other
equilibria (0, 0), (λ/a, 0), (0, µ/d) are unstable.
In the obligatory case with ad > bc, since λ < 0, µ < 0 the only equilib-
rium in the first quadrant is (0, 0), asymptotically stable since the community
matrix is  
λ 0
,
0 µ
and asymptotic stability follows from λ < 0, µ < 0. Thus the only asymptoti-
cally stable equilibrium is the origin, and every orbit tends to the origin. In the
obligatory case neither species survives. While our model may be acceptable
in the facultative case, it is clear that the possibility of obligatory mutualism
is not described by this model. If we consider the obligatory case with ad < bc,
there is an equilibrium (x∞ , y∞ ) with x∞ > 0, y∞ > 0, which may be shown
to be a saddle point whose stable separatrices separate the phase plane into
a region of mutual extinction and a region of unbounded growth. Such a sep-
aration is plausible biologically, but it would be necessary to alter the model
so as to rule out the possibility of unbounded growth in order to give a more
realistic model.

Exercises
In each of Exercises 1–4, determine the outcome of the competition modeled
by the given system.

1. y ′ = y(120 − 3y − z), z ′ = z(80 − y − z)

2. y ′ = y(75 − y − 4z), z ′ = z(60 − y − 3z)
3. y ′ = y(90 − 2y − z), z ′ = z(40 − y − z)
4. y ′ = y(80 − y − z), z ′ = z(80 − 2y − 3z)
In each of Exercises 5–10, determine the qualitative behavior of the predator-
prey system modeled by the given system.
 
y yz y
5. y ′ = y 1 − 20 , z ′ = 3z y+10 − 21


− y+10
348 Dynamical Systems for Biological Modeling: An Introduction
 
y 2yz 2y
6. y ′ = 3y 1 − 40 , z ′ = z y+15 − 56


− y+15
 
y yz y 1
7. y ′ = y 1 − z ′ = 3z


20 − y+10 , y+10 − 6
 
y 2yz 2y 1
8. y ′ = 3y 1 − z′ = z


40 − y+15 , y+15 − 2
 
y yz y 3
9. y ′ = y 1 − z ′ = 3z


20 − y+20 , y+10 − 4
 
y 2yz 2y 8
10. y ′ = 3y 1 − z′ = z


40 − y+15 , y+15 − 5

11. * Show that the coexistence equilibrium (y∞ , z∞ ) of the system

 
 y ayz ay aJ
y ′ = ry 1 − − , z ′ = sz −
K y+A y+A J +A

is unstable if K > A + 2J and asymptotically stable if J < K < A + 2J.

12. * Determine the behavior as t → ∞ of solutions of the system
   
′ y ′ z
y = ry 1 − , z = sz 1 −
K + az M + by

with a, b, K, M , r, s positive constants. [Warning: The behavior if ab < 1

is different from the behavior if ab > 1.]
13. * Show that every first-quadrant orbit of the system y ′ = yf (y)−yzφ(y),
z ′ = z[syφ(y) − c], under the assumptions f ′ (y) < 0 [y ≥ 0], f (K) = 0,
and φ(y) > 0, φ′ (y) ≤ 0, [yφ(y)]′ ≥ 0, is bounded, by showing that
the function sy + z is decreasing except in a bounded set, much as in
Example 5.
In Exercises 14 through 17, find all equilibria of the given mutualistic system
and determine their stability. Are there any unbounded orbits?
14. y ′ = y(−50 + z − y), z ′ = z(−20 − z + 2y)
15. y ′ = y(−50 + z − y), z ′ = z(−20 − z + 2y)
16. y ′ = y(10 − y + z/(1 + y)), z ′ = z(y/(1 + z) − 1)
17. y ′ = y(1 − y/20), z ′ = z(1 − z/10) · ((2y + 1)/(1 + y))
18. For a system of the form (6.23) with λ > 0, µ < 0, show that both species
survive, even though the z-species would go extinct in the absence of the
y-species.
 Topics in Modeling Systems of Populations 349

19. * Models for cell growth often assume that a cell can be in various
states, with switching (transfer) between one state an another.4 Assume
that there are two states, in only one of which there is proliferation
(cell division). If P (t) and Q(t) represent the concentrations of cells in
the two states, the following equations can be assumed, with all Greek
letters representing positive constants

P ′ = (γ − δP )P − αP + βQ, Q′ = αP − (λ + β)Q. (6.24)

(a) Is P or Q the proliferating state? Why?

(b) Find all biologically meaningful steady states of (6.24) and determine
their stability. You should distinguish the cases (α − γ)/α < β/(λ + β)
and (α − γ)/α > β/(λ + β).
For further reading about this type of model, see Eisen and Schiller [20].
20. * This problem5 is concerned with a theory of liver regeneration due to
Bard [2, 3]. Normally, the rate of cell division in the liver is very low, but
if up to two thirds of the liver of a rat is removed, then the liver grows
back to its original size in about a week. Bard discusses two theories. One
theory, based on the assumed existence of a growth stimulator, predicts
that the liver volume V will overshoot its normal value before finally
settling down to a steady state. Such an overshoot has not been observed.
Here we will show something of how an alternative inhibitor model can
account for the facts. Bard assumes that liver cells are prevented from
dividing by an inhibitor of short half-life. The inhibitor is synthesized
by the liver at a rate proportional to its size and is secreted into the
blood, where its concentration is the same as in the liver. Let V (t) be
the volume of the liver and S(t) the concentration of the inhibitor. Bard
postulates the equations
pV
V ′ = V [f (S) − r], S′ = − qS, (6.25)
W +V
with W (blood volume), r, p, q constants. (a) Why should the function
f be assumed to have a negative derivative?
(b) Show that the system (6.25) has a unique positive equilibrium, and
that the linearization of (6.25) at this equilibrium has the form

u′ = −γv, v ′ = αu − qv,

with positive constants γ, α.

4 This problem is taken from Lee A. Segel, Modeling dynamic phenomena in molecular

and cellular biology, Cambridge University Press, Cambridge (1984), p. 267.

5 Taken from Lee A. Segel, Modeling dynamic phenomena in molecular and cellular

biology, Cambridge University Press, Cambridge (1984), p. 269.

350 Dynamical Systems for Biological Modeling: An Introduction
6
21. * Consider the system

A′ = αA − a1 A3 − a2 aB 2 , B ′ = βB − b1 BA2 − b2 B 3 .

All coefficients are assumed to be positive., and we are interested only

in non-negative values of A(t), B(t).
(a) Discuss the suitability of this system as a model for the interaction
of two bacteria populations.
(b) Show that there are four possible equilibria,
(i) I = (0, 0),
p
(ii) II = (0, β/b2 ),
(iii) III = (α/a1 , 0),
p p
(iv) IV = ( (a2 β − αb2 )/(a2 b1 − a1 b2 ), (αb1 − βa1 )/(a2 b1 − a1 b2 )).
(c) Examine the stability of the equilibria and show that there are four
possibilities, namely
(i) II and III are unstable and if IV exists it is stable.
(ii) II is stable, III is unstable, and IV cannot exist.
(iii) II is unstable, III is stable, and IV cannot exist.
(iv) II and III are stable and IV exists but is unstable.
22. Consider a simple age-structured model in which individuals are classi-
fied as juveniles J or adults A (discrete age structure applies naturally to
species with distinct life stages, or more arbitrarily to species with con-
tinuous development for whom a maturation age is defined). We assume
that juveniles are born at a rate dependent on the number of adults but
limited (logistically) by the resources used by juveniles and adults; that
juveniles mature at a certain rate; and that juvenile and adult mortality
rates may differ. This leads to the model
 
ǫJ + A
J ′ = bA 1 − − (µ + γ)J, A′ = γJ − dA,
K
for constant rates b, γ, µ, d, carrying capacity K, and a dimensionless
constant ǫ which represents how many resources a juvenile uses relative
to an adult (we might then assume 0 ≤ ǫ ≤ 1).
Perform a complete qualitative analysis for this model, identifying any
equilibria and the conditions for their stability. As with many population
models, the long-term survival of the population can be determined by a
demographic reproductive number Rd similar to the basic reproductive
number used in epidemiological models. Derive and interpret biologically
an expression for Rd .

6 This problem is taken from Lee A. Segel, Modeling dynamic phenomena in molecular

and cellular biology, Cambridge University Press, Cambridge (1984), p. 270.

 Topics in Modeling Systems of Populations 351

6.3 Numerical approximation to solutions of systems

Just as for single differential equations, the numerical approximation of
solutions of a system of differential equations is often a useful approach. The
qualitative methods on which we have concentrated in this chapter are di-
rected at obtaining information about the behavior of solutions for large t;
numerical approximation methods may be the only way to obtain information
about the behavior of solutions on a finite interval. There are many computer
programs available commercially, either as part of a computer algebra system
or as stand-alone programs. These programs can generate both numerical data
and a graphic portrayal of solutions either by showing both y and z as func-
tions of t or by showing orbits in the phase plane. We will not discuss the
implementation of approximation methods in detail but will confine ourselves
to brief descriptions of the analogues of the methods described in Section 4.5
for first-order differential equations.
Our goal is to estimate the values y(T ) and z(T ) for t = T of the solution
[y(t), z(t)] of an initial value problem

y ′ = F (y, z), y(0) = y0 , (6.26)

′
z = G(y, z), z(0) = z0 .

We divide the interval 0 ≤ t ≤ T into N subintervals of length h by defining

t1 = h, t2 = 2h, . . ., tN = N h. The number N of subintervals and the length
h of each subinterval are related by T = N h. Just as for first-order equations,
the fundamental idea is to replace the system of differential equations (6.26) by
a system of difference equations which can be solved recursively. In the Euler
method we achieve this by replacing y ′ and z ′ by the respective difference
quotients
y(t + h) − y(t) z(t + h) − z(t)
and
h h
to obtain the approximation scheme

yk+1 = yk + hF (yk , zk ), (6.27)

zk+1 = zk + hG(yk , zk ), [k = 0, . . . , N − 1].

As noted in Section 4.5, it can be shown via Taylor approximations that the
truncation error of the Euler method at each step is no greater than a constant
multiple of h2 , and that the accumulated truncation error is no more than a
constant multiple of h.
The modified Euler method is given by

yk+2 = yk + 2h F (yk+1 , zk+1 ), (6.28)

zk+2 = zk + 2h G(yk+1 , zk+1 ), [k = 0, . . . , N − 1].
352 Dynamical Systems for Biological Modeling: An Introduction

As in the first-order case, it requires the use of some starting method to give
y1 and z1 . It can be shown that the truncation error of the modified Euler
method in each step is no more than a constant multiple of h3 , and that the
accumulated truncation error is no more than a constant multiple of h2 . This
represents a considerable improvement over the Euler method.
The Runge-Kutta method is still more accurate, having a truncation error
in each step of no more than a constant multiple of h5 and an accumulated
truncation error of no more than a constant multiple of h4 . It is given by

yk+1 = yk + h6 [K1 + 2K2 + 2K3 + K4 ], (6.29)

h
zk+1 = zk + 6 [L1 + 2L2 + 2L3 + L4 ],

where

K1 = F (yk , zk ), L1 = G(yk , zk ),

hK1 hL1 hK1 hL1

K2 = F (yk + , zk + ), L2 = G(yk + , zk + ),
2 2 2 2
hK2 hL2 hK2 hL2
K3 = F (yk + , zk + ), L3 = G(yk + , zk + ),
2 2 2 2
K4 = F (yk + hK3 , zk + hL3 ), L4 = G(yk + hK3 , zk + hL3 ).

The warnings in Section 4.5 about the proper choice of step size h are also
relevant here. Some commercial software uses a variable step size to minimize
problems.

6.3.1 Example: A two-sex model

The population models studied up to this point have ignored the fact that
humans and most other animal species reproduce sexually, requiring both
males and females. Ignoring gender and treating the population as a single
group is a modeling assumption made to simplify aspects of the biological
system which do not directly address the research question under investiga-
tion. However, there are certainly many questions which do require separating
a population into males and females. We shall now develop a reasonably simple
two-sex model, and quickly see that even simple systems may be too complex
to complete qualitative analysis, in which case numerical analysis can provide
helpful insights.
Let us denote the numbers of females and males in a population by F (t)
and M (t), respectively. If we assume (i) that the total birth rate is a function
r(F, M ) of F and M , (ii) that the proportions of those births which are female
or male, denoted f and m (so that f + m = 1), are independent of population
size, and (iii) that per capita death rates µf and µm may differ by gender but
 Topics in Modeling Systems of Populations 353

are also independent of population density, then we can write the system

F ′ = f r(F, M ) − µf F, M ′ = m r(F, M ) − µm M. (6.30)

One idea for describing the reproduction rate is to consider it a contact

rate between females and males, and to base it on the law of mass action,
r(F, M ) = cF M for some contact rate c. However, such a model turns out (see
Exercise 5) to have solutions (for any parameter values) which become infinite
in finite time. The problem is that the growth rates for such a model are not
only unconstrained by resource limitations, but increase too quickly even for
small populations: this choice of r makes growth increase quadratically, rather
than linearly, in population size (recall from Chapter 3 that growth rates linear
in population size, leading to exponential population growth over time, do a
good job of describing observed initial population growth). Therefore, a more
realistic reproduction rate should increase roughly linearly in N = F + M .
A ratio-dependent approach can provide a roughly linear increase in repro-
duction. If we suppose females’ maximum reproduction rate is cf while males’
is cm , the function
(cf F )(cm M )
r(F, M ) =
(cf F ) + (cm M )
gives a rate which is bounded above by both cf F and cm M : that is,
cm M cf F
r = (cf F ) < (cf F ) and r = (cm M ) < (cm M ).
cf F + cm M cf F + cm M
This choice of r does still allow unbounded growth, like a simple linear ODE,
but the population does not become infinite in finite time. Substituting this
function into (6.30) and solving the equilibrium conditions yields only the ex-
tinction equilibrium (0,0). However, the usual qualitative analysis techniques
fail to provide conditions for its stability, because the Jacobian matrix is in-
determinate at (0,0):

cf f (1 − x)2 − µf cm f x2 cf F ∗
 
J= 2 2 , where x = ,
cf m(1 − x) cm mx − µm cf F + cm M ∗
∗

but the quantity x can take on any value from 0 to 1 as (F, M ) → (0, 0),
depending upon the path taken to approach the origin (as students of multi-
variable calculus will recall). [The equilibrium stability theorem of Chapter 5
does not apply here since the growth rates are not differentiable at (0,0).]
Therefore, in order to explore what determines the long-term behavior of
this system, we turn to numerical analysis. Computer systems such as Math-
ematica, Maple, and MATLAB can perform symbolic calculations in some
cases (such as computing the Jacobian matrix above), but here we illustrate
their ability to compute and graph numerical solutions (approximations) for
systems. Since exploration is inherently ad hoc rather than formulaic, we here
present a sample of commands and results representative of the problem solv-
ing process. Using Mathematica syntax, we may first define the equations and
354 Dynamical Systems for Biological Modeling: An Introduction

their numerical solution, as well as a specialized command to plot them (both

variables F (t) and M (t) on the same graph, with F (t) in thick light gray and
M (t) dashed to distinguish them) as follows, with input commands given in
bold):
In[1]:= twosexeqns :=
{F’[t] == f*cf*cm*M[t]*F[t]/(cm*M[t]+cf*F[t]) - muf*F[t],
M’[t] == m*cm*cf*M[t]*F[t]/(cm*M[t]+cf*F[t]) - mum*M[t],
F[0] == F0, M[0] == M0}
In[2]:= twosexsoln := NDSolve[twosexeqns, {F[t],M[t]}, {t,0,tf}][[1]]
In[3]:= PopPlot[x ] := Plot[{F[t]/.x, M[t]/.x}, {t,0,10}, AxesLabel→
{”t”, ”P(t)”}, PlotStyle→{{Thickness[0.02], LightGray},
{Thickness[0.01], Dashing[0.1], Black}}]
We must now set parameter values (here arbitrarily setting all parameters to
1 for simplicity, except that since f + m = 1 by definition, we set them both
equal to 1/2):
In[4]:= cf=1; cm=1; f=1/2; m=1/2; muf=1; mum=1; K=1;
...as well as the initial conditions and final time:
In[5]:= F0 = 1; M0 = 1; tf = 100;
We now request a solution using the given values, and plot it:
In[6]:= set1 = twosexsoln
Out[6]= {F[t] → InterpolatingFunction[{{0.,100.}},<>][t],
M[t] → InterpolatingFunction[{{0.,100.}},<>][t]}
In[7]:= plot1 = PopPlot[set1]
PHtL

0.6

0.5

0.4

0.3

0.2

0.1

t
Out[7]= 2 4 6 8 10

It should not be surprising that F (t) and M (t) are equal (their graphs su-
perimposed) since we have given symmetric parameter values. However, we
see that for the given values, both populations die out quickly. Perhaps larger
initial population sizes would make a difference?
 Topics in Modeling Systems of Populations 355

In[8]:= F0=10; M0=10;

In[9]:= set2=twosexsoln;
In[10]:= plot2 = PopPlot[set2]
PHtL

t
Out[10]= 2 4 6 8 10

This graph is identical to the previous one except for scale (the numbers are 10
times as great). So initial conditions may not matter. A partial phase portrait
may make this clearer, showing that trajectories from the two different sets
of initial conditions lead to the same equilibrium: the origin.
In[11]:= pp1=ParametricPlot[{F[t]/.set1, M[t]/.set1}, {t,0,tf},
AxesLabel→{”F”, ”M”}, PlotStyle→Thickness[0.01], PlotRange→All];
In[12]:= pp2=ParametricPlot[{F[t]/.set2, M[t]/.set2}, {t,0,tf},
AxesLabel→{”F”, ”M”}, PlotStyle→Dashing[0.05]];
In[13]:= Show[pp1,pp2]
M
10

Out[13]= 2 4 6 8 10
F

(The trajectories are superimposed since they both fall on the line F = M .)
Now what if we increase the reproduction rates, say by an order of mag-
nitude?
In[14]:= cf=10; cm=10;
356 Dynamical Systems for Biological Modeling: An Introduction

In[15]:= set3=twosexsoln;
In[16]:= plot3 = PopPlot[set3]
PHtL
6 ´ 107

5 ´ 107

4 ´ 107

3 ´ 107

2 ´ 107

1 ´ 107

t
Out[16]= 2 4 6 8 10

The outcome has now changed qualitatively: the population grows exponen-
tially instead of dying out exponentially. What if we now reduce one of the
reproduction rates, creating an asymmetry?
In[17]:= cf=3;
In[18]:= set4=twosexsoln;
In[19]:= plot4 = PopPlot[set4]
PHtL

450

400

350

300

250

200

150

t
Out[19]= 2 4 6 8 10

The overall growth rate is certainly slowed, but the population still grows,
and both populations remain matched. What if we force an asymmetry in the
population values by changing one of the initial conditions?
In[20]:= F0=50;
In[21]:= set5=twosexsoln;
In[22]:= plot5 = PopPlot[set5]
 Topics in Modeling Systems of Populations 357
PHtL

250

200

150

100

t
Out[22]= 2 4 6 8 10

Despite the initial disparity in sizes, the two sexes’ populations quickly return
to even before growing, because all new births are evenly divided between
females and males.
Here we have used arbitrary values rather than realistic estimates because
the goal is to uncover qualitative insights about the model’s behavior. The
original question of the precise threshold condition between the two possible
behaviors (growth and decay) has not yet been answered precisely, although a
few data points have been established. A more methodical approach would fix
all but one or two key parameters (say cf and cm ), vary those independently
over a regular grid (say each taking on values between 1 and 10), then mark
on the grid which combinations led to extinction and which to growth, and
finally abstract from the results an overall pattern describing the influence of
these two parameters. We invite the reader to try such an experiment.
As one final note in this example, we observe that the unbounded growth of
the model above can be eliminated by introducing a logistic term in each equa-

tion, that is, by multiplying each sex’s growth rate by a factor of 1 − F +M K
for some carrying capacity K. While the resulting model is more complicated
to analyze (and still has the problem of not being linearizable at the origin),
even a partial analysis (see Exercise 6) sheds some light on the threshold
between survival and extinction for both that model and the one explored
above.

Exercises
Use a computer and available software to approximate the solution of each of
the following initial value problems on the given interval.

1. The two-species [moth] competition model y ′ = y(100 − 4y − z), z ′ =

z(60−y−2z) of Example 1, Section 6.2, with symmetric initial conditions
y(0) = 10, z(0) = 10, over the time range 0 ≤ t ≤ 10.
y yz y
2. The [pond] predator-prey system y ′ = y(1 − 30 )− y+10 , z ′ = z( y+10 − 35 )
358 Dynamical Systems for Biological Modeling: An Introduction

of Example 4, Section 6.2, with initial conditions y(0) = 40, z(0) = 10,
over the time range 0 ≤ t ≤ 10.
y yz y
3. The [pond] predator-prey system y ′ = y(1 − 30 )− y+10 , z ′ = z( y+10 − 13 )
of Example 5, Section 6.2, with initial conditions y(0) = 40, z(0) = 10,
over the time range 0 ≤ t ≤ 10.
1 1
4. The SIR disease model S ′ = − 200 SI, I ′ = 200 SI − 16 I of Section 6.1,
with initial conditions S(0) = 1190, I(0) = 10, over the time range
0 ≤ t ≤ 20.

In each case, compare the numerical approximation with the qualitative in-
sights obtained in Sections 6.1 and 6.2 where these systems are studied as
examples.

5. Analyze the [discarded] mass-action two-sex model described in the text,

substituting r(F, M ) = cF M into (6.30), as follows:
(a) Identify all equilibria for the system.
(b) Analyze each equilibrium’s local stability using the Jacobian ma-
trix.
(c) Draw a phase portrait, using a computer and available software as
necessary. Include some sample solutions, showing that some lead
to extinction while others lead to unbounded growth.
6. Analyze the logistic two-sex model
 
(cf F )(cm M ) F +M
F′ = f 1− − µf F,
(cf F ) + (cm M ) K
 
(cf F )(cm M ) F +M
M′ = m 1− − µm M
(cf F ) + (cm M ) K

(with f + m = 1), as follows:

(a) Identify analytically all equilibria of the system. Rewrite all nonzero
expressions in terms of the dimensionless reproductive quantities
Rf = cf f /µf , Rm = cm m/µm and the ratio cf /cm .
(b) Since the Jacobian matrix is indeterminate at the origin, use the
condition that the nonzero equilibrium be positive to conjecture
stability conditions for the equilibria.
(c) Explore these stability conditions by graphing (with the aid of a
computer and appropriate software) numerical solutions for several
sets of parameter values that obey each condition.
Chapter 7
Systems with Sustained Oscillations
and Singularities

In this chapter we discuss some processes in which oscillations occur naturally

as well as some situations in which singularities and multiple time scales arise.
These lead to some new methods of analysis more advanced than those of
previous chapters and some new biological fields of study.

7.1 Oscillations in neural activity

Neurons are cells in the body which transmit information to the brain and
the body by amplifying an incoming stimulus (electrical charge input) and
transmitting it to neighboring neurons and then turning off to be ready for the
next stimulus. Neurons have fast and slow mechanisms to open ion channels
in response to electrical charges. The key quantities are the concentration of
sodium ions and potassium ions (both positively charged). A resting neuron
has an excess of potassium and a deficit of sodium, and a negative resting
potential (an excess of negative ions). Neurons use exchanges of sodium and
potassium ions across the cell membrane to amplify and transmit information.
There are voltage-gated channels for each kind of ion which open and close
in response to voltage differences and which are closed in a resting neuron.
When a burst of positive charge enters the cell, making the potential less
negative, the voltage gated sodium channels open. Since there is an excess
of sodium outside the cell, more sodium ions enter, increasing the potential
until it eventually becomes positive. Next, a slow mechanism acts to block
the voltage-gated sodium channels and another slow mechanism begins to
open voltage-gated potassium channels. Both of these diminish the buildup
of positive charge by blocking sodium ions from entering and by allowing
excess potassium ions to leave. When the potential decreases to or below the
resting potential, these slow mechanisms turn off, and then the process can
start over. If the electrical excitation reaches a sufficiently high level, called an
action potential, the neuron fires, transmitting the excitation to other neurons.

359
360 Dynamical Systems for Biological Modeling: An Introduction

7.1.1 The Fitzhugh-Nagumo equations

In order to describe a simple model for this neuron firing process, we denote
the potential by v, scaled so that v = 0 is the resting potential. We let v = a
be the potential above which the neuron fires and v = 1 the potential with
sodium channels open (0 < a < 1). A model of the form

v ′ = −v(v − a)(v − 1),

which has1 asymptotically stable equilibria at 0 and 1 and an unstable equilib-

rium at a, would explain part of the observed behavior. If the initial potential
is above a, the potential increases to 1, and if the initial potential is below a,
the potential decreases to 0. Thus the model allows the signal amplification
of the neuron, but stops at v = 1.
We must also build in a blocking mechanism, to lower the voltage by
switching the ion channels. Let w denote the strength of this blocking mech-
anism, with w = 0 (turned off) when v = 0. As v → 1 the mechanism gets
stronger but remains bounded, and we assume an equation of the form

w′ = ǫ(v − γw),

with a limiting value for w of γv if v is fixed. If v = 0, then w → 0, and if

v = 1, then w → γ1 (the maximum strength of the blocking mechanism). The
scaling parameter ǫ does not affect w’s equilibrium value but does influence
its rate of approach to equilibrium. A small value of ǫ indicates a slow-acting
mechanism (relative to the potential v).
Finally, to incorporate the effect of the blocking mechanism on v, we add
a term −w to the rate of change of v. The model we shall examine, known as
the Fitzhugh-Nagumo system,2 is the two-dimensional system

v ′ = −v(v − a)(v − 1) − w, (7.1)

′
w = ǫ(v − γw).

Alternatively, to develop this model in electrical terms, we may think of

the cell membrane as consisting of three components: a capacitor representing
the membrane capacity, a nonlinear current voltage device for the fast current,
and a resistor, inductor and battery in series for the recovery current. Using
Kirchhoff’s laws, we may write equations for this membrane circuit, namely
dV
Cm = −F (V ) − i,
dτ (7.2)
di
L = −Ri + V − V0 .
dτ
1 We leave the straightforward analysis of this solo equation as an exercise for the reader;

see Exercise 12.

2 R. Fitzhugh, Impulses and physiological states in theoretical models of nerve membrane,

Biophy. J. 1 (1961), 445–466; J. S. Nagumo, S. Arimoto, and S. Yoshizawa, An active pulse

transmission line simulating nerve axon, Proc. Inst. Radio Engineers 50 (1962), 2061–2071.
 Systems with Sustained Oscillations and Singularities 361

Here, i is the current through the resistor and inductance, V = Vi − Ve is

the membrane potential, and V0 is the potential gain across the battery. The
variable τ is used to represent time, because we will presently want to use t as
a dimensionless time variable. The first equation gives the fast current across
the membrane; the second gives the voltage drops across the inductor and
across the resistor and battery (completing the slow current series circuit).
The function F (V ) which describes the current through the voltage device
is assumed to have three zeros, with the smallest, V = V0 , and the largest,
V = V0 + V1 , being asymptotically stable equilibria of the differential equation
dV /dτ = −F (V ). We take R1 to be the “passive” resistance of the nonlinear
current voltage device, R1 = 1/F ′ (V0 ).
To simplify the model, we reformulate the system (7.2) in dimensionless
variables, by letting
V − V0 R1 i R1 F (V1 v) Lτ
v= , w= , f (v) = , t= .
V1 V1 V1 R1
This transforms the system (7.2) to
dv
ǫ = −f (v) − w,
dt (7.3)
dw
= v − γw
dt
with
R12 Cm V0 R
ǫ= , vo = , γ= .
L V1 R1
The simplest choice for the function f (v) with three zeros is the cubic poly-
nomial
f (v) = v(v − 1)(v − a)
and this gives the Fitzhugh-Nagumo model.
The Fitzhugh-Nagumo model is a simplification of the four-dimensional
Hodgkin-Huxley model proposed in the early 1950’s by Sir Alan Hodgkin and
Sir Andrew Huxley,3 for which they received the Nobel Prize in Physiology
and Medicine in 1963 and which is still used now for studying neurons and
other kinds of cells. The model describes an excitable system, and these occur
in a large variety of biological systems.
If γ is not too large, the only equilibrium of the system (7.1) is v = 0,
w = 0, and the matrix of the linearization at this equilibrium is
 
−a −1
;
ǫ −ǫγ

as this matrix has negative trace and positive determinant, the equilibrium
3 A.L. Hodgkin and A.F. Huxley, A quantitative description of membrane current and

its application to conduction and excitation in nerve, J. Physiology 117 (1952), 500–544.
362 Dynamical Systems for Biological Modeling: An Introduction

(0,0) is asymptotically stable. However, it is interesting to see how a solution

which starts at a point (v0 ,0), v0 > a, approaches the equilibrium. We may
do so by examining the nullclines, but it is perhaps more convincing, if less
rigorous, to use a computer algebra system to draw this orbit. Figure 7.1 shows
this orbit, traversed counterclockwise, with a = 0.3, γ = 1, ǫ = 0.01, v0 = 0.4.
Thus v amplifies quickly and then returns to zero. An unexpected property
of this orbit is that after the neuron fires and the potential drops, it overshoots
zero. This property can be derived from the nullcline analysis, and is also
observed in practice.

0.2 0.3

w 0.1 0.2

0.1
–0.2 0.2 0.4 0.6 0.8
v

–0.2 0.2 0.4 0.6 0.8 1

–0.1 v

FIGURE 7.1: One orbit of system FIGURE 7.2: An orbit of system

(7.1). (7.4) with stable equilibrium.

Another experiment gives the cell a constant input of positive ions instead
of a single pulse. If we apply a constant current J, we add J to the rate of
change of potential (assuming that 1 unit of current raises the potential by 1
unit in unit time), so that the modification of (7.1) is

v ′ = −v(v − a)(v − 1) − w + J, (7.4)

w′ = ǫ(v − γw).

This moves the equilibrium (0,0) into the first quadrant. For small values of
J this equilibrium is asymptotically stable (Figure 7.2). For larger inputs it
becomes unstable and a periodic orbit is set up (Figure 7.3).
Thus a steady input leads to a periodic solution. Examination of v as
a function of t (Figure 7.4) shows a “bursting” behavior, similar to what is
observed in real neurons, with the potential rising close to 1 and then dropping
below zero.
 Systems with Sustained Oscillations and Singularities 363
0.6

0.4

w v

0.5

0.2

100 200 300 400

–0.2 0.2 0.4 0.6 0.8 1 1.2
v t

FIGURE 7.3: An orbit of system FIGURE 7.4: Periodic “bursting”

(7.4) with stable limit cycle. behavior of system (7.4) for large J.

7.1.2 A model for cat neurons

A model by McCarley and Hobson4 seeks to explain regular oscillations
in cat neuron activity associated with the sleep cycle (in particular the deep-
est part of sleep, called REM for rapid eye movement), measured in number
of electrical discharges per second. The model assumes the sleep cycle can
be described primarily in terms of two cell groups, aminergic and cholinergic
cells, with activity levels (signal strengths) x(t) and y(t). Aminergic neurons
fire steadily during waking, inhibiting the activity of REM-activating cholin-
ergic cells, but during sleep the aminergic neurons decrease their firing, and
the cholinergic neurons activate. This causes REM episodes but also begins
to excite the aminergic cells, until eventually they resume firing, once more
inhibiting the cholinergic cells and ending the REM period. During sleep this
cycle repeats several times (cf. Figure 7.5).

FIGURE 7.5: Aminergic cell activity is high while cats are awake but low
during REM sleep.
4 R.W. McCarley and J.A. Hobson, Neuronal excitability modulation over the sleep cycle:

A structured and mathematical model, Science 189: 58–60 (1975).

364 Dynamical Systems for Biological Modeling: An Introduction

The rate of change of neuron activity can thus be described as proportional

to the current level of activity, with nonlinear interaction between cell groups.
The model suggested is

x′ = ax − bxy, (7.5)
y′ = −cy + dxy,

where a, b, c, d are positive constants. The equilibrium conditions are

x(a − by) = 0, y(dx − c) = 0.

These imply that there are two equilibria: (0, 0) and (c/d, a/b). The matrix of
the linearization at an equilibrium (x, y) is
 
a − by −bx
.
dy dx − c

At the equilibrium (0, 0) this matrix is

 
a 0
,
0 −c

and it is clear that since the eigenvalues of this matrix are a, −c, this equilib-
rium is unstable.
At the equilibrium (c/d, a/b) the matrix of the linearization is
 
0 −bc/d
.
ad/b 0

Since the√ eigenvalues of this matrix are complex conjugates with real part
zero, ±i ac, the linearization at this equilibrium has periodic solutions. In
fact, the model (7.5) is the same as the model (5.1) for the Lotka-Volterra
system studied in Section 5.1, where it is shown that all solutions are given by
periodic orbits around the equilibrium. Thus the simple model (7.5) predicts
regular oscillations as observed.

Exercises
In each of Exercises 1–9, use a computer algebra system to display the behavior
of solutions of the Fitzhugh-Nagumo system (7.4) with the given values of the
parameters.

1. a = 0.3, v0 = 0.2, γ = 1, ǫ = 0.01, J = 0

2. a = 0.3, v0 = 0.4, γ = 1, ǫ = 0.01, J = 0
3. a = 0.3, v0 = 0.4, γ = 1, ǫ = 1, J = 0
4. a = 0.3, v0 = 0.2, γ = 1, ǫ = 0.01, J = 0.3
 Systems with Sustained Oscillations and Singularities 365

5. a = 0.3, v0 = 0.2, γ = 1, ǫ = 1, J = 0.3

6. a = 0.3, v0 = 0.4, γ = 1, ǫ = 1, J = 0.3
7. a = 0.3, v0 = 0.2, γ = 10, ǫ = 0.01, J = 0
8. a = 0.3, v0 = 0.2, γ = 10, ǫ = 0.01, J = 0.3
9. a = 0.3, v0 = 0.4, γ = 10, ǫ = 1, J = 0.3
10. Summarize the results of the above exercises: which cases showed peri-
odic “bursting” behavior? Which parameter(s) appear to be most im-
portant in determining the qualitative behavior of the system?
11. Show that the Fitzhugh-Nagumo system (7.1) has positive equilibria if
and only if γ(1−a)2 ≥ 4. Use a computer algebra system to display some
solutions of the systems (7.1) and (7.4) with a = 0.3, γ = 20, ǫ = 0.01
and J = 0, J = 0.1, J = 0.3.

12. (a) Use a phase line or linearization to analyze the behavior of the solo
equation v ′ = −v(v − a)(v − 1), which represents neuron voltage in the
absence of a blocking mechanism.
*(b) Analyze the behavior of the following model, in which the blocking
mechanism w exists but fails to act on the voltage: v ′ = −v(v −a)(v −1),
w′ = ǫ(v − γw).
366 Dynamical Systems for Biological Modeling: An Introduction

7.2 Singular perturbations and enzyme kinetics

For differential equations or systems of differential equations which depend
on a parameter there is a general theorem to the effect that solutions are
continuous functions of the parameter on any finite interval. However, if a
derivative is multiplied by a parameter which may be allowed to tend to zero,
this is not necessarily true. Such a situation is called a singular perturbation.
The mathematical analysis of singular perturbation problems was devel-
oped long after the idea appeared in applications in the physical and biological
sciences, where it manifests in phenomena which act on multiple spatial or
time scales. In the analysis of the flow of a fluid with small viscosity there are
boundary layer effects. These are large changes in the behavior of the flow in
a small spatial region at the boundary of the region of flow, first studied in
1905. In such cases it may be helpful to study the flow on two different spatial
scales: for instance, fluid passing through a tube may flow freely in the center
of the tube but not close to the edges of the tube. Many problems in the
biological sciences involve actions on very different time scales, and these may
lead to a rapid change in some of the variables on a very short initial time in-
terval while other variables act more slowly (in fact, we described this kind of
behavior in the last section, in developing the Fitzhugh-Nagumo model). One
early example was the study of enzyme kinetics, where the chemical reactions
which occur take place at vastly different rates. In this section we shall study
a typical enzyme kinetics problem, and revisit models for neuron bursting,
through the lens of singular perturbations, which allows us to isolate the two
(or more) scales on which the process under study operates and deal with
them one at a time.
Singular perturbation problems arise in models (systems of differential
equations) containing a small parameter ǫ, of the form

ǫy ′ = f (y, z, ǫ), y(0) = y0 (7.6)

z′ = g(y, z, ǫ), z(0) = z0

with solution (y(t, ǫ), z(t, ǫ)). There is a corresponding reduced system ob-
tained by setting ǫ = 0,

f (y, z, 0) = 0 (7.7)
′
z = g(y, z, 0), z(0) = z0

with solution (y0 (t), z0 (t)).

Since ǫ is assumed to be small, the form (7.6) suggests that the y reaction
time is much faster than the z reaction time. Thus y goes to its equilibrium
value rapidly, and at its equilibrium f (y, z, 0) = 0. Then we might expect
that the reduced problem (7.7) is a good approximation to the full problem
(7.6) after a short initial time interval near t = 0 during which y moves to its
 Systems with Sustained Oscillations and Singularities 367

equilibrium value. Because (7.7) is a first-order differential equation (which

requires one initial condition to identify a unique solution) and (7.6) is a two
dimensional system (requiring two initial conditions for a unique solution) we
must expect to lose an initial condition in the reduction, and this suggests
that the solutions of (7.7) and (7.6) (each derived on a different time scale)
may not agree close to t = 0.
If the partial derivative fy (y, z, 0) 6= 0 we may solve the equation
f (y, z, 0) = 0 for y as a function of z, y = φ(z). Thus the reduced system
(7.7) is equivalent to the first-order initial value problem
z ′ = g(φ(z), z, 0), z(0) = z0 . (7.8)
Then we have the solution of (7.7) with z0 (t) the solution of (7.8) and y0 (t) =
φ(z0 (t)). Then y0 (0) = φ(z0 ). If y0 6= φ(z0 ) it is not possible for the solution
of the reduced problem (7.7) to satisfy the two initial conditions of the full
problem (7.6). The solution (y0 (t), z0 (t)) of the reduced problem is called the
outer solution. In order to use the solution of the reduced problem (7.7) as
an approximation to the solution of the full problem (7.6), we would need a
result to the effect that for each t away from t = 0 the solution of the reduced
problem (7.7) is the limit as ǫ → 0 of the solution of the full problem (7.6).
There is such a result, and we will state it shortly.
In applications one often makes a quasi-steady-state hypothesis, that y
remains almost constant, so that y ′ ≈ 0. This hypothesis is expressed as
f (y, z, 0) = 0; in singular perturbation language the hypothesis is just that
the full problem is approximated by the reduced problem.
Of course, because y(t, ǫ) (the solution to the full problem) and y0 (t) (the
solution to the reduced problem) do not match at t = 0, we should expect
that y(t, ǫ) changes rapidly for t close to 0. To analyze this, we change the
time scale by making the change of independent variable t = ǫs (making s
“fast” time, relative to “slow” time t). Then for any function u we have, by
the Chain Rule of calculus,
du du ds 1 du
= · = ·
dt ds dt ǫ ds
and the system (7.6) is transformed to
dy
= f (y, z, ǫ), y(0) = y0 ; (7.9)
ds
dz
= ǫg(y, z, ǫ), z(0) = z0 .
ds
A very small t-interval, called a boundary layer, corresponds to a much longer
s-interval. Since ǫ is small, the second equation of (7.9) says that the second
variable z remains almost constant initially and may be replaced by z0 . This
suggests solving the initial value problem
dy
= f (y, z0 , 0), y(0) = y0 , (7.10)
ds
368 Dynamical Systems for Biological Modeling: An Introduction

called the boundary layer system, to give an approximation, called the inner
solution to the behavior of the system near t = 0. Away from t = 0 we
hope that the solution of the full problem is approximated well by the outer
solution.
Sometimes a problem is given in the form (7.9) from the start. The under-
lying idea in a singular perturbation problem is that there are two different
time scales inherent in the problem, and this makes it possible to analyze the
problem separately on each time scale. The reduction in dimension because of
this separation simplifies the analysis.
The mathematical treatment of singular perturbations began in the 1940’s
from the perspective of asymptotic expansions. A few years later the qualita-
tive result which justifies the use of the reduced system as an approximation
to the full system was obtained independently in the U.S.A. and the Soviet
Union:5

THEOREM (Levinson-Tihonov): Suppose that

1. f, g are smooth functions,
2. the equation f (y, z, 0) = 0 can be solved for y as a smooth function
of z, y = φ(z),
3. the reduced system (7.7) has a solution on an interval 0 ≤ t ≤ T ,
4. the boundary layer system (7.10) has an asymptotically stable equi-
librium.
Then y(t, ǫ) → y0 (t), z(t, ǫ) → z0 (t) as ǫ → 0+ for 0 < t ≤ T . The
convergence of y is non-uniform at t = 0.

There is an extension of this result to infinite time intervals.6

THEOREM: Suppose, in addition to the hypotheses of the Levinson-

Tihonov theorem, that the reduced system (7.7) has a solution which is
asymptotically stable and that the boundary layer system (7.10) has a so-
lution which is asymptotically stable uniformly in z0 . Then the convergence
is uniform on closed subsets of 0 < t < ∞.

The essential content of these results is that if ǫ is sufficiently small the

solution of the reduced system is a good approximation to the solution of
the singularly perturbed system except very close to t = 0. The relation
5 N. Levinson, Perturbations of discontinuous solutions of nonlinear systems of differential

equations, Acta Math. 82 (1950), 71–106; A.N. Tihonov, On the dependence of the solutions
of differential equations on a small parameter, Mat. Sbornik NS 22 (1948), 193–204.
6 F.C. Hoppensteadt, Singular perturbations on the infinite interval, Trans. Amer. Math.

Soc. 123 (1966), 521–535.

 Systems with Sustained Oscillations and Singularities 369

f (y, z, 0) = 0 is called the quasi-steady-state hypothesis. Close to t = 0 the so-

lution of the boundary layer system (7.10) describes the behavior of solutions.
Thus the analysis of a singular perturbation problem can be decomposed into
the analysis of two simpler problems, namely the boundary layer system and
the reduced problem. Curiously, in fluid dynamic applications the focus of
attention has been on the boundary layer system, whereas in most biological
applications the primary interest has been in the long-term behavior, that is,
the reduced problem.

Example 1.
Describe the solution of the first-order differential equation
ǫy ′ = −y, y(0) = 1.

Solution: The solution of this initial value problem may be obtained easily by
separation of variables and is
y(t, ǫ) = e−t/ǫ .
We may calculate (
1 for t = 0
lim y(t, ǫ) =
ǫ→0 0 for t > 0.
This limit is discontinuous at t = 0. When ǫ = 0 the problem is no longer
an initial value problem but is just the relation y = 0 together with the
(incompatible) initial condition y(0) = 1. This indicates that the solution for
ǫ > 0 begins with the value 1 at t = 0 and then decreases rapidly to 0. A
graph of the solution for a small value of ǫ would indicate this boundary layer
(see Exercise 3 below).
Another way to approach this problem is to change the time scale by
making the change of independent variable t = ǫs to transform the problem
to
dy
= −y, y(0) = 1,
ds
whose solution is y(s) = e−s = e−t/ǫ .

Example 2.
Describe the behavior of solutions of the initial value problem
ǫy ′ = −y, y(0) = 1, (7.11)
′
z = −yz, z(0) = 1.

Solution: We let (y(t, ǫ), z(t, ǫ)) be the solution of this problem and we let
(y0 (t), z0 (t)) be the solution of the reduced problem
y = 0, (7.12)
z′ = −yz, z(0) = 1,
370 Dynamical Systems for Biological Modeling: An Introduction

which is obtained by setting ǫ = 0 in (7.11). It is easy to see that the solution

of the reduced problem (7.12) is given by y0 (t) = 0, z0 (t) = 1. In order to
solve the full problem (7.11) we may solve its first equation as in Example
1 to obtain y(t, ǫ) = e−t/ǫ and then substitute this solution into the second
equation to obtain z ′ = −e−t/ǫ z, z(0) = 1. The solution of this is
−t/ǫ)
z(t, ǫ) = e−ǫ(1−e

(see Exercise 1 below). Since limǫ→0 z(t, ǫ) = 1 for all t ≥ 0 (see Exercise 2
below) we see that z(t, ǫ) → y0 (t) as ǫ → 0 for all t ≥ 0. However, as we have
seen in Example 1, (
1 for t = 0,
lim y(t, ǫ) =
ǫ→0 0 for t > 0,
and
lim y(t, ǫ) 6= y0 (t) ≡ 0
ǫ→0

for t = 0. It is possible, just as in Example 1, to stretch the time variable by

setting t = ǫs to give the system
dy
= −y, y(0) = 1
ds
dz
= −ǫyz, z(0) = 1
ds
and to solve this version of the problem (see Exercise 4 below). 

Example 3: The Fitzhugh-Nagumo model.

In the previous section we described the Fitzhugh-Nagumo model, and re-
marked that the two variables in this system operated on different time scales.
However, we did not make any use of this in our description of the behavior
of the model. Now we revisit this model thinking of ǫ as a small parameter
and view it as a singular perturbation problem.
Solution: We recall the Fitzhugh-Nagumo model
dv
ǫ = −f (v) − w, (7.13)
dt
dw
= v − γw
dt
with
f (v) = v(v − 1)(v − a).
Then v is a fast variable and w is a slow variable if ǫ is small, with the quasi-
steady-state (dv/dt = 0) given by the cubic

w = −f (v) = −v(v − a)(v − 1).

 Systems with Sustained Oscillations and Singularities 371

0.1
B

w 0.1
B
w

–0.2 0.2 0.4 0.6 0.8 1

v
A

–0.2 0 0.2 0.4 0.6 0.8 1

v
A

FIGURE 7.6: Quasi-steady state FIGURE 7.7: An orbit of the

for the Fitzhugh-Nagumo system. Fitzhugh-Nagumo system.

If we represent the system graphically, the quasi-steady state (QSS) is a

curve in the (v, w) plane with three monotone branches which we denote

v = v1 (w), v = v2 (w), v = v3 (w),

seen in Figure 7.6 as the branches to the left of the point A, between the
points A and B, and to the right of the point B, respectively. The fast-time
dynamics in v make solutions move (horizontally) toward [the nearest stable
branch of] the QSS, and then the slow-time dynamics in w send solutions up
or down the QSS curve toward the w-coordinate v/γ. To see this, one can first
analyze the behavior of the “fast” dynamics in v (for fixed w, see Exercise 11)
and then the behavior of the “slow” dynamics in w (see Exercise 12). Analysis
of the v equation shows that the branch v2 is unstable while v1 and v3 are
locally asymptotically stable (in the v direction), so if we take w(0) = 0, then
v = a, where a is the middle intersection of the quasi-steady-state curve (the
unstable v2 ) with the v axis, separates solution behaviors into two distinct
groups, depending on the initial value of v. If v(0) < a, then v goes directly
to the equilibrium (0, 0). If v(0) > a, then v goes rapidly to the right branch
v = v3 (w) while w remains constant. Then, in the slow time, the orbit follows
the curve v = v3 (w) until it comes near the point B; this is the excited phase
of the motion. If γ is small enough, say γ < 4.5 (this seems arbitrary but it’s
not, see Exercise 13), then near B we have dw/dt = v − γw > 0, indicating
a continued upward trajectory, but at B the graph of the quasi-steady-state
w = −f (v) turns downward. The orbit thus cannot continue to follow the
quasi-steady-state curve past B; it is then governed by a fast system

ǫv ′ = −f (v) − w, w′ = 0

and moves rapidly horizontally until it reaches the branch v = v1 (w). It then
follows this branch of the quasi-steady-state curve to the equilibrium (0, 0).
372 Dynamical Systems for Biological Modeling: An Introduction

This behavior is illustrated in Figure 7.7, which shows an orbit for the system
(7.13) with parameter values a = 0.3, γ = 1, ǫ = 0.002 along with the quasi-
steady-state curve.
Now suppose a constant current is applied, changing the model to
dv
ǫ = −f (v) − w + J, (7.14)
dt
dw
= v − γw.
dt
Then the effect is to move the quasi-steady-state curve upward (it is now
w = −f (v) + J) and the equilibrium into the first quadrant (since the line
v = γw where dw/dt = 0 only passes through the first and third quadrants).
It is possible to show that when J is large enough for the equilibrium to reach
the portion v = v2 (w) of the quasi-steady-state curve (at the point labeled A)
the equilibrium becomes unstable and the orbit becomes periodic, oscillating
between the branches v = v1 (w) and v = v3 (w). 

7.2.1 Bursting
In the Fitzhugh-Nagumo model, and also in the four-dimensional Hodgkin-
Huxley model, which models the behavior of a neuron somewhat more closely,
neurons may fire periodically if stimulated by a constant applied current.
Many cells exhibit a more complicated behavior called bursting, in which
periods during which the potential changes slowly alternate with periods of
rapid oscillation. Such behavior is observed, for example, in groups of pan-
creatic β-cells. Action potential bursting in these cells, clustered together in
the pancreas in groups called the islets of Langerhans (see Figure 7.8), plays
a critical role in secreting the hormone insulin, used for maintaining glucose
levels in the body. (Type II diabetes, hyperglycemia, develops when this burst-
ing behavior does not compensate correctly in response to glucose levels in
blood plasma.) Here we will not explore the rather complicated models which
have been formulated to attempt to explain bursting phenomena,7 but we
will sketch out a way to use multiple timescales and singular perturbations to
design a model that exhibits bursting.
The basic idea is to build a model which can exhibit either oscillation or
quiescence (a stable equilibrium), and then use a slow timescale to switch back
and forth between the two behaviors, to replicate bursting. In order to exhibit
oscillation within the fast timescale, we will need at least two dimensions
there, so we suggest a three-dimensional model with two fast variables and
7 One review of models for pancreatic β-cell bursting is Arthur Sherman and Richard

Bertram, Integrative modeling of the pancreatic β-cell, in Wiley Interscience Encyclopedia

of Genetics, Genomics, Proteomics, and Bioinformatics, Part 3 Proteomics, M. Dunn, ed.,
Section 3.8 Systems Biology, R. L. Winslow, ed., John Wiley and Sons, Ltd., 2005. DOI:
10.1002/047001153X.g308213.
 Systems with Sustained Oscillations and Singularities 373

Image courtesy National Diabetes Information Clearinghouse

FIGURE 7.8: β-cells in the pancreas, clustered in islets, exhibit bursting

behavior central to insulin secretion.

one slow variable, say

ǫv ′ = f1 (v, w, z),
ǫw′ = f2 (v, w, z), (7.15)
′
z = g(v, w, z).

We would like to arrange the fast variable system to have three equilibria:
an asymptotically stable node, a saddle point, and an unstable node with an
asymptotically stable limit cycle around it. In this way the unstable equilib-
rium (saddle point) serves to separate the fast variable state space into two
different regions, one where solutions approach the stable equilibrium and
one where solutions approach the limit cycle; the dividing curve (or surface)
that extends from the unstable equilibrium to do this is called a separatrix,
formed by the stable manifold of the saddle point, separating the domain of
attraction of the asymptotically stable node from the domain of attraction
of the limit cycle. (The unstable equilibrium branch v2 in the fast dynamics
of the Fitzhugh-Nagumo example serves similarly to separate the domains of
attraction for the stable branches v1 and v3 .)
The slow variable z then follows the quasi-steady state curve

f1 (v, w, z) = 0, f2 (v, w, z) = 0.

We would like to arrange for the slow variable to move solutions back and
forth across the separatrix. A move into the domain of attraction of the limit
cycle will trigger a rapid oscillation, and then a move back into the domain of
attraction of the asymptotically stable node will lead to a quiescent period.
This cycle will be repeated to give bursting.
374 Dynamical Systems for Biological Modeling: An Introduction

An (admittedly artificial) example is a fast system

v′ = w − v 3 + 3v 2 , (7.16)
w′ = 1 − 5v 2 − w

coupled with a slow variable z and an applied voltage J to give the system

v′ = w − v 3 + 3v 2 + J − z,
w′ = 1 − 5v 2 − w, (7.17)
′
z = ǫ[s(v − v1 ) − z]
√
where v1 = ( 5 − 1)/2. In the slow dynamics, when v is large, z increases
and becomes large, but then v ′ is decreased by the larger z, until eventually v
decreases, and then so does z. With the parameter values J = 2, ǫ = 0.002, s =
4 the graph of v for this system is shown in Figure 7.9.
2

v 1

0 200 400 600 800 1000

–1

–2

FIGURE 7.9: Bursting.

The phenomenon of bursting arises in many real cell situations, and the
above example is an extremely simplified version of one of the types of model
that displays bursting. Since study of the behavior of such models requires
very detailed examination of the phase portrait of the fast subsystem, we will
not go further into this subject.

7.2.2 An example from enzyme kinetics

Most biochemical reactions in living organisms involve proteins, called en-
zymes, which act as catalysts. Enzymes react on certain compounds, called
substrates. For example, hemoglobin in red blood cells is an enzyme, and the
oxygen with which it combines is a substrate. A large part of enzyme kinetics
is concerned with the study of rates of reaction and the behavior of the various
reactants.
 Systems with Sustained Oscillations and Singularities 375

A basic enzymatic reaction, proposed by Michaelis and Menten,8 involves

a substrate S reacting with an enzyme E to form a complex SE which is con-
verted into a product P and the enzyme. We let s, e, c, p denote the concen-
trations of substrate, enzyme, complex, and product, respectively. We assume
that the substrate and enzyme combine according to a mass action law with
rate k1 se, and that the complex decomposes into substrate and enzyme at a
rate k−1 c and converts to product at a rate k2 c. This leads to the system of
differential equations

s′ = −k1 se + k−1 c,
e′ = −k1 se + (k−1 + k2 )c, (7.18)
′
c = k1 se − (k−1 + k2 )c,
p′ = k2 c,

with initial conditions at the beginning of the process

s(0) = s0 , e(0) = e0 , c(0) = 0, p(0) = 0.

Since (c+e)′ = 0, c+e is a constant e0 , and we may replace e by e0 −c in (7.18).

Also, if s and c are known, then e is determined, and we may calculate p by
integration. Thus the system (7.18) may be reduced to the two-dimensional
system

s′ = −k1 (e0 − c)s + k−1 c, (7.19)

′
c = k1 (e0 − c)s − (k−1 + k2 )c

with initial conditions s(0) = s0 , c(0) = 0.

It is easy to verify (see Exercise 9 below) that the only equilibrium of
(7.19) is s = 0, c = 0 and that this equilibrium is asymptotically stable.
We are interested in more detailed qualitative information. At t = 0, since
c = 0, s = s0 , we have s′ = −k1 e0 s0 < 0, c′ = k1 e0 s0 > 0. Thus initially s
decreases from s0 while c increases from zero, and c continues to increase until
c′ = 0, or k1 (e0 − c)s − (k−1 + k2 )c = 0. Thus c increases until
k1 e0 s
c= . (7.20)
k−1 + k2 + k1 s
Similarly, we may see that s decreases until
k1 e0 s
c= . (7.21)
k−1 + k1 s
Since the value of c defined by (7.21) is greater than the value defined by
(7.20), s continues to decrease for all t (since c never reaches that higher
threshold) while c increases from zero to a maximum and then decreases.
8 L. Michaelis and M.I. Menten, Die Kinetik der Invertinwirkung, Biochem. Z. 49 (1913),

333–369.
376 Dynamical Systems for Biological Modeling: An Introduction

The hallmark of a singular perturbation problem is that it contains differ-

ential equations with very different reaction rates. In order to identify (7.19)
as a singular perturbation problem we put it into dimensionless form. Be-
cause each term in (7.19) must have the same dimension we see that k1 has
dimension 1/(cell concentration)(time) and k−1 , k2 have dimension 1/(time).
We let
c s
y= , z= ,
e0 s0
both dimensionless, and then define the dimensionless time τ = k1 e0 t. Then
dc dy dτ dy
= e0 dy
dt = e0 = k1 e20
dt dτ dt dτ
ds dz dτ dz
= s0 dz
dt = s0 = k1 e0 s0
dt dτ dt dτ
and the system (7.18) is transformed to

dy
k1 e20 = k1 e0 (1 − y)s0 z − (k−1 + k2 )e0 y, (7.22)
dτ
dz
k1 e0 s0 = −k1 e0 (1 − y)s0 z + k−1 e0 y.
dτ
We now define the parameters
k2 k−1 + k2 e0
λ= , K= , ǫ=
k1 s0 k1 s0 s0
so that K − λ > 0. Then the system (7.22) becomes

dy
ǫ = z(1 − y) + Ky, y(0) = 0, (7.23)
dτ
dz
= −z(1 − y) + (K − λ)y, z(0) = 1.
dτ
In many enzyme reactions the enzymes are very effective catalysts and
the concentration of enzyme needed is very small compared to the substrate
concentration. This means that e0 is much smaller than s0 and thus that ǫ is
very small. Often ǫ is between 10−2 and 10−7 . Thus we may view (7.23) as a
singular perturbation problem of the form (7.6) with

f (y, z) = z(1 − y) − Ky, y0 = 0, g(y, z) = −z(1 − y), z0 = 1.

This means that except in the boundary layer very close to t = 0, where c may
change rapidly, we may approximate the solution of (7.19) by the solution of
the reduced problem

s′ = −k1 (e0 − c)s + k−1 c, (7.24)

0 = k1 (e0 − c)s − (k−1 + k2 )c.
 Systems with Sustained Oscillations and Singularities 377

This gives
k1 e0 s
c= ,
k1 s + k−1 + k2
and then (7.20) reduces to the single differential equation

s′ = −k1 e0 s + c(k1 s + k−1 )

k1 e0 s(k1 s + k−1
= −k1 e0 s +
k1 s + k−1 + k2
k2 e0 s
= − .
s + k−1k+k
1
2

k−1 +k2
If we let Km = k1 ,Q = k2 e0 , this becomes

Qs
s′ = − . (7.25)
s + Km
This substrate uptake rate is called a Michaelis-Menten uptake, and reaction
rates of this form occur in other problems. For example, the predator func-
tional response in predator-prey models is often assumed to be of this form.
We recall that in the original model (7.18) the rate of formation of product is
given by p′ = k2 c. For the reduced problem we may replace c in this expression
by (7.20), and then we obtain p′ = −s′ . Thus the rate of reaction is given by
(7.25).

Exercises
−t/ǫ)
1. Show that the solution of z ′ = −e−t/ǫ z, z(0) = 1 is z(t, ǫ) = e−ǫ(1−e .
−t/ǫ)
2. Show that limǫ→0 e−ǫ(1−e = 1 for all t ≥ 0. [Hint: Show that
limǫ→0 ǫ(1 − e−t/ǫ ) = 0.]
3. Graph the solution of the initial value problem in Example 1 with the
values ǫ = 0.2, ǫ = 0.1, ǫ = 0.01.
4. Find the solution of the stretched version of the initial value problem of
Example 2.
5. Solve the initial value problem ǫy ′ = y − y 3 , y(0) = 1/2 and compare the
solution with the solution of the corresponding reduced problem.
6. Solve the initial value problem

ǫy ′ = y − y3, y(0) = 1/2,

′
z = yz, z(0) = 1.

7. Find the equilibria and analyze their stability for the system (7.16).
378 Dynamical Systems for Biological Modeling: An Introduction

8. Use a computer algebra system to sketch the graph of v as a function of

t for the system (7.17) with J = 1.5, J = 2, and J = 2.5.
9. Show that the only equilibrium of (7.19) is s = 0, c = 0 and that this
equilibrium is asymptotically stable.
10. Sketch in the phase plane the orbit with initial value s = s0 , c = 0 and
identify the portion of the orbit which corresponds to the initial fast
reaction.
11. Complete a qualitative analysis of the behavior of solutions of the “fast”
dynamics of the Fitzhugh-Nagumo model
dv
= −v(v − a)(v − 1) − w,
dt
considering w as a constant. In particular, without attempting to solve
explicitly for equilibrium values v ∗ , find the points A and B shown in
Figure 7.6 and discussed in Example 3, and show that any equilibrium
located between them must be unstable, while all other equilibria outside
[A, B] must be locally asymptotically stable.
12. Complete a qualitative analysis of the behavior of solutions of the “slow”
dynamics in the Fitzhugh-Nagumo model, dw dt = v − γw, considering v
as a constant (even though it’s really not constant, this can give an
indication of the direction of change of w at any point in time).
13. * Use the results of the previous two exercises (in particular, the expres-
sion for B in terms of a, and the fact that dw/dt > 0 when v > γw) to
find a condition on γ that makes dw/dt > 0 near the point (B, −f (B))
on the QSS curve. [Hint: first write the condition in terms of a and B,
then just in terms of a, and finally consider the range of values of the
bound on γ as a varies from 0 to 1.]
14. Use a computer and software such as Maple, Mathematica or MATLAB
to verify numerically the claim made at the end of Example 3 regarding
the behavior of the Fitzhugh-Nagumo model as a current J is applied
across the cell membrane. For fixed values of the parameters a, γ and
ǫ (such as those used to generate Figure 7.7), graph the two nullclines
of system (7.14) and the solution to the system, for several values of
J beginning at 0. How does the behavior of the solutions change as J
increases? (There is a second change as J continues increasing, beyond
the change mentioned in the text.)
15. To review the section, list terms associated with fast vs. slow systems.
 Systems with Sustained Oscillations and Singularities 379

7.3 HIV: An example from immunology

The human immunodeficiency virus (HIV) is the virus which causes Ac-
quired Immune Deficiency Syndrome (AIDS). Since the early 1980s there has
been a major effort to combat this disease, and one part of this effort has been
to understand the immunology of HIV as a pathogen.
Immunology is the study of the body’s immune system, which protects the
body from external attacks, such as by pathogens. When a pathogen is intro-
duced into the body, there is an immune response which attempts to clear
it. Pathogens are captured by macrophages. These are cells which present
digested pieces of the pathogen, called antigens, to the CD4 positive lympho-
cytes (CD4+ T cells). In response these T cells reproduce by subdivision and
activate a second type of T cell, the CD8 positive T lymphocytes (CD8+ T
cells) which can seek out and destroy cells which are infected with pathogens.
A second type of response is the antibody response, in which the CD4+ T
cells signal the B lymphocyte cells in the blood which produce antibodies to
destroy the pathogen. In the remainder of this section we will use the term T
cell to mean a CD4+ T cell unless otherwise specified.
HIV and AIDS both refer to immunodeficiency because HIV attacks the
immune system, infecting and destroying the T cells, removing the body’s abil-
ity to defend itself against not only HIV but other, more common pathogens.
Progression to AIDS is typically defined in terms of a patient’s T cell count.
For these reasons this section will focus on describing and explaining the long-
term interaction between HIV and T cells within an average individual.
When HIV infects the body (see Figure 7.10) it attacks the T cells. It causes
an infected host cell to produce new virus particles, either slowly during the
lifetime of the cell or (eventually) in a rapid burst which also kills the cell
(lysis). Immediately after infection the amount of virus detected in the blood
rises rapidly. This acute stage is accompanied by flu-like symptoms. After a
few weeks the virus concentration decreases and the symptoms disappear (the
infection is then said to be clinically latent). An immune response to the virus
occurs, and antibodies against the virus can be detected in the blood. If these
antibodies are detected in a person’s blood, that person is said to be HIV-
positive. The level to which the virus falls after the primary infection is called
the set point. The virus concentration then remains almost constant for several
years, and although there are no disease symptoms, there is a gradual decrease
in the T cell count. This asymptomatic period may last as long as 10 years.
This is followed by a passage to AIDS, in which the T cell count drops to a
very low level and the body is unable to fight off infections. For an uninfected
individual the T cell count is approximately 1000/mm3. Progression to AIDS
is usually measured by a drop below 200/mm3 in the T cell count. Figure
7.11 shows the decrease in T cell count over ten years for a typical patient,
superimposed on a graph of HIV viral load over the same time frame.
380 Dynamical Systems for Biological Modeling: An Introduction
Acute Clinically latent phase
1000.à
phase AIDS æ
æ 6.
à

log10 HHIV RNA copiesmL plasmaL

800. æ

CD4+ count Hcellsmm3 L

5.
à
600. à à à
æ à à æ
æ 4.
æ æ æ
400. æ æ à à
æ
æ à
æ
à
3.
200. à

à
æ à à2.
1 1 1 1 3 5 7 9 11
12 6 4

years
Photo courtesy CDC/Maureen Metcalfe, Tom
Hodge
FIGURE 7.11: This graph, adapted from
FIGURE 7.10: HIV-1 viri- widely circulated ones in Pantaleo et al.
ons. (1993) and Fauci et al. (1996), shows the pro-
gression of within-host viral load (circles) and
corresponding CD4+ T cell count (squares)
[vs. time since infection] in a typical HIV pa-
tient, over the initial acute phase, the clini-
cally latent phase, and eventual progression
to AIDS. Note the change in timescale after
3 months.

In modeling the interaction between T cells and HIV, one must make deci-
sions about which processes and approaches to include. Must all the immune
responses be modeled separately? Infected T cells spawn HIV virions inter-
nally during a quiescent or latent period, followed by lysis and the sudden
release of free virions into the system; since much of the viral load comes
from other types of infected cells, need each of these stages (or indeed the
infected T cells at all) be represented explicitly? Is the initial acute phase im-
portant in determining the long-term outcome? Do viral and immune system
dynamics take place on different enough timescales that a singular pertur-
bation approach would help? Some answers are offered by the specificity of
one’s hypothesis (which often claims one mechanism to be most important in
explaining observed behavior), while others can only be determined by trying
several models and/or analysis approaches, and then selecting the simplest
one which explains observations well. We shall now employ both ways, via
multiple tries beginning with a simple model for a healthy immune system.9
9 Much of the pioneering work in modeling HIV immunology builds on work by Perelsen

and colleagues; one set of models whose structure parallels that developed in this section
is discussed in Alan S. Perelson and Patrick W. Nelson, Mathematical analysis of HIV-1
dynamics in vivo, SIAM Review 41(1): 3–44, March 1999. Another helpful reference for this
topic is M.A. Nowak and R.M. May, Virus dynamics: mathematical principles of immunol-
ogy and virology, Princeton University Press, 2003.
 Systems with Sustained Oscillations and Singularities 381

7.3.1 A basic model

T cells are produced mainly in the thymus. The production is not well
understood, but it is plausible to assume a constant production rate in an
uninfected person, and a proportional death rate. (The T cell production rate
is also known to be a decreasing function of viral load.) We let T denote the
concentration of T cells and assume a production rate s1 and a proportional
death rate µ in the absence of infection (making the average cell lifetime 1/µ).
Then the T cell concentration in a healthy individual is described by the initial
value problem
T ′ = s1 − µT, T (0) = T0
which has an asymptotically stable equilibrium s1 /µ.
We attempt to model the immune system’s interaction with HIV in terms
of the concentrations of healthy T cells, infected T cells (denoted by I), and
free virus (denoted by V ). We begin with a relatively simple model which adds
only the presence and direct influence of free virions V , incorporating just four
of the forces described earlier. The virus’s presence decreases the production
of T cells, altering the production function; if we assume that the maximal
decrease still leaves the net production rate positive (regardless of viral load),
we can describe it via a Michaelis-Menten type saturation term:
s2 V
s1 −
a+V
with s1 > s2 (and some positive half-saturation constant a). We also assume
a similar form for the baseline growth rate of the virus due to growth from
other infected cells such as macrophages and infected thymocytes,
gV
,
b+V
with maximum growth rate g and half-saturation constant b. (We justify the
dependence on V by taking it as an indicator of the overall number of infected
cells producing virions, which does not explicitly appear in the model.) Finally,
we assume the encounter rate between the immune system and free virus
has a mass action form T V ; such encounters result in the virus infecting
T cells, say at a rate k1 T V , and (taking T as an indicator of the immune
system’s capacity to eliminate the virus) in clearance of free virus, say at a
rate k2 T V . Since this latter term includes the loss of free virus in the process
of infecting T cells, we take k2 > k1 . (Since the T cell level is not directly
responsible for clearance of free virus, an alternative assumption would be a
proportional clearing rate of free virus, but without this interaction term the
virus is completely unaffected by the immune system in the model.) These
assumptions yield the two-dimensional system
s2 V
T′ = s1 − − µT − k1 T V, T (0) = T0 , (7.26)
a+V
gV
V′ = − k2 T V, V (0) = V0 .
b+V
382 Dynamical Systems for Biological Modeling: An Introduction

A reasonable initial condition for T is the stable equilibrium of the healthy T

cell model, T0 = s1 /µ.
Note that (7.26) includes the healthy T cell model as a special case where
V0 = 0, and therefore has the virus-free equilibrium T = T0 , V = 0. The
stability of the virus-free equilibrium (T0 , 0) is straightforward to determine.
The matrix of the linearization of system (7.26) at (T0 , 0) is

−µ − sa2 − k1 T0
 
g ,
0 b − k2 T0

which has eigenvalues −µ and gb −k2 T0 . Thus this equilibrium is asymptotically

stable if and only if g < k2 T0 b (the maximum virus growth rate is less than
the clearance rate when T = T0 , V = b).
The existence of any endemic equilibrium is more complicated to determine
analytically (the condition is quadratic, so there are 0, 1, or 2 of them) and
best considered graphically, using a phase portrait. It is straightforward to
solve both equilibrium conditions for T as functions of V , so we write the
nullclines as
(s1 − s2 )V + as1 g
T = FT (V ) = , T = FV (V ) = .
(V + a)(k1 V + µ) k2 (V + b)

Both of these are positive, decreasing, concave up functions of V (for V ≥ 0).

Examining the T nullcline (dT /dt = 0), we see FT (0) = sµ1 = T0 and
limV →∞ FT (V ) = 0; for the V nullcline (dV /dt = 0), FV (0) = kg2 b and
limV →∞ FV (V ) = 0. Comparing the two values at V = 0, we see their
order is determined by the stability condition for the virus-free equilib-
rium: FV (0) < FT (0) if and only if g < k2 T0 b. Since both functions ap-
proach zero as V → ∞, to determine their order as V → ∞ we instead
compare limV →∞ V F (V ). We find limV →∞ V FT (V ) = (s1 − s2 )/k1 while
limV →∞ V FV (V ) = g/k2 . Thus FV (V ) < FT (V ) as V → ∞ if and only if
s1 −s2
k1 > kg2 . There are four combinations of these two conditions, and there-
fore at least as many different possible phase portraits for the system. Here
we shall consider only two of them (see Exercise 2 for the full set).
If g > bk2 T0 , so that the virus-free equilibrium is unstable, and in addition
s1 −s2
k1 > kg2 , then FV begins above FT (at V = 0) but ends up beneath it (as
V → ∞), so the two nullclines must cross, with an endemic steady state at
their intersection. The phase portrait (Figure 7.12) shows this equilibrium to
be [globally] asymptotically stable. (This proves that these two conditions are
sufficient to bring about a globally stable endemic state; in theory it might be
possible for one to exist some other way, but later in this section we will show
that these conditions are indeed necessary as well.)
If instead g > bk2 T0 and kg2 > s1k−s 1
2
, then FV begins and ends above FT .
The simplest possible behavior here is for the two nullclines never to meet,
with the V nullcline lying above the T nullcline for 0 ≤ V < ∞. In this case
all crossings of the V nullcline are downward, and crossings of the T nullcline
 Systems with Sustained Oscillations and Singularities 383
FHVL FHVL

V V

FIGURE 7.12: Nullclines of the sys- FIGURE 7.13: Nullclines of the sys-
tem (7.26) in the V -T plane, in the tem (7.26) in the V -T plane, in the
case where kg2 > bT0 and kg2 < s1k−s
1
2
. case where kg2 > bT0 , kg2 > s1k−s
1
2
and
the nullclines do not cross.

are to the right (Figure 7.13). Thus solutions will enter the region between the
curves, either from above or from below; any solution lying between the two
nullclines must remain between them for all further time (moving down and
to the right). Thus T will decrease and V will increase for all t, with V → ∞
and T → 0, indicating progression to AIDS.
Given the relatively complicated analysis, a simplification technique may
help us better understand the model system’s behavior. Since the changes in
V take place on a much faster time scale (a few weeks) than the changes in
T (several years), it is reasonable to consider rewriting (7.26) as a singular
perturbation problem of the form
 
′ s2 V
T = cT s1 − − µT − k1 T V ,
a+V
 
′ gV
ǫV = cV − k2 T V ,
b+V
with proportionality coefficients cT and cV , allowing us to rescale the time
variable to study either fast (inner system) or slow (outer system) time. Let
us suppose that the values of the coefficients in (7.26) support this notion.
Then our analysis of (7.26) can be decomposed into the analysis of the inner
system with T constant, T = T0 :
dV gV
= − k2 T0 V, V (0) = V0 ≥ 0, (7.27)
dτ b+V
in the boundary layer near t = 0 (where τ is the rescaled, “fast” time), and
an outer system for T (with V given in terms of T ).
For the inner system, the first-order equation (7.27) can easily be shown
384 Dynamical Systems for Biological Modeling: An Introduction

to have a unique asymptotically stable equilibrium: V = 0 if g < bk2 T0 , and

V = k2gT0 − b > 0 if g > bk2 T0 (see Exercise 1 below). Thus if g < bk2 T0
the virus is eradicated quickly while if g > bk2 T0 the viral load has a positive
limit.
The outer system, meanwhile, is
s2 V
T ′ = s1 − − µT − k1 T V, T (0) = T0 , (7.28)
a+V
with
gV
− k2 T V = 0. (7.29)
b+V
In order to analyze the outer system we first solve (7.29) for V as a function
of T , obtaining either V = 0 or (if T < g/k2 b)
g
V = − b, (7.30)
k2 T
next substitute this solution into (7.28), and then analyze the resulting first-
order equation. If V = 0, the outer system is the healthy T cell equation
T ′ = s1 − µT for which T (t) → T0 as t → ∞. This is the case in which the
virus is eliminated rapidly. Note that if, at any point in time, T (t) > g/k2 b,
then V = 0 is the only nonnegative solution for V , and (revisiting the inner
system with this value of T ) V then quickly (i.e., in fast time) approaches 0,
leaving T to approach T0 .
To treat the case V = k2gT −b, we substitute (7.30) for V into (7.28), which
yields, after some simplification, the outer system
g − k2 bT g
T ′ = F (T ) = s1 −s2 −(µ−k1 b)T −k1 , T (0) = T0 , (7.31)
g − k2 (b − a)T k2

with g > k2 bT0 and, since the virus terms should make F (T ) < s1 − µT
(and thus T (t) ≤ T0 ), T (t) < g/k2 b more broadly. Thus in analyzing the
outer system we restrict our attention to the interval [0, g/k2 b]. Given the
complexity of F , we proceed using basic properties of F rather than finding
its roots. We observe
   
s1 − s2 g g
F (0) = k1 − , F (g/k2 b) = µ T0 − < 0,
k1 k2 k2 b

s2 agk2
and calculate F ′ (T ) = k1 b − µ + ,
[g + (a − b)k2 T ]2

s2 ak2 s2 k2 b2
F ′ (0) = k1 b − µ + , F ′ (g/k2 b) = k1 b − µ + .
g ga
g
Both F and F ′ have a pole (i.e., a vertical asymptote) at T = k2 (b−a) , but
it does not fall within [0, g/k2 b]: if a > b the pole is negative and F ′ (which
 Systems with Sustained Oscillations and Singularities 385

decreases away from the pole) is then monotone decreasing on [0, g/k2 b], while
if a < b some algebra shows that the pole is greater than g/k2 b, making F ′
monotone increasing on [0, g/k2 b]. In either case F ′ is monotone, so F ′ has at
most one root (F ′ = 0) in [0, g/k2 b], meaning F can “turn around” at most
once.
Thus if F (0) > 0 there is a unique equilibrium in (0, g/k2 b) (since
F (g/k2 b) < 0), and since F crosses from positive to negative there, F ′ < 0
there, making the equilibrium asymptotically stable (locally, and also globally
since there are no other stable equilibria). This indicates an endemic state.
If a > b (making F ′ monotone decreasing and thus F concave down) and
′
F does have a root, Tc , in [0, g/k2 b], then Tc is a local maximum for F , and
it is possible to have F (0) < 0 but F (Tc ) > 0. In this case there are two
equilibria: one in (0, Tc ) (call it E0 ) which is unstable (since F crosses from
negative to positive there, making F ′ > 0), and one in (Tc , g/k2 b) (call it E1 ),
asymptotically stable (since F crosses from positive to negative there, making
F ′ < 0 as before). Here the dynamics are more complex, as the unstable E0
acts as a separatrix: If T0 < E0 , then T → 0 (and then V → ∞) and the
infection progresses to AIDS, but if T0 > E0 , then T → E1 , and the infection
persists in an endemic state.
If neither of the two sets of conditions above hold (i.e., F (0) < 0 and either
Tc does not exist in [0, g/k2 b] or F (Tc ) < 0), then F < 0 on all of [0, g/k2 b],
making T ′ < 0 so that T → 0, V → ∞ and the infection progresses to AIDS
regardless of initial conditions.
The condition F (0) > 0 (for a single, globally stable endemic equilibrium)
is equivalent to
g s1 − s2
< , (7.32)
k2 k1
seen in the analysis of the original (full) system (7.26). The more complicated
conditions that lead to two equilibria can be condensed (see Exercise 3) to
s r s s   
b k2 µ − k1 b k2 k1 b s1 − s2 k2
< < − 1− 1− . (7.33)
a g s2 a g s2 a k1 g

By viewing (7.26) as a singular perturbation problem we have been able to

obtain some information that we could not obtain directly. In the first place,
we can see that if the virus is eradicated, this takes place in the inner system
and thus happens very quickly. Second, the analysis of the outer system shows
that the conditions g > bk2 T0 and kg2 < s1k−s 1
2
obtained graphically using
nullclines are not only sufficient for the existence of a unique, globally stable
endemic equilibrium but necessary as well. Also, the relation (7.29) between
T and V is useful information obtained from the outer system.
The information that we have obtained about the model (7.26) is that
if g < bk2 T0 the virus is eradicated quickly, if g > bk2 T0 and kg2 < s1k−s 1
2

there is an asymptotically stable infected steady state without progression

to AIDS (as these two inequalities may be contradictory for some parameter
386 Dynamical Systems for Biological Modeling: An Introduction
0.001

0.0008

0.0006

V V

2
0.0004

1
0.0002

1 2 3 4 5 1 2 3 4 5
t t

FIGURE 7.14: V (t) for 0 ≤ t ≤ 5, FIGURE 7.15: V (t) for 0 ≤ t ≤ 5,

g = 1: the virus is cleared quickly. g = 125: the virus quickly reaches a
plateau.
1000
80

800

60

600

T V
40

400

20
200

500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000
t t

FIGURE 7.16: T (t) for 0 ≤ t ≤ FIGURE 7.17: V (t) for 0 ≤ t ≤

3000, g = 100: a long-term endemic 3000, g = 100: a long-term endemic
state. state.
1000

12000

800
10000

600 8000

T V

6000
400

4000

200
2000

500 1000 1500 2000 2500 3000 500 1000 1500 2000 2500 3000
t t

FIGURE 7.18: T (t) for 0 ≤ t ≤ FIGURE 7.19: V (t) for 0 ≤ t ≤

3000, g = 125: long-term progression 3000, g = 125: long-term progression
to AIDS. to AIDS.

values, this possibility may not arise), and under most other conditions there
is progression to AIDS.
For numerical simulations with given sets of parameter values, the rapid
changes in V near t = 0 may cause difficulties in the approximation. It is
 Systems with Sustained Oscillations and Singularities 387

advisable to use the equilibrium of the inner system as an initial value. This is
essentially equivalent to finding a solution of the outer problem and matching
it to the solution of the inner problem, thus using the singular perturbation
approach at least implicitly. We will give the results of some numerical simu-
lations, using the parameter values
cells  cells  virus virus
s1 = 10 day, s2 = 7 day, a = 12 , b=8 , (7.34)
mm3 mm3 mm3 mm3
 −1  −1
virus cells
k1 = 2.5 × 10−4 day , k2 = 0.01 day , µ = 0.01day −1 .
mm3 mm3

The values for s1 , s2 , µ, k1 , k2 have been derived from experimental data. The
graphs are very sensitive to changes in the parameters a and b. We use different
values of g to illustrate the possibilities of eradication of the virus, an infected
steady state, and progression to AIDS. Figures 7.14 and 7.15 show the short
term behavior with g = 1 and g = 125, respectively (units for g in all cases are
virus
mm3 day but will hereinafter be omitted for space constraints). Figures 7.16
and 7.17 show the long term behavior of T and V for g = 100, illustrating an
asymptotically stable infected steady state, while Figures 7.18 and 7.19 show
the long term behavior for g = 125, illustrating progression to AIDS.

7.3.2 Including infected cells

The system (7.26) neglects the impact of infected T cells. Since the fraction
of T cells which are infected is very small, perhaps of the order of 10−4 or
10−5 , this may be a plausible simplification. We can determine the extent to
which this simplification affects the model’s ability to describe the immune
system–virus interaction by developing and analyzing a model which includes
infected T cells. Infected cells I are created by the infection encounters already
described as occurring at a rate k1 T V ; we now also assume that they die
(through lysis) at a proportional death rate δ (the reciprocal of how long an
average infected T cell takes to rupture). Finally, we assume that, when lysis
occurs, on average each infected T cell produces N virions in its lifetime.
These virions are released when the cell dies, adding to the concentration of
free virus. These assumptions lead to the three-dimensional model
s2 V
T′ = s1 − − µT − k1 T V, T (0) = T0 ,
a+V
I′ = k1 T V − δI, I(0) = 0, (7.35)
gV
V′ = N δI + − k2 T V, V (0) = V0 .
b+V
The analysis of this more complicated system follows the same lines as
that of (7.26). Like the simpler system, (7.35) has a virus-free equilibrium
(T0 , 0, 0), and the matrix of the linearization at this equilibrium is
388 Dynamical Systems for Biological Modeling: An Introduction

− sa2 − k1 T0
 
−µ 0
 0 −δ k1 T0 ,
g
0 δN b − k T
2 0

whose eigenvalues are −µ and the eigenvaluesof the 2 × 2 matrix

 
−δ k1 T0
,
δN gb − k2 T0

which we denote by B. Thus the virus-free equilibrium is asymptotically stable

if and only if
g
trB = − k2 T0 − δ < 0,
b
g 
det B = δ − + k2 T0 − k1 N T0 > 0.
b
Since the determinant condition is stronger than the trace condition, the equi-
librium is asymptotically stable if and only if the determinant is positive, or

g < bT0 (k2 − N k1 ). (7.36)

In particular, stability requires that k2 > N k1 . The interpretation is again that

the maximum virus growth rate g must be outweighed by the net clearance
rate [when T = T0 , V = b], here bT0 (k2 − N k1 ), in order for the infection to
die out; the clearance rate due to T cell–virus encounters is reduced by the
indirect virus production that also results from such encounters (each of which
eventually produces N virions).
It is possible to show by an elementary but technically complicated argu-
ment that if (7.36) is not satisfied there is an asymptotically stable infected
equilibrium for g in a suitable range. We will omit the argument, however,
since once again we will be able to obtain a stronger result by viewing the
problem as a singular perturbation.
We recast (7.35) as a singular perturbation with I and V as fast variables
and T as a slow variable. Then the singular perturbation approach would give
a two-dimensional system for I and V as an inner system and a first-order
equation for T as an a outer problem. For this model the mean number N
of virions produced by an infected T cell turns out to be significant in the
long-term behavior of the model.
The inner problem is the system
dI
= k1 T0 V − δI, (7.37)
dτ
dV gV
= N δI + − k2 T0 V
dτ b+V
(where τ again denotes rescaled, fast time). The Jacobian matrix of this system
at the equilibrium I = 0, V = 0 is just the matrix B that we obtained in
 Systems with Sustained Oscillations and Singularities 389

analyzing the full three-dimensional model (7.35), and thus this equilibrium
is asymptotically stable if and only (7.36) is satisfied. It is not difficult to
show that the system (7.37) has another equilibrium, which is positive and
asymptotically stable, if (7.36) is not satisfied but 0 < k2 − N k1 < Tg0 b , given
by  
k1 T0 g g
I= −b , V = − b.
δ (k2 − N k1 )T0 (k2 − N k1 )T0
If, finally, k2 < N k1 , so that the virus clearance rate is outweighed by the
indirect virus creation rate from infected T cells alone, then one can show
(Exercise 6) that I, V → ∞, corresponding to AIDS.
The outer problem is
s2 V
T ′ = s1 − − µT − k1 T V
a+V
with V the appropriate limiting value of the inner system. If k2 − N k1 >
g/T0 b (so V = 0), which is the case of virus eradication, then the outer
system reduces to the healthy T cell equation and T → T0 . If k2 < N k1 (so
V → ∞), then T → 0 and the infection progresses to AIDS and death. If
0 < k2 − N k1 < g/T0 b, so that
g
V (T ) = − b,
(k2 − N k1 )T

then we may write the outer problem in the form (7.31) with k2 replaced by
k2 − N k1 in the expression for F (T ). The argument used in analyzing (7.31)
shows there is a positive asymptotically stable equilibrium if and only if
g s1 − s2
< (7.38)
k2 − N k1 k1
g s1 −s2
(analogous to (7.32)). For larger values of g ( k2 −N k1 > k1 ) two endemic
equilibria may exist if (analogous to (7.33))
r r r s   
b k2 −N k1 µ−k1 b k2 −N k1 k1 b s1 −s2 k2 −N k1
< < − 1− 1− .
a g s2 a g s2 a k1 g

Otherwise f (T ) < 0 for all T , and once again the solutions of the outer
problem are monotone decreasing, and without a positive equilibrium every
solution must approach zero, so that we have progression to AIDS.
For numerical simulations, it is reasonable to use the parameter values of
(7.34) together with δ = 0.5/day, N = 10 virions per infected T cell. However,
the value of N is not well determined experimentally, and different values may
give quite different behavior.
390 Dynamical Systems for Biological Modeling: An Introduction

Another way to improve the model (7.35) would be to incorporate the

fact that T cells do not become infective immediately on being infected but
go first into a latent stage. This can be modeled most simply by adding a
class L of latent T cells which die at the same rate as uninfected T cells and
move into the infected class at a proportional rate k3 , corresponding to an
exponential distribution of latent periods with mean 1/k3 . This would give a
four-dimensional model
s2 V
T′ = s1 − − µT − k1 T V, T (0) = T0 ,
a+V
L′ = k1 T V − µT L − k3 L, L(0) = 0, (7.39)
I′ = k3 L − δI, I(0) = 0,
gV
V′ = N µI I − k2 T V + , V (0) = V0 .
b+V
This system may be analyzed in the same way as the simpler system (7.35).
The behavior of solutions is similar to that of solutions of (7.35) but agrees
somewhat more closely with experimental data.
The models we have proposed in this section are radical simplifications of
reality. Although the concentrations in the blood are thought to be represen-
tative, only about 2% of the T cells and virions are in the blood plasma. More
accurate models would also include the concentrations in the lymphorecticu-
lar system and would distinguish between naive and memory cells, which have
different roles. Nevertheless, the very simple models we have introduced do
show many of the features observed in real life. One observation which is not
seen in our models is that after the initial rapid increase in viral load there
is a decrease with a relatively long period in which the viral load varies little
before the sharp increase in progression to AIDS (Figure 7.11).
One of the main purposes of AIDS modeling is to evaluate and compare
treatment strategies. The main treatments being used currently are combina-
tions of drugs such as AZT which are inhibitors of reverse transcriptase and
act to decrease k1 . Treatment models might consider at what point in the pro-
gression one should begin treatment. A problem is the development of drug
resistant virus strains. Early treatment is considered to be treatment begin-
ning before the T cell count has dropped below 500, but then drug-resistant
strains of virus tend to replace drug-sensitive strains rapidly. Treatment de-
layed until the T cell count falls below 200 is too late to have much effect.
It now appears that treatment beginning when the T cell count is between
200 and 500 may be the most effective choice with the treatments currently
available. Models which incorporate treatment and drug resistance appear to
corroborate this, but the question is by no means decided.
 Systems with Sustained Oscillations and Singularities 391

Exercises

1. Determine the behavior of solutions of the first-order equation (7.27).

2. For the system (7.26) draw all possible phase portraits for the cases:
s1 −s2 g
(a) k1 < k2 < bT0
g s1 −s2
(b) bT0 < k2 < k1
g s1 −s2
(c) k2 < bT0 , k1
g
(d) k2 > bT0 , s1k−s
1
2

Use the fact that the endemic equilibrium condition is quadratic to limit
the number of crossings of the two nullclines (two of the four cases above
will nevertheless have two possible portraits each).
3. Derive the inequality (7.33) from the set of conditions that lead to two
endemic equilibria for the system (7.31), as follows:
(a) Find Tc such that F ′ (Tc ) = 0. (The equation has 2 roots, but one
is always outside the interval [0, g/k2 b].) Find the condition that
Tc exist (be a real number.)
(b) Find conditions that 0 < Tc < g/k2 b when a > b. Verify that these
conditions imply that a > b and Tc is real. Show also that these
conditions are equivalent to F ′ (g/k2 b) < 0 < F ′ (0).
√ q 2 
(c) Show that F (Tc ) = F (0) + s2 a − (µ − k1 b) kg2 (a − b).
(d) Find conditions that F (0) < 0 < F (Tc ) by solving F (Tc ) > 0 for
µ−k1 b
s2 a . Simplify to (7.33), and verify that it implies all the previous
conditions (i.e., that F (0) < 0, 0 < Tc < kg2 b , a > b, Tc is real).

4. Suppose it were possible to decrease k1 to a value of 2.5 ×

−1
10−5 virus
mm3 day . Use the model (7.26) to predict the values of g for
which the behavior would change.
5. Would being able to decrease the value of s2 be helpful in treatment for
the model (7.26)? Explain.
6. Show that if k2 < N k1 there is progression to AIDS (I, V → ∞) for all
g ≥ 0, (a) in system (7.37), (b) in system (7.35). [Hint: Write first an
inequality for V ′ , and then an inequality for the derivative of the sum
of V and a certain multiple of I.]
7. Determine conditions for the stability of the virus-free equilibrium of
(7.39).
392 Dynamical Systems for Biological Modeling: An Introduction

7.4 Slow selection in population genetics

In Exercise 9 of Section 2.3 we established the Hardy-Weinberg law of

genetics, that the frequencies of genes (and their constituent alleles) remain
constant from one generation to the next if mating is random with respect to
genotype and if all genotypes have equal fitness (as measured through birth
and death rates). The assumption of equal fitness may be true for some genes,
such as those that determine widow’s peak hairlines or attached earlobes in
humans, but it is clearly not so for others, such as the gene which causes
sickle-cell disease, in which the hemoglobin molecules in red blood cells are
deformed and rigid (Figure 7.20). Now, we wish to investigate whether the
Hardy-Weinberg principle remains true if there are differences in fitness. We
shall use modeling to examine gene frequencies in a population with density-
dependent population dynamics and a genotype-dependent death rate. Incor-
porating ongoing population dynamics shifts the context from the discrete-
time models of Chapter 2, difference equations with distinct generations, to
the continuous-time models provided by differential equations. In keeping with
the theme of this chapter, we shall use a singular perturbation approach to
simplify the eventual analysis, by assuming that the differences in death rates
among different genotypes are small, and separating the time scales of the
population dynamics and the genetic selection.
To establish a baseline result and distinguish the effects of continuous
population dynamics from those of genotype-dependent fitness, we begin with
the case in which all genotypes have equal fitness.
P A R E N T 2
D H R
A A A a a a
A
D
A
p p /2 P
D H
P A R E N T 1

A
H
a
p /2 p
H R
a 1−P
R
a

Photo by Janice Haney Carr, courtesy CDC/ P 1−P

Sickle Cell Foundation of Georgia: Jackie George,
Beverly Sinclair
FIGURE 7.21: Proportions pi of
FIGURE 7.20: This electron micro- offspring with each genotype in a
graph shows a deformed red blood population with genotype popula-
cell (left) alongside normal ones in a tions yi and genotype-independent
person with sickle-cell anemia. mating.
 Systems with Sustained Oscillations and Singularities 393

7.4.1 Equally fit genotypes

To develop a model we must first define some notation. We consider diploid
organisms with alleles A, a in a population of size N (t). Let y1 (t), y2 (t), y3 (t)
be the number of members of genotype AA, Aa, aa, respectively, so that

y1 + y2 + y3 = N.

Suppose that the density-dependent birth rate Λ(N ) is the same for all geno-
types, while proportional death rates d1 , d2 , d3 for genotypes AA, Aa, aa, re-
spectively, may differ. Initially, we will assume that these death rates are
equal, d1 = d2 = d3 = d. We also make some assumptions on the form of
Λ(N ): namely, that Λ(0) = 0 and that Λ(N ) is differentiable, with

Λ(N )
Λ′ (N ) ≤ for N > 0. (7.40)
N
That is, Λ increases slower than linearly as N increases. This bound prevents
unbounded growth (since the total death rate does increase linearly in N ) and
is satisfied by most common growth functions, including logistic.
To track the quantities of interest, we define the genotype frequencies D
(for dominant), H (for heterozygous) and R (for recessive):
y1 y2 y3
D= , H= , R= ,
N N N
so that D + H + R = 1, and the allele frequencies P for A and 1 − P for a.
Note that P = D + H2 since all of the alleles of the dominant (AA) genotype
and half those of the heterozygous (Aa) genotype are As. Correspondingly
1−P =R+ H 2.
We retain the assumption from Section 2.3 that mating is random with
respect to these alleles and their associated genotypes. As noted above, we
also retain the assumption of genotype-independent fecundity, so that the
birth rate Λ(N ) applies to all genotypes. Then Figure 7.21 shows the possible
offspring for all combination of parent genotypes, and we can write expressions
for the proportions pD , pH , pR of offspring of each genotype, in terms of the
parent frequencies D, H, R. Note that in the figure, y1 , y2 , y3 are deliberately
shown as different in size to emphasize that variations in area within the square
correspond to proportions of offspring with different genetic compositions.
The dark shading shows how the proportion pD of homozygous dominant (D)
offspring can be computed as the number of offspring from two D parents,
half the offspring from a D parent and an H parent, and one quarter of the
offspring of two H parents, that is,
 2
1 1 H
pD = D2 + 2 · DH + H 2 = D+ = P 2. (7.41)
2 4 2

Likewise the light shading in Figure 7.21, which represents homozygous reces-
394 Dynamical Systems for Biological Modeling: An Introduction

sive (R) offspring, leads to the equation

 2
1 1 H
2
pR = R + 2 · RH + H 2 = R+ = (1 − P )2 . (7.42)
2 4 2

Finally, the two unshaded areas, both equal in proportion, representing het-
erozygous offspring, yield
  
H H
pH = 2 D + R+ = 2P (1 − P ). (7.43)
2 2

2
(Note that thus pD + pH + pR = D + H2 + H2 + R = P 2 + 2P (1 − P ) +


2
(1 − P ) = 1, as it should.)
The dynamics of the frequencies D, H, R (and P ) follow from those of
the populations with each genotype. Under the assumptions that mating and
fecundity are genotype-independent, we have

y1′ = pD Λ(N ) − d1 y1 ,
y2′ = pH Λ(N ) − d2 y2 , (7.44)
y3′ = pR Λ(N ) − d3 y3 .

If in addition we assume genotype-independent death, d1 = d2 = d3 = d, then

the population dynamics decouple from the genetics, and taking the sum of
the three equations, we have

N ′ = Λ(N ) − dN. (7.45)

To convert these to differential equations for the frequencies, we write

y1 = DN , so that y1′ = D′ N + DN ′ and D′ N = y1′ − DN ′ . Into this last
equation we now substitute y1′ from (7.44), pD = P 2 from (7.41), and N ′ from
(7.45):

D′ N = (pD Λ(N ) − d1 y1 ) − D(Λ(N ) − dN )

= P 2 Λ(N ) − dDN − DΛ(N ) + dDN = Λ(N )(P 2 − D),

so that
Λ(N ) 2
D′ = (P − D). (7.46)
N
Similarly y2 = HN , so y2′ = H ′ N + HN ′ and H ′ N = y2′ − HN ′ , from which,
by (7.43),(7.44),(7.45),

H ′ N = (pH Λ(N ) − d2 y2 ) − H(Λ(N ) − dN )

= 2P (1 − P )Λ(N ) − dHN − HΛ(N ) + dHN = Λ(N )(2P (1 − P ) − H),

so that
Λ(N )
H′ = (2P (1 − P ) − H). (7.47)
N
 Systems with Sustained Oscillations and Singularities 395

Since by definition D + H + R = 1, we need not write an equation for R′ ;

instead we may deduce R = 1 − D − H from D and H. In fact, however, we
may reduce the dimension of the frequency dynamics model even further, by
recalling that P = D + H2 , and calculating

H′ Λ(N ) 2 Λ(N )
P ′ = D′ + = (P − D) + (P (1 − P ) − H/2)
2 N  N 
Λ(N ) H Λ(N ) 2
= P 2 + P (1 − P ) − D + = (P + P − P 2 − P ) = 0,
N 2 N
from which we see that the frequency of allele A remains constant over time
(P ′ = 0), and thus so does the frequency (1 − P ) of allele a. This makes sense
given that all genotypes reproduce at the same rate and die at the same rate:
each member of y1 reproduces pairs of A’s at the same rate as each member
of y3 reproduces pairs of a’s and each member of y2 reproduces sets of A and
a.
The result that P ′ = 0 answers one of the original questions: continuous,
density-dependent population dynamics does not disturb the constancy of the
allele frequencies. To track the genotype frequencies over time, we need only
adjoin equations (7.45),(7.46), since H can be found from D and P , and then
R from D and H (or directly as R = 1 + D − 2P ). The system for both
frequency types can thus be summarized as
Λ(N ) 2
N ′ = Λ(N ) − dN, D′ = (P − D), P ′ = 0. (7.48)
N
Since the first equation, (7.45), is independent of the other two, the pop-
ulation dynamics can be analyzed separately. The assumptions made on the
form of Λ can be used (see Exercise 1) to show that (7.45) has a unique,
globally stable positive equilibrium if and only if Λ′ (0) > d; otherwise the
extinction equilibrium N ∗ = 0 is globally stable. In what follows, therefore,
we assume that Λ′ (0) > d, so that N (t) approaches a stable positive limit, say
K.
If we consider the two-dimensional system in N and D in (7.48) with
P = P0 , it has the unique positive equilibrium (K, P02 ). Linearization produces
the Jacobian matrix
 ′
Λ (N ) − d 0

J(N, D) = Λ (N ) Λ(N )
′
) ,
N − N2 − Λ(N
N

which has, for any N, D > 0,

 
′ Λ(N ) ′ Λ(N )
trJ(N, D) = Λ (N )− −d < −d < 0, det J(N, D) = (Λ (N )−d) − ,
N N

and, for the equilibrium (K, P02 ),

det J(K, P02 ) = (Λ′ (K) − d)(−d) > 0

396 Dynamical Systems for Biological Modeling: An Introduction

since from (7.40) Λ′ (K) < Λ(K)/K = dK/K = d, making both factors of
the determinant negative. Thus, by the equilibrium stability theorem of Sec-
tion 5.4, the equilibrium (K, P02 ) is locally asymptotically stable. In addition,
since trJ(N, D) < 0 for all N, D > 0, a theorem called Bendixson’s Crite-
rion implies that the system (7.45),(7.46) does not have a periodic orbit in
the first quadrant. Since 0 ≤ N ≤ K, 0 ≤ D ≤ 1, solutions of (7.45),(7.46)
are bounded, and now by the Poincaré-Bendixson theorem the equilibrium
(K, P02 ) is globally asymptotically stable (see Exercise 2 for discussion of
N ∗ = 0). Since H = P − D and R = 1 + D − 2P the genotype frequen-
cies D, H, and R approach P02 , P0 (1 − P0 ), and (1 − P0 )2 , respectively.
In the discrete generation model of Section 2.3 the genotype frequencies
were constant. With nonlinear population dynamics included this is not neces-
sarily true (unless the population starts out already at this equilibrium), but
still the Hardy-Weinberg distribution is approached in the limit. This result
serves as a baseline for studying the effects of varying genotype fitness.

7.4.2 Slow genetic selection

To study the consequences of genotype-dependent death rates on genotype
frequencies, we maintain the hypotheses of genotype-independent mating and
reproduction but assume the genotype affects survival; in particular, we as-
sume homozygous individuals’ death rates differ from those of heterozygous
individuals as follows:
d1 = d + ǫ∆D , d2 = d, d2 = d + ǫ∆R . (7.49)
If ∆D > 0, dominant homozygous (AA) individuals are less fit (die sooner)
than heterozygous (Aa); if ∆D < 0, they are instead more fit (die later)
than heterozygous. Likewise for ∆R and recessive homozygous (aa). Both
differences are multiplied by a dimensionless parameter ǫ << 1 to indicate
that the differences in death rates are small compared to the overall death
rates and the overall population dynamics.
Under the new set of hypotheses, system (7.44) becomes, using (7.41),
(7.42), (7.43), (7.49),
y1′ = P 2 Λ(N ) − (d + ǫ∆D )y1 ,
y2′ = 2P (1 − P )Λ(N ) − dy2 , (7.50)
y3′ = (1 − P )2 Λ(N ) − (d + ǫ∆R )y3 ,
with sum
N ′ = Λ(N ) − dN − ǫ(∆D y1 + ∆R y3 ). (7.51)
′
The equation D N = y1′ ′
− DN now becomes instead
D′ N = [P 2 Λ(N ) − (d + ǫ∆D )DN ] − D[Λ(N ) − dN − ǫ(∆D DN + ∆R RN )]
= Λ(N )(P 2 − D) + ǫDN [∆D (D − 1) + ∆R R]
Λ(N ) 2
D′ = (P − D) + ǫD[∆D (D − 1) + ∆R (1 + D − 2P )]. (7.52)
N
 Systems with Sustained Oscillations and Singularities 397

Similarly H ′ N = y2′ − HN ′ now becomes

H ′ N = [2P (1 − P )Λ(N ) − dHN ] − H[Λ(N ) − dN − ǫ(∆D DN + ∆R RN )]

= Λ(N )[2P (1 − P ) − H] + ǫHN [∆D D + ∆R (1 + D − 2P )]
Λ(N )
H′ = (2P (1 − P ) − H) + ǫH[∆D D + ∆R (1 + D − 2P )]. (7.53)
N
Again we omit an equation for R′ in lieu of D+H +R = 1 and focus instead on
′
P = D + H2 . As before, the terms without ǫ in the equation for P ′ = D′ + H2
cancel out, leaving

P ′ = ǫ[∆D D(P − 1) + ∆R P (1 + D − 2P )]. (7.54)

We collect the differential equations for population dynamics (7.51), genotype

frequency (7.52), and allele frequency (7.54) in the system

N′ = Λ(N ) − dN − ǫN (∆D D + ∆R (1 + D − 2P )),

Λ(N ) 2
D′ = (P − D) + ǫD[∆D (D − 1) + ∆R (1 + D − 2P )], (7.55)
N
P′ = ǫ[∆D D(P − 1) + ∆R P (1 + D − 2P )].

Note that, for small ǫ, P is a slow variable and N and D are fast variables;
in particular, the inner system (ǫ = 0) is given by (7.48), the system with
genotype-independent death(!). As seen in the previous subsection, the inner
solution approaches the Hardy-Weinberg proportions, D → P 2 (with P a
constant, say P0 ). The outer system, which governs the long-term evolution
of the slow variable P , is obtained using the slow-time variable10 s = ǫt, so
that dP 1 dP 2
ds = ǫ dt , and substituting D = P in (7.54) (see Exercise 3), to get

dP
= f (P ) = (∆D + ∆R )P (1 − P )(P ∗ − P ), (7.56)
ds
where
∆R
P∗ = .
∆D + ∆R
The equation (7.56) has three equilibria, P = 0, P = P ∗ , P = 1; the equilib-
rium P ∗ is biologically meaningful only if 0 ≤ P ∗ ≤ 1. The question now is
which of the meaningful equilibria are asymptotically stable, and this depends
on the signs of ∆D and ∆R .
We consider four separate cases depending on these signs, cf. Figure 7.22.
We sketch out the results in each case and leave the details for the reader to
work through (Exercises 4–7):
10 The variable s is here used to represent slow time relative to t, in contrast to its use in

Section 7.2, where it was used as fast time.

398 Dynamical Systems for Biological Modeling: An Introduction

I If ∆D < 0, ∆R > 0 then d1 < d2 < d3 and the dominant allele A is

clearly an advantage, since members of y1 live longer than those in y2 ,
who live longer than homozygous recessive y3 . This is classic dominance;
we expect natural selection to eliminate the recessive a allele over time,
and model analysis shows (Exercise 4) that this is precisely what should
happen: as t → ∞, P → 1, D = P 2 → 1, H = P − D → 0, R =
1 + D − 2P → 0.
In practice, many mutations are of the “loss-of-function” type, harmful
but recessive because one copy of the normal allele is enough to maintain
some level of proper functionality (often a protein synthesis) of whatever
part of the organism is affected by the given gene. Case I includes such
scenarios.
II If ∆D > 0, ∆R < 0 then d1 > d2 > d3 and the recessive allele a clearly
confers a fitness advantage. This is the reverse situation of Case I, and
consistent with our intuition, model analysis (Exercise 5) predicts fixa-
tion of the recessive allele: as t → ∞, P → 0, D = P 2 → 0, H = P −D →
0, R = 1 + D − 2P → 0.
A well-known example of a recessive allele which confers a clear survival
advantage, more to homozygotes than to heterozygotes, is the CCR5-
∆32 allele, which guards against both bubonic plague and HIV. CCR5-
∆32 homozygotes appear to be immune to both one type of plague11
and one variant of HIV; heterozygotes (with one copy) have partial re-
sistance to infection, and when infected appear to have milder or slower-
progressing cases. This occurs with HIV because many types of HIV use
the protein CCR5 to enter host cells, but individuals with the mutation
produce smaller (defective) versions of this protein, which HIV cannot
use. Over the centuries, the periodic waves of plague which ravaged
Europe gradually increased the frequency of this allele (since dispropor-
tionately many of the survivors had it) to a point where it now occurs at
a roughly 10% frequency in modern humans of European descent. The
fact that despite conferring a clear advantage the allele only occurs at
10% frequency after many centuries underscores the difference in time
scales between population dynamics and natural selection.
III If ∆D > 0, ∆R > 0 then both types of homozygote are disadvantaged
(have higher death rates and thus shorter lifetimes) relative to heterozy-
gotes Aa. This scenario is called overdominance and occurs with the
aforementioned sickle-cell allele as well as that for cystic fibrosis: one
copy of the [defective] recessive allele does not impede normal func-
tionality (i.e., no disease) but confers protection against an infectious
disease—malaria in the case of sickle-cell, and cholera in the case of
cystic fibrosis—while two copies of the recessive allele produce a lethal
11 It has been widely speculated that CCR5-∆32 protected against the bubonic plague,

but some studies suggest that it instead conferred protection against smallpox.
 Systems with Sustained Oscillations and Singularities 399

genetic disease. Natural selection cannot select for exclusively heterozy-

gous individuals, because even if one started with such a population,
half of their offspring would be homozygotes of one or the other type.
Overdominance instead results in an asymptotic equilibrium distribution
of all three genotypes, and analysis shows (Exercise 6) that this equi-
librium is the Hardy-Weinberg distribution: as t → ∞, P → P ∗ , D =
P 2 → P ∗2 , H = P − D → P ∗ (1 − P ∗ ), R = 1 + D − 2P → (1 − P ∗ )2 .
IV If ∆D < 0, ∆R < 0 then both homozygous genotypes AA, aa have lower
death rate (greater fitness) than the heterozygous genotype Aa. This sit-
uation is referred to as underdominance; the selection against heterozy-
gotes promotes survival of homozygotes only, but in eliminating het-
erozygotes, only one type of homozygote may survive (otherwise there
will continue to be heterozygotes under random mating). Which ho-
mozygote (and allele) survive depends on both the relative advantages
−∆D , −∆R and the initial frequency of each type. Model analysis (Ex-
ercise 7) shows that P ∗ is an unstable equilibrium here which separates
the interval (0,1) into two basins of attraction (for P = 0 (A) and P = 1
(a)): if the initial frequency of the A allele is less than P ∗ , it will disap-
pear, replaced by a, but if its initial frequency is greater than P ∗ , then
it will become fixated and the recessive allele a will gradually disappear
from the population.
An example of underdominance occurs in the African butterfly Pseu-
dacraea eurytus, which employs several different forms of mimicry in
order to avoid predators (see Figure 7.23). Two alleles each duplicate
the appearance of a different local species of butterfly, one orange and
one blue, which are toxic to predators (so predators leave them alone).
However, heterozygotes have an intermediate appearance which resem-
bles nothing in particular, and thus are subject to predation. A less
understood example of underdominance is the interaction of two alleles
called HLA-DR3 and HLA-DR4 which increase a human’s risk of devel-
oping type 1 diabetes, especially in heterozygous DR3/DR4 individuals.
Despite underdominance promoting elimination of one allele (histori-
cally, diabetes was nearly universally fatal prior to the development of
insulin treatments), both persist in some populations, but they are not
the only alleles for the given gene.
In the first two cases the fittest genotype, a homozygote, is the only one that
survives long-term. However, this is not true in the other cases. Genotype-
dependent fitness (in particular, death) therefore clearly affects long-term
gene frequencies, with both allele and genotype frequencies time-varying, and
asymptotic to Hardy-Weinberg proportions only in the case of overdominance.
In many singular perturbation problems which model problems of viscous
fluid flow, the boundary layer (inner solution, associated with fast time) has
more significance. In biological models, the quasi-steady-state solution (outer
solution, associated with slow time) often has more significance. The situation
400 Dynamical Systems for Biological Modeling: An Introduction

∆
R
Case I Case III
D fittest H fittest
A dominates overdominance
0 ∆D
Case IV Case II
H least fit R fittest
underdominance a dominates

FIGURE 7.22: The slow dynamics of equation (7.56) are determined by the
signs of the differential mortality rates ∆D and ∆R .

Sketches by W.C. Hewitson

FIGURE 7.23: Three variants of the African mimic butterfly Pseudacraea

eurytus, called the False Wanderer, as drawn by William Chapman Hewitson
in Illustrations of new species of exotic butterflies: selected chiefly from the
collections of W. Wilson Saunders and William C. Hewitson, Volume III.
John van Voorst, London, 1868 (where it is referred to as Diadema eurytus).

described in this section is one where both the short-term and long-term
behavior are of interest.

Exercises

1. Prove that the differential equation (7.45) has a globally asymptotically

stable equilibrium which is positive if Λ′ (0) > d and zero otherwise, as
follows:
(a) Suppose that Λ′ (0) > d. Explain why (7.40) implies that the graph
of y = Λ(x) will nevertheless eventually cross the line y = d · x
(from above to below). [Recall Λ(0) = 0.]
(b) Suppose that for some x0 Λ(x0 ) < d·x0 . Explain why (7.40) implies
that the graphs of y = Λ(x) and y = d · x cannot intersect for
x > x0 (this covers both the case Λ′ (0) < d and the case where
 Systems with Sustained Oscillations and Singularities 401

Λ′ (0) > d but the graph of y = Λ(x) has already crossed below the
line y = d · x).
(c) The two previous parts prove that a unique positive equilibrium
exists iff Λ′ (0) > d. The extinction equilibrium N ∗ = 0 of (7.45)
always exists, by inspection. Use linearization to determine stabil-
ity conditions for each equilibrium and complete the proof of the
desired result.
2. In this section we analyzed the stability of the equilibrium (K, P02 ) of sys-
tem (7.45),(7.46). Although the right-hand side of (7.46) is undefined
at N = 0, we can extend the system to this point by observing that
limN →0 Λ(N )/N = Λ′ (0) from the definition of the derivative, and sup-
posing that Λ′ (0) is positive (and finite). Then let D′ = h(N )(P 2 − D),
where h(N ) = Λ(N N
)
if N > 0 and h(0) = Λ′ (0). Then (0, P02 ) is an
equilibrium of this system. Determine the condition for its stability by
taking the limit of the expressions related to the Jacobian in the text,
as N → 0. How does this fit with the stability of (K, P02 )?
3. Derive the differential equation (7.56) from (7.54) using the change of
independent variable s = ǫt and the slow-time equilibrium condition
D = P 2.
4. Complete the analysis of Case I for equation (7.56), in which ∆D <
0, ∆R > 0. You will need to consider two subcases, depending on the
sign of ∆D + ∆R , but the result should be the same for both.
5. Complete the analysis of Case II for equation (7.56), in which ∆D >
0, ∆R < 0.
6. Complete the analysis of Case III for equation (7.56), in which ∆D >
0, ∆R > 0.
7. Complete the analysis of Case IV for equation (7.56), in which ∆D <
0, ∆R < 0.
8. Determine the asymptotic behavior of the system (7.56) in the case
∆D = 0, and interpret the results in terms of the genotype distribution.
9. Determine the asymptotic behavior of the system (7.56) in the case
∆R = 0, and interpret the results in terms of the genotype distribution.
10. * Rewrite system (7.48) to incorporate genotype-dependent reproduc-
tion (while maintaining genotype-independent mating preferences and
death): Let AA homozygous individuals have 1 + ǫ∆D times as many
offspring as heterozygous individuals, and aa homozygous individuals
1 + ǫ∆R times as many. Derive new differential equations for N, D, P .
402 Dynamical Systems for Biological Modeling: An Introduction

7.5 Second-order differential equations: Acceleration

7.5.1 The harmonic oscillator
The simple harmonic oscillator, described by the second-order differential
equation
y ′′ + ω 2 y = 0,
is an elementary description of motion under a restoring force proportional
to displacement. It arises in many physical situations such as mechanical
springs,12 and is a prototypical example of periodic motion described by its
general solution
y = c1 cos ωt + c2 sin ωt.
Every second-order differential equation may be viewed as a system of two
first-order differential equations (as explained immediately below), and thus
the techniques we have developed to study first-order equations, including
the machinery (singular perturbations) we have developed in this chapter, are
also applicable to second-order equations. In fact, as the first-order system
which corresponds to a second-order equation has a particular form, some
simplifications are possible for second-order equations.
The general second-order autonomous differential equation
y ′′ = F (y, y ′ ), (7.57)
with F a function of two variables, may be converted to a system by defining
z = y ′ . Then y ′′ = z ′ = F (y, z), and the equation (7.57) is converted to the
system
y ′ = z, (7.58)
z ′ = F (y, z).
It is easy to verify that if (y(t), z(t)) is a solution of (7.58), then y(t) is a
solution of (7.57), and vice versa: if y(t) is a solution of (7.57), then (y(t), y ′ (t))
is a solution of (7.58). Thus we can switch between the second-order equation
(7.57) and the system (7.58) at will, and we could develop all the theory of
second-order equations as a special case of the theory for first-order systems.
We shall consider below the simplest special case, the linear homogeneous
equation with constant coefficients.

The general second-order linear homogeneous differential equation with

constant coefficients has the form
y ′′ + py ′ + qy = 0, (7.59)
12 Consider, for example, Hooke’s Law F = −kx in conjunction with Newton’s second law

F = ma = mx′′ , which produce the equation mx′′ = −kx equivalent to the equation given
above.
 Systems with Sustained Oscillations and Singularities 403

where p and q are constants. To transform this into a system, we let y ′ = z,

so that y ′′ = z ′ = −py ′ − qy = −pz − qy. Thus the corresponding system is

y ′ = z, (7.60)
z ′ = −qy − pz,

with coefficient matrix  

0 1
.
−q −p
The trace of this matrix is −p, and the determinant of this matrix is q. To
solve the system (7.60), we write its characteristic equation, which takes the
form
λ2 + pλ + q = 0. (7.61)
In contrast to the solution of systems described in Section 5.3, now we need
only the solution for y; z may be obtained from y by differentiation. Thus
(paralleling the discussion in Section 5.3) if the characteristic equation (7.61)
has two distinct roots λ1 and λ2 , we obtain the solutions y = eλ1 t and eλ2 t of
(7.59), and the general solution is y = K1 eλ1 t + K2 eλ2 t . If the characteristic
equation (7.61) has a double root λ1 , the general solution is y = K1 eλ1 t +
K2 teλ1 t . If the characteristic equation (7.61) has complex conjugate roots
α ± iβ, the general solution of (7.59) is y = eαt (K1 cos βt + K2 sin βt). In
each case, if a solution satisfying prescribed initial conditions is sought, the
constants K1 and K2 may be determined from the initial conditions. For
systems, it is appropriate to prescribe initial values for y and z; for a second-
order equation the analogue is to prescribe initial values y(0) and y ′ (0).

Example 1.
Find the general solution of the differential equation y ′′ + 4y ′ + 3y = 0, and
the particular solution satisfying the initial conditions y(0) = 1, y ′ (0) = 0.
Solution: The characteristic equation is λ2 +4λ+3 = 0, with roots −3 and −1.
Thus the general solution of the differential equation is y = K1 e−3t + K2 e−t .
To satisfy the given initial conditions, we differentiate this general solution,
obtaining y ′ = −3K1 e−3t − K2 e−t and then substitute t = 0, y = 1, y ′ = 0.
This gives the conditions 1 = K1 + K2 , 0 = −3K1 − K2 , whose solution is
K1 = − 21 , K2 = 23 . Thus the solution of the initial value problem is y =
− 21 e−3t + 32 e−t . 

Example 2.
Find the solution of the initial value problem

y ′′ + 2y ′ + y = 0, y(0) = 0, y ′ (0) = 1.

Solution: The characteristic equation λ2 +2λ+1 = 0 has a double root λ = −1.

Thus the general solution of the differential equation is y = K1 e−t + K2 te−t ,
404 Dynamical Systems for Biological Modeling: An Introduction

and y ′ = −K1 e−t + K2 e−t − K2 te−t = (K2 − K1 )e−t − K2 te−t . Substitution

of t = 0, y = 0, z = 1 gives K1 = 0, K2 = 1, and the solution is y = −te−t . 

Example 3.
Find the general solution of the differential equation y ′′ + ω 2 y = 0.
Solution: The characteristic equation λ2 + ω 2 = 0 has complex conjugate
roots ±iω, and thus the general solution of the differential equation is y =
K1 cos ωt + K2 sin ωt. 

Example 4.
We can consider the motion of the human diaphragm in breathing as that of a
massive spring being moved up and down by actuating forces: muscles which
pull it down during inhalation (and, to some extent, up during exhalation),
the natural springlike restoring force of the diaphragm itself, and a frictionlike
resistance proportional to (and opposite) the movement. If we define y to be
the downward displacement of the diaphragm (y = 0 at rest), we can write a
simple model for the net force F acting on the diaphragm:

F = f − ky − ry ′ ,

where f is the external muscular forcing, k is the proportionality constant

for the restoring force, and r is the proportionality constant for the resistive
force.
Using Newton’s second law, F = ma = my ′′ , write a second-order differen-
tial equation for y, the corresponding first-order system, and the characteristic
equation for the homogeneous (f = 0) system.
Solution: We substitute for F in the first equation given above, and move all
the terms involving y to the left side, to obtain the model

my ′′ + ry ′ + ky = f.

We can write this as the following system, defining z = y ′ :

y ′ = z,
k r f
z′ = − y− z+ .
m m m
Finally, the characteristic equation (which can be read from the single second-
order equation with f = 0, cf. (7.59), (7.61)) is mλ2 + rλ + k = 0.
Note, however, that because of the external forcing f , this model is not ho-
mogeneous, so finding the solution is somewhat more complicated than finding
the solution to a homogeneous equation, as done in the previous examples of
this section. 
 Systems with Sustained Oscillations and Singularities 405

7.5.2 The van der Pol oscillator

Nonlinear differential equations of second-order may also be reduced to
systems, and their equilibria and stability may be studied as for systems. In
the remainder of this section, we consider an equation known as the van der
Pol equation, which arose originally as a model for cardiac oscillations, but
has also had many applications in mechanical and electrical problems. The
Dutch scientist Balthazar van der Pol discovered in the 1920’s an electrical
circuit which produced sustained oscillations, and proposed it as a model for
the pacemaker of the heart.13 In this equation, y ′′ + µ(y 2 − 1)y ′ + y = 0, y
originally represented the level of electric potential, with µ a scaling factor
regulating the strength of the nonlinearity (note that µ = 0 gives us a special
case of the simple linear oscillator with which we began this section).

Example 5.
Write the van der Pol equation as a system.
Solution: We let z = y ′ , and then z ′ = y ′′ = −µ(y 2 −1)y ′ −y = −µ(y 2 −1)z −y,
and we have the system

y ′ = z, (7.62)
′ 2
z = −µ(y − 1)z − y. 

We note that because of the special form of the converted system (7.58),
any equilibrium of such a system satisfies z = 0, and thus may be found
simply by solving the equation F (y, 0) = 0 for y. The linearization of the
system (7.58) at an equilibrium (y∞ ,0) is the system

u′ = v,
v ′ = Fy (y∞ , 0)u + Fz (y∞ , 0)v,

which is equivalent to the second-order differential equation

u′′ = Fz (y∞ , 0)u′ + Fy (y∞ , 0)u,

so we could write down this linearization directly, without transforming to a

system and back.

Example 6.
Find the equilibria of the van der Pol equation, the linearization at each
equilibrium, and the stability of each equilibrium.
Solution: An equilibrium of (7.62) is found by solving z = 0, −µ(y 2 −1)z −y =
13 B. van der Pol and J. van der Mark, The heartbeat considered as a relaxation oscillation,

and an electrical model of the heart, Phil. Mag. 6 (1928), 763–775.

406 Dynamical Systems for Biological Modeling: An Introduction

0, and thus the only equilibrium is y = 0, z = 0. From above, the linearization

of the van der Pol equation about this equilibrium is

u′′ = µu′ + (−1)u,

or
u′′ − µu′ + u = 0,
since the linear approximation to y 2 − 1 at y =  0 isp−1. The roots of the
characteristic equation r2 − µr + 1 = 0 are r = 21 µ ± µ2 − 4 . Since µ > 0,
these roots are both positive if µ > 2, and complex conjugates with positive
real part if µ < 2; in either case, the origin is an unstable equilibrium. 

It is possible to show, although not without some difficulty, that every orbit
of the system (7.62) is bounded, and this implies by the Poincaré-Bendixson
theorem that there is a limit cycle. In fact, there is a unique limit cycle ap-
proached by every orbit, as suggested by the phase portrait in Figure 7.24
with µ = 1.5.

z(t)
2

–2 –1 1 2
y(t)

–2

–4

FIGURE 7.24: Phase portrait for system (7.62).

If we wish to treat the van der Pol equation as a second-order differential

equation y ′′ +µ(y 2 −1)y ′ +y = 0, we would find equilibria as constant solutions
(with y ′ = 0, y ′′ = 0), and thus y = 0 is the only equilibrium. The linearization
at this equilibrium is the second-order differential equation y ′′ − µy ′ + y = 0
identified above.
 Systems with Sustained Oscillations and Singularities 407

The van der Pol equation is usually treated as a system, but more infor-
mation can be derived by using a different system from (7.62) which allows
us to see more clearly the role of the parameter µ in determining the shape
of solutions (recall µ determines the strength of the nonlinear term). We first
define the function f (y) = y 2 − 1, so that the van der Pol equation may be
written as y ′′ + µf (y)y ′ + y = 0. We then define
Z y
y3
F (y) = f (u) du = − y, w = y ′ + µF (y).
0 3
Then by the Chain Rule

w′ = y ′′ + µF ′ (y)y ′ = −(µf (y)y ′ + y) + µf (y)y ′ = −y,

and we have a system

y ′ = w − µF (y),
w′ = −y.

It is also convenient to make a change of scale by letting w = µz, to give the

new system

y ′ = µ(z − F (y)), (7.63)

1
z ′ = − y.
µ
The advantage of the form (7.63), discussed in more detail below, is that for
large µ the two variables operate on quite different time scales if µ is large.
It is easy to see that the only equilibrium of (7.63) is (0,0). Since both
(7.62) and (7.63) are systems equivalent to the van der Pol equation, we al-
ready know that the origin must be unstable for (7.63) since we have shown
that it is unstable for (7.62). It is possible to show that all orbits of (7.63)
are bounded, so that we may apply the Poincaré-Bendixson theorem. The
Poincaré-Bendixson theorem gives the existence of a limit cycle. A more re-
fined argument shows that this limit cycle is unique, and is approached by
every orbit except the constant orbit y = 0, z = 0.
If µ is large, we can give a more precise description of the orbit. The y-
nullcline is the cubic curve z = F (y), and the z-nullcline is the axis y = 0.
Except near the y-nullcline, y ′ is large and z ′ is small, and thus the motion is
essentially horizontal in the y-z plane (i.e., in the y direction). Above z = F (y),
the trajectory moves to the right. When it reaches the y-nullcline (and crosses
it vertically), y ′ and z ′ are comparable in magnitude, and the orbit follows the
nullcline to its minimum at (1, − 23 ). Then it goes horizontally to the left until
it reaches the nullcline at (−2, − 32 ), after which it follows the nullcline up to
its maximum at (−1, − 32 ), and then repeats the cycle. Figure 7.25 shows this
orbit, traversed clockwise, superimposed on the y-nullcline.
For large µ, the oscillations consist of time intervals in which y increases
408 Dynamical Systems for Biological Modeling: An Introduction

y
z 1
1.5

1.0

10 20 30 40 50
0.5 t

y
-2 -1 1 2
–1
-0.5

-1.0

–2
-1.5

FIGURE 7.25: Representative solu- FIGURE 7.26: A solution of system

tion (in gray) for system (7.63) begin- (7.63) with µ = 8.
ning at (0,1) for µ = 5, with y-nullcline
(in black) superimposed.

or decreases rapidly (primarily horizontal movement in Figure 7.25), alternat-

ing with time intervals in which y remains almost constant (primarily vertical
movement in Figure 7.25). Such oscillations are called relaxation oscillations;
Figure 7.26 displays a solution with µ = 8. Such oscillations, with relatively
slow changes in y interrupted by sudden changes, are similar to cardiac elec-
trical activity, and modifications of the van der Pol equation have been used
as models for the study of oscillations in cardiac tissues.

7.5.3 A model of oxygen diffusion in muscle fibers

The muscles of the body require a continuous supply of oxygen even when
at rest, due to the need to maintain certain conditions (chemical and elec-
trical) within the cells. Most muscles are formed of long, thin fibers, which
oxygen enters from the periphery (delivered via the blood system). Oxygen
and oxymyoglobin (oxygenated myoglobin) diffuse throughout the fibers to
transport oxygen to the cells in the (axial) centers of the fibers. We can make
a simple model to describe this process in the following way.
In reality, diffusion occurs in all three spatial dimensions within the muscle
fibers. However, to simplify our model, we will consider the fiber as a long,
thin cylinder of uniform radius a, with all diffusion occurring radially. That is,
we suppose that there is no significant axial component of diffusion (parallel
to the fiber axis), and that the concentrations of oxygen and oxymyoglobin
at any given point inside the fiber depend only on the distance r from the
center of the fiber (or, equivalently, from the boundary). In mathematical
terms, we use cylindrical coordinates (r, θ, z) to describe location inside the
muscle fiber, but assume that the diffusion of oxygen and oxymyoglobin is
 Systems with Sustained Oscillations and Singularities 409

independent of axial height z and angle θ (around the axis, see Figure 7.27).
Assuming these two symmetries allows us to consider only the matter of how
oxygen is transported from the boundary of the fiber (at r = a) inward to the
center (r = 0).

z
θ
r=0 r=a

Photo courtesy Howard Swatland

FIGURE 7.27: Viewing a sin-

FIGURE 7.28: A single muscle fiber (my-
gle muscle fiber in a bundle as
ofibre) is roughly cylindrical in shape. As
a cylinder, taking a cross-section
they are formed by the fusion of myoblasts,
fixes z, with r = 0 at the center
these long cells have many nuclei (visible as
(axis) of the fiber and r = a at
dark spots on the surface).
the edge.

What we need to do now is account for the radial diffusion of oxygen

for 0 ≤ r ≤ a. There are three factors which change the concentration of
oxygen: the uptake of oxygen by myoglobin to form oxymyoglobin, the ongoing
consumption of oxygen by the muscle cells, and the diffusion process. We will
suppose that the oxygenization of myoglobin takes place at a rate f , and that
oxygen is being consumed at a constant rate g (constant, at least, with respect
to r: let us consider a muscle either at rest or at a given level of activity).
This leaves the diffusion term. The function commonly used to describe the
diffusion of a substance is the Laplacian operator, which, when converted into
the cylindrical coordinates of our model, has radial component 1r dr d d·


r dr .
Once the concentration of oxygen within the muscle reaches a steady-state
distribution, these three factors (the first two negative, since oxygen is being
removed via them, and the third positive) balance out, so that their sum is
zero.
This gives us the model
 
1 d dy
K1 r −f −g =0 (7.64)
r dr dr
for the amount y of oxygen (O2 ) at a distance r from the center of the fiber,
410 Dynamical Systems for Biological Modeling: An Introduction

where the constant K1 describes the rate of diffusion. Since (under our as-
sumptions of symmetry) the diffusion at the center is even, the rate of change
of the concentration should level off as one approaches the center; this gives
us one boundary condition, dy dr (0) = 0. The concentration of oxygen entering
the fiber from outside (at r = a) is known, giving us our second boundary
condition, y(a) = Ya . (Since the diffusion begins outside the fiber boundaries,
we cannot know that dy dr (a) = 0.)
By completing the differentiation in (7.64) we can obtain a second-order
differential equation, which can be rewritten as a first-order system and ana-
lyzed using the methods illustrated in this section. Exercises 19–22 below out-
line some extensions and analysis of model (7.64). The simplest model gives
a quadratic distribution for y(r), reflecting the higher concentration closer to
the source (r = a). Mathematically the most interesting feature of the model
and its extensions is that the nature of the first boundary condition, specifying
dy/dr rather than r at 0, means there are no boundary layers for y: a slight
correction (corresponding to a fast-time or inner solution) is needed only for
dy/dr(0), not for y(0).
A more realistic model would consider the specific locations (in terms of
angle θ) of the surface capillaries from which diffusion originates, and might
further consider what happens when the muscle contracts, increasing the ra-
dius and changing the locations of the capillaries.

Exercises
In each of Exercises 1–12, find the general solution and the solution satisfying
the initial conditions y(0) = 1, y ′ (0) = −1.

1. y ′′ = 0 7. y ′′ + 5y ′ + 10y = 0
2. y ′′ − 9y = 0 8. 4y ′′ + 4y ′ + 13y = 0
3. y ′′ − 5y ′ + 6y = 0 9. y ′′ − 4y ′ + 13y = 0
4. y ′′ + 9y = 0 10. y ′′ − 4y ′ + y = 0
5. y ′′ + 10y ′ + 25y = 0 11. y ′′ − 2y ′ + 2y = 0
6. 4y ′′ − y = 0 12. ǫy ′′ + 2y ′ + y = 0 (0< ǫ < 1)

13. Find all equilibria of the differential equation y ′′ + ay ′ + by + y 2 = 0,

where a and b are real constants and b 6= 0.
In each of Exercises 14–17 find all equilibria of the given differential equation,
and the linearization at each equilibrium.
14. y ′′ + y(1 − y 2 ) = 0
15. y ′′ + y ′ + y 3 = 0
 Systems with Sustained Oscillations and Singularities 411

16. y ′′ + g(y) = 0, where g(y) is a smooth function with yg(y) > 0 for y 6= 0.
17. y ′′ + f (y)y ′ + g(y) = 0, where f (y) is a smooth function with f (y) > 0
for all y and g(y) is a smooth function with yg(y) > 0 for y 6= 0.
18. * To illustrate the complexity of the structure when the origin is not an
isolated equilibrium, consider the differential equation y ′′ + y ′ = 0.
(a) Show that there is a line of equilibria.
(b) Sketch the phase portrait in the y-y ′ phase plane.
19. In the model (7.64), treating the conversion rate f as a constant, com-
plete the differentiation in the above model, and write it first as a second-
order differential equation, and then as a system of two first-order equa-
tions. Can you solve the resulting equation for y ′ ?
20. * Given that oxymyoglobin is not consumed within the muscle fiber (it
acts to transport the oxygen), but that the oxygenization process men-
tioned above increases the amount of oxymyoglobin present, and also
that the oxymyoglobin does not leave the fiber (i.e., diffusion is limited
to 0 ≤ r ≤ a), write an equation modeling the radial diffusion of oxymyo-
globin in the muscle fiber, analogous to the one given above for oxygen,
using w(r) to represent the [radial] concentration of oxymyoglobin.
21. * In fact, f is not a constant, but depends on the amounts of reactants
(oxygen, myoglobin and oxymyoglobin) present, as well as on some con-
stants of proportionality. Let us take f = K3 xy − K4 w, where w is the
concentration of oxymyoglobin, x and y are the concentrations of myo-
globin and oxygen, respectively, and K3 and K4 are the constants of
proportionality. Rewrite the second-order equations for y and w using
this expression for f , and transform them into a system of four first-order
differential equations.
22. * Approximate values for the parameters in this model are as follows:
a = 2.5 × 10−3 cm, Ya = 3.5 × 10−8 mol/cm3 , g = 5 × 10−8 mol/cm3 sec,
K1 = 10−5 cm2 /sec, K2 = 5 × 10−7 cm2 /sec (diffusion rate constant
for oxymyoglobin), K3 = 2.4 × 1010 cm3 /mol sec, K4 = 65/sec, x =
2.8 × 10−7 mol/cm3 (the concentration of myoglobin in muscle fibers).14
Use a computer to find the distributions the two models predict for
oxygen and myoglobin as functions of r.

14 J. Keener and J. Sneyd, Mathematical Physiology, Springer-Verlag, New York (1998),

p. 42.
This page intentionally left blank
 Part III

Appendices

413
This page intentionally left blank
Appendix A
An Introduction to the Use of
MapleTM

Maple is a computer algebra system which can assist you in solving math-
ematical problems which would be quite inaccessible without its help. This
introduction is not intended to make you an expert on Maple but it should be
a reasonably self-contained start which will enable you to handle many prob-
lems in the analysis of ordinary differential equations and difference equations
which are found in this book. Maple is intended as an aid, not as a replacement
for thinking. Ideally, you will begin by working on a problem to understand
what is required and will turn to Maple to provide graphic output where ap-
propriate. While Maple has considerable capacity for calculation (good enough
to pass most calculus courses), this is an aspect which we shall try not to em-
phasize. Our use of Maple will be concentrated on
(i) plotting graphs of functions or pairs of functions and thus finding inter-
sections of two graphs,
(ii) graphical solution of ordinary differential equations, including first-order
equations and systems of first-order differential equations,
(iii) difference equations, including the cobweb method for first-order differ-
ence equations and the graphing of solutions of difference equations of
first or higher order by iterative calculations.
Maple can be run on many different kinds of computers, including Win-
dows, Macintosh, and many UNIX systems. On many university campuses it is
installed on computers in student labs. In addition, for students planning on a
career which will use mathematics, the “academic version” of Maple available
for sale at a moderate price in many universities is an excellent investment.
Maple can be used for many “word-processor” functions to dress up your
worksheets and convert them into more formal reports. You can toggle between
Maple ([>) and text (T) input on the toolbar in order to add text. Inside a
Maple computation you can add comments; any Maple input beginning with
# is treated as a comment and is ignored as far as calculation is concerned.
Note that every time you hit the “Enter” key you begin a new input and if
you want a comment to extend for more than a paragraph you need to begin
each paragraph with “#”.

415
416 Dynamical Systems for Biological Modeling: An Introduction

Here are a few general instructions for using Maple. They are not meant
to be complete, and a manual on the use of Maple or the online Help available
in Maple contains much more useful information.
Every input for which you want Maple to do something must end with a
semicolon. If you want Maple to act but not display the output you would use
a colon.
The command restart at the beginning of a worksheet tells Maple to
forget any previous instructions. If you have been working on one problem
and defined a symbol and then begin working on another problem which uses
the same symbol, this instruction will avoid confusion.
The command with tells Maple to read in a package containing some
additional definitions. For example, when we solve differential equations we
will always use the package DEtools and thus will begin the worksheet with
the command with(DEtools):. If we use a semicolon rather than a colon at
the end of the command, Maple will print out a list of the commands included
in the package.
The syntax for defining a quantity for later use is : =. Thus the command
a: = 1 assigns the value 1 to the quantity a. This syntax is also used to define
functions.

A.1 Plotting graphs of functions

The basic command to plot the graph of a given function on a given interval
is
> plot(f(x),x=a..b);
Options include specification of the range of the ordinate, specification of
an infinite interval, and plotting of more than one function on the same set of
axes, using the formats
> plot(f(x),x=a..b,y=c..d);
> plot(f(x),x=0..infinity);
> plot([f(x),g(x)],x=a..b);
The command to plot a graph given parametrically is
> plot([f(t),g(t),t=a..b]);
The option scaling=constrained may be used to make the units on the
two axes the same. This may give a more accurate geometric picture but may
hide important information if the units are on a different scale. Figures drawn
with Maple can be exported as .eps files.
We give a few examples to illustrate.
> restart:
> f(x):=x*exp(-x): g(x):=0.2*x:
> plot(f(x),x=0..5);
(see Figure A.1)
 An Introduction to the Use of MapleTM 417
1
0.36
0.34
0.32
0.3 0.8
0.28
0.26
0.24
0.22 0.6
0.2
0.18
0.16
0.4
0.14
0.12
0.1
0.08 0.2
0.06
0.04
0.02
0 1 2 3 4 5 0 1 2 3 4 5
x x

FIGURE A.1: Graph of f (x) = FIGURE A.2: Graph of f (x) =

xe−x . xe−x , g(x) = 0.2x.
0.5

0.4

0.3

0.2

0.1
0.5
y

0 0 1 2 3 4 5
1 2 x 3 4 5 x

FIGURE A.3: Graph of f (x) = FIGURE A.4: Graph of f (x) =

xe−x , g(x) = 0.2x,y-axis specified. xe−x , g(x) = 0.2x, scaling con-
strained.

> plot([f(x),g(x)],x=0..5);
(see Figure A.2)
> plot([f(x),g(x)],x=0..5,y=0..0.5);
(see Figure A.3)
> plot([f(x),g(x)],x=0..5,y=0..0.5,scaling=constrained):
(see Figure A.4)

A.2 Graphical solution of first-order differential equa-

tions
The Maple package DEtools includes commands which draw the direction
field for a first-order differential equation, the direction field together with
solution curves corresponding to prescribed initial values, or solution curves
418 Dynamical Systems for Biological Modeling: An Introduction

corresponding to prescribed initial conditions without the direction field. Each

of these can be carried out using the command DEplot. Our description here
will be restricted to its use for first-order differential equations but the com-
mand may be used for systems of differential equations and equations of order
higher than one. We shall describe its use for systems later.
The command DEplot is followed by up to six arguments, namely the
differential equation, the variable to be plotted, the range of the independent
variable, the initial conditions (if solution curves are to be drawn but not if
only the direction field is needed), the range of the dependent variable, and
(possibly) options such as the arrows to be drawn.
A differential equation of the form y ′ = f (t, y) may be specified by
diff(y(t),t)=f(t,y). There are other possible forms which may be used.
However, a differential equation like y ′ + y = g(t) may not be specified as
diff(y(t),t)+y=g(t); it should be written diff(y(t),t)=-y+g(t).
The variable to be plotted is given as y(t) and the range of the variables
to be plotted is given in the form t=a..b, or y=c..d. Initial conditions are
specified as lists given in the form {[y(0)=a], [y(0) = b],...}.
The default is to draw thin arrows. To draw solution curves without the
direction field using DEplot, it is necessary to specify arrows = none. Other
possible specifications for the kind of arrows include small, medium, large,
and line. In the examples below, colors and linecolors have been set to
black except where there are both arrows and solutions. This makes the print-
out clearer but is not necessary. Incidentally, as Maple is a system developed in
Canada, it understands both the spellings “color” and “colour”. We illustrate
these procedures with the differential equation y ′ = y(1 − y) on the interval [0,
4], using the initial conditions y(0) = −0.5, y(0) = 0, y(0) = 0.2, y(0) = 0.4,
y(0) = 0.6, y(0) = 0.8, y(0) = 1.0, y(0) = 1.2, y(0) = 1.4.

> restart: with(DEtools):

To draw a direction field:
> DEplot(diff(y(t),t)=y*(1-y), y(t), t=0..4, y=-1.0..1.4,
arrows=small,color=black);
(see Figure A.5)
To draw solution curves:
> DEplot(diff(y(t),t)=y*(1-y), y(t), t =0..4, {[y(0)=-0.5],
[y(0)=0], [y(0)=0.2], [y(0)=0.4], [y(0)=0.6], [y(0)=0.8],
[y(0)=1.0], [y(0)=1.2], [y(0)=1.4]}, y=-1.0..1.4, arrows=none,
thickness=0, linecolor=black);
(see Figure A.6)
To draw a direction field with solution curves:
> DEplot(diff(y(t),t)=y*(1-y), y(t), t =0..4, {[y(0)=-0.5],
[y(0)=0], [y(0)=0.2], [y(0)=0.4], [y(0)=0.6], [y(0)=0.8],
[y(0)=1.0], [y(0)=1.2], [y(0)=1.4]},y=-1.0..1.4, arrows=line,
thickness=0, linecolor=black);
(see Figure A.7)
 An Introduction to the Use of MapleTM 419

1.4 1.4
1.2 1.2
1 1

y(t)0.8 y(t)0.8
0.6 0.6
0.4 0.4
0.2 0.2

0 1 2t 3 4 0 1 2t 3 4
–0.2 –0.2
–0.4 –0.4
–0.6 –0.6
–0.8 –0.8
–1 –1

FIGURE A.5: Direction field FIGURE A.6: Solution curves

y ′ = y(1 − y). y ′ = y(1 − y).

1.4
1.2
1

y(t)0.8
0.6
0.4
0.2

0 1 2t 3 4
–0.2
–0.4
–0.6
–0.8
–1

FIGURE A.7: Direction field with solution curves, y ′ = y(1 − y).

A.3 Graphical solution of systems of differential equa-

tions
The solution of an autonomous system of two first-order differential equa-
tions is similar to the procedure used for a first-order differential equation. The
default is to plot the phase portrait of the two dependent variables but the
“scene” option allows plotting either of the dependent variables as a function
of the independent variables. The default is to show arrows in the direction
field. One example should suffice to give the idea.
> restart: with (DEtools):
DEplot({diff(x(t),t)=1*x*(1-x/40)-x*y/(x+10),
diff(y(t),t)=1*y*10*(x-20)/(30*(x+10))-0.4},
{x(t),y(t)},t=0..25,[[x(0)=50,y(0)=10],[x(0)=50,y(0)=20],
[x(0)=50,y(0)=30],[x(0)=50,y(0)=15],[x(0)=50,y(0)=5]],
420 Dynamical Systems for Biological Modeling: An Introduction

x=0..50,y=0..40,scene=[x,y],stepsize=0.05,linecolour=black,
thickness=0,arrows=none);
(see Figure A.8)

40

30

y
20

10

0 10 20 30 40 50
x

FIGURE A.8: Phase portrait.

> DEplot({diff(x(t),t)=1*x*(1-x/40)
-x*y/(x+10),diff(y(t),t)=1*y*10*(x-20)/(30*(x+10))-0.4},
{x(t),y(t)},t=0..25,[[x(0)=50,y(0)=10],[x(0)=50,y(0)=20],
[x(0)=50,y(0)=30],[x(0)=50,y(0)=15],[x(0)=50,y(0)=5]],
x=0..50,y=0..40,scene=[t,x],stepsize=0.05,linecolour=black,
thickness=0,arrows=none);
(see Figure A.9)
> DEplot({diff(x(t),t)=1*x*(1-x/40)
-x*y/(x+10),diff(y(t),t)=1*y*10*(x-20)/(30*(x+10))0.4},
{x(t),y(t)},t=0..25,[[x(0)=50,y(0)=10],[x(0)=50,y(0)=20],
[x(0)=50,y(0)=30],[x(0)=50,y(0)=15],[x(0)=50,y(0)=5]],
x=0..50,y=0..40,scene=[t,y],stepsize=0.05,linecolour=black,
thickness=0,arrows=none);
(see Figure A.10)

A.4 The cobwebbing method for graphical solution of

first-order difference equations
In Section 2.2 we described the cobwebbing method for solving first-order
difference equations. This graphical method requires repeated calculation of
functional values and is ideally suited to execution by means of a computer
program. Here we define a program called “cobweb” which needs input of
a specified function (func), an initial value (x0), a number of iterations to
 An Introduction to the Use of MapleTM 421

50 40

40
30

30
x y
20

20

10
10

0 2 4 6 8 10 12 t 14 16 18 20 22 24 26 0 2 4 6 8 10 12 t 14 16 18 20 22 24 26

FIGURE A.9: x as function of t. FIGURE A.10: y as function of t.

be performed (numb), a minimum value of the independent variable (xmin),

and a maximum value of the independent variable (xmax). When these are
defined and entered into the program, the cobweb method will run and plot
the output.
> restart: with(plots): with(plottools):
> cobweb :=proc(func, x0, numb, xmin, xmax)
> local curve,diagonal, oldx, newx, lines, i, l1, l2:
> curve := plot(func(x),x=xmin..xmax):
> diagonal := line([xmin,xmin],[xmax,xmax]):
> oldx := x0: newx := func(oldx):
> lines := [line([oldx,0],[oldx, newx])]:
> for i from 0 to numb do;
> l1 := line([oldx, newx],[newx,newx]):
> oldx := newx: newx := func(oldx):
> l2 := line([oldx, oldx],[oldx,newx]):
> lines := [op(lines),l1,l2]:
> od:
> display(lines,curve,diagonal,scaling=
constrained, thickness=2);
> end;
Here is an example using the logistic difference equation. We begin by
defining the function and its parameter.
> r:=2.9:
> logistic := x -> r*x*(1-x):
> cobweb(logistic,.45,10,0,1);
(see Figure A.11)
We suggest some additional examples using the logistic function all on the
interval [0,1] but with different parameter values, initial points, and number
of steps, namely:
1. r = 3.4, x0= 0.45, numb = 20
422 Dynamical Systems for Biological Modeling: An Introduction
1

0.8

0.6

0.4

0.2

0 0.2 0.4 0.6 0.8 1

x

FIGURE A.11: Cobwebbing for xn+1 = 3.62xn (1 − xn ).

2. r = 3.5, x0 = 0.385, numb = 20

3. r = 3.58, x0 = 0.48, numb = 20
4. r = 3.8, x0 = 0.3, numb = 20
5. r = 3.8, x0 = 0.3, numb = 40
6. r = 3.8, x0 = 0.3, numb = 100
7. r = 3.8, x0 = 0.3, numb = 200.

A.5 Solution of difference equations and systems of dif-

ference equations
Here is a simple program for the iterative calculations required in solving a
difference equation and plotting the resulting solution as a line segment graph.
> restart: with(plots):
> logistic:=y->r*y*(1-y): r:=3.62: y(0):=0.5:
> for k from 0 to 25 do
> y(k+1):=logistic(y(k)):
> od:
> Q:=[seq([k,y(k)], k=0..25)]: > pointplot(Q,style=line);
(see Figure A.12)
The program is easily adapted to systems of difference equations, as we
illustrate with an example from Section 2.8:
> restart: with(plots):
> r:=1: b:=0.5: p:=0.3:
 An Introduction to the Use of MapleTM 423
0.9

0.8

0.7

0.6

0.5

0.4

0 2 4 6 8 10 12 14 16 18 20 22 24

FIGURE A.12: Solution plot, xn+1 = 3.62xn (1 − xn ).

> y(0):=1: z(0):=1:

> for k from 0 to 35 do
> y(k+1):=r*y(k)*exp(-b*z(k))+p*y(k):
> z(k+1):=r*y(k)*(1-exp(-b*z(k))):
> od:
> Q:=seq([k,y(k)],k=0..35):
> R:=seq([k,z(k)],k=0..35):
> pointplot(Q,style=line);
> pointplot(R,style=line);

(see Figures A.13 and A.14)

10 8

8
6

0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32

FIGURE A.13: Solution plot for FIGURE A.14: Solution plot for
yn in system. zn in system.
424 Dynamical Systems for Biological Modeling: An Introduction

A.6 A bifurcation program

In Section 2.6, we described the process of bifurcation for the logistic differ-
ence equation and gave a bifurcation diagram. Here, we give a Maple program
to generate a bifurcation diagram. A bifurcation diagram shows the periodic
orbits for a range of values of the parameter r. The idea behind the program
is to iterate the logistic function for a given value of r and to plot the values
obtained after enough iterations for the orbit to have been approached. Then,
for example, if two values are obtained there is a solution of period 2. The
program does this for a range of values of r and plots x values as a function
of r.
1

0.8

0.6

0.4

0.2

0 2 2.2 2.4 2.6 2.8 3 3.2 3.4 3.6 3.8 4

FIGURE A.15: Bifurcation diagram for the logistic difference equation.

> restart: with(plots):

> x(1):=0.5: r:=200:
> for r from 200 to 400 do
> for j from 1 to 100 do
> x(j+1):=r*x(j)*(1-x(j))/100:
> L(r):=seq([r/100,x(j)],j=20..100):
> od:
> od:
> for r from 200 to 400 do
> A(r):=pointplot (L(r)):
> od:
> display (seq(A(r),r=200..400),symbol=point,view=[2..4,0..1]);
(see Figure A.15)
Appendix B
Taylor’s Theorem and Linearization

The linearization of a dynamical, either discrete- or continuous-time, system

at an equilibrium is an essential tool in the study of the equilibrium’s stability.
For one-dimensional systems the stability of an equilibrium can be determined
without recourse to the linearization. In fact, we did not use the linearization
in Section 4.1 in discussing stability for first-order differential equations, and
one can give an alternative to the approach used in Section 2.3 for stability
of equilibria of first-order difference equations.
If y∞ is an equilibrium of a first-order difference equation yk+1 = g(yk ) or a
first-order differential equation y ′ = g(y), we make the change of variable u =
y −y∞ , so that u represents deviation from the equilibrium. Then substitution
into the difference or differential equation produces an expression g(y∞ + u).
We wish to expand this expression in powers of u, because near the equilibrium
u is small and higher powers of u are small compared to lower powers. Taylor’s
theorem provides such an expansion together with an estimate of how close
the expansion is to the function being expanded. We need only the special
case of the theorem covering expansions limited to first-order terms, that is,
to approximate a function near a base point y∞ by a linear expression in u.
We assume that g is a sufficiently smooth function, specifically that it is
continuous with continuous first derivative and a bounded second derivative
in an interval I with y∞ in its interior. Then for any u such that y∞ + u is
also in I, we may write g(y∞ + u) as the integral of its derivative,
Z y∞ +u
g(y∞ + u) = g(y∞ ) + g ′ (t)dt. (B.1)
y∞

′
Likewise, we may write g (t) as the integral of its derivative,
Z t
g ′ (t) = g ′ (y∞ ) + g ′′ (s)ds. (B.2)
y∞

Substitution of (B.2) into (B.1) gives

Z y∞ +u Z y∞ +u Z t 
′ ′
g(y∞ + u) = g(y∞ ) + g (y)dt + g (s)ds dt
y∞ y∞ y∞
Z y∞ +u Z t 
′ ′
= g(y∞ ) + g (y∞ )u + g (s)ds dt. (B.3)
y∞ y∞

425
426 Dynamical Systems for Biological Modeling: An Introduction

If the second derivative g ′′ (s) is bounded below by m and above by M on the

Rt
interval I, then the integral y∞ g ′ (s)ds is between m(t − y∞ ) and M (t − y∞ )
for every t in the interval, and this implies that the iterated integral (which
is the error term in our expansion)
Z y∞ +u Z t 
′
g (s)ds dt
y∞ y∞
R y +u 2 R y +u 2
is between y∞∞ m(t − y∞ )dt = m u2 and y∞∞ M (t − y∞ )dt = M u2 . Now
we have a version of Taylor’s theorem which meets our needs.

TAYLOR’S THEOREM: Let g be a function which is continuous and

has a continuous first derivative and a bounded second derivative on an in-
terval I containing a point y∞ in its interior. Suppose the second derivative
is at least m and at most M in this interval. Then the linear Taylor approx-
imation to g(y∞ +u) is g(y∞ )+g ′ (y∞ )u and the error in this approximation
2 2
is between m u2 and M u2 .

2
The error term can be expressed as g ′′ (c) u2 , with c some (unknown) point
between y∞ and y∞ + u. For our purposes, the factor u2 , which for small u
makes the error term negligible compared to the linear term, is what matters.
There is a two-dimensional version of Taylor’s theorem, which is needed
for linearization of two-dimensional systems (Sections 2.8, 5.4). We wish to
expand a function of two variables, g(y∞ + u, z∞ + v) in powers of u and v
with coefficients which depend on the value of the function g and its partial
derivatives at the base point (y∞ , z∞ ). In order to reduce this problem to
Taylor’s theorem in one variable, we define a function of one variable t,

G(t) = g(y∞ + tu, z∞ + tv).

Then G(0) = g(y∞ , z∞ ) and G(1) = g(y∞ + u, z∞ + v). Taylor’s theorem for
one variable says that

G(1) = G(0) + G′ (0) + R (B.4)

where R is the error term. We may use the chain rule for partial derivatives
to calculate G′ (t) and G′′ (t):

G′ (t) = ugy (y∞ + tu, z∞ + tv) + vgz (y∞ + tu, z∞ + tv)

G′′ (t) = u2 gyy (y∞ +tu, z∞ +tv)+2uvgyz (y∞ +tu, z∞ +tv)+v 2 gzz (y∞ +tu, z∞ +tv)
Substitution of t = 0 gives

G′ (0) = ugy (y∞ , z∞ ) + vgz (y∞ , z∞ ).

The expressions for G(0), G(1), and G′ (0) substituted into (B.4) give

g(y∞ + u, z∞ + v) = g(y∞ , z∞ ) + ugy (y∞ , z∞ ) + vgz (y∞ , z∞ ) + R, (B.5)

and it is possible to check that the error term R is quadratic in u and v, that
is, it is a sum of terms which are multiples of u2 , uv, and v 2 . This is the
expression which is needed for linearization of a two-dimensional system at an
equilibrium (y∞ , z∞ ).

Appendix C
Location of Roots of Polynomial
Equations

We consider the quadratic polynomial equation

f (λ) = λ2 + a1 λ + a2 = 0. (C.1)

In Section 2.8, in order to determine stability of equilibria of systems of differ-

ence equations, we needed a criterion to tell when all roots of a characteristic
equation (C.1) satisfy |λ| < 1. In Section 5.3, in order to determine stability
of equilibria of systems of differential equations, we needed a criterion to tell
when all roots of a characteristic equation (C.1) satisfy Rλ < 0. In this ap-
pendix we shall establish both of these criteria. There are more complicated
criteria of both types for higher order polynomial equations but we shall not
consider these.
The equation (C.1) has two roots; if a21 − 4a2 ≥ 0 there are two real roots
λ1 and λ2 , and if a21 − 4a2 < 0 the roots are complex conjugates, α ± iβ. It is
useful to note that in either case the sum of the roots is −a1 and the product
of the roots is a2 .
Our first result, needed in the study of systems of differential equations, is
that the roots of (C.1) satisfy Rλ < 0 if and only if a1 > 0 and a2 > 0. To see
that these conditions are necessary we observe that if the roots are conjugate
complex their product a2 is automatically positive and since their sum is twice
their common real part a1 must be positive for these real parts to be negative.

427
428 Dynamical Systems for Biological Modeling: An Introduction

If the roots are real, in order for both to be negative their product must be
positive and their sum must be negative. On the other hand, the conditions
a1 > 0, a2 > 0 are sufficient. To see this, we note that if a21 − 4a2 ≥ 0, so that
the roots are real, a2 > 0 implies that they have the same sign and a1 > 0
implies that both are negative. If a21 − 4a2 < 0, so that the roots are complex
conjugates, a1 > 0 implies that they have negative real part. To sum up, we
have obtained the following result:

THEOREM: The roots of the quadratic equation λ2 + a1 λ + a2 = 0 have

negative real part if and only if a1 > 0 and a2 > 0.

Our other result, needed in the study of systems of difference equations,

is that the roots of (C.1) satisfy |λ| < 1. To obtain necessary conditions, we
note that the product of the roots must have absolute value less than 1, so
that |a2 | < 1. Also, we must have f (−1) > 0 since otherwise there would be
a root less than −1, and we must have f (1) > 0 since otherwise there would
be a root greater than 1. Thus we must have

f (−1) = 1 − a1 + a2 > 0, f (1) = 1 + a1 + a2 > 0

or −(a2 + 1) < a1 < (a2 + 1), and this is equivalent to |a1 | < a2 + 1. The
condition |a2 | < 1 may be written −1 < a2 < 1, or 0 < a2 + 1 < 2, and we
may combine our two necessary conditions into the double inequality

|a1 | < a2 + 1 < 2. (C.2)

These conditions are also sufficient. If (C.2) is satisfied, then f (−1) > 0
and f (1) > 0. If the roots are real, this implies that either both are less than
−1 or both between −1 and +1, or both greater than 1. However, if both
are less than −1 or greater than 1, their product is greater than 1, which
contradicts a2 < 1. Thus if the roots are real they must be between −1 and
1. If the roots are conjugate complex, their product a2 is the product of their
absolute values and thus less than 1. Since both roots have the same absolute
value, they must both have absolute value less than 1.
To sum up, we have obtained the following result:

THEOREM (Jury criterion): The roots of the quadratic equation λ2 +a1 λ+

a2 = 0 satisfy |λ| < 1 if and only if

|a1 | < a2 + 1 < 2.

Appendix D
Stability of Equilibrium of Difference
Equations

In Section 2.3 we stated a theorem on asymptotic stability of an equilibrium

of a difference equation as a consequence of the idea of linearization at an
equilibrium. For a first order difference equation it is possible to establish this
theorem directly, without using the linearization. The proof is an application
of the mean value theorem, which is presented in most calculus courses as
one of the most important results in calculus but then used mainly only to
establish precisely results which might appear to be obvious.
We are concerned with the asymptotic stability of an equilibrium y∞ of a
difference equation
yk+1 = g(yk ). (D.1)
The first result we must establish is that if a solution of (D.1) approaches a
limit as k → ∞, this limit must be an equilibrium of (D.1). We will not give
a precise proof of this fact, but the underlying idea is that if yk and yk+1 are
arbitrarily close to a limit y ∗ , then by the continuity of the function g, g(yk )
is very close to g(y ∗ ) and therefore (since yk+1 = g(yk ) is thus close to both
y ∗ and g(y ∗ )) we must have g(y ∗ ) = y ∗ .
The mean value theorem states that if a function f (x) is continuous and
has a continuous derivative in an interval I, then for any two points a and b
in this interval we have

f (b) − f (a) = (b − a)f ′ (c)

for some point c between a and b (actually, the requirement of continuity of

f ′ can be relaxed somewhat). A consequence of this is that if we know that
|f ′ (x)| < M for some number M and all points in the interval, then we have
the estimate
|f (b) − f (a)| < M |b − a| (D.2)
for all points a and b in the interval.
Now we are ready to prove our main result, the following stability of equi-
librium theorem of Section 2.3.

429
430 Dynamical Systems for Biological Modeling: An Introduction

STABILITY OF EQUILIBRIUM THEOREM: Let y∞ be an equi-

librium of the difference equation yk+1 = g(yk ). If |g ′ (y∞ )| < 1, the equi-
librium is asymptotically stable and if |g ′ (y∞ )| > 1, the equilibrium is
unstable.

Proof: Suppose y∞ is an equilibrium, so that y∞ = g(y∞ ) and let g ′ (y∞ ) = L,

with |L| < 1. We choose a number ρ with |L| ≤ ρ < 1, and then there is
an interval I centered at y∞ such that |g ′ (y)| < ρ for all points y in this
interval. We take the initial value y0 in this interval. Then since y1 = g(y0 )
and y∞ = g(y∞ ) we may apply (D.2) with the function g, the points y0 and
y∞ , and M = ρ, and we have

|y1 − y∞ | = |g(y0 ) − g(y∞ )| < ρ|y0 − y∞ |.

Since ρ < 1, y1 is closer to y∞ than y0 is, and therefore y1 is also in the

interval I. We may repeat the argument to give

|y2 − y∞ | = |g(y1 ) − g(y∞ )| < ρ|y1 − y∞ | < ρ2 |y0 − y∞ |

showing that y2 is also in the interval I and yet closer to y∞ than is y1 . In

the same way (or by induction), we may establish that

|yk − y∞ | < ρk |y0 − y∞ |

for k = 1, 2, . . .. This shows that each of the terms yk is in the interval I, that
the successive terms come closer to y∞ , and that yk approaches the limit y∞
as k → ∞. Thus every solution starting close enough to y∞ remains close to
y∞ and approaches y∞ as k → ∞, and this establishes asymptotic stability
of y∞ .
If |g ′ (y∞ )| > 1, we may pick a number ρ > 1 and an interval centered at
y∞ on which |g ′ (y)| > ρ and use the Mean Value Theorem in much the same
way but with the inequalities reversed to give

|yk − y∞ | > ρk |y0 − y∞ |.

From this we see that the terms yk move away from y∞ and thus that the
equilibrium y∞ is unstable. 

In the case of asymptotic stability, with ρ < 1, ρ gives a measure of the

rapidity of approach to the equilibrium. A small value of ρ means that the
equilibrium value is approached rapidly. An interesting application of this is
Newton’s method for approximating the solution of an equation f (y) = 0.
The method consists of choosing a starting value y0 sufficiently close to the
desired root and iterating according to the scheme
f (yk )
yk+1 = yk − .
f ′ (yk )
 Stability of Equilibrium of Difference Equations 431

Geometrically, this means drawing the tangent line to the curve z = f (y)
at the point y0 and taking y1 to be the point where this tangent line meets
the y-axis (i.e., taking the tangent line as an approximation to the function
near y0 and finding where this approximation is zero). This process is then
repeated.
In order to view Newton’s method as a stability of equilibrium question,
we define
f (y)
g(y) = y − ′ .
f (y)
Then an equilibrium y∞ of the difference equation yk+1 = g(yk ) is a solution
of the equation f (y) = 0 (provided f ′ (y) 6= 0). Since

[f ′ (y)]2 − f (y)f ′′ (y) f (y)f ′′ (y)

g ′ (y) = 1 − =
[f ′ (y)]2 [f ′ (y)]2

(and f (y∞ ) = 0) we have g ′ (y∞ ) = 0. Thus g ′ (y) is small on an interval close

to y∞ and the approach to a root by Newton’s method is very rapid once a
suitable starting approximation has been found.
This page intentionally left blank
Answers to Selected Exercises

Chapter 2
Section 2.1
2. a.  
µ µ
wn = w0 − Fn +
1−F 1−F
µ
b. If F < 1, then eventually wn → 1−F . This means that if the
weed is less fit than its competitor, then it will approach the given
equilibrium density thanks to the constant growth due to mutation.
c. If F > 1, meaning the weed is more fit than its competitor, then wn
grows without bound, i.e., weed density becomes arbitrarily high.
d. Unbounded growth in a single patch is unrealistic, so the model’s
operating range is limited (in w) when F > 1.
3. y1 = 3/16, y2 = 39/256, y3 = 8463/65536.
5. y1 = 1/2, y2 = 1/3, y3 = 1/4.
11. yk = (1.1)k .
13. yk = 1/2k .
15. yk = 1.
17. yk = (1/2)k − 1.
19. yk → 0.
21. yk → 5/4.

433
434 Answers to Selected Exercises

Section 2.2

FIGURE S.1: Solution of Exercise FIGURE S.2: Solution of Exercise

1 3

FIGURE S.3: Solution of Exercise 5

8. yk → 0.

Section 2.3
1. Equilibrium y = 2, unstable.
3. Equilibrium y = β/(1 − α), unstable.
5. Equilibrium y = 1 + r if r > 1/e, asymptotically stable for r < e.
10. a. (2.21) becomes
F pn + (1 − pn )
pn+1 = .
F pn + 2(1 − pn )
The p = 0 equilibrium vanishes, and the coexistence eqm becomes
1/(2 − F ). The equilibrium p = 1 is again stable for F > 1, and
1/(2 − F ) is for F < 1.
 Answers to Selected Exercises 435

b. Coexistence occurs only for F < 1, i.e., the a allele only survives if
the Aa zygote is more viable than the AA zygote.
c. We can rewrite the G = 0 model as pn+1 = K+1 K+2 , where K =
F pn /(1 − pn ) > 0. Since K+1
K+2 > 1
2 for K > 0, this guarantees A
alleles will always be in the majority. This makes sense because all
individuals who reproduce must have at least one A allele, so at
least half the alleles contributed to the next generation are As.
e.

13. a.
F αn + 14 (1 − αn − βn )
αn+1 =
F αn + (1 − αn − βn ) + Gβn
Gβn + 41 (1 − αn − βn )
βn+1 =
F αn + (1 − αn − βn ) + Gβn

b. We can rewrite the equations for αn+1 and βn+1 as

1 1
+K
4 4 +L
αn+1 = , βn+1 = ,
1+K+L 1+K +L
where K = F αn /(1−αn −βn ) > 0 and L = Gβn /(1−αn −βn ) > 0.
This makes (1 − αn − βn ) = 12 /(1 + K + L) ≤ 21 . This is biologically
sound under 100% like-with-like mating because two homozygotes
of the same type never reproduce heterozygotes, and only half of
the heterozygotes’ descendants will also be heterozygotes.
436 Answers to Selected Exercises

Section 2.4
1. Equilibrium y = 0 is asymptotically stable for all r.
3. r < e.
√ √
5. H < ( r − A)2 .
7. H < 1/10.

 ayk , byk < 1;
a
14. yk+1 = yk , byk = 1;
 2
0, byk > 1.

Section 2.5
1. a. 233.3 (naturally occurring), 182.56 (field experiment)
b. 158.68 (naturally occurring), 124.15 (field experiment)
c. 154, 62.1 (naturally occurring), 104.9, 0 (field experiment)
2. a. 80.687
b. 78.92
4. √ √
( a − 1)2 a−1 ( a + 1)2
< <
b b b
5. Multiply the desired double inequality by 2, subtract ( a−1
b − H), and
rewrite as a single inequality
√
s
2
( a − 1)2

a−1
H− < − H − 4H/b.
b b

Then square and simplify.

10. r − 1 < E < r + 1.

Section 2.6
1. r = e.
8. a. yk+1 = g((1 − d)yk )
b. Both are the same, r(1 − d).
 Answers to Selected Exercises 437

Section 2.7
1. y = 0, z = 0, unstable.
3. y = 0, z = 4, unstable.
5. y = 0, z = 0, unstable.
a−µ
7. y = 0, asymptotically stable if 0 < µ − a < 2 and y = bµ , asymptoti-
cally stable if µ < a, µa (a − µ) < 2.

Section 2.8
2. The ragwort is the factor limiting the moth population size and/or repro-
ductive ability; a certain amount of ragwort is needed to provide energy
to produce each egg, and the amount available is less than the total
reproductive capacity of the moths in the absence of any limitations.
3. The equilibrium (B, 0) is stable if and only if aB √ < 1; the equilibrium
( a1 , ab
1
log aB) is stable if and only if 1 < aB < e.

5c.
r
10

p
0.2 0.4 0.6 0.8 1
438 Answers to Selected Exercises

Chapter 3
Section 3.1
1. 1830.
3. 211.66
5. 17.46.
7. 199 days.
9. 4.5 billion years.

Section 3.2
3. (a) r = 1.0668 × 10−5 , (b)205(1990), 206(2000), 208(2010).
5. a. World War I and the ”Spanish” flu epidemic of 1918-19
b. The model is pretty good through 1970. Changing immigration pat-
terns, better census counts, and better food production technology
might explain changes, and changing some parameter values might
give a better fit.
7. Biomass after i year is 32.5 × 106 kg. It takes 1.55 years for biomass to
reach K
2 .

13. R0 = 1.4. Reduce contact rate to 1/14 person per day.

15. a. I ′ ≥ 0 with I ′ > 0 as long as S, I > 0 so I will continue to increase
[asymptotically] toward the entire population K, slowing down only
as S → 0.
b. “Eventually” here refers to the mathematical limit as time becomes
infinite—that is, the conclusion in (a) deals with long-term behav-
ior. In the long term, individuals will die, and others will be born to
replace them. The model above does not include such demographic
renewal. If we add a demographic renewal term to the model, the
conclusion in (a) is no longer valid.
y
17. y ′ = ry 1 − K


.

Section 3.3
5. y = (1 − t2 )−1/2 .
1−et 1−et−2
8. y = 1, y = 1+et , y = 1+et−2
 Answers to Selected Exercises 439

FIGURE S.4: Solution of Exercise FIGURE S.5: Solution of Exercise

11 13

Section 3.4
2
1. y = cet .
3. y = (c + 2t)1/2 .
2
5. y = ce−t .
3
7. y = 4 + cet /3
.
3
9. y = e(t −125)/3
.
2
11. y = y0 e−t .
2
13. y = 1−t2 .

17. a. The pacemaker signal y follows the control u, with the difference
y − u between them decaying over time at a rate proportional to
that difference.
b. y(t) − u(t) = z(t) = z(0)e−at so y(t) = u(t) + [y(0) − u(0)]e−at .
18. a. The heart beats more slowly (or at least firing rate decreases).
b. Firing rate decreases toward zero: no beating!
c. Solve equation (3.29) for yH and substitute.
19. T1 = T2
20. The solution changes only in the exponent of the exponential in the
r1
denominator of (3.37), which goes from −rt to −r0 t + 2π cos 2πt. As
r1
t → ∞ the oscillation dies out since r0 t >> 2π , so y → K anyway.
21. a. u′ = −µ1 u leads to u(t) = u(0)e−µ1 t and u(tc ) = ryk e−µ1 tc .
440 Answers to Selected Exercises

b. u′ = −µ2 u leads to u(t) = u(tc )e−µ2 (t−tc ) = ryk e−µ1 tc e−µ2 (t−tc )
and u(T ) = ryk e−(µ1 −µ2 )tc e−µ2 T
27. 99.
31. f (t) ≡ 0,and f (t) = t/2.
32. f (t) = 0(t < c), f (t) = (t − c)/2(t > c) for every c ≥ 0.

Section 3.5
1. Let Q(t) denote the weight of salt in the tank at time t, so that Q(0) = 3.
The rate at which salt is flowing
 in  is 2.5 lb./minute, and the rate at
Q
which salt is flowing out is 5 100 , since the concentration of salt at
Q
time t is 100 lb./gal. and the solution is flowing out at 5 gal./minute.
This leads to the initial value problem
Q
Q′ = 2.5 − , Q(0) = 3,
20
which is of the form (3.46) with k = 0.05, b = 2.5, and has solution

Q = 50 − 47 e−0.05t.

In particular, Q(60) = 50 − 47 e−3 = 47.66 lb. Also, it is clear from the

expression for Q that limt→∞ Q(t) = 50.
5. The concentration is 4% times (1−exp(−(0.6/1800)t)), so set that equal
to 10−4 and solve for t. Answer: 7.5 minutes.
7. Let y(t) be the amount of drug in the bloodstream at time t. Since the
drug leaves the blood at a rate proportional to y(t), we have y ′ (t) =
−ky(t) for some constant k > 0. We are given that y(0) = d. The
solution of the initial value problem y ′ = −ky, y(0) = d is y(t) = de−kt .
We are given also that y(1) = d/2. Therefore de−k = d/2, and k = log 2.
Now we have y(t) = de−(log 2)t = d 2−t . (Alternatively, e−k = 21 , so

t
y(t) = d 12 .)
The desired time is the value of t such that y(t) = d 2−t = d/5. Thus
2−t = 1/5, −t log 2 = log( 15 ) = − log 5, or

log 5
t= = 2.32 hours.
log 2

11. 40 minutes.
13. 58.3 minutes.
 Answers to Selected Exercises 441

Section 3.6
1. y = 1 − e−2t /2.
3. Y = e−2t + 4e−3t .
t2 +1
5. y = 2t .

7. y = t2 /4 + 4/t2 .
et +1−e
9. y = t .
11. y = t3 − 23 t2 + 32 t − 43 [1 − e−2t ].
p
13. y = 3t/2.
15. Y = e2t(t−1) .
19. y = (a + 1)et − 1.
100
24. x(t) = 500/(t + 3 ), x(50) = 6lb.
25. a.
272
s(t) = 1440(25 − t) − (25 − t)5/2
25
b.
s(t)
V (t) = 1000 − 40t = 40(25 − t) so = 36 − 0.272(25 − t)3/2
V (t)
The poor fish only have 3.56 minutes in which to be rescued.
29. The substitution y = 1/z, y ′ = −z ′ /z 2 gives
1 ′ 1 r 1
− z =r − ,
z2 z K z2
r
z ′ + rz = .
K
Using the integrating factor ert we obtain
r rt
ert z ′ + rert z = (ert z)′ = e ,
K
and integration gives
1
ert z = ert + c,
K
1
z= + ce −rt .
K
Returning to the original dependent variable y, we find the family of
solutions
1 K
y= 1 −rt
= −rt
.
K + ce 1 + cKe
This is the same family that we found before, in Section 3.3, except
that the arbitrary constant c in Section 3.3 is replaced by the arbitrary
constant cK here.
442 Answers to Selected Exercises

Chapter 4
Section 4.1

0
FIGURE S.6: Solution of Exercise 1

0
FIGURE S.7: Solution of Exercise 3

0 1 2
FIGURE S.8: Solution of Exercise 5

−1 0
FIGURE S.9: Solution of Exercise 7

9. y = 0 is unstable,y = K is asymptotically stable.

11. y = 0 is unstable,y = K is asymptotically stable.
15. a.
x+K
Z Z
dx = dt
(α + βx)(n − x)(x + K) − sx
In the case where B(x) has three real roots, integration gives

A1 log |x − r1 | + A2 log |x − r2 | + A3 log |x − r3 | = t + c;

exponentiation then yields

(x − r1 )A1 (x − r2 )A2 (x − r3 )A3 = ke−t .

 Answers to Selected Exercises 443

The only further simplification available in general is to take the

Ai th root of both sides of the equation, so that the ith factor on
the left is linear.
17. A single food source lasts much less than a single day for ants such as
these which feed on ephemeral sources. Long-term here refers to a matter
of hours or even minutes. On this scale the colony size is constant: births
and deaths would accumulate significantly only over a period of days or
weeks (barring catastrophic events).
18. a. β = 0
b. s = 0
19. Since 0 < x/(x + K) < 1, raising it to a higher power makes it smaller.
Therefore the rate would saturate more slowly, i.e., fewer ants would be
leaving the trail in unit time, though the rate still approaches s.
20. This allows the ants to optimize the trail (cut off roundabout loops), as
well as find other food sources.

Section 4.2
r
1. In case (i), where H < rK/4, we can factor dy/dt = − K (y − y1 )(y − y2 )
which leads to
 
y(t) − y2 y(0) − y2 y2 − y1
= exp −rt ,
y(t) − y1 y(0) − y1 K
from which
y2 − y1 Q exp −rt y2K
−y1


y(0) − y2
y(t) = , where Q = .
1 − Q exp −rt y2K−y1


y(0) − y1
r
In case (ii), where H = rK/4, we have dy/dt = − K (y − K/2)2 , the
solution to which is
K y(0) − K
2
y(t) = + rt K

.
2 1+ K y(0) − 2

Finally, in case (iii), where H > rK/4, there are no equilibria, so the
quadratic cannot be factored over the reals. Instead we complete the
square: " 2  #
dy r K HK K2
=− y− + − ,
dt K 2 r 4
so that after integration and solving for y(t) we have
" r #
K2
 
K −1 K rt HK
y(t) = + tan tan y(0) − − − .
2 2 K r 4
444 Answers to Selected Exercises
∗
2. Factoring dy/dt = (r/K)y(yH − y) and the partial fraction expansion
 
1 1 1 1
∗ − y) = ∗ − ∗ )
y(yH yH y (y − yH

lead to a solution (after integration and simplification) of

∗
y(0)yH
y(t) = ∗ .
y(0) + (yH − y(0)) exp[−(r − E)t]

4. a. From (4.12) we have y ∗ = K(1 − E/r) = 3(1 − 0.023/0.5) =

3(0.954) = 2.86 million deer. Deer-vehicle collisions (now 66,000
per year) reduce the deer population by E/r =4.6% or 140,000
deer.
b. dy/dt = ry(1 − y/K) − H − Ey
c.  
  s  2
1 E E 4HK 
y ∗ = K 1 − + K 1− −
2 r r r

gives an equilibrium value of 2.47 million deer, a much higher

(17.5%) impact than the simpler estimate predicts.
5. In the absence of harvesting, we assume a logistic law, so that the pop-
ulation over time is given by equation (2.13). Solving this for r as we
did for t in Example 4.2.1 yields the expression
1 K − y0
r= ln  .
tf y0 y(tKf ) − 1

For the given parameter values, we get r = 0.097/yr. A simple expo-

nential (rather than logistic) estimate would give r = t1f ln (yf /y0 ) =
0.094/yr, a difference of 3.2%. The critical harvest rate is rK/4 = 5812
whales/yr.
6. a. From (4.9), we can solve for r in terms of y2 : r = H/y(1 − y/K) =
577/174, 000(1 − 174, 000/265, 000) = 0.00966/yr, an order of mag-
nitude lower than commonly reported values for whale species. This
would mean that a very weak reproductive rate is required in order
for Norway’s harvesting rate to hold the population so far below
carrying capacity. Since the true value of r is likely much higher,
the population is probably still rising back toward equilibrium.
b. We substitute into (4.11) to find y2 at the different r values, ob-
taining 244,000, 255,000, and 258,000 whales, respectively.
 Answers to Selected Exercises 445

7. The MSY is rK/4 = 960 whales/yr. Estimating the population in 2000

as 11,700, there was a drop from the 1900 value of 120,000 of 108,300,
or an average 1083 whales/yr. Even if whales died from no reason but
whaling, the net death rate exceeds the MSY.
8. We calculate M SY = rK/4 = (.025)24000/4 = 150 whales/yr, well un-
der the actual catch, which would drive even a post-recovery population
extinct.
9. It would reduce it by a certain amount: calculate y2 and subtract from
100K.
12. The logistic CYH model with r = 0.05/yr, K = 10, 000 acres, H = 100
acres/yr has a stable equilibrium y2 = 7236 acres.
14. The equilibrium size is K(1 − E/r) = 194600(1 − 0.04/0.098) ≈ 115, 000
cranes.
15. a. 12,500/(20–100) allows space for between 125 and 625 breeding
pairs. We translate this to 250–1250 breeding adults, so 500–2500
cranes total; take an average and say 1500.
18. a. From the sustainability criterion H < rK/4 we obtain K > 4H/r =
4(344, 000)/0.07 = 19.7 million seals as a minimum.
b. For sustainability, the present population of 4.8 million must exceed
the minimum threshold, y(t) > y1 . From Exercise 10, y1 > H/r.
Therefore a necessary condition is that y(t) > H/r. But here
4.8 million < H/r = (344, 000)/0.07 = 4.9 million, so the hunt is
unsustainable given the present numbers. (It is close enough that
the uncertainty in the data allows for either possibility, though.)
19. a. You’re always in CEH mode, below the switch.
b. Here you begin in CYH mode, assuming the population is initially
near K, so the dynamics depend on whether the stable CYH equi-
librium y2∗ is greater or less than H/E. If y2∗ > H/E, then you re-
main always in CYH mode. If instead y2∗ < H/E, then CYH drives
the population down into the CEH range, where the equilibrium
population size is instead the CEH equilibrium.
20. (1 − p) of the equilibrium value y2 must exceed y1 . Thus H < rK(1 −
p)/(2 − p)2 .

Section 4.3
αKA
1. y = A+(K−A)e−αt .
a
3. y = a − αat−1 .
446 Answers to Selected Exercises

6. First solve (4.18) for Kd to get Kd = (KA − C ∗ )(KB − C ∗ )/C ∗ and then
substitute KA = A∗ + C ∗ , KB = B ∗ + C ∗ .
8.
β(N ) I ∗ 1
R0 = 1, N = 1− R ,
µ+ τ 0

and the disease persists if and only if R0 > 1, that is, the disease-
free equilibrium is globally asymptotically stable if R0 < 1, while the
endemic equilibrium is if R0 > 1.
9. The model formulation would only change inasmuch as β(N ) becomes
β(I), but the equilibrium analysis would be complicated considerably
with an additional function of I in the nonlinear term. If people change
their behavior in response to observed prevalence, β(I) should be a non-
increasing (perhaps even sharply decreasing) function of I. If the health-
care system breaks down beyond some critical level Ic , then β(I) should
be constant below Ic and increasing above it (perhaps up to some max-
imum).
10. The new effective population size for contacts is S + σI, so we get
dI σI 1
= β(S + σI)S − I.
dt S + σI τ
We reduce to one variable as before with S = N −I, S+σI = N −(1−σ)I.
1
11. a. R0 = σβ1 N τ , and I ∗ /N = 1 − R0 as before.
∗
b. R0 = σβ0 τ , and I /N = (R0 − 1)/(R0 − 1 + σ).
12. a. The recovery term disappears and R0 → ∞.
b. The proportion of infectives at the endemic equilibrium approaches
1: everyone is infected. This is unrealistic primarily because it ig-
nores timescale issues, such as demographic renewal.
13. Of course, dI/dt gets a new term −δI, but more significantly N is no
longer constant, so we cannot represent the system with a single equa-
tion. We must include an equation for either dS/dt or dN/dt.
14. a. In addition to the DFE, endemic equilibria
∗
r
I±
 
1 4
= 1± 1−
N 2 βτ N 2

exist iff βτ N 2 > 4. To analyze stability, we calculate f ′ (I) =

βN 2 [2(I/N ) − 3(I/N )2 − (1/βτ N 2 )], so that f ′ (0) < 0 and the
DFE is always LAS. If the two endemic equilibria exist, then
 r 
′ ∗ 1 2 4 4
f (I± ) = − βN 1 − ± 1− ;
2 βτ N 2 βτ N 2
 Answers to Selected Exercises 447

since 0 < 1 − βτ4N 2 < 1, the square root is larger than the radi-
cand, and we conclude that f ′ (I−∗
) > 0 while f ′ (I+∗
) < 0, so that the
upper endemic equilibrium is LAS and the lower one is unstable.
This model behaves differently than the basic mass-action epidemic
model in two important ways: first, the DFE is always LAS, which
implies that R0 = 0 (that is, a single infective introduced into a
population of susceptibles never causes an outbreak, because infec-
tion requires many infectives (I 2 )); and second, for βτ N 2 > 4, the
persistence of the infection depends on the initial number of infec-
∗
tives: if it is high enough (I(0) > I− ), it will establish an endemic
state; otherwise it will die out quickly.
b. This model behaves
√ mostly like the usual mass-action incidence
model, with R0 = βτ .
16. a. dy/dt = ry(1 − y/K) − kxy
b. dx/dt = cxy − µx
c. With y constant, the model becomes linear: dy/dt = (cY − µ)x,
and the predators either grow unchecked (if cY > µ) or die out (if
cY < µ).
d. With x constant, the model becomes logistic, dx/dt = r2 y(1 −
y/K2 ), where r2 = r − kX and K2 = K(r − kX)/r, so that the
prey population x approaches a smaller carrying capacity K2 as
long as r > kX, and dies out if r < kX.

Section 4.4
1. If r > 0, no equilibrium. If r = 0,y = 0 is unstable. If r < 0, there are
two equilibria, one unstable and one asymptotically stable.
3. y = 0 is asymptotically stable if and only if r < 0. y = −r is asymptot-
ically stable if and only if r > 0.
p
5. If r ≥ 0, y = 0 is unstable. If r < 0, y = 0 is unstable and y = ± −1/r
are asymptotically stable.
7. If −2 < r < 2 there are no equilibria. If r < −2 or r > 2, there are two
equilibria, one unstable and one asymptotically stable.
y
14. Let y = x/K and τ = ts/K. Then dy/dτ = a + by − cy 2 − y+1 , where
a = αn/s, b = (βn − α)K/s, and c = βK 2 /s.
15. a. We use A as the unit of population size and make the change of
y
dependent variable u = A . This transforms the equation (4.29) to

u2
 
′ Au
Au = rAu 1 − −H
K u2 + 1
448 Answers to Selected Exercises

or
u2
 
A ′ A Au
u =r u 1− − 2 .
H H K u +1
Ar K
We define the new parameters R and Q by R = H , Q= A and
then we have the model
u2
 
A ′ u
u = Ru 1 − − 2 . (S.3)
H Q u +1
H
We could also make the change of independent variable s = A t,
which would replace (S.3) by

u2
 
du u
= Ru 1 − − 2 ,
ds Q u +1

but we need not do this in order to carry out the equilibrium analy-
sis. In terms of the original independent variable t we have the same
equilibria and stability; the change of dependent variable changes
only the exponential rate of approach to equilibrium. Effectively, we
have replaced the original four parameters by the two parameters
R and Q. An examination of the forest changes indicates that un-
der change of the forest, the parameter Q remains almost constant,
but the parameter R grows with the development of the forest. The
process of change in the parameter in many applications is the rea-
son for our study of bifurcations. In this case we will consider R to
be our bifurcation parameter.
b. Equilibria
 of (S.3) are u = 0 and the intersections of the line v =
u u
R 1 − Q and the curve v = 1+u 2 , solutions of the equation

 
u u
R 1− = . (S.4)
Q 1 + u2

In order to understand the dynamics of the model, we will need to

u
investigate how the equilibria depend on R. The curve v = 1+u 2

starts at the origin, increases to a maximum at (1, 21 ), and then

decreases, approaching zeroas u → ∞, with an inflection point at
√ √
( 3, 43 ). The line v = R 1 − Q u
passes through the points (0,
R) and (Q, 0) in the u − v plane. A typical situation is as shown in
Figure S.11. There we see that, in addition to the u = 0 equilibrium,
equation (S.4) gives three more, labelled K, L, and M , where the
line and the curve intersect. We can determine their stability by
applying the theorem from Section
h  4.1. Here (S.3) i has the form
u′ = uf (u), where f (u) = H u u
A R 1 − Q − u2 +1 — in terms of
Figure S.11, “the line minus the curve”. Therefore the equilibria
 Answers to Selected Exercises 449

8 food
supply B
limitations
(M)

u
4

(L)

predator limitations (K) D

0.5 1 1.5 2
R
r

FIGURE S.10: Bifurcation diagram for equation (S.3)

will be stable which have f ′ (u) < 0 there, i.e., where the slope of
the line is less than the slope of the curve. From the graph we can
see that this occurs at K and M , but not at 0 or L. Therefore
we have an unstable equilibrium at u = 0 and two asymptotically
stable equilibria K and M separated by an unstable equilibrium L.
The domain of attraction of the small equilibrium K is the interval
[0, L), and the domain of attraction of the large equilibrium M is
the interval (L, ∞).
If R (the slope of the line) increases until K and L coalesce, as in
Figure S.12, then only the equilibrium M remains. Thus if a system
is in equilibrium at K and R increases, there would be a jump from
the small equilibrium K to the large equilibrium M. Conversely, if
R is decreased until L and M coalesce, as in Figure S.13, then
only the equilibrium K would remain, and a system which is in
equilibrium at the large equilibrium M would crash to the small
equilibrium K. These two situations therefore involve bifurcations
like the saddle-node bifurcation seen earlier in this section.
c. We think of the parameter R as representing resources of the forest
as food supply for the budworms, and the equilibria K and M as
corresponding to budworm limitation by predators and food sup-
ply, respectively. In a young forest, R is small with only the small
equilibrium K (limited by predators). As the forest develops, the
food supply becomes so great that budworm growth exceeds con-
trol by predators, and R increases to the stage where there is only
the large equilibrium M , signifying an outbreak of the budworm
population. If this outbreak destroys the forest, the predators may
regain control as R decreases, and the population may crash to the
small equilibrium K. The model exhibits hysteresis; the outbreak
and crash will generally occur at different population levels. This
450 Answers to Selected Exercises
1.6 1.6 1.6

1.4 1.4 1.4

K

1.2 1.2 1.2

L K

1 1 1

v0.8 v0.8 v0.8

0.6 0.6 0.6

M M

0.4 0.4 0.4

0.2 0.2 0.2

0 2 4 6 8 10 12 0 2 4 6 8 10 12 0 2 4 6 8 10 12
u u u

FIGURE S.11: The FIGURE S.12: Equi- FIGURE S.13: Equi-

two sides of equation libria K and L coalesce libria L and M coalesce
(S.4) and their intersec-
tions

may also be seen from the model (S.3) by drawing the graph of
the equilibria as a function of the bifurcation parameter R. The
equation (S.4) is easily solved for R as a function of y, and the
bifurcation curve is just the (multi-valued) inverse function (Fig-
ure S.10). When R grows from a small value, the equilibrium moves
along the bifurcation curve until it reaches the point A, and then
jumps to the higher point B and continues upwards. When R de-
creases from a high value, the equilibrium moves down the bifurca-
tion curve past B to C before crashing to the lower value D. Thus
the discontinuity in equilibrium on the way up is at a higher value
of R than the discontinuity in equilibrium on the way down.

Section 4.5
1. The modified Euler method gives 2.70796 instead of 2.5937, an error of
0.38% rather than 4.6%.
3. The Runge-Kutta approximation gives

t 0 0.2 0.4 0.6 0.8 1.0

y 1 1.177 1.368 1.566 1.765 1.956

5. The Runge-Kutta approximation gives

 Answers to Selected Exercises 451

t 0 0.2 0.4 0.6 0.8 1.0

y 1 0.9803 0.9251 0.8441 0.7500 0.6544

8. h < 2.
10. 4.6 minutes
11. 8.0 minutes; there are fewer ants wandering around, so it takes longer
for the pheromone trail to become established enough to attract many
ants
13.
rmq qP q−1
f1′ (P ) = 2,
(mq + P q )
rqmq P q−2 [(q − 1)mq − (q + 1)P q ]
f1′′ (P ) = 3
(P q + mq )
so f1′′ (P ) = 0 when P = 0 or
 1/q
q−1
P = P̄ = m .
q+1
Now
  q−1
q
q−1
r q q+1
′
f1 (P̄ ) = 2 .
m

q−1
1 + q+1

For q ≥ 2 this is f1′ (P̄ ) ≈ rq/2m from above.

14. a. Let p = P/m and τ = rt/m. Then dp/dτ = a − bp + pq /(1 + pq ),
where a = L/r and b = sm/r.
b.
qp2q − pq − 1 qpq−1 (1 − pq )
a= 2 , b= 2
(1 + pq ) (1 + pq )
c.
qpq−1 (2q + 1)p2q + 2(q + 1)pq + 1
 
da
= 4 ,
dp (1 + pq )
qpq−2 p2q − 3qpq + (q − 1)
 
db
= 3 ,
dp (1 + pq )
so da/dp > 0 for p > 0, i.e., a is increasing in p, with a(0) = −1
and limp→∞ a(p) = q, while db/dp = 0 at
r !1/q
3 9 2
p± = q± q −q+1 ,
2 4
452 Answers to Selected Exercises

with b(0) = 0, b increasing in p for 0 < p < p− , decreasing on

[p− , p+ ], and approaching 0 again as p → ∞. Meanwhile, by in-
spection we see that b(p)’s only zeroes are at p = 0 and 1.
d. b(1) = 0, while a(1) = (q − 2)/4 > 0 ⇔ q > 2. Thus the unique
x-intercept is positive iff q > 2. The significance of this result is
that bifurcations (and hysteresis) exist only for q > 2. For q >
2, there exist three equilibria when (a, b) falls below the reverse-
parametrized curve.
15. a. r up by 10, m down by 100; irreversible lake, eutrophic equilibrium
b. r down by 10, m up by 100. yields reversible lake, oligotrophic
equilibrium

Chapter 5
Section 5.1
1. Orbits are circles with center at the origin.
3. Orbits are curves sin z = y 2 /2 + c.
5. Orbits are curves z = ceβy/α .

Section 5.2
1. Linearization at (1, 1) is u′ = u + v, v ′ = −u + v.
3. Linearization at (−1, 1) is u′ = u + 2v, v ′ = u. Linearization at (−1, −1)
is u′ = u − 2v, v ′ = u.
5. System has no equilibria.
9. Equilibrium (0, 0) is asymptotically stable.
11. Equilibrium (0, 0) is asymptotically stable.

Section 5.3
1. y = 5K1 e2t + K2 e−4t , z = K1 e2t − K2 e−4t .
3. y = K2 et , z = K1 e2t − K2 et .
5. y = K1 e4t , z = 2K1 te4t , z = 2K1 te4t + K2 e4t .
7. y = K1 e−t + K2 e−2t , z = −K1 e−t − 2K2 e−2t .
 Answers to Selected Exercises 453

9. y = K1 tet + K2 et , z = K1 et .
11. y = K1 e3t sin 5t + K2 e3t cos 5t, y = K2 e3t sin 5t + K1 e3t cos 5t.
13. y = K1 tet + K2 et , z = K1 et .
cK1 at
17. y = K1 eat , z = a−d e + K2 eat , (a 6= d).

18. y = K1 eat , z = cK1 eat + K2 eat .

Section 5.4
1. (1, 1) is unstable.
3. Both (−1, 1) and (−1, −1) are unstable.
5. System has no equilibria.
7. (0, 0) and (1/3, 1/3) are unstable; (0, 1) and (1, 0) are asymptotically
stable.
9. The origin is unstable (saddle point). The equilibrium (a/λ, 0) is asymp-
totically stable if and only if λc < aµ. If λc > aµ, there is an asymptot-
ically stable equilibrium with y > 0, z > 0.

Chapter 6
Section 6.1
1. 1.44.
4. 17.56 %.
5. 2.83 %

9. a. New model is

S ′ = −βSI + cR, I ′ = βSI − γI, R′ = γI − cR.

b. Reduced model is

S ′ = −βSI + c[N − S − I], I ′ = βSI − γI.

c. Equilibria are (N, 0) and (γ/β, c(N − γ/β)/(γ + c)).

d. Threshold condition is at R0 = βN/γ = 1.
454 Answers to Selected Exercises

Section 6.2
1. Equilibrium (60, 20) is asymptotically stable; species coexist.
3. Equilibrium (0, 45) asymptotically stable; y species goes extinct and z
species wins competition.
5. All orbits approach the asymptotically stable equilibrium (10, 10).
7. All orbits approach a limit cycle about the unstable equilibrium
(20, 10.8).
9. All orbits approach a limit cycle about the unstable equilibrium (30, 25).
12. There are three equilibria (0, 0), (0, M ), and (K, 0). They are all un-
stable. If ab≥ 1, orbits areunbounded. If ab < 1, a stable equilibrium
K+aM Mb+K
(y∞ , z∞ ) = 1−ab , 1−ab appears and all trajectories approach this
equilibrium.
15. (0, 0) is asymptotically stable and (70, 120) is unstable.
17. (20, 10 is asymptotically stable, and (0, 0), (20, 0), (0, 10) are unstable.
22. 
The system has  two equilibria: extinction (0,0) and persistence
d γ 1
γ+ǫd , γ+ǫd × 1 − Rd K; the extinction equilibrium is LAS iff Rd <
1, and the persistence equilibrium exists and is LAS iff Rd > 1. Here
γ
Rd = db µ+γ gives the birth-to-adult-death ratio, multiplied by the pro-
portion of juveniles who survive to maturity. Stability follows from stan-
dard methods using the Jacobian matrix and Routh-Hurwitz criteria.

Section 6.3
1. The Runge-Kutta approximation gives

t 0 0.01 0.02 0.05 1.0 10.0

y 10 14.74 18.00 20.39 20.15 20.00
z 10 12.80 14.96 18.19 19.63 20.00

3. The Runge-Kutta approximation gives

t 0 2 4 6 8 10
y 40 22.99 17.27 13.71 11.92 11.94
z 10 13.25 15.07 15.21 14.00 12.44
 Answers to Selected Exercises 455

5. (a) (F ∗ , M ∗ ) = (0, 0) and (µm /mc, µf /f c)

f cM ∗ − µf f cF ∗
   
−µf 0
(b) J = , so J(0, 0) =
mcM ∗ mcF ∗ − µm 0 −µm
which has both eigenvalues negative, so (0,0) is LAS, but
 
0 µm f /m
J(µm /mc, µf /f c) =
µf m/f 0
√
which has eigenvalues ± µf µm so (µm /mc, µf /f c) is a saddle point
(hence unstable).
(c) Note one trajectory leads directly from the saddle point to the ori-
gin; perpendicular to this at the saddle point, a hyperbola-like curve (2
trajectories approaching the saddle point) divides the state space into
2 regions: below it, all solutions lead to extinction, while above it all
trajectories lead to infinity.
!
Rf Rm −Rf −Rm R R −R −R
6. (a) (F ∗ , M ∗ ) = E0 (0, 0) or E1 cf , f“ m f m”
Rm (Rf + cm Rm ) Rf ccm Rf +Rm
f

(b) E1 is positive iff Rf Rm > Rf + Rm , so conjecture E1 is LAS when

it exists, and E0 is LAS otherwise (i.e., when Rf Rm < Rf + Rm ).
Numerical explorations should confirm this.

Chapter 7
Section 7.1
In Exercises 1–9, only Exercise 4 exhibits a limit cycle and bursting be-
havior. Bursting behavior requires J > 0 and ǫ and γ small.

Section 7.2
q √ t/ǫ
5. 2y 1+y
1−y = 3e . Solution of reduced problem is y = 0(t = 0), y =
1/2(t > 0).
7. Equation (6.16) has a unique equilibrium (v, w) with 1/2 < v < 2/3, w <
0, and this equilibrium is unstable.
11. For dv/dt = −f (v) − w, an equilibrium v ∗ is LAS iff −f ′ (v ∗ ) < 0. But
this occurs precisely outside the interval [A, B]. We can find A and B
by setting f ′ (v) = 0, which yields
1h p i
A, B = (a + 1) ± a2 − a + 1 .
3
456 Answers to Selected Exercises

12. The unique equilibrium w∗ = v/γ can be seen to be LAS since

d dw


dw dt = −γ which is negative as long as γ > 0, so w always heads
toward w∗ .
13. dw/dt > 0 iff v > γw; at (B, −f (B)), this yields B > −γf (B). Substi-
tuting for f and then B, this condition becomes
1 9
γ< = √ .
(B − a)(1 − B) a − 4a + 1 + (a + 1) a2 − a + 1
2

The function on the right is an increasing function of a, starting at 4.5

when a = 0, and increasing without bound as a → 1. Thus any γ < 4.5
satisfies this criterion regardless of the value of a.
14. For parameter values with γ sufficiently low (as described in the text),
when the equilibrium reaches A the orbit becomes periodic. The figure
below shows the graph for a = 0.3, γ = 1, ǫ = 0.002, and J = 0.5.
When the equilibrium reaches B the periodic orbit collapses to a stable
equilibrium again.
w
0.8

0.6

0.4

0.2

v
0.5 1.0

15. Fast: inner solution, boundary layer. Slow: outer solution, reduced prob-
lem, quasi-steady state.

Section 7.3
1. V → 0 if g < bK2 T0 ; V → g/K2 T0 − b if g > bK2 T0 .
2. (a) Solutions approach either the virus-free equilibrium (VFE) or pro-
gression to AIDS, depending on initial conditions (an unstable endemic
equilibrium provides a separatrix). (b) A unique endemic equilibrium
(EE) is globally asymptotically stable (GAS). (c) Two possible por-
traits: the VFE is GAS if the nullcline curves do not cross, but if they
do cross [twice], the VFE and an EE are both LAS (an unstable EE
again providing a separatrix). (d) Two possible portraits: progression to
AIDS if the nullcline curves do not cross (Figure 7.13), while if they do
cross [twice] there is also a LAS EE (with an unstable EE providing a
separatrix).
 Answers to Selected Exercises 457
r 
s2 a g g 
3. (a) Tc = − (a − b), requires µ > k1 b
µ − k1 b k2 k2
s2 a b 2 g s2 a
(b) < <
µ − k1 b a2 k2 µ − k1 b
4. Behavior would change to progression to AIDS for g = 1200.
6. V ′ > VgV ′ ′ gV
+b − N I , so (V + N I) > V +b and thus increases without bound
in both systems. In the system with T , the growth of V pulls down
T ′ < 0 so eventually T → 0.
7. The virus-free equilibrium is asymptotically stable if g < bK2 s1 /µ.

Section 7.4
1. (a) The condition (7.40) provides an upper bound on the slope of
Λ(N ) which decreases as N increases: for any given N0 , Λ′ (N0 ) <
Λ(N0 )/N0 , the slope of the line segment connecting (0,0) with
(N, Λ(N )). Thus for any N1 > N0 , Λ(N1 ) < (Λ(N0 )/N0 )N1 , which
implies that the bound on Λ′ (N1 ) is even lower:

Λ′ (N1 ) < Λ(N1 )/N1 < Λ(N0 )/N0 .

Thus the slope of Λ(N ) will continue to decrease until it is lower

than d, and at some point the curve will cross below d · N .
(b) If Λ(x0 ) < d · x0 , then by (7.40) Λ′ (x0 ) < d, and as discussed in
(a), since the bound decreases as x increases, the slope of the curve
can never again increase past d, which would be necessary in order
to cross above the line y = d · x.
(c) If we write (7.45) in the form N ′ = g(N ), then g ′ (N ) = Λ′ (N ) − d.
Thus for an equilibrium K > 0 we have g ′ (K) = Λ′ (K) − d <
Λ(K)/K −d = dK/K −d = 0 (since at an equilibrium Λ(K) = dK),
and since g ′ (K) < 0 K must be asymptotically stable. Meanwhile,
g ′ (0) = Λ(0) − d, which is asymptotically stable iff Λ′ (0) < d.
2. (0, P02 ) is locally asymptotically stable iff Λ′ (0) < d, consistent with the
stability of (K, P02 ) iff Λ′ (0) > d.
4. For all cases we first compute f ′ (P ) = (∆D +∆R )[3P 2 −2P (P ∗ +1)+P ∗],
from which f ′ (0) = (∆D + ∆R )P ∗ , f ′ (1) = (∆D + ∆R )(1 − P ∗ ), and
f ′ (P ∗ ) = (∆D + ∆R )P ∗ (P ∗ − 1).
Now for Case I we first consider the subcase ∆D + ∆R > 0. Here P ∗ > 1
since ∆D < 0, so it is not of interest. If instead ∆D + ∆R < 0, then
P ∗ < 0, so it is again not of interest. Either way, f ′ (0) > 0 and f ′ (1) < 0,
so P = 0 is unstable and P = 1 is stable; thus the A allele approaches
100% frequency.
458 Answers to Selected Exercises

5. Case II is the reverse of Case I: if ∆D + ∆R > 0 then P ∗ < 0, if

∆D + ∆R < 0 then P ∗ > 1, and either way f ′ (0) < 0, f ′ (1) > 0,
making P = 0 stable and P = 1 unstable. The a allele approaches 100%
frequency.
6. With both ∆D > 0, ∆R > 0, P ∗ ∈ (0, 1), so that f ′ (0) > 0, f ′ (1) > 0,
and f ′ (P ∗ ) < 0. Thus P ∗ is globally stable on the interval (0,1).
7. With ∆D < 0, ∆R < 0, P ∗ ∈ (0, 1), and f ′ (0) < 0, f ′ (1) < 0, and
f ′ (P ∗ ) > 0. Thus the unstable P ∗ divides the interval (0,1) into two
basins of attraction, with trajectories beginning on either side of it ap-
proaching the extreme on that side.
8. If ∆R > 0, P → 1, and asymptotically all members are of genotype AA.
If ∆R < 0, P → 0, and asymptotically all members are of genotype aa.

Section 7.5
1. y = K1 t + K2 , y = −t + 1.
3. y = K1 e3t + K2 e2t , y = −3e3t + 4e2t .
5. y = K1 e5t + K2 te5t , y = e5t − 6tE 5t .
√ √ √
y = K1 e−5t/2 cos√ 15t/2 + K2 e−5t/2 sin 15t/2, y = e−5t/2 cos 15/2 +
7. √
15/2e−5t/2 sin 15t/2.
9. y = K1 e2t cos 3t + K2 e2t sin 3t, y = e2t cos 3t − e2t sin 3t.
11. y = K1 et cos t + K2 et sin t, y = et cos t − 2et sin t.
√
13. Equilibria are (0, 0) and (0, ± −b) if b < 0.
15. Linearization at (0, 0) is u′ = v, v ′ = −v.
17. Linearization at (0, 0) is u′ = v, v ′ = −g ′ (0) − f (0)v.
Bibliography

Caroline Bampfylde, Modelling rainforests, M.Sc. thesis, Oxford University

(1999).
J.B.I. Bard, A quantitative model of liver regeneration in the rat, Journal of
Theoretical Biology 73(3): 509–530 (1978).
J.B.I. Bard, A quantitative theory of of liver regeneration in the rat II: Match-
ing an improved mitotic inhibitor model to the data, Journal of Theo-
retical Biology 79(1): 121–136 (1979).
Madeleine Beekman, David J. T. Sumpter, and Francis L. W. Ratnieks,
Phase transition between disordered and ordered foraging in Pharaoh’s
ants, Proceedings of the National Academy of Sciences USA, 98: 9703–
9706 (2001).
Hassan Benchekroun and Ngo Van Long, Transboundary fishery: a differen-
tial game model, Economica, 69: 207–221 (2002).
Samuel Bernard, Jacques Bélair and Michael C. Mackey, Bifurcations in
a white-blood-cell production model, Comptes Rendus Biologies, 327:
201–210 (2004).
Raymond J.H. Beverton and Sidney J. Holt, On the dynamics of exploited fish
populations, Fishery Investigations Series 2 (19), Great Britain Ministry
of Agriculture, London (1954) (reprinted by Chapman & Hall (1993)).
R.J.H. Beverton and S.J. Holt, The theory of fishing, in Sea fisheries: their
investigation in the United Kingdom (M. Graham, ed.), Edward Arnold,
London, pp. 372–441 (1956).
G.E.P. Box, Science and statistics, Journal of the American Statistical As-
sociation, 71: 791–799 (1976).
Fred Brauer and David A. Sánchez, Constant rate population harvest-
ing: equilibrium and stability, Theoretical Population Biology, 8: 12–30
(1975).
T. Carlson, Über Geschwindigkeit und Grösse der Hefevermehrung in Würze,
Biochemische Zeitschrift, 57: 313–334 (1913).

459
460 Bibliography

Robert J. Carman and Mary Alice Woodburn, Effects of low levels of

ciprofloxacin on a chemostat model of the human colonic microflora,
Regulatory Toxicology and Pharmacology, 33: 276–284 (2001).
S. R. Carpenter, D. Ludwig, W. A. Brock, Management of eutrophication for
lakes subject to potentially irreversible change, Ecological Applications,
9: 751–771 (1999).
Hal Caswell, Masami Fujiwara, and Solange Brault, Declining survival prob-
ability threatens the north Atlantic right whale, Proceedings of the Na-
tional Academy of Sciences USA, 96: 3308–3313 (1999).
J. Chervinski, Salinity tolerance of young catfish, Clarias lazera (Burchell),
Journal of Fish Biology, 25: 147 (1984).
J. Davidson, On the ecology of the growth of the sheep population in South
Australia, Transactions of the Royal Society of South Australia, 62: 141–
148 (1938).
J. Davidson, On the growth of the sheep population in Tasmania, Transac-
tions of the Royal Society of South Australia 62: 342–346 (1938).
M. Doebeli, Dispersal and dynamics, Theoretical Population Biology, 47: 82–
106 (1995).
J. Dumas and P. Prouzet, Variability of demographic parameters and popula-
tion dynamics of Atlantic salmon Salmo salar L. in a south-west French
river, ICES Journal of Marine Science, 60: 356–370 (2003).
M. Eisen and J. Schiller, Stability analysis of normal and neoplastic growth,
Bulletin of Mathematical Biology 39(5): 597–605 (1977).
K.E. Emmert and L.J.S. Allen, Population persistence and extinction in a
discrete-time, stage-structured epidemic model, Journal of Difference
Equations and Applications, 10: 1177–1199 (2004).
A.S. Evans, Viral Infections of Humans, 2nd ed., Plenum Press, New York
(1982), reported by H.W. Hethcote, Three basic epidemiological models,
Applied Mathematical Ecology, S.A. Levin, T.G. Hallam, and L. J. Gross
(eds), Biomathematics 18 Springer-Verlag, New York-Heidelberg, pp.
119–144 (1989).
A.S. Fauci, G. Pantaleo, and D. Weissman, Immunopathogenic mechanisms
of HIV infection, Annals of Internal Medicine 124: 654–663 (1996).
U. Feldmann and B. Schneider, A general approach to multicompartment
analysis and models for the pharmacodynamics, in J. Berger et al.
(eds.), Mathematical models in medicine: workshop, Mainz, March 1976
(Berlin/New York: Springer-Verlag), Lecture Notes in Biomathematics
11: 243–277 (1976).
 Bibliography 461

R. Fitzhugh, Impulses and physiological states in theoretical models of nerve

membrane, Biophysics Journal, 1: 445–466 (1961).
G.P. Garnett, An introduction to mathematical models in sexually trans-
mitted disease epidemiology, Sexually Transmitted Infections, 78: 7–12
(2002).
G.F. Gause, The Struggle for Existence, Williams and Wilkins, Baltimore
(1934).
Tim Gerrodette, The tuna-dolphin issue, in W.F. Perrin, B. Würsig and
J.G.M. Thewissen, eds. Encyclopedia of Marine Mammals (2nd edition).
Elsevier, Amsterdam, pp. 1192–1195 (2009).
T. Gerrodette and J. Forcada, Non-recovery of two spotted and spinner dol-
phin populations in the eastern tropical Pacific Ocean, Marine Ecology
Progress Series, 291: 1–21 (2005).
Tim Gerrodette, George Watters, Wayne Perryman, Lisa Ballance,
Estimates of 2006 dolphin abundance in the Eastern Tropical
Pacific, with revised estimated from 1986–2003, NOAA Techni-
cal Memorandum NMFS-SWFSC-422, April 2008, available at
https://swfsc.noaa.gov/publications/TM/SWFSC/NOAA-TM-NMFS-
SWFSC-422.pdf.
L. Glass and M.C. Mackey, Pathological conditions resulting from instabil-
ities in physiological control systems, Annals of the N.Y. Academy of
Science, 316: 214–235 (1979).
J. Gressel and L.A. Segel, The paucity of plants evolving genetic resistance
to herbicides: possible reasons and implications, Journal of Theoretical
Biology, 75: 349–371 (1978).
G.H. Hardy, Mendelian proportions in a mixed population, Science 28(706):
49–50, July 1908 (letter to the editor).
James D. Hardy, The physical laws of heat loss from the human body, Pro-
ceedings of the National Academy of Science USA, 23: 631–637 (1937).
Craig A. Harms, Mark G. Papich, M. Andrew Stamper, Patricia M. Ross,
Mauricio X. Rodriguez, and Aleta A. Holm, Pharmacokinetics of oxyte-
tracycline in loggerhead sea turtles (Caretta caretta) after single in-
travenous and intramuscular injections, Journal of Zoo and Wildlife
Medicine, 35: 477–488 (2004).
M.P. Hassell, Density dependence in single species populations, Journal of
Animal Ecology, 44: 283–295, 1975.
462 Bibliography

H.W. Hethcote and J.A. Yorke, Gonorrhea Transmission Dynamics and Con-
trol. Lecture Notes in Biomathematics 56, Springer-Verlag, New York
(1984).
A. L. Hodgkin and A. F. Huxley, A quantitative description of membrane
current and its application to conduction and excitation in nerve, J.
Physiology, 117: 500–544 (1952).
F.C. Hoppensteadt, Singular perturbations on the infinite interval, Trans.
Amer. Math. Soc., 123: 521–535 (1966).
G. Evelyn Hutchinson, An introduction to population ecology, Yale University
Press, New Haven (1978).
Catherine Ryan Hyde, Pay it forward, Simon and Schuster, New York (1999).
International Dolphin Conservation Program Scientific Advisory Board, 7th
meeting, Updated estimates of NMIN and stock mortality limits, Doc-
ument SAB-07-05, La Jolla, California, 30 October 2009, available
at http://www.iattc.org/PDFFiles2/SAB-07-05-Nmin-and-Stock-
Mortality-Limits.pdf.
Holger W. Jannasch, Steady state and the chemostat in ecology, Limnology
and Oceanography, 19: 716–720 (1974).
Emmanuel Jolivet, Introduction aux modèles mathématiques en biologie,
INRA/Masson, Paris (1983).
D.S. Jones and B.D. Sleeman, Differential Equations and Mathematical Bi-
ology, CRC Press, Boca Raton, FL (2003).
Sukgeun Jung and Edward D. Houde, Recruitment and spawning-stock
biomass distribution of bay anchovy (Anchoa mitchilli) in Chesapeake
Bay, Fishery Bulletin, 102: 63–77 (2004).
J. Keener and J. Sneyd, Mathematical Physiology, Springer-Verlag, New York
(1998).
W. O. Kermack and A. G. McKendrick, A contribution to the mathematical
theory of epidemics, Proceedings of the Royal Society of London, 115:
700–721 (1927).
W.O. Kermack and A.G. McKendrick, Contributions to the mathematical
theory of epidemics, Part II, Proceedings of the Royal Society of London,
138: 55–83 (1932).
J.M. Kienzler, P.F. Dahm, W.A. Fuller, A.F. Ritter, Temperature-based es-
timation for the time of death in white-tailed deer, Biometrics, 40: 849–
854 (1984).
 Bibliography 463

B.W. Knight, Dynamics of encoding in a population of neurons, J. General

Physiology, 59: 734–766 (1972).
G. A. Knox, The key role of krill in the ecosystem of the southern ocean
with special reference to the convention on the conservation of antarctic
marine living resources, Ocean Management, 9: 113–156 (1984).
Mark Kot, Elements of Mathematical Ecology, Cambridge University Press,
Cambridge (2001).
C.M. Kribs-Zaleta, To switch or taper off: the dynamics of saturation, Math.
Biosci., 192: 137–152 (2004).
S.D. Lane and N.J. Mills, Intraspecific competition and density dependence
in an Ephesia kuehniella–Venturia canescens laboratory system, OIKOS
101: 578–590 (2003).
Richard C. Lathrop, Stephen R. Carpenter, Craig A. Stow, Patricia A. So-
ranno, and John C. Panuska, Phosphorus loading reductions needed to
control blue-green algal blooms in Lake Mendota, Canadian Journal of
Fisheries and Aquatic Sciences, 55: 1169–1178 (1998).
R. Levins, Some demographic and genetic consequences of environmental
heterogeneity for biological control, Bulletin of the Entomological Society
of America 15: 237–240 (1969).
N. Levinson, Perturbations of discontinuous solutions of nonlinear systems
of differential equations, Acta Mathematica, 82: 71–106 (1950).
D. Ludwig, D.D. Jones and C.S. Holling, Qualitative analysis of insect out-
break systems: the spruce budworm and forest, Journal of Animal Ecol-
ogy, 47: 315–332 (1978).
M.C. Mackey and L. Glass, Oscillation and chaos in physiological control
systems, Science, 197: 287–289 (1977).
T.R. Malthus, An essay on the principle of population, Harmondsworth,
Middlesex (1798) (republished by Penguin, New York (1970)).
Robert M. May, Thresholds and breakpoints in ecosystems with a multiplicity
of stable states, Nature, 269: 471–477 (1977).
J. Maynard Smith, Mathematical ideas in biology, Cambridge University
Press, Cambridge, 1968.
J. Maynard Smith and M. Slatkin, The stability of predator-prey systems,
Ecology, 54: 384–391 (1973).
R.W. McCarley and J.A. Hobson, Neuronal excitability modulation over the
sleep cycle: A structured and mathematical model, Science, 189: 58–60
(1975).
464 Bibliography

Curt D. Meine and George W. Archibald (Eds)., The cranes: Status survey
and conservation action plan. IUCN, Gland, Switzerland, and Cam-
bridge, U.K. 294pp. Northern Prairie Wildlife Research Center Online.
http://www.npwrc.usgs.gov/resource/birds/cranes/index.htm
(Version 02MAR98) (1996).
L. Michaelis and M.I. Menten, Die Kinetik der Invertinwirkung, Biochem.
Z., 49: 333–369 (1913).
R.S. Miller and D.B. Botkin, Endangered species: models and predictions,
American Scientist, 62: 172–181 (1974).
Jacques Monod, La technique de culture continue: théorie et applications,
Annales de l’Institut Pasteur, 79: 390–410 (1950).
J.S. Nagumo, S. Arimoto, and S. Yoshizawa, An active pulse transmission
line simulating nerve axon, Proc. Inst. Radio Engineers, 50: 2061–2071
(1962).
M.E.J. Newman, The structure and function of complex networks, SIAM
Review, 45: 167–256 (2003).
A.J. Nicholson, An outline of the dynamics of animal populations, Australian
Journal of Zoology, 3: 9–65 (1954).
A.J. Nicholson and V.A. Bailey, The balance of animal populations, Proc.
Zoological Society of London, 3: 551–598 (1935).
Aaron Novick and Leo Szilard, Description of the chemostat, Science, 112:
715–716 (1950).
M.A. Nowak and R.M. May, Virus dynamics: Mathematical principles of
immunology and virology, Princeton University Press, 2003.
R. Pearl, Introduction of medical biometry and statistics, Saunders, Philadel-
phia (1930).
Alan S. Perelson and P.W. Nelson, Mathematical analysis of HIV-1 dynamics
in vivo, SIAM Review 41: 3–44 (1999).
J.D. Reeve, D.J. Rhodes and P. Turchin, Scramble competition in the south-
ern pine beetle, Dendroctonus frontalis, Ecological Entomology, 23: 433–
443 (1998).
R.S. Rempel and C.K. Kaufman, Spatial modeling of harvest constraints on
wood supply versus wildlife habitat objectives, Environmental Manage-
ment, 32: 646–659 (2003). doi: 10.1007/s00267-003-0056-8.
Eric Renshaw, Modelling biological populations in space and time, Cambridge
Studies in Mathematical Biology 11, Cambridge University Press, Cam-
bridge (1995).
 Bibliography 465

W.E. Ricker, Stock and recruitment, J. Fisheries Research Board Canada,

11: 559–623 (1954).
Alex H. Ross, William S. C. Gurney, Michael R. Heath, Steven J. Hay, and
Eric W. Henderson, A strategic simulation model of a fjord ecosystem,
Limnology and Oceanography, 38: 128–153 (1993).
W.M. Roth, Unspecified things, signs, and ’natural objects’: towards a phe-
nomenological hermeneutic of graphing, In S. B. Berenson, K. R. Daw-
son, M. Blanton, W. N. Coulombe, J. Kolb, K. Norwood, and L. Stiff
(Eds.), Proceedings of the Twentieth Annual Meeting of the North Amer-
ican Chapter of the International Group for the Psychology of Mathe-
matics Education (Vol. I), ERIC Clearinghouse for Science, Mathemat-
ics, and Environmental Education, Columbus, OH, pp. 291–297, (1998).
M.B. Schaefer, Some aspects of the dynamics of populations important to the
management of commercial marine fisheries, Bulletin of the Inter-Amer.
Trop. Tuna Comm. I: 25–56 (1954).
Russell J. Schmidt, Sally J. Holbrook, and Craig W. Osenberg, Quantifying
the effects of multiple processes on local abundance: a cohort approach
for open populations, Ecology Letters, 2: 294–303 (1999).
M.D. Scott, S.J. Chivers, R.J. Olson, P.C. Fiedler, K. Holland, Pelagic preda-
tor associations: tuna and dolphins in the eastern tropical Pacific Ocean,
Marine Ecology Progress Series 458: 283–302 (2012).
Lee A. Segel, Modeling dynamic phenomena in molecular and cellular biology,
Cambridge University Press, Cambridge (1984).
Arthur Sherman and Richard Bertram, Integrative modeling of the pancre-
atic β-cell, in Wiley Interscience Encyclopedia of Genetics, Genomics,
Proteomics, and Bioinformatics, Part 3 Proteomics, M. Dunn, ed., Sec-
tion 3.8 Systems Biology, R. L. Winslow, ed., John Wiley and Sons, Ltd.
(2005). DOI: 10.1002/047001153X.g308213.
E. Smith, S. Haarer, J. Confrey, Seeking diversity in mathematics education:
mathematical modeling in the practice of biologists and mathematicians,
Science and Education, 6(5): 441–472 (1997).
E.M. Stock, K.E. Emmert, Deterministic discrete-time epidemic models with
applications to amphibians, masters thesis, Tarleton State University
(2006).

S.H. Strogatz, Exploring complex networks, Nature, 410: 268–276 (2001).

David J.T. Sumpter and Madeleine Beekman, From nonlinearity to optimal-
ity: pheromone foraging by ants, Animal Behaviour, 66: 273–280 (2003).
466 Bibliography

D.J.T. Sumpter and S.C. Pratt, A modelling framework for understanding

social insect foraging, Behav. Ecol. Sociobiol., 53: 131–144 (2003).
D’Arcy Wentworth Thompson, On growth and form, Vol. I, 2nd ed., Cam-
bridge University Press, Cambridge (1942).
A.N. Tihonov, On the dependence of the solutions of differential equations
on a small parameter, Mat. Sbornik NS, 22: 193–204 (1948).
David Tilman, Competition and biodiversity in spatially structured habitats,
Ecology, 75: 2–16, (1994).
David Tilman, Robert M. May, Clarence M. Lehman, and Martin A. Nowak,
Habitat destruction and the extinction debt, Nature, 371: 65–66 (1994).
Pauline van den Driessche and James Watmough, A simple SIS epidemic
model with a backward bifurcation, Journal of Mathematical Biology,
40: 525–540 (2000).
E. van der Meijden, M. J. Crawley, and R. M. Nisbet, The dynamics of a
herbivore-plant interaction, in Insect Populations: in Theory and Prac-
tice (J. P. Dempster and I. F. G. McLean, eds.), Chapman and Hall,
London (1998).
B. van der Pol and J. van der Mark, The heartbeat considered as a relaxation
oscillation, and an electrical model of the heart, Phil. Mag., 6: 763–775
(1928).
P.F. Verhulst, Notice sur la loi que la population suit dans son accroissement,
Corr. Math. et Phys., 10: 113–121 (1838).
P.F. Verhulst, Récherches mathématiques sur la loi d’accroissement de la
population, Mem. Acad. Roy. Brussels, 18: 1–38 (1845).
P.R. Wade, G.W. Watters, T. Gerrodette, and S.R. Reilly, Depletion of spot-
ted and spinner dolphins in the eastern tropical Pacific: modeling hy-
potheses for their lack of recovery, Marine Ecology Progress Series, 343:
1–14 (2007).
Wilhelm Weinberg, Uber den Nachweis der Vererbung beim Men-
schen, Jahreshefte des Vereins f ur vaterl andische Naturkunde in
W urttemberg, 64: 368–382 (1908).
 Mathematics
Advances in Applied Mathematics

FOR
DYNAMICAL SYSTEMS
BIOLOGICAL MODELING
DYNAMICAL SYSTEMS
FOR BIOLOGICAL
MODELING
Dynamical Systems for Biological Modeling: An Introduction pro-
vides both biology and mathematics students with the understanding
and techniques necessary to undertake basic modeling of biological
systems. It achieves this through the development and analysis of
dynamical systems. AN INTRODUCTION
The approach emphasizes qualitative ideas rather than explicit com-
putations. Some technical details are necessary, but a qualitative ap-
proach emphasizing ideas is essential for understanding. The model-
ing approach helps students focus on essentials rather than extensive
mathematical details, which is helpful for students whose primary in-
terests are in sciences other than mathematics.
The book discusses a variety of biological modeling topics, including
population biology, epidemiology, immunology, intraspecies competi-
tion, harvesting, predator–prey systems, structured populations, and
more.
The authors also include examples of problems with solutions and
some exercises that follow the examples quite closely. In addition,
problems are included that go beyond the examples, both in math-
ematical analysis and in the development of mathematical models for
Brauer • Kribs Fred Brauer
biological problems, in order to encourage deeper understanding and
an eagerness to use mathematics in learning about biology. Christopher Kribs

C664X

w w w. c rc p r e s s . c o m

C664X_cover.indd 1 11/9/15 3:23 PM